feat(asm): encode the arm64 Q-width FP load and store

FMOVQ routes through the MOV load/store machinery in the plain, post-index,
pre-index and static-symbol forms.  The Q width carries its size in the opc
field, so the store spelling is opc=10 and the access scales by sixteen;
both come from helpers now instead of the size exponent.  The static-symbol
form takes the toolchain's twelve-byte ADRP + ADD + access fallback with the
R_ADDRARM64 pair.  The register-to-register and immediate forms stay
rejected, matching the toolchain's own table.

Assisted-by: GLM 5.3 Flash

asm/arm64_assemble.go

@@ -228,11 +228,12 @@ func arm64PCAlignPad(pos int, instr *ast.Instr) int {
 }
 
 // isARM64MovMnemonic reports whether m is one of the MOV-family spellings the
-// arm64 encoder treats as the MOV pseudo-instruction.
+// arm64 encoder treats as the MOV pseudo-instruction.  FMOVQ rides the same
+// load/store machinery but carries no register-move or immediate form.
 func isARM64MovMnemonic(m string) bool {
 	switch m {
 	case "MOV", "MOVD", "MOVW", "MOVWU", "MOVH", "MOVHU", "MOVB", "MOVBU",
-		"FMOVS", "FMOVD":
+		"FMOVS", "FMOVD", "FMOVQ":
 		return true
 	}
 	return false
@@ -279,7 +280,7 @@ func arm64InstrSize(instr *ast.Instr, fi arm64FrameInfo, pos int) int {
 	}
 	switch mnem {
 	case "MOV", "MOVD", "MOVW", "MOVWU", "MOVH", "MOVHU", "MOVB", "MOVBU",
-		"FMOVS", "FMOVD":
+		"FMOVS", "FMOVD", "FMOVQ":
 		return arm64MovSize(mnem, ops, fi)
 	case "ADD", "ADDW", "SUB", "SUBW", "CMP", "CMPW", "CMN", "CMNW",
 		"ADDS", "ADDSW", "SUBS", "SUBSW":
@@ -354,7 +355,7 @@ func encodeARM64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi arm64
 	case "BL", "CALL":
 		return encodeARM64Branch(mnem, ops, pc, offsets, true, relocs, resolve)
 	case "MOV", "MOVD", "MOVW", "MOVWU", "MOVH", "MOVHU", "MOVB", "MOVBU",
-		"FMOVS", "FMOVD":
+		"FMOVS", "FMOVD", "FMOVQ":
 		return encodeARM64Mov(instr, mnem, "", fi, relocs)
 	}
 
@@ -1479,6 +1480,20 @@ func encodeARM64Mov(instr *ast.Instr, mnem string, wb string, fi arm64FrameInfo,
 	}
 	src, dst := ops[0], ops[1]
 
+	if mnem == "FMOVQ" {
+		// The 128-bit FP move is a pure memory access in the toolchain: its
+		// table carries no register-to-register or immediate row, and both
+		// spellings are rejected ("illegal combination"), so only the
+		// load/store and static-symbol forms exist here either.
+		loadOK := arm64IsMemOperand(src) && !arm64IsMemOperand(dst) && !isImmOperand(src)
+		storeOK := arm64IsMemOperand(dst) && !arm64IsMemOperand(src) && !isImmOperand(src)
+		sbLoad := src.Addr.Sym != nil && src.Addr.Sym.Pseudo == "SB"
+		sbStore := dst.Addr.Sym != nil && dst.Addr.Sym.Pseudo == "SB"
+		if !loadOK && !storeOK && !sbLoad && !sbStore {
+			return nil, fmt.Errorf("%s: only the memory load and store forms exist", mnem)
+		}
+	}
+
 	if wb != "" {
 		switch {
 		case isMemOperand(src) && !isMemOperand(dst):
@@ -1609,8 +1624,14 @@ func arm64MovSize(mnem string, ops []*ast.Operand, fi arm64FrameInfo) int {
 		}
 		return len(b)
 	case src.Addr.Sym != nil && src.Addr.Sym.Pseudo == "SB":
+		if mnem == "FMOVQ" {
+			return 12 // ADRP + ADD + LDR: the unaligned-access fallback
+		}
 		return 8 // ADRP + LDR
 	case dst.Addr.Sym != nil && dst.Addr.Sym.Pseudo == "SB":
+		if mnem == "FMOVQ" {
+			return 12 // ADRP + ADD + STR: the unaligned-access fallback
+		}
 		return 8 // ADRP + STR
 	case isMemOperand(src) || isMemOperand(dst):
 		mem := src
@@ -1623,7 +1644,7 @@ func arm64MovSize(mnem string, ops []*ast.Operand, fi arm64FrameInfo) int {
 		if !ok {
 			lt = a64LoadTable["MOVD"] // the MOV pseudo is a 64-bit access
 		}
-		scale := int64(1) << uint(lt.size)
+		scale := a64LSScale(lt)
 		if off >= 0 && off%scale == 0 && off/scale < 4096 {
 			return 4
 		}
@@ -1876,7 +1897,7 @@ func encodeARM64MemOp(mnem string, mem *ast.Operand, reg int, load bool, fi arm6
 		lt = a64LoadTable["MOVD"]
 	}
 
