From: Krzysztof Parzyszek <kparzysz@codeaurora.org>
Date: Thu, 14 Dec 2017 21:28:48 +0000 (+0000)
Subject: [Hexagon] Generate HVX code for comparisons and selects
X-Git-Tag: android-x86-7.1-r4~7210
X-Git-Url: http://git.osdn.net/view?a=commitdiff_plain;h=a9bb60c600bfec513cc00b7b126dfade550a3356;p=android-x86%2Fexternal-llvm.git

[Hexagon] Generate HVX code for comparisons and selects


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@320744 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/Hexagon/HexagonISelLowering.cpp b/lib/Target/Hexagon/HexagonISelLowering.cpp
index 0b7765c4391..be6d13a9fd0 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.cpp
+++ b/lib/Target/Hexagon/HexagonISelLowering.cpp
@@ -1272,6 +1272,9 @@ SDValue HexagonTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
 
   SDValue LHS = Op.getOperand(0);
   SDValue RHS = Op.getOperand(1);
+  if (Subtarget.useHVXOps() && Subtarget.isHVXVectorType(ty(LHS)))
+    return LowerHvxSetCC(Op, DAG);
+
   SDValue Cmp = Op.getOperand(2);
   ISD::CondCode CC = cast<CondCodeSDNode>(Cmp)->get();
 
@@ -1732,6 +1735,9 @@ HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
       addRegisterClass(MVT::v128i8, &Hexagon::HvxWRRegClass);
       addRegisterClass(MVT::v64i16, &Hexagon::HvxWRRegClass);
       addRegisterClass(MVT::v32i32, &Hexagon::HvxWRRegClass);
+      addRegisterClass(MVT::v16i1, &Hexagon::HvxQRRegClass);
+      addRegisterClass(MVT::v32i1, &Hexagon::HvxQRRegClass);
+      addRegisterClass(MVT::v64i1, &Hexagon::HvxQRRegClass);
       addRegisterClass(MVT::v512i1, &Hexagon::HvxQRRegClass);
     } else if (Subtarget.useHVX128BOps()) {
       addRegisterClass(MVT::v128i8,  &Hexagon::HvxVRRegClass);
@@ -1740,6 +1746,9 @@ HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
       addRegisterClass(MVT::v256i8,  &Hexagon::HvxWRRegClass);
       addRegisterClass(MVT::v128i16, &Hexagon::HvxWRRegClass);
       addRegisterClass(MVT::v64i32,  &Hexagon::HvxWRRegClass);
+      addRegisterClass(MVT::v32i1, &Hexagon::HvxQRRegClass);
+      addRegisterClass(MVT::v64i1, &Hexagon::HvxQRRegClass);
+      addRegisterClass(MVT::v128i1, &Hexagon::HvxQRRegClass);
       addRegisterClass(MVT::v1024i1, &Hexagon::HvxQRRegClass);
     }
   }
@@ -2001,10 +2010,12 @@ HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
       setIndexedLoadAction(ISD::POST_INC,  T, Legal);
       setIndexedStoreAction(ISD::POST_INC, T, Legal);
 
-      setOperationAction(ISD::ADD, T, Legal);
-      setOperationAction(ISD::SUB, T, Legal);
-      setOperationAction(ISD::MUL, T, Custom);
+      setOperationAction(ISD::ADD,     T, Legal);
+      setOperationAction(ISD::SUB,     T, Legal);
+      setOperationAction(ISD::VSELECT, T, Legal);
 
+      setOperationAction(ISD::MUL,                T, Custom);
+      setOperationAction(ISD::SETCC,              T, Custom);
       setOperationAction(ISD::BUILD_VECTOR,       T, Custom);
       setOperationAction(ISD::INSERT_SUBVECTOR,   T, Custom);
       setOperationAction(ISD::INSERT_VECTOR_ELT,  T, Custom);
diff --git a/lib/Target/Hexagon/HexagonISelLowering.h b/lib/Target/Hexagon/HexagonISelLowering.h
index fca48d4cca6..0a5771ec883 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.h
+++ b/lib/Target/Hexagon/HexagonISelLowering.h
