[release-branch.go1.9] cmd/compile: fix regression in PPC64.rules move zero

When a MOVDstorezero (8 bytes) is used the offset field
in the instruction must be a multiple of 4. This situation
had been corrected in the rules for other types of stores
but not for the zero case.

This also removes some of the special MOVDstorezero cases since
they can be handled by the general LowerZero case.

Updates made to the ssa test for lowering zero moves to
include cases where the target is not aligned to at least 4.

Fixes #21947

Change-Id: I7cceceb1be4898c77cd3b5e78b58dce0a7e28edd
Reviewed-on: https://go-review.googlesource.com/64970
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Carlos Eduardo Seo <cseo@linux.vnet.ibm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-on: https://go-review.googlesource.com/70978
Run-TryBot: Russ Cox <rsc@golang.org>

4 files changed, 350 insertions, 159 deletions

diff --git a/src/cmd/compile/internal/gc/testdata/gen/zeroGen.go b/src/cmd/compile/internal/gc/testdata/gen/zeroGen.go
index fa70b16495..c764c369e6 100644
--- a/src/cmd/compile/internal/gc/testdata/gen/zeroGen.go
+++ b/src/cmd/compile/internal/gc/testdata/gen/zeroGen.go
@@ -19,6 +19,7 @@ import (
 // will be written into the parent directory containing the tests.
 
 var sizes = [...]int{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 16, 17, 23, 24, 25, 31, 32, 33, 63, 64, 65, 1023, 1024, 1025}
+var usizes = [...]int{8, 16, 24, 32, 64, 256}
 
 func main() {
 	w := new(bytes.Buffer)
@@ -61,12 +62,74 @@ func main() {
 		fmt.Fprintf(w, "}\n")
 	}
 
+	for _, s := range usizes {
+		// type for test
+		fmt.Fprintf(w, "type T%du1 struct {\n", s)
+		fmt.Fprintf(w, "  b   bool\n")
+		fmt.Fprintf(w, "  val [%d]byte\n", s)
+		fmt.Fprintf(w, "}\n")
+
+		fmt.Fprintf(w, "type T%du2 struct {\n", s)
+		fmt.Fprintf(w, "  i   uint16\n")
+		fmt.Fprintf(w, "  val [%d]byte\n", s)
+		fmt.Fprintf(w, "}\n")
+
+		// function being tested
+		fmt.Fprintf(w, "//go:noinline\n")
+		fmt.Fprintf(w, "func zero%du1_ssa(t *T%du1) {\n", s, s)
