[Comgr] Add -relink-builtin-bitcode-postop to device library linking

Compiling with the ROCm device libraries is currently possible via
the *_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC action. However, with
this action it is possible to encounter LLD errors if new device
library calls are introduced during optimization.

In this patch, we add a newly introduced LLVM option,
-relink-builtin-bitcode-postop (llvm#69371),
to re-link the device libraries after optimization, resolving any
newly introduced definitions. We also update the relevant test
to reflect this improvement

Change-Id: I3889b7a5ffab795065684c59486c6a9965a038f2

Author: lamb-j

amd/comgr/docs/ReleaseNotes.md

@@ -59,7 +59,10 @@ lld::elf::link in Comgr's linkWithLLD()
 - Added -x assembler option to assembly compilation. Before, if an assembly file
 did not end with a .s file extension, it was not handled properly by the Comgr
 ASSEMBLE\_SOURCE\_TO\_RELOCATABLE action.
-
+- Added new -relink-builtin-bitcode-postop LLVM option to device library. This
+fixes an issue with the \*COMPILE\_SOURCE\_WITH\_DEVICE\_LIBRARIES\_TO\_BC where
+OpenCL applications that leveraged AMDGPUSimplifyLibCalls optimizations would
+need to re-link bitcodes separately to avoid errors at runtime.
 
 New APIs
 --------

amd/comgr/src/comgr-compiler.cpp

@@ -1082,6 +1082,8 @@ amd_comgr_status_t AMDGPUCompiler::addDeviceLibraries() {
     return AMD_COMGR_STATUS_ERROR;
   }
   Args.push_back(Saver.save(Twine("--rocm-path=") + FakeRocmDir).data());
+  Args.push_back("-mllvm");
+  Args.push_back("-relink-builtin-bitcode-postop");
   NoGpuLib = false;
 
   for (auto DeviceLib : getDeviceLibraries()) {

amd/comgr/test/compile_source_with_device_libs_to_bc_test.c

@@ -48,7 +48,7 @@ int main(int argc, char *argv[]) {
   amd_comgr_action_info_t DataAction;
   amd_comgr_status_t Status;
   const char *CodeGenOptions[] = {"-mllvm", "-amdgpu-early-inline-all",
-    "-mcode-object-version=5"};
+    "-mcode-object-version=5", "-mllvm", "-amdgpu-prelink"};
   size_t CodeGenOptionsCount =
       sizeof(CodeGenOptions) / sizeof(CodeGenOptions[0]);