@@ -347,6 +347,7 @@ namespace HexagonISD {
     SDValue LowerHvxExtractSubvector(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerHvxInsertSubvector(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerHvxMul(SDValue Op, SelectionDAG &DAG) const;
+    SDValue LowerHvxSetCC(SDValue Op, SelectionDAG &DAG) const;
 
     std::pair<const TargetRegisterClass*, uint8_t>
     findRepresentativeClass(const TargetRegisterInfo *TRI, MVT VT)
diff --git a/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp b/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
index cfa0c003e41..652dffce841 100644
--- a/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
+++ b/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
@@ -383,3 +383,74 @@ HexagonTargetLowering::LowerHvxMul(SDValue Op, SelectionDAG &DAG) const {
   }
   return SDValue();
 }
+
+SDValue
+HexagonTargetLowering::LowerHvxSetCC(SDValue Op, SelectionDAG &DAG) const {
+  MVT VecTy = ty(Op.getOperand(0));
+  assert(VecTy == ty(Op.getOperand(1)));
+
+  SDValue Cmp = Op.getOperand(2);
+  ISD::CondCode CC = cast<CondCodeSDNode>(Cmp)->get();
+  bool Negate = false, Swap = false;
+
+  // HVX has instructions for SETEQ, SETGT, SETUGT. The other comparisons
+  // can be arranged as operand-swapped/negated versions of these. Since
+  // the generated code will have the original CC expressed as
+  //   (negate (swap-op NewCmp)),
+  // the condition code for the NewCmp should be calculated from the original
+  // CC by applying these operations in the reverse order.
+
+  switch (CC) {
+    case ISD::SETNE:    // !eq
+    case ISD::SETLE:    // !gt
+    case ISD::SETGE:    // !lt
+    case ISD::SETULE:   // !ugt
+    case ISD::SETUGE:   // !ult
+      CC = ISD::getSetCCInverse(CC, true);
+      Negate = true;
+      break;
+    default:
+      break;
+  }
+
+  switch (CC) {
+    case ISD::SETLT:    // swap gt
+    case ISD::SETULT:   // swap ugt
+      CC = ISD::getSetCCSwappedOperands(CC);
+      Swap = true;
+      break;
+    default:
+      break;
+  }
+
+  assert(CC == ISD::SETEQ || CC == ISD::SETGT || CC == ISD::SETUGT);
+
+  MVT ElemTy = VecTy.getVectorElementType();
+  unsigned ElemWidth = ElemTy.getSizeInBits();
+  assert(isPowerOf2_32(ElemWidth));
+
+  auto getIdx = [] (unsigned Code) {
+    static const unsigned Idx[] = { ISD::SETEQ, ISD::SETGT, ISD::SETUGT };
+    for (unsigned I = 0, E = array_lengthof(Idx); I != E; ++I)
+      if (Code == Idx[I])
+        return I;
+    llvm_unreachable("Unhandled CondCode");
+  };
+
+  static unsigned OpcTable[3][3] = {
+    //           SETEQ             SETGT,            SETUGT
+    /* Byte */ { Hexagon::V6_veqb, Hexagon::V6_vgtb, Hexagon::V6_vgtub },
+    /* Half */ { Hexagon::V6_veqh, Hexagon::V6_vgth, Hexagon::V6_vgtuh },
+    /* Word */ { Hexagon::V6_veqw, Hexagon::V6_vgtw, Hexagon::V6_vgtuw }
+  };
+
+  unsigned CmpOpc = OpcTable[Log2_32(ElemWidth)-3][getIdx(CC)];
+
+  MVT ResTy = ty(Op);
+  const SDLoc &dl(Op);
+  SDValue OpL = Swap ? Op.getOperand(1) : Op.getOperand(0);
+  SDValue OpR = Swap ? Op.getOperand(0) : Op.getOperand(1);
+  SDValue CmpV = getNode(CmpOpc, dl, ResTy, {OpL, OpR}, DAG);
+  return Negate ? getNode(Hexagon::V6_pred_not, dl, ResTy, {CmpV}, DAG)
+                : CmpV;
+}
