feat(parser): read the TEXT and GLOBL operands the toolchain counts them

Assisted-by: GLM 5.3 Flash

parser/flags.go

@@ -62,23 +62,14 @@ var flagOrder = []struct {
 	{"ABIWRAPPER", 4096},
 }
 
-// flagsRun returns the flags operand of a TEXT or GLOBL directive: the
+// flagsRun prepares one flags operand for evaluation: the comments are
-// tokens after the symbol up to the frame operand's '$', with comments and
+// dropped, they are layout the expression never sees.
-// the one trailing comma that separated the operand from the '$' removed.
 func flagsRun(g []token.Token) []token.Token {
-	n := 0
+	out := make([]token.Token, 0, len(g))
-	for n < len(g) && g[n].Kind != token.Dollar {
+	for _, t := range g {
-		n++
+		if t.Kind != token.Comment {
-	}
-	out := make([]token.Token, 0, n)
-	for _, t := range g[:n] {
-		if t.Kind == token.Comment {
-			continue
-		}
 			out = append(out, t)
 		}
-	if len(out) > 0 && out[len(out)-1].Kind == token.Comma {
-		out = out[:len(out)-1]
 	}
 	return out
 }

parser/flags_test.go

@@ -0,0 +1,225 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package parser
+
+import (
+	"slices"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+)
+
+// parseTextHeader parses the one TEXT directive of src, raw (no expansion).
+func parseTextHeader(t *testing.T, header string) *ast.Text {
+	t.Helper()
+	f, errs := Parse("t.s", header+"\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse %q: %v", header, errs)
+	}
+	return f.Decls[0].(*ast.Text)
+}
+
+// parseTextHeaderExpanded parses the one TEXT directive of src on the
+// assembly path, where the toolchain's rejections apply.
+func parseTextHeaderExpanded(t *testing.T, header string) (*ast.Text, []error) {
+	t.Helper()
+	f, errs := ParseWithOptions("t.s", header+"\n\tRET\n", Options{Expand: true})
+	if len(f.Decls) != 1 {
+		t.Fatalf("parse %q: %d decls, want 1", header, len(f.Decls))
+	}
+	return f.Decls[0].(*ast.Text), errs
+}
+
+// TestTextFlagsExpression covers the flags operand as one expression: names
+// joined by '|', the legacy numeric spellings, and the arithmetic forms the
+// toolchain's evalInteger folds.  Names written purely keep their written
+// order and duplicates; an operand with literals in it expands to the
+// canonical ascending name list.
+func TestTextFlagsExpression(t *testing.T) {
+	for _, c := range []struct {
+		operand  string
+		wantFlag []string
+		wantVal  int64
+	}{
+		{"NOSPLIT", []string{"NOSPLIT"}, 4},
+		{"NOSPLIT|NOFRAME|DUPOK", []string{"NOSPLIT", "NOFRAME", "DUPOK"}, 518},
+		{"TOPFRAME|NOSPLIT", []string{"TOPFRAME", "NOSPLIT"}, 2052},
+		{"NOSPLIT|NOSPLIT", []string{"NOSPLIT", "NOSPLIT"}, 4},
+		{"4", []string{"NOSPLIT"}, 4},
+		{"512", []string{"NOFRAME"}, 512},
+		{"2|4", []string{"DUPOK", "NOSPLIT"}, 6},
+		{"(NOSPLIT|NOFRAME)", []string{"NOSPLIT", "NOFRAME"}, 516},
+		{"NOSPLIT|4", []string{"NOSPLIT"}, 4},
+		{"NOSPLIT&DUPOK", nil, 0},
+		{"$NOSPLIT", []string{"NOSPLIT"}, 4},
+	} {
+		txt := parseTextHeader(t, "TEXT \u00b7f(SB), "+c.operand+", $0")
+		if !slices.Equal(txt.Flags, c.wantFlag) {
+			t.Errorf("%s: flags = %v, want %v", c.operand, txt.Flags, c.wantFlag)
+		}
+		if txt.FlagVal != c.wantVal {
+			t.Errorf("%s: FlagVal = %d, want %d", c.operand, txt.FlagVal, c.wantVal)
+		}
