fix(asm): match go tool asm encodings and strictness

asm/instrs.go

@@ -51,13 +51,23 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 	// Integer scalar XMM moves: MOVQ with an XMM operand is the SSE2
 	// packed-quadword move, NOT a GPR move: mem→xmm encodes as F3 0F 7E
 	// (reg = dst, no REX.W — the Go assembler's form), xmm→mem as
-	// 66 0F D6 (rm = xmm).  MOVL is the packed-dword move instead:
-	// 66 0F 6E load, 66 0F 7E store.  A GPR-move fallback would silently
-	// emit REX.W 8B with the wrong operand meaning.
+	// 66 0F D6 (rm = xmm).  Register forms against a GPR use the MOVD
+	// opcodes with REX.W instead: 66 REX.W 0F 6E (gpr→xmm) and
+	// 66 REX.W 0F 7E (xmm→gpr); the memory opcodes with a register r/m
+	// would be undefined forms.  MOVL is the packed-dword move:
+	// 66 0F 6E load, 66 0F 7E store, no REX.W.  A GPR-move fallback would
+	// silently emit REX.W 8B with the wrong operand meaning.
 	_, srcVec := vecReg(src)
 	dstReg, dstVec := vecReg(dst)
 	if srcVec || dstVec {
 		if dstVec {
+			if g, ok := src.(Reg); ok && !g.isVec() {
+				i := &instr{prefix: 0x66, opcode: []byte{0x0F, 0x6E}, modrm: -1, sib: -1, rexW: size == 8}
+				if err := setRM(i, dstReg, src, 8); err != nil {
+					return err
+				}
+				return e.emit(i)
+			}
 			i := &instr{prefix: 0xF3, opcode: []byte{0x0F, 0x7E}, modrm: -1, sib: -1}
 			if size == 4 {
 				i.prefix = 0x66
@@ -72,6 +82,13 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 		if !srcIsXMM || !srcXMM.isVec() {
 			return fmt.Errorf("MOV: store needs an XMM source")
 		}
+		if g, ok := dst.(Reg); ok && !g.isVec() {
+			i := &instr{prefix: 0x66, opcode: []byte{0x0F, 0x7E}, modrm: -1, sib: -1, rexW: size == 8}
+			if err := setRM(i, srcXMM, dst, 8); err != nil {
+				return err
+			}
+			return e.emit(i)
+		}
 		i := &instr{prefix: 0x66, opcode: []byte{0x0F, 0xD6}, modrm: -1, sib: -1}
 		if size == 4 {
 			i.opcode = []byte{0x0F, 0x7E}
@@ -199,7 +216,13 @@ func (e *enc) encodeALU(op struct {
 	}
 	src, dst := ops[0], ops[1]
 
+	// CMP never takes its immediate first: the Go assembler rejects
+	// CMPL $0, AX outright (only CMPL AX, $0 is legal, unlike TEST and the
+	// writing ALU ops whose immediate is naturally the source).
 	if imm, ok := src.(Imm); ok {
+		if op.digit == 7 {
+			return fmt.Errorf("CMP immediate must be the second operand (reg, $imm)")
+		}
 		return e.encodeALUImm(op.digit, dst, int64(imm), size)
 	}
 
@@ -290,6 +313,15 @@ func (e *enc) encodeALUImm(digit int, dst Operand, imm int64, size int) error {
 		i.imm = []byte{byte(int8(imm))}
 		return e.emit(i)
 	}
+	// 0x81 /digit, imm16/imm32 — or the Go assembler's accumulator short
+	// form (opcode+5, no ModR/M) when the destination is AX/AL, which it
+	// prefers over the generic form exactly here.
+	if r, ok := dst.(Reg); ok && r.idx == 0 {
+		accOp := map[int]byte{0: 0x05, 1: 0x0D, 2: 0x15, 3: 0x1D, 4: 0x25, 5: 0x2D, 6: 0x35, 7: 0x3D}[digit]
+		i := newInstr(size, []byte{accOp})
+		i.imm = immediate(imm, size, false)
+		return e.emit(i)
+	}
 	// 0x81 /digit, imm16/imm32.
 	i := newInstr(size, []byte{0x81})
 	if err := setRMDigit(i, digit, dst, size); err != nil {
@@ -307,7 +339,18 @@ func (e *enc) encodeTest(ops []Operand, size int) error {
 	}
 	src, dst := ops[0], ops[1]
 	if imm, ok := src.(Imm); ok {
-		// TEST r/m, imm: 0xF6 (8-bit) / 0xF7 /0.
+		// TEST r/m, imm: 0xF6 (8-bit) / 0xF7 /0 — but the Go assembler
+		// always uses the accumulator forms (A8/A9, no ModR/M) when the
+		// register operand is AL/AX, whatever the immediate's width.
+		if r, ok := dst.(Reg); ok && r.idx == 0 {
+			op := byte(0xA9)
+			if size == 1 {
+				op = 0xA8
+			}
+			i := newInstr(size, []byte{op})
+			i.imm = immediate(int64(imm), size, false)
+			return e.emit(i)
+		}
 		op := byte(0xF7)
 		if size == 1 {
 			op = 0xF6