cmd/compile: use STP/LDP around morestack on arm64

The spill/restore code around morestack is almost never exectued, so
we should make it as small as possible. Using 2-register loads/stores
makes sense here. Also, the offsets from SP are pretty small so the
offset almost always fits in the (smaller than a normal load/store)
offset field of the instruction.

Makes cmd/go 0.6% smaller.

Change-Id: I8845283c1b269a259498153924428f6173bda293
Reviewed-on: https://go-review.googlesource.com/c/go/+/621556
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

1 files changed, 31 insertions, 0 deletions

diff --git a/test/codegen/spills.go b/test/codegen/spills.go
new file mode 100644
index 0000000000..c8ac9859a4
--- /dev/null
+++ b/test/codegen/spills.go
@@ -0,0 +1,31 @@
+// asmcheck
+
+// Copyright 2024 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 codegen
+
+func i64(a, b int64) int64 { // arm64:`STP\s`,`LDP\s`
+	g()
+	return a + b
+}
+
+func i32(a, b int32) int32 { // arm64:`STPW`,`LDPW`
+	g()
+	return a + b
+}
+
+func f64(a, b float64) float64 { // arm64:`FSTPD`,`FLDPD`
+	g()
+	return a + b
+}
+
+func f32(a, b float32) float32 { // arm64:`FSTPS`,`FLDPS`
+	g()
+	return a + b
+}
+
+//go:noinline
+func g() {
+}