+		if txt.Frame == nil || !txt.Frame.Imm.HasVal || txt.Frame.Imm.Val != 0 {
+			t.Errorf("%s: frame = %+v, want $0", c.operand, txt.Frame)
+		}
+	}
+
+	// No flags operand at all: the toolchain's zero.
+	txt := parseTextHeader(t, "TEXT \u00b7f(SB), $0")
+	if txt.Flags != nil || txt.FlagVal != 0 {
+		t.Errorf("no flags: flags = %v val = %d, want nil 0", txt.Flags, txt.FlagVal)
+	}
+}
+
+// TestTextFlagValues pins the flag table against the constants
+// cmd/internal/obj/textflag.go defines and the toolchain's textflag.h
+// ships; the assembler, linker and compiler must all agree on them.
+func TestTextFlagValues(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		val  int64
+	}{
+		{"NOPROF", 1}, {"DUPOK", 2}, {"NOSPLIT", 4}, {"RODATA", 8},
+		{"NOPTR", 16}, {"WRAPPER", 32}, {"NEEDCTXT", 64}, {"TLSBSS", 256},
+		{"NOFRAME", 512}, {"REFLECTMETHOD", 1024}, {"TOPFRAME", 2048},
+		{"ABIWRAPPER", 4096},
+	} {
+		txt := parseTextHeader(t, "TEXT \u00b7f(SB), "+c.name+", $0")
+		if txt.FlagVal != c.val {
+			t.Errorf("%s = %d, want %d", c.name, txt.FlagVal, c.val)
+		}
+		if !slices.Equal(txt.Flags, []string{c.name}) {
+			t.Errorf("%s: flags = %v, want [%s]", c.name, txt.Flags, c.name)
+		}
+	}
+}
+
+// TestFlagsUnknownName covers the identifier the table does not know: the
+// toolchain's assembler rejects the line ("unexpected TYPO evaluating
+// expression") and so does the assembly path of this parser, while the
+// tooling view stays tolerant, keeps the scanned names and carries no
+// value the assembler could trust.
+func TestFlagsUnknownName(t *testing.T) {
+	const header = "TEXT \u00b7f(SB), NOSPLIT|TYPO, $0"
+
+	txt := parseTextHeader(t, header)
+	if !slices.Equal(txt.Flags, []string{"NOSPLIT", "TYPO"}) {
+		t.Errorf("raw flags = %v, want the scanned names", txt.Flags)
+	}
+	if txt.FlagVal != 0 {
+		t.Errorf("raw FlagVal = %d, want 0", txt.FlagVal)
+	}
+
+	_, errs := parseTextHeaderExpanded(t, header)
+	if len(errs) != 1 {
+		t.Fatalf("assembly path: %d errors, want 1: %v", len(errs), errs)
+	}
+	perr, ok := errs[0].(Error)
+	if !ok {
+		t.Fatalf("error %v is not a parser.Error", errs[0])
+	}
+	if !strings.Contains(perr.Msg, "unexpected TYPO evaluating expression") {
+		t.Errorf("message = %q, want the toolchain's rejection", perr.Msg)
+	}
+	if perr.Pos.Line != 1 || perr.Pos.Column != 22 {
+		t.Errorf("position = %d:%d, want 1:22", perr.Pos.Line, perr.Pos.Column)
+	}
+}
+
+// TestFlagsLeftoverExpression covers the operand that never closes into
+// one expression: the toolchain rejects it where the expression stops.
+func TestFlagsLeftoverExpression(t *testing.T) {
+	_, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f(SB), NOSPLIT TYPO, $0")
+	if len(errs) != 1 || !strings.Contains(errs[0].Error(), "unexpected TYPO evaluating expression") {
+		t.Fatalf("errors = %v, want the toolchain's rejection at TYPO", errs)
+	}
+	txt := parseTextHeader(t, "TEXT \u00b7f(SB), NOSPLIT TYPO, $0")
+	if txt.FlagVal != 0 {
+		t.Errorf("raw FlagVal = %d, want 0", txt.FlagVal)
+	}
+}
+
+// TestNoFrameDemandsZeroFrame covers the toolchain's frame check its arm64
+// backend raises: NOFRAME reserves no frame, so a declared positive frame
+// contradicts the flag.  The zero and negative frames are the flag's own
+// shapes (the arm64 BSD syscall-stub pattern writes $-8).
