[OpenMP][NFC] Move the `noinline` to the parallel entry point

[jdoerfert]

Summary

The noinline for non-SPMD parallel functions is probably not necessary
but as long as we use it we should put it on the outermost parallel
function, which is the wrapper, not the actual outlined function.

Resolves PR49752

[JonChesterfield]

Context in https://bugs.llvm.org/show_bug.cgi?id=49752 is that this resolves a regression in stack usage from D94315. This change looks good. I'm not totally sold on using a function call boundary to convey invariants on ICV, but that's an existing property.

[fhahn]

Is it possible to add a test?

[jdoerfert]

In D99506#2655948, @fhahn wrote:

Is it possible to add a test?

There is a test for the presence of noinline, let me make it more explicit.

[jdoerfert]

Add test for nvptx codegen, including wrapper attribute check

clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

 void CGOpenMPRuntimeGPU::emitNonSPMDParallelCall(
     CodeGenFunction &CGF, SourceLocation Loc, llvm::Value *OutlinedFn,
     ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
   llvm::Function *Fn = cast<llvm::Function>(OutlinedFn);
 
   // Force inline this outlined function at its call site.
   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
 
-  // Ensure we do not inline the function. This is trivially true for the ones
-  // passed to __kmpc_fork_call but the ones calles in serialized regions
-  // could be inlined. This is not a perfect but it is closer to the invariant
-  // we want, namely, every data environment starts with a new function.
-  // TODO: We should pass the if condition to the runtime function and do the
-  //       handling there. Much cleaner code.
-  cast<llvm::Function>(OutlinedFn)->addFnAttr(llvm::Attribute::NoInline);
-
   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                                       /*Name=*/".zero.addr");
   CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
   // ThreadId for serialized parallels is 0.
   Address ThreadIDAddr = ZeroAddr;
   auto &&CodeGen = [this, Fn, CapturedVars, Loc, &ThreadIDAddr](
                        CodeGenFunction &CGF, PrePostActionTy &Action) {
     Action.Enter(CGF);
llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
   WrapperArgs.emplace_back(&WrapperArg);
 
   const CGFunctionInfo &CGFI =
       CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
 
   auto *Fn = llvm::Function::Create(
       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
       Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
+
+  // Ensure we do not inline the function. This is trivially true for the ones
+  // passed to __kmpc_fork_call but the ones calles in serialized regions
+  // could be inlined. This is not a perfect but it is closer to the invariant
+  // we want, namely, every data environment starts with a new function.
+  // TODO: We should pass the if condition to the runtime function and do the
+  //       handling there. Much cleaner code.
+  Fn->addFnAttr(llvm::Attribute::NoInline);
+
   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
   Fn->setDoesNotRecurse();
 
   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
   CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
                     D.getBeginLoc(), D.getBeginLoc());
 

clang/test/OpenMP/nvptx_parallel_codegen.cpp

 // Test target codegen - host bc file has to be created first.
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm-bc %s -o %t-ppc-host.bc
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - -disable-llvm-optzns | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 --check-prefix SEQ
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - -disable-llvm-optzns -fopenmp-cuda-parallel-target-regions | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 --check-prefix PAR
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm-bc %s -o %t-x86-host.bc
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - -disable-llvm-optzns | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 --check-prefix SEQ
-// RUN: %clang_cc1 -verify -fopenmp -fexceptions -fcxx-exceptions -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - -disable-llvm-optzns | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 --check-prefix SEQ
+// RUN: %clang_cc1 -verify -fopenmp -fexceptions -fcxx-exceptions -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - -disable-llvm-optzns -disable-O0-optnone | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 --check-prefix SEQ
 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - -disable-llvm-optzns -fopenmp-cuda-parallel-target-regions | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 --check-prefix PAR
 // RUN: %clang_cc1 -verify -fopenmp -fexceptions -fcxx-exceptions -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - -disable-llvm-optzns -fopenmp-cuda-parallel-target-regions | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 --check-prefix PAR
 // expected-no-diagnostics
 #ifndef HEADER
 #define HEADER
 
 template<typename tx>
 tx ftemplate(int n) {

 // CHECK: [[TERMINATE]]
 // CHECK: call void @__kmpc_kernel_deinit(
 // CHECK: call void @__kmpc_barrier_simple_spmd(%struct.ident_t* null, i32 0)
 // CHECK: br label {{%?}}[[EXIT]]
 //
 // CHECK: [[EXIT]]
 // CHECK: ret void
 
-// CHECK: define internal void [[PARALLEL_FN4]](
+// CHECK: noinline
+// CHECK-NEXT: define internal void [[PARALLEL_FN4]](
 // CHECK: [[A:%.+]] = alloca i[[SZ:32|64]],
 // CHECK: store i[[SZ]] 45, i[[SZ]]* %a,
 // CHECK: call void @__kmpc_barrier(%struct.ident_t* @{{.+}}, i32 %{{.+}})
 // CHECK: ret void
 
 // CHECK: declare void @__kmpc_barrier(%struct.ident_t*, i32) #[[#CONVERGENT:]]
 
+// CHECK: Function Attrs: convergent noinline norecurse nounwind
+// CHECK-NEXT: [[PARALLEL_FN4]]_wrapper
+
 // CHECK-LABEL: define {{.*}}void {{@__omp_offloading_.+template.+l58}}_worker()
 // CHECK-LABEL: define {{.*}}void {{@__omp_offloading_.+template.+l58}}(
 // CHECK-32: [[A_ADDR:%.+]] = alloca i32,
 // CHECK-64: [[A_ADDR:%.+]] = alloca i64,
 // CHECK-64: [[CONV:%.+]] = bitcast i64* [[A_ADDR]] to i32*
 // SEQ: [[IS_SHARED:%.+]] = load i16, i16* [[KERNEL_SHARED]],
 // SEQ: [[SIZE:%.+]] = load i{{64|32}}, i{{64|32}}* [[KERNEL_SIZE]],
 // SEQ: call void @__kmpc_get_team_static_memory(i16 0, i8* addrspacecast (i8 addrspace(3)* getelementptr inbounds ([[MEM_TY]], [[MEM_TY]] addrspace(3)* [[SHARED_GLOBAL_RD]], i32 0, i32 0, i32 0) to i8*), i{{64|32}} [[SIZE]], i16 [[IS_SHARED]], i8** addrspacecast (i8* addrspace(3)* [[KERNEL_PTR]] to i8**))

 // CHECK:  store i32
 // CHECK:  call void @__kmpc_end_critical(
 
 // CHECK:  call void @__kmpc_syncwarp(i32 [[MASK]]){{$}}
 // CHECK:  [[NEW_CC_VAL:%.+]] = add nsw i32 [[CC_VAL]], 1
 // CHECK:  store i32 [[NEW_CC_VAL]], i32* [[CC]],
 // CHECK:  br label
 
-
 // CHECK: declare i32 @__kmpc_warp_active_thread_mask() #[[#CONVERGENT:]]
 // CHECK: declare void @__kmpc_syncwarp(i32) #[[#CONVERGENT:]]
 
 // CHECK: attributes #[[#CONVERGENT:]] = {{.*}} convergent {{.*}}
 
 #endif