feat(asm): the operand forms and defines GOROOT writes

Assisted-by: GLM 5.3 Flash

CHANGELOG.md

@@ -7,6 +7,22 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [development]
 
+### Fixed
+
+- **The operand forms GOROOT writes.**  Numeric PC-relative jumps
+  (`JEQ 2(PC)`, the park loop `JMP 0(PC)`) resolve with the toolchain's
+  own instruction counting and fold jump-to-jump chains exactly as its
+  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
+  asm` predefines the `GOARCH_<arch>` and `GOOS_<goos>` macros the go
+  command passes to `go tool asm`, so GOROOT headers' `#ifdef
+  GOARCH_amd64` platform blocks (`go_tls.h`'s `get_tls` and friends)
+  select as intended.  The GOROOT corpus measure moves to 261 of 353
+  files assembling for every target architecture (73.9 %), 87.5 % of
+  the real-code corpus, from 70.8 % and 82.2 %.
+
 ### Added
 
 - **Per-architecture reference pages.**  [docs/asm/](docs/asm/README.md)

asm/assemble.go

@@ -86,6 +86,10 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	// outgrows the short form.
 	long := make([]bool, len(t.Body))
 	sizes := make([]int, len(t.Body))
+	numTargets := make([]int, len(t.Body))
+	for i := range numTargets {
+		numTargets[i] = -1
+	}
 	offsets := map[string]int{}
 	pcs := make([]int, len(t.Body))
 	var guardJBlong, guardJBElong, moreJMPlong bool
@@ -94,6 +98,10 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	for {
 		guard := fi.guardLen(guardJBlong, guardJBElong)
 		pos := guard + len(fi.prologue)
+		for i := range numTargets {
+			numTargets[i] = -1
+		}
+		idxAtPc := map[int]int{}
 		for i, stmt := range t.Body {
 			switch s := stmt.(type) {
 			case *ast.Label:
@@ -105,12 +113,78 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 				}
 				sizes[i] = sz
 				pcs[i] = pos
+				idxAtPc[pos] = i
 				pos += sz
 			}
 		}
 		bodyLen := pos - (guard + len(fi.prologue))
 		// Expand any short jump whose displacement no longer fits rel8.
 		changed := false
+		// Numeric ±N(PC) jumps resolve against this iteration's layout; the
+		// emission pass reads the same table after the loop converges.  A
+		// target that is itself an unconditional local JMP is chased to the
+		// ultimate target: the toolchain's brloop pass collapses branch-to-
+		// branch chains before it encodes, so matching its bytes requires
+		// the same redirection.
+		for i := range numTargets {
+			numTargets[i] = -1
+		}
+		for i, stmt := range t.Body {
+			s, ok := stmt.(*ast.Instr)
+			if !ok {
+				continue
+			}
+			if len(s.Operands) == 1 {
+				if n, isNum := pcJumpOffset(s.Operands[0]); isNum {
+					if target, okT := pcJumpTarget(t, i, n, pcs); okT {
+						numTargets[i] = target
+					}
+				}
+			}
+		}
+		for i := range numTargets {
+			if numTargets[i] < 0 {
+				continue
+			}
+			tgt := numTargets[i]
+			for hop := 0; hop < len(t.Body); hop++ {
+				idx, ok := idxAtPc[tgt]
+				if !ok {
+					break
+				}
+				in, ok := t.Body[idx].(*ast.Instr)
+				if !ok || strings.ToUpper(in.Mnemonic.Text) != "JMP" || len(in.Operands) != 1 {
+					break
+				}
+				if name, isLabel := labelName(in.Operands[0]); isLabel {
+					tgt = offsets[resolve(name)]
+					continue
+				}
+				if n, isNum := pcJumpOffset(in.Operands[0]); isNum {
+					next, okT := pcJumpTarget(t, idx, n, pcs)
+					if !okT {
+						break
+					}
+					tgt = next
+					continue
+				}
+				break // JMP through a register or memory: the chain ends
+			}
+			numTargets[i] = tgt
+		}
+		for i, stmt := range t.Body {
+			s, ok := stmt.(*ast.Instr)
+			if !ok {
+				continue
+			}
+			if numTargets[i] >= 0 && !long[i] {
+				rel := int64(numTargets[i] - (pcs[i] + jumpSize(strings.ToUpper(s.Mnemonic.Text), false)))
+				if !fits8(rel) {
+					long[i] = true
+					changed = true
+				}
+			}
+		}
 		for i, stmt := range t.Body {
 			s, ok := stmt.(*ast.Instr)
 			if !ok {
@@ -232,7 +306,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 				spadjStep{pos + epi, 0},
 			)
 		}
-		code, ps, pool, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
+		code, ps, pool, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link, numTargets[i])
 		if err != nil {
 			return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 		}
@@ -428,6 +502,37 @@ func computeFrame(t *ast.Text) frameInfo {
 	return fi
 }
 