+func TestNoFrameDemandsZeroFrame(t *testing.T) {
+	_, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f(SB), NOSPLIT|NOFRAME, $8")
+	if len(errs) != 1 || !strings.Contains(errs[0].Error(),
+		"NOFRAME functions must have a frame size of 0, not 8") {
+		t.Fatalf("errors = %v, want the toolchain's frame check", errs)
+	}
+	for _, frame := range []string{"$0", "$-8", "$-0"} {
+		if _, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f(SB), NOSPLIT|NOFRAME, "+frame); len(errs) != 0 {
+			t.Errorf("%s: errors = %v, want none", frame, errs)
+		}
+	}
+	// The tooling view leaves the case to the linter's advisory.
+	txt := parseTextHeader(t, "TEXT \u00b7f(SB), NOSPLIT|NOFRAME, $8")
+	if txt.Frame == nil {
+		t.Fatal("raw parse lost the frame")
+	}
+}
+
+// TestABIInternalRequiresNoSplit covers the toolchain's own TEXT check: an
+// ABIInternal symbol must not need the stack the ABI wrapper would have to
+// bridge, so the flag is mandatory.
+func TestABIInternalRequiresNoSplit(t *testing.T) {
+	_, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f<ABIInternal>(SB), $0")
+	if len(errs) != 1 || !strings.Contains(errs[0].Error(),
+		`TEXT "f": ABIInternal requires NOSPLIT`) {
+		t.Fatalf("errors = %v, want the toolchain's ABI check", errs)
+	}
+	if _, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f<ABIInternal>(SB), NOSPLIT, $0"); len(errs) != 0 {
+		t.Errorf("NOSPLIT present: errors = %v, want none", errs)
+	}
+	// The numeric spelling of the flag satisfies the check the same way
+	// the toolchain's integer test does.
+	if _, errs := parseTextHeaderExpanded(t, "TEXT \u00b7f<ABIInternal>(SB), 4, $0"); len(errs) != 0 {
+		t.Errorf("numeric NOSPLIT: errors = %v, want none", errs)
+	}
+}
+
+// TestGloblFlagsOperand covers the GLOBL side: the atoms stay as written
+// (the numeric combinations are what the link layer reads back) while the
+// operand still folds to one value and validates on the assembly path.
+func TestGloblFlagsOperand(t *testing.T) {
+	for _, c := range []struct {
+		operand  string
+		wantFlag []string
+		wantVal  int64
+		wantSize int64
+	}{
+		{"RODATA|NOPTR, $8", []string{"RODATA", "NOPTR"}, 24, 8},
+		{"10, $8", []string{"10"}, 10, 8},
+		{"8|2, $8", []string{"8", "2"}, 10, 8},
+		{"RODATA, $8", []string{"RODATA"}, 8, 8},
+		{"$8", nil, 0, 8},
+	} {
+		f, errs := Parse("t.s", "GLOBL \u00b7x(SB), "+c.operand+"\n")
+		if len(errs) > 0 {
+			t.Fatalf("parse %q: %v", c.operand, errs)
+		}
+		g := f.Decls[0].(*ast.Globl)
+		if !slices.Equal(g.Flags, c.wantFlag) {
+			t.Errorf("%s: flags = %v, want %v", c.operand, g.Flags, c.wantFlag)
+		}
+		if g.FlagVal != c.wantVal {
+			t.Errorf("%s: FlagVal = %d, want %d", c.operand, g.FlagVal, c.wantVal)
+		}
+		if c.wantSize != 0 && (g.Size == nil || !g.Size.Imm.HasVal || g.Size.Imm.Val != c.wantSize) {
+			t.Errorf("%s: size = %+v, want $%d", c.operand, g.Size, c.wantSize)
+		}
+	}
+
+	f, errs := ParseWithOptions("t.s", "GLOBL \u00b7x(SB), RODATA|TYPO, $8\n", Options{Expand: true})
+	if len(errs) != 1 || !strings.Contains(errs[0].Error(), "unexpected TYPO evaluating expression") {
+		t.Fatalf("errors = %v, want the toolchain's rejection", errs)
+	}
+	if f.Decls[0].(*ast.Globl).FlagVal != 0 {
+		t.Errorf("failed operand must carry no value")
+	}
+}

