feat(riscv64,loong64): operand tail, float DATA and honest port classification

Assisted-by: GLM 5.3 Flash

asm/riscv_assemble.go

@@ -6,6 +6,7 @@ package asm
 import (
 	"errors"
 	"fmt"
+	"math/bits"
 	"slices"
 	"strings"
 
@@ -14,13 +15,14 @@ import (
 
 // assembleRISCV assembles a RISC-V TEXT function body into machine code.
 // It handles the full RV64IMAFDC instruction set including RVC compression.
-func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, []SpadjStep, error) {
+func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, []SpadjStep, []RiscvLiteral, error) {
 	fi := riscvComputeFrame(t)
 	prologue := riscvPrologue(fi)
 	guardLen, err := riscvGuardLen(fi)
 	if err != nil {
-		return nil, nil, nil, nil, nil, err
+		return nil, nil, nil, nil, nil, nil, err
 	}
+	lits := &riscvLiterals{}
 
 	var relocs []Reloc
 	var spadj []SpadjStep
@@ -74,9 +76,9 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 	pc := len(prologue)
 	for i := range recs {
 		branchLike := isBranchLike(recs[i].instr.Mnemonic.Text) || riscvIsCondBranch(recs[i].instr.Mnemonic.Text)
-		code, err := encodeRISCVInstr(recs[i].instr, pc, offsets, fi, nil, nil) // no relocs in Pass 2
+		code, err := encodeRISCVInstr(recs[i].instr, pc, offsets, fi, nil, nil, lits) // no relocs in Pass 2
 		if err != nil && !(branchLike && riscvIsRangeError(err)) {
-			return nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", recs[i].instr.Mnemonic.Text, err)
+			return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", recs[i].instr.Mnemonic.Text, err)
 		}
 		if err != nil {
 			code = make([]byte, 4)
@@ -178,13 +180,20 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 			}
 		}
 		if !changed {
-			// Capture the final pcs for the N(PC) branch forms: their target
-			// is the instruction N source slots away, resolved by index.
+			// Capture the final pcs for the N(PC) branch and jump forms: the
+			// target is the instruction N source slots away (N=0 the branch
+			// itself, N negative backwards), resolved by index against the
+			// final layout.
 			pcRelPcs = map[*ast.Instr]int{}
 			for i := range recs {
-				if _, ok := riscvPCRelOffset(recs[i].instr); ok {
-					pcRelPcs[recs[i].instr] = pcs[i]
+				n, ok := riscvPCRelOffset(recs[i].instr)
+				if !ok {
+					continue
 				}
+				if i+n < 0 || i+n >= len(recs) {
+					continue
+				}
+				pcRelPcs[recs[i].instr] = pcs[i+n]
 			}
 			break
 		}
@@ -198,7 +207,7 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 	var out []byte
 	guardBytes, guardReloc, err := riscvGuard(fi)
 	if err != nil {
-		return nil, nil, nil, nil, nil, err
+		return nil, nil, nil, nil, nil, nil, err
 	}
 	if fi.needSplit {
 		out = append(out, guardBytes...)
@@ -220,11 +229,11 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 			// The JMP a relaxation inserted: JAL X0 to the original target.
 			targetOff, ok := offsets[r.jmpTo]
 			if !ok {
-				return nil, nil, nil, nil, nil, fmt.Errorf("undefined label %q", r.jmpTo)
+				return nil, nil, nil, nil, nil, nil, fmt.Errorf("undefined label %q", r.jmpTo)
 			}
 			offset := int32(targetOff - pc)
 			if err := riscvCheckJumpOffset(r.jmpTo, offset); err != nil {
-				return nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, err
 			}
 			word := riscvJType(0, offset)
 			code = []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}
@@ -233,7 +242,7 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 			// JMP, always the very next instruction (offset 4).
 			enc, rs1, rs2, ok := riscvInvertedBranchEnc(strings.ToUpper(r.instr.Mnemonic.Text), r.instr.Operands)
 			if !ok {
-				return nil, nil, nil, nil, nil, fmt.Errorf("%s: cannot relax branch", r.instr.Mnemonic.Text)
+				return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: cannot relax branch", r.instr.Mnemonic.Text)
 			}
 			word := riscvBType(enc, rs1, rs2, 4)
 			code = []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}
@@ -241,9 +250,9 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 			code = r.code
 		default:
 			var err error
-			code, err = encodeRISCVInstr(r.instr, pc, offsets, fi, &relocs, pcRelPcs)
+			code, err = encodeRISCVInstr(r.instr, pc, offsets, fi, &relocs, pcRelPcs, lits)
 			if err != nil {
-				return nil, nil, nil, nil, nil, err
+				return nil, nil, nil, nil, nil, nil, err
 			}
 			if c16, ok := tryCompressRVC(r.instr, fi); ok {
 				code = []byte{byte(c16), byte(c16 >> 8)}
@@ -270,7 +279,7 @@ func assembleRISCV(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, [
 	if fi.needSplit {
 		relocs = append(relocs, guardReloc)
 	}
-	return out, offsets, relocs, lines, spadj, nil
+	return out, offsets, relocs, lines, spadj, lits.list(), nil
 }
 
