#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86TargetMachine.h"
+#include "X86TargetObjectFile.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Target/TargetLoweringObjectFile.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/StringExtras.h"
static cl::opt<bool>
DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
+// Disable16Bit - 16-bit operations typically have a larger encoding than
+// corresponding 32-bit instructions, and 16-bit code is slow on some
+// processors. This is an experimental flag to disable 16-bit operations
+// (which forces them to be Legalized to 32-bit operations).
+static cl::opt<bool>
+Disable16Bit("disable-16bit", cl::Hidden,
+ cl::desc("Disable use of 16-bit instructions"));
+
// Forward declarations.
static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
SDValue V2);
switch (TM.getSubtarget<X86Subtarget>().TargetType) {
default: llvm_unreachable("unknown subtarget type");
case X86Subtarget::isDarwin:
+ if (TM.getSubtarget<X86Subtarget>().is64Bit())
+ return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
case X86Subtarget::isELF:
return new TargetLoweringObjectFileELF();
case X86Subtarget::isWindows:
return new TargetLoweringObjectFileCOFF();
}
-
+
}
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
// Set up the register classes.
addRegisterClass(MVT::i8, X86::GR8RegisterClass);
- addRegisterClass(MVT::i16, X86::GR16RegisterClass);
+ if (!Disable16Bit)
+ addRegisterClass(MVT::i16, X86::GR16RegisterClass);
addRegisterClass(MVT::i32, X86::GR32RegisterClass);
if (Subtarget->is64Bit())
addRegisterClass(MVT::i64, X86::GR64RegisterClass);
// We don't accept any truncstore of integer registers.
setTruncStoreAction(MVT::i64, MVT::i32, Expand);
- setTruncStoreAction(MVT::i64, MVT::i16, Expand);
+ if (!Disable16Bit)
+ setTruncStoreAction(MVT::i64, MVT::i16, Expand);
setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
- setTruncStoreAction(MVT::i32, MVT::i16, Expand);
+ if (!Disable16Bit)
+ setTruncStoreAction(MVT::i32, MVT::i16, Expand);
setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
setTruncStoreAction(MVT::i16, MVT::i8, Expand);
setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
- setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
- setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
+ if (Disable16Bit) {
+ setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
+ setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
+ } else {
+ setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
+ setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
+ }
setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
// These should be promoted to a larger select which is supported.
- setOperationAction(ISD::SELECT , MVT::i1 , Promote);
- setOperationAction(ISD::SELECT , MVT::i8 , Promote);
+ setOperationAction(ISD::SELECT , MVT::i1 , Promote);
// X86 wants to expand cmov itself.
- setOperationAction(ISD::SELECT , MVT::i16 , Custom);
+ setOperationAction(ISD::SELECT , MVT::i8 , Custom);
+ if (Disable16Bit)
+ setOperationAction(ISD::SELECT , MVT::i16 , Expand);
+ else
+ setOperationAction(ISD::SELECT , MVT::i16 , Custom);
setOperationAction(ISD::SELECT , MVT::i32 , Custom);
setOperationAction(ISD::SELECT , MVT::f32 , Custom);
setOperationAction(ISD::SELECT , MVT::f64 , Custom);
setOperationAction(ISD::SELECT , MVT::f80 , Custom);
setOperationAction(ISD::SETCC , MVT::i8 , Custom);
- setOperationAction(ISD::SETCC , MVT::i16 , Custom);
+ if (Disable16Bit)
+ setOperationAction(ISD::SETCC , MVT::i16 , Expand);
+ else
+ setOperationAction(ISD::SETCC , MVT::i16 , Custom);
setOperationAction(ISD::SETCC , MVT::i32 , Custom);
setOperationAction(ISD::SETCC , MVT::f32 , Custom);
setOperationAction(ISD::SETCC , MVT::f64 , Custom);
setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
}
- // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
+ // Use the default ISD::DBG_STOPPOINT.
setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
// FIXME - use subtarget debug flags
if (!Subtarget->isTargetDarwin() &&
if (Subtarget->is64Bit()) {
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
- }
+ }
#endif
#if 0
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
- allowUnalignedMemoryAccesses = true; // x86 supports it!
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
}
/// getFunctionAlignment - Return the Log2 alignment of this function.
unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
- return F->hasFnAttr(Attribute::OptimizeForSize) ? 1 : 4;
+ return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
}
//===----------------------------------------------------------------------===//
SDValue
X86TargetLowering::LowerReturn(SDValue Chain,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
DebugLoc dl, SelectionDAG &DAG) {
SmallVector<SDValue, 6> RetOps;
RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
// Operand #1 = Bytes To Pop
- RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
+ RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
// Copy the result values into the output registers.
for (unsigned i = 0; i != RVLocs.size(); ++i) {
///
SDValue
X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) {
/// IsCalleePop - Determines whether the callee is required to pop its
/// own arguments. Callee pop is necessary to support tail calls.
-bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
+bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
if (IsVarArg)
return false;
/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
/// given CallingConvention value.
-CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
+CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
if (Subtarget->is64Bit()) {
if (Subtarget->isTargetWin64())
return CC_X86_Win64_C;
/// NameDecorationForCallConv - Selects the appropriate decoration to
/// apply to a MachineFunction containing a given calling convention.
