ARM64: handle v1i1 types arising from setcc properly.

There were several overlapping problems here, and this solution is
closely inspired by the one adopted in AArch64 in r201381.

Firstly, scalarisation of v1i1 setcc operations simply fails if the
input types are legal. This is fixed in LegalizeVectorTypes.cpp this
time, and allows AArch64 code to be simplified slightly.

Second, vselect with such a setcc feeding into it ends up in
ScalarizeVectorOperand, where it's not handled. I experimented with an
implementation, but found that whatever DAG came out was rather
horrific. I think Hao's DAG combine approach is a good one for
quality, though there are edge cases it won't catch (to be fixed
separately).

Should fix PR19335.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205625 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

@@ -331,12 +331,24 @@ SDValue DAGTypeLegalizer::ScalarizeVecRes_VSETCC(SDNode *N) {
   assert(N->getValueType(0).isVector() &&
          N->getOperand(0).getValueType().isVector() &&
          "Operand types must be vectors");
-
-  SDValue LHS = GetScalarizedVector(N->getOperand(0));
-  SDValue RHS = GetScalarizedVector(N->getOperand(1));
+  SDValue LHS = N->getOperand(0);
+  SDValue RHS = N->getOperand(1);
+  EVT OpVT = LHS.getValueType();
   EVT NVT = N->getValueType(0).getVectorElementType();
   SDLoc DL(N);
 
+  // The result needs scalarizing, but it's not a given that the source does.
+  if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {
+    LHS = GetScalarizedVector(LHS);
+    RHS = GetScalarizedVector(RHS);
+  } else {
+    EVT VT = OpVT.getVectorElementType();
+    LHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, LHS,
+                      DAG.getConstant(0, TLI.getVectorIdxTy()));
+    RHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, RHS,
+                      DAG.getConstant(0, TLI.getVectorIdxTy()));
+  }
+
   // Turn it into a scalar SETCC.
   SDValue Res = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS,
                             N->getOperand(2));

lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -538,7 +538,6 @@ AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
     setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
   }
 
-  setTargetDAGCombine(ISD::SETCC);
   setTargetDAGCombine(ISD::SIGN_EXTEND);
   setTargetDAGCombine(ISD::VSELECT);
 }
@@ -4284,32 +4283,6 @@ static SDValue CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   return SDValue(N, 0);
 }
 
-// v1i1 setcc ->
-//     v1i1 (bitcast (i1 setcc (extract_vector_elt, extract_vector_elt))
-// FIXME: Currently the type legalizer can't handle SETCC having v1i1 as result.
-// If it can legalize "v1i1 SETCC" correctly, no need to combine such SETCC.
-static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
-  EVT ResVT = N->getValueType(0);
-
-  if (!ResVT.isVector() || ResVT.getVectorNumElements() != 1 ||
-      ResVT.getVectorElementType() != MVT::i1)
-    return SDValue();
-
-  SDValue LHS = N->getOperand(0);
-  SDValue RHS = N->getOperand(1);
-  EVT CmpVT = LHS.getValueType();
-  LHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
-                    CmpVT.getVectorElementType(), LHS,
-                    DAG.getConstant(0, MVT::i64));
-  RHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
-                    CmpVT.getVectorElementType(), RHS,
-                    DAG.getConstant(0, MVT::i64));
-  SDValue SetCC =
-      DAG.getSetCC(SDLoc(N), MVT::i1, LHS, RHS,
-                   cast<CondCodeSDNode>(N->getOperand(2))->get());
-  return DAG.getNode(ISD::BITCAST, SDLoc(N), ResVT, SetCC);
-}
-
 // vselect (v1i1 setcc) ->
 //     vselect (v1iXX setcc)  (XX is the size of the compared operand type)
 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
@@ -4378,7 +4351,6 @@ AArch64TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::SRA:
   case ISD::SRL:
     return PerformShiftCombine(N, DCI, getSubtarget());
-  case ISD::SETCC: return PerformSETCCCombine(N, DCI.DAG);
   case ISD::VSELECT: return PerformVSelectCombine(N, DCI.DAG);
   case ISD::SIGN_EXTEND: return PerformSignExtendCombine(N, DCI.DAG);
   case ISD::INTRINSIC_WO_CHAIN:

lib/Target/ARM64/ARM64ISelLowering.cpp

@@ -434,6 +434,8 @@ ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
 
   setTargetDAGCombine(ISD::MUL);
 
+  setTargetDAGCombine(ISD::VSELECT);
+
   MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
   MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
   MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
@@ -7227,6 +7229,36 @@ static SDValue performBRCONDCombine(SDNode *N,
   return SDValue();
 }
 
