#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
-#include "X86MachineFunctionInfo.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/Intrinsics.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/VectorExtras.h"
-#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
-#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/TargetOptions.h"
DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
// Forward declarations.
-static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl);
+static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
+ SDValue V2);
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
: TargetLowering(TM) {
X86ScalarSSEf32 = Subtarget->hasSSE1();
X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
- bool Fast = false;
-
RegInfo = TM.getRegisterInfo();
TD = getTargetData();
setUseUnderscoreSetJmp(true);
setUseUnderscoreLongJmp(true);
}
-
+
// Set up the register classes.
addRegisterClass(MVT::i8, X86::GR8RegisterClass);
addRegisterClass(MVT::i16, X86::GR16RegisterClass);
setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
- // We don't accept any truncstore of integer registers.
+ // We don't accept any truncstore of integer registers.
setTruncStoreAction(MVT::i64, MVT::i32, Expand);
setTruncStoreAction(MVT::i64, MVT::i16, Expand);
setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
if (Subtarget->is64Bit()) {
- setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
- } else {
+ setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
+ } else if (!UseSoftFloat) {
if (X86ScalarSSEf64) {
// We have an impenetrably clever algorithm for ui64->double only.
setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
-
- // We have faster algorithm for ui32->single only.
- setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
- } else
- setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
+ }
+ // We have an algorithm for SSE2, and we turn this into a 64-bit
+ // FILD for other targets.
+ setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
}
// Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
// this operation.
setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
- // SSE has no i16 to fp conversion, only i32
- if (X86ScalarSSEf32) {
- setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
- // f32 and f64 cases are Legal, f80 case is not
- setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
+
+ if (!UseSoftFloat) {
+ // SSE has no i16 to fp conversion, only i32
+ if (X86ScalarSSEf32) {
+ setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
+ // f32 and f64 cases are Legal, f80 case is not
+ setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
+ } else {
+ setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
+ setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
+ }
} else {
- setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
- setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
+ setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
+ setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
}
// In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
if (Subtarget->is64Bit()) {
setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
- } else {
+ } else if (!UseSoftFloat) {
if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
// Expand FP_TO_UINT into a select.
// FIXME: We would like to use a Custom expander here eventually to do
// the optimal thing for SSE vs. the default expansion in the legalizer.
setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
else
- // With SSE3 we can use fisttpll to convert to a signed i64.
- setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
+ // With SSE3 we can use fisttpll to convert to a signed i64; without
+ // SSE, we're stuck with a fistpll.
+ setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
}
// TODO: when we have SSE, these could be more efficient, by using movd/movq.
setOperationAction(ISD::FREM , MVT::f64 , Expand);
setOperationAction(ISD::FREM , MVT::f80 , Expand);
setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
-
+
setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
// cases we handle.
addLegalFPImmediate(APFloat(+0.0)); // xorpd
addLegalFPImmediate(APFloat(+0.0f)); // xorps
-
- // Floating truncations from f80 and extensions to f80 go through memory.
- // If optimizing, we lie about this though and handle it in
- // InstructionSelectPreprocess so that dagcombine2 can hack on these.
- if (Fast) {
- setConvertAction(MVT::f32, MVT::f80, Expand);
- setConvertAction(MVT::f64, MVT::f80, Expand);
- setConvertAction(MVT::f80, MVT::f32, Expand);
- setConvertAction(MVT::f80, MVT::f64, Expand);
- }
} else if (!UseSoftFloat && X86ScalarSSEf32) {
// Use SSE for f32, x87 for f64.
// Set up the FP register classes.
addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
- // SSE <-> X87 conversions go through memory. If optimizing, we lie about
- // this though and handle it in InstructionSelectPreprocess so that
- // dagcombine2 can hack on these.
- if (Fast) {
- setConvertAction(MVT::f32, MVT::f64, Expand);
- setConvertAction(MVT::f32, MVT::f80, Expand);
- setConvertAction(MVT::f80, MVT::f32, Expand);
- setConvertAction(MVT::f64, MVT::f32, Expand);
- // And x87->x87 truncations also.
