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

[Mekhanoshin, Stanislav via llvm-commits]

Thanks! Forgot about it.
Review updated.

Stas

From: Daniel Berlin [mailto:dberlin at dberlin.org]
Sent: Friday, April 14, 2017 5:19 PM
To: reviews+D32101+public+305894f3d2e6edbd at reviews.llvm.org; Stanislav Mekhanoshin via Phabricator
Cc: Mekhanoshin, Stanislav; Arsenault, Matthew; Ding, Wei; llvm-commits; dorit.nuzman at intel.com; Anna Thomas
Subject: Re: [PATCH] D32101: Skip bitcasts while looking for GEP in LoadStoreVectorizer

Ignoring the issue of neededness or not, we have a function to do this anyway, you would want to use ->stripPointerCasts(), not do it yourself :)


On Fri, Apr 14, 2017 at 4:15 PM, Stanislav Mekhanoshin via Phabricator via llvm-commits <llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>> wrote:
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;



_______________________________________________
llvm-commits mailing list
llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170415/25e86bd8/attachment.html>