feat(asm): the TLS access pair in the toolchain's one-instruction form

Assisted-by: GLM 5.3 Flash

CHANGELOG.md

@@ -15,13 +15,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   branch optimiser does; symbol immediates (`MOVQ $sym(SB), AX`)
   assemble to the toolchain's RIP-relative LEA with an R_PCREL
   relocation; negated constant expressions in operands (`ADJSP
-  $-(REGS - 8)`, the shape the cgo ABI macros write) fold; and `gasm
+  $-(REGS - 8)`, the shape the cgo ABI macros write) fold; the TLS
-  asm` predefines the `GOARCH_<arch>` and `GOOS_<goos>` macros the go
+  access pair assembles as the toolchain's one-instruction form (the
-  command passes to `go tool asm`, so GOROOT headers' `#ifdef
+  bare `MOVQ TLS, r` load nops out and `off(r)(TLS*1)` folds to the
-  GOARCH_amd64` platform blocks (`go_tls.h`'s `get_tls` and friends)
+  segment-prefixed absolute whose disp32 carries the R_TLSLE
-  select as intended.  The GOROOT corpus measure moves to 261 of 353
+  relocation, per-GOOS); and `gasm asm` predefines the `GOARCH_<arch>`
-  files assembling for every target architecture (73.9 %), 87.5 % of
+  and `GOOS_<goos>` macros the go command passes to `go tool asm`, so
-  the real-code corpus, from 70.8 % and 82.2 %.
+  GOROOT headers' `#ifdef GOARCH_amd64` platform blocks (`go_tls.h`'s
+  `get_tls` and friends) select as intended.  The GOROOT corpus measure
+  moves to 267 of 353 files assembling for every target architecture
+  (75.6 %), 89.4 % of the real-code corpus, from 70.8 % and 82.2 %.
 
 ### Added
 

asm/assemble.go

@@ -38,10 +38,25 @@ func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
 // rejects SB operands outright (single-function assembly cannot resolve
 // them).  When allowExternal is set, a reference to a symbol no GLOBL in the
 // file defines is recorded as an external relocation instead of failing
-// the object-file emitters resolve it at link time.
+// the object-file emitters resolve it at link time.  goos selects the TLS
+// access form: the empty default behaves as linux.
 type linkInfo struct {
 	symbols       map[string]bool
 	allowExternal bool
+	goos          string
+}
+
+// tlsOneInsn reports the one-instruction TLS form, obj6.go's
+// CanUse1InsnTLS for the GOOS gasm supports: the bare TLS load nops out and
+// the (TLS*1) index folds to a segment-absolute access.  Windows and plan9
+// keep the two-instruction form; shared linux does too, which gasm's raw
+// path does not model and therefore does not select.
+func (l *linkInfo) tlsOneInsn() bool {
+	switch l.goos {
+	case "", "linux", "freebsd":
+		return true
+	}
+	return false
 }
 
 // sbPatch is a function-relative static-symbol relocation: the disp32 field
@@ -834,6 +849,15 @@ func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch
 		}
 		return e.out, ps, nil, nil
 	}
+	// MOVQ/MOVL TLS, r: the bare TLS load.  The toolchain's progedit nops
+	// it out on the one-instruction TLS systems (linux and freebsd, not
+	// shared) and encodes the segment-prefixed load elsewhere; get_tls(r),
+	// the macro GOROOT's go_tls.h defines, expands to exactly this
+	// statement, and the toolchain's pairing pass removes it whenever the
+	// following instruction's (TLS*1) index folds.
+	if (mnemUpper == "MOVQ" || mnemUpper == "MOVL") && len(s.Operands) == 2 && isBareTLS(s.Operands[0]) {
+		return encodeTLSBaseLoad(s, fi, link)
+	}
 	_, size := splitSize(mnemUpper)
 	if size == 0 {
 		size = 8
@@ -853,10 +877,67 @@ func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch
 	ps := make([]sbPatch, len(e.patches))
 	for i, p := range e.patches {
 		ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend}
+		if p.tls {
+			ps[i].kind = RelTLSLE
+		}
 	}
 	return e.out, ps, e.floatPoolList(), nil
 }
 
