[X86] -- Fix fptoui i64 conversions for IA32 (performance and correctness)

[mbodart]

Summary

Fixes assertion failures when using fptosi or fptoui with f80 and AVX512. The operation actions need to be left as Custom, not Legal. The legality for f32/f64 is handled by having FP_TO_INTHelper do nothing for those cases.

Suppresses use of the MSVC ftol2 library function for fptoui i64. That function performs a conversion to *signed* i64, so the results were incorrect for source values >= 2^63. I didn't rip out the now dead references to ftol2, but that can be done with a small followup change set if desired.

Implements an inline sequence for fptoui i64 for 32-bit X86. This is mostly in FP_TO_INTHelper, replacing the ftol2 usage on Windows, and replacing the calls to fixuns{sf,df,xf}di for non-windows. Improves performance by 6X under SSE3, 3X otherwise.

[mbodart]

Hi Michael,

Just a ping to remind you of this review request.

I selected you as you were involved in the original ftol2 introduction, which I am disabling.

If you have suggestions for other reviewers, please pass them along.

thanks,

• mitch

[mbodart]

Per Michael Spencer's suggestion, added Nadav as a reviewer.

Also adding Michael Kuperstein, who consulted on these changes.

[mkuper]

lib/Target/X86/X86ISelLowering.cpp

26289

I'm ok with just deleting the FTOL code immediately, unless anyone objects.
(Can be a follow-up patch)

test/CodeGen/X86/scalar-fp-to-i64.ll

18

Can you use -mattr here instead of an explicit -mcpu?

test/CodeGen/X86/win_ftol2.ll

4

What happened to this RUN line?

[majnemer]

lib/Target/X86/X86ISelLowering.cpp

26289

I think it makes sense to remove it.

[mbodart]

lib/Target/X86/X86ISelLowering.cpp

26289

I'm fine with removing it, but need a little guidance as to the depth of removal.

Certainly isIntegerTypeFTOL and isTargetFTOL, and their use in FP_TO_INTHelper, can be removed.

But further deletion of the WIN_FTOL instructions in the .td files would prevent any future use of that library routine. How can I be certain that isn't being used, or won't be needed?

test/CodeGen/X86/scalar-fp-to-i64.ll

18

I see both -mcpu and -mattr being used for skx and avx512 testing in lit tests.
-mcpu seemed simpler and more "bullet proof" if you will in that it captures
all relevant attributes (sse, sse2, ... avx, skx).

But if -mattr is the preferred method, I can certainly change to that.

What is the BKM?

test/CodeGen/X86/win_ftol2.ll

4

Oops, forgot to reset it before submitting the review.
I was playing around with different ways to disable this test.
It's left as RUN in my current version.

[mbodart]

I uploaded a new revision, removing all usage of isTargetFTOL and isIntegerTypeFTOL, which I think is sufficient for this change set. I left the WIN_FTOL instructions, as conceivably those could still be used for conversion to signed int64, though they aren't currently.

I'm still waiting on a recommendation as to whether lit tests should lean towards using -mattr instead of -mcpu.
Please let me know your preference, and whether that applies to all RUN lines. Specifically, what would be the appropriate -mattr replacement for -mcpu=generic?

thanks!
Mitch Bodart

[majnemer]

lib/Target/X86/X86ISelLowering.cpp

12399

Could we use ANY_EXTEND for High32? We will shift out the top 32-bits anyway.

12697

Please format the return onto its own line.

[mbodart]

Applied fixes for David's recent comments (using ANY_EXTEND, and formatting).

[mbodart]

lib/Target/X86/X86ISelLowering.cpp

12399

Yes, thanks for the suggestion!

Fixed in latest upload.

12697

Fixed, both here and up above in LowerFP_TO_SINT, where this was copied from.

[mbodart]

Updated new test to use -mattr instead of -mcpu, per MK's recommendation.

With that I believe all review comments have been addressed.

