fix(asm): carry the riscv64 scalar pseudos and the swapped branches

Assisted-by: GLM 5.3 Flash

asm/riscv_assemble.go

@@ -437,7 +437,7 @@ func riscvExtendedSize(mnem string, ops []*ast.Operand) int {
 	case "NOP":
 		// The toolchain drops a bare NOP entirely.
 		return 0
-	case "ANDN", "ORN", "XNOR":
+	case "ANDN", "ORN", "XNOR", "FNES", "FNED":
 		return 8
 	case "MAX", "MAXU", "MIN", "MINU":
 		if riscvIdenticalMinMax(mnem, ops) {
@@ -894,7 +894,7 @@ func riscvPCRelOffset(instr *ast.Instr) (int, bool) {
 		if len(instr.Operands) != 1 && len(instr.Operands) != 2 {
 			return 0, false
 		}
-	case "BEQ", "BNE", "BLT", "BGE", "BLTU", "BGEU",
+	case "BEQ", "BNE", "BLT", "BGE", "BLTU", "BGEU", "BGT", "BLE", "BGTU", "BLEU",
 		"BEQZ", "BNEZ", "BLTZ", "BGEZ", "BLEZ", "BGTZ":
 		if len(instr.Operands) < 2 {
 			return 0, false
@@ -1138,16 +1138,22 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 				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).
+			// JMP (X5) and JMP 4(X5): an indirect branch, the toolchain's
+			// JALR X0, imm(X5) with the offset carried in the I-type
+			// immediate (JMP 4(X5) is 0x67804200).
 			if ops[0].Addr.Sym == nil && ops[0].Addr.Base != "" {
-				if ops[0].Addr.Offset != 0 || ops[0].Addr.Index != "" {
+				if ops[0].Addr.Index != "" {
 					return nil, fmt.Errorf("JMP: invalid indirect operand %q", ops[0].Raw)
 				}
 				rs1 := riscvRegNum(ops[0].Addr.Base)
 				if rs1 < 0 {
 					return nil, fmt.Errorf("JMP: unknown branch register %q", ops[0].Addr.Base)
 				}
-				word = riscvIType(riscvEnc{0x67, 0x0, 0x00}, 0, rs1, 0)
+				imm := int32(ops[0].Addr.Offset)
+				if imm < -2048 || imm > 2047 {
+					return nil, fmt.Errorf("JMP: displacement %d does not fit in 12 bits", imm)
+				}
+				word = riscvIType(riscvEnc{0x67, 0x0, 0x00}, 0, rs1, imm)
 				return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
 			}
 		}
@@ -1301,6 +1307,47 @@ func encodeRISCVInstr(instr *ast.Instr, pc int, offsets map[string]int, fi riscv
 		}
 		word = riscvRType(riscvEnc{0x73, 0x0, 0x09}, 0, rs1, rs2)
 		return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
+
+	case "NEG", "NEGW":
+		// INSTR rs [, rd]: SUB/SUBW with X0 in the rs1 field, the
+		// one-operand form negating in place (ANEG: NEG rs, rd -> SUB rs,
+		// X0, rd).  The toolchain pins the bytes: NEG X5 is 0x405002b3.
+		if len(ops) != 1 && len(ops) != 2 {
+			return nil, fmt.Errorf("%s expects 1 or 2 operands, got %d", mnem, len(ops))
+		}
+		rs := regFromOperand(ops[0])
+		rd := rs
+		if len(ops) == 2 {
+			rd = regFromOperand(ops[1])
+		}
+		if rs < 0 || rd < 0 {
+			return nil, fmt.Errorf("%s: invalid register operand", mnem)
+		}
+		enc := riscvEnc{0x33, 0x0, 0x20} // sub
+		if mnem == "NEGW" {
+			enc = riscvEnc{0x3B, 0x0, 0x20} // subw
+		}
+		word = riscvRType(enc, rd, 0, rs)
+		return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
+
+	case "SEQZ", "SNEZ":
+		// INSTR rs, rd: the set-equal and set-not-equal pseudos read as
+		// SLTIU $1 and SLTU against X0 (ASEQZ/ASNEZ).  The toolchain pins
+		// the bytes: SEQZ X14, X15 is 0x00173793.
+		if len(ops) != 2 {
+			return nil, fmt.Errorf("%s expects 2 operands, got %d", mnem, len(ops))
+		}
+		rs := regFromOperand(ops[0])
+		rd := regFromOperand(ops[1])
+		if rs < 0 || rd < 0 {
+			return nil, fmt.Errorf("%s: invalid register operand", mnem)
+		}
+		if mnem == "SEQZ" {
+			word = riscvIType(riscvEnc{0x13, 0x3, 0x00}, rd, rs, 1) // sltiu $1
+		} else {
+			word = riscvRType(riscvEnc{0x33, 0x3, 0x00}, rd, 0, rs) // sltu rd, x0, rs
+		}
+		return []byte{byte(word), byte(word >> 8), byte(word >> 16), byte(word >> 24)}, nil
 	}
 