+// isBareTLS reports whether the operand is the bare TLS pseudo-register
+// load source, the expansion of go_tls.h's get_tls(r) macro.
+func isBareTLS(op *ast.Operand) bool {
+	return op.Kind == ast.OpAddr && op.Addr.Sym != nil &&
+		op.Addr.Sym.Pseudo == "" && op.Addr.Sym.Name == "TLS" &&
+		op.Addr.Base == "" && op.Addr.Index == ""
+}
+
+// encodeTLSBaseLoad assembles MOVQ/MOVL TLS, r.  On the one-instruction TLS
+// systems (linux and freebsd outside -shared, obj6.go's CanUse1InsnTLS) the
+// statement nops out: the following (TLS*1) access folds to a direct
+// segment-absolute load.  The two-instruction systems keep the segment load,
+// nine bytes with the R_TLSLE patch site at the disp32.
+func encodeTLSBaseLoad(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, []floatPoolEntry, error) {
+	_, size := splitSize(strings.ToUpper(s.Mnemonic.Text))
+	if size == 0 {
+		size = 8
+	}
+	dst, err := operandFromAST("MOVQ", s.Operands[1], 8, fi, link)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	reg, ok := dst.(Reg)
+	if !ok || reg.isVec() {
+		return nil, nil, nil, fmt.Errorf("TLS: destination must be a general register")
+	}
+	if link == nil || link.tlsOneInsn() {
+		return nil, nil, nil, nil // noped out
+	}
+	seg := byte(0x64) // FS
+	if link.goos == "windows" {
+		seg = 0x65 // GS
+	}
+	e := &enc{}
+	i := &instr{
+		prefix: seg,
+		rexW:   size == 8,
+		rexR:   reg.idx >= 8,
+		opcode: []byte{0x8B},
+		modrm:  0x04 | (reg.idx&7)<<3,
+		sib:    0x25,
+		disp:   le32(0),
+		tls:    true,
+	}
+	if err := e.emit(i); err != nil {
+		return nil, nil, nil, err
+	}
+	ps := make([]sbPatch, len(e.patches))
+	for i, p := range e.patches {
+		ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend, kind: RelTLSLE}
+	}
+	return e.out, ps, nil, nil
+}
+
 // encodeBookkeeping accepts-and-ignores FUNCDATA and PCDATA at the statement
 // level, before operand conversion: the toolchain's shapes are FUNCDATA
 // $n, sym(SB) and PCDATA $n, $m, and neither contributes a byte to the
@@ -1108,6 +1189,19 @@ func operandFromAST(mnemUpper string, op *ast.Operand, size int, fi frameInfo, l
 			}
 			m := Mem{Base: base, Disp: a.Offset, HasBase: true, Size: size}
 			if a.Index != "" {
+				if a.Index == "TLS" {
+					// off(base)(TLS*1): the thread-local annotation.  The
+					// one-instruction TLS form folds it to off(TLS), the
+					// segment-prefixed absolute whose disp32 carries an
+					// R_TLS_LE patch site; the base register disappears
+					// from the encoding, exactly as the toolchain's
+					// progedit rewrites the address.
+					seg := byte(0x64) // FS on linux, freebsd, plan9
+					if link != nil && link.goos == "windows" {
+						seg = 0x65 // GS
+					}
+					return TLSMem{Disp: a.Offset, Size: size, Seg: seg}, nil
+				}
 				idx, ok := ParseReg(a.Index)
 				if !ok {
 					return nil, fmt.Errorf("unknown index register %q", a.Index)

asm/encode.go

@@ -64,6 +64,7 @@ type encPatch struct {
 	off    int
 	name   string
 	addend int64
+	tls    bool // a TLS slot offset: the patch is R_TLSLE with no symbol
 }
 
 func (e *enc) encode(mnem string, ops []Operand) error {
@@ -578,6 +579,7 @@ type instr struct {
 	disp      []byte
 	imm       []byte
 	sb        *sbRef // static-symbol displacement in disp, awaiting resolution
+	tls       bool   // the displacement is a TLS slot offset, patched R_TLSLE
 }
 