Can I please get an LGTM?

thanks,
Mitch

[mkuper]

LGTM

[mkuper]

I'll rip out the rest of the FTOL stuff in a follow-up commit, no reason to leave dead code lying around.

lib/Target/X86/X86ISelLowering.h

public:
 
     /// Return true if the specified scalar FP type is computed in an SSE
     /// register, not on the X87 floating point stack.
     bool isScalarFPTypeInSSEReg(EVT VT) const {
       return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
       (VT == MVT::f32 && X86ScalarSSEf32);   // f32 is when SSE1
     }
 
-    /// Return true if the target uses the MSVC _ftol2 routine for fptoui.
-    bool isTargetFTOL() const;
-
-    /// Return true if the MSVC _ftol2 routine should be used for fptoui to the
-    /// given type.
-    bool isIntegerTypeFTOL(EVT VT) const {
-      return isTargetFTOL() && VT == MVT::i64;
-    }
-
     /// \brief Returns true if it is beneficial to convert a load of a constant
     /// to just the constant itself.
     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                            Type *Ty) const override;
 
     /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
     /// with this index.
     bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;

lib/Target/X86/X86ISelLowering.cpp

if (Subtarget->isTargetKnownWindowsMSVC()) {
     setLibcallName(RTLIB::SREM_I64, "_allrem");
     setLibcallName(RTLIB::UREM_I64, "_aullrem");
     setLibcallName(RTLIB::MUL_I64, "_allmul");
     setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
     setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
     setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
     setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
     setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
-
-    // The _ftol2 runtime function has an unusual calling conv, which
-    // is modeled by a special pseudo-instruction.
-    setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
-    setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
-    setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
-    setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
   }
 
   if (Subtarget->isTargetDarwin()) {
     // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
     setUseUnderscoreSetJmp(false);
     setUseUnderscoreLongJmp(false);
   } else if (Subtarget->isTargetWindowsGNU()) {
     // MS runtime is weird: it exports _setjmp, but longjmp!
X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
 
   // Handle FP_TO_UINT by promoting the destination to a larger signed
   // conversion.
   setOperationAction(ISD::FP_TO_UINT       , MVT::i1   , Promote);
   setOperationAction(ISD::FP_TO_UINT       , MVT::i8   , Promote);
   setOperationAction(ISD::FP_TO_UINT       , MVT::i16  , Promote);
 
   if (Subtarget->is64Bit()) {
-    setOperationAction(ISD::FP_TO_UINT     , MVT::i64  , Expand);
-    setOperationAction(ISD::FP_TO_UINT     , MVT::i32  , Promote);
+    if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
+      // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
+      setOperationAction(ISD::FP_TO_UINT   , MVT::i32  , Custom);
+      setOperationAction(ISD::FP_TO_UINT   , MVT::i64  , Custom);
+    } else {
+      setOperationAction(ISD::FP_TO_UINT   , MVT::i32  , Promote);
+      setOperationAction(ISD::FP_TO_UINT   , MVT::i64  , Expand);
+    }
   } else if (!Subtarget->useSoftFloat()) {
     // Since AVX is a superset of SSE3, only check for SSE here.
     if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
       // Expand FP_TO_UINT into a select.
       // FIXME: We would like to use a Custom expander here eventually to do
       // the optimal thing for SSE vs. the default expansion in the legalizer.
       setOperationAction(ISD::FP_TO_UINT   , MVT::i32  , Expand);
     else
+      // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
       // With SSE3 we can use fisttpll to convert to a signed i64; without
       // SSE, we're stuck with a fistpll.
       setOperationAction(ISD::FP_TO_UINT   , MVT::i32  , Custom);
-  }
 
-  if (isTargetFTOL()) {
-    // Use the _ftol2 runtime function, which has a pseudo-instruction
-    // to handle its weird calling convention.
     setOperationAction(ISD::FP_TO_UINT     , MVT::i64  , Custom);
   }
 
   // TODO: when we have SSE, these could be more efficient, by using movd/movq.
   if (!X86ScalarSSEf64) {
     setOperationAction(ISD::BITCAST        , MVT::f32  , Expand);
     setOperationAction(ISD::BITCAST        , MVT::i32  , Expand);
     if (Subtarget->is64Bit()) {
if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
     setOperationAction(ISD::FSUB,               MVT::v8f64, Legal);
     setOperationAction(ISD::FMUL,               MVT::v8f64, Legal);
     setOperationAction(ISD::FDIV,               MVT::v8f64, Legal);
     setOperationAction(ISD::FSQRT,              MVT::v8f64, Legal);
     setOperationAction(ISD::FNEG,               MVT::v8f64, Custom);
     setOperationAction(ISD::FMA,                MVT::v8f64, Legal);
     setOperationAction(ISD::FMA,                MVT::v16f32, Legal);
 
