init

.gitignore

@@ -0,0 +1,8 @@
+.env
+.git
+.svn
+.idea
+.vscode
+*.log
+gitmod.sh
+*_test.go

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2018 茂名聚合科技有限公司
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,25 @@
+<h1>
+<a href="https://www.dtapp.net/">Golang MongoDB CRUD</a>
+</h1>
+
+📦 Golang MongoDB CRUD 组件
+
+[comment]: <> (go)
+[![godoc](https://pkg.go.dev/badge/github.com/dtapps/gomongo?status.svg)](https://pkg.go.dev/github.com/dtapps/gomongo)
+[![goproxy.cn](https://goproxy.cn/stats/github.com/dtapps/gomongo/badges/download-count.svg)](https://goproxy.cn/stats/github.com/dtapps/gomongo)
+[![goreportcard.com](https://goreportcard.com/badge/github.com/dtapps/gomongo)](https://goreportcard.com/report/github.com/dtapps/gomongo)
+[![deps.dev](https://img.shields.io/badge/deps-go-red.svg)](https://deps.dev/go/github.com%2Fdtapps%2Fgomongo)
+
+#### 安装使用
+
+```go
+go get -v -u github.com/dtapps/gomongo
+```
+
+#### 导入
+
+```go
+import (
+    "github.com/dtapps/gomongo"
+)
+```

app.go

@@ -0,0 +1,49 @@
+package gomongo
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"go.mongodb.org/mongo-driver/mongo"
+	"go.mongodb.org/mongo-driver/mongo/options"
+)
+
+const Version = "1.0.0"
+
+type app struct {
+	db             *mongo.Client // 驱动
+	Dns            string        // 连接地址
+	databaseName   string        // 库名
+	collectionName string        // 表名
+}
+
+// NewApp 实例化并链接数据库
+func NewApp(dns string) *app {
+	// 连接到MongoDB
+	db, err := mongo.Connect(context.TODO(), options.Client().ApplyURI(dns))
+	if err != nil {
+		panic(fmt.Sprintf("数据库【mongo】连接失败：%v", err))
+	}
+
+	// 检查连接
+	err = db.Ping(context.TODO(), nil)
+	if err != nil {
+		panic(fmt.Sprintf("数据库【mongo】连接服务器失败：%v", err))
+	}
+
+	return &app{db: db, Dns: dns}
+}
+
+// NewDb 实例化并传入链接
+func NewDb(db *mongo.Client) *app {
+	return &app{db: db}
+}
+
+// Close 关闭
+func (app *app) Close() {
+	err := app.db.Disconnect(context.TODO())
+	if err != nil {
+		panic(errors.New(fmt.Sprintf("数据库【mongo】关闭失败：%v", err)))
+	}
+	return
+}

curd.go

@@ -0,0 +1,190 @@
+package gomongo
+
+import (
+	"context"
+	"go.mongodb.org/mongo-driver/bson"
+	"go.mongodb.org/mongo-driver/mongo"
+	"go.mongodb.org/mongo-driver/mongo/options"
+	"reflect"
+)
+
+// Database 设置库名
+func (app *app) Database(database string) *app {
+	app.databaseName = database
+	return app
+}
+
+// Collection 设置表名
+func (app *app) Collection(collection string) *app {
+	app.collectionName = collection
+	return app
+}
+
+// Model 传入模型自动获取库名和表名
+func (app *app) Model(value interface{}) *app {
+	// https://studygolang.com/articles/896
+	val := reflect.ValueOf(value)
+	if methodValue := val.MethodByName("Database"); methodValue.IsValid() {
+		app.databaseName = methodValue.Call(nil)[0].String()
+	}
+	if methodValue := val.MethodByName("TableName"); methodValue.IsValid() {
+		app.collectionName = methodValue.Call(nil)[0].String()
+	}
+	return app
+}
+
+func (app *app) Session() (session mongo.Session, err error) {
+	session, err = app.db.StartSession()
+	return
+}
+
+// InsertOne 插入单个文档
+func (app *app) InsertOne(value interface{}) (result *mongo.InsertOneResult, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err = collection.InsertOne(context.TODO(), value)
+	return result, err
+}
+
+// InsertMany 插入多个文档
+func (app *app) InsertMany(values []interface{}) (result *mongo.InsertManyResult, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err = collection.InsertMany(context.TODO(), values)
+	return result, err
+}
+
+// Delete 删除文档
+func (app *app) Delete(filter interface{}) (int64, error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	count, err := collection.DeleteOne(context.TODO(), filter, nil)
+	return count.DeletedCount, err
+}
+
+// DeleteMany 删除多个文档
+func (app *app) DeleteMany(key string, value interface{}) (int64, error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	filter := bson.D{{key, value}}
+	count, err := collection.DeleteMany(context.TODO(), filter)
+	return count.DeletedCount, err
+}
+
+// UpdateOne 更新单个文档
+// 修改字段的值($set)
+// 字段增加值 inc($inc)
+// 从数组中增加一个元素 push($push)
+// 从数组中删除一个元素 pull($pull)
+func (app *app) UpdateOne(filter, update interface{}) (int64, error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err := collection.UpdateOne(context.TODO(), filter, update)
+	return result.UpsertedCount, err
+}
+
+// UpdateMany 更新多个文档
+// 修改字段的值($set)
+// 字段增加值 inc($inc)
+// 从数组中增加一个元素 push($push)
+// 从数组中删除一个元素 pull($pull)
+func (app *app) UpdateMany(filter, update interface{}) (int64, error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err := collection.UpdateMany(context.TODO(), filter, update)
+	return result.UpsertedCount, err
+}
+
+// Find 查询
+func (app *app) Find(filter interface{}, opts ...*options.FindOptions) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err = collection.Find(context.TODO(), filter, opts...)
+	return result, err
+}
+
+// FindOne 查询单个文档
+func (app *app) FindOne(filter interface{}) (result *mongo.SingleResult) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result = collection.FindOne(context.TODO(), filter)
+	return result
+}
+
+// FindMany 查询多个文档
+func (app *app) FindMany(filter interface{}) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err = collection.Find(context.TODO(), filter)
+	return result, err
+}
+
+// FindManyByFilters 多条件查询
+func (app *app) FindManyByFilters(filter interface{}) (result *mongo.Cursor, err error) {
+	collection, err := app.db.Database(app.databaseName).Collection(app.collectionName).Clone()
+	result, err = collection.Find(context.TODO(), bson.M{"$and": filter})
+	return result, err
+}
+
+// FindManyByFiltersSort 多条件查询支持排序
+func (app *app) FindManyByFiltersSort(filter interface{}, Sort interface{}) (result *mongo.Cursor, err error) {
+	collection, err := app.db.Database(app.databaseName).Collection(app.collectionName).Clone()
+	findOptions := options.Find()
+	findOptions.SetSort(Sort)
+	result, err = collection.Find(context.TODO(), filter, findOptions)
+	return result, err
+}
+
+// FindCollection 查询集合文档
+func (app *app) FindCollection(Limit int64) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	findOptions := options.Find()
+	findOptions.SetLimit(Limit)
+	result, err = collection.Find(context.TODO(), bson.D{{}}, findOptions)
+	return result, err
+}
+
+// FindCollectionSort 查询集合文档支持排序
+func (app *app) FindCollectionSort(Sort interface{}, Limit int64) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	findOptions := options.Find()
+	findOptions.SetSort(Sort)
+	findOptions.SetLimit(Limit)
+	result, err = collection.Find(context.TODO(), bson.D{{}}, findOptions)
+	return result, err
+}
+
+// FindManyCollectionSort 查询集合文档支持排序支持条件
+func (app *app) FindManyCollectionSort(filter interface{}, Sort interface{}) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	findOptions := options.Find()
+	findOptions.SetSort(Sort)
+	result, err = collection.Find(context.TODO(), filter, findOptions)
+	return result, err
+}
+
+// CollectionCount 查询集合里有多少数据
+func (app *app) CollectionCount() (name string, size int64) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	name = collection.Name()
+	size, _ = collection.EstimatedDocumentCount(context.TODO())
+	return name, size
+}
+
+// CollectionDocuments 按选项查询集合
+// Skip 跳过
+// Limit 读取数量
+// sort 1 ，-1 . 1 为升序 ， -1 为降序
+func (app *app) CollectionDocuments(Skip, Limit int64, sort int, key string, value interface{}) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	SORT := bson.D{{"_id", sort}}
+	filter := bson.D{{key, value}}
+	findOptions := options.Find().SetSort(SORT).SetLimit(Limit).SetSkip(Skip)
+	result, err = collection.Find(context.Background(), filter, findOptions)
+	return result, err
+}
+
+// AggregateByFiltersSort 统计分析
+func (app *app) AggregateByFiltersSort(pipeline interface{}, opts ...*options.AggregateOptions) (result *mongo.Cursor, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	result, err = collection.Aggregate(context.TODO(), pipeline, opts...)
+	return result, err
+}
+
+// CountDocumentsByFilters 统计数量
+func (app *app) CountDocumentsByFilters(filter interface{}) (count int64, err error) {
+	collection := app.db.Database(app.databaseName).Collection(app.collectionName)
+	count, err = collection.CountDocuments(context.TODO(), filter)
+	return count, err
+}

go.mod

@@ -0,0 +1,19 @@
+module github.com/dtapps/gomongo
+
+go 1.18
+
+require go.mongodb.org/mongo-driver v1.9.0
+
+require (
+	github.com/go-stack/stack v1.8.1 // indirect
+	github.com/golang/snappy v0.0.4 // indirect
+	github.com/klauspost/compress v1.15.2 // indirect
+	github.com/pkg/errors v0.9.1 // indirect
+	github.com/xdg-go/pbkdf2 v1.0.0 // indirect
+	github.com/xdg-go/scram v1.1.1 // indirect
+	github.com/xdg-go/stringprep v1.0.3 // indirect
+	github.com/youmark/pkcs8 v0.0.0-20201027041543-1326539a0a0a // indirect
+	golang.org/x/crypto v0.0.0-20220427172511-eb4f295cb31f // indirect
+	golang.org/x/sync v0.0.0-20210220032951-036812b2e83c // indirect
+	golang.org/x/text v0.3.7 // indirect
+)

go.sum

@@ -0,0 +1,67 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/go-stack/stack v1.8.0/go.mod h1:v0f6uXyyMGvRgIKkXu+yp6POWl0qKG85gN/melR3HDY=
+github.com/go-stack/stack v1.8.1 h1:ntEHSVwIt7PNXNpgPmVfMrNhLtgjlmnZha2kOpuRiDw=
+github.com/go-stack/stack v1.8.1/go.mod h1:dcoOX6HbPZSZptuspn9bctJ+N/CnF5gGygcUP3XYfe4=
+github.com/golang/snappy v0.0.1/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
+github.com/golang/snappy v0.0.4 h1:yAGX7huGHXlcLOEtBnF4w7FQwA26wojNCwOYAEhLjQM=
+github.com/golang/snappy v0.0.4/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
+github.com/google/go-cmp v0.5.2 h1:X2ev0eStA3AbceY54o37/0PQ/UWqKEiiO2dKL5OPaFM=
+github.com/google/go-cmp v0.5.2/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
+github.com/klauspost/compress v1.13.6/go.mod h1:/3/Vjq9QcHkK5uEr5lBEmyoZ1iFhe47etQ6QUkpK6sk=
+github.com/klauspost/compress v1.15.2 h1:3WH+AG7s2+T8o3nrM/8u2rdqUEcQhmga7smjrT41nAw=
+github.com/klauspost/compress v1.15.2/go.mod h1:PhcZ0MbTNciWF3rruxRgKxI5NkcHHrHUDtV4Yw2GlzU=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/montanaflynn/stats v0.0.0-20171201202039-1bf9dbcd8cbe/go.mod h1:wL8QJuTMNUDYhXwkmfOly8iTdp5TEcJFWZD2D7SIkUc=
+github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
+github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/testify v1.6.1 h1:hDPOHmpOpP40lSULcqw7IrRb/u7w6RpDC9399XyoNd0=
+github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/tidwall/pretty v1.0.0 h1:HsD+QiTn7sK6flMKIvNmpqz1qrpP3Ps6jOKIKMooyg4=
+github.com/tidwall/pretty v1.0.0/go.mod h1:XNkn88O1ChpSDQmQeStsy+sBenx6DDtFZJxhVysOjyk=
+github.com/xdg-go/pbkdf2 v1.0.0 h1:Su7DPu48wXMwC3bs7MCNG+z4FhcyEuz5dlvchbq0B0c=
+github.com/xdg-go/pbkdf2 v1.0.0/go.mod h1:jrpuAogTd400dnrH08LKmI/xc1MbPOebTwRqcT5RDeI=
+github.com/xdg-go/scram v1.0.2/go.mod h1:1WAq6h33pAW+iRreB34OORO2Nf7qel3VV3fjBj+hCSs=
+github.com/xdg-go/scram v1.1.1 h1:VOMT+81stJgXW3CpHyqHN3AXDYIMsx56mEFrB37Mb/E=
+github.com/xdg-go/scram v1.1.1/go.mod h1:RaEWvsqvNKKvBPvcKeFjrG2cJqOkHTiyTpzz23ni57g=
+github.com/xdg-go/stringprep v1.0.2/go.mod h1:8F9zXuvzgwmyT5DUm4GUfZGDdT3W+LCvS6+da4O5kxM=
+github.com/xdg-go/stringprep v1.0.3 h1:kdwGpVNwPFtjs98xCGkHjQtGKh86rDcRZN17QEMCOIs=
+github.com/xdg-go/stringprep v1.0.3/go.mod h1:W3f5j4i+9rC0kuIEJL0ky1VpHXQU3ocBgklLGvcBnW8=
+github.com/youmark/pkcs8 v0.0.0-20181117223130-1be2e3e5546d/go.mod h1:rHwXgn7JulP+udvsHwJoVG1YGAP6VLg4y9I5dyZdqmA=
+github.com/youmark/pkcs8 v0.0.0-20201027041543-1326539a0a0a h1:fZHgsYlfvtyqToslyjUt3VOPF4J7aK/3MPcK7xp3PDk=
+github.com/youmark/pkcs8 v0.0.0-20201027041543-1326539a0a0a/go.mod h1:ul22v+Nro/R083muKhosV54bj5niojjWZvU8xrevuH4=
+go.mongodb.org/mongo-driver v1.9.0 h1:f3aLGJvQmBl8d9S40IL+jEyBC6hfLPbJjv9t5hEM9ck=
+go.mongodb.org/mongo-driver v1.9.0/go.mod h1:0sQWfOeY63QTntERDJJ/0SuKK0T1uVSgKCuAROlKEPY=
+golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
+golang.org/x/crypto v0.0.0-20200302210943-78000ba7a073/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
+golang.org/x/crypto v0.0.0-20201216223049-8b5274cf687f/go.mod h1:jdWPYTVW3xRLrWPugEBEK3UY2ZEsg3UU495nc5E+M+I=
+golang.org/x/crypto v0.0.0-20220427172511-eb4f295cb31f h1:OeJjE6G4dgCY4PIXvIRQbE8+RX+uXZyGhUy/ksMGJoc=
+golang.org/x/crypto v0.0.0-20220427172511-eb4f295cb31f/go.mod h1:IxCIyHEi3zRg3s0A5j5BB6A9Jmi73HwBIUl50j+osU4=
+golang.org/x/net v0.0.0-20190311183353-d8887717615a/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
+golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
+golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20210220032951-036812b2e83c h1:5KslGYwFpkhGh+Q16bwMP3cOontH8FOep7tGV86Y7SQ=
+golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
+golang.org/x/sys v0.0.0-20190412213103-97732733099d/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20191026070338-33540a1f6037/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/term v0.0.0-20201117132131-f5c789dd3221/go.mod h1:Nr5EML6q2oocZ2LXRh80K7BxOlk5/8JxuGnuhpl+muw=
+golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
+golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
+golang.org/x/text v0.3.7 h1:olpwvP2KacW1ZWvsR7uQhoyTYvKAupfQrRGBFM352Gk=
+golang.org/x/text v0.3.7/go.mod h1:u+2+/6zg+i71rQMx5EYifcz6MCKuco9NR6JIITiCfzQ=
+golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
+golang.org/x/tools v0.0.0-20190531172133-b3315ee88b7d/go.mod h1:/rFqwRUd4F7ZHNgwSSTFct+R/Kf4OFW1sUzUTQQTgfc=
+golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543 h1:E7g+9GITq07hpfrRu66IVDexMakfv52eLZ2CXBWiKr4=
+golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c h1:dUUwHk2QECo/6vqA44rthZ8ie2QXMNeKRTHCNY2nXvo=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

gomod.sh

@@ -0,0 +1,3 @@
+go get -u && go mod tidy
+go get -u all
+go mod vendor

vendor/github.com/go-stack/stack/LICENSE.md

@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 Chris Hines
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

vendor/github.com/go-stack/stack/README.md

@@ -0,0 +1,38 @@
+[![GoDoc](https://godoc.org/github.com/go-stack/stack?status.svg)](https://godoc.org/github.com/go-stack/stack)
+[![Go Report Card](https://goreportcard.com/badge/go-stack/stack)](https://goreportcard.com/report/go-stack/stack)
+[![TravisCI](https://travis-ci.org/go-stack/stack.svg?branch=master)](https://travis-ci.org/go-stack/stack)
+[![Coverage Status](https://coveralls.io/repos/github/go-stack/stack/badge.svg?branch=master)](https://coveralls.io/github/go-stack/stack?branch=master)
+
+# stack
+
+Package stack implements utilities to capture, manipulate, and format call
+stacks. It provides a simpler API than package runtime.
+
+The implementation takes care of the minutia and special cases of interpreting
+the program counter (pc) values returned by runtime.Callers.
+
+## Versioning
+
+Package stack publishes releases via [semver](http://semver.org/) compatible Git
+tags prefixed with a single 'v'. The master branch always contains the latest
+release. The develop branch contains unreleased commits.
+
+## Formatting
+
+Package stack's types implement fmt.Formatter, which provides a simple and
+flexible way to declaratively configure formatting when used with logging or
+error tracking packages.
+
+```go
+func DoTheThing() {
+    c := stack.Caller(0)
+    log.Print(c)          // "source.go:10"
+    log.Printf("%+v", c)  // "pkg/path/source.go:10"
+    log.Printf("%n", c)   // "DoTheThing"
+
+    s := stack.Trace().TrimRuntime()
+    log.Print(s)          // "[source.go:15 caller.go:42 main.go:14]"
+}
+```
+
+See the docs for all of the supported formatting options.

vendor/github.com/go-stack/stack/stack.go

@@ -0,0 +1,400 @@
+// +build go1.7
+
+// Package stack implements utilities to capture, manipulate, and format call
+// stacks. It provides a simpler API than package runtime.
+//
+// The implementation takes care of the minutia and special cases of
+// interpreting the program counter (pc) values returned by runtime.Callers.
+//
+// Package stack's types implement fmt.Formatter, which provides a simple and
+// flexible way to declaratively configure formatting when used with logging
+// or error tracking packages.
+package stack
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"io"
+	"runtime"
+	"strconv"
+	"strings"
+)
+
+// Call records a single function invocation from a goroutine stack.
+type Call struct {
+	frame runtime.Frame
+}
+
+// Caller returns a Call from the stack of the current goroutine. The argument
+// skip is the number of stack frames to ascend, with 0 identifying the
+// calling function.
+func Caller(skip int) Call {
+	// As of Go 1.9 we need room for up to three PC entries.
+	//
+	// 0. An entry for the stack frame prior to the target to check for
+	//    special handling needed if that prior entry is runtime.sigpanic.
+	// 1. A possible second entry to hold metadata about skipped inlined
+	//    functions. If inline functions were not skipped the target frame
+	//    PC will be here.
+	// 2. A third entry for the target frame PC when the second entry
+	//    is used for skipped inline functions.
+	var pcs [3]uintptr
+	n := runtime.Callers(skip+1, pcs[:])
+	frames := runtime.CallersFrames(pcs[:n])
+	frame, _ := frames.Next()
+	frame, _ = frames.Next()
+
+	return Call{
+		frame: frame,
+	}
+}
+
+// String implements fmt.Stinger. It is equivalent to fmt.Sprintf("%v", c).
+func (c Call) String() string {
+	return fmt.Sprint(c)
+}
+
+// MarshalText implements encoding.TextMarshaler. It formats the Call the same
+// as fmt.Sprintf("%v", c).
+func (c Call) MarshalText() ([]byte, error) {
+	if c.frame == (runtime.Frame{}) {
+		return nil, ErrNoFunc
+	}
+
+	buf := bytes.Buffer{}
+	fmt.Fprint(&buf, c)
+	return buf.Bytes(), nil
+}
+
+// ErrNoFunc means that the Call has a nil *runtime.Func. The most likely
+// cause is a Call with the zero value.
+var ErrNoFunc = errors.New("no call stack information")
+
+// Format implements fmt.Formatter with support for the following verbs.
+//
+//    %s    source file
+//    %d    line number
+//    %n    function name
+//    %k    last segment of the package path
+//    %v    equivalent to %s:%d
+//
+// It accepts the '+' and '#' flags for most of the verbs as follows.
+//
+//    %+s   path of source file relative to the compile time GOPATH,
+//          or the module path joined to the path of source file relative
+//          to module root
+//    %#s   full path of source file
+//    %+n   import path qualified function name
+//    %+k   full package path
+//    %+v   equivalent to %+s:%d
+//    %#v   equivalent to %#s:%d
+func (c Call) Format(s fmt.State, verb rune) {
+	if c.frame == (runtime.Frame{}) {
+		fmt.Fprintf(s, "%%!%c(NOFUNC)", verb)
+		return
+	}
+
+	switch verb {
+	case 's', 'v':
+		file := c.frame.File
+		switch {
+		case s.Flag('#'):
+			// done
+		case s.Flag('+'):
+			file = pkgFilePath(&c.frame)
+		default:
+			const sep = "/"
+			if i := strings.LastIndex(file, sep); i != -1 {
+				file = file[i+len(sep):]
+			}
+		}
+		io.WriteString(s, file)
+		if verb == 'v' {
+			buf := [7]byte{':'}
+			s.Write(strconv.AppendInt(buf[:1], int64(c.frame.Line), 10))
+		}
+
+	case 'd':
+		buf := [6]byte{}
+		s.Write(strconv.AppendInt(buf[:0], int64(c.frame.Line), 10))
+
+	case 'k':
+		name := c.frame.Function
+		const pathSep = "/"
+		start, end := 0, len(name)
+		if i := strings.LastIndex(name, pathSep); i != -1 {
+			start = i + len(pathSep)
+		}
+		const pkgSep = "."
+		if i := strings.Index(name[start:], pkgSep); i != -1 {
+			end = start + i
+		}
+		if s.Flag('+') {
+			start = 0
+		}
+		io.WriteString(s, name[start:end])
+
+	case 'n':
+		name := c.frame.Function
+		if !s.Flag('+') {
+			const pathSep = "/"
+			if i := strings.LastIndex(name, pathSep); i != -1 {
+				name = name[i+len(pathSep):]
+			}
+			const pkgSep = "."
+			if i := strings.Index(name, pkgSep); i != -1 {
+				name = name[i+len(pkgSep):]
+			}
+		}
+		io.WriteString(s, name)
+	}
+}
+
+// Frame returns the call frame infomation for the Call.
+func (c Call) Frame() runtime.Frame {
+	return c.frame
+}
+
+// PC returns the program counter for this call frame; multiple frames may
+// have the same PC value.
+//
+// Deprecated: Use Call.Frame instead.
+func (c Call) PC() uintptr {
+	return c.frame.PC
+}
+
+// CallStack records a sequence of function invocations from a goroutine
+// stack.
+type CallStack []Call
+
+// String implements fmt.Stinger. It is equivalent to fmt.Sprintf("%v", cs).
+func (cs CallStack) String() string {
+	return fmt.Sprint(cs)
+}
+
+var (
+	openBracketBytes  = []byte("[")
+	closeBracketBytes = []byte("]")
+	spaceBytes        = []byte(" ")
+)
+
+// MarshalText implements encoding.TextMarshaler. It formats the CallStack the
+// same as fmt.Sprintf("%v", cs).
+func (cs CallStack) MarshalText() ([]byte, error) {
+	buf := bytes.Buffer{}
+	buf.Write(openBracketBytes)
+	for i, pc := range cs {
+		if i > 0 {
+			buf.Write(spaceBytes)
+		}
+		fmt.Fprint(&buf, pc)
+	}
+	buf.Write(closeBracketBytes)
+	return buf.Bytes(), nil
+}
+
+// Format implements fmt.Formatter by printing the CallStack as square brackets
+// ([, ]) surrounding a space separated list of Calls each formatted with the
+// supplied verb and options.
+func (cs CallStack) Format(s fmt.State, verb rune) {
+	s.Write(openBracketBytes)
+	for i, pc := range cs {
+		if i > 0 {
+			s.Write(spaceBytes)
+		}
+		pc.Format(s, verb)
+	}
+	s.Write(closeBracketBytes)
+}
+
+// Trace returns a CallStack for the current goroutine with element 0
+// identifying the calling function.
+func Trace() CallStack {
+	var pcs [512]uintptr
+	n := runtime.Callers(1, pcs[:])
+
+	frames := runtime.CallersFrames(pcs[:n])
+	cs := make(CallStack, 0, n)
+
+	// Skip extra frame retrieved just to make sure the runtime.sigpanic
+	// special case is handled.
+	frame, more := frames.Next()
+
+	for more {
+		frame, more = frames.Next()
+		cs = append(cs, Call{frame: frame})
+	}
+
+	return cs
+}
+
+// TrimBelow returns a slice of the CallStack with all entries below c
+// removed.
+func (cs CallStack) TrimBelow(c Call) CallStack {
+	for len(cs) > 0 && cs[0] != c {
+		cs = cs[1:]
+	}
+	return cs
+}
+
+// TrimAbove returns a slice of the CallStack with all entries above c
+// removed.
+func (cs CallStack) TrimAbove(c Call) CallStack {
+	for len(cs) > 0 && cs[len(cs)-1] != c {
+		cs = cs[:len(cs)-1]
+	}
+	return cs
+}
+
+// pkgIndex returns the index that results in file[index:] being the path of
+// file relative to the compile time GOPATH, and file[:index] being the
+// $GOPATH/src/ portion of file. funcName must be the name of a function in
+// file as returned by runtime.Func.Name.
+func pkgIndex(file, funcName string) int {
+	// As of Go 1.6.2 there is no direct way to know the compile time GOPATH
+	// at runtime, but we can infer the number of path segments in the GOPATH.
+	// We note that runtime.Func.Name() returns the function name qualified by
+	// the import path, which does not include the GOPATH. Thus we can trim
+	// segments from the beginning of the file path until the number of path
+	// separators remaining is one more than the number of path separators in
+	// the function name. For example, given:
+	//
+	//    GOPATH     /home/user
+	//    file       /home/user/src/pkg/sub/file.go
+	//    fn.Name()  pkg/sub.Type.Method
+	//
+	// We want to produce:
+	//
+	//    file[:idx] == /home/user/src/
+	//    file[idx:] == pkg/sub/file.go
+	//
+	// From this we can easily see that fn.Name() has one less path separator
+	// than our desired result for file[idx:]. We count separators from the
+	// end of the file path until it finds two more than in the function name
+	// and then move one character forward to preserve the initial path
+	// segment without a leading separator.
+	const sep = "/"
+	i := len(file)
+	for n := strings.Count(funcName, sep) + 2; n > 0; n-- {
+		i = strings.LastIndex(file[:i], sep)
+		if i == -1 {
+			i = -len(sep)
+			break
+		}
+	}
+	// get back to 0 or trim the leading separator
+	return i + len(sep)
+}
+
+// pkgFilePath returns the frame's filepath relative to the compile-time GOPATH,
+// or its module path joined to its path relative to the module root.
+//
+// As of Go 1.11 there is no direct way to know the compile time GOPATH or
+// module paths at runtime, but we can piece together the desired information
+// from available information. We note that runtime.Frame.Function contains the
+// function name qualified by the package path, which includes the module path
+// but not the GOPATH. We can extract the package path from that and append the
+// last segments of the file path to arrive at the desired package qualified
+// file path. For example, given:
+//
+//    GOPATH          /home/user
+//    import path     pkg/sub
+//    frame.File      /home/user/src/pkg/sub/file.go
+//    frame.Function  pkg/sub.Type.Method
+//    Desired return  pkg/sub/file.go
+//
+// It appears that we simply need to trim ".Type.Method" from frame.Function and
+// append "/" + path.Base(file).
+//
+// But there are other wrinkles. Although it is idiomatic to do so, the internal
+// name of a package is not required to match the last segment of its import
+// path. In addition, the introduction of modules in Go 1.11 allows working
+// without a GOPATH. So we also must make these work right:
+//
+//    GOPATH          /home/user
+//    import path     pkg/go-sub
+//    package name    sub
+//    frame.File      /home/user/src/pkg/go-sub/file.go
+//    frame.Function  pkg/sub.Type.Method
+//    Desired return  pkg/go-sub/file.go
+//
+//    Module path     pkg/v2
+//    import path     pkg/v2/go-sub
+//    package name    sub
+//    frame.File      /home/user/cloned-pkg/go-sub/file.go
+//    frame.Function  pkg/v2/sub.Type.Method
+//    Desired return  pkg/v2/go-sub/file.go
+//
+// We can handle all of these situations by using the package path extracted
+// from frame.Function up to, but not including, the last segment as the prefix
+// and the last two segments of frame.File as the suffix of the returned path.
+// This preserves the existing behavior when working in a GOPATH without modules
+// and a semantically equivalent behavior when used in module aware project.
+func pkgFilePath(frame *runtime.Frame) string {
+	pre := pkgPrefix(frame.Function)
+	post := pathSuffix(frame.File)
+	if pre == "" {
+		return post
+	}
+	return pre + "/" + post
+}
+
+// pkgPrefix returns the import path of the function's package with the final
+// segment removed.
+func pkgPrefix(funcName string) string {
+	const pathSep = "/"
+	end := strings.LastIndex(funcName, pathSep)
+	if end == -1 {
+		return ""
+	}
+	return funcName[:end]
+}
+
+// pathSuffix returns the last two segments of path.
+func pathSuffix(path string) string {
+	const pathSep = "/"
+	lastSep := strings.LastIndex(path, pathSep)
+	if lastSep == -1 {
+		return path
+	}
+	return path[strings.LastIndex(path[:lastSep], pathSep)+1:]
+}
+
+var runtimePath string
+
+func init() {
+	var pcs [3]uintptr
+	runtime.Callers(0, pcs[:])
+	frames := runtime.CallersFrames(pcs[:])
+	frame, _ := frames.Next()
+	file := frame.File
+
+	idx := pkgIndex(frame.File, frame.Function)
+
+	runtimePath = file[:idx]
+	if runtime.GOOS == "windows" {
+		runtimePath = strings.ToLower(runtimePath)
+	}
+}
+
+func inGoroot(c Call) bool {
+	file := c.frame.File
+	if len(file) == 0 || file[0] == '?' {
+		return true
+	}
+	if runtime.GOOS == "windows" {
+		file = strings.ToLower(file)
+	}
+	return strings.HasPrefix(file, runtimePath) || strings.HasSuffix(file, "/_testmain.go")
+}
+
+// TrimRuntime returns a slice of the CallStack with the topmost entries from
+// the go runtime removed. It considers any calls originating from unknown
+// files, files under GOROOT, or _testmain.go as part of the runtime.
+func (cs CallStack) TrimRuntime() CallStack {
+	for len(cs) > 0 && inGoroot(cs[len(cs)-1]) {
+		cs = cs[:len(cs)-1]
+	}
+	return cs
+}

vendor/github.com/golang/snappy/.gitignore

@@ -0,0 +1,16 @@
+cmd/snappytool/snappytool
+testdata/bench
+
+# These explicitly listed benchmark data files are for an obsolete version of
+# snappy_test.go.
+testdata/alice29.txt
+testdata/asyoulik.txt
+testdata/fireworks.jpeg
+testdata/geo.protodata
+testdata/html
+testdata/html_x_4
+testdata/kppkn.gtb
+testdata/lcet10.txt
+testdata/paper-100k.pdf
+testdata/plrabn12.txt
+testdata/urls.10K

vendor/github.com/golang/snappy/AUTHORS

@@ -0,0 +1,18 @@
+# This is the official list of Snappy-Go authors for copyright purposes.
+# This file is distinct from the CONTRIBUTORS files.
+# See the latter for an explanation.
+
+# Names should be added to this file as
+#	Name or Organization <email address>
+# The email address is not required for organizations.
+
+# Please keep the list sorted.
+
+Amazon.com, Inc
+Damian Gryski <dgryski@gmail.com>
+Eric Buth <eric@topos.com>
+Google Inc.
+Jan Mercl <0xjnml@gmail.com>
+Klaus Post <klauspost@gmail.com>
+Rodolfo Carvalho <rhcarvalho@gmail.com>
+Sebastien Binet <seb.binet@gmail.com>

vendor/github.com/golang/snappy/CONTRIBUTORS

@@ -0,0 +1,41 @@
+# This is the official list of people who can contribute
+# (and typically have contributed) code to the Snappy-Go repository.
+# The AUTHORS file lists the copyright holders; this file
+# lists people.  For example, Google employees are listed here
+# but not in AUTHORS, because Google holds the copyright.
+#
+# The submission process automatically checks to make sure
+# that people submitting code are listed in this file (by email address).
+#
+# Names should be added to this file only after verifying that
+# the individual or the individual's organization has agreed to
+# the appropriate Contributor License Agreement, found here:
+#
+#     http://code.google.com/legal/individual-cla-v1.0.html
+#     http://code.google.com/legal/corporate-cla-v1.0.html
+#
+# The agreement for individuals can be filled out on the web.
+#
+# When adding J Random Contributor's name to this file,
+# either J's name or J's organization's name should be
+# added to the AUTHORS file, depending on whether the
+# individual or corporate CLA was used.
+
+# Names should be added to this file like so:
+#     Name <email address>
+
+# Please keep the list sorted.
+
+Alex Legg <alexlegg@google.com>
+Damian Gryski <dgryski@gmail.com>
+Eric Buth <eric@topos.com>
+Jan Mercl <0xjnml@gmail.com>
+Jonathan Swinney <jswinney@amazon.com>
+Kai Backman <kaib@golang.org>
+Klaus Post <klauspost@gmail.com>
+Marc-Antoine Ruel <maruel@chromium.org>
+Nigel Tao <nigeltao@golang.org>
+Rob Pike <r@golang.org>
+Rodolfo Carvalho <rhcarvalho@gmail.com>
+Russ Cox <rsc@golang.org>
+Sebastien Binet <seb.binet@gmail.com>

vendor/github.com/golang/snappy/LICENSE

@@ -0,0 +1,27 @@
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

vendor/github.com/golang/snappy/README

@@ -0,0 +1,107 @@
+The Snappy compression format in the Go programming language.
+
+To download and install from source:
+$ go get github.com/golang/snappy
+
+Unless otherwise noted, the Snappy-Go source files are distributed
+under the BSD-style license found in the LICENSE file.
+
+
+
+Benchmarks.
+
+The golang/snappy benchmarks include compressing (Z) and decompressing (U) ten
+or so files, the same set used by the C++ Snappy code (github.com/google/snappy
+and note the "google", not "golang"). On an "Intel(R) Core(TM) i7-3770 CPU @
+3.40GHz", Go's GOARCH=amd64 numbers as of 2016-05-29:
+
+"go test -test.bench=."
+
+_UFlat0-8         2.19GB/s ± 0%  html
+_UFlat1-8         1.41GB/s ± 0%  urls
+_UFlat2-8         23.5GB/s ± 2%  jpg
+_UFlat3-8         1.91GB/s ± 0%  jpg_200
+_UFlat4-8         14.0GB/s ± 1%  pdf
+_UFlat5-8         1.97GB/s ± 0%  html4
+_UFlat6-8          814MB/s ± 0%  txt1
+_UFlat7-8          785MB/s ± 0%  txt2
+_UFlat8-8          857MB/s ± 0%  txt3
+_UFlat9-8          719MB/s ± 1%  txt4
+_UFlat10-8        2.84GB/s ± 0%  pb
+_UFlat11-8        1.05GB/s ± 0%  gaviota
+
+_ZFlat0-8         1.04GB/s ± 0%  html
+_ZFlat1-8          534MB/s ± 0%  urls
+_ZFlat2-8         15.7GB/s ± 1%  jpg
+_ZFlat3-8          740MB/s ± 3%  jpg_200
+_ZFlat4-8         9.20GB/s ± 1%  pdf
+_ZFlat5-8          991MB/s ± 0%  html4
+_ZFlat6-8          379MB/s ± 0%  txt1
+_ZFlat7-8          352MB/s ± 0%  txt2
+_ZFlat8-8          396MB/s ± 1%  txt3
+_ZFlat9-8          327MB/s ± 1%  txt4
+_ZFlat10-8        1.33GB/s ± 1%  pb
+_ZFlat11-8         605MB/s ± 1%  gaviota
+
+
+
+"go test -test.bench=. -tags=noasm"
+
+_UFlat0-8          621MB/s ± 2%  html
+_UFlat1-8          494MB/s ± 1%  urls
+_UFlat2-8         23.2GB/s ± 1%  jpg
+_UFlat3-8         1.12GB/s ± 1%  jpg_200
+_UFlat4-8         4.35GB/s ± 1%  pdf
+_UFlat5-8          609MB/s ± 0%  html4
+_UFlat6-8          296MB/s ± 0%  txt1
+_UFlat7-8          288MB/s ± 0%  txt2
+_UFlat8-8          309MB/s ± 1%  txt3
+_UFlat9-8          280MB/s ± 1%  txt4
+_UFlat10-8         753MB/s ± 0%  pb
+_UFlat11-8         400MB/s ± 0%  gaviota
+
+_ZFlat0-8          409MB/s ± 1%  html
+_ZFlat1-8          250MB/s ± 1%  urls
+_ZFlat2-8         12.3GB/s ± 1%  jpg
+_ZFlat3-8          132MB/s ± 0%  jpg_200
+_ZFlat4-8         2.92GB/s ± 0%  pdf
+_ZFlat5-8          405MB/s ± 1%  html4
+_ZFlat6-8          179MB/s ± 1%  txt1
+_ZFlat7-8          170MB/s ± 1%  txt2
+_ZFlat8-8          189MB/s ± 1%  txt3
+_ZFlat9-8          164MB/s ± 1%  txt4
+_ZFlat10-8         479MB/s ± 1%  pb
+_ZFlat11-8         270MB/s ± 1%  gaviota
+
+
+
+For comparison (Go's encoded output is byte-for-byte identical to C++'s), here
+are the numbers from C++ Snappy's
+
+make CXXFLAGS="-O2 -DNDEBUG -g" clean snappy_unittest.log && cat snappy_unittest.log
+
+BM_UFlat/0     2.4GB/s  html
+BM_UFlat/1     1.4GB/s  urls
+BM_UFlat/2    21.8GB/s  jpg
+BM_UFlat/3     1.5GB/s  jpg_200
+BM_UFlat/4    13.3GB/s  pdf
+BM_UFlat/5     2.1GB/s  html4
+BM_UFlat/6     1.0GB/s  txt1
+BM_UFlat/7   959.4MB/s  txt2
+BM_UFlat/8     1.0GB/s  txt3
+BM_UFlat/9   864.5MB/s  txt4
+BM_UFlat/10    2.9GB/s  pb
+BM_UFlat/11    1.2GB/s  gaviota
+
+BM_ZFlat/0   944.3MB/s  html (22.31 %)
+BM_ZFlat/1   501.6MB/s  urls (47.78 %)
+BM_ZFlat/2    14.3GB/s  jpg (99.95 %)
+BM_ZFlat/3   538.3MB/s  jpg_200 (73.00 %)
+BM_ZFlat/4     8.3GB/s  pdf (83.30 %)
+BM_ZFlat/5   903.5MB/s  html4 (22.52 %)
+BM_ZFlat/6   336.0MB/s  txt1 (57.88 %)
+BM_ZFlat/7   312.3MB/s  txt2 (61.91 %)
+BM_ZFlat/8   353.1MB/s  txt3 (54.99 %)
+BM_ZFlat/9   289.9MB/s  txt4 (66.26 %)
+BM_ZFlat/10    1.2GB/s  pb (19.68 %)
+BM_ZFlat/11  527.4MB/s  gaviota (37.72 %)

vendor/github.com/golang/snappy/decode.go

@@ -0,0 +1,264 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt reports that the input is invalid.
+	ErrCorrupt = errors.New("snappy: corrupt input")
+	// ErrTooLarge reports that the uncompressed length is too large.
+	ErrTooLarge = errors.New("snappy: decoded block is too large")
+	// ErrUnsupported reports that the input isn't supported.
+	ErrUnsupported = errors.New("snappy: unsupported input")
+
+	errUnsupportedLiteralLength = errors.New("snappy: unsupported literal length")
+)
+
+// DecodedLen returns the length of the decoded block.
+func DecodedLen(src []byte) (int, error) {
+	v, _, err := decodedLen(src)
+	return v, err
+}
+
+// decodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func decodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n <= 0 || v > 0xffffffff {
+		return 0, 0, ErrCorrupt
+	}
+
+	const wordSize = 32 << (^uint(0) >> 32 & 1)
+	if wordSize == 32 && v > 0x7fffffff {
+		return 0, 0, ErrTooLarge
+	}
+	return int(v), n, nil
+}
+
+const (
+	decodeErrCodeCorrupt                  = 1
+	decodeErrCodeUnsupportedLiteralLength = 2
+)
+
+// Decode returns the decoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire decoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Decode handles the Snappy block format, not the Snappy stream format.
+func Decode(dst, src []byte) ([]byte, error) {
+	dLen, s, err := decodedLen(src)
+	if err != nil {
+		return nil, err
+	}
+	if dLen <= len(dst) {
+		dst = dst[:dLen]
+	} else {
+		dst = make([]byte, dLen)
+	}
+	switch decode(dst, src[s:]) {
+	case 0:
+		return dst, nil
+	case decodeErrCodeUnsupportedLiteralLength:
+		return nil, errUnsupportedLiteralLength
+	}
+	return nil, ErrCorrupt
+}
+
+// NewReader returns a new Reader that decompresses from r, using the framing
+// format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r:       r,
+		decoded: make([]byte, maxBlockSize),
+		buf:     make([]byte, maxEncodedLenOfMaxBlockSize+checksumSize),
+	}
+}
+
+// Reader is an io.Reader that can read Snappy-compressed bytes.
+//
+// Reader handles the Snappy stream format, not the Snappy block format.
+type Reader struct {
+	r       io.Reader
+	err     error
+	decoded []byte
+	buf     []byte
+	// decoded[i:j] contains decoded bytes that have not yet been passed on.
+	i, j       int
+	readHeader bool
+}
+
+// Reset discards any buffered data, resets all state, and switches the Snappy
+// reader to read from r. This permits reusing a Reader rather than allocating
+// a new one.
+func (r *Reader) Reset(reader io.Reader) {
+	r.r = reader
+	r.err = nil
+	r.i = 0
+	r.j = 0
+	r.readHeader = false
+}
+
+func (r *Reader) readFull(p []byte, allowEOF bool) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF || (r.err == io.EOF && !allowEOF) {
+			r.err = ErrCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+func (r *Reader) fill() error {
+	for r.i >= r.j {
+		if !r.readFull(r.buf[:4], true) {
+			return r.err
+		}
+		chunkType := r.buf[0]
+		if !r.readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			r.err = ErrUnsupported
+			return r.err
+		}
+
+		// The chunk types are specified at
+		// https://github.com/google/snappy/blob/master/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[checksumSize:]
+
+			n, err := DecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return r.err
+			}
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if _, err := Decode(r.decoded, buf); err != nil {
+				r.err = err
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeUncompressedData:
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:checksumSize]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - checksumSize
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.decoded[:n], false) {
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeStreamIdentifier:
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(magicBody) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.buf[:len(magicBody)], false) {
+				return r.err
+			}
+			for i := 0; i < len(magicBody); i++ {
+				if r.buf[i] != magicBody[i] {
+					r.err = ErrCorrupt
+					return r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			r.err = ErrUnsupported
+			return r.err
+		}
+		// Section 4.4 Padding (chunk type 0xfe).
+		// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+		if !r.readFull(r.buf[:chunkLen], false) {
+			return r.err
+		}
+	}
+
+	return nil
+}
+
+// Read satisfies the io.Reader interface.
+func (r *Reader) Read(p []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	n := copy(p, r.decoded[r.i:r.j])
+	r.i += n
+	return n, nil
+}
+
+// ReadByte satisfies the io.ByteReader interface.
+func (r *Reader) ReadByte() (byte, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	c := r.decoded[r.i]
+	r.i++
+	return c, nil
+}

vendor/github.com/golang/snappy/decode_amd64.s

@@ -0,0 +1,490 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+
+// The asm code generally follows the pure Go code in decode_other.go, except
+// where marked with a "!!!".
+
+// func decode(dst, src []byte) int
+//
+// All local variables fit into registers. The non-zero stack size is only to
+// spill registers and push args when issuing a CALL. The register allocation:
+//	- AX	scratch
+//	- BX	scratch
+//	- CX	length or x
+//	- DX	offset
+//	- SI	&src[s]
+//	- DI	&dst[d]
+//	+ R8	dst_base
+//	+ R9	dst_len
+//	+ R10	dst_base + dst_len
+//	+ R11	src_base
+//	+ R12	src_len
+//	+ R13	src_base + src_len
+//	- R14	used by doCopy
+//	- R15	used by doCopy
+//
+// The registers R8-R13 (marked with a "+") are set at the start of the
+// function, and after a CALL returns, and are not otherwise modified.
+//
+// The d variable is implicitly DI - R8,  and len(dst)-d is R10 - DI.
+// The s variable is implicitly SI - R11, and len(src)-s is R13 - SI.
+TEXT ·decode(SB), NOSPLIT, $48-56
+	// Initialize SI, DI and R8-R13.
+	MOVQ dst_base+0(FP), R8
+	MOVQ dst_len+8(FP), R9
+	MOVQ R8, DI
+	MOVQ R8, R10
+	ADDQ R9, R10
+	MOVQ src_base+24(FP), R11
+	MOVQ src_len+32(FP), R12
+	MOVQ R11, SI
+	MOVQ R11, R13
+	ADDQ R12, R13
+
+loop:
+	// for s < len(src)
+	CMPQ SI, R13
+	JEQ  end
+
+	// CX = uint32(src[s])
+	//
+	// switch src[s] & 0x03
+	MOVBLZX (SI), CX
+	MOVL    CX, BX
+	ANDL    $3, BX
+	CMPL    BX, $1
+	JAE     tagCopy
+
+	// ----------------------------------------
+	// The code below handles literal tags.
+
+	// case tagLiteral:
+	// x := uint32(src[s] >> 2)
+	// switch
+	SHRL $2, CX
+	CMPL CX, $60
+	JAE  tagLit60Plus
+
+	// case x < 60:
+	// s++
+	INCQ SI
+
+doLit:
+	// This is the end of the inner "switch", when we have a literal tag.
+	//
+	// We assume that CX == x and x fits in a uint32, where x is the variable
+	// used in the pure Go decode_other.go code.
+
+	// length = int(x) + 1
+	//
+	// Unlike the pure Go code, we don't need to check if length <= 0 because
+	// CX can hold 64 bits, so the increment cannot overflow.
+	INCQ CX
+
+	// Prepare to check if copying length bytes will run past the end of dst or
+	// src.
+	//
+	// AX = len(dst) - d
+	// BX = len(src) - s
+	MOVQ R10, AX
+	SUBQ DI, AX
+	MOVQ R13, BX
+	SUBQ SI, BX
+
+	// !!! Try a faster technique for short (16 or fewer bytes) copies.
+	//
+	// if length > 16 || len(dst)-d < 16 || len(src)-s < 16 {
+	//   goto callMemmove // Fall back on calling runtime·memmove.
+	// }
+	//
+	// The C++ snappy code calls this TryFastAppend. It also checks len(src)-s
+	// against 21 instead of 16, because it cannot assume that all of its input
+	// is contiguous in memory and so it needs to leave enough source bytes to
+	// read the next tag without refilling buffers, but Go's Decode assumes
+	// contiguousness (the src argument is a []byte).
+	CMPQ CX, $16
+	JGT  callMemmove
+	CMPQ AX, $16
+	JLT  callMemmove
+	CMPQ BX, $16
+	JLT  callMemmove
+
+	// !!! Implement the copy from src to dst as a 16-byte load and store.
+	// (Decode's documentation says that dst and src must not overlap.)
+	//
+	// This always copies 16 bytes, instead of only length bytes, but that's
+	// OK. If the input is a valid Snappy encoding then subsequent iterations
+	// will fix up the overrun. Otherwise, Decode returns a nil []byte (and a
+	// non-nil error), so the overrun will be ignored.
+	//
+	// Note that on amd64, it is legal and cheap to issue unaligned 8-byte or
+	// 16-byte loads and stores. This technique probably wouldn't be as
+	// effective on architectures that are fussier about alignment.
+	MOVOU 0(SI), X0
+	MOVOU X0, 0(DI)
+
+	// d += length
+	// s += length
+	ADDQ CX, DI
+	ADDQ CX, SI
+	JMP  loop
+
+callMemmove:
+	// if length > len(dst)-d || length > len(src)-s { etc }
+	CMPQ CX, AX
+	JGT  errCorrupt
+	CMPQ CX, BX
+	JGT  errCorrupt
+
+	// copy(dst[d:], src[s:s+length])
+	//
+	// This means calling runtime·memmove(&dst[d], &src[s], length), so we push
+	// DI, SI and CX as arguments. Coincidentally, we also need to spill those
+	// three registers to the stack, to save local variables across the CALL.
+	MOVQ DI, 0(SP)
+	MOVQ SI, 8(SP)
+	MOVQ CX, 16(SP)
+	MOVQ DI, 24(SP)
+	MOVQ SI, 32(SP)
+	MOVQ CX, 40(SP)
+	CALL runtime·memmove(SB)
+
+	// Restore local variables: unspill registers from the stack and
+	// re-calculate R8-R13.
+	MOVQ 24(SP), DI
+	MOVQ 32(SP), SI
+	MOVQ 40(SP), CX
+	MOVQ dst_base+0(FP), R8
+	MOVQ dst_len+8(FP), R9
+	MOVQ R8, R10
+	ADDQ R9, R10
+	MOVQ src_base+24(FP), R11
+	MOVQ src_len+32(FP), R12
+	MOVQ R11, R13
+	ADDQ R12, R13
+
+	// d += length
+	// s += length
+	ADDQ CX, DI
+	ADDQ CX, SI
+	JMP  loop
+
+tagLit60Plus:
+	// !!! This fragment does the
+	//
+	// s += x - 58; if uint(s) > uint(len(src)) { etc }
+	//
+	// checks. In the asm version, we code it once instead of once per switch case.
+	ADDQ CX, SI
+	SUBQ $58, SI
+	MOVQ SI, BX
+	SUBQ R11, BX
+	CMPQ BX, R12
+	JA   errCorrupt
+
+	// case x == 60:
+	CMPL CX, $61
+	JEQ  tagLit61
+	JA   tagLit62Plus
+
+	// x = uint32(src[s-1])
+	MOVBLZX -1(SI), CX
+	JMP     doLit
+
+tagLit61:
+	// case x == 61:
+	// x = uint32(src[s-2]) | uint32(src[s-1])<<8
+	MOVWLZX -2(SI), CX
+	JMP     doLit
+
+tagLit62Plus:
+	CMPL CX, $62
+	JA   tagLit63
+
+	// case x == 62:
+	// x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+	MOVWLZX -3(SI), CX
+	MOVBLZX -1(SI), BX
+	SHLL    $16, BX
+	ORL     BX, CX
+	JMP     doLit
+
+tagLit63:
+	// case x == 63:
+	// x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+	MOVL -4(SI), CX
+	JMP  doLit
+
+// The code above handles literal tags.
+// ----------------------------------------
+// The code below handles copy tags.
+
+tagCopy4:
+	// case tagCopy4:
+	// s += 5
+	ADDQ $5, SI
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVQ SI, BX
+	SUBQ R11, BX
+	CMPQ BX, R12
+	JA   errCorrupt
+
+	// length = 1 + int(src[s-5])>>2
+	SHRQ $2, CX
+	INCQ CX
+
+	// offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24)
+	MOVLQZX -4(SI), DX
+	JMP     doCopy
+
+tagCopy2:
+	// case tagCopy2:
+	// s += 3
+	ADDQ $3, SI
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVQ SI, BX
+	SUBQ R11, BX
+	CMPQ BX, R12
+	JA   errCorrupt
+
+	// length = 1 + int(src[s-3])>>2
+	SHRQ $2, CX
+	INCQ CX
+
+	// offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8)
+	MOVWQZX -2(SI), DX
+	JMP     doCopy
+
+tagCopy:
+	// We have a copy tag. We assume that:
+	//	- BX == src[s] & 0x03
+	//	- CX == src[s]
+	CMPQ BX, $2
+	JEQ  tagCopy2
+	JA   tagCopy4
+
+	// case tagCopy1:
+	// s += 2
+	ADDQ $2, SI
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVQ SI, BX
+	SUBQ R11, BX
+	CMPQ BX, R12
+	JA   errCorrupt
+
+	// offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1]))
+	MOVQ    CX, DX
+	ANDQ    $0xe0, DX
+	SHLQ    $3, DX
+	MOVBQZX -1(SI), BX
+	ORQ     BX, DX
+
+	// length = 4 + int(src[s-2])>>2&0x7
+	SHRQ $2, CX
+	ANDQ $7, CX
+	ADDQ $4, CX
+
+doCopy:
+	// This is the end of the outer "switch", when we have a copy tag.
+	//
+	// We assume that:
+	//	- CX == length && CX > 0
+	//	- DX == offset
+
+	// if offset <= 0 { etc }
+	CMPQ DX, $0
+	JLE  errCorrupt
+
+	// if d < offset { etc }
+	MOVQ DI, BX
+	SUBQ R8, BX
+	CMPQ BX, DX
+	JLT  errCorrupt
+
+	// if length > len(dst)-d { etc }
+	MOVQ R10, BX
+	SUBQ DI, BX
+	CMPQ CX, BX
+	JGT  errCorrupt
+
+	// forwardCopy(dst[d:d+length], dst[d-offset:]); d += length
+	//
+	// Set:
+	//	- R14 = len(dst)-d
+	//	- R15 = &dst[d-offset]
+	MOVQ R10, R14
+	SUBQ DI, R14
+	MOVQ DI, R15
+	SUBQ DX, R15
+
+	// !!! Try a faster technique for short (16 or fewer bytes) forward copies.
+	//
+	// First, try using two 8-byte load/stores, similar to the doLit technique
+	// above. Even if dst[d:d+length] and dst[d-offset:] can overlap, this is
+	// still OK if offset >= 8. Note that this has to be two 8-byte load/stores
+	// and not one 16-byte load/store, and the first store has to be before the
+	// second load, due to the overlap if offset is in the range [8, 16).
+	//
+	// if length > 16 || offset < 8 || len(dst)-d < 16 {
+	//   goto slowForwardCopy
+	// }
+	// copy 16 bytes
+	// d += length
+	CMPQ CX, $16
+	JGT  slowForwardCopy
+	CMPQ DX, $8
+	JLT  slowForwardCopy
+	CMPQ R14, $16
+	JLT  slowForwardCopy
+	MOVQ 0(R15), AX
+	MOVQ AX, 0(DI)
+	MOVQ 8(R15), BX
+	MOVQ BX, 8(DI)
+	ADDQ CX, DI
+	JMP  loop
+
+slowForwardCopy:
+	// !!! If the forward copy is longer than 16 bytes, or if offset < 8, we
+	// can still try 8-byte load stores, provided we can overrun up to 10 extra
+	// bytes. As above, the overrun will be fixed up by subsequent iterations
+	// of the outermost loop.
+	//
+	// The C++ snappy code calls this technique IncrementalCopyFastPath. Its
+	// commentary says:
+	//
+	// ----
+	//
+	// The main part of this loop is a simple copy of eight bytes at a time
+	// until we've copied (at least) the requested amount of bytes.  However,
+	// if d and d-offset are less than eight bytes apart (indicating a
+	// repeating pattern of length < 8), we first need to expand the pattern in
+	// order to get the correct results. For instance, if the buffer looks like
+	// this, with the eight-byte <d-offset> and <d> patterns marked as
+	// intervals:
+	//
+	//    abxxxxxxxxxxxx
+	//    [------]           d-offset
+	//      [------]         d
+	//
+	// a single eight-byte copy from <d-offset> to <d> will repeat the pattern
+	// once, after which we can move <d> two bytes without moving <d-offset>:
+	//
+	//    ababxxxxxxxxxx
+	//    [------]           d-offset
+	//        [------]       d
+	//
+	// and repeat the exercise until the two no longer overlap.
+	//
+	// This allows us to do very well in the special case of one single byte
+	// repeated many times, without taking a big hit for more general cases.
+	//
+	// The worst case of extra writing past the end of the match occurs when
+	// offset == 1 and length == 1; the last copy will read from byte positions
+	// [0..7] and write to [4..11], whereas it was only supposed to write to
+	// position 1. Thus, ten excess bytes.
+	//
+	// ----
+	//
+	// That "10 byte overrun" worst case is confirmed by Go's
+	// TestSlowForwardCopyOverrun, which also tests the fixUpSlowForwardCopy
+	// and finishSlowForwardCopy algorithm.
+	//
+	// if length > len(dst)-d-10 {
+	//   goto verySlowForwardCopy
+	// }
+	SUBQ $10, R14
+	CMPQ CX, R14
+	JGT  verySlowForwardCopy
+
+makeOffsetAtLeast8:
+	// !!! As above, expand the pattern so that offset >= 8 and we can use
+	// 8-byte load/stores.
+	//
+	// for offset < 8 {
+	//   copy 8 bytes from dst[d-offset:] to dst[d:]
+	//   length -= offset
+	//   d      += offset
+	//   offset += offset
+	//   // The two previous lines together means that d-offset, and therefore
+	//   // R15, is unchanged.
+	// }
+	CMPQ DX, $8
+	JGE  fixUpSlowForwardCopy
+	MOVQ (R15), BX
+	MOVQ BX, (DI)
+	SUBQ DX, CX
+	ADDQ DX, DI
+	ADDQ DX, DX
+	JMP  makeOffsetAtLeast8
+
+fixUpSlowForwardCopy:
+	// !!! Add length (which might be negative now) to d (implied by DI being
+	// &dst[d]) so that d ends up at the right place when we jump back to the
+	// top of the loop. Before we do that, though, we save DI to AX so that, if
+	// length is positive, copying the remaining length bytes will write to the
+	// right place.
+	MOVQ DI, AX
+	ADDQ CX, DI
+
+finishSlowForwardCopy:
+	// !!! Repeat 8-byte load/stores until length <= 0. Ending with a negative
+	// length means that we overrun, but as above, that will be fixed up by
+	// subsequent iterations of the outermost loop.
+	CMPQ CX, $0
+	JLE  loop
+	MOVQ (R15), BX
+	MOVQ BX, (AX)
+	ADDQ $8, R15
+	ADDQ $8, AX
+	SUBQ $8, CX
+	JMP  finishSlowForwardCopy
+
+verySlowForwardCopy:
+	// verySlowForwardCopy is a simple implementation of forward copy. In C
+	// parlance, this is a do/while loop instead of a while loop, since we know
+	// that length > 0. In Go syntax:
+	//
+	// for {
+	//   dst[d] = dst[d - offset]
+	//   d++
+	//   length--
+	//   if length == 0 {
+	//     break
+	//   }
+	// }
+	MOVB (R15), BX
+	MOVB BX, (DI)
+	INCQ R15
+	INCQ DI
+	DECQ CX
+	JNZ  verySlowForwardCopy
+	JMP  loop
+
+// The code above handles copy tags.
+// ----------------------------------------
+
+end:
+	// This is the end of the "for s < len(src)".
+	//
+	// if d != len(dst) { etc }
+	CMPQ DI, R10
+	JNE  errCorrupt
+
+	// return 0
+	MOVQ $0, ret+48(FP)
+	RET
+
+errCorrupt:
+	// return decodeErrCodeCorrupt
+	MOVQ $1, ret+48(FP)
+	RET

vendor/github.com/golang/snappy/decode_arm64.s

@@ -0,0 +1,494 @@
+// Copyright 2020 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+
+// The asm code generally follows the pure Go code in decode_other.go, except
+// where marked with a "!!!".
+
+// func decode(dst, src []byte) int
+//
+// All local variables fit into registers. The non-zero stack size is only to
+// spill registers and push args when issuing a CALL. The register allocation:
+//	- R2	scratch
+//	- R3	scratch
+//	- R4	length or x
+//	- R5	offset
+//	- R6	&src[s]
+//	- R7	&dst[d]
+//	+ R8	dst_base
+//	+ R9	dst_len
+//	+ R10	dst_base + dst_len
+//	+ R11	src_base
+//	+ R12	src_len
+//	+ R13	src_base + src_len
+//	- R14	used by doCopy
+//	- R15	used by doCopy
+//
+// The registers R8-R13 (marked with a "+") are set at the start of the
+// function, and after a CALL returns, and are not otherwise modified.
+//
+// The d variable is implicitly R7 - R8,  and len(dst)-d is R10 - R7.
+// The s variable is implicitly R6 - R11, and len(src)-s is R13 - R6.
+TEXT ·decode(SB), NOSPLIT, $56-56
+	// Initialize R6, R7 and R8-R13.
+	MOVD dst_base+0(FP), R8
+	MOVD dst_len+8(FP), R9
+	MOVD R8, R7
+	MOVD R8, R10
+	ADD  R9, R10, R10
+	MOVD src_base+24(FP), R11
+	MOVD src_len+32(FP), R12
+	MOVD R11, R6
+	MOVD R11, R13
+	ADD  R12, R13, R13
+
+loop:
+	// for s < len(src)
+	CMP R13, R6
+	BEQ end
+
+	// R4 = uint32(src[s])
+	//
+	// switch src[s] & 0x03
+	MOVBU (R6), R4
+	MOVW  R4, R3
+	ANDW  $3, R3
+	MOVW  $1, R1
+	CMPW  R1, R3
+	BGE   tagCopy
+
+	// ----------------------------------------
+	// The code below handles literal tags.
+
+	// case tagLiteral:
+	// x := uint32(src[s] >> 2)
+	// switch
+	MOVW $60, R1
+	LSRW $2, R4, R4
+	CMPW R4, R1
+	BLS  tagLit60Plus
+
+	// case x < 60:
+	// s++
+	ADD $1, R6, R6
+
+doLit:
+	// This is the end of the inner "switch", when we have a literal tag.
+	//
+	// We assume that R4 == x and x fits in a uint32, where x is the variable
+	// used in the pure Go decode_other.go code.
+
+	// length = int(x) + 1
+	//
+	// Unlike the pure Go code, we don't need to check if length <= 0 because
+	// R4 can hold 64 bits, so the increment cannot overflow.
+	ADD $1, R4, R4
+
+	// Prepare to check if copying length bytes will run past the end of dst or
+	// src.
+	//
+	// R2 = len(dst) - d
+	// R3 = len(src) - s
+	MOVD R10, R2
+	SUB  R7, R2, R2
+	MOVD R13, R3
+	SUB  R6, R3, R3
+
+	// !!! Try a faster technique for short (16 or fewer bytes) copies.
+	//
+	// if length > 16 || len(dst)-d < 16 || len(src)-s < 16 {
+	//   goto callMemmove // Fall back on calling runtime·memmove.
+	// }
+	//
+	// The C++ snappy code calls this TryFastAppend. It also checks len(src)-s
+	// against 21 instead of 16, because it cannot assume that all of its input
+	// is contiguous in memory and so it needs to leave enough source bytes to
+	// read the next tag without refilling buffers, but Go's Decode assumes
+	// contiguousness (the src argument is a []byte).
+	CMP $16, R4
+	BGT callMemmove
+	CMP $16, R2
+	BLT callMemmove
+	CMP $16, R3
+	BLT callMemmove
+
+	// !!! Implement the copy from src to dst as a 16-byte load and store.
+	// (Decode's documentation says that dst and src must not overlap.)
+	//
+	// This always copies 16 bytes, instead of only length bytes, but that's
+	// OK. If the input is a valid Snappy encoding then subsequent iterations
+	// will fix up the overrun. Otherwise, Decode returns a nil []byte (and a
+	// non-nil error), so the overrun will be ignored.
+	//
+	// Note that on arm64, it is legal and cheap to issue unaligned 8-byte or
+	// 16-byte loads and stores. This technique probably wouldn't be as
+	// effective on architectures that are fussier about alignment.
+	LDP 0(R6), (R14, R15)
+	STP (R14, R15), 0(R7)
+
+	// d += length
+	// s += length
+	ADD R4, R7, R7
+	ADD R4, R6, R6
+	B   loop
+
+callMemmove:
+	// if length > len(dst)-d || length > len(src)-s { etc }
+	CMP R2, R4
+	BGT errCorrupt
+	CMP R3, R4
+	BGT errCorrupt
+
+	// copy(dst[d:], src[s:s+length])
+	//
+	// This means calling runtime·memmove(&dst[d], &src[s], length), so we push
+	// R7, R6 and R4 as arguments. Coincidentally, we also need to spill those
+	// three registers to the stack, to save local variables across the CALL.
+	MOVD R7, 8(RSP)
+	MOVD R6, 16(RSP)
+	MOVD R4, 24(RSP)
+	MOVD R7, 32(RSP)
+	MOVD R6, 40(RSP)
+	MOVD R4, 48(RSP)
+	CALL runtime·memmove(SB)
+
+	// Restore local variables: unspill registers from the stack and
+	// re-calculate R8-R13.
+	MOVD 32(RSP), R7
+	MOVD 40(RSP), R6
+	MOVD 48(RSP), R4
+	MOVD dst_base+0(FP), R8
+	MOVD dst_len+8(FP), R9
+	MOVD R8, R10
+	ADD  R9, R10, R10
+	MOVD src_base+24(FP), R11
+	MOVD src_len+32(FP), R12
+	MOVD R11, R13
+	ADD  R12, R13, R13
+
+	// d += length
+	// s += length
+	ADD R4, R7, R7
+	ADD R4, R6, R6
+	B   loop
+
+tagLit60Plus:
+	// !!! This fragment does the
+	//
+	// s += x - 58; if uint(s) > uint(len(src)) { etc }
+	//
+	// checks. In the asm version, we code it once instead of once per switch case.
+	ADD  R4, R6, R6
+	SUB  $58, R6, R6
+	MOVD R6, R3
+	SUB  R11, R3, R3
+	CMP  R12, R3
+	BGT  errCorrupt
+
+	// case x == 60:
+	MOVW $61, R1
+	CMPW R1, R4
+	BEQ  tagLit61
+	BGT  tagLit62Plus
+
+	// x = uint32(src[s-1])
+	MOVBU -1(R6), R4
+	B     doLit
+
+tagLit61:
+	// case x == 61:
+	// x = uint32(src[s-2]) | uint32(src[s-1])<<8
+	MOVHU -2(R6), R4
+	B     doLit
+
+tagLit62Plus:
+	CMPW $62, R4
+	BHI  tagLit63
+
+	// case x == 62:
+	// x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+	MOVHU -3(R6), R4
+	MOVBU -1(R6), R3
+	ORR   R3<<16, R4
+	B     doLit
+
+tagLit63:
+	// case x == 63:
+	// x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+	MOVWU -4(R6), R4
+	B     doLit
+
+	// The code above handles literal tags.
+	// ----------------------------------------
+	// The code below handles copy tags.
+
+tagCopy4:
+	// case tagCopy4:
+	// s += 5
+	ADD $5, R6, R6
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVD R6, R3
+	SUB  R11, R3, R3
+	CMP  R12, R3
+	BGT  errCorrupt
+
+	// length = 1 + int(src[s-5])>>2
+	MOVD $1, R1
+	ADD  R4>>2, R1, R4
+
+	// offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24)
+	MOVWU -4(R6), R5
+	B     doCopy
+
+tagCopy2:
+	// case tagCopy2:
+	// s += 3
+	ADD $3, R6, R6
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVD R6, R3
+	SUB  R11, R3, R3
+	CMP  R12, R3
+	BGT  errCorrupt
+
+	// length = 1 + int(src[s-3])>>2
+	MOVD $1, R1
+	ADD  R4>>2, R1, R4
+
+	// offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8)
+	MOVHU -2(R6), R5
+	B     doCopy
+
+tagCopy:
+	// We have a copy tag. We assume that:
+	//	- R3 == src[s] & 0x03
+	//	- R4 == src[s]
+	CMP $2, R3
+	BEQ tagCopy2
+	BGT tagCopy4
+
+	// case tagCopy1:
+	// s += 2
+	ADD $2, R6, R6
+
+	// if uint(s) > uint(len(src)) { etc }
+	MOVD R6, R3
+	SUB  R11, R3, R3
+	CMP  R12, R3
+	BGT  errCorrupt
+
+	// offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1]))
+	MOVD  R4, R5
+	AND   $0xe0, R5
+	MOVBU -1(R6), R3
+	ORR   R5<<3, R3, R5
+
+	// length = 4 + int(src[s-2])>>2&0x7
+	MOVD $7, R1
+	AND  R4>>2, R1, R4
+	ADD  $4, R4, R4
+
+doCopy:
+	// This is the end of the outer "switch", when we have a copy tag.
+	//
+	// We assume that:
+	//	- R4 == length && R4 > 0
+	//	- R5 == offset
+
+	// if offset <= 0 { etc }
+	MOVD $0, R1
+	CMP  R1, R5
+	BLE  errCorrupt
+
+	// if d < offset { etc }
+	MOVD R7, R3
+	SUB  R8, R3, R3
+	CMP  R5, R3
+	BLT  errCorrupt
+
+	// if length > len(dst)-d { etc }
+	MOVD R10, R3
+	SUB  R7, R3, R3
+	CMP  R3, R4
+	BGT  errCorrupt
+
+	// forwardCopy(dst[d:d+length], dst[d-offset:]); d += length
+	//
+	// Set:
+	//	- R14 = len(dst)-d
+	//	- R15 = &dst[d-offset]
+	MOVD R10, R14
+	SUB  R7, R14, R14
+	MOVD R7, R15
+	SUB  R5, R15, R15
+
+	// !!! Try a faster technique for short (16 or fewer bytes) forward copies.
+	//
+	// First, try using two 8-byte load/stores, similar to the doLit technique
+	// above. Even if dst[d:d+length] and dst[d-offset:] can overlap, this is
+	// still OK if offset >= 8. Note that this has to be two 8-byte load/stores
+	// and not one 16-byte load/store, and the first store has to be before the
+	// second load, due to the overlap if offset is in the range [8, 16).
+	//
+	// if length > 16 || offset < 8 || len(dst)-d < 16 {
+	//   goto slowForwardCopy
+	// }
+	// copy 16 bytes
+	// d += length
+	CMP  $16, R4
+	BGT  slowForwardCopy
+	CMP  $8, R5
+	BLT  slowForwardCopy
+	CMP  $16, R14
+	BLT  slowForwardCopy
+	MOVD 0(R15), R2
+	MOVD R2, 0(R7)
+	MOVD 8(R15), R3
+	MOVD R3, 8(R7)
+	ADD  R4, R7, R7
+	B    loop
+
+slowForwardCopy:
+	// !!! If the forward copy is longer than 16 bytes, or if offset < 8, we
+	// can still try 8-byte load stores, provided we can overrun up to 10 extra
+	// bytes. As above, the overrun will be fixed up by subsequent iterations
+	// of the outermost loop.
+	//
+	// The C++ snappy code calls this technique IncrementalCopyFastPath. Its
+	// commentary says:
+	//
+	// ----
+	//
+	// The main part of this loop is a simple copy of eight bytes at a time
+	// until we've copied (at least) the requested amount of bytes.  However,
+	// if d and d-offset are less than eight bytes apart (indicating a
+	// repeating pattern of length < 8), we first need to expand the pattern in
+	// order to get the correct results. For instance, if the buffer looks like
+	// this, with the eight-byte <d-offset> and <d> patterns marked as
+	// intervals:
+	//
+	//    abxxxxxxxxxxxx
+	//    [------]           d-offset
+	//      [------]         d
+	//
+	// a single eight-byte copy from <d-offset> to <d> will repeat the pattern
+	// once, after which we can move <d> two bytes without moving <d-offset>:
+	//
+	//    ababxxxxxxxxxx
+	//    [------]           d-offset
+	//        [------]       d
+	//
+	// and repeat the exercise until the two no longer overlap.
+	//
+	// This allows us to do very well in the special case of one single byte
+	// repeated many times, without taking a big hit for more general cases.
+	//
+	// The worst case of extra writing past the end of the match occurs when
+	// offset == 1 and length == 1; the last copy will read from byte positions
+	// [0..7] and write to [4..11], whereas it was only supposed to write to
+	// position 1. Thus, ten excess bytes.
+	//
+	// ----
+	//
+	// That "10 byte overrun" worst case is confirmed by Go's
+	// TestSlowForwardCopyOverrun, which also tests the fixUpSlowForwardCopy
+	// and finishSlowForwardCopy algorithm.
+	//
+	// if length > len(dst)-d-10 {
+	//   goto verySlowForwardCopy
+	// }
+	SUB $10, R14, R14
+	CMP R14, R4
+	BGT verySlowForwardCopy
+
+makeOffsetAtLeast8:
+	// !!! As above, expand the pattern so that offset >= 8 and we can use
+	// 8-byte load/stores.
+	//
+	// for offset < 8 {
+	//   copy 8 bytes from dst[d-offset:] to dst[d:]
+	//   length -= offset
+	//   d      += offset
+	//   offset += offset
+	//   // The two previous lines together means that d-offset, and therefore
+	//   // R15, is unchanged.
+	// }
+	CMP  $8, R5
+	BGE  fixUpSlowForwardCopy
+	MOVD (R15), R3
+	MOVD R3, (R7)
+	SUB  R5, R4, R4
+	ADD  R5, R7, R7
+	ADD  R5, R5, R5
+	B    makeOffsetAtLeast8
+
+fixUpSlowForwardCopy:
+	// !!! Add length (which might be negative now) to d (implied by R7 being
+	// &dst[d]) so that d ends up at the right place when we jump back to the
+	// top of the loop. Before we do that, though, we save R7 to R2 so that, if
+	// length is positive, copying the remaining length bytes will write to the
+	// right place.
+	MOVD R7, R2
+	ADD  R4, R7, R7
+
+finishSlowForwardCopy:
+	// !!! Repeat 8-byte load/stores until length <= 0. Ending with a negative
+	// length means that we overrun, but as above, that will be fixed up by
+	// subsequent iterations of the outermost loop.
+	MOVD $0, R1
+	CMP  R1, R4
+	BLE  loop
+	MOVD (R15), R3
+	MOVD R3, (R2)
+	ADD  $8, R15, R15
+	ADD  $8, R2, R2
+	SUB  $8, R4, R4
+	B    finishSlowForwardCopy
+
+verySlowForwardCopy:
+	// verySlowForwardCopy is a simple implementation of forward copy. In C
+	// parlance, this is a do/while loop instead of a while loop, since we know
+	// that length > 0. In Go syntax:
+	//
+	// for {
+	//   dst[d] = dst[d - offset]
+	//   d++
+	//   length--
+	//   if length == 0 {
+	//     break
+	//   }
+	// }
+	MOVB (R15), R3
+	MOVB R3, (R7)
+	ADD  $1, R15, R15
+	ADD  $1, R7, R7
+	SUB  $1, R4, R4
+	CBNZ R4, verySlowForwardCopy
+	B    loop
+
+	// The code above handles copy tags.
+	// ----------------------------------------
+
+end:
+	// This is the end of the "for s < len(src)".
+	//
+	// if d != len(dst) { etc }
+	CMP R10, R7
+	BNE errCorrupt
+
+	// return 0
+	MOVD $0, ret+48(FP)
+	RET
+
+errCorrupt:
+	// return decodeErrCodeCorrupt
+	MOVD $1, R2
+	MOVD R2, ret+48(FP)
+	RET

vendor/github.com/golang/snappy/decode_asm.go

@@ -0,0 +1,15 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+// +build amd64 arm64
+
+package snappy
+
+// decode has the same semantics as in decode_other.go.
+//
+//go:noescape
+func decode(dst, src []byte) int

