setOperationAction(ISD::SRA, MVT::v4i32, Custom);
}
+ if (Subtarget->hasXOP()) {
+ setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
+ setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
+ setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
+ setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
+ setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
+ setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
+ setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
+ setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
+ }
+
if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
setOperationAction(ISD::BR_CC, MVT::i1, Expand);
setOperationAction(ISD::SETCC, MVT::i1, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
setOperationAction(ISD::XOR, MVT::i1, Legal);
setOperationAction(ISD::OR, MVT::i1, Legal);
setOperationAction(ISD::AND, MVT::i1, Legal);
setOperationAction(ISD::XOR, MVT::v16i32, Legal);
if (Subtarget->hasCDI()) {
- setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
+ setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
- setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
- setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
+
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
- }
- if (Subtarget->hasVLX() && Subtarget->hasCDI()) {
- setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
- setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
- setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
- setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
- setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
- setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
- setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
- setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
-
- setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
- setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
- setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
- setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
- }
+
+ if (Subtarget->hasVLX()) {
+ setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
+ setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
+ setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
+ setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
+
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
+ } else {
+ setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
+ }
+ } // Subtarget->hasCDI()
+
if (Subtarget->hasDQI()) {
setOperationAction(ISD::MUL, MVT::v2i64, Legal);
setOperationAction(ISD::MUL, MVT::v4i64, Legal);
setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
+ setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
+ setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
if (Subtarget->hasVLX())
setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
+ if (Subtarget->hasCDI()) {
+ setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
+ setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
+ }
+
for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
const MVT VT = (MVT::SimpleValueType)i;
MaxStoresPerMemmoveOptSize = 4;
setPrefLoopAlignment(4); // 2^4 bytes.
- // Predictable cmov don't hurt on atom because it's in-order.
+ // A predictable cmov does not hurt on an in-order CPU.
+ // FIXME: Use a CPU attribute to trigger this, not a CPU model.
PredictableSelectIsExpensive = !Subtarget->isAtom();
EnableExtLdPromotion = true;
setPrefFunctionAlignment(4); // 2^4 bytes.
return true;
}
+/// Android provides a fixed TLS slot for the SafeStack pointer.
+/// See the definition of TLS_SLOT_SAFESTACK in
+/// https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
+bool X86TargetLowering::getSafeStackPointerLocation(unsigned &AddressSpace,
+ unsigned &Offset) const {
+ if (!Subtarget->isTargetAndroid())
+ return false;
+
+ if (Subtarget->is64Bit()) {
+ // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
+ Offset = 0x48;
+ if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
+ AddressSpace = 256;
+ else
+ AddressSpace = 257;
+ } else {
+ // %gs:0x24 on i386
+ Offset = 0x24;
+ AddressSpace = 256;
+ }
+ return true;
+}
+
bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
unsigned DestAS) const {
assert(SrcAS != DestAS && "Expected different address spaces!");
MachinePointerInfo(), MachinePointerInfo());
}
-/// Return true if the calling convention is one that
-/// supports tail call optimization.
-static bool IsTailCallConvention(CallingConv::ID CC) {
+/// Return true if the calling convention is one that we can guarantee TCO for.
+static bool canGuaranteeTCO(CallingConv::ID CC) {
return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
- CC == CallingConv::HiPE);
+ CC == CallingConv::HiPE || CC == CallingConv::HHVM);
+}
+
+/// Return true if we might ever do TCO for calls with this calling convention.
+static bool mayTailCallThisCC(CallingConv::ID CC) {
+ switch (CC) {
+ // C calling conventions:
+ case CallingConv::C:
+ case CallingConv::X86_64_Win64:
+ case CallingConv::X86_64_SysV:
+ // Callee pop conventions:
+ case CallingConv::X86_ThisCall:
+ case CallingConv::X86_StdCall:
+ case CallingConv::X86_VectorCall:
+ case CallingConv::X86_FastCall:
+ return true;
+ default:
+ return canGuaranteeTCO(CC);
+ }
}
-/// \brief Return true if the calling convention is a C calling convention.
-static bool IsCCallConvention(CallingConv::ID CC) {
- return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
- CC == CallingConv::X86_64_SysV);
+/// Return true if the function is being made into a tailcall target by
+/// changing its ABI.
+static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
+ return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
}
bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
CallSite CS(CI);
CallingConv::ID CalleeCC = CS.getCallingConv();
- if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
+ if (!mayTailCallThisCC(CalleeCC))
return false;
return true;
}
-/// Return true if the function is being made into
-/// a tailcall target by changing its ABI.
-static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
- bool GuaranteedTailCallOpt) {
- return GuaranteedTailCallOpt && IsTailCallConvention(CC);
-}
-
SDValue
X86TargetLowering::LowerMemArgument(SDValue Chain,
CallingConv::ID CallConv,
unsigned i) const {
// Create the nodes corresponding to a load from this parameter slot.
