[InstCombine] hoist vector cast above splat as scalar cast?

[rotateright]

Bugzilla Link	32865
Version	trunk
OS	All
CC	@RKSimon,@ZviRackover

Extended Description

Forking this off from bug 14657.

One of these should be canonical:

define <8 x i32> @​sext_last(i16 %x) {
%ins = insertelement <8 x i16> undef, i16 %x, i32 0
%splat = shufflevector <8 x i16> %ins, <8 x i16> undef, <8 x i32> zeroinitializer
%sext = sext <8 x i16> %splat to <8 x i32>
ret <8 x i32> %sext
}

define <8 x i32> @​sext_first(i16 %x) {
%sext = sext i16 %x to i32
%ins = insertelement <8 x i32> undef, i32 %sext, i32 0
%splat = shufflevector <8 x i32> %ins, <8 x i32> undef, <8 x i32> zeroinitializer
ret <8 x i32> %splat
}

We may need to prepare the backend to expect the canonical pattern to avoid regressions. Eg, with AVX1, there's a big difference in codegen:

$ ./llc -o - splatsext.ll -mattr=avx
_sext_last: ## @​sext_last
vmovd %edi, %xmm0
vpshuflw $0, %xmm0, %xmm0 ## xmm0 = xmm0[0,0,0,0,4,5,6,7]
vpshufd $80, %xmm0, %xmm0 ## xmm0 = xmm0[0,0,1,1]
vpmovsxwd %xmm0, %xmm1
vpshufd $78, %xmm0, %xmm0 ## xmm0 = xmm0[2,3,0,1]
vpmovsxwd %xmm0, %xmm0
vinsertf128 $1, %xmm0, %ymm1, %ymm0

_sext_first: ## @​sext_first
movswl %di, %eax
vmovd %eax, %xmm0
vpshufd $0, %xmm0, %xmm0 ## xmm0 = xmm0[0,0,0,0]
vinsertf128 $1, %xmm0, %ymm0, %ymm0

[rotateright]

The AVX1 codegen for 'sext_last' has improved (might even be optimal?):
vmovd %edi, %xmm0
vpshuflw $224, %xmm0, %xmm0 ## xmm0 = xmm0[0,0,2,3,4,5,6,7]
vpshufd $0, %xmm0, %xmm0 ## xmm0 = xmm0[0,0,0,0]
vpmovsxwd %xmm0, %xmm0
vinsertf128 $1, %xmm0, %ymm0, %ymm0

---

But looking at this again, I think we should canonicalize IR to 'sext_first' because we should always favor a scalar op to a vector op for its analysis/transform potential.

In this case, the codegen also looks better, but that should be confirmed with other casts (zext/trunc//fpext/fptrunc) and other targets.