vendor/github.com/golang/snappy/decode_other.go

@@ -0,0 +1,115 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !amd64,!arm64 appengine !gc noasm
+
+package snappy
+
+// decode writes the decoding of src to dst. It assumes that the varint-encoded
+// length of the decompressed bytes has already been read, and that len(dst)
+// equals that length.
+//
+// It returns 0 on success or a decodeErrCodeXxx error code on failure.
+func decode(dst, src []byte) int {
+	var d, s, offset, length int
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case tagLiteral:
+			x := uint32(src[s] >> 2)
+			switch {
+			case x < 60:
+				s++
+			case x == 60:
+				s += 2
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-1])
+			case x == 61:
+				s += 3
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-2]) | uint32(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+			}
+			length = int(x) + 1
+			if length <= 0 {
+				return decodeErrCodeUnsupportedLiteralLength
+			}
+			if length > len(dst)-d || length > len(src)-s {
+				return decodeErrCodeCorrupt
+			}
+			copy(dst[d:], src[s:s+length])
+			d += length
+			s += length
+			continue
+
+		case tagCopy1:
+			s += 2
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1]))
+
+		case tagCopy2:
+			s += 3
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8)
+
+		case tagCopy4:
+			s += 5
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-5])>>2
+			offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24)
+		}
+
+		if offset <= 0 || d < offset || length > len(dst)-d {
+			return decodeErrCodeCorrupt
+		}
+		// Copy from an earlier sub-slice of dst to a later sub-slice.
+		// If no overlap, use the built-in copy:
+		if offset >= length {
+			copy(dst[d:d+length], dst[d-offset:])
+			d += length
+			continue
+		}
+
+		// Unlike the built-in copy function, this byte-by-byte copy always runs
+		// forwards, even if the slices overlap. Conceptually, this is:
+		//
+		// d += forwardCopy(dst[d:d+length], dst[d-offset:])
+		//
+		// We align the slices into a and b and show the compiler they are the same size.
+		// This allows the loop to run without bounds checks.
+		a := dst[d : d+length]
+		b := dst[d-offset:]
+		b = b[:len(a)]
+		for i := range a {
+			a[i] = b[i]
+		}
+		d += length
+	}
+	if d != len(dst) {
+		return decodeErrCodeCorrupt
+	}
+	return 0
+}

vendor/github.com/golang/snappy/encode.go

@@ -0,0 +1,289 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// Encode returns the encoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire encoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Encode handles the Snappy block format, not the Snappy stream format.
+func Encode(dst, src []byte) []byte {
+	if n := MaxEncodedLen(len(src)); n < 0 {
+		panic(ErrTooLarge)
+	} else if len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	// The block starts with the varint-encoded length of the decompressed bytes.
+	d := binary.PutUvarint(dst, uint64(len(src)))
+
+	for len(src) > 0 {
+		p := src
+		src = nil
+		if len(p) > maxBlockSize {
+			p, src = p[:maxBlockSize], p[maxBlockSize:]
+		}
+		if len(p) < minNonLiteralBlockSize {
+			d += emitLiteral(dst[d:], p)
+		} else {
+			d += encodeBlock(dst[d:], p)
+		}
+	}
+	return dst[:d]
+}
+
+// inputMargin is the minimum number of extra input bytes to keep, inside
+// encodeBlock's inner loop. On some architectures, this margin lets us
+// implement a fast path for emitLiteral, where the copy of short (<= 16 byte)
+// literals can be implemented as a single load to and store from a 16-byte
+// register. That literal's actual length can be as short as 1 byte, so this
+// can copy up to 15 bytes too much, but that's OK as subsequent iterations of
+// the encoding loop will fix up the copy overrun, and this inputMargin ensures
+// that we don't overrun the dst and src buffers.
+const inputMargin = 16 - 1
+
+// minNonLiteralBlockSize is the minimum size of the input to encodeBlock that
+// could be encoded with a copy tag. This is the minimum with respect to the
+// algorithm used by encodeBlock, not a minimum enforced by the file format.
+//
+// The encoded output must start with at least a 1 byte literal, as there are
+// no previous bytes to copy. A minimal (1 byte) copy after that, generated
+// from an emitCopy call in encodeBlock's main loop, would require at least
+// another inputMargin bytes, for the reason above: we want any emitLiteral
+// calls inside encodeBlock's main loop to use the fast path if possible, which
+// requires being able to overrun by inputMargin bytes. Thus,
+// minNonLiteralBlockSize equals 1 + 1 + inputMargin.
+//
+// The C++ code doesn't use this exact threshold, but it could, as discussed at
+// https://groups.google.com/d/topic/snappy-compression/oGbhsdIJSJ8/discussion
+// The difference between Go (2+inputMargin) and C++ (inputMargin) is purely an
+// optimization. It should not affect the encoded form. This is tested by
+// TestSameEncodingAsCppShortCopies.
+const minNonLiteralBlockSize = 1 + 1 + inputMargin
+
+// MaxEncodedLen returns the maximum length of a snappy block, given its
+// uncompressed length.
+//
+// It will return a negative value if srcLen is too large to encode.
+func MaxEncodedLen(srcLen int) int {
+	n := uint64(srcLen)
+	if n > 0xffffffff {
+		return -1
+	}
+	// Compressed data can be defined as:
+	//    compressed := item* literal*
+	//    item       := literal* copy
+	//
+	// The trailing literal sequence has a space blowup of at most 62/60
+	// since a literal of length 60 needs one tag byte + one extra byte
+	// for length information.
+	//
+	// Item blowup is trickier to measure. Suppose the "copy" op copies
+	// 4 bytes of data. Because of a special check in the encoding code,
+	// we produce a 4-byte copy only if the offset is < 65536. Therefore
+	// the copy op takes 3 bytes to encode, and this type of item leads
+	// to at most the 62/60 blowup for representing literals.
+	//
+	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
+	// enough, it will take 5 bytes to encode the copy op. Therefore the
+	// worst case here is a one-byte literal followed by a five-byte copy.
+	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
+	//
+	// This last factor dominates the blowup, so the final estimate is:
+	n = 32 + n + n/6
+	if n > 0xffffffff {
+		return -1
+	}
+	return int(n)
+}
+
+var errClosed = errors.New("snappy: Writer is closed")
+
+// NewWriter returns a new Writer that compresses to w.
+//
+// The Writer returned does not buffer writes. There is no need to Flush or
+// Close such a Writer.
+//
+// Deprecated: the Writer returned is not suitable for many small writes, only
+// for few large writes. Use NewBufferedWriter instead, which is efficient
+// regardless of the frequency and shape of the writes, and remember to Close
+// that Writer when done.
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// NewBufferedWriter returns a new Writer that compresses to w, using the
+// framing format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+//
+// The Writer returned buffers writes. Users must call Close to guarantee all
+// data has been forwarded to the underlying io.Writer. They may also call
+// Flush zero or more times before calling Close.
+func NewBufferedWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		ibuf: make([]byte, 0, maxBlockSize),
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// Writer is an io.Writer that can write Snappy-compressed bytes.
+//
+// Writer handles the Snappy stream format, not the Snappy block format.
+type Writer struct {
+	w   io.Writer
+	err error
+
+	// ibuf is a buffer for the incoming (uncompressed) bytes.
+	//
+	// Its use is optional. For backwards compatibility, Writers created by the
+	// NewWriter function have ibuf == nil, do not buffer incoming bytes, and
+	// therefore do not need to be Flush'ed or Close'd.
+	ibuf []byte
+
+	// obuf is a buffer for the outgoing (compressed) bytes.
+	obuf []byte
+
+	// wroteStreamHeader is whether we have written the stream header.
+	wroteStreamHeader bool
+}
+
+// Reset discards the writer's state and switches the Snappy writer to write to
+// w. This permits reusing a Writer rather than allocating a new one.
+func (w *Writer) Reset(writer io.Writer) {
+	w.w = writer
+	w.err = nil
+	if w.ibuf != nil {
+		w.ibuf = w.ibuf[:0]
+	}
+	w.wroteStreamHeader = false
+}
+
+// Write satisfies the io.Writer interface.
+func (w *Writer) Write(p []byte) (nRet int, errRet error) {
+	if w.ibuf == nil {
+		// Do not buffer incoming bytes. This does not perform or compress well
+		// if the caller of Writer.Write writes many small slices. This
+		// behavior is therefore deprecated, but still supported for backwards
+		// compatibility with code that doesn't explicitly Flush or Close.
+		return w.write(p)
+	}
+
+	// The remainder of this method is based on bufio.Writer.Write from the
+	// standard library.
+
+	for len(p) > (cap(w.ibuf)-len(w.ibuf)) && w.err == nil {
+		var n int
+		if len(w.ibuf) == 0 {
+			// Large write, empty buffer.
+			// Write directly from p to avoid copy.
+			n, _ = w.write(p)
+		} else {
+			n = copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+			w.ibuf = w.ibuf[:len(w.ibuf)+n]
+			w.Flush()
+		}
+		nRet += n
+		p = p[n:]
+	}
+	if w.err != nil {
+		return nRet, w.err
+	}
+	n := copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+	w.ibuf = w.ibuf[:len(w.ibuf)+n]
+	nRet += n
+	return nRet, nil
+}
+
+func (w *Writer) write(p []byte) (nRet int, errRet error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	for len(p) > 0 {
+		obufStart := len(magicChunk)
+		if !w.wroteStreamHeader {
+			w.wroteStreamHeader = true
+			copy(w.obuf, magicChunk)
+			obufStart = 0
+		}
+
+		var uncompressed []byte
+		if len(p) > maxBlockSize {
+			uncompressed, p = p[:maxBlockSize], p[maxBlockSize:]
+		} else {
+			uncompressed, p = p, nil
+		}
+		checksum := crc(uncompressed)
+
+		// Compress the buffer, discarding the result if the improvement
+		// isn't at least 12.5%.
+		compressed := Encode(w.obuf[obufHeaderLen:], uncompressed)
+		chunkType := uint8(chunkTypeCompressedData)
+		chunkLen := 4 + len(compressed)
+		obufEnd := obufHeaderLen + len(compressed)
+		if len(compressed) >= len(uncompressed)-len(uncompressed)/8 {
+			chunkType = chunkTypeUncompressedData
+			chunkLen = 4 + len(uncompressed)
+			obufEnd = obufHeaderLen
+		}
+
+		// Fill in the per-chunk header that comes before the body.
+		w.obuf[len(magicChunk)+0] = chunkType
+		w.obuf[len(magicChunk)+1] = uint8(chunkLen >> 0)
+		w.obuf[len(magicChunk)+2] = uint8(chunkLen >> 8)
+		w.obuf[len(magicChunk)+3] = uint8(chunkLen >> 16)
+		w.obuf[len(magicChunk)+4] = uint8(checksum >> 0)
+		w.obuf[len(magicChunk)+5] = uint8(checksum >> 8)
+		w.obuf[len(magicChunk)+6] = uint8(checksum >> 16)
+		w.obuf[len(magicChunk)+7] = uint8(checksum >> 24)
+
+		if _, err := w.w.Write(w.obuf[obufStart:obufEnd]); err != nil {
+			w.err = err
+			return nRet, err
+		}
+		if chunkType == chunkTypeUncompressedData {
+			if _, err := w.w.Write(uncompressed); err != nil {
+				w.err = err
+				return nRet, err
+			}
+		}
+		nRet += len(uncompressed)
+	}
+	return nRet, nil
+}
+
+// Flush flushes the Writer to its underlying io.Writer.
+func (w *Writer) Flush() error {
+	if w.err != nil {
+		return w.err
+	}
+	if len(w.ibuf) == 0 {
+		return nil
+	}
+	w.write(w.ibuf)
+	w.ibuf = w.ibuf[:0]
+	return w.err
+}
+
+// Close calls Flush and then closes the Writer.
+func (w *Writer) Close() error {
+	w.Flush()
+	ret := w.err
+	if w.err == nil {
+		w.err = errClosed
+	}
+	return ret
+}

vendor/github.com/golang/snappy/encode_amd64.s

@@ -0,0 +1,730 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+
+// The XXX lines assemble on Go 1.4, 1.5 and 1.7, but not 1.6, due to a
+// Go toolchain regression. See https://github.com/golang/go/issues/15426 and
+// https://github.com/golang/snappy/issues/29
+//
+// As a workaround, the package was built with a known good assembler, and
+// those instructions were disassembled by "objdump -d" to yield the
+//	4e 0f b7 7c 5c 78       movzwq 0x78(%rsp,%r11,2),%r15
+// style comments, in AT&T asm syntax. Note that rsp here is a physical
+// register, not Go/asm's SP pseudo-register (see https://golang.org/doc/asm).
+// The instructions were then encoded as "BYTE $0x.." sequences, which assemble
+// fine on Go 1.6.
+
+// The asm code generally follows the pure Go code in encode_other.go, except
+// where marked with a "!!!".
+
+// ----------------------------------------------------------------------------
+
+// func emitLiteral(dst, lit []byte) int
+//
+// All local variables fit into registers. The register allocation:
+//	- AX	len(lit)
+//	- BX	n
+//	- DX	return value
+//	- DI	&dst[i]
+//	- R10	&lit[0]
+//
+// The 24 bytes of stack space is to call runtime·memmove.
+//
+// The unusual register allocation of local variables, such as R10 for the
+// source pointer, matches the allocation used at the call site in encodeBlock,
+// which makes it easier to manually inline this function.
+TEXT ·emitLiteral(SB), NOSPLIT, $24-56
+	MOVQ dst_base+0(FP), DI
+	MOVQ lit_base+24(FP), R10
+	MOVQ lit_len+32(FP), AX
+	MOVQ AX, DX
+	MOVL AX, BX
+	SUBL $1, BX
+
+	CMPL BX, $60
+	JLT  oneByte
+	CMPL BX, $256
+	JLT  twoBytes
+
+threeBytes:
+	MOVB $0xf4, 0(DI)
+	MOVW BX, 1(DI)
+	ADDQ $3, DI
+	ADDQ $3, DX
+	JMP  memmove
+
+twoBytes:
+	MOVB $0xf0, 0(DI)
+	MOVB BX, 1(DI)
+	ADDQ $2, DI
+	ADDQ $2, DX
+	JMP  memmove
+
+oneByte:
+	SHLB $2, BX
+	MOVB BX, 0(DI)
+	ADDQ $1, DI
+	ADDQ $1, DX
+
+memmove:
+	MOVQ DX, ret+48(FP)
+
+	// copy(dst[i:], lit)
+	//
+	// This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push
+	// DI, R10 and AX as arguments.
+	MOVQ DI, 0(SP)
+	MOVQ R10, 8(SP)
+	MOVQ AX, 16(SP)
+	CALL runtime·memmove(SB)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func emitCopy(dst []byte, offset, length int) int
+//
+// All local variables fit into registers. The register allocation:
+//	- AX	length
+//	- SI	&dst[0]
+//	- DI	&dst[i]
+//	- R11	offset
+//
+// The unusual register allocation of local variables, such as R11 for the
+// offset, matches the allocation used at the call site in encodeBlock, which
+// makes it easier to manually inline this function.
+TEXT ·emitCopy(SB), NOSPLIT, $0-48
+	MOVQ dst_base+0(FP), DI
+	MOVQ DI, SI
+	MOVQ offset+24(FP), R11
+	MOVQ length+32(FP), AX
+
+loop0:
+	// for length >= 68 { etc }
+	CMPL AX, $68
+	JLT  step1
+
+	// Emit a length 64 copy, encoded as 3 bytes.
+	MOVB $0xfe, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+	SUBL $64, AX
+	JMP  loop0
+
+step1:
+	// if length > 64 { etc }
+	CMPL AX, $64
+	JLE  step2
+
+	// Emit a length 60 copy, encoded as 3 bytes.
+	MOVB $0xee, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+	SUBL $60, AX
+
+step2:
+	// if length >= 12 || offset >= 2048 { goto step3 }
+	CMPL AX, $12
+	JGE  step3
+	CMPL R11, $2048
+	JGE  step3
+
+	// Emit the remaining copy, encoded as 2 bytes.
+	MOVB R11, 1(DI)
+	SHRL $8, R11
+	SHLB $5, R11
+	SUBB $4, AX
+	SHLB $2, AX
+	ORB  AX, R11
+	ORB  $1, R11
+	MOVB R11, 0(DI)
+	ADDQ $2, DI
+
+	// Return the number of bytes written.
+	SUBQ SI, DI
+	MOVQ DI, ret+40(FP)
+	RET
+
+step3:
+	// Emit the remaining copy, encoded as 3 bytes.
+	SUBL $1, AX
+	SHLB $2, AX
+	ORB  $2, AX
+	MOVB AX, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+
+	// Return the number of bytes written.
+	SUBQ SI, DI
+	MOVQ DI, ret+40(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func extendMatch(src []byte, i, j int) int
+//
+// All local variables fit into registers. The register allocation:
+//	- DX	&src[0]
+//	- SI	&src[j]
+//	- R13	&src[len(src) - 8]
+//	- R14	&src[len(src)]
+//	- R15	&src[i]
+//
+// The unusual register allocation of local variables, such as R15 for a source
+// pointer, matches the allocation used at the call site in encodeBlock, which
+// makes it easier to manually inline this function.
+TEXT ·extendMatch(SB), NOSPLIT, $0-48
+	MOVQ src_base+0(FP), DX
+	MOVQ src_len+8(FP), R14
+	MOVQ i+24(FP), R15
+	MOVQ j+32(FP), SI
+	ADDQ DX, R14
+	ADDQ DX, R15
+	ADDQ DX, SI
+	MOVQ R14, R13
+	SUBQ $8, R13
+
+cmp8:
+	// As long as we are 8 or more bytes before the end of src, we can load and
+	// compare 8 bytes at a time. If those 8 bytes are equal, repeat.
+	CMPQ SI, R13
+	JA   cmp1
+	MOVQ (R15), AX
+	MOVQ (SI), BX
+	CMPQ AX, BX
+	JNE  bsf
+	ADDQ $8, R15
+	ADDQ $8, SI
+	JMP  cmp8
+
+bsf:
+	// If those 8 bytes were not equal, XOR the two 8 byte values, and return
+	// the index of the first byte that differs. The BSF instruction finds the
+	// least significant 1 bit, the amd64 architecture is little-endian, and
+	// the shift by 3 converts a bit index to a byte index.
+	XORQ AX, BX
+	BSFQ BX, BX
+	SHRQ $3, BX
+	ADDQ BX, SI
+
+	// Convert from &src[ret] to ret.
+	SUBQ DX, SI
+	MOVQ SI, ret+40(FP)
+	RET
+
+cmp1:
+	// In src's tail, compare 1 byte at a time.
+	CMPQ SI, R14
+	JAE  extendMatchEnd
+	MOVB (R15), AX
+	MOVB (SI), BX
+	CMPB AX, BX
+	JNE  extendMatchEnd
+	ADDQ $1, R15
+	ADDQ $1, SI
+	JMP  cmp1
+
+extendMatchEnd:
+	// Convert from &src[ret] to ret.
+	SUBQ DX, SI
+	MOVQ SI, ret+40(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func encodeBlock(dst, src []byte) (d int)
+//
+// All local variables fit into registers, other than "var table". The register
+// allocation:
+//	- AX	.	.
+//	- BX	.	.
+//	- CX	56	shift (note that amd64 shifts by non-immediates must use CX).
+//	- DX	64	&src[0], tableSize
+//	- SI	72	&src[s]
+//	- DI	80	&dst[d]
+//	- R9	88	sLimit
+//	- R10	.	&src[nextEmit]
+//	- R11	96	prevHash, currHash, nextHash, offset
+//	- R12	104	&src[base], skip
+//	- R13	.	&src[nextS], &src[len(src) - 8]
+//	- R14	.	len(src), bytesBetweenHashLookups, &src[len(src)], x
+//	- R15	112	candidate
+//
+// The second column (56, 64, etc) is the stack offset to spill the registers
+// when calling other functions. We could pack this slightly tighter, but it's
+// simpler to have a dedicated spill map independent of the function called.
+//
+// "var table [maxTableSize]uint16" takes up 32768 bytes of stack space. An
+// extra 56 bytes, to call other functions, and an extra 64 bytes, to spill
+// local variables (registers) during calls gives 32768 + 56 + 64 = 32888.
+TEXT ·encodeBlock(SB), 0, $32888-56
+	MOVQ dst_base+0(FP), DI
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R14
+
+	// shift, tableSize := uint32(32-8), 1<<8
+	MOVQ $24, CX
+	MOVQ $256, DX
+
+calcShift:
+	// for ; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 {
+	//	shift--
+	// }
+	CMPQ DX, $16384
+	JGE  varTable
+	CMPQ DX, R14
+	JGE  varTable
+	SUBQ $1, CX
+	SHLQ $1, DX
+	JMP  calcShift
+
+varTable:
+	// var table [maxTableSize]uint16
+	//
+	// In the asm code, unlike the Go code, we can zero-initialize only the
+	// first tableSize elements. Each uint16 element is 2 bytes and each MOVOU
+	// writes 16 bytes, so we can do only tableSize/8 writes instead of the
+	// 2048 writes that would zero-initialize all of table's 32768 bytes.
+	SHRQ $3, DX
+	LEAQ table-32768(SP), BX
+	PXOR X0, X0
+
+memclr:
+	MOVOU X0, 0(BX)
+	ADDQ  $16, BX
+	SUBQ  $1, DX
+	JNZ   memclr
+
+	// !!! DX = &src[0]
+	MOVQ SI, DX
+
+	// sLimit := len(src) - inputMargin
+	MOVQ R14, R9
+	SUBQ $15, R9
+
+	// !!! Pre-emptively spill CX, DX and R9 to the stack. Their values don't
+	// change for the rest of the function.
+	MOVQ CX, 56(SP)
+	MOVQ DX, 64(SP)
+	MOVQ R9, 88(SP)
+
+	// nextEmit := 0
+	MOVQ DX, R10
+
+	// s := 1
+	ADDQ $1, SI
+
+	// nextHash := hash(load32(src, s), shift)
+	MOVL  0(SI), R11
+	IMULL $0x1e35a7bd, R11
+	SHRL  CX, R11
+
+outer:
+	// for { etc }
+
+	// skip := 32
+	MOVQ $32, R12
+
+	// nextS := s
+	MOVQ SI, R13
+
+	// candidate := 0
+	MOVQ $0, R15
+
+inner0:
+	// for { etc }
+
+	// s := nextS
+	MOVQ R13, SI
+
+	// bytesBetweenHashLookups := skip >> 5
+	MOVQ R12, R14
+	SHRQ $5, R14
+
+	// nextS = s + bytesBetweenHashLookups
+	ADDQ R14, R13
+
+	// skip += bytesBetweenHashLookups
+	ADDQ R14, R12
+
+	// if nextS > sLimit { goto emitRemainder }
+	MOVQ R13, AX
+	SUBQ DX, AX
+	CMPQ AX, R9
+	JA   emitRemainder
+
+	// candidate = int(table[nextHash])
+	// XXX: MOVWQZX table-32768(SP)(R11*2), R15
+	// XXX: 4e 0f b7 7c 5c 78       movzwq 0x78(%rsp,%r11,2),%r15
+	BYTE $0x4e
+	BYTE $0x0f
+	BYTE $0xb7
+	BYTE $0x7c
+	BYTE $0x5c
+	BYTE $0x78
+
+	// table[nextHash] = uint16(s)
+	MOVQ SI, AX
+	SUBQ DX, AX
+
+	// XXX: MOVW AX, table-32768(SP)(R11*2)
+	// XXX: 66 42 89 44 5c 78       mov    %ax,0x78(%rsp,%r11,2)
+	BYTE $0x66
+	BYTE $0x42
+	BYTE $0x89
+	BYTE $0x44
+	BYTE $0x5c
+	BYTE $0x78
+
+	// nextHash = hash(load32(src, nextS), shift)
+	MOVL  0(R13), R11
+	IMULL $0x1e35a7bd, R11
+	SHRL  CX, R11
+
+	// if load32(src, s) != load32(src, candidate) { continue } break
+	MOVL 0(SI), AX
+	MOVL (DX)(R15*1), BX
+	CMPL AX, BX
+	JNE  inner0
+
+fourByteMatch:
+	// As per the encode_other.go code:
+	//
+	// A 4-byte match has been found. We'll later see etc.
+
+	// !!! Jump to a fast path for short (<= 16 byte) literals. See the comment
+	// on inputMargin in encode.go.
+	MOVQ SI, AX
+	SUBQ R10, AX
+	CMPQ AX, $16
+	JLE  emitLiteralFastPath
+
+	// ----------------------------------------
+	// Begin inline of the emitLiteral call.
+	//
+	// d += emitLiteral(dst[d:], src[nextEmit:s])
+
+	MOVL AX, BX
+	SUBL $1, BX
+
+	CMPL BX, $60
+	JLT  inlineEmitLiteralOneByte
+	CMPL BX, $256
+	JLT  inlineEmitLiteralTwoBytes
+
+inlineEmitLiteralThreeBytes:
+	MOVB $0xf4, 0(DI)
+	MOVW BX, 1(DI)
+	ADDQ $3, DI
+	JMP  inlineEmitLiteralMemmove
+
+inlineEmitLiteralTwoBytes:
+	MOVB $0xf0, 0(DI)
+	MOVB BX, 1(DI)
+	ADDQ $2, DI
+	JMP  inlineEmitLiteralMemmove
+
+inlineEmitLiteralOneByte:
+	SHLB $2, BX
+	MOVB BX, 0(DI)
+	ADDQ $1, DI
+
+inlineEmitLiteralMemmove:
+	// Spill local variables (registers) onto the stack; call; unspill.
+	//
+	// copy(dst[i:], lit)
+	//
+	// This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push
+	// DI, R10 and AX as arguments.
+	MOVQ DI, 0(SP)
+	MOVQ R10, 8(SP)
+	MOVQ AX, 16(SP)
+	ADDQ AX, DI              // Finish the "d +=" part of "d += emitLiteral(etc)".
+	MOVQ SI, 72(SP)
+	MOVQ DI, 80(SP)
+	MOVQ R15, 112(SP)
+	CALL runtime·memmove(SB)
+	MOVQ 56(SP), CX
+	MOVQ 64(SP), DX
+	MOVQ 72(SP), SI
+	MOVQ 80(SP), DI
+	MOVQ 88(SP), R9
+	MOVQ 112(SP), R15
+	JMP  inner1
+
+inlineEmitLiteralEnd:
+	// End inline of the emitLiteral call.
+	// ----------------------------------------
+
+emitLiteralFastPath:
+	// !!! Emit the 1-byte encoding "uint8(len(lit)-1)<<2".
+	MOVB AX, BX
+	SUBB $1, BX
+	SHLB $2, BX
+	MOVB BX, (DI)
+	ADDQ $1, DI
+
+	// !!! Implement the copy from lit to dst as a 16-byte load and store.
+	// (Encode's documentation says that dst and src must not overlap.)
+	//
+	// This always copies 16 bytes, instead of only len(lit) bytes, but that's
+	// OK. Subsequent iterations will fix up the overrun.
+	//
+	// Note that on amd64, it is legal and cheap to issue unaligned 8-byte or
+	// 16-byte loads and stores. This technique probably wouldn't be as
+	// effective on architectures that are fussier about alignment.
+	MOVOU 0(R10), X0
+	MOVOU X0, 0(DI)
+	ADDQ  AX, DI
+
+inner1:
+	// for { etc }
+
+	// base := s
+	MOVQ SI, R12
+
+	// !!! offset := base - candidate
+	MOVQ R12, R11
+	SUBQ R15, R11
+	SUBQ DX, R11
+
+	// ----------------------------------------
+	// Begin inline of the extendMatch call.
+	//
+	// s = extendMatch(src, candidate+4, s+4)
+
+	// !!! R14 = &src[len(src)]
+	MOVQ src_len+32(FP), R14
+	ADDQ DX, R14
+
+	// !!! R13 = &src[len(src) - 8]
+	MOVQ R14, R13
+	SUBQ $8, R13
+
+	// !!! R15 = &src[candidate + 4]
+	ADDQ $4, R15
+	ADDQ DX, R15
+
+	// !!! s += 4
+	ADDQ $4, SI
+
+inlineExtendMatchCmp8:
+	// As long as we are 8 or more bytes before the end of src, we can load and
+	// compare 8 bytes at a time. If those 8 bytes are equal, repeat.
+	CMPQ SI, R13
+	JA   inlineExtendMatchCmp1
+	MOVQ (R15), AX
+	MOVQ (SI), BX
+	CMPQ AX, BX
+	JNE  inlineExtendMatchBSF
+	ADDQ $8, R15
+	ADDQ $8, SI
+	JMP  inlineExtendMatchCmp8
+
+inlineExtendMatchBSF:
+	// If those 8 bytes were not equal, XOR the two 8 byte values, and return
+	// the index of the first byte that differs. The BSF instruction finds the
+	// least significant 1 bit, the amd64 architecture is little-endian, and
+	// the shift by 3 converts a bit index to a byte index.
+	XORQ AX, BX
+	BSFQ BX, BX
+	SHRQ $3, BX
+	ADDQ BX, SI
+	JMP  inlineExtendMatchEnd
+
+inlineExtendMatchCmp1:
+	// In src's tail, compare 1 byte at a time.
+	CMPQ SI, R14
+	JAE  inlineExtendMatchEnd
+	MOVB (R15), AX
+	MOVB (SI), BX
+	CMPB AX, BX
+	JNE  inlineExtendMatchEnd
+	ADDQ $1, R15
+	ADDQ $1, SI
+	JMP  inlineExtendMatchCmp1
+
+inlineExtendMatchEnd:
+	// End inline of the extendMatch call.
+	// ----------------------------------------
+
+	// ----------------------------------------
+	// Begin inline of the emitCopy call.
+	//
+	// d += emitCopy(dst[d:], base-candidate, s-base)
+
+	// !!! length := s - base
+	MOVQ SI, AX
+	SUBQ R12, AX
+
+inlineEmitCopyLoop0:
+	// for length >= 68 { etc }
+	CMPL AX, $68
+	JLT  inlineEmitCopyStep1
+
+	// Emit a length 64 copy, encoded as 3 bytes.
+	MOVB $0xfe, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+	SUBL $64, AX
+	JMP  inlineEmitCopyLoop0
+
+inlineEmitCopyStep1:
+	// if length > 64 { etc }
+	CMPL AX, $64
+	JLE  inlineEmitCopyStep2
+
+	// Emit a length 60 copy, encoded as 3 bytes.
+	MOVB $0xee, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+	SUBL $60, AX
+
+inlineEmitCopyStep2:
+	// if length >= 12 || offset >= 2048 { goto inlineEmitCopyStep3 }
+	CMPL AX, $12
+	JGE  inlineEmitCopyStep3
+	CMPL R11, $2048
+	JGE  inlineEmitCopyStep3
+
+	// Emit the remaining copy, encoded as 2 bytes.
+	MOVB R11, 1(DI)
+	SHRL $8, R11
+	SHLB $5, R11
+	SUBB $4, AX
+	SHLB $2, AX
+	ORB  AX, R11
+	ORB  $1, R11
+	MOVB R11, 0(DI)
+	ADDQ $2, DI
+	JMP  inlineEmitCopyEnd
+
+inlineEmitCopyStep3:
+	// Emit the remaining copy, encoded as 3 bytes.
+	SUBL $1, AX
+	SHLB $2, AX
+	ORB  $2, AX
+	MOVB AX, 0(DI)
+	MOVW R11, 1(DI)
+	ADDQ $3, DI
+
+inlineEmitCopyEnd:
+	// End inline of the emitCopy call.
+	// ----------------------------------------
+
+	// nextEmit = s
+	MOVQ SI, R10
+
+	// if s >= sLimit { goto emitRemainder }
+	MOVQ SI, AX
+	SUBQ DX, AX
+	CMPQ AX, R9
+	JAE  emitRemainder
+
+	// As per the encode_other.go code:
+	//
+	// We could immediately etc.
+
+	// x := load64(src, s-1)
+	MOVQ -1(SI), R14
+
+	// prevHash := hash(uint32(x>>0), shift)
+	MOVL  R14, R11
+	IMULL $0x1e35a7bd, R11
+	SHRL  CX, R11
+
+	// table[prevHash] = uint16(s-1)
+	MOVQ SI, AX
+	SUBQ DX, AX
+	SUBQ $1, AX
+
+	// XXX: MOVW AX, table-32768(SP)(R11*2)
+	// XXX: 66 42 89 44 5c 78       mov    %ax,0x78(%rsp,%r11,2)
+	BYTE $0x66
+	BYTE $0x42
+	BYTE $0x89
+	BYTE $0x44
+	BYTE $0x5c
+	BYTE $0x78
+
+	// currHash := hash(uint32(x>>8), shift)
+	SHRQ  $8, R14
+	MOVL  R14, R11
+	IMULL $0x1e35a7bd, R11
+	SHRL  CX, R11
+
+	// candidate = int(table[currHash])
+	// XXX: MOVWQZX table-32768(SP)(R11*2), R15
+	// XXX: 4e 0f b7 7c 5c 78       movzwq 0x78(%rsp,%r11,2),%r15
+	BYTE $0x4e
+	BYTE $0x0f
+	BYTE $0xb7
+	BYTE $0x7c
+	BYTE $0x5c
+	BYTE $0x78
+
+	// table[currHash] = uint16(s)
+	ADDQ $1, AX
+
+	// XXX: MOVW AX, table-32768(SP)(R11*2)
+	// XXX: 66 42 89 44 5c 78       mov    %ax,0x78(%rsp,%r11,2)
+	BYTE $0x66
+	BYTE $0x42
+	BYTE $0x89
+	BYTE $0x44
+	BYTE $0x5c
+	BYTE $0x78
+
+	// if uint32(x>>8) == load32(src, candidate) { continue }
+	MOVL (DX)(R15*1), BX
+	CMPL R14, BX
+	JEQ  inner1
+
+	// nextHash = hash(uint32(x>>16), shift)
+	SHRQ  $8, R14
+	MOVL  R14, R11
+	IMULL $0x1e35a7bd, R11
+	SHRL  CX, R11
+
+	// s++
+	ADDQ $1, SI
+
+	// break out of the inner1 for loop, i.e. continue the outer loop.
+	JMP outer
+
+emitRemainder:
+	// if nextEmit < len(src) { etc }
+	MOVQ src_len+32(FP), AX
+	ADDQ DX, AX
+	CMPQ R10, AX
+	JEQ  encodeBlockEnd
+
+	// d += emitLiteral(dst[d:], src[nextEmit:])
+	//
+	// Push args.
+	MOVQ DI, 0(SP)
+	MOVQ $0, 8(SP)   // Unnecessary, as the callee ignores it, but conservative.
+	MOVQ $0, 16(SP)  // Unnecessary, as the callee ignores it, but conservative.
+	MOVQ R10, 24(SP)
+	SUBQ R10, AX
+	MOVQ AX, 32(SP)
+	MOVQ AX, 40(SP)  // Unnecessary, as the callee ignores it, but conservative.
+
+	// Spill local variables (registers) onto the stack; call; unspill.
+	MOVQ DI, 80(SP)
+	CALL ·emitLiteral(SB)
+	MOVQ 80(SP), DI
+
+	// Finish the "d +=" part of "d += emitLiteral(etc)".
+	ADDQ 48(SP), DI
+
+encodeBlockEnd:
+	MOVQ dst_base+0(FP), AX
+	SUBQ AX, DI
+	MOVQ DI, d+48(FP)
+	RET

vendor/github.com/golang/snappy/encode_arm64.s

@@ -0,0 +1,722 @@
+// Copyright 2020 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+
+// The asm code generally follows the pure Go code in encode_other.go, except
+// where marked with a "!!!".
+
+// ----------------------------------------------------------------------------
+
+// func emitLiteral(dst, lit []byte) int
+//
+// All local variables fit into registers. The register allocation:
+//	- R3	len(lit)
+//	- R4	n
+//	- R6	return value
+//	- R8	&dst[i]
+//	- R10	&lit[0]
+//
+// The 32 bytes of stack space is to call runtime·memmove.
+//
+// The unusual register allocation of local variables, such as R10 for the
+// source pointer, matches the allocation used at the call site in encodeBlock,
+// which makes it easier to manually inline this function.
+TEXT ·emitLiteral(SB), NOSPLIT, $32-56
+	MOVD dst_base+0(FP), R8
+	MOVD lit_base+24(FP), R10
+	MOVD lit_len+32(FP), R3
+	MOVD R3, R6
+	MOVW R3, R4
+	SUBW $1, R4, R4
+
+	CMPW $60, R4
+	BLT  oneByte
+	CMPW $256, R4
+	BLT  twoBytes
+
+threeBytes:
+	MOVD $0xf4, R2
+	MOVB R2, 0(R8)
+	MOVW R4, 1(R8)
+	ADD  $3, R8, R8
+	ADD  $3, R6, R6
+	B    memmove
+
+twoBytes:
+	MOVD $0xf0, R2
+	MOVB R2, 0(R8)
+	MOVB R4, 1(R8)
+	ADD  $2, R8, R8
+	ADD  $2, R6, R6
+	B    memmove
+
+oneByte:
+	LSLW $2, R4, R4
+	MOVB R4, 0(R8)
+	ADD  $1, R8, R8
+	ADD  $1, R6, R6
+
+memmove:
+	MOVD R6, ret+48(FP)
+
+	// copy(dst[i:], lit)
+	//
+	// This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push
+	// R8, R10 and R3 as arguments.
+	MOVD R8, 8(RSP)
+	MOVD R10, 16(RSP)
+	MOVD R3, 24(RSP)
+	CALL runtime·memmove(SB)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func emitCopy(dst []byte, offset, length int) int
+//
+// All local variables fit into registers. The register allocation:
+//	- R3	length
+//	- R7	&dst[0]
+//	- R8	&dst[i]
+//	- R11	offset
+//
+// The unusual register allocation of local variables, such as R11 for the
+// offset, matches the allocation used at the call site in encodeBlock, which
+// makes it easier to manually inline this function.
+TEXT ·emitCopy(SB), NOSPLIT, $0-48
+	MOVD dst_base+0(FP), R8
+	MOVD R8, R7
+	MOVD offset+24(FP), R11
+	MOVD length+32(FP), R3
+
+loop0:
+	// for length >= 68 { etc }
+	CMPW $68, R3
+	BLT  step1
+
+	// Emit a length 64 copy, encoded as 3 bytes.
+	MOVD $0xfe, R2
+	MOVB R2, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+	SUB  $64, R3, R3
+	B    loop0
+
+step1:
+	// if length > 64 { etc }
+	CMP $64, R3
+	BLE step2
+
+	// Emit a length 60 copy, encoded as 3 bytes.
+	MOVD $0xee, R2
+	MOVB R2, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+	SUB  $60, R3, R3
+
+step2:
+	// if length >= 12 || offset >= 2048 { goto step3 }
+	CMP  $12, R3
+	BGE  step3
+	CMPW $2048, R11
+	BGE  step3
+
+	// Emit the remaining copy, encoded as 2 bytes.
+	MOVB R11, 1(R8)
+	LSRW $3, R11, R11
+	AND  $0xe0, R11, R11
+	SUB  $4, R3, R3
+	LSLW $2, R3
+	AND  $0xff, R3, R3
+	ORRW R3, R11, R11
+	ORRW $1, R11, R11
+	MOVB R11, 0(R8)
+	ADD  $2, R8, R8
+
+	// Return the number of bytes written.
+	SUB  R7, R8, R8
+	MOVD R8, ret+40(FP)
+	RET
+
+step3:
+	// Emit the remaining copy, encoded as 3 bytes.
+	SUB  $1, R3, R3
+	AND  $0xff, R3, R3
+	LSLW $2, R3, R3
+	ORRW $2, R3, R3
+	MOVB R3, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+
+	// Return the number of bytes written.
+	SUB  R7, R8, R8
+	MOVD R8, ret+40(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func extendMatch(src []byte, i, j int) int
+//
+// All local variables fit into registers. The register allocation:
+//	- R6	&src[0]
+//	- R7	&src[j]
+//	- R13	&src[len(src) - 8]
+//	- R14	&src[len(src)]
+//	- R15	&src[i]
+//
+// The unusual register allocation of local variables, such as R15 for a source
+// pointer, matches the allocation used at the call site in encodeBlock, which
+// makes it easier to manually inline this function.
+TEXT ·extendMatch(SB), NOSPLIT, $0-48
+	MOVD src_base+0(FP), R6
+	MOVD src_len+8(FP), R14
+	MOVD i+24(FP), R15
+	MOVD j+32(FP), R7
+	ADD  R6, R14, R14
+	ADD  R6, R15, R15
+	ADD  R6, R7, R7
+	MOVD R14, R13
+	SUB  $8, R13, R13
+
+cmp8:
+	// As long as we are 8 or more bytes before the end of src, we can load and
+	// compare 8 bytes at a time. If those 8 bytes are equal, repeat.
+	CMP  R13, R7
+	BHI  cmp1
+	MOVD (R15), R3
+	MOVD (R7), R4
+	CMP  R4, R3
+	BNE  bsf
+	ADD  $8, R15, R15
+	ADD  $8, R7, R7
+	B    cmp8
+
+bsf:
+	// If those 8 bytes were not equal, XOR the two 8 byte values, and return
+	// the index of the first byte that differs.
+	// RBIT reverses the bit order, then CLZ counts the leading zeros, the
+	// combination of which finds the least significant bit which is set.
+	// The arm64 architecture is little-endian, and the shift by 3 converts
+	// a bit index to a byte index.
+	EOR  R3, R4, R4
+	RBIT R4, R4
+	CLZ  R4, R4
+	ADD  R4>>3, R7, R7
+
+	// Convert from &src[ret] to ret.
+	SUB  R6, R7, R7
+	MOVD R7, ret+40(FP)
+	RET
+
+cmp1:
+	// In src's tail, compare 1 byte at a time.
+	CMP  R7, R14
+	BLS  extendMatchEnd
+	MOVB (R15), R3
+	MOVB (R7), R4
+	CMP  R4, R3
+	BNE  extendMatchEnd
+	ADD  $1, R15, R15
+	ADD  $1, R7, R7
+	B    cmp1
+
+extendMatchEnd:
+	// Convert from &src[ret] to ret.
+	SUB  R6, R7, R7
+	MOVD R7, ret+40(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func encodeBlock(dst, src []byte) (d int)
+//
+// All local variables fit into registers, other than "var table". The register
+// allocation:
+//	- R3	.	.
+//	- R4	.	.
+//	- R5	64	shift
+//	- R6	72	&src[0], tableSize
+//	- R7	80	&src[s]
+//	- R8	88	&dst[d]
+//	- R9	96	sLimit
+//	- R10	.	&src[nextEmit]
+//	- R11	104	prevHash, currHash, nextHash, offset
+//	- R12	112	&src[base], skip
+//	- R13	.	&src[nextS], &src[len(src) - 8]
+//	- R14	.	len(src), bytesBetweenHashLookups, &src[len(src)], x
+//	- R15	120	candidate
+//	- R16	.	hash constant, 0x1e35a7bd
+//	- R17	.	&table
+//	- .  	128	table
+//
+// The second column (64, 72, etc) is the stack offset to spill the registers
+// when calling other functions. We could pack this slightly tighter, but it's
+// simpler to have a dedicated spill map independent of the function called.
+//
+// "var table [maxTableSize]uint16" takes up 32768 bytes of stack space. An
+// extra 64 bytes, to call other functions, and an extra 64 bytes, to spill
+// local variables (registers) during calls gives 32768 + 64 + 64 = 32896.
+TEXT ·encodeBlock(SB), 0, $32896-56
+	MOVD dst_base+0(FP), R8
+	MOVD src_base+24(FP), R7
+	MOVD src_len+32(FP), R14
+
+	// shift, tableSize := uint32(32-8), 1<<8
+	MOVD  $24, R5
+	MOVD  $256, R6
+	MOVW  $0xa7bd, R16
+	MOVKW $(0x1e35<<16), R16
+
+calcShift:
+	// for ; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 {
+	//	shift--
+	// }
+	MOVD $16384, R2
+	CMP  R2, R6
+	BGE  varTable
+	CMP  R14, R6
+	BGE  varTable
+	SUB  $1, R5, R5
+	LSL  $1, R6, R6
+	B    calcShift
+
+varTable:
+	// var table [maxTableSize]uint16
+	//
+	// In the asm code, unlike the Go code, we can zero-initialize only the
+	// first tableSize elements. Each uint16 element is 2 bytes and each
+	// iterations writes 64 bytes, so we can do only tableSize/32 writes
+	// instead of the 2048 writes that would zero-initialize all of table's
+	// 32768 bytes. This clear could overrun the first tableSize elements, but
+	// it won't overrun the allocated stack size.
+	ADD  $128, RSP, R17
+	MOVD R17, R4
+
+	// !!! R6 = &src[tableSize]
+	ADD R6<<1, R17, R6
+
+memclr:
+	STP.P (ZR, ZR), 64(R4)
+	STP   (ZR, ZR), -48(R4)
+	STP   (ZR, ZR), -32(R4)
+	STP   (ZR, ZR), -16(R4)
+	CMP   R4, R6
+	BHI   memclr
+
+	// !!! R6 = &src[0]
+	MOVD R7, R6
+
+	// sLimit := len(src) - inputMargin
+	MOVD R14, R9
+	SUB  $15, R9, R9
+
+	// !!! Pre-emptively spill R5, R6 and R9 to the stack. Their values don't
+	// change for the rest of the function.
+	MOVD R5, 64(RSP)
+	MOVD R6, 72(RSP)
+	MOVD R9, 96(RSP)
+
+	// nextEmit := 0
+	MOVD R6, R10
+
+	// s := 1
+	ADD $1, R7, R7
+
+	// nextHash := hash(load32(src, s), shift)
+	MOVW 0(R7), R11
+	MULW R16, R11, R11
+	LSRW R5, R11, R11
+
+outer:
+	// for { etc }
+
+	// skip := 32
+	MOVD $32, R12
+
+	// nextS := s
+	MOVD R7, R13
+
+	// candidate := 0
+	MOVD $0, R15
+
+inner0:
+	// for { etc }
+
+	// s := nextS
+	MOVD R13, R7
+
+	// bytesBetweenHashLookups := skip >> 5
+	MOVD R12, R14
+	LSR  $5, R14, R14
+
+	// nextS = s + bytesBetweenHashLookups
+	ADD R14, R13, R13
+
+	// skip += bytesBetweenHashLookups
+	ADD R14, R12, R12
+
+	// if nextS > sLimit { goto emitRemainder }
+	MOVD R13, R3
+	SUB  R6, R3, R3
+	CMP  R9, R3
+	BHI  emitRemainder
+
+	// candidate = int(table[nextHash])
+	MOVHU 0(R17)(R11<<1), R15
+
+	// table[nextHash] = uint16(s)
+	MOVD R7, R3
+	SUB  R6, R3, R3
+
+	MOVH R3, 0(R17)(R11<<1)
+
+	// nextHash = hash(load32(src, nextS), shift)
+	MOVW 0(R13), R11
+	MULW R16, R11
+	LSRW R5, R11, R11
+
+	// if load32(src, s) != load32(src, candidate) { continue } break
+	MOVW 0(R7), R3
+	MOVW (R6)(R15), R4
+	CMPW R4, R3
+	BNE  inner0
+
+fourByteMatch:
+	// As per the encode_other.go code:
+	//
+	// A 4-byte match has been found. We'll later see etc.
+
+	// !!! Jump to a fast path for short (<= 16 byte) literals. See the comment
+	// on inputMargin in encode.go.
+	MOVD R7, R3
+	SUB  R10, R3, R3
+	CMP  $16, R3
+	BLE  emitLiteralFastPath
+
+	// ----------------------------------------
+	// Begin inline of the emitLiteral call.
+	//
+	// d += emitLiteral(dst[d:], src[nextEmit:s])
+
+	MOVW R3, R4
+	SUBW $1, R4, R4
+
+	MOVW $60, R2
+	CMPW R2, R4
+	BLT  inlineEmitLiteralOneByte
+	MOVW $256, R2
+	CMPW R2, R4
+	BLT  inlineEmitLiteralTwoBytes
+
+inlineEmitLiteralThreeBytes:
+	MOVD $0xf4, R1
+	MOVB R1, 0(R8)
+	MOVW R4, 1(R8)
+	ADD  $3, R8, R8
+	B    inlineEmitLiteralMemmove
+
+inlineEmitLiteralTwoBytes:
+	MOVD $0xf0, R1
+	MOVB R1, 0(R8)
+	MOVB R4, 1(R8)
+	ADD  $2, R8, R8
+	B    inlineEmitLiteralMemmove
+
+inlineEmitLiteralOneByte:
+	LSLW $2, R4, R4
+	MOVB R4, 0(R8)
+	ADD  $1, R8, R8
+
+inlineEmitLiteralMemmove:
+	// Spill local variables (registers) onto the stack; call; unspill.
+	//
+	// copy(dst[i:], lit)
+	//
+	// This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push
+	// R8, R10 and R3 as arguments.
+	MOVD R8, 8(RSP)
+	MOVD R10, 16(RSP)
+	MOVD R3, 24(RSP)
+
+	// Finish the "d +=" part of "d += emitLiteral(etc)".
+	ADD   R3, R8, R8
+	MOVD  R7, 80(RSP)
+	MOVD  R8, 88(RSP)
+	MOVD  R15, 120(RSP)
+	CALL  runtime·memmove(SB)
+	MOVD  64(RSP), R5
+	MOVD  72(RSP), R6
+	MOVD  80(RSP), R7
+	MOVD  88(RSP), R8
+	MOVD  96(RSP), R9
+	MOVD  120(RSP), R15
+	ADD   $128, RSP, R17
+	MOVW  $0xa7bd, R16
+	MOVKW $(0x1e35<<16), R16
+	B     inner1
+
+inlineEmitLiteralEnd:
+	// End inline of the emitLiteral call.
+	// ----------------------------------------
+
+emitLiteralFastPath:
+	// !!! Emit the 1-byte encoding "uint8(len(lit)-1)<<2".
+	MOVB R3, R4
+	SUBW $1, R4, R4
+	AND  $0xff, R4, R4
+	LSLW $2, R4, R4
+	MOVB R4, (R8)
+	ADD  $1, R8, R8
+
+	// !!! Implement the copy from lit to dst as a 16-byte load and store.
+	// (Encode's documentation says that dst and src must not overlap.)
+	//
+	// This always copies 16 bytes, instead of only len(lit) bytes, but that's
+	// OK. Subsequent iterations will fix up the overrun.
+	//
+	// Note that on arm64, it is legal and cheap to issue unaligned 8-byte or
+	// 16-byte loads and stores. This technique probably wouldn't be as
+	// effective on architectures that are fussier about alignment.
+	LDP 0(R10), (R0, R1)
+	STP (R0, R1), 0(R8)
+	ADD R3, R8, R8
+
+inner1:
+	// for { etc }
+
+	// base := s
+	MOVD R7, R12
+
+	// !!! offset := base - candidate
+	MOVD R12, R11
+	SUB  R15, R11, R11
+	SUB  R6, R11, R11
+
+	// ----------------------------------------
+	// Begin inline of the extendMatch call.
+	//
+	// s = extendMatch(src, candidate+4, s+4)
+
+	// !!! R14 = &src[len(src)]
+	MOVD src_len+32(FP), R14
+	ADD  R6, R14, R14
+
+	// !!! R13 = &src[len(src) - 8]
+	MOVD R14, R13
+	SUB  $8, R13, R13
+
+	// !!! R15 = &src[candidate + 4]
+	ADD $4, R15, R15
+	ADD R6, R15, R15
+
+	// !!! s += 4
+	ADD $4, R7, R7
+
+inlineExtendMatchCmp8:
+	// As long as we are 8 or more bytes before the end of src, we can load and
+	// compare 8 bytes at a time. If those 8 bytes are equal, repeat.
+	CMP  R13, R7
+	BHI  inlineExtendMatchCmp1
+	MOVD (R15), R3
+	MOVD (R7), R4
+	CMP  R4, R3
+	BNE  inlineExtendMatchBSF
+	ADD  $8, R15, R15
+	ADD  $8, R7, R7
+	B    inlineExtendMatchCmp8
+
+inlineExtendMatchBSF:
+	// If those 8 bytes were not equal, XOR the two 8 byte values, and return
+	// the index of the first byte that differs.
+	// RBIT reverses the bit order, then CLZ counts the leading zeros, the
+	// combination of which finds the least significant bit which is set.
+	// The arm64 architecture is little-endian, and the shift by 3 converts
+	// a bit index to a byte index.
+	EOR  R3, R4, R4
+	RBIT R4, R4
+	CLZ  R4, R4
+	ADD  R4>>3, R7, R7
+	B    inlineExtendMatchEnd
+
+inlineExtendMatchCmp1:
+	// In src's tail, compare 1 byte at a time.
+	CMP  R7, R14
+	BLS  inlineExtendMatchEnd
+	MOVB (R15), R3
+	MOVB (R7), R4
+	CMP  R4, R3
+	BNE  inlineExtendMatchEnd
+	ADD  $1, R15, R15
+	ADD  $1, R7, R7
+	B    inlineExtendMatchCmp1
+
+inlineExtendMatchEnd:
+	// End inline of the extendMatch call.
+	// ----------------------------------------
+
+	// ----------------------------------------
+	// Begin inline of the emitCopy call.
+	//
+	// d += emitCopy(dst[d:], base-candidate, s-base)
+
+	// !!! length := s - base
+	MOVD R7, R3
+	SUB  R12, R3, R3
+
+inlineEmitCopyLoop0:
+	// for length >= 68 { etc }
+	MOVW $68, R2
+	CMPW R2, R3
+	BLT  inlineEmitCopyStep1
+
+	// Emit a length 64 copy, encoded as 3 bytes.
+	MOVD $0xfe, R1
+	MOVB R1, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+	SUBW $64, R3, R3
+	B    inlineEmitCopyLoop0
+
+inlineEmitCopyStep1:
+	// if length > 64 { etc }
+	MOVW $64, R2
+	CMPW R2, R3
+	BLE  inlineEmitCopyStep2
+
+	// Emit a length 60 copy, encoded as 3 bytes.
+	MOVD $0xee, R1
+	MOVB R1, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+	SUBW $60, R3, R3
+
+inlineEmitCopyStep2:
+	// if length >= 12 || offset >= 2048 { goto inlineEmitCopyStep3 }
+	MOVW $12, R2
+	CMPW R2, R3
+	BGE  inlineEmitCopyStep3
+	MOVW $2048, R2
+	CMPW R2, R11
+	BGE  inlineEmitCopyStep3
+
+	// Emit the remaining copy, encoded as 2 bytes.
+	MOVB R11, 1(R8)
+	LSRW $8, R11, R11
+	LSLW $5, R11, R11
+	SUBW $4, R3, R3
+	AND  $0xff, R3, R3
+	LSLW $2, R3, R3
+	ORRW R3, R11, R11
+	ORRW $1, R11, R11
+	MOVB R11, 0(R8)
+	ADD  $2, R8, R8
+	B    inlineEmitCopyEnd
+
+inlineEmitCopyStep3:
+	// Emit the remaining copy, encoded as 3 bytes.
+	SUBW $1, R3, R3
+	LSLW $2, R3, R3
+	ORRW $2, R3, R3
+	MOVB R3, 0(R8)
+	MOVW R11, 1(R8)
+	ADD  $3, R8, R8
+
+inlineEmitCopyEnd:
+	// End inline of the emitCopy call.
+	// ----------------------------------------
+
+	// nextEmit = s
+	MOVD R7, R10
+
+	// if s >= sLimit { goto emitRemainder }
+	MOVD R7, R3
+	SUB  R6, R3, R3
+	CMP  R3, R9
+	BLS  emitRemainder
+
+	// As per the encode_other.go code:
+	//
+	// We could immediately etc.
+
+	// x := load64(src, s-1)
+	MOVD -1(R7), R14
+
+	// prevHash := hash(uint32(x>>0), shift)
+	MOVW R14, R11
+	MULW R16, R11, R11
+	LSRW R5, R11, R11
+
+	// table[prevHash] = uint16(s-1)
+	MOVD R7, R3
+	SUB  R6, R3, R3
+	SUB  $1, R3, R3
+
+	MOVHU R3, 0(R17)(R11<<1)
+
+	// currHash := hash(uint32(x>>8), shift)
+	LSR  $8, R14, R14
+	MOVW R14, R11
+	MULW R16, R11, R11
+	LSRW R5, R11, R11
+
+	// candidate = int(table[currHash])
+	MOVHU 0(R17)(R11<<1), R15
+
+	// table[currHash] = uint16(s)
+	ADD   $1, R3, R3
+	MOVHU R3, 0(R17)(R11<<1)
+
+	// if uint32(x>>8) == load32(src, candidate) { continue }
+	MOVW (R6)(R15), R4
+	CMPW R4, R14
+	BEQ  inner1
+
+	// nextHash = hash(uint32(x>>16), shift)
+	LSR  $8, R14, R14
+	MOVW R14, R11
+	MULW R16, R11, R11
+	LSRW R5, R11, R11
+
+	// s++
+	ADD $1, R7, R7
+
+	// break out of the inner1 for loop, i.e. continue the outer loop.
+	B outer
+
+emitRemainder:
+	// if nextEmit < len(src) { etc }
+	MOVD src_len+32(FP), R3
+	ADD  R6, R3, R3
+	CMP  R3, R10
+	BEQ  encodeBlockEnd
+
+	// d += emitLiteral(dst[d:], src[nextEmit:])
+	//
+	// Push args.
+	MOVD R8, 8(RSP)
+	MOVD $0, 16(RSP)  // Unnecessary, as the callee ignores it, but conservative.
+	MOVD $0, 24(RSP)  // Unnecessary, as the callee ignores it, but conservative.
+	MOVD R10, 32(RSP)
+	SUB  R10, R3, R3
+	MOVD R3, 40(RSP)
+	MOVD R3, 48(RSP)  // Unnecessary, as the callee ignores it, but conservative.
+
+	// Spill local variables (registers) onto the stack; call; unspill.
+	MOVD R8, 88(RSP)
+	CALL ·emitLiteral(SB)
+	MOVD 88(RSP), R8
+
+	// Finish the "d +=" part of "d += emitLiteral(etc)".
+	MOVD 56(RSP), R1
+	ADD  R1, R8, R8
+
+encodeBlockEnd:
+	MOVD dst_base+0(FP), R3
+	SUB  R3, R8, R8
+	MOVD R8, d+48(FP)
+	RET

vendor/github.com/golang/snappy/encode_asm.go

@@ -0,0 +1,30 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build !noasm
+// +build amd64 arm64
+
+package snappy
+
+// emitLiteral has the same semantics as in encode_other.go.
+//
+//go:noescape
+func emitLiteral(dst, lit []byte) int
+
+// emitCopy has the same semantics as in encode_other.go.
+//
+//go:noescape
+func emitCopy(dst []byte, offset, length int) int
+
+// extendMatch has the same semantics as in encode_other.go.
+//
+//go:noescape
+func extendMatch(src []byte, i, j int) int
+
+// encodeBlock has the same semantics as in encode_other.go.
+//
+//go:noescape
+func encodeBlock(dst, src []byte) (d int)

vendor/github.com/golang/snappy/encode_other.go

@@ -0,0 +1,238 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !amd64,!arm64 appengine !gc noasm
+
+package snappy
+
+func load32(b []byte, i int) uint32 {
+	b = b[i : i+4 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+}
+
+func load64(b []byte, i int) uint64 {
+	b = b[i : i+8 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 |
+		uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+}
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+//
+// It assumes that:
+//	dst is long enough to hold the encoded bytes
+//	1 <= len(lit) && len(lit) <= 65536
+func emitLiteral(dst, lit []byte) int {
+	i, n := 0, uint(len(lit)-1)
+	switch {
+	case n < 60:
+		dst[0] = uint8(n)<<2 | tagLiteral
+		i = 1
+	case n < 1<<8:
+		dst[0] = 60<<2 | tagLiteral
+		dst[1] = uint8(n)
+		i = 2
+	default:
+		dst[0] = 61<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		i = 3
+	}
+	return i + copy(dst[i:], lit)
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+//
+// It assumes that:
+//	dst is long enough to hold the encoded bytes
+//	1 <= offset && offset <= 65535
+//	4 <= length && length <= 65535
+func emitCopy(dst []byte, offset, length int) int {
+	i := 0
+	// The maximum length for a single tagCopy1 or tagCopy2 op is 64 bytes. The
+	// threshold for this loop is a little higher (at 68 = 64 + 4), and the
+	// length emitted down below is is a little lower (at 60 = 64 - 4), because
+	// it's shorter to encode a length 67 copy as a length 60 tagCopy2 followed
+	// by a length 7 tagCopy1 (which encodes as 3+2 bytes) than to encode it as
+	// a length 64 tagCopy2 followed by a length 3 tagCopy2 (which encodes as
+	// 3+3 bytes). The magic 4 in the 64±4 is because the minimum length for a
+	// tagCopy1 op is 4 bytes, which is why a length 3 copy has to be an
+	// encodes-as-3-bytes tagCopy2 instead of an encodes-as-2-bytes tagCopy1.
+	for length >= 68 {
+		// Emit a length 64 copy, encoded as 3 bytes.
+		dst[i+0] = 63<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 64
+	}
+	if length > 64 {
+		// Emit a length 60 copy, encoded as 3 bytes.
+		dst[i+0] = 59<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 60
+	}
+	if length >= 12 || offset >= 2048 {
+		// Emit the remaining copy, encoded as 3 bytes.
+		dst[i+0] = uint8(length-1)<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		return i + 3
+	}
+	// Emit the remaining copy, encoded as 2 bytes.
+	dst[i+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
+	dst[i+1] = uint8(offset)
+	return i + 2
+}
+
+// extendMatch returns the largest k such that k <= len(src) and that
+// src[i:i+k-j] and src[j:k] have the same contents.
+//
+// It assumes that:
+//	0 <= i && i < j && j <= len(src)
+func extendMatch(src []byte, i, j int) int {
+	for ; j < len(src) && src[i] == src[j]; i, j = i+1, j+1 {
+	}
+	return j
+}
+
+func hash(u, shift uint32) uint32 {
+	return (u * 0x1e35a7bd) >> shift
+}
+
+// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It
+// assumes that the varint-encoded length of the decompressed bytes has already
+// been written.
+//
+// It also assumes that:
+//	len(dst) >= MaxEncodedLen(len(src)) &&
+// 	minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
+func encodeBlock(dst, src []byte) (d int) {
+	// Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive.
+	// The table element type is uint16, as s < sLimit and sLimit < len(src)
+	// and len(src) <= maxBlockSize and maxBlockSize == 65536.
+	const (
+		maxTableSize = 1 << 14
+		// tableMask is redundant, but helps the compiler eliminate bounds
+		// checks.
+		tableMask = maxTableSize - 1
+	)
+	shift := uint32(32 - 8)
+	for tableSize := 1 << 8; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 {
+		shift--
+	}
+	// In Go, all array elements are zero-initialized, so there is no advantage
+	// to a smaller tableSize per se. However, it matches the C++ algorithm,
+	// and in the asm versions of this code, we can get away with zeroing only
+	// the first tableSize elements.
+	var table [maxTableSize]uint16
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := len(src) - inputMargin
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := 0
+
+	// The encoded form must start with a literal, as there are no previous
+	// bytes to copy, so we start looking for hash matches at s == 1.
+	s := 1
+	nextHash := hash(load32(src, s), shift)
+
+	for {
+		// Copied from the C++ snappy implementation:
+		//
+		// Heuristic match skipping: If 32 bytes are scanned with no matches
+		// found, start looking only at every other byte. If 32 more bytes are
+		// scanned (or skipped), look at every third byte, etc.. When a match
+		// is found, immediately go back to looking at every byte. This is a
+		// small loss (~5% performance, ~0.1% density) for compressible data
+		// due to more bookkeeping, but for non-compressible data (such as
+		// JPEG) it's a huge win since the compressor quickly "realizes" the
+		// data is incompressible and doesn't bother looking for matches
+		// everywhere.
+		//
+		// The "skip" variable keeps track of how many bytes there are since
+		// the last match; dividing it by 32 (ie. right-shifting by five) gives
+		// the number of bytes to move ahead for each iteration.
+		skip := 32
+
+		nextS := s
+		candidate := 0
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidate = int(table[nextHash&tableMask])
+			table[nextHash&tableMask] = uint16(s)
+			nextHash = hash(load32(src, nextS), shift)
+			if load32(src, s) == load32(src, candidate) {
+				break
+			}
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+		d += emitLiteral(dst[d:], src[nextEmit:s])
+
+		// Call emitCopy, and then see if another emitCopy could be our next
+		// move. Repeat until we find no match for the input immediately after
+		// what was consumed by the last emitCopy call.
+		//
+		// If we exit this loop normally then we need to call emitLiteral next,
+		// though we don't yet know how big the literal will be. We handle that
+		// by proceeding to the next iteration of the main loop. We also can
+		// exit this loop via goto if we get close to exhausting the input.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+			base := s
+
+			// Extend the 4-byte match as long as possible.
+			//
+			// This is an inlined version of:
+			//	s = extendMatch(src, candidate+4, s+4)
+			s += 4
+			for i := candidate + 4; s < len(src) && src[i] == src[s]; i, s = i+1, s+1 {
+			}
+
+			d += emitCopy(dst[d:], base-candidate, s-base)
+			nextEmit = s
+			if s >= sLimit {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-1 and at s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load64(src, s-1)
+			prevHash := hash(uint32(x>>0), shift)
+			table[prevHash&tableMask] = uint16(s - 1)
+			currHash := hash(uint32(x>>8), shift)
+			candidate = int(table[currHash&tableMask])
+			table[currHash&tableMask] = uint16(s)
+			if uint32(x>>8) != load32(src, candidate) {
+				nextHash = hash(uint32(x>>16), shift)
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if nextEmit < len(src) {
+		d += emitLiteral(dst[d:], src[nextEmit:])
+	}
+	return d
+}

vendor/github.com/golang/snappy/snappy.go

@@ -0,0 +1,98 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package snappy implements the Snappy compression format. It aims for very
+// high speeds and reasonable compression.
+//
+// There are actually two Snappy formats: block and stream. They are related,
+// but different: trying to decompress block-compressed data as a Snappy stream
+// will fail, and vice versa. The block format is the Decode and Encode
+// functions and the stream format is the Reader and Writer types.
+//
+// The block format, the more common case, is used when the complete size (the
+// number of bytes) of the original data is known upfront, at the time
+// compression starts. The stream format, also known as the framing format, is
+// for when that isn't always true.
+//
+// The canonical, C++ implementation is at https://github.com/google/snappy and
+// it only implements the block format.
+package snappy // import "github.com/golang/snappy"
+
+import (
+	"hash/crc32"
+)
+
+/*
+Each encoded block begins with the varint-encoded length of the decoded data,
+followed by a sequence of chunks. Chunks begin and end on byte boundaries. The
+first byte of each chunk is broken into its 2 least and 6 most significant bits
+called l and m: l ranges in [0, 4) and m ranges in [0, 64). l is the chunk tag.
+Zero means a literal tag. All other values mean a copy tag.
+
+For literal tags:
+  - If m < 60, the next 1 + m bytes are literal bytes.
+  - Otherwise, let n be the little-endian unsigned integer denoted by the next
+    m - 59 bytes. The next 1 + n bytes after that are literal bytes.
+
+For copy tags, length bytes are copied from offset bytes ago, in the style of
+Lempel-Ziv compression algorithms. In particular:
+  - For l == 1, the offset ranges in [0, 1<<11) and the length in [4, 12).
+    The length is 4 + the low 3 bits of m. The high 3 bits of m form bits 8-10
+    of the offset. The next byte is bits 0-7 of the offset.
+  - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65).
+    The length is 1 + m. The offset is the little-endian unsigned integer
+    denoted by the next 2 bytes.
+  - For l == 3, this tag is a legacy format that is no longer issued by most
+    encoders. Nonetheless, the offset ranges in [0, 1<<32) and the length in
+    [1, 65). The length is 1 + m. The offset is the little-endian unsigned
+    integer denoted by the next 4 bytes.
+*/
+const (
+	tagLiteral = 0x00
+	tagCopy1   = 0x01
+	tagCopy2   = 0x02
+	tagCopy4   = 0x03
+)
+
+const (
+	checksumSize    = 4
+	chunkHeaderSize = 4
+	magicChunk      = "\xff\x06\x00\x00" + magicBody
+	magicBody       = "sNaPpY"
+
+	// maxBlockSize is the maximum size of the input to encodeBlock. It is not
+	// part of the wire format per se, but some parts of the encoder assume
+	// that an offset fits into a uint16.
+	//
+	// Also, for the framing format (Writer type instead of Encode function),
+	// https://github.com/google/snappy/blob/master/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536
+	// bytes".
+	maxBlockSize = 65536
+
+	// maxEncodedLenOfMaxBlockSize equals MaxEncodedLen(maxBlockSize), but is
+	// hard coded to be a const instead of a variable, so that obufLen can also
+	// be a const. Their equivalence is confirmed by
+	// TestMaxEncodedLenOfMaxBlockSize.
+	maxEncodedLenOfMaxBlockSize = 76490
+
+	obufHeaderLen = len(magicChunk) + checksumSize + chunkHeaderSize
+	obufLen       = obufHeaderLen + maxEncodedLenOfMaxBlockSize
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func crc(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return uint32(c>>15|c<<17) + 0xa282ead8
+}

vendor/github.com/klauspost/compress/.gitattributes

@@ -0,0 +1,2 @@
+* -text
+*.bin -text -diff

vendor/github.com/klauspost/compress/.gitignore

@@ -0,0 +1,32 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+
+# Folders
+_obj
+_test
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+*.test
+*.prof
+/s2/cmd/_s2sx/sfx-exe
+
+# Linux perf files
+perf.data
+perf.data.old
+
+# gdb history
+.gdb_history

vendor/github.com/klauspost/compress/.goreleaser.yml

@@ -0,0 +1,141 @@
+# This is an example goreleaser.yaml file with some sane defaults.
+# Make sure to check the documentation at http://goreleaser.com
+before:
+  hooks:
+    - ./gen.sh
+    - go install mvdan.cc/garble@latest
+
+builds:
+  -
+    id: "s2c"
+    binary: s2c
+    main: ./s2/cmd/s2c/main.go
+    flags:
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+    gobinary: garble
+  -
+    id: "s2d"
+    binary: s2d
+    main: ./s2/cmd/s2d/main.go
+    flags:
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+    gobinary: garble
+  -
+    id: "s2sx"
+    binary: s2sx
+    main: ./s2/cmd/_s2sx/main.go
+    flags:
+      - -modfile=s2sx.mod
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+    gobinary: garble
+
+archives:
+  -
+    id: s2-binaries
+    name_template: "s2-{{ .Os }}_{{ .Arch }}_{{ .Version }}"
+    replacements:
+      aix: AIX
+      darwin: OSX
+      linux: Linux
+      windows: Windows
+      386: i386
+      amd64: x86_64
+      freebsd: FreeBSD
+      netbsd: NetBSD
+    format_overrides:
+      - goos: windows
+        format: zip
+    files:
+      - unpack/*
+      - s2/LICENSE
+      - s2/README.md
+checksum:
+  name_template: 'checksums.txt'
+snapshot:
+  name_template: "{{ .Tag }}-next"
+changelog:
+  sort: asc
+  filters:
+    exclude:
+    - '^doc:'
+    - '^docs:'
+    - '^test:'
+    - '^tests:'
+    - '^Update\sREADME.md'
+
+nfpms:
+  -
+    file_name_template: "s2_package_{{ .Version }}_{{ .Os }}_{{ .Arch }}"
+    vendor: Klaus Post
+    homepage: https://github.com/klauspost/compress
+    maintainer: Klaus Post <klauspost@gmail.com>
+    description: S2 Compression Tool
+    license: BSD 3-Clause
+    formats:
+      - deb
+      - rpm
+    replacements:
+      darwin: Darwin
+      linux: Linux
+      freebsd: FreeBSD
+      amd64: x86_64

vendor/github.com/klauspost/compress/LICENSE

@@ -0,0 +1,304 @@
+Copyright (c) 2012 The Go Authors. All rights reserved.
+Copyright (c) 2019 Klaus Post. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+------------------
+
+Files: gzhttp/*
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2016-2017 The New York Times Company
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+------------------
+
+Files: s2/cmd/internal/readahead/*
+
+The MIT License (MIT)
+
+Copyright (c) 2015 Klaus Post
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+---------------------
+Files: snappy/*
+Files: internal/snapref/*
+
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+-----------------
+
+Files: s2/cmd/internal/filepathx/*
+
+Copyright 2016 The filepathx Authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

