[X86][SSE] Bitcast assertion in XFormVExtractWithShuffleIntoLoad

[RKSimon]

Summary

Minor patch to fix an issue in XFormVExtractWithShuffleIntoLoad where a load is unary shuffled, then bitcast (to a type with the same number of elements) before extracting an element.

An undef was created for the second shuffle operand using the original (post-bitcasted) vector type instead of the pre-bitcasted type like the rest of the shuffle node - this was then causing an assertion on the different types later on inside SelectionDAG::getVectorShuffle.

[qcolombet]

Hi Simon,

LGTM.

Thanks,
-Quentin

[RKSimon]

Closed by commit rL220592 (authored by @RKSimon).

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       const X86Subtarget *Subtarget)  {
   return SDValue();
 }
 
 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
 /// specific shuffle of a load can be folded into a single element load.
 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
-/// shuffles have been customed lowered so we need to handle those here.
+/// shuffles have been custom lowered so we need to handle those here.
 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
                                          TargetLowering::DAGCombinerInfo &DCI) {
   if (DCI.isBeforeLegalizeOps())
     return SDValue();
 
   SDValue InVec = N->getOperand(0);
   SDValue EltNo = N->getOperand(1);
 
   if (!isa<ConstantSDNode>(EltNo))
     return SDValue();
 
-  EVT VT = InVec.getValueType();
+  EVT OriginalVT = InVec.getValueType();
 
   if (InVec.getOpcode() == ISD::BITCAST) {
     // Don't duplicate a load with other uses.
     if (!InVec.hasOneUse())
       return SDValue();
     EVT BCVT = InVec.getOperand(0).getValueType();
-    if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
+    if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
       return SDValue();
     InVec = InVec.getOperand(0);
   }
 
+  EVT CurrentVT = InVec.getValueType();
+
   if (!isTargetShuffle(InVec.getOpcode()))
     return SDValue();
 
   // Don't duplicate a load with other uses.
   if (!InVec.hasOneUse())
     return SDValue();
 
   SmallVector<int, 16> ShuffleMask;
   bool UnaryShuffle;
-  if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
-                            UnaryShuffle))
+  if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
+                            ShuffleMask, UnaryShuffle))
     return SDValue();
 
   // Select the input vector, guarding against out of range extract vector.
-  unsigned NumElems = VT.getVectorNumElements();
+  unsigned NumElems = CurrentVT.getVectorNumElements();
   int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
   int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
   SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
                                          : InVec.getOperand(1);
 
   // If inputs to shuffle are the same for both ops, then allow 2 uses
   unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
 
static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
   if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
     return SDValue();
 
   // All checks match so transform back to vector_shuffle so that DAG combiner
   // can finish the job
   SDLoc dl(N);
 
   // Create shuffle node taking into account the case that its a unary shuffle
-  SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
-  Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
+  SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
+                                   : InVec.getOperand(1);
+  Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
                                  InVec.getOperand(0), Shuffle,
                                  &ShuffleMask[0]);
-  Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
+  Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
                      EltNo);
 }
 
 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
 /// generation and convert it from being a bunch of shuffles and extracts
 /// to a simple store and scalar loads to extract the elements.
 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,

llvm/trunk/test/CodeGen/X86/extractelement-load.ll

 ; CHECK: movlpd
 
 bb:
   %tmp13 = load <2 x double>* undef, align 1
   %.sroa.3.24.vec.extract = extractelement <2 x double> %tmp13, i32 1
   store double %.sroa.3.24.vec.extract, double* undef, align 8
   unreachable
 }
+
+; Case where a load is unary shuffled, then bitcast (to a type with the same
+; number of elements) before extractelement.
+; This is testing for an assertion - the extraction was assuming that the undef
+; second shuffle operand was a post-bitcast type instead of a pre-bitcast type.
+define i64 @t4(<2 x double>* %a) {
+; CHECK-LABEL: t4:
+; CHECK: mov
+; CHECK: ret
+  %b = load <2 x double>* %a, align 16
+  %c = shufflevector <2 x double> %b, <2 x double> %b, <2 x i32> <i32 1, i32 0>
+  %d = bitcast <2 x double> %c to <2 x i64>
+  %e = extractelement <2 x i64> %d, i32 1
+  ret i64 %e
+}
+