AArch64: skip select/setcc combine in complex case.

In an llvm-stress generated test, we were trying to create a v0iN type and
asserting when that failed. This case could probably be handled by the
function, but not without added complexity and the situation it arises in is
sufficiently odd that there's probably no benefit anyway.

Should fix PR20775.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216725 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index d3ec172c9d4c6f2e71459899c72e823925e4a3e0..430938e2f430b56c09f6ebd4d17e9bcbb0a54988 100644 (file)
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -7994,22 +7994,24 @@ static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
   SDValue N0 = N->getOperand(0);
   EVT ResVT = N->getValueType(0);
+  EVT SrcVT = N0.getOperand(0).getValueType();
+  int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
 
-  if (!N->getOperand(1).getValueType().isVector())
+  // If NumMaskElts == 0, the comparison is larger than select result. The
+  // largest real NEON comparison is 64-bits per lane, which means the result is
+  // at most 32-bits and an illegal vector. Just bail out for now.
+  if (!ResVT.isVector() || NumMaskElts == 0)
     return SDValue();
 
   if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
     return SDValue();
 
-  SDLoc DL(N0);
-
-  EVT SrcVT = N0.getOperand(0).getValueType();
-  SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT,
-                           ResVT.getSizeInBits() / SrcVT.getSizeInBits());
+  SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
   EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
 
   // First perform a vector comparison, where lane 0 is the one we're interested
   // in.
+  SDLoc DL(N0);
   SDValue LHS =
       DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
   SDValue RHS =
@@ -8019,8 +8021,8 @@ static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
   // Now duplicate the comparison mask we want across all other lanes.
   SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
   SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
-  Mask = DAG.getNode(ISD::BITCAST, DL, ResVT.changeVectorElementTypeToInteger(),
-                     Mask);
+  Mask = DAG.getNode(ISD::BITCAST, DL,
+                     ResVT.changeVectorElementTypeToInteger(), Mask);
 
   return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
 }

diff --git a/test/CodeGen/AArch64/cond-sel.ll b/test/CodeGen/AArch64/cond-sel.ll
index 5f81cba66cbc85f135103416958a427ddcecdd53..333f24361336b79dd40a5a050d514d081b58aeb9 100644 (file)
--- a/test/CodeGen/AArch64/cond-sel.ll
+++ b/test/CodeGen/AArch64/cond-sel.ll
@@ -214,3 +214,13 @@ define void @test_csetm(i32 %lhs, i32 %rhs, i64 %lhs64) {
   ret void
 ; CHECK: ret
 }
+
+define <1 x i1> @test_wide_comparison(i32 %in) {
+; CHECK-LABEL: test_wide_comparison:
+; CHECK: cmp w0, #1234
+; CHECK: cset
+
+  %tmp = icmp sgt i32 %in, 1234
+  %res = select i1 %tmp, <1 x i1> <i1 1>, <1 x i1> zeroinitializer
+  ret <1 x i1> %res
+}