[release-branch.go1.21] cmd/compile: fix memcombine pass for big endian, > 1 byte elements

The shift amounts were wrong in this case, leading to miscompilation
of load combining.

Also the store combining was not triggering when it should.

Fixes #64472

Change-Id: Iaeb08972c5fc1d6f628800334789c6af7216e87b
Reviewed-on: https://go-review.googlesource.com/c/go/+/546355
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Mauri de Souza Meneguzzo <mauri870@gmail.com>
Reviewed-by: Keith Randall <khr@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-on: https://go-review.googlesource.com/c/go/+/546356

3 files changed, 169 insertions, 6 deletions

diff --git a/src/cmd/compile/internal/ssa/memcombine.go b/src/cmd/compile/internal/ssa/memcombine.go
index c1346434c9..26fb3f5454 100644
--- a/src/cmd/compile/internal/ssa/memcombine.go
+++ b/src/cmd/compile/internal/ssa/memcombine.go
@@ -313,8 +313,8 @@ func combineLoads(root *Value, n int64) bool {
 	if isLittleEndian && shift0 != 0 {
 		v = leftShift(loadBlock, pos, v, shift0)
 	}
-	if isBigEndian && shift0-(n-1)*8 != 0 {
-		v = leftShift(loadBlock, pos, v, shift0-(n-1)*8)
+	if isBigEndian && shift0-(n-1)*size*8 != 0 {
+		v = leftShift(loadBlock, pos, v, shift0-(n-1)*size*8)
 	}
 
