| // Copyright 2026 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. |
| |
| package midway |
| |
| import ( |
| "cmd/compile/internal/base" |
| "cmd/compile/internal/syntax" |
| "cmd/compile/internal/types2" |
| "fmt" |
| "internal/buildcfg" |
| "strings" |
| ) |
| |
| // "Midway" rewriting |
| // |
| // Go attempts to provide a package similar to the the "Highway" library |
| // for C++ (https://google.github.io/highway). The library package is "simd" |
| // and defines vector types with unspecified widths that are bound to particular |
| // machine dependent types as late as program execution. This is accomplished |
| // by rewriting code that depends on these types into code that references |
| // architecture-specific types, perhaps more than once, and if necessary |
| // dynamically choosing which version to execute based on hardware attributes. |
| // |
| // The rewriting takes place early in the compiler, after type checking but |
| // before conversion to "unified" IR. To ensure that types are correctly set |
| // on the modified version of the code, type checking information is reset and |
| // the type checking phase is re-run. The places some limits on the shape of |
| // the rewrites, but it also ensures that the rewritten code is well-formed. |
| // |
| // Rewritten code does not reference "archsimd" types directly, but instead |
| // references types in a "bridge" package that filters the available methods |
| // and adds a few more. The package used relies on a builder/compiler hack; |
| // the compiler's type checker enforces export naming conventions, but the |
| // build system limits visibility to unrelated "internal" packages and can be |
| // modified to allow access in special cases (like this one). This allows the |
| // rewritten code to reference types, functions, and methods that are not |
| // accessible otherwise. |
| // |
| // The rewrite works in phases. The first is "analysis", to discover functions, |
| // types, methods, and variables that depend on "simd" types. "Depend on" means |
| // any mention of a simd type, and for types, also includes types that have a |
| // simd-dependent method. Dependent functions are split into two categories; |
| // those whose dependence includes their signature, and those that do not. |
| // The second category forms the boundary between code that depends on simd and |
| // code that does not. Notice that there cannot be a boundary method, because |
| // (by design) the receiver type is simd-dependent and thus a dependent method |
| // also has a dependent type in its signature. |
| // |
| // The second phase rewrites such "boundary" functions into a "dispatch" version |
| // and (later, third phase) "specialized" versions. The dispatch function |
| // will choose which specialized version to call based on which simd implementation |
| // has been chosen, and forward parameters and results to/from that specialized version |
| // of the function. The dispatch version shares the same name as the original function. |
| // Note that this applies to functions only, and not methods. |
| |
| // The third phase specializes dependent functions (both kinds), methods, |
| // global variables, and types into size/emulation/feature-specific variants. |
| // Except for methods, this is done by adding a suffix beginning with "@" to |
| // the name. Because "@" cannot appear in legal Go identifiers this removes |
