[OpenMP] Run an extra 'OpenMPOpt' pass in LTO-mode

[jhuber6]

Summary

The OpenMPOpt pass is pivotal to the performance of many OpenMP
offloading programs. When we perform non-LTO builds with OpenMP we used
to link the OpenMP deviceRTL individually for each TU. This lead to us
getting an additional attributor run on the combined runtime and user
code. When we used LTO we lost a run and suffered a large performance
degradation. This patch simply adds in the extra OpenMPOpt pass that
we miss into the LTO pipeline. This patch fixes the performance
regression shown in applications that used OpenMP offloading in LTO
mode.

Previously, this wasn't legal to do as we could emit new runtime calls
into the module. That was fixed by D142646.

Depends on D142646

Fixes https://github.com/llvm/llvm-project/issues/60300

[jhuber6]

Fix test

[jdoerfert]

LG, one nit.

llvm/lib/Passes/PassBuilderPipelines.cpp

1723

Move it before DCE, maybe even before the GlobalOpt pass.

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

std::unique_ptr<lto::LTO> createLTO(
 
   StringRef OptLevel = Args.getLastArgValue(OPT_opt_level, "O2");
   Conf.MAttrs = Features;
   std::optional<CodeGenOpt::Level> CGOptLevelOrNone =
       CodeGenOpt::parseLevel(OptLevel[1]);
   assert(CGOptLevelOrNone && "Invalid optimization level");
   Conf.CGOptLevel = *CGOptLevelOrNone;
   Conf.OptLevel = OptLevel[1] - '0';
-  if (Conf.OptLevel > 0)
-    Conf.UseDefaultPipeline = true;
   Conf.DefaultTriple = Triple.getTriple();
 
   LTOError = false;
   Conf.DiagHandler = diagnosticHandler;
 
   Conf.PTO.LoopVectorization = Conf.OptLevel > 1;
   Conf.PTO.SLPVectorization = Conf.OptLevel > 1;
 

llvm/lib/Passes/PassBuilderPipelines.cpp

MPM.addPass(ModuleInlinerWrapperPass(
       getInlineParamsFromOptLevel(Level),
       /* MandatoryFirst */ true,
       InlineContext{ThinOrFullLTOPhase::FullLTOPostLink,
                     InlinePass::CGSCCInliner}));
 
   // Optimize globals again after we ran the inliner.
   MPM.addPass(GlobalOptPass());
 
+  // Run the OpenMPOpt pass again after global optimizations.
+  MPM.addPass(OpenMPOptPass(ThinOrFullLTOPhase::FullLTOPostLink));
+
   // Garbage collect dead functions.
   MPM.addPass(GlobalDCEPass());
 
   // If we didn't decide to inline a function, check to see if we can
   // transform it to pass arguments by value instead of by reference.
   MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(ArgumentPromotionPass()));
 
   // Remove unused arguments from functions.
   MPM.addPass(DeadArgumentEliminationPass());
 
   FunctionPassManager FPM;
   // The IPO Passes may leave cruft around. Clean up after them.
   FPM.addPass(InstCombinePass());
   invokePeepholeEPCallbacks(FPM, Level);

llvm/test/Other/new-pm-lto-defaults.ll

 ; CHECK-O3-NEXT: Running pass: AggressiveInstCombinePass
 ; CHECK-EP-Peephole-NEXT: Running pass: NoOpFunctionPass
 ; CHECK-O23SZ-NEXT: Running pass: ModuleInlinerWrapperPass
 ; CHECK-O23SZ-NEXT: Running analysis: InlineAdvisorAnalysis
 ; CHECK-O23SZ-NEXT: Running pass: InlinerPass
 ; CHECK-O23SZ-NEXT: Running pass: InlinerPass
 ; CHECK-O23SZ-NEXT: Invalidating analysis: InlineAdvisorAnalysis
 ; CHECK-O23SZ-NEXT: Running pass: GlobalOptPass
+; CHECK-O23SZ-NEXT: Running pass: OpenMPOptPass
 ; CHECK-O23SZ-NEXT: Running pass: GlobalDCEPass
 ; CHECK-O23SZ-NEXT: Running pass: ArgumentPromotionPass
 ; CHECK-O23SZ-NEXT: Running pass: DeadArgumentEliminationPass
 ; CHECK-O23SZ-NEXT: Running pass: InstCombinePass
 ; CHECK-EP-Peephole-NEXT: Running pass: NoOpFunctionPass
 ; CHECK-O23SZ-NEXT: Running pass: JumpThreadingPass
 ; CHECK-O23SZ-NEXT: Running analysis: LazyValueAnalysis
 ; CHECK-O23SZ-NEXT: Running pass: SROAPass on foo