[PATCH] D13740: Catch combine opportunities for redundant imuls

[Zia Ansari via llvm-commits]

zansari updated this revision to Diff 37820.
zansari added a comment.

New patch includes changes to use "isConstOrConstSplat" to consolidate the 2 constant checks, as suggested.

Also, as suggested, I added a test to make sure we catch the vector multiply by constant case. I verified that before this patch we generate 4 pmulduqs, compared to just 2 now (i.e. we eliminate an extra vector multiply after the patch).

Thanks,
Zia.

Before:

  movdqa  .LCPI1_0, %xmm1         # xmm1 = [11,11,11,11]
  paddd   %xmm0, %xmm1
  movdqa  .LCPI1_1, %xmm2         # xmm2 = [22,22,22,22]
  pshufd  $245, %xmm1, %xmm3      # xmm3 = xmm1[1,1,3,3]
  movdqa  %xmm1, x
  pmuludq %xmm2, %xmm1
  pshufd  $232, %xmm1, %xmm1      # xmm1 = xmm1[0,2,2,3]
  pmuludq %xmm2, %xmm3
  pshufd  $232, %xmm3, %xmm3      # xmm3 = xmm3[0,2,2,3]
  punpckldq       %xmm3, %xmm1    # xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
  movdqa  %xmm1, v2
  pshufd  $245, %xmm0, %xmm1      # xmm1 = xmm0[1,1,3,3]
  pmuludq %xmm2, %xmm0
  pshufd  $232, %xmm0, %xmm0      # xmm0 = xmm0[0,2,2,3]
  pmuludq %xmm2, %xmm1
  pshufd  $232, %xmm1, %xmm1      # xmm1 = xmm1[0,2,2,3]
  punpckldq       %xmm1, %xmm0    # xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
  paddd   .LCPI1_2, %xmm0
  movdqa  %xmm0, v3
  retl

After:

  movdqa  .LCPI1_0, %xmm1         # xmm1 = [11,11,11,11]
  paddd   %xmm0, %xmm1
  movdqa  .LCPI1_1, %xmm2         # xmm2 = [22,22,22,22]
  pshufd  $245, %xmm0, %xmm3      # xmm3 = xmm0[1,1,3,3]
  pmuludq %xmm2, %xmm0
  pshufd  $232, %xmm0, %xmm0      # xmm0 = xmm0[0,2,2,3]
  pmuludq %xmm2, %xmm3
  pshufd  $232, %xmm3, %xmm2      # xmm2 = xmm3[0,2,2,3]
  punpckldq       %xmm2, %xmm0    # xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
  movdqa  .LCPI1_2, %xmm2         # xmm2 = [242,242,242,242]
  paddd   %xmm0, %xmm2
  movdqa  %xmm2, v2
  paddd   .LCPI1_3, %xmm0
  movdqa  %xmm0, v3
  movdqa  %xmm1, x
  retl


http://reviews.llvm.org/D13740

Files:
  lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  test/CodeGen/X86/combine-multiplies.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D13740.37820.patch
Type: text/x-patch
Size: 10025 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20151020/50c00b8e/attachment.bin>