[PATCH] D65612: Handle casts changing pointer size in the vectorizer

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

This revision was automatically updated to reflect the committed changes.
Closed by commit rL367646: Handle casts changing pointer size in the vectorizer (authored by rampitec, committed by ).
Herald added a project: LLVM.

Changed prior to commit:
  https://reviews.llvm.org/D65612?vs=212929&id=212963#toc

Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65612/new/

https://reviews.llvm.org/D65612

Files:
  llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
  llvm/trunk/test/Transforms/LoadStoreVectorizer/AMDGPU/vect-ptr-ptr-size-mismatch.ll


Index: llvm/trunk/test/Transforms/LoadStoreVectorizer/AMDGPU/vect-ptr-ptr-size-mismatch.ll
===================================================================
--- llvm/trunk/test/Transforms/LoadStoreVectorizer/AMDGPU/vect-ptr-ptr-size-mismatch.ll
+++ llvm/trunk/test/Transforms/LoadStoreVectorizer/AMDGPU/vect-ptr-ptr-size-mismatch.ll
@@ -54,4 +54,43 @@
   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 = !{}
Index: llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
===================================================================
--- llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -147,7 +147,7 @@
   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;
@@ -336,18 +336,29 @@
 }
 
 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


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D65612.212963.patch
Type: text/x-patch
Size: 4017 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190802/3ad2bb95/attachment.bin>