-	scale := int64(1) << uint(lt.size)
+	scale := a64LSScale(lt)
 	storeOpc := a64StoreOpc(lt)
 	var opc int
 	if load {
@@ -1963,12 +1984,27 @@ func encodeARM64SBAddr(sym *ast.Symbol, rd int, relocs *[]Reloc) []byte {
 
 // encodeARM64SBLoad emits ADRP R27, 0; LDR Rd, [R27, 0] with relocations,
 // matching the toolchain: the scratch register is REGTMP (R27) and the pair
-// carries R_ARM64_PCREL_LDST64.
+// carries R_ARM64_PCREL_LDST64.  FMOVQ has no LDST relocation width, so it
+// takes the toolchain's unaligned-access fallback instead (asm7.go case 65):
+// ADRP R27, 0; ADD R27, R27, 0; LDR with the R_ADDRARM64 pair.
 func encodeARM64SBLoad(sym *ast.Symbol, rd int, mnem string, relocs *[]Reloc) ([]byte, error) {
 	lt, ok := a64LoadTable[mnem]
 	if !ok {
 		lt = a64LoadTable["MOVD"]
 	}
+	if mnem == "FMOVQ" {
+		if relocs != nil {
+			*relocs = append(*relocs,
+				Reloc{Off: 0, After: 0, Name: sym.Name, Kind: RelArm64Addr, Addend: sym.Offset},
+				Reloc{Off: 4, After: 4, Name: sym.Name, Kind: RelArm64Addr, Addend: sym.Offset},
+			)
+		}
+		return a64WordsLE(
+			a64ADR(1, 0, 0, 27),              // ADRP R27, 0
+			a64AddSub(1, 0, 0, 0, 0, 27, 27), // ADD $0, R27, R27
+			a64LSU(uint32(lt.size), uint32(lt.V), uint32(lt.opc), 0, 27, uint32(rd)), // LDR Rd, (R27)
+		), nil
+	}
 	if relocs != nil {
 		*relocs = append(*relocs,
 			Reloc{Off: 0, After: 8, Name: sym.Name, Kind: RelArm64LDST64, Addend: sym.Offset},
@@ -1981,13 +2017,28 @@ func encodeARM64SBLoad(sym *ast.Symbol, rd int, mnem string, relocs *[]Reloc) ([
 }
 
 // encodeARM64SBStore emits ADRP R27, 0; STR Rs, [R27, 0] with relocations,
-// matching the toolchain's R27 scratch and R_ARM64_PCREL_LDST64 pair.
+// matching the toolchain's R27 scratch and R_ARM64_PCREL_LDST64 pair.  FMOVQ
+// takes the twelve-byte ADRP + ADD + STR fallback with the R_ADDRARM64 pair,
+// the store-side twin of the load above (asm7.go case 64).
 func encodeARM64SBStore(sym *ast.Symbol, rs int, mnem string, relocs *[]Reloc) ([]byte, error) {
 	lt, ok := a64LoadTable[mnem]
 	if !ok {
 		lt = a64LoadTable["MOVD"]
 	}
 	storeOpc := a64StoreOpc(lt)
+	if mnem == "FMOVQ" {
+		if relocs != nil {
+			*relocs = append(*relocs,
+				Reloc{Off: 0, After: 0, Name: sym.Name, Kind: RelArm64Addr, Addend: sym.Offset},
+				Reloc{Off: 4, After: 4, Name: sym.Name, Kind: RelArm64Addr, Addend: sym.Offset},
+			)
+		}
+		return a64WordsLE(
+			a64ADR(1, 0, 0, 27),              // ADRP R27, 0
+			a64AddSub(1, 0, 0, 0, 0, 27, 27), // ADD $0, R27, R27
+			a64LSU(uint32(lt.size), uint32(lt.V), uint32(storeOpc), 0, 27, uint32(rs)), // STR Rs, (R27)
+		), nil
+	}
 	if relocs != nil {
 		*relocs = append(*relocs,
 			Reloc{Off: 0, After: 8, Name: sym.Name, Kind: RelArm64LDST64, Addend: sym.Offset},

asm/arm64_encode.go

@@ -1518,11 +1518,26 @@ var a64LoadTable = map[string]a64LSType{
 
 // a64StoreOpc returns the store opc for a given load type: integer and FP
 // stores both encode opc=00 (the load's signedness bit sits in opc[1], which
-// the store form clears; FP registers are selected by V, not opc).
+// the store form clears; FP registers are selected by V, not opc).  The one
+// exception is the 128-bit Q width, whose store is the opc=10 spelling: with
+// opc=00 the size field selects STR B instead (STR Qt is size=00, opc=10).
 func a64StoreOpc(t a64LSType) int {
+	if t.size == 0 && t.V == 1 {
+		return 2
+	}
 	return 0
 }
 