+		fmt.Fprintf(w, "  t.val = [%d]byte{}\n", s)
+		fmt.Fprintf(w, "}\n")
+
+		// function being tested
+		fmt.Fprintf(w, "//go:noinline\n")
+		fmt.Fprintf(w, "func zero%du2_ssa(t *T%du2) {\n", s, s)
+		fmt.Fprintf(w, "  t.val = [%d]byte{}\n", s)
+		fmt.Fprintf(w, "}\n")
+
+		// testing harness
+		fmt.Fprintf(w, "func testZero%du() {\n", s)
+		fmt.Fprintf(w, "  a := T%du1{false, [%d]byte{", s, s)
+		for i := 0; i < s; i++ {
+			fmt.Fprintf(w, "255,")
+		}
+		fmt.Fprintf(w, "}}\n")
+		fmt.Fprintf(w, "  zero%du1_ssa(&a)\n", s)
+		fmt.Fprintf(w, "  want := T%du1{false, [%d]byte{", s, s)
+		for i := 0; i < s; i++ {
+			fmt.Fprintf(w, "0,")
+		}
+		fmt.Fprintf(w, "}}\n")
+		fmt.Fprintf(w, "  if a != want {\n")
+		fmt.Fprintf(w, "    fmt.Printf(\"zero%du2 got=%%v, want %%v\\n\", a, want)\n", s)
+		fmt.Fprintf(w, "    failed=true\n")
+		fmt.Fprintf(w, "  }\n")
+		fmt.Fprintf(w, "  b := T%du2{15, [%d]byte{", s, s)
+		for i := 0; i < s; i++ {
+			fmt.Fprintf(w, "255,")
+		}
+		fmt.Fprintf(w, "}}\n")
+		fmt.Fprintf(w, "  zero%du2_ssa(&b)\n", s)
+		fmt.Fprintf(w, "  wantb := T%du2{15, [%d]byte{", s, s)
+		for i := 0; i < s; i++ {
+			fmt.Fprintf(w, "0,")
+		}
+		fmt.Fprintf(w, "}}\n")
+		fmt.Fprintf(w, "  if b != wantb {\n")
+		fmt.Fprintf(w, "    fmt.Printf(\"zero%du2 got=%%v, want %%v\\n\", b, wantb)\n", s)
+		fmt.Fprintf(w, "    failed=true\n")
+		fmt.Fprintf(w, "  }\n")
+		fmt.Fprintf(w, "}\n")
+	}
+
 	// boilerplate at end
 	fmt.Fprintf(w, "var failed bool\n")
 	fmt.Fprintf(w, "func main() {\n")
 	for _, s := range sizes {
 		fmt.Fprintf(w, "  testZero%d()\n", s)
 	}
+	for _, s := range usizes {
+		fmt.Fprintf(w, "  testZero%du()\n", s)
+	}
 	fmt.Fprintf(w, "  if failed {\n")
 	fmt.Fprintf(w, "    panic(\"failed\")\n")
 	fmt.Fprintf(w, "  }\n")
diff --git a/src/cmd/compile/internal/gc/testdata/zero.go b/src/cmd/compile/internal/gc/testdata/zero.go
index f6354868cb..9d261aa401 100644
--- a/src/cmd/compile/internal/gc/testdata/zero.go
+++ b/src/cmd/compile/internal/gc/testdata/zero.go
@@ -505,6 +505,216 @@ func testZero1025() {
 	}
 }
 
