[PATCH] D32101: Skip bitcasts while looking for GEP in LoadStoreVectorizer

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec created this revision.
Herald added subscribers: nhaehnle, wdng, mzolotukhin.

Repository:
  rL LLVM

https://reviews.llvm.org/D32101

Files:
  lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
  test/Transforms/LoadStoreVectorizer/AMDGPU/gep-bitcast.ll


Index: test/Transforms/LoadStoreVectorizer/AMDGPU/gep-bitcast.ll
===================================================================
--- /dev/null
+++ test/Transforms/LoadStoreVectorizer/AMDGPU/gep-bitcast.ll
@@ -0,0 +1,27 @@
+; RUN: opt -S -mtriple=amdgcn--amdhsa -load-store-vectorizer < %s | FileCheck %s
+
+; Check that vectorizer can find a GEP through bitcast
+; CHECK: load <4 x i32>
+define void @vect_zext_bitcast_idx(float addrspace(1)* %arg1, i32 %base) {
+  %add1 = add nuw i32 %base, 0
+  %zext1 = zext i32 %add1 to i64
+  %gep1 = getelementptr inbounds float, float addrspace(1)* %arg1, i64 %zext1
+  %f2i1 = bitcast float addrspace(1)* %gep1 to i32 addrspace(1)*
+  %load1 = load i32, i32 addrspace(1)* %f2i1, align 4
+  %add2 = add nuw i32 %base, 1
+  %zext2 = zext i32 %add2 to i64
+  %gep2 = getelementptr inbounds float, float addrspace(1)* %arg1, i64 %zext2
+  %f2i2 = bitcast float addrspace(1)* %gep2 to i32 addrspace(1)*
+  %load2 = load i32, i32 addrspace(1)* %f2i2, align 4
+  %add3 = add nuw i32 %base, 2
+  %zext3 = zext i32 %add3 to i64
+  %gep3 = getelementptr inbounds float, float addrspace(1)* %arg1, i64 %zext3
+  %f2i3 = bitcast float addrspace(1)* %gep3 to i32 addrspace(1)*
+  %load3 = load i32, i32 addrspace(1)* %f2i3, align 4
+  %add4 = add nuw i32 %base, 3
+  %zext4 = zext i32 %add4 to i64
+  %gep4 = getelementptr inbounds float, float addrspace(1)* %arg1, i64 %zext4
+  %f2i4 = bitcast float addrspace(1)* %gep4 to i32 addrspace(1)*
+  %load4 = load i32, i32 addrspace(1)* %f2i4, align 4
+  ret void
+}
Index: lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
===================================================================
--- lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -283,8 +283,14 @@
 
   // Look through GEPs after checking they're the same except for the last
   // index.
-  GetElementPtrInst *GEPA = dyn_cast<GetElementPtrInst>(getPointerOperand(A));
-  GetElementPtrInst *GEPB = dyn_cast<GetElementPtrInst>(getPointerOperand(B));
+  Value *SrcPTRA = getPointerOperand(A);
+  Value *SrcPTRB = getPointerOperand(B);
+  while (auto *C = dyn_cast<BitCastInst>(SrcPTRA))
+    SrcPTRA = C->getOperand(0);
+  while (auto *C = dyn_cast<BitCastInst>(SrcPTRB))
+    SrcPTRB = C->getOperand(0);
+  GetElementPtrInst *GEPA = dyn_cast<GetElementPtrInst>(SrcPTRA);
+  GetElementPtrInst *GEPB = dyn_cast<GetElementPtrInst>(SrcPTRB);
   if (!GEPA || !GEPB || GEPA->getNumOperands() != GEPB->getNumOperands())
     return false;
   unsigned FinalIndex = GEPA->getNumOperands() - 1;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D32101.95363.patch
Type: text/x-patch
Size: 2583 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170414/84c7ee83/attachment.bin>