feat(amd64): floating-point immediates through a synthesised pool

Assisted-by: GLM 5.3 Flash

asm/assemble.go

@@ -5,6 +5,7 @@ package asm
 
 import (
 	"fmt"
+	"strconv"
 	"strings"
 
 	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
@@ -28,7 +29,7 @@ import (
 // emitted: the bytes match go tool asm only for NOSPLIT functions or
 // zero-frame leaves, where the toolchain emits no guard either.
 func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
-	code, _, labels, _, _, err := assemble(t, nil)
+	code, _, labels, _, _, _, err := assemble(t, nil)
 	return code, labels, err
 }
 
@@ -66,9 +67,9 @@ type spadjStep struct {
 // assemble encodes a TEXT body, returning the machine code, the static-symbol
 // patch sites (for the file-level layout to resolve), the label table and the
 // stack-adjustment boundaries.
-func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, []spadjStep, []LineEntry, error) {
+func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, []spadjStep, []LineEntry, []floatPoolEntry, error) {
 	if err := checkAdjspBalance(t); err != nil {
-		return nil, nil, nil, nil, nil, err
+		return nil, nil, nil, nil, nil, nil, err
 	}
 	fi := computeFrame(t)
 	chain := jumpChain(t)
@@ -88,6 +89,8 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	offsets := map[string]int{}
 	pcs := make([]int, len(t.Body))
 	var guardJBlong, guardJBElong, moreJMPlong bool
+	poolSeen := map[string]bool{}
+	var poolList []floatPoolEntry
 	for {
 		guard := fi.guardLen(guardJBlong, guardJBElong)
 		pos := guard + len(fi.prologue)
@@ -98,7 +101,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 			case *ast.Instr:
 				sz, err := instrSize(s, fi, long[i], link)
 				if err != nil {
-					return nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+					return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 				}
 				sizes[i] = sz
 				pcs[i] = pos
@@ -229,12 +232,18 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 				spadjStep{pos + epi, 0},
 			)
 		}
-		code, ps, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
+		code, ps, pool, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
 		if err != nil {
-			return nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+			return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+		}
+		for _, entry := range pool {
+			if !poolSeen[entry.name] {
+				poolSeen[entry.name] = true
+				poolList = append(poolList, entry)
+			}
 		}
 		if len(code) != sizes[i] {
-			return nil, nil, nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
+			return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
 		}
 		if strings.ToUpper(s.Mnemonic.Text) == "CALL" {
 			for k := range ps {
@@ -272,7 +281,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 		pos += len(suffix)
 	}
 	_ = pos
-	return out, patches, offsets, steps, lines, nil
+	return out, patches, offsets, steps, lines, poolList, nil
 }
 
 // jumpChain precomputes jump-to-jump folding: a label whose first instruction
@@ -593,7 +602,7 @@ 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)
 	if err != nil {
 		return 0, err
 	}
@@ -628,7 +637,7 @@ 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, error) {
+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) {
 	mnem := strings.ToUpper(s.Mnemonic.Text)
 
 	var prefix []byte
@@ -638,6 +647,7 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 
 	var code []byte
 	var ps []sbPatch
+	var pool []floatPoolEntry
 	var err error
 	if isJumpMnemonic(mnem) {
 		if (mnem == "CALL" || mnem == "JMP") && isSBCall(s) {
@@ -646,7 +656,7 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 			// or the linker.
 			code, ps, err = encodeSBCall(s, link)
 			if err != nil {
-				return nil, nil, err
+				return nil, nil, nil, err
 			}
 			for i := range ps {
 				ps[i].kind = RelCall
@@ -656,23 +666,23 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 				ps[i].off += body
 				ps[i].after = body + len(code)
 			}
-			return append(prefix, code...), ps, nil
+			return append(prefix, code...), ps, nil, nil
 		}
 		if (mnem == "CALL" || mnem == "JMP") && indirectJumpTarget(s) {
 			// JMP/CALL through a register or memory: no relocation and no
 			// label to resolve, the operand fully determines the bytes.
 			code, err = encodeIndirectJump(s, mnem)
 			if err != nil {
-				return nil, nil, err
+				return nil, nil, nil, err
 			}
-			return append(prefix, code...), nil, nil
+			return append(prefix, code...), nil, nil, nil
 		}
 		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve)
 	} else {
-		code, ps, err = encodeNormal(s, fi, link)
+		code, ps, pool, err = encodeNormal(s, fi, link)
 	}
 	if err != nil {
-		return nil, nil, err
+		return nil, nil, nil, err
 	}
 	// Anchor the patch fields at function-relative positions: off indexes the
 	// disp32 field, after is the address just past the instruction.
@@ -681,31 +691,65 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 		ps[i].off += body
 		ps[i].after = body + len(code)
 	}
-	return append(prefix, code...), ps, nil
+	return append(prefix, code...), ps, pool, nil
 }
 
-func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, error) {
-	_, size := splitSize(strings.ToUpper(s.Mnemonic.Text))
+func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, []floatPoolEntry, error) {
+	mnemUpper := strings.ToUpper(s.Mnemonic.Text)
+	if mnemUpper == "FUNCDATA" || mnemUpper == "PCDATA" {
+		code, err := encodeBookkeeping(mnemUpper, s)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		return code, nil, nil, nil
+	}
+	_, size := splitSize(mnemUpper)
 	if size == 0 {
 		size = 8
 	}
 	ops := make([]Operand, len(s.Operands))
 	for i, op := range s.Operands {
-		o, err := operandFromAST(op, size, fi, link)
+		o, err := operandFromAST(mnemUpper, op, size, fi, link)
 		if err != nil {
-			return nil, nil, err
+			return nil, nil, nil, err
 		}
 		ops[i] = o
 	}
 	e := &enc{}
 	if err := e.encode(s.Mnemonic.Text, ops); err != nil {
-		return nil, nil, err
+		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
+	return e.out, ps, e.floatPoolList(), 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
+// function body.  The symbol reference must not run through the SB-operand
+// path, which demands file-level resolution the statement never needs.
+func encodeBookkeeping(upper string, s *ast.Instr) ([]byte, error) {
+	if len(s.Operands) != 2 {
+		return nil, fmt.Errorf("%s expects 2 operands, got %d", upper, len(s.Operands))
+	}
+	a, b := s.Operands[0], s.Operands[1]
+	if a.Kind != ast.OpImmediate || !a.Imm.HasVal {
+		return nil, fmt.Errorf("%s: first operand must be an integer immediate", upper)
+	}
+	switch upper {
+	case "FUNCDATA":
+		if b.Kind != ast.OpAddr || b.Addr.Sym == nil || b.Addr.Sym.Pseudo != "SB" {
+			return nil, fmt.Errorf("FUNCDATA: second operand must be a symbol reference")
+		}
+	case "PCDATA":
+		if b.Kind != ast.OpImmediate || !b.Imm.HasVal {
+			return nil, fmt.Errorf("PCDATA: second operand must be an integer immediate")
+		}
+	}
+	return nil, nil
 }
 