vendor/github.com/klauspost/compress/README.md

@@ -0,0 +1,496 @@
+# compress
+
+This package provides various compression algorithms.
+
+* [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and decompression in pure Go.
+* [S2](https://github.com/klauspost/compress/tree/master/s2#s2-compression) is a high performance replacement for Snappy.
+* Optimized [deflate](https://godoc.org/github.com/klauspost/compress/flate) packages which can be used as a dropin replacement for [gzip](https://godoc.org/github.com/klauspost/compress/gzip), [zip](https://godoc.org/github.com/klauspost/compress/zip) and [zlib](https://godoc.org/github.com/klauspost/compress/zlib).
+* [snappy](https://github.com/klauspost/compress/tree/master/snappy) is a drop-in replacement for `github.com/golang/snappy` offering better compression and concurrent streams.
+* [huff0](https://github.com/klauspost/compress/tree/master/huff0) and [FSE](https://github.com/klauspost/compress/tree/master/fse) implementations for raw entropy encoding.
+* [gzhttp](https://github.com/klauspost/compress/tree/master/gzhttp) Provides client and server wrappers for handling gzipped requests efficiently.
+* [pgzip](https://github.com/klauspost/pgzip) is a separate package that provides a very fast parallel gzip implementation.
+* [fuzz package](https://github.com/klauspost/compress-fuzz) for fuzz testing all compressors/decompressors here.
+
+[![Go Reference](https://pkg.go.dev/badge/klauspost/compress.svg)](https://pkg.go.dev/github.com/klauspost/compress?tab=subdirectories)
+[![Go](https://github.com/klauspost/compress/actions/workflows/go.yml/badge.svg)](https://github.com/klauspost/compress/actions/workflows/go.yml)
+[![Sourcegraph Badge](https://sourcegraph.com/github.com/klauspost/compress/-/badge.svg)](https://sourcegraph.com/github.com/klauspost/compress?badge)
+
+# changelog
+
+* Mar 11, 2022 (v1.15.1)
+	* huff0: Add x86 assembly of Decode4X by @WojciechMula in [#512](https://github.com/klauspost/compress/pull/512)
+	* zstd: Reuse zip decoders in [#514](https://github.com/klauspost/compress/pull/514)
+	* zstd: Detect extra block data and report as corrupted in [#520](https://github.com/klauspost/compress/pull/520)
+	* zstd: Handle zero sized frame content size stricter in [#521](https://github.com/klauspost/compress/pull/521)
+	* zstd: Add stricter block size checks in [#523](https://github.com/klauspost/compress/pull/523)
+
+* Mar 3, 2022 (v1.15.0)
+	* zstd: Refactor decoder by @klauspost in [#498](https://github.com/klauspost/compress/pull/498)
+	* zstd: Add stream encoding without goroutines by @klauspost in [#505](https://github.com/klauspost/compress/pull/505)
+	* huff0: Prevent single blocks exceeding 16 bits by @klauspost in[#507](https://github.com/klauspost/compress/pull/507)
+	* flate: Inline literal emission by @klauspost in [#509](https://github.com/klauspost/compress/pull/509)
+	* gzhttp: Add zstd to transport by @klauspost in [#400](https://github.com/klauspost/compress/pull/400)
+	* gzhttp: Make content-type optional by @klauspost in [#510](https://github.com/klauspost/compress/pull/510)
+
+<details>
+	<summary>See  Details</summary>
+Both compression and decompression now supports "synchronous" stream operations. This means that whenever "concurrency" is set to 1, they will operate without spawning goroutines.
+
+Stream decompression is now faster on asynchronous, since the goroutine allocation much more effectively splits the workload. On typical streams this will typically use 2 cores fully for decompression. When a stream has finished decoding no goroutines will be left over, so decoders can now safely be pooled and still be garbage collected.
+
+While the release has been extensively tested, it is recommended to testing when upgrading.
+</details>
+
+* Feb 22, 2022 (v1.14.4)
+	* flate: Fix rare huffman only (-2) corruption. [#503](https://github.com/klauspost/compress/pull/503)
+	* zip: Update deprecated CreateHeaderRaw to correctly call CreateRaw by @saracen in [#502](https://github.com/klauspost/compress/pull/502)
+	* zip: don't read data descriptor early by @saracen in [#501](https://github.com/klauspost/compress/pull/501)  #501
+	* huff0: Use static decompression buffer up to 30% faster by @klauspost in [#499](https://github.com/klauspost/compress/pull/499) [#500](https://github.com/klauspost/compress/pull/500)
+
+* Feb 17, 2022 (v1.14.3)
+	* flate: Improve fastest levels compression speed ~10% more throughput. [#482](https://github.com/klauspost/compress/pull/482) [#489](https://github.com/klauspost/compress/pull/489) [#490](https://github.com/klauspost/compress/pull/490) [#491](https://github.com/klauspost/compress/pull/491) [#494](https://github.com/klauspost/compress/pull/494)  [#478](https://github.com/klauspost/compress/pull/478)
+	* flate: Faster decompression speed, ~5-10%. [#483](https://github.com/klauspost/compress/pull/483)
+	* s2: Faster compression with Go v1.18 and amd64 microarch level 3+. [#484](https://github.com/klauspost/compress/pull/484) [#486](https://github.com/klauspost/compress/pull/486)
+
+* Jan 25, 2022 (v1.14.2)
+	* zstd: improve header decoder by @dsnet  [#476](https://github.com/klauspost/compress/pull/476)
+	* zstd: Add bigger default blocks  [#469](https://github.com/klauspost/compress/pull/469)
+	* zstd: Remove unused decompression buffer [#470](https://github.com/klauspost/compress/pull/470)
+	* zstd: Fix logically dead code by @ningmingxiao [#472](https://github.com/klauspost/compress/pull/472)
+	* flate: Improve level 7-9 [#471](https://github.com/klauspost/compress/pull/471) [#473](https://github.com/klauspost/compress/pull/473)
+	* zstd: Add noasm tag for xxhash [#475](https://github.com/klauspost/compress/pull/475)
+
+* Jan 11, 2022 (v1.14.1)
+	* s2: Add stream index in [#462](https://github.com/klauspost/compress/pull/462)
+	* flate: Speed and efficiency improvements in [#439](https://github.com/klauspost/compress/pull/439) [#461](https://github.com/klauspost/compress/pull/461) [#455](https://github.com/klauspost/compress/pull/455) [#452](https://github.com/klauspost/compress/pull/452) [#458](https://github.com/klauspost/compress/pull/458)
+	* zstd: Performance improvement in [#420]( https://github.com/klauspost/compress/pull/420) [#456](https://github.com/klauspost/compress/pull/456) [#437](https://github.com/klauspost/compress/pull/437) [#467](https://github.com/klauspost/compress/pull/467) [#468](https://github.com/klauspost/compress/pull/468)
+	* zstd: add arm64 xxhash assembly in [#464](https://github.com/klauspost/compress/pull/464)
+	* Add garbled for binaries for s2 in [#445](https://github.com/klauspost/compress/pull/445)
+
+* Aug 30, 2021 (v1.13.5)
+	* gz/zlib/flate: Alias stdlib errors [#425](https://github.com/klauspost/compress/pull/425)
+	* s2: Add block support to commandline tools [#413](https://github.com/klauspost/compress/pull/413)
+	* zstd: pooledZipWriter should return Writers to the same pool [#426](https://github.com/klauspost/compress/pull/426)
+	* Removed golang/snappy as external dependency for tests [#421](https://github.com/klauspost/compress/pull/421)
+
+* Aug 12, 2021 (v1.13.4)
+	* Add [snappy replacement package](https://github.com/klauspost/compress/tree/master/snappy).
+	* zstd: Fix incorrect encoding in "best" mode [#415](https://github.com/klauspost/compress/pull/415)
+
+* Aug 3, 2021 (v1.13.3) 
+	* zstd: Improve Best compression [#404](https://github.com/klauspost/compress/pull/404)
+	* zstd: Fix WriteTo error forwarding [#411](https://github.com/klauspost/compress/pull/411)
+	* gzhttp: Return http.HandlerFunc instead of http.Handler. Unlikely breaking change. [#406](https://github.com/klauspost/compress/pull/406)
+	* s2sx: Fix max size error [#399](https://github.com/klauspost/compress/pull/399)
+	* zstd: Add optional stream content size on reset [#401](https://github.com/klauspost/compress/pull/401)
+	* zstd: use SpeedBestCompression for level >= 10 [#410](https://github.com/klauspost/compress/pull/410)
+
+* Jun 14, 2021 (v1.13.1)
+	* s2: Add full Snappy output support  [#396](https://github.com/klauspost/compress/pull/396)
+	* zstd: Add configurable [Decoder window](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithDecoderMaxWindow) size [#394](https://github.com/klauspost/compress/pull/394)
+	* gzhttp: Add header to skip compression  [#389](https://github.com/klauspost/compress/pull/389)
+	* s2: Improve speed with bigger output margin  [#395](https://github.com/klauspost/compress/pull/395)
+
+* Jun 3, 2021 (v1.13.0)
+	* Added [gzhttp](https://github.com/klauspost/compress/tree/master/gzhttp#gzip-handler) which allows wrapping HTTP servers and clients with GZIP compressors.
+	* zstd: Detect short invalid signatures [#382](https://github.com/klauspost/compress/pull/382)
+	* zstd: Spawn decoder goroutine only if needed. [#380](https://github.com/klauspost/compress/pull/380)
+
+<details>
+	<summary>See changes to v1.12.x</summary>
+	
+* May 25, 2021 (v1.12.3)
+	* deflate: Better/faster Huffman encoding [#374](https://github.com/klauspost/compress/pull/374)
+	* deflate: Allocate less for history. [#375](https://github.com/klauspost/compress/pull/375)
+	* zstd: Forward read errors [#373](https://github.com/klauspost/compress/pull/373) 
+
+* Apr 27, 2021 (v1.12.2)
+	* zstd: Improve better/best compression [#360](https://github.com/klauspost/compress/pull/360) [#364](https://github.com/klauspost/compress/pull/364) [#365](https://github.com/klauspost/compress/pull/365)
+	* zstd: Add helpers to compress/decompress zstd inside zip files [#363](https://github.com/klauspost/compress/pull/363)
+	* deflate: Improve level 5+6 compression [#367](https://github.com/klauspost/compress/pull/367)
+	* s2: Improve better/best compression [#358](https://github.com/klauspost/compress/pull/358) [#359](https://github.com/klauspost/compress/pull/358)
+	* s2: Load after checking src limit on amd64. [#362](https://github.com/klauspost/compress/pull/362)
+	* s2sx: Limit max executable size [#368](https://github.com/klauspost/compress/pull/368) 
+
+* Apr 14, 2021 (v1.12.1)
+	* snappy package removed. Upstream added as dependency.
+	* s2: Better compression in "best" mode [#353](https://github.com/klauspost/compress/pull/353)
+	* s2sx: Add stdin input and detect pre-compressed from signature [#352](https://github.com/klauspost/compress/pull/352)
+	* s2c/s2d: Add http as possible input [#348](https://github.com/klauspost/compress/pull/348)
+	* s2c/s2d/s2sx: Always truncate when writing files [#352](https://github.com/klauspost/compress/pull/352)
+	* zstd: Reduce memory usage further when using [WithLowerEncoderMem](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithLowerEncoderMem) [#346](https://github.com/klauspost/compress/pull/346)
+	* s2: Fix potential problem with amd64 assembly and profilers [#349](https://github.com/klauspost/compress/pull/349)
+</details>
+
+<details>
+	<summary>See changes to v1.11.x</summary>
+	
+* Mar 26, 2021 (v1.11.13)
+	* zstd: Big speedup on small dictionary encodes [#344](https://github.com/klauspost/compress/pull/344) [#345](https://github.com/klauspost/compress/pull/345)
+	* zstd: Add [WithLowerEncoderMem](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithLowerEncoderMem) encoder option [#336](https://github.com/klauspost/compress/pull/336)
+	* deflate: Improve entropy compression [#338](https://github.com/klauspost/compress/pull/338)
+	* s2: Clean up and minor performance improvement in best [#341](https://github.com/klauspost/compress/pull/341)
+
+* Mar 5, 2021 (v1.11.12)
+	* s2: Add `s2sx` binary that creates [self extracting archives](https://github.com/klauspost/compress/tree/master/s2#s2sx-self-extracting-archives).
+	* s2: Speed up decompression on non-assembly platforms [#328](https://github.com/klauspost/compress/pull/328)
+
+* Mar 1, 2021 (v1.11.9)
+	* s2: Add ARM64 decompression assembly. Around 2x output speed. [#324](https://github.com/klauspost/compress/pull/324)
+	* s2: Improve "better" speed and efficiency. [#325](https://github.com/klauspost/compress/pull/325)
+	* s2: Fix binaries.
+
+* Feb 25, 2021 (v1.11.8)
+	* s2: Fixed occational out-of-bounds write on amd64. Upgrade recommended.
+	* s2: Add AMD64 assembly for better mode. 25-50% faster. [#315](https://github.com/klauspost/compress/pull/315)
+	* s2: Less upfront decoder allocation. [#322](https://github.com/klauspost/compress/pull/322)
+	* zstd: Faster "compression" of incompressible data. [#314](https://github.com/klauspost/compress/pull/314)
+	* zip: Fix zip64 headers. [#313](https://github.com/klauspost/compress/pull/313)
+  
+* Jan 14, 2021 (v1.11.7)
+	* Use Bytes() interface to get bytes across packages. [#309](https://github.com/klauspost/compress/pull/309)
+	* s2: Add 'best' compression option.  [#310](https://github.com/klauspost/compress/pull/310)
+	* s2: Add ReaderMaxBlockSize, changes `s2.NewReader` signature to include varargs. [#311](https://github.com/klauspost/compress/pull/311)
+	* s2: Fix crash on small better buffers. [#308](https://github.com/klauspost/compress/pull/308)
+	* s2: Clean up decoder. [#312](https://github.com/klauspost/compress/pull/312)
+
+* Jan 7, 2021 (v1.11.6)
+	* zstd: Make decoder allocations smaller [#306](https://github.com/klauspost/compress/pull/306)
+	* zstd: Free Decoder resources when Reset is called with a nil io.Reader  [#305](https://github.com/klauspost/compress/pull/305)
+
+* Dec 20, 2020 (v1.11.4)
+	* zstd: Add Best compression mode [#304](https://github.com/klauspost/compress/pull/304)
+	* Add header decoder [#299](https://github.com/klauspost/compress/pull/299)
+	* s2: Add uncompressed stream option [#297](https://github.com/klauspost/compress/pull/297)
+	* Simplify/speed up small blocks with known max size. [#300](https://github.com/klauspost/compress/pull/300)
+	* zstd: Always reset literal dict encoder [#303](https://github.com/klauspost/compress/pull/303)
+
+* Nov 15, 2020 (v1.11.3)
+	* inflate: 10-15% faster decompression  [#293](https://github.com/klauspost/compress/pull/293)
+	* zstd: Tweak DecodeAll default allocation [#295](https://github.com/klauspost/compress/pull/295)
+
+* Oct 11, 2020 (v1.11.2)
+	* s2: Fix out of bounds read in "better" block compression [#291](https://github.com/klauspost/compress/pull/291)
+
+* Oct 1, 2020 (v1.11.1)
+	* zstd: Set allLitEntropy true in default configuration [#286](https://github.com/klauspost/compress/pull/286)
+
+* Sept 8, 2020 (v1.11.0)
+	* zstd: Add experimental compression [dictionaries](https://github.com/klauspost/compress/tree/master/zstd#dictionaries) [#281](https://github.com/klauspost/compress/pull/281)
+	* zstd: Fix mixed Write and ReadFrom calls [#282](https://github.com/klauspost/compress/pull/282)
+	* inflate/gz: Limit variable shifts, ~5% faster decompression [#274](https://github.com/klauspost/compress/pull/274)
+</details>
+
+<details>
+	<summary>See changes to v1.10.x</summary>
+ 
+* July 8, 2020 (v1.10.11) 
+	* zstd: Fix extra block when compressing with ReadFrom. [#278](https://github.com/klauspost/compress/pull/278)
+	* huff0: Also populate compression table when reading decoding table. [#275](https://github.com/klauspost/compress/pull/275)
+	
+* June 23, 2020 (v1.10.10) 
+	* zstd: Skip entropy compression in fastest mode when no matches. [#270](https://github.com/klauspost/compress/pull/270)
+	
+* June 16, 2020 (v1.10.9): 
+	* zstd: API change for specifying dictionaries. See [#268](https://github.com/klauspost/compress/pull/268)
+	* zip: update CreateHeaderRaw to handle zip64 fields. [#266](https://github.com/klauspost/compress/pull/266)
+	* Fuzzit tests removed. The service has been purchased and is no longer available.
+	
+* June 5, 2020 (v1.10.8): 
+	* 1.15x faster zstd block decompression. [#265](https://github.com/klauspost/compress/pull/265)
+	
+* June 1, 2020 (v1.10.7): 
+	* Added zstd decompression [dictionary support](https://github.com/klauspost/compress/tree/master/zstd#dictionaries)
+	* Increase zstd decompression speed up to 1.19x.  [#259](https://github.com/klauspost/compress/pull/259)
+	* Remove internal reset call in zstd compression and reduce allocations. [#263](https://github.com/klauspost/compress/pull/263)
+	
+* May 21, 2020: (v1.10.6) 
+	* zstd: Reduce allocations while decoding. [#258](https://github.com/klauspost/compress/pull/258), [#252](https://github.com/klauspost/compress/pull/252)
+	* zstd: Stricter decompression checks.
+	
+* April 12, 2020: (v1.10.5)
+	* s2-commands: Flush output when receiving SIGINT. [#239](https://github.com/klauspost/compress/pull/239)
+	
+* Apr 8, 2020: (v1.10.4) 
+	* zstd: Minor/special case optimizations. [#251](https://github.com/klauspost/compress/pull/251),  [#250](https://github.com/klauspost/compress/pull/250),  [#249](https://github.com/klauspost/compress/pull/249),  [#247](https://github.com/klauspost/compress/pull/247)
+* Mar 11, 2020: (v1.10.3) 
+	* s2: Use S2 encoder in pure Go mode for Snappy output as well. [#245](https://github.com/klauspost/compress/pull/245)
+	* s2: Fix pure Go block encoder. [#244](https://github.com/klauspost/compress/pull/244)
+	* zstd: Added "better compression" mode. [#240](https://github.com/klauspost/compress/pull/240)
+	* zstd: Improve speed of fastest compression mode by 5-10% [#241](https://github.com/klauspost/compress/pull/241)
+	* zstd: Skip creating encoders when not needed. [#238](https://github.com/klauspost/compress/pull/238)
+	
+* Feb 27, 2020: (v1.10.2) 
+	* Close to 50% speedup in inflate (gzip/zip decompression). [#236](https://github.com/klauspost/compress/pull/236) [#234](https://github.com/klauspost/compress/pull/234) [#232](https://github.com/klauspost/compress/pull/232)
+	* Reduce deflate level 1-6 memory usage up to 59%. [#227](https://github.com/klauspost/compress/pull/227)
+	
+* Feb 18, 2020: (v1.10.1)
+	* Fix zstd crash when resetting multiple times without sending data. [#226](https://github.com/klauspost/compress/pull/226)
+	* deflate: Fix dictionary use on level 1-6. [#224](https://github.com/klauspost/compress/pull/224)
+	* Remove deflate writer reference when closing. [#224](https://github.com/klauspost/compress/pull/224)
+	
+* Feb 4, 2020: (v1.10.0) 
+	* Add optional dictionary to [stateless deflate](https://pkg.go.dev/github.com/klauspost/compress/flate?tab=doc#StatelessDeflate). Breaking change, send `nil` for previous behaviour. [#216](https://github.com/klauspost/compress/pull/216)
+	* Fix buffer overflow on repeated small block deflate.  [#218](https://github.com/klauspost/compress/pull/218)
+	* Allow copying content from an existing ZIP file without decompressing+compressing. [#214](https://github.com/klauspost/compress/pull/214)
+	* Added [S2](https://github.com/klauspost/compress/tree/master/s2#s2-compression) AMD64 assembler and various optimizations. Stream speed >10GB/s.  [#186](https://github.com/klauspost/compress/pull/186)
+
+</details>
+
+<details>
+	<summary>See changes prior to v1.10.0</summary>
+
+* Jan 20,2020 (v1.9.8) Optimize gzip/deflate with better size estimates and faster table generation. [#207](https://github.com/klauspost/compress/pull/207) by [luyu6056](https://github.com/luyu6056),  [#206](https://github.com/klauspost/compress/pull/206).
+* Jan 11, 2020: S2 Encode/Decode will use provided buffer if capacity is big enough. [#204](https://github.com/klauspost/compress/pull/204) 
+* Jan 5, 2020: (v1.9.7) Fix another zstd regression in v1.9.5 - v1.9.6 removed.
+* Jan 4, 2020: (v1.9.6) Regression in v1.9.5 fixed causing corrupt zstd encodes in rare cases.
+* Jan 4, 2020: Faster IO in [s2c + s2d commandline tools](https://github.com/klauspost/compress/tree/master/s2#commandline-tools) compression/decompression. [#192](https://github.com/klauspost/compress/pull/192)
+* Dec 29, 2019: Removed v1.9.5 since fuzz tests showed a compatibility problem with the reference zstandard decoder.
+* Dec 29, 2019: (v1.9.5) zstd: 10-20% faster block compression. [#199](https://github.com/klauspost/compress/pull/199)
+* Dec 29, 2019: [zip](https://godoc.org/github.com/klauspost/compress/zip) package updated with latest Go features
+* Dec 29, 2019: zstd: Single segment flag condintions tweaked. [#197](https://github.com/klauspost/compress/pull/197)
+* Dec 18, 2019: s2: Faster compression when ReadFrom is used. [#198](https://github.com/klauspost/compress/pull/198)
+* Dec 10, 2019: s2: Fix repeat length output when just above at 16MB limit.
+* Dec 10, 2019: zstd: Add function to get decoder as io.ReadCloser. [#191](https://github.com/klauspost/compress/pull/191)
+* Dec 3, 2019: (v1.9.4) S2: limit max repeat length. [#188](https://github.com/klauspost/compress/pull/188)
+* Dec 3, 2019: Add [WithNoEntropyCompression](https://godoc.org/github.com/klauspost/compress/zstd#WithNoEntropyCompression) to zstd [#187](https://github.com/klauspost/compress/pull/187)
+* Dec 3, 2019: Reduce memory use for tests. Check for leaked goroutines.
+* Nov 28, 2019 (v1.9.3) Less allocations in stateless deflate.
+* Nov 28, 2019: 5-20% Faster huff0 decode. Impacts zstd as well. [#184](https://github.com/klauspost/compress/pull/184)
+* Nov 12, 2019 (v1.9.2) Added [Stateless Compression](#stateless-compression) for gzip/deflate.
+* Nov 12, 2019: Fixed zstd decompression of large single blocks. [#180](https://github.com/klauspost/compress/pull/180)
+* Nov 11, 2019: Set default  [s2c](https://github.com/klauspost/compress/tree/master/s2#commandline-tools) block size to 4MB.
+* Nov 11, 2019: Reduce inflate memory use by 1KB.
+* Nov 10, 2019: Less allocations in deflate bit writer.
+* Nov 10, 2019: Fix inconsistent error returned by zstd decoder.
+* Oct 28, 2019 (v1.9.1) ztsd: Fix crash when compressing blocks. [#174](https://github.com/klauspost/compress/pull/174)
+* Oct 24, 2019 (v1.9.0) zstd: Fix rare data corruption [#173](https://github.com/klauspost/compress/pull/173)
+* Oct 24, 2019 zstd: Fix huff0 out of buffer write [#171](https://github.com/klauspost/compress/pull/171) and always return errors [#172](https://github.com/klauspost/compress/pull/172) 
+* Oct 10, 2019: Big deflate rewrite, 30-40% faster with better compression [#105](https://github.com/klauspost/compress/pull/105)
+
+</details>
+
+<details>
+	<summary>See changes prior to v1.9.0</summary>
+
+* Oct 10, 2019: (v1.8.6) zstd: Allow partial reads to get flushed data. [#169](https://github.com/klauspost/compress/pull/169)
+* Oct 3, 2019: Fix inconsistent results on broken zstd streams.
+* Sep 25, 2019: Added `-rm` (remove source files) and `-q` (no output except errors) to `s2c` and `s2d` [commands](https://github.com/klauspost/compress/tree/master/s2#commandline-tools)
+* Sep 16, 2019: (v1.8.4) Add `s2c` and `s2d` [commandline tools](https://github.com/klauspost/compress/tree/master/s2#commandline-tools).
+* Sep 10, 2019: (v1.8.3) Fix s2 decoder [Skip](https://godoc.org/github.com/klauspost/compress/s2#Reader.Skip).
+* Sep 7, 2019: zstd: Added [WithWindowSize](https://godoc.org/github.com/klauspost/compress/zstd#WithWindowSize), contributed by [ianwilkes](https://github.com/ianwilkes).
+* Sep 5, 2019: (v1.8.2) Add [WithZeroFrames](https://godoc.org/github.com/klauspost/compress/zstd#WithZeroFrames) which adds full zero payload block encoding option.
+* Sep 5, 2019: Lazy initialization of zstandard predefined en/decoder tables.
+* Aug 26, 2019: (v1.8.1) S2: 1-2% compression increase in "better" compression mode.
+* Aug 26, 2019: zstd: Check maximum size of Huffman 1X compressed literals while decoding.
+* Aug 24, 2019: (v1.8.0) Added [S2 compression](https://github.com/klauspost/compress/tree/master/s2#s2-compression), a high performance replacement for Snappy. 
+* Aug 21, 2019: (v1.7.6) Fixed minor issues found by fuzzer. One could lead to zstd not decompressing.
+* Aug 18, 2019: Add [fuzzit](https://fuzzit.dev/) continuous fuzzing.
+* Aug 14, 2019: zstd: Skip incompressible data 2x faster.  [#147](https://github.com/klauspost/compress/pull/147)
+* Aug 4, 2019 (v1.7.5): Better literal compression. [#146](https://github.com/klauspost/compress/pull/146)
+* Aug 4, 2019: Faster zstd compression. [#143](https://github.com/klauspost/compress/pull/143) [#144](https://github.com/klauspost/compress/pull/144)
+* Aug 4, 2019: Faster zstd decompression. [#145](https://github.com/klauspost/compress/pull/145) [#143](https://github.com/klauspost/compress/pull/143) [#142](https://github.com/klauspost/compress/pull/142)
+* July 15, 2019 (v1.7.4): Fix double EOF block in rare cases on zstd encoder.
+* July 15, 2019 (v1.7.3): Minor speedup/compression increase in default zstd encoder.
+* July 14, 2019: zstd decoder: Fix decompression error on multiple uses with mixed content.
+* July 7, 2019 (v1.7.2): Snappy update, zstd decoder potential race fix.
+* June 17, 2019: zstd decompression bugfix.
+* June 17, 2019: fix 32 bit builds.
+* June 17, 2019: Easier use in modules (less dependencies).
+* June 9, 2019: New stronger "default" [zstd](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression mode. Matches zstd default compression ratio.
+* June 5, 2019: 20-40% throughput in [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and better compression.
+* June 5, 2019: deflate/gzip compression: Reduce memory usage of lower compression levels.
+* June 2, 2019: Added [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression!
+* May 25, 2019: deflate/gzip: 10% faster bit writer, mostly visible in lower levels.
+* Apr 22, 2019: [zstd](https://github.com/klauspost/compress/tree/master/zstd#zstd) decompression added.
+* Aug 1, 2018: Added [huff0 README](https://github.com/klauspost/compress/tree/master/huff0#huff0-entropy-compression).
+* Jul 8, 2018: Added [Performance Update 2018](#performance-update-2018) below.
+* Jun 23, 2018: Merged [Go 1.11 inflate optimizations](https://go-review.googlesource.com/c/go/+/102235). Go 1.9 is now required. Backwards compatible version tagged with [v1.3.0](https://github.com/klauspost/compress/releases/tag/v1.3.0).
+* Apr 2, 2018: Added [huff0](https://godoc.org/github.com/klauspost/compress/huff0) en/decoder. Experimental for now, API may change.
+* Mar 4, 2018: Added [FSE Entropy](https://godoc.org/github.com/klauspost/compress/fse) en/decoder. Experimental for now, API may change.
+* Nov 3, 2017: Add compression [Estimate](https://godoc.org/github.com/klauspost/compress#Estimate) function.
+* May 28, 2017: Reduce allocations when resetting decoder.
+* Apr 02, 2017: Change back to official crc32, since changes were merged in Go 1.7.
+* Jan 14, 2017: Reduce stack pressure due to array copies. See [Issue #18625](https://github.com/golang/go/issues/18625).
+* Oct 25, 2016: Level 2-4 have been rewritten and now offers significantly better performance than before.
+* Oct 20, 2016: Port zlib changes from Go 1.7 to fix zlib writer issue. Please update.
+* Oct 16, 2016: Go 1.7 changes merged. Apples to apples this package is a few percent faster, but has a significantly better balance between speed and compression per level. 
+* Mar 24, 2016: Always attempt Huffman encoding on level 4-7. This improves base 64 encoded data compression.
+* Mar 24, 2016: Small speedup for level 1-3.
+* Feb 19, 2016: Faster bit writer, level -2 is 15% faster, level 1 is 4% faster.
+* Feb 19, 2016: Handle small payloads faster in level 1-3.
+* Feb 19, 2016: Added faster level 2 + 3 compression modes.
+* Feb 19, 2016: [Rebalanced compression levels](https://blog.klauspost.com/rebalancing-deflate-compression-levels/), so there is a more even progresssion in terms of compression. New default level is 5.
+* Feb 14, 2016: Snappy: Merge upstream changes. 
+* Feb 14, 2016: Snappy: Fix aggressive skipping.
+* Feb 14, 2016: Snappy: Update benchmark.
+* Feb 13, 2016: Deflate: Fixed assembler problem that could lead to sub-optimal compression.
+* Feb 12, 2016: Snappy: Added AMD64 SSE 4.2 optimizations to matching, which makes easy to compress material run faster. Typical speedup is around 25%.
+* Feb 9, 2016: Added Snappy package fork. This version is 5-7% faster, much more on hard to compress content.
+* Jan 30, 2016: Optimize level 1 to 3 by not considering static dictionary or storing uncompressed. ~4-5% speedup.
+* Jan 16, 2016: Optimization on deflate level 1,2,3 compression.
+* Jan 8 2016: Merge [CL 18317](https://go-review.googlesource.com/#/c/18317): fix reading, writing of zip64 archives.
+* Dec 8 2015: Make level 1 and -2 deterministic even if write size differs.
+* Dec 8 2015: Split encoding functions, so hashing and matching can potentially be inlined. 1-3% faster on AMD64. 5% faster on other platforms.
+* Dec 8 2015: Fixed rare [one byte out-of bounds read](https://github.com/klauspost/compress/issues/20). Please update!
+* Nov 23 2015: Optimization on token writer. ~2-4% faster. Contributed by [@dsnet](https://github.com/dsnet).
+* Nov 20 2015: Small optimization to bit writer on 64 bit systems.
+* Nov 17 2015: Fixed out-of-bound errors if the underlying Writer returned an error. See [#15](https://github.com/klauspost/compress/issues/15).
+* Nov 12 2015: Added [io.WriterTo](https://golang.org/pkg/io/#WriterTo) support to gzip/inflate.
+* Nov 11 2015: Merged [CL 16669](https://go-review.googlesource.com/#/c/16669/4): archive/zip: enable overriding (de)compressors per file
+* Oct 15 2015: Added skipping on uncompressible data. Random data speed up >5x.
+
+</details>
+
+# deflate usage
+
+The packages are drop-in replacements for standard libraries. Simply replace the import path to use them:
+
+| old import         | new import                              | Documentation
+|--------------------|-----------------------------------------|--------------------|
+| `compress/gzip`    | `github.com/klauspost/compress/gzip`    | [gzip](https://pkg.go.dev/github.com/klauspost/compress/gzip?tab=doc)
+| `compress/zlib`    | `github.com/klauspost/compress/zlib`    | [zlib](https://pkg.go.dev/github.com/klauspost/compress/zlib?tab=doc)
+| `archive/zip`      | `github.com/klauspost/compress/zip`     | [zip](https://pkg.go.dev/github.com/klauspost/compress/zip?tab=doc)
+| `compress/flate`   | `github.com/klauspost/compress/flate`   | [flate](https://pkg.go.dev/github.com/klauspost/compress/flate?tab=doc)
+
+* Optimized [deflate](https://godoc.org/github.com/klauspost/compress/flate) packages which can be used as a dropin replacement for [gzip](https://godoc.org/github.com/klauspost/compress/gzip), [zip](https://godoc.org/github.com/klauspost/compress/zip) and [zlib](https://godoc.org/github.com/klauspost/compress/zlib).
+
+You may also be interested in [pgzip](https://github.com/klauspost/pgzip), which is a drop in replacement for gzip, which support multithreaded compression on big files and the optimized [crc32](https://github.com/klauspost/crc32) package used by these packages.
+
+The packages contains the same as the standard library, so you can use the godoc for that: [gzip](http://golang.org/pkg/compress/gzip/), [zip](http://golang.org/pkg/archive/zip/),  [zlib](http://golang.org/pkg/compress/zlib/), [flate](http://golang.org/pkg/compress/flate/).
+
+Currently there is only minor speedup on decompression (mostly CRC32 calculation).
+
+Memory usage is typically 1MB for a Writer. stdlib is in the same range. 
+If you expect to have a lot of concurrently allocated Writers consider using 
+the stateless compress described below.
+
+For compression performance, see: [this spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing).
+
+# Stateless compression
+
+This package offers stateless compression as a special option for gzip/deflate. 
+It will do compression but without maintaining any state between Write calls.
+
+This means there will be no memory kept between Write calls, but compression and speed will be suboptimal.
+
+This is only relevant in cases where you expect to run many thousands of compressors concurrently, 
+but with very little activity. This is *not* intended for regular web servers serving individual requests.  
+
+Because of this, the size of actual Write calls will affect output size.
+
+In gzip, specify level `-3` / `gzip.StatelessCompression` to enable.
+
+For direct deflate use, NewStatelessWriter and StatelessDeflate are available. See [documentation](https://godoc.org/github.com/klauspost/compress/flate#NewStatelessWriter)
+
+A `bufio.Writer` can of course be used to control write sizes. For example, to use a 4KB buffer:
+
+```
+	// replace 'ioutil.Discard' with your output.
+	gzw, err := gzip.NewWriterLevel(ioutil.Discard, gzip.StatelessCompression)
+	if err != nil {
+		return err
+	}
+	defer gzw.Close()
+
+	w := bufio.NewWriterSize(gzw, 4096)
+	defer w.Flush()
+	
+	// Write to 'w' 
+```
+
+This will only use up to 4KB in memory when the writer is idle. 
+
+Compression is almost always worse than the fastest compression level 
+and each write will allocate (a little) memory. 
+
+# Performance Update 2018
+
+It has been a while since we have been looking at the speed of this package compared to the standard library, so I thought I would re-do my tests and give some overall recommendations based on the current state. All benchmarks have been performed with Go 1.10 on my Desktop Intel(R) Core(TM) i7-2600 CPU @3.40GHz. Since I last ran the tests, I have gotten more RAM, which means tests with big files are no longer limited by my SSD.
+
+The raw results are in my [updated spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing). Due to cgo changes and upstream updates i could not get the cgo version of gzip to compile. Instead I included the [zstd](https://github.com/datadog/zstd) cgo implementation. If I get cgo gzip to work again, I might replace the results in the sheet.
+
+The columns to take note of are: *MB/s* - the throughput. *Reduction* - the data size reduction in percent of the original. *Rel Speed* relative speed compared to the standard library at the same level. *Smaller* - how many percent smaller is the compressed output compared to stdlib. Negative means the output was bigger. *Loss* means the loss (or gain) in compression as a percentage difference of the input.
+
+The `gzstd` (standard library gzip) and `gzkp` (this package gzip) only uses one CPU core. [`pgzip`](https://github.com/klauspost/pgzip), [`bgzf`](https://github.com/biogo/hts/tree/master/bgzf) uses all 4 cores. [`zstd`](https://github.com/DataDog/zstd) uses one core, and is a beast (but not Go, yet).
+
+
+## Overall differences.
+
+There appears to be a roughly 5-10% speed advantage over the standard library when comparing at similar compression levels.
+
+The biggest difference you will see is the result of [re-balancing](https://blog.klauspost.com/rebalancing-deflate-compression-levels/) the compression levels. I wanted by library to give a smoother transition between the compression levels than the standard library.
+
+This package attempts to provide a more smooth transition, where "1" is taking a lot of shortcuts, "5" is the reasonable trade-off and "9" is the "give me the best compression", and the values in between gives something reasonable in between. The standard library has big differences in levels 1-4, but levels 5-9 having no significant gains - often spending a lot more time than can be justified by the achieved compression.
+
+There are links to all the test data in the [spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing) in the top left field on each tab.
+
+## Web Content
+
+This test set aims to emulate typical use in a web server. The test-set is 4GB data in 53k files, and is a mixture of (mostly) HTML, JS, CSS.
+
+Since level 1 and 9 are close to being the same code, they are quite close. But looking at the levels in-between the differences are quite big.
+
+Looking at level 6, this package is 88% faster, but will output about 6% more data. For a web server, this means you can serve 88% more data, but have to pay for 6% more bandwidth. You can draw your own conclusions on what would be the most expensive for your case.
+
+## Object files
+
+This test is for typical data files stored on a server. In this case it is a collection of Go precompiled objects. They are very compressible.
+
+The picture is similar to the web content, but with small differences since this is very compressible. Levels 2-3 offer good speed, but is sacrificing quite a bit of compression. 
+
+The standard library seems suboptimal on level 3 and 4 - offering both worse compression and speed than level 6 & 7 of this package respectively.
+
+## Highly Compressible File
+
+This is a JSON file with very high redundancy. The reduction starts at 95% on level 1, so in real life terms we are dealing with something like a highly redundant stream of data, etc.
+
+It is definitely visible that we are dealing with specialized content here, so the results are very scattered. This package does not do very well at levels 1-4, but picks up significantly at level 5 and levels 7 and 8 offering great speed for the achieved compression.
+
+So if you know you content is extremely compressible you might want to go slightly higher than the defaults. The standard library has a huge gap between levels 3 and 4 in terms of speed (2.75x slowdown), so it offers little "middle ground".
+
+## Medium-High Compressible
+
+This is a pretty common test corpus: [enwik9](http://mattmahoney.net/dc/textdata.html). It contains the first 10^9 bytes of the English Wikipedia dump on Mar. 3, 2006. This is a very good test of typical text based compression and more data heavy streams.
+
+We see a similar picture here as in "Web Content". On equal levels some compression is sacrificed for more speed. Level 5 seems to be the best trade-off between speed and size, beating stdlib level 3 in both.
+
+## Medium Compressible
+
+I will combine two test sets, one [10GB file set](http://mattmahoney.net/dc/10gb.html) and a VM disk image (~8GB). Both contain different data types and represent a typical backup scenario.
+
+The most notable thing is how quickly the standard library drops to very low compression speeds around level 5-6 without any big gains in compression. Since this type of data is fairly common, this does not seem like good behavior.
+
+
+## Un-compressible Content
+
+This is mainly a test of how good the algorithms are at detecting un-compressible input. The standard library only offers this feature with very conservative settings at level 1. Obviously there is no reason for the algorithms to try to compress input that cannot be compressed.  The only downside is that it might skip some compressible data on false detections.
+
+
+## Huffman only compression
+
+This compression library adds a special compression level, named `HuffmanOnly`, which allows near linear time compression. This is done by completely disabling matching of previous data, and only reduce the number of bits to represent each character. 
+
+This means that often used characters, like 'e' and ' ' (space) in text use the fewest bits to represent, and rare characters like '¤' takes more bits to represent. For more information see [wikipedia](https://en.wikipedia.org/wiki/Huffman_coding) or this nice [video](https://youtu.be/ZdooBTdW5bM).
+
+Since this type of compression has much less variance, the compression speed is mostly unaffected by the input data, and is usually more than *180MB/s* for a single core.
+
+The downside is that the compression ratio is usually considerably worse than even the fastest conventional compression. The compression ratio can never be better than 8:1 (12.5%). 
+
+The linear time compression can be used as a "better than nothing" mode, where you cannot risk the encoder to slow down on some content. For comparison, the size of the "Twain" text is *233460 bytes* (+29% vs. level 1) and encode speed is 144MB/s (4.5x level 1). So in this case you trade a 30% size increase for a 4 times speedup.
+
+For more information see my blog post on [Fast Linear Time Compression](http://blog.klauspost.com/constant-time-gzipzip-compression/).
+
+This is implemented on Go 1.7 as "Huffman Only" mode, though not exposed for gzip.
+
+# Other packages
+
+Here are other packages of good quality and pure Go (no cgo wrappers or autoconverted code):
+
+* [github.com/pierrec/lz4](https://github.com/pierrec/lz4) - strong multithreaded LZ4 compression.
+* [github.com/cosnicolaou/pbzip2](https://github.com/cosnicolaou/pbzip2) - multithreaded bzip2 decompression.
+* [github.com/dsnet/compress](https://github.com/dsnet/compress) - brotli decompression, bzip2 writer.
+
+# license
+
+This code is licensed under the same conditions as the original Go code. See LICENSE file.

vendor/github.com/klauspost/compress/compressible.go

@@ -0,0 +1,85 @@
+package compress
+
+import "math"
+
+// Estimate returns a normalized compressibility estimate of block b.
+// Values close to zero are likely uncompressible.
+// Values above 0.1 are likely to be compressible.
+// Values above 0.5 are very compressible.
+// Very small lengths will return 0.
+func Estimate(b []byte) float64 {
+	if len(b) < 16 {
+		return 0
+	}
+
+	// Correctly predicted order 1
+	hits := 0
+	lastMatch := false
+	var o1 [256]byte
+	var hist [256]int
+	c1 := byte(0)
+	for _, c := range b {
+		if c == o1[c1] {
+			// We only count a hit if there was two correct predictions in a row.
+			if lastMatch {
+				hits++
+			}
+			lastMatch = true
+		} else {
+			lastMatch = false
+		}
+		o1[c1] = c
+		c1 = c
+		hist[c]++
+	}
+
+	// Use x^0.6 to give better spread
+	prediction := math.Pow(float64(hits)/float64(len(b)), 0.6)
+
+	// Calculate histogram distribution
+	variance := float64(0)
+	avg := float64(len(b)) / 256
+
+	for _, v := range hist {
+		Δ := float64(v) - avg
+		variance += Δ * Δ
+	}
+
+	stddev := math.Sqrt(float64(variance)) / float64(len(b))
+	exp := math.Sqrt(1 / float64(len(b)))
+
+	// Subtract expected stddev
+	stddev -= exp
+	if stddev < 0 {
+		stddev = 0
+	}
+	stddev *= 1 + exp
+
+	// Use x^0.4 to give better spread
+	entropy := math.Pow(stddev, 0.4)
+
+	// 50/50 weight between prediction and histogram distribution
+	return math.Pow((prediction+entropy)/2, 0.9)
+}
+
+// ShannonEntropyBits returns the number of bits minimum required to represent
+// an entropy encoding of the input bytes.
+// https://en.wiktionary.org/wiki/Shannon_entropy
+func ShannonEntropyBits(b []byte) int {
+	if len(b) == 0 {
+		return 0
+	}
+	var hist [256]int
+	for _, c := range b {
+		hist[c]++
+	}
+	shannon := float64(0)
+	invTotal := 1.0 / float64(len(b))
+	for _, v := range hist[:] {
+		if v > 0 {
+			n := float64(v)
+			shannon += math.Ceil(-math.Log2(n*invTotal) * n)
+		}
+	}
+	return int(math.Ceil(shannon))
+}

vendor/github.com/klauspost/compress/fse/README.md

@@ -0,0 +1,79 @@
+# Finite State Entropy
+
+This package provides Finite State Entropy encoding and decoding.
+            
+Finite State Entropy (also referenced as [tANS](https://en.wikipedia.org/wiki/Asymmetric_numeral_systems#tANS)) 
+encoding provides a fast near-optimal symbol encoding/decoding
+for byte blocks as implemented in [zstandard](https://github.com/facebook/zstd).
+
+This can be used for compressing input with a lot of similar input values to the smallest number of bytes.
+This does not perform any multi-byte [dictionary coding](https://en.wikipedia.org/wiki/Dictionary_coder) as LZ coders,
+but it can be used as a secondary step to compressors (like Snappy) that does not do entropy encoding. 
+
+* [Godoc documentation](https://godoc.org/github.com/klauspost/compress/fse)
+
+## News
+
+ * Feb 2018: First implementation released. Consider this beta software for now.
+
+# Usage
+
+This package provides a low level interface that allows to compress single independent blocks. 
+
+Each block is separate, and there is no built in integrity checks. 
+This means that the caller should keep track of block sizes and also do checksums if needed.  
+
+Compressing a block is done via the [`Compress`](https://godoc.org/github.com/klauspost/compress/fse#Compress) function.
+You must provide input and will receive the output and maybe an error.
+
+These error values can be returned:
+
+| Error               | Description                                                                 |
+|---------------------|-----------------------------------------------------------------------------|
+| `<nil>`             | Everything ok, output is returned                                           |
+| `ErrIncompressible` | Returned when input is judged to be too hard to compress                    |
+| `ErrUseRLE`         | Returned from the compressor when the input is a single byte value repeated |
+| `(error)`           | An internal error occurred.                                                 |
+
+As can be seen above there are errors that will be returned even under normal operation so it is important to handle these.
+
+To reduce allocations you can provide a [`Scratch`](https://godoc.org/github.com/klauspost/compress/fse#Scratch) object 
+that can be re-used for successive calls. Both compression and decompression accepts a `Scratch` object, and the same 
+object can be used for both.   
+
+Be aware, that when re-using a `Scratch` object that the *output* buffer is also re-used, so if you are still using this
+you must set the `Out` field in the scratch to nil. The same buffer is used for compression and decompression output.
+
+Decompressing is done by calling the [`Decompress`](https://godoc.org/github.com/klauspost/compress/fse#Decompress) function.
+You must provide the output from the compression stage, at exactly the size you got back. If you receive an error back
+your input was likely corrupted. 
+
+It is important to note that a successful decoding does *not* mean your output matches your original input. 
+There are no integrity checks, so relying on errors from the decompressor does not assure your data is valid.
+
+For more detailed usage, see examples in the [godoc documentation](https://godoc.org/github.com/klauspost/compress/fse#pkg-examples).
+
+# Performance
+
+A lot of factors are affecting speed. Block sizes and compressibility of the material are primary factors.  
+All compression functions are currently only running on the calling goroutine so only one core will be used per block.  
+
+The compressor is significantly faster if symbols are kept as small as possible. The highest byte value of the input
+is used to reduce some of the processing, so if all your input is above byte value 64 for instance, it may be 
+beneficial to transpose all your input values down by 64.   
+
+With moderate block sizes around 64k speed are typically 200MB/s per core for compression and 
+around 300MB/s decompression speed. 
+
+The same hardware typically does Huffman (deflate) encoding at 125MB/s and decompression at 100MB/s. 
+
+# Plans
+
+At one point, more internals will be exposed to facilitate more "expert" usage of the components. 
+
+A streaming interface is also likely to be implemented. Likely compatible with [FSE stream format](https://github.com/Cyan4973/FiniteStateEntropy/blob/dev/programs/fileio.c#L261).  
+
+# Contributing
+
+Contributions are always welcome. Be aware that adding public functions will require good justification and breaking 
+changes will likely not be accepted. If in doubt open an issue before writing the PR.  

vendor/github.com/klauspost/compress/fse/bitreader.go

@@ -0,0 +1,122 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package fse
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// bitReader reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReader struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReader) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.bitsRead += 8 - uint8(highBits(uint32(v)))
+	return nil
+}
+
+// getBits will return n bits. n can be 0.
+func (b *bitReader) getBits(n uint8) uint16 {
+	if n == 0 || b.bitsRead >= 64 {
+		return 0
+	}
+	return b.getBitsFast(n)
+}
+
+// getBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReader) getBitsFast(n uint8) uint16 {
+	const regMask = 64 - 1
+	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
+	b.bitsRead += n
+	return v
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReader) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+	// 2 bounds checks.
+	v := b.in[b.off-4:]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value = (b.value << 32) | uint64(low)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReader) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value = (b.value << 32) | uint64(low)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value = (b.value << 8) | uint64(b.in[b.off-1])
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
+func (b *bitReader) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReader) finished() bool {
+	return b.bitsRead >= 64 && b.off == 0
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReader) close() error {
+	// Release reference.
+	b.in = nil
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}

vendor/github.com/klauspost/compress/fse/bitwriter.go

@@ -0,0 +1,168 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package fse
+
+import "fmt"
+
+// bitWriter will write bits.
+// First bit will be LSB of the first byte of output.
+type bitWriter struct {
+	bitContainer uint64
+	nBits        uint8
+	out          []byte
+}
+
+// bitMask16 is bitmasks. Has extra to avoid bounds check.
+var bitMask16 = [32]uint16{
+	0, 1, 3, 7, 0xF, 0x1F,
+	0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
+	0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF} /* up to 16 bits */
+
+// addBits16NC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16NC(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16ZeroNC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+// This is fastest if bits can be zero.
+func (b *bitWriter) addBits16ZeroNC(value uint16, bits uint8) {
+	if bits == 0 {
+		return
+	}
+	value <<= (16 - bits) & 15
+	value >>= (16 - bits) & 15
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// flush will flush all pending full bytes.
+// There will be at least 56 bits available for writing when this has been called.
+// Using flush32 is faster, but leaves less space for writing.
+func (b *bitWriter) flush() {
+	v := b.nBits >> 3
+	switch v {
+	case 0:
+	case 1:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+		)
+	case 2:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+		)
+	case 3:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+		)
+	case 4:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+		)
+	case 5:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+		)
+	case 6:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+		)
+	case 7:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+		)
+	case 8:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+			byte(b.bitContainer>>56),
+		)
+	default:
+		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
+	}
+	b.bitContainer >>= v << 3
+	b.nBits &= 7
+}
+
+// flush32 will flush out, so there are at least 32 bits available for writing.
+func (b *bitWriter) flush32() {
+	if b.nBits < 32 {
+		return
+	}
+	b.out = append(b.out,
+		byte(b.bitContainer),
+		byte(b.bitContainer>>8),
+		byte(b.bitContainer>>16),
+		byte(b.bitContainer>>24))
+	b.nBits -= 32
+	b.bitContainer >>= 32
+}
+
+// flushAlign will flush remaining full bytes and align to next byte boundary.
+func (b *bitWriter) flushAlign() {
+	nbBytes := (b.nBits + 7) >> 3
+	for i := uint8(0); i < nbBytes; i++ {
+		b.out = append(b.out, byte(b.bitContainer>>(i*8)))
+	}
+	b.nBits = 0
+	b.bitContainer = 0
+}
+
+// close will write the alignment bit and write the final byte(s)
+// to the output.
+func (b *bitWriter) close() error {
+	// End mark
+	b.addBits16Clean(1, 1)
+	// flush until next byte.
+	b.flushAlign()
+	return nil
+}
+
+// reset and continue writing by appending to out.
+func (b *bitWriter) reset(out []byte) {
+	b.bitContainer = 0
+	b.nBits = 0
+	b.out = out
+}

vendor/github.com/klauspost/compress/fse/bytereader.go

@@ -0,0 +1,47 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package fse
+
+// byteReader provides a byte reader that reads
+// little endian values from a byte stream.
+// The input stream is manually advanced.
+// The reader performs no bounds checks.
+type byteReader struct {
+	b   []byte
+	off int
+}
+
+// init will initialize the reader and set the input.
+func (b *byteReader) init(in []byte) {
+	b.b = in
+	b.off = 0
+}
+
+// advance the stream b n bytes.
+func (b *byteReader) advance(n uint) {
+	b.off += int(n)
+}
+
+// Uint32 returns a little endian uint32 starting at current offset.
+func (b byteReader) Uint32() uint32 {
+	b2 := b.b[b.off:]
+	b2 = b2[:4]
+	v3 := uint32(b2[3])
+	v2 := uint32(b2[2])
+	v1 := uint32(b2[1])
+	v0 := uint32(b2[0])
+	return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24)
+}
+
+// unread returns the unread portion of the input.
+func (b byteReader) unread() []byte {
+	return b.b[b.off:]
+}
+
+// remain will return the number of bytes remaining.
+func (b byteReader) remain() int {
+	return len(b.b) - b.off
+}

vendor/github.com/klauspost/compress/fse/compress.go

@@ -0,0 +1,683 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package fse
+
+import (
+	"errors"
+	"fmt"
+)
+
+// Compress the input bytes. Input must be < 2GB.
+// Provide a Scratch buffer to avoid memory allocations.
+// Note that the output is also kept in the scratch buffer.
+// If input is too hard to compress, ErrIncompressible is returned.
+// If input is a single byte value repeated ErrUseRLE is returned.
+func Compress(in []byte, s *Scratch) ([]byte, error) {
+	if len(in) <= 1 {
+		return nil, ErrIncompressible
+	}
+	if len(in) > (2<<30)-1 {
+		return nil, errors.New("input too big, must be < 2GB")
+	}
+	s, err := s.prepare(in)
+	if err != nil {
+		return nil, err
+	}
+
+	// Create histogram, if none was provided.
+	maxCount := s.maxCount
+	if maxCount == 0 {
+		maxCount = s.countSimple(in)
+	}
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount == len(in) {
+		// One symbol, use RLE
+		return nil, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return nil, ErrIncompressible
+	}
+	s.optimalTableLog()
+	err = s.normalizeCount()
+	if err != nil {
+		return nil, err
+	}
+	err = s.writeCount()
+	if err != nil {
+		return nil, err
+	}
+
+	if false {
+		err = s.validateNorm()
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	err = s.buildCTable()
+	if err != nil {
+		return nil, err
+	}
+	err = s.compress(in)
+	if err != nil {
+		return nil, err
+	}
+	s.Out = s.bw.out
+	// Check if we compressed.
+	if len(s.Out) >= len(in) {
+		return nil, ErrIncompressible
+	}
+	return s.Out, nil
+}
+
+// cState contains the compression state of a stream.
+type cState struct {
+	bw         *bitWriter
+	stateTable []uint16
+	state      uint16
+}
+
+// init will initialize the compression state to the first symbol of the stream.
+func (c *cState) init(bw *bitWriter, ct *cTable, tableLog uint8, first symbolTransform) {
+	c.bw = bw
+	c.stateTable = ct.stateTable
+
+	nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16
+	im := int32((nbBitsOut << 16) - first.deltaNbBits)
+	lu := (im >> nbBitsOut) + first.deltaFindState
+	c.state = c.stateTable[lu]
+}
+
+// encode the output symbol provided and write it to the bitstream.
+func (c *cState) encode(symbolTT symbolTransform) {
+	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
+	dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState
+	c.bw.addBits16NC(c.state, uint8(nbBitsOut))
+	c.state = c.stateTable[dstState]
+}
+
+// encode the output symbol provided and write it to the bitstream.
+func (c *cState) encodeZero(symbolTT symbolTransform) {
+	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
+	dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState
+	c.bw.addBits16ZeroNC(c.state, uint8(nbBitsOut))
+	c.state = c.stateTable[dstState]
+}
+
+// flush will write the tablelog to the output and flush the remaining full bytes.
+func (c *cState) flush(tableLog uint8) {
+	c.bw.flush32()
+	c.bw.addBits16NC(c.state, tableLog)
+	c.bw.flush()
+}
+
+// compress is the main compression loop that will encode the input from the last byte to the first.
+func (s *Scratch) compress(src []byte) error {
+	if len(src) <= 2 {
+		return errors.New("compress: src too small")
+	}
+	tt := s.ct.symbolTT[:256]
+	s.bw.reset(s.Out)
+
+	// Our two states each encodes every second byte.
+	// Last byte encoded (first byte decoded) will always be encoded by c1.
+	var c1, c2 cState
+
+	// Encode so remaining size is divisible by 4.
+	ip := len(src)
+	if ip&1 == 1 {
+		c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]])
+		c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]])
+		c1.encodeZero(tt[src[ip-3]])
+		ip -= 3
+	} else {
+		c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]])
+		c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]])
+		ip -= 2
+	}
+	if ip&2 != 0 {
+		c2.encodeZero(tt[src[ip-1]])
+		c1.encodeZero(tt[src[ip-2]])
+		ip -= 2
+	}
+
+	// Main compression loop.
+	switch {
+	case !s.zeroBits && s.actualTableLog <= 8:
+		// We can encode 4 symbols without requiring a flush.
+		// We do not need to check if any output is 0 bits.
+		for ip >= 4 {
+			s.bw.flush32()
+			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			c2.encode(tt[v0])
+			c1.encode(tt[v1])
+			c2.encode(tt[v2])
+			c1.encode(tt[v3])
+			ip -= 4
+		}
+	case !s.zeroBits:
+		// We do not need to check if any output is 0 bits.
+		for ip >= 4 {
+			s.bw.flush32()
+			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			c2.encode(tt[v0])
+			c1.encode(tt[v1])
+			s.bw.flush32()
+			c2.encode(tt[v2])
+			c1.encode(tt[v3])
+			ip -= 4
+		}
+	case s.actualTableLog <= 8:
+		// We can encode 4 symbols without requiring a flush
+		for ip >= 4 {
+			s.bw.flush32()
+			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			c2.encodeZero(tt[v0])
+			c1.encodeZero(tt[v1])
+			c2.encodeZero(tt[v2])
+			c1.encodeZero(tt[v3])
+			ip -= 4
+		}
+	default:
+		for ip >= 4 {
+			s.bw.flush32()
+			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			c2.encodeZero(tt[v0])
+			c1.encodeZero(tt[v1])
+			s.bw.flush32()
+			c2.encodeZero(tt[v2])
+			c1.encodeZero(tt[v3])
+			ip -= 4
+		}
+	}
+
+	// Flush final state.
+	// Used to initialize state when decoding.
+	c2.flush(s.actualTableLog)
+	c1.flush(s.actualTableLog)
+
+	return s.bw.close()
+}
+
+// writeCount will write the normalized histogram count to header.
+// This is read back by readNCount.
+func (s *Scratch) writeCount() error {
+	var (
+		tableLog  = s.actualTableLog
+		tableSize = 1 << tableLog
+		previous0 bool
+		charnum   uint16
+
+		maxHeaderSize = ((int(s.symbolLen) * int(tableLog)) >> 3) + 3
+
+		// Write Table Size
+		bitStream = uint32(tableLog - minTablelog)
+		bitCount  = uint(4)
+		remaining = int16(tableSize + 1) /* +1 for extra accuracy */
+		threshold = int16(tableSize)
+		nbBits    = uint(tableLog + 1)
+	)
+	if cap(s.Out) < maxHeaderSize {
+		s.Out = make([]byte, 0, s.br.remain()+maxHeaderSize)
+	}
+	outP := uint(0)
+	out := s.Out[:maxHeaderSize]
+
+	// stops at 1
+	for remaining > 1 {
+		if previous0 {
+			start := charnum
+			for s.norm[charnum] == 0 {
+				charnum++
+			}
+			for charnum >= start+24 {
+				start += 24
+				bitStream += uint32(0xFFFF) << bitCount
+				out[outP] = byte(bitStream)
+				out[outP+1] = byte(bitStream >> 8)
+				outP += 2
+				bitStream >>= 16
+			}
+			for charnum >= start+3 {
+				start += 3
+				bitStream += 3 << bitCount
+				bitCount += 2
+			}
+			bitStream += uint32(charnum-start) << bitCount
+			bitCount += 2
+			if bitCount > 16 {
+				out[outP] = byte(bitStream)
+				out[outP+1] = byte(bitStream >> 8)
+				outP += 2
+				bitStream >>= 16
+				bitCount -= 16
+			}
+		}
+
+		count := s.norm[charnum]
+		charnum++
+		max := (2*threshold - 1) - remaining
+		if count < 0 {
+			remaining += count
+		} else {
+			remaining -= count
+		}
+		count++ // +1 for extra accuracy
+		if count >= threshold {
+			count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[
+		}
+		bitStream += uint32(count) << bitCount
+		bitCount += nbBits
+		if count < max {
+			bitCount--
+		}
+
+		previous0 = count == 1
+		if remaining < 1 {
+			return errors.New("internal error: remaining<1")
+		}
+		for remaining < threshold {
+			nbBits--
+			threshold >>= 1
+		}
+
+		if bitCount > 16 {
+			out[outP] = byte(bitStream)
+			out[outP+1] = byte(bitStream >> 8)
+			outP += 2
+			bitStream >>= 16
+			bitCount -= 16
+		}
+	}
+
+	out[outP] = byte(bitStream)
+	out[outP+1] = byte(bitStream >> 8)
+	outP += (bitCount + 7) / 8
+
+	if charnum > s.symbolLen {
+		return errors.New("internal error: charnum > s.symbolLen")
+	}
+	s.Out = out[:outP]
+	return nil
+}
+
+// symbolTransform contains the state transform for a symbol.
+type symbolTransform struct {
+	deltaFindState int32
+	deltaNbBits    uint32
+}
+
+// String prints values as a human readable string.
+func (s symbolTransform) String() string {
+	return fmt.Sprintf("dnbits: %08x, fs:%d", s.deltaNbBits, s.deltaFindState)
+}
+
+// cTable contains tables used for compression.
+type cTable struct {
+	tableSymbol []byte
+	stateTable  []uint16
+	symbolTT    []symbolTransform
+}
+
+// allocCtable will allocate tables needed for compression.
+// If existing tables a re big enough, they are simply re-used.
+func (s *Scratch) allocCtable() {
+	tableSize := 1 << s.actualTableLog
+	// get tableSymbol that is big enough.
+	if cap(s.ct.tableSymbol) < tableSize {
+		s.ct.tableSymbol = make([]byte, tableSize)
+	}
+	s.ct.tableSymbol = s.ct.tableSymbol[:tableSize]
+
+	ctSize := tableSize
+	if cap(s.ct.stateTable) < ctSize {
+		s.ct.stateTable = make([]uint16, ctSize)
+	}
+	s.ct.stateTable = s.ct.stateTable[:ctSize]
+
+	if cap(s.ct.symbolTT) < 256 {
+		s.ct.symbolTT = make([]symbolTransform, 256)
+	}
+	s.ct.symbolTT = s.ct.symbolTT[:256]
+}
+
+// buildCTable will populate the compression table so it is ready to be used.
+func (s *Scratch) buildCTable() error {
+	tableSize := uint32(1 << s.actualTableLog)
+	highThreshold := tableSize - 1
+	var cumul [maxSymbolValue + 2]int16
+
+	s.allocCtable()
+	tableSymbol := s.ct.tableSymbol[:tableSize]
+	// symbol start positions
+	{
+		cumul[0] = 0
+		for ui, v := range s.norm[:s.symbolLen-1] {
+			u := byte(ui) // one less than reference
+			if v == -1 {
+				// Low proba symbol
+				cumul[u+1] = cumul[u] + 1
+				tableSymbol[highThreshold] = u
+				highThreshold--
+			} else {
+				cumul[u+1] = cumul[u] + v
+			}
+		}
+		// Encode last symbol separately to avoid overflowing u
+		u := int(s.symbolLen - 1)
+		v := s.norm[s.symbolLen-1]
+		if v == -1 {
+			// Low proba symbol
+			cumul[u+1] = cumul[u] + 1
+			tableSymbol[highThreshold] = byte(u)
+			highThreshold--
+		} else {
+			cumul[u+1] = cumul[u] + v
+		}
+		if uint32(cumul[s.symbolLen]) != tableSize {
+			return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize)
+		}
+		cumul[s.symbolLen] = int16(tableSize) + 1
+	}
+	// Spread symbols
+	s.zeroBits = false
+	{
+		step := tableStep(tableSize)
+		tableMask := tableSize - 1
+		var position uint32
+		// if any symbol > largeLimit, we may have 0 bits output.
+		largeLimit := int16(1 << (s.actualTableLog - 1))
+		for ui, v := range s.norm[:s.symbolLen] {
+			symbol := byte(ui)
+			if v > largeLimit {
+				s.zeroBits = true
+			}
+			for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ {
+				tableSymbol[position] = symbol
+				position = (position + step) & tableMask
+				for position > highThreshold {
+					position = (position + step) & tableMask
+				} /* Low proba area */
+			}
+		}
+
+		// Check if we have gone through all positions
+		if position != 0 {
+			return errors.New("position!=0")
+		}
+	}
+
+	// Build table
+	table := s.ct.stateTable
+	{
+		tsi := int(tableSize)
+		for u, v := range tableSymbol {
+			// TableU16 : sorted by symbol order; gives next state value
+			table[cumul[v]] = uint16(tsi + u)
+			cumul[v]++
+		}
+	}
+
+	// Build Symbol Transformation Table
+	{
+		total := int16(0)
+		symbolTT := s.ct.symbolTT[:s.symbolLen]
+		tableLog := s.actualTableLog
+		tl := (uint32(tableLog) << 16) - (1 << tableLog)
+		for i, v := range s.norm[:s.symbolLen] {
+			switch v {
+			case 0:
+			case -1, 1:
+				symbolTT[i].deltaNbBits = tl
+				symbolTT[i].deltaFindState = int32(total - 1)
+				total++
+			default:
+				maxBitsOut := uint32(tableLog) - highBits(uint32(v-1))
+				minStatePlus := uint32(v) << maxBitsOut
+				symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus
+				symbolTT[i].deltaFindState = int32(total - v)
+				total += v
+			}
+		}
+		if total != int16(tableSize) {
+			return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize)
+		}
+	}
+	return nil
+}
+
+// countSimple will create a simple histogram in s.count.
+// Returns the biggest count.
+// Does not update s.clearCount.
+func (s *Scratch) countSimple(in []byte) (max int) {
+	for _, v := range in {
+		s.count[v]++
+	}
+	m := uint32(0)
+	for i, v := range s.count[:] {
+		if v > m {
+			m = v
+		}
+		if v > 0 {
+			s.symbolLen = uint16(i) + 1
+		}
+	}
+	return int(m)
+}
+
+// minTableLog provides the minimum logSize to safely represent a distribution.
+func (s *Scratch) minTableLog() uint8 {
+	minBitsSrc := highBits(uint32(s.br.remain()-1)) + 1
+	minBitsSymbols := highBits(uint32(s.symbolLen-1)) + 2
+	if minBitsSrc < minBitsSymbols {
+		return uint8(minBitsSrc)
+	}
+	return uint8(minBitsSymbols)
+}
+
+// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog
+func (s *Scratch) optimalTableLog() {
+	tableLog := s.TableLog
+	minBits := s.minTableLog()
+	maxBitsSrc := uint8(highBits(uint32(s.br.remain()-1))) - 2
+	if maxBitsSrc < tableLog {
+		// Accuracy can be reduced
+		tableLog = maxBitsSrc
+	}
+	if minBits > tableLog {
+		tableLog = minBits
+	}
+	// Need a minimum to safely represent all symbol values
+	if tableLog < minTablelog {
+		tableLog = minTablelog
+	}
+	if tableLog > maxTableLog {
+		tableLog = maxTableLog
+	}
+	s.actualTableLog = tableLog
+}
+
+var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000}
+
+// normalizeCount will normalize the count of the symbols so
+// the total is equal to the table size.
+func (s *Scratch) normalizeCount() error {
+	var (
+		tableLog          = s.actualTableLog
+		scale             = 62 - uint64(tableLog)
+		step              = (1 << 62) / uint64(s.br.remain())
+		vStep             = uint64(1) << (scale - 20)
+		stillToDistribute = int16(1 << tableLog)
+		largest           int
+		largestP          int16
+		lowThreshold      = (uint32)(s.br.remain() >> tableLog)
+	)
+
+	for i, cnt := range s.count[:s.symbolLen] {
+		// already handled
+		// if (count[s] == s.length) return 0;   /* rle special case */
+
+		if cnt == 0 {
+			s.norm[i] = 0
+			continue
+		}
+		if cnt <= lowThreshold {
+			s.norm[i] = -1
+			stillToDistribute--
+		} else {
+			proba := (int16)((uint64(cnt) * step) >> scale)
+			if proba < 8 {
+				restToBeat := vStep * uint64(rtbTable[proba])
+				v := uint64(cnt)*step - (uint64(proba) << scale)
+				if v > restToBeat {
+					proba++
+				}
+			}
+			if proba > largestP {
+				largestP = proba
+				largest = i
+			}
+			s.norm[i] = proba
+			stillToDistribute -= proba
+		}
+	}
+
+	if -stillToDistribute >= (s.norm[largest] >> 1) {
+		// corner case, need another normalization method
+		return s.normalizeCount2()
+	}
+	s.norm[largest] += stillToDistribute
+	return nil
+}
+
+// Secondary normalization method.
+// To be used when primary method fails.
+func (s *Scratch) normalizeCount2() error {
+	const notYetAssigned = -2
+	var (
+		distributed  uint32
+		total        = uint32(s.br.remain())
+		tableLog     = s.actualTableLog
+		lowThreshold = total >> tableLog
+		lowOne       = (total * 3) >> (tableLog + 1)
+	)
+	for i, cnt := range s.count[:s.symbolLen] {
+		if cnt == 0 {
+			s.norm[i] = 0
+			continue
+		}
+		if cnt <= lowThreshold {
+			s.norm[i] = -1
+			distributed++
+			total -= cnt
+			continue
+		}
+		if cnt <= lowOne {
+			s.norm[i] = 1
+			distributed++
+			total -= cnt
+			continue
+		}
+		s.norm[i] = notYetAssigned
+	}
+	toDistribute := (1 << tableLog) - distributed
+
+	if (total / toDistribute) > lowOne {
+		// risk of rounding to zero
+		lowOne = (total * 3) / (toDistribute * 2)
+		for i, cnt := range s.count[:s.symbolLen] {
+			if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) {
+				s.norm[i] = 1
+				distributed++
+				total -= cnt
+				continue
+			}
+		}
+		toDistribute = (1 << tableLog) - distributed
+	}
+	if distributed == uint32(s.symbolLen)+1 {
+		// all values are pretty poor;
+		//   probably incompressible data (should have already been detected);
+		//   find max, then give all remaining points to max
+		var maxV int
+		var maxC uint32
+		for i, cnt := range s.count[:s.symbolLen] {
+			if cnt > maxC {
+				maxV = i
+				maxC = cnt
+			}
+		}
+		s.norm[maxV] += int16(toDistribute)
+		return nil
+	}
+
+	if total == 0 {
+		// all of the symbols were low enough for the lowOne or lowThreshold
+		for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) {
+			if s.norm[i] > 0 {
+				toDistribute--
+				s.norm[i]++
+			}
+		}
+		return nil
+	}
+
+	var (
+		vStepLog = 62 - uint64(tableLog)
+		mid      = uint64((1 << (vStepLog - 1)) - 1)
+		rStep    = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining
+		tmpTotal = mid
+	)
+	for i, cnt := range s.count[:s.symbolLen] {
+		if s.norm[i] == notYetAssigned {
+			var (
+				end    = tmpTotal + uint64(cnt)*rStep
+				sStart = uint32(tmpTotal >> vStepLog)
+				sEnd   = uint32(end >> vStepLog)
+				weight = sEnd - sStart
+			)
+			if weight < 1 {
+				return errors.New("weight < 1")
+			}
+			s.norm[i] = int16(weight)
+			tmpTotal = end
+		}
+	}
+	return nil
+}
+
+// validateNorm validates the normalized histogram table.
+func (s *Scratch) validateNorm() (err error) {
+	var total int
+	for _, v := range s.norm[:s.symbolLen] {
+		if v >= 0 {
+			total += int(v)
+		} else {
+			total -= int(v)
+		}
+	}
+	defer func() {
+		if err == nil {
+			return
+		}
+		fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen)
+		for i, v := range s.norm[:s.symbolLen] {
+			fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v)
+		}
+	}()
+	if total != (1 << s.actualTableLog) {
+		return fmt.Errorf("warning: Total == %d != %d", total, 1<<s.actualTableLog)
+	}
+	for i, v := range s.count[s.symbolLen:] {
+		if v != 0 {
+			return fmt.Errorf("warning: Found symbol out of range, %d after cut", i)
+		}
+	}
+	return nil
+}

vendor/github.com/klauspost/compress/fse/decompress.go

@@ -0,0 +1,374 @@
+package fse
+
+import (
+	"errors"
+	"fmt"
+)
+
+const (
+	tablelogAbsoluteMax = 15
+)
+
+// Decompress a block of data.
+// You can provide a scratch buffer to avoid allocations.
+// If nil is provided a temporary one will be allocated.
+// It is possible, but by no way guaranteed that corrupt data will
+// return an error.
+// It is up to the caller to verify integrity of the returned data.
+// Use a predefined Scrach to set maximum acceptable output size.
+func Decompress(b []byte, s *Scratch) ([]byte, error) {
+	s, err := s.prepare(b)
+	if err != nil {
+		return nil, err
+	}
+	s.Out = s.Out[:0]
+	err = s.readNCount()
+	if err != nil {
+		return nil, err
+	}
+	err = s.buildDtable()
+	if err != nil {
+		return nil, err
+	}
+	err = s.decompress()
+	if err != nil {
+		return nil, err
+	}
+
+	return s.Out, nil
+}
+
+// readNCount will read the symbol distribution so decoding tables can be constructed.
+func (s *Scratch) readNCount() error {
+	var (
+		charnum   uint16
+		previous0 bool
+		b         = &s.br
+	)
+	iend := b.remain()
+	if iend < 4 {
+		return errors.New("input too small")
+	}
+	bitStream := b.Uint32()
+	nbBits := uint((bitStream & 0xF) + minTablelog) // extract tableLog
+	if nbBits > tablelogAbsoluteMax {
+		return errors.New("tableLog too large")
+	}
+	bitStream >>= 4
+	bitCount := uint(4)
+
+	s.actualTableLog = uint8(nbBits)
+	remaining := int32((1 << nbBits) + 1)
+	threshold := int32(1 << nbBits)
+	gotTotal := int32(0)
+	nbBits++
+
+	for remaining > 1 {
+		if previous0 {
+			n0 := charnum
+			for (bitStream & 0xFFFF) == 0xFFFF {
+				n0 += 24
+				if b.off < iend-5 {
+					b.advance(2)
+					bitStream = b.Uint32() >> bitCount
+				} else {
+					bitStream >>= 16
+					bitCount += 16
+				}
+			}
+			for (bitStream & 3) == 3 {
+				n0 += 3
+				bitStream >>= 2
+				bitCount += 2
+			}
+			n0 += uint16(bitStream & 3)
+			bitCount += 2
+			if n0 > maxSymbolValue {
+				return errors.New("maxSymbolValue too small")
+			}
+			for charnum < n0 {
+				s.norm[charnum&0xff] = 0
+				charnum++
+			}
+
+			if b.off <= iend-7 || b.off+int(bitCount>>3) <= iend-4 {
+				b.advance(bitCount >> 3)
+				bitCount &= 7
+				bitStream = b.Uint32() >> bitCount
+			} else {
+				bitStream >>= 2
+			}
+		}
+
+		max := (2*(threshold) - 1) - (remaining)
+		var count int32
+
+		if (int32(bitStream) & (threshold - 1)) < max {
+			count = int32(bitStream) & (threshold - 1)
+			bitCount += nbBits - 1
+		} else {
+			count = int32(bitStream) & (2*threshold - 1)
+			if count >= threshold {
+				count -= max
+			}
+			bitCount += nbBits
+		}
+
+		count-- // extra accuracy
+		if count < 0 {
+			// -1 means +1
+			remaining += count
+			gotTotal -= count
+		} else {
+			remaining -= count
+			gotTotal += count
+		}
+		s.norm[charnum&0xff] = int16(count)
+		charnum++
+		previous0 = count == 0
+		for remaining < threshold {
+			nbBits--
+			threshold >>= 1
+		}
+		if b.off <= iend-7 || b.off+int(bitCount>>3) <= iend-4 {
+			b.advance(bitCount >> 3)
+			bitCount &= 7
+		} else {
+			bitCount -= (uint)(8 * (len(b.b) - 4 - b.off))
+			b.off = len(b.b) - 4
+		}
+		bitStream = b.Uint32() >> (bitCount & 31)
+	}
+	s.symbolLen = charnum
+
+	if s.symbolLen <= 1 {
+		return fmt.Errorf("symbolLen (%d) too small", s.symbolLen)
+	}
+	if s.symbolLen > maxSymbolValue+1 {
+		return fmt.Errorf("symbolLen (%d) too big", s.symbolLen)
+	}
+	if remaining != 1 {
+		return fmt.Errorf("corruption detected (remaining %d != 1)", remaining)
+	}
+	if bitCount > 32 {
+		return fmt.Errorf("corruption detected (bitCount %d > 32)", bitCount)
+	}
+	if gotTotal != 1<<s.actualTableLog {
+		return fmt.Errorf("corruption detected (total %d != %d)", gotTotal, 1<<s.actualTableLog)
+	}
+	b.advance((bitCount + 7) >> 3)
+	return nil
+}
+
+// decSymbol contains information about a state entry,
+// Including the state offset base, the output symbol and
+// the number of bits to read for the low part of the destination state.
+type decSymbol struct {
+	newState uint16
+	symbol   uint8
+	nbBits   uint8
+}
+
+// allocDtable will allocate decoding tables if they are not big enough.
+func (s *Scratch) allocDtable() {
+	tableSize := 1 << s.actualTableLog
+	if cap(s.decTable) < tableSize {
+		s.decTable = make([]decSymbol, tableSize)
+	}
+	s.decTable = s.decTable[:tableSize]
+
+	if cap(s.ct.tableSymbol) < 256 {
+		s.ct.tableSymbol = make([]byte, 256)
+	}
+	s.ct.tableSymbol = s.ct.tableSymbol[:256]
+
+	if cap(s.ct.stateTable) < 256 {
+		s.ct.stateTable = make([]uint16, 256)
+	}
+	s.ct.stateTable = s.ct.stateTable[:256]
+}
+
+// buildDtable will build the decoding table.
+func (s *Scratch) buildDtable() error {
+	tableSize := uint32(1 << s.actualTableLog)
+	highThreshold := tableSize - 1
+	s.allocDtable()
+	symbolNext := s.ct.stateTable[:256]
+
+	// Init, lay down lowprob symbols
+	s.zeroBits = false
+	{
+		largeLimit := int16(1 << (s.actualTableLog - 1))
+		for i, v := range s.norm[:s.symbolLen] {
+			if v == -1 {
+				s.decTable[highThreshold].symbol = uint8(i)
+				highThreshold--
+				symbolNext[i] = 1
+			} else {
+				if v >= largeLimit {
+					s.zeroBits = true
+				}
+				symbolNext[i] = uint16(v)
+			}
+		}
+	}
+	// Spread symbols
+	{
+		tableMask := tableSize - 1
+		step := tableStep(tableSize)
+		position := uint32(0)
+		for ss, v := range s.norm[:s.symbolLen] {
+			for i := 0; i < int(v); i++ {
+				s.decTable[position].symbol = uint8(ss)
+				position = (position + step) & tableMask
+				for position > highThreshold {
+					// lowprob area
+					position = (position + step) & tableMask
+				}
+			}
+		}
+		if position != 0 {
+			// position must reach all cells once, otherwise normalizedCounter is incorrect
+			return errors.New("corrupted input (position != 0)")
+		}
+	}
+
+	// Build Decoding table
+	{
+		tableSize := uint16(1 << s.actualTableLog)
+		for u, v := range s.decTable {
+			symbol := v.symbol
+			nextState := symbolNext[symbol]
+			symbolNext[symbol] = nextState + 1
+			nBits := s.actualTableLog - byte(highBits(uint32(nextState)))
+			s.decTable[u].nbBits = nBits
+			newState := (nextState << nBits) - tableSize
+			if newState >= tableSize {
+				return fmt.Errorf("newState (%d) outside table size (%d)", newState, tableSize)
+			}
+			if newState == uint16(u) && nBits == 0 {
+				// Seems weird that this is possible with nbits > 0.
+				return fmt.Errorf("newState (%d) == oldState (%d) and no bits", newState, u)
+			}
+			s.decTable[u].newState = newState
+		}
+	}
+	return nil
+}
+
+// decompress will decompress the bitstream.
+// If the buffer is over-read an error is returned.
+func (s *Scratch) decompress() error {
+	br := &s.bits
+	br.init(s.br.unread())
+
+	var s1, s2 decoder
+	// Initialize and decode first state and symbol.
+	s1.init(br, s.decTable, s.actualTableLog)
+	s2.init(br, s.decTable, s.actualTableLog)
+
+	// Use temp table to avoid bound checks/append penalty.
+	var tmp = s.ct.tableSymbol[:256]
+	var off uint8
+
+	// Main part
+	if !s.zeroBits {
+		for br.off >= 8 {
+			br.fillFast()
+			tmp[off+0] = s1.nextFast()
+			tmp[off+1] = s2.nextFast()
+			br.fillFast()
+			tmp[off+2] = s1.nextFast()
+			tmp[off+3] = s2.nextFast()
+			off += 4
+			// When off is 0, we have overflowed and should write.
+			if off == 0 {
+				s.Out = append(s.Out, tmp...)
+				if len(s.Out) >= s.DecompressLimit {
+					return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit)
+				}
+			}
+		}
+	} else {
+		for br.off >= 8 {
+			br.fillFast()
+			tmp[off+0] = s1.next()
+			tmp[off+1] = s2.next()
+			br.fillFast()
+			tmp[off+2] = s1.next()
+			tmp[off+3] = s2.next()
+			off += 4
+			if off == 0 {
+				s.Out = append(s.Out, tmp...)
+				// When off is 0, we have overflowed and should write.
+				if len(s.Out) >= s.DecompressLimit {
+					return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit)
+				}
+			}
+		}
+	}
+	s.Out = append(s.Out, tmp[:off]...)
+
+	// Final bits, a bit more expensive check
+	for {
+		if s1.finished() {
+			s.Out = append(s.Out, s1.final(), s2.final())
+			break
+		}
+		br.fill()
+		s.Out = append(s.Out, s1.next())
+		if s2.finished() {
+			s.Out = append(s.Out, s2.final(), s1.final())
+			break
+		}
+		s.Out = append(s.Out, s2.next())
+		if len(s.Out) >= s.DecompressLimit {
+			return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit)
+		}
+	}
+	return br.close()
+}
+
+// decoder keeps track of the current state and updates it from the bitstream.
+type decoder struct {
+	state uint16
+	br    *bitReader
+	dt    []decSymbol
+}
+
+// init will initialize the decoder and read the first state from the stream.
+func (d *decoder) init(in *bitReader, dt []decSymbol, tableLog uint8) {
+	d.dt = dt
+	d.br = in
+	d.state = in.getBits(tableLog)
+}
+
+// next returns the next symbol and sets the next state.
+// At least tablelog bits must be available in the bit reader.
+func (d *decoder) next() uint8 {
+	n := &d.dt[d.state]
+	lowBits := d.br.getBits(n.nbBits)
+	d.state = n.newState + lowBits
+	return n.symbol
+}
+
+// finished returns true if all bits have been read from the bitstream
+// and the next state would require reading bits from the input.
+func (d *decoder) finished() bool {
+	return d.br.finished() && d.dt[d.state].nbBits > 0
+}
+
+// final returns the current state symbol without decoding the next.
+func (d *decoder) final() uint8 {
+	return d.dt[d.state].symbol
+}
+
+// nextFast returns the next symbol and sets the next state.
+// This can only be used if no symbols are 0 bits.
+// At least tablelog bits must be available in the bit reader.
+func (d *decoder) nextFast() uint8 {
+	n := d.dt[d.state]
+	lowBits := d.br.getBitsFast(n.nbBits)
+	d.state = n.newState + lowBits
+	return n.symbol
+}

vendor/github.com/klauspost/compress/fse/fse.go

@@ -0,0 +1,144 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+// Package fse provides Finite State Entropy encoding and decoding.
+//
+// Finite State Entropy encoding provides a fast near-optimal symbol encoding/decoding
+// for byte blocks as implemented in zstd.
+//
+// See https://github.com/klauspost/compress/tree/master/fse for more information.
+package fse
+
+import (
+	"errors"
+	"fmt"
+	"math/bits"
+)
+
+const (
+	/*!MEMORY_USAGE :
+	 *  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+	 *  Increasing memory usage improves compression ratio
+	 *  Reduced memory usage can improve speed, due to cache effect
+	 *  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+	maxMemoryUsage     = 14
+	defaultMemoryUsage = 13
+
+	maxTableLog     = maxMemoryUsage - 2
+	maxTablesize    = 1 << maxTableLog
+	defaultTablelog = defaultMemoryUsage - 2
+	minTablelog     = 5
+	maxSymbolValue  = 255
+)
+
+var (
+	// ErrIncompressible is returned when input is judged to be too hard to compress.
+	ErrIncompressible = errors.New("input is not compressible")
+
+	// ErrUseRLE is returned from the compressor when the input is a single byte value repeated.
+	ErrUseRLE = errors.New("input is single value repeated")
+)
+
+// Scratch provides temporary storage for compression and decompression.
+type Scratch struct {
+	// Private
+	count    [maxSymbolValue + 1]uint32
+	norm     [maxSymbolValue + 1]int16
+	br       byteReader
+	bits     bitReader
+	bw       bitWriter
+	ct       cTable      // Compression tables.
+	decTable []decSymbol // Decompression table.
+	maxCount int         // count of the most probable symbol
+
+	// Per block parameters.
+	// These can be used to override compression parameters of the block.
+	// Do not touch, unless you know what you are doing.
+
+	// Out is output buffer.
+	// If the scratch is re-used before the caller is done processing the output,
+	// set this field to nil.
+	// Otherwise the output buffer will be re-used for next Compression/Decompression step
+	// and allocation will be avoided.
+	Out []byte
+
+	// DecompressLimit limits the maximum decoded size acceptable.
+	// If > 0 decompression will stop when approximately this many bytes
+	// has been decoded.
+	// If 0, maximum size will be 2GB.
+	DecompressLimit int
+
+	symbolLen      uint16 // Length of active part of the symbol table.
+	actualTableLog uint8  // Selected tablelog.
+	zeroBits       bool   // no bits has prob > 50%.
+	clearCount     bool   // clear count
+
+	// MaxSymbolValue will override the maximum symbol value of the next block.
+	MaxSymbolValue uint8
+
+	// TableLog will attempt to override the tablelog for the next block.
+	TableLog uint8
+}
+
+// Histogram allows to populate the histogram and skip that step in the compression,
+// It otherwise allows to inspect the histogram when compression is done.
+// To indicate that you have populated the histogram call HistogramFinished
+// with the value of the highest populated symbol, as well as the number of entries
+// in the most populated entry. These are accepted at face value.
+// The returned slice will always be length 256.
+func (s *Scratch) Histogram() []uint32 {
+	return s.count[:]
+}
+
+// HistogramFinished can be called to indicate that the histogram has been populated.
+// maxSymbol is the index of the highest set symbol of the next data segment.
+// maxCount is the number of entries in the most populated entry.
+// These are accepted at face value.
+func (s *Scratch) HistogramFinished(maxSymbol uint8, maxCount int) {
+	s.maxCount = maxCount
+	s.symbolLen = uint16(maxSymbol) + 1
+	s.clearCount = maxCount != 0
+}
+
+// prepare will prepare and allocate scratch tables used for both compression and decompression.
+func (s *Scratch) prepare(in []byte) (*Scratch, error) {
+	if s == nil {
+		s = &Scratch{}
+	}
+	if s.MaxSymbolValue == 0 {
+		s.MaxSymbolValue = 255
+	}
+	if s.TableLog == 0 {
+		s.TableLog = defaultTablelog
+	}
+	if s.TableLog > maxTableLog {
+		return nil, fmt.Errorf("tableLog (%d) > maxTableLog (%d)", s.TableLog, maxTableLog)
+	}
+	if cap(s.Out) == 0 {
+		s.Out = make([]byte, 0, len(in))
+	}
+	if s.clearCount && s.maxCount == 0 {
+		for i := range s.count {
+			s.count[i] = 0
+		}
+		s.clearCount = false
+	}
+	s.br.init(in)
+	if s.DecompressLimit == 0 {
+		// Max size 2GB.
+		s.DecompressLimit = (2 << 30) - 1
+	}
+
+	return s, nil
+}
+
+// tableStep returns the next table index.
+func tableStep(tableSize uint32) uint32 {
+	return (tableSize >> 1) + (tableSize >> 3) + 3
+}
+
+func highBits(val uint32) (n uint32) {
+	return uint32(bits.Len32(val) - 1)
+}

vendor/github.com/klauspost/compress/gen.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+
+cd s2/cmd/_s2sx/ || exit 1
+go generate .

