[CallGraphUpdater] Properly remove strongly connected components (old PM)

[jdoerfert]

Summary

The old code did eliminate references from and to functions that were
about to be deleted only just before we deleted them. This can cause
references from other functions that are supposed to be deleted to still
exist, depending on the order. If the functions form a strongly
connected component the problem manifests regardless of the order in
which we try to actually delete the functions.

This patch introduces a two step deletion. First we remove all
references and then we delete the function. Note that this only affects
the old call graph. There should not be any functional changes if no old
style call graph was given.

To test this we delete two strongly connected functions instead of one
in an existing test.

[jdoerfert]

I accidentally influenced an existing test, will address this.

[jdoerfert]

Avoid affecting other tests

[hfinkel]

LGTM

llvm/lib/Transforms/Utils/CallGraphUpdater.cpp

 bool CallGraphUpdater::finalize() {
   if (!DeadFunctionsInComdats.empty()) {
     filterDeadComdatFunctions(*DeadFunctionsInComdats.front()->getParent(),
                               DeadFunctionsInComdats);
     DeadFunctions.append(DeadFunctionsInComdats.begin(),
                          DeadFunctionsInComdats.end());
   }
 
+  if (CG) {
+    // First remove all references, e.g., outgoing via called functions. This is
+    // necessary as we can delete functions that have circular references.
     for (Function *DeadFn : DeadFunctions) {
       DeadFn->removeDeadConstantUsers();
-
-    if (CG) {
-      CallGraphNode *OldCGN = CG->getOrInsertFunction(DeadFn);
-      CG->getExternalCallingNode()->removeAnyCallEdgeTo(OldCGN);
-      OldCGN->removeAllCalledFunctions();
+      CallGraphNode *DeadCGN = (*CG)[DeadFn];
+      DeadCGN->removeAllCalledFunctions();
+      CG->getExternalCallingNode()->removeAnyCallEdgeTo(DeadCGN);
       DeadFn->replaceAllUsesWith(UndefValue::get(DeadFn->getType()));
-
-      assert(OldCGN->getNumReferences() == 0);
-
-      delete CG->removeFunctionFromModule(OldCGN);
-      continue;
     }
 
-    // The old style call graph (CG) has a value handle we do not want to
-    // replace with undef so we do this here.
+    // Then remove the node and function from the module.
+    for (Function *DeadFn : DeadFunctions) {
+      CallGraphNode *DeadCGN = CG->getOrInsertFunction(DeadFn);
+      assert(DeadCGN->getNumReferences() == 0 &&
+             "References should have been handled by now");
+      delete CG->removeFunctionFromModule(DeadCGN);
+    }
+  } else {
+    // This is the code path for the new lazy call graph and for the case were
+    // no call graph was provided.
+    for (Function *DeadFn : DeadFunctions) {
+      DeadFn->removeDeadConstantUsers();
       DeadFn->replaceAllUsesWith(UndefValue::get(DeadFn->getType()));
 
       if (LCG && !ReplacedFunctions.count(DeadFn)) {
         // Taken mostly from the inliner:
         LazyCallGraph::Node &N = LCG->get(*DeadFn);
         auto *DeadSCC = LCG->lookupSCC(N);
         assert(DeadSCC && DeadSCC->size() == 1 &&
                &DeadSCC->begin()->getFunction() == DeadFn);
         auto &DeadRC = DeadSCC->getOuterRefSCC();
 
         FunctionAnalysisManager &FAM =
             AM->getResult<FunctionAnalysisManagerCGSCCProxy>(*DeadSCC, *LCG)
                 .getManager();
 
         FAM.clear(*DeadFn, DeadFn->getName());
         AM->clear(*DeadSCC, DeadSCC->getName());
         LCG->removeDeadFunction(*DeadFn);
 
-      // Mark the relevant parts of the call graph as invalid so we don't visit
-      // them.
+        // Mark the relevant parts of the call graph as invalid so we don't
+        // visit them.
         UR->InvalidatedSCCs.insert(DeadSCC);
         UR->InvalidatedRefSCCs.insert(&DeadRC);
       }
 
       // The function is now really dead and de-attached from everything.
       DeadFn->eraseFromParent();
     }
+  }
 
   bool Changed = !DeadFunctions.empty();
   DeadFunctionsInComdats.clear();
   DeadFunctions.clear();
   return Changed;
 }
 
 void CallGraphUpdater::reanalyzeFunction(Function &Fn) {

llvm/unittests/IR/LegacyPassManagerTest.cpp

Module *makeLLVMModule(LLVMContext &Context) {
         BranchInst::Create(label_bb1, label_return, int1_f, label_bb1);
 
