cmd/compile: suboptimal bits.Rotate* on arm64

[greatroar]

What version of Go are you using (go version)?

$ go version
go version go1.17.1 linux/amd64

Does this issue reproduce with the latest release?

Reproduces with

$ gotip version
go version devel go1.18-771b8ea4f4c Sun Sep 19 02:43:09 2021 +0000 linux/amd64

What operating system and processor architecture are you using (go env)?

GOARCH=arm64

What did you do?

func rotxor1(x, y uint64) uint64 {
	y = (y << 7) | (y >> 57)
	return x ^ y
}

func rotxor2(x, y uint64) uint64 {
	return x ^ bits.RotateLeft64(y, 7)
}

What did you expect to see?

Would have been nice to see a fused rotate-xor (#48002), but at least the generated asm should be the same.

What did you see instead?

For rotxor1:

MOVD	"".y+8(FP), R0
ROR	$57, R0, R0
MOVD	"".x(FP), R1
EOR	R0, R1, R0

For rotxor2:

MOVD	"".y+8(FP), R0
MOVD	$-7, R1
ROR	R1, R0, R0
MOVD	"".x(FP), R1
EOR	R0, R1, R0

The compiler inserts a mov and allocates a register for the shift constant when the intrinsic is used. This is a regression introduced in Go 1.12. It's not limited to these tiny functions: I spotted this when inspecting the generated asm for SipHash.

[gopherbot]

Change https://golang.org/cl/350909 mentions this issue: cmd/compile: implement constant rotates on arm64