vendor/github.com/klauspost/compress/huff0/.gitignore

@@ -0,0 +1 @@
+/huff0-fuzz.zip

vendor/github.com/klauspost/compress/huff0/README.md

@@ -0,0 +1,89 @@
+# Huff0 entropy compression
+
+This package provides Huff0 encoding and decoding as used in zstd.
+            
+[Huff0](https://github.com/Cyan4973/FiniteStateEntropy#new-generation-entropy-coders), 
+a Huffman codec designed for modern CPU, featuring OoO (Out of Order) operations on multiple ALU 
+(Arithmetic Logic Unit), achieving extremely fast compression and decompression speeds.
+
+This can be used for compressing input with a lot of similar input values to the smallest number of bytes.
+This does not perform any multi-byte [dictionary coding](https://en.wikipedia.org/wiki/Dictionary_coder) as LZ coders,
+but it can be used as a secondary step to compressors (like Snappy) that does not do entropy encoding. 
+
+* [Godoc documentation](https://godoc.org/github.com/klauspost/compress/huff0)
+
+## News
+
+This is used as part of the [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and decompression package.
+
+This ensures that most functionality is well tested.
+
+# Usage
+
+This package provides a low level interface that allows to compress single independent blocks. 
+
+Each block is separate, and there is no built in integrity checks. 
+This means that the caller should keep track of block sizes and also do checksums if needed.  
+
+Compressing a block is done via the [`Compress1X`](https://godoc.org/github.com/klauspost/compress/huff0#Compress1X) and 
+[`Compress4X`](https://godoc.org/github.com/klauspost/compress/huff0#Compress4X) functions.
+You must provide input and will receive the output and maybe an error.
+
+These error values can be returned:
+
+| Error               | Description                                                                 |
+|---------------------|-----------------------------------------------------------------------------|
+| `<nil>`             | Everything ok, output is returned                                           |
+| `ErrIncompressible` | Returned when input is judged to be too hard to compress                    |
+| `ErrUseRLE`         | Returned from the compressor when the input is a single byte value repeated |
+| `ErrTooBig`         | Returned if the input block exceeds the maximum allowed size (128 Kib)      |
+| `(error)`           | An internal error occurred.                                                 |
+
+
+As can be seen above some of there are errors that will be returned even under normal operation so it is important to handle these.
+
+To reduce allocations you can provide a [`Scratch`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch) object 
+that can be re-used for successive calls. Both compression and decompression accepts a `Scratch` object, and the same 
+object can be used for both.   
+
+Be aware, that when re-using a `Scratch` object that the *output* buffer is also re-used, so if you are still using this
+you must set the `Out` field in the scratch to nil. The same buffer is used for compression and decompression output.
+
+The `Scratch` object will retain state that allows to re-use previous tables for encoding and decoding.  
+
+## Tables and re-use
+
+Huff0 allows for reusing tables from the previous block to save space if that is expected to give better/faster results. 
+
+The Scratch object allows you to set a [`ReusePolicy`](https://godoc.org/github.com/klauspost/compress/huff0#ReusePolicy) 
+that controls this behaviour. See the documentation for details. This can be altered between each block.
+
+Do however note that this information is *not* stored in the output block and it is up to the users of the package to
+record whether [`ReadTable`](https://godoc.org/github.com/klauspost/compress/huff0#ReadTable) should be called,
+based on the boolean reported back from the CompressXX call. 
+
+If you want to store the table separate from the data, you can access them as `OutData` and `OutTable` on the 
+[`Scratch`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch) object.
+
+## Decompressing
+
+The first part of decoding is to initialize the decoding table through [`ReadTable`](https://godoc.org/github.com/klauspost/compress/huff0#ReadTable).
+This will initialize the decoding tables. 
+You can supply the complete block to `ReadTable` and it will return the data part of the block 
+which can be given to the decompressor. 
+
+Decompressing is done by calling the [`Decompress1X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress1X) 
+or [`Decompress4X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress4X) function.
+
+For concurrently decompressing content with a fixed table a stateless [`Decoder`](https://godoc.org/github.com/klauspost/compress/huff0#Decoder) can be requested which will remain correct as long as the scratch is unchanged. The capacity of the provided slice indicates the expected output size.
+
+You must provide the output from the compression stage, at exactly the size you got back. If you receive an error back
+your input was likely corrupted. 
+
+It is important to note that a successful decoding does *not* mean your output matches your original input. 
+There are no integrity checks, so relying on errors from the decompressor does not assure your data is valid.
+
+# Contributing
+
+Contributions are always welcome. Be aware that adding public functions will require good justification and breaking 
+changes will likely not be accepted. If in doubt open an issue before writing the PR.

vendor/github.com/klauspost/compress/huff0/autogen.go

@@ -0,0 +1,5 @@
+package huff0
+
+//go:generate go run generate.go
+//go:generate asmfmt -w decompress_amd64.s
+//go:generate asmfmt -w decompress_8b_amd64.s

vendor/github.com/klauspost/compress/huff0/bitreader.go

@@ -0,0 +1,243 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+// bitReader reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReaderBytes struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReaderBytes) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.advance(8 - uint8(highBit32(uint32(v))))
+	return nil
+}
+
+// peekBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReaderBytes) peekByteFast() uint8 {
+	got := uint8(b.value >> 56)
+	return got
+}
+
+func (b *bitReaderBytes) advance(n uint8) {
+	b.bitsRead += n
+	b.value <<= n & 63
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReaderBytes) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+
+	// 2 bounds checks.
+	v := b.in[b.off-4 : b.off]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value |= uint64(low) << (b.bitsRead - 32)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fillFastStart() assumes the bitReaderBytes is empty and there is at least 8 bytes to read.
+func (b *bitReaderBytes) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReaderBytes) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value |= uint64(low) << (b.bitsRead - 32)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value |= uint64(b.in[b.off-1]) << (b.bitsRead - 8)
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReaderBytes) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+func (b *bitReaderBytes) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReaderBytes) close() error {
+	// Release reference.
+	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+
+// bitReaderShifted reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReaderShifted struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReaderShifted) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.advance(8 - uint8(highBit32(uint32(v))))
+	return nil
+}
+
+// peekBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReaderShifted) peekBitsFast(n uint8) uint16 {
+	return uint16(b.value >> ((64 - n) & 63))
+}
+
+// peekTopBits(n) is equvialent to peekBitFast(64 - n)
+func (b *bitReaderShifted) peekTopBits(n uint8) uint16 {
+	return uint16(b.value >> n)
+}
+
+func (b *bitReaderShifted) advance(n uint8) {
+	b.bitsRead += n
+	b.value <<= n & 63
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReaderShifted) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+
+	// 2 bounds checks.
+	v := b.in[b.off-4 : b.off]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fillFastStart() assumes the bitReaderShifted is empty and there is at least 8 bytes to read.
+func (b *bitReaderShifted) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReaderShifted) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value |= uint64(b.in[b.off-1]) << ((b.bitsRead - 8) & 63)
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReaderShifted) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+func (b *bitReaderShifted) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReaderShifted) close() error {
+	// Release reference.
+	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}

vendor/github.com/klauspost/compress/huff0/bitwriter.go

@@ -0,0 +1,210 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+import "fmt"
+
+// bitWriter will write bits.
+// First bit will be LSB of the first byte of output.
+type bitWriter struct {
+	bitContainer uint64
+	nBits        uint8
+	out          []byte
+}
+
+// bitMask16 is bitmasks. Has extra to avoid bounds check.
+var bitMask16 = [32]uint16{
+	0, 1, 3, 7, 0xF, 0x1F,
+	0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
+	0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF} /* up to 16 bits */
+
+// addBits16NC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16NC(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// encSymbol will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) encSymbol(ct cTable, symbol byte) {
+	enc := ct[symbol]
+	b.bitContainer |= uint64(enc.val) << (b.nBits & 63)
+	if false {
+		if enc.nBits == 0 {
+			panic("nbits 0")
+		}
+	}
+	b.nBits += enc.nBits
+}
+
+// encTwoSymbols will add up to 32 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) encTwoSymbols(ct cTable, av, bv byte) {
+	encA := ct[av]
+	encB := ct[bv]
+	sh := b.nBits & 63
+	combined := uint64(encA.val) | (uint64(encB.val) << (encA.nBits & 63))
+	b.bitContainer |= combined << sh
+	if false {
+		if encA.nBits == 0 {
+			panic("nbitsA 0")
+		}
+		if encB.nBits == 0 {
+			panic("nbitsB 0")
+		}
+	}
+	b.nBits += encA.nBits + encB.nBits
+}
+
+// addBits16ZeroNC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+// This is fastest if bits can be zero.
+func (b *bitWriter) addBits16ZeroNC(value uint16, bits uint8) {
+	if bits == 0 {
+		return
+	}
+	value <<= (16 - bits) & 15
+	value >>= (16 - bits) & 15
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// flush will flush all pending full bytes.
+// There will be at least 56 bits available for writing when this has been called.
+// Using flush32 is faster, but leaves less space for writing.
+func (b *bitWriter) flush() {
+	v := b.nBits >> 3
+	switch v {
+	case 0:
+		return
+	case 1:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+		)
+		b.bitContainer >>= 1 << 3
+	case 2:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+		)
+		b.bitContainer >>= 2 << 3
+	case 3:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+		)
+		b.bitContainer >>= 3 << 3
+	case 4:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+		)
+		b.bitContainer >>= 4 << 3
+	case 5:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+		)
+		b.bitContainer >>= 5 << 3
+	case 6:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+		)
+		b.bitContainer >>= 6 << 3
+	case 7:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+		)
+		b.bitContainer >>= 7 << 3
+	case 8:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+			byte(b.bitContainer>>56),
+		)
+		b.bitContainer = 0
+		b.nBits = 0
+		return
+	default:
+		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
+	}
+	b.nBits &= 7
+}
+
+// flush32 will flush out, so there are at least 32 bits available for writing.
+func (b *bitWriter) flush32() {
+	if b.nBits < 32 {
+		return
+	}
+	b.out = append(b.out,
+		byte(b.bitContainer),
+		byte(b.bitContainer>>8),
+		byte(b.bitContainer>>16),
+		byte(b.bitContainer>>24))
+	b.nBits -= 32
+	b.bitContainer >>= 32
+}
+
+// flushAlign will flush remaining full bytes and align to next byte boundary.
+func (b *bitWriter) flushAlign() {
+	nbBytes := (b.nBits + 7) >> 3
+	for i := uint8(0); i < nbBytes; i++ {
+		b.out = append(b.out, byte(b.bitContainer>>(i*8)))
+	}
+	b.nBits = 0
+	b.bitContainer = 0
+}
+
+// close will write the alignment bit and write the final byte(s)
+// to the output.
+func (b *bitWriter) close() error {
+	// End mark
+	b.addBits16Clean(1, 1)
+	// flush until next byte.
+	b.flushAlign()
+	return nil
+}
+
+// reset and continue writing by appending to out.
+func (b *bitWriter) reset(out []byte) {
+	b.bitContainer = 0
+	b.nBits = 0
+	b.out = out
+}

vendor/github.com/klauspost/compress/huff0/bytereader.go

@@ -0,0 +1,54 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+// byteReader provides a byte reader that reads
+// little endian values from a byte stream.
+// The input stream is manually advanced.
+// The reader performs no bounds checks.
+type byteReader struct {
+	b   []byte
+	off int
+}
+
+// init will initialize the reader and set the input.
+func (b *byteReader) init(in []byte) {
+	b.b = in
+	b.off = 0
+}
+
+// advance the stream b n bytes.
+func (b *byteReader) advance(n uint) {
+	b.off += int(n)
+}
+
+// Int32 returns a little endian int32 starting at current offset.
+func (b byteReader) Int32() int32 {
+	v3 := int32(b.b[b.off+3])
+	v2 := int32(b.b[b.off+2])
+	v1 := int32(b.b[b.off+1])
+	v0 := int32(b.b[b.off])
+	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
+}
+
+// Uint32 returns a little endian uint32 starting at current offset.
+func (b byteReader) Uint32() uint32 {
+	v3 := uint32(b.b[b.off+3])
+	v2 := uint32(b.b[b.off+2])
+	v1 := uint32(b.b[b.off+1])
+	v0 := uint32(b.b[b.off])
+	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
+}
+
+// unread returns the unread portion of the input.
+func (b byteReader) unread() []byte {
+	return b.b[b.off:]
+}
+
+// remain will return the number of bytes remaining.
+func (b byteReader) remain() int {
+	return len(b.b) - b.off
+}

vendor/github.com/klauspost/compress/huff0/compress.go

@@ -0,0 +1,729 @@
+package huff0
+
+import (
+	"fmt"
+	"math"
+	"runtime"
+	"sync"
+)
+
+// Compress1X will compress the input.
+// The output can be decoded using Decompress1X.
+// Supply a Scratch object. The scratch object contains state about re-use,
+// So when sharing across independent encodes, be sure to set the re-use policy.
+func Compress1X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return nil, false, err
+	}
+	return compress(in, s, s.compress1X)
+}
+
+// Compress4X will compress the input. The input is split into 4 independent blocks
+// and compressed similar to Compress1X.
+// The output can be decoded using Decompress4X.
+// Supply a Scratch object. The scratch object contains state about re-use,
+// So when sharing across independent encodes, be sure to set the re-use policy.
+func Compress4X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return nil, false, err
+	}
+	if false {
+		// TODO: compress4Xp only slightly faster.
+		const parallelThreshold = 8 << 10
+		if len(in) < parallelThreshold || runtime.GOMAXPROCS(0) == 1 {
+			return compress(in, s, s.compress4X)
+		}
+		return compress(in, s, s.compress4Xp)
+	}
+	return compress(in, s, s.compress4X)
+}
+
+func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)) (out []byte, reUsed bool, err error) {
+	// Nuke previous table if we cannot reuse anyway.
+	if s.Reuse == ReusePolicyNone {
+		s.prevTable = s.prevTable[:0]
+	}
+
+	// Create histogram, if none was provided.
+	maxCount := s.maxCount
+	var canReuse = false
+	if maxCount == 0 {
+		maxCount, canReuse = s.countSimple(in)
+	} else {
+		canReuse = s.canUseTable(s.prevTable)
+	}
+
+	// We want the output size to be less than this:
+	wantSize := len(in)
+	if s.WantLogLess > 0 {
+		wantSize -= wantSize >> s.WantLogLess
+	}
+
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount >= len(in) {
+		if maxCount > len(in) {
+			return nil, false, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
+		}
+		if len(in) == 1 {
+			return nil, false, ErrIncompressible
+		}
+		// One symbol, use RLE
+		return nil, false, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return nil, false, ErrIncompressible
+	}
+	if s.Reuse == ReusePolicyMust && !canReuse {
+		// We must reuse, but we can't.
+		return nil, false, ErrIncompressible
+	}
+	if (s.Reuse == ReusePolicyPrefer || s.Reuse == ReusePolicyMust) && canReuse {
+		keepTable := s.cTable
+		keepTL := s.actualTableLog
+		s.cTable = s.prevTable
+		s.actualTableLog = s.prevTableLog
+		s.Out, err = compressor(in)
+		s.cTable = keepTable
+		s.actualTableLog = keepTL
+		if err == nil && len(s.Out) < wantSize {
+			s.OutData = s.Out
+			return s.Out, true, nil
+		}
+		if s.Reuse == ReusePolicyMust {
+			return nil, false, ErrIncompressible
+		}
+		// Do not attempt to re-use later.
+		s.prevTable = s.prevTable[:0]
+	}
+
+	// Calculate new table.
+	err = s.buildCTable()
+	if err != nil {
+		return nil, false, err
+	}
+
+	if false && !s.canUseTable(s.cTable) {
+		panic("invalid table generated")
+	}
+
+	if s.Reuse == ReusePolicyAllow && canReuse {
+		hSize := len(s.Out)
+		oldSize := s.prevTable.estimateSize(s.count[:s.symbolLen])
+		newSize := s.cTable.estimateSize(s.count[:s.symbolLen])
+		if oldSize <= hSize+newSize || hSize+12 >= wantSize {
+			// Retain cTable even if we re-use.
+			keepTable := s.cTable
+			keepTL := s.actualTableLog
+
+			s.cTable = s.prevTable
+			s.actualTableLog = s.prevTableLog
+			s.Out, err = compressor(in)
+
+			// Restore ctable.
+			s.cTable = keepTable
+			s.actualTableLog = keepTL
+			if err != nil {
+				return nil, false, err
+			}
+			if len(s.Out) >= wantSize {
+				return nil, false, ErrIncompressible
+			}
+			s.OutData = s.Out
+			return s.Out, true, nil
+		}
+	}
+
+	// Use new table
+	err = s.cTable.write(s)
+	if err != nil {
+		s.OutTable = nil
+		return nil, false, err
+	}
+	s.OutTable = s.Out
+
+	// Compress using new table
+	s.Out, err = compressor(in)
+	if err != nil {
+		s.OutTable = nil
+		return nil, false, err
+	}
+	if len(s.Out) >= wantSize {
+		s.OutTable = nil
+		return nil, false, ErrIncompressible
+	}
+	// Move current table into previous.
+	s.prevTable, s.prevTableLog, s.cTable = s.cTable, s.actualTableLog, s.prevTable[:0]
+	s.OutData = s.Out[len(s.OutTable):]
+	return s.Out, false, nil
+}
+
+// EstimateSizes will estimate the data sizes
+func EstimateSizes(in []byte, s *Scratch) (tableSz, dataSz, reuseSz int, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	// Create histogram, if none was provided.
+	tableSz, dataSz, reuseSz = -1, -1, -1
+	maxCount := s.maxCount
+	var canReuse = false
+	if maxCount == 0 {
+		maxCount, canReuse = s.countSimple(in)
+	} else {
+		canReuse = s.canUseTable(s.prevTable)
+	}
+
+	// We want the output size to be less than this:
+	wantSize := len(in)
+	if s.WantLogLess > 0 {
+		wantSize -= wantSize >> s.WantLogLess
+	}
+
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount >= len(in) {
+		if maxCount > len(in) {
+			return 0, 0, 0, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
+		}
+		if len(in) == 1 {
+			return 0, 0, 0, ErrIncompressible
+		}
+		// One symbol, use RLE
+		return 0, 0, 0, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return 0, 0, 0, ErrIncompressible
+	}
+
+	// Calculate new table.
+	err = s.buildCTable()
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	if false && !s.canUseTable(s.cTable) {
+		panic("invalid table generated")
+	}
+
+	tableSz, err = s.cTable.estTableSize(s)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+	if canReuse {
+		reuseSz = s.prevTable.estimateSize(s.count[:s.symbolLen])
+	}
+	dataSz = s.cTable.estimateSize(s.count[:s.symbolLen])
+
+	// Restore
+	return tableSz, dataSz, reuseSz, nil
+}
+
+func (s *Scratch) compress1X(src []byte) ([]byte, error) {
+	return s.compress1xDo(s.Out, src)
+}
+
+func (s *Scratch) compress1xDo(dst, src []byte) ([]byte, error) {
+	var bw = bitWriter{out: dst}
+
+	// N is length divisible by 4.
+	n := len(src)
+	n -= n & 3
+	cTable := s.cTable[:256]
+
+	// Encode last bytes.
+	for i := len(src) & 3; i > 0; i-- {
+		bw.encSymbol(cTable, src[n+i-1])
+	}
+	n -= 4
+	if s.actualTableLog <= 8 {
+		for ; n >= 0; n -= 4 {
+			tmp := src[n : n+4]
+			// tmp should be len 4
+			bw.flush32()
+			bw.encTwoSymbols(cTable, tmp[3], tmp[2])
+			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
+		}
+	} else {
+		for ; n >= 0; n -= 4 {
+			tmp := src[n : n+4]
+			// tmp should be len 4
+			bw.flush32()
+			bw.encTwoSymbols(cTable, tmp[3], tmp[2])
+			bw.flush32()
+			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
+		}
+	}
+	err := bw.close()
+	return bw.out, err
+}
+
+var sixZeros [6]byte
+
+func (s *Scratch) compress4X(src []byte) ([]byte, error) {
+	if len(src) < 12 {
+		return nil, ErrIncompressible
+	}
+	segmentSize := (len(src) + 3) / 4
+
+	// Add placeholder for output length
+	offsetIdx := len(s.Out)
+	s.Out = append(s.Out, sixZeros[:]...)
+
+	for i := 0; i < 4; i++ {
+		toDo := src
+		if len(toDo) > segmentSize {
+			toDo = toDo[:segmentSize]
+		}
+		src = src[len(toDo):]
+
+		var err error
+		idx := len(s.Out)
+		s.Out, err = s.compress1xDo(s.Out, toDo)
+		if err != nil {
+			return nil, err
+		}
+		if len(s.Out)-idx > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
+		}
+		// Write compressed length as little endian before block.
+		if i < 3 {
+			// Last length is not written.
+			length := len(s.Out) - idx
+			s.Out[i*2+offsetIdx] = byte(length)
+			s.Out[i*2+offsetIdx+1] = byte(length >> 8)
+		}
+	}
+
+	return s.Out, nil
+}
+
+// compress4Xp will compress 4 streams using separate goroutines.
+func (s *Scratch) compress4Xp(src []byte) ([]byte, error) {
+	if len(src) < 12 {
+		return nil, ErrIncompressible
+	}
+	// Add placeholder for output length
+	s.Out = s.Out[:6]
+
+	segmentSize := (len(src) + 3) / 4
+	var wg sync.WaitGroup
+	var errs [4]error
+	wg.Add(4)
+	for i := 0; i < 4; i++ {
+		toDo := src
+		if len(toDo) > segmentSize {
+			toDo = toDo[:segmentSize]
+		}
+		src = src[len(toDo):]
+
+		// Separate goroutine for each block.
+		go func(i int) {
+			s.tmpOut[i], errs[i] = s.compress1xDo(s.tmpOut[i][:0], toDo)
+			wg.Done()
+		}(i)
+	}
+	wg.Wait()
+	for i := 0; i < 4; i++ {
+		if errs[i] != nil {
+			return nil, errs[i]
+		}
+		o := s.tmpOut[i]
+		if len(o) > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
+		}
+		// Write compressed length as little endian before block.
+		if i < 3 {
+			// Last length is not written.
+			s.Out[i*2] = byte(len(o))
+			s.Out[i*2+1] = byte(len(o) >> 8)
+		}
+
+		// Write output.
+		s.Out = append(s.Out, o...)
+	}
+	return s.Out, nil
+}
+
+// countSimple will create a simple histogram in s.count.
+// Returns the biggest count.
+// Does not update s.clearCount.
+func (s *Scratch) countSimple(in []byte) (max int, reuse bool) {
+	reuse = true
+	for _, v := range in {
+		s.count[v]++
+	}
+	m := uint32(0)
+	if len(s.prevTable) > 0 {
+		for i, v := range s.count[:] {
+			if v > m {
+				m = v
+			}
+			if v > 0 {
+				s.symbolLen = uint16(i) + 1
+				if i >= len(s.prevTable) {
+					reuse = false
+				} else {
+					if s.prevTable[i].nBits == 0 {
+						reuse = false
+					}
+				}
+			}
+		}
+		return int(m), reuse
+	}
+	for i, v := range s.count[:] {
+		if v > m {
+			m = v
+		}
+		if v > 0 {
+			s.symbolLen = uint16(i) + 1
+		}
+	}
+	return int(m), false
+}
+
+func (s *Scratch) canUseTable(c cTable) bool {
+	if len(c) < int(s.symbolLen) {
+		return false
+	}
+	for i, v := range s.count[:s.symbolLen] {
+		if v != 0 && c[i].nBits == 0 {
+			return false
+		}
+	}
+	return true
+}
+
+func (s *Scratch) validateTable(c cTable) bool {
+	if len(c) < int(s.symbolLen) {
+		return false
+	}
+	for i, v := range s.count[:s.symbolLen] {
+		if v != 0 {
+			if c[i].nBits == 0 {
+				return false
+			}
+			if c[i].nBits > s.actualTableLog {
+				return false
+			}
+		}
+	}
+	return true
+}
+
+// minTableLog provides the minimum logSize to safely represent a distribution.
+func (s *Scratch) minTableLog() uint8 {
+	minBitsSrc := highBit32(uint32(s.br.remain())) + 1
+	minBitsSymbols := highBit32(uint32(s.symbolLen-1)) + 2
+	if minBitsSrc < minBitsSymbols {
+		return uint8(minBitsSrc)
+	}
+	return uint8(minBitsSymbols)
+}
+
+// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog
+func (s *Scratch) optimalTableLog() {
+	tableLog := s.TableLog
+	minBits := s.minTableLog()
+	maxBitsSrc := uint8(highBit32(uint32(s.br.remain()-1))) - 1
+	if maxBitsSrc < tableLog {
+		// Accuracy can be reduced
+		tableLog = maxBitsSrc
+	}
+	if minBits > tableLog {
+		tableLog = minBits
+	}
+	// Need a minimum to safely represent all symbol values
+	if tableLog < minTablelog {
+		tableLog = minTablelog
+	}
+	if tableLog > tableLogMax {
+		tableLog = tableLogMax
+	}
+	s.actualTableLog = tableLog
+}
+
+type cTableEntry struct {
+	val   uint16
+	nBits uint8
+	// We have 8 bits extra
+}
+
+const huffNodesMask = huffNodesLen - 1
+
+func (s *Scratch) buildCTable() error {
+	s.optimalTableLog()
+	s.huffSort()
+	if cap(s.cTable) < maxSymbolValue+1 {
+		s.cTable = make([]cTableEntry, s.symbolLen, maxSymbolValue+1)
+	} else {
+		s.cTable = s.cTable[:s.symbolLen]
+		for i := range s.cTable {
+			s.cTable[i] = cTableEntry{}
+		}
+	}
+
+	var startNode = int16(s.symbolLen)
+	nonNullRank := s.symbolLen - 1
+
+	nodeNb := startNode
+	huffNode := s.nodes[1 : huffNodesLen+1]
+
+	// This overlays the slice above, but allows "-1" index lookups.
+	// Different from reference implementation.
+	huffNode0 := s.nodes[0 : huffNodesLen+1]
+
+	for huffNode[nonNullRank].count == 0 {
+		nonNullRank--
+	}
+
+	lowS := int16(nonNullRank)
+	nodeRoot := nodeNb + lowS - 1
+	lowN := nodeNb
+	huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count
+	huffNode[lowS].parent, huffNode[lowS-1].parent = uint16(nodeNb), uint16(nodeNb)
+	nodeNb++
+	lowS -= 2
+	for n := nodeNb; n <= nodeRoot; n++ {
+		huffNode[n].count = 1 << 30
+	}
+	// fake entry, strong barrier
+	huffNode0[0].count = 1 << 31
+
+	// create parents
+	for nodeNb <= nodeRoot {
+		var n1, n2 int16
+		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+			n1 = lowS
+			lowS--
+		} else {
+			n1 = lowN
+			lowN++
+		}
+		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+			n2 = lowS
+			lowS--
+		} else {
+			n2 = lowN
+			lowN++
+		}
+
+		huffNode[nodeNb].count = huffNode0[n1+1].count + huffNode0[n2+1].count
+		huffNode0[n1+1].parent, huffNode0[n2+1].parent = uint16(nodeNb), uint16(nodeNb)
+		nodeNb++
+	}
+
+	// distribute weights (unlimited tree height)
+	huffNode[nodeRoot].nbBits = 0
+	for n := nodeRoot - 1; n >= startNode; n-- {
+		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+	}
+	for n := uint16(0); n <= nonNullRank; n++ {
+		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+	}
+	s.actualTableLog = s.setMaxHeight(int(nonNullRank))
+	maxNbBits := s.actualTableLog
+
+	// fill result into tree (val, nbBits)
+	if maxNbBits > tableLogMax {
+		return fmt.Errorf("internal error: maxNbBits (%d) > tableLogMax (%d)", maxNbBits, tableLogMax)
+	}
+	var nbPerRank [tableLogMax + 1]uint16
+	var valPerRank [16]uint16
+	for _, v := range huffNode[:nonNullRank+1] {
+		nbPerRank[v.nbBits]++
+	}
+	// determine stating value per rank
+	{
+		min := uint16(0)
+		for n := maxNbBits; n > 0; n-- {
+			// get starting value within each rank
+			valPerRank[n] = min
+			min += nbPerRank[n]
+			min >>= 1
+		}
+	}
+
+	// push nbBits per symbol, symbol order
+	for _, v := range huffNode[:nonNullRank+1] {
+		s.cTable[v.symbol].nBits = v.nbBits
+	}
+
+	// assign value within rank, symbol order
+	t := s.cTable[:s.symbolLen]
+	for n, val := range t {
+		nbits := val.nBits & 15
+		v := valPerRank[nbits]
+		t[n].val = v
+		valPerRank[nbits] = v + 1
+	}
+
+	return nil
+}
+
+// huffSort will sort symbols, decreasing order.
+func (s *Scratch) huffSort() {
+	type rankPos struct {
+		base    uint32
+		current uint32
+	}
+
+	// Clear nodes
+	nodes := s.nodes[:huffNodesLen+1]
+	s.nodes = nodes
+	nodes = nodes[1 : huffNodesLen+1]
+
+	// Sort into buckets based on length of symbol count.
+	var rank [32]rankPos
+	for _, v := range s.count[:s.symbolLen] {
+		r := highBit32(v+1) & 31
+		rank[r].base++
+	}
+	// maxBitLength is log2(BlockSizeMax) + 1
+	const maxBitLength = 18 + 1
+	for n := maxBitLength; n > 0; n-- {
+		rank[n-1].base += rank[n].base
+	}
+	for n := range rank[:maxBitLength] {
+		rank[n].current = rank[n].base
+	}
+	for n, c := range s.count[:s.symbolLen] {
+		r := (highBit32(c+1) + 1) & 31
+		pos := rank[r].current
+		rank[r].current++
+		prev := nodes[(pos-1)&huffNodesMask]
+		for pos > rank[r].base && c > prev.count {
+			nodes[pos&huffNodesMask] = prev
+			pos--
+			prev = nodes[(pos-1)&huffNodesMask]
+		}
+		nodes[pos&huffNodesMask] = nodeElt{count: c, symbol: byte(n)}
+	}
+}
+
+func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
+	maxNbBits := s.actualTableLog
+	huffNode := s.nodes[1 : huffNodesLen+1]
+	//huffNode = huffNode[: huffNodesLen]
+
+	largestBits := huffNode[lastNonNull].nbBits
+
+	// early exit : no elt > maxNbBits
+	if largestBits <= maxNbBits {
+		return largestBits
+	}
+	totalCost := int(0)
+	baseCost := int(1) << (largestBits - maxNbBits)
+	n := uint32(lastNonNull)
+
+	for huffNode[n].nbBits > maxNbBits {
+		totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits))
+		huffNode[n].nbBits = maxNbBits
+		n--
+	}
+	// n stops at huffNode[n].nbBits <= maxNbBits
+
+	for huffNode[n].nbBits == maxNbBits {
+		n--
+	}
+	// n end at index of smallest symbol using < maxNbBits
+
+	// renorm totalCost
+	totalCost >>= largestBits - maxNbBits /* note : totalCost is necessarily a multiple of baseCost */
+
+	// repay normalized cost
+	{
+		const noSymbol = 0xF0F0F0F0
+		var rankLast [tableLogMax + 2]uint32
+
+		for i := range rankLast[:] {
+			rankLast[i] = noSymbol
+		}
+
+		// Get pos of last (smallest) symbol per rank
+		{
+			currentNbBits := maxNbBits
+			for pos := int(n); pos >= 0; pos-- {
+				if huffNode[pos].nbBits >= currentNbBits {
+					continue
+				}
+				currentNbBits = huffNode[pos].nbBits // < maxNbBits
+				rankLast[maxNbBits-currentNbBits] = uint32(pos)
+			}
+		}
+
+		for totalCost > 0 {
+			nBitsToDecrease := uint8(highBit32(uint32(totalCost))) + 1
+
+			for ; nBitsToDecrease > 1; nBitsToDecrease-- {
+				highPos := rankLast[nBitsToDecrease]
+				lowPos := rankLast[nBitsToDecrease-1]
+				if highPos == noSymbol {
+					continue
+				}
+				if lowPos == noSymbol {
+					break
+				}
+				highTotal := huffNode[highPos].count
+				lowTotal := 2 * huffNode[lowPos].count
+				if highTotal <= lowTotal {
+					break
+				}
+			}
+			// only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !)
+			// HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary
+			// FIXME: try to remove
+			for (nBitsToDecrease <= tableLogMax) && (rankLast[nBitsToDecrease] == noSymbol) {
+				nBitsToDecrease++
+			}
+			totalCost -= 1 << (nBitsToDecrease - 1)
+			if rankLast[nBitsToDecrease-1] == noSymbol {
+				// this rank is no longer empty
+				rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]
+			}
+			huffNode[rankLast[nBitsToDecrease]].nbBits++
+			if rankLast[nBitsToDecrease] == 0 {
+				/* special case, reached largest symbol */
+				rankLast[nBitsToDecrease] = noSymbol
+			} else {
+				rankLast[nBitsToDecrease]--
+				if huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease {
+					rankLast[nBitsToDecrease] = noSymbol /* this rank is now empty */
+				}
+			}
+		}
+
+		for totalCost < 0 { /* Sometimes, cost correction overshoot */
+			if rankLast[1] == noSymbol { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
+				for huffNode[n].nbBits == maxNbBits {
+					n--
+				}
+				huffNode[n+1].nbBits--
+				rankLast[1] = n + 1
+				totalCost++
+				continue
+			}
+			huffNode[rankLast[1]+1].nbBits--
+			rankLast[1]++
+			totalCost++
+		}
+	}
+	return maxNbBits
+}
+
+type nodeElt struct {
+	count  uint32
+	parent uint16
+	symbol byte
+	nbBits uint8
+}

vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s

@@ -0,0 +1,488 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#define bufoff      256 // see decompress.go, we're using [4][256]byte table
+
+// func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+	MOVQ BP, 0(SP)
+
+	XORQ exhausted, exhausted // exhausted = false
+	XORQ off, off             // off = 0
+
+	MOVBQZX peekBits+32(FP), peek_bits
+	MOVQ    buf+40(FP), buffer
+	MOVQ    tbl+48(FP), table
+
+	MOVQ pbr0+0(FP), br0
+	MOVQ pbr1+8(FP), br1
+	MOVQ pbr2+16(FP), br2
+	MOVQ pbr3+24(FP), br3
+
+main_loop:
+
+	// const stream = 0
+	// br0.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br0), br_bits_read
+	MOVQ    bitReaderShifted_value(br0), br_value
+	MOVQ    bitReaderShifted_off(br0), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill0
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br0), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br0.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill0:
+
+	// val0 := br0.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 0(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br0.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 0+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br0)
+	MOVQ br_value, bitReaderShifted_value(br0)
+	MOVQ br_offset, bitReaderShifted_off(br0)
+
+	// const stream = 1
+	// br1.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br1), br_bits_read
+	MOVQ    bitReaderShifted_value(br1), br_value
+	MOVQ    bitReaderShifted_off(br1), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill1
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br1), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br1.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill1:
+
+	// val0 := br1.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 256(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br1.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 256+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br1)
+	MOVQ br_value, bitReaderShifted_value(br1)
+	MOVQ br_offset, bitReaderShifted_off(br1)
+
+	// const stream = 2
+	// br2.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br2), br_bits_read
+	MOVQ    bitReaderShifted_value(br2), br_value
+	MOVQ    bitReaderShifted_off(br2), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill2
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br2), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br2.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill2:
+
+	// val0 := br2.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 512(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br2.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 512+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br2)
+	MOVQ br_value, bitReaderShifted_value(br2)
+	MOVQ br_offset, bitReaderShifted_off(br2)
+
+	// const stream = 3
+	// br3.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br3), br_bits_read
+	MOVQ    bitReaderShifted_value(br3), br_value
+	MOVQ    bitReaderShifted_off(br3), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill3
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br3), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br3.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill3:
+
+	// val0 := br3.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 768(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br3.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 768+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br3)
+	MOVQ br_value, bitReaderShifted_value(br3)
+	MOVQ br_offset, bitReaderShifted_off(br3)
+
+	ADDQ $4, off // off += 2
+
+	TESTB DH, DH // any br[i].ofs < 4?
+	JNZ   end
+
+	CMPQ off, $bufoff
+	JL   main_loop
+
+end:
+	MOVQ 0(SP), BP
+
+	MOVB off, ret+56(FP)
+	RET
+
+#undef off
+#undef buffer
+#undef table
+
+#undef br_bits_read
+#undef br_value
+#undef br_offset
+#undef peek_bits
+#undef exhausted
+
+#undef br0
+#undef br1
+#undef br2
+#undef br3

vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s.in

@@ -0,0 +1,197 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+
+#define bufoff      256     // see decompress.go, we're using [4][256]byte table
+
+//func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+    MOVQ    BP, 0(SP)
+
+    XORQ    exhausted, exhausted    // exhausted = false
+    XORQ    off, off                // off = 0
+
+    MOVBQZX peekBits+32(FP), peek_bits
+    MOVQ    buf+40(FP), buffer
+    MOVQ    tbl+48(FP), table
+
+    MOVQ    pbr0+0(FP), br0
+    MOVQ    pbr1+8(FP), br1
+    MOVQ    pbr2+16(FP), br2
+    MOVQ    pbr3+24(FP), br3
+
+main_loop:
+{{ define "decode_2_values_x86" }}
+    // const stream = {{ var "id" }}
+    // br{{ var "id"}}.fillFast()
+    MOVBQZX bitReaderShifted_bitsRead(br{{ var "id" }}), br_bits_read
+    MOVQ    bitReaderShifted_value(br{{ var "id" }}), br_value
+    MOVQ    bitReaderShifted_off(br{{ var "id" }}), br_offset
+
+	// if b.bitsRead >= 32 {
+    CMPQ    br_bits_read, $32
+    JB      skip_fill{{ var "id" }}
+
+    SUBQ    $32, br_bits_read       // b.bitsRead -= 32
+    SUBQ    $4, br_offset           // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+    MOVQ    bitReaderShifted_in(br{{ var "id" }}), AX
+    MOVL    0(br_offset)(AX*1), AX  // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+    MOVQ    br_bits_read, CX
+    SHLQ    CL, AX
+    ORQ     AX, br_value
+
+    // exhausted = exhausted || (br{{ var "id"}}.off < 4)
+    CMPQ    br_offset, $4
+    SETLT   DL
+    ORB     DL, DH
+    // }
+skip_fill{{ var "id" }}:
+
+    // val0 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v0 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+    // val1 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v1 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CX, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off] = uint8(v0.entry >> 8)
+    // buf[stream][off+1] = uint8(v1.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}(buffer)(off*1)
+
+    // SECOND PART:
+    // val2 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v2 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+    // val3 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v3 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CX, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off+2] = uint8(v2.entry >> 8)
+    // buf[stream][off+3] = uint8(v3.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}+2(buffer)(off*1)
+
+    // update the bitrader reader structure
+    MOVB    br_bits_read, bitReaderShifted_bitsRead(br{{ var "id" }})
+    MOVQ    br_value, bitReaderShifted_value(br{{ var "id" }})
+    MOVQ    br_offset, bitReaderShifted_off(br{{ var "id" }})
+{{ end }}
+
+    {{ set "id" "0" }}
+    {{ set "ofs" "0" }}
+    {{ set "bufofs" "0" }} {{/* id * bufoff */}}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "1" }}
+    {{ set "ofs" "8" }}
+    {{ set "bufofs" "256" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "2" }}
+    {{ set "ofs" "16" }}
+    {{ set "bufofs" "512" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "3" }}
+    {{ set "ofs" "24" }}
+    {{ set "bufofs" "768" }}
+    {{ template "decode_2_values_x86" . }}
+
+    ADDQ    $4, off     // off += 2
+
+    TESTB   DH, DH      // any br[i].ofs < 4?
+    JNZ     end
+
+    CMPQ    off, $bufoff
+    JL      main_loop
+end:
+    MOVQ    0(SP), BP
+
+    MOVB    off, ret+56(FP)
+    RET
+#undef  off
+#undef  buffer
+#undef  table
+
+#undef  br_bits_read
+#undef  br_value
+#undef  br_offset
+#undef  peek_bits
+#undef  exhausted
+
+#undef  br0
+#undef  br1
+#undef  br2
+#undef  br3

vendor/github.com/klauspost/compress/huff0/decompress_amd64.go

@@ -0,0 +1,181 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+// This file contains the specialisation of Decoder.Decompress4X
+// that uses an asm implementation of its main loop.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+)
+
+// decompress4x_main_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog > 8.
+//go:noescape
+func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
+
+// decompress4x_8b_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog <= 8 which decodes 4 entries
+// per loop.
+//go:noescape
+func decompress4x_8b_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
+
+// fallback8BitSize is the size where using Go version is faster.
+const fallback8BitSize = 800
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+
+	use8BitTables := d.actualTableLog <= 8
+	if cap(dst) < fallback8BitSize && use8BitTables {
+		return d.decompress4X8bit(dst, src)
+	}
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	buf := d.buffer()
+	var off uint8
+	var decoded int
+
+	const debug = false
+
+	// see: bitReaderShifted.peekBitsFast()
+	peekBits := uint8((64 - d.actualTableLog) & 63)
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256
+	for {
+		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
+			break
+		}
+
+		if use8BitTables {
+			off = decompress4x_8b_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+		} else {
+			off = decompress4x_main_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+		}
+		if debug {
+			fmt.Print("DEBUG: ")
+			fmt.Printf("off=%d,", off)
+			for i := 0; i < 4; i++ {
+				fmt.Printf(" br[%d]={bitsRead=%d, value=%x, off=%d}",
+					i, br[i].bitsRead, br[i].value, br[i].off)
+			}
+			fmt.Println("")
+		}
+
+		if off != 0 {
+			break
+		}
+
+		if bufoff > dstEvery {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 1")
+		}
+		copy(out, buf[0][:])
+		copy(out[dstEvery:], buf[1][:])
+		copy(out[dstEvery*2:], buf[2][:])
+		copy(out[dstEvery*3:], buf[3][:])
+		out = out[bufoff:]
+		decoded += bufoff * 4
+		// There must at least be 3 buffers left.
+		if len(out) < dstEvery*3 {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 2")
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
+		decoded += int(off) * 4
+		out = out[off:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.bufs.Put(buf)
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}

vendor/github.com/klauspost/compress/huff0/decompress_amd64.s

@@ -0,0 +1,506 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#ifdef GOAMD64_v4
+#ifndef GOAMD64_v3
+#define GOAMD64_v3
+#endif
+#endif
+
+#define bufoff      256 // see decompress.go, we're using [4][256]byte table
+
+// func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_main_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+	MOVQ BP, 0(SP)
+
+	XORQ exhausted, exhausted // exhausted = false
+	XORQ off, off             // off = 0
+
+	MOVBQZX peekBits+32(FP), peek_bits
+	MOVQ    buf+40(FP), buffer
+	MOVQ    tbl+48(FP), table
+
+	MOVQ pbr0+0(FP), br0
+	MOVQ pbr1+8(FP), br1
+	MOVQ pbr2+16(FP), br2
+	MOVQ pbr3+24(FP), br3
+
+main_loop:
+
+	// const stream = 0
+	// br0.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br0), br_bits_read
+	MOVQ    bitReaderShifted_value(br0), br_value
+	MOVQ    bitReaderShifted_off(br0), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill0
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br0), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br0.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill0:
+
+	// val0 := br0.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 0(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br0)
+	MOVQ br_value, bitReaderShifted_value(br0)
+	MOVQ br_offset, bitReaderShifted_off(br0)
+
+	// const stream = 1
+	// br1.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br1), br_bits_read
+	MOVQ    bitReaderShifted_value(br1), br_value
+	MOVQ    bitReaderShifted_off(br1), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill1
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br1), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br1.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill1:
+
+	// val0 := br1.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 256(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br1)
+	MOVQ br_value, bitReaderShifted_value(br1)
+	MOVQ br_offset, bitReaderShifted_off(br1)
+
+	// const stream = 2
+	// br2.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br2), br_bits_read
+	MOVQ    bitReaderShifted_value(br2), br_value
+	MOVQ    bitReaderShifted_off(br2), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill2
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br2), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br2.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill2:
+
+	// val0 := br2.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 512(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br2)
+	MOVQ br_value, bitReaderShifted_value(br2)
+	MOVQ br_offset, bitReaderShifted_off(br2)
+
+	// const stream = 3
+	// br3.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br3), br_bits_read
+	MOVQ    bitReaderShifted_value(br3), br_value
+	MOVQ    bitReaderShifted_off(br3), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill3
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br3), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br3.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill3:
+
+	// val0 := br3.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 768(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br3)
+	MOVQ br_value, bitReaderShifted_value(br3)
+	MOVQ br_offset, bitReaderShifted_off(br3)
+
+	ADDQ $2, off // off += 2
+
+	TESTB DH, DH // any br[i].ofs < 4?
+	JNZ   end
+
+	CMPQ off, $bufoff
+	JL   main_loop
+
+end:
+	MOVQ 0(SP), BP
+
+	MOVB off, ret+56(FP)
+	RET
+
+#undef off
+#undef buffer
+#undef table
+
+#undef br_bits_read
+#undef br_value
+#undef br_offset
+#undef peek_bits
+#undef exhausted
+
+#undef br0
+#undef br1
+#undef br2
+#undef br3

vendor/github.com/klauspost/compress/huff0/decompress_amd64.s.in

@@ -0,0 +1,195 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#ifdef GOAMD64_v4
+#ifndef GOAMD64_v3
+#define GOAMD64_v3
+#endif
+#endif
+
+#define bufoff      256     // see decompress.go, we're using [4][256]byte table
+
+//func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_main_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+    MOVQ    BP, 0(SP)
+
+    XORQ    exhausted, exhausted    // exhausted = false
+    XORQ    off, off                // off = 0
+
+    MOVBQZX peekBits+32(FP), peek_bits
+    MOVQ    buf+40(FP), buffer
+    MOVQ    tbl+48(FP), table
+
+    MOVQ    pbr0+0(FP), br0
+    MOVQ    pbr1+8(FP), br1
+    MOVQ    pbr2+16(FP), br2
+    MOVQ    pbr3+24(FP), br3
+
+main_loop:
+{{ define "decode_2_values_x86" }}
+    // const stream = {{ var "id" }}
+    // br{{ var "id"}}.fillFast()
+    MOVBQZX bitReaderShifted_bitsRead(br{{ var "id" }}), br_bits_read
+    MOVQ    bitReaderShifted_value(br{{ var "id" }}), br_value
+    MOVQ    bitReaderShifted_off(br{{ var "id" }}), br_offset
+
+    // We must have at least 2 * max tablelog left
+    CMPQ    br_bits_read, $64-22
+    JBE     skip_fill{{ var "id" }}
+
+    SUBQ    $32, br_bits_read       // b.bitsRead -= 32
+    SUBQ    $4, br_offset           // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+    MOVQ    bitReaderShifted_in(br{{ var "id" }}), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+    SHLXQ   br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+#else
+    MOVL    0(br_offset)(AX*1), AX  // AX = uint32(b.in[b.off:b.off+4])
+    MOVQ    br_bits_read, CX
+    SHLQ    CL, AX
+#endif
+
+    ORQ     AX, br_value
+
+    // exhausted = exhausted || (br{{ var "id"}}.off < 4)
+    CMPQ    br_offset, $4
+    SETLT   DL
+    ORB     DL, DH
+    // }
+skip_fill{{ var "id" }}:
+
+    // val0 := br{{ var "id"}}.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+    SHRXQ   peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+#else
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+#endif
+
+    // v0 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+    MOVBQZX AL, CX
+    SHLXQ   AX, br_value, br_value // value <<= n
+#else
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+#endif
+
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+#ifdef GOAMD64_v3
+    SHRXQ    peek_bits, br_value, AX  // AX = (value >> peek_bits) & mask
+#else
+    // val1 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+#endif
+
+    // v1 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+    MOVBQZX AL, CX
+    SHLXQ   AX, br_value, br_value // value <<= n
+#else
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+#endif
+
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off] = uint8(v0.entry >> 8)
+    // buf[stream][off+1] = uint8(v1.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}(buffer)(off*1)
+
+    // update the bitrader reader structure
+    MOVB    br_bits_read, bitReaderShifted_bitsRead(br{{ var "id" }})
+    MOVQ    br_value, bitReaderShifted_value(br{{ var "id" }})
+    MOVQ    br_offset, bitReaderShifted_off(br{{ var "id" }})
+{{ end }}
+
+    {{ set "id" "0" }}
+    {{ set "ofs" "0" }}
+    {{ set "bufofs" "0" }} {{/* id * bufoff */}}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "1" }}
+    {{ set "ofs" "8" }}
+    {{ set "bufofs" "256" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "2" }}
+    {{ set "ofs" "16" }}
+    {{ set "bufofs" "512" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "3" }}
+    {{ set "ofs" "24" }}
+    {{ set "bufofs" "768" }}
+    {{ template "decode_2_values_x86" . }}
+
+    ADDQ    $2, off     // off += 2
+
+    TESTB   DH, DH      // any br[i].ofs < 4?
+    JNZ     end
+
+    CMPQ    off, $bufoff
+    JL      main_loop
+end:
+    MOVQ    0(SP), BP
+
+    MOVB    off, ret+56(FP)
+    RET
+#undef  off
+#undef  buffer
+#undef  table
+
+#undef  br_bits_read
+#undef  br_value
+#undef  br_offset
+#undef  peek_bits
+#undef  exhausted
+
+#undef  br0
+#undef  br1
+#undef  br2
+#undef  br3

vendor/github.com/klauspost/compress/huff0/decompress_generic.go

@@ -0,0 +1,193 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+// This file contains a generic implementation of Decoder.Decompress4X.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+)
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+	if use8BitTables && d.actualTableLog <= 8 {
+		return d.decompress4X8bit(dst, src)
+	}
+
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	buf := d.buffer()
+	var off uint8
+	var decoded int
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256
+	for {
+		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
+			break
+		}
+
+		{
+			const stream = 0
+			const stream2 = 1
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		{
+			const stream = 2
+			const stream2 = 3
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		off += 2
+
+		if off == 0 {
+			if bufoff > dstEvery {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 1")
+			}
+			copy(out, buf[0][:])
+			copy(out[dstEvery:], buf[1][:])
+			copy(out[dstEvery*2:], buf[2][:])
+			copy(out[dstEvery*3:], buf[3][:])
+			out = out[bufoff:]
+			decoded += bufoff * 4
+			// There must at least be 3 buffers left.
+			if len(out) < dstEvery*3 {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 2")
+			}
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
+		decoded += int(off) * 4
+		out = out[off:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.bufs.Put(buf)
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}

vendor/github.com/klauspost/compress/huff0/huff0.go

@@ -0,0 +1,337 @@
+// Package huff0 provides fast huffman encoding as used in zstd.
+//
+// See README.md at https://github.com/klauspost/compress/tree/master/huff0 for details.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"math/bits"
+	"sync"
+
+	"github.com/klauspost/compress/fse"
+)
+
+const (
+	maxSymbolValue = 255
+
+	// zstandard limits tablelog to 11, see:
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#huffman-tree-description
+	tableLogMax     = 11
+	tableLogDefault = 11
+	minTablelog     = 5
+	huffNodesLen    = 512
+
+	// BlockSizeMax is maximum input size for a single block uncompressed.
+	BlockSizeMax = 1<<18 - 1
+)
+
+var (
+	// ErrIncompressible is returned when input is judged to be too hard to compress.
+	ErrIncompressible = errors.New("input is not compressible")
+
+	// ErrUseRLE is returned from the compressor when the input is a single byte value repeated.
+	ErrUseRLE = errors.New("input is single value repeated")
+
+	// ErrTooBig is return if input is too large for a single block.
+	ErrTooBig = errors.New("input too big")
+
+	// ErrMaxDecodedSizeExceeded is return if input is too large for a single block.
+	ErrMaxDecodedSizeExceeded = errors.New("maximum output size exceeded")
+)
+
+type ReusePolicy uint8
+
+const (
+	// ReusePolicyAllow will allow reuse if it produces smaller output.
+	ReusePolicyAllow ReusePolicy = iota
+
+	// ReusePolicyPrefer will re-use aggressively if possible.
+	// This will not check if a new table will produce smaller output,
+	// except if the current table is impossible to use or
+	// compressed output is bigger than input.
+	ReusePolicyPrefer
+
+	// ReusePolicyNone will disable re-use of tables.
+	// This is slightly faster than ReusePolicyAllow but may produce larger output.
+	ReusePolicyNone
+
+	// ReusePolicyMust must allow reuse and produce smaller output.
+	ReusePolicyMust
+)
+
+type Scratch struct {
+	count [maxSymbolValue + 1]uint32
+
+	// Per block parameters.
+	// These can be used to override compression parameters of the block.
+	// Do not touch, unless you know what you are doing.
+
+	// Out is output buffer.
+	// If the scratch is re-used before the caller is done processing the output,
+	// set this field to nil.
+	// Otherwise the output buffer will be re-used for next Compression/Decompression step
+	// and allocation will be avoided.
+	Out []byte
+
+	// OutTable will contain the table data only, if a new table has been generated.
+	// Slice of the returned data.
+	OutTable []byte
+
+	// OutData will contain the compressed data.
+	// Slice of the returned data.
+	OutData []byte
+
+	// MaxDecodedSize will set the maximum allowed output size.
+	// This value will automatically be set to BlockSizeMax if not set.
+	// Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded.
+	MaxDecodedSize int
+
+	br byteReader
+
+	// MaxSymbolValue will override the maximum symbol value of the next block.
+	MaxSymbolValue uint8
+
+	// TableLog will attempt to override the tablelog for the next block.
+	// Must be <= 11 and >= 5.
+	TableLog uint8
+
+	// Reuse will specify the reuse policy
+	Reuse ReusePolicy
+
+	// WantLogLess allows to specify a log 2 reduction that should at least be achieved,
+	// otherwise the block will be returned as incompressible.
+	// The reduction should then at least be (input size >> WantLogLess)
+	// If WantLogLess == 0 any improvement will do.
+	WantLogLess uint8
+
+	symbolLen      uint16 // Length of active part of the symbol table.
+	maxCount       int    // count of the most probable symbol
+	clearCount     bool   // clear count
+	actualTableLog uint8  // Selected tablelog.
+	prevTableLog   uint8  // Tablelog for previous table
+	prevTable      cTable // Table used for previous compression.
+	cTable         cTable // compression table
+	dt             dTable // decompression table
+	nodes          []nodeElt
+	tmpOut         [4][]byte
+	fse            *fse.Scratch
+	decPool        sync.Pool // *[4][256]byte buffers.
+	huffWeight     [maxSymbolValue + 1]byte
+}
+
+// TransferCTable will transfer the previously used compression table.
+func (s *Scratch) TransferCTable(src *Scratch) {
+	if cap(s.prevTable) < len(src.prevTable) {
+		s.prevTable = make(cTable, 0, maxSymbolValue+1)
+	}
+	s.prevTable = s.prevTable[:len(src.prevTable)]
+	copy(s.prevTable, src.prevTable)
+	s.prevTableLog = src.prevTableLog
+}
+
+func (s *Scratch) prepare(in []byte) (*Scratch, error) {
+	if len(in) > BlockSizeMax {
+		return nil, ErrTooBig
+	}
+	if s == nil {
+		s = &Scratch{}
+	}
+	if s.MaxSymbolValue == 0 {
+		s.MaxSymbolValue = maxSymbolValue
+	}
+	if s.TableLog == 0 {
+		s.TableLog = tableLogDefault
+	}
+	if s.TableLog > tableLogMax || s.TableLog < minTablelog {
+		return nil, fmt.Errorf(" invalid tableLog %d (%d -> %d)", s.TableLog, minTablelog, tableLogMax)
+	}
+	if s.MaxDecodedSize <= 0 || s.MaxDecodedSize > BlockSizeMax {
+		s.MaxDecodedSize = BlockSizeMax
+	}
+	if s.clearCount && s.maxCount == 0 {
+		for i := range s.count {
+			s.count[i] = 0
+		}
+		s.clearCount = false
+	}
+	if cap(s.Out) == 0 {
+		s.Out = make([]byte, 0, len(in))
+	}
+	s.Out = s.Out[:0]
+
+	s.OutTable = nil
+	s.OutData = nil
+	if cap(s.nodes) < huffNodesLen+1 {
+		s.nodes = make([]nodeElt, 0, huffNodesLen+1)
+	}
+	s.nodes = s.nodes[:0]
+	if s.fse == nil {
+		s.fse = &fse.Scratch{}
+	}
+	s.br.init(in)
+
+	return s, nil
+}
+
+type cTable []cTableEntry
+
+func (c cTable) write(s *Scratch) error {
+	var (
+		// precomputed conversion table
+		bitsToWeight [tableLogMax + 1]byte
+		huffLog      = s.actualTableLog
+		// last weight is not saved.
+		maxSymbolValue = uint8(s.symbolLen - 1)
+		huffWeight     = s.huffWeight[:256]
+	)
+	const (
+		maxFSETableLog = 6
+	)
+	// convert to weight
+	bitsToWeight[0] = 0
+	for n := uint8(1); n < huffLog+1; n++ {
+		bitsToWeight[n] = huffLog + 1 - n
+	}
+
+	// Acquire histogram for FSE.
+	hist := s.fse.Histogram()
+	hist = hist[:256]
+	for i := range hist[:16] {
+		hist[i] = 0
+	}
+	for n := uint8(0); n < maxSymbolValue; n++ {
+		v := bitsToWeight[c[n].nBits] & 15
+		huffWeight[n] = v
+		hist[v]++
+	}
+
+	// FSE compress if feasible.
+	if maxSymbolValue >= 2 {
+		huffMaxCnt := uint32(0)
+		huffMax := uint8(0)
+		for i, v := range hist[:16] {
+			if v == 0 {
+				continue
+			}
+			huffMax = byte(i)
+			if v > huffMaxCnt {
+				huffMaxCnt = v
+			}
+		}
+		s.fse.HistogramFinished(huffMax, int(huffMaxCnt))
+		s.fse.TableLog = maxFSETableLog
+		b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse)
+		if err == nil && len(b) < int(s.symbolLen>>1) {
+			s.Out = append(s.Out, uint8(len(b)))
+			s.Out = append(s.Out, b...)
+			return nil
+		}
+		// Unable to compress (RLE/uncompressible)
+	}
+	// write raw values as 4-bits (max : 15)
+	if maxSymbolValue > (256 - 128) {
+		// should not happen : likely means source cannot be compressed
+		return ErrIncompressible
+	}
+	op := s.Out
+	// special case, pack weights 4 bits/weight.
+	op = append(op, 128|(maxSymbolValue-1))
+	// be sure it doesn't cause msan issue in final combination
+	huffWeight[maxSymbolValue] = 0
+	for n := uint16(0); n < uint16(maxSymbolValue); n += 2 {
+		op = append(op, (huffWeight[n]<<4)|huffWeight[n+1])
+	}
+	s.Out = op
+	return nil
+}
+
+func (c cTable) estTableSize(s *Scratch) (sz int, err error) {
+	var (
+		// precomputed conversion table
+		bitsToWeight [tableLogMax + 1]byte
+		huffLog      = s.actualTableLog
+		// last weight is not saved.
+		maxSymbolValue = uint8(s.symbolLen - 1)
+		huffWeight     = s.huffWeight[:256]
+	)
+	const (
+		maxFSETableLog = 6
+	)
+	// convert to weight
+	bitsToWeight[0] = 0
+	for n := uint8(1); n < huffLog+1; n++ {
+		bitsToWeight[n] = huffLog + 1 - n
+	}
+
+	// Acquire histogram for FSE.
+	hist := s.fse.Histogram()
+	hist = hist[:256]
+	for i := range hist[:16] {
+		hist[i] = 0
+	}
+	for n := uint8(0); n < maxSymbolValue; n++ {
+		v := bitsToWeight[c[n].nBits] & 15
+		huffWeight[n] = v
+		hist[v]++
+	}
+
+	// FSE compress if feasible.
+	if maxSymbolValue >= 2 {
+		huffMaxCnt := uint32(0)
+		huffMax := uint8(0)
+		for i, v := range hist[:16] {
+			if v == 0 {
+				continue
+			}
+			huffMax = byte(i)
+			if v > huffMaxCnt {
+				huffMaxCnt = v
+			}
+		}
+		s.fse.HistogramFinished(huffMax, int(huffMaxCnt))
+		s.fse.TableLog = maxFSETableLog
+		b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse)
+		if err == nil && len(b) < int(s.symbolLen>>1) {
+			sz += 1 + len(b)
+			return sz, nil
+		}
+		// Unable to compress (RLE/uncompressible)
+	}
+	// write raw values as 4-bits (max : 15)
+	if maxSymbolValue > (256 - 128) {
+		// should not happen : likely means source cannot be compressed
+		return 0, ErrIncompressible
+	}
+	// special case, pack weights 4 bits/weight.
+	sz += 1 + int(maxSymbolValue/2)
+	return sz, nil
+}
+
+// estimateSize returns the estimated size in bytes of the input represented in the
+// histogram supplied.
+func (c cTable) estimateSize(hist []uint32) int {
+	nbBits := uint32(7)
+	for i, v := range c[:len(hist)] {
+		nbBits += uint32(v.nBits) * hist[i]
+	}
+	return int(nbBits >> 3)
+}
+
+// minSize returns the minimum possible size considering the shannon limit.
+func (s *Scratch) minSize(total int) int {
+	nbBits := float64(7)
+	fTotal := float64(total)
+	for _, v := range s.count[:s.symbolLen] {
+		n := float64(v)
+		if n > 0 {
+			nbBits += math.Log2(fTotal/n) * n
+		}
+	}
+	return int(nbBits) >> 3
+}
+
+func highBit32(val uint32) (n uint32) {
+	return uint32(bits.Len32(val) - 1)
+}

vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo.go

@@ -0,0 +1,34 @@
+// Package cpuinfo gives runtime info about the current CPU.
+//
+// This is a very limited module meant for use internally
+// in this project. For more versatile solution check
+// https://github.com/klauspost/cpuid.
+package cpuinfo
+
+// HasBMI1 checks whether an x86 CPU supports the BMI1 extension.
+func HasBMI1() bool {
+	return hasBMI1
+}
+
+// HasBMI2 checks whether an x86 CPU supports the BMI2 extension.
+func HasBMI2() bool {
+	return hasBMI2
+}
+
+// DisableBMI2 will disable BMI2, for testing purposes.
+// Call returned function to restore previous state.
+func DisableBMI2() func() {
+	old := hasBMI2
+	hasBMI2 = false
+	return func() {
+		hasBMI2 = old
+	}
+}
+
+// HasBMI checks whether an x86 CPU supports both BMI1 and BMI2 extensions.
+func HasBMI() bool {
+	return HasBMI1() && HasBMI2()
+}
+
+var hasBMI1 bool
+var hasBMI2 bool

vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.go

@@ -0,0 +1,11 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+package cpuinfo
+
+// go:noescape
+func x86extensions() (bmi1, bmi2 bool)
+
+func init() {
+	hasBMI1, hasBMI2 = x86extensions()
+}

vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.s

@@ -0,0 +1,36 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+TEXT ·x86extensions(SB), NOSPLIT, $0
+	// 1. determine max EAX value
+	XORQ AX, AX
+	CPUID
+
+	CMPQ AX, $7
+	JB   unsupported
+
+	// 2. EAX = 7, ECX = 0 --- see Table 3-8 "Information Returned by CPUID Instruction"
+	MOVQ $7, AX
+	MOVQ $0, CX
+	CPUID
+
+	BTQ   $3, BX // bit 3 = BMI1
+	SETCS AL
+
+	BTQ   $8, BX // bit 8 = BMI2
+	SETCS AH
+
+	MOVB AL, bmi1+0(FP)
+	MOVB AH, bmi2+1(FP)
+	RET
+
+unsupported:
+	XORQ AX, AX
+	MOVB AL, bmi1+0(FP)
+	MOVB AL, bmi2+1(FP)
+	RET

vendor/github.com/klauspost/compress/internal/snapref/LICENSE

@@ -0,0 +1,27 @@
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

vendor/github.com/klauspost/compress/internal/snapref/decode.go

@@ -0,0 +1,264 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt reports that the input is invalid.
+	ErrCorrupt = errors.New("snappy: corrupt input")
+	// ErrTooLarge reports that the uncompressed length is too large.
+	ErrTooLarge = errors.New("snappy: decoded block is too large")
+	// ErrUnsupported reports that the input isn't supported.
+	ErrUnsupported = errors.New("snappy: unsupported input")
+
+	errUnsupportedLiteralLength = errors.New("snappy: unsupported literal length")
+)
+
+// DecodedLen returns the length of the decoded block.
+func DecodedLen(src []byte) (int, error) {
+	v, _, err := decodedLen(src)
+	return v, err
+}
+
+// decodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func decodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n <= 0 || v > 0xffffffff {
+		return 0, 0, ErrCorrupt
+	}
+
+	const wordSize = 32 << (^uint(0) >> 32 & 1)
+	if wordSize == 32 && v > 0x7fffffff {
+		return 0, 0, ErrTooLarge
+	}
+	return int(v), n, nil
+}
+
+const (
+	decodeErrCodeCorrupt                  = 1
+	decodeErrCodeUnsupportedLiteralLength = 2
+)
+
+// Decode returns the decoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire decoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Decode handles the Snappy block format, not the Snappy stream format.
+func Decode(dst, src []byte) ([]byte, error) {
+	dLen, s, err := decodedLen(src)
+	if err != nil {
+		return nil, err
+	}
+	if dLen <= len(dst) {
+		dst = dst[:dLen]
+	} else {
+		dst = make([]byte, dLen)
+	}
+	switch decode(dst, src[s:]) {
+	case 0:
+		return dst, nil
+	case decodeErrCodeUnsupportedLiteralLength:
+		return nil, errUnsupportedLiteralLength
+	}
+	return nil, ErrCorrupt
+}
+
+// NewReader returns a new Reader that decompresses from r, using the framing
+// format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r:       r,
+		decoded: make([]byte, maxBlockSize),
+		buf:     make([]byte, maxEncodedLenOfMaxBlockSize+checksumSize),
+	}
+}
+
+// Reader is an io.Reader that can read Snappy-compressed bytes.
+//
+// Reader handles the Snappy stream format, not the Snappy block format.
+type Reader struct {
+	r       io.Reader
+	err     error
+	decoded []byte
+	buf     []byte
+	// decoded[i:j] contains decoded bytes that have not yet been passed on.
+	i, j       int
+	readHeader bool
+}
+
+// Reset discards any buffered data, resets all state, and switches the Snappy
+// reader to read from r. This permits reusing a Reader rather than allocating
+// a new one.
+func (r *Reader) Reset(reader io.Reader) {
+	r.r = reader
+	r.err = nil
+	r.i = 0
+	r.j = 0
+	r.readHeader = false
+}
+
+func (r *Reader) readFull(p []byte, allowEOF bool) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF || (r.err == io.EOF && !allowEOF) {
+			r.err = ErrCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+func (r *Reader) fill() error {
+	for r.i >= r.j {
+		if !r.readFull(r.buf[:4], true) {
+			return r.err
+		}
+		chunkType := r.buf[0]
+		if !r.readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			r.err = ErrUnsupported
+			return r.err
+		}
+
+		// The chunk types are specified at
+		// https://github.com/google/snappy/blob/master/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[checksumSize:]
+
+			n, err := DecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return r.err
+			}
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if _, err := Decode(r.decoded, buf); err != nil {
+				r.err = err
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeUncompressedData:
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:checksumSize]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - checksumSize
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.decoded[:n], false) {
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeStreamIdentifier:
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(magicBody) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.buf[:len(magicBody)], false) {
+				return r.err
+			}
+			for i := 0; i < len(magicBody); i++ {
+				if r.buf[i] != magicBody[i] {
+					r.err = ErrCorrupt
+					return r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			r.err = ErrUnsupported
+			return r.err
+		}
+		// Section 4.4 Padding (chunk type 0xfe).
+		// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+		if !r.readFull(r.buf[:chunkLen], false) {
+			return r.err
+		}
+	}
+
+	return nil
+}
+
+// Read satisfies the io.Reader interface.
+func (r *Reader) Read(p []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	n := copy(p, r.decoded[r.i:r.j])
+	r.i += n
+	return n, nil
+}
+
+// ReadByte satisfies the io.ByteReader interface.
+func (r *Reader) ReadByte() (byte, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	c := r.decoded[r.i]
+	r.i++
+	return c, nil
+}