diff --git a/lib/Target/Hexagon/HexagonPatterns.td b/lib/Target/Hexagon/HexagonPatterns.td
index e6c8ad4deea..8ed9a8d01ce 100644
--- a/lib/Target/Hexagon/HexagonPatterns.td
+++ b/lib/Target/Hexagon/HexagonPatterns.td
@@ -88,6 +88,10 @@ def V8I8:   PatLeaf<(v8i8    DoubleRegs:$R)>;
 def V4I16:  PatLeaf<(v4i16   DoubleRegs:$R)>;
 def V2I32:  PatLeaf<(v2i32   DoubleRegs:$R)>;
 
+def HQ8:    PatLeaf<(VecQ8   HvxQR:$R)>;
+def HQ16:   PatLeaf<(VecQ16  HvxQR:$R)>;
+def HQ32:   PatLeaf<(VecQ32  HvxQR:$R)>;
+
 def HVI8:   PatLeaf<(VecI8   HvxVR:$R)>;
 def HVI16:  PatLeaf<(VecI16  HvxVR:$R)>;
 def HVI32:  PatLeaf<(VecI32  HvxVR:$R)>;
@@ -2932,7 +2936,14 @@ let Predicates = [UseHVX] in {
   def: Pat<(sub HVI16:$Vs, HVI16:$Vt),  (V6_vsubh HvxVR:$Vs, HvxVR:$Vt)>;
   def: Pat<(sub HVI32:$Vs, HVI32:$Vt),  (V6_vsubw HvxVR:$Vs, HvxVR:$Vt)>;
 
-  def: Pat<(and HVI8:$Vs, HVI8:$Vt),    (V6_vand HvxVR:$Vs, HvxVR:$Vt)>;
-  def: Pat<(or  HVI8:$Vs, HVI8:$Vt),    (V6_vor  HvxVR:$Vs, HvxVR:$Vt)>;
-  def: Pat<(xor HVI8:$Vs, HVI8:$Vt),    (V6_vxor HvxVR:$Vs, HvxVR:$Vt)>;
+  def: Pat<(and HVI8:$Vs,  HVI8:$Vt),   (V6_vand  HvxVR:$Vs, HvxVR:$Vt)>;
+  def: Pat<(or  HVI8:$Vs,  HVI8:$Vt),   (V6_vor   HvxVR:$Vs, HvxVR:$Vt)>;
+  def: Pat<(xor HVI8:$Vs,  HVI8:$Vt),   (V6_vxor  HvxVR:$Vs, HvxVR:$Vt)>;
+
+  def: Pat<(vselect HQ8:$Qu, HVI8:$Vs, HVI8:$Vt),
+           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;
+  def: Pat<(vselect HQ16:$Qu, HVI16:$Vs, HVI16:$Vt),
+           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;
+  def: Pat<(vselect HQ32:$Qu, HVI32:$Vs, HVI32:$Vt),
+           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;
 }
diff --git a/lib/Target/Hexagon/HexagonRegisterInfo.td b/lib/Target/Hexagon/HexagonRegisterInfo.td
index aa8ecb8d314..2ceed70c249 100644
--- a/lib/Target/Hexagon/HexagonRegisterInfo.td
+++ b/lib/Target/Hexagon/HexagonRegisterInfo.td
@@ -225,6 +225,15 @@ def VecPI16
 def VecPI32
   : ValueTypeByHwMode<[Hvx64,  Hvx64old,  Hvx128,  Hvx128old, DefaultMode],
                       [v32i32, v32i32,    v64i32,  v64i32,    v32i32]>;
+def VecQ8
+  : ValueTypeByHwMode<[Hvx64,  Hvx64old,  Hvx128,  Hvx128old, DefaultMode],
+                      [v64i1,  v64i1,     v128i1,  v128i1,    v64i1]>;
+def VecQ16
+  : ValueTypeByHwMode<[Hvx64,  Hvx64old,  Hvx128,  Hvx128old, DefaultMode],
+                      [v32i1,  v32i1,     v64i1,   v64i1,     v32i1]>;
+def VecQ32
+  : ValueTypeByHwMode<[Hvx64,  Hvx64old,  Hvx128,  Hvx128old, DefaultMode],
+                      [v16i1,  v16i1,     v32i1,   v32i1,     v16i1]>;
 
 // HVX register classes
 
@@ -263,7 +272,8 @@ def HvxWR : RegisterClass<"Hexagon", [VecPI8, VecPI16, VecPI32], 1024,
     [RegInfo<1024,1024,1024>, RegInfo<2048,2048,2048>, RegInfo<1024,1024,1024>]>;
 }
 
-def HvxQR : RegisterClass<"Hexagon", [VecI1], 512, (add Q0, Q1, Q2, Q3)> {
+def HvxQR : RegisterClass<"Hexagon", [VecI1, VecQ8, VecQ16, VecQ32], 512,
+  (add Q0, Q1, Q2, Q3)> {
   let RegInfos = RegInfoByHwMode<[Hvx64, Hvx128, DefaultMode],
     [RegInfo<512,512,512>, RegInfo<1024,1024,1024>, RegInfo<512,512,512>]>;
 }
diff --git a/test/CodeGen/Hexagon/autohvx/vector-compare-128b.ll b/test/CodeGen/Hexagon/autohvx/vector-compare-128b.ll
new file mode 100644
index 00000000000..b4d78d5e50f
--- /dev/null
+++ b/test/CodeGen/Hexagon/autohvx/vector-compare-128b.ll
