Merge "Update aosp/master LLVM for rebase to r230699."

1 files changed, 345 insertions, 135 deletions

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 2f95b74..4707279 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -102,11 +102,9 @@ EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
   return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
 }
 
-AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
-  TargetLowering(TM) {
-
-  Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
-
+AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM,
+                                           const AMDGPUSubtarget &STI)
+    : TargetLowering(TM), Subtarget(&STI) {
   setOperationAction(ISD::Constant, MVT::i32, Legal);
   setOperationAction(ISD::Constant, MVT::i64, Legal);
   setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
@@ -127,12 +125,21 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::FABS,   MVT::f32, Legal);
   setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
   setOperationAction(ISD::FRINT,  MVT::f32, Legal);
-  setOperationAction(ISD::FROUND, MVT::f32, Legal);
   setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
 
+  setOperationAction(ISD::FROUND, MVT::f32, Custom);
+  setOperationAction(ISD::FROUND, MVT::f64, Custom);
+
   setOperationAction(ISD::FREM, MVT::f32, Custom);
   setOperationAction(ISD::FREM, MVT::f64, Custom);
 
+  // v_mad_f32 does not support denormals according to some sources.
+  if (!Subtarget->hasFP32Denormals())
+    setOperationAction(ISD::FMAD, MVT::f32, Legal);
+
+  // Expand to fneg + fadd.
+  setOperationAction(ISD::FSUB, MVT::f64, Expand);
+
   // Lower floating point store/load to integer store/load to reduce the number
   // of patterns in tablegen.
   setOperationAction(ISD::STORE, MVT::f32, Promote);
@@ -141,9 +148,6 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::STORE, MVT::v2f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
 
-  setOperationAction(ISD::STORE, MVT::i64, Promote);
-  AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
-
   setOperationAction(ISD::STORE, MVT::v4f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
 
@@ -162,9 +166,6 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   // Custom lowering of vector stores is required for local address space
   // stores.
   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
-  // XXX: Native v2i32 local address space stores are possible, but not
-  // currently implemented.
-  setOperationAction(ISD::STORE, MVT::v2i32, Custom);
 
   setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
   setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
@@ -187,9 +188,6 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
 
-  setOperationAction(ISD::LOAD, MVT::i64, Promote);
-  AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
-
   setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
 
@@ -216,18 +214,28 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
 
-  setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
-  setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
-  setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
-  setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
-  setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
-  setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
-  setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
-  setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
-  setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
-  setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
-  setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
-  setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
+  // There are no 64-bit extloads. These should be done as a 32-bit extload and
+  // an extension to 64-bit.
+  for (MVT VT : MVT::integer_valuetypes()) {
+    setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
+    setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
+  }
+
+  for (MVT VT : MVT::integer_vector_valuetypes()) {
+    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
+    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
+    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
+    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
+    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
+    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
+    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
+    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
+  }
 
   setOperationAction(ISD::BR_CC, MVT::i1, Expand);
 
@@ -246,7 +254,8 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
 
   setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
 
-  setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
 
@@ -382,6 +391,12 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setTargetDAGCombine(ISD::SELECT_CC);
   setTargetDAGCombine(ISD::STORE);
 
+  setTargetDAGCombine(ISD::FADD);
+  setTargetDAGCombine(ISD::FSUB);
+
+  setBooleanContents(ZeroOrNegativeOneBooleanContent);
+  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
+
   setSchedulingPreference(Sched::RegPressure);
   setJumpIsExpensive(true);
 
@@ -397,6 +412,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   // large sequence of instructions.
   setIntDivIsCheap(false);
   setPow2SDivIsCheap(false);
+  setFsqrtIsCheap(true);
 
   // FIXME: Need to really handle these.
   MaxStoresPerMemcpy  = 4096;
@@ -429,6 +445,29 @@ bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
   return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
 }
 
+bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
+                                                 ISD::LoadExtType,
+                                                 EVT NewVT) const {
+
+  unsigned NewSize = NewVT.getStoreSizeInBits();
+
+  // If we are reducing to a 32-bit load, this is always better.
+  if (NewSize == 32)
+    return true;
+
+  EVT OldVT = N->getValueType(0);
+  unsigned OldSize = OldVT.getStoreSizeInBits();
+
+  // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
+  // extloads, so doing one requires using a buffer_load. In cases where we
+  // still couldn't use a scalar load, using the wider load shouldn't really
+  // hurt anything.
+
+  // If the old size already had to be an extload, there's no harm in continuing
+  // to reduce the width.
+  return (OldSize < 32);
+}
+
 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
                                                    EVT CastTy) const {
   if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
@@ -442,6 +481,18 @@ bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
           (LScalarSize < 32));
 }
 
+// SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
+// profitable with the expansion for 64-bit since it's generally good to
+// speculate things.
+// FIXME: These should really have the size as a parameter.
+bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
+  return true;
+}
+
+bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
+  return true;
+}
+
 //===---------------------------------------------------------------------===//
 // Target Properties
 //===---------------------------------------------------------------------===//
@@ -560,6 +611,7 @@ SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
   case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
   case ISD::FRINT: return LowerFRINT(Op, DAG);
   case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
+  case ISD::FROUND: return LowerFROUND(Op, DAG);
   case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
   case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
   case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
@@ -619,7 +671,7 @@ SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
                                                        const SDValue &InitPtr,
                                                        SDValue Chain,
                                                        SelectionDAG &DAG) const {
-  const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
+  const DataLayout *TD = getDataLayout();
   SDLoc DL(InitPtr);
   Type *InitTy = Init->getType();
 
@@ -707,7 +759,7 @@ SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
                                                  SDValue Op,
                                                  SelectionDAG &DAG) const {
 
-  const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
+  const DataLayout *TD = getDataLayout();
   GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
   const GlobalValue *GV = G->getGlobal();
 
@@ -810,8 +862,7 @@ SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
                                               SelectionDAG &DAG) const {
 
   MachineFunction &MF = DAG.getMachineFunction();
-  const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
-      getTargetMachine().getSubtargetImpl()->getFrameLowering());
+  const AMDGPUFrameLowering *TFL = Subtarget->getFrameLowering();
 
   FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
 
@@ -866,10 +917,9 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     }
 
     case Intrinsic::AMDGPU_div_fmas:
-      // FIXME: Dropping bool parameter. Work is needed to support the implicit
-      // read from VCC.
       return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
-                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
+                         Op.getOperand(4));
 
     case Intrinsic::AMDGPU_div_fixup:
       return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
@@ -889,7 +939,19 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
       return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
 
     case Intrinsic::AMDGPU_rsq_clamped:
-      return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
+      if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+        Type *Type = VT.getTypeForEVT(*DAG.getContext());
+        APFloat Max = APFloat::getLargest(Type->getFltSemantics());
+        APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
+
+        SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
+        SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
+                                  DAG.getConstantFP(Max, VT));
+        return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
+                           DAG.getConstantFP(Min, VT));
+      } else {
+        return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
+      }
 
     case Intrinsic::AMDGPU_ldexp:
       return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
@@ -962,6 +1024,10 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     case AMDGPUIntrinsic::AMDGPU_brev:
       return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
 
+  case Intrinsic::AMDGPU_class:
+    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
+                       Op.getOperand(1), Op.getOperand(2));
+
     case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
       return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
 
@@ -1000,17 +1066,21 @@ SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
 }
 
 /// \brief Generate Min/Max node
-SDValue AMDGPUTargetLowering::CombineFMinMax(SDLoc DL,
-                                             EVT VT,
-                                             SDValue LHS,
-                                             SDValue RHS,
-                                             SDValue True,
-                                             SDValue False,
-                                             SDValue CC,
-                                             SelectionDAG &DAG) const {
+SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
+                                                   EVT VT,
+                                                   SDValue LHS,
+                                                   SDValue RHS,
+                                                   SDValue True,
+                                                   SDValue False,
+                                                   SDValue CC,
+                                                   DAGCombinerInfo &DCI) const {
+  if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
+    return SDValue();
+
   if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
     return SDValue();
 