vendor/github.com/klauspost/compress/internal/snapref/decode_other.go

@@ -0,0 +1,113 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+// decode writes the decoding of src to dst. It assumes that the varint-encoded
+// length of the decompressed bytes has already been read, and that len(dst)
+// equals that length.
+//
+// It returns 0 on success or a decodeErrCodeXxx error code on failure.
+func decode(dst, src []byte) int {
+	var d, s, offset, length int
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case tagLiteral:
+			x := uint32(src[s] >> 2)
+			switch {
+			case x < 60:
+				s++
+			case x == 60:
+				s += 2
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-1])
+			case x == 61:
+				s += 3
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-2]) | uint32(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+			}
+			length = int(x) + 1
+			if length <= 0 {
+				return decodeErrCodeUnsupportedLiteralLength
+			}
+			if length > len(dst)-d || length > len(src)-s {
+				return decodeErrCodeCorrupt
+			}
+			copy(dst[d:], src[s:s+length])
+			d += length
+			s += length
+			continue
+
+		case tagCopy1:
+			s += 2
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1]))
+
+		case tagCopy2:
+			s += 3
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8)
+
+		case tagCopy4:
+			s += 5
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-5])>>2
+			offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24)
+		}
+
+		if offset <= 0 || d < offset || length > len(dst)-d {
+			return decodeErrCodeCorrupt
+		}
+		// Copy from an earlier sub-slice of dst to a later sub-slice.
+		// If no overlap, use the built-in copy:
+		if offset >= length {
+			copy(dst[d:d+length], dst[d-offset:])
+			d += length
+			continue
+		}
+
+		// Unlike the built-in copy function, this byte-by-byte copy always runs
+		// forwards, even if the slices overlap. Conceptually, this is:
+		//
+		// d += forwardCopy(dst[d:d+length], dst[d-offset:])
+		//
+		// We align the slices into a and b and show the compiler they are the same size.
+		// This allows the loop to run without bounds checks.
+		a := dst[d : d+length]
+		b := dst[d-offset:]
+		b = b[:len(a)]
+		for i := range a {
+			a[i] = b[i]
+		}
+		d += length
+	}
+	if d != len(dst) {
+		return decodeErrCodeCorrupt
+	}
+	return 0
+}

vendor/github.com/klauspost/compress/internal/snapref/encode.go

@@ -0,0 +1,289 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// Encode returns the encoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire encoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Encode handles the Snappy block format, not the Snappy stream format.
+func Encode(dst, src []byte) []byte {
+	if n := MaxEncodedLen(len(src)); n < 0 {
+		panic(ErrTooLarge)
+	} else if len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	// The block starts with the varint-encoded length of the decompressed bytes.
+	d := binary.PutUvarint(dst, uint64(len(src)))
+
+	for len(src) > 0 {
+		p := src
+		src = nil
+		if len(p) > maxBlockSize {
+			p, src = p[:maxBlockSize], p[maxBlockSize:]
+		}
+		if len(p) < minNonLiteralBlockSize {
+			d += emitLiteral(dst[d:], p)
+		} else {
+			d += encodeBlock(dst[d:], p)
+		}
+	}
+	return dst[:d]
+}
+
+// inputMargin is the minimum number of extra input bytes to keep, inside
+// encodeBlock's inner loop. On some architectures, this margin lets us
+// implement a fast path for emitLiteral, where the copy of short (<= 16 byte)
+// literals can be implemented as a single load to and store from a 16-byte
+// register. That literal's actual length can be as short as 1 byte, so this
+// can copy up to 15 bytes too much, but that's OK as subsequent iterations of
+// the encoding loop will fix up the copy overrun, and this inputMargin ensures
+// that we don't overrun the dst and src buffers.
+const inputMargin = 16 - 1
+
+// minNonLiteralBlockSize is the minimum size of the input to encodeBlock that
+// could be encoded with a copy tag. This is the minimum with respect to the
+// algorithm used by encodeBlock, not a minimum enforced by the file format.
+//
+// The encoded output must start with at least a 1 byte literal, as there are
+// no previous bytes to copy. A minimal (1 byte) copy after that, generated
+// from an emitCopy call in encodeBlock's main loop, would require at least
+// another inputMargin bytes, for the reason above: we want any emitLiteral
+// calls inside encodeBlock's main loop to use the fast path if possible, which
+// requires being able to overrun by inputMargin bytes. Thus,
+// minNonLiteralBlockSize equals 1 + 1 + inputMargin.
+//
+// The C++ code doesn't use this exact threshold, but it could, as discussed at
+// https://groups.google.com/d/topic/snappy-compression/oGbhsdIJSJ8/discussion
+// The difference between Go (2+inputMargin) and C++ (inputMargin) is purely an
+// optimization. It should not affect the encoded form. This is tested by
+// TestSameEncodingAsCppShortCopies.
+const minNonLiteralBlockSize = 1 + 1 + inputMargin
+
+// MaxEncodedLen returns the maximum length of a snappy block, given its
+// uncompressed length.
+//
+// It will return a negative value if srcLen is too large to encode.
+func MaxEncodedLen(srcLen int) int {
+	n := uint64(srcLen)
+	if n > 0xffffffff {
+		return -1
+	}
+	// Compressed data can be defined as:
+	//    compressed := item* literal*
+	//    item       := literal* copy
+	//
+	// The trailing literal sequence has a space blowup of at most 62/60
+	// since a literal of length 60 needs one tag byte + one extra byte
+	// for length information.
+	//
+	// Item blowup is trickier to measure. Suppose the "copy" op copies
+	// 4 bytes of data. Because of a special check in the encoding code,
+	// we produce a 4-byte copy only if the offset is < 65536. Therefore
+	// the copy op takes 3 bytes to encode, and this type of item leads
+	// to at most the 62/60 blowup for representing literals.
+	//
+	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
+	// enough, it will take 5 bytes to encode the copy op. Therefore the
+	// worst case here is a one-byte literal followed by a five-byte copy.
+	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
+	//
+	// This last factor dominates the blowup, so the final estimate is:
+	n = 32 + n + n/6
+	if n > 0xffffffff {
+		return -1
+	}
+	return int(n)
+}
+
+var errClosed = errors.New("snappy: Writer is closed")
+
+// NewWriter returns a new Writer that compresses to w.
+//
+// The Writer returned does not buffer writes. There is no need to Flush or
+// Close such a Writer.
+//
+// Deprecated: the Writer returned is not suitable for many small writes, only
+// for few large writes. Use NewBufferedWriter instead, which is efficient
+// regardless of the frequency and shape of the writes, and remember to Close
+// that Writer when done.
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// NewBufferedWriter returns a new Writer that compresses to w, using the
+// framing format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+//
+// The Writer returned buffers writes. Users must call Close to guarantee all
+// data has been forwarded to the underlying io.Writer. They may also call
+// Flush zero or more times before calling Close.
+func NewBufferedWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		ibuf: make([]byte, 0, maxBlockSize),
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// Writer is an io.Writer that can write Snappy-compressed bytes.
+//
+// Writer handles the Snappy stream format, not the Snappy block format.
+type Writer struct {
+	w   io.Writer
+	err error
+
+	// ibuf is a buffer for the incoming (uncompressed) bytes.
+	//
+	// Its use is optional. For backwards compatibility, Writers created by the
+	// NewWriter function have ibuf == nil, do not buffer incoming bytes, and
+	// therefore do not need to be Flush'ed or Close'd.
+	ibuf []byte
+
+	// obuf is a buffer for the outgoing (compressed) bytes.
+	obuf []byte
+
+	// wroteStreamHeader is whether we have written the stream header.
+	wroteStreamHeader bool
+}
+
+// Reset discards the writer's state and switches the Snappy writer to write to
+// w. This permits reusing a Writer rather than allocating a new one.
+func (w *Writer) Reset(writer io.Writer) {
+	w.w = writer
+	w.err = nil
+	if w.ibuf != nil {
+		w.ibuf = w.ibuf[:0]
+	}
+	w.wroteStreamHeader = false
+}
+
+// Write satisfies the io.Writer interface.
+func (w *Writer) Write(p []byte) (nRet int, errRet error) {
+	if w.ibuf == nil {
+		// Do not buffer incoming bytes. This does not perform or compress well
+		// if the caller of Writer.Write writes many small slices. This
+		// behavior is therefore deprecated, but still supported for backwards
+		// compatibility with code that doesn't explicitly Flush or Close.
+		return w.write(p)
+	}
+
+	// The remainder of this method is based on bufio.Writer.Write from the
+	// standard library.
+
+	for len(p) > (cap(w.ibuf)-len(w.ibuf)) && w.err == nil {
+		var n int
+		if len(w.ibuf) == 0 {
+			// Large write, empty buffer.
+			// Write directly from p to avoid copy.
+			n, _ = w.write(p)
+		} else {
+			n = copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+			w.ibuf = w.ibuf[:len(w.ibuf)+n]
+			w.Flush()
+		}
+		nRet += n
+		p = p[n:]
+	}
+	if w.err != nil {
+		return nRet, w.err
+	}
+	n := copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+	w.ibuf = w.ibuf[:len(w.ibuf)+n]
+	nRet += n
+	return nRet, nil
+}
+
+func (w *Writer) write(p []byte) (nRet int, errRet error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	for len(p) > 0 {
+		obufStart := len(magicChunk)
+		if !w.wroteStreamHeader {
+			w.wroteStreamHeader = true
+			copy(w.obuf, magicChunk)
+			obufStart = 0
+		}
+
+		var uncompressed []byte
+		if len(p) > maxBlockSize {
+			uncompressed, p = p[:maxBlockSize], p[maxBlockSize:]
+		} else {
+			uncompressed, p = p, nil
+		}
+		checksum := crc(uncompressed)
+
+		// Compress the buffer, discarding the result if the improvement
+		// isn't at least 12.5%.
+		compressed := Encode(w.obuf[obufHeaderLen:], uncompressed)
+		chunkType := uint8(chunkTypeCompressedData)
+		chunkLen := 4 + len(compressed)
+		obufEnd := obufHeaderLen + len(compressed)
+		if len(compressed) >= len(uncompressed)-len(uncompressed)/8 {
+			chunkType = chunkTypeUncompressedData
+			chunkLen = 4 + len(uncompressed)
+			obufEnd = obufHeaderLen
+		}
+
+		// Fill in the per-chunk header that comes before the body.
+		w.obuf[len(magicChunk)+0] = chunkType
+		w.obuf[len(magicChunk)+1] = uint8(chunkLen >> 0)
+		w.obuf[len(magicChunk)+2] = uint8(chunkLen >> 8)
+		w.obuf[len(magicChunk)+3] = uint8(chunkLen >> 16)
+		w.obuf[len(magicChunk)+4] = uint8(checksum >> 0)
+		w.obuf[len(magicChunk)+5] = uint8(checksum >> 8)
+		w.obuf[len(magicChunk)+6] = uint8(checksum >> 16)
+		w.obuf[len(magicChunk)+7] = uint8(checksum >> 24)
+
+		if _, err := w.w.Write(w.obuf[obufStart:obufEnd]); err != nil {
+			w.err = err
+			return nRet, err
+		}
+		if chunkType == chunkTypeUncompressedData {
+			if _, err := w.w.Write(uncompressed); err != nil {
+				w.err = err
+				return nRet, err
+			}
+		}
+		nRet += len(uncompressed)
+	}
+	return nRet, nil
+}
+
+// Flush flushes the Writer to its underlying io.Writer.
+func (w *Writer) Flush() error {
+	if w.err != nil {
+		return w.err
+	}
+	if len(w.ibuf) == 0 {
+		return nil
+	}
+	w.write(w.ibuf)
+	w.ibuf = w.ibuf[:0]
+	return w.err
+}
+
+// Close calls Flush and then closes the Writer.
+func (w *Writer) Close() error {
+	w.Flush()
+	ret := w.err
+	if w.err == nil {
+		w.err = errClosed
+	}
+	return ret
+}

vendor/github.com/klauspost/compress/internal/snapref/encode_other.go

@@ -0,0 +1,236 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+func load32(b []byte, i int) uint32 {
+	b = b[i : i+4 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+}
+
+func load64(b []byte, i int) uint64 {
+	b = b[i : i+8 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 |
+		uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+}
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+//
+// It assumes that:
+//	dst is long enough to hold the encoded bytes
+//	1 <= len(lit) && len(lit) <= 65536
+func emitLiteral(dst, lit []byte) int {
+	i, n := 0, uint(len(lit)-1)
+	switch {
+	case n < 60:
+		dst[0] = uint8(n)<<2 | tagLiteral
+		i = 1
+	case n < 1<<8:
+		dst[0] = 60<<2 | tagLiteral
+		dst[1] = uint8(n)
+		i = 2
+	default:
+		dst[0] = 61<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		i = 3
+	}
+	return i + copy(dst[i:], lit)
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+//
+// It assumes that:
+//	dst is long enough to hold the encoded bytes
+//	1 <= offset && offset <= 65535
+//	4 <= length && length <= 65535
+func emitCopy(dst []byte, offset, length int) int {
+	i := 0
+	// The maximum length for a single tagCopy1 or tagCopy2 op is 64 bytes. The
+	// threshold for this loop is a little higher (at 68 = 64 + 4), and the
+	// length emitted down below is is a little lower (at 60 = 64 - 4), because
+	// it's shorter to encode a length 67 copy as a length 60 tagCopy2 followed
+	// by a length 7 tagCopy1 (which encodes as 3+2 bytes) than to encode it as
+	// a length 64 tagCopy2 followed by a length 3 tagCopy2 (which encodes as
+	// 3+3 bytes). The magic 4 in the 64±4 is because the minimum length for a
+	// tagCopy1 op is 4 bytes, which is why a length 3 copy has to be an
+	// encodes-as-3-bytes tagCopy2 instead of an encodes-as-2-bytes tagCopy1.
+	for length >= 68 {
+		// Emit a length 64 copy, encoded as 3 bytes.
+		dst[i+0] = 63<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 64
+	}
+	if length > 64 {
+		// Emit a length 60 copy, encoded as 3 bytes.
+		dst[i+0] = 59<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 60
+	}
+	if length >= 12 || offset >= 2048 {
+		// Emit the remaining copy, encoded as 3 bytes.
+		dst[i+0] = uint8(length-1)<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		return i + 3
+	}
+	// Emit the remaining copy, encoded as 2 bytes.
+	dst[i+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
+	dst[i+1] = uint8(offset)
+	return i + 2
+}
+
+// extendMatch returns the largest k such that k <= len(src) and that
+// src[i:i+k-j] and src[j:k] have the same contents.
+//
+// It assumes that:
+//	0 <= i && i < j && j <= len(src)
+func extendMatch(src []byte, i, j int) int {
+	for ; j < len(src) && src[i] == src[j]; i, j = i+1, j+1 {
+	}
+	return j
+}
+
+func hash(u, shift uint32) uint32 {
+	return (u * 0x1e35a7bd) >> shift
+}
+
+// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It
+// assumes that the varint-encoded length of the decompressed bytes has already
+// been written.
+//
+// It also assumes that:
+//	len(dst) >= MaxEncodedLen(len(src)) &&
+// 	minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
+func encodeBlock(dst, src []byte) (d int) {
+	// Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive.
+	// The table element type is uint16, as s < sLimit and sLimit < len(src)
+	// and len(src) <= maxBlockSize and maxBlockSize == 65536.
+	const (
+		maxTableSize = 1 << 14
+		// tableMask is redundant, but helps the compiler eliminate bounds
+		// checks.
+		tableMask = maxTableSize - 1
+	)
+	shift := uint32(32 - 8)
+	for tableSize := 1 << 8; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 {
+		shift--
+	}
+	// In Go, all array elements are zero-initialized, so there is no advantage
+	// to a smaller tableSize per se. However, it matches the C++ algorithm,
+	// and in the asm versions of this code, we can get away with zeroing only
+	// the first tableSize elements.
+	var table [maxTableSize]uint16
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := len(src) - inputMargin
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := 0
+
+	// The encoded form must start with a literal, as there are no previous
+	// bytes to copy, so we start looking for hash matches at s == 1.
+	s := 1
+	nextHash := hash(load32(src, s), shift)
+
+	for {
+		// Copied from the C++ snappy implementation:
+		//
+		// Heuristic match skipping: If 32 bytes are scanned with no matches
+		// found, start looking only at every other byte. If 32 more bytes are
+		// scanned (or skipped), look at every third byte, etc.. When a match
+		// is found, immediately go back to looking at every byte. This is a
+		// small loss (~5% performance, ~0.1% density) for compressible data
+		// due to more bookkeeping, but for non-compressible data (such as
+		// JPEG) it's a huge win since the compressor quickly "realizes" the
+		// data is incompressible and doesn't bother looking for matches
+		// everywhere.
+		//
+		// The "skip" variable keeps track of how many bytes there are since
+		// the last match; dividing it by 32 (ie. right-shifting by five) gives
+		// the number of bytes to move ahead for each iteration.
+		skip := 32
+
+		nextS := s
+		candidate := 0
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidate = int(table[nextHash&tableMask])
+			table[nextHash&tableMask] = uint16(s)
+			nextHash = hash(load32(src, nextS), shift)
+			if load32(src, s) == load32(src, candidate) {
+				break
+			}
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+		d += emitLiteral(dst[d:], src[nextEmit:s])
+
+		// Call emitCopy, and then see if another emitCopy could be our next
+		// move. Repeat until we find no match for the input immediately after
+		// what was consumed by the last emitCopy call.
+		//
+		// If we exit this loop normally then we need to call emitLiteral next,
+		// though we don't yet know how big the literal will be. We handle that
+		// by proceeding to the next iteration of the main loop. We also can
+		// exit this loop via goto if we get close to exhausting the input.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+			base := s
+
+			// Extend the 4-byte match as long as possible.
+			//
+			// This is an inlined version of:
+			//	s = extendMatch(src, candidate+4, s+4)
+			s += 4
+			for i := candidate + 4; s < len(src) && src[i] == src[s]; i, s = i+1, s+1 {
+			}
+
+			d += emitCopy(dst[d:], base-candidate, s-base)
+			nextEmit = s
+			if s >= sLimit {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-1 and at s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load64(src, s-1)
+			prevHash := hash(uint32(x>>0), shift)
+			table[prevHash&tableMask] = uint16(s - 1)
+			currHash := hash(uint32(x>>8), shift)
+			candidate = int(table[currHash&tableMask])
+			table[currHash&tableMask] = uint16(s)
+			if uint32(x>>8) != load32(src, candidate) {
+				nextHash = hash(uint32(x>>16), shift)
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if nextEmit < len(src) {
+		d += emitLiteral(dst[d:], src[nextEmit:])
+	}
+	return d
+}

vendor/github.com/klauspost/compress/internal/snapref/snappy.go

@@ -0,0 +1,98 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package snapref implements the Snappy compression format. It aims for very
+// high speeds and reasonable compression.
+//
+// There are actually two Snappy formats: block and stream. They are related,
+// but different: trying to decompress block-compressed data as a Snappy stream
+// will fail, and vice versa. The block format is the Decode and Encode
+// functions and the stream format is the Reader and Writer types.
+//
+// The block format, the more common case, is used when the complete size (the
+// number of bytes) of the original data is known upfront, at the time
+// compression starts. The stream format, also known as the framing format, is
+// for when that isn't always true.
+//
+// The canonical, C++ implementation is at https://github.com/google/snappy and
+// it only implements the block format.
+package snapref
+
+import (
+	"hash/crc32"
+)
+
+/*
+Each encoded block begins with the varint-encoded length of the decoded data,
+followed by a sequence of chunks. Chunks begin and end on byte boundaries. The
+first byte of each chunk is broken into its 2 least and 6 most significant bits
+called l and m: l ranges in [0, 4) and m ranges in [0, 64). l is the chunk tag.
+Zero means a literal tag. All other values mean a copy tag.
+
+For literal tags:
+  - If m < 60, the next 1 + m bytes are literal bytes.
+  - Otherwise, let n be the little-endian unsigned integer denoted by the next
+    m - 59 bytes. The next 1 + n bytes after that are literal bytes.
+
+For copy tags, length bytes are copied from offset bytes ago, in the style of
+Lempel-Ziv compression algorithms. In particular:
+  - For l == 1, the offset ranges in [0, 1<<11) and the length in [4, 12).
+    The length is 4 + the low 3 bits of m. The high 3 bits of m form bits 8-10
+    of the offset. The next byte is bits 0-7 of the offset.
+  - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65).
+    The length is 1 + m. The offset is the little-endian unsigned integer
+    denoted by the next 2 bytes.
+  - For l == 3, this tag is a legacy format that is no longer issued by most
+    encoders. Nonetheless, the offset ranges in [0, 1<<32) and the length in
+    [1, 65). The length is 1 + m. The offset is the little-endian unsigned
+    integer denoted by the next 4 bytes.
+*/
+const (
+	tagLiteral = 0x00
+	tagCopy1   = 0x01
+	tagCopy2   = 0x02
+	tagCopy4   = 0x03
+)
+
+const (
+	checksumSize    = 4
+	chunkHeaderSize = 4
+	magicChunk      = "\xff\x06\x00\x00" + magicBody
+	magicBody       = "sNaPpY"
+
+	// maxBlockSize is the maximum size of the input to encodeBlock. It is not
+	// part of the wire format per se, but some parts of the encoder assume
+	// that an offset fits into a uint16.
+	//
+	// Also, for the framing format (Writer type instead of Encode function),
+	// https://github.com/google/snappy/blob/master/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536
+	// bytes".
+	maxBlockSize = 65536
+
+	// maxEncodedLenOfMaxBlockSize equals MaxEncodedLen(maxBlockSize), but is
+	// hard coded to be a const instead of a variable, so that obufLen can also
+	// be a const. Their equivalence is confirmed by
+	// TestMaxEncodedLenOfMaxBlockSize.
+	maxEncodedLenOfMaxBlockSize = 76490
+
+	obufHeaderLen = len(magicChunk) + checksumSize + chunkHeaderSize
+	obufLen       = obufHeaderLen + maxEncodedLenOfMaxBlockSize
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func crc(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return uint32(c>>15|c<<17) + 0xa282ead8
+}

vendor/github.com/klauspost/compress/s2sx.mod

@@ -0,0 +1,4 @@
+module github.com/klauspost/compress
+
+go 1.16
+

vendor/github.com/klauspost/compress/zstd/README.md

@@ -0,0 +1,439 @@
+# zstd 
+
+[Zstandard](https://facebook.github.io/zstd/) is a real-time compression algorithm, providing high compression ratios. 
+It offers a very wide range of compression / speed trade-off, while being backed by a very fast decoder.
+A high performance compression algorithm is implemented. For now focused on speed. 
+
+This package provides [compression](#Compressor) to and [decompression](#Decompressor) of Zstandard content. 
+
+This package is pure Go and without use of "unsafe". 
+
+The `zstd` package is provided as open source software using a Go standard license.
+
+Currently the package is heavily optimized for 64 bit processors and will be significantly slower on 32 bit processors.
+
+## Installation
+
+Install using `go get -u github.com/klauspost/compress`. The package is located in `github.com/klauspost/compress/zstd`.
+
+[![Go Reference](https://pkg.go.dev/badge/github.com/klauspost/compress/zstd.svg)](https://pkg.go.dev/github.com/klauspost/compress/zstd)
+
+## Compressor
+
+### Status: 
+
+STABLE - there may always be subtle bugs, a wide variety of content has been tested and the library is actively 
+used by several projects. This library is being [fuzz-tested](https://github.com/klauspost/compress-fuzz) for all updates.
+
+There may still be specific combinations of data types/size/settings that could lead to edge cases, 
+so as always, testing is recommended.  
+
+For now, a high speed (fastest) and medium-fast (default) compressor has been implemented. 
+
+* The "Fastest" compression ratio is roughly equivalent to zstd level 1. 
+* The "Default" compression ratio is roughly equivalent to zstd level 3 (default).
+* The "Better" compression ratio is roughly equivalent to zstd level 7.
+* The "Best" compression ratio is roughly equivalent to zstd level 11.
+
+In terms of speed, it is typically 2x as fast as the stdlib deflate/gzip in its fastest mode. 
+The compression ratio compared to stdlib is around level 3, but usually 3x as fast.
+
+ 
+### Usage
+
+An Encoder can be used for either compressing a stream via the
+`io.WriteCloser` interface supported by the Encoder or as multiple independent
+tasks via the `EncodeAll` function.
+Smaller encodes are encouraged to use the EncodeAll function.
+Use `NewWriter` to create a new instance that can be used for both.
+
+To create a writer with default options, do like this:
+
+```Go
+// Compress input to output.
+func Compress(in io.Reader, out io.Writer) error {
+    enc, err := zstd.NewWriter(out)
+    if err != nil {
+        return err
+    }
+    _, err = io.Copy(enc, in)
+    if err != nil {
+        enc.Close()
+        return err
+    }
+    return enc.Close()
+}
+```
+
+Now you can encode by writing data to `enc`. The output will be finished writing when `Close()` is called.
+Even if your encode fails, you should still call `Close()` to release any resources that may be held up.  
+
+The above is fine for big encodes. However, whenever possible try to *reuse* the writer.
+
+To reuse the encoder, you can use the `Reset(io.Writer)` function to change to another output. 
+This will allow the encoder to reuse all resources and avoid wasteful allocations. 
+
+Currently stream encoding has 'light' concurrency, meaning up to 2 goroutines can be working on part 
+of a stream. This is independent of the `WithEncoderConcurrency(n)`, but that is likely to change 
+in the future. So if you want to limit concurrency for future updates, specify the concurrency
+you would like.
+
+If you would like stream encoding to be done without spawning async goroutines, use `WithEncoderConcurrency(1)`
+which will compress input as each block is completed, blocking on writes until each has completed.
+
+You can specify your desired compression level using `WithEncoderLevel()` option. Currently only pre-defined 
+compression settings can be specified.
+
+#### Future Compatibility Guarantees
+
+This will be an evolving project. When using this package it is important to note that both the compression efficiency and speed may change.
+
+The goal will be to keep the default efficiency at the default zstd (level 3). 
+However the encoding should never be assumed to remain the same, 
+and you should not use hashes of compressed output for similarity checks.
+
+The Encoder can be assumed to produce the same output from the exact same code version.
+However, the may be modes in the future that break this, 
+although they will not be enabled without an explicit option.   
+
+This encoder is not designed to (and will probably never) output the exact same bitstream as the reference encoder.
+
+Also note, that the cgo decompressor currently does not [report all errors on invalid input](https://github.com/DataDog/zstd/issues/59),
+[omits error checks](https://github.com/DataDog/zstd/issues/61), [ignores checksums](https://github.com/DataDog/zstd/issues/43) 
+and seems to ignore concatenated streams, even though [it is part of the spec](https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frames).
+
+#### Blocks
+
+For compressing small blocks, the returned encoder has a function called `EncodeAll(src, dst []byte) []byte`.
+
+`EncodeAll` will encode all input in src and append it to dst.
+This function can be called concurrently. 
+Each call will only run on a same goroutine as the caller.
+
+Encoded blocks can be concatenated and the result will be the combined input stream.
+Data compressed with EncodeAll can be decoded with the Decoder, using either a stream or `DecodeAll`.
+
+Especially when encoding blocks you should take special care to reuse the encoder. 
+This will effectively make it run without allocations after a warmup period. 
+To make it run completely without allocations, supply a destination buffer with space for all content.   
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+// Create a writer that caches compressors.
+// For this operation type we supply a nil Reader.
+var encoder, _ = zstd.NewWriter(nil)
+
+// Compress a buffer. 
+// If you have a destination buffer, the allocation in the call can also be eliminated.
+func Compress(src []byte) []byte {
+    return encoder.EncodeAll(src, make([]byte, 0, len(src)))
+} 
+```
+
+You can control the maximum number of concurrent encodes using the `WithEncoderConcurrency(n)` 
+option when creating the writer.
+
+Using the Encoder for both a stream and individual blocks concurrently is safe. 
+
+### Performance
+
+I have collected some speed examples to compare speed and compression against other compressors.
+
+* `file` is the input file.
+* `out` is the compressor used. `zskp` is this package. `zstd` is the Datadog cgo library. `gzstd/gzkp` is gzip standard and this library.
+* `level` is the compression level used. For `zskp` level 1 is "fastest", level 2 is "default"; 3 is "better", 4 is "best".
+* `insize`/`outsize` is the input/output size.
+* `millis` is the number of milliseconds used for compression.
+* `mb/s` is megabytes (2^20 bytes) per second.
+
+```
+Silesia Corpus:
+http://sun.aei.polsl.pl/~sdeor/corpus/silesia.zip
+
+This package:
+file    out     level   insize      outsize     millis  mb/s
+silesia.tar zskp    1   211947520   73821326    634     318.47
+silesia.tar zskp    2   211947520   67655404    1508    133.96
+silesia.tar zskp    3   211947520   64746933    3000    67.37
+silesia.tar zskp    4   211947520   60073508    16926   11.94
+
+cgo zstd:
+silesia.tar zstd    1   211947520   73605392    543     371.56
+silesia.tar zstd    3   211947520   66793289    864     233.68
+silesia.tar zstd    6   211947520   62916450    1913    105.66
+silesia.tar zstd    9   211947520   60212393    5063    39.92
+
+gzip, stdlib/this package:
+silesia.tar gzstd   1   211947520   80007735    1498    134.87
+silesia.tar gzkp    1   211947520   80088272    1009    200.31
+
+GOB stream of binary data. Highly compressible.
+https://files.klauspost.com/compress/gob-stream.7z
+
+file        out     level   insize  outsize     millis  mb/s
+gob-stream  zskp    1   1911399616  233948096   3230    564.34
+gob-stream  zskp    2   1911399616  203997694   4997    364.73
+gob-stream  zskp    3   1911399616  173526523   13435   135.68
+gob-stream  zskp    4   1911399616  162195235   47559   38.33
+
+gob-stream  zstd    1   1911399616  249810424   2637    691.26
+gob-stream  zstd    3   1911399616  208192146   3490    522.31
+gob-stream  zstd    6   1911399616  193632038   6687    272.56
+gob-stream  zstd    9   1911399616  177620386   16175   112.70
+
+gob-stream  gzstd   1   1911399616  357382013   9046    201.49
+gob-stream  gzkp    1   1911399616  359136669   4885    373.08
+
+The test data for the Large Text Compression Benchmark is the first
+10^9 bytes of the English Wikipedia dump on Mar. 3, 2006.
+http://mattmahoney.net/dc/textdata.html
+
+file    out level   insize      outsize     millis  mb/s
+enwik9  zskp    1   1000000000  343833605   3687    258.64
+enwik9  zskp    2   1000000000  317001237   7672    124.29
+enwik9  zskp    3   1000000000  291915823   15923   59.89
+enwik9  zskp    4   1000000000  261710291   77697   12.27
+
+enwik9  zstd    1   1000000000  358072021   3110    306.65
+enwik9  zstd    3   1000000000  313734672   4784    199.35
+enwik9  zstd    6   1000000000  295138875   10290   92.68
+enwik9  zstd    9   1000000000  278348700   28549   33.40
+
+enwik9  gzstd   1   1000000000  382578136   8608    110.78
+enwik9  gzkp    1   1000000000  382781160   5628    169.45
+
+Highly compressible JSON file.
+https://files.klauspost.com/compress/github-june-2days-2019.json.zst
+
+file                        out level   insize      outsize     millis  mb/s
+github-june-2days-2019.json zskp    1   6273951764  697439532   9789    611.17
+github-june-2days-2019.json zskp    2   6273951764  610876538   18553   322.49
+github-june-2days-2019.json zskp    3   6273951764  517662858   44186   135.41
+github-june-2days-2019.json zskp    4   6273951764  464617114   165373  36.18
+
+github-june-2days-2019.json zstd    1   6273951764  766284037   8450    708.00
+github-june-2days-2019.json zstd    3   6273951764  661889476   10927   547.57
+github-june-2days-2019.json zstd    6   6273951764  642756859   22996   260.18
+github-june-2days-2019.json zstd    9   6273951764  601974523   52413   114.16
+
+github-june-2days-2019.json gzstd   1   6273951764  1164397768  26793   223.32
+github-june-2days-2019.json gzkp    1   6273951764  1120631856  17693   338.16
+
+VM Image, Linux mint with a few installed applications:
+https://files.klauspost.com/compress/rawstudio-mint14.7z
+
+file                    out level   insize      outsize     millis  mb/s
+rawstudio-mint14.tar    zskp    1   8558382592  3718400221  18206   448.29
+rawstudio-mint14.tar    zskp    2   8558382592  3326118337  37074   220.15
+rawstudio-mint14.tar    zskp    3   8558382592  3163842361  87306   93.49
+rawstudio-mint14.tar    zskp    4   8558382592  2970480650  783862  10.41
+
+rawstudio-mint14.tar    zstd    1   8558382592  3609250104  17136   476.27
+rawstudio-mint14.tar    zstd    3   8558382592  3341679997  29262   278.92
+rawstudio-mint14.tar    zstd    6   8558382592  3235846406  77904   104.77
+rawstudio-mint14.tar    zstd    9   8558382592  3160778861  140946  57.91
+
+rawstudio-mint14.tar    gzstd   1   8558382592  3926234992  51345   158.96
+rawstudio-mint14.tar    gzkp    1   8558382592  3960117298  36722   222.26
+
+CSV data:
+https://files.klauspost.com/compress/nyc-taxi-data-10M.csv.zst
+
+file                    out level   insize      outsize     millis  mb/s
+nyc-taxi-data-10M.csv   zskp    1   3325605752  641319332   9462    335.17
+nyc-taxi-data-10M.csv   zskp    2   3325605752  588976126   17570   180.50
+nyc-taxi-data-10M.csv   zskp    3   3325605752  529329260   32432   97.79
+nyc-taxi-data-10M.csv   zskp    4   3325605752  474949772   138025  22.98
+
+nyc-taxi-data-10M.csv   zstd    1   3325605752  687399637   8233    385.18
+nyc-taxi-data-10M.csv   zstd    3   3325605752  598514411   10065   315.07
+nyc-taxi-data-10M.csv   zstd    6   3325605752  570522953   20038   158.27
+nyc-taxi-data-10M.csv   zstd    9   3325605752  517554797   64565   49.12
+
+nyc-taxi-data-10M.csv   gzstd   1   3325605752  928654908   21270   149.11
+nyc-taxi-data-10M.csv   gzkp    1   3325605752  922273214   13929   227.68
+```
+
+## Decompressor
+
+Staus: STABLE - there may still be subtle bugs, but a wide variety of content has been tested.
+
+This library is being continuously [fuzz-tested](https://github.com/klauspost/compress-fuzz),
+kindly supplied by [fuzzit.dev](https://fuzzit.dev/). 
+The main purpose of the fuzz testing is to ensure that it is not possible to crash the decoder, 
+or run it past its limits with ANY input provided.  
+ 
+### Usage
+
+The package has been designed for two main usages, big streams of data and smaller in-memory buffers. 
+There are two main usages of the package for these. Both of them are accessed by creating a `Decoder`.
+
+For streaming use a simple setup could look like this:
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+func Decompress(in io.Reader, out io.Writer) error {
+    d, err := zstd.NewReader(in)
+    if err != nil {
+        return err
+    }
+    defer d.Close()
+    
+    // Copy content...
+    _, err = io.Copy(out, d)
+    return err
+}
+```
+
+It is important to use the "Close" function when you no longer need the Reader to stop running goroutines, 
+when running with default settings.
+Goroutines will exit once an error has been returned, including `io.EOF` at the end of a stream.
+
+Streams are decoded concurrently in 4 asynchronous stages to give the best possible throughput.
+However, if you prefer synchronous decompression, use `WithDecoderConcurrency(1)` which will decompress data 
+as it is being requested only.
+
+For decoding buffers, it could look something like this:
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+// Create a reader that caches decompressors.
+// For this operation type we supply a nil Reader.
+var decoder, _ = zstd.NewReader(nil, WithDecoderConcurrency(0))
+
+// Decompress a buffer. We don't supply a destination buffer,
+// so it will be allocated by the decoder.
+func Decompress(src []byte) ([]byte, error) {
+    return decoder.DecodeAll(src, nil)
+} 
+```
+
+Both of these cases should provide the functionality needed. 
+The decoder can be used for *concurrent* decompression of multiple buffers.
+By default 4 decompressors will be created. 
+
+It will only allow a certain number of concurrent operations to run. 
+To tweak that yourself use the `WithDecoderConcurrency(n)` option when creating the decoder.
+It is possible to use `WithDecoderConcurrency(0)` to create GOMAXPROCS decoders.
+
+### Dictionaries
+
+Data compressed with [dictionaries](https://github.com/facebook/zstd#the-case-for-small-data-compression) can be decompressed.
+
+Dictionaries are added individually to Decoders.
+Dictionaries are generated by the `zstd --train` command and contains an initial state for the decoder.
+To add a dictionary use the `WithDecoderDicts(dicts ...[]byte)` option with the dictionary data.
+Several dictionaries can be added at once.
+
+The dictionary will be used automatically for the data that specifies them.
+A re-used Decoder will still contain the dictionaries registered.
+
+When registering multiple dictionaries with the same ID, the last one will be used.
+
+It is possible to use dictionaries when compressing data.
+
+To enable a dictionary use `WithEncoderDict(dict []byte)`. Here only one dictionary will be used 
+and it will likely be used even if it doesn't improve compression. 
+
+The used dictionary must be used to decompress the content.
+
+For any real gains, the dictionary should be built with similar data. 
+If an unsuitable dictionary is used the output may be slightly larger than using no dictionary.
+Use the [zstd commandline tool](https://github.com/facebook/zstd/releases) to build a dictionary from sample data.
+For information see [zstd dictionary information](https://github.com/facebook/zstd#the-case-for-small-data-compression). 
+
+For now there is a fixed startup performance penalty for compressing content with dictionaries. 
+This will likely be improved over time. Just be aware to test performance when implementing.  
+
+### Allocation-less operation
+
+The decoder has been designed to operate without allocations after a warmup. 
+
+This means that you should *store* the decoder for best performance. 
+To re-use a stream decoder, use the `Reset(r io.Reader) error` to switch to another stream.
+A decoder can safely be re-used even if the previous stream failed.
+
+To release the resources, you must call the `Close()` function on a decoder.
+After this it can *no longer be reused*, but all running goroutines will be stopped.
+So you *must* use this if you will no longer need the Reader.
+
+For decompressing smaller buffers a single decoder can be used.
+When decoding buffers, you can supply a destination slice with length 0 and your expected capacity.
+In this case no unneeded allocations should be made. 
+
+### Concurrency
+
+The buffer decoder does everything on the same goroutine and does nothing concurrently.
+It can however decode several buffers concurrently. Use `WithDecoderConcurrency(n)` to limit that.
+
+The stream decoder will create goroutines that:
+
+1) Reads input and splits the input into blocks.
+2) Decompression of literals.
+3) Decompression of sequences.
+4) Reconstruction of output stream.
+
+So effectively this also means the decoder will "read ahead" and prepare data to always be available for output.
+
+The concurrency level will, for streams, determine how many blocks ahead the compression will start.
+
+Since "blocks" are quite dependent on the output of the previous block stream decoding will only have limited concurrency.
+
+In practice this means that concurrency is often limited to utilizing about 3 cores effectively.
+  
+### Benchmarks
+
+The first two are streaming decodes and the last are smaller inputs. 
+
+Running on AMD Ryzen 9 3950X 16-Core Processor. AMD64 assembly used.
+
+```
+BenchmarkDecoderSilesia-32    	                   5	 206878840 ns/op	1024.50 MB/s	   49808 B/op	      43 allocs/op
+BenchmarkDecoderEnwik9-32                          1	1271809000 ns/op	 786.28 MB/s	   72048 B/op	      52 allocs/op
+
+Concurrent blocks, performance:
+
+BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-32         	   67356	     17857 ns/op	10321.96 MB/s	        22.48 pct	     102 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-32     	  266656	      4421 ns/op	26823.21 MB/s	        11.89 pct	      19 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-32      	   20992	     56842 ns/op	8477.17 MB/s	        39.90 pct	     754 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-32        	   27456	     43932 ns/op	9714.01 MB/s	        33.27 pct	     524 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-32      	   78432	     15047 ns/op	8319.15 MB/s	        40.34 pct	      66 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-32       	   65800	     18436 ns/op	8249.63 MB/s	        37.75 pct	      88 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-32          	  102993	     11523 ns/op	35546.09 MB/s	         3.637 pct	     143 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-32    	 1000000	      1070 ns/op	95720.98 MB/s	        80.53 pct	       3 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-32    	  749802	      1752 ns/op	70272.35 MB/s	       100.0 pct	       5 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-32          	   22640	     52934 ns/op	13263.37 MB/s	        26.25 pct	    1014 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html.zst-32              	  226412	      5232 ns/op	19572.27 MB/s	        14.49 pct	      20 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-32     	  923041	      1276 ns/op	3194.71 MB/s	        31.26 pct	       0 B/op	       0 allocs/op
+```
+
+This reflects the performance around May 2022, but this may be out of date.
+
+## Zstd inside ZIP files
+
+It is possible to use zstandard to compress individual files inside zip archives.
+While this isn't widely supported it can be useful for internal files.
+
+To support the compression and decompression of these files you must register a compressor and decompressor.
+
+It is highly recommended registering the (de)compressors on individual zip Reader/Writer and NOT
+use the global registration functions. The main reason for this is that 2 registrations from 
+different packages will result in a panic.
+
+It is a good idea to only have a single compressor and decompressor, since they can be used for multiple zip
+files concurrently, and using a single instance will allow reusing some resources.
+
+See [this example](https://pkg.go.dev/github.com/klauspost/compress/zstd#example-ZipCompressor) for 
+how to compress and decompress files inside zip archives.
+
+# Contributions
+
+Contributions are always welcome. 
+For new features/fixes, remember to add tests and for performance enhancements include benchmarks.
+
+For general feedback and experience reports, feel free to open an issue or write me on [Twitter](https://twitter.com/sh0dan).
+
+This package includes the excellent [`github.com/cespare/xxhash`](https://github.com/cespare/xxhash) package Copyright (c) 2016 Caleb Spare.

vendor/github.com/klauspost/compress/zstd/bitreader.go

@@ -0,0 +1,147 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"math/bits"
+)
+
+// bitReader reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReader struct {
+	in       []byte
+	off      uint   // next byte to read is at in[off - 1]
+	value    uint64 // Maybe use [16]byte, but shifting is awkward.
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReader) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.bitsRead += 8 - uint8(highBits(uint32(v)))
+	return nil
+}
+
+// getBits will return n bits. n can be 0.
+func (b *bitReader) getBits(n uint8) int {
+	if n == 0 /*|| b.bitsRead >= 64 */ {
+		return 0
+	}
+	return int(b.get32BitsFast(n))
+}
+
+// get32BitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReader) get32BitsFast(n uint8) uint32 {
+	const regMask = 64 - 1
+	v := uint32((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
+	b.bitsRead += n
+	return v
+}
+
+func (b *bitReader) get16BitsFast(n uint8) uint16 {
+	const regMask = 64 - 1
+	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
+	b.bitsRead += n
+	return v
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReader) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+	// 2 bounds checks.
+	v := b.in[b.off-4:]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value = (b.value << 32) | uint64(low)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
+func (b *bitReader) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReader) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off >= 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value = (b.value << 32) | uint64(low)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value = (b.value << 8) | uint64(b.in[b.off-1])
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReader) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+// overread returns true if more bits have been requested than is on the stream.
+func (b *bitReader) overread() bool {
+	return b.bitsRead > 64
+}
+
+// remain returns the number of bits remaining.
+func (b *bitReader) remain() uint {
+	return b.off*8 + 64 - uint(b.bitsRead)
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReader) close() error {
+	// Release reference.
+	b.in = nil
+	if !b.finished() {
+		return fmt.Errorf("%d extra bits on block, should be 0", b.remain())
+	}
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+
+func highBits(val uint32) (n uint32) {
+	return uint32(bits.Len32(val) - 1)
+}

vendor/github.com/klauspost/compress/zstd/bitwriter.go

@@ -0,0 +1,189 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package zstd
+
+import "fmt"
+
+// bitWriter will write bits.
+// First bit will be LSB of the first byte of output.
+type bitWriter struct {
+	bitContainer uint64
+	nBits        uint8
+	out          []byte
+}
+
+// bitMask16 is bitmasks. Has extra to avoid bounds check.
+var bitMask16 = [32]uint16{
+	0, 1, 3, 7, 0xF, 0x1F,
+	0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
+	0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF} /* up to 16 bits */
+
+var bitMask32 = [32]uint32{
+	0, 1, 3, 7, 0xF, 0x1F, 0x3F, 0x7F, 0xFF,
+	0x1FF, 0x3FF, 0x7FF, 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF,
+	0x1ffff, 0x3ffff, 0x7FFFF, 0xfFFFF, 0x1fFFFF, 0x3fFFFF, 0x7fFFFF, 0xffFFFF,
+	0x1ffFFFF, 0x3ffFFFF, 0x7ffFFFF, 0xfffFFFF, 0x1fffFFFF, 0x3fffFFFF, 0x7fffFFFF,
+} // up to 32 bits
+
+// addBits16NC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16NC(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits32NC will add up to 31 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits32NC(value uint32, bits uint8) {
+	b.bitContainer |= uint64(value&bitMask32[bits&31]) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits64NC will add up to 64 bits.
+// There must be space for 32 bits.
+func (b *bitWriter) addBits64NC(value uint64, bits uint8) {
+	if bits <= 31 {
+		b.addBits32Clean(uint32(value), bits)
+		return
+	}
+	b.addBits32Clean(uint32(value), 32)
+	b.flush32()
+	b.addBits32Clean(uint32(value>>32), bits-32)
+}
+
+// addBits32Clean will add up to 32 bits.
+// It will not check if there is space for them.
+// The input must not contain more bits than specified.
+func (b *bitWriter) addBits32Clean(value uint32, bits uint8) {
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// flush will flush all pending full bytes.
+// There will be at least 56 bits available for writing when this has been called.
+// Using flush32 is faster, but leaves less space for writing.
+func (b *bitWriter) flush() {
+	v := b.nBits >> 3
+	switch v {
+	case 0:
+	case 1:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+		)
+	case 2:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+		)
+	case 3:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+		)
+	case 4:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+		)
+	case 5:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+		)
+	case 6:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+		)
+	case 7:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+		)
+	case 8:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+			byte(b.bitContainer>>56),
+		)
+	default:
+		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
+	}
+	b.bitContainer >>= v << 3
+	b.nBits &= 7
+}
+
+// flush32 will flush out, so there are at least 32 bits available for writing.
+func (b *bitWriter) flush32() {
+	if b.nBits < 32 {
+		return
+	}
+	b.out = append(b.out,
+		byte(b.bitContainer),
+		byte(b.bitContainer>>8),
+		byte(b.bitContainer>>16),
+		byte(b.bitContainer>>24))
+	b.nBits -= 32
+	b.bitContainer >>= 32
+}
+
+// flushAlign will flush remaining full bytes and align to next byte boundary.
+func (b *bitWriter) flushAlign() {
+	nbBytes := (b.nBits + 7) >> 3
+	for i := uint8(0); i < nbBytes; i++ {
+		b.out = append(b.out, byte(b.bitContainer>>(i*8)))
+	}
+	b.nBits = 0
+	b.bitContainer = 0
+}
+
+// close will write the alignment bit and write the final byte(s)
+// to the output.
+func (b *bitWriter) close() error {
+	// End mark
+	b.addBits16Clean(1, 1)
+	// flush until next byte.
+	b.flushAlign()
+	return nil
+}
+
+// reset and continue writing by appending to out.
+func (b *bitWriter) reset(out []byte) {
+	b.bitContainer = 0
+	b.nBits = 0
+	b.out = out
+}

vendor/github.com/klauspost/compress/zstd/blockdec.go

@@ -0,0 +1,728 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/klauspost/compress/huff0"
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+type blockType uint8
+
+//go:generate stringer -type=blockType,literalsBlockType,seqCompMode,tableIndex
+
+const (
+	blockTypeRaw blockType = iota
+	blockTypeRLE
+	blockTypeCompressed
+	blockTypeReserved
+)
+
+type literalsBlockType uint8
+
+const (
+	literalsBlockRaw literalsBlockType = iota
+	literalsBlockRLE
+	literalsBlockCompressed
+	literalsBlockTreeless
+)
+
+const (
+	// maxCompressedBlockSize is the biggest allowed compressed block size (128KB)
+	maxCompressedBlockSize = 128 << 10
+
+	compressedBlockOverAlloc    = 16
+	maxCompressedBlockSizeAlloc = 128<<10 + compressedBlockOverAlloc
+
+	// Maximum possible block size (all Raw+Uncompressed).
+	maxBlockSize = (1 << 21) - 1
+
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#literals_section_header
+	maxCompressedLiteralSize = 1 << 18
+	maxRLELiteralSize        = 1 << 20
+	maxMatchLen              = 131074
+	maxSequences             = 0x7f00 + 0xffff
+
+	// We support slightly less than the reference decoder to be able to
+	// use ints on 32 bit archs.
+	maxOffsetBits = 30
+)
+
+var (
+	huffDecoderPool = sync.Pool{New: func() interface{} {
+		return &huff0.Scratch{}
+	}}
+
+	fseDecoderPool = sync.Pool{New: func() interface{} {
+		return &fseDecoder{}
+	}}
+)
+
+type blockDec struct {
+	// Raw source data of the block.
+	data        []byte
+	dataStorage []byte
+
+	// Destination of the decoded data.
+	dst []byte
+
+	// Buffer for literals data.
+	literalBuf []byte
+
+	// Window size of the block.
+	WindowSize uint64
+
+	err error
+
+	// Check against this crc
+	checkCRC []byte
+
+	// Frame to use for singlethreaded decoding.
+	// Should not be used by the decoder itself since parent may be another frame.
+	localFrame *frameDec
+
+	sequence []seqVals
+
+	async struct {
+		newHist  *history
+		literals []byte
+		seqData  []byte
+		seqSize  int // Size of uncompressed sequences
+		fcs      uint64
+	}
+
+	// Block is RLE, this is the size.
+	RLESize uint32
+	tmp     [4]byte
+
+	Type blockType
+
+	// Is this the last block of a frame?
+	Last bool
+
+	// Use less memory
+	lowMem bool
+}
+
+func (b *blockDec) String() string {
+	if b == nil {
+		return "<nil>"
+	}
+	return fmt.Sprintf("Steam Size: %d, Type: %v, Last: %t, Window: %d", len(b.data), b.Type, b.Last, b.WindowSize)
+}
+
+func newBlockDec(lowMem bool) *blockDec {
+	b := blockDec{
+		lowMem: lowMem,
+	}
+	return &b
+}
+
+// reset will reset the block.
+// Input must be a start of a block and will be at the end of the block when returned.
+func (b *blockDec) reset(br byteBuffer, windowSize uint64) error {
+	b.WindowSize = windowSize
+	tmp, err := br.readSmall(3)
+	if err != nil {
+		println("Reading block header:", err)
+		return err
+	}
+	bh := uint32(tmp[0]) | (uint32(tmp[1]) << 8) | (uint32(tmp[2]) << 16)
+	b.Last = bh&1 != 0
+	b.Type = blockType((bh >> 1) & 3)
+	// find size.
+	cSize := int(bh >> 3)
+	maxSize := maxCompressedBlockSizeAlloc
+	switch b.Type {
+	case blockTypeReserved:
+		return ErrReservedBlockType
+	case blockTypeRLE:
+		if cSize > maxCompressedBlockSize || cSize > int(b.WindowSize) {
+			if debugDecoder {
+				printf("rle block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrWindowSizeExceeded
+		}
+		b.RLESize = uint32(cSize)
+		if b.lowMem {
+			maxSize = cSize
+		}
+		cSize = 1
+	case blockTypeCompressed:
+		if debugDecoder {
+			println("Data size on stream:", cSize)
+		}
+		b.RLESize = 0
+		maxSize = maxCompressedBlockSizeAlloc
+		if windowSize < maxCompressedBlockSize && b.lowMem {
+			maxSize = int(windowSize) + compressedBlockOverAlloc
+		}
+		if cSize > maxCompressedBlockSize || uint64(cSize) > b.WindowSize {
+			if debugDecoder {
+				printf("compressed block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrCompressedSizeTooBig
+		}
+		// Empty compressed blocks must at least be 2 bytes
+		// for Literals_Block_Type and one for Sequences_Section_Header.
+		if cSize < 2 {
+			return ErrBlockTooSmall
+		}
+	case blockTypeRaw:
+		if cSize > maxCompressedBlockSize || cSize > int(b.WindowSize) {
+			if debugDecoder {
+				printf("rle block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrWindowSizeExceeded
+		}
+
+		b.RLESize = 0
+		// We do not need a destination for raw blocks.
+		maxSize = -1
+	default:
+		panic("Invalid block type")
+	}
+
+	// Read block data.
+	if cap(b.dataStorage) < cSize {
+		if b.lowMem || cSize > maxCompressedBlockSize {
+			b.dataStorage = make([]byte, 0, cSize+compressedBlockOverAlloc)
+		} else {
+			b.dataStorage = make([]byte, 0, maxCompressedBlockSizeAlloc)
+		}
+	}
+	if cap(b.dst) <= maxSize {
+		b.dst = make([]byte, 0, maxSize+1)
+	}
+	b.data, err = br.readBig(cSize, b.dataStorage)
+	if err != nil {
+		if debugDecoder {
+			println("Reading block:", err, "(", cSize, ")", len(b.data))
+			printf("%T", br)
+		}
+		return err
+	}
+	return nil
+}
+
+// sendEOF will make the decoder send EOF on this frame.
+func (b *blockDec) sendErr(err error) {
+	b.Last = true
+	b.Type = blockTypeReserved
+	b.err = err
+}
+
+// Close will release resources.
+// Closed blockDec cannot be reset.
+func (b *blockDec) Close() {
+}
+
+// decodeBuf
+func (b *blockDec) decodeBuf(hist *history) error {
+	switch b.Type {
+	case blockTypeRLE:
+		if cap(b.dst) < int(b.RLESize) {
+			if b.lowMem {
+				b.dst = make([]byte, b.RLESize)
+			} else {
+				b.dst = make([]byte, maxBlockSize)
+			}
+		}
+		b.dst = b.dst[:b.RLESize]
+		v := b.data[0]
+		for i := range b.dst {
+			b.dst[i] = v
+		}
+		hist.appendKeep(b.dst)
+		return nil
+	case blockTypeRaw:
+		hist.appendKeep(b.data)
+		return nil
+	case blockTypeCompressed:
+		saved := b.dst
+		// Append directly to history
+		if hist.ignoreBuffer == 0 {
+			b.dst = hist.b
+			hist.b = nil
+		} else {
+			b.dst = b.dst[:0]
+		}
+		err := b.decodeCompressed(hist)
+		if debugDecoder {
+			println("Decompressed to total", len(b.dst), "bytes, hash:", xxhash.Sum64(b.dst), "error:", err)
+		}
+		if hist.ignoreBuffer == 0 {
+			hist.b = b.dst
+			b.dst = saved
+		} else {
+			hist.appendKeep(b.dst)
+		}
+		return err
+	case blockTypeReserved:
+		// Used for returning errors.
+		return b.err
+	default:
+		panic("Invalid block type")
+	}
+}
+
+func (b *blockDec) decodeLiterals(in []byte, hist *history) (remain []byte, err error) {
+	// There must be at least one byte for Literals_Block_Type and one for Sequences_Section_Header
+	if len(in) < 2 {
+		return in, ErrBlockTooSmall
+	}
+
+	litType := literalsBlockType(in[0] & 3)
+	var litRegenSize int
+	var litCompSize int
+	sizeFormat := (in[0] >> 2) & 3
+	var fourStreams bool
+	var literals []byte
+	switch litType {
+	case literalsBlockRaw, literalsBlockRLE:
+		switch sizeFormat {
+		case 0, 2:
+			// Regenerated_Size uses 5 bits (0-31). Literals_Section_Header uses 1 byte.
+			litRegenSize = int(in[0] >> 3)
+			in = in[1:]
+		case 1:
+			// Regenerated_Size uses 12 bits (0-4095). Literals_Section_Header uses 2 bytes.
+			litRegenSize = int(in[0]>>4) + (int(in[1]) << 4)
+			in = in[2:]
+		case 3:
+			//  Regenerated_Size uses 20 bits (0-1048575). Literals_Section_Header uses 3 bytes.
+			if len(in) < 3 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			litRegenSize = int(in[0]>>4) + (int(in[1]) << 4) + (int(in[2]) << 12)
+			in = in[3:]
+		}
+	case literalsBlockCompressed, literalsBlockTreeless:
+		switch sizeFormat {
+		case 0, 1:
+			// Both Regenerated_Size and Compressed_Size use 10 bits (0-1023).
+			if len(in) < 3 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12)
+			litRegenSize = int(n & 1023)
+			litCompSize = int(n >> 10)
+			fourStreams = sizeFormat == 1
+			in = in[3:]
+		case 2:
+			fourStreams = true
+			if len(in) < 4 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12) + (uint64(in[3]) << 20)
+			litRegenSize = int(n & 16383)
+			litCompSize = int(n >> 14)
+			in = in[4:]
+		case 3:
+			fourStreams = true
+			if len(in) < 5 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12) + (uint64(in[3]) << 20) + (uint64(in[4]) << 28)
+			litRegenSize = int(n & 262143)
+			litCompSize = int(n >> 18)
+			in = in[5:]
+		}
+	}
+	if debugDecoder {
+		println("literals type:", litType, "litRegenSize:", litRegenSize, "litCompSize:", litCompSize, "sizeFormat:", sizeFormat, "4X:", fourStreams)
+	}
+	if litRegenSize > int(b.WindowSize) || litRegenSize > maxCompressedBlockSize {
+		return in, ErrWindowSizeExceeded
+	}
+
+	switch litType {
+	case literalsBlockRaw:
+		if len(in) < litRegenSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litRegenSize)
+			return in, ErrBlockTooSmall
+		}
+		literals = in[:litRegenSize]
+		in = in[litRegenSize:]
+		//printf("Found %d uncompressed literals\n", litRegenSize)
+	case literalsBlockRLE:
+		if len(in) < 1 {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", 1)
+			return in, ErrBlockTooSmall
+		}
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, litRegenSize)
+			} else {
+				if litRegenSize > maxCompressedLiteralSize {
+					// Exceptional
+					b.literalBuf = make([]byte, litRegenSize)
+				} else {
+					b.literalBuf = make([]byte, litRegenSize, maxCompressedLiteralSize)
+				}
+			}
+		}
+		literals = b.literalBuf[:litRegenSize]
+		v := in[0]
+		for i := range literals {
+			literals[i] = v
+		}
+		in = in[1:]
+		if debugDecoder {
+			printf("Found %d RLE compressed literals\n", litRegenSize)
+		}
+	case literalsBlockTreeless:
+		if len(in) < litCompSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litCompSize)
+			return in, ErrBlockTooSmall
+		}
+		// Store compressed literals, so we defer decoding until we get history.
+		literals = in[:litCompSize]
+		in = in[litCompSize:]
+		if debugDecoder {
+			printf("Found %d compressed literals\n", litCompSize)
+		}
+		huff := hist.huffTree
+		if huff == nil {
+			return in, errors.New("literal block was treeless, but no history was defined")
+		}
+		// Ensure we have space to store it.
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, 0, litRegenSize)
+			} else {
+				b.literalBuf = make([]byte, 0, maxCompressedLiteralSize)
+			}
+		}
+		var err error
+		// Use our out buffer.
+		huff.MaxDecodedSize = maxCompressedBlockSize
+		if fourStreams {
+			literals, err = huff.Decoder().Decompress4X(b.literalBuf[:0:litRegenSize], literals)
+		} else {
+			literals, err = huff.Decoder().Decompress1X(b.literalBuf[:0:litRegenSize], literals)
+		}
+		// Make sure we don't leak our literals buffer
+		if err != nil {
+			println("decompressing literals:", err)
+			return in, err
+		}
+		if len(literals) != litRegenSize {
+			return in, fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
+		}
+
+	case literalsBlockCompressed:
+		if len(in) < litCompSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litCompSize)
+			return in, ErrBlockTooSmall
+		}
+		literals = in[:litCompSize]
+		in = in[litCompSize:]
+		// Ensure we have space to store it.
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, 0, litRegenSize)
+			} else {
+				b.literalBuf = make([]byte, 0, maxCompressedBlockSize)
+			}
+		}
+		huff := hist.huffTree
+		if huff == nil || (hist.dict != nil && huff == hist.dict.litEnc) {
+			huff = huffDecoderPool.Get().(*huff0.Scratch)
+			if huff == nil {
+				huff = &huff0.Scratch{}
+			}
+		}
+		var err error
+		huff, literals, err = huff0.ReadTable(literals, huff)
+		if err != nil {
+			println("reading huffman table:", err)
+			return in, err
+		}
+		hist.huffTree = huff
+		huff.MaxDecodedSize = maxCompressedBlockSize
+		// Use our out buffer.
+		if fourStreams {
+			literals, err = huff.Decoder().Decompress4X(b.literalBuf[:0:litRegenSize], literals)
+		} else {
+			literals, err = huff.Decoder().Decompress1X(b.literalBuf[:0:litRegenSize], literals)
+		}
+		if err != nil {
+			println("decoding compressed literals:", err)
+			return in, err
+		}
+		// Make sure we don't leak our literals buffer
+		if len(literals) != litRegenSize {
+			return in, fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
+		}
+		if debugDecoder {
+			printf("Decompressed %d literals into %d bytes\n", litCompSize, litRegenSize)
+		}
+	}
+	hist.decoders.literals = literals
+	return in, nil
+}
+
+// decodeCompressed will start decompressing a block.
+func (b *blockDec) decodeCompressed(hist *history) error {
+	in := b.data
+	in, err := b.decodeLiterals(in, hist)
+	if err != nil {
+		return err
+	}
+	err = b.prepareSequences(in, hist)
+	if err != nil {
+		return err
+	}
+	if hist.decoders.nSeqs == 0 {
+		b.dst = append(b.dst, hist.decoders.literals...)
+		return nil
+	}
+	before := len(hist.decoders.out)
+	err = hist.decoders.decodeSync(hist.b[hist.ignoreBuffer:])
+	if err != nil {
+		return err
+	}
+	if hist.decoders.maxSyncLen > 0 {
+		hist.decoders.maxSyncLen += uint64(before)
+		hist.decoders.maxSyncLen -= uint64(len(hist.decoders.out))
+	}
+	b.dst = hist.decoders.out
+	hist.recentOffsets = hist.decoders.prevOffset
+	return nil
+}
+
+func (b *blockDec) prepareSequences(in []byte, hist *history) (err error) {
+	if debugDecoder {
+		printf("prepareSequences: %d byte(s) input\n", len(in))
+	}
+	// Decode Sequences
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#sequences-section
+	if len(in) < 1 {
+		return ErrBlockTooSmall
+	}
+	var nSeqs int
+	seqHeader := in[0]
+	switch {
+	case seqHeader < 128:
+		nSeqs = int(seqHeader)
+		in = in[1:]
+	case seqHeader < 255:
+		if len(in) < 2 {
+			return ErrBlockTooSmall
+		}
+		nSeqs = int(seqHeader-128)<<8 | int(in[1])
+		in = in[2:]
+	case seqHeader == 255:
+		if len(in) < 3 {
+			return ErrBlockTooSmall
+		}
+		nSeqs = 0x7f00 + int(in[1]) + (int(in[2]) << 8)
+		in = in[3:]
+	}
+	if nSeqs == 0 && len(in) != 0 {
+		// When no sequences, there should not be any more data...
+		if debugDecoder {
+			printf("prepareSequences: 0 sequences, but %d byte(s) left on stream\n", len(in))
+		}
+		return ErrUnexpectedBlockSize
+	}
+
+	var seqs = &hist.decoders
+	seqs.nSeqs = nSeqs
+	if nSeqs > 0 {
+		if len(in) < 1 {
+			return ErrBlockTooSmall
+		}
+		br := byteReader{b: in, off: 0}
+		compMode := br.Uint8()
+		br.advance(1)
+		if debugDecoder {
+			printf("Compression modes: 0b%b", compMode)
+		}
+		for i := uint(0); i < 3; i++ {
+			mode := seqCompMode((compMode >> (6 - i*2)) & 3)
+			if debugDecoder {
+				println("Table", tableIndex(i), "is", mode)
+			}
+			var seq *sequenceDec
+			switch tableIndex(i) {
+			case tableLiteralLengths:
+				seq = &seqs.litLengths
+			case tableOffsets:
+				seq = &seqs.offsets
+			case tableMatchLengths:
+				seq = &seqs.matchLengths
+			default:
+				panic("unknown table")
+			}
+			switch mode {
+			case compModePredefined:
+				if seq.fse != nil && !seq.fse.preDefined {
+					fseDecoderPool.Put(seq.fse)
+				}
+				seq.fse = &fsePredef[i]
+			case compModeRLE:
+				if br.remain() < 1 {
+					return ErrBlockTooSmall
+				}
+				v := br.Uint8()
+				br.advance(1)
+				if seq.fse == nil || seq.fse.preDefined {
+					seq.fse = fseDecoderPool.Get().(*fseDecoder)
+				}
+				symb, err := decSymbolValue(v, symbolTableX[i])
+				if err != nil {
+					printf("RLE Transform table (%v) error: %v", tableIndex(i), err)
+					return err
+				}
+				seq.fse.setRLE(symb)
+				if debugDecoder {
+					printf("RLE set to %+v, code: %v", symb, v)
+				}
+			case compModeFSE:
+				println("Reading table for", tableIndex(i))
+				if seq.fse == nil || seq.fse.preDefined {
+					seq.fse = fseDecoderPool.Get().(*fseDecoder)
+				}
+				err := seq.fse.readNCount(&br, uint16(maxTableSymbol[i]))
+				if err != nil {
+					println("Read table error:", err)
+					return err
+				}
+				err = seq.fse.transform(symbolTableX[i])
+				if err != nil {
+					println("Transform table error:", err)
+					return err
+				}
+				if debugDecoder {
+					println("Read table ok", "symbolLen:", seq.fse.symbolLen)
+				}
+			case compModeRepeat:
+				seq.repeat = true
+			}
+			if br.overread() {
+				return io.ErrUnexpectedEOF
+			}
+		}
+		in = br.unread()
+	}
+	if debugDecoder {
+		println("Literals:", len(seqs.literals), "hash:", xxhash.Sum64(seqs.literals), "and", seqs.nSeqs, "sequences.")
+	}
+
+	if nSeqs == 0 {
+		if len(b.sequence) > 0 {
+			b.sequence = b.sequence[:0]
+		}
+		return nil
+	}
+	br := seqs.br
+	if br == nil {
+		br = &bitReader{}
+	}
+	if err := br.init(in); err != nil {
+		return err
+	}
+
+	if err := seqs.initialize(br, hist, b.dst); err != nil {
+		println("initializing sequences:", err)
+		return err
+	}
+	// Extract blocks...
+	if false && hist.dict == nil {
+		fatalErr := func(err error) {
+			if err != nil {
+				panic(err)
+			}
+		}
+		fn := fmt.Sprintf("n-%d-lits-%d-prev-%d-%d-%d-win-%d.blk", hist.decoders.nSeqs, len(hist.decoders.literals), hist.recentOffsets[0], hist.recentOffsets[1], hist.recentOffsets[2], hist.windowSize)
+		var buf bytes.Buffer
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.litLengths.fse))
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.matchLengths.fse))
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.offsets.fse))
+		buf.Write(in)
+		ioutil.WriteFile(filepath.Join("testdata", "seqs", fn), buf.Bytes(), os.ModePerm)
+	}
+
+	return nil
+}
+
+func (b *blockDec) decodeSequences(hist *history) error {
+	if cap(b.sequence) < hist.decoders.nSeqs {
+		if b.lowMem {
+			b.sequence = make([]seqVals, 0, hist.decoders.nSeqs)
+		} else {
+			b.sequence = make([]seqVals, 0, 0x7F00+0xffff)
+		}
+	}
+	b.sequence = b.sequence[:hist.decoders.nSeqs]
+	if hist.decoders.nSeqs == 0 {
+		hist.decoders.seqSize = len(hist.decoders.literals)
+		return nil
+	}
+	hist.decoders.windowSize = hist.windowSize
+	hist.decoders.prevOffset = hist.recentOffsets
+
+	err := hist.decoders.decode(b.sequence)
+	hist.recentOffsets = hist.decoders.prevOffset
+	return err
+}
+
+func (b *blockDec) executeSequences(hist *history) error {
+	hbytes := hist.b
+	if len(hbytes) > hist.windowSize {
+		hbytes = hbytes[len(hbytes)-hist.windowSize:]
+		// We do not need history anymore.
+		if hist.dict != nil {
+			hist.dict.content = nil
+		}
+	}
+	hist.decoders.windowSize = hist.windowSize
+	hist.decoders.out = b.dst[:0]
+	err := hist.decoders.execute(b.sequence, hbytes)
+	if err != nil {
+		return err
+	}
+	return b.updateHistory(hist)
+}
+
+func (b *blockDec) updateHistory(hist *history) error {
+	if len(b.data) > maxCompressedBlockSize {
+		return fmt.Errorf("compressed block size too large (%d)", len(b.data))
+	}
+	// Set output and release references.
+	b.dst = hist.decoders.out
+	hist.recentOffsets = hist.decoders.prevOffset
+
+	if b.Last {
+		// if last block we don't care about history.
+		println("Last block, no history returned")
+		hist.b = hist.b[:0]
+		return nil
+	} else {
+		hist.append(b.dst)
+		if debugDecoder {
+			println("Finished block with ", len(b.sequence), "sequences. Added", len(b.dst), "to history, now length", len(hist.b))
+		}
+	}
+	hist.decoders.out, hist.decoders.literals = nil, nil
+
+	return nil
+}