ISD::ArgFlagsTy Flags = Ins[i].Flags;
- bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
+ bool AlwaysUseMutable = shouldGuaranteeTCO(
CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
EVT ValVT;
bool Is64Bit = Subtarget->is64Bit();
bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
- assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
+ assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
"Var args not supported with calling convention fastcc, ghc or hipe");
// Assign locations to all of the incoming arguments.
unsigned StackSize = CCInfo.getNextStackOffset();
// Align stack specially for tail calls.
- if (FuncIsMadeTailCallSafe(CallConv,
- MF.getTarget().Options.GuaranteedTailCallOpt))
+ if (shouldGuaranteeTCO(CallConv,
+ MF.getTarget().Options.GuaranteedTailCallOpt))
StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
// If the function takes variable number of arguments, make a frame index for
}
MachineModuleInfo &MMI = MF.getMMI();
- const Function *WinEHParent = nullptr;
- if (MMI.hasWinEHFuncInfo(Fn))
- WinEHParent = MMI.getWinEHParent(Fn);
- bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
// Figure out if XMM registers are in use.
assert(!(Subtarget->useSoftFloat() &&
} else {
FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
// If this is an sret function, the return should pop the hidden pointer.
- if (!Is64Bit && !IsTailCallConvention(CallConv) &&
+ if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
!Subtarget->getTargetTriple().isOSMSVCRT() &&
argsAreStructReturn(Ins) == StackStructReturn)
FuncInfo->setBytesToPopOnReturn(4);
FuncInfo->setArgumentStackSize(StackSize);
- if (IsWinEHParent) {
+ if (MMI.hasWinEHFuncInfo(Fn)) {
if (Is64Bit) {
int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
++NumTailCalls;
}
- assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
+ assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
"Var args not supported with calling convention fastcc, ghc or hipe");
// Analyze operands of the call, assigning locations to each operand.
CCInfo.AnalyzeCallOperands(Outs, CC_X86);
// Get a count of how many bytes are to be pushed on the stack.
- unsigned NumBytes = CCInfo.getNextStackOffset();
+ unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
if (IsSibcall)
// This is a sibcall. The memory operands are available in caller's
// own caller's stack.
NumBytes = 0;
else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
- IsTailCallConvention(CallConv))
+ canGuaranteeTCO(CallConv))
NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
int FPDiff = 0;
const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
- // If this is an invoke in a 32-bit function using an MSVC personality, assume
- // the function clobbers all registers. If an exception is thrown, the runtime
- // will not restore CSRs.
+ // If this is an invoke in a 32-bit function using a funclet-based
+ // personality, assume the function clobbers all registers. If an exception
+ // is thrown, the runtime will not restore CSRs.
// FIXME: Model this more precisely so that we can register allocate across
// the normal edge and spill and fill across the exceptional edge.
if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
CallerFn->hasPersonalityFn()
? classifyEHPersonality(CallerFn->getPersonalityFn())
: EHPersonality::Unknown;
- if (isMSVCEHPersonality(Pers))
+ if (isFuncletEHPersonality(Pers))
Mask = RegInfo->getNoPreservedMask();
}
if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
DAG.getTarget().Options.GuaranteedTailCallOpt))
NumBytesForCalleeToPop = NumBytes; // Callee pops everything
- else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
+ else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
!Subtarget->getTargetTriple().isOSMSVCRT() &&
SR == StackStructReturn)
// If this is a call to a struct-return function, the callee
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
- if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
+ if (!mayTailCallThisCC(CalleeCC))
return false;
// If -tailcallopt is specified, make fastcc functions tail-callable.
- const MachineFunction &MF = DAG.getMachineFunction();
+ MachineFunction &MF = DAG.getMachineFunction();
const Function *CallerF = MF.getFunction();
// If the function return type is x86_fp80 and the callee return type is not,
return false;
if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
- if (IsTailCallConvention(CalleeCC) && CCMatch)
+ if (canGuaranteeTCO(CalleeCC) && CCMatch)
return true;
return false;
}
if (isCalleeStructRet || isCallerStructRet)
return false;
- // An stdcall/thiscall caller is expected to clean up its arguments; the
- // callee isn't going to do that.
- // FIXME: this is more restrictive than needed. We could produce a tailcall
- // when the stack adjustment matches. For example, with a thiscall that takes
- // only one argument.
- if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
- CallerCC == CallingConv::X86_ThisCall))
- return false;
-
// Do not sibcall optimize vararg calls unless all arguments are passed via
// registers.
if (isVarArg && !Outs.empty()) {
-
// Optimizing for varargs on Win64 is unlikely to be safe without
// additional testing.
if (IsCalleeWin64 || IsCallerWin64)
}
}
+ unsigned StackArgsSize = 0;
+
// If the callee takes no arguments then go on to check the results of the
// call.
if (!Outs.empty()) {
CCInfo.AllocateStack(32, 8);
CCInfo.AnalyzeCallOperands(Outs, CC_X86);
- if (CCInfo.getNextStackOffset()) {
- MachineFunction &MF = DAG.getMachineFunction();
- if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
- return false;
+ StackArgsSize = CCInfo.getNextStackOffset();
+ if (CCInfo.getNextStackOffset()) {
// Check if the arguments are already laid out in the right way as
// the caller's fixed stack objects.