-    setOperationAction(ISD::FP_TO_SINT,         MVT::i32, Legal);
-    setOperationAction(ISD::FP_TO_UINT,         MVT::i32, Legal);
+    // FIXME:  [US]INT_TO_FP are not legal for f80.
     setOperationAction(ISD::SINT_TO_FP,         MVT::i32, Legal);
     setOperationAction(ISD::UINT_TO_FP,         MVT::i32, Legal);
     if (Subtarget->is64Bit()) {
-      setOperationAction(ISD::FP_TO_UINT,       MVT::i64, Legal);
-      setOperationAction(ISD::FP_TO_SINT,       MVT::i64, Legal);
       setOperationAction(ISD::SINT_TO_FP,       MVT::i64, Legal);
       setOperationAction(ISD::UINT_TO_FP,       MVT::i64, Legal);
     }
     setOperationAction(ISD::FP_TO_SINT,         MVT::v16i32, Legal);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v16i32, Legal);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v8i32, Legal);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v4i32, Legal);
     setOperationAction(ISD::SINT_TO_FP,         MVT::v16i32, Legal);
SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
                                  FudgePtr, MachinePointerInfo::getConstantPool(),
                                  MVT::f32, false, false, false, 4);
   // Extend everything to 80 bits to force it to be done on x87.
   SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
   return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
                      DAG.getIntPtrConstant(0, dl));
 }
 
+// If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
+// is legal, or has an f16 source (which needs to be promoted to f32),
+// just return an <SDValue(), SDValue()> pair.
+// Otherwise it is assumed to be a conversion from one of f32, f64 or f80
+// to i16, i32 or i64, and we lower it to a legal sequence.
+// If lowered to the final integer result we return a <result, SDValue()> pair.
+// Otherwise we lower it to a sequence ending with a FIST, return a
+// <FIST, StackSlot> pair, and the caller is responsible for loading
+// the final integer result from StackSlot.
 std::pair<SDValue,SDValue>
-X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
-                                    bool IsSigned, bool IsReplace) const {
+X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
   SDLoc DL(Op);
 
   EVT DstTy = Op.getValueType();
+  EVT TheVT = Op.getOperand(0).getValueType();
   auto PtrVT = getPointerTy(DAG.getDataLayout());
 
-  if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
+  if (TheVT == MVT::f16)
+    // We need to promote the f16 to f32 before using the lowering
+    // in this routine.
+    return std::make_pair(SDValue(), SDValue());
+
+  assert((TheVT == MVT::f32 ||
+          TheVT == MVT::f64 ||
+          TheVT == MVT::f80) &&
+         "Unexpected FP operand type in FP_TO_INTHelper");
+
+  // If using FIST to compute an unsigned i64, we'll need some fixup
+  // to handle values above the maximum signed i64.  A FIST is always
+  // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
+  bool UnsignedFixup = !IsSigned &&
+                       DstTy == MVT::i64 &&
+                       (!Subtarget->is64Bit() ||
+                        !isScalarFPTypeInSSEReg(TheVT));
+
+  if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
+    // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
+    // The low 32 bits of the fist result will have the correct uint32 result.
     assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
     DstTy = MVT::i64;
   }
 
   assert(DstTy.getSimpleVT() <= MVT::i64 &&
          DstTy.getSimpleVT() >= MVT::i16 &&
          "Unknown FP_TO_INT to lower!");
 
   // These are really Legal.
   if (DstTy == MVT::i32 &&
       isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
     return std::make_pair(SDValue(), SDValue());
   if (Subtarget->is64Bit() &&
       DstTy == MVT::i64 &&
       isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
     return std::make_pair(SDValue(), SDValue());
 
-  // We lower FP->int64 either into FISTP64 followed by a load from a temporary
-  // stack slot, or into the FTOL runtime function.
+  // We lower FP->int64 into FISTP64 followed by a load from a temporary
+  // stack slot.
   MachineFunction &MF = DAG.getMachineFunction();
   unsigned MemSize = DstTy.getSizeInBits()/8;
   int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
   SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
 
   unsigned Opc;
-  if (!IsSigned && isIntegerTypeFTOL(DstTy))
-    Opc = X86ISD::WIN_FTOL;
-  else
-    switch (DstTy.getSimpleVT().SimpleTy) {
-    default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
-    case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
-    case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
-    case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
-    }
 
