[DAGCombine] Fix a bug in a BUILD_VECTOR combine

[mkuper]

Summary

When trying to convert a BUILD_VECTOR into a shuffle, we try to split vectors that are twice as wide as the destination.
We can not do this when we also need the zero vector to create a blend.

This fixes PR22774.

[nadav]

LGTM.

[andreadb]

Hi Michael,

The patch looks good to me. Thanks for the quick fix!

Is this a regression from 224759?
If so, then your fix may be a good candidate for 3.6.1.

[mkuper]

Thanks, Nadav, Andrea.

Yes, this is a regression from r224759 and should probably go into 3.6.1.

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if ((VT != VecIn1.getValueType())) {
           // If we have two small inputs of the same type, try to concat them.
           VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
           VecIn2 = SDValue(nullptr, 0);
         } else
           return SDValue();
       } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
         // If the input vector is too large, try to split it.
         // We don't support having two input vectors that are too large.
-        if (VecIn2.getNode())
+        // If the zero vector was used, we can not split the vector,
+        // since we'd need 3 inputs.
+        if (UsesZeroVector || VecIn2.getNode())
           return SDValue();
 
         if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
           return SDValue();
 
         // Try to replace VecIn1 with two extract_subvectors
         // No need to update the masks, they should still be correct.
         VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
           DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
         VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
           DAG.getConstant(0, TLI.getVectorIdxTy()));
-        UsesZeroVector = false;
       } else
         return SDValue();
     }
 
     if (UsesZeroVector)
       VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
                                 DAG.getConstantFP(0.0, VT);
     else

llvm/trunk/test/CodeGen/X86/pr22774.ll

+; RUN: llc -mattr=avx %s -o - | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-linux-gnu"
+
+@in = global <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>, align 32
+@out = global <2 x i64> zeroinitializer, align 16
+
+define i32 @_Z3foov() {
+entry:
+; CHECK: vmovdqa in(%rip), %ymm0
+; CHECK-NEXT: vmovq %xmm0, %xmm0
+; CHECK-NEXT: vmovdqa %xmm0, out(%rip)
+  %0 = load <4 x i64>, <4 x i64>* @in, align 32
+  %vecext = extractelement <4 x i64> %0, i32 0
+  %vecinit = insertelement <2 x i64> undef, i64 %vecext, i32 0
+  %vecinit1 = insertelement <2 x i64> %vecinit, i64 0, i32 1
+  store <2 x i64> %vecinit1, <2 x i64>* @out, align 16
+  ret i32 0
+}