[LTO] Add configuartion option to use default optimization pipeline

[jhuber6]

Summary

This patch adds a configuration option to simply use the default pass
pipeline in favor of the LTO-specific one. We observed some severe
performance penalties when uding device-side LTO for OpenMP offloading
applications caused by the LTO-pass pipeline. This is primarily because
OpenMP uses an LLVM bitcode library to implement a GPU runtime library.
In a standard compilation we link this bitcode library into each source
file and optimize it with the default pipeline. When performing LTO we
link it late with all the files, but the bitcode library never has the
regular optimization pipeline applied to it so we miss a few
optimizations just using the LTO pipeline to optimize it.

I'm not committed to this solution, but it's the easiest method to solve
this performance regression when using LTO without changing the
optimizatin pipeline for other users.

[tianshilei1992]

I tested it locally, and it fixes the performance regression. Since it is internal flag, it will not break any existing code. I'm okay with the change. Not sure if we want to ask the gate keeper of LTOBackend.

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

 std::unique_ptr<lto::LTO> createLTO(
     const Triple &TheTriple, StringRef Arch, bool WholeProgram,
     ModuleHook Hook = [](size_t, const Module &) { return true; }) {
   lto::Config Conf;
   lto::ThinBackend Backend;
   // TODO: Handle index-only thin-LTO
   Backend = lto::createInProcessThinBackend(
       llvm::heavyweight_hardware_concurrency(1));
+  Conf.UseDefaultPipeline = true;
 
   Conf.CPU = Arch.str();
   Conf.Options = codegen::InitTargetOptionsFromCodeGenFlags(TheTriple);
 
   Conf.MAttrs = getTargetFeatures(TheTriple);
   Conf.CGOptLevel = getCGOptLevel(OptLevel[1] - '0');
   Conf.OptLevel = OptLevel[1] - '0';
   Conf.DefaultTriple = TheTriple.getTriple();

llvm/include/llvm/LTO/Config.h

struct Config {
   CodeGenOpt::Level CGOptLevel = CodeGenOpt::Default;
   CodeGenFileType CGFileType = CGFT_ObjectFile;
   unsigned OptLevel = 2;
   bool DisableVerify = false;
 
   /// Use the new pass manager
   bool UseNewPM = LLVM_ENABLE_NEW_PASS_MANAGER;
 
+  /// Use the standard optimization pipeline.
+  bool UseDefaultPipeline = false;
+
   /// Flag to indicate that the optimizer should not assume builtins are present
   /// on the target.
   bool Freestanding = false;
 
   /// Disable entirely the optimizer, including importing for ThinLTO
   bool CodeGenOnly = false;
 
   /// Run PGO context sensitive IR instrumentation.

llvm/lib/LTO/LTOBackend.cpp

static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
   }
 
   // Parse a custom pipeline if asked to.
   if (!Conf.OptPipeline.empty()) {
     if (auto Err = PB.parsePassPipeline(MPM, Conf.OptPipeline)) {
       report_fatal_error(Twine("unable to parse pass pipeline description '") +
                          Conf.OptPipeline + "': " + toString(std::move(Err)));
     }
+  } else if (Conf.UseDefaultPipeline) {
+    MPM.addPass(PB.buildPerModuleDefaultPipeline(OL));
   } else if (IsThinLTO) {
     MPM.addPass(PB.buildThinLTODefaultPipeline(OL, ImportSummary));
   } else {
     MPM.addPass(PB.buildLTODefaultPipeline(OL, ExportSummary));
   }
 
   if (!Conf.DisableVerify)
     MPM.addPass(VerifierPass());