   SDValue Chain = DAG.getEntryNode();
   SDValue Value = Op.getOperand(0);
-  EVT TheVT = Op.getOperand(0).getValueType();
+  SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
+
+  if (UnsignedFixup) {
+    //
+    // Conversion to unsigned i64 is implemented with a select,
+    // depending on whether the source value fits in the range
+    // of a signed i64.  Let Thresh be the FP equivalent of
+    // 0x8000000000000000ULL.
+    //
+    //  Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
+    //  FistSrc    = (Value < Thresh) ? Value : (Value - Thresh);
+    //  Fist-to-mem64 FistSrc
+    //  Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
+    //  to XOR'ing the high 32 bits with Adjust.
+    //
+    // Being a power of 2, Thresh is exactly representable in all FP formats.
+    // For X87 we'd like to use the smallest FP type for this constant, but
+    // for DAG type consistency we have to match the FP operand type.
+
+    APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
+    APFloat::opStatus Status = APFloat::opOK;
+    bool LosesInfo = false;
+    if (TheVT == MVT::f64)
+      // The rounding mode is irrelevant as the conversion should be exact.
+      Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
+                              &LosesInfo);
+    else if (TheVT == MVT::f80)
+      Status = Thresh.convert(APFloat::x87DoubleExtended,
+                              APFloat::rmNearestTiesToEven, &LosesInfo);
+
+    assert(Status == APFloat::opOK && !LosesInfo &&
+           "FP conversion should have been exact");
+
+    SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
+
+    SDValue Cmp = DAG.getSetCC(DL,
+                               getSetCCResultType(DAG.getDataLayout(),
+                                                  *DAG.getContext(), TheVT),
+                               Value, ThreshVal, ISD::SETLT);
+    Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
+                           DAG.getConstant(0, DL, MVT::i32),
+                           DAG.getConstant(0x80000000, DL, MVT::i32));
+    SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
+    Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
+                                              *DAG.getContext(), TheVT),
+                       Value, ThreshVal, ISD::SETLT);
+    Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
+  }
+
   // FIXME This causes a redundant load/store if the SSE-class value is already
   // in memory, such as if it is on the callstack.
   if (isScalarFPTypeInSSEReg(TheVT)) {
     assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
     Chain = DAG.getStore(Chain, DL, Value, StackSlot,
                          MachinePointerInfo::getFixedStack(SSFI),
                          false, false, 0);
     SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
if (isScalarFPTypeInSSEReg(TheVT)) {
     SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
     StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
   }
 
   MachineMemOperand *MMO =
     MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
                             MachineMemOperand::MOStore, MemSize, MemSize);
 
-  if (Opc != X86ISD::WIN_FTOL) {
+  if (UnsignedFixup) {
+
+    // Insert the FIST, load its result as two i32's,
+    // and XOR the high i32 with Adjust.
+
+    SDValue FistOps[] = { Chain, Value, StackSlot };
+    SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
+                                           FistOps, DstTy, MMO);
+
+    SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
+                                MachinePointerInfo(),
+                                false, false, false, 0);
+    SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
+                                   DAG.getConstant(4, DL, PtrVT));
+
+    SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
+                                 MachinePointerInfo(),
+                                 false, false, false, 0);
+    High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
+
+    if (Subtarget->is64Bit()) {
+      // Join High32 and Low32 into a 64-bit result.
+      // (High32 << 32) | Low32
+      Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
+      High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
+      High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
+                           DAG.getConstant(32, DL, MVT::i8));
+      SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
+      return std::make_pair(Result, SDValue());
+    }
+
+    SDValue ResultOps[] = { Low32, High32 };
+
+    SDValue pair = IsReplace
+      ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
+      : DAG.getMergeValues(ResultOps, DL);
+    return std::make_pair(pair, SDValue());
+  } else {
     // Build the FP_TO_INT*_IN_MEM
     SDValue Ops[] = { Chain, Value, StackSlot };
     SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
                                            Ops, DstTy, MMO);
     return std::make_pair(FIST, StackSlot);
-  } else {
-    SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
-      DAG.getVTList(MVT::Other, MVT::Glue),
-      Chain, Value);
-    SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
-      MVT::i32, ftol.getValue(1));
-    SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
-      MVT::i32, eax.getValue(2));
-    SDValue Ops[] = { eax, edx };
-    SDValue pair = IsReplace
-      ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
-      : DAG.getMergeValues(Ops, DL);
-    return std::make_pair(pair, SDValue());
   }
 }
 