 // encodeJump encodes a JMP/CALL/Jcc with a relative offset resolved from the
@@ -756,7 +800,7 @@ func isSBCall(s *ast.Instr) bool {
 
 // encodeSBCall encodes CALL sym(SB) as E8 rel32 with a patch site.
 func encodeSBCall(s *ast.Instr, link *linkInfo) ([]byte, []sbPatch, error) {
-	o, err := operandFromAST(s.Operands[0], 8, frameInfo{}, link)
+	o, err := operandFromAST(strings.ToUpper(s.Mnemonic.Text), s.Operands[0], 8, frameInfo{}, link)
 	if err != nil {
 		return nil, nil, err
 	}
@@ -813,7 +857,7 @@ func indirectJumpTarget(s *ast.Instr) bool {
 func encodeIndirectJump(s *ast.Instr, mnem string) ([]byte, error) {
 	ops := make([]Operand, len(s.Operands))
 	for i, op := range s.Operands {
-		o, err := operandFromAST(op, 8, frameInfo{}, nil)
+		o, err := operandFromAST(mnem, op, 8, frameInfo{}, nil)
 		if err != nil {
 			return nil, err
 		}
@@ -830,8 +874,11 @@ func encodeIndirectJump(s *ast.Instr, mnem string) ([]byte, error) {
 var spReg = Reg{idx: 4, size: 8}
 
 // operandFromAST converts a parsed operand into an encoder Operand, applying
-// the frame translation to FP/SP pseudo-register operands.
-func operandFromAST(op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Operand, error) {
+// the frame translation to FP/SP pseudo-register operands.  mnemUpper is the
+// instruction's upper-case mnemonic, which the floating-point immediate gate
+// needs: only the SSE mnemonics whose encoding takes an XMM/memory source
+// accept one.
+func operandFromAST(mnemUpper string, op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Operand, error) {
 	switch op.Kind {
 	case ast.OpImmediate:
 		if op.Imm.HasVal {
@@ -841,17 +888,43 @@ func operandFromAST(op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Op
 			}
 			return Imm(v), nil
 		}
+		// A floating-point immediate: $1.5, $-1.0 or the parenthesised
+		// $(-1.0) spelling (the constant-expression folder only folds
+		// integers, so that shape arrives with an empty Immediate and only
+		// the raw spelling carries the value).  The toolchain rewrites it
+		// into a pooled-constant read on the SSE scalar paths and rejects
+		// it everywhere else.
+		if text, neg, ok := floatImmText(op); ok {
+			if !sseFloatImm[mnemUpper] {
+				return nil, fmt.Errorf("%s does not take a floating-point immediate", mnemUpper)
+			}
+			return FloatImm{Text: text, Neg: neg}, nil
+		}
 		return nil, fmt.Errorf("non-integer immediate not supported")
 
 	case ast.OpAddr:
 		a := op.Addr
 
 		// A bracketed register range, [Z0-Z3]: the four-register source of
-		// the 4FMAPS/4VNNIW families.  The EVEX quad-register emit path
-		// needs an encoder operand of its own, so the shape stays a named
-		// gap rather than an encoding.
+		// the 4FMAPS/4VNNIW families.  The range must span four consecutive
+		// same-width vector registers, exactly what the toolchain's parser
+		// takes; the EVEX quad-register emit path reads the low end.
 		if a.Range != nil {
-			return nil, fmt.Errorf("register range %q needs quad-register encoder support", op.Raw)
+			lo, ok := ParseReg(a.Range.Lo)
+			if !ok {
+				return nil, fmt.Errorf("unknown register %q in range", a.Range.Lo)
+			}
+			hi, ok := ParseReg(a.Range.Hi)
+			if !ok {
+				return nil, fmt.Errorf("unknown register %q in range", a.Range.Hi)
+			}
+			if !lo.isVec() || lo.size != hi.size {
+				return nil, fmt.Errorf("register range %q must span four same-width vector registers", op.Raw)
+			}
+			if hi.idx != lo.idx+3 {
+				return nil, fmt.Errorf("register range %q must span four consecutive registers", op.Raw)
+			}
+			return RegList{Lo: lo, Hi: hi}, nil
 		}
 
 		// FP-relative: x+N(FP) → (N + fpAdjust)(SP).  The offset N lives in the
@@ -921,3 +994,37 @@ func operandFromAST(op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Op
 	}
 	return nil, fmt.Errorf("unsupported operand")
 }
+
+// floatImmText recovers a floating-point immediate's magnitude and sign from
+// the parsed operand.  The ordinary spellings arrive in Imm.Float; the
+// parenthesised $(-1.0) leaves the Immediate empty, because the integer
+// folder cannot read it, and only the verbatim operand text still carries
+// the value.  Anything that is not a number a float parser accepts reports
+// not-ok, so every other shape keeps its existing diagnostic.
+func floatImmText(op *ast.Operand) (text string, neg bool, ok bool) {
+	if op.Imm.Float != "" {
+		return op.Imm.Float, op.Imm.Neg, true
+	}
+	if op.Imm.HasVal || op.Imm.Str != "" || op.Imm.Sym != nil {
+		return "", false, false
+	}
+	// joinRaw spaced the token texts; the compact spelling is what matters.
+	compact := strings.ReplaceAll(op.Raw, " ", "")
+	inner, ok := strings.CutPrefix(compact, "$(")
+	if !ok || !strings.HasSuffix(inner, ")") {
+		return "", false, false
+	}
+	inner = strings.TrimSuffix(inner, ")")
+	inner = strings.TrimPrefix(inner, "+")
+	if s, ok := strings.CutPrefix(inner, "-"); ok {
+		neg = true
+		inner = s
+	}
+	if inner == "" || !strings.ContainsAny(inner, "0123456789") {
+		return "", false, false
+	}
+	if _, err := strconv.ParseFloat(inner, 64); err != nil {
+		return "", false, false
+	}
+	return inner, neg, true
+}