| // the risk of a naming overlap. Methods are specialized, but not renamed, |
| // because their receiver type is renamed instead. Not changing method names |
| // preserves interface satisfaction, for example in the case of generic interfaces. |
| // |
| // Non-boundary dependent function and methods are not rewritten into dispatch |
| // functions/methods, but remain in the generated code because they must be |
| // present in the export data so that other packages that import them will still |
| // compile before rewriting. Their bodies are replaced with panic(...) to allow |
| // compilation while preventing even worse chaos in the event of a bug either in |
| // the compiler or through ambitious use of reflection or assembly language. |
| // |
| |
| /* Example rewrites |
| |
| // Type alias, global variable, and init function: |
| |
| // before: |
| type MyInt8s = simd.Int8s |
| func Generic[T haslen](x int) int { |
| var v T |
| return x + v.Len() |
| } |
| var VL int |
| func init() { |
| VL = Generic[MyInt8s](1) |
| } |
| // dispatch: |
| func init() { |
| switch simd.VectorBitSize() { |
| case |
| 128: |
| init@simd128() |
| return |
| case 256: |
| init@simd256() |
| return |
| case 512: |
| init@simd512() |
| return |
| default: |
| panic("unsupported vector size") |
| } |
| } |
| // specialized (128) |
| type MyInt8s@simd128 = archsimd.Int8x16 |
| func init@simd128() { |
| VL = Generic[MyInt8s@simd128](1) |
| } |
| |
| |
| // structure containing simd fields, and with simd methods |
| |
| // before |
| // A struct dependent on SIMD |
| type VectorC struct { |
| Field simd.Float32s |
| } |
| func (v *VectorC) MethodOfSimd() bool { |
| return false |
| } |
| func (v VectorC) Data() simd.Float32s { |
| return v.Field |
| } |
| func (v VectorC) Foo(x VectorC) VectorC { |
| return VectorC{Field: v.Field.Add(x.Field)} |
| } |
| |
| // dispatch |
| // technically there is none, but functions with panicking bodies |
| // remain because code must pass type checking before rewriting. |
| type VectorC struct { |
| Field simd.Float32s |
| } |
| func (v *VectorC) MethodOfSimd() bool { |
| panic(...) |
| } |
| func (v VectorC) Data() simd.Float32s { |
| panic(...) |
| } |
| func (v VectorC) Foo(x VectorC) VectorC { |
| panic(...) |
| } |
| |
| // specialized (128) |
| |
| // A struct dependent on SIMD |
| type VectorC@simd128 struct { |
| Field bridge.Float32x4 |
| } |
| func (v *VectorC@simd128) MethodOfSimd() bool { |
| return false |
| } |
| func (v VectorC@simd128) Data() bridge.Float32x4 { |
| return v.Field |
| } |
| func (v VectorC@simd128) Foo(x VectorC@simd128) VectorC@simd128 { |
| return VectorC@simd128{Field: v.Field.Add(x.Field)} |
| } |
| |
| */ |
| |
| type Rewriter struct { |
| pkg *types2.Package |
| analyzer *Analyzer |
| info *types2.Info |
| sizes []int |
| } |
| |
| func NewRewriter(pkg *types2.Package, info *types2.Info, analyzer *Analyzer, sizes []int) *Rewriter { |
| return &Rewriter{ |
| pkg: pkg, |
| info: info, |
| analyzer: analyzer, |
| sizes: sizes, |
| } |
| } |
| |
| func (r *Rewriter) Rewrite(files []*syntax.File) { |
| |
| // First duplicate and specialize all dependent functions and variables. |
| for _, fileAST := range files { |
| |
| var newDecls []syntax.Decl |
| for _, k := range r.sizes { |
| newDecls = r.generateForSize(fileAST, k, newDecls) |
| } |
| |
| // Then replace original functions with dispatchers. |