NameDecorationStyle
-X86TargetLowering::NameDecorationForCallConv(unsigned CallConv) {
+X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
if (CallConv == CallingConv::X86_FastCall)
return FastCall;
else if (CallConv == CallingConv::X86_StdCall)
SDValue
X86TargetLowering::LowerMemArgument(SDValue Chain,
- unsigned CallConv,
+ CallingConv::ID CallConv,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
const CCValAssign &VA,
SDValue
X86TargetLowering::LowerFormalArguments(SDValue Chain,
- unsigned CallConv,
+ CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl,
Offset += 8;
}
- if (!MemOps.empty())
- Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
- &MemOps[0], MemOps.size());
+ if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
+ // Now store the XMM (fp + vector) parameter registers.
+ SmallVector<SDValue, 11> SaveXMMOps;
+ SaveXMMOps.push_back(Chain);
- // Now store the XMM (fp + vector) parameter registers.
- SmallVector<SDValue, 11> SaveXMMOps;
- SaveXMMOps.push_back(Chain);
+ unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
+ SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
+ SaveXMMOps.push_back(ALVal);
- unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
- SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
- SaveXMMOps.push_back(ALVal);
+ SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
+ SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
- SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
- SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
-
- for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
- unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
- X86::VR128RegisterClass);
- SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
- SaveXMMOps.push_back(Val);
+ for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
+ unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
+ X86::VR128RegisterClass);
+ SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
+ SaveXMMOps.push_back(Val);
+ }
+ MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
+ MVT::Other,
+ &SaveXMMOps[0], SaveXMMOps.size()));
}
- Chain = DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl, MVT::Other,
- &SaveXMMOps[0], SaveXMMOps.size());
+
+ if (!MemOps.empty())
+ Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &MemOps[0], MemOps.size());
}
}
SDValue
X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
- unsigned CallConv, bool isVarArg, bool isTailCall,
+ CallingConv::ID CallConv, bool isVarArg,
+ bool isTailCall,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
InFlag = Chain.getValue(1);
}
-
+
if (Subtarget->isPICStyleGOT()) {
// ELF / PIC requires GOT in the EBX register before function calls via PLT
// GOT pointer.
GlobalValue *GV = G->getGlobal();
if (!GV->hasDLLImportLinkage()) {
unsigned char OpFlags = 0;
-
+
// On ELF targets, in both X86-64 and X86-32 mode, direct calls to
// external symbols most go through the PLT in PIC mode. If the symbol
// has hidden or protected visibility, or if it is static or local, then
// automatically synthesizes these stubs.
OpFlags = X86II::MO_DARWIN_STUB;
}
-
+
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
OpFlags);
} else if (isTailCall) {
/// optimization should implement this function.
bool
X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
- unsigned CalleeCC,
+ CallingConv::ID CalleeCC,
bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
SelectionDAG& DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
- unsigned CallerCC = MF.getFunction()->getCallingConv();
+ CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
}
}
bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
- SmallVector<int, 8> M;
+ SmallVector<int, 8> M;
N->getMask(M);
return ::isPSHUFDMask(M, N->getValueType(0));
}
static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
if (VT != MVT::v8i16)
return false;
-
+
// Lower quadword copied in order or undef.
for (int i = 0; i != 4; ++i)
if (Mask[i] >= 0 && Mask[i] != i)
return false;
-
+
// Upper quadword shuffled.
for (int i = 4; i != 8; ++i)
if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
return false;
-
+
return true;
}
bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
- SmallVector<int, 8> M;
+ SmallVector<int, 8> M;
N->getMask(M);
return ::isPSHUFHWMask(M, N->getValueType(0));
}
static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
if (VT != MVT::v8i16)
return false;
-
+
// Upper quadword copied in order.
for (int i = 4; i != 8; ++i)
if (Mask[i] >= 0 && Mask[i] != i)
return false;
-
+
// Lower quadword shuffled.
for (int i = 0; i != 4; ++i)
if (Mask[i] >= 4)
return false;
-
+
return true;
}
bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
- SmallVector<int, 8> M;
+ SmallVector<int, 8> M;
N->getMask(M);
return ::isPSHUFLWMask(M, N->getValueType(0));
}
int NumElems = VT.getVectorNumElements();
if (NumElems != 2 && NumElems != 4)
return false;
-
+
int Half = NumElems / 2;
for (int i = 0; i < Half; ++i)
if (!isUndefOrInRange(Mask[i], 0, NumElems))
for (int i = Half; i < NumElems; ++i)
if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
return false;
-
+
return true;
}
/// the upper half to come from vector 2.
static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
int NumElems = VT.getVectorNumElements();
-
- if (NumElems != 2 && NumElems != 4)
+
+ if (NumElems != 2 && NumElems != 4)
return false;
-
+
int Half = NumElems / 2;
for (int i = 0; i < Half; ++i)
if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
/// <2, 3, 2, 3>
bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
unsigned NumElems = N->getValueType(0).getVectorNumElements();
-
+
if (NumElems != 4)
return false;
-
- return isUndefOrEqual(N->getMaskElt(0), 2) &&
+
+ return isUndefOrEqual(N->getMaskElt(0), 2) &&
isUndefOrEqual(N->getMaskElt(1), 3) &&
- isUndefOrEqual(N->getMaskElt(2), 2) &&
+ isUndefOrEqual(N->getMaskElt(2), 2) &&
isUndefOrEqual(N->getMaskElt(3), 3);
}
int NumElts = VT.getVectorNumElements();
if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
return false;
-
+
for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKH.