+// vselect (v1i1 setcc) ->
+//     vselect (v1iXX setcc)  (XX is the size of the compared operand type)
+// FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
+// condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
+// such VSELECT.
+static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
+  SDValue N0 = N->getOperand(0);
+  EVT CCVT = N0.getValueType();
+
+  if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
+      CCVT.getVectorElementType() != MVT::i1)
+    return SDValue();
+
+  EVT ResVT = N->getValueType(0);
+  EVT CmpVT = N0.getOperand(0).getValueType();
+  // Only combine when the result type is of the same size as the compared
+  // operands.
+  if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
+    return SDValue();
+
+  SDValue IfTrue = N->getOperand(1);
+  SDValue IfFalse = N->getOperand(2);
+  SDValue SetCC =
+      DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
+                   N0.getOperand(0), N0.getOperand(1),
+                   cast<CondCodeSDNode>(N0.getOperand(2))->get());
+  return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
+                     IfTrue, IfFalse);
+}
+
 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -7255,6 +7287,8 @@ SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
     return performBitcastCombine(N, DCI, DAG);
   case ISD::CONCAT_VECTORS:
     return performConcatVectorsCombine(N, DCI, DAG);
+  case ISD::VSELECT:
+    return performVSelectCombine(N, DCI.DAG);
   case ISD::STORE:
     return performSTORECombine(N, DCI, DAG, Subtarget);
   case ARM64ISD::BRCOND:

test/CodeGen/ARM64/neon-v1i1-setcc.ll

@@ -0,0 +1,65 @@
+; RUN: llc %s -o - -verify-machineinstrs -mtriple=arm64-none-linux-gnu | FileCheck %s
+
+; This is the analogue of AArch64's file of the same name. It's mostly testing
+; some form of correct lowering occurs, the tests are a little artificial but I
+; strongly suspect there's room for improved CodeGen (FIXME).
+
+define i64 @test_sext_extr_cmp_0(<1 x i64> %v1, <1 x i64> %v2) {
+; CHECK-LABEL: test_sext_extr_cmp_0:
+; CHECK: cmp {{x[0-9]+}}, {{x[0-9]+}}
+; CHECK: csinc
+  %1 = icmp sge <1 x i64> %v1, %v2
+  %2 = extractelement <1 x i1> %1, i32 0
+  %vget_lane = sext i1 %2 to i64
+  ret i64 %vget_lane
+}
+
+define i64 @test_sext_extr_cmp_1(<1 x double> %v1, <1 x double> %v2) {
+; CHECK-LABEL: test_sext_extr_cmp_1:
+; CHECK: fcmp {{d[0-9]+}}, {{d[0-9]+}}
+  %1 = fcmp oeq <1 x double> %v1, %v2
+  %2 = extractelement <1 x i1> %1, i32 0
+  %vget_lane = sext i1 %2 to i64
+  ret i64 %vget_lane
+}
+
+define <1 x i64> @test_select_v1i1_0(<1 x i64> %v1, <1 x i64> %v2, <1 x i64> %v3) {
+; CHECK-LABEL: test_select_v1i1_0:
+; CHECK: cmeq d{{[0-9]+}}, d{{[0-9]+}}, d{{[0-9]+}}
+; CHECK: bic v{{[0-9]+}}.8b, v{{[0-9]+}}.8b, v{{[0-9]+}}.8b
+  %1 = icmp eq <1 x i64> %v1, %v2
+  %res = select <1 x i1> %1, <1 x i64> zeroinitializer, <1 x i64> %v3
+  ret <1 x i64> %res
+}
+
+define <1 x i64> @test_select_v1i1_1(<1 x double> %v1, <1 x double> %v2, <1 x i64> %v3) {
+; CHECK-LABEL: test_select_v1i1_1:
+; CHECK: fcmeq d{{[0-9]+}}, d{{[0-9]+}}, d{{[0-9]+}}
+; CHECK: bic v{{[0-9]+}}.8b, v{{[0-9]+}}.8b, v{{[0-9]+}}.8b
+  %1 = fcmp oeq <1 x double> %v1, %v2
+  %res = select <1 x i1> %1, <1 x i64> zeroinitializer, <1 x i64> %v3
+  ret <1 x i64> %res
+}
+
+define <1 x double> @test_select_v1i1_2(<1 x i64> %v1, <1 x i64> %v2, <1 x double> %v3) {
+; CHECK-LABEL: test_select_v1i1_2:
+; CHECK: cmeq d{{[0-9]+}}, d{{[0-9]+}}, d{{[0-9]+}}
+; CHECK: bic v{{[0-9]+}}.8b, v{{[0-9]+}}.8b, v{{[0-9]+}}.8b
+  %1 = icmp eq <1 x i64> %v1, %v2
+  %res = select <1 x i1> %1, <1 x double> zeroinitializer, <1 x double> %v3
+  ret <1 x double> %res
+}
+
+define i32 @test_br_extr_cmp(<1 x i64> %v1, <1 x i64> %v2) {
+; CHECK-LABEL: test_br_extr_cmp:
+; CHECK: cmp x{{[0-9]+}}, x{{[0-9]+}}
+  %1 = icmp eq <1 x i64> %v1, %v2
+  %2 = extractelement <1 x i1> %1, i32 0
+  br i1 %2, label %if.end, label %if.then
+
+if.then:
+  ret i32 0;
+
+if.end:
+  ret i32 1;
+}