- setConvertAction(MVT::f80, MVT::f64, Expand);
- }
-
if (!UnsafeFPMath) {
setOperationAction(ISD::FSIN , MVT::f64 , Expand);
setOperationAction(ISD::FCOS , MVT::f64 , Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
- // Floating truncations go through memory. If optimizing, we lie about
- // this though and handle it in InstructionSelectPreprocess so that
- // dagcombine2 can hack on these.
- if (Fast) {
- setConvertAction(MVT::f80, MVT::f32, Expand);
- setConvertAction(MVT::f64, MVT::f32, Expand);
- setConvertAction(MVT::f80, MVT::f64, Expand);
- }
-
if (!UnsafeFPMath) {
setOperationAction(ISD::FSIN , MVT::f64 , Expand);
setOperationAction(ISD::FCOS , MVT::f64 , Expand);
TmpFlt2.changeSign();
addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
}
-
+
if (!UnsafeFPMath) {
setOperationAction(ISD::FSIN , MVT::f80 , Expand);
setOperationAction(ISD::FCOS , MVT::f80 , Expand);
setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
+ setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
}
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
- // FIXME: add MMX packed arithmetics
-
setOperationAction(ISD::ADD, MVT::v8i8, Legal);
setOperationAction(ISD::ADD, MVT::v4i16, Legal);
setOperationAction(ISD::ADD, MVT::v2i32, Legal);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
- setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
+ setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
if (!UseSoftFloat && Subtarget->hasSSE2()) {
addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
- // FIXME: Unfortunately -soft-float means XMM registers cannot be used even
- // for integer operations.
+ // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
+ // registers cannot be used even for integer operations.
addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
}
+
setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
+
if (Subtarget->is64Bit()) {
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
-
+
+ setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
+ setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
+ if (!DisableMMX && Subtarget->hasMMX()) {
+ setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
+ setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
+ }
}
if (Subtarget->hasSSE41()) {
if (Subtarget->hasSSE42()) {
setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
}
-
+
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::USUBO, MVT::i64, Custom);
setOperationAction(ISD::SMULO, MVT::i32, Custom);
setOperationAction(ISD::SMULO, MVT::i64, Custom);
- setOperationAction(ISD::UMULO, MVT::i32, Custom);
- setOperationAction(ISD::UMULO, MVT::i64, Custom);
+
+ if (!Subtarget->is64Bit()) {
+ // These libcalls are not available in 32-bit.
+ setLibcallName(RTLIB::SHL_I128, 0);
+ setLibcallName(RTLIB::SRL_I128, 0);
+ setLibcallName(RTLIB::SRA_I128, 0);
+ }
// We have target-specific dag combine patterns for the following nodes:
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
setTargetDAGCombine(ISD::SRA);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::STORE);
+ if (Subtarget->is64Bit())
+ setTargetDAGCombine(ISD::MUL);
computeRegisterProperties();
maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
allowUnalignedMemoryAccesses = true; // x86 supports it!
setPrefLoopAlignment(16);
+ benefitFromCodePlacementOpt = true;
}
/// determining it.
MVT
X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
- bool isSrcConst, bool isSrcStr) const {
+ bool isSrcConst, bool isSrcStr,
+ SelectionDAG &DAG) const {
// FIXME: This turns off use of xmm stores for memset/memcpy on targets like
// linux. This is because the stack realignment code can't handle certain
// cases like PR2962. This should be removed when PR2962 is fixed.
- if (Subtarget->getStackAlignment() >= 16) {
+ const Function *F = DAG.getMachineFunction().getFunction();
+ bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
+ if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
return MVT::v4i32;
if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
return MVT::i32;
}
-
/// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
/// jumptable.
SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
-
+
SmallVector<CCValAssign, 16> RVLocs;
unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
-
+
// If this is the first return lowered for this function, add the regs to the
// liveout set for the function.
if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
}
SDValue Chain = Op.getOperand(0);
-
+
// Handle tail call return.
Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
if (Chain.getOpcode() == X86ISD::TAILCALL) {
SDValue StackAdjustment = TailCall.getOperand(2);
assert(((TargetAddress.getOpcode() == ISD::Register &&
(cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
- cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
+ cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
- TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
+ TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
"Expecting an global address, external symbol, or register");
assert(StackAdjustment.getOpcode() == ISD::Constant &&
"Expecting a const value");
for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
Operands.push_back(Chain.getOperand(i));
}
- return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
+ return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
Operands.size());
}
-
+
// Regular return.
SDValue Flag;
RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
// Operand #1 = Bytes To Pop
RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
-
+
// Copy the result values into the output registers.
for (unsigned i = 0; i != RVLocs.size(); ++i) {
CCValAssign &VA = RVLocs[i];
assert(VA.isRegLoc() && "Can only return in registers!");
SDValue ValToCopy = Op.getOperand(i*2+1);
-
+
// Returns in ST0/ST1 are handled specially: these are pushed as operands to
// the RET instruction and handled by the FP Stackifier.
if (VA.getLocReg() == X86::ST0 ||
continue;
}
+ // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
+ // which is returned in RAX / RDX.
+ if (Subtarget->is64Bit()) {
+ MVT ValVT = ValToCopy.getValueType();
+ if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
+ ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
+ if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
+ ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
+ }
+ }
+
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
Flag = Chain.getValue(1);
}
Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
Flag = Chain.getValue(1);
}
-
+
RetOps[0] = Chain; // Update chain.
// Add the flag if we have it.
if (Flag.getNode())
RetOps.push_back(Flag);
-
- return DAG.getNode(X86ISD::RET_FLAG, dl,
+
+ return DAG.getNode(X86ISD::RET_FLAG, dl,
MVT::Other, &RetOps[0], RetOps.size());
}
/// being lowered. The returns a SDNode with the same number of values as the
/// ISD::CALL.
SDNode *X86TargetLowering::
-LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
+LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
unsigned CallingConv, SelectionDAG &DAG) {
- DebugLoc dl = TheCall->getDebugLoc();
+ DebugLoc dl = TheCall->getDebugLoc();
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RVLocs;
bool isVarArg = TheCall->isVarArg();
CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
SmallVector<SDValue, 8> ResultVals;
-
+
// Copy all of the result registers out of their specified physreg.
for (unsigned i = 0; i != RVLocs.size(); ++i) {
CCValAssign &VA = RVLocs[i];
MVT CopyVT = VA.getValVT();
-
+
// If this is x86-64, and we disabled SSE, we can't return FP values
- if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
+ if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
cerr << "SSE register return with SSE disabled\n";
exit(1);
isScalarFPTypeInSSEReg(VA.getValVT())) {
CopyVT = MVT::f80;
}
-
- Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
- CopyVT, InFlag).getValue(1);
- SDValue Val = Chain.getValue(0);
+
+ SDValue Val;
+ if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
+ // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
+ if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
+ Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
+ MVT::v2i64, InFlag).getValue(1);
+ Val = Chain.getValue(0);
+ Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
+ Val, DAG.getConstant(0, MVT::i64));
+ } else {
+ Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
+ MVT::i64, InFlag).getValue(1);
+ Val = Chain.getValue(0);
+ }
+ Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
+ } else {
+ Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
+ CopyVT, InFlag).getValue(1);
+ Val = Chain.getValue(0);
+ }
InFlag = Chain.getValue(2);
if (CopyVT != VA.getValVT()) {
// This truncation won't change the value.
DAG.getIntPtrConstant(1));
}
-
+
ResultVals.push_back(Val);
}
// For info on fast calling convention see Fast Calling Convention (tail call)
// implementation LowerX86_32FastCCCallTo.
-/// AddLiveIn - This helper function adds the specified physical register to the
-/// MachineFunction as a live in value. It also creates a corresponding virtual
-/// register for it.
