feat(asm): add the EVEX floating-point and conversion set

Assisted-by: Qwen 3.8 Max Preview

arch/arch.go

@@ -156,6 +156,16 @@ func (t *Table) Lookup(mnemonic string) (Instr, bool) {
 			}
 		}
 	}
+	// amd64 EVEX instructions take a .Z zeroing suffix (masking is written as
+	// an explicit K operand rather than a suffix); strip it so the base
+	// instruction is still recognised.
+	if t.Arch == AMD64 {
+		if base, ok := strings.CutSuffix(key, ".Z"); ok {
+			if in, found := t.instrs[base]; found {
+				return in, true
+			}
+		}
+	}
 	return Instr{}, false
 }
 

asm/encode.go

@@ -51,9 +51,16 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 
 	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
 	// B/W/L/Q/D is part of the mnemonic, not a size suffix, so dispatch
-	// before splitSize.
+	// before splitSize.  A ".Z" suffix requests EVEX zeroing.
-	if isVex(upper) || isEvex(upper) || upper == "KMOVW" {
+	base, zeroing, err := stripEvexSuffix(upper)
-		return e.encodeVec(upper, ops)
+	if err != nil {
+		return err
+	}
+	if isVex(base) || isEvex(base) || base == "KMOVW" {
+		return e.encodeVec(base, ops, zeroing)
+	}
+	if zeroing {
+		return fmt.Errorf("%s: the .Z suffix requires an EVEX instruction", mnem)
 	}
 
 	// CMOVcc and SETcc carry the condition in the mnemonic (CMOVLGT, SETNE).
@@ -121,14 +128,20 @@ func splitSize(upper string) (base string, size int) {
 // its own direction-dependent opcodes; KTESTW is always VEX; everything else
 // takes EVEX when an operand demands it (a ZMM or K register, or an
 // EVEX-only mnemonic) and VEX otherwise.
-func (e *enc) encodeVec(upper string, ops []Operand) error {
+func (e *enc) encodeVec(upper string, ops []Operand, zeroing bool) error {
 	if upper == "KMOVW" {
+		if zeroing {
+			return fmt.Errorf("KMOVW takes no .Z suffix")
+		}
 		return e.encodeKmovw(ops)
 	}
 	if upper == "KTESTW" || !evexRequired(upper, ops) {
+		if zeroing {
+			return fmt.Errorf("%s: the .Z suffix requires an EVEX instruction", upper)
+		}
 		return e.encodeVex(upper, ops)
 	}
-	return e.encodeEvex(upper, ops)
+	return e.encodeEvex(upper, ops, zeroing)
 }
 
 // --- instruction components -------------------------------------------------

asm/evex.go

@@ -3,14 +3,19 @@
 
 package asm
 
-import "fmt"
+import (
+	"fmt"
+	"strings"
+)
 
 // This file implements EVEX (AVX-512) instruction encoding: the four-byte
 // EVEX prefix with 5-bit vector register fields (Z0–Z31, X/Y 16–31), the
 // compressed disp8×N displacement, and the operand shapes the go-flac
-// AVX-512 kernels use.  Masking ({k}) and zeroing ({z}) are not supported —
+// AVX-512 kernels use plus the common floating-point and conversion set.
-// the kernels do not use them.  K-register operands (mask destinations,
+// Masking follows the Go assembler's spelling: an explicit K1–K7 operand
-// KMOVW, KTESTW) are.
+// anywhere among the operands (merging) plus a ".Z" mnemonic suffix for
+// zeroing.  K-register operands (mask destinations, KMOVW, KTESTW) are
+// supported too.
 
 // evexSpec describes one EVEX instruction's encoding parameters.  The form
 // field reuses the vexForm shapes, which carry over unchanged.
@@ -46,6 +51,31 @@ var evexTable = map[string]evexSpec{
 	// EVEX.128/256/512.66.0F.W1 — packed double arithmetic.
 	"VADDPD": {1, 0x58, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VMULPD": {1, 0x59, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VSUBPD": {1, 0x5C, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VDIVPD": {1, 0x5E, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VMINPD": {1, 0x5D, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VMAXPD": {1, 0x5F, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.66.0F.W1 — packed double unpack.
+	"VUNPCKLPD": {1, 0x14, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VUNPCKHPD": {1, 0x15, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
+	// EVEX.128.F2.0F.W1 — scalar double arithmetic (the packed opcodes with
+	// an F2 pp; the EVEX forms exist for masked and zeroing use).  The
+	// memory operand is a single double, so disp8×N = 8.
+	"VADDSD": {1, 0x58, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VSUBSD": {1, 0x5C, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VMULSD": {1, 0x59, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VDIVSD": {1, 0x5E, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VMINSD": {1, 0x5D, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VMAXSD": {1, 0x5F, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+
+	// EVEX.128.F3.0F.W0 — scalar single arithmetic (disp8×N = 4).
+	"VADDSS": {1, 0x58, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VSUBSS": {1, 0x5C, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VMULSS": {1, 0x59, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VDIVSS": {1, 0x5E, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VMINSS": {1, 0x5D, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VMAXSS": {1, 0x5F, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
 
 	// EVEX.512.66.0F3A — align (NDS + imm8).
 	"VALIGND": {3, 0x03, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
@@ -59,6 +89,36 @@ var evexTable = map[string]evexSpec{
 	// EVEX.128/256/512.F3.0F.W1 — signed qword to packed double (reg=dst,
 	// rm=src, no vvvv).
 	"VCVTQQ2PD": {1, 0xE6, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.F2.0F.W1 — duplicate the low double (reg=dst,
+	// rm=src, no vvvv): a 128-bit destination reads a single double from
+	// memory (disp8×8), the wider ones read the full operand.
+	"VMOVDDUP": {1, 0x12, 1, 3, -1, vexRM, [3]int{8, 32, 64}},
+	// EVEX.128/256/512.0F.W0 — signed dword to packed single (reg=dst,
+	// rm=src, no vvvv, no mandatory prefix — as in the VEX form).
+	"VCVTDQ2PS": {1, 0x5B, 0, 0, -1, vexRM, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.0F.W0 — packed single to packed double: the
+	// destination is twice the source width and sets the length; disp8×N
+	// follows the narrow memory source.  No F3 prefix: the Go assembler
+	// emits this instruction with pp = 00 (Intel's maps would call that
+	// undefined) and gasm reproduces the Go assembler's bytes — its machine
+	// code is the oracle, not the manual.
+	"VCVTPS2PD": {1, 0x5A, 0, 0, -1, vexRM, [3]int{8, 16, 32}},
+	// EVEX.128/256/512.F3.0F.W0 — signed dword to packed double (the EVEX
+	// form of the VEX instruction; the destination sets the length, disp8×N
+	// follows the narrow memory source).
+	"VCVTDQ2PD": {1, 0xE6, 0, 2, -1, vexRM, [3]int{8, 16, 32}},
+	// EVEX packed double → dword conversions: the source is the wide
+	// operand and the mnemonic fixes the length — the bare names are
+	// 512-bit only (ZMM source, XMM destination), the X/Y spellings are
+	// EVEX-128/256.  Exactly one slot of n is valid; it names the vector
+	// length (and the disp8×N multiplier) a register or memory source
+	// encodes.
+	"VCVTPD2DQ":   {1, 0xE6, 1, 3, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTTPD2DQ":  {1, 0xE6, 1, 1, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTPD2DQX":  {1, 0xE6, 1, 3, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTPD2DQY":  {1, 0xE6, 1, 3, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+	"VCVTTPD2DQX": {1, 0xE6, 1, 1, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTTPD2DQY": {1, 0xE6, 1, 1, -1, vexRMSrcLen, [3]int{0, 32, 0}},
 	// EVEX.128/256/512.66.0F38.W0 — sign-extend dwords to qwords; the memory
 	// operand is the narrow source, so disp8×N follows its size (8/16/32 for
 	// the xmm/ymm/zmm destination lengths).
@@ -74,6 +134,48 @@ var evexTable = map[string]evexSpec{
 	"VPMULLQ": {2, 0x40, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMD":  {2, 0x36, 0, 1, -1, vexNDS3, [3]int{0, 32, 64}},
 
+	// EVEX.128/256/512 — the wider integer set (AVX-512 F/BW): byte/word
+	// arithmetic, the bitwise ops with D/Q suffixes, min/max, averages and
+	// variable shifts.  All NDS form; W distinguishes element size.
+	"VPADDB":  {1, 0xFC, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPADDW":  {1, 0xFD, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSUBB":  {1, 0xF8, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSUBW":  {1, 0xF9, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMULLW": {1, 0xD5, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPAVGB":  {1, 0xE0, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPAVGW":  {1, 0xE3, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINUB": {1, 0xDA, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXUB": {1, 0xDE, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINSW": {1, 0xEA, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXSW": {1, 0xEE, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPANDD":  {1, 0xDB, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPANDQ":  {1, 0xDB, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPANDND": {1, 0xDF, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPANDNQ": {1, 0xDF, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINSB": {2, 0x38, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXSB": {2, 0x3C, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINSQ": {2, 0x39, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXSQ": {2, 0x3D, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINUW": {2, 0x3A, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXUW": {2, 0x3E, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINSD": {2, 0x39, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXSD": {2, 0x3D, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINUD": {2, 0x3B, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXUD": {2, 0x3F, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMINUQ": {2, 0x3B, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMAXUQ": {2, 0x3F, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSLLVD": {2, 0x47, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSLLVQ": {2, 0x47, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSRLVD": {2, 0x45, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSRLVQ": {2, 0x45, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSRAVD": {2, 0x46, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSRAVQ": {2, 0x46, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
+	// EVEX forms of instructions that also exist in VEX (selected when a ZMM
+	// or K register, or indices 16–31, demand EVEX).
+	"VPSHUFD": {1, 0x70, 0, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VPSHUFB": {2, 0x00, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
 	// EVEX.66.0F — immediate shift (VPSLLD /6).
 	"VPSLLD": {1, 0x72, 0, 1, 6, vexShiftImm, [3]int{16, 32, 64}},
 
@@ -117,6 +219,13 @@ var evexMoveTable = map[string]evexMoveSpec{
 	"VMOVDQU32": {1, 2, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
 	// EVEX.128/256/512.F3.0F.W1 — unaligned qword move.
 	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.F2.0F.W0 — unaligned byte move (byte/word moves use the
+	// F2 prefix, dword/qword moves F3; the element size only changes the tuple
+	// semantics).
+	"VMOVDQU8": {1, 3, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.F2.0F.W1 — unaligned word move (shares the qword
+	// encoding).
+	"VMOVDQU16": {1, 3, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
 	// EVEX.128/256/512.66.0F.W1 — unaligned packed double move.
 	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}},
 }
@@ -151,13 +260,75 @@ func evexRequired(upper string, ops []Operand) bool {
 	return false
 }
 
