[Clang] Do not pass -shared when using -r for AMDGPU (#100760)

Summary:
We can use `-r` and `--lto-emit-llvm` to get the LTO pass to optimize +
link LLVM-IR. Currently this doesn't work on AMDGPU because it always
passes `-shared` which is incompatible with `-r`. Fix that so we can use
it.

Author: jhuber6

clang/lib/Driver/ToolChains/AMDGPU.cpp

@@ -620,8 +620,10 @@ void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA,
                                   const char *LinkingOutput) const {
   std::string Linker = getToolChain().GetLinkerPath();
   ArgStringList CmdArgs;
-  CmdArgs.push_back("--no-undefined");
-  CmdArgs.push_back("-shared");
+  if (!Args.hasArg(options::OPT_r)) {
+    CmdArgs.push_back("--no-undefined");
+    CmdArgs.push_back("-shared");
+  }
 
   addLinkerCompressDebugSectionsOption(getToolChain(), Args, CmdArgs);
   Args.AddAllArgs(CmdArgs, options::OPT_L);

clang/test/Driver/amdgpu-toolchain.c

@@ -28,3 +28,7 @@
 // RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx906 -nogpulib \
 // RUN:   -fuse-ld=ld %s 2>&1 | FileCheck -check-prefixes=LD %s
 // LD: ld.lld
+
+// RUN: %clang -### --target=amdgcn-amd-amdhsa -mcpu=gfx906 -nogpulib \
+// RUN:   -r %s 2>&1 | FileCheck -check-prefixes=RELO %s
+// RELO-NOT: -shared