[PATCH] Canonicalize splats as build_vectors (PR22283)

[Sanjay Patel]

Hi rengolin, andreadb, mkuper, RKSimon,

This is a minimal follow-on patch to:
http://reviews.llvm.org/D7093

That patch canonicalized constant splats as build_vectors, and this patch just removes the constant check so we can canonicalize all splats as build_vectors.

This fixes the 2nd test case in PR22283 ( http://llvm.org/bugs/show_bug.cgi?id=22283 ).

The unfortunate code duplication between SelectionDAG and DAGCombiner is discussed in the earlier patch. At least this patch is just removing code...

This improves an existing x86 AVX test and changes codegen in an ARM test. I don't know enough about ARM implementations to know if the new codegen is an improvement or not. 

For the changed ARM tests, we currently generate something like this:
  ldrsh	r12, [sp]   <--- load scalar
  vmov	d17, r2, r3
  vmov	d16, r0, r1
  vmov	s0, r12
  vcvt.f32.s32	s0, s0
  vdup.32	q9, d0[0]     <--- splat register value
  vsub.f32	q8, q9, q8
  vmov	r0, r1, d16
  vmov	r2, r3, d17
  mov	pc, lr

With this patch, we'll generate a 'vld1' to splat instead:
  mov	r12, sp
  vmov	d19, r2, r3
  vld1.16	{d16[]}, [r12:16] <--- load and splat
  vmov	d18, r0, r1
  vmovl.s16	q8, d16
  vcvt.f32.s32	q8, q8
  vsub.f32	q8, q8, q9
  vmov	r0, r1, d16
  vmov	r2, r3, d17
  mov	pc, lr

http://reviews.llvm.org/D7389

Files:
  lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  lib/CodeGen/SelectionDAG/SelectionDAG.cpp
  test/CodeGen/ARM/vdup.ll
  test/CodeGen/X86/vector-shuffle-256-v4.ll

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D7389.19261.patch
Type: text/x-patch
Size: 4904 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150203/2c4234bf/attachment.bin>