MachineFrameInfo *MFI = MF.getFrameInfo();
}
}
+ bool CalleeWillPop =
+ X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
+ MF.getTarget().Options.GuaranteedTailCallOpt);
+
+ if (unsigned BytesToPop =
+ MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
+ // If we have bytes to pop, the callee must pop them.
+ bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
+ if (!CalleePopMatches)
+ return false;
+ } else if (CalleeWillPop && StackArgsSize > 0) {
+ // If we don't have bytes to pop, make sure the callee doesn't pop any.
+ return false;
+ }
+
return true;
}
/// Determines whether the callee is required to pop its own arguments.
/// Callee pop is necessary to support tail calls.
bool X86::isCalleePop(CallingConv::ID CallingConv,
- bool is64Bit, bool IsVarArg, bool TailCallOpt) {
+ bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
+ // If GuaranteeTCO is true, we force some calls to be callee pop so that we
+ // can guarantee TCO.
+ if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
+ return true;
+
switch (CallingConv) {
default:
return false;
case CallingConv::X86_StdCall:
case CallingConv::X86_FastCall:
case CallingConv::X86_ThisCall:
+ case CallingConv::X86_VectorCall:
return !is64Bit;
- case CallingConv::Fast:
- case CallingConv::GHC:
- case CallingConv::HiPE:
- if (IsVarArg)
- return false;
- return TailCallOpt;
}
}
return getInsertVINSERTImmediate(N, 256);
}
-/// Returns true if Elt is a constant integer zero
+/// Returns true if V is a constant integer zero.
static bool isZero(SDValue V) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
return C && C->isNullValue();
return false;
}
+// Build a vector of constants
+// Use an UNDEF node if MaskElt == -1.
+// Spilt 64-bit constants in the 32-bit mode.
+static SDValue getConstVector(ArrayRef<int> Values, EVT VT,
+ SelectionDAG &DAG,
+ SDLoc dl, bool IsMask = false) {
+
+ SmallVector<SDValue, 32> Ops;
+ bool Split = false;
+
+ EVT ConstVecVT = VT;
+ unsigned NumElts = VT.getVectorNumElements();
+ bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
+ if (!In64BitMode && VT.getScalarType() == MVT::i64) {
+ ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
+ Split = true;
+ }
+
+ EVT EltVT = ConstVecVT.getScalarType();
+ for (unsigned i = 0; i < NumElts; ++i) {
+ bool IsUndef = Values[i] < 0 && IsMask;
+ SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
+ DAG.getConstant(Values[i], dl, EltVT);
+ Ops.push_back(OpNode);
+ if (Split)
+ Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
+ DAG.getConstant(0, dl, EltVT));
+ }
+ SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
+ if (Split)
+ ConstsNode = DAG.getBitcast(VT, ConstsNode);
+ return ConstsNode;
+}
+
/// Returns a vector of specified type with all zero elements.
static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
SelectionDAG &DAG, SDLoc dl) {
MVT::v16i8, PSHUFBMask)));
}
- // If we are extending from an (odd)offset, shuffle them by 1 element.
- if (Offset & 1) {
+ // If we are extending from an offset, ensure we start on a boundary that
+ // we can unpack from.
+ int AlignToUnpack = Offset % (NumElements / Scale);
+ if (AlignToUnpack) {
SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
- for (int i = 1; i < NumElements; ++i)
- ShMask[i - 1] = i;
+ for (int i = AlignToUnpack; i < NumElements; ++i)
+ ShMask[i - AlignToUnpack] = i;
InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
- Offset--;
+ Offset -= AlignToUnpack;
}
// Otherwise emit a sequence of unpacks.
}
}
+/// \brief Try to lower a vector shuffle as a 128-bit shuffles.
+static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
+ ArrayRef<int> Mask,
+ SDValue V1, SDValue V2,
+ SelectionDAG &DAG) {
+ assert(VT.getScalarSizeInBits() == 64 &&
+ "Unexpected element type size for 128bit shuffle.");
+
+ // To handle 256 bit vector requires VLX and most probably
+ // function lowerV2X128VectorShuffle() is better solution.
+ assert(VT.getSizeInBits() == 512 &&
+ "Unexpected vector size for 128bit shuffle.");
+
+ SmallVector<int, 4> WidenedMask;
+ if (!canWidenShuffleElements(Mask, WidenedMask))
+ return SDValue();
+
+ // Form a 128-bit permutation.
+ // Convert the 64-bit shuffle mask selection values into 128-bit selection
+ // bits defined by a vshuf64x2 instruction's immediate control byte.
+ unsigned PermMask = 0, Imm = 0;
+ unsigned ControlBitsNum = WidenedMask.size() / 2;
+
+ for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
+ if (WidenedMask[i] == SM_SentinelZero)
+ return SDValue();
+
+ // Use first element in place of undef mask.
+ Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
+ PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
+ }
+
+ return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
+ DAG.getConstant(PermMask, DL, MVT::i8));
+}
+
static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
ArrayRef<int> Mask, SDValue V1,
SDValue V2, SelectionDAG &DAG) {
MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
- SmallVector<SDValue, 32> VPermMask;
- for (unsigned i = 0; i < VT.getVectorNumElements(); ++i)
- VPermMask.push_back(Mask[i] < 0 ? DAG.getUNDEF(MaskEltVT) :
- DAG.getConstant(Mask[i], DL, MaskEltVT));
- SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecVT,
- VPermMask);
+ SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
if (isSingleInputShuffleMask(Mask))
return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
+ if (SDValue Shuf128 =
+ lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
+ return Shuf128;
+
if (SDValue Unpck =
lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
return Unpck;
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
+ if (SDValue Shuf128 =
+ lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
+ return Shuf128;
+
if (SDValue Unpck =
lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
return Unpck;
unsigned &MaskValue) {
MaskValue = 0;
unsigned NumElems = BuildVector->getNumOperands();
+
// There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
+ // We don't handle the >2 lanes case right now.
unsigned NumLanes = (NumElems - 1) / 8 + 1;
+ if (NumLanes > 2)
+ return false;
+
unsigned NumElemsInLane = NumElems / NumLanes;
// Blend for v16i16 should be symmetric for the both lanes.
if (isa<ConstantSDNode>(SndLaneEltCond))
Lane2Cond = !isZero(SndLaneEltCond);
+ unsigned LaneMask = 0;
if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
// Lane1Cond != 0, means we want the first argument.
// Lane1Cond == 0, means we want the second argument.
// The encoding of this argument is 0 for the first argument, 1
// for the second. Therefore, invert the condition.
- MaskValue |= !Lane1Cond << i;
+ LaneMask = !Lane1Cond << i;
else if (Lane1Cond < 0)
- MaskValue |= !Lane2Cond << i;
+ LaneMask = !Lane2Cond << i;
else
return false;
+
+ MaskValue |= LaneMask;
+ if (NumLanes == 2)
+ MaskValue |= LaneMask << NumElemsInLane;
}
return true;
}
// float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
// return (float4) lo + fhi;
+ // We shouldn't use it when unsafe-fp-math is enabled though: we might later
+ // reassociate the two FADDs, and if we do that, the algorithm fails
+ // spectacularly (PR24512).
+ // FIXME: If we ever have some kind of Machine FMF, this should be marked
+ // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
+ // there's also the MachineCombiner reassociations happening on Machine IR.
+ if (DAG.getTarget().Options.UnsafeFPMath)
+ return SDValue();
+
SDLoc DL(Op);
SDValue V = Op->getOperand(0);
EVT VecIntVT = V.getValueType();
}
// 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),
+// is legal, or has an fp128 or 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.
EVT TheVT = Op.getOperand(0).getValueType();
auto PtrVT = getPointerTy(DAG.getDataLayout());
- if (TheVT == MVT::f16)
- // We need to promote the f16 to f32 before using the lowering
- // in this routine.
+ if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
+ // f16 must be promoted before using the lowering in this routine.
+ // fp128 does not use this lowering.
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
// for DAG type consistency we have to match the FP operand type.
APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
- APFloat::opStatus Status = APFloat::opOK;
+ LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
bool LosesInfo = false;
if (TheVT == MVT::f64)
// The rounding mode is irrelevant as the conversion should be exact.
DAG.getConstant(SSECC, dl, MVT::i8));
}
+ MVT VTOp0 = Op0.getSimpleValueType();
+ assert(VTOp0 == Op1.getSimpleValueType() &&
+ "Expected operands with same type!");
+ assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
+ "Invalid number of packed elements for source and destination!");
+
+ if (VT.is128BitVector() && VTOp0.is256BitVector()) {
+ // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
+ // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
+ // legalizer firstly checks if the first operand in input to the setcc has
+ // a legal type. If so, then it promotes the return type to that same type.
+ // Otherwise, the return type is promoted to the 'next legal type' which,
+ // for a vector of MVT::i1 is always a 128-bit integer vector type.
+ //
+ // We reach this code only if the following two conditions are met:
+ // 1. Both return type and operand type have been promoted to wider types
+ // by the type legalizer.
+ // 2. The original operand type has been promoted to a 256-bit vector.
+ //
+ // Note that condition 2. only applies for AVX targets.
+ SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
+ return DAG.getZExtOrTrunc(NewOp, dl, VT);
+ }
+
+ // The non-AVX512 code below works under the assumption that source and
+ // destination types are the same.
+ assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
+ "Value types for source and destination must be the same!");
+
// Break 256-bit integer vector compare into smaller ones.
if (VT.is256BitVector() && !Subtarget->hasInt256())
return Lower256IntVSETCC(Op, DAG);
DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
}
+ // Lower using XOP integer comparisons.
+ if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
+ VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
+ // Translate compare code to XOP PCOM compare mode.