 // riscvImmAlias maps the R-type ALU mnemonics onto their I-type immediate
@@ -348,6 +357,11 @@ func riscvPadBytes(pad int) []byte {
 func riscvInstrSize(instr *ast.Instr, fi riscvFrameInfo) int {
 	mnem := instr.Mnemonic.Text
 	ops := instr.Operands
+	mnem = riscvNormalisePseudo(mnem)
+	if mnem == "FUNCDATA" || mnem == "PCDATA" {
+		// The bookkeeping statements contribute no bytes.
+		return 0
+	}
 	var immNeg bool
 	mnem, immNeg = riscvNormaliseImmAlias(mnem, ops)
 	if mnem == "RET" {
@@ -368,7 +382,25 @@ func riscvInstrSize(instr *ast.Instr, fi riscvFrameInfo) int {
 		}
 		// MOV $imm, rd → size depends on the immediate and RVC compression.
 		if isImmOperand(ops[0]) && ops[0].Imm.Sym == nil {
-			return riscvMovImmSize(regFromOperand(ops[1]), immFromOperand(ops[0]))
+			imm := riscvOperandImm64(ops[0])
+			if int64(int32(imm)) != imm {
+				return riscvMovImm64Size(regFromOperand(ops[1]), imm)
+			}
+			return riscvMovImmSize(regFromOperand(ops[1]), int32(imm))
+		}
+		// MOV $sym+off(FP|SP), rd → the frame-adjusted offset as an ADDI,
+		// compressed like riscvSPAddiBytes encodes it.
+		if isImmOperand(ops[0]) && ops[0].Imm.Sym != nil &&
+			(ops[0].Imm.Sym.Pseudo == "FP" || ops[0].Imm.Sym.Pseudo == "SP") {
+			rd := regFromOperand(ops[1])
+			_, off := riscvResolvePseudo(ops[0].Imm.Sym, fi)
+			if rd > 0 && off == 0 {
+				return 2 // C.MV rd, SP
+			}
+			if isRVCIntReg(rd) && off > 0 && off < 1024 && off%4 == 0 {
+				return 2 // C.ADDI4SPN
+			}
+			return riscvItypeImmediateSize("ADDI", off)
 		}
 		// Frame-relative loads and stores: a frame offset beyond the signed
 		// 12-bit range materialises the address in X31 first.
@@ -664,9 +696,10 @@ func riscvCheckJumpOffset(target string, off int32) error {
 }
 
 // encodeRISCVInstr encodes a single RISC-V instruction.
-func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscvFrameInfo, relocs *[]Reloc, pcRelPcs map[*ast.Instr]int) ([]byte, error) {
+func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscvFrameInfo, relocs *[]Reloc, pcRelPcs map[*ast.Instr]int, lits *riscvLiterals) ([]byte, error) {
 	mnem := instr.Mnemonic.Text
 	ops := instr.Operands
+	mnem = riscvNormalisePseudo(mnem)
 	var immNeg bool
 	mnem, immNeg = riscvNormaliseImmAlias(mnem, ops)
 	var word uint32
@@ -677,6 +710,22 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 		// RET = epilogue (restore LR and close the frame when present) +
 		// uncompressed JALR X0, 0(X1) (the toolchain never compresses RET).
 		return riscvReturn(fi), nil
+	case "FUNCDATA":
+		// The assembler's bookkeeping statement, the expanded form of the
+		// GO_ARGS and NO_LOCAL_POINTERS macros: FUNCDATA $n, sym(SB)
+		// contributes no bytes, exactly as the toolchain's listing shows
+		// (the FUNCDATA entries and the instruction after them share a PC).
+		if len(ops) != 2 || !isImmOperand(ops[0]) {
+			return nil, fmt.Errorf("FUNCDATA expects $n, sym(SB)")
+		}
+		return nil, nil
+	case "PCDATA":
+		// The other bookkeeping statement, the expanded form of
+		// GO_RESULTS_INITIALIZED: PCDATA $n, $m contributes no bytes too.
+		if len(ops) != 2 || !isImmOperand(ops[0]) || !isImmOperand(ops[1]) {
+			return nil, fmt.Errorf("PCDATA expects $n, $m")
+		}
+		return nil, nil
 	case "WORD":
 		// WORD $w lays down a raw 32-bit little-endian word.
 		if len(ops) != 1 {
@@ -739,6 +788,22 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 				return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
 			}
 			target = labelFromOperand(ops[0])
+			// JMP N(PC): the PC-relative slot form, resolved like the
+			// branches (the toolchain counts source instructions at a
+			// uniform 4 bytes, so JMP 0(PC) is a self-loop and JMP -3(PC)
+			// reaches twelve bytes back).  It must be recognised before the
+			// indirect-register form, whose operand it resembles.
+			if off, isPCRel, err := riscvPCRelTargetOff(instr, pc, pcRelPcs); isPCRel {
+				if err != nil {
+					return nil, err
+				}
+				offset := int32(off - pc)
+				if err := riscvCheckJumpOffset("", offset); err != nil {
+					return nil, err
+				}
+				word = riscvJType(0, offset)
+				return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
+			}
 			// JMP (X5): an indirect branch, the toolchain's JALR X0, 0(X5).
 			if ops[0].Addr.Sym == nil && ops[0].Addr.Base != "" {
 				if ops[0].Addr.Offset != 0 || ops[0].Addr.Index != "" {
@@ -751,17 +816,6 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 				word = riscvIType(riscvEnc{0x67, 0x0, 0x00}, 0, rs1, 0)
 				return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
 			}
-			if off, isPCRel, err := riscvPCRelTargetOff(instr, pc, pcRelPcs); isPCRel {
-				if err != nil {
-					return nil, err
-				}
-				offset := int32(off - pc)
-				if err := riscvCheckJumpOffset("", offset); err != nil {
-					return nil, err
-				}
-				word = riscvJType(0, offset)
-				return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
-			}
 		}
 		targetOff, ok := offsets[target]
 		if !ok {
@@ -810,7 +864,7 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 	// (MOVB/MOVH/MOVW and unsigned forms) select the access width, and
 	// MOVD/MOVF address the FP registers.
 	case "MOV", "MOVB", "MOVBU", "MOVH", "MOVHU", "MOVW", "MOVWU", "MOVF", "MOVD":
-		return encodeRISCVMov(instr, fi, relocs)
+		return encodeRISCVMov(instr, fi, relocs, lits)
 
