[AMDGPU] Preserve dom-tree analysis in atomic optimizer.

[pravinjagtap]

Summary

AMDGPUAtomicOptimizer updates the dominator tree whenever
it modified the control flow. Therefore preserving the
analysis similar to legacy PM.

[yassingh]

llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp

150–152

Nit: Can reverse the if condition to be more consistent with other passes.

[pravinjagtap]

Addressed review comment.

[arsenm]

Why doesn’t this need an llc-pipeline or opt-pipeline update? Do we not have new PM pass testing?

[pravinjagtap]

In D153349#4437207, @arsenm wrote:

Why doesn’t this need an llc-pipeline or opt-pipeline update? Do we not have new PM pass testing?

Atomic optimizer pass is not configured with opt and llc uses legacy PM which already preserve this analysis. Therefore opt/llc pipeline doesn't need update. I think, There is no test for NewPM with pre-codegen (IR) passes.

[arsenm]

In D153349#4437262, @pravinjagtap wrote:

In D153349#4437207, @arsenm wrote:

Why doesn’t this need an llc-pipeline or opt-pipeline update? Do we not have new PM pass testing?

Atomic optimizer pass is not configured with opt and llc uses legacy PM which already preserve this analysis. Therefore opt/llc pipeline doesn't need update. I think, There is no test for NewPM with pre-codegen (IR) passes.

We should have one. I thought we had an opt pipeline test but I don't see one, other targets do

[pravinjagtap]

In D153349#4437321, @arsenm wrote:

In D153349#4437262, @pravinjagtap wrote:

In D153349#4437207, @arsenm wrote:

Why doesn’t this need an llc-pipeline or opt-pipeline update? Do we not have new PM pass testing?

Atomic optimizer pass is not configured with opt and llc uses legacy PM which already preserve this analysis. Therefore opt/llc pipeline doesn't need update. I think, There is no test for NewPM with pre-codegen (IR) passes.

We should have one. I thought we had an opt pipeline test but I don't see one, other targets do

I am not finding opt pipeline test for other targets. Can you please link it here. Interestingly, X86 and loongarch have filenames as opt-pipeline but these tests are invoking llc and not opt

[foad]

In D153349#4437321, @arsenm wrote:

In D153349#4437262, @pravinjagtap wrote:

In D153349#4437207, @arsenm wrote:

Why doesn’t this need an llc-pipeline or opt-pipeline update? Do we not have new PM pass testing?

Atomic optimizer pass is not configured with opt and llc uses legacy PM which already preserve this analysis. Therefore opt/llc pipeline doesn't need update. I think, There is no test for NewPM with pre-codegen (IR) passes.

We should have one. I thought we had an opt pipeline test but I don't see one, other targets do

See comments on https://github.com/llvm/llvm-project/commit/6f73bd781305266a747055875ce8352e5a36c809

[arsenm]

In D153349#4437431, @foad wrote:

See comments on https://github.com/llvm/llvm-project/commit/6f73bd781305266a747055875ce8352e5a36c809

So it was just delete and never replaced, as suggested should follow through with that. I also don't see why we wouldn't just duplicate the test per target

llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp

PreservedAnalyses AMDGPUAtomicOptimizerPass::run(Function &F,
   const DataLayout *DL = &F.getParent()->getDataLayout();
 
   DomTreeUpdater DTU(&AM.getResult<DominatorTreeAnalysis>(F),
                      DomTreeUpdater::UpdateStrategy::Lazy);
   const GCNSubtarget *ST = &TM.getSubtarget<GCNSubtarget>(F);
 
   bool IsPixelShader = F.getCallingConv() == CallingConv::AMDGPU_PS;
 
-  return AMDGPUAtomicOptimizerImpl(UA, DL, DTU, ST, IsPixelShader, ScanImpl)
-                 .run(F)
-             ? PreservedAnalyses::none()
-             : PreservedAnalyses::all();
+  if (AMDGPUAtomicOptimizerImpl(UA, DL, DTU, ST, IsPixelShader, ScanImpl)
+          .run(F)) {
+    PreservedAnalyses PA;
+    PA.preserve<DominatorTreeAnalysis>();
+    return PA;
+  }
+  return PreservedAnalyses::all();
 }
 
 bool AMDGPUAtomicOptimizerImpl::run(Function &F) {
 
   visit(F);
 
   const bool Changed = !ToReplace.empty();
 

llvm/test/CodeGen/AMDGPU/global-atomic-scan.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN:  opt -S -mtriple=amdgcn-- -passes=amdgpu-atomic-optimizer %s | FileCheck -check-prefix=IR %s
+; RUN:  opt -S -mtriple=amdgcn-- -passes='amdgpu-atomic-optimizer,verify<domtree>' %s | FileCheck -check-prefix=IR %s
 
 define amdgpu_kernel void @atomic_add_i32_offset(ptr addrspace(1) %out, i32 %in) {
 ; IR-LABEL: @atomic_add_i32_offset(
 ; IR-NEXT:  entry:
 ; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
 ; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
 ; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
 ; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0