[OpenMP] Fix: opposite attributes could be set by -fno-inline

[igor.kirillov]

Summary

After the changes introduced by D106799 it is possible to tag
outlined function with both AlwaysInline and NoInline attributes using
-fno-inline command line options.
This issue is similiar to D107649.

[igor.kirillov]

clang/lib/CodeGen/CGStmtOpenMP.cpp

5370

I noticed that here we set this attribute on the same function that is created by emitOutlinedFunctionPrologue but there AlwaysInline attribute is guaranteed to be set under same condition. So I suppose we could remove this piece of code?

[jhuber6]

LGTM.

clang/lib/CodeGen/CGStmtOpenMP.cpp

5370

Yeah, you can remove this.

clang/test/OpenMP/parallel_for_noinline.cpp

1

This can probably be merged with the other test added in D107649 since it's the same problem, just add another run line and run the update script.

[igor.kirillov]

clang/test/OpenMP/parallel_for_noinline.cpp

1

It was my first thought too but I don't see how to check this bug with #pragma omp parallel if(0). If I add function with #pragma omp parallel for there it goes contrary to the test file name (parallel_if_codegen_PR51349). More than that, checks generated by update_cc_test_checks.py are too overwhelming and hide whatever we were going to test there. I can do it this way or may be you have any other ideas?

What if I rename my file this test file name to parallel_for_codegen_PR51349.cpp instead? They are related and thus we don't loose sight of them.

[jhuber6]

clang/test/OpenMP/parallel_for_noinline.cpp

1

Just adding this triggered the bug for me and hardly generates anything.

#pragma omp parallel
  ;

[igor.kirillov]

Update test, remove redundant code

[igor.kirillov]

Remove my old test for the change

clang/lib/CodeGen/CGStmtOpenMP.cpp

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->removeFnAttr(llvm::Attribute::NoInline);
     F->addFnAttr(llvm::Attribute::AlwaysInline);
+  }
 
   // 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();
static llvm::Function *emitOutlinedOrderedFunction(CodeGenModule &CGM,
                                                    const CapturedStmt *S,
                                                    SourceLocation Loc) {
   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
   CodeGenFunction::CGCapturedStmtInfo CapStmtInfo;
   CGF.CapturedStmtInfo = &CapStmtInfo;
   llvm::Function *Fn = CGF.GenerateOpenMPCapturedStmtFunction(*S, Loc);
   Fn->setDoesNotRecurse();
   if (CGM.getCodeGenOpts().OptimizationLevel != 0)
     Fn->addFnAttr(llvm::Attribute::AlwaysInline);
   return Fn;
 }
 
 void CodeGenFunction::EmitOMPOrderedDirective(const OMPOrderedDirective &S) {
   if (CGM.getLangOpts().OpenMPIRBuilder) {
     llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder();
     using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
 

clang/test/OpenMP/parallel_for_noinline.cpp

+// RUN: %clang -O0 -fopenmp -fno-inline %s -S -emit-llvm -o - | FileCheck %s --check-prefixes=CHECK,CHECK-O0
+// RUN: %clang -O1 -fopenmp -fno-inline %s -S -emit-llvm -o - | FileCheck %s --check-prefixes=CHECK,CHECK-O1
+
+// CHECK-LABEL: define dso_local void @_Z3foov
+// CHECK: Function Attrs:
+// CHECK-O0-SAME: inline
+// CHECK-O0-NOT: awaysinline
+// CHECK-O1-SAME: alwaysinline
+// CHECK-O1-NOT: noinline
+// CHECK-LABEL: define internal void @.omp_outlined
+void foo() {
+   #pragma omp parallel for
+   for (int i = 0; i < 100; ++i)
+     ;
+}