-static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
+static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool V2IsSplat = false) {
int NumElts = VT.getVectorNumElements();
if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
return false;
-
+
for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
int NumElems = VT.getVectorNumElements();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
-
+
for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
int NumElems = VT.getVectorNumElements();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
-
+
for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return false;
int NumElts = VT.getVectorNumElements();
-
+
if (!isUndefOrEqual(Mask[0], NumElts))
return false;
-
+
for (int i = 1; i < NumElts; ++i)
if (!isUndefOrEqual(Mask[i], i))
return false;
-
+
return true;
}
int NumOps = VT.getVectorNumElements();
if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
return false;
-
+
if (!isUndefOrEqual(Mask[0], 0))
return false;
-
+
for (int i = 1; i < NumOps; ++i)
if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
(V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
(V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
return false;
-
+
return true;
}
/// specifies a shuffle of elements that is suitable for input to MOVDDUP.
bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
int e = N->getValueType(0).getVectorNumElements() / 2;
-
+
for (int i = 0; i < e; ++i)
if (!isUndefOrEqual(N->getMaskElt(i), i))
return false;
EVT VT = SVOp->getValueType(0);
unsigned NumElems = VT.getVectorNumElements();
SmallVector<int, 8> MaskVec;
-
+
for (unsigned i = 0; i != NumElems; ++i) {
int idx = SVOp->getMaskElt(i);
if (idx < 0)
return false;
unsigned NumElems = Op->getValueType(0).getVectorNumElements();
-
+
if (NumElems != 2 && NumElems != 4)
return false;
for (unsigned i = 0, e = NumElems/2; i != e; ++i)
}
/// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
-/// to an zero vector.
+/// to an zero vector.
/// FIXME: move to dag combiner / method on ShuffleVectorSDNode
static bool isZeroShuffle(ShuffleVectorSDNode *N) {
SDValue V1 = N->getOperand(0);
static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
EVT VT = SVOp->getValueType(0);
unsigned NumElems = VT.getVectorNumElements();
-
+
bool Changed = false;
SmallVector<int, 8> MaskVec;
SVOp->getMask(MaskVec);
-
+
for (unsigned i = 0; i != NumElems; ++i) {
if (MaskVec[i] > (int)NumElems) {
MaskVec[i] = NumElems;
}
/// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
-static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
+static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
bool HasSSE2) {
if (SV->getValueType(0).getVectorNumElements() <= 4)
return SDValue(SV, 0);
-
+
EVT PVT = MVT::v4f32;
EVT VT = SV->getValueType(0);
DebugLoc dl = SV->getDebugLoc();
}
NumElems >>= 1;
}
-
+
// Perform the splat.
int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
for (unsigned i = 1; i != VecElts; ++i)
Mask.push_back(i);
Item = DAG.getVectorShuffle(VecVT, dl, Item,
- DAG.getUNDEF(Item.getValueType()),
+ DAG.getUNDEF(Item.getValueType()),
&Mask[0]);
}
return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
// If we have SSE 4.1, Expand into a number of inserts unless the number of
// values to be inserted is equal to the number of elements, in which case
// use the unpack code below in the hopes of matching the consecutive elts
- // load merge pattern for shuffles.
+ // load merge pattern for shuffles.
// FIXME: We could probably just check that here directly.
- if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
+ if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
getSubtarget()->hasSSE41()) {
V[0] = DAG.getUNDEF(VT);
for (unsigned i = 0; i < NumElems; ++i)
}
// For SSSE3, If all 8 words of the result come from only 1 quadword of each
- // of the two input vectors, shuffle them into one input vector so only a
+ // of the two input vectors, shuffle them into one input vector so only a
// single pshufb instruction is necessary. If There are more than 2 input
// quads, disable the next transformation since it does not help SSSE3.
bool V1Used = InputQuads[0] || InputQuads[1];
SmallVector<int, 8> MaskV;
MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
- NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
+ NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
int idx = MaskVals[i];
if (idx < 0)
continue;
- idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
+ idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
if ((idx != i) && idx < 4)
pshufhw = false;
if ((idx != i) && idx > 3)
// If we've eliminated the use of V2, and the new mask is a pshuflw or
// pshufhw, that's as cheap as it gets. Return the new shuffle.
if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
- return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
+ return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
}
}
-
+
// If we have SSSE3, and all words of the result are from 1 input vector,
// case 2 is generated, otherwise case 3 is generated. If no SSSE3
// is present, fall back to case 4.
if (TLI.getSubtarget()->hasSSSE3()) {
SmallVector<SDValue,16> pshufbMask;
-
+
// If we have elements from both input vectors, set the high bit of the
- // shuffle mask element to zero out elements that come from V2 in the V1
+ // shuffle mask element to zero out elements that come from V2 in the V1
// mask, and elements that come from V1 in the V2 mask, so that the two
// results can be OR'd together.
bool TwoInputs = V1Used && V2Used;
pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
}
V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
- V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
+ V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
DAG.getNode(ISD::BUILD_VECTOR, dl,
MVT::v16i8, &pshufbMask[0], 16));
if (!TwoInputs)
return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
-
+
// Calculate the shuffle mask for the second input, shuffle it, and
// OR it with the first shuffled input.