-// encodeEvex encodes an EVEX instruction with operands in Plan 9 order.
+// stripEvexSuffix splits a ".Z" zeroing suffix off the mnemonic.  It is the
-func (e *enc) encodeEvex(mnemUpper string, ops []Operand) error {
+// only EVEX suffix supported; Go writes masking as an explicit K operand, not
+// a suffix.
+func stripEvexSuffix(mnem string) (base string, zeroing bool, err error) {
+	i := strings.LastIndexByte(mnem, '.')
+	if i < 0 {
+		return mnem, false, nil
+	}
+	if mnem[i+1:] == "Z" {
+		return mnem[:i], true, nil
+	}
+	return "", false, fmt.Errorf("unsupported EVEX suffix %q", mnem[i+1:])
+}
+
+// splitMask extracts an explicit mask register (K1–K7) from the operand list,
+// returning the remaining operands and the mask index.  K0 is not a usable
+// mask (aaa = 0 means "no mask"), matching the assembler.
+func splitMask(ops []Operand) ([]Operand, int, error) {
+	var rest []Operand
+	mask := 0
+	for _, op := range ops {
+		if r, ok := op.(Reg); ok && r.mask {
+			if mask != 0 {
+				return nil, 0, fmt.Errorf("at most one mask register operand")
+			}
+			if r.idx == 0 {
+				return nil, 0, fmt.Errorf("K0 is not a usable mask register")
+			}
+			mask = r.idx
+			continue
+		}
+		rest = append(rest, op)
+	}
+	return rest, mask, nil
+}
+
+// encodeEvex encodes an EVEX instruction with operands in Plan 9 order.  The
+// mask, when present, is an explicit K1–K7 operand anywhere among the
+// operands; zeroing comes from the .Z mnemonic suffix and requires a mask.
+func (e *enc) encodeEvex(mnemUpper string, ops []Operand, zeroing bool) error {
+	// Mask-destination comparisons (VPCMPEQD …, K1): the last operand is the
+	// destination K register, and any mask sits among the preceding operands.
+	if spec, ok := evexTable[mnemUpper]; ok && spec.form == vexNDS3 && len(ops) > 0 {
+		if dst, ok := ops[len(ops)-1].(Reg); ok && dst.mask {
+			rest, mask, err := splitMask(ops[:len(ops)-1])
+			if err != nil {
+				return err
+			}
+			if zeroing && mask == 0 {
+				return fmt.Errorf("%s: zeroing (.Z) requires a mask register", mnemUpper)
+			}
+			return e.encodeEvexNDS3(spec, append(rest, dst), mask, zeroing)
+		}
+	}
+
+	rest, mask, err := splitMask(ops)
+	if err != nil {
+		return err
+	}
+	if zeroing && mask == 0 {
+		return fmt.Errorf("%s: zeroing (.Z) requires a mask register", mnemUpper)
+	}
+	ops = rest
+
 	if bs, ok := evexBcastTable[mnemUpper]; ok {
-		return e.encodeEvexBcast(bs, ops)
+		return e.encodeEvexBcast(bs, ops, mask, zeroing)
 	}
 	if ms, ok := evexMoveTable[mnemUpper]; ok {
-		return e.encodeEvexMove(mnemUpper, ms, ops)
+		return e.encodeEvexMove(mnemUpper, ms, ops, mask, zeroing)
 	}
 	spec, ok := evexTable[mnemUpper]
 	if !ok {
@@ -165,17 +336,21 @@ func (e *enc) encodeEvex(mnemUpper string, ops []Operand) error {
 	}
 	switch spec.form {
 	case vexNDS3:
-		return e.encodeEvexNDS3(spec, ops)
+		return e.encodeEvexNDS3(spec, ops, mask, zeroing)
 	case vexRM:
-		return e.encodeEvexRM(spec, ops)
+		return e.encodeEvexRM(spec, ops, mask, zeroing)
 	case vexRMRev:
-		return e.encodeEvexRMRev(spec, ops)
+		return e.encodeEvexRMRev(spec, ops, mask, zeroing)
+	case vexImmRM:
+		return e.encodeEvexImmRM(spec, ops, mask, zeroing)
 	case vexShiftImm:
-		return e.encodeEvexShiftImm(spec, ops)
+		return e.encodeEvexShiftImm(spec, ops, mask, zeroing)
 	case vexNDS3Imm:
-		return e.encodeEvexNDS3Imm(spec, ops)
+		return e.encodeEvexNDS3Imm(spec, ops, mask, zeroing)
 	case vexExtract:
-		return e.encodeEvexExtract(spec, ops)
+		return e.encodeEvexExtract(spec, ops, mask, zeroing)
+	case vexRMSrcLen:
+		return e.encodeEvexRMSrcLen(spec, ops, mask, zeroing)
 	}
 	return fmt.Errorf("unhandled EVEX form for %s", mnemUpper)
 }
@@ -183,7 +358,7 @@ func (e *enc) encodeEvex(mnemUpper string, ops []Operand) error {
 // encodeEvexNDS3 encodes the three-operand NDS form: OP src2, src1, dst.  The
 // destination may be an opmask register (VPCMPEQD), in which case the vector
 // length comes from the sources.
-func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 3 {
 		return fmt.Errorf("EVEX NDS instruction expects 3 operands, got %d", len(ops))
 	}
@@ -203,12 +378,12 @@ func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand) error {
 			ll = r.vecLenBit()
 		}
 	}
-	return e.emitEvexFields(spec, ll, dstReg.idx, vvvvReg.idx, src2)
+	return e.emitEvexFields(spec, ll, dstReg.idx, vvvvReg.idx, src2, mask, zeroing)
 }
 
 // encodeEvexRM encodes the two-operand form: OP src, dst (reg=dst, rm=src,
 // no vvvv), e.g. VCVTQQ2PD.
-func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 2 {
 		return fmt.Errorf("EVEX two-operand instruction expects 2 operands, got %d", len(ops))
 	}
@@ -217,12 +392,42 @@ func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand) error {
 	if !ok || !dstReg.isVec() {
 		return fmt.Errorf("EVEX destination must be a vector register")
 	}
-	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src)
+	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src, mask, zeroing)
+}
+
+// encodeEvexImmRM encodes the immediate shuffle form: OP $imm, src, dst
+// (reg = dst, rm = src, imm8), e.g. VPSHUFD.
+func (e *enc) encodeEvexImmRM(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("shuffle expects 3 operands ($imm, src, dst), got %d", len(ops))
+	}
+	imm, src, dst := ops[0], ops[1], ops[2]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("shuffle control must be an immediate")
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("shuffle destination must be a vector register")
+	}
+	ll := dstReg.vecLenBit()
+	if r, ok := src.(Reg); ok && r.isVec() {
+		ll = r.vecLenBit()
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	if err := e.emitEvexFields(spec, ll, dstReg.idx, -1, src, mask, zeroing); err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
 }
 
 // encodeEvexShiftImm encodes an immediate shift: OP $imm, src, dst
 // (ModRM.reg = /digit, vvvv = dst, rm = src, imm8), e.g. VPSRAD $31, Z3, Z5.
-func (e *enc) encodeEvexShiftImm(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexShiftImm(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 3 {
 		return fmt.Errorf("EVEX shift expects 3 operands ($imm, src, dst), got %d", len(ops))
 	}
@@ -243,7 +448,7 @@ func (e *enc) encodeEvexShiftImm(spec evexSpec, ops []Operand) error {
 	if err != nil {
 		return err
 	}
-	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), spec.opdigit, dstReg.idx, srcReg); err != nil {
+	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), spec.opdigit, dstReg.idx, srcReg, mask, zeroing); err != nil {
 		return err
 	}
 	e.out = append(e.out, immByte)
@@ -252,7 +457,7 @@ func (e *enc) encodeEvexShiftImm(spec evexSpec, ops []Operand) error {
 
 // encodeEvexNDS3Imm encodes OP $imm, src2, src1, dst (reg=dst, vvvv=src1,
 // rm=src2, imm8), e.g. VALIGND.
-func (e *enc) encodeEvexNDS3Imm(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexNDS3Imm(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 4 {
 		return fmt.Errorf("instruction expects 4 operands ($imm, src2, src1, dst), got %d", len(ops))
 	}
@@ -273,7 +478,7 @@ func (e *enc) encodeEvexNDS3Imm(spec evexSpec, ops []Operand) error {
 	if err != nil {
 		return err
 	}
-	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, vvvvReg.idx, src2); err != nil {
+	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, vvvvReg.idx, src2, mask, zeroing); err != nil {
 		return err
 	}
 	e.out = append(e.out, immByte)
@@ -282,7 +487,7 @@ func (e *enc) encodeEvexNDS3Imm(spec evexSpec, ops []Operand) error {
 
 // encodeEvexExtract encodes OP $imm, zsrc, ydst (reg=ZMM source, rm=YMM/memory
 // destination, imm8), e.g. VEXTRACTI64X4.
-func (e *enc) encodeEvexExtract(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexExtract(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 3 {
 		return fmt.Errorf("extract expects 3 operands ($imm, zsrc, ydst), got %d", len(ops))
 	}
@@ -299,7 +504,7 @@ func (e *enc) encodeEvexExtract(spec evexSpec, ops []Operand) error {
 	if err != nil {
 		return err
 	}
-	if err := e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst); err != nil {
+	if err := e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst, mask, zeroing); err != nil {
 		return err
 	}
 	e.out = append(e.out, immByte)
@@ -309,7 +514,7 @@ func (e *enc) encodeEvexExtract(spec evexSpec, ops []Operand) error {
 // encodeEvexMove encodes a two-operand EVEX move; a vector→vector move uses
 // the store-form opcode (reg = source, rm = destination), matching the Go
 // assembler.
-func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand) error {
+func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 2 {
 		return fmt.Errorf("EVEX move expects 2 operands, got %d", len(ops))
 	}
@@ -338,7 +543,47 @@ func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand) error
 		return fmt.Errorf("%s needs a vector register operand", mnem)
 	}
 	spec := evexSpec{mapSel: ms.mapSel, opcode: op, w: ms.w, pp: ms.pp, opdigit: -1, n: ms.n}
-	return e.emitEvexFields(spec, reg.vecLenBit(), reg.idx, -1, rm)
+	return e.emitEvexFields(spec, reg.vecLenBit(), reg.idx, -1, rm, mask, zeroing)
+}
+
+// encodeEvexRMSrcLen encodes a length-narrowing conversion: OP src, dst with
+// the destination always XMM and the length fixed by the mnemonic — the
+// single valid slot of spec.n names the vector length (and the disp8×N
+// multiplier) a register or memory source encodes.
+func (e *enc) encodeEvexRMSrcLen(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("conversion expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("EVEX destination must be a vector register")
+	}
+	ll, err := soleLen(spec.n)
+	if err != nil {
+		return err
+	}
+	return e.emitEvexFields(spec, ll, dstReg.idx, -1, src, mask, zeroing)
+}
+
+// soleLen returns the vector-length index of the single valid slot of n —
+// the length a length-fixed mnemonic (the EVEX conversion spellings) encodes
+// regardless of its operands.
+func soleLen(n [3]int) (int, error) {
+	ll := -1
+	for i, v := range n {
+		if v == 0 {
+			continue
+		}
+		if ll >= 0 {
+			return 0, fmt.Errorf("ambiguous vector-length table %v", n)
+		}
+		ll = i
+	}
+	if ll < 0 {
+		return 0, fmt.Errorf("empty vector-length table")
+	}
+	return ll, nil
 }
 
