[AMDGPU] Fixed function traversal in attribute propagation

[rampitec]

Summary

AMDGPUPropagateAttributes pass was skipping some of the functions
when clonning. Functions were added to root set and then skipped
on the next interation because they are already in the root set,
while were meant to be processed with different features.

[arsenm]

Typo clonning in commit message

llvm/lib/Target/AMDGPU/AMDGPUPropagateAttributes.cpp

if (Roots.empty())
     return false;
   Module &M = *(*Roots.begin())->getParent();
 
   do {
     Roots.insert(NewRoots.begin(), NewRoots.end());
     NewRoots.clear();
 
     for (auto &F : M.functions()) {
-      if (F.isDeclaration() || Roots.count(&F) || Roots.count(&F))
+      if (F.isDeclaration())
         continue;
 
       const FeatureBitset &CalleeBits =
         TM->getSubtargetImpl(F)->getFeatureBits();
       SmallVector<std::pair<CallBase *, Function *>, 32> ToReplace;
+      SmallSet<CallBase *, 32> Visited;
 
       for (User *U : F.users()) {
         Instruction *I = dyn_cast<Instruction>(U);
         if (!I)
           continue;
         CallBase *CI = dyn_cast<CallBase>(I);
         if (!CI)
           continue;
         Function *Caller = CI->getCaller();
-        if (!Caller)
+        if (!Caller || !Visited.insert(CI).second)
           continue;
-        if (!Roots.count(Caller))
+        if (!Roots.count(Caller) && !NewRoots.count(Caller))
           continue;
 
         const FeatureBitset &CallerBits =
           TM->getSubtargetImpl(*Caller)->getFeatureBits() & TargetFeatures;
 
         if (CallerBits == (CalleeBits  & TargetFeatures)) {
+          if (!Roots.count(&F))
             NewRoots.insert(&F);
           continue;
         }
 
         Function *NewF = findFunction(CallerBits, &F);
         if (!NewF) {
           FeatureBitset NewFeatures((CalleeBits & ~TargetFeatures) |
                                     CallerBits);
           if (!AllowClone) {
do {
     }
   } while (!NewRoots.empty());
 
   for (Function *F : Replaced) {
     if (F->use_empty())
       F->eraseFromParent();
   }
 
+  Roots.clear();
+  Clones.clear();
+
   return Changed;
 }
 
 Function *
 AMDGPUPropagateAttributes::cloneWithFeatures(Function &F,
                                              const FeatureBitset &NewFeatures) {
   LLVM_DEBUG(dbgs() << "Cloning " << F.getName() << '\n');
 

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

 ; 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 @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-INT: define internal fastcc void @foo2.3() unnamed_addr #1
 ; OPT-EXT: define void @foo1() local_unnamed_addr #1
+; OPT-EXT:  tail call void @foo4()
+; OPT-EXT:  tail call void @foo3()
+; OPT-EXT:  tail call void @foo2()
+; OPT-EXT:  tail call void @foo2()
+; OPT-EXT:  tail call void @foo1()
+; OPT-EXT:  tail call fastcc void @0()
 ; OPT-INT: define internal fastcc void @foo1.1() unnamed_addr #1
+; OPT-INT:  tail call void @foo4()
+; OPT-INT:  tail call fastcc void @foo3.2()
+; OPT-INT:  tail call fastcc void @foo2.3()
+; OPT-INT:  tail call fastcc void @foo2.3()
+; OPT-INT:  tail call fastcc void @foo1.1()
+; OPT-INT:  tail call fastcc void @0()
+; OPT:      ret void
 ; OPT: define amdgpu_kernel void @kernel1() local_unnamed_addr #2
+; OPT-EXT:  tail call fastcc void @foo1.1()
+; OPT-INT:  tail call fastcc void @foo1()
+; OPT:      ret void
 ; OPT: define amdgpu_kernel void @kernel2() local_unnamed_addr #3
+; OPT-EXT:  tail call void @foo2()
+; OPT-INT:  tail call fastcc void @foo2.3()
+; OPT:      ret void
 ; OPT: define amdgpu_kernel void @kernel3() local_unnamed_addr #3
+; OPT-EXT:  tail call void @foo1()
+; OPT-INT:  tail call fastcc void @foo1.1()
+; OPT:      ret void
 ; OPT-EXT: define internal fastcc void @foo1.1() unnamed_addr #4
+; OPT-EXT:  tail call void @foo4()
+; OPT-EXT:  tail call fastcc void @foo3.2()
+; OPT-EXT:  tail call fastcc void @foo2.3()
+; OPT-EXT:  tail call fastcc void @foo2.3()
+; OPT-EXT:  tail call fastcc void @foo1.1()
+; OPT-EXT:  tail call fastcc void @1()
 ; OPT-INT: define internal fastcc void @foo1() unnamed_addr #4
+; OPT-INT:  tail call void @foo4()
+; OPT-INT:  tail call fastcc void @foo3()
+; OPT-INT:  tail call fastcc void @foo2()
+; OPT-INT:  tail call fastcc void @foo2()
+; OPT-INT:  tail call fastcc void @foo1()
+; OPT-INT:  tail call fastcc void @1()
+; OPT:      ret void
 ; 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-EXT: define internal fastcc void @foo2.3() unnamed_addr #4
+; OPT-INT: define internal fastcc void @foo2() 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