Revert "SWDEV-371628 - Shift device lib linking into clang driver"

This reverts commit a4c5c44008.

Reason for revert: Performance regressions and failures observed. Need to investigate those and before re-applying patch

Change-Id: I42ba0605797f9bdcfb5d5102927dd01405cf05e3

rocclr/device/devprogram.cpp

@@ -348,7 +348,7 @@ amd_comgr_status_t Program::createAction(const amd_comgr_language_t oclver,
 bool Program::linkLLVMBitcode(const amd_comgr_data_set_t inputs,
                               const std::vector<std::string>& options,
                               amd::option::Options* amdOptions, amd_comgr_data_set_t* output,
-                              char* binaryData[], size_t* binarySize) {
+                              char* binaryData[], size_t* binarySize, const bool link_dev_libs) {
 
   amd_comgr_language_t langver = getCOMGRLanguage(isHIP(), *amdOptions);
   if (langver == AMD_COMGR_LANGUAGE_NONE) {
@@ -357,13 +357,28 @@ bool Program::linkLLVMBitcode(const amd_comgr_data_set_t inputs,
 
   //  Create the action for linking
   amd_comgr_action_info_t action;
+  amd_comgr_data_set_t dataSetDevLibs;
   bool hasAction = false;
+  bool hasDataSetDevLibs = false;
 
   amd_comgr_status_t status = createAction(langver, options, &action, &hasAction);
 
+  if (link_dev_libs) {
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      status = amd::Comgr::create_data_set(&dataSetDevLibs);
+    }
+
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      hasDataSetDevLibs = true;
+      status = amd::Comgr::do_action(AMD_COMGR_ACTION_ADD_DEVICE_LIBRARIES, action, inputs,
+                                    dataSetDevLibs);
+      extractBuildLog(dataSetDevLibs);
+    }
+  }
+
   if (status == AMD_COMGR_STATUS_SUCCESS) {
     status = amd::Comgr::do_action(AMD_COMGR_ACTION_LINK_BC_TO_BC, action,
-                                   inputs, *output);
+      (link_dev_libs) ? dataSetDevLibs : inputs, *output);
     extractBuildLog(*output);
   }
 
@@ -380,14 +395,17 @@ bool Program::linkLLVMBitcode(const amd_comgr_data_set_t inputs,
     amd::Comgr::destroy_action_info(action);
   }
 
+  if (hasDataSetDevLibs) {
+    amd::Comgr::destroy_data_set(dataSetDevLibs);
+  }
+
   return (status == AMD_COMGR_STATUS_SUCCESS);
 }
 
 bool Program::compileToLLVMBitcode(const amd_comgr_data_set_t compileInputs,
                                    const std::vector<std::string>& options,
                                    amd::option::Options* amdOptions,
-                                   char* binaryData[], size_t* binarySize,
+                                   char* binaryData[], size_t* binarySize) {
-                                   const bool link_dev_libs) {
 
   amd_comgr_language_t langver = getCOMGRLanguage(isHIP(), *amdOptions);
   if (langver == AMD_COMGR_LANGUAGE_NONE) {
@@ -461,14 +479,8 @@ bool Program::compileToLLVMBitcode(const amd_comgr_data_set_t compileInputs,
 
   //  Compiling the source codes with precompiled headers or directly compileInputs
   if (status == AMD_COMGR_STATUS_SUCCESS) {
-    if (link_dev_libs) {
+    status = amd::Comgr::do_action(AMD_COMGR_ACTION_COMPILE_SOURCE_TO_BC,
-      status = amd::Comgr::do_action(AMD_COMGR_ACTION_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC,
+                                 action, input, output);
-                                     action, input, output);
-    }
-    else {
-      status = amd::Comgr::do_action(AMD_COMGR_ACTION_COMPILE_SOURCE_TO_BC,
-                                     action, input, output);
-    }
     extractBuildLog(output);
   }
 
@@ -665,11 +677,8 @@ bool Program::compileImplLC(const std::string& sourceCode,
   driverOptions.push_back("-mllvm");
   driverOptions.push_back("-amdgpu-early-inline-all");
 #endif
-
+  driverOptions.push_back("-mllvm");
-  // TODO: Re-enable prelink once folding optimizations no longer
+  driverOptions.push_back("-amdgpu-prelink");
-  // introduce undefined symbol errors
-  //driverOptions.push_back("-mllvm");
-  //driverOptions.push_back("-amdgpu-prelink");
 
   if (!device().settings().enableWgpMode_) {
     driverOptions.push_back("-mcumode");
@@ -959,8 +968,9 @@ bool Program::linkImplLC(const std::vector<Program*>& inputPrograms,
   char* binaryData = nullptr;
   size_t binarySize = 0;
   std::vector<std::string> linkOptions;
+  constexpr bool kLinkDevLibs = false;
   bool ret = linkLLVMBitcode(inputs, linkOptions, options, &output, &binaryData,
-                             &binarySize);
+                             &binarySize, kLinkDevLibs);
 
   amd::Comgr::destroy_data_set(output);
   amd::Comgr::destroy_data_set(inputs);

rocclr/device/devprogram.hpp

@@ -457,13 +457,13 @@ class Program : public amd::HeapObject {
   bool linkLLVMBitcode(const amd_comgr_data_set_t inputs,
     const std::vector<std::string>& options,
     amd::option::Options* amdOptions, amd_comgr_data_set_t* output,
-    char* binaryData[] = nullptr, size_t* binarySize = nullptr);
+    char* binaryData[] = nullptr, size_t* binarySize = nullptr,
+    const bool link_dev_libs = true);
 
   //! Create the bitcode of the compiled input dataset
   bool compileToLLVMBitcode(const amd_comgr_data_set_t compileInputs,
     const std::vector<std::string>& options, amd::option::Options* amdOptions,
-    char* binaryData[], size_t* binarySize,
+    char* binaryData[], size_t* binarySize);
-    const bool link_dev_libs = true);
 
   //! Compile and create the excutable of the input dataset
   bool compileAndLinkExecutable(const amd_comgr_data_set_t inputs,