cmd/compile: fold offsets into memory ops

[randall77]

func f(x, y int) bool {
    return x < y
}

generates the assembly:

	0x0000 00000 (tmp5.go:3)	MOVQ	"".x+8(SP), AX
	0x0005 00005 (tmp5.go:3)	MOVQ	"".y+16(SP), CX
	0x000a 00010 (tmp5.go:4)	CMPQ	AX, CX
	0x000d 00013 (tmp5.go:4)	LEAQ	"".~r2+24(SP), AX
	0x0012 00018 (tmp5.go:4)	SETLT	(AX)

We should fold the LEAQ into the SETLT. We do so in many other cases, the SETXX cases were missed. We should audit to make sure that all ops for which we can fold address computations into, we do so.

[randall77]

Need to do (for amd64):

{ADD,SUB,MUL}[SD]mem
ADD[LQ]constmem
{ADD,SUB,AND,OR,XOR}[LQ]mem
SET{L,LE,G,GE,EQ,NE,A,AE,B,BE}mem

For each, fold an LEAQ or an ADDQ in the pointer slot into the instruction itself. Example:

(SETLmem [off1] {sym1} (LEAQ [off2] {sym2} ptr) x mem) && is32Bit(off1+off2) && canMergeSym(sym1, sym2) ->
	(SETLmem  [off1+off2] {mergeSym(sym1,sym2)} ptr x mem)
(SETLmem  [off1] {sym} (ADDQconst [off2] ptr) x mem) && is32Bit(off1+off2) -> (SETLmem  [off1+off2] {sym} ptr x mem)

[gopherbot]

Change https://golang.org/cl/86055 mentions this issue: cmd/compile: fold LEAQ and ADDQ into memory ops

[rasky]

SET*mem was done in 0cacc4d.

[gopherbot]

Change https://golang.org/cl/98035 mentions this issue: cmd/compile: fold offsets into memory ops