+  SelectionDAG &DAG = DCI.DAG;
   ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
   switch (CCOpcode) {
   case ISD::SETOEQ:
@@ -1027,27 +1097,47 @@ SDValue AMDGPUTargetLowering::CombineFMinMax(SDLoc DL,
   case ISD::SETO:
     break;
   case ISD::SETULE:
-  case ISD::SETULT:
+  case ISD::SETULT: {
+    if (LHS == True)
+      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
+    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
+  }
   case ISD::SETOLE:
   case ISD::SETOLT:
   case ISD::SETLE:
   case ISD::SETLT: {
+    // Ordered. Assume ordered for undefined.
+
+    // Only do this after legalization to avoid interfering with other combines
+    // which might occur.
+    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
+        !DCI.isCalledByLegalizer())
+      return SDValue();
+
     // We need to permute the operands to get the correct NaN behavior. The
     // selected operand is the second one based on the failing compare with NaN,
     // so permute it based on the compare type the hardware uses.
     if (LHS == True)
-      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
-    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
+      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
+    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
+  }
+  case ISD::SETUGE:
+  case ISD::SETUGT: {
+    if (LHS == True)
+      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
+    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
   }
   case ISD::SETGT:
   case ISD::SETGE:
-  case ISD::SETUGE:
   case ISD::SETOGE:
-  case ISD::SETUGT:
   case ISD::SETOGT: {
+    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
+        !DCI.isCalledByLegalizer())
+      return SDValue();
+
     if (LHS == True)
-      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
-    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
+      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
+    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
   }
   case ISD::SETCC_INVALID:
     llvm_unreachable("Invalid setcc condcode!");
@@ -1330,24 +1420,6 @@ SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
   EVT MemVT = Load->getMemoryVT();
 
-  if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
-    // We can do the extload to 32-bits, and then need to separately extend to
-    // 64-bits.
-
-    SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
-                                       Load->getChain(),
-                                       Load->getBasePtr(),
-                                       MemVT,
-                                       Load->getMemOperand());
-
-    SDValue Ops[] = {
-      DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
-      ExtLoad32.getValue(1)
-    };
-
-    return DAG.getMergeValues(Ops, DL);
-  }
-
   if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
     assert(VT == MVT::i1 && "Only i1 non-extloads expected");
     // FIXME: Copied from PPC
@@ -1586,12 +1658,26 @@ void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
   SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
   SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
 
+  if (VT == MVT::i64 &&
+    DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
+    DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
+
+    SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
+                              LHS_Lo, RHS_Lo);
+
+    SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(0), zero);
+    SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(1), zero);
+    Results.push_back(DIV);
+    Results.push_back(REM);
+    return;
+  }
+
   // Get Speculative values
   SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
   SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
 
-  SDValue REM_Hi = zero;
   SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
+  SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, zero);
 
   SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
   SDValue DIV_Lo = zero;
@@ -1599,8 +1685,10 @@ void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
   const unsigned halfBitWidth = HalfVT.getSizeInBits();
 
   for (unsigned i = 0; i < halfBitWidth; ++i) {
-    SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
-    // Get Value of high bit
+    const unsigned bitPos = halfBitWidth - i - 1;
+    SDValue POS = DAG.getConstant(bitPos, HalfVT);
+    // Get value of high bit
+    // TODO: Remove the BFE part when the optimization is fixed
     SDValue HBit;
     if (halfBitWidth == 32 && Subtarget->hasBFE()) {
       HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
@@ -1608,33 +1696,23 @@ void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
       HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
       HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
     }
+    HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
 
-    SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
-      DAG.getConstant(halfBitWidth - 1, HalfVT));
-    REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
-    REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
-
-    REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
-    REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
+    // Shift
+    REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, VT));
+    // Add LHS high bit
+    REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
 
-
-    SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
-
-    SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
+    SDValue BIT = DAG.getConstant(1 << bitPos, HalfVT);
     SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
 
     DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
 
     // Update REM
-
     SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
-
     REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
-    REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
-    REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
   }
 
-  SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
   SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
   Results.push_back(DIV);
   Results.push_back(REM);
@@ -1655,8 +1733,8 @@ SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
   SDValue Den = Op.getOperand(1);
 
   if (VT == MVT::i32) {
-    if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
-        DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
+    if (DAG.MaskedValueIsZero(Num, APInt::getHighBitsSet(32, 8)) &&
+        DAG.MaskedValueIsZero(Den, APInt::getHighBitsSet(32, 8))) {
       // TODO: We technically could do this for i64, but shouldn't that just be
       // handled by something generally reducing 64-bit division on 32-bit
       // values to 32-bit?
@@ -1768,19 +1846,31 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
   SDValue LHS = Op.getOperand(0);
   SDValue RHS = Op.getOperand(1);
 