 // memOperand reports whether op is a memory reference (including a
@@ -353,7 +598,7 @@ func memOperand(op Operand) bool {
 
 // encodeEvexRMRev encodes the narrowing-store form: OP src, dst with the wide
 // source in the reg field and the narrow destination in r/m (VPMOVDW/QD).
-func (e *enc) encodeEvexRMRev(spec evexSpec, ops []Operand) error {
+func (e *enc) encodeEvexRMRev(spec evexSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 2 {
 		return fmt.Errorf("EVEX store instruction expects 2 operands, got %d", len(ops))
 	}
@@ -362,12 +607,12 @@ func (e *enc) encodeEvexRMRev(spec evexSpec, ops []Operand) error {
 	if !ok || !srcReg.isVec() {
 		return fmt.Errorf("EVEX source must be a vector register")
 	}
-	return e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst)
+	return e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst, mask, zeroing)
 }
 
 // encodeEvexBcast encodes VPBROADCASTD/Q: OP src, dst with the GPR or memory
 // source broadcast to every lane of the vector destination.
-func (e *enc) encodeEvexBcast(bs evexBcastSpec, ops []Operand) error {
+func (e *enc) encodeEvexBcast(bs evexBcastSpec, ops []Operand, mask int, zeroing bool) error {
 	if len(ops) != 2 {
 		return fmt.Errorf("broadcast expects 2 operands, got %d", len(ops))
 	}
@@ -386,14 +631,15 @@ func (e *enc) encodeEvexBcast(bs evexBcastSpec, ops []Operand) error {
 	default:
 		return fmt.Errorf("broadcast source must be a register or memory")
 	}
-	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src)
+	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src, mask, zeroing)
 }
 
 // emitEvexFields emits the EVEX prefix, opcode, ModR/M, SIB and displacement
 // (disp8×N compressed) for the given precomputed fields.  regIdx is the
 // unextended reg-field register index, or a /digit (0–7); vvvvIdx is the
-// vvvv register index, or -1 when unused.
+// vvvv register index, or -1 when unused.  mask (K1–K7, 0 = unmasked) and
-func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand) error {
+// zeroing fill the aaa and z bits of the P2 byte.
+func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand, mask int, zeroing bool) error {
 	if ll > 2 {
 		return fmt.Errorf("invalid vector length")
 	}
@@ -420,7 +666,8 @@ func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand)
 	var sb *sbRef
 	switch r := rm.(type) {
 	case Reg:
-		// ModRM.mod = 11: rm[3] extends via B̄, rm[4] via X̄.
+		// ModRM.mod = 11: rm[3] extends via B̄, and rm[4] via X̄ (the EVEX
+		// register-register quirk).
 		modrm = 0xC0 | (regIdx&7)<<3 | (r.idx & 7)
 		sib = -1
 		if r.idx&8 != 0 {
@@ -429,6 +676,9 @@ func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand)
 		if r.idx&16 != 0 {
 			xBar = 0
 		}
+		if r.idx&16 != 0 {
+			xBar = 0
+		}
 	case Mem:
 		var err error
 		modrm, sib, disp, xBar, bBar, err = memComponentsEvex(regIdx&7, r, spec.n[ll])
@@ -451,9 +701,13 @@ func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand)
 		return fmt.Errorf("invalid EVEX r/m operand")
 	}
 
+	z := 0
+	if zeroing {
+		z = 1
+	}
 	p0 := byte(rBar<<7 | xBar<<6 | bBar<<5 | rPrimeBar<<4 | spec.mapSel)
 	p1 := byte(spec.w<<7 | vBar<<3 | 1<<2 | spec.pp)
-	p2 := byte(ll<<5 | vPrimeBar<<3) // z = 0, b = 0, aaa = 0
+	p2 := byte(z<<7 | ll<<5 | vPrimeBar<<3 | mask) // z, L'L, b=0, V', aaa
 	e.out = append(e.out, 0x62, p0, p1, p2, spec.opcode, byte(modrm))
 	if sib >= 0 {
 		e.out = append(e.out, byte(sib))

asm/evex_test.go

@@ -65,6 +65,23 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VMOVDQU64 (SI)(R15*4),Z3", "VMOVDQU64", []Operand{Idx(SI, vreg(t, "R15"), 4, 0, 64), vreg(t, "Z3")}, "62b1fe486f1cbe"},
 		{"VMOVDQU64 Z0,4(SI)(AX*1)", "VMOVDQU64", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f1fe487f840604000000"},
 		{"VMOVDQU64 Z1,Z2", "VMOVDQU64", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fe487fca"},
+		// The wider AVX-512 F/BW integer set.
+		{"VPADDB Z1,Z2,Z3", "VPADDB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48fcd9"},
+		{"VPSUBW Z1,Z2,Z3", "VPSUBW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48f9d9"},
+		{"VPANDQ Z1,Z2,Z3", "VPANDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed48dbd9"},
+		{"VPANDND Z1,Z2,Z3", "VPANDND", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48dfd9"},
+		{"VPMULLW Z1,Z2,Z3", "VPMULLW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48d5d9"},
+		{"VPMINUB Z1,Z2,Z3", "VPMINUB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48dad9"},
+		{"VPMAXUQ Z1,Z2,Z3", "VPMAXUQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed483fd9"},
+		{"VPAVGW Z1,Z2,Z3", "VPAVGW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48e3d9"},
+		{"VPSLLVQ Z3,Z1,Z2", "VPSLLVQ", []Operand{vreg(t, "Z3"), vreg(t, "Z1"), vreg(t, "Z2")}, "62f2f54847d3"},
+		{"VPSRAVQ Z3,Z1,Z2", "VPSRAVQ", []Operand{vreg(t, "Z3"), vreg(t, "Z1"), vreg(t, "Z2")}, "62f2f54846d3"},
+		{"VPSHUFD $0x1B,Z1,Z2", "VPSHUFD", []Operand{Imm(0x1B), vreg(t, "Z1"), vreg(t, "Z2")}, "62f17d4870d11b"},
+		{"VPSHUFB Z1,Z2,Z3", "VPSHUFB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4800d9"},
+		{"VMOVDQU8 Z1,Z2", "VMOVDQU8", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17f487fca"},
+		{"VMOVDQU16 Z1,Z2", "VMOVDQU16", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff487fca"},
+		// Indices 16–31: rm[4] rides in X̄ for register operands.
+		{"VPSHUFD $1,X16,X17", "VPSHUFD", []Operand{Imm(1), vreg(t, "X16"), vreg(t, "X17")}, "62a17d0870c801"},
 		{"VMOVUPD (DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 0, 64), vreg(t, "Z14")}, "6271fd481037"},
 		{"VMOVUPD 64(DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 64, 64), vreg(t, "Z14")}, "6271fd48107701"},
 		// Conversions and narrowing stores (reg = wide source).
@@ -84,6 +101,32 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VPBROADCASTQ AX,Z9", "VPBROADCASTQ", []Operand{AX, vreg(t, "Z9")}, "6272fd487cc8"},
 		// Register indices 16–31 exist only in EVEX encodings.
 		{"VPBROADCASTD AX,Y30", "VPBROADCASTD", []Operand{AX, vreg(t, "Y30")}, "62627d287cf0"},
+		// Packed double arithmetic / unpack (EVEX forms carry W=1).
+		{"VSUBPD Z1,Z2,Z3", "VSUBPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed485cd9"},
+		{"VDIVPD Z4,Z5,Z6", "VDIVPD", []Operand{vreg(t, "Z4"), vreg(t, "Z5"), vreg(t, "Z6")}, "62f1d5485ef4"},
+		{"VMINPD Z7,Z8,Z9", "VMINPD", []Operand{vreg(t, "Z7"), vreg(t, "Z8"), vreg(t, "Z9")}, "6271bd485dcf"},
+		{"VMAXPD Z10,Z11,Z12", "VMAXPD", []Operand{vreg(t, "Z10"), vreg(t, "Z11"), vreg(t, "Z12")}, "6251a5485fe2"},
+		{"VUNPCKLPD Z1,Z2,Z3", "VUNPCKLPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed4814d9"},
+		{"VUNPCKHPD Z1,Z2,Z3", "VUNPCKHPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed4815d9"},
+		{"VSUBPD 64(AX),Z1,Z2", "VSUBPD", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5485c5001"},
+		{"VSUBPD Z17,Z18,Z19", "VSUBPD", []Operand{vreg(t, "Z17"), vreg(t, "Z18"), vreg(t, "Z19")}, "62a1ed405cd9"},
+		// VMOVDDUP — duplicate the low double; disp8×N = 64 at 512 bits, and
+		// X16/X17 force EVEX (the mod=11 rm[4] extension rides in X̄).
+		{"VMOVDDUP Z1,Z2", "VMOVDDUP", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff4812d1"},
+		{"VMOVDDUP 64(AX),Z1", "VMOVDDUP", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1")}, "62f1ff48124801"},
+		{"VMOVDDUP X16,X17", "VMOVDDUP", []Operand{vreg(t, "X16"), vreg(t, "X17")}, "62a1ff0812c8"},
+		// Conversions: DQ→PS, PS→PD (pp = 00, the Go assembler's choice),
+		// DQ→PD (the destination sets the length).
+		{"VCVTDQ2PS Z1,Z2", "VCVTDQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c485bd1"},
+		{"VCVTPS2PD Y1,Z2", "VCVTPS2PD", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17c485ad1"},
+		{"VCVTPS2PD 32(AX),Z2", "VCVTPS2PD", []Operand{Ptr(AX, 32, 32), vreg(t, "Z2")}, "62f17c485a5001"},
+		{"VCVTDQ2PD Y1,Z2", "VCVTDQ2PD", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17e48e6d1"},
+		// PD→DQ conversions: the source is the wide operand and fixes the
+		// length (ZMM source → L'L = 10 even with an XMM destination; a
+		// memory source takes the length the mnemonic's spelling implies).
+		{"VCVTPD2DQ Z1,Y2", "VCVTPD2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1ff48e6d1"},
+		{"VCVTPD2DQ 64(AX),Y2", "VCVTPD2DQ", []Operand{Ptr(AX, 64, 64), vreg(t, "Y2")}, "62f1ff48e65001"},
+		{"VCVTTPD2DQ Z3,Y4", "VCVTTPD2DQ", []Operand{vreg(t, "Z3"), vreg(t, "Y4")}, "62f1fd48e6e3"},
 	}
 	for _, c := range cases {
 		want := strings.ReplaceAll(c.want, " ", "")
@@ -110,6 +153,93 @@ func TestEvexGroundTruth(t *testing.T) {
 	}
 }
 