| // This also edits the DeclList of fileAST. |
| r.generateDispatchers(fileAST) |
| |
| fileAST.DeclList = append(fileAST.DeclList, newDecls...) |
| } |
| } |
| |
| func (r *Rewriter) generateDispatchers(fileAST *syntax.File) { |
| var newDecls []syntax.Decl |
| |
| change := false |
| |
| for _, decl := range fileAST.DeclList { |
| switch d := decl.(type) { |
| case *syntax.FuncDecl: |
| if d.Name == nil { |
| newDecls = append(newDecls, d) |
| continue |
| } |
| obj := r.info.Defs[d.Name] |
| if !r.analyzer.isDependentObj[obj] || r.analyzer.inSimd { |
| newDecls = append(newDecls, d) |
| continue |
| } |
| |
| sig, ok := obj.Type().(*types2.Signature) |
| if !ok { |
| newDecls = append(newDecls, d) |
| continue |
| } |
| |
| change = true |
| if r.analyzer.HasDependentSignature(sig) { |
| if base.Debug.Simd > 0 { |
| base.Warn("%s: removing body of dependent-sig original function %v", d.Pos().String(), d.Name.Value) |
| } |
| d.Body = r.blockOf(d.Pos(), r.panicStmt(d.Pos(), |
| "unexpected call of original function rewritten to specialized SIMD")) |
| newDecls = append(newDecls, d) |
| continue |
| } |
| |
| // Clean signature -> Replace body with dispatcher |
| d.Body = r.createDispatcherBody(d, sig) |
| newDecls = append(newDecls, d) |
| |
| case *syntax.VarDecl: |
| // Keep var decls even if rewritten, so that pre-rewrite code parses correctly. |
| // TODO figure out how to deal with side-effects in initializers. |
| newDecls = append(newDecls, d) |
| |
| case *syntax.TypeDecl: |
| // Keep all types; we need the untranslated copy if a method referencing it |
| // needs to typecheck pre-translation. |
| newDecls = append(newDecls, d) |
| default: |
| newDecls = append(newDecls, decl) |
| } |
| } |
| |
| if !change { |
| return |
| } |
| |
| fileAST.DeclList = newDecls |
| |
| if !r.analyzer.inSimd { |
| // Inject an import to the bridge package (if not exists) |
| hasArchSimd := false |
| var simdImport *syntax.ImportDecl |
| p := fileAST.Pos() |
| for _, decl := range fileAST.DeclList { |
| if imp, ok := decl.(*syntax.ImportDecl); ok { |
| if imp.Path.Value == `"`+archFullPkg+`"` { |
| hasArchSimd = true |
| if simdImport == nil { |
| p = imp.Pos() |
| } |
| } |
| if imp.Path.Value == `"`+simdPkg+`"` { |
| simdImport = imp |
| p = imp.Pos() |
| } |
| } |
| } |
| |
| if !hasArchSimd { |
| r.injectImport(fileAST, archFullPkg, p) |
| } |
| |
| // Ensure at least one use of "simd" |
| // var _ = simd.VectorBitLen() |
| fun := &syntax.SelectorExpr{ |
| X: syntax.NewName(p, simdPkg), // Assume this is resolvable |
| Sel: syntax.NewName(p, vectorSizeFn), |
| } |
| fun.SetPos(p) |
| call := &syntax.CallExpr{Fun: fun} |
| call.SetPos(p) |
| |
| name := syntax.NewName(p, "_") |
| |
| varDecl := &syntax.VarDecl{NameList: []*syntax.Name{name}, Values: call} |
| varDecl.SetPos(p) |
| fileAST.DeclList = append(fileAST.DeclList, varDecl) |
| } |
| } |
| |
| func (r *Rewriter) injectImport(fileAST *syntax.File, toImport string, simdImportPos syntax.Pos) { |
| importDecl := &syntax.ImportDecl{ |
| Path: &syntax.BasicLit{Value: `"` + toImport + `"`, Kind: syntax.StringLit}, |
| } |
| importDecl.Path.SetPos(simdImportPos) |
| importDecl.SetPos(simdImportPos) |
| fileAST.DeclList = append([]syntax.Decl{importDecl}, fileAST.DeclList...) |
| } |
| |
| func (r *Rewriter) createDispatcherBody(d *syntax.FuncDecl, sig *types2.Signature) *syntax.BlockStmt { |
| |
| // Build call arguments from the function parameters |