 // sbRef records that an instruction's displacement refers to a static symbol
@@ -620,6 +622,9 @@ func (e *enc) emit(i *instr) error {
 	if i.sb != nil {
 		e.patches = append(e.patches, encPatch{off: len(e.out), name: i.sb.name, addend: i.sb.addend})
 	}
+	if i.tls {
+		e.patches = append(e.patches, encPatch{off: len(e.out), tls: true})
+	}
 	e.out = append(e.out, i.disp...)
 	e.out = append(e.out, i.imm...)
 	return nil
@@ -674,6 +679,16 @@ func setRMReg(i *instr, regField int, rexR, regForced bool, rm Operand, opSize i
 		i.disp = le32(0)
 		i.sb = &sbRef{name: r.name, addend: r.addend}
 		return nil
+	case TLSMem:
+		// off(TLS): the segment-prefixed absolute access, mod=00 with the
+		// SIB escape's disp32 absolute form.  The displacement is the TLS
+		// slot offset, patched by the linker's TLS relocation.
+		i.prefix = r.Seg
+		i.modrm = 0x04 | regField<<3
+		i.sib = 0x25
+		i.disp = le32(r.Disp)
+		i.tls = true
+		return nil
 	default:
 		return fmt.Errorf("invalid r/m operand %T", rm)
 	}

asm/guard_test.go

@@ -307,12 +307,12 @@ func TestStackGuardBytesLOONG64(t *testing.T) {
 func TestStackGuardGOObjInternalCall(t *testing.T) {
 	for _, tt := range []struct {
 		src      string
-		assemble func(*ast.File) (*Image, error)
+		assemble func(*ast.File, ...AssembleOption) (*Image, error)
 	}{
 		{"g_amd64.s", AssembleFile},
-		{"g_arm64.s", AssembleFileARM64},
+		{"g_arm64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileARM64(f) }},
-		{"g_riscv64.s", AssembleFileRISCV},
+		{"g_riscv64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileRISCV(f) }},
-		{"g_loong64.s", AssembleFileLOONG64},
+		{"g_loong64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileLOONG64(f) }},
 	} {
 		f, errs := parser.Parse(tt.src, "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
 		if len(errs) > 0 {

asm/instrs.go

@@ -192,6 +192,19 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 		}
 		return e.emit(i)
 
+	case TLSMem:
+		if !dstIsReg {
+			return fmt.Errorf("MOV: two memory operands")
+		}
+		// MOV r, off(TLS): the segment-prefixed absolute load, reg=dst,
+		// rm=src(tlsMem) through the SIB escape; the disp32 is the TLS slot
+		// offset with its R_TLSLE patch site.
+		i := newInstr(size, []byte{movRR(size)})
+		if err := setRM(i, dstReg, src, size); err != nil {
+			return err
+		}
+		return e.emit(i)
+
 	case Imm:
 		if dstIsReg {
 			v := int64(src)

asm/link.go

@@ -150,6 +150,18 @@ func (img *Image) Bytes() []byte {
 	return append(out, img.Data...)
 }
 
+// AssembleOption adjusts the file-level assembly context.
+type AssembleOption func(*linkInfo)
+
+// WithGOOS selects the target operating system for the forms that depend on
+// it, the TLS access shape above all: linux and freebsd take the
+// one-instruction form, windows and plan9 keep the two-instruction load.
+func WithGOOS(goos string) AssembleOption {
+	return func(l *linkInfo) {
+		l.goos = goos
+	}
+}
+
 // AssembleFile assembles every TEXT function of a parsed file and lays out
 // its static symbols (GLOBL/DATA) in a data section behind the code.  Each
 // reference to a file-local static symbol becomes a RIP-relative load whose
@@ -157,7 +169,7 @@ func (img *Image) Bytes() []byte {
 // GLOBL defines is recorded as an external relocation (Externals) with its
 // displacement left zero, the object-file emitters resolve it at link
 // time, while the raw image (Bytes) cannot represent it.
-func AssembleFile(f *ast.File) (*Image, error) {
+func AssembleFile(f *ast.File, opts ...AssembleOption) (*Image, error) {
 	dataSyms, err := collectData(f)
 	if err != nil {
 		return nil, err
@@ -174,6 +186,9 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		}
 	}
 	link := &linkInfo{symbols: known, allowExternal: true}
+	for _, o := range opts {
+		o(link)
+	}
 	poolSeen := map[string]bool{}
 