 	// JALR: indirect jump/call.  Plan 9: JALR rs1, rd or JALR offset(rs1).
 	case "JALR":
@@ -1316,7 +1370,7 @@ func isImmOperand(op *ast.Operand) bool {
 //   - MOV Rs, (Rd)          register-relative store
 //   - MOV Rs, Rd            register-to-register move (ADDI $0)
 //   - MOV $imm, Rd          load immediate (ADDI or LUI+ADDIW)
-func encodeRISCVMov(instr *ast.Instr, fi riscvFrameInfo, relocs *[]Reloc) ([]byte, error) {
+func encodeRISCVMov(instr *ast.Instr, fi riscvFrameInfo, relocs *[]Reloc, lits *riscvLiterals) ([]byte, error) {
 	ops := instr.Operands
 	if len(ops) != 2 {
 		return nil, fmt.Errorf("MOV expects 2 operands, got %d", len(ops))
@@ -1335,8 +1389,21 @@ func encodeRISCVMov(instr *ast.Instr, fi riscvFrameInfo, relocs *[]Reloc) ([]byt
 			}
 			return encodeRISCVSBAddr(src.Imm.Sym, rd, relocs), nil
 		}
+		// MOV $sym+off(FP|SP), rd: the address of a frame slot as an
+		// immediate is the frame-adjusted offset against the hardware SP,
+		// the toolchain's ADDI $adj, SP, rd (argframe+0(FP) in the runtime's
+		// reflect trampolines is the spelling).
+		if src.Imm.Sym != nil && (src.Imm.Sym.Pseudo == "FP" || src.Imm.Sym.Pseudo == "SP") {
+			rd := regFromOperand(dst)
+			if rd < 0 {
+				return nil, fmt.Errorf("MOV $%s(%s): invalid destination register", src.Imm.Sym.Name, src.Imm.Sym.Pseudo)
+			}
+			_, off := riscvResolvePseudo(src.Imm.Sym, fi)
+			return riscvSPAddiBytes(rd, off), nil
+		}
 		// MOV $sym(FP/SP), rd, not supported: immediate symbol references
-		// other than SB cannot be encoded as a simple immediate.
+		// other than the frame pseudos cannot be encoded as a simple
+		// immediate.
 		if src.Imm.Sym != nil && src.Imm.Sym.Pseudo != "" {
 			return nil, fmt.Errorf("MOV $%s(%s): unsupported immediate symbol reference (only SB is supported)", src.Imm.Sym.Name, src.Imm.Sym.Pseudo)
 		}
@@ -1344,11 +1411,14 @@ func encodeRISCVMov(instr *ast.Instr, fi riscvFrameInfo, relocs *[]Reloc) ([]byt
 		if rd < 0 {
 			return nil, fmt.Errorf("MOV $imm: invalid destination register")
 		}
-		imm, err := riscvImm32FromOperand(src, false)
-		if err != nil {
-			return nil, err
+		imm := riscvOperandImm64(src)
+		if int64(int32(imm)) != imm {
+			// Beyond the signed 32-bit span the toolchain either builds the
+			// value from a shifted 32-bit part or loads it from the pooled
+			// $i64 constant it synthesises for the purpose.
+			return riscvLoadImm64(rd, imm, lits, relocs), nil
 		}
-		return encodeRISCVLoadImm(rd, imm), nil
+		return encodeRISCVLoadImm(rd, int32(imm)), nil
 	}
 
 	// Memory → register (load).
@@ -1557,6 +1627,186 @@ func splitRISCV32Imm(imm int32) (low, high int32) {
 	return low, high
 }
 