 	// FP conversion / move instructions use a separate table (rs2 encodes
@@ -2767,7 +2814,9 @@ func encodeRISCVExtended(mnem string, instr *ast.Instr, pc int, offsets map[stri
 		}
 		return wordLE(riscvIType(riscvEnc{0x73, 0x2, 0x00}, rd, 0, 0xC01)), true, nil
 
-	case "NEG", "NOT", "SEQZ":
+	case "NOT":
+		// NEG and SEQZ have their own cases in encodeRISCVInstr's pseudo
+		// switch; NOT reads as XORI $-1.
 		if len(ops) != 1 && len(ops) != 2 {
 			return nil, true, fmt.Errorf("%s expects 1 or 2 operands, got %d", mnem, len(ops))
 		}
@@ -2781,12 +2830,8 @@ func encodeRISCVExtended(mnem string, instr *ast.Instr, pc int, offsets map[stri
 		}
 		var word uint32
 		switch mnem {
-		case "NEG":
-			word = riscvRType(riscvInstrTable["SUB"], rd, 0, rs)
 		case "NOT":
 			word = riscvIType(riscvInstrTable["XORI"], rd, rs, -1)
-		case "SEQZ":
-			word = riscvIType(riscvInstrTable["SLTIU"], rd, rs, 1)
 		}
 		return wordLE(word), true, nil
 
@@ -2975,49 +3020,49 @@ func encodeRISCVExtended(mnem string, instr *ast.Instr, pc int, offsets map[stri
 		out = append(out, wordLE(riscvRType(riscvInstrTable["OR"], rd, 31, rd))...)
 		return out, true, nil
 
-	case "BGT", "BGTU", "BLE", "BLEU":
+		// BGT/BGTU/BLE/BLEU have no extended handler: the main table's
-		// The reversed conditional branches: BGT a, b, label is BLT b, a.
+		// branch path owns them, including the N(PC) forms.
-		if len(ops) != 3 {
-			return nil, true, fmt.Errorf("%s expects 3 operands, got %d", mnem, len(ops))
-		}
-		a := regFromOperand(ops[0])
-		b := regFromOperand(ops[1])
-		if a < 0 || b < 0 {
-			return nil, true, fmt.Errorf("%s: invalid register", mnem)
-		}
-		target := labelFromOperand(ops[2])
-		targetOff, ok := offsets[target]
-		if !ok {
-			return nil, true, fmt.Errorf("undefined label %q%s", target, suggestLabel(target, offsets))
-		}
-		offset := int32(targetOff - pc)
-		if err := riscvCheckBranchOffset(target, offset); err != nil {
-			return nil, true, err
-		}
-		var enc riscvEnc
-		switch mnem {
-		case "BGT":
-			enc = riscvEnc{0x63, 0x4, 0x00} // blt b, a
-		case "BGTU":
-			enc = riscvEnc{0x63, 0x6, 0x00} // bltu b, a
-		case "BLE":
-			enc = riscvEnc{0x63, 0x5, 0x00} // bge b, a
-		case "BLEU":
-			enc = riscvEnc{0x63, 0x7, 0x00} // bgeu b, a
-		}
-		return wordLE(riscvBType(enc, b, a, offset)), true, nil
 
