From: Hans Wennborg <hans@hanshq.net>
Date: Fri, 29 Jan 2016 21:33:02 +0000 (+0000)
Subject: Merging r259236:
X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=commitdiff_plain;h=5ad5d2c5359a4e878c732db59ee7fc6e0a25dc00

Merging r259236:
------------------------------------------------------------------------
r259236 | spatel | 2016-01-29 12:21:02 -0800 (Fri, 29 Jan 2016) | 8 lines

[InstCombine] avoid an insertelement transformation that induces the opposite extractelement fold (PR26354)

We would infinite loop because we created a shufflevector that was wider than
needed and then failed to combine that with the insertelement. When subsequently
visiting the extractelement from that shuffle, we see that it's unnecessary,
delete it, and trigger another visit to the insertelement.


------------------------------------------------------------------------


git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_38@259245 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index 5cde31a9162..bc4c0ebae79 100644
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -380,6 +380,23 @@ static void replaceExtractElements(InsertElementInst *InsElt,
     ExtendMask.push_back(UndefValue::get(IntType));
 
   Value *ExtVecOp = ExtElt->getVectorOperand();
+  auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
+  BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
+                                   ? ExtVecOpInst->getParent()
+                                   : ExtElt->getParent();
+
+  // TODO: This restriction matches the basic block check below when creating
+  // new extractelement instructions. If that limitation is removed, this one
+  // could also be removed. But for now, we just bail out to ensure that we
+  // will replace the extractelement instruction that is feeding our
+  // insertelement instruction. This allows the insertelement to then be
+  // replaced by a shufflevector. If the insertelement is not replaced, we can
+  // induce infinite looping because there's an optimization for extractelement
+  // that will delete our widening shuffle. This would trigger another attempt
+  // here to create that shuffle, and we spin forever.
+  if (InsertionBlock != InsElt->getParent())
+    return;
+
   auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
                                         ConstantVector::get(ExtendMask));
 
@@ -387,7 +404,6 @@ static void replaceExtractElements(InsertElementInst *InsElt,
   // (as long as it's not a PHI) or at the start of the basic block of the
   // extract, so any subsequent extracts in the same basic block can use it.
   // TODO: Insert before the earliest ExtractElementInst that is replaced.
-  auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
   if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
     WideVec->insertAfter(ExtVecOpInst);
   else
diff --git a/test/Transforms/InstCombine/insert-extract-shuffle.ll b/test/Transforms/InstCombine/insert-extract-shuffle.ll
index 47c2a139a47..8ed4db8bbbc 100644
--- a/test/Transforms/InstCombine/insert-extract-shuffle.ll
+++ b/test/Transforms/InstCombine/insert-extract-shuffle.ll
@@ -175,3 +175,33 @@ bb3:
   ret <4 x double> %tmp4
 }
 
+; PR26354: https://llvm.org/bugs/show_bug.cgi?id=26354
+; Don't create a shufflevector if we know that we're not going to replace the insertelement.
+
+define double @pr26354(<2 x double>* %tmp, i1 %B) {
+; CHECK-LABEL: @pr26354(
+; CHECK:       %ld = load <2 x double>, <2 x double>* %tmp
+; CHECK-NEXT:  %e1 = extractelement <2 x double> %ld, i32 0
+; CHECK-NEXT:  br i1 %B, label %if, label %end
+; CHECK:       if:
+; CHECK-NEXT:  %e2 = extractelement <2 x double> %ld, i32 1
+; CHECK-NEXT:  %i1 = insertelement <4 x double>
+; CHECK-NEXT:  br label %end
+
+entry:
+  %ld = load <2 x double>, <2 x double>* %tmp
+  %e1 = extractelement <2 x double> %ld, i32 0
+  %e2 = extractelement <2 x double> %ld, i32 1
+  br i1 %B, label %if, label %end
+
+if:
+  %i1 = insertelement <4 x double> zeroinitializer, double %e2, i32 3
+  br label %end
+
+end:
+  %ph = phi <4 x double> [ undef, %entry ], [ %i1, %if ]
+  %e3 = extractelement <4 x double> %ph, i32 1
+  %mu = fmul double %e1, %e3
+  ret double %mu
+}
+