+// riscvNormalisePseudo rewrites the toolchain's UNDEF spelling onto EBREAK:
+// the assembler accepts UNDEF where the hardware wants the trap instruction
+// and emits ebreak (compressed to C.EBREAK under RVC), so every pass sees the
+// canonical name.
+func riscvNormalisePseudo(mnem string) string {
+	if strings.EqualFold(mnem, "UNDEF") {
+		return "EBREAK"
+	}
+	return mnem
+}
+
+// riscvOperandImm64 reads an immediate operand as a full signed 64-bit value,
+// where immFromOperand would truncate to int32; the MOV immediate path uses
+// it to classify the wide constants.
+func riscvOperandImm64(op *ast.Operand) int64 {
+	if !op.Imm.HasVal {
+		return 0
+	}
+	v := op.Imm.Val
+	if op.Imm.Neg {
+		v = -v
+	}
+	return v
+}
+
+// riscvSplitShiftConst mirrors cmd/internal/obj/riscv's splitShiftConst: it
+// looks for the signed 32-bit integer a constant can be rebuilt from with a
+// left shift, a left-and-right shift pair (a run of ones), or a zero-extended
+// 32-bit pattern.  A constant that fits none of the shapes is materialised
+// from the pooled $i64 data symbol instead.
+func riscvSplitShiftConst(v int64) (imm int64, lsh int, rsh int, ok bool) {
+	// Rebuild from a signed 32-bit integer shifted left.
+	lsh = bits.TrailingZeros64(uint64(v))
+	c := v >> lsh
+	if int64(int32(c)) == c {
+		return c, lsh, 0, true
+	}
+
+	// Rebuild from a small negative constant: shift left into place, then
+	// shift the sign-extended ones run right.
+	rsh = bits.LeadingZeros64(uint64(v))
+	ones := bits.OnesCount64((uint64(v) >> lsh) >> 11)
+	if rsh+ones+lsh+11 == 64 {
+		c = (1<<11 | ((v >> lsh) & 0x7ff)) << 52 >> 52 // sign extend 12 bits
+		if lsh > 0 || c != -1 {
+			lsh += rsh
+		}
+		return c, lsh, rsh, true
+	}
+
+	// Rebuild from a zero-extended signed 32-bit integer.
+	if int64(uint32(c)) == c {
+		c = int64(int32(c))
+		lsh, rsh = 32, 32-lsh
+		return c, lsh, rsh, true
+	}
+
+	return 0, 0, 0, false
+}
+
+// riscvSPAddiBytes encodes ADDI rd, SP, imm for the frame-address immediates
+// (the MOV $sym+off(FP|SP) form), using the compressed forms the toolchain
+// picks under RVC: C.ADDI4SPN for a positive 4-byte multiple that fits, C.MV
+// for the zero offset, the plain ADDI otherwise.
+func riscvSPAddiBytes(rd int, imm int32) []byte {
+	if rd != 0 && imm == 0 {
+		return word16(rvcCR(0x8, uint32(rd), 2)) // C.MV rd, SP
+	}
+	if isRVCIntReg(rd) && imm > 0 && imm < 1024 && imm%4 == 0 {
+		return word16(rvcCIW(0x0, rvcReg3(rd), uint32(imm)))
+	}
+	return wordLE(riscvIType(riscvInstrTable["ADDI"], rd, 2, imm))
+}
+
+// riscvMovImm64Size returns the encoded byte length of MOV $imm, rd when the
+// immediate sits outside the signed 32-bit span: the shifted-part sequences
+// of riscvLoadImm64, or the 8-byte AUIPC+LD pool load.
+func riscvMovImm64Size(rd int, imm int64) int {
+	c, lsh, rsh, ok := riscvSplitShiftConst(imm)
+	if !ok {
+		return 8 // AUIPC + LD against the $i64 pool symbol
+	}
+	size := riscvMovImmSize(rd, int32(c))
+	if lsh > 0 {
+		size += riscvShiftImmSize(rd, true)
+	}
+	if rsh > 0 {
+		size += riscvShiftImmSize(rd, false)
+	}
+	return size
+}
+
+// riscvShiftImmSize returns the encoded size of one SLLI/SRLI expansion
+// part: two bytes under RVC when the destination can carry a compressed
+// shift (C.SLLI admits every register but X0, C.SRLI only X8 to X15), four
+// otherwise.
+func riscvShiftImmSize(rd int, left bool) int {
+	if rd != 0 && (left || isRVCIntReg(rd)) {
+		return 2