vendor/github.com/klauspost/compress/zstd/blockenc.go

@@ -0,0 +1,871 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"math/bits"
+
+	"github.com/klauspost/compress/huff0"
+)
+
+type blockEnc struct {
+	size       int
+	literals   []byte
+	sequences  []seq
+	coders     seqCoders
+	litEnc     *huff0.Scratch
+	dictLitEnc *huff0.Scratch
+	wr         bitWriter
+
+	extraLits         int
+	output            []byte
+	recentOffsets     [3]uint32
+	prevRecentOffsets [3]uint32
+
+	last   bool
+	lowMem bool
+}
+
+// init should be used once the block has been created.
+// If called more than once, the effect is the same as calling reset.
+func (b *blockEnc) init() {
+	if b.lowMem {
+		// 1K literals
+		if cap(b.literals) < 1<<10 {
+			b.literals = make([]byte, 0, 1<<10)
+		}
+		const defSeqs = 20
+		if cap(b.sequences) < defSeqs {
+			b.sequences = make([]seq, 0, defSeqs)
+		}
+		// 1K
+		if cap(b.output) < 1<<10 {
+			b.output = make([]byte, 0, 1<<10)
+		}
+	} else {
+		if cap(b.literals) < maxCompressedBlockSize {
+			b.literals = make([]byte, 0, maxCompressedBlockSize)
+		}
+		const defSeqs = 2000
+		if cap(b.sequences) < defSeqs {
+			b.sequences = make([]seq, 0, defSeqs)
+		}
+		if cap(b.output) < maxCompressedBlockSize {
+			b.output = make([]byte, 0, maxCompressedBlockSize)
+		}
+	}
+
+	if b.coders.mlEnc == nil {
+		b.coders.mlEnc = &fseEncoder{}
+		b.coders.mlPrev = &fseEncoder{}
+		b.coders.ofEnc = &fseEncoder{}
+		b.coders.ofPrev = &fseEncoder{}
+		b.coders.llEnc = &fseEncoder{}
+		b.coders.llPrev = &fseEncoder{}
+	}
+	b.litEnc = &huff0.Scratch{WantLogLess: 4}
+	b.reset(nil)
+}
+
+// initNewEncode can be used to reset offsets and encoders to the initial state.
+func (b *blockEnc) initNewEncode() {
+	b.recentOffsets = [3]uint32{1, 4, 8}
+	b.litEnc.Reuse = huff0.ReusePolicyNone
+	b.coders.setPrev(nil, nil, nil)
+}
+
+// reset will reset the block for a new encode, but in the same stream,
+// meaning that state will be carried over, but the block content is reset.
+// If a previous block is provided, the recent offsets are carried over.
+func (b *blockEnc) reset(prev *blockEnc) {
+	b.extraLits = 0
+	b.literals = b.literals[:0]
+	b.size = 0
+	b.sequences = b.sequences[:0]
+	b.output = b.output[:0]
+	b.last = false
+	if prev != nil {
+		b.recentOffsets = prev.prevRecentOffsets
+	}
+	b.dictLitEnc = nil
+}
+
+// reset will reset the block for a new encode, but in the same stream,
+// meaning that state will be carried over, but the block content is reset.
+// If a previous block is provided, the recent offsets are carried over.
+func (b *blockEnc) swapEncoders(prev *blockEnc) {
+	b.coders.swap(&prev.coders)
+	b.litEnc, prev.litEnc = prev.litEnc, b.litEnc
+}
+
+// blockHeader contains the information for a block header.
+type blockHeader uint32
+
+// setLast sets the 'last' indicator on a block.
+func (h *blockHeader) setLast(b bool) {
+	if b {
+		*h = *h | 1
+	} else {
+		const mask = (1 << 24) - 2
+		*h = *h & mask
+	}
+}
+
+// setSize will store the compressed size of a block.
+func (h *blockHeader) setSize(v uint32) {
+	const mask = 7
+	*h = (*h)&mask | blockHeader(v<<3)
+}
+
+// setType sets the block type.
+func (h *blockHeader) setType(t blockType) {
+	const mask = 1 | (((1 << 24) - 1) ^ 7)
+	*h = (*h & mask) | blockHeader(t<<1)
+}
+
+// appendTo will append the block header to a slice.
+func (h blockHeader) appendTo(b []byte) []byte {
+	return append(b, uint8(h), uint8(h>>8), uint8(h>>16))
+}
+
+// String returns a string representation of the block.
+func (h blockHeader) String() string {
+	return fmt.Sprintf("Type: %d, Size: %d, Last:%t", (h>>1)&3, h>>3, h&1 == 1)
+}
+
+// literalsHeader contains literals header information.
+type literalsHeader uint64
+
+// setType can be used to set the type of literal block.
+func (h *literalsHeader) setType(t literalsBlockType) {
+	const mask = math.MaxUint64 - 3
+	*h = (*h & mask) | literalsHeader(t)
+}
+
+// setSize can be used to set a single size, for uncompressed and RLE content.
+func (h *literalsHeader) setSize(regenLen int) {
+	inBits := bits.Len32(uint32(regenLen))
+	// Only retain 2 bits
+	const mask = 3
+	lh := uint64(*h & mask)
+	switch {
+	case inBits < 5:
+		lh |= (uint64(regenLen) << 3) | (1 << 60)
+		if debugEncoder {
+			got := int(lh>>3) & 0xff
+			if got != regenLen {
+				panic(fmt.Sprint("litRegenSize = ", regenLen, "(want) != ", got, "(got)"))
+			}
+		}
+	case inBits < 12:
+		lh |= (1 << 2) | (uint64(regenLen) << 4) | (2 << 60)
+	case inBits < 20:
+		lh |= (3 << 2) | (uint64(regenLen) << 4) | (3 << 60)
+	default:
+		panic(fmt.Errorf("internal error: block too big (%d)", regenLen))
+	}
+	*h = literalsHeader(lh)
+}
+
+// setSizes will set the size of a compressed literals section and the input length.
+func (h *literalsHeader) setSizes(compLen, inLen int, single bool) {
+	compBits, inBits := bits.Len32(uint32(compLen)), bits.Len32(uint32(inLen))
+	// Only retain 2 bits
+	const mask = 3
+	lh := uint64(*h & mask)
+	switch {
+	case compBits <= 10 && inBits <= 10:
+		if !single {
+			lh |= 1 << 2
+		}
+		lh |= (uint64(inLen) << 4) | (uint64(compLen) << (10 + 4)) | (3 << 60)
+		if debugEncoder {
+			const mmask = (1 << 24) - 1
+			n := (lh >> 4) & mmask
+			if int(n&1023) != inLen {
+				panic(fmt.Sprint("regensize:", int(n&1023), "!=", inLen, inBits))
+			}
+			if int(n>>10) != compLen {
+				panic(fmt.Sprint("compsize:", int(n>>10), "!=", compLen, compBits))
+			}
+		}
+	case compBits <= 14 && inBits <= 14:
+		lh |= (2 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (14 + 4)) | (4 << 60)
+		if single {
+			panic("single stream used with more than 10 bits length.")
+		}
+	case compBits <= 18 && inBits <= 18:
+		lh |= (3 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (18 + 4)) | (5 << 60)
+		if single {
+			panic("single stream used with more than 10 bits length.")
+		}
+	default:
+		panic("internal error: block too big")
+	}
+	*h = literalsHeader(lh)
+}
+
+// appendTo will append the literals header to a byte slice.
+func (h literalsHeader) appendTo(b []byte) []byte {
+	size := uint8(h >> 60)
+	switch size {
+	case 1:
+		b = append(b, uint8(h))
+	case 2:
+		b = append(b, uint8(h), uint8(h>>8))
+	case 3:
+		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16))
+	case 4:
+		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24))
+	case 5:
+		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24), uint8(h>>32))
+	default:
+		panic(fmt.Errorf("internal error: literalsHeader has invalid size (%d)", size))
+	}
+	return b
+}
+
+// size returns the output size with currently set values.
+func (h literalsHeader) size() int {
+	return int(h >> 60)
+}
+
+func (h literalsHeader) String() string {
+	return fmt.Sprintf("Type: %d, SizeFormat: %d, Size: 0x%d, Bytes:%d", literalsBlockType(h&3), (h>>2)&3, h&((1<<60)-1)>>4, h>>60)
+}
+
+// pushOffsets will push the recent offsets to the backup store.
+func (b *blockEnc) pushOffsets() {
+	b.prevRecentOffsets = b.recentOffsets
+}
+
+// pushOffsets will push the recent offsets to the backup store.
+func (b *blockEnc) popOffsets() {
+	b.recentOffsets = b.prevRecentOffsets
+}
+
+// matchOffset will adjust recent offsets and return the adjusted one,
+// if it matches a previous offset.
+func (b *blockEnc) matchOffset(offset, lits uint32) uint32 {
+	// Check if offset is one of the recent offsets.
+	// Adjusts the output offset accordingly.
+	// Gives a tiny bit of compression, typically around 1%.
+	if true {
+		if lits > 0 {
+			switch offset {
+			case b.recentOffsets[0]:
+				offset = 1
+			case b.recentOffsets[1]:
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset = 2
+			case b.recentOffsets[2]:
+				b.recentOffsets[2] = b.recentOffsets[1]
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset = 3
+			default:
+				b.recentOffsets[2] = b.recentOffsets[1]
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset += 3
+			}
+		} else {
+			switch offset {
+			case b.recentOffsets[1]:
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset = 1
+			case b.recentOffsets[2]:
+				b.recentOffsets[2] = b.recentOffsets[1]
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset = 2
+			case b.recentOffsets[0] - 1:
+				b.recentOffsets[2] = b.recentOffsets[1]
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset = 3
+			default:
+				b.recentOffsets[2] = b.recentOffsets[1]
+				b.recentOffsets[1] = b.recentOffsets[0]
+				b.recentOffsets[0] = offset
+				offset += 3
+			}
+		}
+	} else {
+		offset += 3
+	}
+	return offset
+}
+
+// encodeRaw can be used to set the output to a raw representation of supplied bytes.
+func (b *blockEnc) encodeRaw(a []byte) {
+	var bh blockHeader
+	bh.setLast(b.last)
+	bh.setSize(uint32(len(a)))
+	bh.setType(blockTypeRaw)
+	b.output = bh.appendTo(b.output[:0])
+	b.output = append(b.output, a...)
+	if debugEncoder {
+		println("Adding RAW block, length", len(a), "last:", b.last)
+	}
+}
+
+// encodeRaw can be used to set the output to a raw representation of supplied bytes.
+func (b *blockEnc) encodeRawTo(dst, src []byte) []byte {
+	var bh blockHeader
+	bh.setLast(b.last)
+	bh.setSize(uint32(len(src)))
+	bh.setType(blockTypeRaw)
+	dst = bh.appendTo(dst)
+	dst = append(dst, src...)
+	if debugEncoder {
+		println("Adding RAW block, length", len(src), "last:", b.last)
+	}
+	return dst
+}
+
+// encodeLits can be used if the block is only litLen.
+func (b *blockEnc) encodeLits(lits []byte, raw bool) error {
+	var bh blockHeader
+	bh.setLast(b.last)
+	bh.setSize(uint32(len(lits)))
+
+	// Don't compress extremely small blocks
+	if len(lits) < 8 || (len(lits) < 32 && b.dictLitEnc == nil) || raw {
+		if debugEncoder {
+			println("Adding RAW block, length", len(lits), "last:", b.last)
+		}
+		bh.setType(blockTypeRaw)
+		b.output = bh.appendTo(b.output)
+		b.output = append(b.output, lits...)
+		return nil
+	}
+
+	var (
+		out            []byte
+		reUsed, single bool
+		err            error
+	)
+	if b.dictLitEnc != nil {
+		b.litEnc.TransferCTable(b.dictLitEnc)
+		b.litEnc.Reuse = huff0.ReusePolicyAllow
+		b.dictLitEnc = nil
+	}
+	if len(lits) >= 1024 {
+		// Use 4 Streams.
+		out, reUsed, err = huff0.Compress4X(lits, b.litEnc)
+	} else if len(lits) > 32 {
+		// Use 1 stream
+		single = true
+		out, reUsed, err = huff0.Compress1X(lits, b.litEnc)
+	} else {
+		err = huff0.ErrIncompressible
+	}
+
+	switch err {
+	case huff0.ErrIncompressible:
+		if debugEncoder {
+			println("Adding RAW block, length", len(lits), "last:", b.last)
+		}
+		bh.setType(blockTypeRaw)
+		b.output = bh.appendTo(b.output)
+		b.output = append(b.output, lits...)
+		return nil
+	case huff0.ErrUseRLE:
+		if debugEncoder {
+			println("Adding RLE block, length", len(lits))
+		}
+		bh.setType(blockTypeRLE)
+		b.output = bh.appendTo(b.output)
+		b.output = append(b.output, lits[0])
+		return nil
+	case nil:
+	default:
+		return err
+	}
+	// Compressed...
+	// Now, allow reuse
+	b.litEnc.Reuse = huff0.ReusePolicyAllow
+	bh.setType(blockTypeCompressed)
+	var lh literalsHeader
+	if reUsed {
+		if debugEncoder {
+			println("Reused tree, compressed to", len(out))
+		}
+		lh.setType(literalsBlockTreeless)
+	} else {
+		if debugEncoder {
+			println("New tree, compressed to", len(out), "tree size:", len(b.litEnc.OutTable))
+		}
+		lh.setType(literalsBlockCompressed)
+	}
+	// Set sizes
+	lh.setSizes(len(out), len(lits), single)
+	bh.setSize(uint32(len(out) + lh.size() + 1))
+
+	// Write block headers.
+	b.output = bh.appendTo(b.output)
+	b.output = lh.appendTo(b.output)
+	// Add compressed data.
+	b.output = append(b.output, out...)
+	// No sequences.
+	b.output = append(b.output, 0)
+	return nil
+}
+
+// fuzzFseEncoder can be used to fuzz the FSE encoder.
+func fuzzFseEncoder(data []byte) int {
+	if len(data) > maxSequences || len(data) < 2 {
+		return 0
+	}
+	enc := fseEncoder{}
+	hist := enc.Histogram()
+	maxSym := uint8(0)
+	for i, v := range data {
+		v = v & 63
+		data[i] = v
+		hist[v]++
+		if v > maxSym {
+			maxSym = v
+		}
+	}
+	if maxSym == 0 {
+		// All 0
+		return 0
+	}
+	maxCount := func(a []uint32) int {
+		var max uint32
+		for _, v := range a {
+			if v > max {
+				max = v
+			}
+		}
+		return int(max)
+	}
+	cnt := maxCount(hist[:maxSym])
+	if cnt == len(data) {
+		// RLE
+		return 0
+	}
+	enc.HistogramFinished(maxSym, cnt)
+	err := enc.normalizeCount(len(data))
+	if err != nil {
+		return 0
+	}
+	_, err = enc.writeCount(nil)
+	if err != nil {
+		panic(err)
+	}
+	return 1
+}
+
+// encode will encode the block and append the output in b.output.
+// Previous offset codes must be pushed if more blocks are expected.
+func (b *blockEnc) encode(org []byte, raw, rawAllLits bool) error {
+	if len(b.sequences) == 0 {
+		return b.encodeLits(b.literals, rawAllLits)
+	}
+	// We want some difference to at least account for the headers.
+	saved := b.size - len(b.literals) - (b.size >> 5)
+	if saved < 16 {
+		if org == nil {
+			return errIncompressible
+		}
+		b.popOffsets()
+		return b.encodeLits(org, rawAllLits)
+	}
+
+	var bh blockHeader
+	var lh literalsHeader
+	bh.setLast(b.last)
+	bh.setType(blockTypeCompressed)
+	// Store offset of the block header. Needed when we know the size.
+	bhOffset := len(b.output)
+	b.output = bh.appendTo(b.output)
+
+	var (
+		out            []byte
+		reUsed, single bool
+		err            error
+	)
+	if b.dictLitEnc != nil {
+		b.litEnc.TransferCTable(b.dictLitEnc)
+		b.litEnc.Reuse = huff0.ReusePolicyAllow
+		b.dictLitEnc = nil
+	}
+	if len(b.literals) >= 1024 && !raw {
+		// Use 4 Streams.
+		out, reUsed, err = huff0.Compress4X(b.literals, b.litEnc)
+	} else if len(b.literals) > 32 && !raw {
+		// Use 1 stream
+		single = true
+		out, reUsed, err = huff0.Compress1X(b.literals, b.litEnc)
+	} else {
+		err = huff0.ErrIncompressible
+	}
+
+	switch err {
+	case huff0.ErrIncompressible:
+		lh.setType(literalsBlockRaw)
+		lh.setSize(len(b.literals))
+		b.output = lh.appendTo(b.output)
+		b.output = append(b.output, b.literals...)
+		if debugEncoder {
+			println("Adding literals RAW, length", len(b.literals))
+		}
+	case huff0.ErrUseRLE:
+		lh.setType(literalsBlockRLE)
+		lh.setSize(len(b.literals))
+		b.output = lh.appendTo(b.output)
+		b.output = append(b.output, b.literals[0])
+		if debugEncoder {
+			println("Adding literals RLE")
+		}
+	case nil:
+		// Compressed litLen...
+		if reUsed {
+			if debugEncoder {
+				println("reused tree")
+			}
+			lh.setType(literalsBlockTreeless)
+		} else {
+			if debugEncoder {
+				println("new tree, size:", len(b.litEnc.OutTable))
+			}
+			lh.setType(literalsBlockCompressed)
+			if debugEncoder {
+				_, _, err := huff0.ReadTable(out, nil)
+				if err != nil {
+					panic(err)
+				}
+			}
+		}
+		lh.setSizes(len(out), len(b.literals), single)
+		if debugEncoder {
+			printf("Compressed %d literals to %d bytes", len(b.literals), len(out))
+			println("Adding literal header:", lh)
+		}
+		b.output = lh.appendTo(b.output)
+		b.output = append(b.output, out...)
+		b.litEnc.Reuse = huff0.ReusePolicyAllow
+		if debugEncoder {
+			println("Adding literals compressed")
+		}
+	default:
+		if debugEncoder {
+			println("Adding literals ERROR:", err)
+		}
+		return err
+	}
+	// Sequence compression
+
+	// Write the number of sequences
+	switch {
+	case len(b.sequences) < 128:
+		b.output = append(b.output, uint8(len(b.sequences)))
+	case len(b.sequences) < 0x7f00: // TODO: this could be wrong
+		n := len(b.sequences)
+		b.output = append(b.output, 128+uint8(n>>8), uint8(n))
+	default:
+		n := len(b.sequences) - 0x7f00
+		b.output = append(b.output, 255, uint8(n), uint8(n>>8))
+	}
+	if debugEncoder {
+		println("Encoding", len(b.sequences), "sequences")
+	}
+	b.genCodes()
+	llEnc := b.coders.llEnc
+	ofEnc := b.coders.ofEnc
+	mlEnc := b.coders.mlEnc
+	err = llEnc.normalizeCount(len(b.sequences))
+	if err != nil {
+		return err
+	}
+	err = ofEnc.normalizeCount(len(b.sequences))
+	if err != nil {
+		return err
+	}
+	err = mlEnc.normalizeCount(len(b.sequences))
+	if err != nil {
+		return err
+	}
+
+	// Choose the best compression mode for each type.
+	// Will evaluate the new vs predefined and previous.
+	chooseComp := func(cur, prev, preDef *fseEncoder) (*fseEncoder, seqCompMode) {
+		// See if predefined/previous is better
+		hist := cur.count[:cur.symbolLen]
+		nSize := cur.approxSize(hist) + cur.maxHeaderSize()
+		predefSize := preDef.approxSize(hist)
+		prevSize := prev.approxSize(hist)
+
+		// Add a small penalty for new encoders.
+		// Don't bother with extremely small (<2 byte gains).
+		nSize = nSize + (nSize+2*8*16)>>4
+		switch {
+		case predefSize <= prevSize && predefSize <= nSize || forcePreDef:
+			if debugEncoder {
+				println("Using predefined", predefSize>>3, "<=", nSize>>3)
+			}
+			return preDef, compModePredefined
+		case prevSize <= nSize:
+			if debugEncoder {
+				println("Using previous", prevSize>>3, "<=", nSize>>3)
+			}
+			return prev, compModeRepeat
+		default:
+			if debugEncoder {
+				println("Using new, predef", predefSize>>3, ". previous:", prevSize>>3, ">", nSize>>3, "header max:", cur.maxHeaderSize()>>3, "bytes")
+				println("tl:", cur.actualTableLog, "symbolLen:", cur.symbolLen, "norm:", cur.norm[:cur.symbolLen], "hist", cur.count[:cur.symbolLen])
+			}
+			return cur, compModeFSE
+		}
+	}
+
+	// Write compression mode
+	var mode uint8
+	if llEnc.useRLE {
+		mode |= uint8(compModeRLE) << 6
+		llEnc.setRLE(b.sequences[0].llCode)
+		if debugEncoder {
+			println("llEnc.useRLE")
+		}
+	} else {
+		var m seqCompMode
+		llEnc, m = chooseComp(llEnc, b.coders.llPrev, &fsePredefEnc[tableLiteralLengths])
+		mode |= uint8(m) << 6
+	}
+	if ofEnc.useRLE {
+		mode |= uint8(compModeRLE) << 4
+		ofEnc.setRLE(b.sequences[0].ofCode)
+		if debugEncoder {
+			println("ofEnc.useRLE")
+		}
+	} else {
+		var m seqCompMode
+		ofEnc, m = chooseComp(ofEnc, b.coders.ofPrev, &fsePredefEnc[tableOffsets])
+		mode |= uint8(m) << 4
+	}
+
+	if mlEnc.useRLE {
+		mode |= uint8(compModeRLE) << 2
+		mlEnc.setRLE(b.sequences[0].mlCode)
+		if debugEncoder {
+			println("mlEnc.useRLE, code: ", b.sequences[0].mlCode, "value", b.sequences[0].matchLen)
+		}
+	} else {
+		var m seqCompMode
+		mlEnc, m = chooseComp(mlEnc, b.coders.mlPrev, &fsePredefEnc[tableMatchLengths])
+		mode |= uint8(m) << 2
+	}
+	b.output = append(b.output, mode)
+	if debugEncoder {
+		printf("Compression modes: 0b%b", mode)
+	}
+	b.output, err = llEnc.writeCount(b.output)
+	if err != nil {
+		return err
+	}
+	start := len(b.output)
+	b.output, err = ofEnc.writeCount(b.output)
+	if err != nil {
+		return err
+	}
+	if false {
+		println("block:", b.output[start:], "tablelog", ofEnc.actualTableLog, "maxcount:", ofEnc.maxCount)
+		fmt.Printf("selected TableLog: %d, Symbol length: %d\n", ofEnc.actualTableLog, ofEnc.symbolLen)
+		for i, v := range ofEnc.norm[:ofEnc.symbolLen] {
+			fmt.Printf("%3d: %5d -> %4d \n", i, ofEnc.count[i], v)
+		}
+	}
+	b.output, err = mlEnc.writeCount(b.output)
+	if err != nil {
+		return err
+	}
+
+	// Maybe in block?
+	wr := &b.wr
+	wr.reset(b.output)
+
+	var ll, of, ml cState
+
+	// Current sequence
+	seq := len(b.sequences) - 1
+	s := b.sequences[seq]
+	llEnc.setBits(llBitsTable[:])
+	mlEnc.setBits(mlBitsTable[:])
+	ofEnc.setBits(nil)
+
+	llTT, ofTT, mlTT := llEnc.ct.symbolTT[:256], ofEnc.ct.symbolTT[:256], mlEnc.ct.symbolTT[:256]
+
+	// We have 3 bounds checks here (and in the loop).
+	// Since we are iterating backwards it is kinda hard to avoid.
+	llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
+	ll.init(wr, &llEnc.ct, llB)
+	of.init(wr, &ofEnc.ct, ofB)
+	wr.flush32()
+	ml.init(wr, &mlEnc.ct, mlB)
+
+	// Each of these lookups also generates a bounds check.
+	wr.addBits32NC(s.litLen, llB.outBits)
+	wr.addBits32NC(s.matchLen, mlB.outBits)
+	wr.flush32()
+	wr.addBits32NC(s.offset, ofB.outBits)
+	if debugSequences {
+		println("Encoded seq", seq, s, "codes:", s.llCode, s.mlCode, s.ofCode, "states:", ll.state, ml.state, of.state, "bits:", llB, mlB, ofB)
+	}
+	seq--
+	// Store sequences in reverse...
+	for seq >= 0 {
+		s = b.sequences[seq]
+
+		ofB := ofTT[s.ofCode]
+		wr.flush32() // tablelog max is below 8 for each, so it will fill max 24 bits.
+		//of.encode(ofB)
+		nbBitsOut := (uint32(of.state) + ofB.deltaNbBits) >> 16
+		dstState := int32(of.state>>(nbBitsOut&15)) + int32(ofB.deltaFindState)
+		wr.addBits16NC(of.state, uint8(nbBitsOut))
+		of.state = of.stateTable[dstState]
+
+		// Accumulate extra bits.
+		outBits := ofB.outBits & 31
+		extraBits := uint64(s.offset & bitMask32[outBits])
+		extraBitsN := outBits
+
+		mlB := mlTT[s.mlCode]
+		//ml.encode(mlB)
+		nbBitsOut = (uint32(ml.state) + mlB.deltaNbBits) >> 16
+		dstState = int32(ml.state>>(nbBitsOut&15)) + int32(mlB.deltaFindState)
+		wr.addBits16NC(ml.state, uint8(nbBitsOut))
+		ml.state = ml.stateTable[dstState]
+
+		outBits = mlB.outBits & 31
+		extraBits = extraBits<<outBits | uint64(s.matchLen&bitMask32[outBits])
+		extraBitsN += outBits
+
+		llB := llTT[s.llCode]
+		//ll.encode(llB)
+		nbBitsOut = (uint32(ll.state) + llB.deltaNbBits) >> 16
+		dstState = int32(ll.state>>(nbBitsOut&15)) + int32(llB.deltaFindState)
+		wr.addBits16NC(ll.state, uint8(nbBitsOut))
+		ll.state = ll.stateTable[dstState]
+
+		outBits = llB.outBits & 31
+		extraBits = extraBits<<outBits | uint64(s.litLen&bitMask32[outBits])
+		extraBitsN += outBits
+
+		wr.flush32()
+		wr.addBits64NC(extraBits, extraBitsN)
+
+		if debugSequences {
+			println("Encoded seq", seq, s)
+		}
+
+		seq--
+	}
+	ml.flush(mlEnc.actualTableLog)
+	of.flush(ofEnc.actualTableLog)
+	ll.flush(llEnc.actualTableLog)
+	err = wr.close()
+	if err != nil {
+		return err
+	}
+	b.output = wr.out
+
+	if len(b.output)-3-bhOffset >= b.size {
+		// Maybe even add a bigger margin.
+		b.litEnc.Reuse = huff0.ReusePolicyNone
+		return errIncompressible
+	}
+
+	// Size is output minus block header.
+	bh.setSize(uint32(len(b.output)-bhOffset) - 3)
+	if debugEncoder {
+		println("Rewriting block header", bh)
+	}
+	_ = bh.appendTo(b.output[bhOffset:bhOffset])
+	b.coders.setPrev(llEnc, mlEnc, ofEnc)
+	return nil
+}
+
+var errIncompressible = errors.New("incompressible")
+
+func (b *blockEnc) genCodes() {
+	if len(b.sequences) == 0 {
+		// nothing to do
+		return
+	}
+	if len(b.sequences) > math.MaxUint16 {
+		panic("can only encode up to 64K sequences")
+	}
+	// No bounds checks after here:
+	llH := b.coders.llEnc.Histogram()
+	ofH := b.coders.ofEnc.Histogram()
+	mlH := b.coders.mlEnc.Histogram()
+	for i := range llH {
+		llH[i] = 0
+	}
+	for i := range ofH {
+		ofH[i] = 0
+	}
+	for i := range mlH {
+		mlH[i] = 0
+	}
+
+	var llMax, ofMax, mlMax uint8
+	for i := range b.sequences {
+		seq := &b.sequences[i]
+		v := llCode(seq.litLen)
+		seq.llCode = v
+		llH[v]++
+		if v > llMax {
+			llMax = v
+		}
+
+		v = ofCode(seq.offset)
+		seq.ofCode = v
+		ofH[v]++
+		if v > ofMax {
+			ofMax = v
+		}
+
+		v = mlCode(seq.matchLen)
+		seq.mlCode = v
+		mlH[v]++
+		if v > mlMax {
+			mlMax = v
+			if debugAsserts && mlMax > maxMatchLengthSymbol {
+				panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d), matchlen: %d", mlMax, seq.matchLen))
+			}
+		}
+	}
+	maxCount := func(a []uint32) int {
+		var max uint32
+		for _, v := range a {
+			if v > max {
+				max = v
+			}
+		}
+		return int(max)
+	}
+	if debugAsserts && mlMax > maxMatchLengthSymbol {
+		panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d)", mlMax))
+	}
+	if debugAsserts && ofMax > maxOffsetBits {
+		panic(fmt.Errorf("ofMax > maxOffsetBits (%d)", ofMax))
+	}
+	if debugAsserts && llMax > maxLiteralLengthSymbol {
+		panic(fmt.Errorf("llMax > maxLiteralLengthSymbol (%d)", llMax))
+	}
+
+	b.coders.mlEnc.HistogramFinished(mlMax, maxCount(mlH[:mlMax+1]))
+	b.coders.ofEnc.HistogramFinished(ofMax, maxCount(ofH[:ofMax+1]))
+	b.coders.llEnc.HistogramFinished(llMax, maxCount(llH[:llMax+1]))
+}

vendor/github.com/klauspost/compress/zstd/blocktype_string.go

@@ -0,0 +1,85 @@
+// Code generated by "stringer -type=blockType,literalsBlockType,seqCompMode,tableIndex"; DO NOT EDIT.
+
+package zstd
+
+import "strconv"
+
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[blockTypeRaw-0]
+	_ = x[blockTypeRLE-1]
+	_ = x[blockTypeCompressed-2]
+	_ = x[blockTypeReserved-3]
+}
+
+const _blockType_name = "blockTypeRawblockTypeRLEblockTypeCompressedblockTypeReserved"
+
+var _blockType_index = [...]uint8{0, 12, 24, 43, 60}
+
+func (i blockType) String() string {
+	if i >= blockType(len(_blockType_index)-1) {
+		return "blockType(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _blockType_name[_blockType_index[i]:_blockType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[literalsBlockRaw-0]
+	_ = x[literalsBlockRLE-1]
+	_ = x[literalsBlockCompressed-2]
+	_ = x[literalsBlockTreeless-3]
+}
+
+const _literalsBlockType_name = "literalsBlockRawliteralsBlockRLEliteralsBlockCompressedliteralsBlockTreeless"
+
+var _literalsBlockType_index = [...]uint8{0, 16, 32, 55, 76}
+
+func (i literalsBlockType) String() string {
+	if i >= literalsBlockType(len(_literalsBlockType_index)-1) {
+		return "literalsBlockType(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _literalsBlockType_name[_literalsBlockType_index[i]:_literalsBlockType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[compModePredefined-0]
+	_ = x[compModeRLE-1]
+	_ = x[compModeFSE-2]
+	_ = x[compModeRepeat-3]
+}
+
+const _seqCompMode_name = "compModePredefinedcompModeRLEcompModeFSEcompModeRepeat"
+
+var _seqCompMode_index = [...]uint8{0, 18, 29, 40, 54}
+
+func (i seqCompMode) String() string {
+	if i >= seqCompMode(len(_seqCompMode_index)-1) {
+		return "seqCompMode(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _seqCompMode_name[_seqCompMode_index[i]:_seqCompMode_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[tableLiteralLengths-0]
+	_ = x[tableOffsets-1]
+	_ = x[tableMatchLengths-2]
+}
+
+const _tableIndex_name = "tableLiteralLengthstableOffsetstableMatchLengths"
+
+var _tableIndex_index = [...]uint8{0, 19, 31, 48}
+
+func (i tableIndex) String() string {
+	if i >= tableIndex(len(_tableIndex_index)-1) {
+		return "tableIndex(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _tableIndex_name[_tableIndex_index[i]:_tableIndex_index[i+1]]
+}

vendor/github.com/klauspost/compress/zstd/bytebuf.go

@@ -0,0 +1,133 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"fmt"
+	"io"
+	"io/ioutil"
+)
+
+type byteBuffer interface {
+	// Read up to 8 bytes.
+	// Returns io.ErrUnexpectedEOF if this cannot be satisfied.
+	readSmall(n int) ([]byte, error)
+
+	// Read >8 bytes.
+	// MAY use the destination slice.
+	readBig(n int, dst []byte) ([]byte, error)
+
+	// Read a single byte.
+	readByte() (byte, error)
+
+	// Skip n bytes.
+	skipN(n int) error
+}
+
+// in-memory buffer
+type byteBuf []byte
+
+func (b *byteBuf) readSmall(n int) ([]byte, error) {
+	if debugAsserts && n > 8 {
+		panic(fmt.Errorf("small read > 8 (%d). use readBig", n))
+	}
+	bb := *b
+	if len(bb) < n {
+		return nil, io.ErrUnexpectedEOF
+	}
+	r := bb[:n]
+	*b = bb[n:]
+	return r, nil
+}
+
+func (b *byteBuf) readBig(n int, dst []byte) ([]byte, error) {
+	bb := *b
+	if len(bb) < n {
+		return nil, io.ErrUnexpectedEOF
+	}
+	r := bb[:n]
+	*b = bb[n:]
+	return r, nil
+}
+
+func (b *byteBuf) remain() []byte {
+	return *b
+}
+
+func (b *byteBuf) readByte() (byte, error) {
+	bb := *b
+	if len(bb) < 1 {
+		return 0, nil
+	}
+	r := bb[0]
+	*b = bb[1:]
+	return r, nil
+}
+
+func (b *byteBuf) skipN(n int) error {
+	bb := *b
+	if len(bb) < n {
+		return io.ErrUnexpectedEOF
+	}
+	*b = bb[n:]
+	return nil
+}
+
+// wrapper around a reader.
+type readerWrapper struct {
+	r   io.Reader
+	tmp [8]byte
+}
+
+func (r *readerWrapper) readSmall(n int) ([]byte, error) {
+	if debugAsserts && n > 8 {
+		panic(fmt.Errorf("small read > 8 (%d). use readBig", n))
+	}
+	n2, err := io.ReadFull(r.r, r.tmp[:n])
+	// We only really care about the actual bytes read.
+	if err != nil {
+		if err == io.EOF {
+			return nil, io.ErrUnexpectedEOF
+		}
+		if debugDecoder {
+			println("readSmall: got", n2, "want", n, "err", err)
+		}
+		return nil, err
+	}
+	return r.tmp[:n], nil
+}
+
+func (r *readerWrapper) readBig(n int, dst []byte) ([]byte, error) {
+	if cap(dst) < n {
+		dst = make([]byte, n)
+	}
+	n2, err := io.ReadFull(r.r, dst[:n])
+	if err == io.EOF && n > 0 {
+		err = io.ErrUnexpectedEOF
+	}
+	return dst[:n2], err
+}
+
+func (r *readerWrapper) readByte() (byte, error) {
+	n2, err := r.r.Read(r.tmp[:1])
+	if err != nil {
+		if err == io.EOF {
+			err = io.ErrUnexpectedEOF
+		}
+		return 0, err
+	}
+	if n2 != 1 {
+		return 0, io.ErrUnexpectedEOF
+	}
+	return r.tmp[0], nil
+}
+
+func (r *readerWrapper) skipN(n int) error {
+	n2, err := io.CopyN(ioutil.Discard, r.r, int64(n))
+	if n2 != int64(n) {
+		err = io.ErrUnexpectedEOF
+	}
+	return err
+}

vendor/github.com/klauspost/compress/zstd/bytereader.go

@@ -0,0 +1,88 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+// byteReader provides a byte reader that reads
+// little endian values from a byte stream.
+// The input stream is manually advanced.
+// The reader performs no bounds checks.
+type byteReader struct {
+	b   []byte
+	off int
+}
+
+// init will initialize the reader and set the input.
+func (b *byteReader) init(in []byte) {
+	b.b = in
+	b.off = 0
+}
+
+// advance the stream b n bytes.
+func (b *byteReader) advance(n uint) {
+	b.off += int(n)
+}
+
+// overread returns whether we have advanced too far.
+func (b *byteReader) overread() bool {
+	return b.off > len(b.b)
+}
+
+// Int32 returns a little endian int32 starting at current offset.
+func (b byteReader) Int32() int32 {
+	b2 := b.b[b.off:]
+	b2 = b2[:4]
+	v3 := int32(b2[3])
+	v2 := int32(b2[2])
+	v1 := int32(b2[1])
+	v0 := int32(b2[0])
+	return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24)
+}
+
+// Uint8 returns the next byte
+func (b *byteReader) Uint8() uint8 {
+	v := b.b[b.off]
+	return v
+}
+
+// Uint32 returns a little endian uint32 starting at current offset.
+func (b byteReader) Uint32() uint32 {
+	if r := b.remain(); r < 4 {
+		// Very rare
+		v := uint32(0)
+		for i := 1; i <= r; i++ {
+			v = (v << 8) | uint32(b.b[len(b.b)-i])
+		}
+		return v
+	}
+	b2 := b.b[b.off:]
+	b2 = b2[:4]
+	v3 := uint32(b2[3])
+	v2 := uint32(b2[2])
+	v1 := uint32(b2[1])
+	v0 := uint32(b2[0])
+	return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24)
+}
+
+// Uint32NC returns a little endian uint32 starting at current offset.
+// The caller must be sure if there are at least 4 bytes left.
+func (b byteReader) Uint32NC() uint32 {
+	b2 := b.b[b.off:]
+	b2 = b2[:4]
+	v3 := uint32(b2[3])
+	v2 := uint32(b2[2])
+	v1 := uint32(b2[1])
+	v0 := uint32(b2[0])
+	return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24)
+}
+
+// unread returns the unread portion of the input.
+func (b byteReader) unread() []byte {
+	return b.b[b.off:]
+}
+
+// remain will return the number of bytes remaining.
+func (b byteReader) remain() int {
+	return len(b.b) - b.off
+}

vendor/github.com/klauspost/compress/zstd/decodeheader.go

@@ -0,0 +1,230 @@
+// Copyright 2020+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// HeaderMaxSize is the maximum size of a Frame and Block Header.
+// If less is sent to Header.Decode it *may* still contain enough information.
+const HeaderMaxSize = 14 + 3
+
+// Header contains information about the first frame and block within that.
+type Header struct {
+	// SingleSegment specifies whether the data is to be decompressed into a
+	// single contiguous memory segment.
+	// It implies that WindowSize is invalid and that FrameContentSize is valid.
+	SingleSegment bool
+
+	// WindowSize is the window of data to keep while decoding.
+	// Will only be set if SingleSegment is false.
+	WindowSize uint64
+
+	// Dictionary ID.
+	// If 0, no dictionary.
+	DictionaryID uint32
+
+	// HasFCS specifies whether FrameContentSize has a valid value.
+	HasFCS bool
+
+	// FrameContentSize is the expected uncompressed size of the entire frame.
+	FrameContentSize uint64
+
+	// Skippable will be true if the frame is meant to be skipped.
+	// This implies that FirstBlock.OK is false.
+	Skippable bool
+
+	// SkippableID is the user-specific ID for the skippable frame.
+	// Valid values are between 0 to 15, inclusive.
+	SkippableID int
+
+	// SkippableSize is the length of the user data to skip following
+	// the header.
+	SkippableSize uint32
+
+	// HeaderSize is the raw size of the frame header.
+	//
+	// For normal frames, it includes the size of the magic number and
+	// the size of the header (per section 3.1.1.1).
+	// It does not include the size for any data blocks (section 3.1.1.2) nor
+	// the size for the trailing content checksum.
+	//
+	// For skippable frames, this counts the size of the magic number
+	// along with the size of the size field of the payload.
+	// It does not include the size of the skippable payload itself.
+	// The total frame size is the HeaderSize plus the SkippableSize.
+	HeaderSize int
+
+	// First block information.
+	FirstBlock struct {
+		// OK will be set if first block could be decoded.
+		OK bool
+
+		// Is this the last block of a frame?
+		Last bool
+
+		// Is the data compressed?
+		// If true CompressedSize will be populated.
+		// Unfortunately DecompressedSize cannot be determined
+		// without decoding the blocks.
+		Compressed bool
+
+		// DecompressedSize is the expected decompressed size of the block.
+		// Will be 0 if it cannot be determined.
+		DecompressedSize int
+
+		// CompressedSize of the data in the block.
+		// Does not include the block header.
+		// Will be equal to DecompressedSize if not Compressed.
+		CompressedSize int
+	}
+
+	// If set there is a checksum present for the block content.
+	// The checksum field at the end is always 4 bytes long.
+	HasCheckSum bool
+}
+
+// Decode the header from the beginning of the stream.
+// This will decode the frame header and the first block header if enough bytes are provided.
+// It is recommended to provide at least HeaderMaxSize bytes.
+// If the frame header cannot be read an error will be returned.
+// If there isn't enough input, io.ErrUnexpectedEOF is returned.
+// The FirstBlock.OK will indicate if enough information was available to decode the first block header.
+func (h *Header) Decode(in []byte) error {
+	*h = Header{}
+	if len(in) < 4 {
+		return io.ErrUnexpectedEOF
+	}
+	h.HeaderSize += 4
+	b, in := in[:4], in[4:]
+	if !bytes.Equal(b, frameMagic) {
+		if !bytes.Equal(b[1:4], skippableFrameMagic) || b[0]&0xf0 != 0x50 {
+			return ErrMagicMismatch
+		}
+		if len(in) < 4 {
+			return io.ErrUnexpectedEOF
+		}
+		h.HeaderSize += 4
+		h.Skippable = true
+		h.SkippableID = int(b[0] & 0xf)
+		h.SkippableSize = binary.LittleEndian.Uint32(in)
+		return nil
+	}
+
+	// Read Window_Descriptor
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#window_descriptor
+	if len(in) < 1 {
+		return io.ErrUnexpectedEOF
+	}
+	fhd, in := in[0], in[1:]
+	h.HeaderSize++
+	h.SingleSegment = fhd&(1<<5) != 0
+	h.HasCheckSum = fhd&(1<<2) != 0
+	if fhd&(1<<3) != 0 {
+		return errors.New("reserved bit set on frame header")
+	}
+
+	if !h.SingleSegment {
+		if len(in) < 1 {
+			return io.ErrUnexpectedEOF
+		}
+		var wd byte
+		wd, in = in[0], in[1:]
+		h.HeaderSize++
+		windowLog := 10 + (wd >> 3)
+		windowBase := uint64(1) << windowLog
+		windowAdd := (windowBase / 8) * uint64(wd&0x7)
+		h.WindowSize = windowBase + windowAdd
+	}
+
+	// Read Dictionary_ID
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id
+	if size := fhd & 3; size != 0 {
+		if size == 3 {
+			size = 4
+		}
+		if len(in) < int(size) {
+			return io.ErrUnexpectedEOF
+		}
+		b, in = in[:size], in[size:]
+		h.HeaderSize += int(size)
+		switch size {
+		case 1:
+			h.DictionaryID = uint32(b[0])
+		case 2:
+			h.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8)
+		case 4:
+			h.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		}
+	}
+
+	// Read Frame_Content_Size
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frame_content_size
+	var fcsSize int
+	v := fhd >> 6
+	switch v {
+	case 0:
+		if h.SingleSegment {
+			fcsSize = 1
+		}
+	default:
+		fcsSize = 1 << v
+	}
+
+	if fcsSize > 0 {
+		h.HasFCS = true
+		if len(in) < fcsSize {
+			return io.ErrUnexpectedEOF
+		}
+		b, in = in[:fcsSize], in[fcsSize:]
+		h.HeaderSize += int(fcsSize)
+		switch fcsSize {
+		case 1:
+			h.FrameContentSize = uint64(b[0])
+		case 2:
+			// When FCS_Field_Size is 2, the offset of 256 is added.
+			h.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) + 256
+		case 4:
+			h.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) | (uint64(b[2]) << 16) | (uint64(b[3]) << 24)
+		case 8:
+			d1 := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+			d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24)
+			h.FrameContentSize = uint64(d1) | (uint64(d2) << 32)
+		}
+	}
+
+	// Frame Header done, we will not fail from now on.
+	if len(in) < 3 {
+		return nil
+	}
+	tmp := in[:3]
+	bh := uint32(tmp[0]) | (uint32(tmp[1]) << 8) | (uint32(tmp[2]) << 16)
+	h.FirstBlock.Last = bh&1 != 0
+	blockType := blockType((bh >> 1) & 3)
+	// find size.
+	cSize := int(bh >> 3)
+	switch blockType {
+	case blockTypeReserved:
+		return nil
+	case blockTypeRLE:
+		h.FirstBlock.Compressed = true
+		h.FirstBlock.DecompressedSize = cSize
+		h.FirstBlock.CompressedSize = 1
+	case blockTypeCompressed:
+		h.FirstBlock.Compressed = true
+		h.FirstBlock.CompressedSize = cSize
+	case blockTypeRaw:
+		h.FirstBlock.DecompressedSize = cSize
+		h.FirstBlock.CompressedSize = cSize
+	default:
+		panic("Invalid block type")
+	}
+
+	h.FirstBlock.OK = true
+	return nil
+}

vendor/github.com/klauspost/compress/zstd/decoder.go

@@ -0,0 +1,926 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"bytes"
+	"context"
+	"encoding/binary"
+	"io"
+	"sync"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+// Decoder provides decoding of zstandard streams.
+// The decoder has been designed to operate without allocations after a warmup.
+// This means that you should store the decoder for best performance.
+// To re-use a stream decoder, use the Reset(r io.Reader) error to switch to another stream.
+// A decoder can safely be re-used even if the previous stream failed.
+// To release the resources, you must call the Close() function on a decoder.
+type Decoder struct {
+	o decoderOptions
+
+	// Unreferenced decoders, ready for use.
+	decoders chan *blockDec
+
+	// Current read position used for Reader functionality.
+	current decoderState
+
+	// sync stream decoding
+	syncStream struct {
+		decodedFrame uint64
+		br           readerWrapper
+		enabled      bool
+		inFrame      bool
+	}
+
+	frame *frameDec
+
+	// Custom dictionaries.
+	// Always uses copies.
+	dicts map[uint32]dict
+
+	// streamWg is the waitgroup for all streams
+	streamWg sync.WaitGroup
+}
+
+// decoderState is used for maintaining state when the decoder
+// is used for streaming.
+type decoderState struct {
+	// current block being written to stream.
+	decodeOutput
+
+	// output in order to be written to stream.
+	output chan decodeOutput
+
+	// cancel remaining output.
+	cancel context.CancelFunc
+
+	// crc of current frame
+	crc *xxhash.Digest
+
+	flushed bool
+}
+
+var (
+	// Check the interfaces we want to support.
+	_ = io.WriterTo(&Decoder{})
+	_ = io.Reader(&Decoder{})
+)
+
+// NewReader creates a new decoder.
+// A nil Reader can be provided in which case Reset can be used to start a decode.
+//
+// A Decoder can be used in two modes:
+//
+// 1) As a stream, or
+// 2) For stateless decoding using DecodeAll.
+//
+// Only a single stream can be decoded concurrently, but the same decoder
+// can run multiple concurrent stateless decodes. It is even possible to
+// use stateless decodes while a stream is being decoded.
+//
+// The Reset function can be used to initiate a new stream, which is will considerably
+// reduce the allocations normally caused by NewReader.
+func NewReader(r io.Reader, opts ...DOption) (*Decoder, error) {
+	initPredefined()
+	var d Decoder
+	d.o.setDefault()
+	for _, o := range opts {
+		err := o(&d.o)
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.current.crc = xxhash.New()
+	d.current.flushed = true
+
+	if r == nil {
+		d.current.err = ErrDecoderNilInput
+	}
+
+	// Transfer option dicts.
+	d.dicts = make(map[uint32]dict, len(d.o.dicts))
+	for _, dc := range d.o.dicts {
+		d.dicts[dc.id] = dc
+	}
+	d.o.dicts = nil
+
+	// Create decoders
+	d.decoders = make(chan *blockDec, d.o.concurrent)
+	for i := 0; i < d.o.concurrent; i++ {
+		dec := newBlockDec(d.o.lowMem)
+		dec.localFrame = newFrameDec(d.o)
+		d.decoders <- dec
+	}
+
+	if r == nil {
+		return &d, nil
+	}
+	return &d, d.Reset(r)
+}
+
+// Read bytes from the decompressed stream into p.
+// Returns the number of bytes written and any error that occurred.
+// When the stream is done, io.EOF will be returned.
+func (d *Decoder) Read(p []byte) (int, error) {
+	var n int
+	for {
+		if len(d.current.b) > 0 {
+			filled := copy(p, d.current.b)
+			p = p[filled:]
+			d.current.b = d.current.b[filled:]
+			n += filled
+		}
+		if len(p) == 0 {
+			break
+		}
+		if len(d.current.b) == 0 {
+			// We have an error and no more data
+			if d.current.err != nil {
+				break
+			}
+			if !d.nextBlock(n == 0) {
+				return n, d.current.err
+			}
+		}
+	}
+	if len(d.current.b) > 0 {
+		if debugDecoder {
+			println("returning", n, "still bytes left:", len(d.current.b))
+		}
+		// Only return error at end of block
+		return n, nil
+	}
+	if d.current.err != nil {
+		d.drainOutput()
+	}
+	if debugDecoder {
+		println("returning", n, d.current.err, len(d.decoders))
+	}
+	return n, d.current.err
+}
+
+// Reset will reset the decoder the supplied stream after the current has finished processing.
+// Note that this functionality cannot be used after Close has been called.
+// Reset can be called with a nil reader to release references to the previous reader.
+// After being called with a nil reader, no other operations than Reset or DecodeAll or Close
+// should be used.
+func (d *Decoder) Reset(r io.Reader) error {
+	if d.current.err == ErrDecoderClosed {
+		return d.current.err
+	}
+
+	d.drainOutput()
+
+	d.syncStream.br.r = nil
+	if r == nil {
+		d.current.err = ErrDecoderNilInput
+		if len(d.current.b) > 0 {
+			d.current.b = d.current.b[:0]
+		}
+		d.current.flushed = true
+		return nil
+	}
+
+	// If bytes buffer and < 5MB, do sync decoding anyway.
+	if bb, ok := r.(byter); ok && bb.Len() < 5<<20 {
+		bb2 := bb
+		if debugDecoder {
+			println("*bytes.Buffer detected, doing sync decode, len:", bb.Len())
+		}
+		b := bb2.Bytes()
+		var dst []byte
+		if cap(d.current.b) > 0 {
+			dst = d.current.b
+		}
+
+		dst, err := d.DecodeAll(b, dst[:0])
+		if err == nil {
+			err = io.EOF
+		}
+		d.current.b = dst
+		d.current.err = err
+		d.current.flushed = true
+		if debugDecoder {
+			println("sync decode to", len(dst), "bytes, err:", err)
+		}
+		return nil
+	}
+	// Remove current block.
+	d.stashDecoder()
+	d.current.decodeOutput = decodeOutput{}
+	d.current.err = nil
+	d.current.flushed = false
+	d.current.d = nil
+
+	// Ensure no-one else is still running...
+	d.streamWg.Wait()
+	if d.frame == nil {
+		d.frame = newFrameDec(d.o)
+	}
+
+	if d.o.concurrent == 1 {
+		return d.startSyncDecoder(r)
+	}
+
+	d.current.output = make(chan decodeOutput, d.o.concurrent)
+	ctx, cancel := context.WithCancel(context.Background())
+	d.current.cancel = cancel
+	d.streamWg.Add(1)
+	go d.startStreamDecoder(ctx, r, d.current.output)
+
+	return nil
+}
+
+// drainOutput will drain the output until errEndOfStream is sent.
+func (d *Decoder) drainOutput() {
+	if d.current.cancel != nil {
+		if debugDecoder {
+			println("cancelling current")
+		}
+		d.current.cancel()
+		d.current.cancel = nil
+	}
+	if d.current.d != nil {
+		if debugDecoder {
+			printf("re-adding current decoder %p, decoders: %d", d.current.d, len(d.decoders))
+		}
+		d.decoders <- d.current.d
+		d.current.d = nil
+		d.current.b = nil
+	}
+	if d.current.output == nil || d.current.flushed {
+		println("current already flushed")
+		return
+	}
+	for v := range d.current.output {
+		if v.d != nil {
+			if debugDecoder {
+				printf("re-adding decoder %p", v.d)
+			}
+			d.decoders <- v.d
+		}
+	}
+	d.current.output = nil
+	d.current.flushed = true
+}
+
+// WriteTo writes data to w until there's no more data to write or when an error occurs.
+// The return value n is the number of bytes written.
+// Any error encountered during the write is also returned.
+func (d *Decoder) WriteTo(w io.Writer) (int64, error) {
+	var n int64
+	for {
+		if len(d.current.b) > 0 {
+			n2, err2 := w.Write(d.current.b)
+			n += int64(n2)
+			if err2 != nil && (d.current.err == nil || d.current.err == io.EOF) {
+				d.current.err = err2
+			} else if n2 != len(d.current.b) {
+				d.current.err = io.ErrShortWrite
+			}
+		}
+		if d.current.err != nil {
+			break
+		}
+		d.nextBlock(true)
+	}
+	err := d.current.err
+	if err != nil {
+		d.drainOutput()
+	}
+	if err == io.EOF {
+		err = nil
+	}
+	return n, err
+}
+
+// DecodeAll allows stateless decoding of a blob of bytes.
+// Output will be appended to dst, so if the destination size is known
+// you can pre-allocate the destination slice to avoid allocations.
+// DecodeAll can be used concurrently.
+// The Decoder concurrency limits will be respected.
+func (d *Decoder) DecodeAll(input, dst []byte) ([]byte, error) {
+	if d.decoders == nil {
+		return dst, ErrDecoderClosed
+	}
+
+	// Grab a block decoder and frame decoder.
+	block := <-d.decoders
+	frame := block.localFrame
+	defer func() {
+		if debugDecoder {
+			printf("re-adding decoder: %p", block)
+		}
+		frame.rawInput = nil
+		frame.bBuf = nil
+		if frame.history.decoders.br != nil {
+			frame.history.decoders.br.in = nil
+		}
+		d.decoders <- block
+	}()
+	frame.bBuf = input
+
+	for {
+		frame.history.reset()
+		err := frame.reset(&frame.bBuf)
+		if err != nil {
+			if err == io.EOF {
+				if debugDecoder {
+					println("frame reset return EOF")
+				}
+				return dst, nil
+			}
+			return dst, err
+		}
+		if frame.DictionaryID != nil {
+			dict, ok := d.dicts[*frame.DictionaryID]
+			if !ok {
+				return nil, ErrUnknownDictionary
+			}
+			if debugDecoder {
+				println("setting dict", frame.DictionaryID)
+			}
+			frame.history.setDict(&dict)
+		}
+		if frame.WindowSize > d.o.maxWindowSize {
+			return dst, ErrWindowSizeExceeded
+		}
+		if frame.FrameContentSize != fcsUnknown {
+			if frame.FrameContentSize > d.o.maxDecodedSize-uint64(len(dst)) {
+				return dst, ErrDecoderSizeExceeded
+			}
+			if cap(dst)-len(dst) < int(frame.FrameContentSize) {
+				dst2 := make([]byte, len(dst), len(dst)+int(frame.FrameContentSize)+compressedBlockOverAlloc)
+				copy(dst2, dst)
+				dst = dst2
+			}
+		}
+
+		if cap(dst) == 0 {
+			// Allocate len(input) * 2 by default if nothing is provided
+			// and we didn't get frame content size.
+			size := len(input) * 2
+			// Cap to 1 MB.
+			if size > 1<<20 {
+				size = 1 << 20
+			}
+			if uint64(size) > d.o.maxDecodedSize {
+				size = int(d.o.maxDecodedSize)
+			}
+			dst = make([]byte, 0, size)
+		}
+
+		dst, err = frame.runDecoder(dst, block)
+		if err != nil {
+			return dst, err
+		}
+		if len(frame.bBuf) == 0 {
+			if debugDecoder {
+				println("frame dbuf empty")
+			}
+			break
+		}
+	}
+	return dst, nil
+}
+
+// nextBlock returns the next block.
+// If an error occurs d.err will be set.
+// Optionally the function can block for new output.
+// If non-blocking mode is used the returned boolean will be false
+// if no data was available without blocking.
+func (d *Decoder) nextBlock(blocking bool) (ok bool) {
+	if d.current.err != nil {
+		// Keep error state.
+		return false
+	}
+	d.current.b = d.current.b[:0]
+
+	// SYNC:
+	if d.syncStream.enabled {
+		if !blocking {
+			return false
+		}
+		ok = d.nextBlockSync()
+		if !ok {
+			d.stashDecoder()
+		}
+		return ok
+	}
+
+	//ASYNC:
+	d.stashDecoder()
+	if blocking {
+		d.current.decodeOutput, ok = <-d.current.output
+	} else {
+		select {
+		case d.current.decodeOutput, ok = <-d.current.output:
+		default:
+			return false
+		}
+	}
+	if !ok {
+		// This should not happen, so signal error state...
+		d.current.err = io.ErrUnexpectedEOF
+		return false
+	}
+	next := d.current.decodeOutput
+	if next.d != nil && next.d.async.newHist != nil {
+		d.current.crc.Reset()
+	}
+	if debugDecoder {
+		var tmp [4]byte
+		binary.LittleEndian.PutUint32(tmp[:], uint32(xxhash.Sum64(next.b)))
+		println("got", len(d.current.b), "bytes, error:", d.current.err, "data crc:", tmp)
+	}
+
+	if len(next.b) > 0 {
+		n, err := d.current.crc.Write(next.b)
+		if err == nil {
+			if n != len(next.b) {
+				d.current.err = io.ErrShortWrite
+			}
+		}
+	}
+	if next.err == nil && next.d != nil && len(next.d.checkCRC) != 0 {
+		got := d.current.crc.Sum64()
+		var tmp [4]byte
+		binary.LittleEndian.PutUint32(tmp[:], uint32(got))
+		if !bytes.Equal(tmp[:], next.d.checkCRC) && !ignoreCRC {
+			if debugDecoder {
+				println("CRC Check Failed:", tmp[:], " (got) !=", next.d.checkCRC, "(on stream)")
+			}
+			d.current.err = ErrCRCMismatch
+		} else {
+			if debugDecoder {
+				println("CRC ok", tmp[:])
+			}
+		}
+	}
+
+	return true
+}
+
+func (d *Decoder) nextBlockSync() (ok bool) {
+	if d.current.d == nil {
+		d.current.d = <-d.decoders
+	}
+	for len(d.current.b) == 0 {
+		if !d.syncStream.inFrame {
+			d.frame.history.reset()
+			d.current.err = d.frame.reset(&d.syncStream.br)
+			if d.current.err != nil {
+				return false
+			}
+			if d.frame.DictionaryID != nil {
+				dict, ok := d.dicts[*d.frame.DictionaryID]
+				if !ok {
+					d.current.err = ErrUnknownDictionary
+					return false
+				} else {
+					d.frame.history.setDict(&dict)
+				}
+			}
+			if d.frame.WindowSize > d.o.maxDecodedSize || d.frame.WindowSize > d.o.maxWindowSize {
+				d.current.err = ErrDecoderSizeExceeded
+				return false
+			}
+
+			d.syncStream.decodedFrame = 0
+			d.syncStream.inFrame = true
+		}
+		d.current.err = d.frame.next(d.current.d)
+		if d.current.err != nil {
+			return false
+		}
+		d.frame.history.ensureBlock()
+		if debugDecoder {
+			println("History trimmed:", len(d.frame.history.b), "decoded already:", d.syncStream.decodedFrame)
+		}
+		histBefore := len(d.frame.history.b)
+		d.current.err = d.current.d.decodeBuf(&d.frame.history)
+
+		if d.current.err != nil {
+			println("error after:", d.current.err)
+			return false
+		}
+		d.current.b = d.frame.history.b[histBefore:]
+		if debugDecoder {
+			println("history after:", len(d.frame.history.b))
+		}
+
+		// Check frame size (before CRC)
+		d.syncStream.decodedFrame += uint64(len(d.current.b))
+		if d.syncStream.decodedFrame > d.frame.FrameContentSize {
+			if debugDecoder {
+				printf("DecodedFrame (%d) > FrameContentSize (%d)\n", d.syncStream.decodedFrame, d.frame.FrameContentSize)
+			}
+			d.current.err = ErrFrameSizeExceeded
+			return false
+		}
+
+		// Check FCS
+		if d.current.d.Last && d.frame.FrameContentSize != fcsUnknown && d.syncStream.decodedFrame != d.frame.FrameContentSize {
+			if debugDecoder {
+				printf("DecodedFrame (%d) != FrameContentSize (%d)\n", d.syncStream.decodedFrame, d.frame.FrameContentSize)
+			}
+			d.current.err = ErrFrameSizeMismatch
+			return false
+		}
+
+		// Update/Check CRC
+		if d.frame.HasCheckSum {
+			d.frame.crc.Write(d.current.b)
+			if d.current.d.Last {
+				d.current.err = d.frame.checkCRC()
+				if d.current.err != nil {
+					println("CRC error:", d.current.err)
+					return false
+				}
+			}
+		}
+		d.syncStream.inFrame = !d.current.d.Last
+	}
+	return true
+}
+
+func (d *Decoder) stashDecoder() {
+	if d.current.d != nil {
+		if debugDecoder {
+			printf("re-adding current decoder %p", d.current.d)
+		}
+		d.decoders <- d.current.d
+		d.current.d = nil
+	}
+}
+
+// Close will release all resources.
+// It is NOT possible to reuse the decoder after this.
+func (d *Decoder) Close() {
+	if d.current.err == ErrDecoderClosed {
+		return
+	}
+	d.drainOutput()
+	if d.current.cancel != nil {
+		d.current.cancel()
+		d.streamWg.Wait()
+		d.current.cancel = nil
+	}
+	if d.decoders != nil {
+		close(d.decoders)
+		for dec := range d.decoders {
+			dec.Close()
+		}
+		d.decoders = nil
+	}
+	if d.current.d != nil {
+		d.current.d.Close()
+		d.current.d = nil
+	}
+	d.current.err = ErrDecoderClosed
+}
+
+// IOReadCloser returns the decoder as an io.ReadCloser for convenience.
+// Any changes to the decoder will be reflected, so the returned ReadCloser
+// can be reused along with the decoder.
+// io.WriterTo is also supported by the returned ReadCloser.
+func (d *Decoder) IOReadCloser() io.ReadCloser {
+	return closeWrapper{d: d}
+}
+
+// closeWrapper wraps a function call as a closer.
+type closeWrapper struct {
+	d *Decoder
+}
+
+// WriteTo forwards WriteTo calls to the decoder.
+func (c closeWrapper) WriteTo(w io.Writer) (n int64, err error) {
+	return c.d.WriteTo(w)
+}
+
+// Read forwards read calls to the decoder.
+func (c closeWrapper) Read(p []byte) (n int, err error) {
+	return c.d.Read(p)
+}
+
+// Close closes the decoder.
+func (c closeWrapper) Close() error {
+	c.d.Close()
+	return nil
+}
+
+type decodeOutput struct {
+	d   *blockDec
+	b   []byte
+	err error
+}
+
+func (d *Decoder) startSyncDecoder(r io.Reader) error {
+	d.frame.history.reset()
+	d.syncStream.br = readerWrapper{r: r}
+	d.syncStream.inFrame = false
+	d.syncStream.enabled = true
+	d.syncStream.decodedFrame = 0
+	return nil
+}
+
+// Create Decoder:
+// ASYNC:
+// Spawn 4 go routines.
+// 0: Read frames and decode blocks.
+// 1: Decode block and literals. Receives hufftree and seqdecs, returns seqdecs and huff tree.
+// 2: Wait for recentOffsets if needed. Decode sequences, send recentOffsets.
+// 3: Wait for stream history, execute sequences, send stream history.
+func (d *Decoder) startStreamDecoder(ctx context.Context, r io.Reader, output chan decodeOutput) {
+	defer d.streamWg.Done()
+	br := readerWrapper{r: r}
+
+	var seqPrepare = make(chan *blockDec, d.o.concurrent)
+	var seqDecode = make(chan *blockDec, d.o.concurrent)
+	var seqExecute = make(chan *blockDec, d.o.concurrent)
+
+	// Async 1: Prepare blocks...
+	go func() {
+		var hist history
+		var hasErr bool
+		for block := range seqPrepare {
+			if hasErr {
+				if block != nil {
+					seqDecode <- block
+				}
+				continue
+			}
+			if block.async.newHist != nil {
+				if debugDecoder {
+					println("Async 1: new history")
+				}
+				hist.reset()
+				if block.async.newHist.dict != nil {
+					hist.setDict(block.async.newHist.dict)
+				}
+			}
+			if block.err != nil || block.Type != blockTypeCompressed {
+				hasErr = block.err != nil
+				seqDecode <- block
+				continue
+			}
+
+			remain, err := block.decodeLiterals(block.data, &hist)
+			block.err = err
+			hasErr = block.err != nil
+			if err == nil {
+				block.async.literals = hist.decoders.literals
+				block.async.seqData = remain
+			} else if debugDecoder {
+				println("decodeLiterals error:", err)
+			}
+			seqDecode <- block
+		}
+		close(seqDecode)
+	}()
+
+	// Async 2: Decode sequences...
+	go func() {
+		var hist history
+		var hasErr bool
+
+		for block := range seqDecode {
+			if hasErr {
+				if block != nil {
+					seqExecute <- block
+				}
+				continue
+			}
+			if block.async.newHist != nil {
+				if debugDecoder {
+					println("Async 2: new history, recent:", block.async.newHist.recentOffsets)
+				}
+				hist.decoders = block.async.newHist.decoders
+				hist.recentOffsets = block.async.newHist.recentOffsets
+				hist.windowSize = block.async.newHist.windowSize
+				if block.async.newHist.dict != nil {
+					hist.setDict(block.async.newHist.dict)
+				}
+			}
+			if block.err != nil || block.Type != blockTypeCompressed {
+				hasErr = block.err != nil
+				seqExecute <- block
+				continue
+			}
+
+			hist.decoders.literals = block.async.literals
+			block.err = block.prepareSequences(block.async.seqData, &hist)
+			if debugDecoder && block.err != nil {
+				println("prepareSequences returned:", block.err)
+			}
+			hasErr = block.err != nil
+			if block.err == nil {
+				block.err = block.decodeSequences(&hist)
+				if debugDecoder && block.err != nil {
+					println("decodeSequences returned:", block.err)
+				}
+				hasErr = block.err != nil
+				//				block.async.sequence = hist.decoders.seq[:hist.decoders.nSeqs]
+				block.async.seqSize = hist.decoders.seqSize
+			}
+			seqExecute <- block
+		}
+		close(seqExecute)
+	}()
+
+	var wg sync.WaitGroup
+	wg.Add(1)
+
+	// Async 3: Execute sequences...
+	frameHistCache := d.frame.history.b
+	go func() {
+		var hist history
+		var decodedFrame uint64
+		var fcs uint64
+		var hasErr bool
+		for block := range seqExecute {
+			out := decodeOutput{err: block.err, d: block}
+			if block.err != nil || hasErr {
+				hasErr = true
+				output <- out
+				continue
+			}
+			if block.async.newHist != nil {
+				if debugDecoder {
+					println("Async 3: new history")
+				}
+				hist.windowSize = block.async.newHist.windowSize
+				hist.allocFrameBuffer = block.async.newHist.allocFrameBuffer
+				if block.async.newHist.dict != nil {
+					hist.setDict(block.async.newHist.dict)
+				}
+
+				if cap(hist.b) < hist.allocFrameBuffer {
+					if cap(frameHistCache) >= hist.allocFrameBuffer {
+						hist.b = frameHistCache
+					} else {
+						hist.b = make([]byte, 0, hist.allocFrameBuffer)
+						println("Alloc history sized", hist.allocFrameBuffer)
+					}
+				}
+				hist.b = hist.b[:0]
+				fcs = block.async.fcs
+				decodedFrame = 0
+			}
+			do := decodeOutput{err: block.err, d: block}
+			switch block.Type {
+			case blockTypeRLE:
+				if debugDecoder {
+					println("add rle block length:", block.RLESize)
+				}
+
+				if cap(block.dst) < int(block.RLESize) {
+					if block.lowMem {
+						block.dst = make([]byte, block.RLESize)
+					} else {
+						block.dst = make([]byte, maxBlockSize)
+					}
+				}
+				block.dst = block.dst[:block.RLESize]
+				v := block.data[0]
+				for i := range block.dst {
+					block.dst[i] = v
+				}
+				hist.append(block.dst)
+				do.b = block.dst
+			case blockTypeRaw:
+				if debugDecoder {
+					println("add raw block length:", len(block.data))
+				}
+				hist.append(block.data)
+				do.b = block.data
+			case blockTypeCompressed:
+				if debugDecoder {
+					println("execute with history length:", len(hist.b), "window:", hist.windowSize)
+				}
+				hist.decoders.seqSize = block.async.seqSize
+				hist.decoders.literals = block.async.literals
+				do.err = block.executeSequences(&hist)
+				hasErr = do.err != nil
+				if debugDecoder && hasErr {
+					println("executeSequences returned:", do.err)
+				}
+				do.b = block.dst
+			}
+			if !hasErr {
+				decodedFrame += uint64(len(do.b))
+				if decodedFrame > fcs {
+					println("fcs exceeded", block.Last, fcs, decodedFrame)
+					do.err = ErrFrameSizeExceeded
+					hasErr = true
+				} else if block.Last && fcs != fcsUnknown && decodedFrame != fcs {
+					do.err = ErrFrameSizeMismatch
+					hasErr = true
+				} else {
+					if debugDecoder {
+						println("fcs ok", block.Last, fcs, decodedFrame)
+					}
+				}
+			}
+			output <- do
+		}
+		close(output)
+		frameHistCache = hist.b
+		wg.Done()
+		if debugDecoder {
+			println("decoder goroutines finished")
+		}
+	}()
+
+decodeStream:
+	for {
+		frame := d.frame
+		if debugDecoder {
+			println("New frame...")
+		}
+		var historySent bool
+		frame.history.reset()
+		err := frame.reset(&br)
+		if debugDecoder && err != nil {
+			println("Frame decoder returned", err)
+		}
+		if err == nil && frame.DictionaryID != nil {
+			dict, ok := d.dicts[*frame.DictionaryID]
+			if !ok {
+				err = ErrUnknownDictionary
+			} else {
+				frame.history.setDict(&dict)
+			}
+		}
+		if err == nil && d.frame.WindowSize > d.o.maxWindowSize {
+			err = ErrDecoderSizeExceeded
+		}
+		if err != nil {
+			select {
+			case <-ctx.Done():
+			case dec := <-d.decoders:
+				dec.sendErr(err)
+				seqPrepare <- dec
+			}
+			break decodeStream
+		}
+
+		// Go through all blocks of the frame.
+		for {
+			var dec *blockDec
+			select {
+			case <-ctx.Done():
+				break decodeStream
+			case dec = <-d.decoders:
+				// Once we have a decoder, we MUST return it.
+			}
+			err := frame.next(dec)
+			if !historySent {
+				h := frame.history
+				if debugDecoder {
+					println("Alloc History:", h.allocFrameBuffer)
+				}
+				dec.async.newHist = &h
+				dec.async.fcs = frame.FrameContentSize
+				historySent = true
+			} else {
+				dec.async.newHist = nil
+			}
+			if debugDecoder && err != nil {
+				println("next block returned error:", err)
+			}
+			dec.err = err
+			dec.checkCRC = nil
+			if dec.Last && frame.HasCheckSum && err == nil {
+				crc, err := frame.rawInput.readSmall(4)
+				if err != nil {
+					println("CRC missing?", err)
+					dec.err = err
+				}
+				var tmp [4]byte
+				copy(tmp[:], crc)
+				dec.checkCRC = tmp[:]
+				if debugDecoder {
+					println("found crc to check:", dec.checkCRC)
+				}
+			}
+			err = dec.err
+			last := dec.Last
+			seqPrepare <- dec
+			if err != nil {
+				break decodeStream
+			}
+			if last {
+				break
+			}
+		}
+	}
+	close(seqPrepare)
+	wg.Wait()
+	d.frame.history.b = frameHistCache
+}

vendor/github.com/klauspost/compress/zstd/decoder_options.go

@@ -0,0 +1,114 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"errors"
+	"runtime"
+)
+
+// DOption is an option for creating a decoder.
+type DOption func(*decoderOptions) error
+
+// options retains accumulated state of multiple options.
+type decoderOptions struct {
+	lowMem         bool
+	concurrent     int
+	maxDecodedSize uint64
+	maxWindowSize  uint64
+	dicts          []dict
+}
+
+func (o *decoderOptions) setDefault() {
+	*o = decoderOptions{
+		// use less ram: true for now, but may change.
+		lowMem:        true,
+		concurrent:    runtime.GOMAXPROCS(0),
+		maxWindowSize: MaxWindowSize,
+	}
+	if o.concurrent > 4 {
+		o.concurrent = 4
+	}
+	o.maxDecodedSize = 64 << 30
+}
+
+// WithDecoderLowmem will set whether to use a lower amount of memory,
+// but possibly have to allocate more while running.
+func WithDecoderLowmem(b bool) DOption {
+	return func(o *decoderOptions) error { o.lowMem = b; return nil }
+}
+
+// WithDecoderConcurrency sets the number of created decoders.
+// When decoding block with DecodeAll, this will limit the number
+// of possible concurrently running decodes.
+// When decoding streams, this will limit the number of
+// inflight blocks.
+// When decoding streams and setting maximum to 1,
+// no async decoding will be done.
+// When a value of 0 is provided GOMAXPROCS will be used.
+// By default this will be set to 4 or GOMAXPROCS, whatever is lower.
+func WithDecoderConcurrency(n int) DOption {
+	return func(o *decoderOptions) error {
+		if n < 0 {
+			return errors.New("concurrency must be at least 1")
+		}
+		if n == 0 {
+			o.concurrent = runtime.GOMAXPROCS(0)
+		} else {
+			o.concurrent = n
+		}
+		return nil
+	}
+}
+
+// WithDecoderMaxMemory allows to set a maximum decoded size for in-memory
+// non-streaming operations or maximum window size for streaming operations.
+// This can be used to control memory usage of potentially hostile content.
+// Maximum is 1 << 63 bytes. Default is 64GiB.
+func WithDecoderMaxMemory(n uint64) DOption {
+	return func(o *decoderOptions) error {
+		if n == 0 {
+			return errors.New("WithDecoderMaxMemory must be at least 1")
+		}
+		if n > 1<<63 {
+			return errors.New("WithDecoderMaxmemory must be less than 1 << 63")
+		}
+		o.maxDecodedSize = n
+		return nil
+	}
+}
+
+// WithDecoderDicts allows to register one or more dictionaries for the decoder.
+// If several dictionaries with the same ID is provided the last one will be used.
+func WithDecoderDicts(dicts ...[]byte) DOption {
+	return func(o *decoderOptions) error {
+		for _, b := range dicts {
+			d, err := loadDict(b)
+			if err != nil {
+				return err
+			}
+			o.dicts = append(o.dicts, *d)
+		}
+		return nil
+	}
+}
+
+// WithDecoderMaxWindow allows to set a maximum window size for decodes.
+// This allows rejecting packets that will cause big memory usage.
+// The Decoder will likely allocate more memory based on the WithDecoderLowmem setting.
+// If WithDecoderMaxMemory is set to a lower value, that will be used.
+// Default is 512MB, Maximum is ~3.75 TB as per zstandard spec.
+func WithDecoderMaxWindow(size uint64) DOption {
+	return func(o *decoderOptions) error {
+		if size < MinWindowSize {
+			return errors.New("WithMaxWindowSize must be at least 1KB, 1024 bytes")
+		}
+		if size > (1<<41)+7*(1<<38) {
+			return errors.New("WithMaxWindowSize must be less than (1<<41) + 7*(1<<38) ~ 3.75TB")
+		}
+		o.maxWindowSize = size
+		return nil
+	}
+}

vendor/github.com/klauspost/compress/zstd/dict.go

@@ -0,0 +1,122 @@
+package zstd
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+
+	"github.com/klauspost/compress/huff0"
+)
+
+type dict struct {
+	id uint32
+
+	litEnc              *huff0.Scratch
+	llDec, ofDec, mlDec sequenceDec
+	//llEnc, ofEnc, mlEnc []*fseEncoder
+	offsets [3]int
+	content []byte
+}
+
+var dictMagic = [4]byte{0x37, 0xa4, 0x30, 0xec}
+
+// ID returns the dictionary id or 0 if d is nil.
+func (d *dict) ID() uint32 {
+	if d == nil {
+		return 0
+	}
+	return d.id
+}
+
+// DictContentSize returns the dictionary content size or 0 if d is nil.
+func (d *dict) DictContentSize() int {
+	if d == nil {
+		return 0
+	}
+	return len(d.content)
+}
+
+// Load a dictionary as described in
+// https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#dictionary-format
+func loadDict(b []byte) (*dict, error) {
+	// Check static field size.
+	if len(b) <= 8+(3*4) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	d := dict{
+		llDec: sequenceDec{fse: &fseDecoder{}},
+		ofDec: sequenceDec{fse: &fseDecoder{}},
+		mlDec: sequenceDec{fse: &fseDecoder{}},
+	}
+	if !bytes.Equal(b[:4], dictMagic[:]) {
+		return nil, ErrMagicMismatch
+	}
+	d.id = binary.LittleEndian.Uint32(b[4:8])
+	if d.id == 0 {
+		return nil, errors.New("dictionaries cannot have ID 0")
+	}
+
+	// Read literal table
+	var err error
+	d.litEnc, b, err = huff0.ReadTable(b[8:], nil)
+	if err != nil {
+		return nil, err
+	}
+	d.litEnc.Reuse = huff0.ReusePolicyMust
+
+	br := byteReader{
+		b:   b,
+		off: 0,
+	}
+	readDec := func(i tableIndex, dec *fseDecoder) error {
+		if err := dec.readNCount(&br, uint16(maxTableSymbol[i])); err != nil {
+			return err
+		}
+		if br.overread() {
+			return io.ErrUnexpectedEOF
+		}
+		err = dec.transform(symbolTableX[i])
+		if err != nil {
+			println("Transform table error:", err)
+			return err
+		}
+		if debugDecoder || debugEncoder {
+			println("Read table ok", "symbolLen:", dec.symbolLen)
+		}
+		// Set decoders as predefined so they aren't reused.
+		dec.preDefined = true
+		return nil
+	}
+
+	if err := readDec(tableOffsets, d.ofDec.fse); err != nil {
+		return nil, err
+	}
+	if err := readDec(tableMatchLengths, d.mlDec.fse); err != nil {
+		return nil, err
+	}
+	if err := readDec(tableLiteralLengths, d.llDec.fse); err != nil {
+		return nil, err
+	}
+	if br.remain() < 12 {
+		return nil, io.ErrUnexpectedEOF
+	}
+
+	d.offsets[0] = int(br.Uint32())
+	br.advance(4)
+	d.offsets[1] = int(br.Uint32())
+	br.advance(4)
+	d.offsets[2] = int(br.Uint32())
+	br.advance(4)
+	if d.offsets[0] <= 0 || d.offsets[1] <= 0 || d.offsets[2] <= 0 {
+		return nil, errors.New("invalid offset in dictionary")
+	}
+	d.content = make([]byte, br.remain())
+	copy(d.content, br.unread())
+	if d.offsets[0] > len(d.content) || d.offsets[1] > len(d.content) || d.offsets[2] > len(d.content) {
+		return nil, fmt.Errorf("initial offset bigger than dictionary content size %d, offsets: %v", len(d.content), d.offsets)
+	}
+
+	return &d, nil
+}