+// a64LSScale returns the byte width an access's unsigned offset divides by.
+// The Q width carries its size in opc (the size field stays 0), yet scales
+// by 16 like any other 128-bit access, so the exponent alone does not answer.
+func a64LSScale(t a64LSType) int64 {
+	if t.size == 0 && t.V == 1 {
+		return 16
+	}
+	return int64(1) << uint(t.size)
+}
+
 // arm64RegClass discriminates integer (R), floating-point (F) registers for
 // the MOV pseudo-instruction.
 type arm64RegClass int

asm/arm64_encode_test.go

@@ -679,6 +679,42 @@ func TestArm64AcquireRelease(t *testing.T) {
 	}
 }
 
+// TestArm64QMove pins the Q-width FP move against the toolchain words the
+// corpus records: the plain, writeback and static-symbol load/store forms.
+// FMOVQ carries no register-to-register or immediate form, and both are
+// rejected the way the toolchain rejects them.
+func TestArm64QMove(t *testing.T) {
+	got := arm64Words(t, "\tFMOVQ.P F13, 11(R10)\n\tFMOVQ.W F15, 11(R20)\n\tFMOVQ.P 11(R10), F13\n\tFMOVQ.W 11(R20), F15\n"+
+		"\tFMOVQ F0, 32(R5)\n\tFMOVQ F10, 65520(R10)\n\tFMOVQ 32(R5), F2\n")
+	want := []uint32{
+		0x3c80b54d, // FMOVQ.P F13, 11(R10)
+		0x3c80be8f, // FMOVQ.W F15, 11(R20)
+		0x3cc0b54d, // FMOVQ.P 11(R10), F13
+		0x3cc0be8f, // FMOVQ.W 11(R20), F15
+		0x3d8008a0, // FMOVQ F0, 32(R5)
+		0x3dbffd4a, // FMOVQ F10, 65520(R10)
+		0x3dc008a2, // FMOVQ 32(R5), F2
+		0xd65f03c0,
+	}
+	if len(got) != len(want) {
+		t.Fatalf("word count = %d, want %d", len(got), len(want))
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("word %d = %08x, want %08x", i, got[i], want[i])
+		}
+	}
+	for _, src := range []string{"\tFMOVQ F1, F2\n", "\tFMOVQ $1, R2\n", "\tFMOVQ $0, 8(R3)\n"} {
+		f, errs := parser.Parse("test_arm64.s", "#include \"textflag.h\"\n\nTEXT ·f(SB), NOSPLIT, $0-0\n"+src+"\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("parse %q: %v", src, errs)
+		}
+		if _, err := AssembleFileARM64(f); err == nil {
+			t.Errorf("%q: assembled, want rejection", strings.TrimSpace(src))
+		}
+	}
+}
+
 // TestArm64BTI pins the landing-pad family against the toolchain words:
 // only the uppercase C/J/JC spellings assemble, and bare BTI is a
 // diagnostic, never a panic.

asm/kernels_differential_test.go

@@ -142,6 +142,7 @@ func TestDifferentialKernels(t *testing.T) {
 		{filepath.Join("..", "testdata", "verify", "forms_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "datarel_arm64.s"), "arm64", true},
 		{filepath.Join("..", "testdata", "verify", "divslash_arm64.s"), "arm64", true},
+		{filepath.Join("..", "testdata", "verify", "qmov_arm64.s"), "arm64", true},
 	} {
 		t.Run(filepath.Base(k.path), func(t *testing.T) {
 			src, err := os.ReadFile(k.path)

testdata/verify/qmov_arm64.s

@@ -0,0 +1,49 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the arm64 Q-width FP move: the FMOVQ load and
+// store in the plain, post-index and pre-index forms, against scaled and
+// unscaled offsets, and against static symbols.  Every function is
+// byte-compared against go tool asm.
+
+#include "textflag.h"
+
+// func writeback()
+TEXT ·writeback(SB), NOSPLIT, $0-0
+	FMOVQ.P F13, 11(R10)
+	FMOVQ.W F15, 11(R20)
+	FMOVS.P F20, 4(R0)
+	FMOVS.W F20, 4(R0)
+	FMOVD.P F20, 8(R1)
+	FMOVD.W F20, 8(R1)
+	FMOVQ.P 11(R10), F13
+	FMOVQ.W 11(R20), F15
+	FMOVS.P 8(R0), F20
+	FMOVS.W 8(R0), F20
+	FMOVD.W 8(R1), F20
+	RET
+
+// func plain()
+TEXT ·plain(SB), NOSPLIT, $0-0
+	FMOVQ F0, 32(R5)
+	FMOVQ F10, 65520(R10)
+	FMOVQ F11, 64(RSP)
+	FMOVQ F11, 8(R20)
+	FMOVQ F11, 4(R20)
+	FMOVQ 32(R5), F2
+	FMOVQ 65520(R10), F10
+	FMOVQ 64(RSP), F11
+	FMOVD F1, 8(R2)
+	FMOVD 8(R2), F1
+	RET
+
+// func symbols()
+TEXT ·symbols(SB), NOSPLIT, $0-0
+	FMOVQ F5, x(SB)
+	FMOVQ F5, x+8(SB)
+	FMOVQ x+8(SB), F5
+	FMOVQ x(SB), F5
+	RET
+
+// data the symbol forms refer to
+GLOBL x(SB), NOPTR, $16