-  if (VT == MVT::i32) {
-    if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
-        DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
-      // TODO: We technically could do this for i64, but shouldn't that just be
-      // handled by something generally reducing 64-bit division on 32-bit
-      // values to 32-bit?
-      return LowerDIVREM24(Op, DAG, true);
-    }
-  }
-
   SDValue Zero = DAG.getConstant(0, VT);
   SDValue NegOne = DAG.getConstant(-1, VT);
 
+  if (VT == MVT::i32 &&
+      DAG.ComputeNumSignBits(LHS) > 8 &&
+      DAG.ComputeNumSignBits(RHS) > 8) {
+    return LowerDIVREM24(Op, DAG, true);
+  }
+  if (VT == MVT::i64 &&
+      DAG.ComputeNumSignBits(LHS) > 32 &&
+      DAG.ComputeNumSignBits(RHS) > 32) {
+    EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
+
+    //HiLo split
+    SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
+    SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
+    SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
+                                 LHS_Lo, RHS_Lo);
+    SDValue Res[2] = {
+      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
+      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
+    };
+    return DAG.getMergeValues(Res, DL);
+  }
+
   SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
   SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
   SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
@@ -1845,6 +1935,20 @@ SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
 }
 
+static SDValue extractF64Exponent(SDValue Hi, SDLoc SL, SelectionDAG &DAG) {
+  const unsigned FractBits = 52;
+  const unsigned ExpBits = 11;
+
+  SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
+                                Hi,
+                                DAG.getConstant(FractBits - 32, MVT::i32),
+                                DAG.getConstant(ExpBits, MVT::i32));
+  SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
+                            DAG.getConstant(1023, MVT::i32));
+
+  return Exp;
+}
+
 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Src = Op.getOperand(0);
@@ -1860,16 +1964,9 @@ SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
   // exponent.
   SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
 
-  const unsigned FractBits = 52;
-  const unsigned ExpBits = 11;
+  SDValue Exp = extractF64Exponent(Hi, SL, DAG);
 
-  // Extract the exponent.
-  SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
-                                Hi,
-                                DAG.getConstant(FractBits - 32, MVT::i32),
-                                DAG.getConstant(ExpBits, MVT::i32));
-  SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
-                            DAG.getConstant(1023, MVT::i32));
+  const unsigned FractBits = 52;
 
   // Extract the sign bit.
   const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
@@ -1932,6 +2029,99 @@ SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) con
   return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
 }
 
+// XXX - May require not supporting f32 denormals?
+SDValue AMDGPUTargetLowering::LowerFROUND32(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc SL(Op);
+  SDValue X = Op.getOperand(0);
+
+  SDValue T = DAG.getNode(ISD::FTRUNC, SL, MVT::f32, X);
+
+  SDValue Diff = DAG.getNode(ISD::FSUB, SL, MVT::f32, X, T);
+
+  SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, MVT::f32, Diff);
+
+  const SDValue Zero = DAG.getConstantFP(0.0, MVT::f32);
+  const SDValue One = DAG.getConstantFP(1.0, MVT::f32);
+  const SDValue Half = DAG.getConstantFP(0.5, MVT::f32);
+
+  SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f32, One, X);
+
+  EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
+
+  SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
+
+  SDValue Sel = DAG.getNode(ISD::SELECT, SL, MVT::f32, Cmp, SignOne, Zero);
+
+  return DAG.getNode(ISD::FADD, SL, MVT::f32, T, Sel);
+}
+
+SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc SL(Op);
+  SDValue X = Op.getOperand(0);
+
+  SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
+
+  const SDValue Zero = DAG.getConstant(0, MVT::i32);
+  const SDValue One = DAG.getConstant(1, MVT::i32);
+  const SDValue NegOne = DAG.getConstant(-1, MVT::i32);
+  const SDValue FiftyOne = DAG.getConstant(51, MVT::i32);
+  EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
+
+
+  SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
+
+  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
+
+  SDValue Exp = extractF64Exponent(Hi, SL, DAG);
+
+  const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), MVT::i64);
+
+  SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
+  SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
+                          DAG.getConstant(INT64_C(0x0008000000000000), MVT::i64),
+                          Exp);
+
+  SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
+  SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
+                              DAG.getConstant(0, MVT::i64), Tmp0,
+                              ISD::SETNE);
+
+  SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
+                             D, DAG.getConstant(0, MVT::i64));
+  SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
+
+  K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
+  K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
+
+  SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
+  SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
+  SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
+
+  SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
+                            ExpEqNegOne,
+                            DAG.getConstantFP(1.0, MVT::f64),
+                            DAG.getConstantFP(0.0, MVT::f64));
+
+  SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
+
+  K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
+  K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
+
+  return K;
+}
+
+SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+
+  if (VT == MVT::f32)
+    return LowerFROUND32(Op, DAG);
+
+  if (VT == MVT::f64)
+    return LowerFROUND64(Op, DAG);
+
+  llvm_unreachable("unhandled type");
+}
+
 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Src = Op.getOperand(0);
@@ -2155,7 +2345,8 @@ SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
   SDValue Value = SN->getValue();
   EVT VT = Value.getValueType();
 