| args := func() []syntax.Expr { |
| var args []syntax.Expr |
| if d.Type.ParamList != nil { |
| for _, field := range d.Type.ParamList { |
| if field.Name != nil { |
| paramName := syntax.NewName(field.Pos(), field.Name.Value) |
| args = append(args, paramName) |
| } |
| } |
| } |
| return args |
| } |
| |
| // Slap a pos on an expression |
| pe := func(e syntax.Expr) syntax.Expr { |
| e.SetPos(d.Pos()) |
| return e |
| } |
| // Slap a pos on a statement |
| ps := func(e syntax.Stmt) syntax.Stmt { |
| e.SetPos(d.Pos()) |
| return e |
| } |
| |
| // switch ast node. |
| // the goal is something like (for now, till there are finer-grained choices) |
| // switch simd.VectorSize() { |
| // case 128: if simd.Emulated() { call the specialize-for-emulation-code(args) } |
| // else { call the specialize-for-128-code(args) } |
| // case 256: call the specialize-for-256-code(args) |
| // etc |
| // } |
| // |
| // the cases above deal with the usual `return call(...)` vs `call(...); return` |
| switchStmt := &syntax.SwitchStmt{ |
| Tag: pe(&syntax.CallExpr{ |
| Fun: pe(&syntax.SelectorExpr{ |
| X: syntax.NewName(d.Pos(), simdPkg), // Assume this is resolvable |
| Sel: syntax.NewName(d.Pos(), vectorSizeFn), |
| }), |
| }), |
| Body: []*syntax.CaseClause{}, |
| } |
| |
| var emulation syntax.Stmt |
| |
| for _, k := range r.sizes { |
| fnName := fmt.Sprintf("%s@simd%d", d.Name.Value, k) |
| fnIdent := syntax.NewName(d.Pos(), fnName) |
| |
| callExpr := pe(&syntax.CallExpr{ |
| Fun: pe(fnIdent), |
| ArgList: args(), |
| }) |
| |
| // callReturnStmt is either `return call(...)` or `call(...); return` |
| var callReturnStmt syntax.Stmt |
| if d.Type.ResultList != nil && len(d.Type.ResultList) > 0 { |
| callReturnStmt = &syntax.ReturnStmt{Results: callExpr} |
| } else { |
| callReturnStmt = &syntax.BlockStmt{ |
| List: []syntax.Stmt{ |
| ps(&syntax.ExprStmt{X: callExpr}), |
| ps(&syntax.ReturnStmt{}), |
| }, |
| Rbrace: d.Pos(), |
| } |
| } |
| callReturnStmt.SetPos(d.Pos()) |
| |
| if k == 0 { |
| // emulation == `if simd.Emulated() { callReturnStmt }` |
| // save it for the first part of the 128 case. |
| cond := pe(&syntax.CallExpr{ |
| Fun: pe(&syntax.SelectorExpr{ |
| X: syntax.NewName(d.Pos(), simdPkg), // Assume this is resolvable |
| Sel: syntax.NewName(d.Pos(), emulatedFn), |
| })}) |
| |
| blockStmt, ok := callReturnStmt.(*syntax.BlockStmt) |
| if !ok { |
| blockStmt = &syntax.BlockStmt{ |
| List: []syntax.Stmt{callReturnStmt}, |
| Rbrace: d.Pos(), |
| } |
| blockStmt.SetPos(d.Pos()) |
| } |
| |
| emulation = ps(&syntax.IfStmt{ |
| Cond: cond, |
| Then: blockStmt, |
| }) |
| continue |
| } |
| |
| var caseBody []syntax.Stmt |
| // assume that 128 is a case; when we do scalable simd, this may change. |
| // For now, if there is emulation, it is 128-bit (only). |
| if emulation != nil && k == 128 { |
| caseBody = append(caseBody, emulation) |
| emulation = nil |
| } |
| |
| caseClause := &syntax.CaseClause{ |
| Cases: pe(&syntax.BasicLit{Kind: syntax.IntLit, Value: fmt.Sprintf("%d", k)}), |
| Body: append(caseBody, callReturnStmt), |
| } |
| caseClause.SetPos(d.Pos()) |
| switchStmt.Body = append(switchStmt.Body, caseClause) |
| } |
| |
| panicStmt := r.panicStmt(d.Pos(), "unsupported vector size in simd-rewritten code") |
| return r.blockOf(d.Pos(), switchStmt, panicStmt) |
| } |
| |
| func (r *Rewriter) blockOf(p syntax.Pos, stmts ...syntax.Stmt) *syntax.BlockStmt { |