+// pcJumpOffset recognises the numeric relative jump operand ±N(PC) and
+// returns N: the toolchain counts instructions, not bytes, so +2(PC) targets
+// the second instruction boundary after the branch.
+func pcJumpOffset(op *ast.Operand) (int, bool) {
+	if op.Kind != ast.OpAddr || op.Addr.Base != "PC" {
+		return 0, false
+	}
+	return int(op.Addr.Offset), true
+}
+
+// pcJumpTarget resolves a numeric jump at statement index j: N counts the
+// instruction statements after the jump itself (N = 0 is the jump's own
+// address, the classic park loop), and the target is the start of the Nth
+// one.  It reports false when the count runs past the end of the function.
+func pcJumpTarget(t *ast.Text, j, n int, pcs []int) (int, bool) {
+	if n == 0 {
+		return pcs[j], true
+	}
+	seen := 0
+	for k := j + 1; k < len(t.Body); k++ {
+		if _, ok := t.Body[k].(*ast.Instr); !ok {
+			continue
+		}
+		seen++
+		if seen == n {
+			return pcs[k], true
+		}
+	}
+	return 0, false
+}
+
 // hasCall reports whether the function body contains a CALL instruction.
 func hasCall(t *ast.Text) bool {
 	for _, stmt := range t.Body {
@@ -602,7 +707,7 @@ func instrSize(s *ast.Instr, fi frameInfo, long bool, link *linkInfo) (int, erro
 		}
 		return jumpSize(mnem, long), nil
 	}
-	code, _, _, err := encodeInstr(s, 0, nil, fi, false, nil, link)
+	code, _, _, err := encodeInstr(s, 0, nil, fi, false, nil, link, -1)
 	if err != nil {
 		return 0, err
 	}
@@ -637,7 +742,7 @@ func jumpSize(mnem string, long bool) int {
 // (relative to pc, the instruction's own offset).  A RET in a frame-pointer
 // function is prefixed with the epilogue.  resolve, when non-nil, redirects a
 // jump label through the jump-to-jump chain before the offset lookup.
-func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string, link *linkInfo) ([]byte, []sbPatch, []floatPoolEntry, error) {
+func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string, link *linkInfo, numTarget int) ([]byte, []sbPatch, []floatPoolEntry, error) {
 	mnem := strings.ToUpper(s.Mnemonic.Text)
 
 	var prefix []byte
@@ -677,7 +782,7 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 			}
 			return append(prefix, code...), nil, nil, nil
 		}
-		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve)
+		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve, numTarget)
 	} else {
 		code, ps, pool, err = encodeNormal(s, fi, link)
 	}
@@ -703,6 +808,32 @@ func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch
 		}
 		return code, nil, nil, nil
 	}
+	// MOVQ $sym±off(SB), r64: the toolchain assembles a symbol immediate as
+	// LEAQ disp32(RIP), r64 with an R_PCREL relocation at the disp32 field,
+	// never as a 64-bit absolute immediate (verified against go tool asm).
+	// MOVD is the MOVQ alias; the narrower widths reject the form outright.
+	if (mnemUpper == "MOVQ" || mnemUpper == "MOVD") && len(s.Operands) == 2 &&
+		s.Operands[0].Kind == ast.OpImmediate && s.Operands[0].Imm.Sym != nil &&
+		s.Operands[0].Imm.Sym.Pseudo == "SB" {
+		mem := &ast.Operand{Kind: ast.OpAddr, Addr: ast.Address{Sym: s.Operands[0].Imm.Sym}}
+		src, err := operandFromAST(mnemUpper, mem, 8, fi, link)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		dst, err := operandFromAST(mnemUpper, s.Operands[1], 8, fi, link)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		e := &enc{}
+		if err := e.encodeLea([]Operand{src, dst}, 8); 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}
+		}
+		return e.out, ps, nil, nil
+	}
 	_, size := splitSize(mnemUpper)
 	if size == 0 {
 		size = 8
@@ -753,21 +884,29 @@ func encodeBookkeeping(upper string, s *ast.Instr) ([]byte, error) {
 }
 