@@ -0,0 +1,294 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; --- Byte
+
+; CHECK-LABEL: test_00:
+; CHECK: q[[Q000:[0-3]]] = vcmp.eq(v0.b,v1.b)
+; CHECK: v0 = vmux(q[[Q000]],v0,v1)
+define <128 x i8> @test_00(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp eq <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_01:
+; CHECK: q[[Q010:[0-3]]] = vcmp.eq(v0.b,v1.b)
+; CHECK: q[[Q011:[0-9]]] = not(q[[Q010]])
+; CHECK: v0 = vmux(q[[Q011]],v0,v1)
+define <128 x i8> @test_01(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp ne <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_02:
+; CHECK: q[[Q020:[0-3]]] = vcmp.gt(v1.b,v0.b)
+; CHECK: v0 = vmux(q[[Q020]],v0,v1)
+define <128 x i8> @test_02(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp slt <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_03:
+; CHECK: q[[Q030:[0-3]]] = vcmp.gt(v0.b,v1.b)
+; CHECK: q[[Q031:[0-9]]] = not(q[[Q030]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <128 x i8> @test_03(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp sle <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_04:
+; CHECK: q[[Q040:[0-3]]] = vcmp.gt(v0.b,v1.b)
+; CHECK: v0 = vmux(q[[Q040]],v0,v1)
+define <128 x i8> @test_04(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp sgt <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_05:
+; CHECK: q[[Q050:[0-3]]] = vcmp.gt(v1.b,v0.b)
+; CHECK: q[[Q051:[0-9]]] = not(q[[Q050]])
+; CHECK: v0 = vmux(q[[Q051]],v0,v1)
+define <128 x i8> @test_05(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp sge <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_06:
+; CHECK: q[[Q060:[0-3]]] = vcmp.gt(v1.ub,v0.ub)
+; CHECK: v0 = vmux(q[[Q060]],v0,v1)
+define <128 x i8> @test_06(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp ult <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_07:
+; CHECK: q[[Q070:[0-3]]] = vcmp.gt(v0.ub,v1.ub)
+; CHECK: q[[Q071:[0-9]]] = not(q[[Q070]])
+; CHECK: v0 = vmux(q[[Q071]],v0,v1)
+define <128 x i8> @test_07(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp ule <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_08:
+; CHECK: q[[Q080:[0-3]]] = vcmp.gt(v0.ub,v1.ub)
+; CHECK: v0 = vmux(q[[Q080]],v0,v1)
+define <128 x i8> @test_08(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp ugt <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; CHECK-LABEL: test_09:
+; CHECK: q[[Q090:[0-3]]] = vcmp.gt(v1.ub,v0.ub)
+; CHECK: q[[Q091:[0-9]]] = not(q[[Q090]])
+; CHECK: v0 = vmux(q[[Q091]],v0,v1)
+define <128 x i8> @test_09(<128 x i8> %v0, <128 x i8> %v1) #0 {
+  %t0 = icmp uge <128 x i8> %v0, %v1
+  %t1 = select <128 x i1> %t0, <128 x i8> %v0, <128 x i8> %v1
+  ret <128 x i8> %t1
+}
+
+; --- Half
+
+; CHECK-LABEL: test_10:
+; CHECK: q[[Q100:[0-3]]] = vcmp.eq(v0.h,v1.h)
+; CHECK: v0 = vmux(q[[Q100]],v0,v1)
+define <64 x i16> @test_10(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp eq <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_11:
+; CHECK: q[[Q110:[0-3]]] = vcmp.eq(v0.h,v1.h)
+; CHECK: q[[Q111:[0-9]]] = not(q[[Q110]])
+; CHECK: v0 = vmux(q[[Q111]],v0,v1)
+define <64 x i16> @test_11(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp ne <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_12:
+; CHECK: q[[Q120:[0-3]]] = vcmp.gt(v1.h,v0.h)
+; CHECK: v0 = vmux(q[[Q120]],v0,v1)
+define <64 x i16> @test_12(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp slt <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_13:
+; CHECK: q[[Q130:[0-3]]] = vcmp.gt(v0.h,v1.h)
+; CHECK: q[[Q131:[0-9]]] = not(q[[Q130]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <64 x i16> @test_13(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp sle <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_14:
+; CHECK: q[[Q140:[0-3]]] = vcmp.gt(v0.h,v1.h)
+; CHECK: v0 = vmux(q[[Q140]],v0,v1)
+define <64 x i16> @test_14(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp sgt <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_15:
+; CHECK: q[[Q150:[0-3]]] = vcmp.gt(v1.h,v0.h)
+; CHECK: q[[Q151:[0-9]]] = not(q[[Q150]])
+; CHECK: v0 = vmux(q[[Q151]],v0,v1)
+define <64 x i16> @test_15(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp sge <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_16:
+; CHECK: q[[Q160:[0-3]]] = vcmp.gt(v1.uh,v0.uh)
+; CHECK: v0 = vmux(q[[Q160]],v0,v1)
+define <64 x i16> @test_16(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp ult <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_17:
+; CHECK: q[[Q170:[0-3]]] = vcmp.gt(v0.uh,v1.uh)
+; CHECK: q[[Q171:[0-9]]] = not(q[[Q170]])
+; CHECK: v0 = vmux(q[[Q171]],v0,v1)
+define <64 x i16> @test_17(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp ule <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_18:
+; CHECK: q[[Q180:[0-3]]] = vcmp.gt(v0.uh,v1.uh)
+; CHECK: v0 = vmux(q[[Q180]],v0,v1)
+define <64 x i16> @test_18(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp ugt <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; CHECK-LABEL: test_19:
+; CHECK: q[[Q190:[0-3]]] = vcmp.gt(v1.uh,v0.uh)
+; CHECK: q[[Q191:[0-9]]] = not(q[[Q190]])
+; CHECK: v0 = vmux(q[[Q191]],v0,v1)
+define <64 x i16> @test_19(<64 x i16> %v0, <64 x i16> %v1) #0 {
+  %t0 = icmp uge <64 x i16> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i16> %v0, <64 x i16> %v1
+  ret <64 x i16> %t1
+}
+
+; --- Word
+
+; CHECK-LABEL: test_20:
+; CHECK: q[[Q200:[0-3]]] = vcmp.eq(v0.w,v1.w)
+; CHECK: v0 = vmux(q[[Q200]],v0,v1)
+define <32 x i32> @test_20(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp eq <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_21:
+; CHECK: q[[Q210:[0-3]]] = vcmp.eq(v0.w,v1.w)
+; CHECK: q[[Q211:[0-9]]] = not(q[[Q210]])
+; CHECK: v0 = vmux(q[[Q211]],v0,v1)
+define <32 x i32> @test_21(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp ne <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_22:
+; CHECK: q[[Q220:[0-3]]] = vcmp.gt(v1.w,v0.w)
+; CHECK: v0 = vmux(q[[Q220]],v0,v1)
+define <32 x i32> @test_22(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp slt <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_23:
+; CHECK: q[[Q230:[0-3]]] = vcmp.gt(v0.w,v1.w)
+; CHECK: q[[Q231:[0-9]]] = not(q[[Q230]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <32 x i32> @test_23(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp sle <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_24:
+; CHECK: q[[Q240:[0-3]]] = vcmp.gt(v0.w,v1.w)
+; CHECK: v0 = vmux(q[[Q240]],v0,v1)
