[opaque pointer types] Alloca: use getAllocatedType() instead of getType()->getElementType().

[mjacob]

lib/Transforms/Scalar/ScalarReplAggregates.cpp

2237 ↗

(On Diff #45111)

Are you sure this is right? I don't think we can assume that AI and OtherPtr have the same type.

2254 ↗

(On Diff #45111)

Why don't you reuse EltPtr?

2486 ↗

(On Diff #45111)

Why don't you reuse SrcField?

[eddyb]

lib/Transforms/Scalar/ScalarReplAggregates.cpp

2237 ↗

(On Diff #45111)

OtherPtr is ensured above to have the type PointerType::get(AI->getAllocatedType(), AddrSpace).

2254 ↗

(On Diff #45111)

Wrong type, NewElts is an array of AllocaInst*, and I believe EltPtr is assigned something that is not an AllocaInst* (or used to, looking again at it doesn't seem to be the case atm).

2486 ↗

(On Diff #45111)

See previous comment. In this case, SrcField *does* get assigned, so I can't change its type to AllocaInst*.

[dblaikie]

Thanks for all the work/patches posted here. Will be getting to the reviews as quick as I can.

This one looks good, please commit!

lib/Transforms/Scalar/ScalarReplAggregates.cpp

2237 ↗

(On Diff #45111)

Could throw in an assert to that effect if you like

2254 ↗

(On Diff #45111)

Feel free to correct the type (since, as you say, it's no longer reused for other things) & reuse it if you like

[eddyb]

Reuse EltPtr variable, with a more precise type.

[eddyb]

Assert that OtherPtr still has the same pointer type as the alloca.

[mjacob]

LGTM.

llvm/trunk/lib/Analysis/ValueTracking.cpp

case Instruction::URem: {
     KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
     break;
   }
 
   case Instruction::Alloca: {
     AllocaInst *AI = cast<AllocaInst>(I);
     unsigned Align = AI->getAlignment();
     if (Align == 0)
-      Align = Q.DL.getABITypeAlignment(AI->getType()->getElementType());
+      Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
 
     if (Align > 0)
       KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
     break;
   }
   case Instruction::GetElementPtr: {
     // Analyze all of the subscripts of this getelementptr instruction
     // to determine if we can prove known low zero bits.

llvm/trunk/lib/CodeGen/GCRootLowering.cpp

if (StoreInst *SI = dyn_cast<StoreInst>(IP))
         InitedRoots.insert(AI);
 
   // Add root initializers.
   bool MadeChange = false;
 
   for (AllocaInst **I = Roots, **E = Roots + Count; I != E; ++I)
     if (!InitedRoots.count(*I)) {
       StoreInst *SI = new StoreInst(
-          ConstantPointerNull::get(cast<PointerType>(
-              cast<PointerType>((*I)->getType())->getElementType())),
+          ConstantPointerNull::get(cast<PointerType>((*I)->getAllocatedType())),
           *I);
       SI->insertAfter(*I);
       MadeChange = true;
     }
 
   return MadeChange;
 }
 

llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp

if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
       }
 
       DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
     } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
       if (AI->isArrayAllocation()) {
         DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
         return false;  // Cannot handle array allocs.
       }
-      Type *Ty = AI->getType()->getElementType();
+      Type *Ty = AI->getAllocatedType();
       AllocaTmps.push_back(
           make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
                                       UndefValue::get(Ty), AI->getName()));
       InstResult = AllocaTmps.back().get();
       DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
     } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
       CallSite CS(&*CurInst);
 

llvm/trunk/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

if (ClobberNonLive) {
         if (VisitedLiveValues.count(Def)) {
           continue;
         }
         ToClobber.push_back(Alloca);
       }
 
       auto InsertClobbersAt = [&](Instruction *IP) {
         for (auto *AI : ToClobber) {
-          auto AIType = cast<PointerType>(AI->getType());
-          auto PT = cast<PointerType>(AIType->getElementType());
+          auto PT = cast<PointerType>(AI->getAllocatedType());
           Constant *CPN = ConstantPointerNull::get(PT);
           StoreInst *Store = new StoreInst(CPN, AI);
           Store->insertBefore(IP);
         }
       };
 
       // Insert the clobbering stores.  These may get intermixed with the
       // gc.results and gc.relocates, but that's fine.