pshufbMask.clear();
pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
}
V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
- V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
+ V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
DAG.getNode(ISD::BUILD_VECTOR, dl,
MVT::v16i8, &pshufbMask[0], 16));
V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
&MaskV[0]);
}
-
+
// If BestHi >= 0, generate a pshufhw to put the high elements in order,
// and update MaskVals with the new element order.
if (BestHiQuad >= 0) {
NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
&MaskV[0]);
}
-
+
// In case BestHi & BestLo were both -1, which means each quadword has a word
// from each of the four input quadwords, calculate the InOrder bitvector now
// before falling through to the insert/extract cleanup.
if (MaskVals[i] < 0 || MaskVals[i] == i)
InOrder.set(i);
}
-
+
// The other elements are put in the right place using pextrw and pinsrw.
for (unsigned i = 0; i != 8; ++i) {
if (InOrder[i])
DebugLoc dl = SVOp->getDebugLoc();
SmallVector<int, 16> MaskVals;
SVOp->getMask(MaskVals);
-
+
// If we have SSSE3, case 1 is generated when all result bytes come from
- // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
+ // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
// present, fall back to case 3.
// FIXME: kill V2Only once shuffles are canonizalized by getNode.
bool V1Only = true;
else
V1Only = false;
}
-
+
// If SSSE3, use 1 pshufb instruction per vector with elements in the result.
if (TLI.getSubtarget()->hasSSSE3()) {
SmallVector<SDValue,16> pshufbMask;
-
+
// If all result elements are from one input vector, then only translate
- // undef mask values to 0x80 (zero out result) in the pshufb mask.
+ // undef mask values to 0x80 (zero out result) in the pshufb mask.
//
// Otherwise, we have elements from both input vectors, and must zero out
// elements that come from V2 in the first mask, and V1 in the second mask
MVT::v16i8, &pshufbMask[0], 16));
if (!TwoInputs)
return V1;
-
+
// Calculate the shuffle mask for the second input, shuffle it, and
// OR it with the first shuffled input.
pshufbMask.clear();
MVT::v16i8, &pshufbMask[0], 16));
return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
}
-
+
// No SSSE3 - Calculate in place words and then fix all out of place words
// With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
// the 16 different words that comprise the two doublequadword input vectors.
for (int i = 0; i != 8; ++i) {
int Elt0 = MaskVals[i*2];
int Elt1 = MaskVals[i*2+1];
-
+
// This word of the result is all undef, skip it.
if (Elt0 < 0 && Elt1 < 0)
continue;
-
+
// This word of the result is already in the correct place, skip it.
if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
continue;
if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
continue;
-
+
SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
SDValue InsElt;
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
EVT VT = SVOp->getValueType(0);
-
+
SmallVector<std::pair<int, int>, 8> Locs;
Locs.resize(4);
SmallVector<int, 8> Mask1(4U, -1);
V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
SmallVector<int, 8> Mask2(4U, -1);
-
+
for (unsigned i = 0; i != 4; ++i) {
if (Locs[i].first == -1)
continue;
// Promote splats to v4f32.
if (SVOp->isSplat()) {
- if (isMMX || NumElems < 4)
+ if (isMMX || NumElems < 4)
return Op;
return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
}
DAG, Subtarget, dl);
}
}
-
+
if (X86::isPSHUFDMask(SVOp))
return Op;
-
+
// Check if this can be converted into a logical shift.
bool isLeft = false;
unsigned ShAmt = 0;
ShAmt *= EVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
-
+
if (X86::isMOVLMask(SVOp)) {
if (V1IsUndef)
return V2;
if (!isMMX)
return Op;
}
-
+
// FIXME: fold these into legal mask.
if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
X86::isMOVSLDUPMask(SVOp) ||
ShAmt *= EVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
-
+
bool Commuted = false;
// FIXME: This should also accept a bitcast of a splat? Be careful, not
// 1,1,1,1 -> v8i16 though.
if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
// Shuffling low element of v1 into undef, just return v1.
- if (V2IsUndef)
+ if (V2IsUndef)
return V1;
// If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
// the instruction selector will not match, so get a canonical MOVL with
SVOp->getMask(PermMask);
if (isShuffleMaskLegal(PermMask, VT))
return Op;
-
+
// Handle v8i16 specifically since SSE can do byte extraction and insertion.
if (VT == MVT::v8i16) {
SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
if (NewOp.getNode())
return NewOp;
}
-
+
// Handle all 4 wide cases with a number of shuffles except for MMX.
if (NumElems == 4 && !isMMX)
return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
if (Idx == 0)
return Op;
-
+
// SHUFPS the element to the lowest double word, then movss.
int Mask[4] = { Idx, -1, -1, -1 };
EVT VVT = Op.getOperand(0).getValueType();
- SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
+ SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
DAG.getUNDEF(VVT), Mask);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
DAG.getIntPtrConstant(0));
// to a f64mem, the whole operation is folded into a single MOVHPDmr.
int Mask[2] = { 1, -1 };
EVT VVT = Op.getOperand(0).getValueType();
- SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
+ SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
DAG.getUNDEF(VVT), Mask);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
DAG.getIntPtrConstant(0));
SDValue
X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
-
+
// In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
// global base reg.
