cmd/compile: storing strings doesn't take advantage of paired stores on arm64

[randall77]

func f(p *string, s string) {
	*p = s
}

Compiles to:

	0x001c 00028 	MOVD	R2, 8(R0) // write length
	0x0020 00032 	MOVWU	runtime.writeBarrier(SB), R2
	0x0028 00040 	CBZW	R2, 56
	0x002c 00044 	MOVD	(R0), R2 // load old pointer
	0x0030 00048 	CALL	runtime.gcWriteBarrier2(SB)
	0x0034 00052 	STP	(R1, R2), (R25) // write new ptr & old ptr to write barrier buffer
	0x0038 00056 	MOVD	R1, (R0) // write new ptr

It would be nice if we wrote the length and ptr together at the end (effectively merging the first and last instructions).

The length and ptr writes get separated because the write barrier pass picks the smallest set of writes to include in the barrier. Which in this case, is just the one pointer write. Maybe we could include one more non-ptr write in the barrier group if we know it could be merged with a pointer write.

[gopherbot]

Change https://go.dev/cl/675715 mentions this issue: cmd/compile: pair stores across write barrier boundaries on arm64