[PATCH] D26014: [ConstantFold] Get the correct vector type when folding a getelementptr instruction with vector indices

Davide Italiano via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 26 15:08:39 PDT 2016


davide created this revision.
davide added reviewers: majnemer, Bigcheese, RKSimon, spatel.
davide added a subscriber: llvm-commits.

When we constant fold `getelementptr` and the indices are vector, the type created is `i64 *` and not `<8 x i64*>` which caseuse an assertion to fire in `replaceAllUsesWith`. I think it's OK just checking for the first index for the reason pointed out in the comment.


https://reviews.llvm.org/D26014

Files:
  lib/IR/ConstantFold.cpp
  test/Transforms/InstCombine/gep-vector.ll


Index: test/Transforms/InstCombine/gep-vector.ll
===================================================================
--- /dev/null
+++ test/Transforms/InstCombine/gep-vector.ll
@@ -0,0 +1,19 @@
+; RUN: opt -instcombine %s -S | FileCheck %s
+
+; CHECK-LABEL: patatino
+; CHECK-NEXT: ret void
+define void @patatino() {
+  %el = getelementptr i64, <8 x i64*> undef, <8 x i64> undef
+  %gather = call <8 x i64> @llvm.masked.gather.v8i64(<8 x i64*> %el, i32 8, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <8 x i64> undef)
+  ret void
+}
+
+; CHECK-LABEL: patatino2
+; CHECK-NEXT: ret void
+define void @patatino2() {
+  %el = getelementptr inbounds i64, i64* undef, <8 x i64> undef
+  %gather = call <8 x i64> @llvm.masked.gather.v8i64(<8 x i64*> %el, i32 8, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <8 x i64> undef)
+  ret void
+}
+
+declare <8 x i64> @llvm.masked.gather.v8i64(<8 x i64*>, i32, <8 x i1>, <8 x i64>)
Index: lib/IR/ConstantFold.cpp
===================================================================
--- lib/IR/ConstantFold.cpp
+++ lib/IR/ConstantFold.cpp
@@ -2062,6 +2062,12 @@
     Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
     if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
       GEPTy = VectorType::get(GEPTy, VT->getNumElements());
+    // When the getelementptr has one of more vector indices, it will
+    // return a vector of pointers. We guarantee that all the vectors
+    // in this case will have the same width, so we can just look at
+    // the first one.
+    else if (VectorType *VT = dyn_cast<VectorType>(Idxs[0]->getType()))
+      GEPTy = VectorType::get(GEPTy, VT->getNumElements());
     return UndefValue::get(GEPTy);
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D26014.75955.patch
Type: text/x-patch
Size: 1774 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161026/6d0a036f/attachment.bin>


More information about the llvm-commits mailing list