vendor/github.com/klauspost/compress/zstd/enc_base.go

@@ -0,0 +1,188 @@
+package zstd
+
+import (
+	"fmt"
+	"math/bits"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+const (
+	dictShardBits = 6
+)
+
+type fastBase struct {
+	// cur is the offset at the start of hist
+	cur int32
+	// maximum offset. Should be at least 2x block size.
+	maxMatchOff int32
+	hist        []byte
+	crc         *xxhash.Digest
+	tmp         [8]byte
+	blk         *blockEnc
+	lastDictID  uint32
+	lowMem      bool
+}
+
+// CRC returns the underlying CRC writer.
+func (e *fastBase) CRC() *xxhash.Digest {
+	return e.crc
+}
+
+// AppendCRC will append the CRC to the destination slice and return it.
+func (e *fastBase) AppendCRC(dst []byte) []byte {
+	crc := e.crc.Sum(e.tmp[:0])
+	dst = append(dst, crc[7], crc[6], crc[5], crc[4])
+	return dst
+}
+
+// WindowSize returns the window size of the encoder,
+// or a window size small enough to contain the input size, if > 0.
+func (e *fastBase) WindowSize(size int64) int32 {
+	if size > 0 && size < int64(e.maxMatchOff) {
+		b := int32(1) << uint(bits.Len(uint(size)))
+		// Keep minimum window.
+		if b < 1024 {
+			b = 1024
+		}
+		return b
+	}
+	return e.maxMatchOff
+}
+
+// Block returns the current block.
+func (e *fastBase) Block() *blockEnc {
+	return e.blk
+}
+
+func (e *fastBase) addBlock(src []byte) int32 {
+	if debugAsserts && e.cur > bufferReset {
+		panic(fmt.Sprintf("ecur (%d) > buffer reset (%d)", e.cur, bufferReset))
+	}
+	// check if we have space already
+	if len(e.hist)+len(src) > cap(e.hist) {
+		if cap(e.hist) == 0 {
+			e.ensureHist(len(src))
+		} else {
+			if cap(e.hist) < int(e.maxMatchOff+maxCompressedBlockSize) {
+				panic(fmt.Errorf("unexpected buffer cap %d, want at least %d with window %d", cap(e.hist), e.maxMatchOff+maxCompressedBlockSize, e.maxMatchOff))
+			}
+			// Move down
+			offset := int32(len(e.hist)) - e.maxMatchOff
+			copy(e.hist[0:e.maxMatchOff], e.hist[offset:])
+			e.cur += offset
+			e.hist = e.hist[:e.maxMatchOff]
+		}
+	}
+	s := int32(len(e.hist))
+	e.hist = append(e.hist, src...)
+	return s
+}
+
+// ensureHist will ensure that history can keep at least this many bytes.
+func (e *fastBase) ensureHist(n int) {
+	if cap(e.hist) >= n {
+		return
+	}
+	l := e.maxMatchOff
+	if (e.lowMem && e.maxMatchOff > maxCompressedBlockSize) || e.maxMatchOff <= maxCompressedBlockSize {
+		l += maxCompressedBlockSize
+	} else {
+		l += e.maxMatchOff
+	}
+	// Make it at least 1MB.
+	if l < 1<<20 && !e.lowMem {
+		l = 1 << 20
+	}
+	// Make it at least the requested size.
+	if l < int32(n) {
+		l = int32(n)
+	}
+	e.hist = make([]byte, 0, l)
+}
+
+// useBlock will replace the block with the provided one,
+// but transfer recent offsets from the previous.
+func (e *fastBase) UseBlock(enc *blockEnc) {
+	enc.reset(e.blk)
+	e.blk = enc
+}
+
+func (e *fastBase) matchlen(s, t int32, src []byte) int32 {
+	if debugAsserts {
+		if s < 0 {
+			err := fmt.Sprintf("s (%d) < 0", s)
+			panic(err)
+		}
+		if t < 0 {
+			err := fmt.Sprintf("s (%d) < 0", s)
+			panic(err)
+		}
+		if s-t > e.maxMatchOff {
+			err := fmt.Sprintf("s (%d) - t (%d) > maxMatchOff (%d)", s, t, e.maxMatchOff)
+			panic(err)
+		}
+		if len(src)-int(s) > maxCompressedBlockSize {
+			panic(fmt.Sprintf("len(src)-s (%d) > maxCompressedBlockSize (%d)", len(src)-int(s), maxCompressedBlockSize))
+		}
+	}
+	a := src[s:]
+	b := src[t:]
+	b = b[:len(a)]
+	end := int32((len(a) >> 3) << 3)
+	for i := int32(0); i < end; i += 8 {
+		if diff := load6432(a, i) ^ load6432(b, i); diff != 0 {
+			return i + int32(bits.TrailingZeros64(diff)>>3)
+		}
+	}
+
+	a = a[end:]
+	b = b[end:]
+	for i := range a {
+		if a[i] != b[i] {
+			return int32(i) + end
+		}
+	}
+	return int32(len(a)) + end
+}
+
+// Reset the encoding table.
+func (e *fastBase) resetBase(d *dict, singleBlock bool) {
+	if e.blk == nil {
+		e.blk = &blockEnc{lowMem: e.lowMem}
+		e.blk.init()
+	} else {
+		e.blk.reset(nil)
+	}
+	e.blk.initNewEncode()
+	if e.crc == nil {
+		e.crc = xxhash.New()
+	} else {
+		e.crc.Reset()
+	}
+	if d != nil {
+		low := e.lowMem
+		if singleBlock {
+			e.lowMem = true
+		}
+		e.ensureHist(d.DictContentSize() + maxCompressedBlockSize)
+		e.lowMem = low
+	}
+
+	// We offset current position so everything will be out of reach.
+	// If above reset line, history will be purged.
+	if e.cur < bufferReset {
+		e.cur += e.maxMatchOff + int32(len(e.hist))
+	}
+	e.hist = e.hist[:0]
+	if d != nil {
+		// Set offsets (currently not used)
+		for i, off := range d.offsets {
+			e.blk.recentOffsets[i] = uint32(off)
+			e.blk.prevRecentOffsets[i] = e.blk.recentOffsets[i]
+		}
+		// Transfer litenc.
+		e.blk.dictLitEnc = d.litEnc
+		e.hist = append(e.hist, d.content...)
+	}
+}

vendor/github.com/klauspost/compress/zstd/enc_best.go

@@ -0,0 +1,558 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"bytes"
+	"fmt"
+
+	"github.com/klauspost/compress"
+)
+
+const (
+	bestLongTableBits = 22                     // Bits used in the long match table
+	bestLongTableSize = 1 << bestLongTableBits // Size of the table
+	bestLongLen       = 8                      // Bytes used for table hash
+
+	// Note: Increasing the short table bits or making the hash shorter
+	// can actually lead to compression degradation since it will 'steal' more from the
+	// long match table and match offsets are quite big.
+	// This greatly depends on the type of input.
+	bestShortTableBits = 18                      // Bits used in the short match table
+	bestShortTableSize = 1 << bestShortTableBits // Size of the table
+	bestShortLen       = 4                       // Bytes used for table hash
+
+)
+
+type match struct {
+	offset int32
+	s      int32
+	length int32
+	rep    int32
+	est    int32
+}
+
+const highScore = 25000
+
+// estBits will estimate output bits from predefined tables.
+func (m *match) estBits(bitsPerByte int32) {
+	mlc := mlCode(uint32(m.length - zstdMinMatch))
+	var ofc uint8
+	if m.rep < 0 {
+		ofc = ofCode(uint32(m.s-m.offset) + 3)
+	} else {
+		ofc = ofCode(uint32(m.rep))
+	}
+	// Cost, excluding
+	ofTT, mlTT := fsePredefEnc[tableOffsets].ct.symbolTT[ofc], fsePredefEnc[tableMatchLengths].ct.symbolTT[mlc]
+
+	// Add cost of match encoding...
+	m.est = int32(ofTT.outBits + mlTT.outBits)
+	m.est += int32(ofTT.deltaNbBits>>16 + mlTT.deltaNbBits>>16)
+	// Subtract savings compared to literal encoding...
+	m.est -= (m.length * bitsPerByte) >> 10
+	if m.est > 0 {
+		// Unlikely gain..
+		m.length = 0
+		m.est = highScore
+	}
+}
+
+// bestFastEncoder uses 2 tables, one for short matches (5 bytes) and one for long matches.
+// The long match table contains the previous entry with the same hash,
+// effectively making it a "chain" of length 2.
+// When we find a long match we choose between the two values and select the longest.
+// When we find a short match, after checking the long, we check if we can find a long at n+1
+// and that it is longer (lazy matching).
+type bestFastEncoder struct {
+	fastBase
+	table         [bestShortTableSize]prevEntry
+	longTable     [bestLongTableSize]prevEntry
+	dictTable     []prevEntry
+	dictLongTable []prevEntry
+}
+
+// Encode improves compression...
+func (e *bestFastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 4
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = prevEntry{}
+			}
+			for i := range e.longTable[:] {
+				e.longTable[i] = prevEntry{}
+			}
+			e.cur = e.maxMatchOff
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			v2 := e.table[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.table[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			v2 := e.longTable[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.longTable[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+
+	s := e.addBlock(src)
+	blk.size = len(src)
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	// Use this to estimate literal cost.
+	// Scaled by 10 bits.
+	bitsPerByte := int32((compress.ShannonEntropyBits(src) * 1024) / len(src))
+	// Huffman can never go < 1 bit/byte
+	if bitsPerByte < 1024 {
+		bitsPerByte = 1024
+	}
+
+	// Override src
+	src = e.hist
+	sLimit := int32(len(src)) - inputMargin
+	const kSearchStrength = 10
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+	cv := load6432(src, s)
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+	offset3 := int32(blk.recentOffsets[2])
+
+	addLiterals := func(s *seq, until int32) {
+		if until == nextEmit {
+			return
+		}
+		blk.literals = append(blk.literals, src[nextEmit:until]...)
+		s.litLen = uint32(until - nextEmit)
+	}
+	_ = addLiterals
+
+	if debugEncoder {
+		println("recent offsets:", blk.recentOffsets)
+	}
+
+encodeLoop:
+	for {
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+
+		if debugAsserts && canRepeat && offset1 == 0 {
+			panic("offset0 was 0")
+		}
+
+		bestOf := func(a, b match) match {
+			if a.est+(a.s-b.s)*bitsPerByte>>10 < b.est+(b.s-a.s)*bitsPerByte>>10 {
+				return a
+			}
+			return b
+		}
+		const goodEnough = 100
+
+		nextHashL := hashLen(cv, bestLongTableBits, bestLongLen)
+		nextHashS := hashLen(cv, bestShortTableBits, bestShortLen)
+		candidateL := e.longTable[nextHashL]
+		candidateS := e.table[nextHashS]
+
+		matchAt := func(offset int32, s int32, first uint32, rep int32) match {
+			if s-offset >= e.maxMatchOff || load3232(src, offset) != first {
+				return match{s: s, est: highScore}
+			}
+			if debugAsserts {
+				if !bytes.Equal(src[s:s+4], src[offset:offset+4]) {
+					panic(fmt.Sprintf("first match mismatch: %v != %v, first: %08x", src[s:s+4], src[offset:offset+4], first))
+				}
+			}
+			m := match{offset: offset, s: s, length: 4 + e.matchlen(s+4, offset+4, src), rep: rep}
+			m.estBits(bitsPerByte)
+			return m
+		}
+
+		best := bestOf(matchAt(candidateL.offset-e.cur, s, uint32(cv), -1), matchAt(candidateL.prev-e.cur, s, uint32(cv), -1))
+		best = bestOf(best, matchAt(candidateS.offset-e.cur, s, uint32(cv), -1))
+		best = bestOf(best, matchAt(candidateS.prev-e.cur, s, uint32(cv), -1))
+
+		if canRepeat && best.length < goodEnough {
+			cv32 := uint32(cv >> 8)
+			spp := s + 1
+			best = bestOf(best, matchAt(spp-offset1, spp, cv32, 1))
+			best = bestOf(best, matchAt(spp-offset2, spp, cv32, 2))
+			best = bestOf(best, matchAt(spp-offset3, spp, cv32, 3))
+			if best.length > 0 {
+				cv32 = uint32(cv >> 24)
+				spp += 2
+				best = bestOf(best, matchAt(spp-offset1, spp, cv32, 1))
+				best = bestOf(best, matchAt(spp-offset2, spp, cv32, 2))
+				best = bestOf(best, matchAt(spp-offset3, spp, cv32, 3))
+			}
+		}
+		// Load next and check...
+		e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: candidateL.offset}
+		e.table[nextHashS] = prevEntry{offset: s + e.cur, prev: candidateS.offset}
+
+		// Look far ahead, unless we have a really long match already...
+		if best.length < goodEnough {
+			// No match found, move forward on input, no need to check forward...
+			if best.length < 4 {
+				s += 1 + (s-nextEmit)>>(kSearchStrength-1)
+				if s >= sLimit {
+					break encodeLoop
+				}
+				cv = load6432(src, s)
+				continue
+			}
+
+			s++
+			candidateS = e.table[hashLen(cv>>8, bestShortTableBits, bestShortLen)]
+			cv = load6432(src, s)
+			cv2 := load6432(src, s+1)
+			candidateL = e.longTable[hashLen(cv, bestLongTableBits, bestLongLen)]
+			candidateL2 := e.longTable[hashLen(cv2, bestLongTableBits, bestLongLen)]
+
+			// Short at s+1
+			best = bestOf(best, matchAt(candidateS.offset-e.cur, s, uint32(cv), -1))
+			// Long at s+1, s+2
+			best = bestOf(best, matchAt(candidateL.offset-e.cur, s, uint32(cv), -1))
+			best = bestOf(best, matchAt(candidateL.prev-e.cur, s, uint32(cv), -1))
+			best = bestOf(best, matchAt(candidateL2.offset-e.cur, s+1, uint32(cv2), -1))
+			best = bestOf(best, matchAt(candidateL2.prev-e.cur, s+1, uint32(cv2), -1))
+			if false {
+				// Short at s+3.
+				// Too often worse...
+				best = bestOf(best, matchAt(e.table[hashLen(cv2>>8, bestShortTableBits, bestShortLen)].offset-e.cur, s+2, uint32(cv2>>8), -1))
+			}
+			// See if we can find a better match by checking where the current best ends.
+			// Use that offset to see if we can find a better full match.
+			if sAt := best.s + best.length; sAt < sLimit {
+				nextHashL := hashLen(load6432(src, sAt), bestLongTableBits, bestLongLen)
+				candidateEnd := e.longTable[nextHashL]
+				if pos := candidateEnd.offset - e.cur - best.length; pos >= 0 {
+					bestEnd := bestOf(best, matchAt(pos, best.s, load3232(src, best.s), -1))
+					if pos := candidateEnd.prev - e.cur - best.length; pos >= 0 {
+						bestEnd = bestOf(bestEnd, matchAt(pos, best.s, load3232(src, best.s), -1))
+					}
+					best = bestEnd
+				}
+			}
+		}
+
+		if debugAsserts {
+			if !bytes.Equal(src[best.s:best.s+best.length], src[best.offset:best.offset+best.length]) {
+				panic(fmt.Sprintf("match mismatch: %v != %v", src[best.s:best.s+best.length], src[best.offset:best.offset+best.length]))
+			}
+		}
+
+		// We have a match, we can store the forward value
+		if best.rep > 0 {
+			s = best.s
+			var seq seq
+			seq.matchLen = uint32(best.length - zstdMinMatch)
+
+			// We might be able to match backwards.
+			// Extend as long as we can.
+			start := best.s
+			// We end the search early, so we don't risk 0 literals
+			// and have to do special offset treatment.
+			startLimit := nextEmit + 1
+
+			tMin := s - e.maxMatchOff
+			if tMin < 0 {
+				tMin = 0
+			}
+			repIndex := best.offset
+			for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
+				repIndex--
+				start--
+				seq.matchLen++
+			}
+			addLiterals(&seq, start)
+
+			// rep 0
+			seq.offset = uint32(best.rep)
+			if debugSequences {
+				println("repeat sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Index match start+1 (long) -> s - 1
+			index0 := s
+			s = best.s + best.length
+
+			nextEmit = s
+			if s >= sLimit {
+				if debugEncoder {
+					println("repeat ended", s, best.length)
+
+				}
+				break encodeLoop
+			}
+			// Index skipped...
+			off := index0 + e.cur
+			for index0 < s-1 {
+				cv0 := load6432(src, index0)
+				h0 := hashLen(cv0, bestLongTableBits, bestLongLen)
+				h1 := hashLen(cv0, bestShortTableBits, bestShortLen)
+				e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+				e.table[h1] = prevEntry{offset: off, prev: e.table[h1].offset}
+				off++
+				index0++
+			}
+			switch best.rep {
+			case 2:
+				offset1, offset2 = offset2, offset1
+			case 3:
+				offset1, offset2, offset3 = offset3, offset1, offset2
+			}
+			cv = load6432(src, s)
+			continue
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// We'll later see if more than 4 bytes.
+		s = best.s
+		t := best.offset
+		offset1, offset2, offset3 = s-t, offset1, offset2
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && int(offset1) > len(src) {
+			panic("invalid offset")
+		}
+
+		// Extend the n-byte match as long as possible.
+		l := best.length
+
+		// Extend backwards
+		tMin := s - e.maxMatchOff
+		if tMin < 0 {
+			tMin = 0
+		}
+		for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength {
+			s--
+			t--
+			l++
+		}
+
+		// Write our sequence
+		var seq seq
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+		if s >= sLimit {
+			break encodeLoop
+		}
+
+		// Index match start+1 (long) -> s - 1
+		index0 := s - l + 1
+		// every entry
+		for index0 < s-1 {
+			cv0 := load6432(src, index0)
+			h0 := hashLen(cv0, bestLongTableBits, bestLongLen)
+			h1 := hashLen(cv0, bestShortTableBits, bestShortLen)
+			off := index0 + e.cur
+			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+			e.table[h1] = prevEntry{offset: off, prev: e.table[h1].offset}
+			index0++
+		}
+
+		cv = load6432(src, s)
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashS := hashLen(cv, bestShortTableBits, bestShortLen)
+			nextHashL := hashLen(cv, bestLongTableBits, bestLongLen)
+
+			// We have at least 4 byte match.
+			// No need to check backwards. We come straight from a match
+			l := 4 + e.matchlen(s+4, o2+4, src)
+
+			e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: e.longTable[nextHashL].offset}
+			e.table[nextHashS] = prevEntry{offset: s + e.cur, prev: e.table[nextHashS].offset}
+			seq.matchLen = uint32(l) - zstdMinMatch
+			seq.litLen = 0
+
+			// Since litlen is always 0, this is offset 1.
+			seq.offset = 1
+			s += l
+			nextEmit = s
+			if debugSequences {
+				println("sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Swap offset 1 and 2.
+			offset1, offset2 = offset2, offset1
+			if s >= sLimit {
+				// Finished
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+	}
+
+	if int(nextEmit) < len(src) {
+		blk.literals = append(blk.literals, src[nextEmit:]...)
+		blk.extraLits = len(src) - int(nextEmit)
+	}
+	blk.recentOffsets[0] = uint32(offset1)
+	blk.recentOffsets[1] = uint32(offset2)
+	blk.recentOffsets[2] = uint32(offset3)
+	if debugEncoder {
+		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
+	}
+}
+
+// EncodeNoHist will encode a block with no history and no following blocks.
+// Most notable difference is that src will not be copied for history and
+// we do not need to check for max match length.
+func (e *bestFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	e.ensureHist(len(src))
+	e.Encode(blk, src)
+}
+
+// Reset will reset and set a dictionary if not nil
+func (e *bestFastEncoder) Reset(d *dict, singleBlock bool) {
+	e.resetBase(d, singleBlock)
+	if d == nil {
+		return
+	}
+	// Init or copy dict table
+	if len(e.dictTable) != len(e.table) || d.id != e.lastDictID {
+		if len(e.dictTable) != len(e.table) {
+			e.dictTable = make([]prevEntry, len(e.table))
+		}
+		end := int32(len(d.content)) - 8 + e.maxMatchOff
+		for i := e.maxMatchOff; i < end; i += 4 {
+			const hashLog = bestShortTableBits
+
+			cv := load6432(d.content, i-e.maxMatchOff)
+			nextHash := hashLen(cv, hashLog, bestShortLen)      // 0 -> 4
+			nextHash1 := hashLen(cv>>8, hashLog, bestShortLen)  // 1 -> 5
+			nextHash2 := hashLen(cv>>16, hashLog, bestShortLen) // 2 -> 6
+			nextHash3 := hashLen(cv>>24, hashLog, bestShortLen) // 3 -> 7
+			e.dictTable[nextHash] = prevEntry{
+				prev:   e.dictTable[nextHash].offset,
+				offset: i,
+			}
+			e.dictTable[nextHash1] = prevEntry{
+				prev:   e.dictTable[nextHash1].offset,
+				offset: i + 1,
+			}
+			e.dictTable[nextHash2] = prevEntry{
+				prev:   e.dictTable[nextHash2].offset,
+				offset: i + 2,
+			}
+			e.dictTable[nextHash3] = prevEntry{
+				prev:   e.dictTable[nextHash3].offset,
+				offset: i + 3,
+			}
+		}
+		e.lastDictID = d.id
+	}
+
+	// Init or copy dict table
+	if len(e.dictLongTable) != len(e.longTable) || d.id != e.lastDictID {
+		if len(e.dictLongTable) != len(e.longTable) {
+			e.dictLongTable = make([]prevEntry, len(e.longTable))
+		}
+		if len(d.content) >= 8 {
+			cv := load6432(d.content, 0)
+			h := hashLen(cv, bestLongTableBits, bestLongLen)
+			e.dictLongTable[h] = prevEntry{
+				offset: e.maxMatchOff,
+				prev:   e.dictLongTable[h].offset,
+			}
+
+			end := int32(len(d.content)) - 8 + e.maxMatchOff
+			off := 8 // First to read
+			for i := e.maxMatchOff + 1; i < end; i++ {
+				cv = cv>>8 | (uint64(d.content[off]) << 56)
+				h := hashLen(cv, bestLongTableBits, bestLongLen)
+				e.dictLongTable[h] = prevEntry{
+					offset: i,
+					prev:   e.dictLongTable[h].offset,
+				}
+				off++
+			}
+		}
+		e.lastDictID = d.id
+	}
+	// Reset table to initial state
+	copy(e.longTable[:], e.dictLongTable)
+
+	e.cur = e.maxMatchOff
+	// Reset table to initial state
+	copy(e.table[:], e.dictTable)
+}

vendor/github.com/klauspost/compress/zstd/enc_fast.go

@@ -0,0 +1,898 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"fmt"
+)
+
+const (
+	tableBits        = 15                               // Bits used in the table
+	tableSize        = 1 << tableBits                   // Size of the table
+	tableShardCnt    = 1 << (tableBits - dictShardBits) // Number of shards in the table
+	tableShardSize   = tableSize / tableShardCnt        // Size of an individual shard
+	tableFastHashLen = 6
+	tableMask        = tableSize - 1 // Mask for table indices. Redundant, but can eliminate bounds checks.
+	maxMatchLength   = 131074
+)
+
+type tableEntry struct {
+	val    uint32
+	offset int32
+}
+
+type fastEncoder struct {
+	fastBase
+	table [tableSize]tableEntry
+}
+
+type fastEncoderDict struct {
+	fastEncoder
+	dictTable       []tableEntry
+	tableShardDirty [tableShardCnt]bool
+	allDirty        bool
+}
+
+// Encode mimmics functionality in zstd_fast.c
+func (e *fastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		inputMargin            = 8
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			e.cur = e.maxMatchOff
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v < minOff {
+				v = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+			}
+			e.table[i].offset = v
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+
+	s := e.addBlock(src)
+	blk.size = len(src)
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	// Override src
+	src = e.hist
+	sLimit := int32(len(src)) - inputMargin
+	// stepSize is the number of bytes to skip on every main loop iteration.
+	// It should be >= 2.
+	const stepSize = 2
+
+	// TEMPLATE
+	const hashLog = tableBits
+	// seems global, but would be nice to tweak.
+	const kSearchStrength = 6
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+	cv := load6432(src, s)
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+
+	addLiterals := func(s *seq, until int32) {
+		if until == nextEmit {
+			return
+		}
+		blk.literals = append(blk.literals, src[nextEmit:until]...)
+		s.litLen = uint32(until - nextEmit)
+	}
+	if debugEncoder {
+		println("recent offsets:", blk.recentOffsets)
+	}
+
+encodeLoop:
+	for {
+		// t will contain the match offset when we find one.
+		// When existing the search loop, we have already checked 4 bytes.
+		var t int32
+
+		// We will not use repeat offsets across blocks.
+		// By not using them for the first 3 matches
+		canRepeat := len(blk.sequences) > 2
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
+			candidate := e.table[nextHash]
+			candidate2 := e.table[nextHash2]
+			repIndex := s - offset1 + 2
+
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)}
+
+			if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) {
+				// Consider history as well.
+				var seq seq
+				var length int32
+				length = 4 + e.matchlen(s+6, repIndex+4, src)
+				seq.matchLen = uint32(length - zstdMinMatch)
+
+				// We might be able to match backwards.
+				// Extend as long as we can.
+				start := s + 2
+				// We end the search early, so we don't risk 0 literals
+				// and have to do special offset treatment.
+				startLimit := nextEmit + 1
+
+				sMin := s - e.maxMatchOff
+				if sMin < 0 {
+					sMin = 0
+				}
+				for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch {
+					repIndex--
+					start--
+					seq.matchLen++
+				}
+				addLiterals(&seq, start)
+
+				// rep 0
+				seq.offset = 1
+				if debugSequences {
+					println("repeat sequence", seq, "next s:", s)
+				}
+				blk.sequences = append(blk.sequences, seq)
+				s += length + 2
+				nextEmit = s
+				if s >= sLimit {
+					if debugEncoder {
+						println("repeat ended", s, length)
+
+					}
+					break encodeLoop
+				}
+				cv = load6432(src, s)
+				continue
+			}
+			coffset0 := s - (candidate.offset - e.cur)
+			coffset1 := s - (candidate2.offset - e.cur) + 1
+			if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val {
+				// found a regular match
+				t = candidate.offset - e.cur
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				break
+			}
+
+			if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val {
+				// found a regular match
+				t = candidate2.offset - e.cur
+				s++
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugAsserts && t < 0 {
+					panic("t<0")
+				}
+				break
+			}
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+		// A 4-byte match has been found. We'll later see if more than 4 bytes.
+		offset2 = offset1
+		offset1 = s - t
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && canRepeat && int(offset1) > len(src) {
+			panic("invalid offset")
+		}
+
+		// Extend the 4-byte match as long as possible.
+		l := e.matchlen(s+4, t+4, src) + 4
+
+		// Extend backwards
+		tMin := s - e.maxMatchOff
+		if tMin < 0 {
+			tMin = 0
+		}
+		for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength {
+			s--
+			t--
+			l++
+		}
+
+		// Write our sequence.
+		var seq seq
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		// Don't use repeat offsets
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+		if s >= sLimit {
+			break encodeLoop
+		}
+		cv = load6432(src, s)
+
+		// Check offset 2
+		if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) {
+			// We have at least 4 byte match.
+			// No need to check backwards. We come straight from a match
+			l := 4 + e.matchlen(s+4, o2+4, src)
+
+			// Store this, since we have it.
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			seq.matchLen = uint32(l) - zstdMinMatch
+			seq.litLen = 0
+			// Since litlen is always 0, this is offset 1.
+			seq.offset = 1
+			s += l
+			nextEmit = s
+			if debugSequences {
+				println("sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Swap offset 1 and 2.
+			offset1, offset2 = offset2, offset1
+			if s >= sLimit {
+				break encodeLoop
+			}
+			// Prepare next loop.
+			cv = load6432(src, s)
+		}
+	}
+
+	if int(nextEmit) < len(src) {
+		blk.literals = append(blk.literals, src[nextEmit:]...)
+		blk.extraLits = len(src) - int(nextEmit)
+	}
+	blk.recentOffsets[0] = uint32(offset1)
+	blk.recentOffsets[1] = uint32(offset2)
+	if debugEncoder {
+		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
+	}
+}
+
+// EncodeNoHist will encode a block with no history and no following blocks.
+// Most notable difference is that src will not be copied for history and
+// we do not need to check for max match length.
+func (e *fastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	const (
+		inputMargin            = 8
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+	if debugEncoder {
+		if len(src) > maxBlockSize {
+			panic("src too big")
+		}
+	}
+
+	// Protect against e.cur wraparound.
+	if e.cur >= bufferReset {
+		for i := range e.table[:] {
+			e.table[i] = tableEntry{}
+		}
+		e.cur = e.maxMatchOff
+	}
+
+	s := int32(0)
+	blk.size = len(src)
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	sLimit := int32(len(src)) - inputMargin
+	// stepSize is the number of bytes to skip on every main loop iteration.
+	// It should be >= 2.
+	const stepSize = 2
+
+	// TEMPLATE
+	const hashLog = tableBits
+	// seems global, but would be nice to tweak.
+	const kSearchStrength = 6
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+	cv := load6432(src, s)
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+
+	addLiterals := func(s *seq, until int32) {
+		if until == nextEmit {
+			return
+		}
+		blk.literals = append(blk.literals, src[nextEmit:until]...)
+		s.litLen = uint32(until - nextEmit)
+	}
+	if debugEncoder {
+		println("recent offsets:", blk.recentOffsets)
+	}
+
+encodeLoop:
+	for {
+		// t will contain the match offset when we find one.
+		// When existing the search loop, we have already checked 4 bytes.
+		var t int32
+
+		// We will not use repeat offsets across blocks.
+		// By not using them for the first 3 matches
+
+		for {
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
+			candidate := e.table[nextHash]
+			candidate2 := e.table[nextHash2]
+			repIndex := s - offset1 + 2
+
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)}
+
+			if len(blk.sequences) > 2 && load3232(src, repIndex) == uint32(cv>>16) {
+				// Consider history as well.
+				var seq seq
+				length := 4 + e.matchlen(s+6, repIndex+4, src)
+
+				seq.matchLen = uint32(length - zstdMinMatch)
+
+				// We might be able to match backwards.
+				// Extend as long as we can.
+				start := s + 2
+				// We end the search early, so we don't risk 0 literals
+				// and have to do special offset treatment.
+				startLimit := nextEmit + 1
+
+				sMin := s - e.maxMatchOff
+				if sMin < 0 {
+					sMin = 0
+				}
+				for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] {
+					repIndex--
+					start--
+					seq.matchLen++
+				}
+				addLiterals(&seq, start)
+
+				// rep 0
+				seq.offset = 1
+				if debugSequences {
+					println("repeat sequence", seq, "next s:", s)
+				}
+				blk.sequences = append(blk.sequences, seq)
+				s += length + 2
+				nextEmit = s
+				if s >= sLimit {
+					if debugEncoder {
+						println("repeat ended", s, length)
+
+					}
+					break encodeLoop
+				}
+				cv = load6432(src, s)
+				continue
+			}
+			coffset0 := s - (candidate.offset - e.cur)
+			coffset1 := s - (candidate2.offset - e.cur) + 1
+			if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val {
+				// found a regular match
+				t = candidate.offset - e.cur
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugAsserts && t < 0 {
+					panic(fmt.Sprintf("t (%d) < 0, candidate.offset: %d, e.cur: %d, coffset0: %d, e.maxMatchOff: %d", t, candidate.offset, e.cur, coffset0, e.maxMatchOff))
+				}
+				break
+			}
+
+			if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val {
+				// found a regular match
+				t = candidate2.offset - e.cur
+				s++
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugAsserts && t < 0 {
+					panic("t<0")
+				}
+				break
+			}
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+		// A 4-byte match has been found. We'll later see if more than 4 bytes.
+		offset2 = offset1
+		offset1 = s - t
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && t < 0 {
+			panic(fmt.Sprintf("t (%d) < 0 ", t))
+		}
+		// Extend the 4-byte match as long as possible.
+		l := e.matchlen(s+4, t+4, src) + 4
+
+		// Extend backwards
+		tMin := s - e.maxMatchOff
+		if tMin < 0 {
+			tMin = 0
+		}
+		for t > tMin && s > nextEmit && src[t-1] == src[s-1] {
+			s--
+			t--
+			l++
+		}
+
+		// Write our sequence.
+		var seq seq
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		// Don't use repeat offsets
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+		if s >= sLimit {
+			break encodeLoop
+		}
+		cv = load6432(src, s)
+
+		// Check offset 2
+		if o2 := s - offset2; len(blk.sequences) > 2 && load3232(src, o2) == uint32(cv) {
+			// We have at least 4 byte match.
+			// No need to check backwards. We come straight from a match
+			l := 4 + e.matchlen(s+4, o2+4, src)
+
+			// Store this, since we have it.
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			seq.matchLen = uint32(l) - zstdMinMatch
+			seq.litLen = 0
+			// Since litlen is always 0, this is offset 1.
+			seq.offset = 1
+			s += l
+			nextEmit = s
+			if debugSequences {
+				println("sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Swap offset 1 and 2.
+			offset1, offset2 = offset2, offset1
+			if s >= sLimit {
+				break encodeLoop
+			}
+			// Prepare next loop.
+			cv = load6432(src, s)
+		}
+	}
+
+	if int(nextEmit) < len(src) {
+		blk.literals = append(blk.literals, src[nextEmit:]...)
+		blk.extraLits = len(src) - int(nextEmit)
+	}
+	if debugEncoder {
+		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
+	}
+	// We do not store history, so we must offset e.cur to avoid false matches for next user.
+	if e.cur < bufferReset {
+		e.cur += int32(len(src))
+	}
+}
+
+// Encode will encode the content, with a dictionary if initialized for it.
+func (e *fastEncoderDict) Encode(blk *blockEnc, src []byte) {
+	const (
+		inputMargin            = 8
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+	if e.allDirty || len(src) > 32<<10 {
+		e.fastEncoder.Encode(blk, src)
+		e.allDirty = true
+		return
+	}
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			e.cur = e.maxMatchOff
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v < minOff {
+				v = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+			}
+			e.table[i].offset = v
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+
+	s := e.addBlock(src)
+	blk.size = len(src)
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	// Override src
+	src = e.hist
+	sLimit := int32(len(src)) - inputMargin
+	// stepSize is the number of bytes to skip on every main loop iteration.
+	// It should be >= 2.
+	const stepSize = 2
+
+	// TEMPLATE
+	const hashLog = tableBits
+	// seems global, but would be nice to tweak.
+	const kSearchStrength = 7
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+	cv := load6432(src, s)
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+
+	addLiterals := func(s *seq, until int32) {
+		if until == nextEmit {
+			return
+		}
+		blk.literals = append(blk.literals, src[nextEmit:until]...)
+		s.litLen = uint32(until - nextEmit)
+	}
+	if debugEncoder {
+		println("recent offsets:", blk.recentOffsets)
+	}
+
+encodeLoop:
+	for {
+		// t will contain the match offset when we find one.
+		// When existing the search loop, we have already checked 4 bytes.
+		var t int32
+
+		// We will not use repeat offsets across blocks.
+		// By not using them for the first 3 matches
+		canRepeat := len(blk.sequences) > 2
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
+			candidate := e.table[nextHash]
+			candidate2 := e.table[nextHash2]
+			repIndex := s - offset1 + 2
+
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.markShardDirty(nextHash)
+			e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)}
+			e.markShardDirty(nextHash2)
+
+			if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) {
+				// Consider history as well.
+				var seq seq
+				var length int32
+				length = 4 + e.matchlen(s+6, repIndex+4, src)
+
+				seq.matchLen = uint32(length - zstdMinMatch)
+
+				// We might be able to match backwards.
+				// Extend as long as we can.
+				start := s + 2
+				// We end the search early, so we don't risk 0 literals
+				// and have to do special offset treatment.
+				startLimit := nextEmit + 1
+
+				sMin := s - e.maxMatchOff
+				if sMin < 0 {
+					sMin = 0
+				}
+				for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch {
+					repIndex--
+					start--
+					seq.matchLen++
+				}
+				addLiterals(&seq, start)
+
+				// rep 0
+				seq.offset = 1
+				if debugSequences {
+					println("repeat sequence", seq, "next s:", s)
+				}
+				blk.sequences = append(blk.sequences, seq)
+				s += length + 2
+				nextEmit = s
+				if s >= sLimit {
+					if debugEncoder {
+						println("repeat ended", s, length)
+
+					}
+					break encodeLoop
+				}
+				cv = load6432(src, s)
+				continue
+			}
+			coffset0 := s - (candidate.offset - e.cur)
+			coffset1 := s - (candidate2.offset - e.cur) + 1
+			if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val {
+				// found a regular match
+				t = candidate.offset - e.cur
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				break
+			}
+
+			if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val {
+				// found a regular match
+				t = candidate2.offset - e.cur
+				s++
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugAsserts && t < 0 {
+					panic("t<0")
+				}
+				break
+			}
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+		// A 4-byte match has been found. We'll later see if more than 4 bytes.
+		offset2 = offset1
+		offset1 = s - t
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && canRepeat && int(offset1) > len(src) {
+			panic("invalid offset")
+		}
+
+		// Extend the 4-byte match as long as possible.
+		l := e.matchlen(s+4, t+4, src) + 4
+
+		// Extend backwards
+		tMin := s - e.maxMatchOff
+		if tMin < 0 {
+			tMin = 0
+		}
+		for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength {
+			s--
+			t--
+			l++
+		}
+
+		// Write our sequence.
+		var seq seq
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		// Don't use repeat offsets
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+		if s >= sLimit {
+			break encodeLoop
+		}
+		cv = load6432(src, s)
+
+		// Check offset 2
+		if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) {
+			// We have at least 4 byte match.
+			// No need to check backwards. We come straight from a match
+			l := 4 + e.matchlen(s+4, o2+4, src)
+
+			// Store this, since we have it.
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.markShardDirty(nextHash)
+			seq.matchLen = uint32(l) - zstdMinMatch
+			seq.litLen = 0
+			// Since litlen is always 0, this is offset 1.
+			seq.offset = 1
+			s += l
+			nextEmit = s
+			if debugSequences {
+				println("sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Swap offset 1 and 2.
+			offset1, offset2 = offset2, offset1
+			if s >= sLimit {
+				break encodeLoop
+			}
+			// Prepare next loop.
+			cv = load6432(src, s)
+		}
+	}
+
+	if int(nextEmit) < len(src) {
+		blk.literals = append(blk.literals, src[nextEmit:]...)
+		blk.extraLits = len(src) - int(nextEmit)
+	}
+	blk.recentOffsets[0] = uint32(offset1)
+	blk.recentOffsets[1] = uint32(offset2)
+	if debugEncoder {
+		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
+	}
+}
+
+// ResetDict will reset and set a dictionary if not nil
+func (e *fastEncoder) Reset(d *dict, singleBlock bool) {
+	e.resetBase(d, singleBlock)
+	if d != nil {
+		panic("fastEncoder: Reset with dict")
+	}
+}
+
+// ResetDict will reset and set a dictionary if not nil
+func (e *fastEncoderDict) Reset(d *dict, singleBlock bool) {
+	e.resetBase(d, singleBlock)
+	if d == nil {
+		return
+	}
+
+	// Init or copy dict table
+	if len(e.dictTable) != len(e.table) || d.id != e.lastDictID {
+		if len(e.dictTable) != len(e.table) {
+			e.dictTable = make([]tableEntry, len(e.table))
+		}
+		if true {
+			end := e.maxMatchOff + int32(len(d.content)) - 8
+			for i := e.maxMatchOff; i < end; i += 3 {
+				const hashLog = tableBits
+
+				cv := load6432(d.content, i-e.maxMatchOff)
+				nextHash := hashLen(cv, hashLog, tableFastHashLen)      // 0 -> 5
+				nextHash1 := hashLen(cv>>8, hashLog, tableFastHashLen)  // 1 -> 6
+				nextHash2 := hashLen(cv>>16, hashLog, tableFastHashLen) // 2 -> 7
+				e.dictTable[nextHash] = tableEntry{
+					val:    uint32(cv),
+					offset: i,
+				}
+				e.dictTable[nextHash1] = tableEntry{
+					val:    uint32(cv >> 8),
+					offset: i + 1,
+				}
+				e.dictTable[nextHash2] = tableEntry{
+					val:    uint32(cv >> 16),
+					offset: i + 2,
+				}
+			}
+		}
+		e.lastDictID = d.id
+		e.allDirty = true
+	}
+
+	e.cur = e.maxMatchOff
+	dirtyShardCnt := 0
+	if !e.allDirty {
+		for i := range e.tableShardDirty {
+			if e.tableShardDirty[i] {
+				dirtyShardCnt++
+			}
+		}
+	}
+
+	const shardCnt = tableShardCnt
+	const shardSize = tableShardSize
+	if e.allDirty || dirtyShardCnt > shardCnt*4/6 {
+		copy(e.table[:], e.dictTable)
+		for i := range e.tableShardDirty {
+			e.tableShardDirty[i] = false
+		}
+		e.allDirty = false
+		return
+	}
+	for i := range e.tableShardDirty {
+		if !e.tableShardDirty[i] {
+			continue
+		}
+
+		copy(e.table[i*shardSize:(i+1)*shardSize], e.dictTable[i*shardSize:(i+1)*shardSize])
+		e.tableShardDirty[i] = false
+	}
+	e.allDirty = false
+}
+
+func (e *fastEncoderDict) markAllShardsDirty() {
+	e.allDirty = true
+}
+
+func (e *fastEncoderDict) markShardDirty(entryNum uint32) {
+	e.tableShardDirty[entryNum/tableShardSize] = true
+}

vendor/github.com/klauspost/compress/zstd/encoder.go

@@ -0,0 +1,641 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"crypto/rand"
+	"fmt"
+	"io"
+	rdebug "runtime/debug"
+	"sync"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+// Encoder provides encoding to Zstandard.
+// An Encoder can be used for either compressing a stream via the
+// io.WriteCloser interface supported by the Encoder or as multiple independent
+// tasks via the EncodeAll function.
+// Smaller encodes are encouraged to use the EncodeAll function.
+// Use NewWriter to create a new instance.
+type Encoder struct {
+	o        encoderOptions
+	encoders chan encoder
+	state    encoderState
+	init     sync.Once
+}
+
+type encoder interface {
+	Encode(blk *blockEnc, src []byte)
+	EncodeNoHist(blk *blockEnc, src []byte)
+	Block() *blockEnc
+	CRC() *xxhash.Digest
+	AppendCRC([]byte) []byte
+	WindowSize(size int64) int32
+	UseBlock(*blockEnc)
+	Reset(d *dict, singleBlock bool)
+}
+
+type encoderState struct {
+	w                io.Writer
+	filling          []byte
+	current          []byte
+	previous         []byte
+	encoder          encoder
+	writing          *blockEnc
+	err              error
+	writeErr         error
+	nWritten         int64
+	nInput           int64
+	frameContentSize int64
+	headerWritten    bool
+	eofWritten       bool
+	fullFrameWritten bool
+
+	// This waitgroup indicates an encode is running.
+	wg sync.WaitGroup
+	// This waitgroup indicates we have a block encoding/writing.
+	wWg sync.WaitGroup
+}
+
+// NewWriter will create a new Zstandard encoder.
+// If the encoder will be used for encoding blocks a nil writer can be used.
+func NewWriter(w io.Writer, opts ...EOption) (*Encoder, error) {
+	initPredefined()
+	var e Encoder
+	e.o.setDefault()
+	for _, o := range opts {
+		err := o(&e.o)
+		if err != nil {
+			return nil, err
+		}
+	}
+	if w != nil {
+		e.Reset(w)
+	}
+	return &e, nil
+}
+
+func (e *Encoder) initialize() {
+	if e.o.concurrent == 0 {
+		e.o.setDefault()
+	}
+	e.encoders = make(chan encoder, e.o.concurrent)
+	for i := 0; i < e.o.concurrent; i++ {
+		enc := e.o.encoder()
+		e.encoders <- enc
+	}
+}
+
+// Reset will re-initialize the writer and new writes will encode to the supplied writer
+// as a new, independent stream.
+func (e *Encoder) Reset(w io.Writer) {
+	s := &e.state
+	s.wg.Wait()
+	s.wWg.Wait()
+	if cap(s.filling) == 0 {
+		s.filling = make([]byte, 0, e.o.blockSize)
+	}
+	if e.o.concurrent > 1 {
+		if cap(s.current) == 0 {
+			s.current = make([]byte, 0, e.o.blockSize)
+		}
+		if cap(s.previous) == 0 {
+			s.previous = make([]byte, 0, e.o.blockSize)
+		}
+		s.current = s.current[:0]
+		s.previous = s.previous[:0]
+		if s.writing == nil {
+			s.writing = &blockEnc{lowMem: e.o.lowMem}
+			s.writing.init()
+		}
+		s.writing.initNewEncode()
+	}
+	if s.encoder == nil {
+		s.encoder = e.o.encoder()
+	}
+	s.filling = s.filling[:0]
+	s.encoder.Reset(e.o.dict, false)
+	s.headerWritten = false
+	s.eofWritten = false
+	s.fullFrameWritten = false
+	s.w = w
+	s.err = nil
+	s.nWritten = 0
+	s.nInput = 0
+	s.writeErr = nil
+	s.frameContentSize = 0
+}
+
+// ResetContentSize will reset and set a content size for the next stream.
+// If the bytes written does not match the size given an error will be returned
+// when calling Close().
+// This is removed when Reset is called.
+// Sizes <= 0 results in no content size set.
+func (e *Encoder) ResetContentSize(w io.Writer, size int64) {
+	e.Reset(w)
+	if size >= 0 {
+		e.state.frameContentSize = size
+	}
+}
+
+// Write data to the encoder.
+// Input data will be buffered and as the buffer fills up
+// content will be compressed and written to the output.
+// When done writing, use Close to flush the remaining output
+// and write CRC if requested.
+func (e *Encoder) Write(p []byte) (n int, err error) {
+	s := &e.state
+	for len(p) > 0 {
+		if len(p)+len(s.filling) < e.o.blockSize {
+			if e.o.crc {
+				_, _ = s.encoder.CRC().Write(p)
+			}
+			s.filling = append(s.filling, p...)
+			return n + len(p), nil
+		}
+		add := p
+		if len(p)+len(s.filling) > e.o.blockSize {
+			add = add[:e.o.blockSize-len(s.filling)]
+		}
+		if e.o.crc {
+			_, _ = s.encoder.CRC().Write(add)
+		}
+		s.filling = append(s.filling, add...)
+		p = p[len(add):]
+		n += len(add)
+		if len(s.filling) < e.o.blockSize {
+			return n, nil
+		}
+		err := e.nextBlock(false)
+		if err != nil {
+			return n, err
+		}
+		if debugAsserts && len(s.filling) > 0 {
+			panic(len(s.filling))
+		}
+	}
+	return n, nil
+}
+
+// nextBlock will synchronize and start compressing input in e.state.filling.
+// If an error has occurred during encoding it will be returned.
+func (e *Encoder) nextBlock(final bool) error {
+	s := &e.state
+	// Wait for current block.
+	s.wg.Wait()
+	if s.err != nil {
+		return s.err
+	}
+	if len(s.filling) > e.o.blockSize {
+		return fmt.Errorf("block > maxStoreBlockSize")
+	}
+	if !s.headerWritten {
+		// If we have a single block encode, do a sync compression.
+		if final && len(s.filling) == 0 && !e.o.fullZero {
+			s.headerWritten = true
+			s.fullFrameWritten = true
+			s.eofWritten = true
+			return nil
+		}
+		if final && len(s.filling) > 0 {
+			s.current = e.EncodeAll(s.filling, s.current[:0])
+			var n2 int
+			n2, s.err = s.w.Write(s.current)
+			if s.err != nil {
+				return s.err
+			}
+			s.nWritten += int64(n2)
+			s.nInput += int64(len(s.filling))
+			s.current = s.current[:0]
+			s.filling = s.filling[:0]
+			s.headerWritten = true
+			s.fullFrameWritten = true
+			s.eofWritten = true
+			return nil
+		}
+
+		var tmp [maxHeaderSize]byte
+		fh := frameHeader{
+			ContentSize:   uint64(s.frameContentSize),
+			WindowSize:    uint32(s.encoder.WindowSize(s.frameContentSize)),
+			SingleSegment: false,
+			Checksum:      e.o.crc,
+			DictID:        e.o.dict.ID(),
+		}
+
+		dst, err := fh.appendTo(tmp[:0])
+		if err != nil {
+			return err
+		}
+		s.headerWritten = true
+		s.wWg.Wait()
+		var n2 int
+		n2, s.err = s.w.Write(dst)
+		if s.err != nil {
+			return s.err
+		}
+		s.nWritten += int64(n2)
+	}
+	if s.eofWritten {
+		// Ensure we only write it once.
+		final = false
+	}
+
+	if len(s.filling) == 0 {
+		// Final block, but no data.
+		if final {
+			enc := s.encoder
+			blk := enc.Block()
+			blk.reset(nil)
+			blk.last = true
+			blk.encodeRaw(nil)
+			s.wWg.Wait()
+			_, s.err = s.w.Write(blk.output)
+			s.nWritten += int64(len(blk.output))
+			s.eofWritten = true
+		}
+		return s.err
+	}
+
+	// SYNC:
+	if e.o.concurrent == 1 {
+		src := s.filling
+		s.nInput += int64(len(s.filling))
+		if debugEncoder {
+			println("Adding sync block,", len(src), "bytes, final:", final)
+		}
+		enc := s.encoder
+		blk := enc.Block()
+		blk.reset(nil)
+		enc.Encode(blk, src)
+		blk.last = final
+		if final {
+			s.eofWritten = true
+		}
+
+		err := errIncompressible
+		// If we got the exact same number of literals as input,
+		// assume the literals cannot be compressed.
+		if len(src) != len(blk.literals) || len(src) != e.o.blockSize {
+			err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
+		}
+		switch err {
+		case errIncompressible:
+			if debugEncoder {
+				println("Storing incompressible block as raw")
+			}
+			blk.encodeRaw(src)
+			// In fast mode, we do not transfer offsets, so we don't have to deal with changing the.
+		case nil:
+		default:
+			s.err = err
+			return err
+		}
+		_, s.err = s.w.Write(blk.output)
+		s.nWritten += int64(len(blk.output))
+		s.filling = s.filling[:0]
+		return s.err
+	}
+
+	// Move blocks forward.
+	s.filling, s.current, s.previous = s.previous[:0], s.filling, s.current
+	s.nInput += int64(len(s.current))
+	s.wg.Add(1)
+	go func(src []byte) {
+		if debugEncoder {
+			println("Adding block,", len(src), "bytes, final:", final)
+		}
+		defer func() {
+			if r := recover(); r != nil {
+				s.err = fmt.Errorf("panic while encoding: %v", r)
+				rdebug.PrintStack()
+			}
+			s.wg.Done()
+		}()
+		enc := s.encoder
+		blk := enc.Block()
+		enc.Encode(blk, src)
+		blk.last = final
+		if final {
+			s.eofWritten = true
+		}
+		// Wait for pending writes.
+		s.wWg.Wait()
+		if s.writeErr != nil {
+			s.err = s.writeErr
+			return
+		}
+		// Transfer encoders from previous write block.
+		blk.swapEncoders(s.writing)
+		// Transfer recent offsets to next.
+		enc.UseBlock(s.writing)
+		s.writing = blk
+		s.wWg.Add(1)
+		go func() {
+			defer func() {
+				if r := recover(); r != nil {
+					s.writeErr = fmt.Errorf("panic while encoding/writing: %v", r)
+					rdebug.PrintStack()
+				}
+				s.wWg.Done()
+			}()
+			err := errIncompressible
+			// If we got the exact same number of literals as input,
+			// assume the literals cannot be compressed.
+			if len(src) != len(blk.literals) || len(src) != e.o.blockSize {
+				err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
+			}
+			switch err {
+			case errIncompressible:
+				if debugEncoder {
+					println("Storing incompressible block as raw")
+				}
+				blk.encodeRaw(src)
+				// In fast mode, we do not transfer offsets, so we don't have to deal with changing the.
+			case nil:
+			default:
+				s.writeErr = err
+				return
+			}
+			_, s.writeErr = s.w.Write(blk.output)
+			s.nWritten += int64(len(blk.output))
+		}()
+	}(s.current)
+	return nil
+}
+
+// ReadFrom reads data from r until EOF or error.
+// The return value n is the number of bytes read.
+// Any error except io.EOF encountered during the read is also returned.
+//
+// The Copy function uses ReaderFrom if available.
+func (e *Encoder) ReadFrom(r io.Reader) (n int64, err error) {
+	if debugEncoder {
+		println("Using ReadFrom")
+	}
+
+	// Flush any current writes.
+	if len(e.state.filling) > 0 {
+		if err := e.nextBlock(false); err != nil {
+			return 0, err
+		}
+	}
+	e.state.filling = e.state.filling[:e.o.blockSize]
+	src := e.state.filling
+	for {
+		n2, err := r.Read(src)
+		if e.o.crc {
+			_, _ = e.state.encoder.CRC().Write(src[:n2])
+		}
+		// src is now the unfilled part...
+		src = src[n2:]
+		n += int64(n2)
+		switch err {
+		case io.EOF:
+			e.state.filling = e.state.filling[:len(e.state.filling)-len(src)]
+			if debugEncoder {
+				println("ReadFrom: got EOF final block:", len(e.state.filling))
+			}
+			return n, nil
+		case nil:
+		default:
+			if debugEncoder {
+				println("ReadFrom: got error:", err)
+			}
+			e.state.err = err
+			return n, err
+		}
+		if len(src) > 0 {
+			if debugEncoder {
+				println("ReadFrom: got space left in source:", len(src))
+			}
+			continue
+		}
+		err = e.nextBlock(false)
+		if err != nil {
+			return n, err
+		}
+		e.state.filling = e.state.filling[:e.o.blockSize]
+		src = e.state.filling
+	}
+}
+
+// Flush will send the currently written data to output
+// and block until everything has been written.
+// This should only be used on rare occasions where pushing the currently queued data is critical.
+func (e *Encoder) Flush() error {
+	s := &e.state
+	if len(s.filling) > 0 {
+		err := e.nextBlock(false)
+		if err != nil {
+			return err
+		}
+	}
+	s.wg.Wait()
+	s.wWg.Wait()
+	if s.err != nil {
+		return s.err
+	}
+	return s.writeErr
+}
+
+// Close will flush the final output and close the stream.
+// The function will block until everything has been written.
+// The Encoder can still be re-used after calling this.
+func (e *Encoder) Close() error {
+	s := &e.state
+	if s.encoder == nil {
+		return nil
+	}
+	err := e.nextBlock(true)
+	if err != nil {
+		return err
+	}
+	if s.frameContentSize > 0 {
+		if s.nInput != s.frameContentSize {
+			return fmt.Errorf("frame content size %d given, but %d bytes was written", s.frameContentSize, s.nInput)
+		}
+	}
+	if e.state.fullFrameWritten {
+		return s.err
+	}
+	s.wg.Wait()
+	s.wWg.Wait()
+
+	if s.err != nil {
+		return s.err
+	}
+	if s.writeErr != nil {
+		return s.writeErr
+	}
+
+	// Write CRC
+	if e.o.crc && s.err == nil {
+		// heap alloc.
+		var tmp [4]byte
+		_, s.err = s.w.Write(s.encoder.AppendCRC(tmp[:0]))
+		s.nWritten += 4
+	}
+
+	// Add padding with content from crypto/rand.Reader
+	if s.err == nil && e.o.pad > 0 {
+		add := calcSkippableFrame(s.nWritten, int64(e.o.pad))
+		frame, err := skippableFrame(s.filling[:0], add, rand.Reader)
+		if err != nil {
+			return err
+		}
+		_, s.err = s.w.Write(frame)
+	}
+	return s.err
+}
+
+// EncodeAll will encode all input in src and append it to dst.
+// This function can be called concurrently, but each call will only run on a single goroutine.
+// If empty input is given, nothing is returned, unless WithZeroFrames is specified.
+// Encoded blocks can be concatenated and the result will be the combined input stream.
+// Data compressed with EncodeAll can be decoded with the Decoder,
+// using either a stream or DecodeAll.
+func (e *Encoder) EncodeAll(src, dst []byte) []byte {
+	if len(src) == 0 {
+		if e.o.fullZero {
+			// Add frame header.
+			fh := frameHeader{
+				ContentSize:   0,
+				WindowSize:    MinWindowSize,
+				SingleSegment: true,
+				// Adding a checksum would be a waste of space.
+				Checksum: false,
+				DictID:   0,
+			}
+			dst, _ = fh.appendTo(dst)
+
+			// Write raw block as last one only.
+			var blk blockHeader
+			blk.setSize(0)
+			blk.setType(blockTypeRaw)
+			blk.setLast(true)
+			dst = blk.appendTo(dst)
+		}
+		return dst
+	}
+	e.init.Do(e.initialize)
+	enc := <-e.encoders
+	defer func() {
+		// Release encoder reference to last block.
+		// If a non-single block is needed the encoder will reset again.
+		e.encoders <- enc
+	}()
+	// Use single segments when above minimum window and below 1MB.
+	single := len(src) < 1<<20 && len(src) > MinWindowSize
+	if e.o.single != nil {
+		single = *e.o.single
+	}
+	fh := frameHeader{
+		ContentSize:   uint64(len(src)),
+		WindowSize:    uint32(enc.WindowSize(int64(len(src)))),
+		SingleSegment: single,
+		Checksum:      e.o.crc,
+		DictID:        e.o.dict.ID(),
+	}
+
+	// If less than 1MB, allocate a buffer up front.
+	if len(dst) == 0 && cap(dst) == 0 && len(src) < 1<<20 && !e.o.lowMem {
+		dst = make([]byte, 0, len(src))
+	}
+	dst, err := fh.appendTo(dst)
+	if err != nil {
+		panic(err)
+	}
+
+	// If we can do everything in one block, prefer that.
+	if len(src) <= maxCompressedBlockSize {
+		enc.Reset(e.o.dict, true)
+		// Slightly faster with no history and everything in one block.
+		if e.o.crc {
+			_, _ = enc.CRC().Write(src)
+		}
+		blk := enc.Block()
+		blk.last = true
+		if e.o.dict == nil {
+			enc.EncodeNoHist(blk, src)
+		} else {
+			enc.Encode(blk, src)
+		}
+
+		// If we got the exact same number of literals as input,
+		// assume the literals cannot be compressed.
+		err := errIncompressible
+		oldout := blk.output
+		if len(blk.literals) != len(src) || len(src) != e.o.blockSize {
+			// Output directly to dst
+			blk.output = dst
+			err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
+		}
+
+		switch err {
+		case errIncompressible:
+			if debugEncoder {
+				println("Storing incompressible block as raw")
+			}
+			dst = blk.encodeRawTo(dst, src)
+		case nil:
+			dst = blk.output
+		default:
+			panic(err)
+		}
+		blk.output = oldout
+	} else {
+		enc.Reset(e.o.dict, false)
+		blk := enc.Block()
+		for len(src) > 0 {
+			todo := src
+			if len(todo) > e.o.blockSize {
+				todo = todo[:e.o.blockSize]
+			}
+			src = src[len(todo):]
+			if e.o.crc {
+				_, _ = enc.CRC().Write(todo)
+			}
+			blk.pushOffsets()
+			enc.Encode(blk, todo)
+			if len(src) == 0 {
+				blk.last = true
+			}
+			err := errIncompressible
+			// If we got the exact same number of literals as input,
+			// assume the literals cannot be compressed.
+			if len(blk.literals) != len(todo) || len(todo) != e.o.blockSize {
+				err = blk.encode(todo, e.o.noEntropy, !e.o.allLitEntropy)
+			}
+
+			switch err {
+			case errIncompressible:
+				if debugEncoder {
+					println("Storing incompressible block as raw")
+				}
+				dst = blk.encodeRawTo(dst, todo)
+				blk.popOffsets()
+			case nil:
+				dst = append(dst, blk.output...)
+			default:
+				panic(err)
+			}
+			blk.reset(nil)
+		}
+	}
+	if e.o.crc {
+		dst = enc.AppendCRC(dst)
+	}
+	// Add padding with content from crypto/rand.Reader
+	if e.o.pad > 0 {
+		add := calcSkippableFrame(int64(len(dst)), int64(e.o.pad))
+		dst, err = skippableFrame(dst, add, rand.Reader)
+		if err != nil {
+			panic(err)
+		}
+	}
+	return dst
+}

vendor/github.com/klauspost/compress/zstd/encoder_options.go

@@ -0,0 +1,317 @@
+package zstd
+
+import (
+	"errors"
+	"fmt"
+	"runtime"
+	"strings"
+)
+
+// EOption is an option for creating a encoder.
+type EOption func(*encoderOptions) error
+
+// options retains accumulated state of multiple options.
+type encoderOptions struct {
+	concurrent      int
+	level           EncoderLevel
+	single          *bool
+	pad             int
+	blockSize       int
+	windowSize      int
+	crc             bool
+	fullZero        bool
+	noEntropy       bool
+	allLitEntropy   bool
+	customWindow    bool
+	customALEntropy bool
+	customBlockSize bool
+	lowMem          bool
+	dict            *dict
+}
+
+func (o *encoderOptions) setDefault() {
+	*o = encoderOptions{
+		concurrent:    runtime.GOMAXPROCS(0),
+		crc:           true,
+		single:        nil,
+		blockSize:     maxCompressedBlockSize,
+		windowSize:    8 << 20,
+		level:         SpeedDefault,
+		allLitEntropy: true,
+		lowMem:        false,
+	}
+}
+
+// encoder returns an encoder with the selected options.
+func (o encoderOptions) encoder() encoder {
+	switch o.level {
+	case SpeedFastest:
+		if o.dict != nil {
+			return &fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}}
+		}
+		return &fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}
+
+	case SpeedDefault:
+		if o.dict != nil {
+			return &doubleFastEncoderDict{fastEncoderDict: fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}}}
+		}
+		return &doubleFastEncoder{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}}
+	case SpeedBetterCompression:
+		if o.dict != nil {
+			return &betterFastEncoderDict{betterFastEncoder: betterFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}}
+		}
+		return &betterFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}
+	case SpeedBestCompression:
+		return &bestFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}
+	}
+	panic("unknown compression level")
+}
+
+// WithEncoderCRC will add CRC value to output.
+// Output will be 4 bytes larger.
+func WithEncoderCRC(b bool) EOption {
+	return func(o *encoderOptions) error { o.crc = b; return nil }
+}
+
+// WithEncoderConcurrency will set the concurrency,
+// meaning the maximum number of encoders to run concurrently.
+// The value supplied must be at least 1.
+// For streams, setting a value of 1 will disable async compression.
+// By default this will be set to GOMAXPROCS.
+func WithEncoderConcurrency(n int) EOption {
+	return func(o *encoderOptions) error {
+		if n <= 0 {
+			return fmt.Errorf("concurrency must be at least 1")
+		}
+		o.concurrent = n
+		return nil
+	}
+}
+
+// WithWindowSize will set the maximum allowed back-reference distance.
+// The value must be a power of two between MinWindowSize and MaxWindowSize.
+// A larger value will enable better compression but allocate more memory and,
+// for above-default values, take considerably longer.
+// The default value is determined by the compression level.
+func WithWindowSize(n int) EOption {
+	return func(o *encoderOptions) error {
+		switch {
+		case n < MinWindowSize:
+			return fmt.Errorf("window size must be at least %d", MinWindowSize)
+		case n > MaxWindowSize:
+			return fmt.Errorf("window size must be at most %d", MaxWindowSize)
+		case (n & (n - 1)) != 0:
+			return errors.New("window size must be a power of 2")
+		}
+
+		o.windowSize = n
+		o.customWindow = true
+		if o.blockSize > o.windowSize {
+			o.blockSize = o.windowSize
+			o.customBlockSize = true
+		}
+		return nil
+	}
+}
+
+// WithEncoderPadding will add padding to all output so the size will be a multiple of n.
+// This can be used to obfuscate the exact output size or make blocks of a certain size.
+// The contents will be a skippable frame, so it will be invisible by the decoder.
+// n must be > 0 and <= 1GB, 1<<30 bytes.
+// The padded area will be filled with data from crypto/rand.Reader.
+// If `EncodeAll` is used with data already in the destination, the total size will be multiple of this.
+func WithEncoderPadding(n int) EOption {
+	return func(o *encoderOptions) error {
+		if n <= 0 {
+			return fmt.Errorf("padding must be at least 1")
+		}
+		// No need to waste our time.
+		if n == 1 {
+			o.pad = 0
+		}
+		if n > 1<<30 {
+			return fmt.Errorf("padding must less than 1GB (1<<30 bytes) ")
+		}
+		o.pad = n
+		return nil
+	}
+}
+
+// EncoderLevel predefines encoder compression levels.
+// Only use the constants made available, since the actual mapping
+// of these values are very likely to change and your compression could change
+// unpredictably when upgrading the library.
+type EncoderLevel int
+
+const (
+	speedNotSet EncoderLevel = iota
+
+	// SpeedFastest will choose the fastest reasonable compression.
+	// This is roughly equivalent to the fastest Zstandard mode.
+	SpeedFastest
+
+	// SpeedDefault is the default "pretty fast" compression option.
+	// This is roughly equivalent to the default Zstandard mode (level 3).
+	SpeedDefault
+
+	// SpeedBetterCompression will yield better compression than the default.
+	// Currently it is about zstd level 7-8 with ~ 2x-3x the default CPU usage.
+	// By using this, notice that CPU usage may go up in the future.
+	SpeedBetterCompression
+
+	// SpeedBestCompression will choose the best available compression option.
+	// This will offer the best compression no matter the CPU cost.
+	SpeedBestCompression
+
+	// speedLast should be kept as the last actual compression option.
+	// The is not for external usage, but is used to keep track of the valid options.
+	speedLast
+)
+
+// EncoderLevelFromString will convert a string representation of an encoding level back
+// to a compression level. The compare is not case sensitive.
+// If the string wasn't recognized, (false, SpeedDefault) will be returned.
+func EncoderLevelFromString(s string) (bool, EncoderLevel) {
+	for l := speedNotSet + 1; l < speedLast; l++ {
+		if strings.EqualFold(s, l.String()) {
+			return true, l
+		}
+	}
+	return false, SpeedDefault
+}
+
+// EncoderLevelFromZstd will return an encoder level that closest matches the compression
+// ratio of a specific zstd compression level.
+// Many input values will provide the same compression level.
+func EncoderLevelFromZstd(level int) EncoderLevel {
+	switch {
+	case level < 3:
+		return SpeedFastest
+	case level >= 3 && level < 6:
+		return SpeedDefault
+	case level >= 6 && level < 10:
+		return SpeedBetterCompression
+	default:
+		return SpeedBestCompression
+	}
+}
+
+// String provides a string representation of the compression level.
+func (e EncoderLevel) String() string {
+	switch e {
+	case SpeedFastest:
+		return "fastest"
+	case SpeedDefault:
+		return "default"
+	case SpeedBetterCompression:
+		return "better"
+	case SpeedBestCompression:
+		return "best"
+	default:
+		return "invalid"
+	}
+}
+
+// WithEncoderLevel specifies a predefined compression level.
+func WithEncoderLevel(l EncoderLevel) EOption {
+	return func(o *encoderOptions) error {
+		switch {
+		case l <= speedNotSet || l >= speedLast:
+			return fmt.Errorf("unknown encoder level")
+		}
+		o.level = l
+		if !o.customWindow {
+			switch o.level {
+			case SpeedFastest:
+				o.windowSize = 4 << 20
+				if !o.customBlockSize {
+					o.blockSize = 1 << 16
+				}
+			case SpeedDefault:
+				o.windowSize = 8 << 20
+			case SpeedBetterCompression:
+				o.windowSize = 16 << 20
+			case SpeedBestCompression:
+				o.windowSize = 32 << 20
+			}
+		}
+		if !o.customALEntropy {
+			o.allLitEntropy = l > SpeedFastest
+		}
+
+		return nil
+	}
+}
+
+// WithZeroFrames will encode 0 length input as full frames.
+// This can be needed for compatibility with zstandard usage,
+// but is not needed for this package.
+func WithZeroFrames(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.fullZero = b
+		return nil
+	}
+}
+
+// WithAllLitEntropyCompression will apply entropy compression if no matches are found.
+// Disabling this will skip incompressible data faster, but in cases with no matches but
+// skewed character distribution compression is lost.
+// Default value depends on the compression level selected.
+func WithAllLitEntropyCompression(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.customALEntropy = true
+		o.allLitEntropy = b
+		return nil
+	}
+}
+
+// WithNoEntropyCompression will always skip entropy compression of literals.
+// This can be useful if content has matches, but unlikely to benefit from entropy
+// compression. Usually the slight speed improvement is not worth enabling this.
+func WithNoEntropyCompression(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.noEntropy = b
+		return nil
+	}
+}
+
+// WithSingleSegment will set the "single segment" flag when EncodeAll is used.
+// If this flag is set, data must be regenerated within a single continuous memory segment.
+// In this case, Window_Descriptor byte is skipped, but Frame_Content_Size is necessarily present.
+// As a consequence, the decoder must allocate a memory segment of size equal or larger than size of your content.
+// In order to preserve the decoder from unreasonable memory requirements,
+// a decoder is allowed to reject a compressed frame which requests a memory size beyond decoder's authorized range.
+// For broader compatibility, decoders are recommended to support memory sizes of at least 8 MB.
+// This is only a recommendation, each decoder is free to support higher or lower limits, depending on local limitations.
+// If this is not specified, block encodes will automatically choose this based on the input size.
+// This setting has no effect on streamed encodes.
+func WithSingleSegment(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.single = &b
+		return nil
+	}
+}
+
+// WithLowerEncoderMem will trade in some memory cases trade less memory usage for
+// slower encoding speed.
+// This will not change the window size which is the primary function for reducing
+// memory usage. See WithWindowSize.
+func WithLowerEncoderMem(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.lowMem = b
+		return nil
+	}
+}
+
+// WithEncoderDict allows to register a dictionary that will be used for the encode.
+// The encoder *may* choose to use no dictionary instead for certain payloads.
+func WithEncoderDict(dict []byte) EOption {
+	return func(o *encoderOptions) error {
+		d, err := loadDict(dict)
+		if err != nil {
+			return err
+		}
+		o.dict = d
+		return nil
+	}
+}

vendor/github.com/klauspost/compress/zstd/framedec.go

@@ -0,0 +1,389 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"bytes"
+	"encoding/hex"
+	"errors"
+	"io"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+type frameDec struct {
+	o   decoderOptions
+	crc *xxhash.Digest
+
+	WindowSize uint64
+
+	// Frame history passed between blocks
+	history history
+
+	rawInput byteBuffer
+
+	// Byte buffer that can be reused for small input blocks.
+	bBuf byteBuf
+
+	FrameContentSize uint64
+
+	DictionaryID  *uint32
+	HasCheckSum   bool
+	SingleSegment bool
+}
+
+const (
+	// MinWindowSize is the minimum Window Size, which is 1 KB.
+	MinWindowSize = 1 << 10
+
+	// MaxWindowSize is the maximum encoder window size
+	// and the default decoder maximum window size.
+	MaxWindowSize = 1 << 29
+)
+
+var (
+	frameMagic          = []byte{0x28, 0xb5, 0x2f, 0xfd}
+	skippableFrameMagic = []byte{0x2a, 0x4d, 0x18}
+)
+
+func newFrameDec(o decoderOptions) *frameDec {
+	if o.maxWindowSize > o.maxDecodedSize {
+		o.maxWindowSize = o.maxDecodedSize
+	}
+	d := frameDec{
+		o: o,
+	}
+	return &d
+}
+
+// reset will read the frame header and prepare for block decoding.
+// If nothing can be read from the input, io.EOF will be returned.
+// Any other error indicated that the stream contained data, but
+// there was a problem.
+func (d *frameDec) reset(br byteBuffer) error {
+	d.HasCheckSum = false
+	d.WindowSize = 0
+	var signature [4]byte
+	for {
+		var err error
+		// Check if we can read more...
+		b, err := br.readSmall(1)
+		switch err {
+		case io.EOF, io.ErrUnexpectedEOF:
+			return io.EOF
+		default:
+			return err
+		case nil:
+			signature[0] = b[0]
+		}
+		// Read the rest, don't allow io.ErrUnexpectedEOF
+		b, err = br.readSmall(3)
+		switch err {
+		case io.EOF:
+			return io.EOF
+		default:
+			return err
+		case nil:
+			copy(signature[1:], b)
+		}
+
+		if !bytes.Equal(signature[1:4], skippableFrameMagic) || signature[0]&0xf0 != 0x50 {
+			if debugDecoder {
+				println("Not skippable", hex.EncodeToString(signature[:]), hex.EncodeToString(skippableFrameMagic))
+			}
+			// Break if not skippable frame.
+			break
+		}
+		// Read size to skip
+		b, err = br.readSmall(4)
+		if err != nil {
+			if debugDecoder {
+				println("Reading Frame Size", err)
+			}
+			return err
+		}
+		n := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		println("Skipping frame with", n, "bytes.")
+		err = br.skipN(int(n))
+		if err != nil {
+			if debugDecoder {
+				println("Reading discarded frame", err)
+			}
+			return err
+		}
+	}
+	if !bytes.Equal(signature[:], frameMagic) {
+		if debugDecoder {
+			println("Got magic numbers: ", signature, "want:", frameMagic)
+		}
+		return ErrMagicMismatch
+	}
+
+	// Read Frame_Header_Descriptor
+	fhd, err := br.readByte()
+	if err != nil {
+		if debugDecoder {
+			println("Reading Frame_Header_Descriptor", err)
+		}
+		return err
+	}
+	d.SingleSegment = fhd&(1<<5) != 0
+
+	if fhd&(1<<3) != 0 {
+		return errors.New("reserved bit set on frame header")
+	}
+
+	// Read Window_Descriptor
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#window_descriptor
+	d.WindowSize = 0
+	if !d.SingleSegment {
+		wd, err := br.readByte()
+		if err != nil {
+			if debugDecoder {
+				println("Reading Window_Descriptor", err)
+			}
+			return err
+		}
+		printf("raw: %x, mantissa: %d, exponent: %d\n", wd, wd&7, wd>>3)
+		windowLog := 10 + (wd >> 3)
+		windowBase := uint64(1) << windowLog
+		windowAdd := (windowBase / 8) * uint64(wd&0x7)
+		d.WindowSize = windowBase + windowAdd
+	}
+
+	// Read Dictionary_ID
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id
+	d.DictionaryID = nil
+	if size := fhd & 3; size != 0 {
+		if size == 3 {
+			size = 4
+		}
+
+		b, err := br.readSmall(int(size))
+		if err != nil {
+			println("Reading Dictionary_ID", err)
+			return err
+		}
+		var id uint32
+		switch size {
+		case 1:
+			id = uint32(b[0])
+		case 2:
+			id = uint32(b[0]) | (uint32(b[1]) << 8)
+		case 4:
+			id = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		}
+		if debugDecoder {
+			println("Dict size", size, "ID:", id)
+		}
+		if id > 0 {
+			// ID 0 means "sorry, no dictionary anyway".
+			// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary-format
+			d.DictionaryID = &id
+		}
+	}
+
+	// Read Frame_Content_Size
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frame_content_size
+	var fcsSize int
+	v := fhd >> 6
+	switch v {
+	case 0:
+		if d.SingleSegment {
+			fcsSize = 1
+		}
+	default:
+		fcsSize = 1 << v
+	}
+	d.FrameContentSize = fcsUnknown
+	if fcsSize > 0 {
+		b, err := br.readSmall(fcsSize)
+		if err != nil {
+			println("Reading Frame content", err)
+			return err
+		}
+		switch fcsSize {
+		case 1:
+			d.FrameContentSize = uint64(b[0])
+		case 2:
+			// When FCS_Field_Size is 2, the offset of 256 is added.
+			d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) + 256
+		case 4:
+			d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) | (uint64(b[2]) << 16) | (uint64(b[3]) << 24)
+		case 8:
+			d1 := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+			d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24)
+			d.FrameContentSize = uint64(d1) | (uint64(d2) << 32)
+		}
+		if debugDecoder {
+			println("Read FCS:", d.FrameContentSize)
+		}
+	}
+
+	// Move this to shared.
+	d.HasCheckSum = fhd&(1<<2) != 0
+	if d.HasCheckSum {
+		if d.crc == nil {
+			d.crc = xxhash.New()
+		}
+		d.crc.Reset()
+	}
+
+	if d.WindowSize == 0 && d.SingleSegment {
+		// We may not need window in this case.
+		d.WindowSize = d.FrameContentSize
+		if d.WindowSize < MinWindowSize {
+			d.WindowSize = MinWindowSize
+		}
+	}
+
+	if d.WindowSize > uint64(d.o.maxWindowSize) {
+		if debugDecoder {
+			printf("window size %d > max %d\n", d.WindowSize, d.o.maxWindowSize)
+		}
+		return ErrWindowSizeExceeded
+	}
+	// The minimum Window_Size is 1 KB.
+	if d.WindowSize < MinWindowSize {
+		if debugDecoder {
+			println("got window size: ", d.WindowSize)
+		}
+		return ErrWindowSizeTooSmall
+	}
+	d.history.windowSize = int(d.WindowSize)
+	if d.o.lowMem && d.history.windowSize < maxBlockSize {
+		d.history.allocFrameBuffer = d.history.windowSize * 2
+		// TODO: Maybe use FrameContent size
+	} else {
+		d.history.allocFrameBuffer = d.history.windowSize + maxBlockSize
+	}
+
+	if debugDecoder {
+		println("Frame: Dict:", d.DictionaryID, "FrameContentSize:", d.FrameContentSize, "singleseg:", d.SingleSegment, "window:", d.WindowSize, "crc:", d.HasCheckSum)
+	}
+
+	// history contains input - maybe we do something
+	d.rawInput = br
+	return nil
+}
+
+// next will start decoding the next block from stream.
+func (d *frameDec) next(block *blockDec) error {
+	if debugDecoder {
+		println("decoding new block")
+	}
+	err := block.reset(d.rawInput, d.WindowSize)
+	if err != nil {
+		println("block error:", err)
+		// Signal the frame decoder we have a problem.
+		block.sendErr(err)
+		return err
+	}
+	return nil
+}
+
+// checkCRC will check the checksum if the frame has one.
+// Will return ErrCRCMismatch if crc check failed, otherwise nil.
+func (d *frameDec) checkCRC() error {
+	if !d.HasCheckSum {
+		return nil
+	}
+	var tmp [4]byte
+	got := d.crc.Sum64()
+	// Flip to match file order.
+	tmp[0] = byte(got >> 0)
+	tmp[1] = byte(got >> 8)
+	tmp[2] = byte(got >> 16)
+	tmp[3] = byte(got >> 24)
+
+	// We can overwrite upper tmp now
+	want, err := d.rawInput.readSmall(4)
+	if err != nil {
+		println("CRC missing?", err)
+		return err
+	}
+
+	if !bytes.Equal(tmp[:], want) && !ignoreCRC {
+		if debugDecoder {
+			println("CRC Check Failed:", tmp[:], "!=", want)
+		}
+		return ErrCRCMismatch
+	}
+	if debugDecoder {
+		println("CRC ok", tmp[:])
+	}
+	return nil
+}
+
+// runDecoder will create a sync decoder that will decode a block of data.
+func (d *frameDec) runDecoder(dst []byte, dec *blockDec) ([]byte, error) {
+	saved := d.history.b
+
+	// We use the history for output to avoid copying it.
+	d.history.b = dst
+	d.history.ignoreBuffer = len(dst)
+	// Store input length, so we only check new data.
+	crcStart := len(dst)
+	d.history.decoders.maxSyncLen = 0
+	if d.FrameContentSize != fcsUnknown {
+		d.history.decoders.maxSyncLen = d.FrameContentSize + uint64(len(dst))
+		if d.history.decoders.maxSyncLen > d.o.maxDecodedSize {
+			return dst, ErrDecoderSizeExceeded
+		}
+		if uint64(cap(dst)) < d.history.decoders.maxSyncLen {
+			// Alloc for output
+			dst2 := make([]byte, len(dst), d.history.decoders.maxSyncLen+compressedBlockOverAlloc)
+			copy(dst2, dst)
+			dst = dst2
+		}
+	}
+	var err error
+	for {
+		err = dec.reset(d.rawInput, d.WindowSize)
+		if err != nil {
+			break
+		}
+		if debugDecoder {
+			println("next block:", dec)
+		}
+		err = dec.decodeBuf(&d.history)
+		if err != nil {
+			break
+		}
+		if uint64(len(d.history.b)) > d.o.maxDecodedSize {
+			err = ErrDecoderSizeExceeded
+			break
+		}
+		if uint64(len(d.history.b)-crcStart) > d.FrameContentSize {
+			println("runDecoder: FrameContentSize exceeded", uint64(len(d.history.b)-crcStart), ">", d.FrameContentSize)
+			err = ErrFrameSizeExceeded
+			break
+		}
+		if dec.Last {
+			break
+		}
+		if debugDecoder {
+			println("runDecoder: FrameContentSize", uint64(len(d.history.b)-crcStart), "<=", d.FrameContentSize)
+		}
+	}
+	dst = d.history.b
+	if err == nil {
+		if d.FrameContentSize != fcsUnknown && uint64(len(d.history.b)-crcStart) != d.FrameContentSize {
+			err = ErrFrameSizeMismatch
+		} else if d.HasCheckSum {
+			var n int
+			n, err = d.crc.Write(dst[crcStart:])
+			if err == nil {
+				if n != len(dst)-crcStart {
+					err = io.ErrShortWrite
+				} else {
+					err = d.checkCRC()
+				}
+			}
+		}
+	}
+	d.history.b = saved
+	return dst, err
+}

