[AA] Move Target Specific AA before BasicAA

llvm/include/llvm/Analysis/AliasAnalysis.h

@@ -1013,6 +1013,18 @@ struct ExternalAAWrapperPass : ImmutablePass {
 
   explicit ExternalAAWrapperPass(CallbackT CB);
 
+  /// Returns whether this external AA should run before Basic AA.
+  ///
+  /// By default, external AA passes are run after Basic AA. If this returns
+  /// true, the external AA will be run before Basic AA during alias analysis.
+  ///
+  /// For some targets, we prefer to run the external AA early to improve
+  /// compile time as it has more target-specific information. This is
+  /// particularly useful when the external AA can provide more precise results
+  /// than Basic AA so that Basic AA does not need to spend time recomputing
+  /// them.
+  virtual bool runEarly() { return false; }
+
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesAll();
   }

llvm/include/llvm/Target/TargetMachine.h

@@ -396,6 +396,11 @@ class TargetMachine {
   // TODO: Populate all pass names by using <Target>PassRegistry.def.
   virtual void registerPassBuilderCallbacks(PassBuilder &) {}
 
+  /// Allow the target to register early alias analyses (AA before BasicAA) with
+  /// the AAManager for use with the new pass manager. Only affects the
+  /// "default" AAManager.
+  virtual void registerEarlyDefaultAliasAnalyses(AAManager &) {}
+
   /// Allow the target to register alias analyses with the AAManager for use
   /// with the new pass manager. Only affects the "default" AAManager.
   virtual void registerDefaultAliasAnalyses(AAManager &) {}

llvm/lib/Analysis/AliasAnalysis.cpp

@@ -736,28 +736,49 @@ bool AAResultsWrapperPass::runOnFunction(Function &F) {
   AAR.reset(
       new AAResults(getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F)));
 
+  // Add any target-specific alias analyses that should be run early.
+  auto *ExtWrapperPass = getAnalysisIfAvailable<ExternalAAWrapperPass>();
+  if (ExtWrapperPass && ExtWrapperPass->runEarly() && ExtWrapperPass->CB) {
+    LLVM_DEBUG(dbgs() << "AAResults register Early ExternalAA: "
+                      << ExtWrapperPass->getPassName() << "\n");
+    ExtWrapperPass->CB(*this, F, *AAR);
+  }
+
   // BasicAA is always available for function analyses. Also, we add it first
   // so that it can trump TBAA results when it proves MustAlias.
   // FIXME: TBAA should have an explicit mode to support this and then we
   // should reconsider the ordering here.
-  if (!DisableBasicAA)
+  if (!DisableBasicAA) {
+    LLVM_DEBUG(dbgs() << "AAResults register BasicAA\n");
     AAR->addAAResult(getAnalysis<BasicAAWrapperPass>().getResult());
+  }
 
   // Populate the results with the currently available AAs.
-  if (auto *WrapperPass = getAnalysisIfAvailable<ScopedNoAliasAAWrapperPass>())
+  if (auto *WrapperPass =
+          getAnalysisIfAvailable<ScopedNoAliasAAWrapperPass>()) {
+    LLVM_DEBUG(dbgs() << "AAResults register ScopedNoAliasAA\n");
     AAR->addAAResult(WrapperPass->getResult());
-  if (auto *WrapperPass = getAnalysisIfAvailable<TypeBasedAAWrapperPass>())
+  }
+  if (auto *WrapperPass = getAnalysisIfAvailable<TypeBasedAAWrapperPass>()) {
+    LLVM_DEBUG(dbgs() << "AAResults register TypeBasedAA\n");
     AAR->addAAResult(WrapperPass->getResult());
-  if (auto *WrapperPass = getAnalysisIfAvailable<GlobalsAAWrapperPass>())
+  }
+  if (auto *WrapperPass = getAnalysisIfAvailable<GlobalsAAWrapperPass>()) {
+    LLVM_DEBUG(dbgs() << "AAResults register GlobalsAA\n");
     AAR->addAAResult(WrapperPass->getResult());
-  if (auto *WrapperPass = getAnalysisIfAvailable<SCEVAAWrapperPass>())
+  }
+  if (auto *WrapperPass = getAnalysisIfAvailable<SCEVAAWrapperPass>()) {
+    LLVM_DEBUG(dbgs() << "AAResults register SCEVAA\n");
     AAR->addAAResult(WrapperPass->getResult());
+  }
 
   // If available, run an external AA providing callback over the results as
   // well.
-  if (auto *WrapperPass = getAnalysisIfAvailable<ExternalAAWrapperPass>())
-    if (WrapperPass->CB)
-      WrapperPass->CB(*this, F, *AAR);
+  if (ExtWrapperPass && !ExtWrapperPass->runEarly() && ExtWrapperPass->CB) {
+    LLVM_DEBUG(dbgs() << "AAResults register Late ExternalAA: "
+                      << ExtWrapperPass->getPassName() << "\n");
+    ExtWrapperPass->CB(*this, F, *AAR);
+  }
 
   // Analyses don't mutate the IR, so return false.
   return false;

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -2320,6 +2320,10 @@ AAManager PassBuilder::buildDefaultAAPipeline() {
   // The order in which these are registered determines their priority when
   // being queried.
 
