[AMDGPU][GlobalISel]selectImpl() may create illegal copy from VGPR to SGPR

[chenzheng1030]

This is found during investigation of issue #61468

After importing more td patterns to global MatchTable, for example the patterns in https://reviews.llvm.org/D147780, selectImpl() will create illegal copy.

For now in https://reviews.llvm.org/D147780, this issue is workaround by not using selectImpl() to do the selection. But maybe a fix in selectImpl() should be made to avoid such issue.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

[nickleus27]

I am interested in this issue. Can someone assign to me.

[arsenm]

There's been a number of changes to boolean handling since this was filed. cc @petar-avramovic

[petar-avramovic]

Is there a .ll test for this?
This probably has to be fixed before instruction-select, maybe something like: #68858
If not and we want to select it intentionally as vgpr to sgpr copy we need to make use of uniformity info analysis first in #73684

[chenzheng1030]

Sorry for the delay. Totally lost track for this issue.

To reproduce the bug, try following change and run ninja check-llvm:

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index dcb0f47973c4..9be4b6744cc1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -3609,8 +3609,9 @@ bool AMDGPUInstructionSelector::select(MachineInstr &I) {
     // This is a workaround. For extension from type i1, `selectImpl()` uses
     // patterns from TD file and generates an illegal VGPR to SGPR COPY as type
     // i1 can only be hold in a SGPR class.
-    if (MRI->getType(I.getOperand(1).getReg()) != LLT::scalar(1) &&
-        selectImpl(I, *CoverageInfo))
+    //if (MRI->getType(I.getOperand(1).getReg()) != LLT::scalar(1) &&
+    //    selectImpl(I, *CoverageInfo))
+    if (selectImpl(I, *CoverageInfo))
       return true;
     return selectG_SZA_EXT(I);
   case TargetOpcode::G_FPEXT:

ninja check-llvm -j128

Failed Tests (9):
  LLVM :: CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll
  LLVM :: CodeGen/AMDGPU/GlobalISel/divergence-temporal-divergent-i1.ll
  LLVM :: CodeGen/AMDGPU/GlobalISel/inst-select-anyext.mir
  LLVM :: CodeGen/AMDGPU/GlobalISel/inst-select-freeze.mir
  LLVM :: CodeGen/AMDGPU/GlobalISel/inst-select-icmp.s64.mir
  LLVM :: CodeGen/AMDGPU/GlobalISel/inst-select-sext.mir
  LLVM :: CodeGen/AMDGPU/GlobalISel/inst-select-zext.mir
  LLVM :: CodeGen/AMDGPU/div_i128.ll
  LLVM :: CodeGen/AMDGPU/div_v2i128.ll

an error:

FileCheck -check-prefix=GFX10 llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-temporal-divergent-i1.ll
error: <unknown>:0:0: in function temporal_divergent_i1_phi void (float, ptr): illegal VGPR to SGPR copy

error: <unknown>:0:0: in function temporal_divergent_i1_phi void (float, ptr): illegal VGPR to SGPR copy

error: <unknown>:0:0: in function temporal_divergent_i1_non_phi void (float, ptr): illegal VGPR to SGPR copy

error: <unknown>:0:0: in function temporal_divergent_i1_non_phi void (float, ptr): illegal VGPR to SGPR copy

[nickleus27]

Thank you @chenzheng1030 ! I have been really busy lately, but when I have a chance I would like to circle back to this issue if no one else has picked it up.