 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
                               const X86Subtarget *Subtarget) {
   MVT VT = Op->getSimpleValueType(0);
   SDValue In = Op->getOperand(0);
   MVT InVT = In.getSimpleValueType();

 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
                                            SelectionDAG &DAG) const {
   assert(!Op.getSimpleValueType().isVector());
 
   std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
     /*IsSigned=*/ true, /*IsReplace=*/ false);
   SDValue FIST = Vals.first, StackSlot = Vals.second;
   // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
-  if (!FIST.getNode()) return Op;
+  if (!FIST.getNode())
+    return Op;
 
   if (StackSlot.getNode())
     // Load the result.
     return DAG.getLoad(Op.getValueType(), SDLoc(Op),
                        FIST, StackSlot, MachinePointerInfo(),
                        false, false, false, 0);
 
   // The node is the result.
   return FIST;
 }
 
 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
                                            SelectionDAG &DAG) const {
   std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
     /*IsSigned=*/ false, /*IsReplace=*/ false);
   SDValue FIST = Vals.first, StackSlot = Vals.second;
-  assert(FIST.getNode() && "Unexpected failure");
+  // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
+  if (!FIST.getNode())
+    return Op;
 
   if (StackSlot.getNode())
     // Load the result.
     return DAG.getLoad(Op.getValueType(), SDLoc(Op),
                        FIST, StackSlot, MachinePointerInfo(),
                        false, false, false, 0);
 
   // The node is the result.
void X86TargetLowering::ReplaceNodeResults(SDNode *N,
   case ISD::UREM:
   case ISD::SDIVREM:
   case ISD::UDIVREM: {
     SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
     Results.push_back(V);
     return;
   }
   case ISD::FP_TO_SINT:
-    // FP_TO_INT*_IN_MEM is not legal for f16 inputs.  Do not convert
-    // (FP_TO_SINT (load f16)) to FP_TO_INT*.
-    if (N->getOperand(0).getValueType() == MVT::f16)
-      break;
-    // fallthrough
   case ISD::FP_TO_UINT: {
     bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
 
-    if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
-      return;
-
     std::pair<SDValue,SDValue> Vals =
         FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
     SDValue FIST = Vals.first, StackSlot = Vals.second;
     if (FIST.getNode()) {
       EVT VT = N->getValueType(0);
       // Return a load from the stack slot.
       if (StackSlot.getNode())
         Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
   // E.g., on Haswell:
   // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
   // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
   if (isLegalAddressingMode(DL, AM, Ty, AS))
     // Scale represents reg2 * scale, thus account for 1
     // as soon as we use a second register.
     return AM.Scale != 0;
   return -1;
 }
-
-bool X86TargetLowering::isTargetFTOL() const {
-  return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();
-}

test/CodeGen/X86/pr17631.ll

define <8 x float> @foo(<8 x float> %y, i64* %p, double %x) {
   %i = fptoui double %x to i64
   store i64 %i, i64* %p
   %ret = fadd <8 x float> %y, %y
   ret <8 x float> %ret
 }
 
 ; CHECK: foo
 ; CHECK-NOT: vzeroupper
-; CHECK: _ftol2
+; CHECK: {{cvtt|fist}}
 ; CHECK: ret

