[InstCombine] avoid infinite loop from shuffle-extract-insert sequence (PR30923)

Removing the limitation in visitInsertElementInst() causes several regressions
because we're not prepared to fold sequences of shuffles or inserts and extracts
separated by shuffles. Fixing that appears to be a difficult mission because we
are purposely trying to avoid creating shuffles with arbitrary shuffle masks
because some targets may choke on those.

https://llvm.org/bugs/show_bug.cgi?id=30923

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@286423 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index 6c3aa3ad6c9cd054463f1254bb3440cb6a0c2df4..732a786ceb7c024be43a80e60319634abf764385 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -413,6 +413,14 @@ static void replaceExtractElements(InsertElementInst *InsElt,
   if (InsertionBlock != InsElt->getParent())
     return;
 
+  // TODO: This restriction matches the check in visitInsertElementInst() and
+  // prevents an infinite loop caused by not turning the extract/insert pair
+  // into a shuffle. We really should not need either check, but we're lacking
+  // folds for shufflevectors because we're afraid to generate shuffle masks
+  // that the backend can't handle.
+  if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
+    return;
+
   auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
                                         ConstantVector::get(ExtendMask));
 

diff --git a/test/Transforms/InstCombine/insert-extract-shuffle.ll b/test/Transforms/InstCombine/insert-extract-shuffle.ll
index 163c336869969c31fb131fbe61daec857ac26a1b..4507deb7f023a78873df235e5e46633aaa153990 100644 (file)
--- a/test/Transforms/InstCombine/insert-extract-shuffle.ll
+++ b/test/Transforms/InstCombine/insert-extract-shuffle.ll
@@ -237,3 +237,30 @@ end:
   ret double %mu
 }
 
+; https://llvm.org/bugs/show_bug.cgi?id=30923
+; Delete the widening shuffle if we're not going to reduce the extract/insert to a shuffle.
+
+define <4 x float> @PR30923(<2 x float> %x) {
+; CHECK-LABEL: @PR30923(
+; CHECK-NEXT:  bb1:
+; CHECK-NEXT:    [[EXT1:%.*]] = extractelement <2 x float> %x, i32 1
+; CHECK-NEXT:    store float [[EXT1]], float* undef, align 4
+; CHECK-NEXT:    br label %bb2
+; CHECK:       bb2:
+; CHECK-NEXT:    [[EXT2:%.*]] = extractelement <2 x float> %x, i32 0
+; CHECK-NEXT:    [[INS1:%.*]] = insertelement <4 x float> <float 0.000000e+00, float 0.000000e+00, float undef, float undef>, float [[EXT2]], i32 2
+; CHECK-NEXT:    [[INS2:%.*]] = insertelement <4 x float> [[INS1]], float [[EXT1]], i32 3
+; CHECK-NEXT:    ret <4 x float> [[INS2]]
+;
+bb1:
+  %ext1 = extractelement <2 x float> %x, i32 1
+  store float %ext1, float* undef, align 4
+  br label %bb2
+
+bb2:
+  %widen = shufflevector <2 x float> %x, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+  %ext2 = extractelement <4 x float> %widen, i32 0
+  %ins1 = insertelement <4 x float> <float 0.0, float 0.0, float undef, float undef>, float %ext2, i32 2
+  %ins2 = insertelement <4 x float> %ins1, float %ext1, i32 3
+  ret <4 x float> %ins2
+}