[PATCH] D38316: [InstCombine] replace bitcast to scalar + insertelement with widening shuffle + vector bitcast

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In https://reviews.llvm.org/D38316#882443, @efriedma wrote:

> > Right - I thought about that case, but I couldn't justify it if we need another IR instruction.
>
> So how do you plan to fix x86 lowering for this case?  If we need a DAGCombine for this anyway, not sure there's much point to handling it in instcombine.


The only thing missing is a load fold. Ie, we'll produce this asm for 32-bit with AVX512f (the motivating target in the bug report) with this patch:

  vmovsd	4(%esp), %xmm0          # xmm0 = mem[0],zero
  vbroadcastsd	%xmm0, %zmm0

I think the ideal code would fuse those into splat of memop:

  vbroadcastsd 4(%esp), %zmm0


https://reviews.llvm.org/D38316