-	case "FABSD":
+	case "FABSS", "FABSD", "FNEGS", "FNEGD":
-		// FABSD rs, rd is FSGNJX.D (sign XOR, funct3 2) with the source in
+		// INSTR fs, fd: the sign-injection pseudos, the source in both the
+		// rs1 and rs2 fields (AFABSS: FSGNJXS rs, rs, rd; AFNEGS: FSGNJNS).
+		// The toolchain pins the bytes: FABSS F0, F1 is 0x200020d3.
 		if len(ops) != 2 {
-			return nil, true, fmt.Errorf("FABSD expects 2 operands, got %d", len(ops))
+			return nil, true, fmt.Errorf("%s expects 2 operands, got %d", mnem, len(ops))
 		}
 		rs := regFromOperand(ops[0])
 		rd := regFromOperand(ops[1])
 		if rs < 0 || rd < 0 {
-			return nil, true, fmt.Errorf("FABSD: invalid register")
+			return nil, true, fmt.Errorf("%s: invalid register", mnem)
 		}
-		return wordLE(riscvRType(riscvEnc{0x53, 0x2, 0x11}, rd, rs, rs)), true, nil
+		enc := map[string]riscvEnc{
+			"FABSS": {0x53, 0x2, 0x10}, // fsgnjx.s
+			"FABSD": {0x53, 0x2, 0x11}, // fsgnjx.d
+			"FNEGS": {0x53, 0x1, 0x10}, // fsgnjn.s
+			"FNEGD": {0x53, 0x1, 0x11}, // fsgnjn.d
+		}[mnem]
+		return wordLE(riscvRType(enc, rd, rs, rs)), true, nil
+
+	case "FNES", "FNED":
+		// INSTR fs1, fs2, xrd: the not-equal pseudos read as FEQ.S/FEQ.D
+		// followed by XORI $1 on the result register (AFNES), two words.
+		if len(ops) != 3 {
+			return nil, true, fmt.Errorf("%s expects 3 operands, got %d", mnem, len(ops))
+		}
+		rs2 := regFromOperand(ops[0])
+		rs1 := regFromOperand(ops[1])
+		rd := regFromOperand(ops[2])
+		if rs1 < 0 || rs2 < 0 || rd < 0 {
+			return nil, true, fmt.Errorf("%s: invalid register", mnem)
+		}
+		feq := riscvEnc{0x53, 0x2, 0x50} // feq.s
+		if mnem == "FNED" {
+			feq = riscvEnc{0x53, 0x2, 0x51} // feq.d
+		}
+		eq := riscvRType(feq, rd, rs1, rs2)
+		not := riscvIType(riscvEnc{0x13, 0x4, 0x00}, rd, rd, 1) // xori $1
+		return []byte{byte(eq), byte(eq >> 8), byte(eq >> 16), byte(eq >> 24),
+			byte(not), byte(not >> 8), byte(not >> 16), byte(not >> 24)}, true, nil
 
 	default:
 		return encodeRISCVVector(mnem, ops)

asm/riscv_vecarith_test.go

@@ -810,3 +810,36 @@ TEXT ·vecarith(SB), NOSPLIT, $0
 	}
 	assertRISCVDifferential(t, path, src, "vecarith")
 }
+
+// TestRISCVScalarPseudos_Differential proves the scalar pseudo tail against
+// the toolchain: the NEG/NEGW and SEQZ/SNEZ one- and two-operand forms, the
+// sign-injection pseudos and the two-word not-equal expansions, exactly as
+// the toolchain's own testdata spells them.
+func TestRISCVScalarPseudos_Differential(t *testing.T) {
+	src := `#include "textflag.h"
+
+TEXT ·pseudos(SB), NOSPLIT, $0-8
+	NEG	X5
+	NEG	X5, X6
+	NEGW	X5
+	NEGW	X5, X6
+	SEQZ	X14, X15
+	SNEZ	X14, X15
+	FABSS	F0, F1
+	FNEGS	F0, F1
+	FNES	F0, F1, X7
+	FABSD	F0, F1
+	FNEGD	F0, F1
+	FNED	F0, F1, X5
+	FLTD	F0, F1, X5
+	FLED	F0, F1, X5
+	FEQD	F0, F1, X5
+	RET
+`
+	dir := t.TempDir()
+	path := filepath.Join(dir, "pseudos_riscv64.s")
+	if err := os.WriteFile(path, []byte(src), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	assertRISCVDifferential(t, path, src, "pseudos")
+}