 // encodeJump encodes a JMP/CALL/Jcc with a relative offset resolved from the
-// target label, in the short (rel8) or long (rel32) form.
-func encodeJump(s *ast.Instr, mnem string, pc int, offsets map[string]int, long bool, resolve func(string) string) ([]byte, error) {
+// target label or from a numeric ±N(PC) instruction count, in the short
+// (rel8) or long (rel32) form.  numTarget is the resolved byte offset of a
+// numeric operand, negative when the operand is not one.
+func encodeJump(s *ast.Instr, mnem string, pc int, offsets map[string]int, long bool, resolve func(string) string, numTarget int) ([]byte, error) {
 	if len(s.Operands) != 1 {
 		return nil, fmt.Errorf("jump expects 1 operand, got %d", len(s.Operands))
 	}
-	name, ok := labelName(s.Operands[0])
-	if !ok {
+	name, isLabel := labelName(s.Operands[0])
+	if !isLabel && numTarget < 0 {
 		return nil, fmt.Errorf("jump target must be a local label")
 	}
-	if resolve != nil && mnem != "CALL" {
-		name = resolve(name)
-	}
-	target, ok := offsets[name]
-	if !ok {
-		return nil, fmt.Errorf("undefined label %q", name)
+	var target int
+	if isLabel {
+		if resolve != nil && mnem != "CALL" {
+			name = resolve(name)
+		}
+		t, ok := offsets[name]
+		if !ok {
+			return nil, fmt.Errorf("undefined label %q", name)
+		}
+		target = t
+	} else {
+		target = numTarget
 	}
 	rel := int64(target - (pc + jumpSize(mnem, long)))
 
@@ -841,6 +980,11 @@ func indirectJumpTarget(s *ast.Instr) bool {
 		return false
 	}
 	a := s.Operands[0].Addr
+	// ±N(PC) is the numeric relative form, the PC counts instructions from
+	// the branch: relative, not indirect.
+	if a.Base == "PC" || a.Index == "PC" {
+		return false
+	}
 	if a.Base != "" || a.Index != "" {
 		return true
 	}

asm/kernels_differential_test.go

@@ -130,6 +130,7 @@ func TestDifferentialKernels(t *testing.T) {
 		{filepath.Join("..", "testdata", "verify", "quadreg_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "floatimm_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "bookkeep_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "forms_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "datarel_arm64.s"), "arm64", true},
 		{filepath.Join("..", "testdata", "verify", "divslash_arm64.s"), "arm64", true},
 	} {

asm/link.go

@@ -166,6 +166,13 @@ func AssembleFile(f *ast.File) (*Image, error) {
 	for _, d := range dataSyms {
 		known[d.name] = true
 	}
+	// TEXT symbols are file-level definitions too: a symbol immediate
+	// ($fn(SB)) may name one, exactly as a data reference names a GLOBL.
+	for _, d := range f.Decls {
+		if t, ok := d.(*ast.Text); ok {
+			known[t.Name.Name] = true
+		}
+	}
 	link := &linkInfo{symbols: known, allowExternal: true}
 	poolSeen := map[string]bool{}
 

cmd/gasm/audit.go

@@ -664,6 +664,7 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 				f, errs := parser.ParseWithOptions(path, src, parser.Options{
 					Expand:      true,
 					IncludeDirs: append(slices.Clone(dirs), hdrDir),
+					Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
 				})
 				if len(errs) > 0 {
 					ok = false
@@ -683,12 +684,19 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 			continue
 		}
 
-		f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
-
 		ok := true
 		for _, i := range wanted {
 			tg, t := targets[i], tallies[i]
 			t.attempted++
+			// The parse carries the target's platform predefines, so it
+			// cannot be shared across targets the way a header-free file's
+			// could: a #ifdef GOARCH_arm block must be live on arm64 and
+			// dead everywhere else.
+			f, errs := parser.ParseWithOptions(path, src, parser.Options{
+				Expand:      true,
+				IncludeDirs: dirs,
+				Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
+			})
 			var err error
 			if len(errs) > 0 {
 				err = errs[0] // a parse failure is a failure for every target

cmd/gasm/main.go

@@ -574,7 +574,7 @@ naming the package.
 		defer cleanup()
 		dirs = append(dirs, hdrDir)
 	}
-	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs, Predefines: platformPredefinesFor(string(targetArch), goos)})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
@@ -789,6 +789,29 @@ e.g. --map wideCopyAVX2=wideCopyAVX512 pairs the two regardless of suffix.
 	return 1
 }
 
