[PATCH] [CodeGen] Combine small-element shuffles of scalar_to_vector in terms of the wider scalar.

[Ahmed Bougacha]

Try combining to BUILD_VECTOR instead, adding to the 2-input scalar shuffle combine.

I didn't look at two bitcasts at the same time because this only fires because of http://reviews.llvm.org/D8884:  otherwise, the vector_shuffle combine is way too early (and the operands are legalized to BUILD_VECTOR/SCALAR_TO_VECTOR right after), and either the operands get lowered between combine rounds, or other combines catch this without needing to peak through bitcasts (usually concat_vectors if you have two vector inputs).

This does change one ARM test, where we replace:

vmov	r0, r1, d16
	vmov.32	d16[0], r0
	vmov.32	d16[1], r1
	vext.8	q8, q8, q8, #4

with:

vmov	r0, r1, d16
	vmov.32	d16[0], r1

Which looks ..interesting, but is an improvement nonetheless.

-Ahmed


http://reviews.llvm.org/D8885

Files:
  lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  test/CodeGen/AArch64/concat_vectors-combines.ll
  test/CodeGen/ARM/vector-DAGCombine.ll

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D8885.23457.patch
Type: text/x-patch
Size: 5816 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150409/13f80851/attachment.bin>