-static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
- const TargetRegisterClass *RC) {
- assert(RC->contains(PReg) && "Not the correct regclass!");
- unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
- MF.getRegInfo().addLiveIn(PReg, VReg);
- return VReg;
-}
-
/// CallIsStructReturn - Determines whether a CALL node uses struct return
/// semantics.
static bool CallIsStructReturn(CallSDNode *TheCall) {
if (Subtarget->is64Bit()) {
if (Subtarget->isTargetWin64())
return CC_X86_Win64_C;
- else if (CC == CallingConv::Fast && PerformTailCallOpt)
- return CC_X86_64_TailCall;
else
return CC_X86_64_C;
}
/// CallRequiresFnAddressInReg - Check whether the call requires the function
/// address to be loaded in a register.
-bool
+bool
X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
- return !Is64Bit && IsTailCall &&
+ return !Is64Bit && IsTailCall &&
getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
Subtarget->isPICStyleGOT();
}
/// by "Src" to address "Dst" with size and alignment information specified by
/// the specific parameter attribute. The copy will be passed as a byval
/// function parameter.
-static SDValue
+static SDValue
CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
DebugLoc dl) {
bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
// FIXME: For now, all byval parameter objects are marked mutable. This can be
- // changed with more analysis.
+ // changed with more analysis.
// In case of tail call optimization mark all arguments mutable. Since they
// could be overwritten by lowering of arguments in case of a tail call.
int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
DebugLoc dl = Op.getDebugLoc();
-
+
const Function* Fn = MF.getFunction();
if (Fn->hasExternalLinkage() &&
Subtarget->isTargetCygMing() &&
// Decorate the function name.
FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
-
+
MachineFrameInfo *MFI = MF.getFrameInfo();
SDValue Root = Op.getOperand(0);
bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
-
+
SmallVector<SDValue, 8> ArgValues;
unsigned LastVal = ~0U;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
assert(VA.getValNo() != LastVal &&
"Don't support value assigned to multiple locs yet");
LastVal = VA.getValNo();
-
+
if (VA.isRegLoc()) {
MVT RegVT = VA.getLocVT();
TargetRegisterClass *RC = NULL;
assert(0 && "Unknown argument type!");
}
- unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
+ unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
-
+
// If this is an 8 or 16-bit value, it is really passed promoted to 32
// bits. Insert an assert[sz]ext to capture this, then truncate to the
// right size.
else if (VA.getLocInfo() == CCValAssign::ZExt)
ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
DAG.getValueType(VA.getValVT()));
-
+
if (VA.getLocInfo() != CCValAssign::Full)
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
-
+
// Handle MMX values passed in GPRs.
if (Is64Bit && RegVT != VA.getLocVT()) {
if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
}
}
-
+
ArgValues.push_back(ArgValue);
} else {
assert(VA.isMemLoc());
unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
TotalNumXMMRegs);
+ bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
"SSE register cannot be used when SSE is disabled!");
- assert(!(NumXMMRegs && UseSoftFloat) &&
+ assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
"SSE register cannot be used when SSE is disabled!");
- if (UseSoftFloat || !Subtarget->hasSSE1()) {
+ if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
// Kernel mode asks for SSE to be disabled, so don't push them
// on the stack.
TotalNumXMMRegs = 0;
- }
+
// For X86-64, if there are vararg parameters that are passed via
// registers, then we must store them to their spots on the stack so they
// may be loaded by deferencing the result of va_next.
SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
DAG.getIntPtrConstant(VarArgsGPOffset));
for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
- unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
- X86::GR64RegisterClass);
+ unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
+ X86::GR64RegisterClass);
SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
DAG.getIntPtrConstant(VarArgsFPOffset));
for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
- unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
- X86::VR128RegisterClass);
+ unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
+ X86::VR128RegisterClass);
SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
&MemOps[0], MemOps.size());
}
}
-
+
ArgValues.push_back(Root);
// Some CCs need callee pop.
BytesToPopOnReturn = 0; // Callee pops nothing.
// If this is an sret function, the return should pop the hidden pointer.
if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
- BytesToPopOnReturn = 4;
+ BytesToPopOnReturn = 4;
BytesCallerReserves = StackSize;
}
/// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
/// optimization is performed and it is required.