+// TestEvexMasking checks the AVX-512 mask operand (K1–K7, placed freely among
+// the operands) and the .Z zeroing suffix, byte for byte against the Go
+// assembler.
+func TestEvexMasking(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Masked arithmetic: K anywhere among the operands; .Z sets the z bit.
+		{"VPADDD.Z merging+zeroing", "VPADDD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K2"), vreg(t, "Z3")}, "62f16dcafed9"},
+		{"VPADDD merging", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f16d49fed9"},
+		{"VADDPD.Z", "VADDPD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1edca58d9"},
+		{"VPMINSD.Z", "VPMINSD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K5"), vreg(t, "Z3")}, "62f26dcd39d9"},
+		{"VPMINSQ.Z", "VPMINSQ.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K5"), vreg(t, "Z3")}, "62f2edcd39d9"},
+		// Masked immediate shift (K before the destination).
+		{"VPSRAD.Z", "VPSRAD.Z", []Operand{Imm(1), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f165c972e201"},
+		{"VPSLLD merge", "VPSLLD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1654a72f104"},
+		// Masked align.
+		{"VALIGND", "VALIGND", []Operand{Imm(12), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z4")}, "62f36d4b03e10c"},
+		// Masked conversion and extract.
+		{"VCVTQQ2PD.Z", "VCVTQQ2PD.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1fecae6d9"},
+		{"VEXTRACTI64X4", "VEXTRACTI64X4", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f3fd4a3bcb01"},
+		// Masked moves: K sits between the register and memory operands.
+		{"VMOVDQU8 store", "VMOVDQU8", []Operand{vreg(t, "Z1"), vreg(t, "K3"), Ptr(SI, 0, 64)}, "62f17f4b7f0e"},
+		{"VMOVDQU32 load", "VMOVDQU32", []Operand{Ptr(SI, 0, 64), vreg(t, "K4"), vreg(t, "Z1")}, "62f17e4c6f0e"},
+		{"VMOVDQU32 store", "VMOVDQU32", []Operand{vreg(t, "Z1"), vreg(t, "K4"), Ptr(DI, 0, 64)}, "62f17e4c7f0f"},
+		// Masked comparison with a K destination: dst K1, mask K2.
+		{"VPCMPEQD k-dst+mask", "VPCMPEQD", []Operand{vreg(t, "Z0"), vreg(t, "Z3"), vreg(t, "K2"), vreg(t, "K1")}, "62f1654a76c8"},
+		// Masked floating point: packed double, the scalar SD/SS forms (which
+		// exist under EVEX only for masked and zeroing use) and conversions.
+		{"VSUBPD.Z", "VSUBPD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z4")}, "62f1edcb5ce1"},
+		{"VADDSD merge", "VADDSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K3"), vreg(t, "X4")}, "62f1ef0b58e1"},
+		{"VSUBSD.Z", "VSUBSD.Z", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K5"), vreg(t, "X3")}, "62f1ef8d5cd9"},
+		{"VADDSS merge", "VADDSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K1"), vreg(t, "X3")}, "62f16e0958d9"},
+		{"VCVTPD2DQ merge", "VCVTPD2DQ", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f1ff4ae6d9"},
+		{"VCVTTPD2DQ.Z", "VCVTTPD2DQ.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f1fdcae6d9"},
+		{"VCVTDQ2PS.Z", "VCVTDQ2PS.Z", []Operand{vreg(t, "Z1"), vreg(t, "K4"), vreg(t, "Z2")}, "62f17ccc5bd1"},
+		{"VCVTDQ2PD merge", "VCVTDQ2PD", []Operand{vreg(t, "X1"), vreg(t, "K2"), vreg(t, "X3")}, "62f17e0ae6d9"},
+		{"VCVTDQ2PD.Z", "VCVTDQ2PD.Z", []Operand{vreg(t, "Y1"), vreg(t, "K2"), vreg(t, "Z2")}, "62f17ecae6d1"},
+		{"VCVTPS2PD.Z", "VCVTPS2PD.Z", []Operand{vreg(t, "Y1"), vreg(t, "K3"), vreg(t, "Z2")}, "62f17ccb5ad1"},
+	}
+	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: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		if i := len(want) - 2; i > 0 && want[i:] == ".Z" {
+			want = want[:i]
+		}
+		if inst.Op.String() != want {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+
+	// Error cases.
+	bad := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"zeroing without mask", "VPADDD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"K0 mask", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K0"), vreg(t, "Z3")}},
+		{"two masks", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "K1"), vreg(t, "K2"), vreg(t, "Z3")}},
+		{".Z on VEX-only", "VPSHUFD.Z", []Operand{Imm(1), vreg(t, "X0"), vreg(t, "X1")}},
+		{"unsupported suffix", "VPADDD.BCST", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"KMOVW.Z", "KMOVW.Z", []Operand{vreg(t, "K1"), vreg(t, "K2")}},
+	}
+	for _, c := range bad {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}
+
 // TestEvexErrors checks the EVEX-specific error paths.
 func TestEvexErrors(t *testing.T) {
 	cases := []struct {
@@ -125,7 +255,7 @@ func TestEvexErrors(t *testing.T) {
 		{"VPMOVDW src", "VPMOVDW", []Operand{AX, vreg(t, "Y0")}},
 		{"align arity", "VALIGND", []Operand{Imm(1), vreg(t, "Z0"), vreg(t, "Z1")}},
 		// VEX-only mnemonics reject registers only EVEX can encode.
-		{"VPSHUFD X16", "VPSHUFD", []Operand{Imm(1), vreg(t, "X16"), vreg(t, "X17")}},
+		{"VMOVMSKPS X16", "VMOVMSKPS", []Operand{vreg(t, "X16"), AX}},
 	}
 	for _, c := range cases {
 		if _, err := Encode(c.mnem, c.ops...); err == nil {

asm/vex.go

@@ -43,6 +43,13 @@ const (
 	// in ModRM.reg and the destination in r/m — the layout of the EVEX
 	// narrowing stores (VPMOVDW, VPMOVQD).
 	vexRMRev
+	// vexRMSrcLen is the two-operand conversion form `OP src, dst` whose
+	// vector length follows the source: the packed-double → dword
+	// conversions (VCVTPD2DQ/VCVTTPD2DQ and their X/Y spellings) narrow into
+	// an XMM destination, so the L bit rides with the wider source.  The
+	// mnemonic's spelling fixes the length (X = 128, Y = 256), which also
+	// covers a memory source.  ModRM.reg = dst, ModRM.rm = src, no vvvv.
+	vexRMSrcLen
 	// vexZero is the no-operand form (VZEROUPPER).
 	vexZero
 )
@@ -86,12 +93,29 @@ var vexTable = map[string]vexSpec{
 	// VEX.128/256.66.0F.WIG — packed double-precision arithmetic / logic.
 	"VADDPD":    {1, 0x58, 0, 1, -1, vexNDS3},
 	"VMULPD":    {1, 0x59, 0, 1, -1, vexNDS3},
+	"VSUBPD":    {1, 0x5C, 0, 1, -1, vexNDS3},
+	"VDIVPD":    {1, 0x5E, 0, 1, -1, vexNDS3},
+	"VMINPD":    {1, 0x5D, 0, 1, -1, vexNDS3},
+	"VMAXPD":    {1, 0x5F, 0, 1, -1, vexNDS3},
 	"VXORPD":    {1, 0x57, 0, 1, -1, vexNDS3},
 	"VUNPCKHPD": {1, 0x15, 0, 1, -1, vexNDS3},
+	"VUNPCKLPD": {1, 0x14, 0, 1, -1, vexNDS3},
 	// VEX.128.F2.0F.WIG — scalar double-precision arithmetic (the packed
 	// opcodes with an F2 pp).
 	"VADDSD": {1, 0x58, 0, 3, -1, vexNDS3},
+	"VSUBSD": {1, 0x5C, 0, 3, -1, vexNDS3},
 	"VMULSD": {1, 0x59, 0, 3, -1, vexNDS3},
+	"VDIVSD": {1, 0x5E, 0, 3, -1, vexNDS3},
+	"VMINSD": {1, 0x5D, 0, 3, -1, vexNDS3},
+	"VMAXSD": {1, 0x5F, 0, 3, -1, vexNDS3},
+	// VEX.128.F3.0F.WIG — scalar single-precision arithmetic (the packed
+	// opcodes with an F3 pp).
+	"VADDSS": {1, 0x58, 0, 2, -1, vexNDS3},
+	"VSUBSS": {1, 0x5C, 0, 2, -1, vexNDS3},
+	"VMULSS": {1, 0x59, 0, 2, -1, vexNDS3},
+	"VDIVSS": {1, 0x5E, 0, 2, -1, vexNDS3},
+	"VMINSS": {1, 0x5D, 0, 2, -1, vexNDS3},
+	"VMAXSS": {1, 0x5F, 0, 2, -1, vexNDS3},
 	// VEX.128/256.66.0F38.W1 — fused multiply-add (NDS form).
 	"VFMADD231PD": {2, 0xB8, 1, 1, -1, vexNDS3},
 
@@ -105,6 +129,19 @@ var vexTable = map[string]vexSpec{
 	// VEX.128/256.F3.0F.WIG — signed dword to packed double conversion
 	// (reg=dst, rm=src, no vvvv; the length follows the destination).
 	"VCVTDQ2PD": {1, 0xE6, 0, 2, -1, vexRM},
+	// VEX.128/256.0F.WIG — signed dword to packed single conversion
+	// (reg=dst, rm=src, no vvvv, no mandatory prefix).
+	"VCVTDQ2PS": {1, 0x5B, 0, 0, -1, vexRM},
+	// VEX.128/256.0F.WIG — packed single to packed double conversion
+	// (reg=dst, rm=src; the destination is the wide operand and sets the
+	// length).  Intel's maps prescribe the F3 prefix here (VEX.pp = 10), but
+	// the Go assembler emits the instruction with pp = 00, and gasm follows
+	// the Go assembler's bytes — its machine code is the oracle, not the
+	// manual.
+	"VCVTPS2PD": {1, 0x5A, 0, 0, -1, vexRM},
+	// VEX.128.F2.0F.WIG — duplicate the low double of each 128-bit lane
+	// (reg=dst, rm=src, no vvvv; the length follows the destination).
+	"VMOVDDUP": {1, 0x12, 0, 3, -1, vexRM},
 	// VEX.128/256.66.0F.WIG — move mask to a GPR (reg=gpr dst, rm=vec src).
 	"VPMOVMSKB": {1, 0xD7, 0, 1, -1, vexRM},
 	"VMOVMSKPS": {1, 0x50, 0, 0, -1, vexRM}, // no 66 prefix (that would be VMOVMSKPD)
@@ -138,6 +175,25 @@ var vexTable = map[string]vexSpec{
 