+type T8u1 struct {
+	b   bool
+	val [8]byte
+}
+type T8u2 struct {
+	i   uint16
+	val [8]byte
+}
+
+//go:noinline
+func zero8u1_ssa(t *T8u1) {
+	t.val = [8]byte{}
+}
+
+//go:noinline
+func zero8u2_ssa(t *T8u2) {
+	t.val = [8]byte{}
+}
+func testZero8u() {
+	a := T8u1{false, [8]byte{255, 255, 255, 255, 255, 255, 255, 255}}
+	zero8u1_ssa(&a)
+	want := T8u1{false, [8]byte{0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero8u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T8u2{15, [8]byte{255, 255, 255, 255, 255, 255, 255, 255}}
+	zero8u2_ssa(&b)
+	wantb := T8u2{15, [8]byte{0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero8u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
+type T16u1 struct {
+	b   bool
+	val [16]byte
+}
+type T16u2 struct {
+	i   uint16
+	val [16]byte
+}
+
+//go:noinline
+func zero16u1_ssa(t *T16u1) {
+	t.val = [16]byte{}
+}
+
+//go:noinline
+func zero16u2_ssa(t *T16u2) {
+	t.val = [16]byte{}
+}
+func testZero16u() {
+	a := T16u1{false, [16]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero16u1_ssa(&a)
+	want := T16u1{false, [16]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero16u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T16u2{15, [16]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero16u2_ssa(&b)
+	wantb := T16u2{15, [16]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero16u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
+type T24u1 struct {
+	b   bool
+	val [24]byte
+}
+type T24u2 struct {
+	i   uint16
+	val [24]byte
+}
+
+//go:noinline
+func zero24u1_ssa(t *T24u1) {
+	t.val = [24]byte{}
+}
+
+//go:noinline
+func zero24u2_ssa(t *T24u2) {
+	t.val = [24]byte{}
+}
+func testZero24u() {
+	a := T24u1{false, [24]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero24u1_ssa(&a)
+	want := T24u1{false, [24]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero24u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T24u2{15, [24]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero24u2_ssa(&b)
+	wantb := T24u2{15, [24]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero24u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
+type T32u1 struct {
+	b   bool
+	val [32]byte
+}
+type T32u2 struct {
+	i   uint16
+	val [32]byte
+}
+
+//go:noinline
+func zero32u1_ssa(t *T32u1) {
+	t.val = [32]byte{}
+}
+
+//go:noinline
+func zero32u2_ssa(t *T32u2) {
+	t.val = [32]byte{}
+}
+func testZero32u() {
+	a := T32u1{false, [32]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero32u1_ssa(&a)
+	want := T32u1{false, [32]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero32u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T32u2{15, [32]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero32u2_ssa(&b)
+	wantb := T32u2{15, [32]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero32u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
+type T64u1 struct {
+	b   bool
+	val [64]byte
+}
+type T64u2 struct {
+	i   uint16
+	val [64]byte
+}
+
+//go:noinline
+func zero64u1_ssa(t *T64u1) {
+	t.val = [64]byte{}
+}
+
+//go:noinline
+func zero64u2_ssa(t *T64u2) {
+	t.val = [64]byte{}
+}
+func testZero64u() {
+	a := T64u1{false, [64]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero64u1_ssa(&a)
+	want := T64u1{false, [64]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero64u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T64u2{15, [64]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero64u2_ssa(&b)
+	wantb := T64u2{15, [64]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero64u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
+type T256u1 struct {
+	b   bool
+	val [256]byte
+}
+type T256u2 struct {
+	i   uint16
+	val [256]byte
+}
+
+//go:noinline
+func zero256u1_ssa(t *T256u1) {
+	t.val = [256]byte{}
+}
+
+//go:noinline
+func zero256u2_ssa(t *T256u2) {
+	t.val = [256]byte{}
+}
+func testZero256u() {
+	a := T256u1{false, [256]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero256u1_ssa(&a)
+	want := T256u1{false, [256]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if a != want {
+		fmt.Printf("zero256u2 got=%v, want %v\n", a, want)
+		failed = true
+	}
+	b := T256u2{15, [256]byte{255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255}}
+	zero256u2_ssa(&b)
+	wantb := T256u2{15, [256]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
+	if b != wantb {
+		fmt.Printf("zero256u2 got=%v, want %v\n", b, wantb)
+		failed = true
+	}
+}
+
 var failed bool
 
