Vendor import of llvm trunk r162107:vendor/llvm/llvm-trunk-r162107

http://llvm.org/svn/llvm-project/llvm/trunk@162107

5 files changed, 154 insertions, 28 deletions

diff --git a/lib/Target/X86/X86.td b/lib/Target/X86/X86.td
index 6c1a816c9ff2..18e6b7c3d9b6 100644
--- a/lib/Target/X86/X86.td
+++ b/lib/Target/X86/X86.td
@@ -17,14 +17,14 @@
 include "llvm/Target/Target.td"
 
 //===----------------------------------------------------------------------===//
-// X86 Subtarget state.
+// X86 Subtarget state
 //
 
 def Mode64Bit : SubtargetFeature<"64bit-mode", "In64BitMode", "true",
                                   "64-bit mode (x86_64)">;
 
 //===----------------------------------------------------------------------===//
-// X86 Subtarget features.
+// X86 Subtarget features
 //===----------------------------------------------------------------------===//
 
 def FeatureCMOV    : SubtargetFeature<"cmov","HasCMov", "true",
@@ -97,7 +97,7 @@ def FeatureFMA4    : SubtargetFeature<"fma4", "HasFMA4", "true",
                                       [FeatureAVX, FeatureSSE4A]>;
 def FeatureXOP     : SubtargetFeature<"xop", "HasXOP", "true",
                                       "Enable XOP instructions",
-                                      [FeatureAVX, FeatureSSE4A]>;
+                                      [FeatureFMA4]>;
 def FeatureVectorUAMem : SubtargetFeature<"vector-unaligned-mem",
                                           "HasVectorUAMem", "true",
                  "Allow unaligned memory operands on vector/SIMD instructions">;
@@ -226,7 +226,7 @@ def : Proc<"bdver1",          [FeatureXOP, FeatureFMA4, FeatureCMPXCHG16B,
 def : Proc<"bdver2",          [FeatureXOP, FeatureFMA4, FeatureCMPXCHG16B,
                                FeatureAES, FeaturePCLMUL,
                                FeatureF16C, FeatureLZCNT,
-                               FeaturePOPCNT, FeatureBMI]>;
+                               FeaturePOPCNT, FeatureBMI, FeatureFMA]>;
 
 def : Proc<"winchip-c6",      [FeatureMMX]>;
 def : Proc<"winchip2",        [Feature3DNow]>;
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ea66a6115d7e..c77355f91796 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5114,6 +5114,82 @@ X86TargetLowering::LowerVectorBroadcast(SDValue &Op, SelectionDAG &DAG) const {
   return SDValue();
 }
 
+// LowerVectorFpExtend - Recognize the scalarized FP_EXTEND from v2f32 to v2f64
+// and convert it into X86ISD::VFPEXT due to the current ISD::FP_EXTEND has the
+// constraint of matching input/output vector elements.
+SDValue
+X86TargetLowering::LowerVectorFpExtend(SDValue &Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  SDNode *N = Op.getNode();
+  EVT VT = Op.getValueType();
+  unsigned NumElts = Op.getNumOperands();
+
+  // Check supported types and sub-targets.
+  //
+  // Only v2f32 -> v2f64 needs special handling.
+  if (VT != MVT::v2f64 || !Subtarget->hasSSE2())
+    return SDValue();
+
+  SDValue VecIn;
+  EVT VecInVT;
+  SmallVector<int, 8> Mask;
+  EVT SrcVT = MVT::Other;
+
+  // Check the patterns could be translated into X86vfpext.
+  for (unsigned i = 0; i < NumElts; ++i) {
+    SDValue In = N->getOperand(i);
+    unsigned Opcode = In.getOpcode();
+
+    // Skip if the element is undefined.
+    if (Opcode == ISD::UNDEF) {
+      Mask.push_back(-1);
+      continue;
+    }
+
+    // Quit if one of the elements is not defined from 'fpext'.
+    if (Opcode != ISD::FP_EXTEND)
+      return SDValue();
+
+    // Check how the source of 'fpext' is defined.
+    SDValue L2In = In.getOperand(0);
+    EVT L2InVT = L2In.getValueType();
+
+    // Check the original type
+    if (SrcVT == MVT::Other)
+      SrcVT = L2InVT;
+    else if (SrcVT != L2InVT) // Quit if non-homogenous typed.
+      return SDValue();
+
+    // Check whether the value being 'fpext'ed is extracted from the same
+    // source.
+    Opcode = L2In.getOpcode();
+
+    // Quit if it's not extracted with a constant index.
+    if (Opcode != ISD::EXTRACT_VECTOR_ELT ||
+        !isa<ConstantSDNode>(L2In.getOperand(1)))
+      return SDValue();
+
+    SDValue ExtractedFromVec = L2In.getOperand(0);
+
+    if (VecIn.getNode() == 0) {
+      VecIn = ExtractedFromVec;
+      VecInVT = ExtractedFromVec.getValueType();
+    } else if (VecIn != ExtractedFromVec) // Quit if built from more than 1 vec.
+      return SDValue();
+
+    Mask.push_back(cast<ConstantSDNode>(L2In.getOperand(1))->getZExtValue());
+  }
+
+  // Fill the remaining mask as undef.
+  for (unsigned i = NumElts; i < VecInVT.getVectorNumElements(); ++i)
+    Mask.push_back(-1);
+
+  return DAG.getNode(X86ISD::VFPEXT, DL, VT,
+                     DAG.getVectorShuffle(VecInVT, DL,
+                                          VecIn, DAG.getUNDEF(VecInVT),
+                                          &Mask[0]));
+}
+
 SDValue
 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
   DebugLoc dl = Op.getDebugLoc();
@@ -5146,6 +5222,10 @@ X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
   if (Broadcast.getNode())
     return Broadcast;
 