+	}
+	return 4
+}
+
+// riscvLoadImm64 encodes MOV $imm, rd for an immediate beyond the signed
+// 32-bit span, mirroring the toolchain's instructionsForMOVConst: when a
+// shifted 32-bit part rebuilds the value it emits that part (compressed like
+// any written MOV) followed by the SLLI and SRLI shifts; otherwise it loads
+// the constant from the pooled read-only $i64.<hex> symbol via AUIPC + LD
+// and registers the literal so the data section carries its bytes.
+func riscvLoadImm64(rd int, imm int64, lits *riscvLiterals, relocs *[]Reloc) []byte {
+	c, lsh, rsh, ok := riscvSplitShiftConst(imm)
+	if !ok {
+		name := fmt.Sprintf("$i64.%016x", uint64(imm))
+		if lits != nil {
+			lits.add(name, riscvLiteralBytes(imm))
+		}
+		return encodeRISCVSBLoad(&ast.Symbol{Name: name}, rd, relocs)
+	}
+	out := encodeRISCVLoadImm(rd, int32(c))
+	if lsh > 0 {
+		out = append(out, riscvShiftImmBytes(rd, lsh, true)...)
+	}
+	if rsh > 0 {
+		out = append(out, riscvShiftImmBytes(rd, rsh, false)...)
+	}
+	return out
+}
+
+// riscvShiftImmBytes encodes one SLLI (left) or SRLI expansion part, using
+// the compressed form the toolchain picks under RVC: C.SLLI admits every
+// register but X0, C.SRLI only X8 to X15.
+func riscvShiftImmBytes(rd, shamt int, left bool) []byte {
+	if rd != 0 && shamt >= 1 && shamt <= 63 && (left || isRVCIntReg(rd)) {
+		if left {
+			return word16(rvcSLLI(uint32(rd), uint32(shamt)&0x3F))
+		}
+		return word16(rvcCBShift(0x0, rvcReg3(rd), uint32(shamt)&0x3F))
+	}
+	enc := riscvEnc{0x13, 0x1, 0x00} // SLLI
+	imm := int32(shamt)
+	if !left {
+		enc = riscvEnc{0x13, 0x5, 0x00} // SRLI: funct6 000000, funct3 101
+	}
+	return wordLE(riscvIType(enc, rd, rd, imm))
+}
+
+// riscvLiteralBytes renders a 64-bit constant as the little-endian bytes the
+// $i64 pool symbol holds.
+func riscvLiteralBytes(v int64) []byte {
+	return []byte{byte(v), byte(v >> 8), byte(v >> 16), byte(v >> 24),
+		byte(v >> 32), byte(v >> 40), byte(v >> 48), byte(v >> 56)}
+}
+
+// RiscvLiteral is one pooled 64-bit constant: a MOV whose immediate sits
+// beyond both the 32-bit span and the shift sequences loads its bits from a
+// read-only data symbol named like the toolchain's $i64 pool.
+type RiscvLiteral struct {
+	Name string
+	Data []byte
+}
+
+// riscvLiterals collects the pooled constants the MOV expansions refer to,
+// deduplicated by name, in first-use order.
+type riscvLiterals struct {
+	order []RiscvLiteral
+	seen  map[string]bool
+}
+
+func (l *riscvLiterals) add(name string, data []byte) {
+	if l.seen == nil {
+		l.seen = map[string]bool{}
+	}
+	if !l.seen[name] {
+		l.seen[name] = true
+		l.order = append(l.order, RiscvLiteral{Name: name, Data: data})
+	}
+}
+
+func (l *riscvLiterals) list() []RiscvLiteral { return l.order }
+
 // encodeRISCVItypeImmediate encodes an I-type arithmetic instruction, expanding
 // large immediates for ADDI/ANDI/ORI/XORI into LUI+ADDIW+op (or two ADDIs for
 // ADDI), matching the Go assembler.
@@ -1734,6 +1984,7 @@ func encodeRISCVJALR(instr *ast.Instr, fi riscvFrameInfo) ([]byte, error) {
 // RVC form.  It returns the compressed instruction word and true on success.
 func tryCompressRVC(instr *ast.Instr, fi riscvFrameInfo) (uint16, bool) {
 	mnem := riscvCompressMnem(instr)
+	mnem = riscvNormalisePseudo(mnem)
 	ops := instr.Operands
 	// The immediate aliases fold onto their I-type mnemonics before
 	// compression: the toolchain compresses ADD $imm, rd as c.addi, exactly

