[clang][codegen] Skip adding default function attributes on intrinsics.

[hliao]

Summary

• After loading builtin bitcode for linking, skip adding default function attributes on LLVM intrinsics as their attributes are well-defined and retrieved directly from internal definitions. Adding extra attributes on intrinsics results in inconsistent result when -save-temps is present. Also, that makes few optimizations conservative.

[rjmccall]

Seems reasonable.

clang/lib/CodeGen/CodeGenAction.cpp

void HandleInterestingDecl(DeclGroupRef D) override {
       if (!IRGenFinished)
         HandleTopLevelDecl(D);
     }
 
     // Links each entry in LinkModules into our module.  Returns true on error.
     bool LinkInModules() {
       for (auto &LM : LinkModules) {
         if (LM.PropagateAttrs)
-          for (Function &F : *LM.Module)
+          for (Function &F : *LM.Module) {
+            // Skip intrinsics. Keep consistent with how intrinsics are created
+            // in LLVM IR.
+            if (F.isIntrinsic())
+              continue;
             Gen->CGM().addDefaultFunctionDefinitionAttributes(F);
+          }
 
         CurLinkModule = LM.Module.get();
 
         bool Err;
         if (LM.Internalize) {
           Err = Linker::linkModules(
               *getModule(), std::move(LM.Module), LM.LinkFlags,
               [](llvm::Module &M, const llvm::StringSet<> &GVS) {

clang/test/CodeGenCUDA/Inputs/device-lib-code.ll

+define linkonce_odr protected float @__ocml_fma_f32(float %0, float %1, float %2) local_unnamed_addr {
+  %4 = tail call float @llvm.fma.f32(float %0, float %1, float %2)
+  ret float %4
+}
+declare float @llvm.fma.f32(float, float, float)

clang/test/CodeGenCUDA/dft-func-attr-skip-intrinsic.hip

+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -x ir -fcuda-is-device -triple amdgcn-amd-amdhsa -emit-llvm-bc -disable-llvm-passes -o %t.bc %S/Inputs/device-lib-code.ll
+// RUN: %clang_cc1 -x hip -fcuda-is-device -triple amdgcn-amd-amdhsa -mlink-builtin-bitcode %t.bc -emit-llvm -disable-llvm-passes -o - %s | FileCheck %s
+
+#include "Inputs/cuda.h"
+
+extern "C" __device__ float __ocml_fma_f32(float x, float y, float z);
+
+__device__ float foo(float x) {
+  return __ocml_fma_f32(x, x, x);
+}
+
+// CHECK: {{^}}define{{.*}} @__ocml_fma_f32{{.*}} [[ATTR1:#[0-9]+]]
+// CHECK: {{^}}declare{{.*}} @llvm.fma.f32{{.*}} [[ATTR2:#[0-9]+]]
+// CHECK: attributes [[ATTR1]] = { convergent
+// CHECK: attributes [[ATTR2]] = {
+// CHECK-NOT: convergent
+// CHECK: }