llvm/trunk/lib/Transforms/Scalar/ScalarReplAggregates.cpp

if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
       if (GetUnderlyingObject(MTI->getSource(), DL, 0) != OrigAI) {
         // Dest must be OrigAI, change this to be a load from the original
         // pointer (bitcasted), then a store to our new alloca.
         assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
         Value *SrcPtr = MTI->getSource();
         PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
         PointerType* AIPTy = cast<PointerType>(NewAI->getType());
         if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
-          AIPTy = PointerType::get(AIPTy->getElementType(),
+          AIPTy = PointerType::get(NewAI->getAllocatedType(),
                                    SPTy->getAddressSpace());
         }
         SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
 
         LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
         SrcVal->setAlignment(MTI->getAlignment());
         Builder.CreateStore(SrcVal, NewAI);
       } else if (GetUnderlyingObject(MTI->getDest(), DL, 0) != OrigAI) {
         // Src must be OrigAI, change this to be a load from NewAI then a store
         // through the original dest pointer (bitcasted).
         assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
         LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
 
         PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
         PointerType* AIPTy = cast<PointerType>(NewAI->getType());
         if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
-          AIPTy = PointerType::get(AIPTy->getElementType(),
+          AIPTy = PointerType::get(NewAI->getAllocatedType(),
                                    DPTy->getAddressSpace());
         }
         Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
 
         StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
         NewStore->setAlignment(MTI->getAlignment());
       } else {
         // Noop transfer. Src == Dst
for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
     // Otherwise, we have a PHI node which allows us to push the loads into the
     // predecessors.
     PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
     if (PN->use_empty()) {
       PN->eraseFromParent();
       continue;
     }
 
-    Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
+    Type *LoadTy = AI->getAllocatedType();
     PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
                                      PN->getName()+".ld", PN);
 
     // Get the AA tags and alignment to use from one of the loads.  It doesn't
     // matter which one we get and if any differ, it doesn't matter.
     LoadInst *SomeLoad = cast<LoadInst>(PN->user_back());
 
     AAMDNodes AATags;
if (OtherPtr == AI || OtherPtr == NewElts[0]) {
              E = DeadInsts.end(); I != E; ++I)
         if (*I == MI) return;
       DeadInsts.push_back(MI);
       return;
     }
 
     // If the pointer is not the right type, insert a bitcast to the right
     // type.
-    Type *NewTy =
-      PointerType::get(AI->getType()->getElementType(), AddrSpace);
+    Type *NewTy = PointerType::get(AI->getAllocatedType(), AddrSpace);
 
     if (OtherPtr->getType() != NewTy)
       OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
   }
 
   // Process each element of the aggregate.
   bool SROADest = MI->getRawDest() == Inst;
 
   Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
   const DataLayout &DL = MI->getModule()->getDataLayout();
 
   for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
     // If this is a memcpy/memmove, emit a GEP of the other element address.
     Value *OtherElt = nullptr;
     unsigned OtherEltAlign = MemAlignment;
 
     if (OtherPtr) {
       Value *Idx[2] = { Zero,
                       ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
       OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
                                               OtherPtr->getName()+"."+Twine(i),
                                                    MI);
       uint64_t EltOffset;
-      PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
-      Type *OtherTy = OtherPtrTy->getElementType();
+      assert(AI->getType() == OtherPtr->getType());
+      Type *OtherTy = AI->getAllocatedType();
       if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
         EltOffset = DL.getStructLayout(ST)->getElementOffset(i);
       } else {
         Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
         EltOffset = DL.getTypeAllocSize(EltTy) * i;
       }
 
       // The alignment of the other pointer is the guaranteed alignment of the
       // element, which is affected by both the known alignment of the whole
       // mem intrinsic and the alignment of the element.  If the alignment of
       // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
       // known alignment is just 4 bytes.
       OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
     }
 
-    Value *EltPtr = NewElts[i];
-    Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
+    AllocaInst *EltPtr = NewElts[i];
+    Type *EltTy = EltPtr->getAllocatedType();
 
     // If we got down to a scalar, insert a load or store as appropriate.
     if (EltTy->isSingleValueType()) {
       if (isa<MemTransferInst>(MI)) {
         if (SROADest) {
           // From Other to Alloca.
           Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
           new StoreInst(Elt, EltPtr, MI);
SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
 
   Value *ResultVal =
     Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
 
   for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
     // Load the value from the alloca.  If the NewElt is an aggregate, cast
     // the pointer to an integer of the same size before doing the load.
     Value *SrcField = NewElts[i];
-    Type *FieldTy =
-      cast<PointerType>(SrcField->getType())->getElementType();
+    Type *FieldTy = NewElts[i]->getAllocatedType();
     uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy);
 
     // Ignore zero sized fields like {}, they obviously contain no data.
     if (FieldSizeBits == 0) continue;
 
     IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
                                                      FieldSizeBits);
     if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&