asm/riscv_encode.go

@@ -63,7 +63,7 @@ func riscvRegNum(name string) int {
 		return 24
 	case "X25", "S9":
 		return 25
-	case "X26", "S10":
+	case "X26", "S10", "CTXT":
 		return 26
 	case "X27", "S11", "g":
 		return 27

asm/riscv_encode_test.go

@@ -33,7 +33,7 @@ func firstTextRISCV(t *testing.T, src string) *ast.Text {
 // assembleRISCVHelper assembles one TEXT function and returns its code bytes.
 func assembleRISCVHelper(t *testing.T, fn *ast.Text) []byte {
 	t.Helper()
-	code, _, _, _, _, err := assembleRISCV(fn)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
 	if err != nil {
 		t.Fatalf("assemble: %v", err)
 	}
@@ -785,16 +785,140 @@ TEXT ·sys(SB), NOSPLIT, $0
 	}
 }
 
-func TestRISCV_MOV_sym_FP_error(t *testing.T) {
-	// MOV $sym(FP), rd should return an error (unsupported).
+func TestRISCV_MOV_sym_FP(t *testing.T) {
+	// MOV $sym(FP), rd lowers to the frame-adjusted ADDI against SP: the
+	// toolchain's argframe spelling.  A zero frame leaves the offset at the
+	// 8-byte link slot, compressed to C.ADDI4SPN.
 	fn := firstTextRISCV(t, `#include "textflag.h"
-TEXT ·badfp(SB), NOSPLIT, $0
+TEXT ·argfp(SB), NOSPLIT, $0
 	MOV $arg(FP), X10
 	RET
 `)
-	_, _, _, _, _, err := assembleRISCV(fn)
-	if err == nil {
-		t.Error("expected error for MOV $arg(FP), got nil")
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// prologue (0: leaf, zero frame) + C.ADDI4SPN (2) + RET (4) = 6
+	want := []byte{0x28, 0x00, 0x67, 0x80, 0x00, 0x00}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_Bookkeeping(t *testing.T) {
+	// FUNCDATA and PCDATA contribute no bytes; UNDEF is the toolchain's
+	// ebreak, compressed to C.EBREAK under RVC.
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·book(SB), NOSPLIT, $0-8
+	FUNCDATA $0, marks<>(SB)
+	PCDATA $1, $1
+	UNDEF
+	MOV $1, X10
+	MOV X10, ret+0(FP)
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// C.EBREAK (2) + C.LI X10, 1 (2) + C.SWSP (2) + RET (4) = 10: the
+	// FUNCDATA and PCDATA statements contribute nothing.
+	want := []byte{0x02, 0x90, 0x05, 0x45, 0x2a, 0xe4, 0x67, 0x80, 0x00, 0x00}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_JMPPCRel(t *testing.T) {
+	// JMP N(PC): the displacement tracks the instruction N source slots
+	// away in the final layout (0 the jump itself, negative backwards).
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·slots(SB), NOSPLIT, $0-0
+	JMP 2(PC)
+	MOV $1, X11
+	MOV $2, X12
+	MOV X12, X11
+	JMP -3(PC)
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// JMP 2(PC) lands on the C.MV six bytes ahead; JMP -3(PC) lands back on
+	// the first C.LI, six bytes behind.
+	want := []byte{
+		0x6f, 0x00, 0x60, 0x00, // JAL X0, 6
+		0x85, 0x45, // C.LI X11, 1
+		0x09, 0x46, // C.LI X12, 2
+		0xb2, 0x85, // C.MV X11, X12
+		0x6f, 0xf0, 0xbf, 0xff, // JAL X0, -6
+		0x67, 0x80, 0x00, 0x00, // RET
+	}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_MOVWideImm(t *testing.T) {
+	// Shift-sequence constants compress like the toolchain's expansion.
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·wide(SB), NOSPLIT, $0-0
+	MOV $0x8000000000000000, X5
+	MOV $0x100000000, X5
+	MOV $0x000fffffffffffda, X5
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// C.LI -1, C.SLLI 63; C.LI 1, C.SLLI 32; C.LI -19, C.SLLI 13, SRLI 12.
+	want := []byte{
+		0xfd, 0x52, 0xfe, 0x12,
+		0x85, 0x42, 0x82, 0x12,
+		0xb5, 0x52, 0xb6, 0x02, 0x93, 0xd2, 0xc2, 0x00,
+		0x67, 0x80, 0x00, 0x00,
+	}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_MOVImmPool(t *testing.T) {
+	// A constant outside the shift shapes loads from the pooled $i64 data
+	// symbol via AUIPC+LD, named like the toolchain's pool.
+	src := `#include "textflag.h"
+TEXT ·pool(SB), NOSPLIT, $0-8
+	MOV $0x0101010101010101, X16
+	MOV X16, ret+0(FP)
+	RET
+`
+	f, errs := parser.Parse("pool_riscv64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileRISCV(f)
+	if err != nil {
+		t.Fatalf("AssembleFileRISCV: %v", err)
+	}
+	// AUIPC X16, 0 + LD X16, 0(X16): the relocation pair carries the symbol.
+	wantCode := []byte{0x17, 0x08, 0x00, 0x00, 0x03, 0x38, 0x08, 0x00}
+	if string(img.Code[0:8]) != string(wantCode) {
+		t.Errorf("pool load: got % x", img.Code[0:8])
+	}
+	var lit *DataSymbol
+	for i := range img.DataSyms {
+		if img.DataSyms[i].Name == "$i64.0101010101010101" {
+			lit = &img.DataSyms[i]
+		}
+	}
+	if lit == nil {
+		t.Fatalf("pool symbol missing: %v", img.DataSyms)
+	}
+	wantData := []byte{0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01}
+	if string(img.Data[lit.Offset:lit.Offset+8]) != string(wantData) {
+		t.Errorf("pool bytes: got % x", img.Data[lit.Offset:lit.Offset+8])
 	}
 }
 
