[AMDGPU] Preserve original symbol during attribute propagation

[rampitec]

Summary

AMDGPUPropagateAttributes can swap names while cloning a function.
Only do it if original symbol was not externally visible.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUPropagateAttributes.cpp

282

The linkage wasn't set before for anonymous functions? That should get a test?

[rampitec]

Added test for anonymous function.
Fixed FileCheck prefixes.

llvm/lib/Target/AMDGPU/AMDGPUPropagateAttributes.cpp

282

It is universally set now on every clone as we do not mess with the original function name anymore.

llvm/lib/Target/AMDGPU/AMDGPUPropagateAttributes.cpp

 Function *
 AMDGPUPropagateAttributes::cloneWithFeatures(Function &F,
                                              const FeatureBitset &NewFeatures) {
   LLVM_DEBUG(dbgs() << "Cloning " << F.getName() << '\n');
 
   ValueToValueMapTy dummy;
   Function *NewF = CloneFunction(&F, dummy);
   setFeatures(*NewF, NewFeatures);
+  NewF->setVisibility(GlobalValue::DefaultVisibility);
+  NewF->setLinkage(GlobalValue::InternalLinkage);
 
   // Swap names. If that is the only clone it will retain the name of now
-  // dead value.
-  if (F.hasName()) {
+  // dead value. Preserve original name for externally visible functions.
+  if (F.hasName() && F.hasLocalLinkage()) {
     std::string NewName = std::string(NewF->getName());
     NewF->takeName(&F);
     F.setName(NewName);
-
-    // Name has changed, it does not need an external symbol.
-    F.setVisibility(GlobalValue::DefaultVisibility);
-    F.setLinkage(GlobalValue::InternalLinkage);
   }
 
   return NewF;
 }
 
 void AMDGPUPropagateAttributes::setFeatures(Function &F,
                                             const FeatureBitset &NewFeatures) {
   std::string NewFeatureStr = getFeatureString(NewFeatures);

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

Builder.addExtension(
     [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
                                                legacy::PassManagerBase &PM) {
       if (AMDGPUAA) {
         PM.add(createAMDGPUAAWrapperPass());
         PM.add(createAMDGPUExternalAAWrapperPass());
       }
       PM.add(createAMDGPUUnifyMetadataPass());
       PM.add(createAMDGPUPrintfRuntimeBinding());
-      PM.add(createAMDGPUPropagateAttributesLatePass(this));
-      if (Internalize) {
+      if (Internalize)
         PM.add(createInternalizePass(mustPreserveGV));
+      PM.add(createAMDGPUPropagateAttributesLatePass(this));
+      if (Internalize)
         PM.add(createGlobalDCEPass());
-      }
       if (EarlyInline)
         PM.add(createAMDGPUAlwaysInlinePass(false));
   });
 
   const auto &Opt = Options;
   Builder.addExtension(
     PassManagerBuilder::EP_EarlyAsPossible,
     [AMDGPUAA, LibCallSimplify, &Opt, this](const PassManagerBuilder &,

llvm/test/CodeGen/AMDGPU/propagate-attributes-clone.ll

-; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -O1 < %s | FileCheck -check-prefix=OPT %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -O1 < %s | FileCheck -check-prefixes=OPT,OPT-EXT %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -O1 --amdgpu-internalize-symbols < %s | FileCheck -check-prefixes=OPT,OPT-INT %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=LLC %s
 
 ; OPT: declare void @foo4() local_unnamed_addr #0
-; OPT: define internal fastcc void @foo3.2() unnamed_addr #1
-; OPT: define void @foo2() local_unnamed_addr #1
-; OPT: define internal fastcc void @foo1.1() unnamed_addr #1
+; OPT: define internal fastcc void @0() unnamed_addr #1
+; OPT-EXT: define void @foo3() local_unnamed_addr #1
+; OPT-INT: define internal fastcc void @foo3.2() unnamed_addr #1
+; OPT-EXT: define void @foo2() local_unnamed_addr #1
+; OPT-INT: define internal fastcc void @foo2() unnamed_addr #1
+; OPT-EXT: define void @foo1() local_unnamed_addr #1
+; OPT-INT: define internal fastcc void @foo1.1() unnamed_addr #1
 ; OPT: define amdgpu_kernel void @kernel1() local_unnamed_addr #2
 ; OPT: define amdgpu_kernel void @kernel2() local_unnamed_addr #3
 ; OPT: define amdgpu_kernel void @kernel3() local_unnamed_addr #3
-; OPT: define void @foo1() local_unnamed_addr #4
-; OPT: define void @foo3() local_unnamed_addr #4
+; OPT-EXT: define internal fastcc void @foo1.1() unnamed_addr #4
+; OPT-INT: define internal fastcc void @foo1() unnamed_addr #4
+; OPT: define internal fastcc void @1() unnamed_addr #4
+; OPT-EXT: define internal fastcc void @foo3.2() unnamed_addr #4
+; OPT-INT: define internal fastcc void @foo3() unnamed_addr #4
 ; OPT: attributes #0 = { {{.*}} "target-features"="+wavefrontsize64" }
 ; OPT: attributes #1 = { {{.*}} "target-features"="{{.*}},-wavefrontsize16,-wavefrontsize32,+wavefrontsize64{{.*}}" }
 ; OPT: attributes #2 = { {{.*}} "target-features"="+wavefrontsize32" }
 ; OPT: attributes #3 = { {{.*}} "target-features"="+wavefrontsize64" }
 ; OPT: attributes #4 = { {{.*}} "target-features"="{{.*}},-wavefrontsize16,+wavefrontsize32,-wavefrontsize64{{.*}}" }
 
 ; LLC: foo3:
 ; LLC: sample asm
 ; LLC: foo2:
 ; LLC: sample asm
 ; LLC: foo1:
 ; LLC: foo4@gotpcrel32@lo+4
 ; LLC: foo4@gotpcrel32@hi+4
 ; LLC: foo3@gotpcrel32@lo+4
 ; LLC: foo3@gotpcrel32@hi+4
 ; LLC: foo2@gotpcrel32@lo+4
 ; LLC: foo2@gotpcrel32@hi+4
 ; LLC: foo1@gotpcrel32@lo+4
 ; LLC: foo1@gotpcrel32@hi+4
+; LLC: __unnamed_1@gotpcrel32@lo+4
+; LLC: __unnamed_1@gotpcrel32@hi+4
 ; LLC: kernel1:
 ; LLC: foo1@gotpcrel32@lo+4
 ; LLC: foo1@gotpcrel32@hi+4
 ; LLC: kernel2:
 ; LLC: foo2@gotpcrel32@lo+4
 ; LLC: foo2@gotpcrel32@hi+4
 ; LLC: kernel3:
 ; LLC: foo1@gotpcrel32@lo+4
 ; LLC: foo1@gotpcrel32@hi+4
 
 declare void @foo4() #1
 
+define void @0() #1 {
+entry:
+  call void asm sideeffect "; sample asm", ""()
+  ret void
+}
+
 define void @foo3() #1 {
 entry:
   call void asm sideeffect "; sample asm", ""()
   ret void
 }
 
 define void @foo2() #1 {
 entry:
   call void asm sideeffect "; sample asm", ""()
   ret void
 }
 
 define void @foo1() #1 {
 entry:
   tail call void @foo4()
   tail call void @foo3()
   tail call void @foo2()
   tail call void @foo2()
   tail call void @foo1()
+  tail call void @0()
   ret void
 }
 
 define amdgpu_kernel void @kernel1() #0 {
 entry:
   tail call void @foo1()
   ret void
 }