[CUDA][HIP] Emit dependent libs for host only

[yaxunl]

Summary

Recently D60274 was introduced to allow lld to handle dependent libs. However current
usage of dependent libs (e.g. pragma comment(lib, *) in windows header files) are intended
for host only. Emitting the metadata in device IR causes link error in device path.

Until there is a way to different it dependent libs for device or host, metadata for dependent
libs should be emitted for host only. This patch enforces that.

[phosek]

This seems to be failing on macOS bots with the following error:

FAIL: Clang :: CodeGenCUDA/dependent-libs.cu (3052 of 14933)
******************** TEST 'Clang :: CodeGenCUDA/dependent-libs.cu' FAILED ********************
Script:
--
: 'RUN: at line 1';   /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/bin/clang -cc1 -internal-isystem /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/lib/clang/9.0.0/include -nostdsysteminc -emit-llvm -o - -fcuda-is-device -x hip /b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu | /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/bin/FileCheck --check-prefix=DEV /b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu
: 'RUN: at line 2';   /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/bin/clang -cc1 -internal-isystem /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/lib/clang/9.0.0/include -nostdsysteminc -emit-llvm -o - -x hip /b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu | /b/s/w/ir/k/recipe_cleanup/clangAoYUvt/llvm_build_dir/bin/FileCheck --check-prefix=HOST /b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu
--
Exit Code: 1

Command Output (stderr):
--
/b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu:5:10: error: HOST: expected string not found in input
// HOST: llvm.dependent-libraries
         ^
<stdin>:1:1: note: scanning from here
; ModuleID = '/b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu'
^
<stdin>:1:58: note: possible intended match here
; ModuleID = '/b/s/w/ir/k/llvm-project/clang/test/CodeGenCUDA/dependent-libs.cu'
                                                         ^

--

The full test log is here: https://logs.chromium.org/logs/fuchsia/buildbucket/cr-buildbucket.appspot.com/8912171419117598448/+/steps/clang/0/steps/test/0/stdout

lib/CodeGen/CodeGenModule.cpp

void CodeGenModule::Release() {
 
   // On ELF we pass the dependent library specifiers directly to the linker
   // without manipulating them. This is in contrast to other platforms where
   // they are mapped to a specific linker option by the compiler. This
   // difference is a result of the greater variety of ELF linkers and the fact
   // that ELF linkers tend to handle libraries in a more complicated fashion
   // than on other platforms. This forces us to defer handling the dependent
   // libs to the linker.
-  if (!ELFDependentLibraries.empty()) {
+  //
+  // CUDA/HIP device and host libraries are different. Currently there is no
+  // way to differentiate dependent libraries for host or device. Existing
+  // usage of #pragma comment(lib, *) is intended for host libraries on
+  // Windows. Therefore emit llvm.dependent-libraries only for host.
+  if (!ELFDependentLibraries.empty() && !Context.getLangOpts().CUDAIsDevice) {
     auto *NMD = getModule().getOrInsertNamedMetadata("llvm.dependent-libraries");
     for (auto *MD : ELFDependentLibraries)
       NMD->addOperand(MD);
   }
 
   // Record mregparm value now so it is visible through rest of codegen.
   if (Context.getTargetInfo().getTriple().getArch() == llvm::Triple::x86)
     getModule().addModuleFlag(llvm::Module::Error, "NumRegisterParameters",

test/CodeGenCUDA/dependent-libs.cu

+// RUN: %clang_cc1 -emit-llvm -o - -fcuda-is-device -x hip %s | FileCheck --check-prefix=DEV %s
+// RUN: %clang_cc1 -emit-llvm -o - -x hip %s | FileCheck --check-prefix=HOST %s
+
+// DEV-NOT: llvm.dependent-libraries
+// HOST: llvm.dependent-libraries
+#pragma comment(lib, "libabc")