asm/assemble_test.go

@@ -568,3 +568,28 @@ TEXT ·framed(SB), $16-8
 		t.Errorf("framed adjsp:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
 	}
 }
+
+// TestAssembleRegRange pins the bracketed register range at the statement
+// level: exactly four consecutive same-width vector registers assemble, the
+// toolchain's rejected shapes all report an error.
+func TestAssembleRegRange(t *testing.T) {
+	asm := func(t *testing.T, op string) ([]byte, error) {
+		t.Helper()
+		f, errs := parser.Parse("f_amd64.s", "TEXT \u00b7f(SB), NOSPLIT, $0\n\tV4FMADDPS 17(SP), "+op+", K2, Z0\n\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("parse %s: %v", op, errs)
+		}
+		code, _, err := Assemble(f.Decls[0].(*ast.Text))
+		return code, err
+	}
+	for _, op := range []string{"[Z0-Z3]", "[Z4-Z7]", "[Z28-Z31]"} {
+		if _, err := asm(t, op); err != nil {
+			t.Errorf("%s: %v", op, err)
+		}
+	}
+	for _, op := range []string{"[Z0-Z4]", "[Z0-Z2]", "[Z0-Z0]", "[Z4-Z0]", "[Z1-Z0]", "[AX-Z3]", "[Z0-AX]"} {
+		if _, err := asm(t, op); err == nil {
+			t.Errorf("%s: assembled, want an error", op)
+		}
+	}
+}

asm/encodable.go

@@ -20,9 +20,9 @@ func Encodable(mnemonic string) bool {
 	switch upper {
 	case "RET", "NOP", "CALL", "JMP",
 		"POPFQ", "PUSHFQ", "INT", "LDMXCSR", "STMXCSR", "CMPSD", "SHA256RNDS2",
-		// The literal-data pseudo-ops, the accepted-and-ignored END and the
-		// SP adjust.
-		"BYTE", "WORD", "LONG", "QUAD", "END", "ADJSP":
+		// The literal-data pseudo-ops, the accepted-and-ignored END and
+		// bookkeeping statements, and the SP adjust.
+		"BYTE", "WORD", "LONG", "QUAD", "END", "ADJSP", "FUNCDATA", "PCDATA":
 		return true
 	}
 	if _, ok := noOperandTable[upper]; ok {

asm/encode.go

@@ -5,6 +5,8 @@ package asm
 
 import (
 	"fmt"
+	"math"
+	"strconv"
 	"strings"
 )
 
@@ -21,6 +23,39 @@ func Encode(mnemonic string, ops ...Operand) ([]byte, error) {
 type enc struct {
 	out     []byte
 	patches []encPatch // disp32 fields awaiting static-symbol resolution
+
+	// FloatPool collects the pooled constants the floating-point
+	// immediates reference, in first-use order.
+	floatPool     []floatPoolEntry
+	floatPoolSeen map[string]bool
+}
+
+// floatPoolEntry is one pooled floating-point constant: the symbol name
+// the emitted RIP-relative load refers to and its IEEE-754 bytes.
+type floatPoolEntry struct {
+	name string
+	data []byte
+}
+
+// addFloatPool records a pooled constant, deduplicated by symbol name.
+func (e *enc) addFloatPool(name string, bits uint64, width int) {
+	if e.floatPoolSeen == nil {
+		e.floatPoolSeen = map[string]bool{}
+	}
+	if e.floatPoolSeen[name] {
+		return
+	}
+	e.floatPoolSeen[name] = true
+	data := make([]byte, width)
+	for i := range width {
+		data[i] = byte(bits >> (8 * i))
+	}
+	e.floatPool = append(e.floatPool, floatPoolEntry{name: name, data: data})
+}
+
+// floatPoolList returns the pooled constants in first-use order.
+func (e *enc) floatPoolList() []floatPoolEntry {
+	return e.floatPool
 }
 
 // encPatch marks a 4-byte displacement field in enc.out that must receive the
@@ -102,6 +137,11 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodeEnd(ops)
 	case "ADJSP":
 		return e.encodeAdjsp(ops)
+	// The runtime's bookkeeping statements carry no text bytes: go tool asm
+	// records FUNCDATA and PCDATA in the program list only, so the encoded
+	// body shows nothing, on every architecture.
+	case "FUNCDATA", "PCDATA":
+		return e.encodeFuncdata(upper, ops)
 	}
 
 	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
@@ -141,11 +181,18 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	}
 	// Legacy SSE packed binaries dispatch on the full name: the packed
 	// integer mnemonics carry real width suffixes (PADDB/PCMPGTW/...),
-	// which the size split must not eat.
+	// which the size split must not eat.  A floating-point immediate
+	// rewrites into a pooled-constant read on the scalar members.
 	if m, ok := sseBinTable[upper]; ok {
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatBin(upper, m, f, ops)
+		}
 		return e.encodeSSEBin(m, ops)
 	}
 	if m, ok := sseBinTable[base]; ok {
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatBin(upper, m, f, ops)
+		}
 		return e.encodeSSEBin(m, ops)
 	}
 	// The imm8-controlled legacy instructions, the lane extracts and inserts
@@ -222,7 +269,12 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodeCvtInt(base, ops, size)
 	case "FMOVD":
 		return e.encodeFmov(ops)
-	case "MOVOU", "MOVO", "MOVOA", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
+	case "MOVSD", "MOVSS":
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatMove(upper, f, ops)
+		}
+		return e.encodeSSEMove(sseMoveTable[base], ops)
+	case "MOVOU", "MOVO", "MOVOA", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD":
 		return e.encodeSSEMove(sseMoveTable[base], ops)
 	}
 	return fmt.Errorf("unsupported instruction %q", mnem)
@@ -285,6 +337,33 @@ func (e *enc) encodeData(mnem string, ops []Operand) error {
 	return nil
 }
 
