cmd/compile: fix sign extension of paired 32-bit loads on arm64

Fixes #71759

Change-Id: Iab05294ac933cc9972949158d3fe2bdc3073df5e
Reviewed-on: https://go-review.googlesource.com/c/go/+/649895
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Cherry Mui <[email protected]>
Reviewed-by: Dmitri Shuralyov <[email protected]>

Author: randall77

src/cmd/compile/internal/arm64/ssa.go

@@ -516,7 +516,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		ssagen.AddAux(&p.From, v)
 		p.To.Type = obj.TYPE_REG
 		p.To.Reg = v.Reg()
-	case ssa.OpARM64LDP, ssa.OpARM64LDPW, ssa.OpARM64FLDPD, ssa.OpARM64FLDPS:
+	case ssa.OpARM64LDP, ssa.OpARM64LDPW, ssa.OpARM64LDPSW, ssa.OpARM64FLDPD, ssa.OpARM64FLDPS:
 		p := s.Prog(v.Op.Asm())
 		p.From.Type = obj.TYPE_MEM
 		p.From.Reg = v.Args[0].Reg()

src/cmd/compile/internal/ssa/_gen/ARM64Ops.go

@@ -388,7 +388,8 @@ func init() {
 		// arg1=mem
 		// Returns the tuple <x,y>.
 		{name: "LDP", argLength: 2, reg: gpload2, aux: "SymOff", asm: "LDP", typ: "(UInt64,UInt64)", faultOnNilArg0: true, symEffect: "Read"},       // T=int64 (gp reg destination)
-		{name: "LDPW", argLength: 2, reg: gpload2, aux: "SymOff", asm: "LDPW", typ: "(UInt32,UInt32)", faultOnNilArg0: true, symEffect: "Read"},     // T=int32 (gp reg destination) ??? extension
+		{name: "LDPW", argLength: 2, reg: gpload2, aux: "SymOff", asm: "LDPW", typ: "(UInt32,UInt32)", faultOnNilArg0: true, symEffect: "Read"},     // T=int32 (gp reg destination) unsigned extension
+		{name: "LDPSW", argLength: 2, reg: gpload2, aux: "SymOff", asm: "LDPSW", typ: "(Int32,Int32)", faultOnNilArg0: true, symEffect: "Read"},     // T=int32 (gp reg destination) signed extension
 		{name: "FLDPD", argLength: 2, reg: fpload2, aux: "SymOff", asm: "FLDPD", typ: "(Float64,Float64)", faultOnNilArg0: true, symEffect: "Read"}, // T=float64 (fp reg destination)
 		{name: "FLDPS", argLength: 2, reg: fpload2, aux: "SymOff", asm: "FLDPS", typ: "(Float32,Float32)", faultOnNilArg0: true, symEffect: "Read"}, // T=float32 (fp reg destination)
 

src/cmd/compile/internal/ssa/opGen.go

@@ -32,7 +32,10 @@ type pairableLoadInfo struct {
 // They must also take an offset in Aux/AuxInt.
 var pairableLoads = map[Op]pairableLoadInfo{
 	OpARM64MOVDload:  {8, OpARM64LDP},
-	OpARM64MOVWload:  {4, OpARM64LDPW},
+	OpARM64MOVWUload: {4, OpARM64LDPW},
+	OpARM64MOVWload:  {4, OpARM64LDPSW},
+	// TODO: conceivably we could pair a signed and unsigned load
+	// if we knew the upper bits of one of them weren't being used.
 	OpARM64FMOVDload: {8, OpARM64FLDPD},
 	OpARM64FMOVSload: {4, OpARM64FLDPS},
 }

src/cmd/compile/internal/ssa/pair.go

@@ -982,6 +982,10 @@ func dwloadI64(p *struct{ a, b int64 }) int64 {
 	return p.a + p.b
 }
 func dwloadI32(p *struct{ a, b int32 }) int32 {
+	// arm64:"LDPSW\t"
+	return p.a + p.b
+}
+func dwloadU32(p *struct{ a, b uint32 }) uint32 {
 	// arm64:"LDPW\t"
 	return p.a + p.b
 }

test/codegen/memcombine.go

@@ -0,0 +1,20 @@
+// run
+
+// Copyright 2025 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+//go:noinline
+func f(p *[2]int32) (int64, int64) {
+	return int64(p[0]), int64(p[1])
+}
+
+func main() {
+	p := [2]int32{-1, -1}
+	x, y := f(&p)
+	if x != -1 || y != -1 {
+		println(x, y)
+	}
+}