+  SDValue FpExt = LowerVectorFpExtend(Op, DAG);
+  if (FpExt.getNode())
+    return FpExt;
+
   unsigned EVTBits = ExtVT.getSizeInBits();
 
   unsigned NumZero  = 0;
@@ -11122,9 +11202,9 @@ static void ReplaceATOMIC_LOAD(SDNode *Node,
   Results.push_back(Swap.getValue(1));
 }
 
-void X86TargetLowering::
+static void
 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
-                        SelectionDAG &DAG, unsigned NewOp) const {
+                        SelectionDAG &DAG, unsigned NewOp) {
   DebugLoc dl = Node->getDebugLoc();
   assert (Node->getValueType(0) == MVT::i64 &&
           "Only know how to expand i64 atomics");
@@ -11245,26 +11325,40 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
     return;
   }
   case ISD::ATOMIC_LOAD_ADD:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
-    return;
   case ISD::ATOMIC_LOAD_AND:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
-    return;
   case ISD::ATOMIC_LOAD_NAND:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
-    return;
   case ISD::ATOMIC_LOAD_OR:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
-    return;
   case ISD::ATOMIC_LOAD_SUB:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
-    return;
   case ISD::ATOMIC_LOAD_XOR:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
-    return;
-  case ISD::ATOMIC_SWAP:
-    ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
+  case ISD::ATOMIC_SWAP: {
+    unsigned Opc;
+    switch (N->getOpcode()) {
+    default: llvm_unreachable("Unexpected opcode");
+    case ISD::ATOMIC_LOAD_ADD:
+      Opc = X86ISD::ATOMADD64_DAG;
+      break;
+    case ISD::ATOMIC_LOAD_AND:
+      Opc = X86ISD::ATOMAND64_DAG;
+      break;
+    case ISD::ATOMIC_LOAD_NAND:
+      Opc = X86ISD::ATOMNAND64_DAG;
+      break;
+    case ISD::ATOMIC_LOAD_OR:
+      Opc = X86ISD::ATOMOR64_DAG;
+      break;
+    case ISD::ATOMIC_LOAD_SUB:
+      Opc = X86ISD::ATOMSUB64_DAG;
+      break;
+    case ISD::ATOMIC_LOAD_XOR:
+      Opc = X86ISD::ATOMXOR64_DAG;
+      break;
+    case ISD::ATOMIC_SWAP:
+      Opc = X86ISD::ATOMSWAP64_DAG;
+      break;
+    }
+    ReplaceATOMIC_BINARY_64(N, Results, DAG, Opc);
     return;
+  }
   case ISD::ATOMIC_LOAD:
     ReplaceATOMIC_LOAD(N, Results, DAG);
   }
@@ -11342,7 +11436,9 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::ATOMAND64_DAG:      return "X86ISD::ATOMAND64_DAG";
   case X86ISD::ATOMNAND64_DAG:     return "X86ISD::ATOMNAND64_DAG";
   case X86ISD::VZEXT_MOVL:         return "X86ISD::VZEXT_MOVL";
+  case X86ISD::VSEXT_MOVL:         return "X86ISD::VSEXT_MOVL";
   case X86ISD::VZEXT_LOAD:         return "X86ISD::VZEXT_LOAD";
+  case X86ISD::VFPEXT:             return "X86ISD::VFPEXT";
   case X86ISD::VSHLDQ:             return "X86ISD::VSHLDQ";
   case X86ISD::VSRLDQ:             return "X86ISD::VSRLDQ";
   case X86ISD::VSHL:               return "X86ISD::VSHL";
@@ -12792,16 +12888,31 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
     // String/text processing lowering.
   case X86::PCMPISTRM128REG:
   case X86::VPCMPISTRM128REG:
-    return EmitPCMP(MI, BB, 3, false /* in-mem */);
   case X86::PCMPISTRM128MEM:
   case X86::VPCMPISTRM128MEM:
-    return EmitPCMP(MI, BB, 3, true /* in-mem */);
   case X86::PCMPESTRM128REG:
   case X86::VPCMPESTRM128REG:
-    return EmitPCMP(MI, BB, 5, false /* in mem */);
   case X86::PCMPESTRM128MEM:
-  case X86::VPCMPESTRM128MEM:
-    return EmitPCMP(MI, BB, 5, true /* in mem */);
+  case X86::VPCMPESTRM128MEM: {
+    unsigned NumArgs;
+    bool MemArg;
+    switch (MI->getOpcode()) {
+    default: llvm_unreachable("illegal opcode!");
+    case X86::PCMPISTRM128REG:
+    case X86::VPCMPISTRM128REG:
+      NumArgs = 3; MemArg = false; break;
+    case X86::PCMPISTRM128MEM:
+    case X86::VPCMPISTRM128MEM:
+      NumArgs = 3; MemArg = true; break;
+    case X86::PCMPESTRM128REG:
+    case X86::VPCMPESTRM128REG:
+      NumArgs = 5; MemArg = false; break;
+    case X86::PCMPESTRM128MEM:
+    case X86::VPCMPESTRM128MEM:
+      NumArgs = 5; MemArg = true; break;
+    }
+    return EmitPCMP(MI, BB, NumArgs, MemArg);
+  }
 
