Automerge: [DropAssumes] Drop dereferenceable assumptions after vectorization. (#166947)

This patch adds another run of DropUnnecessaryAssumes after
vectorization, to clean up assumes that are not longer needed after this
point.

The main example of such an assume is currently dereferenceable
assumptions. This complements
llvm/llvm-project#166945, which avoids sinking
code if it would mean remove a dereferenceable assumption.

There are a few additional cases where some unneeded assumes are left
over after vectorization that also get cleaned up.

The main motivation is to work together with
llvm/llvm-project#166945, but there may be a
better solution.

Adding another instance of this pass to the pipeline is not great, but
compile-time impact seems in the noise:
https://llvm-compile-time-tracker.com/compare.php?from=55e71fe08b6406ec7ce2c81ce042e48717acf204&to=85da4ee3a74126f557cdc74c7b40e048dacb3fc4&stat=instructions:u

PR: llvm/llvm-project#166947

Author: fhahn

llvm/include/llvm/Passes/PassBuilder.h

@@ -742,7 +742,7 @@ class PassBuilder {
   void addRequiredLTOPreLinkPasses(ModulePassManager &MPM);
 
   void addVectorPasses(OptimizationLevel Level, FunctionPassManager &FPM,
-                       bool IsFullLTO);
+                       ThinOrFullLTOPhase LTOPhase);
 
   static std::optional<std::vector<PipelineElement>>
   parsePipelineText(StringRef Text);

llvm/include/llvm/Transforms/Scalar/DropUnnecessaryAssumes.h

@@ -19,7 +19,13 @@ namespace llvm {
 
 struct DropUnnecessaryAssumesPass
     : public PassInfoMixin<DropUnnecessaryAssumesPass> {
+  DropUnnecessaryAssumesPass(bool DropDereferenceable = false)
+      : DropDereferenceable(DropDereferenceable) {}
+
   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+
+private:
+  bool DropDereferenceable;
 };
 
 } // end namespace llvm

llvm/lib/Passes/PassBuilder.cpp

@@ -900,6 +900,11 @@ Expected<bool> parseEntryExitInstrumenterPassOptions(StringRef Params) {
                                             "EntryExitInstrumenter");
 }
 
+Expected<bool> parseDropUnnecessaryAssumesPassOptions(StringRef Params) {
+  return PassBuilder::parseSinglePassOption(Params, "drop-deref",
+                                            "DropUnnecessaryAssumes");
+}
+
 Expected<bool> parseLoopExtractorPassOptions(StringRef Params) {
   return PassBuilder::parseSinglePassOption(Params, "single", "LoopExtractor");
 }

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -1303,10 +1303,18 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
 
 /// TODO: Should LTO cause any differences to this set of passes?
 void PassBuilder::addVectorPasses(OptimizationLevel Level,
-                                  FunctionPassManager &FPM, bool IsFullLTO) {
+                                  FunctionPassManager &FPM,
+                                  ThinOrFullLTOPhase LTOPhase) {
+  const bool IsFullLTO = LTOPhase == ThinOrFullLTOPhase::FullLTOPostLink;
+
   FPM.addPass(LoopVectorizePass(
       LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization)));
 
+  // Drop dereferenceable assumes after vectorization, as they are no longer
+  // needed and can inhibit further optimization.
+  if (!isLTOPreLink(LTOPhase))
+    FPM.addPass(DropUnnecessaryAssumesPass(/*DropDereferenceable=*/true));
+
   FPM.addPass(InferAlignmentPass());
   if (IsFullLTO) {
     // The vectorizer may have significantly shortened a loop body; unroll
@@ -1577,7 +1585,7 @@ PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
   // from the TargetLibraryInfo.
   OptimizePM.addPass(InjectTLIMappings());
 
-  addVectorPasses(Level, OptimizePM, /* IsFullLTO */ false);
+  addVectorPasses(Level, OptimizePM, LTOPhase);
 
   invokeVectorizerEndEPCallbacks(OptimizePM, Level);
 
@@ -2179,7 +2187,7 @@ PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
 
   MainFPM.addPass(LoopDistributePass());
 