@@ -805,7 +929,7 @@ TEXT ·calltest(SB), NOSPLIT, $0
 	CALL ext(SB)
 	RET
 `)
-	code, _, relocs, _, _, err := assembleRISCV(fn)
+	code, _, relocs, _, _, _, err := assembleRISCV(fn)
 	if err != nil {
 		t.Fatalf("assemble: %v", err)
 	}
@@ -834,7 +958,7 @@ TEXT ·calllocal(SB), NOSPLIT, $0
 sub:
 	RET
 `)
-	_, _, _, _, _, err := assembleRISCV(fn)
+	_, _, _, _, _, _, err := assembleRISCV(fn)
 	if err == nil {
 		t.Error("expected error for CALL to local label, got nil")
 	}
@@ -868,7 +992,7 @@ func encodeOneInstrRISCV(t *testing.T, src string, pc int, offsets map[string]in
 	t.Helper()
 	fn := firstTextRISCV(t, "#include \"textflag.h\"\n"+src)
 	instr := fn.Body[0].(*ast.Instr)
-	return encodeRISCVInstr(instr, pc, offsets, riscvFrameInfo{}, nil, nil)
+	return encodeRISCVInstr(instr, pc, offsets, riscvFrameInfo{}, nil, nil, nil)
 }
 
 // TestRISCVBranchJumpRange checks that displacements beyond the B-type span
@@ -917,7 +1041,7 @@ func TestRISCVBranchFarBody(t *testing.T) {
 	}
 	sb.WriteString("done:\n\tRET\n")
 	fn := firstTextRISCV(t, sb.String())
-	out, _, _, _, _, err := assembleRISCV(fn)
+	out, _, _, _, _, _, err := assembleRISCV(fn)
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -943,7 +1067,7 @@ TEXT ·csrhi(SB), NOSPLIT, $0
 	CSRRW $4096, X10, X11
 	RET
 `)
-	if _, _, _, _, _, err := assembleRISCV(fn); err == nil {
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err == nil {
 		t.Error("expected an out-of-range error for CSR $4096, got none")
 	}
 	fn = firstTextRISCV(t, `#include "textflag.h"
@@ -951,25 +1075,24 @@ TEXT ·csrmax(SB), NOSPLIT, $0
 	CSRRW $4095, X10, X11
 	RET
 `)
-	if _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
 		t.Errorf("CSR $4095 must assemble: %v", err)
 	}
 }
 
 // TestRISCV_Imm64Rejected checks that immediates outside the signed 32-bit
-// span are diagnosed instead of silently truncated to their low 32 bits (the
-// toolchain materialises such constants via SLLI expansion, which this
-// assembler does not implement).
+// span are diagnosed instead of silently truncated to their low 32 bits for
+// the I-type arithmetic; the MOV forms materialise the wide constant instead
+// (shift sequence or pooled load), like the toolchain.
 func TestRISCV_Imm64Rejected(t *testing.T) {
 	cases := []string{
-		"MOV $0x123456789, X10",
 		"ADDI $0x100000000, X10, X11",
 		"ANDI $-0x800000001, X10, X11",
 		"SUB $0x100000000, X10, X11",
 	}
 	for _, src := range cases {
 		fn := firstTextRISCV(t, "#include \"textflag.h\"\nTEXT ·wide(SB), NOSPLIT, $0\n\t"+src+"\n\tRET\n")
-		if _, _, _, _, _, err := assembleRISCV(fn); err == nil {
+		if _, _, _, _, _, _, err := assembleRISCV(fn); err == nil {
 			t.Errorf("%s: expected an out-of-range error, got none", src)
 		}
 	}
@@ -982,9 +1105,19 @@ TEXT ·edge(SB), NOSPLIT, $0
 	SUB  $0x80000000, X12, X13
 	RET
 `)
-	if _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
 		t.Errorf("int32-span immediates must assemble: %v", err)
 	}
+	// Beyond the span the MOV forms materialise the constant like the
+	// toolchain instead of diagnosing it.
+	fn = firstTextRISCV(t, `#include "textflag.h"
+TEXT ·pool(SB), NOSPLIT, $0
+	MOV $0x123456789, X10
+	RET
+`)
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+		t.Errorf("MOV with a 64-bit immediate must assemble: %v", err)
+	}
 }
 
 // riscvWants decodes code as little-endian words and pins each one; the

cmd/gasm/audit.go

@@ -475,8 +475,8 @@ type corpusStats struct {
 // set, even when gasm does not support the architecture.
 var goPortSuffixes = []string{
 	"386", "amd64", "arm", "arm64", "loong64", "mips", "mips64",
-	"mips64le", "mipsle", "ppc64", "ppc64le", "riscv", "riscv64",
-	"s390x", "wasm",
+	"mips64le", "mipsle", "mips64x", "mipsx", "ppc64", "ppc64le",
+	"ppc64x", "riscv", "riscv64", "s390x", "wasm",
 }
 