-  if (isTypeLegal(VT) || SN->isVolatile() || !ISD::isNormalLoad(Value.getNode()))
+  if (isTypeLegal(VT) || SN->isVolatile() ||
+      !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
     return SDValue();
 
   LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
@@ -2231,27 +2422,9 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
       simplifyI24(N1, DCI);
       return SDValue();
     }
-  case ISD::SELECT_CC: {
-    SDLoc DL(N);
-    EVT VT = N->getValueType(0);
-
-    if (VT == MVT::f32 ||
-        (VT == MVT::f64 &&
-         Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
-      SDValue LHS = N->getOperand(0);
-      SDValue RHS = N->getOperand(1);
-      SDValue True = N->getOperand(2);
-      SDValue False = N->getOperand(3);
-      SDValue CC = N->getOperand(4);
-
-      return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
-    }
-
-    break;
-  }
   case ISD::SELECT: {
     SDValue Cond = N->getOperand(0);
-    if (Cond.getOpcode() == ISD::SETCC) {
+    if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
       SDLoc DL(N);
       EVT VT = N->getValueType(0);
       SDValue LHS = Cond.getOperand(0);
@@ -2261,11 +2434,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
       SDValue True = N->getOperand(1);
       SDValue False = N->getOperand(2);
 
-      if (VT == MVT::f32 ||
-          (VT == MVT::f64 &&
-           Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
-        return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
-      }
+      if (VT == MVT::f32)
+        return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
 
       // TODO: Implement min / max Evergreen instructions.
       if (VT == MVT::i32 &&
@@ -2451,7 +2621,6 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(DWORDADDR)
   NODE_NAME_CASE(FRACT)
   NODE_NAME_CASE(CLAMP)
-  NODE_NAME_CASE(MAD)
   NODE_NAME_CASE(FMAX_LEGACY)
   NODE_NAME_CASE(SMAX)
   NODE_NAME_CASE(UMAX)
@@ -2474,6 +2643,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(RSQ_LEGACY)
   NODE_NAME_CASE(RSQ_CLAMPED)
   NODE_NAME_CASE(LDEXP)
+  NODE_NAME_CASE(FP_CLASS)
   NODE_NAME_CASE(DOT4)
   NODE_NAME_CASE(BFE_U32)
   NODE_NAME_CASE(BFE_I32)
@@ -2505,6 +2675,46 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   }
 }
 
+SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
+                                               DAGCombinerInfo &DCI,
+                                               unsigned &RefinementSteps,
+                                               bool &UseOneConstNR) const {
+  SelectionDAG &DAG = DCI.DAG;
+  EVT VT = Operand.getValueType();
+
+  if (VT == MVT::f32) {
+    RefinementSteps = 0;
+    return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
+  }
+
+  // TODO: There is also f64 rsq instruction, but the documentation is less
+  // clear on its precision.
+
+  return SDValue();
+}
+
+SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
+                                               DAGCombinerInfo &DCI,
+                                               unsigned &RefinementSteps) const {
+  SelectionDAG &DAG = DCI.DAG;
+  EVT VT = Operand.getValueType();
+
+  if (VT == MVT::f32) {
+    // Reciprocal, < 1 ulp error.
+    //
+    // This reciprocal approximation converges to < 0.5 ulp error with one
+    // newton rhapson performed with two fused multiple adds (FMAs).
+
+    RefinementSteps = 0;
+    return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
+  }
+
+  // TODO: There is also f64 rcp instruction, but the documentation is less
+  // clear on its precision.
+
+  return SDValue();
+}
+
 static void computeKnownBitsForMinMax(const SDValue Op0,
                                       const SDValue Op1,
                                       APInt &KnownZero,