[all-commits] [llvm/llvm-project] 18bbd9: GlobalISel: Artifact combine unmerge of unmerge

[Matt Arsenault via All-commits]

  Branch: refs/heads/master
  Home:   https://github.com/llvm/llvm-project
  Commit: 18bbd9f15eb031c5c7e58ebe0692f87fa8d5954f
      https://github.com/llvm/llvm-project/commit/18bbd9f15eb031c5c7e58ebe0692f87fa8d5954f
  Author: Matt Arsenault <Matthew.Arsenault at amd.com>
  Date:   2020-09-01 (Tue, 01 Sep 2020)

  Changed paths:
    M llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h
    M llvm/lib/Target/Mips/MipsRegisterBankInfo.cpp
    M llvm/test/CodeGen/AMDGPU/GlobalISel/artifact-combiner-unmerge-values.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-and.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ashr.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-bitcast.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-concat-vectors.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fabs.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fadd.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fcanonicalize.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fcos.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fdiv.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ffloor.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fma.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fmaxnum.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fmul.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fneg.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fpext.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-freeze.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fsin.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fsqrt.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fsub.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-implicit-def.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-llvm.amdgcn.image.load.2d.d16.ll
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-llvm.amdgcn.image.store.2d.d16.ll
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-lshr.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-or.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-phi.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-saddo.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-saddsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-select.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-sext-inreg.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-shl.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-smax.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-smin.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-sshlsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ssubo.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ssubsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-store-global.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-uaddo.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-uaddsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-umax.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-umin.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-unmerge-values.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ushlsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-usubo.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-usubsat.mir
    M llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-xor.mir

  Log Message:
  -----------
  GlobalISel: Artifact combine unmerge of unmerge

Unmerges have the same fundamental problem as G_TRUNC, and G_TRUNC
could be implemented in terms of G_UNMERGE_VALUES. Reducing the number
of elements in unmerge results ends up producing the original unmerge
type profile, so the artifact combiner needs to eliminate the
intermediate illegal registers. This avoids infinite looping in the
legalizer in a future change.

Assuming an unmerge has each result unmerged the same way, this ends
up producing a new unmerge of the source for every definition. I'm not
sure if the artifact combiner should either insert temporary merges
here and erase the original merge, or if the combiner should look at
uses from defs rather than defs from uses for unmerges.

In a few cases this regresses from using 16-bit shifts for 8-bit
values to using 32-bit shifts, but I think these can be legalized
later (the other legalization rules don't try very hard to use 16-bit
shifts either).