unsigned char OpFlag = 0;
OpFlag = X86II::MO_GOTOFF;
else if (Subtarget->isPICStyleStubPIC())
OpFlag = X86II::MO_PIC_BASE_OFFSET;
-
+
SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
CP->getAlignment(),
CP->getOffset(), OpFlag);
SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
-
+
// In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
// global base reg.
unsigned char OpFlag = 0;
OpFlag = X86II::MO_GOTOFF;
else if (Subtarget->isPICStyleStubPIC())
OpFlag = X86II::MO_PIC_BASE_OFFSET;
-
+
SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
OpFlag);
DebugLoc DL = JT->getDebugLoc();
Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
-
+
// With PIC, the address is actually $g + Offset.
if (OpFlag) {
Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DebugLoc::getUnknownLoc(), getPointerTy()),
Result);
}
-
+
return Result;
}
SDValue
X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
-
+
// In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
// global base reg.
unsigned char OpFlag = 0;
OpFlag = X86II::MO_GOTOFF;
else if (Subtarget->isPICStyleStubPIC())
OpFlag = X86II::MO_PIC_BASE_OFFSET;
-
+
SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
-
+
DebugLoc DL = Op.getDebugLoc();
Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
-
-
+
+
// With PIC, the address is actually $g + Offset.
if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
!Subtarget->is64Bit()) {
getPointerTy()),
Result);
}
-
+
return Result;
}
} else {
Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
}
-
+
if (Subtarget->isPICStyleRIPRel() &&
(M == CodeModel::Small || M == CodeModel::Kernel))
Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
assert(model == TLSModel::InitialExec);
OperandFlags = X86II::MO_INDNTPOFF;
}
-
+
// emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
// exec)
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
"TLS not implemented for non-ELF targets");
GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
const GlobalValue *GV = GA->getGlobal();
-
+
// If GV is an alias then use the aliasee for determining
// thread-localness.
if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
GV = GA->resolveAliasedGlobal(false);
-
+
TLSModel::Model model = getTLSModel(GV,
getTargetMachine().getRelocationModel());
-
+
switch (model) {
case TLSModel::GeneralDynamic:
case TLSModel::LocalDynamic: // not implemented
if (Subtarget->is64Bit())
return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
-
+
case TLSModel::InitialExec:
case TLSModel::LocalExec:
return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
Subtarget->is64Bit());
}
-
+
llvm_unreachable("Unreachable");
return SDValue();
}
unsigned MemSize = DstTy.getSizeInBits()/8;
int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
-
+
unsigned Opc;
switch (DstTy.getSimpleVT().SimpleTy) {
default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
EVT EltVT = VT;
- unsigned EltNum = 1;
- if (VT.isVector()) {
+ if (VT.isVector())
EltVT = VT.getVectorElementType();
- EltNum = VT.getVectorNumElements();
- }
std::vector<Constant*> CV;
if (EltVT == MVT::f64) {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
// CF = 1
X86CC = X86::COND_B;
break;
- case Intrinsic::x86_sse41_ptestnzc:
+ case Intrinsic::x86_sse41_ptestnzc:
// ZF and CF = 0
X86CC = X86::COND_A;
break;
}
-
+
SDValue LHS = Op.getOperand(1);
SDValue RHS = Op.getOperand(2);
SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
break;
}
}
+
+ // The vector shift intrinsics with scalars uses 32b shift amounts but
+ // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
+ // to be zero.
+ SDValue ShOps[4];
+ ShOps[0] = ShAmt;
+ ShOps[1] = DAG.getConstant(0, MVT::i32);
+ if (ShAmtVT == MVT::v4i32) {
+ ShOps[2] = DAG.getUNDEF(MVT::i32);
+ ShOps[3] = DAG.getUNDEF(MVT::i32);
+ ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
+ } else {
+ ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
+ }
+
EVT VT = Op.getValueType();
- ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
+ ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(NewIntNo, MVT::i32),
Op.getOperand(1), ShAmt);
} else {
const Function *Func =
cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
- unsigned CC = Func->getCallingConv();
+ CallingConv::ID CC = Func->getCallingConv();
unsigned NestReg;
switch (CC) {
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
/// are assumed to be legal.
bool
-X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
+X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
EVT VT) const {
// Only do shuffles on 128-bit vector types for now.
if (VT.getSizeInBits() == 64)
return nextMBB;
}
+// FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
+// all of this code can be replaced with that in the .td file.
+MachineBasicBlock *
+X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
+ unsigned numArgs, bool memArg) const {
+
+ MachineFunction *F = BB->getParent();
+ DebugLoc dl = MI->getDebugLoc();
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ unsigned Opc;
+
+ if (memArg) {
+ Opc = numArgs == 3 ?
+ X86::PCMPISTRM128rm :
+ X86::PCMPESTRM128rm;
+ } else {
+ Opc = numArgs == 3 ?