 	// VEX.128.0F.W0 — mask-register test (KTESTW k1, k2: reg = dst, rm = src).
 	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
+
+	// VEX.F2.0F — packed double to packed dword conversions, truncating and
+	// non-truncating.  The destination is always XMM; the X/Y spellings fix
+	// the source length (XMM/YMM), and VEX.L follows it — see vexSrcLen.
+	"VCVTPD2DQX":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
+	"VCVTPD2DQY":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
+	"VCVTTPD2DQX": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
+	"VCVTTPD2DQY": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
+}
+
+// vexSrcLen maps a source-length conversion mnemonic (the X/Y spellings of
+// the packed-double → dword conversions) to its fixed vector length:
+// X = 128 (L = 0), Y = 256 (L = 1).  The spelling fixes the length even for
+// a memory source, matching the Go assembler's ytab.
+var vexSrcLen = map[string]int{
+	"VCVTPD2DQX":  0,
+	"VCVTPD2DQY":  1,
+	"VCVTTPD2DQX": 0,
+	"VCVTTPD2DQY": 1,
 }
 
 // vexVarShift maps the shift mnemonics to their variable-count opcode — the
@@ -230,6 +286,8 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 		return e.encodeVexNDS3Imm(spec, ops)
 	case vexExtract:
 		return e.encodeVexExtract(spec, ops)
+	case vexRMSrcLen:
+		return e.encodeVexRMSrcLen(mnemUpper, spec, ops)
 	case vexZero:
 		return e.encodeVexZero(mnemUpper, spec, ops)
 	}
@@ -295,6 +353,32 @@ func (e *enc) encodeVexRM(spec vexSpec, ops []Operand) error {
 	return e.emitVexFields(spec, l, regField, rBit, 15, src)
 }
 
+// encodeVexRMSrcLen encodes a length-narrowing conversion: OP src, dst with
+// the destination always XMM and the VEX.L bit following the source — fixed
+// by the mnemonic's spelling (VCVTPD2DQX = 128, VCVTPD2DQY = 256) even when
+// the source is memory.
+func (e *enc) encodeVexRMSrcLen(mnem string, spec vexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("conversion expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a vector register")
+	}
+	ll, ok := vexSrcLen[mnem]
+	if !ok {
+		return fmt.Errorf("no fixed vector length for %s", mnem)
+	}
+	regField := dstReg.idx & 7
+	rBit := 0
+	if dstReg.idx >= 8 {
+		rBit = 1
+	}
+	// An unused vvvv field must be stored as all ones (v̄vvv = 1111).
+	return e.emitVexFields(spec, ll, regField, rBit, 15, src)
+}
+
 // encodeVexShiftImm encodes an immediate-shift instruction: OP $imm, src, dst.
 // The destination is carried in VEX.vvvv, the source in ModRM.rm, and the
 // shift kind in the ModRM.reg /digit.