+define <32 x i32> @test_24(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp sgt <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_25:
+; CHECK: q[[Q250:[0-3]]] = vcmp.gt(v1.w,v0.w)
+; CHECK: q[[Q251:[0-9]]] = not(q[[Q250]])
+; CHECK: v0 = vmux(q[[Q251]],v0,v1)
+define <32 x i32> @test_25(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp sge <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_26:
+; CHECK: q[[Q260:[0-3]]] = vcmp.gt(v1.uw,v0.uw)
+; CHECK: v0 = vmux(q[[Q260]],v0,v1)
+define <32 x i32> @test_26(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp ult <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_27:
+; CHECK: q[[Q270:[0-3]]] = vcmp.gt(v0.uw,v1.uw)
+; CHECK: q[[Q271:[0-9]]] = not(q[[Q270]])
+; CHECK: v0 = vmux(q[[Q271]],v0,v1)
+define <32 x i32> @test_27(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp ule <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_28:
+; CHECK: q[[Q280:[0-3]]] = vcmp.gt(v0.uw,v1.uw)
+; CHECK: v0 = vmux(q[[Q280]],v0,v1)
+define <32 x i32> @test_28(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp ugt <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+; CHECK-LABEL: test_29:
+; CHECK: q[[Q290:[0-3]]] = vcmp.gt(v1.uw,v0.uw)
+; CHECK: q[[Q291:[0-9]]] = not(q[[Q290]])
+; CHECK: v0 = vmux(q[[Q291]],v0,v1)
+define <32 x i32> @test_29(<32 x i32> %v0, <32 x i32> %v1) #0 {
+  %t0 = icmp uge <32 x i32> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i32> %v0, <32 x i32> %v1
+  ret <32 x i32> %t1
+}
+
+attributes #0 = { nounwind readnone "target-cpu"="hexagonv60" "target-features"="+hvx,+hvx-length128b" }
diff --git a/test/CodeGen/Hexagon/autohvx/vector-compare-64b.ll b/test/CodeGen/Hexagon/autohvx/vector-compare-64b.ll
new file mode 100644
index 00000000000..4c0e19791d4
--- /dev/null
+++ b/test/CodeGen/Hexagon/autohvx/vector-compare-64b.ll
@@ -0,0 +1,294 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; --- Byte
+
+; CHECK-LABEL: test_00:
+; CHECK: q[[Q000:[0-3]]] = vcmp.eq(v0.b,v1.b)
+; CHECK: v0 = vmux(q[[Q000]],v0,v1)
+define <64 x i8> @test_00(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp eq <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_01:
+; CHECK: q[[Q010:[0-3]]] = vcmp.eq(v0.b,v1.b)
+; CHECK: q[[Q011:[0-9]]] = not(q[[Q010]])
+; CHECK: v0 = vmux(q[[Q011]],v0,v1)
+define <64 x i8> @test_01(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp ne <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_02:
+; CHECK: q[[Q020:[0-3]]] = vcmp.gt(v1.b,v0.b)
+; CHECK: v0 = vmux(q[[Q020]],v0,v1)
+define <64 x i8> @test_02(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp slt <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_03:
+; CHECK: q[[Q030:[0-3]]] = vcmp.gt(v0.b,v1.b)
+; CHECK: q[[Q031:[0-9]]] = not(q[[Q030]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <64 x i8> @test_03(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp sle <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_04:
+; CHECK: q[[Q040:[0-3]]] = vcmp.gt(v0.b,v1.b)
+; CHECK: v0 = vmux(q[[Q040]],v0,v1)
+define <64 x i8> @test_04(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp sgt <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_05:
+; CHECK: q[[Q050:[0-3]]] = vcmp.gt(v1.b,v0.b)
+; CHECK: q[[Q051:[0-9]]] = not(q[[Q050]])
+; CHECK: v0 = vmux(q[[Q051]],v0,v1)