+ unsigned CmpMode = 0;
+ switch (SetCCOpcode) {
+ default: llvm_unreachable("Unexpected SETCC condition");
+ case ISD::SETULT:
+ case ISD::SETLT: CmpMode = 0x00; break;
+ case ISD::SETULE:
+ case ISD::SETLE: CmpMode = 0x01; break;
+ case ISD::SETUGT:
+ case ISD::SETGT: CmpMode = 0x02; break;
+ case ISD::SETUGE:
+ case ISD::SETGE: CmpMode = 0x03; break;
+ case ISD::SETEQ: CmpMode = 0x04; break;
+ case ISD::SETNE: CmpMode = 0x05; break;
+ }
+
+ // Are we comparing unsigned or signed integers?
+ unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
+ ? X86ISD::VPCOMU : X86ISD::VPCOM;
+
+ return DAG.getNode(Opc, dl, VT, Op0, Op1,
+ DAG.getConstant(CmpMode, dl, MVT::i8));
+ }
+
// We are handling one of the integer comparisons here. Since SSE only has
// GT and EQ comparisons for integer, swapping operands and multiple
// operations may be required for some comparisons.
return Sext;
}
- // Otherwise we'll shuffle the small elements in the high bits of the
- // larger type and perform an arithmetic shift. If the shift is not legal
- // it's better to scalarize.
- assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
- "We can't implement a sext load without an arithmetic right shift!");
-
- // Redistribute the loaded elements into the different locations.
- SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
- for (unsigned i = 0; i != NumElems; ++i)
- ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
-
- SDValue Shuff = DAG.getVectorShuffle(
- WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
-
- Shuff = DAG.getBitcast(RegVT, Shuff);
-
- // Build the arithmetic shift.
- unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
- MemVT.getVectorElementType().getSizeInBits();
- Shuff =
- DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
- DAG.getConstant(Amt, dl, RegVT));
+ // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
+ // lanes.
+ assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
+ "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
+ SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
return Shuff;
}
if (Op.getOpcode() == X86ISD::FSETCC)
return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
+ if (Op.getOpcode() == X86ISD::VFPCLASS)
+ return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
if (PreservedSrc.getOpcode() == ISD::UNDEF)
PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
RoundingMode, Sae),
Mask, Src0, Subtarget, DAG);
}
- case INTR_TYPE_2OP_MASK: {
+ case INTR_TYPE_2OP_MASK:
+ case INTR_TYPE_2OP_IMM8_MASK: {
SDValue Src1 = Op.getOperand(1);
SDValue Src2 = Op.getOperand(2);
SDValue PassThru = Op.getOperand(3);
SDValue Mask = Op.getOperand(4);
+
+ if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
+ Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
+
// We specify 2 possible opcodes for intrinsics with rounding modes.
// First, we check if the intrinsic may have non-default rounding mode,
// (IntrData->Opc1 != 0), then we check the rounding mode operand.
Src1, Src2, Src3),
Mask, PassThru, Subtarget, DAG);
}
+ case TERLOG_OP_MASK:
+ case TERLOG_OP_MASKZ: {
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue Src3 = Op.getOperand(3);
+ SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
+ SDValue Mask = Op.getOperand(5);
+ EVT VT = Op.getValueType();
+ SDValue PassThru = Src1;
+ // Set PassThru element.
+ if (IntrData->Type == TERLOG_OP_MASKZ)
+ PassThru = getZeroVector(VT, Subtarget, DAG, dl);
+
+ return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
+ Src1, Src2, Src3, Src4),
+ Mask, PassThru, Subtarget, DAG);
+ }
case FPCLASS: {
// FPclass intrinsics with mask
SDValue Src1 = Op.getOperand(1);
DAG.getIntPtrConstant(0, dl));
return DAG.getBitcast(Op.getValueType(), Res);
}
+ case FPCLASSS: {
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Imm = Op.getOperand(2);
+ SDValue Mask = Op.getOperand(3);
+ SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
+ SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
+ DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
+ return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
+ }
case CMP_MASK:
case CMP_MASK_CC: {
// Comparison intrinsics with masks.
ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
}
-static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
+/// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
+//
+// 1. i32/i64 128/256-bit vector (native support require VLX) are expended
+// to 512-bit vector.
+// 2. i8/i16 vector implemented using dword LZCNT vector instruction
+// ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
+// split the vector, perform operation on it's Lo a Hi part and
+// concatenate the results.
+static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
+ SDLoc dl(Op);
+ MVT VT = Op.getSimpleValueType();
+ MVT EltVT = VT.getVectorElementType();
+ unsigned NumElems = VT.getVectorNumElements();
+
+ if (EltVT == MVT::i64 || EltVT == MVT::i32) {
+ // Extend to 512 bit vector.
+ assert((VT.is256BitVector() || VT.is128BitVector()) &&
+ "Unsupported value type for operation");
+
+ MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
+ SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
+ DAG.getUNDEF(NewVT),
+ Op.getOperand(0),
+ DAG.getIntPtrConstant(0, dl));
+ SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
+
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
+ DAG.getIntPtrConstant(0, dl));
+ }
+
+ assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
+ "Unsupported element type");
+
+ if (16 < NumElems) {
+ // Split vector, it's Lo and Hi parts will be handled in next iteration.