asm/vex_test.go

@@ -160,78 +160,117 @@ func TestVexGroundTruth(t *testing.T) {
 		mnem   string
 		ops    []Operand
 		want   string
+		wantOp string // decoded mnemonic, when it differs from mnem (the X/Y spellings)
 	}{
 		// Three-operand NDS form.
-		{"VPADDQ Y8,Y9,Y8", "VPADDQ", []Operand{vreg(t, "Y8"), vreg(t, "Y9"), vreg(t, "Y8")}, "c44135d4c0"},
+		{"VPADDQ Y8,Y9,Y8", "VPADDQ", []Operand{vreg(t, "Y8"), vreg(t, "Y9"), vreg(t, "Y8")}, "c44135d4c0", ""},
-		{"VPADDQ X9,X8,X8", "VPADDQ", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c44139d4c1"},
+		{"VPADDQ X9,X8,X8", "VPADDQ", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c44139d4c1", ""},
-		{"VPXOR X7,X7,X7", "VPXOR", []Operand{vreg(t, "X7"), vreg(t, "X7"), vreg(t, "X7")}, "c5c1efff"},
+		{"VPXOR X7,X7,X7", "VPXOR", []Operand{vreg(t, "X7"), vreg(t, "X7"), vreg(t, "X7")}, "c5c1efff", ""},
-		{"VPSHUFB Y1,Y2,Y3", "VPSHUFB", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d00d9"},
+		{"VPSHUFB Y1,Y2,Y3", "VPSHUFB", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d00d9", ""},
-		{"VPMULLD Y1,Y2,Y3", "VPMULLD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d40d9"},
+		{"VPMULLD Y1,Y2,Y3", "VPMULLD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d40d9", ""},
-		{"VPUNPCKLDQ Y4,Y3,Y5", "VPUNPCKLDQ", []Operand{vreg(t, "Y4"), vreg(t, "Y3"), vreg(t, "Y5")}, "c5e562ec"},
+		{"VPUNPCKLDQ Y4,Y3,Y5", "VPUNPCKLDQ", []Operand{vreg(t, "Y4"), vreg(t, "Y3"), vreg(t, "Y5")}, "c5e562ec", ""},
-		{"VPERMD Y1,Y2,Y3", "VPERMD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d36d9"},
+		{"VPERMD Y1,Y2,Y3", "VPERMD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d36d9", ""},
 		// Floating point (packed and scalar) and FMA — same NDS form, the pp
 		// bits and map select the operation.
-		{"VADDPD Y9,Y8,Y8", "VADDPD", []Operand{vreg(t, "Y9"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d58c1"},
+		{"VADDPD Y9,Y8,Y8", "VADDPD", []Operand{vreg(t, "Y9"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d58c1", ""},
-		{"VADDPD X1,X2,X3", "VADDPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e958d9"},
+		{"VADDPD X1,X2,X3", "VADDPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e958d9", ""},
-		{"VMULPD Y12,Y12,Y12", "VMULPD", []Operand{vreg(t, "Y12"), vreg(t, "Y12"), vreg(t, "Y12")}, "c4411d59e4"},
+		{"VMULPD Y12,Y12,Y12", "VMULPD", []Operand{vreg(t, "Y12"), vreg(t, "Y12"), vreg(t, "Y12")}, "c4411d59e4", ""},
-		{"VXORPD Y8,Y8,Y8", "VXORPD", []Operand{vreg(t, "Y8"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d57c0"},
+		{"VXORPD Y8,Y8,Y8", "VXORPD", []Operand{vreg(t, "Y8"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d57c0", ""},
-		{"VUNPCKHPD X8,X8,X9", "VUNPCKHPD", []Operand{vreg(t, "X8"), vreg(t, "X8"), vreg(t, "X9")}, "c4413915c8"},
+		{"VUNPCKHPD X8,X8,X9", "VUNPCKHPD", []Operand{vreg(t, "X8"), vreg(t, "X8"), vreg(t, "X9")}, "c4413915c8", ""},
-		{"VADDSD X9,X8,X8", "VADDSD", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c4413b58c1"},
+		{"VADDSD X9,X8,X8", "VADDSD", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c4413b58c1", ""},
-		{"VMULSD X0,X1,X1", "VMULSD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X1")}, "c5f359c8"},
+		{"VMULSD X0,X1,X1", "VMULSD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X1")}, "c5f359c8", ""},
-		{"VFMADD231PD Y14,Y12,Y8", "VFMADD231PD", []Operand{vreg(t, "Y14"), vreg(t, "Y12"), vreg(t, "Y8")}, "c4429db8c6"},
+		{"VFMADD231PD Y14,Y12,Y8", "VFMADD231PD", []Operand{vreg(t, "Y14"), vreg(t, "Y12"), vreg(t, "Y8")}, "c4429db8c6", ""},
-		{"VFMADD231PD (DI),Y12,Y8", "VFMADD231PD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y12"), vreg(t, "Y8")}, "c4629db807"},
+		{"VFMADD231PD (DI),Y12,Y8", "VFMADD231PD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y12"), vreg(t, "Y8")}, "c4629db807", ""},
 		// Two-operand reg/rm form (v̄vvv must be 1111).
-		{"VPMOVSXDQ X0,Y4", "VPMOVSXDQ", []Operand{vreg(t, "X0"), vreg(t, "Y4")}, "c4e27d25e0"},
+		{"VPMOVSXDQ X0,Y4", "VPMOVSXDQ", []Operand{vreg(t, "X0"), vreg(t, "Y4")}, "c4e27d25e0", ""},
-		{"VPMOVSXWD (SI),Y0", "VPMOVSXWD", []Operand{Ptr(SI, 0, 8), vreg(t, "Y0")}, "c4e27d2306"},
+		{"VPMOVSXWD (SI),Y0", "VPMOVSXWD", []Operand{Ptr(SI, 0, 8), vreg(t, "Y0")}, "c4e27d2306", ""},
-		{"VPBROADCASTD X0,Y15", "VPBROADCASTD", []Operand{vreg(t, "X0"), vreg(t, "Y15")}, "c4627d58f8"},
+		{"VPBROADCASTD X0,Y15", "VPBROADCASTD", []Operand{vreg(t, "X0"), vreg(t, "Y15")}, "c4627d58f8", ""},
-		{"VCVTDQ2PD X12,Y12", "VCVTDQ2PD", []Operand{vreg(t, "X12"), vreg(t, "Y12")}, "c4417ee6e4"},
+		{"VCVTDQ2PD X12,Y12", "VCVTDQ2PD", []Operand{vreg(t, "X12"), vreg(t, "Y12")}, "c4417ee6e4", ""},
-		{"VCVTDQ2PD (SI),Y4", "VCVTDQ2PD", []Operand{Ptr(SI, 0, 16), vreg(t, "Y4")}, "c5fee626"},
+		{"VCVTDQ2PD (SI),Y4", "VCVTDQ2PD", []Operand{Ptr(SI, 0, 16), vreg(t, "Y4")}, "c5fee626", ""},
-		{"VPMOVMSKB X11,AX", "VPMOVMSKB", []Operand{vreg(t, "X11"), AX}, "c4c179d7c3"},
+		{"VPMOVMSKB X11,AX", "VPMOVMSKB", []Operand{vreg(t, "X11"), AX}, "c4c179d7c3", ""},
-		{"VMOVMSKPS Y7,AX", "VMOVMSKPS", []Operand{vreg(t, "Y7"), AX}, "c5fc50c7"},
+		{"VMOVMSKPS Y7,AX", "VMOVMSKPS", []Operand{vreg(t, "Y7"), AX}, "c5fc50c7", ""},
 		// Immediate shifts.
-		{"VPSLLD $1,Y3,Y4", "VPSLLD", []Operand{Imm(1), vreg(t, "Y3"), vreg(t, "Y4")}, "c5dd72f301"},
+		{"VPSLLD $1,Y3,Y4", "VPSLLD", []Operand{Imm(1), vreg(t, "Y3"), vreg(t, "Y4")}, "c5dd72f301", ""},
-		{"VPSRLQ $2,Y5,Y6", "VPSRLQ", []Operand{Imm(2), vreg(t, "Y5"), vreg(t, "Y6")}, "c5cd73d502"},
+		{"VPSRLQ $2,Y5,Y6", "VPSRLQ", []Operand{Imm(2), vreg(t, "Y5"), vreg(t, "Y6")}, "c5cd73d502", ""},
 		// Variable-count shifts: the count lives in an XMM register or memory
 		// and the instruction takes the NDS form.
-		{"VPSRLQ X0,Y8,Y8", "VPSRLQ", []Operand{vreg(t, "X0"), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd3c0"},
+		{"VPSRLQ X0,Y8,Y8", "VPSRLQ", []Operand{vreg(t, "X0"), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd3c0", ""},
-		{"VPSRLQ (AX),Y8,Y8", "VPSRLQ", []Operand{Ptr(AX, 0, 16), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd300"},
+		{"VPSRLQ (AX),Y8,Y8", "VPSRLQ", []Operand{Ptr(AX, 0, 16), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd300", ""},
-		{"VPSLLD X0,Y1,Y2", "VPSLLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f2d0"},
+		{"VPSLLD X0,Y1,Y2", "VPSLLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f2d0", ""},
-		{"VPSRLD X0,Y1,Y2", "VPSRLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5d2d0"},
+		{"VPSRLD X0,Y1,Y2", "VPSRLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5d2d0", ""},
-		{"VPSRAD X0,Y1,Y2", "VPSRAD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5e2d0"},
+		{"VPSRAD X0,Y1,Y2", "VPSRAD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5e2d0", ""},
-		{"VPSLLQ X0,Y1,Y2", "VPSLLQ", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f3d0"},
+		{"VPSLLQ X0,Y1,Y2", "VPSLLQ", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f3d0", ""},
 		// Immediate shuffle (reg=dst, rm=src, imm8).
-		{"VPSHUFD $0xEE,X8,X9", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "X8"), vreg(t, "X9")}, "c4417970c8ee"},
+		{"VPSHUFD $0xEE,X8,X9", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "X8"), vreg(t, "X9")}, "c4417970c8ee", ""},
-		{"VPSHUFD $0xEE,Y1,Y2", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "Y1"), vreg(t, "Y2")}, "c5fd70d1ee"},
+		{"VPSHUFD $0xEE,Y1,Y2", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "Y1"), vreg(t, "Y2")}, "c5fd70d1ee", ""},
-		{"VPERMQ $0x1B,Y1,Y2", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e3fd00d11b"},
+		{"VPERMQ $0x1B,Y1,Y2", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e3fd00d11b", ""},
-		{"VPERMQ $0x1B,Y11,Y12", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y11"), vreg(t, "Y12")}, "c443fd00e31b"},
+		{"VPERMQ $0x1B,Y11,Y12", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y11"), vreg(t, "Y12")}, "c443fd00e31b", ""},
 		// Three-operand + immediate (reg=dst, vvvv=src1, rm=src2, imm8).
-		{"VSHUFPD $1,X1,X2,X3", "VSHUFPD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e9c6d901"},
+		{"VSHUFPD $1,X1,X2,X3", "VSHUFPD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e9c6d901", ""},
-		{"VSHUFPD $1,Y1,Y2,Y3", "VSHUFPD", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5edc6d901"},
+		{"VSHUFPD $1,Y1,Y2,Y3", "VSHUFPD", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5edc6d901", ""},
-		{"VPERM2I128 $0x31,Y1,Y2,Y3", "VPERM2I128", []Operand{Imm(0x31), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e36d46d931"},
+		{"VPERM2I128 $0x31,Y1,Y2,Y3", "VPERM2I128", []Operand{Imm(0x31), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e36d46d931", ""},
-		{"VINSERTI128 $1,X5,Y1,Y2", "VINSERTI128", []Operand{Imm(1), vreg(t, "X5"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37538d501"},
+		{"VINSERTI128 $1,X5,Y1,Y2", "VINSERTI128", []Operand{Imm(1), vreg(t, "X5"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37538d501", ""},
 		// Lane extract (reg=YMM source, rm=XMM/memory destination, imm8).
-		{"VEXTRACTI128 $1,Y8,X9", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d39c101"},
+		{"VEXTRACTI128 $1,Y8,X9", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d39c101", ""},
-		{"VEXTRACTI128 $1,Y8,(DI)", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), Ptr(DI, 0, 16)}, "c4637d390701"},
+		{"VEXTRACTI128 $1,Y8,(DI)", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), Ptr(DI, 0, 16)}, "c4637d390701", ""},
-		{"VEXTRACTF128 $1,Y8,X9", "VEXTRACTF128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d19c101"},
+		{"VEXTRACTF128 $1,Y8,X9", "VEXTRACTF128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d19c101", ""},
 		// Moves — each direction picks its own opcode and VEX.W.
-		{"VMOVDQU (SI),Y1", "VMOVDQU", []Operand{Ptr(SI, 0, 32), vreg(t, "Y1")}, "c5fe6f0e"},
+		{"VMOVDQU (SI),Y1", "VMOVDQU", []Operand{Ptr(SI, 0, 32), vreg(t, "Y1")}, "c5fe6f0e", ""},
-		{"VMOVDQU Y3,(DI)", "VMOVDQU", []Operand{vreg(t, "Y3"), Ptr(DI, 0, 32)}, "c5fe7f1f"},
+		{"VMOVDQU Y3,(DI)", "VMOVDQU", []Operand{vreg(t, "Y3"), Ptr(DI, 0, 32)}, "c5fe7f1f", ""},
-		{"VMOVDQU X1,X2", "VMOVDQU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa7fca"},
+		{"VMOVDQU X1,X2", "VMOVDQU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa7fca", ""},
-		{"VMOVUPD (DI),Y14", "VMOVUPD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y14")}, "c57d1037"},
+		{"VMOVUPD (DI),Y14", "VMOVUPD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y14")}, "c57d1037", ""},
-		{"VMOVUPD Y14,(DI)", "VMOVUPD", []Operand{vreg(t, "Y14"), Ptr(DI, 0, 32)}, "c57d1137"},
+		{"VMOVUPD Y14,(DI)", "VMOVUPD", []Operand{vreg(t, "Y14"), Ptr(DI, 0, 32)}, "c57d1137", ""},
-		{"VMOVUPD X1,X2", "VMOVUPD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f911ca"},
+		{"VMOVUPD X1,X2", "VMOVUPD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f911ca", ""},
-		{"VMOVQ X8,AX", "VMOVQ", []Operand{vreg(t, "X8"), AX}, "c461f97ec0"},
+		{"VMOVQ X8,AX", "VMOVQ", []Operand{vreg(t, "X8"), AX}, "c461f97ec0", ""},
-		{"VMOVQ AX,X9", "VMOVQ", []Operand{AX, vreg(t, "X9")}, "c461f96ec8"},
+		{"VMOVQ AX,X9", "VMOVQ", []Operand{AX, vreg(t, "X9")}, "c461f96ec8", ""},
-		{"VMOVQ X8,(DI)", "VMOVQ", []Operand{vreg(t, "X8"), Ptr(DI, 0, 8)}, "c461f97e07"},
+		{"VMOVQ X8,(DI)", "VMOVQ", []Operand{vreg(t, "X8"), Ptr(DI, 0, 8)}, "c461f97e07", ""},
-		{"VMOVQ (SI),X9", "VMOVQ", []Operand{Ptr(SI, 0, 8), vreg(t, "X9")}, "c461f96e0e"},
+		{"VMOVQ (SI),X9", "VMOVQ", []Operand{Ptr(SI, 0, 8), vreg(t, "X9")}, "c461f96e0e", ""},
-		{"VMOVQ X8,X2", "VMOVQ", []Operand{vreg(t, "X8"), vreg(t, "X2")}, "c579d6c2"},
+		{"VMOVQ X8,X2", "VMOVQ", []Operand{vreg(t, "X8"), vreg(t, "X2")}, "c579d6c2", ""},
-		{"VMOVQ X2,X8", "VMOVQ", []Operand{vreg(t, "X2"), vreg(t, "X8")}, "c4c179d6d0"},
+		{"VMOVQ X2,X8", "VMOVQ", []Operand{vreg(t, "X2"), vreg(t, "X8")}, "c4c179d6d0", ""},
-		{"VMOVD X0,(SI)", "VMOVD", []Operand{vreg(t, "X0"), Ptr(SI, 0, 4)}, "c5f97e06"},
+		{"VMOVD X0,(SI)", "VMOVD", []Operand{vreg(t, "X0"), Ptr(SI, 0, 4)}, "c5f97e06", ""},
-		{"VMOVD AX,X0", "VMOVD", []Operand{AX, vreg(t, "X0")}, "c5f96ec0"},
+		{"VMOVD AX,X0", "VMOVD", []Operand{AX, vreg(t, "X0")}, "c5f96ec0", ""},
-		{"VMOVSD (SI),X8", "VMOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X8")}, "c57b1006"},
+		{"VMOVSD (SI),X8", "VMOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X8")}, "c57b1006", ""},
-		{"VMOVSD X8,(SI)", "VMOVSD", []Operand{vreg(t, "X8"), Ptr(SI, 0, 8)}, "c57b1106"},
+		{"VMOVSD X8,(SI)", "VMOVSD", []Operand{vreg(t, "X8"), Ptr(SI, 0, 8)}, "c57b1106", ""},
+		// Packed double arithmetic and unpack — the NDS form, the opcode
+		// selects the operation.
+		{"VSUBPD Y1,Y2,Y3", "VSUBPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed5cd9", ""},
+		{"VDIVPD X1,X2,X3", "VDIVPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e95ed9", ""},
+		{"VMINPD Y1,Y2,Y3", "VMINPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed5dd9", ""},
+		{"VMAXPD X4,X5,X6", "VMAXPD", []Operand{vreg(t, "X4"), vreg(t, "X5"), vreg(t, "X6")}, "c5d15ff4", ""},
+		{"VUNPCKLPD X1,X2,X3", "VUNPCKLPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e914d9", ""},
+		{"VUNPCKLPD Y1,Y2,Y3", "VUNPCKLPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed14d9", ""},
+		{"VSUBPD (AX),X1,X2", "VSUBPD", []Operand{Ptr(AX, 0, 16), vreg(t, "X1"), vreg(t, "X2")}, "c5f15c10", ""},
+		// Scalar double and single arithmetic (F2 / F3 pp, 128-bit only).
+		{"VSUBSD X1,X2,X3", "VSUBSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5eb5cd9", ""},
+		{"VDIVSD X7,X1,X2", "VDIVSD", []Operand{vreg(t, "X7"), vreg(t, "X1"), vreg(t, "X2")}, "c5f35ed7", ""},
+		{"VMINSD X1,X2,X3", "VMINSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5eb5dd9", ""},
+		{"VMAXSD X3,X4,X5", "VMAXSD", []Operand{vreg(t, "X3"), vreg(t, "X4"), vreg(t, "X5")}, "c5db5feb", ""},
+		{"VADDSS X1,X2,X3", "VADDSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea58d9", ""},
+		{"VSUBSS X1,X2,X3", "VSUBSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5cd9", ""},
+		{"VMULSS X9,X10,X11", "VMULSS", []Operand{vreg(t, "X9"), vreg(t, "X10"), vreg(t, "X11")}, "c4412a59d9", ""},
+		{"VDIVSS X1,X2,X3", "VDIVSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5ed9", ""},
+		{"VMINSS X6,X7,X8", "VMINSS", []Operand{vreg(t, "X6"), vreg(t, "X7"), vreg(t, "X8")}, "c5425dc6", ""},
+		{"VMAXSS X1,X2,X3", "VMAXSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5fd9", ""},
+		{"VADDSD 8(AX),X1,X2", "VADDSD", []Operand{Ptr(AX, 8, 8), vreg(t, "X1"), vreg(t, "X2")}, "c5f3585008", ""},
+		// VMOVDDUP — duplicate the low double (reg=dst, rm=src, F2 pp).
+		{"VMOVDDUP X1,X2", "VMOVDDUP", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fb12d1", ""},
+		{"VMOVDDUP Y1,Y2", "VMOVDDUP", []Operand{vreg(t, "Y1"), vreg(t, "Y2")}, "c5ff12d1", ""},
+		{"VMOVDDUP 8(AX),X1", "VMOVDDUP", []Operand{Ptr(AX, 8, 8), vreg(t, "X1")}, "c5fb124808", ""},
+		// Conversions: DQ→PS (no prefix), PS→PD (Go emits it without the F3
+		// prefix — see the table comment), DQ→PD.
+		{"VCVTDQ2PS X1,X2", "VCVTDQ2PS", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85bd1", ""},
+		{"VCVTDQ2PS Y3,Y4", "VCVTDQ2PS", []Operand{vreg(t, "Y3"), vreg(t, "Y4")}, "c5fc5be3", ""},
+		{"VCVTPS2PD X1,X2", "VCVTPS2PD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85ad1", ""},
+		{"VCVTPS2PD X1,Y2", "VCVTPS2PD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c5fc5ad1", ""},
+		// PD→DQ conversions: the X/Y spellings fix the source length and the
+		// destination is always XMM; the decoder reports the base mnemonic.
+		{"VCVTPD2DQX X1,X2", "VCVTPD2DQX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fbe6d1", "VCVTPD2DQ"},
+		{"VCVTPD2DQY Y1,X2", "VCVTPD2DQY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "c5ffe6d1", "VCVTPD2DQ"},
+		{"VCVTTPD2DQX X3,X4", "VCVTTPD2DQX", []Operand{vreg(t, "X3"), vreg(t, "X4")}, "c5f9e6e3", "VCVTTPD2DQ"},
+		{"VCVTTPD2DQY Y5,X6", "VCVTTPD2DQY", []Operand{vreg(t, "Y5"), vreg(t, "X6")}, "c5fde6f5", "VCVTTPD2DQ"},
+		{"VCVTPD2DQY (AX),X1", "VCVTPD2DQY", []Operand{Ptr(AX, 0, 32), vreg(t, "X1")}, "c5ffe608", "VCVTPD2DQ"},
 		// No-operand.
-		{"VZEROUPPER", "VZEROUPPER", nil, "c5f877"},
+		{"VZEROUPPER", "VZEROUPPER", nil, "c5f877", ""},
 	}
 	for _, c := range cases {
 		code, err := Encode(c.mnem, c.ops...)
@@ -251,7 +290,11 @@ func TestVexGroundTruth(t *testing.T) {
 		if inst.Len != len(code) {
 			t.Errorf("%s: Decode consumed %d of %d bytes", c.name, inst.Len, len(code))
 		}
-		if inst.Op.String() != c.mnem {
+		wantOp := c.wantOp
+		if wantOp == "" {
+			wantOp = c.mnem
+		}
+		if inst.Op.String() != wantOp {
 			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
 		}
 	}