+define <64 x i8> @test_05(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp sge <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_06:
+; CHECK: q[[Q060:[0-3]]] = vcmp.gt(v1.ub,v0.ub)
+; CHECK: v0 = vmux(q[[Q060]],v0,v1)
+define <64 x i8> @test_06(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp ult <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_07:
+; CHECK: q[[Q070:[0-3]]] = vcmp.gt(v0.ub,v1.ub)
+; CHECK: q[[Q071:[0-9]]] = not(q[[Q070]])
+; CHECK: v0 = vmux(q[[Q071]],v0,v1)
+define <64 x i8> @test_07(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp ule <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_08:
+; CHECK: q[[Q080:[0-3]]] = vcmp.gt(v0.ub,v1.ub)
+; CHECK: v0 = vmux(q[[Q080]],v0,v1)
+define <64 x i8> @test_08(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp ugt <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; CHECK-LABEL: test_09:
+; CHECK: q[[Q090:[0-3]]] = vcmp.gt(v1.ub,v0.ub)
+; CHECK: q[[Q091:[0-9]]] = not(q[[Q090]])
+; CHECK: v0 = vmux(q[[Q091]],v0,v1)
+define <64 x i8> @test_09(<64 x i8> %v0, <64 x i8> %v1) #0 {
+  %t0 = icmp uge <64 x i8> %v0, %v1
+  %t1 = select <64 x i1> %t0, <64 x i8> %v0, <64 x i8> %v1
+  ret <64 x i8> %t1
+}
+
+; --- Half
+
+; CHECK-LABEL: test_10:
+; CHECK: q[[Q100:[0-3]]] = vcmp.eq(v0.h,v1.h)
+; CHECK: v0 = vmux(q[[Q100]],v0,v1)
+define <32 x i16> @test_10(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp eq <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_11:
+; CHECK: q[[Q110:[0-3]]] = vcmp.eq(v0.h,v1.h)
+; CHECK: q[[Q111:[0-9]]] = not(q[[Q110]])
+; CHECK: v0 = vmux(q[[Q111]],v0,v1)
+define <32 x i16> @test_11(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp ne <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_12:
+; CHECK: q[[Q120:[0-3]]] = vcmp.gt(v1.h,v0.h)
+; CHECK: v0 = vmux(q[[Q120]],v0,v1)
+define <32 x i16> @test_12(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp slt <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_13:
+; CHECK: q[[Q130:[0-3]]] = vcmp.gt(v0.h,v1.h)
+; CHECK: q[[Q131:[0-9]]] = not(q[[Q130]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <32 x i16> @test_13(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp sle <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_14:
+; CHECK: q[[Q140:[0-3]]] = vcmp.gt(v0.h,v1.h)
+; CHECK: v0 = vmux(q[[Q140]],v0,v1)
+define <32 x i16> @test_14(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp sgt <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_15:
+; CHECK: q[[Q150:[0-3]]] = vcmp.gt(v1.h,v0.h)
+; CHECK: q[[Q151:[0-9]]] = not(q[[Q150]])
+; CHECK: v0 = vmux(q[[Q151]],v0,v1)
+define <32 x i16> @test_15(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp sge <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_16:
+; CHECK: q[[Q160:[0-3]]] = vcmp.gt(v1.uh,v0.uh)
+; CHECK: v0 = vmux(q[[Q160]],v0,v1)
+define <32 x i16> @test_16(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp ult <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_17:
+; CHECK: q[[Q170:[0-3]]] = vcmp.gt(v0.uh,v1.uh)
+; CHECK: q[[Q171:[0-9]]] = not(q[[Q170]])
+; CHECK: v0 = vmux(q[[Q171]],v0,v1)
+define <32 x i16> @test_17(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp ule <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_18:
+; CHECK: q[[Q180:[0-3]]] = vcmp.gt(v0.uh,v1.uh)