 	// Install with (Copy v).
@@ -588,14 +588,14 @@ func combineStores(root *Value, n int64) bool {
 	isLittleEndian := true
 	shift0 := shift(a[0].store, shiftBase)
 	for i := int64(1); i < n; i++ {
-		if shift(a[i].store, shiftBase) != shift0+i*8 {
+		if shift(a[i].store, shiftBase) != shift0+i*size*8 {
 			isLittleEndian = false
 			break
 		}
 	}
 	isBigEndian := true
 	for i := int64(1); i < n; i++ {
-		if shift(a[i].store, shiftBase) != shift0-i*8 {
+		if shift(a[i].store, shiftBase) != shift0-i*size*8 {
 			isBigEndian = false
 			break
 		}
@@ -618,8 +618,8 @@ func combineStores(root *Value, n int64) bool {
 	if isLittleEndian && shift0 != 0 {
 		sv = rightShift(root.Block, root.Pos, sv, shift0)
 	}
-	if isBigEndian && shift0-(n-1)*8 != 0 {
-		sv = rightShift(root.Block, root.Pos, sv, shift0-(n-1)*8)
+	if isBigEndian && shift0-(n-1)*size*8 != 0 {
+		sv = rightShift(root.Block, root.Pos, sv, shift0-(n-1)*size*8)
 	}
 	if sv.Type.Size() > size*n {
 		sv = truncate(root.Block, root.Pos, sv, sv.Type.Size(), size*n)
diff --git a/src/cmd/compile/internal/test/memcombine_test.go b/src/cmd/compile/internal/test/memcombine_test.go
index c7e7a208dd..3fc4a004a3 100644
--- a/src/cmd/compile/internal/test/memcombine_test.go
+++ b/src/cmd/compile/internal/test/memcombine_test.go
@@ -71,3 +71,129 @@ func readUint32be(b []byte) uint64 {
 //go:noinline
 func nop() {
 }
+
+type T32 struct {
+	a, b uint32
+}
+
+//go:noinline
+func (t *T32) bigEndianLoad() uint64 {
+	return uint64(t.a)<<32 | uint64(t.b)
+}
+
+//go:noinline
+func (t *T32) littleEndianLoad() uint64 {
+	return uint64(t.a) | (uint64(t.b) << 32)
+}
+
+//go:noinline
+func (t *T32) bigEndianStore(x uint64) {
+	t.a = uint32(x >> 32)
+	t.b = uint32(x)
+}
+
+//go:noinline
+func (t *T32) littleEndianStore(x uint64) {
+	t.a = uint32(x)
+	t.b = uint32(x >> 32)
+}
+
+type T16 struct {
+	a, b uint16
+}
+
+//go:noinline
+func (t *T16) bigEndianLoad() uint32 {
+	return uint32(t.a)<<16 | uint32(t.b)
+}
+
+//go:noinline
+func (t *T16) littleEndianLoad() uint32 {
+	return uint32(t.a) | (uint32(t.b) << 16)
+}
+
+//go:noinline
+func (t *T16) bigEndianStore(x uint32) {
+	t.a = uint16(x >> 16)
+	t.b = uint16(x)
+}
+
+//go:noinline
+func (t *T16) littleEndianStore(x uint32) {
+	t.a = uint16(x)
+	t.b = uint16(x >> 16)
+}
+
+type T8 struct {
+	a, b uint8
+}
+
+//go:noinline
+func (t *T8) bigEndianLoad() uint16 {
+	return uint16(t.a)<<8 | uint16(t.b)
+}
+
+//go:noinline
+func (t *T8) littleEndianLoad() uint16 {
+	return uint16(t.a) | (uint16(t.b) << 8)
+}
+
+//go:noinline
+func (t *T8) bigEndianStore(x uint16) {
+	t.a = uint8(x >> 8)
+	t.b = uint8(x)
+}
+
+//go:noinline
+func (t *T8) littleEndianStore(x uint16) {
+	t.a = uint8(x)
+	t.b = uint8(x >> 8)
+}
+
+func TestIssue64468(t *testing.T) {
+	t32 := T32{1, 2}
+	if got, want := t32.bigEndianLoad(), uint64(1<<32+2); got != want {
+		t.Errorf("T32.bigEndianLoad got %x want %x\n", got, want)
+	}
+	if got, want := t32.littleEndianLoad(), uint64(1+2<<32); got != want {
+		t.Errorf("T32.littleEndianLoad got %x want %x\n", got, want)
+	}
+	t16 := T16{1, 2}
+	if got, want := t16.bigEndianLoad(), uint32(1<<16+2); got != want {
+		t.Errorf("T16.bigEndianLoad got %x want %x\n", got, want)
+	}
+	if got, want := t16.littleEndianLoad(), uint32(1+2<<16); got != want {
+		t.Errorf("T16.littleEndianLoad got %x want %x\n", got, want)
+	}
+	t8 := T8{1, 2}
+	if got, want := t8.bigEndianLoad(), uint16(1<<8+2); got != want {
+		t.Errorf("T8.bigEndianLoad got %x want %x\n", got, want)
+	}
+	if got, want := t8.littleEndianLoad(), uint16(1+2<<8); got != want {
+		t.Errorf("T8.littleEndianLoad got %x want %x\n", got, want)
+	}
+	t32.bigEndianStore(1<<32 + 2)
+	if got, want := t32, (T32{1, 2}); got != want {
+		t.Errorf("T32.bigEndianStore got %x want %x\n", got, want)
+	}
+	t32.littleEndianStore(1<<32 + 2)
+	if got, want := t32, (T32{2, 1}); got != want {
+		t.Errorf("T32.littleEndianStore got %x want %x\n", got, want)
+	}
+	t16.bigEndianStore(1<<16 + 2)
+	if got, want := t16, (T16{1, 2}); got != want {
+		t.Errorf("T16.bigEndianStore got %x want %x\n", got, want)
+	}
+	t16.littleEndianStore(1<<16 + 2)
+	if got, want := t16, (T16{2, 1}); got != want {
+		t.Errorf("T16.littleEndianStore got %x want %x\n", got, want)
+	}
+	t8.bigEndianStore(1<<8 + 2)
+	if got, want := t8, (T8{1, 2}); got != want {
+		t.Errorf("T8.bigEndianStore got %x want %x\n", got, want)
+	}
+	t8.littleEndianStore(1<<8 + 2)
+	if got, want := t8, (T8{2, 1}); got != want {
+		t.Errorf("T8.littleEndianStore got %x want %x\n", got, want)
+	}
+}
diff --git a/test/codegen/memcombine.go b/test/codegen/memcombine.go
index 0d1c390dfc..3cf4fbc0ae 100644
--- a/test/codegen/memcombine.go
+++ b/test/codegen/memcombine.go
@@ -836,3 +836,40 @@ func zero_uint64_2(d1, d2 []uint64) {
 	d1[0], d1[1] = 0, 0 // arm64:"STP",-"MOVB",-"MOVH"
 	d2[1], d2[0] = 0, 0 // arm64:"STP",-"MOVB",-"MOVH"
 }
+
+func store32le(p *struct{ a, b uint32 }, x uint64) {
+	// amd64:"MOVQ",-"MOVL",-"SHRQ"
+	// arm64:"MOVD",-"MOVW",-"LSR"
+	// ppc64le:"MOVD",-"MOVW",-"SRD"
+	p.a = uint32(x)
+	// amd64:-"MOVL",-"SHRQ"
+	// arm64:-"MOVW",-"LSR"
+	// ppc64le:-"MOVW",-"SRD"
+	p.b = uint32(x >> 32)
+}
+func store32be(p *struct{ a, b uint32 }, x uint64) {
+	// ppc64:"MOVD",-"MOVW",-"SRD"
+	// s390x:"MOVD",-"MOVW",-"SRD"
+	p.a = uint32(x >> 32)
+	// ppc64:-"MOVW",-"SRD"
+	// s390x:-"MOVW",-"SRD"
+	p.b = uint32(x)
+}
+func store16le(p *struct{ a, b uint16 }, x uint32) {
+	// amd64:"MOVL",-"MOVW",-"SHRL"
+	// arm64:"MOVW",-"MOVH",-"UBFX"
+	// ppc64le:"MOVW",-"MOVH",-"SRW"
+	p.a = uint16(x)
+	// amd64:-"MOVW",-"SHRL"
+	// arm64:-"MOVH",-"UBFX"
+	// ppc64le:-"MOVH",-"SRW"
+	p.b = uint16(x >> 16)
+}
+func store16be(p *struct{ a, b uint16 }, x uint32) {
+	// ppc64:"MOVW",-"MOVH",-"SRW"
+	// s390x:"MOVW",-"MOVH",-"SRW"
+	p.a = uint16(x >> 16)
+	// ppc64:-"MOVH",-"SRW"
+	// s390x:-"MOVH",-"SRW"
+	p.b = uint16(x)
+}