 // otherPortFile reports whether the file belongs to a build no supported

testdata/verify/bookkeep_riscv64.s

@@ -0,0 +1,40 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the riscv64 bookkeeping statements and the
+// slot-relative branches: FUNCDATA and PCDATA (the expanded forms of the
+// funcdata.h macros, contributing no bytes), UNDEF (the toolchain's ebreak),
+// and the JMP N(PC) slot jumps including the self-loop and the backward form.
+
+#include "textflag.h"
+
+TEXT ·bookkeep(SB), NOSPLIT, $8-8
+	FUNCDATA $1, marks<>(SB)
+	PCDATA   $1, $-1
+	MOV      ZERO, ret+0(FP)
+	PCDATA   $1, $1
+	UNDEF
+	MOV      $1, X10
+	RET
+
+TEXT ·slots(SB), NOSPLIT, $0-0
+	MOV $1, X10
+	JMP 2(PC)
+	MOV $64, X11
+	MOV $128, X12
+	MOV $2, X11
+	MOV $3, X12
+	BEQ X10, X11, skip
+	JMP -2(PC)
+
+skip:
+	JMP 0(PC)
+
+TEXT ·marksreader(SB), NOSPLIT, $0-8
+	MOV $marks<>(SB), X10
+	MOV (X10), X11
+	MOV X11, ret+0(FP)
+	RET
+
+GLOBL marks<>(SB), RODATA, $8
+DATA marks<>+0(SB)/8, $1234605616436508552

testdata/verify/branch2_loong64.s

@@ -0,0 +1,21 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the loong64 two-operand BEQ/BNE spellings the
+// msan trampolines use: BEQ Rj, target compares against R0 (the beqz form).
+
+#include "textflag.h"
+
+TEXT ·branch2(SB), NOSPLIT, $0-8
+	MOVV arg+0(FP), R4
+	BEQ  R4, zero
+	ADDV $1, R4, R4
+
+zero:
+	MOVV $16, R5
+	BNE  R4, done
+	ADDV $2, R4, R4
+
+done:
+	MOVV R4, ret+0(FP)
+	RET

testdata/verify/datainit_loong64.s

@@ -0,0 +1,27 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the loong64 DATA value forms the runtime's exp and
+// asm files use: floating-point initialisers stored as IEEE-754 bits and
+// string initialisers zero-padded within their declared width.
+
+#include "textflag.h"
+
+TEXT ·floatbits(SB), NOSPLIT, $0-8
+	MOVV $floats<>(SB), R12
+	MOVD 8(R12), F0
+	MOVD F0, ret+0(FP)
+	RET
+
+TEXT ·stringhead(SB), NOSPLIT, $0-8
+	MOVV $msg<>(SB), R12
+	MOVV (R12), R13
+	MOVV R13, ret+0(FP)
+	RET
+
+GLOBL floats<>(SB), RODATA, $16
+DATA floats<>+0(SB)/8, $0.0
+DATA floats<>+8(SB)/8, $0.5
+
+GLOBL msg<>(SB), RODATA, $20
+DATA msg<>+0(SB)/20, $"call frame too large"

testdata/verify/wideimm_riscv64.s

@@ -0,0 +1,40 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the riscv64 wide MOV immediates: the 64-bit
+// constants the toolchain materialises through shifted 32-bit parts or loads
+// from its pooled $i64 data symbols, and the frame-address immediates
+// ($name+off(FP)) lowered to ADDI against the hardware stack pointer.
+
+#include "textflag.h"
+
+TEXT ·poolload(SB), NOSPLIT, $0-16
+	MOV arg+0(FP), X10
+	MOV $0x0101010101010101, X16
+	MUL X10, X16, X17
+	MOV $0x123456789, X18
+	ADD X18, X17, X17
+	MOV X17, ret+8(FP)
+	RET
+
+TEXT ·shiftload(SB), NOSPLIT, $0-16
+	MOV arg+0(FP), X10
+	MOV $0x8000000000000000, X20
+	MOV $0x100000000, X21
+	ADD X21, X20, X20
+	MOV $0x000fffffffffffda, X22
+	XOR X22, X20, X20
+	ADD X20, X10, X10
+	MOV X10, ret+8(FP)
+	RET
+
+TEXT ·frameaddr(SB), NOSPLIT, $32-16
+	MOV $ret+8(FP), X12
+	MOV $scratch+0(FP), X13
+	ADD X12, X13, X13
+	MOV $slot+16(SP), X14
+	SUB X14, X13, X13
+	MOV X13, ret+8(FP)
+	RET
+
+GLOBL scratch<>(SB), NOPTR, $8

verify/l64_groundtruth_test.go

@@ -31,6 +31,8 @@ func TestGroundTruthLOONG64(t *testing.T) {
 		"../testdata/verify/vector_loong64.s",
 		"../testdata/verify/pcalign_loong64.s",
 		"../testdata/verify/l64forms_loong64.s",
+		"../testdata/verify/datainit_loong64.s",
+		"../testdata/verify/branch2_loong64.s",
 		"../testdata/verify/vector_arith_add_loong64.s",
 		"../testdata/verify/vector_arith_sub_loong64.s",
 		"../testdata/verify/vector_arith_sat_loong64.s",

verify/riscv_groundtruth_test.go

@@ -35,6 +35,8 @@ func TestGroundTruthRISCV(t *testing.T) {
 		"../testdata/verify/bitmanip_riscv64.s",
 		"../testdata/verify/pcalign_riscv64.s",
 		"../testdata/verify/branch_far_riscv64.s",
+		"../testdata/verify/wideimm_riscv64.s",
+		"../testdata/verify/bookkeep_riscv64.s",
 		"trampoline_riscv64.s",
 	} {
 		t.Run(path, func(t *testing.T) {