[InstCombine] Iterative replacement in PtrReplacer

This patch enhances the PtrReplacer as follows:

1. Users are now collected iteratively to be
generous on the stack. In the case of PHIs with
incoming values which have not yet been visited,
they are pushed back into the stack for
reconsideration.

2. Replace users of the pointer root in a
reverse-postorder traversal, instead of a simple
traversal over the collected users. This reordering
ensures that the uses of an instruction are
replaced before replacing the instruction itself.

3. During the replacement of PHI, use the same
incoming value if it does not have a replacement.

This patch specifically fixes the case when an
incoming value of a PHI is addrspacecasted.

Author: gandhi56

llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

@@ -244,11 +244,10 @@ class PointerReplacer {
   void replacePointer(Value *V);
 
 private:
-  bool collectUsersRecursive(Instruction &I);
   void replace(Instruction *I);
-  Value *getReplacement(Value *I);
+  Value *getReplacement(Value *V) const { return WorkMap.lookup(V); }
   bool isAvailable(Instruction *I) const {
-    return I == &Root || Worklist.contains(I);
+    return I == &Root || UsersToReplace.contains(I);
   }
 
   bool isEqualOrValidAddrSpaceCast(const Instruction *I,
@@ -260,8 +259,7 @@ class PointerReplacer {
     return (FromAS == ToAS) || IC.isValidAddrSpaceCast(FromAS, ToAS);
   }
 
-  SmallPtrSet<Instruction *, 32> ValuesToRevisit;
-  SmallSetVector<Instruction *, 4> Worklist;
+  SmallSetVector<Instruction *, 4> UsersToReplace;
   MapVector<Value *, Value *> WorkMap;
   InstCombinerImpl &IC;
   Instruction &Root;
@@ -270,80 +268,128 @@ class PointerReplacer {
 } // end anonymous namespace
 
 bool PointerReplacer::collectUsers() {
-  if (!collectUsersRecursive(Root))
-    return false;
-
-  // Ensure that all outstanding (indirect) users of I
-  // are inserted into the Worklist. Return false
-  // otherwise.
-  return llvm::set_is_subset(ValuesToRevisit, Worklist);
-}
+  SmallVector<Instruction *> Worklist;
+  SmallSetVector<Instruction *, 4> ValuesToRevisit;
+
+  auto PushUsersToWorklist = [&](Instruction *Inst) {
+    for (auto *U : Inst->users()) {
+      if (auto *I = dyn_cast<Instruction>(U)) {
+        if (!isAvailable(I) && !ValuesToRevisit.contains(I))
+          Worklist.emplace_back(I);
+      }
+    }
+  };
 
-bool PointerReplacer::collectUsersRecursive(Instruction &I) {
-  for (auto *U : I.users()) {
-    auto *Inst = cast<Instruction>(&*U);
+  PushUsersToWorklist(&Root);
+  while (!Worklist.empty()) {
+    Instruction *Inst = Worklist.pop_back_val();
+    if (!Inst)
+      return false;
+    if (isAvailable(Inst))
+      continue;
     if (auto *Load = dyn_cast<LoadInst>(Inst)) {
       if (Load->isVolatile())
         return false;
-      Worklist.insert(Load);
+      UsersToReplace.insert(Load);
     } else if (auto *PHI = dyn_cast<PHINode>(Inst)) {
       // All incoming values must be instructions for replacability
       if (any_of(PHI->incoming_values(),
                  [](Value *V) { return !isa<Instruction>(V); }))
         return false;
 
-      // If at least one incoming value of the PHI is not in Worklist,
-      // store the PHI for revisiting and skip this iteration of the
-      // loop.
-      if (any_of(PHI->incoming_values(), [this](Value *V) {
-            return !isAvailable(cast<Instruction>(V));
-          })) {
-        ValuesToRevisit.insert(Inst);
+      // If all incoming values are available, mark this PHI as
+      // replacable and push it's users into the worklist.
+      if (all_of(PHI->incoming_values(),
+                 [&](Value *V) { return isAvailable(cast<Instruction>(V)); })) {
+        UsersToReplace.insert(PHI);
+        PushUsersToWorklist(PHI);
         continue;
       }
 
-      Worklist.insert(PHI);
-      if (!collectUsersRecursive(*PHI))
-        return false;
-    } else if (auto *SI = dyn_cast<SelectInst>(Inst)) {
-      if (!isa<Instruction>(SI->getTrueValue()) ||
-          !isa<Instruction>(SI->getFalseValue()))
+      // Not all incoming values are available. If this PHI was already
+      // visited prior to this iteration, return false.
+      if (!ValuesToRevisit.insert(PHI))
         return false;
 
