[llvm] [PassBuilder] Add extra ModuleInliner/SROA/SimplifyCFG/LICM to late FullLTO passes (PR #197678)

[Momchil Velikov via llvm-commits]

================
@@ -2250,6 +2250,18 @@ PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
   MPM.addPass(createModuleToFunctionPassAdaptor(std::move(MainFPM),
                                                 PTO.EagerlyInvalidateAnalyses));
 
+  if (EnableModuleInliner) {
+    MPM.addPass(ModuleInlinerPass(::getInlineParamsFromOptLevel(Level),
+                                  UseInlineAdvisor,
+                                  ThinOrFullLTOPhase::FullLTOPostLink));
+  } else {
+    MPM.addPass(ModuleInlinerWrapperPass(
+        ::getInlineParamsFromOptLevel(Level),
+        /* MandatoryFirst */ true,
+        InlineContext{ThinOrFullLTOPhase::FullLTOPostLink,
+                      InlinePass::CGSCCInliner}));
+  }
----------------
momchil-velikov wrote:

No, it is not running after the vectorisation passes when they are disabled in the pre-LTO pipeline.

A few hot functions in `femflow` contain big composites passed as arguments or returned as return values. In the current `main` pipeline pre-link SLP reduces the size of some functions, LTO inliner inlines them, and then SROA triggers, optimising several memset/memcpy calls. Without the pre-link SLP, the LTO inliner does not trigger, and then SROA does not trigger (due to `alloca` addresses escaping to non-inline callees) and then we get 9% regression.

Adding the Inliner + SROA after the LTO SLP recreates the above optimisation sequence, restoring about 8% of the regression.

https://github.com/llvm/llvm-project/pull/197678