parser/parser.go

@@ -251,22 +251,35 @@ func (p *state) parseText(line []token.Token) {
 	text.Name = sym
 	rest = rest[n:]
 
-	// Consume the flags operand: everything between the symbol and the
+	// The tail splits into the toolchain's comma-separated operands: with
-	// frame '$' is one operand, which the toolchain evaluates to a single
+	// two, the first is the flags expression and the last the frame; with
-	// integer (identifiers joined by '|', each a known textflag.h name,
+	// three or more, the toolchain's own complaint.  One operand is the
-	// literals and constant arithmetic beside them).  A single trailing
+	// frame alone, whatever it starts with, so a malformed frame takes the
-	// comma is the separator that stood before the '$'.
+	// frame diagnostic and never reads as flags.
 	rest = skipComma(rest)
-	text.Flags, text.FlagVal = p.evalFlags(flagsRun(rest), false)
+	ops := splitOperands(rest)
-	for len(rest) > 0 && rest[0].Kind != token.Dollar {
+	var frame []token.Token
-		rest = rest[1:]
+	switch {
+	case len(ops) >= 2:
+		text.Flags, text.FlagVal = p.evalFlags(flagsRun(ops[0]), false)
+		frame = ops[1]
+		if len(ops) > 2 {
+			p.errorf(line[0].Pos, "expect two or three operands for TEXT")
+		}
+	case len(ops) == 1:
+		frame = ops[0]
+	}
+	if len(frame) > 0 && frame[0].Kind != token.Dollar {
+		p.errorf(line[0].Pos, "TEXT frame size must be an immediate constant")
+		frame = nil
 	}
 
 	// Frame: $[-]number ; optional args: -number.  The Go runtime writes
 	// zero frames with an explicit sign ("$-0-24"), so the number may carry
 	// one.  Whatever remains after the header is the body and is parsed by
 	// the caller.
-	if len(rest) > 0 && rest[0].Kind == token.Dollar {
+	if len(frame) > 0 && frame[0].Kind == token.Dollar {
+		rest = frame
 		n := 1
 		neg := false
 		if n < len(rest) && (rest[n].Kind == token.Minus || rest[n].Kind == token.Plus) {
@@ -337,13 +350,15 @@ func (p *state) parseGlobl(line []token.Token) *ast.Globl {
 	// operand as one constant expression, so every name is validated here
 	// and the whole operand folds to g.FlagVal; Flags keeps the atoms as
 	// written, the numeric spellings being data-side combinations the
-	// link layer reads back.
+	// link layer reads back.  The size is the last operand; a third
-	g.Flags, g.FlagVal = p.evalFlags(flagsRun(rest), true)
+	// operand's excess is the toolchain's own silence.
-	for i, t := range rest {
+	ops := splitOperands(rest)
-		if t.Kind == token.Dollar {
+	switch {
-			g.Size = parseOperand(rest[i:], false)
+	case len(ops) >= 2:
-			break
+		g.Flags, g.FlagVal = p.evalFlags(flagsRun(ops[0]), true)
-		}
+		g.Size = parseOperand(ops[1], false)
+	case len(ops) == 1:
+		g.Size = parseOperand(ops[0], false)
 	}
 	return g
 }