+ SDValue Lo, Hi;
+ std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
+ MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
+
+ Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
+ Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
+ }
+
+ MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
+
+ assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
+ "Unsupported value type for operation");
+
+ // Use native supported vector instruction vplzcntd.
+ Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
+ SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
+ SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
+ SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
+
+ return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
+}
+
+static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
EVT OpVT = VT;
unsigned NumBits = VT.getSizeInBits();
SDLoc dl(Op);
+ if (VT.isVector() && Subtarget->hasAVX512())
+ return LowerVectorCTLZ_AVX512(Op, DAG);
+
Op = Op.getOperand(0);
if (VT == MVT::i8) {
// Zero extend to i32 since there is not an i8 bsr.
return Op;
}
-static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
+static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
EVT OpVT = VT;
unsigned NumBits = VT.getSizeInBits();
SDLoc dl(Op);
+ if (VT.isVector() && Subtarget->hasAVX512())
+ return LowerVectorCTLZ_AVX512(Op, DAG);
+
Op = Op.getOperand(0);
if (VT == MVT::i8) {
// Zero extend to i32 since there is not an i8 bsr.
// i64 SRA needs to be performed as partial shifts.
if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
- Op.getOpcode() == ISD::SRA)
+ Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
return ArithmeticShiftRight64(ShiftAmt);
if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
unsigned NumElts = VT.getVectorNumElements();
MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
- if (Op.getOpcode() == ISD::SHL) {
- // Simple i8 add case
- if (ShiftAmt == 1)
- return DAG.getNode(ISD::ADD, dl, VT, R, R);
+ // Simple i8 add case
+ if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
+ return DAG.getNode(ISD::ADD, dl, VT, R, R);
+
+ // ashr(R, 7) === cmp_slt(R, 0)
+ if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
+ SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
+ return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
+ }
+
+ // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
+ if (VT == MVT::v16i8 && Subtarget->hasXOP())
+ return SDValue();
+ if (Op.getOpcode() == ISD::SHL) {
// Make a large shift.
SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
R, ShiftAmt, DAG);
DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
}
if (Op.getOpcode() == ISD::SRA) {
- if (ShiftAmt == 7) {
- // ashr(R, 7) === cmp_slt(R, 0)
- SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
- return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
- }
-
// ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
SmallVector<SDValue, 32> V(NumElts,
}
// Special case in 32-bit mode, where i64 is expanded into high and low parts.
- if (!Subtarget->is64Bit() &&
+ if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
(VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
// Peek through any splat that was introduced for i64 shift vectorization.
return V;
if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
- return V;
+ return V;
if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
return Op;
+ // XOP has 128-bit variable logical/arithmetic shifts.
+ // +ve/-ve Amt = shift left/right.
+ if (Subtarget->hasXOP() &&
+ (VT == MVT::v2i64 || VT == MVT::v4i32 ||
+ VT == MVT::v8i16 || VT == MVT::v16i8)) {
+ if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
+ SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
+ Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
+ }
+ if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
+ return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
+ if (Op.getOpcode() == ISD::SRA)
+ return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
+ }
+
// 2i64 vector logical shifts can efficiently avoid scalarization - do the
// shifts per-lane and then shuffle the partial results back together.
if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
}
- if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
+ if (VT == MVT::v16i8 ||
+ (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
unsigned ShiftOpcode = Op->getOpcode();
DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
}
- if (Subtarget->hasInt256() && VT == MVT::v16i16) {
+ if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
MVT ExtVT = MVT::v8i32;
SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
return SDValue();
}
+static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ MVT VT = Op.getSimpleValueType();
+ SDLoc DL(Op);
+ SDValue R = Op.getOperand(0);
+ SDValue Amt = Op.getOperand(1);
+
+ assert(VT.isVector() && "Custom lowering only for vector rotates!");
+ assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
+ assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
+
+ // XOP has 128-bit vector variable + immediate rotates.
+ // +ve/-ve Amt = rotate left/right.
+
+ // Split 256-bit integers.
+ if (VT.getSizeInBits() == 256)
+ return Lower256IntArith(Op, DAG);
+
+ assert(VT.getSizeInBits() == 128 && "Only rotate 128-bit vectors!");
+
+ // Attempt to rotate by immediate.
+ if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
+ if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
+ uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
+ assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
+ return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
+ DAG.getConstant(RotateAmt, DL, MVT::i8));
+ }
+ }
+
+ // Use general rotate by variable (per-element).