+// encodeFuncdata accepts-and-ignores the runtime bookkeeping statements:
+// FUNCDATA $n, sym(SB) and PCDATA $n, $m.  go tool asm emits no text bytes
+// for either (the entries live in the object's ancillary tables, not the
+// function body), and the operand shapes it takes are exactly these: an
+// integer count first, then a symbol reference for FUNCDATA and an integer
+// value for PCDATA.  The other architectures accept-and-ignore the same
+// statements; amd64 now matches.
+func (e *enc) encodeFuncdata(upper string, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("%s expects 2 operands, got %d", upper, len(ops))
+	}
+	if _, ok := ops[0].(Imm); !ok {
+		return fmt.Errorf("%s: first operand must be an integer immediate", upper)
+	}
+	switch upper {
+	case "FUNCDATA":
+		if _, ok := ops[1].(sbMem); !ok {
+			return fmt.Errorf("FUNCDATA: second operand must be a symbol reference")
+		}
+	case "PCDATA":
+		if _, ok := ops[1].(Imm); !ok {
+			return fmt.Errorf("PCDATA: second operand must be an integer immediate")
+		}
+	}
+	return nil
+}
+
 // encodeEnd accepts-and-ignores END.  go tool asm drops the statement
 // entirely: the AEND Prog is skipped when the program list is flushed, so
 // the statements after an END still belong to the same function and the
@@ -319,6 +398,120 @@ func (e *enc) encodeAdjsp(ops []Operand) error {
 	return nil
 }
 
+// --- floating-point immediates ----------------------------------------------
+
+// sseFloatImm lists the mnemonics whose first operand may be a floating-point
+// immediate, the set go tool asm rewrites into a pooled-constant read: the
+// scalar moves, the four scalar arithmetic pairs and the scalar compares.
+// The packed members and the uniform forms (MAXSD, MINSD, SQRTSD, CMPSD)
+// reject the immediate in the toolchain and are absent here on purpose.
+var sseFloatImm = map[string]bool{
+	"MOVSD": true, "MOVSS": true,
+	"ADDSD": true, "ADDSS": true,
+	"SUBSD": true, "SUBSS": true,
+	"MULSD": true, "MULSS": true,
+	"DIVSD": true, "DIVSS": true,
+	"COMISD": true, "COMISS": true,
+	"UCOMISD": true, "UCOMISS": true,
+}
+
+// floatImmOperand reports whether the operand list opens with a
+// floating-point immediate in the two-operand spelling (imm, dst).
+func floatImmOperand(ops []Operand) (FloatImm, bool) {
+	if len(ops) != 2 {
+		return FloatImm{}, false
+	}
+	f, ok := ops[0].(FloatImm)
+	return f, ok
+}
+
+// floatPoolValue evaluates a floating-point immediate at the width its
+// mnemonic encodes and names the pool constant the toolchain synthesises:
+// $f64.<16 hex> for the doubles, $f32.<8 hex> for the singles (the float32
+// rounding of the parsed value).  The name carries the IEEE-754 bits; the
+// section holds them little-endian.
+func floatPoolValue(mnem string, f FloatImm) (bits uint64, name string, err error) {
+	v, err := strconv.ParseFloat(f.Text, 64)
+	if err != nil {
+		return 0, "", fmt.Errorf("invalid floating-point immediate %q", f.Text)
+	}
+	if f.Neg {
+		v = -v
+	}
+	if strings.HasSuffix(mnem, "D") {
+		bits = math.Float64bits(v)
+		return bits, fmt.Sprintf("$f64.%016x", bits), nil
+	}
+	bits = uint64(math.Float32bits(float32(v)))
+	return bits, fmt.Sprintf("$f32.%08x", bits), nil
+}
+
+// encodeSSEFloatMove encodes MOVSD/MOVSS with a floating-point immediate
+// source.  A positive zero needs no memory read: the toolchain emits
+// XORPS dst, dst.  Anything else loads the pooled constant RIP-relative
+// ($f64.<hex>(SB) / $f32.<hex>(SB)), the displacement a patch site the
+// file-level layout or the linker resolves.
+func (e *enc) encodeSSEFloatMove(mnem string, f FloatImm, ops []Operand) error {
+	if !sseFloatImm[mnem] {
+		return fmt.Errorf("%s does not take a floating-point immediate", mnem)
+	}
+	dst, ok := ops[1].(Reg)
+	if !ok || !dst.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", mnem)
+	}
+	bits, name, err := floatPoolValue(mnem, f)
+	if err != nil {
+		return err
+	}
+	e.addFloatPool(name, bits, mwidth(mnem))
+	if bits == 0 {
+		i := &instr{opcode: []byte{0x0F, 0x57}, modrm: -1, sib: -1} // XORPS
+		if err := setRM(i, dst, dst, 8); err != nil {
+			return err
+		}
+		return e.emit(i)
+	}
+	m := sseMoveTable[mnem]
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.load}, modrm: -1, sib: -1}
+	if err := setRM(i, dst, sbMem{size: mwidth(mnem), name: name}, 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
+// encodeSSEFloatBin encodes the scalar arithmetic and compare mnemonics with
+// a floating-point immediate source: the constant is read from the pool into
+// the instruction's r/m side (reg = destination), the rewrite go tool asm
+// performs at the source level.
+func (e *enc) encodeSSEFloatBin(mnem string, m sseBin, f FloatImm, ops []Operand) error {
+	if !sseFloatImm[mnem] {
+		return fmt.Errorf("%s does not take a floating-point immediate", mnem)
+	}
+	dst, ok := ops[1].(Reg)
+	if !ok || !dst.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", mnem)
+	}
+	bits, name, err := floatPoolValue(mnem, f)
+	if err != nil {
+		return err
+	}
+	e.addFloatPool(name, bits, mwidth(mnem))
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.op}, modrm: -1, sib: -1}
+	if err := setRM(i, dst, sbMem{size: mwidth(mnem), name: name}, 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
+// mwidth returns the operand width a scalar SSE mnemonic encodes: the double
+// spellings end in D, the single spellings in S.
+func mwidth(mnem string) int {
+	if strings.HasSuffix(mnem, "D") {
+		return 8
+	}
+	return 4
+}
+
 // splitSize separates a trailing B/W/L/Q size suffix from the mnemonic.
 func splitSize(upper string) (base string, size int) {
 	if upper == "" {

asm/encode_test.go

@@ -9,6 +9,9 @@ import (
 	"testing"
 