+ X86::PCMPISTRM128rr :
+ X86::PCMPESTRM128rr;
+ }
+
+ MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
+
+ for (unsigned i = 0; i < numArgs; ++i) {
+ MachineOperand &Op = MI->getOperand(i+1);
+
+ if (!(Op.isReg() && Op.isImplicit()))
+ MIB.addOperand(Op);
+ }
+
+ BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
+ .addReg(X86::XMM0);
+
+ F->DeleteMachineInstr(MI);
+
+ return BB;
+}
+
MachineBasicBlock *
X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
MachineInstr *MI,
return EndMBB;
}
+MachineBasicBlock *
+X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
+ MachineBasicBlock *BB) const {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc DL = MI->getDebugLoc();
+
+ // To "insert" a SELECT_CC instruction, we actually have to insert the
+ // diamond control-flow pattern. The incoming instruction knows the
+ // 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;
+
+ // thisMBB:
+ // ...
+ // TrueVal = ...
+ // cmpTY ccX, r1, r2
+ // bCC copy1MBB
+ // fallthrough --> copy0MBB
+ MachineBasicBlock *thisMBB = BB;
+ MachineFunction *F = BB->getParent();
+ MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
+ unsigned Opc =
+ X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
+ BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
+ F->insert(It, copy0MBB);
+ F->insert(It, sinkMBB);
+ // Update machine-CFG edges by transferring all successors of the current
+ // block to the new block which will contain the Phi node for the select.
+ sinkMBB->transferSuccessors(BB);
+
+ // Add the true and fallthrough blocks as its successors.
+ BB->addSuccessor(copy0MBB);
+ BB->addSuccessor(sinkMBB);
+
+ // copy0MBB:
+ // %FalseValue = ...
+ // # fallthrough to sinkMBB
+ BB = copy0MBB;
+
+ // Update machine-CFG edges
+ BB->addSuccessor(sinkMBB);
+
+ // sinkMBB:
+ // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
+ // ...
+ BB = sinkMBB;
+ BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
+ .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
+ .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
+
+ F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ return BB;
+}
+
+
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
MachineBasicBlock *BB) const {
- DebugLoc dl = MI->getDebugLoc();
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
+ case X86::CMOV_GR8:
case X86::CMOV_V1I64:
case X86::CMOV_FR32:
case X86::CMOV_FR64:
case X86::CMOV_V4F32:
case X86::CMOV_V2F64:
- case X86::CMOV_V2I64: {
- // To "insert" a SELECT_CC instruction, we actually have to insert the
- // diamond control-flow pattern. The incoming instruction knows the
- // 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;
-
- // thisMBB:
- // ...
- // TrueVal = ...
- // cmpTY ccX, r1, r2
- // bCC copy1MBB
- // fallthrough --> copy0MBB
- MachineBasicBlock *thisMBB = BB;
- MachineFunction *F = BB->getParent();
- MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
- MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
- unsigned Opc =
- X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
- BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
- F->insert(It, copy0MBB);
- F->insert(It, sinkMBB);
- // Update machine-CFG edges by transferring all successors of the current
- // block to the new block which will contain the Phi node for the select.
- sinkMBB->transferSuccessors(BB);
-
- // Add the true and fallthrough blocks as its successors.
- BB->addSuccessor(copy0MBB);
- BB->addSuccessor(sinkMBB);
-
- // copy0MBB:
- // %FalseValue = ...
- // # fallthrough to sinkMBB
- BB = copy0MBB;
-
- // Update machine-CFG edges
- BB->addSuccessor(sinkMBB);
-
- // sinkMBB:
- // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
- // ...
- BB = sinkMBB;
- BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
- .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
- .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
-
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
- return BB;
- }
+ case X86::CMOV_V2I64:
+ return EmitLoweredSelect(MI, BB);
case X86::FP32_TO_INT16_IN_MEM:
case X86::FP32_TO_INT32_IN_MEM:
case X86::FP80_TO_INT16_IN_MEM:
case X86::FP80_TO_INT32_IN_MEM:
case X86::FP80_TO_INT64_IN_MEM: {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc DL = MI->getDebugLoc();
+
// Change the floating point control register to use "round towards zero"
// mode when truncating to an integer value.
MachineFunction *F = BB->getParent();
int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
- addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
// Load the old value of the high byte of the control word...
unsigned OldCW =
F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
- addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
CWFrameIdx);
// Set the high part to be round to zero...
- addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
.addImm(0xC7F);
// Reload the modified control word now...
- addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
// Restore the memory image of control word to original value
- addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
.addReg(OldCW);
// Get the X86 opcode to use.
} else {
AM.Disp = Op.getImm();
}
- addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
+ addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
.addReg(MI->getOperand(X86AddrNumOperands).getReg());
// Reload the original control word now.
- addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
return BB;
}
+ // String/text processing lowering.
+ case X86::PCMPISTRM128REG:
+ return EmitPCMP(MI, BB, 3, false /* in-mem */);
+ case X86::PCMPISTRM128MEM:
+ return EmitPCMP(MI, BB, 3, true /* in-mem */);
+ case X86::PCMPESTRM128REG:
+ return EmitPCMP(MI, BB, 5, false /* in mem */);
+ case X86::PCMPESTRM128MEM:
+ return EmitPCMP(MI, BB, 5, true /* in mem */);
+
+ // Atomic Lowering.
case X86::ATOMAND32:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
X86::AND32ri, X86::MOV32rm,
// Get the LHS/RHS of the select.