+; CHECK: v0 = vmux(q[[Q180]],v0,v1)
+define <32 x i16> @test_18(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp ugt <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; CHECK-LABEL: test_19:
+; CHECK: q[[Q190:[0-3]]] = vcmp.gt(v1.uh,v0.uh)
+; CHECK: q[[Q191:[0-9]]] = not(q[[Q190]])
+; CHECK: v0 = vmux(q[[Q191]],v0,v1)
+define <32 x i16> @test_19(<32 x i16> %v0, <32 x i16> %v1) #0 {
+  %t0 = icmp uge <32 x i16> %v0, %v1
+  %t1 = select <32 x i1> %t0, <32 x i16> %v0, <32 x i16> %v1
+  ret <32 x i16> %t1
+}
+
+; --- Word
+
+; CHECK-LABEL: test_20:
+; CHECK: q[[Q200:[0-3]]] = vcmp.eq(v0.w,v1.w)
+; CHECK: v0 = vmux(q[[Q200]],v0,v1)
+define <16 x i32> @test_20(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp eq <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_21:
+; CHECK: q[[Q210:[0-3]]] = vcmp.eq(v0.w,v1.w)
+; CHECK: q[[Q211:[0-9]]] = not(q[[Q210]])
+; CHECK: v0 = vmux(q[[Q211]],v0,v1)
+define <16 x i32> @test_21(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp ne <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_22:
+; CHECK: q[[Q220:[0-3]]] = vcmp.gt(v1.w,v0.w)
+; CHECK: v0 = vmux(q[[Q220]],v0,v1)
+define <16 x i32> @test_22(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp slt <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_23:
+; CHECK: q[[Q230:[0-3]]] = vcmp.gt(v0.w,v1.w)
+; CHECK: q[[Q231:[0-9]]] = not(q[[Q230]])
+; CHECK: v0 = vmux(q[[Q031]],v0,v1)
+define <16 x i32> @test_23(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp sle <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_24:
+; CHECK: q[[Q240:[0-3]]] = vcmp.gt(v0.w,v1.w)
+; CHECK: v0 = vmux(q[[Q240]],v0,v1)
+define <16 x i32> @test_24(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp sgt <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_25:
+; CHECK: q[[Q250:[0-3]]] = vcmp.gt(v1.w,v0.w)
+; CHECK: q[[Q251:[0-9]]] = not(q[[Q250]])
+; CHECK: v0 = vmux(q[[Q251]],v0,v1)
+define <16 x i32> @test_25(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp sge <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_26:
+; CHECK: q[[Q260:[0-3]]] = vcmp.gt(v1.uw,v0.uw)
+; CHECK: v0 = vmux(q[[Q260]],v0,v1)
+define <16 x i32> @test_26(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp ult <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_27:
+; CHECK: q[[Q270:[0-3]]] = vcmp.gt(v0.uw,v1.uw)
+; CHECK: q[[Q271:[0-9]]] = not(q[[Q270]])
+; CHECK: v0 = vmux(q[[Q271]],v0,v1)
+define <16 x i32> @test_27(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp ule <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_28:
+; CHECK: q[[Q280:[0-3]]] = vcmp.gt(v0.uw,v1.uw)
+; CHECK: v0 = vmux(q[[Q280]],v0,v1)
+define <16 x i32> @test_28(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp ugt <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+; CHECK-LABEL: test_29:
+; CHECK: q[[Q290:[0-3]]] = vcmp.gt(v1.uw,v0.uw)
+; CHECK: q[[Q291:[0-9]]] = not(q[[Q290]])
+; CHECK: v0 = vmux(q[[Q291]],v0,v1)
+define <16 x i32> @test_29(<16 x i32> %v0, <16 x i32> %v1) #0 {
+  %t0 = icmp uge <16 x i32> %v0, %v1
+  %t1 = select <16 x i1> %t0, <16 x i32> %v0, <16 x i32> %v1
+  ret <16 x i32> %t1
+}
+
+attributes #0 = { nounwind readnone "target-cpu"="hexagonv60" "target-features"="+hvx,+hvx-length64b" }