[PATCH] D69211: AMDGPU: Enable shouldConsiderGEPOffsetSplit

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, vpykhtin.
Herald added subscribers: kristof.beyls, t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely, kzhuravl.

The results in the new test (copied from the AArch64 test for the
hook) are slightly worse, due to now using a scalar phi for the loop
bounds add which requires copying to VGPRs for the pointer operand.

     

schedule-regpressure-limit2 is somewhat degenerate looking, but this
gives a much better result (the function ends up about half the size
as before). However it is no longer matching the tested register
limits, and uses a lot more VGPRs in the minreg case.


https://reviews.llvm.org/D69211

Files:
  lib/Target/AMDGPU/AMDGPUISelLowering.cpp
  lib/Target/AMDGPU/AMDGPUISelLowering.h
  test/CodeGen/AMDGPU/large-offset-gep.ll
  test/CodeGen/AMDGPU/schedule-regpressure-limit2.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D69211.225730.patch
Type: text/x-patch
Size: 11081 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20191019/72bedfb0/attachment.bin>