cmd/gasm/main.go

@@ -28,7 +28,7 @@ import (
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.7.0"
+var version = "0.10.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -49,31 +49,71 @@ func main() {
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
 	case "version", "--version", "-V":
-		fmt.Printf("gasm %s\n", version)
+		os.Exit(cmdVersion())
-	case "help", "-h", "--help":
+	case "help", "--help", "-h":
 		usage(os.Stdout)
 	default:
-		fmt.Fprintf(os.Stderr, "gasm: unknown command %q\n\n", os.Args[1])
+		fmt.Fprintf(os.Stderr, "gasm: unknown command %q — run \"gasm --help\" for usage\n", os.Args[1])
-		usage(os.Stderr)
 		os.Exit(2)
 	}
 }
 
+// cmdVersion prints the release version.
+func cmdVersion() int {
+	fmt.Printf("gasm %s\n", version)
+	return 0
+}
+
 func usage(w io.Writer) {
-	fmt.Fprintf(w, `gasm %s — developer tooling for Go's Plan 9 assembler
+	fmt.Fprintf(w, `gasm %s — developer tooling for Go's Plan 9 assembler (GAsm)
+
+gasm bundles a lexer, parser, formatter, linter, standalone assembler and
+language server for Plan 9 assembly into one self-contained binary.
 
 Usage:
-  gasm tokens <file>          print the lexical token stream
+  gasm <command> [arguments]
-  gasm parse  <file>          parse and report syntax errors
+  gasm [flags]
-  gasm fmt    [-w] [path...]  canonicalise formatting (no path or a directory:
+
-                              reformat every .s below it in place, like go fmt)
+Commands:
-  gasm lint   <file...>       run static checks
+  tokens    print the lexical token stream
-  gasm asm    [-o out.bin] <file>  assemble to machine code (amd64, Phase 2)
+  parse     parse and report syntax errors
-  gasm lsp                    run the language server over stdio
+  fmt       canonicalise formatting (gofmt for assembly)
-  gasm version                print the version
+  lint      run static checks
+  asm       assemble .s files to machine code (amd64)
+  lsp       run the language server over stdio
+  version   print the version (same as --version)
+
+Flags:
+  -h, --help      show this help
+  -V, --version   print the version
+
+Run "gasm <command> -h" for a command's usage and flags.
+
+Examples:
+  gasm fmt                  reformat every .s below the current directory
+  gasm lint go-flac/*.s     run static checks over the kernels
+  gasm asm -o k.bin kern_amd64.s
 `, version)
 }
 
+// newCommand returns the FlagSet of a subcommand whose -h/--help prints a
+// proper usage block: the one-line usage, the long description and the flag
+// defaults.  The flag package routes -h/--help to fs.Usage and exits 0.
+func newCommand(name, usageLine, long string) *flag.FlagSet {
+	fs := flag.NewFlagSet(name, flag.ExitOnError)
+	fs.Usage = func() {
+		w := fs.Output()
+		fmt.Fprintf(w, "Usage: %s\n\n%s\n", usageLine, strings.TrimSpace(long))
+		hasFlags := false
+		fs.VisitAll(func(*flag.Flag) { hasFlags = true })
+		if hasFlags {
+			fmt.Fprintln(w, "\nFlags:")
+			fs.PrintDefaults()
+		}
+	}
+	return fs
+}
+
 // readSource returns the contents of path, or stdin when path is "-".
 func readSource(path string) (string, error) {
 	if path == "-" {
@@ -85,7 +125,10 @@ func readSource(path string) (string, error) {
 }
 
 func cmdTokens(args []string) int {
-	fs := flag.NewFlagSet("tokens", flag.ExitOnError)
+	fs := newCommand("tokens", "gasm tokens <file>", `
+Print the lexical token stream of FILE: position, token kind and text, one
+token per line.  FILE may be "-" to read standard input.
+`)
 	fs.Parse(args)
 	if fs.NArg() != 1 {
 		fmt.Fprintln(os.Stderr, "usage: gasm tokens <file>")
@@ -103,7 +146,11 @@ func cmdTokens(args []string) int {
 }
 
 func cmdParse(args []string) int {
-	fs := flag.NewFlagSet("parse", flag.ExitOnError)
+	fs := newCommand("parse", "gasm parse <file>", `
+Parse FILE and report syntax errors on stderr.  On success, print how many
+declarations and TEXT functions the file contains.  FILE may be "-" to read
+standard input.
+`)
 	fs.Parse(args)
 	if fs.NArg() != 1 {
 		fmt.Fprintln(os.Stderr, "usage: gasm parse <file>")
@@ -133,14 +180,24 @@ func cmdParse(args []string) int {
 }
 
 func cmdFmt(args []string) int {
-	fs_ := flag.NewFlagSet("fmt", flag.ExitOnError)
+	fs := newCommand("fmt", "gasm fmt [-w] [path...]", `
-	write := fs_.Bool("w", false, "write result to the source file")
+Canonicalise the formatting of Plan 9 assembly sources: indentation, operand
-	fs_.Parse(args)
+spacing, per-function mnemonic alignment and blank-line layout (exactly one
+blank line before each label, TEXT and GLOBL block).  Formatting is
+idempotent and preserves every line, comments included.
+
+With no paths — or a directory path — every .s file below it is reformatted
+in place and the changed files are listed, the way go fmt does; "." and "_"
+directories are skipped.  Explicit file paths print to stdout unless -w is
+given.
+`)
+	write := fs.Bool("w", false, "write result to the source file")
+	fs.Parse(args)
 	// Like go fmt: with no arguments, or with a directory argument, every .s
 	// file below the directory is formatted in place and the names of the
 	// changed files are listed; explicit file arguments keep the -w / stdout
 	// behaviour.
-	paths := fs_.Args()
+	paths := fs.Args()
 	dirMode := len(paths) == 0
 	if dirMode {
 		paths = []string{"."}
@@ -215,7 +272,16 @@ func asmFiles(dir string) ([]string, error) {
 }
 