vendor/github.com/klauspost/compress/zstd/frameenc.go

@@ -0,0 +1,137 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"math"
+	"math/bits"
+)
+
+type frameHeader struct {
+	ContentSize   uint64
+	WindowSize    uint32
+	SingleSegment bool
+	Checksum      bool
+	DictID        uint32
+}
+
+const maxHeaderSize = 14
+
+func (f frameHeader) appendTo(dst []byte) ([]byte, error) {
+	dst = append(dst, frameMagic...)
+	var fhd uint8
+	if f.Checksum {
+		fhd |= 1 << 2
+	}
+	if f.SingleSegment {
+		fhd |= 1 << 5
+	}
+
+	var dictIDContent []byte
+	if f.DictID > 0 {
+		var tmp [4]byte
+		if f.DictID < 256 {
+			fhd |= 1
+			tmp[0] = uint8(f.DictID)
+			dictIDContent = tmp[:1]
+		} else if f.DictID < 1<<16 {
+			fhd |= 2
+			binary.LittleEndian.PutUint16(tmp[:2], uint16(f.DictID))
+			dictIDContent = tmp[:2]
+		} else {
+			fhd |= 3
+			binary.LittleEndian.PutUint32(tmp[:4], f.DictID)
+			dictIDContent = tmp[:4]
+		}
+	}
+	var fcs uint8
+	if f.ContentSize >= 256 {
+		fcs++
+	}
+	if f.ContentSize >= 65536+256 {
+		fcs++
+	}
+	if f.ContentSize >= 0xffffffff {
+		fcs++
+	}
+
+	fhd |= fcs << 6
+
+	dst = append(dst, fhd)
+	if !f.SingleSegment {
+		const winLogMin = 10
+		windowLog := (bits.Len32(f.WindowSize-1) - winLogMin) << 3
+		dst = append(dst, uint8(windowLog))
+	}
+	if f.DictID > 0 {
+		dst = append(dst, dictIDContent...)
+	}
+	switch fcs {
+	case 0:
+		if f.SingleSegment {
+			dst = append(dst, uint8(f.ContentSize))
+		}
+		// Unless SingleSegment is set, framessizes < 256 are nto stored.
+	case 1:
+		f.ContentSize -= 256
+		dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8))
+	case 2:
+		dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8), uint8(f.ContentSize>>16), uint8(f.ContentSize>>24))
+	case 3:
+		dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8), uint8(f.ContentSize>>16), uint8(f.ContentSize>>24),
+			uint8(f.ContentSize>>32), uint8(f.ContentSize>>40), uint8(f.ContentSize>>48), uint8(f.ContentSize>>56))
+	default:
+		panic("invalid fcs")
+	}
+	return dst, nil
+}
+
+const skippableFrameHeader = 4 + 4
+
+// calcSkippableFrame will return a total size to be added for written
+// to be divisible by multiple.
+// The value will always be > skippableFrameHeader.
+// The function will panic if written < 0 or wantMultiple <= 0.
+func calcSkippableFrame(written, wantMultiple int64) int {
+	if wantMultiple <= 0 {
+		panic("wantMultiple <= 0")
+	}
+	if written < 0 {
+		panic("written < 0")
+	}
+	leftOver := written % wantMultiple
+	if leftOver == 0 {
+		return 0
+	}
+	toAdd := wantMultiple - leftOver
+	for toAdd < skippableFrameHeader {
+		toAdd += wantMultiple
+	}
+	return int(toAdd)
+}
+
+// skippableFrame will add a skippable frame with a total size of bytes.
+// total should be >= skippableFrameHeader and < math.MaxUint32.
+func skippableFrame(dst []byte, total int, r io.Reader) ([]byte, error) {
+	if total == 0 {
+		return dst, nil
+	}
+	if total < skippableFrameHeader {
+		return dst, fmt.Errorf("requested skippable frame (%d) < 8", total)
+	}
+	if int64(total) > math.MaxUint32 {
+		return dst, fmt.Errorf("requested skippable frame (%d) > max uint32", total)
+	}
+	dst = append(dst, 0x50, 0x2a, 0x4d, 0x18)
+	f := uint32(total - skippableFrameHeader)
+	dst = append(dst, uint8(f), uint8(f>>8), uint8(f>>16), uint8(f>>24))
+	start := len(dst)
+	dst = append(dst, make([]byte, f)...)
+	_, err := io.ReadFull(r, dst[start:])
+	return dst, err
+}

vendor/github.com/klauspost/compress/zstd/fse_decoder.go

@@ -0,0 +1,410 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+const (
+	tablelogAbsoluteMax = 9
+)
+
+const (
+	/*!MEMORY_USAGE :
+	 *  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+	 *  Increasing memory usage improves compression ratio
+	 *  Reduced memory usage can improve speed, due to cache effect
+	 *  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+	maxMemoryUsage = tablelogAbsoluteMax + 2
+
+	maxTableLog    = maxMemoryUsage - 2
+	maxTablesize   = 1 << maxTableLog
+	maxTableMask   = (1 << maxTableLog) - 1
+	minTablelog    = 5
+	maxSymbolValue = 255
+)
+
+// fseDecoder provides temporary storage for compression and decompression.
+type fseDecoder struct {
+	dt             [maxTablesize]decSymbol // Decompression table.
+	symbolLen      uint16                  // Length of active part of the symbol table.
+	actualTableLog uint8                   // Selected tablelog.
+	maxBits        uint8                   // Maximum number of additional bits
+
+	// used for table creation to avoid allocations.
+	stateTable [256]uint16
+	norm       [maxSymbolValue + 1]int16
+	preDefined bool
+}
+
+// tableStep returns the next table index.
+func tableStep(tableSize uint32) uint32 {
+	return (tableSize >> 1) + (tableSize >> 3) + 3
+}
+
+// readNCount will read the symbol distribution so decoding tables can be constructed.
+func (s *fseDecoder) readNCount(b *byteReader, maxSymbol uint16) error {
+	var (
+		charnum   uint16
+		previous0 bool
+	)
+	if b.remain() < 4 {
+		return errors.New("input too small")
+	}
+	bitStream := b.Uint32NC()
+	nbBits := uint((bitStream & 0xF) + minTablelog) // extract tableLog
+	if nbBits > tablelogAbsoluteMax {
+		println("Invalid tablelog:", nbBits)
+		return errors.New("tableLog too large")
+	}
+	bitStream >>= 4
+	bitCount := uint(4)
+
+	s.actualTableLog = uint8(nbBits)
+	remaining := int32((1 << nbBits) + 1)
+	threshold := int32(1 << nbBits)
+	gotTotal := int32(0)
+	nbBits++
+
+	for remaining > 1 && charnum <= maxSymbol {
+		if previous0 {
+			//println("prev0")
+			n0 := charnum
+			for (bitStream & 0xFFFF) == 0xFFFF {
+				//println("24 x 0")
+				n0 += 24
+				if r := b.remain(); r > 5 {
+					b.advance(2)
+					// The check above should make sure we can read 32 bits
+					bitStream = b.Uint32NC() >> bitCount
+				} else {
+					// end of bit stream
+					bitStream >>= 16
+					bitCount += 16
+				}
+			}
+			//printf("bitstream: %d, 0b%b", bitStream&3, bitStream)
+			for (bitStream & 3) == 3 {
+				n0 += 3
+				bitStream >>= 2
+				bitCount += 2
+			}
+			n0 += uint16(bitStream & 3)
+			bitCount += 2
+
+			if n0 > maxSymbolValue {
+				return errors.New("maxSymbolValue too small")
+			}
+			//println("inserting ", n0-charnum, "zeroes from idx", charnum, "ending before", n0)
+			for charnum < n0 {
+				s.norm[uint8(charnum)] = 0
+				charnum++
+			}
+
+			if r := b.remain(); r >= 7 || r-int(bitCount>>3) >= 4 {
+				b.advance(bitCount >> 3)
+				bitCount &= 7
+				// The check above should make sure we can read 32 bits
+				bitStream = b.Uint32NC() >> bitCount
+			} else {
+				bitStream >>= 2
+			}
+		}
+
+		max := (2*threshold - 1) - remaining
+		var count int32
+
+		if int32(bitStream)&(threshold-1) < max {
+			count = int32(bitStream) & (threshold - 1)
+			if debugAsserts && nbBits < 1 {
+				panic("nbBits underflow")
+			}
+			bitCount += nbBits - 1
+		} else {
+			count = int32(bitStream) & (2*threshold - 1)
+			if count >= threshold {
+				count -= max
+			}
+			bitCount += nbBits
+		}
+
+		// extra accuracy
+		count--
+		if count < 0 {
+			// -1 means +1
+			remaining += count
+			gotTotal -= count
+		} else {
+			remaining -= count
+			gotTotal += count
+		}
+		s.norm[charnum&0xff] = int16(count)
+		charnum++
+		previous0 = count == 0
+		for remaining < threshold {
+			nbBits--
+			threshold >>= 1
+		}
+
+		if r := b.remain(); r >= 7 || r-int(bitCount>>3) >= 4 {
+			b.advance(bitCount >> 3)
+			bitCount &= 7
+			// The check above should make sure we can read 32 bits
+			bitStream = b.Uint32NC() >> (bitCount & 31)
+		} else {
+			bitCount -= (uint)(8 * (len(b.b) - 4 - b.off))
+			b.off = len(b.b) - 4
+			bitStream = b.Uint32() >> (bitCount & 31)
+		}
+	}
+	s.symbolLen = charnum
+	if s.symbolLen <= 1 {
+		return fmt.Errorf("symbolLen (%d) too small", s.symbolLen)
+	}
+	if s.symbolLen > maxSymbolValue+1 {
+		return fmt.Errorf("symbolLen (%d) too big", s.symbolLen)
+	}
+	if remaining != 1 {
+		return fmt.Errorf("corruption detected (remaining %d != 1)", remaining)
+	}
+	if bitCount > 32 {
+		return fmt.Errorf("corruption detected (bitCount %d > 32)", bitCount)
+	}
+	if gotTotal != 1<<s.actualTableLog {
+		return fmt.Errorf("corruption detected (total %d != %d)", gotTotal, 1<<s.actualTableLog)
+	}
+	b.advance((bitCount + 7) >> 3)
+	// println(s.norm[:s.symbolLen], s.symbolLen)
+	return s.buildDtable()
+}
+
+func (s *fseDecoder) mustReadFrom(r io.Reader) {
+	fatalErr := func(err error) {
+		if err != nil {
+			panic(err)
+		}
+	}
+	// 	dt             [maxTablesize]decSymbol // Decompression table.
+	//	symbolLen      uint16                  // Length of active part of the symbol table.
+	//	actualTableLog uint8                   // Selected tablelog.
+	//	maxBits        uint8                   // Maximum number of additional bits
+	//	// used for table creation to avoid allocations.
+	//	stateTable [256]uint16
+	//	norm       [maxSymbolValue + 1]int16
+	//	preDefined bool
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.dt))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.symbolLen))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.actualTableLog))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.maxBits))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.stateTable))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.norm))
+	fatalErr(binary.Read(r, binary.LittleEndian, &s.preDefined))
+}
+
+// decSymbol contains information about a state entry,
+// Including the state offset base, the output symbol and
+// the number of bits to read for the low part of the destination state.
+// Using a composite uint64 is faster than a struct with separate members.
+type decSymbol uint64
+
+func newDecSymbol(nbits, addBits uint8, newState uint16, baseline uint32) decSymbol {
+	return decSymbol(nbits) | (decSymbol(addBits) << 8) | (decSymbol(newState) << 16) | (decSymbol(baseline) << 32)
+}
+
+func (d decSymbol) nbBits() uint8 {
+	return uint8(d)
+}
+
+func (d decSymbol) addBits() uint8 {
+	return uint8(d >> 8)
+}
+
+func (d decSymbol) newState() uint16 {
+	return uint16(d >> 16)
+}
+
+func (d decSymbol) baseline() uint32 {
+	return uint32(d >> 32)
+}
+
+func (d decSymbol) baselineInt() int {
+	return int(d >> 32)
+}
+
+func (d *decSymbol) set(nbits, addBits uint8, newState uint16, baseline uint32) {
+	*d = decSymbol(nbits) | (decSymbol(addBits) << 8) | (decSymbol(newState) << 16) | (decSymbol(baseline) << 32)
+}
+
+func (d *decSymbol) setNBits(nBits uint8) {
+	const mask = 0xffffffffffffff00
+	*d = (*d & mask) | decSymbol(nBits)
+}
+
+func (d *decSymbol) setAddBits(addBits uint8) {
+	const mask = 0xffffffffffff00ff
+	*d = (*d & mask) | (decSymbol(addBits) << 8)
+}
+
+func (d *decSymbol) setNewState(state uint16) {
+	const mask = 0xffffffff0000ffff
+	*d = (*d & mask) | decSymbol(state)<<16
+}
+
+func (d *decSymbol) setBaseline(baseline uint32) {
+	const mask = 0xffffffff
+	*d = (*d & mask) | decSymbol(baseline)<<32
+}
+
+func (d *decSymbol) setExt(addBits uint8, baseline uint32) {
+	const mask = 0xffff00ff
+	*d = (*d & mask) | (decSymbol(addBits) << 8) | (decSymbol(baseline) << 32)
+}
+
+// decSymbolValue returns the transformed decSymbol for the given symbol.
+func decSymbolValue(symb uint8, t []baseOffset) (decSymbol, error) {
+	if int(symb) >= len(t) {
+		return 0, fmt.Errorf("rle symbol %d >= max %d", symb, len(t))
+	}
+	lu := t[symb]
+	return newDecSymbol(0, lu.addBits, 0, lu.baseLine), nil
+}
+
+// setRLE will set the decoder til RLE mode.
+func (s *fseDecoder) setRLE(symbol decSymbol) {
+	s.actualTableLog = 0
+	s.maxBits = symbol.addBits()
+	s.dt[0] = symbol
+}
+
+// buildDtable will build the decoding table.
+func (s *fseDecoder) buildDtable() error {
+	tableSize := uint32(1 << s.actualTableLog)
+	highThreshold := tableSize - 1
+	symbolNext := s.stateTable[:256]
+
+	// Init, lay down lowprob symbols
+	{
+		for i, v := range s.norm[:s.symbolLen] {
+			if v == -1 {
+				s.dt[highThreshold].setAddBits(uint8(i))
+				highThreshold--
+				symbolNext[i] = 1
+			} else {
+				symbolNext[i] = uint16(v)
+			}
+		}
+	}
+	// Spread symbols
+	{
+		tableMask := tableSize - 1
+		step := tableStep(tableSize)
+		position := uint32(0)
+		for ss, v := range s.norm[:s.symbolLen] {
+			for i := 0; i < int(v); i++ {
+				s.dt[position].setAddBits(uint8(ss))
+				position = (position + step) & tableMask
+				for position > highThreshold {
+					// lowprob area
+					position = (position + step) & tableMask
+				}
+			}
+		}
+		if position != 0 {
+			// position must reach all cells once, otherwise normalizedCounter is incorrect
+			return errors.New("corrupted input (position != 0)")
+		}
+	}
+
+	// Build Decoding table
+	{
+		tableSize := uint16(1 << s.actualTableLog)
+		for u, v := range s.dt[:tableSize] {
+			symbol := v.addBits()
+			nextState := symbolNext[symbol]
+			symbolNext[symbol] = nextState + 1
+			nBits := s.actualTableLog - byte(highBits(uint32(nextState)))
+			s.dt[u&maxTableMask].setNBits(nBits)
+			newState := (nextState << nBits) - tableSize
+			if newState > tableSize {
+				return fmt.Errorf("newState (%d) outside table size (%d)", newState, tableSize)
+			}
+			if newState == uint16(u) && nBits == 0 {
+				// Seems weird that this is possible with nbits > 0.
+				return fmt.Errorf("newState (%d) == oldState (%d) and no bits", newState, u)
+			}
+			s.dt[u&maxTableMask].setNewState(newState)
+		}
+	}
+	return nil
+}
+
+// transform will transform the decoder table into a table usable for
+// decoding without having to apply the transformation while decoding.
+// The state will contain the base value and the number of bits to read.
+func (s *fseDecoder) transform(t []baseOffset) error {
+	tableSize := uint16(1 << s.actualTableLog)
+	s.maxBits = 0
+	for i, v := range s.dt[:tableSize] {
+		add := v.addBits()
+		if int(add) >= len(t) {
+			return fmt.Errorf("invalid decoding table entry %d, symbol %d >= max (%d)", i, v.addBits(), len(t))
+		}
+		lu := t[add]
+		if lu.addBits > s.maxBits {
+			s.maxBits = lu.addBits
+		}
+		v.setExt(lu.addBits, lu.baseLine)
+		s.dt[i] = v
+	}
+	return nil
+}
+
+type fseState struct {
+	dt    []decSymbol
+	state decSymbol
+}
+
+// Initialize and decodeAsync first state and symbol.
+func (s *fseState) init(br *bitReader, tableLog uint8, dt []decSymbol) {
+	s.dt = dt
+	br.fill()
+	s.state = dt[br.getBits(tableLog)]
+}
+
+// next returns the current symbol and sets the next state.
+// At least tablelog bits must be available in the bit reader.
+func (s *fseState) next(br *bitReader) {
+	lowBits := uint16(br.getBits(s.state.nbBits()))
+	s.state = s.dt[s.state.newState()+lowBits]
+}
+
+// finished returns true if all bits have been read from the bitstream
+// and the next state would require reading bits from the input.
+func (s *fseState) finished(br *bitReader) bool {
+	return br.finished() && s.state.nbBits() > 0
+}
+
+// final returns the current state symbol without decoding the next.
+func (s *fseState) final() (int, uint8) {
+	return s.state.baselineInt(), s.state.addBits()
+}
+
+// final returns the current state symbol without decoding the next.
+func (s decSymbol) final() (int, uint8) {
+	return s.baselineInt(), s.addBits()
+}
+
+// nextFast returns the next symbol and sets the next state.
+// This can only be used if no symbols are 0 bits.
+// At least tablelog bits must be available in the bit reader.
+func (s *fseState) nextFast(br *bitReader) (uint32, uint8) {
+	lowBits := br.get16BitsFast(s.state.nbBits())
+	s.state = s.dt[s.state.newState()+lowBits]
+	return s.state.baseline(), s.state.addBits()
+}

vendor/github.com/klauspost/compress/zstd/fse_encoder.go

@@ -0,0 +1,724 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"errors"
+	"fmt"
+	"math"
+)
+
+const (
+	// For encoding we only support up to
+	maxEncTableLog    = 8
+	maxEncTablesize   = 1 << maxTableLog
+	maxEncTableMask   = (1 << maxTableLog) - 1
+	minEncTablelog    = 5
+	maxEncSymbolValue = maxMatchLengthSymbol
+)
+
+// Scratch provides temporary storage for compression and decompression.
+type fseEncoder struct {
+	symbolLen      uint16 // Length of active part of the symbol table.
+	actualTableLog uint8  // Selected tablelog.
+	ct             cTable // Compression tables.
+	maxCount       int    // count of the most probable symbol
+	zeroBits       bool   // no bits has prob > 50%.
+	clearCount     bool   // clear count
+	useRLE         bool   // This encoder is for RLE
+	preDefined     bool   // This encoder is predefined.
+	reUsed         bool   // Set to know when the encoder has been reused.
+	rleVal         uint8  // RLE Symbol
+	maxBits        uint8  // Maximum output bits after transform.
+
+	// TODO: Technically zstd should be fine with 64 bytes.
+	count [256]uint32
+	norm  [256]int16
+}
+
+// cTable contains tables used for compression.
+type cTable struct {
+	tableSymbol []byte
+	stateTable  []uint16
+	symbolTT    []symbolTransform
+}
+
+// symbolTransform contains the state transform for a symbol.
+type symbolTransform struct {
+	deltaNbBits    uint32
+	deltaFindState int16
+	outBits        uint8
+}
+
+// String prints values as a human readable string.
+func (s symbolTransform) String() string {
+	return fmt.Sprintf("{deltabits: %08x, findstate:%d outbits:%d}", s.deltaNbBits, s.deltaFindState, s.outBits)
+}
+
+// Histogram allows to populate the histogram and skip that step in the compression,
+// It otherwise allows to inspect the histogram when compression is done.
+// To indicate that you have populated the histogram call HistogramFinished
+// with the value of the highest populated symbol, as well as the number of entries
+// in the most populated entry. These are accepted at face value.
+func (s *fseEncoder) Histogram() *[256]uint32 {
+	return &s.count
+}
+
+// HistogramFinished can be called to indicate that the histogram has been populated.
+// maxSymbol is the index of the highest set symbol of the next data segment.
+// maxCount is the number of entries in the most populated entry.
+// These are accepted at face value.
+func (s *fseEncoder) HistogramFinished(maxSymbol uint8, maxCount int) {
+	s.maxCount = maxCount
+	s.symbolLen = uint16(maxSymbol) + 1
+	s.clearCount = maxCount != 0
+}
+
+// prepare will prepare and allocate scratch tables used for both compression and decompression.
+func (s *fseEncoder) prepare() (*fseEncoder, error) {
+	if s == nil {
+		s = &fseEncoder{}
+	}
+	s.useRLE = false
+	if s.clearCount && s.maxCount == 0 {
+		for i := range s.count {
+			s.count[i] = 0
+		}
+		s.clearCount = false
+	}
+	return s, nil
+}
+
+// allocCtable will allocate tables needed for compression.
+// If existing tables a re big enough, they are simply re-used.
+func (s *fseEncoder) allocCtable() {
+	tableSize := 1 << s.actualTableLog
+	// get tableSymbol that is big enough.
+	if cap(s.ct.tableSymbol) < tableSize {
+		s.ct.tableSymbol = make([]byte, tableSize)
+	}
+	s.ct.tableSymbol = s.ct.tableSymbol[:tableSize]
+
+	ctSize := tableSize
+	if cap(s.ct.stateTable) < ctSize {
+		s.ct.stateTable = make([]uint16, ctSize)
+	}
+	s.ct.stateTable = s.ct.stateTable[:ctSize]
+
+	if cap(s.ct.symbolTT) < 256 {
+		s.ct.symbolTT = make([]symbolTransform, 256)
+	}
+	s.ct.symbolTT = s.ct.symbolTT[:256]
+}
+
+// buildCTable will populate the compression table so it is ready to be used.
+func (s *fseEncoder) buildCTable() error {
+	tableSize := uint32(1 << s.actualTableLog)
+	highThreshold := tableSize - 1
+	var cumul [256]int16
+
+	s.allocCtable()
+	tableSymbol := s.ct.tableSymbol[:tableSize]
+	// symbol start positions
+	{
+		cumul[0] = 0
+		for ui, v := range s.norm[:s.symbolLen-1] {
+			u := byte(ui) // one less than reference
+			if v == -1 {
+				// Low proba symbol
+				cumul[u+1] = cumul[u] + 1
+				tableSymbol[highThreshold] = u
+				highThreshold--
+			} else {
+				cumul[u+1] = cumul[u] + v
+			}
+		}
+		// Encode last symbol separately to avoid overflowing u
+		u := int(s.symbolLen - 1)
+		v := s.norm[s.symbolLen-1]
+		if v == -1 {
+			// Low proba symbol
+			cumul[u+1] = cumul[u] + 1
+			tableSymbol[highThreshold] = byte(u)
+			highThreshold--
+		} else {
+			cumul[u+1] = cumul[u] + v
+		}
+		if uint32(cumul[s.symbolLen]) != tableSize {
+			return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize)
+		}
+		cumul[s.symbolLen] = int16(tableSize) + 1
+	}
+	// Spread symbols
+	s.zeroBits = false
+	{
+		step := tableStep(tableSize)
+		tableMask := tableSize - 1
+		var position uint32
+		// if any symbol > largeLimit, we may have 0 bits output.
+		largeLimit := int16(1 << (s.actualTableLog - 1))
+		for ui, v := range s.norm[:s.symbolLen] {
+			symbol := byte(ui)
+			if v > largeLimit {
+				s.zeroBits = true
+			}
+			for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ {
+				tableSymbol[position] = symbol
+				position = (position + step) & tableMask
+				for position > highThreshold {
+					position = (position + step) & tableMask
+				} /* Low proba area */
+			}
+		}
+
+		// Check if we have gone through all positions
+		if position != 0 {
+			return errors.New("position!=0")
+		}
+	}
+
+	// Build table
+	table := s.ct.stateTable
+	{
+		tsi := int(tableSize)
+		for u, v := range tableSymbol {
+			// TableU16 : sorted by symbol order; gives next state value
+			table[cumul[v]] = uint16(tsi + u)
+			cumul[v]++
+		}
+	}
+
+	// Build Symbol Transformation Table
+	{
+		total := int16(0)
+		symbolTT := s.ct.symbolTT[:s.symbolLen]
+		tableLog := s.actualTableLog
+		tl := (uint32(tableLog) << 16) - (1 << tableLog)
+		for i, v := range s.norm[:s.symbolLen] {
+			switch v {
+			case 0:
+			case -1, 1:
+				symbolTT[i].deltaNbBits = tl
+				symbolTT[i].deltaFindState = total - 1
+				total++
+			default:
+				maxBitsOut := uint32(tableLog) - highBit(uint32(v-1))
+				minStatePlus := uint32(v) << maxBitsOut
+				symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus
+				symbolTT[i].deltaFindState = total - v
+				total += v
+			}
+		}
+		if total != int16(tableSize) {
+			return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize)
+		}
+	}
+	return nil
+}
+
+var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000}
+
+func (s *fseEncoder) setRLE(val byte) {
+	s.allocCtable()
+	s.actualTableLog = 0
+	s.ct.stateTable = s.ct.stateTable[:1]
+	s.ct.symbolTT[val] = symbolTransform{
+		deltaFindState: 0,
+		deltaNbBits:    0,
+	}
+	if debugEncoder {
+		println("setRLE: val", val, "symbolTT", s.ct.symbolTT[val])
+	}
+	s.rleVal = val
+	s.useRLE = true
+}
+
+// setBits will set output bits for the transform.
+// if nil is provided, the number of bits is equal to the index.
+func (s *fseEncoder) setBits(transform []byte) {
+	if s.reUsed || s.preDefined {
+		return
+	}
+	if s.useRLE {
+		if transform == nil {
+			s.ct.symbolTT[s.rleVal].outBits = s.rleVal
+			s.maxBits = s.rleVal
+			return
+		}
+		s.maxBits = transform[s.rleVal]
+		s.ct.symbolTT[s.rleVal].outBits = s.maxBits
+		return
+	}
+	if transform == nil {
+		for i := range s.ct.symbolTT[:s.symbolLen] {
+			s.ct.symbolTT[i].outBits = uint8(i)
+		}
+		s.maxBits = uint8(s.symbolLen - 1)
+		return
+	}
+	s.maxBits = 0
+	for i, v := range transform[:s.symbolLen] {
+		s.ct.symbolTT[i].outBits = v
+		if v > s.maxBits {
+			// We could assume bits always going up, but we play safe.
+			s.maxBits = v
+		}
+	}
+}
+
+// normalizeCount will normalize the count of the symbols so
+// the total is equal to the table size.
+// If successful, compression tables will also be made ready.
+func (s *fseEncoder) normalizeCount(length int) error {
+	if s.reUsed {
+		return nil
+	}
+	s.optimalTableLog(length)
+	var (
+		tableLog          = s.actualTableLog
+		scale             = 62 - uint64(tableLog)
+		step              = (1 << 62) / uint64(length)
+		vStep             = uint64(1) << (scale - 20)
+		stillToDistribute = int16(1 << tableLog)
+		largest           int
+		largestP          int16
+		lowThreshold      = (uint32)(length >> tableLog)
+	)
+	if s.maxCount == length {
+		s.useRLE = true
+		return nil
+	}
+	s.useRLE = false
+	for i, cnt := range s.count[:s.symbolLen] {
+		// already handled
+		// if (count[s] == s.length) return 0;   /* rle special case */
+
+		if cnt == 0 {
+			s.norm[i] = 0
+			continue
+		}
+		if cnt <= lowThreshold {
+			s.norm[i] = -1
+			stillToDistribute--
+		} else {
+			proba := (int16)((uint64(cnt) * step) >> scale)
+			if proba < 8 {
+				restToBeat := vStep * uint64(rtbTable[proba])
+				v := uint64(cnt)*step - (uint64(proba) << scale)
+				if v > restToBeat {
+					proba++
+				}
+			}
+			if proba > largestP {
+				largestP = proba
+				largest = i
+			}
+			s.norm[i] = proba
+			stillToDistribute -= proba
+		}
+	}
+
+	if -stillToDistribute >= (s.norm[largest] >> 1) {
+		// corner case, need another normalization method
+		err := s.normalizeCount2(length)
+		if err != nil {
+			return err
+		}
+		if debugAsserts {
+			err = s.validateNorm()
+			if err != nil {
+				return err
+			}
+		}
+		return s.buildCTable()
+	}
+	s.norm[largest] += stillToDistribute
+	if debugAsserts {
+		err := s.validateNorm()
+		if err != nil {
+			return err
+		}
+	}
+	return s.buildCTable()
+}
+
+// Secondary normalization method.
+// To be used when primary method fails.
+func (s *fseEncoder) normalizeCount2(length int) error {
+	const notYetAssigned = -2
+	var (
+		distributed  uint32
+		total        = uint32(length)
+		tableLog     = s.actualTableLog
+		lowThreshold = total >> tableLog
+		lowOne       = (total * 3) >> (tableLog + 1)
+	)
+	for i, cnt := range s.count[:s.symbolLen] {
+		if cnt == 0 {
+			s.norm[i] = 0
+			continue
+		}
+		if cnt <= lowThreshold {
+			s.norm[i] = -1
+			distributed++
+			total -= cnt
+			continue
+		}
+		if cnt <= lowOne {
+			s.norm[i] = 1
+			distributed++
+			total -= cnt
+			continue
+		}
+		s.norm[i] = notYetAssigned
+	}
+	toDistribute := (1 << tableLog) - distributed
+
+	if (total / toDistribute) > lowOne {
+		// risk of rounding to zero
+		lowOne = (total * 3) / (toDistribute * 2)
+		for i, cnt := range s.count[:s.symbolLen] {
+			if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) {
+				s.norm[i] = 1
+				distributed++
+				total -= cnt
+				continue
+			}
+		}
+		toDistribute = (1 << tableLog) - distributed
+	}
+	if distributed == uint32(s.symbolLen)+1 {
+		// all values are pretty poor;
+		//   probably incompressible data (should have already been detected);
+		//   find max, then give all remaining points to max
+		var maxV int
+		var maxC uint32
+		for i, cnt := range s.count[:s.symbolLen] {
+			if cnt > maxC {
+				maxV = i
+				maxC = cnt
+			}
+		}
+		s.norm[maxV] += int16(toDistribute)
+		return nil
+	}
+
+	if total == 0 {
+		// all of the symbols were low enough for the lowOne or lowThreshold
+		for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) {
+			if s.norm[i] > 0 {
+				toDistribute--
+				s.norm[i]++
+			}
+		}
+		return nil
+	}
+
+	var (
+		vStepLog = 62 - uint64(tableLog)
+		mid      = uint64((1 << (vStepLog - 1)) - 1)
+		rStep    = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining
+		tmpTotal = mid
+	)
+	for i, cnt := range s.count[:s.symbolLen] {
+		if s.norm[i] == notYetAssigned {
+			var (
+				end    = tmpTotal + uint64(cnt)*rStep
+				sStart = uint32(tmpTotal >> vStepLog)
+				sEnd   = uint32(end >> vStepLog)
+				weight = sEnd - sStart
+			)
+			if weight < 1 {
+				return errors.New("weight < 1")
+			}
+			s.norm[i] = int16(weight)
+			tmpTotal = end
+		}
+	}
+	return nil
+}
+
+// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog
+func (s *fseEncoder) optimalTableLog(length int) {
+	tableLog := uint8(maxEncTableLog)
+	minBitsSrc := highBit(uint32(length)) + 1
+	minBitsSymbols := highBit(uint32(s.symbolLen-1)) + 2
+	minBits := uint8(minBitsSymbols)
+	if minBitsSrc < minBitsSymbols {
+		minBits = uint8(minBitsSrc)
+	}
+
+	maxBitsSrc := uint8(highBit(uint32(length-1))) - 2
+	if maxBitsSrc < tableLog {
+		// Accuracy can be reduced
+		tableLog = maxBitsSrc
+	}
+	if minBits > tableLog {
+		tableLog = minBits
+	}
+	// Need a minimum to safely represent all symbol values
+	if tableLog < minEncTablelog {
+		tableLog = minEncTablelog
+	}
+	if tableLog > maxEncTableLog {
+		tableLog = maxEncTableLog
+	}
+	s.actualTableLog = tableLog
+}
+
+// validateNorm validates the normalized histogram table.
+func (s *fseEncoder) validateNorm() (err error) {
+	var total int
+	for _, v := range s.norm[:s.symbolLen] {
+		if v >= 0 {
+			total += int(v)
+		} else {
+			total -= int(v)
+		}
+	}
+	defer func() {
+		if err == nil {
+			return
+		}
+		fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen)
+		for i, v := range s.norm[:s.symbolLen] {
+			fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v)
+		}
+	}()
+	if total != (1 << s.actualTableLog) {
+		return fmt.Errorf("warning: Total == %d != %d", total, 1<<s.actualTableLog)
+	}
+	for i, v := range s.count[s.symbolLen:] {
+		if v != 0 {
+			return fmt.Errorf("warning: Found symbol out of range, %d after cut", i)
+		}
+	}
+	return nil
+}
+
+// writeCount will write the normalized histogram count to header.
+// This is read back by readNCount.
+func (s *fseEncoder) writeCount(out []byte) ([]byte, error) {
+	if s.useRLE {
+		return append(out, s.rleVal), nil
+	}
+	if s.preDefined || s.reUsed {
+		// Never write predefined.
+		return out, nil
+	}
+
+	var (
+		tableLog  = s.actualTableLog
+		tableSize = 1 << tableLog
+		previous0 bool
+		charnum   uint16
+
+		// maximum header size plus 2 extra bytes for final output if bitCount == 0.
+		maxHeaderSize = ((int(s.symbolLen) * int(tableLog)) >> 3) + 3 + 2
+
+		// Write Table Size
+		bitStream = uint32(tableLog - minEncTablelog)
+		bitCount  = uint(4)
+		remaining = int16(tableSize + 1) /* +1 for extra accuracy */
+		threshold = int16(tableSize)
+		nbBits    = uint(tableLog + 1)
+		outP      = len(out)
+	)
+	if cap(out) < outP+maxHeaderSize {
+		out = append(out, make([]byte, maxHeaderSize*3)...)
+		out = out[:len(out)-maxHeaderSize*3]
+	}
+	out = out[:outP+maxHeaderSize]
+
+	// stops at 1
+	for remaining > 1 {
+		if previous0 {
+			start := charnum
+			for s.norm[charnum] == 0 {
+				charnum++
+			}
+			for charnum >= start+24 {
+				start += 24
+				bitStream += uint32(0xFFFF) << bitCount
+				out[outP] = byte(bitStream)
+				out[outP+1] = byte(bitStream >> 8)
+				outP += 2
+				bitStream >>= 16
+			}
+			for charnum >= start+3 {
+				start += 3
+				bitStream += 3 << bitCount
+				bitCount += 2
+			}
+			bitStream += uint32(charnum-start) << bitCount
+			bitCount += 2
+			if bitCount > 16 {
+				out[outP] = byte(bitStream)
+				out[outP+1] = byte(bitStream >> 8)
+				outP += 2
+				bitStream >>= 16
+				bitCount -= 16
+			}
+		}
+
+		count := s.norm[charnum]
+		charnum++
+		max := (2*threshold - 1) - remaining
+		if count < 0 {
+			remaining += count
+		} else {
+			remaining -= count
+		}
+		count++ // +1 for extra accuracy
+		if count >= threshold {
+			count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[
+		}
+		bitStream += uint32(count) << bitCount
+		bitCount += nbBits
+		if count < max {
+			bitCount--
+		}
+
+		previous0 = count == 1
+		if remaining < 1 {
+			return nil, errors.New("internal error: remaining < 1")
+		}
+		for remaining < threshold {
+			nbBits--
+			threshold >>= 1
+		}
+
+		if bitCount > 16 {
+			out[outP] = byte(bitStream)
+			out[outP+1] = byte(bitStream >> 8)
+			outP += 2
+			bitStream >>= 16
+			bitCount -= 16
+		}
+	}
+
+	if outP+2 > len(out) {
+		return nil, fmt.Errorf("internal error: %d > %d, maxheader: %d, sl: %d, tl: %d, normcount: %v", outP+2, len(out), maxHeaderSize, s.symbolLen, int(tableLog), s.norm[:s.symbolLen])
+	}
+	out[outP] = byte(bitStream)
+	out[outP+1] = byte(bitStream >> 8)
+	outP += int((bitCount + 7) / 8)
+
+	if charnum > s.symbolLen {
+		return nil, errors.New("internal error: charnum > s.symbolLen")
+	}
+	return out[:outP], nil
+}
+
+// Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits)
+// note 1 : assume symbolValue is valid (<= maxSymbolValue)
+// note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits *
+func (s *fseEncoder) bitCost(symbolValue uint8, accuracyLog uint32) uint32 {
+	minNbBits := s.ct.symbolTT[symbolValue].deltaNbBits >> 16
+	threshold := (minNbBits + 1) << 16
+	if debugAsserts {
+		if !(s.actualTableLog < 16) {
+			panic("!s.actualTableLog < 16")
+		}
+		// ensure enough room for renormalization double shift
+		if !(uint8(accuracyLog) < 31-s.actualTableLog) {
+			panic("!uint8(accuracyLog) < 31-s.actualTableLog")
+		}
+	}
+	tableSize := uint32(1) << s.actualTableLog
+	deltaFromThreshold := threshold - (s.ct.symbolTT[symbolValue].deltaNbBits + tableSize)
+	// linear interpolation (very approximate)
+	normalizedDeltaFromThreshold := (deltaFromThreshold << accuracyLog) >> s.actualTableLog
+	bitMultiplier := uint32(1) << accuracyLog
+	if debugAsserts {
+		if s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold {
+			panic("s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold")
+		}
+		if normalizedDeltaFromThreshold > bitMultiplier {
+			panic("normalizedDeltaFromThreshold > bitMultiplier")
+		}
+	}
+	return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold
+}
+
+// Returns the cost in bits of encoding the distribution in count using ctable.
+// Histogram should only be up to the last non-zero symbol.
+// Returns an -1 if ctable cannot represent all the symbols in count.
+func (s *fseEncoder) approxSize(hist []uint32) uint32 {
+	if int(s.symbolLen) < len(hist) {
+		// More symbols than we have.
+		return math.MaxUint32
+	}
+	if s.useRLE {
+		// We will never reuse RLE encoders.
+		return math.MaxUint32
+	}
+	const kAccuracyLog = 8
+	badCost := (uint32(s.actualTableLog) + 1) << kAccuracyLog
+	var cost uint32
+	for i, v := range hist {
+		if v == 0 {
+			continue
+		}
+		if s.norm[i] == 0 {
+			return math.MaxUint32
+		}
+		bitCost := s.bitCost(uint8(i), kAccuracyLog)
+		if bitCost > badCost {
+			return math.MaxUint32
+		}
+		cost += v * bitCost
+	}
+	return cost >> kAccuracyLog
+}
+
+// maxHeaderSize returns the maximum header size in bits.
+// This is not exact size, but we want a penalty for new tables anyway.
+func (s *fseEncoder) maxHeaderSize() uint32 {
+	if s.preDefined {
+		return 0
+	}
+	if s.useRLE {
+		return 8
+	}
+	return (((uint32(s.symbolLen) * uint32(s.actualTableLog)) >> 3) + 3) * 8
+}
+
+// cState contains the compression state of a stream.
+type cState struct {
+	bw         *bitWriter
+	stateTable []uint16
+	state      uint16
+}
+
+// init will initialize the compression state to the first symbol of the stream.
+func (c *cState) init(bw *bitWriter, ct *cTable, first symbolTransform) {
+	c.bw = bw
+	c.stateTable = ct.stateTable
+	if len(c.stateTable) == 1 {
+		// RLE
+		c.stateTable[0] = uint16(0)
+		c.state = 0
+		return
+	}
+	nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16
+	im := int32((nbBitsOut << 16) - first.deltaNbBits)
+	lu := (im >> nbBitsOut) + int32(first.deltaFindState)
+	c.state = c.stateTable[lu]
+}
+
+// encode the output symbol provided and write it to the bitstream.
+func (c *cState) encode(symbolTT symbolTransform) {
+	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
+	dstState := int32(c.state>>(nbBitsOut&15)) + int32(symbolTT.deltaFindState)
+	c.bw.addBits16NC(c.state, uint8(nbBitsOut))
+	c.state = c.stateTable[dstState]
+}
+
+// flush will write the tablelog to the output and flush the remaining full bytes.
+func (c *cState) flush(tableLog uint8) {
+	c.bw.flush32()
+	c.bw.addBits16NC(c.state, tableLog)
+}

vendor/github.com/klauspost/compress/zstd/fse_predefined.go

@@ -0,0 +1,158 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"fmt"
+	"math"
+	"sync"
+)
+
+var (
+	// fsePredef are the predefined fse tables as defined here:
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions
+	// These values are already transformed.
+	fsePredef [3]fseDecoder
+
+	// fsePredefEnc are the predefined encoder based on fse tables as defined here:
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions
+	// These values are already transformed.
+	fsePredefEnc [3]fseEncoder
+
+	// symbolTableX contain the transformations needed for each type as defined in
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#the-codes-for-literals-lengths-match-lengths-and-offsets
+	symbolTableX [3][]baseOffset
+
+	// maxTableSymbol is the biggest supported symbol for each table type
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#the-codes-for-literals-lengths-match-lengths-and-offsets
+	maxTableSymbol = [3]uint8{tableLiteralLengths: maxLiteralLengthSymbol, tableOffsets: maxOffsetLengthSymbol, tableMatchLengths: maxMatchLengthSymbol}
+
+	// bitTables is the bits table for each table.
+	bitTables = [3][]byte{tableLiteralLengths: llBitsTable[:], tableOffsets: nil, tableMatchLengths: mlBitsTable[:]}
+)
+
+type tableIndex uint8
+
+const (
+	// indexes for fsePredef and symbolTableX
+	tableLiteralLengths tableIndex = 0
+	tableOffsets        tableIndex = 1
+	tableMatchLengths   tableIndex = 2
+
+	maxLiteralLengthSymbol = 35
+	maxOffsetLengthSymbol  = 30
+	maxMatchLengthSymbol   = 52
+)
+
+// baseOffset is used for calculating transformations.
+type baseOffset struct {
+	baseLine uint32
+	addBits  uint8
+}
+
+// fillBase will precalculate base offsets with the given bit distributions.
+func fillBase(dst []baseOffset, base uint32, bits ...uint8) {
+	if len(bits) != len(dst) {
+		panic(fmt.Sprintf("len(dst) (%d) != len(bits) (%d)", len(dst), len(bits)))
+	}
+	for i, bit := range bits {
+		if base > math.MaxInt32 {
+			panic("invalid decoding table, base overflows int32")
+		}
+
+		dst[i] = baseOffset{
+			baseLine: base,
+			addBits:  bit,
+		}
+		base += 1 << bit
+	}
+}
+
+var predef sync.Once
+
+func initPredefined() {
+	predef.Do(func() {
+		// Literals length codes
+		tmp := make([]baseOffset, 36)
+		for i := range tmp[:16] {
+			tmp[i] = baseOffset{
+				baseLine: uint32(i),
+				addBits:  0,
+			}
+		}
+		fillBase(tmp[16:], 16, 1, 1, 1, 1, 2, 2, 3, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16)
+		symbolTableX[tableLiteralLengths] = tmp
+
+		// Match length codes
+		tmp = make([]baseOffset, 53)
+		for i := range tmp[:32] {
+			tmp[i] = baseOffset{
+				// The transformation adds the 3 length.
+				baseLine: uint32(i) + 3,
+				addBits:  0,
+			}
+		}
+		fillBase(tmp[32:], 35, 1, 1, 1, 1, 2, 2, 3, 3, 4, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16)
+		symbolTableX[tableMatchLengths] = tmp
+
+		// Offset codes
+		tmp = make([]baseOffset, maxOffsetBits+1)
+		tmp[1] = baseOffset{
+			baseLine: 1,
+			addBits:  1,
+		}
+		fillBase(tmp[2:], 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30)
+		symbolTableX[tableOffsets] = tmp
+
+		// Fill predefined tables and transform them.
+		// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions
+		for i := range fsePredef[:] {
+			f := &fsePredef[i]
+			switch tableIndex(i) {
+			case tableLiteralLengths:
+				// https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L243
+				f.actualTableLog = 6
+				copy(f.norm[:], []int16{4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1,
+					2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1,
+					-1, -1, -1, -1})
+				f.symbolLen = 36
+			case tableOffsets:
+				// https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L281
+				f.actualTableLog = 5
+				copy(f.norm[:], []int16{
+					1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+					1, 1, 1, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1})
+				f.symbolLen = 29
+			case tableMatchLengths:
+				//https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L304
+				f.actualTableLog = 6
+				copy(f.norm[:], []int16{
+					1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+					1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+					1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, -1, -1,
+					-1, -1, -1, -1, -1})
+				f.symbolLen = 53
+			}
+			if err := f.buildDtable(); err != nil {
+				panic(fmt.Errorf("building table %v: %v", tableIndex(i), err))
+			}
+			if err := f.transform(symbolTableX[i]); err != nil {
+				panic(fmt.Errorf("building table %v: %v", tableIndex(i), err))
+			}
+			f.preDefined = true
+
+			// Create encoder as well
+			enc := &fsePredefEnc[i]
+			copy(enc.norm[:], f.norm[:])
+			enc.symbolLen = f.symbolLen
+			enc.actualTableLog = f.actualTableLog
+			if err := enc.buildCTable(); err != nil {
+				panic(fmt.Errorf("building encoding table %v: %v", tableIndex(i), err))
+			}
+			enc.setBits(bitTables[i])
+			enc.preDefined = true
+		}
+	})
+}

vendor/github.com/klauspost/compress/zstd/fuzz.go

@@ -0,0 +1,11 @@
+//go:build ignorecrc
+// +build ignorecrc
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+// ignoreCRC can be used for fuzz testing to ignore CRC values...
+const ignoreCRC = true

vendor/github.com/klauspost/compress/zstd/fuzz_none.go

@@ -0,0 +1,11 @@
+//go:build !ignorecrc
+// +build !ignorecrc
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+// ignoreCRC can be used for fuzz testing to ignore CRC values...
+const ignoreCRC = false

vendor/github.com/klauspost/compress/zstd/hash.go

@@ -0,0 +1,41 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+const (
+	prime3bytes = 506832829
+	prime4bytes = 2654435761
+	prime5bytes = 889523592379
+	prime6bytes = 227718039650203
+	prime7bytes = 58295818150454627
+	prime8bytes = 0xcf1bbcdcb7a56463
+)
+
+// hashLen returns a hash of the lowest mls bytes of with length output bits.
+// mls must be >=3 and <=8. Any other value will return hash for 4 bytes.
+// length should always be < 32.
+// Preferably length and mls should be a constant for inlining.
+func hashLen(u uint64, length, mls uint8) uint32 {
+	switch mls {
+	case 3:
+		return (uint32(u<<8) * prime3bytes) >> (32 - length)
+	case 5:
+		return uint32(((u << (64 - 40)) * prime5bytes) >> (64 - length))
+	case 6:
+		return uint32(((u << (64 - 48)) * prime6bytes) >> (64 - length))
+	case 7:
+		return uint32(((u << (64 - 56)) * prime7bytes) >> (64 - length))
+	case 8:
+		return uint32((u * prime8bytes) >> (64 - length))
+	default:
+		return (uint32(u) * prime4bytes) >> (32 - length)
+	}
+}
+
+// hash3 returns the hash of the lower 3 bytes of u to fit in a hash table with h bits.
+// Preferably h should be a constant and should always be <32.
+func hash3(u uint32, h uint8) uint32 {
+	return ((u << (32 - 24)) * prime3bytes) >> ((32 - h) & 31)
+}

vendor/github.com/klauspost/compress/zstd/history.go

@@ -0,0 +1,119 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+// Based on work by Yann Collet, released under BSD License.
+
+package zstd
+
+import (
+	"github.com/klauspost/compress/huff0"
+)
+
+// history contains the information transferred between blocks.
+type history struct {
+	// Literal decompression
+	huffTree *huff0.Scratch
+
+	// Sequence decompression
+	decoders      sequenceDecs
+	recentOffsets [3]int
+
+	// History buffer...
+	b []byte
+
+	// ignoreBuffer is meant to ignore a number of bytes
+	// when checking for matches in history
+	ignoreBuffer int
+
+	windowSize       int
+	allocFrameBuffer int // needed?
+	error            bool
+	dict             *dict
+}
+
+// reset will reset the history to initial state of a frame.
+// The history must already have been initialized to the desired size.
+func (h *history) reset() {
+	h.b = h.b[:0]
+	h.ignoreBuffer = 0
+	h.error = false
+	h.recentOffsets = [3]int{1, 4, 8}
+	if f := h.decoders.litLengths.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+	}
+	if f := h.decoders.offsets.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+	}
+	if f := h.decoders.matchLengths.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+	}
+	h.decoders = sequenceDecs{br: h.decoders.br}
+	if h.huffTree != nil {
+		if h.dict == nil || h.dict.litEnc != h.huffTree {
+			huffDecoderPool.Put(h.huffTree)
+		}
+	}
+	h.huffTree = nil
+	h.dict = nil
+	//printf("history created: %+v (l: %d, c: %d)", *h, len(h.b), cap(h.b))
+}
+
+func (h *history) setDict(dict *dict) {
+	if dict == nil {
+		return
+	}
+	h.dict = dict
+	h.decoders.litLengths = dict.llDec
+	h.decoders.offsets = dict.ofDec
+	h.decoders.matchLengths = dict.mlDec
+	h.decoders.dict = dict.content
+	h.recentOffsets = dict.offsets
+	h.huffTree = dict.litEnc
+}
+
+// append bytes to history.
+// This function will make sure there is space for it,
+// if the buffer has been allocated with enough extra space.
+func (h *history) append(b []byte) {
+	if len(b) >= h.windowSize {
+		// Discard all history by simply overwriting
+		h.b = h.b[:h.windowSize]
+		copy(h.b, b[len(b)-h.windowSize:])
+		return
+	}
+
+	// If there is space, append it.
+	if len(b) < cap(h.b)-len(h.b) {
+		h.b = append(h.b, b...)
+		return
+	}
+
+	// Move data down so we only have window size left.
+	// We know we have less than window size in b at this point.
+	discard := len(b) + len(h.b) - h.windowSize
+	copy(h.b, h.b[discard:])
+	h.b = h.b[:h.windowSize]
+	copy(h.b[h.windowSize-len(b):], b)
+}
+
+// ensureBlock will ensure there is space for at least one block...
+func (h *history) ensureBlock() {
+	if cap(h.b) < h.allocFrameBuffer {
+		h.b = make([]byte, 0, h.allocFrameBuffer)
+		return
+	}
+
+	avail := cap(h.b) - len(h.b)
+	if avail >= h.windowSize || avail > maxCompressedBlockSize {
+		return
+	}
+	// Move data down so we only have window size left.
+	// We know we have less than window size in b at this point.
+	discard := len(h.b) - h.windowSize
+	copy(h.b, h.b[discard:])
+	h.b = h.b[:h.windowSize]
+}
+
+// append bytes to history without ever discarding anything.
+func (h *history) appendKeep(b []byte) {
+	h.b = append(h.b, b...)
+}

vendor/github.com/klauspost/compress/zstd/internal/xxhash/LICENSE.txt

@@ -0,0 +1,22 @@
+Copyright (c) 2016 Caleb Spare
+
+MIT License
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md

@@ -0,0 +1,58 @@
+# xxhash
+
+VENDORED: Go to [github.com/cespare/xxhash](https://github.com/cespare/xxhash) for original package.
+
+
+[![GoDoc](https://godoc.org/github.com/cespare/xxhash?status.svg)](https://godoc.org/github.com/cespare/xxhash)
+[![Build Status](https://travis-ci.org/cespare/xxhash.svg?branch=master)](https://travis-ci.org/cespare/xxhash)
+
+xxhash is a Go implementation of the 64-bit
+[xxHash](http://cyan4973.github.io/xxHash/) algorithm, XXH64. This is a
+high-quality hashing algorithm that is much faster than anything in the Go
+standard library.
+
+This package provides a straightforward API:
+
+```
+func Sum64(b []byte) uint64
+func Sum64String(s string) uint64
+type Digest struct{ ... }
+    func New() *Digest
+```
+
+The `Digest` type implements hash.Hash64. Its key methods are:
+
+```
+func (*Digest) Write([]byte) (int, error)
+func (*Digest) WriteString(string) (int, error)
+func (*Digest) Sum64() uint64
+```
+
+This implementation provides a fast pure-Go implementation and an even faster
+assembly implementation for amd64.
+
+## Benchmarks
+
+Here are some quick benchmarks comparing the pure-Go and assembly
+implementations of Sum64.
+
+| input size | purego | asm |
+| --- | --- | --- |
+| 5 B   |  979.66 MB/s |  1291.17 MB/s  |
+| 100 B | 7475.26 MB/s | 7973.40 MB/s  |
+| 4 KB  | 17573.46 MB/s | 17602.65 MB/s |
+| 10 MB | 17131.46 MB/s | 17142.16 MB/s |
+
+These numbers were generated on Ubuntu 18.04 with an Intel i7-8700K CPU using
+the following commands under Go 1.11.2:
+
+```
+$ go test -tags purego -benchtime 10s -bench '/xxhash,direct,bytes'
+$ go test -benchtime 10s -bench '/xxhash,direct,bytes'
+```
+
+## Projects using this package
+
+- [InfluxDB](https://github.com/influxdata/influxdb)
+- [Prometheus](https://github.com/prometheus/prometheus)
+- [FreeCache](https://github.com/coocood/freecache)

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go

@@ -0,0 +1,237 @@
+// Package xxhash implements the 64-bit variant of xxHash (XXH64) as described
+// at http://cyan4973.github.io/xxHash/.
+// THIS IS VENDORED: Go to github.com/cespare/xxhash for original package.
+
+package xxhash
+
+import (
+	"encoding/binary"
+	"errors"
+	"math/bits"
+)
+
+const (
+	prime1 uint64 = 11400714785074694791
+	prime2 uint64 = 14029467366897019727
+	prime3 uint64 = 1609587929392839161
+	prime4 uint64 = 9650029242287828579
+	prime5 uint64 = 2870177450012600261
+)
+
+// NOTE(caleb): I'm using both consts and vars of the primes. Using consts where
+// possible in the Go code is worth a small (but measurable) performance boost
+// by avoiding some MOVQs. Vars are needed for the asm and also are useful for
+// convenience in the Go code in a few places where we need to intentionally
+// avoid constant arithmetic (e.g., v1 := prime1 + prime2 fails because the
+// result overflows a uint64).
+var (
+	prime1v = prime1
+	prime2v = prime2
+	prime3v = prime3
+	prime4v = prime4
+	prime5v = prime5
+)
+
+// Digest implements hash.Hash64.
+type Digest struct {
+	v1    uint64
+	v2    uint64
+	v3    uint64
+	v4    uint64
+	total uint64
+	mem   [32]byte
+	n     int // how much of mem is used
+}
+
+// New creates a new Digest that computes the 64-bit xxHash algorithm.
+func New() *Digest {
+	var d Digest
+	d.Reset()
+	return &d
+}
+
+// Reset clears the Digest's state so that it can be reused.
+func (d *Digest) Reset() {
+	d.v1 = prime1v + prime2
+	d.v2 = prime2
+	d.v3 = 0
+	d.v4 = -prime1v
+	d.total = 0
+	d.n = 0
+}
+
+// Size always returns 8 bytes.
+func (d *Digest) Size() int { return 8 }
+
+// BlockSize always returns 32 bytes.
+func (d *Digest) BlockSize() int { return 32 }
+
+// Write adds more data to d. It always returns len(b), nil.
+func (d *Digest) Write(b []byte) (n int, err error) {
+	n = len(b)
+	d.total += uint64(n)
+
+	if d.n+n < 32 {
+		// This new data doesn't even fill the current block.
+		copy(d.mem[d.n:], b)
+		d.n += n
+		return
+	}
+
+	if d.n > 0 {
+		// Finish off the partial block.
+		copy(d.mem[d.n:], b)
+		d.v1 = round(d.v1, u64(d.mem[0:8]))
+		d.v2 = round(d.v2, u64(d.mem[8:16]))
+		d.v3 = round(d.v3, u64(d.mem[16:24]))
+		d.v4 = round(d.v4, u64(d.mem[24:32]))
+		b = b[32-d.n:]
+		d.n = 0
+	}
+
+	if len(b) >= 32 {
+		// One or more full blocks left.
+		nw := writeBlocks(d, b)
+		b = b[nw:]
+	}
+
+	// Store any remaining partial block.
+	copy(d.mem[:], b)
+	d.n = len(b)
+
+	return
+}
+
+// Sum appends the current hash to b and returns the resulting slice.
+func (d *Digest) Sum(b []byte) []byte {
+	s := d.Sum64()
+	return append(
+		b,
+		byte(s>>56),
+		byte(s>>48),
+		byte(s>>40),
+		byte(s>>32),
+		byte(s>>24),
+		byte(s>>16),
+		byte(s>>8),
+		byte(s),
+	)
+}
+
+// Sum64 returns the current hash.
+func (d *Digest) Sum64() uint64 {
+	var h uint64
+
+	if d.total >= 32 {
+		v1, v2, v3, v4 := d.v1, d.v2, d.v3, d.v4
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = d.v3 + prime5
+	}
+
+	h += d.total
+
+	i, end := 0, d.n
+	for ; i+8 <= end; i += 8 {
+		k1 := round(0, u64(d.mem[i:i+8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if i+4 <= end {
+		h ^= uint64(u32(d.mem[i:i+4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		i += 4
+	}
+	for i < end {
+		h ^= uint64(d.mem[i]) * prime5
+		h = rol11(h) * prime1
+		i++
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+const (
+	magic         = "xxh\x06"
+	marshaledSize = len(magic) + 8*5 + 32
+)
+
+// MarshalBinary implements the encoding.BinaryMarshaler interface.
+func (d *Digest) MarshalBinary() ([]byte, error) {
+	b := make([]byte, 0, marshaledSize)
+	b = append(b, magic...)
+	b = appendUint64(b, d.v1)
+	b = appendUint64(b, d.v2)
+	b = appendUint64(b, d.v3)
+	b = appendUint64(b, d.v4)
+	b = appendUint64(b, d.total)
+	b = append(b, d.mem[:d.n]...)
+	b = b[:len(b)+len(d.mem)-d.n]
+	return b, nil
+}
+
+// UnmarshalBinary implements the encoding.BinaryUnmarshaler interface.
+func (d *Digest) UnmarshalBinary(b []byte) error {
+	if len(b) < len(magic) || string(b[:len(magic)]) != magic {
+		return errors.New("xxhash: invalid hash state identifier")
+	}
+	if len(b) != marshaledSize {
+		return errors.New("xxhash: invalid hash state size")
+	}
+	b = b[len(magic):]
+	b, d.v1 = consumeUint64(b)
+	b, d.v2 = consumeUint64(b)
+	b, d.v3 = consumeUint64(b)
+	b, d.v4 = consumeUint64(b)
+	b, d.total = consumeUint64(b)
+	copy(d.mem[:], b)
+	d.n = int(d.total % uint64(len(d.mem)))
+	return nil
+}
+
+func appendUint64(b []byte, x uint64) []byte {
+	var a [8]byte
+	binary.LittleEndian.PutUint64(a[:], x)
+	return append(b, a[:]...)
+}
+
+func consumeUint64(b []byte) ([]byte, uint64) {
+	x := u64(b)
+	return b[8:], x
+}
+
+func u64(b []byte) uint64 { return binary.LittleEndian.Uint64(b) }
+func u32(b []byte) uint32 { return binary.LittleEndian.Uint32(b) }
+
+func round(acc, input uint64) uint64 {
+	acc += input * prime2
+	acc = rol31(acc)
+	acc *= prime1
+	return acc
+}
+
+func mergeRound(acc, val uint64) uint64 {
+	val = round(0, val)
+	acc ^= val
+	acc = acc*prime1 + prime4
+	return acc
+}
+
+func rol1(x uint64) uint64  { return bits.RotateLeft64(x, 1) }
+func rol7(x uint64) uint64  { return bits.RotateLeft64(x, 7) }
+func rol11(x uint64) uint64 { return bits.RotateLeft64(x, 11) }
+func rol12(x uint64) uint64 { return bits.RotateLeft64(x, 12) }
+func rol18(x uint64) uint64 { return bits.RotateLeft64(x, 18) }
+func rol23(x uint64) uint64 { return bits.RotateLeft64(x, 23) }
+func rol27(x uint64) uint64 { return bits.RotateLeft64(x, 27) }
+func rol31(x uint64) uint64 { return bits.RotateLeft64(x, 31) }

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s

@@ -0,0 +1,216 @@
+// +build !appengine
+// +build gc
+// +build !purego
+// +build !noasm
+
+#include "textflag.h"
+
+// Register allocation:
+// AX	h
+// SI	pointer to advance through b
+// DX	n
+// BX	loop end
+// R8	v1, k1
+// R9	v2
+// R10	v3
+// R11	v4
+// R12	tmp
+// R13	prime1v
+// R14	prime2v
+// DI	prime4v
+
+// round reads from and advances the buffer pointer in SI.
+// It assumes that R13 has prime1v and R14 has prime2v.
+#define round(r) \
+	MOVQ  (SI), R12 \
+	ADDQ  $8, SI    \
+	IMULQ R14, R12  \
+	ADDQ  R12, r    \
+	ROLQ  $31, r    \
+	IMULQ R13, r
+
+// mergeRound applies a merge round on the two registers acc and val.
+// It assumes that R13 has prime1v, R14 has prime2v, and DI has prime4v.
+#define mergeRound(acc, val) \
+	IMULQ R14, val \
+	ROLQ  $31, val \
+	IMULQ R13, val \
+	XORQ  val, acc \
+	IMULQ R13, acc \
+	ADDQ  DI, acc
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT, $0-32
+	// Load fixed primes.
+	MOVQ ·prime1v(SB), R13
+	MOVQ ·prime2v(SB), R14
+	MOVQ ·prime4v(SB), DI
+
+	// Load slice.
+	MOVQ b_base+0(FP), SI
+	MOVQ b_len+8(FP), DX
+	LEAQ (SI)(DX*1), BX
+
+	// The first loop limit will be len(b)-32.
+	SUBQ $32, BX
+
+	// Check whether we have at least one block.
+	CMPQ DX, $32
+	JLT  noBlocks
+
+	// Set up initial state (v1, v2, v3, v4).
+	MOVQ R13, R8
+	ADDQ R14, R8
+	MOVQ R14, R9
+	XORQ R10, R10
+	XORQ R11, R11
+	SUBQ R13, R11
+
+	// Loop until SI > BX.
+blockLoop:
+	round(R8)
+	round(R9)
+	round(R10)
+	round(R11)
+
+	CMPQ SI, BX
+	JLE  blockLoop
+
+	MOVQ R8, AX
+	ROLQ $1, AX
+	MOVQ R9, R12
+	ROLQ $7, R12
+	ADDQ R12, AX
+	MOVQ R10, R12
+	ROLQ $12, R12
+	ADDQ R12, AX
+	MOVQ R11, R12
+	ROLQ $18, R12
+	ADDQ R12, AX
+
+	mergeRound(AX, R8)
+	mergeRound(AX, R9)
+	mergeRound(AX, R10)
+	mergeRound(AX, R11)
+
+	JMP afterBlocks
+
+noBlocks:
+	MOVQ ·prime5v(SB), AX
+
+afterBlocks:
+	ADDQ DX, AX
+
+	// Right now BX has len(b)-32, and we want to loop until SI > len(b)-8.
+	ADDQ $24, BX
+
+	CMPQ SI, BX
+	JG   fourByte
+
+wordLoop:
+	// Calculate k1.
+	MOVQ  (SI), R8
+	ADDQ  $8, SI
+	IMULQ R14, R8
+	ROLQ  $31, R8
+	IMULQ R13, R8
+
+	XORQ  R8, AX
+	ROLQ  $27, AX
+	IMULQ R13, AX
+	ADDQ  DI, AX
+
+	CMPQ SI, BX
+	JLE  wordLoop
+
+fourByte:
+	ADDQ $4, BX
+	CMPQ SI, BX
+	JG   singles
+
+	MOVL  (SI), R8
+	ADDQ  $4, SI
+	IMULQ R13, R8
+	XORQ  R8, AX
+
+	ROLQ  $23, AX
+	IMULQ R14, AX
+	ADDQ  ·prime3v(SB), AX
+
+singles:
+	ADDQ $4, BX
+	CMPQ SI, BX
+	JGE  finalize
+
+singlesLoop:
+	MOVBQZX (SI), R12
+	ADDQ    $1, SI
+	IMULQ   ·prime5v(SB), R12
+	XORQ    R12, AX
+
+	ROLQ  $11, AX
+	IMULQ R13, AX
+
+	CMPQ SI, BX
+	JL   singlesLoop
+
+finalize:
+	MOVQ  AX, R12
+	SHRQ  $33, R12
+	XORQ  R12, AX
+	IMULQ R14, AX
+	MOVQ  AX, R12
+	SHRQ  $29, R12
+	XORQ  R12, AX
+	IMULQ ·prime3v(SB), AX
+	MOVQ  AX, R12
+	SHRQ  $32, R12
+	XORQ  R12, AX
+
+	MOVQ AX, ret+24(FP)
+	RET
+
+// writeBlocks uses the same registers as above except that it uses AX to store
+// the d pointer.
+
+// func writeBlocks(d *Digest, b []byte) int
+TEXT ·writeBlocks(SB), NOSPLIT, $0-40
+	// Load fixed primes needed for round.
+	MOVQ ·prime1v(SB), R13
+	MOVQ ·prime2v(SB), R14
+
+	// Load slice.
+	MOVQ b_base+8(FP), SI
+	MOVQ b_len+16(FP), DX
+	LEAQ (SI)(DX*1), BX
+	SUBQ $32, BX
+
+	// Load vN from d.
+	MOVQ d+0(FP), AX
+	MOVQ 0(AX), R8   // v1
+	MOVQ 8(AX), R9   // v2
+	MOVQ 16(AX), R10 // v3
+	MOVQ 24(AX), R11 // v4
+
+	// We don't need to check the loop condition here; this function is
+	// always called with at least one block of data to process.
+blockLoop:
+	round(R8)
+	round(R9)
+	round(R10)
+	round(R11)
+
+	CMPQ SI, BX
+	JLE  blockLoop
+
+	// Copy vN back to d.
+	MOVQ R8, 0(AX)
+	MOVQ R9, 8(AX)
+	MOVQ R10, 16(AX)
+	MOVQ R11, 24(AX)
+
+	// The number of bytes written is SI minus the old base pointer.
+	SUBQ b_base+8(FP), SI
+	MOVQ SI, ret+32(FP)
+
+	RET

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_arm64.s

@@ -0,0 +1,186 @@
+// +build gc,!purego,!noasm
+
+#include "textflag.h"
+
+// Register allocation.
+#define digest	R1
+#define h	R2 // Return value.
+#define p	R3 // Input pointer.
+#define len	R4
+#define nblocks	R5 // len / 32.
+#define prime1	R7
+#define prime2	R8
+#define prime3	R9
+#define prime4	R10
+#define prime5	R11
+#define v1	R12
+#define v2	R13
+#define v3	R14
+#define v4	R15
+#define x1	R20
+#define x2	R21
+#define x3	R22
+#define x4	R23
+
+#define round(acc, x) \
+	MADD prime2, acc, x, acc \
+	ROR  $64-31, acc         \
+	MUL  prime1, acc         \
+
+// x = round(0, x).
+#define round0(x) \
+	MUL prime2, x \
+	ROR $64-31, x \
+	MUL prime1, x \
+
+#define mergeRound(x) \
+	round0(x)                 \
+	EOR  x, h                 \
+	MADD h, prime4, prime1, h \
+
+// Update v[1-4] with 32-byte blocks. Assumes len >= 32.
+#define blocksLoop() \
+	LSR     $5, len, nblocks \
+	PCALIGN $16              \
+	loop:                    \
+	LDP.P   32(p), (x1, x2)  \
+	round(v1, x1)            \
+	LDP     -16(p), (x3, x4) \
+	round(v2, x2)            \
+	SUB     $1, nblocks      \
+	round(v3, x3)            \
+	round(v4, x4)            \
+	CBNZ    nblocks, loop    \
+
+// The primes are repeated here to ensure that they're stored
+// in a contiguous array, so we can load them with LDP.
+DATA primes<> +0(SB)/8, $11400714785074694791
+DATA primes<> +8(SB)/8, $14029467366897019727
+DATA primes<>+16(SB)/8, $1609587929392839161
+DATA primes<>+24(SB)/8, $9650029242287828579
+DATA primes<>+32(SB)/8, $2870177450012600261
+GLOBL primes<>(SB), NOPTR+RODATA, $40
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOFRAME+NOSPLIT, $0-32
+	LDP b_base+0(FP), (p, len)
+
+	LDP  primes<> +0(SB), (prime1, prime2)
+	LDP  primes<>+16(SB), (prime3, prime4)
+	MOVD primes<>+32(SB), prime5
+
+	CMP  $32, len
+	CSEL LO, prime5, ZR, h // if len < 32 { h = prime5 } else { h = 0 }
+	BLO  afterLoop
+
+	ADD  prime1, prime2, v1
+	MOVD prime2, v2
+	MOVD $0, v3
+	NEG  prime1, v4
+
+	blocksLoop()
+
+	ROR $64-1, v1, x1
+	ROR $64-7, v2, x2
+	ADD x1, x2
+	ROR $64-12, v3, x3
+	ROR $64-18, v4, x4
+	ADD x3, x4
+	ADD x2, x4, h
+
+	mergeRound(v1)
+	mergeRound(v2)
+	mergeRound(v3)
+	mergeRound(v4)
+
+afterLoop:
+	ADD len, h
+
+	TBZ   $4, len, try8
+	LDP.P 16(p), (x1, x2)
+
+	round0(x1)
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+	round0(x2)
+	ROR  $64-27, h
+	EOR  x2 @> 64-27, h
+	MADD h, prime4, prime1, h
+
+try8:
+	TBZ    $3, len, try4
+	MOVD.P 8(p), x1
+
+	round0(x1)
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h
+	MADD h, prime4, prime1, h
+
+try4:
+	TBZ     $2, len, try2
+	MOVWU.P 4(p), x2
+
+	MUL  prime1, x2
+	ROR  $64-23, h
+	EOR  x2 @> 64-23, h
+	MADD h, prime3, prime2, h
+
+try2:
+	TBZ     $1, len, try1
+	MOVHU.P 2(p), x3
+	AND     $255, x3, x1
+	LSR     $8, x3, x2
+
+	MUL prime5, x1
+	ROR $64-11, h
+	EOR x1 @> 64-11, h
+	MUL prime1, h
+
+	MUL prime5, x2
+	ROR $64-11, h
+	EOR x2 @> 64-11, h
+	MUL prime1, h
+
+try1:
+	TBZ   $0, len, end
+	MOVBU (p), x4
+
+	MUL prime5, x4
+	ROR $64-11, h
+	EOR x4 @> 64-11, h
+	MUL prime1, h
+
+end:
+	EOR h >> 33, h
+	MUL prime2, h
+	EOR h >> 29, h
+	MUL prime3, h
+	EOR h >> 32, h
+
+	MOVD h, ret+24(FP)
+	RET
+
+// func writeBlocks(d *Digest, b []byte) int
+//
+// Assumes len(b) >= 32.
+TEXT ·writeBlocks(SB), NOFRAME+NOSPLIT, $0-40
+	LDP primes<>(SB), (prime1, prime2)
+
+	// Load state. Assume v[1-4] are stored contiguously.
+	MOVD d+0(FP), digest
+	LDP  0(digest), (v1, v2)
+	LDP  16(digest), (v3, v4)
+
+	LDP b_base+8(FP), (p, len)
+
+	blocksLoop()
+
+	// Store updated state.
+	STP (v1, v2), 0(digest)
+	STP (v3, v4), 16(digest)
+
+	BIC  $31, len
+	MOVD len, ret+32(FP)
+	RET