feat(asm): add EVEX/AVX-512 encoding and assemble the AVX-512 kernel byte-identically

Assisted-by: Qwen 3.8 Max Preview

asm/vex.go

@@ -39,6 +39,10 @@ const (
 	// source lives in the reg field, the destination in r/m — the PEXTR-style
 	// layout.  VEXTRACTI128 and VEXTRACTF128 use this shape.
 	vexExtract
+	// vexRMRev is the reversed two-operand form `OP src, dst` with the source
+	// in ModRM.reg and the destination in r/m — the layout of the EVEX
+	// narrowing stores (VPMOVDW, VPMOVQD).
+	vexRMRev
 	// vexZero is the no-operand form (VZEROUPPER).
 	vexZero
 )
@@ -131,6 +135,9 @@ var vexTable = map[string]vexSpec{
 
 	// VEX.128.0F.W0 — no operands.
 	"VZEROUPPER": {1, 0x77, 0, 0, -1, vexZero},
+
+	// VEX.128.0F.W0 — mask-register test (KTESTW k1, k2: reg = dst, rm = src).
+	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
 }
 
 // vexVarShift maps the shift mnemonics to their variable-count opcode — the
@@ -188,6 +195,13 @@ func isVex(mnemUpper string) bool {
 
 // encodeVex encodes a VEX instruction with operands in Plan 9 order.
 func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
+	// Vector register indices 16–31 exist only in EVEX encodings; fail
+	// loudly rather than silently truncating the index.
+	for _, op := range ops {
+		if r, ok := op.(Reg); ok && r.isVec() && r.idx >= 16 {
+			return fmt.Errorf("%s: vector register index %d needs an EVEX (AVX-512) instruction", mnemUpper, r.idx)
+		}
+	}
 	if ms, ok := vexMoveTable[mnemUpper]; ok {
 		return e.encodeVexMove(mnemUpper, ms, ops)
 	}
@@ -531,6 +545,9 @@ func validMoveOther(ms vexMoveSpec, op Operand) bool {
 // the given precomputed fields.  It is shared by every register/rm VEX form;
 // immediate bytes are appended by the caller.
 func (e *enc) emitVexFields(spec vexSpec, l, regField, rBit, vvvvBar int, rm Operand) error {
+	if l > 1 {
+		return fmt.Errorf("ZMM operand requires an EVEX instruction")
+	}
 	var modrm, sib int
 	var disp []byte
 	var xBit, bBit int