 	"golang.org/x/arch/x86/x86asm"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
 )
 
 // decode encodes an instruction and decodes it back, returning the decoded
@@ -1064,3 +1067,157 @@ func TestAdjsp(t *testing.T) {
 		t.Error("ADJSP AX assembled, want an error")
 	}
 }
+
+// TestFloatImmediateGroundTruth pins the floating-point immediate rewrite
+// byte for byte against go tool asm: the scalar moves and the scalar
+// arithmetic read the constant from a synthesised read-only pool symbol
+// ($f64.<hex>, $f32.<hex>) RIP-relative with the displacement left to the
+// relocation, and a positive zero on the moves collapses to XORPS dst, dst.
+func TestFloatImmediateGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"MOVSD -1.0", "MOVSD", []Operand{FloatImm{Text: "1.0", Neg: true}, vreg(t, "X2")}, "f20f101500000000"},
+		{"MOVSD 1.5", "MOVSD", []Operand{FloatImm{Text: "1.5"}, vreg(t, "X3")}, "f20f101d00000000"},
+		{"MOVSS 2.5", "MOVSS", []Operand{FloatImm{Text: "2.5"}, vreg(t, "X4")}, "f30f102500000000"},
+		{"MOVSS -0.5", "MOVSS", []Operand{FloatImm{Text: "0.5", Neg: true}, vreg(t, "X5")}, "f30f102d00000000"},
+		{"MOVSS +0.0 is XORPS", "MOVSS", []Operand{FloatImm{Text: "0.0"}, vreg(t, "X10")}, "450f57d2"},
+		{"MOVSD +0.0 is XORPS", "MOVSD", []Operand{FloatImm{Text: "0.0"}, vreg(t, "X6")}, "0f57f6"},
+		{"ADDSD 1.0", "ADDSD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}, "f20f580500000000"},
+		{"ADDSS 0.5", "ADDSS", []Operand{FloatImm{Text: "0.5"}, vreg(t, "X1")}, "f30f580d00000000"},
+		{"SUBSD 2.0", "SUBSD", []Operand{FloatImm{Text: "2.0"}, vreg(t, "X3")}, "f20f5c1d00000000"},
+		{"MULSD -2.5", "MULSD", []Operand{FloatImm{Text: "2.5", Neg: true}, vreg(t, "X3")}, "f20f591d00000000"},
+		{"DIVSD 1.0", "DIVSD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}, "f20f5e0500000000"},
+		{"COMISD 1.0", "COMISD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}, "660f2f0500000000"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: got %s, want %s", c.name, got, c.want)
+		}
+	}
+
+	// The pool names carry the IEEE-754 bits, the float32 narrowing for the
+	// single spellings; negative zero keeps its sign bit and never takes the
+	// XORPS shortcut.
+	for _, c := range []struct {
+		mnem string
+		imm  FloatImm
+		want string
+	}{
+		{"MOVSD", FloatImm{Text: "1.0", Neg: true}, "$f64.bff0000000000000"},
+		{"MOVSD", FloatImm{Text: "0.5"}, "$f64.3fe0000000000000"},
+		{"MOVSS", FloatImm{Text: "2.5"}, "$f32.40200000"},
+		{"MOVSS", FloatImm{Text: "0.5", Neg: true}, "$f32.bf000000"},
+		{"MOVSD", FloatImm{Text: "0.0", Neg: true}, "$f64.8000000000000000"},
+	} {
+		_, name, err := floatPoolValue(c.mnem, c.imm)
+		if err != nil {
+			t.Errorf("%s %s: %v", c.mnem, c.imm.Text, err)
+			continue
+		}
+		if name != c.want {
+			t.Errorf("%s $%s: pool name %s, want %s", c.mnem, c.imm.Text, name, c.want)
+		}
+	}
+
+	// The shapes the toolchain's parser rejects: the packed and uniform
+	// forms, a non-vector destination, and the integer spellings.
+	for _, c := range []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"MAXSD rejects the immediate", "MAXSD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}},
+		{"MINSD rejects the immediate", "MINSD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}},
+		{"SQRTSD rejects the immediate", "SQRTSD", []Operand{FloatImm{Text: "1.0"}, vreg(t, "X0")}},
+		{"integer destination", "MOVSD", []Operand{FloatImm{Text: "1.0"}, AX}},
+	} {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}
+
+// TestBookkeepingGroundTruth pins FUNCDATA and PCDATA as accept-and-ignore:
+// go tool asm emits no text bytes for either, on every architecture.
+func TestBookkeepingGroundTruth(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"FUNCDATA", "FUNCDATA", []Operand{Imm(3), sbMem{name: "\u00b7f.arginfo0"}}},
+		{"PCDATA", "PCDATA", []Operand{Imm(1), Imm(-1)}},
+	} {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if len(code) != 0 {
+			t.Errorf("%s: emitted %x, want no bytes", c.name, code)
+		}
+	}
+	for _, c := range []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"FUNCDATA arity", "FUNCDATA", []Operand{Imm(3)}},
+		{"FUNCDATA missing the count", "FUNCDATA", []Operand{sbMem{name: "x"}}},
+		{"FUNCDATA integer value", "FUNCDATA", []Operand{Imm(3), Imm(4)}},
+		{"PCDATA arity", "PCDATA", []Operand{Imm(1)}},
+		{"PCDATA register value", "PCDATA", []Operand{Imm(1), AX}},
+	} {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+
+	// At the statement level the bookkeeping lines sit between real
+	// instructions and contribute nothing to the body, symbol reference
+	// included: the FUNCDATA operand never needs file-level resolution.
+	f, errs := parser.Parse("t_amd64.s", "TEXT \u00b7f(SB), NOSPLIT, $0\n\tNOP\n\tFUNCDATA $3, \u00b7f.arginfo0(SB)\n\tPCDATA $1, $-1\n\tFUNCDATA $0, x<>(SB)\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	want := "90c3"
+	if got := hexCompact(img.Code); got != want {
+		t.Errorf("body %s, want %s (the bookkeeping lines contribute nothing)", got, want)
+	}
+	if _, err := AssembleFile(mustParse(t, "TEXT \u00b7f(SB), NOSPLIT, $0\n\tFUNCDATA $1, X0\n\tRET\n")); err == nil {
+		t.Error("FUNCDATA $1, X0 assembled, want an error")
+	}
+	if _, err := AssembleFile(mustParse(t, "TEXT \u00b7f(SB), NOSPLIT, $0\n\tPCDATA $1, X0\n\tRET\n")); err == nil {
+		t.Error("PCDATA $1, X0 assembled, want an error")
+	}
+
+	// Encodable mirrors Encode for the names this work touched.
+	for _, mnem := range []string{"FUNCDATA", "PCDATA", "V4FMADDPS", "V4FMADDSS", "V4FNMADDPS", "V4FNMADDSS", "VP4DPWSSD", "VP4DPWSSDS"} {
+		if !Encodable(mnem) {
+			t.Errorf("Encodable(%s) = false, want true", mnem)
+		}
+	}
+}
+
+// mustParse parses src or fails the test.
+func mustParse(t *testing.T, src string) *ast.File {
+	t.Helper()
+	f, errs := parser.Parse("t_amd64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	return f
+}