+  // Add any target-specific alias analyses that should be run early.
+  if (TM)
+    TM->registerEarlyDefaultAliasAnalyses(AA);
+
   // First we register the basic alias analysis that provides the majority of
   // per-function local AA logic. This is a stateless, on-demand local set of
   // AA techniques.

llvm/lib/Target/NVPTX/NVPTXAliasAnalysis.h

@@ -85,16 +85,23 @@ class NVPTXAAWrapperPass : public ImmutablePass {
 
 // Wrapper around ExternalAAWrapperPass so that the default
 // constructor gets the callback.
+// Note that NVPTXAA will run before BasicAA for compile time considerations.
 class NVPTXExternalAAWrapper : public ExternalAAWrapperPass {
 public:
   static char ID;
 
+  bool runEarly() override { return true; }
+
   NVPTXExternalAAWrapper()
       : ExternalAAWrapperPass([](Pass &P, Function &, AAResults &AAR) {
           if (auto *WrapperPass =
                   P.getAnalysisIfAvailable<NVPTXAAWrapperPass>())
             AAR.addAAResult(WrapperPass->getResult());
         }) {}
+
+  StringRef getPassName() const override {
+    return "NVPTX Address space based Alias Analysis Wrapper";
+  }
 };
 
 ImmutablePass *createNVPTXAAWrapperPass();

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

@@ -232,7 +232,7 @@ MachineFunctionInfo *NVPTXTargetMachine::createMachineFunctionInfo(
                                                                     F, STI);
 }
 
-void NVPTXTargetMachine::registerDefaultAliasAnalyses(AAManager &AAM) {
+void NVPTXTargetMachine::registerEarlyDefaultAliasAnalyses(AAManager &AAM) {
   AAM.registerFunctionAnalysis<NVPTXAA>();
 }
 
@@ -347,10 +347,7 @@ void NVPTXPassConfig::addIRPasses() {
   disablePass(&RemoveLoadsIntoFakeUsesID);
 
   addPass(createNVPTXAAWrapperPass());
-  addPass(createExternalAAWrapperPass([](Pass &P, Function &, AAResults &AAR) {
-    if (auto *WrapperPass = P.getAnalysisIfAvailable<NVPTXAAWrapperPass>())
-      AAR.addAAResult(WrapperPass->getResult());
-  }));
+  addPass(createNVPTXExternalAAWrapperPass());
 
   // NVVMReflectPass is added in addEarlyAsPossiblePasses, so hopefully running
   // it here does nothing.  But since we need it for correctness when lowering

llvm/lib/Target/NVPTX/NVPTXTargetMachine.h

@@ -64,7 +64,7 @@ class NVPTXTargetMachine : public CodeGenTargetMachineImpl {
   createMachineFunctionInfo(BumpPtrAllocator &Allocator, const Function &F,
                             const TargetSubtargetInfo *STI) const override;
 
-  void registerDefaultAliasAnalyses(AAManager &AAM) override;
+  void registerEarlyDefaultAliasAnalyses(AAManager &AAM) override;
 
   void registerPassBuilderCallbacks(PassBuilder &PB) override;
 

llvm/test/Analysis/NVPTXAA/NVPTXAA_before_BasicAA.ll

@@ -0,0 +1,17 @@
+; REQUIRES: asserts
+; RUN: opt -aa-pipeline=default -passes='require<aa>' -debug-pass-manager -disable-output -S < %s 2>&1 | FileCheck %s
+; RUN: llc --debug-only='aa' -o /dev/null %s 2>&1 | FileCheck %s -check-prefix=LEGACY
+
+; In default AA pipeline, NVPTXAA should run before BasicAA to reduce compile time for NVPTX backend
+target triple = "nvptx64-nvidia-cuda"
+
+; CHECK: Running analysis: NVPTXAA on foo
+; CHECK-NEXT: Running analysis: BasicAA on foo
+
+; LEGACY: AAResults register Early ExternalAA: NVPTX Address space based Alias Analysis Wrapper
+; LEGACY-NEXT: AAResults register BasicAA
+define void @foo(){
+entry:
+  ret void
+}
+