[AMDGPU][OpenMP] Remove optnone from outlined functions

[pdhaliwal]

Summary

This depends on D102107 and unblocks the failing amdgcn runtime
tests in the latter.

There seems to be some issue with function pointers handling
in the middle/back end. Removing optnone attribute makes the
optimizer to replace indirect call with direct call.

[JonChesterfield]

Not pretty but unblocking D102107 is important. Could you write up your current understanding of what we're miscompiling for function calls as a comment on this diff (i.e. not in the code, just in the review discussion)? I'm guessing it's still attribute related but am a few days out of date.

I won't tag it as accepted because it's taking on debt in clang openmp to work around a bug in amdgpu middle/back end and I have a conflict of interest there. Hopefully an external person will accept.

[pdhaliwal]

I don't have any concrete evidence but I have some doubt on presence of function pointers causing backend to behave improperly. Also, here removing optnone alone suffices to fix the issue.

[pdhaliwal]

Only removing optnone.

[jdoerfert]

What if we track down the problem instead? This will simply pop up again in O0 user code, no?

[JonChesterfield]

Root cause would definitely be better but we don't have one at present and would like to unblock the linked diff. O0 is likely to have more problems than just this.

clang/lib/CodeGen/CGStmtOpenMP.cpp

static llvm::Function *emitOutlinedFunctionPrologueAggregate(
   auto *F =
       llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
                              FunctionName, &CGM.getModule());
   CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
   if (CD->isNothrow())
     F->setDoesNotThrow();
   F->setDoesNotRecurse();
 
+  // TODO: should not need this once amdgcn handles function pointers properly.
+  if (CGM.getTriple().isAMDGCN()) {
+    F->removeFnAttr(llvm::Attribute::OptimizeNone);
+  }
+
   // Generate the function.
   CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args, Loc, Loc);
   Address ContextAddr = CGF.GetAddrOfLocalVar(CD->getContextParam());
   llvm::Value *ContextV = CGF.Builder.CreateLoad(ContextAddr);
   LValue ContextLV = CGF.MakeNaturalAlignAddrLValue(
       ContextV, CGM.getContext().getTagDeclType(RD));
   const auto *I = CS.captures().begin();
   for (const FieldDecl *FD : RD->fields()) {
auto *F =
       llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
                              FO.FunctionName, &CGM.getModule());
   CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
   if (CD->isNothrow())
     F->setDoesNotThrow();
   F->setDoesNotRecurse();
 
   // Always inline the outlined function if optimizations are enabled.
-  if (CGM.getCodeGenOpts().OptimizationLevel != 0)
+  if (CGM.getCodeGenOpts().OptimizationLevel != 0) {
     F->addFnAttr(llvm::Attribute::AlwaysInline);
+  }
+
+  // TODO: remove this once amdgcn handles function pointers properly.
+  if (CGM.getTriple().isAMDGCN()) {
+    F->removeFnAttr(llvm::Attribute::OptimizeNone);
+  }
 
   // Generate the function.
   CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, TargetArgs,
                     FO.UIntPtrCastRequired ? FO.Loc : FO.S->getBeginLoc(),
                     FO.UIntPtrCastRequired ? FO.Loc
                                            : CD->getBody()->getBeginLoc());
   unsigned Cnt = CD->getContextParamPosition();
   I = FO.S->captures().begin();