asm/evex.go

@@ -737,6 +737,91 @@ var evexTable = map[string]evexSpec{
 	"VMOVLHPS": {1, 0x16, 0, 0, -1, vexNDS3, [3]int{8, 0, 0}},
 }
 
+// evexQuad describes one quad-register instruction: the opcode under
+// EVEX.0F38.W0 with the F2 mandatory prefix, and the width of the vector
+// registers the bracketed list and the destination take (512-bit ZMM for
+// the packed forms, 128-bit XMM for the scalar ones).
+type evexQuad struct {
+	opcode byte
+	width  int // register width in bytes: 64 (ZMM) or 16 (XMM)
+}
+
+// evexQuadTable maps the quad-register instructions (the 4FMAPS and 4VNNIW
+// families) to their encoding.  The operand shape is fixed: a single memory
+// source in r/m, the bracketed register list whose LOW register travels the
+// inverted 5-bit V'VVVV field, an optional opmask in aaa and the vector
+// destination in reg.  The vector length follows the destination (512-bit
+// for the ZMM list forms, 128-bit for the scalar ones) while the disp8×N
+// multiplier stays 16 for every member, the toolchain's own tuple choice.
+var evexQuadTable = map[string]evexQuad{
+	"V4FMADDPS":  {0x9A, 64},
+	"V4FMADDSS":  {0x9B, 16},
+	"V4FNMADDPS": {0xAA, 64},
+	"V4FNMADDSS": {0xAB, 16},
+	"VP4DPWSSD":  {0x52, 64},
+	"VP4DPWSSDS": {0x53, 64},
+}
+
+// isEvexQuad reports whether the mnemonic is a quad-register instruction.
+func isEvexQuad(upper string) bool {
+	_, ok := evexQuadTable[upper]
+	return ok
+}
+
+// encodeEvexQuad encodes the quad-register form: OP mem, [Zn-Zn+3], (K), dst.
+// The register list is the VVVV-side source: its low register fills the
+// inverted V'VVVV bits, which is why an indexed memory source above Z15 (no
+// spare EVEX.X bit once V' is taken) is refused.  Masking rides the standard
+// aaa field, zeroing keeps the usual requires-a-mask rule, and no other
+// suffix applies.
+func (e *enc) encodeEvexQuad(mnem string, q evexQuad, ops []Operand, sfx evexSuffix) error {
+	if len(ops) != 3 && len(ops) != 4 {
+		return fmt.Errorf("%s expects 3 or 4 operands (mem, [Zn-Zn+3], (K), dst), got %d", mnem, len(ops))
+	}
+	mem, lst := ops[0], ops[1]
+	dst := ops[len(ops)-1]
+	mask := 0
+	if len(ops) == 4 {
+		k, ok := ops[2].(Reg)
+		if !ok || !k.mask {
+			return fmt.Errorf("%s: third operand must be an opmask register", mnem)
+		}
+		if k.idx == 0 {
+			return fmt.Errorf("k0 is not a usable mask register")
+		}
+		mask = k.idx
+	}
+	list, ok := lst.(RegList)
+	if !ok {
+		return fmt.Errorf("%s: second operand must be a four-register list", mnem)
+	}
+	if list.Lo.size != q.width {
+		return fmt.Errorf("%s: the register list must hold %d-bit vector registers", mnem, q.width*8)
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", mnem)
+	}
+	if dstReg.size != q.width {
+		return fmt.Errorf("%s: the destination must be a %d-bit vector register", mnem, q.width*8)
+	}
+	if !memOperand(mem) {
+		return fmt.Errorf("%s: the source must be a memory operand", mnem)
+	}
+	// The list owns V'VVVV; a scaled index in the EVEX-only half would fold
+	// its fifth bit into the same field the list's low register occupies.
+	if m, ok := mem.(Mem); ok && m.HasIndex && m.Index.idx >= 16 {
+		return fmt.Errorf("%s: an index register above Z15 has no EVEX bit free", mnem)
+	}
+	if sfx.zeroing && mask == 0 {
+		return fmt.Errorf("%s: zeroing (.Z) requires a mask register", mnem)
+	}
+	spec := evexSpec{mapSel: 2, opcode: q.opcode, w: 0, pp: 3, opdigit: -1, n: [3]int{16, 16, 16}}
+	// The vector length follows the destination (512-bit for the ZMM forms,
+	// 128-bit for the scalar ones), exactly as the oracle encodes it.
+	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, list.Lo.idx, mem, mask, sfx)
+}
+
 // evexBcastSpec describes an EVEX broadcast (VPBROADCASTD/Q): the opcode
 // depends on the source kind, a GPR source uses opReg, a memory source uses
 // opMem with a disp8×N of n.
@@ -812,8 +897,10 @@ func isEvex(mnemUpper string) bool {
 	if _, ok := evexBcastTable[mnemUpper]; ok {
 		return true
 	}
-	_, ok := evexMoveTable[mnemUpper]
-	return ok
+	if _, ok := evexMoveTable[mnemUpper]; ok {
+		return true
+	}
+	return isEvexQuad(mnemUpper)
 }
 
