Handle casts changing pointer size in the vectorizer

[rampitec]

Summary

Added code to truncate or shrink offsets so that we can continue
base pointer search if size has changed along the way.

[rampitec]

This one will go through our internal extended testing first, just to be sure it did not break anything.

[tra]

lib/Transforms/Vectorize/LoadStoreVectorizer.cpp

351–366 ↗

(On Diff #212914)

This can probably be collapsed to OffsetA = OffsetA.sextOrTrunk(NewPtrBitWidth);
same for OffsetB & PtrDelta.

[rampitec]

Changes to sextOrTrunc().

llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp

if (Align != 0)
       return Align;
 
     return DL.getABITypeAlignment(SI->getValueOperand()->getType());
   }
 
   static const unsigned MaxDepth = 3;
 
   bool isConsecutiveAccess(Value *A, Value *B);
-  bool areConsecutivePointers(Value *PtrA, Value *PtrB, const APInt &PtrDelta,
+  bool areConsecutivePointers(Value *PtrA, Value *PtrB, APInt PtrDelta,
                               unsigned Depth = 0) const;
   bool lookThroughComplexAddresses(Value *PtrA, Value *PtrB, APInt PtrDelta,
                                    unsigned Depth) const;
   bool lookThroughSelects(Value *PtrA, Value *PtrB, const APInt &PtrDelta,
                           unsigned Depth) const;
 
   /// After vectorization, reorder the instructions that I depends on
   /// (the instructions defining its operands), to ensure they dominate I.
bool Vectorizer::isConsecutiveAccess(Value *A, Value *B) {
 
   unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA);
   APInt Size(PtrBitWidth, DL.getTypeStoreSize(PtrATy));
 
   return areConsecutivePointers(PtrA, PtrB, Size);
 }
 
 bool Vectorizer::areConsecutivePointers(Value *PtrA, Value *PtrB,
-                                        const APInt &PtrDelta,
-                                        unsigned Depth) const {
+                                        APInt PtrDelta, unsigned Depth) const {
   unsigned PtrBitWidth = DL.getPointerTypeSizeInBits(PtrA->getType());
   APInt OffsetA(PtrBitWidth, 0);
   APInt OffsetB(PtrBitWidth, 0);
   PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
   PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
 
-  if (DL.getTypeStoreSizeInBits(PtrA->getType()) != PtrBitWidth ||
-      DL.getTypeStoreSizeInBits(PtrB->getType()) != PtrBitWidth)
+  unsigned NewPtrBitWidth = DL.getTypeStoreSizeInBits(PtrA->getType());
+
+  if (NewPtrBitWidth != DL.getTypeStoreSizeInBits(PtrB->getType()))
     return false;
 
+  // In case if we have to shrink the pointer
+  // stripAndAccumulateInBoundsConstantOffsets should properly handle a
+  // possible overflow and the value should fit into a smallest data type
+  // used in the cast/gep chain.
+  assert(OffsetA.getMinSignedBits() <= NewPtrBitWidth &&
+         OffsetB.getMinSignedBits() <= NewPtrBitWidth);
+
+  OffsetA = OffsetA.sextOrTrunc(NewPtrBitWidth);
+  OffsetB = OffsetB.sextOrTrunc(NewPtrBitWidth);
+  PtrDelta = PtrDelta.sextOrTrunc(NewPtrBitWidth);
+
   APInt OffsetDelta = OffsetB - OffsetA;
 
   // Check if they are based on the same pointer. That makes the offsets
   // sufficient.
   if (PtrA == PtrB)
     return OffsetDelta == PtrDelta;
 
   // Compute the necessary base pointer delta to have the necessary final delta

llvm/trunk/test/Transforms/LoadStoreVectorizer/AMDGPU/vect-ptr-ptr-size-mismatch.ll

entry:
   %tmp16 = bitcast i8 addrspace(1)* %tmp15 to float addrspace(1)*
   %tmp17 = load float, float addrspace(1)* %tmp16, align 8, !invariant.load !0
   %tmp24 = getelementptr inbounds i8, i8 addrspace(1)* %alloc16, i64 12
   %tmp25 = bitcast i8 addrspace(1)* %tmp24 to float addrspace(1)*
   %tmp26 = load float, float addrspace(1)* %tmp25, align 4, !invariant.load !0
   ret void
 }
 
+; CHECK-LABEL: @ext_ptr
+; CHECK  load <2 x i32>
+define void @ext_ptr(i32 addrspace(5)* %p) {
+entry:
+  %gep1 = getelementptr inbounds i32, i32 addrspace(5)* %p, i64 0
+  %gep2 = getelementptr inbounds i32, i32 addrspace(5)* %p, i64 1
+  %a.ascast = addrspacecast i32 addrspace(5)* %gep1 to i32*
+  %b.ascast = addrspacecast i32 addrspace(5)* %gep2 to i32*
+  %tmp1 = load i32, i32* %a.ascast, align 8
+  %tmp2 = load i32, i32* %b.ascast, align 8
+  unreachable
+}
+
+; CHECK-LABEL: @shrink_ptr
+; CHECK  load <2 x i32>
+define void @shrink_ptr(i32* %p) {
+entry:
+  %gep1 = getelementptr inbounds i32, i32* %p, i64 0
+  %gep2 = getelementptr inbounds i32, i32* %p, i64 1
+  %a.ascast = addrspacecast i32* %gep1 to i32 addrspace(5)*
+  %b.ascast = addrspacecast i32* %gep2 to i32 addrspace(5)*
+  %tmp1 = load i32, i32 addrspace(5)* %a.ascast, align 8
+  %tmp2 = load i32, i32 addrspace(5)* %b.ascast, align 8
+  unreachable
+}
+
+; CHECK-LABEL: @ext_ptr_wrap
+; CHECK: load <2 x i8>
+define void @ext_ptr_wrap(i8 addrspace(5)* %p) {
+entry:
+  %gep1 = getelementptr inbounds i8, i8 addrspace(5)* %p, i64 0
+  %gep2 = getelementptr inbounds i8, i8 addrspace(5)* %p, i64 4294967295
+  %a.ascast = addrspacecast i8 addrspace(5)* %gep1 to i8*
+  %b.ascast = addrspacecast i8 addrspace(5)* %gep2 to i8*
+  %tmp1 = load i8, i8* %a.ascast, align 1
+  %tmp2 = load i8, i8* %b.ascast, align 1
+  unreachable
+}
+
 !0 = !{}