 func main() {
@@ -533,6 +743,12 @@ func main() {
 	testZero1023()
 	testZero1024()
 	testZero1025()
+	testZero8u()
+	testZero16u()
+	testZero24u()
+	testZero32u()
+	testZero64u()
+	testZero256u()
 	if failed {
 		panic("failed")
 	}
diff --git a/src/cmd/compile/internal/ssa/gen/PPC64.rules b/src/cmd/compile/internal/ssa/gen/PPC64.rules
index 81ac3c26af..45853b4b48 100644
--- a/src/cmd/compile/internal/ssa/gen/PPC64.rules
+++ b/src/cmd/compile/internal/ssa/gen/PPC64.rules
@@ -526,49 +526,29 @@
 	(MOVBstorezero [6] destptr
 		(MOVHstorezero [4] destptr
 			(MOVWstorezero destptr mem)))
-(Zero [8] destptr mem) ->
-	(MOVDstorezero destptr mem)
 
-// Zero small numbers of words directly.
-(Zero [12] destptr mem) ->
+// MOVD for store with DS must have offsets that are multiple of 4
+(Zero [8] {t} destptr mem) && t.(*types.Type).Alignment()%4 == 0 ->
+        (MOVDstorezero destptr mem)
+(Zero [8] destptr mem) ->
+        (MOVWstorezero [4] destptr
+                (MOVWstorezero [0] destptr mem))
+// Handle these cases only if aligned properly, otherwise use general case below
+(Zero [12] {t} destptr mem) && t.(*types.Type).Alignment()%4 == 0 ->
         (MOVWstorezero [8] destptr
                 (MOVDstorezero [0] destptr mem))
-(Zero [16] destptr mem) ->
-	(MOVDstorezero [8] destptr
+(Zero [16] {t} destptr mem) && t.(*types.Type).Alignment()%4 == 0 ->
+       (MOVDstorezero [8] destptr
                 (MOVDstorezero [0] destptr mem))
-(Zero [24] destptr mem) ->
-	(MOVDstorezero [16] destptr
-		(MOVDstorezero [8] destptr
-			(MOVDstorezero [0] destptr mem)))
-(Zero [32] destptr mem) ->
-	(MOVDstorezero [24] destptr
-		(MOVDstorezero [16] destptr
-			(MOVDstorezero [8] destptr
-				(MOVDstorezero [0] destptr mem))))
-
-(Zero [40] destptr mem) ->
-	(MOVDstorezero [32] destptr
-		(MOVDstorezero [24] destptr
-			(MOVDstorezero [16] destptr
-				(MOVDstorezero [8] destptr
-					(MOVDstorezero [0] destptr mem)))))
-
-(Zero [48] destptr mem) ->
-	(MOVDstorezero [40] destptr
-		(MOVDstorezero [32] destptr
-			(MOVDstorezero [24] destptr
-				(MOVDstorezero [16] destptr
-					(MOVDstorezero [8] destptr
-						(MOVDstorezero [0] destptr mem))))))
-
-(Zero [56] destptr mem) ->
-	(MOVDstorezero [48] destptr
-		(MOVDstorezero [40] destptr
-			(MOVDstorezero [32] destptr
-				(MOVDstorezero [24] destptr
-					(MOVDstorezero [16] destptr
-						(MOVDstorezero [8] destptr
-							(MOVDstorezero [0] destptr mem)))))))
+(Zero [24] {t} destptr mem) && t.(*types.Type).Alignment()%4 == 0 ->
+       (MOVDstorezero [16] destptr
+               (MOVDstorezero [8] destptr
+                       (MOVDstorezero [0] destptr mem)))
+(Zero [32] {t} destptr mem) && t.(*types.Type).Alignment()%4 == 0 ->
+       (MOVDstorezero [24] destptr
+               (MOVDstorezero [16] destptr
+                       (MOVDstorezero [8] destptr
+                               (MOVDstorezero [0] destptr mem))))
 
