[llvm] [clang] [clang][llvm][fatlto] Avoid cloning modules in FatLTO (PR #72180)

[Nikita Popov via cfe-commits]

================
@@ -1530,14 +1530,11 @@ PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level,
 }
 
 ModulePassManager
-PassBuilder::buildFatLTODefaultPipeline(OptimizationLevel Level, bool ThinLTO,
-                                        bool EmitSummary) {
+PassBuilder::buildFatLTODefaultPipeline(OptimizationLevel Level) {
   ModulePassManager MPM;
-  MPM.addPass(EmbedBitcodePass(ThinLTO, EmitSummary,
-                               ThinLTO
-                                   ? buildThinLTOPreLinkDefaultPipeline(Level)
-                                   : buildLTOPreLinkDefaultPipeline(Level)));
-  MPM.addPass(buildPerModuleDefaultPipeline(Level));
+  MPM.addPass(buildThinLTOPreLinkDefaultPipeline(Level));
+  MPM.addPass(EmbedBitcodePass());
+  MPM.addPass(buildThinLTODefaultPipeline(Level, /*ImportSummary=*/nullptr));
----------------
nikic wrote:

Yes, ModuleOptimization (plus I guess a call to addAnnotationRemarksPass) is what I would substitute here, and then add a note to the FatLTO docs that says something alone the lines of:

> If FatLTO is used together with SamplePGO (as opposed to normal PGO), some profile-based optimizations will only be applied when linking with LTO.

I believe the other discrepancies that we discussed at the time have already been addressed in the meantime and SamplePGO is the only remaining issue.

https://github.com/llvm/llvm-project/pull/72180