-      if (!isAvailable(cast<Instruction>(SI->getTrueValue())) ||
-          !isAvailable(cast<Instruction>(SI->getFalseValue()))) {
-        ValuesToRevisit.insert(Inst);
-        continue;
+      // Push PHI back into the stack, followed by unavailable
+      // incoming values.
+      Worklist.emplace_back(PHI);
+      for (unsigned Idx = 0; Idx < PHI->getNumIncomingValues(); ++Idx) {
+        auto *IncomingValue = cast<Instruction>(PHI->getIncomingValue(Idx));
+        if (UsersToReplace.contains(IncomingValue))
+          continue;
+        if (!ValuesToRevisit.insert(IncomingValue))
+          return false;
+        Worklist.emplace_back(IncomingValue);
       }
-      Worklist.insert(SI);
-      if (!collectUsersRecursive(*SI))
-        return false;
-    } else if (isa<GetElementPtrInst>(Inst)) {
-      Worklist.insert(Inst);
-      if (!collectUsersRecursive(*Inst))
+    } else if (auto *SI = dyn_cast<SelectInst>(Inst)) {
+      if (isa<Constant>(SI->getTrueValue()) &&
+          isa<Constant>(SI->getFalseValue()))
+        continue;
+
+      auto *TrueInst = dyn_cast<Instruction>(SI->getTrueValue());
+      auto *FalseInst = dyn_cast<Instruction>(SI->getFalseValue());
+      if (!TrueInst || !FalseInst)
         return false;
+
+      UsersToReplace.insert(SI);
+      PushUsersToWorklist(SI);
+    } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
+      UsersToReplace.insert(GEP);
+      PushUsersToWorklist(GEP);
     } else if (auto *MI = dyn_cast<MemTransferInst>(Inst)) {
       if (MI->isVolatile())
         return false;
-      Worklist.insert(Inst);
+      UsersToReplace.insert(Inst);
     } else if (isEqualOrValidAddrSpaceCast(Inst, FromAS)) {
-      Worklist.insert(Inst);
-      if (!collectUsersRecursive(*Inst))
-        return false;
+      UsersToReplace.insert(Inst);
+      PushUsersToWorklist(Inst);
     } else if (Inst->isLifetimeStartOrEnd()) {
       continue;
     } else {
       // TODO: For arbitrary uses with address space mismatches, should we check
       // if we can introduce a valid addrspacecast?
-      LLVM_DEBUG(dbgs() << "Cannot handle pointer user: " << *U << '\n');
+      LLVM_DEBUG(dbgs() << "Cannot handle pointer user: " << *Inst << '\n');
       return false;
     }
   }
 
-  return true;
+  return llvm::set_is_subset(ValuesToRevisit, UsersToReplace);
 }
 
-Value *PointerReplacer::getReplacement(Value *V) { return WorkMap.lookup(V); }
+void PointerReplacer::replacePointer(Value *V) {
+  assert(cast<PointerType>(Root.getType()) != cast<PointerType>(V->getType()) &&
+         "Invalid usage");
+  WorkMap[&Root] = V;
+  SmallVector<Instruction *> Worklist;
+  SetVector<Instruction *> PostOrderWorklist;
+  SmallPtrSet<Instruction *, 4> Visited;
+
+  // Perform a postorder traversal of the users of Root.
+  Worklist.push_back(&Root);
+  while (!Worklist.empty()) {
+    Instruction *I = Worklist.back();
+
+    // If I has not been processed before, push each of its
+    // replacable users into the worklist.
+    if (Visited.insert(I).second) {
+      for (auto *U : I->users()) {
+        auto *UserInst = cast<Instruction>(U);
+        if (UsersToReplace.contains(UserInst))
+          Worklist.push_back(UserInst);
+      }
+      // Otherwise, users of I have already been pushed into
+      // the PostOrderWorklist. Push I as well.
+    } else {
+      PostOrderWorklist.insert(I);
+      Worklist.pop_back();
+    }
+  }
+
+  // Replace pointers in reverse-postorder.
+  for (Instruction *I : reverse(PostOrderWorklist))
+    replace(I);
+}
 
 void PointerReplacer::replace(Instruction *I) {
   if (getReplacement(I))
@@ -365,12 +411,20 @@ void PointerReplacer::replace(Instruction *I) {
     // replacement (new value).
     WorkMap[NewI] = NewI;
   } else if (auto *PHI = dyn_cast<PHINode>(I)) {
-    Type *NewTy = getReplacement(PHI->getIncomingValue(0))->getType();
-    auto *NewPHI = PHINode::Create(NewTy, PHI->getNumIncomingValues(),
-                                   PHI->getName(), PHI->getIterator());
+    // Create a new PHI by replacing any incoming value that is a user of the
+    // root pointer and has a replacement.
+    SmallVector<Value *, 4> IncomingValues;
+    for (unsigned int I = 0; I < PHI->getNumIncomingValues(); ++I) {
+      Value *V = getReplacement(PHI->getIncomingValue(I));
+      if (!V)
+        V = PHI->getIncomingValue(I);
+      IncomingValues.push_back(V);
+    }
+    auto *NewPHI = PHINode::Create(IncomingValues[0]->getType(),
+                                   PHI->getNumIncomingValues(),
+                                   PHI->getName() + ".r", PHI->getIterator());
     for (unsigned int I = 0; I < PHI->getNumIncomingValues(); ++I)
-      NewPHI->addIncoming(getReplacement(PHI->getIncomingValue(I)),
-                          PHI->getIncomingBlock(I));
+      NewPHI->addIncoming(IncomingValues[I], PHI->getIncomingBlock(I));
     WorkMap[PHI] = NewPHI;
   } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
     auto *V = getReplacement(GEP->getPointerOperand());
@@ -435,18 +489,6 @@ void PointerReplacer::replace(Instruction *I) {
   }
 }
 