 	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}

asm/operand.go

@@ -36,6 +36,18 @@ type FloatImm struct {
 
 func (FloatImm) isOperand() {}
 
+// TLSMem is a thread-local access, the source form off(base)(TLS*1) with the
+// base dropped: the toolchain's one-instruction TLS rewrite assembles it as
+// the segment-prefixed absolute whose disp32 carries an R_TLS_LE patch site
+// (the linker fills the TLS slot offset).
+type TLSMem struct {
+	Disp int64
+	Size int
+	Seg  byte // the segment override: FS (0x64) or GS (0x65) on windows
+}
+
+func (TLSMem) isOperand() {}
+
 // Mem is a memory operand of the form disp(base)(index*scale).
 type Mem struct {
 	Base     Reg

cmd/gasm/audit.go

@@ -671,7 +671,7 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 					t.fail(path, errs[0])
 					continue
 				}
-				if _, err := assembleFile(tg.a, f); err != nil {
+				if _, err := assembleFile(tg.a, f, goos); err != nil {
 					ok = false
 					t.fail(path, err)
 					continue
@@ -701,7 +701,7 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 			if len(errs) > 0 {
 				err = errs[0] // a parse failure is a failure for every target
 			} else {
-				_, err = assembleFile(tg.a, f)
+				_, err = assembleFile(tg.a, f, goos)
 			}
 			if err != nil {
 				ok = false

cmd/gasm/dis.go

@@ -84,7 +84,7 @@ func disSource(path string, target arch.Arch) int {
 	if len(errs) > 0 {
 		return 1
 	}
-	img, err := assembleFile(target, f)
+	img, err := assembleFile(target, f, "")
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm dis: %v\n", err)
 		return 1

cmd/gasm/main.go

@@ -582,7 +582,7 @@ naming the package.
 		return 1
 	}
 
-	img, err := assembleFile(targetArch, f)
+	img, err := assembleFile(targetArch, f, goos)
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, err)
 		return 1
@@ -813,10 +813,10 @@ func platformPredefinesFor(goarch string, fileGoos string) map[string]string {
 }
 
 // assembleFile assembles a parsed file for the given architecture and returns the image.
-func assembleFile(targetArch arch.Arch, f *ast.File) (*asm.Image, error) {
+func assembleFile(targetArch arch.Arch, f *ast.File, goos string) (*asm.Image, error) {
 	switch targetArch {
 	case arch.AMD64:
-		return asm.AssembleFile(f)
+		return asm.AssembleFile(f, asm.WithGOOS(goos))
 	case arch.RISCV:
 		return asm.AssembleFileRISCV(f)
 	case arch.ARM64:
@@ -847,7 +847,7 @@ func assemblePath(path string, forced arch.Arch, dirs includeDirs) (*asm.Image,
 	if len(errs) > 0 {
 		return nil, fmt.Errorf("parse errors")
 	}
-	return assembleFile(target, f)
+	return assembleFile(target, f, "")
 }
 
 // printByteDiff shows the first few byte differences between two code blocks.
@@ -943,7 +943,7 @@ func cmdVerifyNonJIT(path string, targetArch arch.Arch, groundTruth, profile boo
 	if len(errs) > 0 {
 		return 1
 	}
-	img, err := assembleFile(targetArch, f)
+	img, err := assembleFile(targetArch, f, "")
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
 		return 1

testdata/verify/forms_amd64.s

@@ -34,3 +34,10 @@ TEXT ·Frame(SB), NOSPLIT, $0
 	ADJSP	$-(64 - 8)
 	POPFQ
 	RET
+
+// func Tls() int64
+TEXT ·Tls(SB), NOSPLIT, $0-8
+	MOVQ	TLS, BX
+	MOVQ	0(BX)(TLS*1), AX
+	MOVQ	AX, ret+0(FP)
+	RET