| for _, s := range stmts { |
| s.SetPos(p) |
| } |
| blockStmt := &syntax.BlockStmt{List: stmts} |
| blockStmt.SetPos(p) |
| return blockStmt |
| } |
| |
| func (r *Rewriter) panicStmt(p syntax.Pos, unquotedMessage string) *syntax.ExprStmt { |
| pe := func(e syntax.Expr) syntax.Expr { |
| e.SetPos(p) |
| return e |
| } |
| fnName := "panic" |
| fnIdent := pe(syntax.NewName(p, fnName)) |
| callExpr := pe(&syntax.CallExpr{ |
| Fun: fnIdent, |
| ArgList: []syntax.Expr{pe(&syntax.BasicLit{Value: `"` + unquotedMessage + `"`, Kind: syntax.StringLit})}, |
| }) |
| panicStmt := &syntax.ExprStmt{X: callExpr} |
| panicStmt.SetPos(p) |
| return panicStmt |
| } |
| |
| func (r *Rewriter) generateForSize(fileAST *syntax.File, k int, newDecls []syntax.Decl) []syntax.Decl { |
| copier := NewDeepCopier(r.pkg, r.info, k, r.analyzer, fmt.Sprintf("@simd%d", k)) |
| for _, decl := range fileAST.DeclList { |
| if r.shouldIncludeDecl(decl) { |
| newDecl := copier.CopyDecl(decl) |
| newDecls = append(newDecls, newDecl) |
| } |
| } |
| return newDecls |
| } |
| |
| func nameToElemBitWidth(name string) int { |
| var width int |
| switch name { |
| case "Int8s", "Uint8s", "Mask8s": |
| width = 8 |
| case "Int16s", "Uint16s", "Mask16s": |
| width = 16 |
| case "Int32s", "Uint32s", "Float32s", "Mask32s": |
| width = 32 |
| case "Int64s", "Uint64s", "Float64s", "Mask64s": |
| width = 64 |
| } |
| return width |
| } |
| |
| func (r *Rewriter) shouldIncludeDecl(decl syntax.Decl) bool { |
| // Files (and declarations) in the simd package are excluded |
| // from processing, except for those that whose name begins |
| // with "tofrom_". |
| if r.analyzer.inSimd { |
| theFile := decl.Pos().Base().Filename() |
| |
| lastSlash := strings.LastIndex(theFile, simdPkg+"/") |
| lastBackslash := strings.LastIndex(theFile, simdPkg+"\\") |
| |
| // Windows paths can be chaos, all we care, is whether the very last part |
| // of the path is any-path-separator + "tofrom_" + anything-else, given that |
| // we already know that we are in the simd package. |
| maxSlash := max(lastSlash, lastBackslash) |
| if maxSlash == -1 { |
| return false |
| } |
| if !strings.HasPrefix(theFile[maxSlash:], simdPkg+"/tofrom_") && |
| !strings.HasPrefix(theFile[maxSlash:], simdPkg+"\\tofrom_") { |
| return false |
| } |
| } |
| |
| switch d := decl.(type) { |
| case *syntax.FuncDecl: |
| if d.Name != nil { |
| return r.analyzer.isDependentObj[r.info.Defs[d.Name]] |
| } |
| case *syntax.TypeDecl: |
| return r.analyzer.isDependentObj[r.info.Defs[d.Name]] |
| case *syntax.VarDecl: |
| for _, name := range d.NameList { |
| if r.analyzer.isDependentObj[r.info.Defs[name]] { |
| return true |
| } |
| } |
| } |
| return false |
| } |
| |
| // Generate an API matching the standalone compilation call |
| func RewriteWrapper(pkg *types2.Package, info *types2.Info, files []*syntax.File) bool { |
| if !buildcfg.Experiment.SIMD { |
| return false |
| } |
| |
| switch buildcfg.GOARCH { |
| case "wasm", "amd64", "arm64": |
| default: |
| return false |
| } |
| |
| sizes := rewriteSizes() |
| if len(sizes) == 0 { |
| return false |
| } |
| analyzer := NewAnalyzer(pkg, info) |
| if !analyzer.Analyze(files) { |
| return false |
| } |
| |
| CheckPositions(files, "before midway") |
| |
| rewriter := NewRewriter(pkg, info, analyzer, sizes) |
| rewriter.Rewrite(files) |
| |
| CheckPositions(files, "after midway") |
| |
| return true |
| } |