 // evexRequired reports whether the operands force the EVEX encoding of a
@@ -995,6 +1082,14 @@ func (e *enc) encodeEvex(mnemUpper string, ops []Operand, sfx evexSuffix) error
 		return e.encodeEvexRM(spec, ops, 0, sfx)
 	}
 	spec, inTable := evexTable[mnemUpper]
+	if q, ok := evexQuadTable[mnemUpper]; ok {
+		// The quad-register family carries no rounding, SAE or broadcast;
+		// only masking and zeroing apply.
+		if sfx.sae || sfx.bcst || sfx.rounding >= 0 {
+			return fmt.Errorf("%s takes no rounding/SAE/broadcast suffix", mnemUpper)
+		}
+		return e.encodeEvexQuad(mnemUpper, q, ops, sfx)
+	}
 	if inTable {
 		if (sfx.rounding >= 0 || sfx.sae) && !evexRound[mnemUpper] {
 			return fmt.Errorf("%s: rounding/SAE is not supported for this instruction", mnemUpper)

asm/kernels_differential_test.go

@@ -63,7 +63,10 @@ func toolAsmObject(t *testing.T, path, goarch string) []byte {
 }
 
 // oracleFuncCode extracts the non-package TEXT functions' code bytes from a
-// toolchain object, keyed by the name the object records (pkg.name).
+// toolchain object, keyed by the name the object records (pkg.name).  Each
+// function's span is its own symbol size: a toolchain object that follows
+// the text with data symbols (the synthesised float-constant pool) would
+// otherwise fold them into the last function's bytes.
 func oracleFuncCode(t *testing.T, obj []byte) map[string][]byte {
 	t.Helper()
 	v := openGoobj(t, obj)
@@ -76,18 +79,13 @@ func oracleFuncCode(t *testing.T, obj []byte) map[string][]byte {
 	for _, bi := range []int{blkSymdef, blkHashed64def, blkHasheddef} {
 		preceding += len(v.blk(bi)) / symSize
 	}
-	total := preceding + len(nps)
 	out := make(map[string][]byte, len(nps))
 	for i, s := range nps {
 		if s.typ != kindSTEXT {
 			continue
 		}
 		start := le.Uint32(didx[4*(preceding+i):])
-		end := uint32(len(data))
-		if preceding+i+1 < total {
-			end = le.Uint32(didx[4*(preceding+i+1):])
-		}
-		out[s.name] = data[start:end]
+		out[s.name] = data[start : start+s.size]
 	}
 	return out
 }
@@ -129,6 +127,9 @@ func TestDifferentialKernels(t *testing.T) {
 		{filepath.Join("..", "testdata", "verify", "datarel_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "divslash_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "semicolons_amd64.s"), "", false},
+		{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", "datarel_arm64.s"), "arm64", true},
 		{filepath.Join("..", "testdata", "verify", "divslash_arm64.s"), "arm64", true},
 	} {

asm/link.go

@@ -167,6 +167,7 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		known[d.name] = true
 	}
 	link := &linkInfo{symbols: known, allowExternal: true}
+	poolSeen := map[string]bool{}
 
 	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
 	textOff := map[string]int{}
@@ -180,7 +181,26 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		if !ok {
 			continue
 		}
-		code, patches, labels, steps, lines, err := assemble(t, link)
+		code, patches, labels, steps, lines, pool, err := assemble(t, link)
+		if err != nil {
+			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
+		}
+		// The pooled floating-point constants join the declared data as
+		// read-only symbols, deduplicated across the file (the toolchain
+		// synthesises the same symbols into its rodata).
+		for _, entry := range pool {
+			if poolSeen[entry.name] {
+				continue
+			}
+			poolSeen[entry.name] = true
+			dataSyms = append(dataSyms, dataSym{
+				name:   entry.name,
+				buf:    entry.data,
+				size:   len(entry.data),
+				rodata: true,
+				dupok:  true,
+			})
+		}
 		if err != nil {
 			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
 		}

asm/operand.go

@@ -14,6 +14,28 @@ type Imm int64
 
 func (Imm) isOperand() {}
 
+// RegList is a bracketed register range, [Z0-Z3]: the four-register source
+// of the 4FMAPS and 4VNNIW families.  The EVEX emit path carries the list's
+// low register through the inverted 5-bit V'VVVV field; the three higher
+// registers are implied by the instruction, so only the pair travels here.
+type RegList struct {
+	Lo Reg
+	Hi Reg // implied by the encoding; Lo.idx+3 by construction
+}
+
+func (RegList) isOperand() {}
+
+// FloatImm is a floating-point immediate ($-1.0).  The SSE mnemonics whose
+// encoding takes an XMM/memory source at that position rewrite it as a read
+// from a read-only pool constant ($f64.<hex> or $f32.<hex>), the toolchain's
+// own behaviour; every other instruction rejects it.
+type FloatImm struct {
+	Text string // the numeric text as written, sign excluded
+	Neg  bool   // a leading minus
+}
+
+func (FloatImm) isOperand() {}
+
 // Mem is a memory operand of the form disp(base)(index*scale).
 type Mem struct {
 	Base     Reg

testdata/verify/bookkeep_amd64.s

@@ -0,0 +1,32 @@
+// The runtime bookkeeping statements: FUNCDATA and PCDATA contribute no
+// text bytes on any architecture, and amd64 now matches.  They sit between
+// real instructions here, with plain, static and offset symbol references
+// on the FUNCDATA lines, so the byte counts prove the zero contribution.
+
+#include "textflag.h"
+
+// func bookkeep(x int64) int64
+TEXT ·bookkeep(SB), NOSPLIT, $0-16
+	PCDATA   $0, $-1
+	MOVQ     x+0(FP), AX
+	PCDATA   $1, $-2
+	FUNCDATA $0, args_stackmap(SB)
+	ADDQ     $1, AX
+	FUNCDATA $5, arginfo0(SB)
+	PCDATA   $1, $3
+	MOVQ     AX, ret+8(FP)
+	FUNCDATA $1, externalfuncdata(SB)
+	PCDATA   $0, $0
+	RET
+
+// func bookkeepstatic() int64
+TEXT ·bookkeepstatic(SB), NOSPLIT, $0-8
+	// A static symbol and a defined data symbol as the funcdata target.
+	// (A symbol+offset target the toolchain itself refuses.)
+	FUNCDATA $2, fdtable<>(SB)
+	FUNCDATA $3, undefsym(SB)
+	MOVQ     $7, AX
+	MOVQ     AX, ret+0(FP)
+	RET
+
+GLOBL fdtable<>(SB), NOPTR, $16