 func cmdLint(args []string) int {
-	fs := flag.NewFlagSet("lint", flag.ExitOnError)
+	fs := newCommand("lint", "gasm lint <file...>", `
+Run the static checks over the given files and print diagnostics as
+"file:line:col: severity: message [code]".  The exit status is non-zero when
+an error-severity diagnostic is found; warnings (e.g. the register-clobber
+audit) do not affect it.
+
+Rules include unknown-instruction, operand-count, undefined-label,
+duplicate-label, missing-ret, missing-textflag-include, abi-argsize,
+unreachable-code, register-clobber and funcdata-pcdata.
+`)
 	disable := fs.String("disable", "", "comma-separated rule codes to disable")
 	fs.Parse(args)
 	if fs.NArg() == 0 {
@@ -256,7 +322,13 @@ func cmdLint(args []string) int {
 }
 
 func cmdLSP(args []string) int {
-	fs := flag.NewFlagSet("lsp", flag.ExitOnError)
+	fs := newCommand("lsp", "gasm lsp", `
+Run the language server over standard input/output: JSON-RPC 2.0 with
+Content-Length framing.  Point an LSP-capable editor at the binary and
+associate it with .s files; the target architecture is inferred from the file
+suffix (_amd64.s, _arm64.s, _riscv64.s, _loong64.s).  Provides completion,
+hover, document symbols, diagnostics and semantic-token highlighting.
+`)
 	fs.Parse(args)
 	srv := lsp.New(os.Stdin, os.Stdout)
 	if err := srv.Run(); err != nil {
@@ -267,7 +339,13 @@ func cmdLSP(args []string) int {
 }
 
 func cmdAsm(args []string) int {
-	fs := flag.NewFlagSet("asm", flag.ExitOnError)
+	fs := newCommand("asm", "gasm asm [-o out.bin] <file>", `
+Assemble FILE (amd64) without the Go toolchain: every TEXT function is
+encoded to machine code — scalar, VEX/AVX2 and EVEX/AVX-512 instructions,
+FP/SP frame mapping, local labels and file-local static symbols (GLOBL/DATA)
+resolved RIP-relative — and printed as a hex dump.  With -o the concatenated
+image (functions followed by the data section) is written to a file instead.
+`)
 	out := fs.String("o", "", "write the concatenated machine code to this file")
 	fs.Parse(args)
 	if fs.NArg() != 1 {

cmd/gasm/main_test.go

@@ -200,8 +200,24 @@ func TestCmdFmtWrite(t *testing.T) {
 func TestUsage(t *testing.T) {
 	var b bytes.Buffer
 	usage(&b)
-	if !strings.Contains(b.String(), "gasm") {
+	out := b.String()
-		t.Errorf("usage text unexpected:\n%s", b.String())
+	for _, want := range []string{
+		"gasm", "Commands:", "Flags:", "--help", "--version",
+		"tokens", "parse", "fmt", "lint", "asm", "lsp", "version",
+	} {
+		if !strings.Contains(out, want) {
+			t.Errorf("usage text missing %q:\n%s", want, out)
+		}
+	}
+}
+
+func TestCmdVersion(t *testing.T) {
+	out, _, code := capture(func() int { return cmdVersion() })
+	if code != 0 {
+		t.Fatalf("code = %d", code)
+	}
+	if !strings.Contains(out, version) {
+		t.Errorf("version output %q does not mention %q", out, version)
 	}
 }
 

docs/ARCHITECTURE.md

@@ -142,7 +142,7 @@ Two deeper analyses sit on top of the AST:
   has no System V style callee-saved registers (amd64 `BX`, `R12`–`R15` and
   the like are caller-saved or permanent scratch, and hand-written kernels may
   clobber them freely).  The audited set is the frame pointer and the
-  goroutine pointer per architecture (amd64 `BP`/`R14`, arm64 `R18`/`R28`/
+  the frame pointer, the goroutine pointer per architecture (amd64 `BP`/`R14`, arm64 `R18`/`R28`/
   `R29`, riscv64 `X27`, loong64 `R22`); the goroutine pointer is reported only
   when the function can reach the runtime — it is not `NOSPLIT` or makes a
   call — since the ABI0 transition machinery restores it on those paths, and
@@ -213,15 +213,26 @@ three-operand-plus-immediate form (`VSHUFPD`,
 `VPERM2I128`, `VINSERTI128`), the lane-extract form (`VEXTRACTI128`,
 `VEXTRACTF128`, where the YMM source occupies the reg field and the XMM or
 memory destination r/m), the direction-sensitive moves (`VMOVDQU`, `VMOVUPD`,
-`VMOVD`, `VMOVQ`, `VMOVSD`), the floating-point and FMA arithmetic (`VADDPD`,
+`VMOVD`, `VMOVQ`, `VMOVSD`), the floating-point and FMA arithmetic — the
-`VMULPD`, `VXORPD`, `VUNPCKHPD`, the scalar `VADDSD`/`VMULSD`, `VCVTDQ2PD`,
+packed double operations (`VADDPD`/`VSUBPD`/`VMULPD`/`VDIVPD`/`VMINPD`/
-`VFMADD231PD`) and the no-operand `VZEROUPPER` — together with `VPERMD` and
+`VMAXPD`), the unpacks (`VUNPCKHPD`/`VUNPCKLPD`), the scalar SD and SS
+operations, `VMOVDDUP`, `VXORPD`, the width-changing conversions
+(`VCVTDQ2PS`, `VCVTPS2PD`, `VCVTDQ2PD`, and the `VCVTPD2DQX`/`Y` and
+`VCVTTPD2DQX`/`Y` spellings, whose length follows the wider source) and
+`VFMADD231PD` — and the no-operand `VZEROUPPER`, together with `VPERMD` and
 the scalar families (`CMOVcc`, `SETcc`, `LZCNT`/`TZCNT`, the extending moves,
 `CVTSx2SD`, `IMUL3`) and the EVEX (AVX-512) prefix — the four-byte prefix with
-5-bit register fields (Z0–Z31, X/Y 16–31), opmask registers as operands and
+5-bit register fields (Z0–Z31, X/Y 16–31, with the mod=11 quirk that carries
-mask destinations, and the compressed disp8×N displacement, whose multiplier
+rm[4] in X̄), opmask registers (K0–K7 as operands, mask destinations and
-follows the memory operand's size — covering every instruction the go-flac
+explicit merging/zeroing masks — written the way Go writes them, as a K
-AVX2 and AVX-512 kernels use.  Every encoding is validated two ways: by
+operand among the operands plus a `.Z` mnemonic suffix), and the compressed
+disp8×N displacement, whose multiplier follows the memory operand's size —
+covering every instruction the go-flac and go-lz4 AVX2/AVX-512 kernels use,
+plus the common AVX-512 F/BW integer set and the floating-point and
+conversion set (the packed double arithmetic, the scalar SD/SS forms —
+whose EVEX encodings serve masked and zeroing use — `VMOVDDUP`, and the
+width-changing conversions, including the `VCVTPD2DQ`/`VCVTTPD2DQ` family
+whose length follows the wider source operand).  Every encoding is validated two ways: by
 round-trip decoding through `golang.org/x/arch`, and byte-for-byte against
 the machine code the real Go assembler emits — a comparison that holds for
 whole functions: all 27 functions of both kernels assemble to exactly the Go
@@ -234,7 +245,8 @@ behind the code and resolves references to them (`mask<>(SB)`) to
 RIP-relative loads whose displacements point inside the resulting image, so
 the bytes are self-consistent at any base address.  External (non-file-local)
 symbols are rejected: they need object-file emission, which — together with
-EVEX masking/zeroing and the other architectures — is the rest of Phase 2.
+the remaining EVEX forms and the other architectures — is the rest of
+Phase 2.
 
 ## Extension points
 

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.7.0"
+version := "0.10.0"
 
 default:
     @just --list

lint/lint.go

@@ -242,7 +242,7 @@ func lintText(t *ast.Text, tab *arch.Table, archKnown bool, cfg Config, macros m
 				}
 			}
 
-			if archKnown && !cfg.Disable[CodeOperandCount] && !isMacroInvocation(mnem, macros) {
+			if archKnown && !cfg.Disable[CodeOperandCount] && !isMacroInvocation(mnem, macros) && !maskedEvex(mnem, st.Operands) {
 				if in, ok := tab.Lookup(mnem); ok && in.MinOps >= 0 {
 					n := len(st.Operands)
 					if n < in.MinOps || n > in.MaxOps {
@@ -438,6 +438,28 @@ func isMacroInvocation(mnem string, macros map[string]bool) bool {
 	return strings.Contains(mnem, "_") || macros[mnem]
 }
 
+// maskedEvex reports whether the instruction is a masked EVEX form: the
+// mnemonic carries a .Z suffix, or the operand list contains an opmask
+// register (K1–K7).  Either way the operand count differs from the unmasked
+// form, so count checks are skipped.
+func maskedEvex(mnem string, ops []*ast.Operand) bool {
+	if strings.Contains(mnem, ".") {
+		return true
+	}
+	for _, op := range ops {
+		if op.Kind == ast.OpAddr && op.Addr.Sym != nil && op.Addr.Base == "" &&
+			op.Addr.Index == "" && op.Addr.Sym.Pseudo == "" && isMaskReg(op.Addr.Sym.Name) {
+			return true
+		}
+	}
+	return false
+}
+
+// isMaskReg reports whether name is an opmask register K0–K7.
+func isMaskReg(name string) bool {
+	return len(name) == 2 && name[0] == 'K' && name[1] >= '0' && name[1] <= '7'
+}
+
 // isConditionalDirective reports whether a preprocessor directive (the text
 // after '#') is a conditional-compilation directive whose branches the parser
 // cannot resolve.

lint/lint_test.go

@@ -187,6 +187,26 @@ done:
 	}
 }
 
+// TestEvexMaskingRecognised checks that masked EVEX forms — the .Z suffix and
+// an explicit K operand — are recognised and exempt from operand-count
+// checks.
+func TestEvexMaskingRecognised(t *testing.T) {
+	diags := lintSrc(t, `
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	VPADDD.Z Z1, Z2, K2, Z3
+	VPMINSD Z1, Z2, K5, Z3
+	VMOVDQU8 Z1, K3, (SI)
+	RET
+`)
+	if codes(diags)[CodeUnknownInstr] != 0 {
+		t.Fatalf("masked EVEX must be recognised: %+v", diags)
+	}
+	if codes(diags)[CodeOperandCount] != 0 {
+		t.Fatalf("masked operand counts must not be flagged: %+v", diags)
+	}
+}
+
 func TestArm64AddressingSuffix(t *testing.T) {
 	// .W (pre-index) and .P (post-index) suffixes must resolve to the base
 	// instruction.