[PATCH] D49262: [DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

@arsenm @uweigand Any commments?



================
Comment at: test/CodeGen/X86/oddshuffles.ll:366
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
 ; SSE2-NEXT:    movd %xmm1, 24(%rdi)
----------------
efriedma wrote:
> It looks like the total instructions is increasing here?  Maybe an issue with x86 shuffle lowering?
Pre-combine we've decreased the the number of shuffles, meaning that it now falls below the threshold for permitting domain swaps to use shufps - on older SSE2 machines we're better off avoiding the domain swap. 


Repository:
  rL LLVM

https://reviews.llvm.org/D49262