+ return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
+}
+
static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
// Lower the "add/sub/mul with overflow" instruction into a regular ins plus
// a "setcc" instruction that checks the overflow flag. The "brcond" lowering
case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
- case ISD::CTLZ: return LowerCTLZ(Op, DAG);
- case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
+ case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
+ case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
case ISD::CTTZ:
case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
case ISD::UMUL_LOHI:
case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
+ case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
case ISD::SRA:
case ISD::SRL:
case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
case X86ISD::VPERMI: return "X86ISD::VPERMI";
+ case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
case X86ISD::VRANGE: return "X86ISD::VRANGE";
case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
case X86ISD::RDSEED: return "X86ISD::RDSEED";
case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
+ case X86ISD::VPROT: return "X86ISD::VPROT";
+ case X86ISD::VPROTI: return "X86ISD::VPROTI";
+ case X86ISD::VPSHA: return "X86ISD::VPSHA";
+ case X86ISD::VPSHL: return "X86ISD::VPSHL";
+ case X86ISD::VPCOM: return "X86ISD::VPCOM";
+ case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
case X86ISD::FMADD: return "X86ISD::FMADD";
case X86ISD::FMSUB: return "X86ISD::FMSUB";
case X86ISD::FNMADD: return "X86ISD::FNMADD";
DebugLoc DL = MI->getDebugLoc();
const BasicBlock *BB = MBB->getBasicBlock();
- MachineFunction::iterator I = MBB;
- ++I;
+ MachineFunction::iterator I = ++MBB->getIterator();
// For the v = xbegin(), we generate
//
offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
- MachineFunction::iterator MBBIter = MBB;
- ++MBBIter;
+ MachineFunction::iterator MBBIter = ++MBB->getIterator();
// Insert the new basic blocks
MF->insert(MBBIter, offsetMBB);
// stores were performed.
const BasicBlock *LLVM_BB = MBB->getBasicBlock();
MachineFunction *F = MBB->getParent();
- MachineFunction::iterator MBBIter = MBB;
- ++MBBIter;
+ MachineFunction::iterator MBBIter = ++MBB->getIterator();
MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
F->insert(MBBIter, XMMSaveMBB);
// destination vreg to set, the condition code register to branch on, the
// true/false values to select between, and a branch opcode to use.
const BasicBlock *LLVM_BB = BB->getBasicBlock();
- MachineFunction::iterator It = BB;
- ++It;
+ MachineFunction::iterator It = ++BB->getIterator();
// thisMBB:
// ...
const X86InstrInfo *TII = Subtarget->getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
- unsigned MSrc = MI->getOperand(0).getReg();
+ MachineOperand MSrc = MI->getOperand(0);
unsigned VSrc = MI->getOperand(5).getReg();
+ const MachineOperand &Disp = MI->getOperand(3);
+ MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
+ bool hasDisp = Disp.isGlobal() || Disp.isImm();
+ if (hasDisp && MSrc.isReg())
+ MSrc.setIsKill(false);
MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
- .addReg(/*Base=*/MSrc)
+ .addOperand(/*Base=*/MSrc)
.addImm(/*Scale=*/1)
.addReg(/*Index=*/0)
- .addImm(0)
+ .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
.addReg(0);
MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
.addReg(VSrc)
- .addReg(/*Base=*/MSrc)
+ .addOperand(/*Base=*/MSrc)
.addImm(/*Scale=*/1)
.addReg(/*Index=*/0)
- .addImm(/*Disp=*/0)
+ .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
.addReg(/*Segment=*/0);
MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
MI->eraseFromParent(); // The pseudo instruction is gone now.
sizeVReg = MI->getOperand(1).getReg(),
physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
- MachineFunction::iterator MBBIter = BB;
- ++MBBIter;
+ MachineFunction::iterator MBBIter = ++BB->getIterator();
MF->insert(MBBIter, bumpMBB);
MF->insert(MBBIter, mallocMBB);
MachineRegisterInfo &MRI = MF->getRegInfo();
const BasicBlock *BB = MBB->getBasicBlock();
- MachineFunction::iterator I = MBB;
- ++I;
+ MachineFunction::iterator I = ++MBB->getIterator();
// Memory Reference
MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
// For v = setjmp(buf), we generate
//
// thisMBB:
- // buf[LabelOffset] = restoreMBB
+ // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
// SjLjSetup restoreMBB
//
// mainMBB:
MF->insert(I, mainMBB);
MF->insert(I, sinkMBB);
MF->push_back(restoreMBB);
+ restoreMBB->setHasAddressTaken();
MachineInstrBuilder MIB;
MVT RootVT = Root.getSimpleValueType();
SDLoc DL(Root);
- // Just remove no-op shuffle masks.
if (Mask.size() == 1) {
- DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
- /*AddTo*/ true);
+ int Index = Mask[0];
+ assert((Index >= 0 || Index == SM_SentinelUndef ||
+ Index == SM_SentinelZero) &&
+ "Invalid shuffle index found!");
+
+ // We may end up with an accumulated mask of size 1 as a result of
+ // widening of shuffle operands (see function canWidenShuffleElements).
+ // If the only shuffle index is equal to SM_SentinelZero then propagate
+ // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
+ // mask, and therefore the entire chain of shuffles can be folded away.
