[PATCH] [ARM] Look through concat when lowering in-place shuffles (VZIP, ..)

[Ahmed Bougacha]

Currently, we canonicalize shuffles that produce a result larger than
their operands with:
  shuffle(concat(v1, undef), concat(v2, undef))
->
  shuffle(concat(v1, v2), undef)
because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).

This is useful in the general case, but there are special cases where
native shuffles produce larger results: the two-result ops.

Look through the concat when lowering them:
  shuffle(concat(v1, v2), undef)
->
  concat(VZIP(v1, v2):0, :1)

This lets us generate the native shuffles instead of scalarizing to
dozens of VMOVs.

I'm a little worried about the disparity between the lowering and
isShuffleMaskLegal, but with the current API we have no way of looking
at the actual operands, and this isn't a problem in practice because
the ARM combine runs last.

The obvious alternative would be to stop doing the combine, but I
think it's useful.  We can also avoid doing it for these masks, but
we'll still need to look through concat(v, undef) to avoid
generating needlessly-wide shuffles.

http://reviews.llvm.org/D10424

Files:
  lib/Target/ARM/ARMISelLowering.cpp
  test/CodeGen/ARM/vtrn.ll
  test/CodeGen/ARM/vuzp.ll
  test/CodeGen/ARM/vzip.ll

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D10424.27613.patch
Type: text/x-patch
Size: 17196 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150612/2b535c2a/attachment.bin>