test/CodeGen/X86/scalar-fp-to-i64.ll

+; Check that scalar FP conversions to signed and unsigned int64 are using
+; reasonable sequences, across platforms and target switches.
+;
+; The signed case is straight forward, and the tests here basically
+; ensure successful compilation (f80 with avx512 was broken at one point).
+;
+; For the unsigned case there are many possible sequences, so to avoid
+; a fragile test we just check for the presence of a few key instructions.
+; AVX512 on Intel64 can use vcvtts[ds]2usi directly for float and double.
+; Otherwise the sequence will involve an FP subtract (fsub, subss or subsd),
+; and a truncating conversion (cvtts[ds]2si, fisttp, or fnstcw+fist).  When
+; both a subtract and fnstcw are needed, they can occur in either order.
+;
+; The interesting subtargets are AVX512F (vcvtts[ds]2usi), SSE3 (fisttp),
+; SSE2 (cvtts[ds]2si) and vanilla X87 (fnstcw+fist, 32-bit only).
+;
+; RUN: llc < %s -mtriple=i386-pc-windows-msvc     -mcpu=skx      | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512_32
+; RUN: llc < %s -mtriple=i386-unknown-linux-gnu   -mcpu=skx      | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512_32
+; RUN: llc < %s -mtriple=x86_64-pc-windows-msvc   -mcpu=skx      | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512_64
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=skx      | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512_64
+; RUN: llc < %s -mtriple=i386-pc-windows-msvc     -mcpu=core2    | FileCheck %s --check-prefix=CHECK --check-prefix=SSE3_32
+; RUN: llc < %s -mtriple=i386-unknown-linux-gnu   -mcpu=core2    | FileCheck %s --check-prefix=CHECK --check-prefix=SSE3_32
+; RUN: llc < %s -mtriple=x86_64-pc-windows-msvc   -mcpu=core2    | FileCheck %s --check-prefix=CHECK --check-prefix=SSE3_64
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=core2    | FileCheck %s --check-prefix=CHECK --check-prefix=SSE3_64
+; RUN: llc < %s -mtriple=i386-pc-windows-msvc     -mcpu=pentium4 | FileCheck %s --check-prefix=CHECK --check-prefix=SSE2_32
+; RUN: llc < %s -mtriple=i386-unknown-linux-gnu   -mcpu=pentium4 | FileCheck %s --check-prefix=CHECK --check-prefix=SSE2_32
+; RUN: llc < %s -mtriple=x86_64-pc-windows-msvc   -mcpu=pentium4 | FileCheck %s --check-prefix=CHECK --check-prefix=SSE2_64
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=pentium4 | FileCheck %s --check-prefix=CHECK --check-prefix=SSE2_64
+; RUN: llc < %s -mtriple=i386-pc-windows-msvc     -mcpu=generic  | FileCheck %s --check-prefix=CHECK --check-prefix=X87
+; RUN: llc < %s -mtriple=i386-unknown-linux-gnu   -mcpu=generic  | FileCheck %s --check-prefix=CHECK --check-prefix=X87
+
+; CHECK-LABEL: f_to_u64
+; X87-DAG: fsub
+; X87-DAG: fnstcw
+; X87: fist
+; SSE2_32-DAG: {{subss|fsub}}
+; SSE2_32-DAG: fnstcw
+; SSE2_32: fist
+; SSE2_64: subss
+; SSE2_64: cvttss2si
+; SSE3_32: {{subss|fsub}}
+; SSE3_32: fistt
+; SSE3_64: subss
+; SSE3_64: cvttss2si
+; AVX512_32: {{subss|fsub}}
+; AVX512_32: fistt
+; AVX512_64: vcvttss2usi
+; CHECK: ret
+define i64 @f_to_u64(float %a) nounwind {
+  %r = fptoui float %a to i64
+  ret i64 %r
+}
+
+; CHECK-LABEL: f_to_s64
+; X87: fnstcw
+; X87: fist
+; SSE2_32: fnstcw
+; SSE2_32: fist
+; SSE2_64: cvttss2si
+; SSE3_32: fistt
+; SSE3_64: cvttss2si
+; AVX512_32: fistt
+; AVX512_64: vcvttss2si
+; CHECK: ret
+define i64 @f_to_s64(float %a) nounwind {
+  %r = fptosi float %a to i64
+  ret i64 %r
+}
+
+; CHECK-LABEL: d_to_u64
+; X87-DAG: fsub
+; X87-DAG: fnstcw
+; X87: fist
+; SSE2_32-DAG: {{subsd|fsub}}
+; SSE2_32-DAG: fnstcw
+; SSE2_32: fist
+; SSE2_64: subsd
+; SSE2_64: cvttsd2si
+; SSE3_32: {{subsd|fsub}}
+; SSE3_32: fistt
+; SSE3_64: subsd
+; SSE3_64: cvttsd2si
+; AVX512_32: {{subsd|fsub}}
+; AVX512_32: fistt
+; AVX512_64: vcvttsd2usi
+; CHECK: ret
+define i64 @d_to_u64(double %a) nounwind {
+  %r = fptoui double %a to i64
+  ret i64 %r
+}
+
+; CHECK-LABEL: d_to_s64
+; X87: fnstcw
+; X87: fist
+; SSE2_32: fnstcw
+; SSE2_32: fist
+; SSE2_64: cvttsd2si
+; SSE3_32: fistt
+; SSE3_64: cvttsd2si
+; AVX512_32: fistt
+; AVX512_64: vcvttsd2si
+; CHECK: ret
+define i64 @d_to_s64(double %a) nounwind {
+  %r = fptosi double %a to i64
+  ret i64 %r
+}
+
+; CHECK-LABEL: x_to_u64
+; CHECK-DAG: fsub
+; X87-DAG: fnstcw
+; SSE2_32-DAG: fnstcw
+; SSE2_64-DAG: fnstcw
+; CHECK: fist
+; CHECK: ret
+define i64 @x_to_u64(x86_fp80 %a) nounwind {
+  %r = fptoui x86_fp80 %a to i64
+  ret i64 %r
+}
+
+; CHECK-LABEL: x_to_s64
+; X87: fnstcw
+; X87: fist
+; SSE2_32: fnstcw
+; SSE2_32: fist
+; SSE2_64: fnstcw
+; SSE2_64: fist
+; SSE3_32: fistt
+; SSE3_64: fistt
+; AVX512_32: fistt
+; AVX512_64: fistt
+; CHECK: ret
+define i64 @x_to_s64(x86_fp80 %a) nounwind {
+  %r = fptosi x86_fp80 %a to i64
+  ret i64 %r
+}