-void PointerReplacer::replacePointer(Value *V) {
-#ifndef NDEBUG
-  auto *PT = cast<PointerType>(Root.getType());
-  auto *NT = cast<PointerType>(V->getType());
-  assert(PT != NT && "Invalid usage");
-#endif
-  WorkMap[&Root] = V;
-
-  for (Instruction *Workitem : Worklist)
-    replace(Workitem);
-}
-
 Instruction *InstCombinerImpl::visitAllocaInst(AllocaInst &AI) {
   if (auto *I = simplifyAllocaArraySize(*this, AI, DT))
     return I;

llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll

@@ -4,15 +4,16 @@
 %struct.type = type { [256 x <2 x i64>] }
 @g1 = external hidden addrspace(3) global %struct.type, align 16
 
+; This test requires the PtrReplacer to replace users in an RPO traversal.
+; Furthermore, %ptr.else need not to be replaced so it must be retained in
+; %ptr.sink.
 define <2 x i64> @func(ptr addrspace(4) byref(%struct.type) align 16 %0, i1 %cmp.0) {
 ; CHECK-LABEL: define <2 x i64> @func(
 ; CHECK-SAME: ptr addrspace(4) byref([[STRUCT_TYPE:%.*]]) align 16 [[TMP0:%.*]], i1 [[CMP_0:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*:]]
-; CHECK-NEXT:    [[COERCE:%.*]] = alloca [[STRUCT_TYPE]], align 16, addrspace(5)
-; CHECK-NEXT:    call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) noundef align 16 dereferenceable(4096) [[COERCE]], ptr addrspace(4) noundef align 16 dereferenceable(4096) [[TMP0]], i64 4096, i1 false)
 ; CHECK-NEXT:    br i1 [[CMP_0]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
 ; CHECK:       [[IF_THEN]]:
-; CHECK-NEXT:    [[VAL_THEN:%.*]] = addrspacecast ptr addrspace(5) [[COERCE]] to ptr
+; CHECK-NEXT:    [[VAL_THEN:%.*]] = addrspacecast ptr addrspace(4) [[TMP0]] to ptr
 ; CHECK-NEXT:    br label %[[SINK:.*]]
 ; CHECK:       [[IF_ELSE]]:
 ; CHECK-NEXT:    [[PTR_ELSE:%.*]] = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 16
@@ -43,5 +44,36 @@ sink:
   ret <2 x i64> %val.sink
 }
 
-; Function Attrs: nocallback nofree nounwind willreturn memory(argmem: readwrite)
+define <2 x i64> @func_phi_loop(ptr addrspace(4) byref(%struct.type) align 16 %0, i1 %cmp.0) {
+; CHECK-LABEL: define <2 x i64> @func_phi_loop(
+; CHECK-SAME: ptr addrspace(4) byref([[STRUCT_TYPE:%.*]]) align 16 [[TMP0:%.*]], i1 [[CMP_0:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[VAL_0:%.*]] = addrspacecast ptr addrspace(4) [[TMP0]] to ptr
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[PTR_PHI_R:%.*]] = phi ptr [ [[PTR_1:%.*]], %[[LOOP]] ], [ [[VAL_0]], %[[ENTRY]] ]
+; CHECK-NEXT:    [[PTR_1]] = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 16
+; CHECK-NEXT:    br i1 [[CMP_0]], label %[[LOOP]], label %[[SINK:.*]]
+; CHECK:       [[SINK]]:
+; CHECK-NEXT:    [[VAL_SINK:%.*]] = load <2 x i64>, ptr [[PTR_PHI_R]], align 16
+; CHECK-NEXT:    ret <2 x i64> [[VAL_SINK]]
+;
+entry:
+  %coerce = alloca %struct.type, align 16, addrspace(5)
+  call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) align 16 %coerce, ptr addrspace(4) align 16 %0, i64 4096, i1 false)
+  %ptr.0 = getelementptr inbounds i8, ptr addrspace(5) %coerce, i64 0
+  %val.0 = addrspacecast ptr addrspace(5) %ptr.0 to ptr
+  br label %loop
+
+loop:
+  %ptr.phi = phi ptr [ %val.1, %loop ], [ %val.0, %entry ]
+  %ptr.1 = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 16
+  %val.1 = getelementptr inbounds nuw i8, ptr %ptr.1, i64 0
+  br i1 %cmp.0, label %loop, label %sink
+
+sink:
+  %val.sink = load <2 x i64>, ptr %ptr.phi, align 16
+  ret <2 x i64> %val.sink
+}
+
 declare void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) noalias writeonly captures(none), ptr addrspace(4) noalias readonly captures(none), i64, i1 immarg) #0