 // Handle cases not handled above
 (Zero [s] ptr mem) -> (LoweredZero [s] ptr mem)
diff --git a/src/cmd/compile/internal/ssa/rewritePPC64.go b/src/cmd/compile/internal/ssa/rewritePPC64.go
index 20e354cb4a..75bf763d12 100644
--- a/src/cmd/compile/internal/ssa/rewritePPC64.go
+++ b/src/cmd/compile/internal/ssa/rewritePPC64.go
@@ -10727,35 +10727,39 @@ func rewriteValuePPC64_OpZero_0(v *Value) bool {
 		v.AddArg(v0)
 		return true
 	}
-	// match: (Zero [8] destptr mem)
-	// cond:
+	// match: (Zero [8] {t} destptr mem)
+	// cond: t.(*types.Type).Alignment()%4 == 0
 	// result: (MOVDstorezero destptr mem)
 	for {
 		if v.AuxInt != 8 {
 			break
 		}
+		t := v.Aux
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
+		if !(t.(*types.Type).Alignment()%4 == 0) {
+			break
+		}
 		v.reset(OpPPC64MOVDstorezero)
 		v.AddArg(destptr)
 		v.AddArg(mem)
 		return true
 	}
-	// match: (Zero [12] destptr mem)
+	// match: (Zero [8] destptr mem)
 	// cond:
-	// result: (MOVWstorezero [8] destptr                 (MOVDstorezero [0] destptr mem))
+	// result: (MOVWstorezero [4] destptr                 (MOVWstorezero [0] destptr mem))
 	for {
-		if v.AuxInt != 12 {
+		if v.AuxInt != 8 {
 			break
 		}
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
 		v.reset(OpPPC64MOVWstorezero)
-		v.AuxInt = 8
+		v.AuxInt = 4
 		v.AddArg(destptr)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVWstorezero, types.TypeMem)
 		v0.AuxInt = 0
 		v0.AddArg(destptr)
 		v0.AddArg(mem)
@@ -10767,17 +10771,21 @@ func rewriteValuePPC64_OpZero_0(v *Value) bool {
 func rewriteValuePPC64_OpZero_10(v *Value) bool {
 	b := v.Block
 	_ = b
-	// match: (Zero [16] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [8] destptr                 (MOVDstorezero [0] destptr mem))
+	// match: (Zero [12] {t} destptr mem)
+	// cond: t.(*types.Type).Alignment()%4 == 0
+	// result: (MOVWstorezero [8] destptr                 (MOVDstorezero [0] destptr mem))
 	for {
-		if v.AuxInt != 16 {
+		if v.AuxInt != 12 {
 			break
 		}
+		t := v.Aux
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		v.reset(OpPPC64MOVDstorezero)
+		if !(t.(*types.Type).Alignment()%4 == 0) {
+			break
+		}
+		v.reset(OpPPC64MOVWstorezero)
 		v.AuxInt = 8
 		v.AddArg(destptr)
 		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
@@ -10787,161 +10795,85 @@ func rewriteValuePPC64_OpZero_10(v *Value) bool {
 		v.AddArg(v0)
 		return true
 	}
-	// match: (Zero [24] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [16] destptr 		(MOVDstorezero [8] destptr 			(MOVDstorezero [0] destptr mem)))
+	// match: (Zero [16] {t} destptr mem)
+	// cond: t.(*types.Type).Alignment()%4 == 0
+	// result: (MOVDstorezero [8] destptr                 (MOVDstorezero [0] destptr mem))
 	for {
-		if v.AuxInt != 24 {
+		if v.AuxInt != 16 {
 			break
 		}
+		t := v.Aux
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		v.reset(OpPPC64MOVDstorezero)
-		v.AuxInt = 16
-		v.AddArg(destptr)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v0.AuxInt = 8
-		v0.AddArg(destptr)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v1.AuxInt = 0
-		v1.AddArg(destptr)
-		v1.AddArg(mem)
-		v0.AddArg(v1)
-		v.AddArg(v0)
-		return true
-	}
-	// match: (Zero [32] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [24] destptr 		(MOVDstorezero [16] destptr 			(MOVDstorezero [8] destptr 				(MOVDstorezero [0] destptr mem))))
-	for {
-		if v.AuxInt != 32 {
+		if !(t.(*types.Type).Alignment()%4 == 0) {
 			break
 		}
-		_ = v.Args[1]
-		destptr := v.Args[0]
-		mem := v.Args[1]
 		v.reset(OpPPC64MOVDstorezero)
-		v.AuxInt = 24
+		v.AuxInt = 8
 		v.AddArg(destptr)
 		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v0.AuxInt = 16
+		v0.AuxInt = 0
 		v0.AddArg(destptr)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v1.AuxInt = 8
-		v1.AddArg(destptr)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v2.AuxInt = 0
-		v2.AddArg(destptr)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v0.AddArg(v1)