test/CodeGen/X86/win_ftol2.ll

-; RUN: llc < %s -mtriple=i686-pc-win32 -mcpu=generic | FileCheck %s -check-prefix=FTOL
+; RUN: llc < %s -mtriple=i686-pc-win32 -mcpu=generic | FileCheck %s -check-prefix=COMPILERRT
 ; RUN: llc < %s -mtriple=i686-pc-mingw32 | FileCheck %s -check-prefix=COMPILERRT
 ; RUN: llc < %s -mtriple=i686-pc-linux | FileCheck %s -check-prefix=COMPILERRT
 ; RUN: llc < %s -mtriple=x86_64-pc-win32 | FileCheck %s -check-prefix=COMPILERRT
 ; RUN: llc < %s -mtriple=x86_64-pc-mingw32 | FileCheck %s -check-prefix=COMPILERRT
 ; RUN: llc < %s -mtriple=x86_64-pc-linux | FileCheck %s -check-prefix=COMPILERRT
-; RUN: llc < %s -mattr=-sse -O0 -mtriple=i686-pc-win32 | FileCheck %s -check-prefix=FTOL_2
+; RUN: llc < %s -mattr=-sse -O0 -mtriple=i686-pc-win32 | FileCheck %s -check-prefix=COMPILERRT
 
-; Win32 targets use the MSVCRT _ftol2 runtime function for fptoui to i64. This
-; function has a nonstandard calling convention: the input value is expected on
-; the x87 stack instead of the callstack. The input value is popped by the
-; callee. Mingw32 uses normal cdecl compiler-rt functions.
+; This test originally used the FTOL and FTOL_2 checks for the i686-pc-win32
+; triples, under the assumption that Win32 targets should use the MSVCRT
+; _ftol2 runtime function for fptoui to i64.  That usage was incorrect,
+; as _ftol2 performs conversion to signed i64.  As of the compiler fix,
+; the FTOL/FTOL_2 checks are no longer used, which basically renders
+; this test meaningless.
 
 define i64 @double_ui64(double %x) nounwind {
 entry:
 ; COMPILERRT: @double_ui64
 ; COMPILERRT-NOT: calll __ftol2
 ; FTOL: @double_ui64
 ; FTOL: fldl
 ; FTOL: calll __ftol2