SDValue LHS = N->getOperand(1);
SDValue RHS = N->getOperand(2);
-
+
// If we have SSE[12] support, try to form min/max nodes.
if (Subtarget->hasSSE2() &&
(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
} else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
switch (CC) {
default: break;
- case ISD::SETOGT: // (X > Y) ? Y : X -> min
- case ISD::SETUGT:
+ case ISD::SETOGT:
+ // This can use a min only if the LHS isn't NaN.
+ if (DAG.isKnownNeverNaN(LHS))
+ Opcode = X86ISD::FMIN;
+ else if (DAG.isKnownNeverNaN(RHS)) {
+ Opcode = X86ISD::FMIN;
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ }
+ break;
+
+ case ISD::SETUGT: // (X > Y) ? Y : X -> min
case ISD::SETGT:
if (!UnsafeFPMath) break;
// FALL THROUGH.
Opcode = X86ISD::FMIN;
break;
- case ISD::SETOLE: // (X <= Y) ? Y : X -> max
case ISD::SETULE:
+ // This can use a max only if the LHS isn't NaN.
+ if (DAG.isKnownNeverNaN(LHS))
+ Opcode = X86ISD::FMAX;
+ else if (DAG.isKnownNeverNaN(RHS)) {
+ Opcode = X86ISD::FMAX;
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ }
+ break;
+
+ case ISD::SETOLE: // (X <= Y) ? Y : X -> max
case ISD::SETLE:
if (!UnsafeFPMath) break;
// FALL THROUGH.
if (Opcode)
return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
}
-
+
// If this is a select between two integer constants, try to do some
// optimizations.
if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
// If this is efficiently invertible, canonicalize the LHSC/RHSC values
// so that TrueC (the true value) is larger than FalseC.
bool NeedsCondInvert = false;
-
+
if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
// Efficiently invertible.
(Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
NeedsCondInvert = true;
std::swap(TrueC, FalseC);
}
-
+
// Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
if (FalseC->getAPIntValue() == 0 &&
TrueC->getAPIntValue().isPowerOf2()) {
if (NeedsCondInvert) // Invert the condition if needed.
Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
DAG.getConstant(1, Cond.getValueType()));
-
+
// Zero extend the condition if needed.
Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
-
+
unsigned ShAmt = TrueC->getAPIntValue().logBase2();
return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
DAG.getConstant(ShAmt, MVT::i8));
}
-
+
// Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
if (NeedsCondInvert) // Invert the condition if needed.
Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
DAG.getConstant(1, Cond.getValueType()));
-
+
// Zero extend the condition if needed.
Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
FalseC->getValueType(0), Cond);
return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
SDValue(FalseC, 0));
}
-
+
// Optimize cases that will turn into an LEA instruction. This requires
// an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
-
+
bool isFastMultiplier = false;
if (Diff < 10) {
switch ((unsigned char)Diff) {
break;
}
}
-
+
if (isFastMultiplier) {
APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
if (NeedsCondInvert) // Invert the condition if needed.
Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
DAG.getConstant(1, Cond.getValueType()));
-
+
// Zero extend the condition if needed.
Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
Cond);
if (Diff != 1)
Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
DAG.getConstant(Diff, Cond.getValueType()));
-
+
// Add the base if non-zero.
if (FalseC->getAPIntValue() != 0)
Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
SDValue(FalseC, 0));
return Cond;
}
- }
+ }
}
}
-
+
return SDValue();
}
static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI) {
DebugLoc DL = N->getDebugLoc();
-
+
// If the flag operand isn't dead, don't touch this CMOV.
if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
return SDValue();
-
+
// If this is a select between two integer constants, try to do some
// optimizations. Note that the operands are ordered the opposite of SELECT
// operands.
// Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
// larger than FalseC (the false value).
X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
-
+
if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
CC = X86::GetOppositeBranchCondition(CC);
std::swap(TrueC, FalseC);
}
-
+
// Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
// This is efficient for any integer data type (including i8/i16) and
// shift amount.
SDValue Cond = N->getOperand(3);
Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
DAG.getConstant(CC, MVT::i8), Cond);
-
+
// Zero extend the condition if needed.
Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
-
+
unsigned ShAmt = TrueC->getAPIntValue().logBase2();
Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
DAG.getConstant(ShAmt, MVT::i8));
return DCI.CombineTo(N, Cond, SDValue());
return Cond;
}
-
+
// Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
// for any integer data type, including i8/i16.
if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
SDValue Cond = N->getOperand(3);
Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
DAG.getConstant(CC, MVT::i8), Cond);
-
+
// Zero extend the condition if needed.
Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
FalseC->getValueType(0), Cond);
Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
SDValue(FalseC, 0));
-
+
if (N->getNumValues() == 2) // Dead flag value?
