[PATCH] D78150: [AMDGPU] Use SDWA for 16 bit subreg copy

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec created this revision.
rampitec added a reviewer: arsenm.
Herald added subscribers: kerbowa, hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely, kzhuravl.
rampitec planned changes to this revision.
rampitec added a comment.

I have realized we still need to model it to protect from post-RA scheduler. I will add imp-use.


This simplifies the logic and allows to use it on GFX8.
The showstopper was a need for tied operand with unused
preserve, but I believe we do not really need it when we
already have physical registers.


https://reviews.llvm.org/D78150

Files:
  llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
  llvm/test/CodeGen/AMDGPU/lo16-hi16-physreg-copy.mir

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D78150.257479.patch
Type: text/x-patch
Size: 10624 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200414/cc700016/attachment.bin>