         // Block return (label_return)
         ReturnInst::Create(Context, label_return);
       }
       return mod;
     }
 
+    /// Split a simple function which contains only a call and a return into two
+    /// such that the first calls the second and the second whoever was called
+    /// initially.
+    Function *splitSimpleFunction(Function &F) {
+      LLVMContext &Context = F.getContext();
+      Function *SF = Function::Create(F.getFunctionType(), F.getLinkage(),
+                                      F.getName() + "b", F.getParent());
+      F.setName(F.getName() + "a");
+      BasicBlock *Entry = BasicBlock::Create(Context, "entry", SF, nullptr);
+      CallInst &CI = cast<CallInst>(F.getEntryBlock().front());
+      CI.clone()->insertBefore(ReturnInst::Create(Context, Entry));
+      CI.setCalledFunction(SF);
+      return SF;
+    }
+
     struct CGModifierPass : public CGPass {
       unsigned NumSCCs = 0;
       unsigned NumFns = 0;
       bool SetupWorked = true;
 
       CallGraphUpdater CGU;
 
       bool runOnSCC(CallGraphSCC &SCMM) override {
         ++NumSCCs;
         for (CallGraphNode *N : SCMM)
           if (N->getFunction())
             ++NumFns;
 
         CGPass::run();
 
         if (SCMM.size() <= 1)
           return false;
 
         CallGraphNode *N = *(SCMM.begin());
         Function *F = N->getFunction();
         Module *M = F->getParent();
         Function *Test1F = M->getFunction("test1");
-        Function *Test2F = M->getFunction("test2");
+        Function *Test2aF = M->getFunction("test2a");
+        Function *Test2bF = M->getFunction("test2b");
         Function *Test3F = M->getFunction("test3");
         auto InSCC = [&](Function *Fn) {
           return llvm::any_of(SCMM, [Fn](CallGraphNode *CGN) {
             return CGN->getFunction() == Fn;
           });
         };
 
-        if (!Test1F || !Test2F || !Test3F || !InSCC(Test1F) || !InSCC(Test2F) ||
-            !InSCC(Test3F))
+        if (!Test1F || !Test2aF || !Test2bF || !Test3F || !InSCC(Test1F) ||
+            !InSCC(Test2aF) || !InSCC(Test2bF) || !InSCC(Test3F))
           return SetupWorked = false;
 
         CallInst *CI = dyn_cast<CallInst>(&Test1F->getEntryBlock().front());
-        if (!CI || CI->getCalledFunction() != Test2F)
+        if (!CI || CI->getCalledFunction() != Test2aF)
           return SetupWorked = false;
 
         CI->setCalledFunction(Test3F);
 
         CGU.initialize(const_cast<CallGraph &>(SCMM.getCallGraph()), SCMM);
-        CGU.removeFunction(*Test2F);
+        CGU.removeFunction(*Test2aF);
+        CGU.removeFunction(*Test2bF);
         CGU.reanalyzeFunction(*Test1F);
         return true;
       }
 
       bool doFinalization(CallGraph &CG) override { return CGU.finalize(); }
     };
 
     TEST(PassManager, CallGraphUpdater0) {
-      // SCC#1: test1->test2->test3->test1
+      // SCC#1: test1->test2a->test2b->test3->test1
       // SCC#2: test4
       // SCC#3: indirect call node
 
       LLVMContext Context;
       std::unique_ptr<Module> M(makeLLVMModule(Context));
       ASSERT_EQ(M->getFunctionList().size(), 4U);
+      Function *F = M->getFunction("test2");
+      Function *SF = splitSimpleFunction(*F);
+      CallInst::Create(F, "", &SF->getEntryBlock());
+      ASSERT_EQ(M->getFunctionList().size(), 5U);
       CGModifierPass *P = new CGModifierPass();
       legacy::PassManager Passes;
       Passes.add(P);
       Passes.run(*M);
       ASSERT_TRUE(P->SetupWorked);
       ASSERT_EQ(P->NumSCCs, 3U);
-      ASSERT_EQ(P->NumFns, 4U);
+      ASSERT_EQ(P->NumFns, 5U);
       ASSERT_EQ(M->getFunctionList().size(), 3U);
     }
   }
 }
 
 INITIALIZE_PASS(ModuleNDM, "mndm", "mndm", false, false)
 INITIALIZE_PASS_BEGIN(CGPass, "cgp","cgp", false, false)
 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
 INITIALIZE_PASS_END(CGPass, "cgp","cgp", false, false)
 INITIALIZE_PASS(FPass, "fp","fp", false, false)
 INITIALIZE_PASS_BEGIN(LPass, "lp","lp", false, false)
 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 INITIALIZE_PASS_END(LPass, "lp","lp", false, false)