+ if (Index == SM_SentinelZero)
+ DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
+ else
+ DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
+ /*AddTo*/ true);
return true;
}
InputVector.getNode()->getOperand(0));
// The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
- SDValue MMXSrcOp = MMXSrc.getOperand(0);
if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
- MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
- MMXSrcOp.getOpcode() == ISD::BITCAST &&
- MMXSrcOp.getValueType() == MVT::v1i64 &&
- MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
- return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
- N->getValueType(0),
- MMXSrcOp.getOperand(0));
+ MMXSrc.getValueType() == MVT::i64) {
+ SDValue MMXSrcOp = MMXSrc.getOperand(0);
+ if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
+ MMXSrcOp.getValueType() == MVT::v1i64 &&
+ MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
+ return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
+ N->getValueType(0), MMXSrcOp.getOperand(0));
+ }
}
EVT VT = N->getValueType(0);
InputVector.getOpcode() == ISD::BITCAST &&
dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
uint64_t ExtractedElt =
- cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+ cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
uint64_t InputValue =
- cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
+ cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
uint64_t Res = (InputValue >> ExtractedElt) & 1;
return DAG.getConstant(Res, dl, MVT::i1);
}
return DAG.getBitcast(N0.getValueType(), NewShuffle);
}
+/// If both input operands of a logic op are being cast from floating point
+/// types, try to convert this into a floating point logic node to avoid
+/// unnecessary moves from SSE to integer registers.
+static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ unsigned FPOpcode = ISD::DELETED_NODE;
+ if (N->getOpcode() == ISD::AND)
+ FPOpcode = X86ISD::FAND;
+ else if (N->getOpcode() == ISD::OR)
+ FPOpcode = X86ISD::FOR;
+ else if (N->getOpcode() == ISD::XOR)
+ FPOpcode = X86ISD::FXOR;
+
+ assert(FPOpcode != ISD::DELETED_NODE &&
+ "Unexpected input node for FP logic conversion");
+
+ EVT VT = N->getValueType(0);
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SDLoc DL(N);
+ if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
+ ((Subtarget->hasSSE1() && VT == MVT::i32) ||
+ (Subtarget->hasSSE2() && VT == MVT::i64))) {
+ SDValue N00 = N0.getOperand(0);
+ SDValue N10 = N1.getOperand(0);
+ EVT N00Type = N00.getValueType();
+ EVT N10Type = N10.getValueType();
+ if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
+ SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
+ return DAG.getBitcast(VT, FPLogic);
+ }
+ }
+ return SDValue();
+}
+
static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
return R;
+ if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
+ return FPLogic;
+
EVT VT = N->getValueType(0);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
return R;
+ if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
+ return FPLogic;
+
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
EVT VT = N->getValueType(0);
if (SDValue RV = performIntegerAbsCombine(N, DAG))
return RV;
+ if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
+ return FPLogic;
+
return SDValue();
}
ShuffleVec[i] = i * SizeRatio;
// Can't shuffle using an illegal type.
- assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
- && "WideVecVT should be legal");
+ assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
+ "WideVecVT should be legal");
WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
}
ShuffleVec[i] = i * SizeRatio;
// Can't shuffle using an illegal type.
- assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
- && "WideVecVT should be legal");
+ assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
+ "WideVecVT should be legal");
SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
DAG.getUNDEF(WideVecVT),
return SDValue();
}
+/// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
+/// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
+/// to combine math ops, use an LEA, or use a complex addressing mode. This can
+/// eliminate extend, add, and shift instructions.
+static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ // TODO: This should be valid for other integer types.
+ EVT VT = Sext->getValueType(0);
+ if (VT != MVT::i64)
+ return SDValue();
+
+ // We need an 'add nsw' feeding into the 'sext'.
+ SDValue Add = Sext->getOperand(0);
+ if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
+ return SDValue();
+
+ // Having a constant operand to the 'add' ensures that we are not increasing
+ // the instruction count because the constant is extended for free below.
+ // A constant operand can also become the displacement field of an LEA.
+ auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
+ if (!AddOp1)
+ return SDValue();
+
+ // Don't make the 'add' bigger if there's no hope of combining it with some
+ // other 'add' or 'shl' instruction.
+ // TODO: It may be profitable to generate simpler LEA instructions in place
+ // of single 'add' instructions, but the cost model for selecting an LEA
+ // currently has a high threshold.
+ bool HasLEAPotential = false;
+ for (auto *User : Sext->uses()) {
+ if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
+ HasLEAPotential = true;
+ break;
+ }
+ }
+ if (!HasLEAPotential)
+ return SDValue();
+
+ // Everything looks good, so pull the 'sext' ahead of the 'add'.
+ int64_t AddConstant = AddOp1->getSExtValue();
+ SDValue AddOp0 = Add.getOperand(0);
+ SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
+ SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
+
+ // The wider add is guaranteed to not wrap because both operands are
+ // sign-extended.
+ SDNodeFlags Flags;
+ Flags.setNoSignedWrap(true);
+ return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
+}
+
static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
}
}
- if (!Subtarget->hasFp256())
- return SDValue();
-
- if (VT.isVector() && VT.getSizeInBits() == 256)
+ if (Subtarget->hasAVX() && VT.isVector() && VT.getSizeInBits() == 256)
if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
return R;
+ if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
+ return NewAdd;
+
return SDValue();
}
projects / oota-llvm.git / blobdiff