testdata/verify/floatimm_amd64.s

@@ -0,0 +1,55 @@
+// Floating-point immediates on the SSE scalar paths: the constant is
+// rewritten into a read from a read-only pool symbol ($f64.<hex> or
+// $f32.<hex>, the IEEE-754 bits in the name), RIP-relative with the
+// displacement left to the relocation.  A positive zero on the moves
+// collapses to XORPS dst, dst; a negative zero keeps its sign bit and
+// takes the pool.  The parenthesised $(-1.0) spelling is the one
+// math/floor_amd64.s uses.  Every result is folded back so no
+// instruction is dead.
+
+#include "textflag.h"
+
+// func floatimm(x float64) float64
+TEXT ·floatimm(SB), NOSPLIT, $0-16
+	MOVQ  x+0(FP), AX
+	MOVQ  AX, X0
+	// The floor kernel's sign fold: the parenthesised negative spelling.
+	MOVSD $ (-1.0), X2
+	ANDPD X2, X0
+	// Positive and fractional constants on the scalar moves.
+	MOVSD $1.5, X3
+	MOVSD $0.5, X4
+	MOVSS $2.5, X5
+	MOVSS $-0.5, X6
+	// A positive zero collapses to XORPS; a negative zero does not.
+	MOVSD $0.0, X7
+	MOVSS $0.0, X8
+	MOVSD $-0.0, X9
+	// The scalar arithmetic reads the pool through r/m (hypot's shape).
+	ADDSD $1.0, X3
+	SUBSD $0.5, X4
+	MULSD $-2.5, X4
+	DIVSD $2.0, X3
+	ADDSS $0.25, X5
+	// Fold everything into one double.
+	ADDSD X5, X3
+	ADDSD X6, X3
+	ADDSD X7, X3
+	ADDSD X8, X3
+	ADDSD X9, X3
+	ADDSD X4, X3
+	ADDSD X0, X3
+	MOVSD X3, ret+8(FP)
+	RET
+
+// func floatimmfloat32() float32
+TEXT ·floatimmfloat32(SB), NOSPLIT, $0-4
+	// The single-width pool constants ride the F3 prefix.
+	MOVSS $1.0, X0
+	MOVSS $-1.0, X1
+	MOVSS $0.0, X2
+	ADDSS $0.5, X0
+	ADDSS X1, X0
+	ADDSS X2, X0
+	MOVSS X0, ret+0(FP)
+	RET

testdata/verify/quadreg_amd64.s

@@ -0,0 +1,44 @@
+// The quad-register instructions: the 4FMAPS family (V4FMADDPS,
+// V4FMADDSS, V4FNMADDPS, V4FNMADDSS) and the 4VNNIW pair (VP4DPWSSD,
+// VP4DPWSSDS).  The bracketed list's low register travels the inverted
+// V'VVVV field, the memory source keeps r/m, the opmask rides aaa and the
+// vector length follows the destination (512-bit for the ZMM forms,
+// 128-bit for the scalar ones) while the disp8xN multiplier stays 16 for
+// every member.  Every result is folded back so no instruction is dead.
+
+#include "textflag.h"
+
+// func quadf4(src *[16]uint32, n int) float32
+TEXT ·quadf4(SB), NOSPLIT, $0-20
+	MOVQ        src+0(FP), SI
+	MOVQ        n+8(FP), CX
+	// The packed 4-FMA form over four consecutive ZMM accumulators,
+	// masked with K2, K3 and unmasked alike; the displacements exercise
+	// the disp32 form and the disp8x16 compressed form.
+	V4FMADDPS   17(SI), [Z0-Z3], K2, Z0
+	V4FMADDPS   64(SI), [Z10-Z13], K2, Z1
+	V4FMADDPS   (SI), [Z20-Z23], Z2
+	V4FNMADDPS  96(SI), [Z1-Z4], K3, Z5
+	// The scalar form reads XMM lists and takes the 128-bit length; the
+	// displacement compresses by 16.
+	V4FMADDSS   7(AX), [X0-X3], K5, X22
+	V4FMADDSS   (DI), [X10-X13], K5, X23
+	V4FNMADDSS  16(SI), [X20-X23], K1, X24
+	// The 4-VNNI dot products, indexed source included.
+	VP4DPWSSD   15(DX)(BX*8), [Z2-Z5], K4, Z17
+	VP4DPWSSDS  -7(DI)(R8*1), [Z4-Z7], K1, Z31
+	VP4DPWSSD   (SI), [Z12-Z15], Z6
+	// Zeroing keeps the usual rule: a mask register must ride along.
+	V4FMADDPS.Z 128(SI), [Z24-Z27], K4, Z3
+	// Fold every accumulator into one scalar.
+	VPADDD      Z0, Z1, Z9
+	VPADDD      Z2, Z5, Z10
+	VPADDD      Z9, Z17, Z11
+	VPADDD      Z10, Z31, Z12
+	VPADDD      Z11, Z12, Z13
+	VPADDD      Z13, Z14, Z15
+	VADDSS      X22, X23, X0
+	VADDSS      X24, X0, X1
+	VADDSS      X1, X2, X3
+	VMOVSS      X3, ret+16(FP)
+	RET

verify/groundtruth_test.go

@@ -124,10 +124,16 @@ func TestGroundTruthAMD64(t *testing.T) {
 		"../testdata/verify/avx_amd64.s",
 		"../testdata/verify/pfx_amd64.s",
 		"../testdata/verify/vsib_amd64.s",
+		"../testdata/verify/floatimm_amd64.s",
+		"../testdata/verify/bookkeep_amd64.s",
+		"../testdata/verify/quadreg_amd64.s",
 		"../testdata/verify/rawdata_amd64.s",
 		"../testdata/verify/avx512_amd64.s",
 		"../testdata/verify/pfx_amd64.s",
 		"../testdata/verify/vsib_amd64.s",
+		"../testdata/verify/floatimm_amd64.s",
+		"../testdata/verify/bookkeep_amd64.s",
+		"../testdata/verify/quadreg_amd64.s",
 		"../testdata/verify/rawdata_amd64.s",
 		"../testdata/verify/avx512_amd64.s",
 		"../testdata/verify/doubleshift_amd64.s",