     // Thread synchronization.
   case X86::MONITOR:
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 9123ebd8ae49..896d067fda75 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -227,6 +227,9 @@ namespace llvm {
       // VSEXT_MOVL - Vector move low and sign extend.
       VSEXT_MOVL,
 
+      // VFPEXT - Vector FP extend.
+      VFPEXT,
+
       // VSHL, VSRL - 128-bit vector logical left / right shift
       VSHLDQ, VSRLDQ,
 
@@ -828,6 +831,8 @@ namespace llvm {
     SDValue LowerVectorBroadcast(SDValue &Op, SelectionDAG &DAG) const;
     SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const;
 
+    SDValue LowerVectorFpExtend(SDValue &Op, SelectionDAG &DAG) const;
+
     virtual SDValue
       LowerFormalArguments(SDValue Chain,
                            CallingConv::ID CallConv, bool isVarArg,
@@ -859,9 +864,6 @@ namespace llvm {
                    const SmallVectorImpl<ISD::OutputArg> &Outs,
                    LLVMContext &Context) const;
 
-    void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results,
-                                 SelectionDAG &DAG, unsigned NewOp) const;
-
     /// Utility function to emit string processing sse4.2 instructions
     /// that return in xmm0.
     /// This takes the instruction to expand, the associated machine basic
diff --git a/lib/Target/X86/X86InstrFragmentsSIMD.td b/lib/Target/X86/X86InstrFragmentsSIMD.td
index d13167bb05db..1db68c86b76d 100644
--- a/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -81,6 +81,11 @@ def X86vsmovl  : SDNode<"X86ISD::VSEXT_MOVL",
 
 def X86vzload  : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
                         [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
+
+def X86vfpext  : SDNode<"X86ISD::VFPEXT",
+                        SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
+                                             SDTCisFP<0>, SDTCisFP<1>]>>;
+
 def X86vshldq  : SDNode<"X86ISD::VSHLDQ",    SDTIntShiftOp>;
 def X86vshrdq  : SDNode<"X86ISD::VSRLDQ",    SDTIntShiftOp>;
 def X86cmpp    : SDNode<"X86ISD::CMPP",      SDTX86VFCMP>;
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index e4c35b9bc556..20dc81eb4a37 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -2101,12 +2101,20 @@ let Predicates = [HasAVX] in {
   def : Pat<(v4f32 (fround (loadv4f64 addr:$src))),
             (VCVTPD2PSYrm addr:$src)>;
 
+  def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
+            (VCVTPS2PDrr VR128:$src)>;
   def : Pat<(v4f64 (fextend (v4f32 VR128:$src))),
             (VCVTPS2PDYrr VR128:$src)>;
   def : Pat<(v4f64 (fextend (loadv4f32 addr:$src))),
             (VCVTPS2PDYrm addr:$src)>;
 }
 
+let Predicates = [HasSSE2] in {
+  // Match fextend for 128 conversions
+  def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
+            (CVTPS2PDrr VR128:$src)>;
+}
+
 //===----------------------------------------------------------------------===//
 // SSE 1 & 2 - Compare Instructions
 //===----------------------------------------------------------------------===//