return DCI.CombineTo(N, Cond, SDValue());
return Cond;
}
-
+
// Optimize cases that will turn into an LEA instruction. This requires
// an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
-
+
bool isFastMultiplier = false;
if (Diff < 10) {
switch ((unsigned char)Diff) {
break;
}
}
-
+
if (isFastMultiplier) {
APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
SDValue Cond = N->getOperand(3);
return DCI.CombineTo(N, Cond, SDValue());
return Cond;
}
- }
+ }
}
}
return SDValue();
std::swap(MulAmt1, MulAmt2);
SDValue NewMul;
- if (isPowerOf2_64(MulAmt1))
+ if (isPowerOf2_64(MulAmt1))
NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
else
NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
DAG.getConstant(MulAmt1, VT));
- if (isPowerOf2_64(MulAmt2))
+ if (isPowerOf2_64(MulAmt2))
NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
- else
+ else
NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
DAG.getConstant(MulAmt2, VT));
SDValue ShAmtOp = N->getOperand(1);
EVT EltVT = VT.getVectorElementType();
DebugLoc DL = N->getDebugLoc();
- SDValue BaseShAmt;
+ SDValue BaseShAmt = SDValue();
if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
unsigned NumElts = VT.getVectorNumElements();
unsigned i = 0;
}
} else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
- BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
- DAG.getIntPtrConstant(0));
+ SDValue InVec = ShAmtOp.getOperand(0);
+ if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
+ unsigned NumElts = InVec.getValueType().getVectorNumElements();
+ unsigned i = 0;
+ for (; i != NumElts; ++i) {
+ SDValue Arg = InVec.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ BaseShAmt = Arg;
+ break;
+ }
+ } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
+ unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
+ if (C->getZExtValue() == SplatIdx)
+ BaseShAmt = InVec.getOperand(1);
+ }
+ }
+ if (BaseShAmt.getNode() == 0)
+ BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
+ DAG.getIntPtrConstant(0));
} else
return SDValue();
+ // The shift amount is an i32.
if (EltVT.bitsGT(MVT::i32))
BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
else if (EltVT.bitsLT(MVT::i32))
- BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
+ BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
// The shift amount is identical so we can do a vector shift.
SDValue ValOp = N->getOperand(0);
const Function *F = DAG.getMachineFunction().getFunction();
bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
- bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
+ bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
&& Subtarget->hasSSE2();
if ((VT.isVector() ||
(VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
Op = Op.getOperand(0);
EVT VT = N->getValueType(0), OpVT = Op.getValueType();
if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
- VT.getVectorElementType().getSizeInBits() ==
+ VT.getVectorElementType().getSizeInBits() ==
OpVT.getVectorElementType().getSizeInBits()) {
return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
}
// On X86 and X86-64, atomic operations are lowered to locked instructions.
// Locked instructions, in turn, have implicit fence semantics (all memory
// operations are flushed before issuing the locked instruction, and the
-// are not buffered), so we can fold away the common pattern of
+// are not buffered), so we can fold away the common pattern of
// fence-atomic-fence.
static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
SDValue atomic = N->getOperand(0);
default:
return SDValue();
}
-
+
SDValue fence = atomic.getOperand(0);
if (fence.getOpcode() != ISD::MEMBARRIER)
return SDValue();
-
+
switch (atomic.getOpcode()) {
case ISD::ATOMIC_CMP_SWAP:
return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
// we will turn this bswap into something that will be lowered to logical ops
// instead of emitting the bswap asm. For now, we don't support 486 or lower
// so don't worry about this.
-
+
// Verify this is a simple bswap.
if (CI->getNumOperands() != 2 ||
CI->getType() != CI->getOperand(1)->getType() ||
!CI->getType()->isInteger())
return false;
-
+
const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
if (!Ty || Ty->getBitWidth() % 16 != 0)
return false;
-
+
// Okay, we can do this xform, do so now.
const Type *Tys[] = { Ty };
Module *M = CI->getParent()->getParent()->getParent();
Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
-
+
Value *Op = CI->getOperand(1);
Op = CallInst::Create(Int, Op, CI->getName(), CI);
-
+
CI->replaceAllUsesWith(Op);
CI->eraseFromParent();
return true;
}
break;
case 3:
- if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
+ if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
Constraints.size() >= 2 &&
Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
// Otherwise, this isn't something we can handle, reject it.
return;
}
-
+
GlobalValue *GV = GA->getGlobal();
// If we require an extra load to get this address, as in PIC mode, we
// can't accept it.
break;
}
- // 32-bit fallthrough
+ // 32-bit fallthrough
case 'Q': // Q_REGS
if (VT == MVT::i32)
return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
// Not found as a standard register?
if (Res.second == 0) {
- // GCC calls "st(0)" just plain "st".
+ // Map st(0) -> st(7) -> ST0
+ if (Constraint.size() == 7 && Constraint[0] == '{' &&
+ tolower(Constraint[1]) == 's' &&
+ tolower(Constraint[2]) == 't' &&
+ Constraint[3] == '(' &&
+ (Constraint[4] >= '0' && Constraint[4] <= '7') &&
+ Constraint[5] == ')' &&
+ Constraint[6] == '}') {
+
+ Res.first = X86::ST0+Constraint[4]-'0';
+ Res.second = X86::RFP80RegisterClass;
+ return Res;
+ }
+
+ // GCC allows "st(0)" to be called just plain "st".
if (StringsEqualNoCase("{st}", Constraint)) {
Res.first = X86::ST0;
Res.second = X86::RFP80RegisterClass;
+ return Res;
}
+
+ // flags -> EFLAGS
+ if (StringsEqualNoCase("{flags}", Constraint)) {
+ Res.first = X86::EFLAGS;
+ Res.second = X86::CCRRegisterClass;
+ return Res;
+ }
+
// 'A' means EAX + EDX.
if (Constraint == "A") {
Res.first = X86::EAX;
Res.second = X86::GR32_ADRegisterClass;
+ return Res;
}
return Res;
}