+		v0.AddArg(mem)
 		v.AddArg(v0)
 		return true
 	}
-	// match: (Zero [40] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [32] destptr 		(MOVDstorezero [24] destptr 			(MOVDstorezero [16] destptr 				(MOVDstorezero [8] destptr 					(MOVDstorezero [0] destptr mem)))))
+	// match: (Zero [24] {t} destptr mem)
+	// cond: t.(*types.Type).Alignment()%4 == 0
+	// result: (MOVDstorezero [16] destptr                (MOVDstorezero [8] destptr                        (MOVDstorezero [0] destptr mem)))
 	for {
-		if v.AuxInt != 40 {
+		if v.AuxInt != 24 {
 			break
 		}
+		t := v.Aux
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		v.reset(OpPPC64MOVDstorezero)
-		v.AuxInt = 32
-		v.AddArg(destptr)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v0.AuxInt = 24
-		v0.AddArg(destptr)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v1.AuxInt = 16
-		v1.AddArg(destptr)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v2.AuxInt = 8
-		v2.AddArg(destptr)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v3.AuxInt = 0
-		v3.AddArg(destptr)
-		v3.AddArg(mem)
-		v2.AddArg(v3)
-		v1.AddArg(v2)
-		v0.AddArg(v1)
-		v.AddArg(v0)
-		return true
-	}
-	// match: (Zero [48] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [40] destptr 		(MOVDstorezero [32] destptr 			(MOVDstorezero [24] destptr 				(MOVDstorezero [16] destptr 					(MOVDstorezero [8] destptr 						(MOVDstorezero [0] destptr mem))))))
-	for {
-		if v.AuxInt != 48 {
+		if !(t.(*types.Type).Alignment()%4 == 0) {
 			break
 		}
-		_ = v.Args[1]
-		destptr := v.Args[0]
-		mem := v.Args[1]
 		v.reset(OpPPC64MOVDstorezero)
-		v.AuxInt = 40
+		v.AuxInt = 16
 		v.AddArg(destptr)
 		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v0.AuxInt = 32
+		v0.AuxInt = 8
 		v0.AddArg(destptr)
 		v1 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v1.AuxInt = 24
+		v1.AuxInt = 0
 		v1.AddArg(destptr)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v2.AuxInt = 16
-		v2.AddArg(destptr)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v3.AuxInt = 8
-		v3.AddArg(destptr)
-		v4 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v4.AuxInt = 0
-		v4.AddArg(destptr)
-		v4.AddArg(mem)
-		v3.AddArg(v4)
-		v2.AddArg(v3)
-		v1.AddArg(v2)
+		v1.AddArg(mem)
 		v0.AddArg(v1)
 		v.AddArg(v0)
 		return true
 	}
-	// match: (Zero [56] destptr mem)
-	// cond:
-	// result: (MOVDstorezero [48] destptr 		(MOVDstorezero [40] destptr 			(MOVDstorezero [32] destptr 				(MOVDstorezero [24] destptr 					(MOVDstorezero [16] destptr 						(MOVDstorezero [8] destptr 							(MOVDstorezero [0] destptr mem)))))))
+	// match: (Zero [32] {t} destptr mem)
+	// cond: t.(*types.Type).Alignment()%4 == 0
+	// result: (MOVDstorezero [24] destptr                (MOVDstorezero [16] destptr                        (MOVDstorezero [8] destptr                                (MOVDstorezero [0] destptr mem))))
 	for {
-		if v.AuxInt != 56 {
+		if v.AuxInt != 32 {
 			break
 		}
+		t := v.Aux
 		_ = v.Args[1]
 		destptr := v.Args[0]
 		mem := v.Args[1]
+		if !(t.(*types.Type).Alignment()%4 == 0) {
+			break
+		}
 		v.reset(OpPPC64MOVDstorezero)
-		v.AuxInt = 48
+		v.AuxInt = 24
 		v.AddArg(destptr)
 		v0 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v0.AuxInt = 40
+		v0.AuxInt = 16
 		v0.AddArg(destptr)
 		v1 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v1.AuxInt = 32
+		v1.AuxInt = 8
 		v1.AddArg(destptr)
 		v2 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v2.AuxInt = 24
+		v2.AuxInt = 0
 		v2.AddArg(destptr)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v3.AuxInt = 16
-		v3.AddArg(destptr)
-		v4 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v4.AuxInt = 8
-		v4.AddArg(destptr)
-		v5 := b.NewValue0(v.Pos, OpPPC64MOVDstorezero, types.TypeMem)
-		v5.AuxInt = 0
-		v5.AddArg(destptr)
-		v5.AddArg(mem)
-		v4.AddArg(v5)
-		v3.AddArg(v4)
-		v2.AddArg(v3)
+		v2.AddArg(mem)
 		v1.AddArg(v2)
 		v0.AddArg(v1)
 		v.AddArg(v0)