-  addVectorPasses(Level, MainFPM, /* IsFullLTO */ true);
+  addVectorPasses(Level, MainFPM, ThinOrFullLTOPhase::FullLTOPostLink);
 
   invokeVectorizerEndEPCallbacks(MainFPM, Level);
 

llvm/lib/Passes/PassRegistry.def

@@ -432,7 +432,6 @@ FUNCTION_PASS("dot-post-dom", PostDomPrinter())
 FUNCTION_PASS("dot-post-dom-only", PostDomOnlyPrinter())
 FUNCTION_PASS("dse", DSEPass())
 FUNCTION_PASS("dwarf-eh-prepare", DwarfEHPreparePass(*TM))
-FUNCTION_PASS("drop-unnecessary-assumes", DropUnnecessaryAssumesPass())
 FUNCTION_PASS("expand-large-div-rem", ExpandLargeDivRemPass(*TM))
 FUNCTION_PASS("expand-memcmp", ExpandMemCmpPass(*TM))
 FUNCTION_PASS("expand-reductions", ExpandReductionsPass())
@@ -584,6 +583,10 @@ FUNCTION_PASS_WITH_PARAMS(
     "early-cse", "EarlyCSEPass",
     [](bool UseMemorySSA) { return EarlyCSEPass(UseMemorySSA); },
     parseEarlyCSEPassOptions, "memssa")
+FUNCTION_PASS_WITH_PARAMS(
+    "drop-unnecessary-assumes", "DropUnnecessaryAssumesPass",
+    [](bool DropDereferenceable) { return DropUnnecessaryAssumesPass(DropDereferenceable); },
+    parseDropUnnecessaryAssumesPassOptions, "drop-deref")
 FUNCTION_PASS_WITH_PARAMS(
     "ee-instrument", "EntryExitInstrumenterPass",
     [](bool PostInlining) { return EntryExitInstrumenterPass(PostInlining); },

llvm/lib/Transforms/Scalar/DropUnnecessaryAssumes.cpp

@@ -78,11 +78,16 @@ DropUnnecessaryAssumesPass::run(Function &F, FunctionAnalysisManager &FAM) {
       SmallVector<OperandBundleDef> KeptBundles;
       unsigned NumBundles = Assume->getNumOperandBundles();
       for (unsigned I = 0; I != NumBundles; ++I) {
-        auto IsDead = [](OperandBundleUse Bundle) {
+        auto IsDead = [&](OperandBundleUse Bundle) {
           // "ignore" operand bundles are always dead.
           if (Bundle.getTagName() == "ignore")
             return true;
 
+          // "dereferenceable" operand bundles are only dropped if requested
+          // (e.g., after loop vectorization has run).
+          if (Bundle.getTagName() == "dereferenceable")
+            return DropDereferenceable;
+
           // Bundles without arguments do not affect any specific values.
           // Always keep them for now.
           if (Bundle.Inputs.empty())

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

@@ -261,6 +261,7 @@
 ; CHECK-O-NEXT: Running analysis: LoopAccessAnalysis on foo
 ; CHECK-O-NEXT: Running pass: InjectTLIMappings
 ; CHECK-O-NEXT: Running pass: LoopVectorizePass
+; CHECK-DEFAULT-NEXT: Running pass: DropUnnecessaryAssumesPass
 ; CHECK-O-NEXT: Running pass: InferAlignmentPass
 ; CHECK-O-NEXT: Running pass: LoopLoadEliminationPass
 ; CHECK-O-NEXT: Running pass: InstCombinePass

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

@@ -129,6 +129,7 @@
 ; CHECK-O23SZ-NEXT: Running analysis: LoopAccessAnalysis on foo
 ; CHECK-O23SZ-NEXT: Running pass: LoopVectorizePass on foo
 ; CHECK-O23SZ-NEXT: Running analysis: DemandedBitsAnalysis on foo
+; CHECK-O23SZ-NEXT: Running pass: DropUnnecessaryAssumesPass on foo
 ; CHECK-O23SZ-NEXT: Running pass: InferAlignmentPass on foo
 ; CHECK-O23SZ-NEXT: Running pass: LoopUnrollPass on foo
 ; CHECK-O23SZ-NEXT: WarnMissedTransformationsPass on foo

llvm/test/Other/new-pm-thinlto-postlink-defaults.ll

@@ -180,6 +180,7 @@
 ; CHECK-POSTLINK-O-NEXT: Running analysis: LoopAccessAnalysis on foo
 ; CHECK-POSTLINK-O-NEXT: Running pass: InjectTLIMappings
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopVectorizePass
+; CHECK-POSTLINK-O-NEXT: Running pass: DropUnnecessaryAssumesPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: InferAlignmentPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: LoopLoadEliminationPass
 ; CHECK-POSTLINK-O-NEXT: Running pass: InstCombinePass

llvm/test/Other/new-pm-thinlto-postlink-pgo-defaults.ll

@@ -165,6 +165,7 @@
 ; CHECK-O-NEXT: Running analysis: LoopAccessAnalysis on foo
 ; CHECK-O-NEXT: Running pass: InjectTLIMappings
 ; CHECK-O-NEXT: Running pass: LoopVectorizePass
+; CHECK-O-NEXT: Running pass: DropUnnecessaryAssumesPass
 ; CHECK-O-NEXT: Running pass: InferAlignmentPass
 ; CHECK-O-NEXT: Running pass: LoopLoadEliminationPass
 ; CHECK-O-NEXT: Running pass: InstCombinePass