fix(asm): key the arm64 immediate class order on the ZR spelling

Assisted-by: GLM 5.3 Flash

asm/arm64_assemble.go

@@ -330,11 +330,11 @@ func arm64InstrSize(instr *ast.Instr, fi arm64FrameInfo, pos int) int {
 				if strings.HasSuffix(mnem, "W") {
 					width = 32
 				}
-				rd := 31
+				zrDest := false
 				if mnem != "TST" && mnem != "TSTW" {
-					rd = arm64RegNum(operandRegName(ops[len(ops)-1]))
+					zrDest = strings.EqualFold(operandRegName(ops[len(ops)-1]), "ZR")
 				}
-				if _, _, _, bc := a64LogicalImm(v, width); bc && (rd != 31 || strings.HasPrefix(mnem, "TST")) {
+				if _, _, _, bc := a64LogicalImm(v, width); bc && (mnem == "TST" || mnem == "TSTW" || !zrDest) {
 					return 4
 				}
 				return 8
@@ -871,13 +871,16 @@ func encodeARM64DPSR(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte, er
 				width = 32
 			}
 			var rn, rd int
+			zrDest := false
 			switch len(ops) {
 			case 3:
 				rn = arm64RegNum(operandRegName(ops[1]))
 				rd = arm64RegNum(operandRegName(ops[2]))
+				zrDest = strings.EqualFold(operandRegName(ops[2]), "ZR")
 			default:
 				rd = arm64RegNum(operandRegName(ops[1]))
 				rn = rd
+				zrDest = strings.EqualFold(operandRegName(ops[1]), "ZR")
 			}
 			if isCmp {
 				// CMP/CMN/TST write the flags alone: the destination is ZR
@@ -891,9 +894,10 @@ func encodeARM64DPSR(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte, er
 			// The toolchain's logical-immediate rows take a real destination
 			// only (omovconst guards the bitmask path with rt != REGZERO): a
 			// non-flag-setting logical to ZR materialises the constant into
-			// REGTMP (R27) and takes the register form.  The flags-only TST
-			// spellings keep the fast path: ANDS ZR, Rn, #imm is their form.
-			if ok && (rd != 31 || isCmp) {
+			// REGTMP (R27) and takes the register form.  RSP is a real
+			// register here, and the flags-only TST spellings keep the fast
+			// path: ANDS ZR, Rn, #imm is their form.
+			if ok && (isCmp || !zrDest) {
 				opc := (baseOp >> 29) & 7
 				sf := (baseOp >> 31) & 1
 				return a64wordLE(sf<<31 | opc<<29 | 0x24<<23 | n<<22 | immr<<16 | imms<<10 |
@@ -1694,6 +1698,11 @@ func encodeARM64Mov(instr *ast.Instr, mnem string, wb string, fi arm64FrameInfo,
 		if rd < 0 {
 			return nil, fmt.Errorf("%s $imm: invalid destination register", mnem)
 		}
+		if strings.EqualFold(operandRegName(dst), "ZR") {
+			// The destination is ZR: omovconst's bitmask path needs a real
+			// register, so the value rides MOVZ/MOVN.
+			return encodeARM64LoadImmClass(rd, arm64Imm64(src), mnem, false)
+		}
 		return encodeARM64LoadImm(rd, arm64Imm64(src), mnem)
 	}
 
@@ -1806,7 +1815,7 @@ func arm64MovSize(mnem string, ops []*ast.Operand, fi arm64FrameInfo) int {
 		// values expand to up to four words and the W forms truncate first.
 		// Anything else would desynchronise the label offsets of pass 1 from
 		// the bytes pass 2 lays down, corrupting every later branch.
-		b, err := encodeARM64LoadImm(31, arm64Imm64(src), mnem)
+		b, err := encodeARM64LoadImmClass(31, arm64Imm64(src), mnem, !strings.EqualFold(operandRegName(dst), "ZR"))
 		if err != nil {
 			return 4
 		}
@@ -1860,6 +1869,14 @@ func arm64MovSize(mnem string, ops []*ast.Operand, fi arm64FrameInfo) int {
 // truncates to 0xFFFFFFFF, whose complement is a single zero chunk, and encodes
 // as MOVN W, #0.
 func encodeARM64LoadImm(rd int, v int64, mnem string) ([]byte, error) {
+	return encodeARM64LoadImmClass(rd, v, mnem, true)
+}
+
+// encodeARM64LoadImmClass is encodeARM64LoadImm with the classification order
+// in hand: bitmaskOK false skips the logical-immediate paths, which the
+// toolchain's omovconst only takes for a real register (rt != REGZERO); an
+// immediate to ZR rides the MOVZ/MOVN sequence carrying the value.
+func encodeARM64LoadImmClass(rd int, v int64, mnem string, bitmaskOK bool) ([]byte, error) {
 	d := v
 	sf := uint32(1) // 64-bit
 	if mnem == "MOVW" || mnem == "MOVWU" {
@@ -1882,10 +1899,7 @@ func encodeARM64LoadImm(rd int, v int64, mnem string) ([]byte, error) {
 	//   `MOVD $4096, R27` is ORR $4096, not MOVZ $(1<<12)
 	// - outside that band: MOVZ/MOVN first (C_MOVCON before C_BITCON), and
 	//   negative values reach MOVN before the bitmask test
-	// The bitmask path exists only for a real register: omovconst guards it
-	// with rt != REGZERO, so an immediate to ZR always rides the MOVZ/MOVN
-	// sequence carrying the value.
-	tryBitmaskFirst := rd != 31 && d > 0 && (d <= 0xFFF || (d&0xFFF == 0 && d <= 0xFFF000))
+	tryBitmaskFirst := bitmaskOK && d > 0 && (d <= 0xFFF || (d&0xFFF == 0 && d <= 0xFFF000))
 
 	if tryBitmaskFirst {
 		// Addcon-band immediate: try bitmask first (Go uses ORR for values
@@ -1914,7 +1928,7 @@ func encodeARM64LoadImm(rd int, v int64, mnem string) ([]byte, error) {
 	}
 
 	// For values outside the bitmask-first range that are not movcon: try bitmask.
-	if !tryBitmaskFirst && rd != 31 {
+	if !tryBitmaskFirst && bitmaskOK {
 		N, immr, imms, ok := arm64Bitmask(uint64(d), int(sf))
 		if ok {
 			return a64wordLE(sf<<31 | 1<<29 | 0x24<<23 | N<<22 | immr<<16 | imms<<10 | 31<<5 | uint32(rd)), nil