+// platformPredefines mirrors the go command's assembler invocation, which
+// defines GOOS_<goos> and GOARCH_<arch> as -D macros: GOROOT headers
+// (go_tls.h, asm_riscv64.h) select their platform blocks with #ifdef on
+// exactly those names, so an assembler without them cannot see the platform
+// definitions at all.
+func platformPredefines(goarch, goos string) map[string]string {
+	return map[string]string{
+		"GOARCH_" + goarch: "1",
+		"GOOS_" + goos:     "1",
+	}
+}
+
+// platformPredefinesFor resolves the ambient GOOS the way a build would: a
+// file whose name carries one (sys_darwin_arm64.s) is compiled for that GOOS
+// and nothing else.
+func platformPredefinesFor(goarch string, fileGoos string) map[string]string {
+	goos := fileGoos
+	if goos == "" {
+		goos = runtime.GOOS
+	}
+	return platformPredefines(goarch, goos)
+}
+
 // assembleFile assembles a parsed file for the given architecture and returns the image.
 func assembleFile(targetArch arch.Arch, f *ast.File) (*asm.Image, error) {
 	switch targetArch {
@@ -813,17 +836,17 @@ func assemblePath(path string, forced arch.Arch, dirs includeDirs) (*asm.Image,
 	if err != nil {
 		return nil, err
 	}
-	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
+	target := forced
+	if target == arch.Unknown {
+		target = arch.FromFilename(path)
+	}
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs, Predefines: platformPredefinesFor(string(target), "")})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
 	if len(errs) > 0 {
 		return nil, fmt.Errorf("parse errors")
 	}
-	target := forced
-	if target == arch.Unknown {
-		target = arch.FromFilename(path)
-	}
 	return assembleFile(target, f)
 }
 

parser/parser.go

@@ -489,6 +489,17 @@ func parseImmediate(g []token.Token) ast.Immediate {
 	} else if g[i].Kind == token.Plus {
 		i++
 	}
+	// A constant expression after the sign: $-(R - 8), $+(32-shift).  The
+	// toolchain folds the negated value in place (the cgo ABI macros write
+	// ADJSP $-(REGS_HOST_TO_ABI0_STACK - 8)), so the sign applies to the
+	// folded value exactly as it does to a bare literal.
+	if i < len(g) && (g[i].Kind == token.LParen || g[i].Kind == token.Tilde) {
+		if v, rest, ok := foldExpr(g[i:]); ok && len(rest) == 0 {
+			imm.Val = v
+			imm.HasVal = true
+			return imm
+		}
+	}
 	if i < len(g) && g[i].Kind == token.Number {
 		text := g[i].Text
 		if v, ok := tryInt(text); ok {

parser/preproc.go

@@ -34,6 +34,12 @@ type Options struct {
 	// Expand enables macro expansion, include splicing and the
 	// statement-separator reading of ';' that the expanded bodies rely on.
 	Expand bool
+	// Predefines names the macros defined before the file is read.  The
+	// go command drives go tool asm with -D GOOS_<goos> -D GOARCH_<arch>,
+	// and GOROOT's own headers (go_tls.h, asm_riscv64.h) select their
+	// platform blocks with #ifdef on exactly those names, so an assembler
+	// without them cannot see the platform definitions at all.
+	Predefines map[string]string
 }
 
 // ParseWithOptions parses src like Parse, optionally preprocessing it first.
@@ -44,6 +50,9 @@ func ParseWithOptions(path, src string, opts Options) (*ast.File, []error) {
 	var errs []error
 	if opts.Expand {
 		pp := &preproc{opts: opts, macros: map[string]*macroDef{}}
+		for name, value := range opts.Predefines {
+			pp.macros[name] = &macroDef{name: name, body: lexer.Tokenize(value)}
+		}
 		lines = pp.fileLines(path, tokens, token.Position{})
 		errs = pp.errs
 	} else {

testdata/verify/forms_amd64.s

@@ -0,0 +1,36 @@
+// Kernel: the operand forms the GOROOT campaign surfaced — numeric
+// PC-relative jumps, symbol-immediate materialisation (the toolchain rewrites
+// MOVQ $sym(SB) into a RIP-relative LEA) and the negated constant-expression
+// ADJSP the cgo ABI macros write.  Bytes are pinned against go tool asm by
+// TestDifferentialKernels.
+#include "textflag.h"
+
+DATA sd<>(SB)/4, $7
+GLOBL sd<>(SB), RODATA, $4
+
+// func Jumps(flag int64) int64
+TEXT ·Jumps(SB), NOSPLIT, $0-16
+	MOVQ	flag+0(FP), AX
+	TESTQ	AX, AX
+	JEQ	2(PC)
+	MOVQ	$1, AX
+	JMP	3(PC)
+	MOVQ	$2, AX
+	MOVQ	AX, ret+0(FP)
+	RET
+
+// func SymImm() int64
+TEXT ·SymImm(SB), NOSPLIT, $0-16
+	MOVQ	$sd<>(SB), AX
+	MOVQ	$·SymImm(SB), CX
+	MOVQ	AX, ret+0(FP)
+	RET
+
+// func Frame()
+TEXT ·Frame(SB), NOSPLIT, $0
+	PUSHFQ
+	CLD
+	ADJSP	$(64 - 8)
+	ADJSP	$-(64 - 8)
+	POPFQ
+	RET