-SDValue
-X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
+SDValue
+X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
SDValue &OutRetAddr,
- SDValue Chain,
- bool IsTailCall,
- bool Is64Bit,
+ SDValue Chain,
+ bool IsTailCall,
+ bool Is64Bit,
int FPDiff,
DebugLoc dl) {
if (!IsTailCall || FPDiff==0) return Chain;
/// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
/// optimization is performed and it is required (FPDiff!=0).
-static SDValue
-EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
+static SDValue
+EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
SDValue Chain, SDValue RetAddrFrIdx,
bool Is64Bit, int FPDiff, DebugLoc dl) {
// Store the return address to the appropriate stack slot.
if (!FPDiff) return Chain;
// Calculate the new stack slot for the return address.
int SlotSize = Is64Bit ? 8 : 4;
- int NewReturnAddrFI =
+ int NewReturnAddrFI =
MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
- Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
+ Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
return Chain;
}
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
-
+
// Get a count of how many bytes are to be pushed on the stack.
unsigned NumBytes = CCInfo.getNextStackOffset();
if (PerformTailCallOpt && CC == CallingConv::Fast)
int FPDiff = 0;
if (IsTailCall) {
// Lower arguments at fp - stackoffset + fpdiff.
- unsigned NumBytesCallerPushed =
+ unsigned NumBytesCallerPushed =
MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
FPDiff = NumBytesCallerPushed - NumBytes;
SDValue Arg = TheCall->getArg(i);
ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
bool isByVal = Flags.isByVal();
-
+
// Promote the value if needed.
switch (VA.getLocInfo()) {
default: assert(0 && "Unknown loc info!");
Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
break;
}
-
+
if (VA.isRegLoc()) {
if (Is64Bit) {
MVT RegVT = VA.getLocVT();
// Special case: passing MMX values in XMM registers.
Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
- Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
- DAG.getUNDEF(MVT::v2i64), Arg,
- getMOVLMask(2, DAG, dl));
+ Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
break;
}
}
assert(VA.isMemLoc());
if (StackPtr.getNode() == 0)
StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
-
+
MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
Chain, Arg, Flags));
}
}
}
-
+
if (!MemOpChains.empty())
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&MemOpChains[0], MemOpChains.size());
// tail call optimization the copies to registers are lowered later.
if (!IsTailCall)
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
+ Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
RegsToPass[i].second, InFlag);
InFlag = Chain.getValue(1);
}
// ELF / PIC requires GOT in the EBX register before function calls via PLT
- // GOT pointer.
+ // GOT pointer.
if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
- DAG.getNode(X86ISD::GlobalBaseReg,
- DebugLoc::getUnknownLoc(),
+ DAG.getNode(X86ISD::GlobalBaseReg,
+ DebugLoc::getUnknownLoc(),
getPointerTy()),
InFlag);
InFlag = Chain.getValue(1);
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
- assert((Subtarget->hasSSE1() || !NumXMMRegs)
+ assert((Subtarget->hasSSE1() || !NumXMMRegs)
&& "SSE registers cannot be used when SSE is disabled");
-
+
Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
InFlag = Chain.getValue(1);
// Copy relative to framepointer.
SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
if (StackPtr.getNode() == 0)
- StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
+ StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
getPointerTy());
Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
MemOpChains2.push_back(
DAG.getStore(Chain, dl, Arg, FIN,
PseudoSourceValue::getFixedStack(FI), 0));
- }
+ }
}
}
// Copy arguments to their registers.
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
+ Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
RegsToPass[i].second, InFlag);
InFlag = Chain.getValue(1);
}
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
} else if (IsTailCall) {
- unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
+ unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
Chain = DAG.getCopyToReg(Chain, dl,
- DAG.getRegister(Opc, getPointerTy()),
+ DAG.getRegister(Opc, getPointerTy()),
Callee,InFlag);
Callee = DAG.getRegister(Opc, getPointerTy());
// Add register as live out.
DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
}
-
+
// Returns a chain & a flag for retval copy to use.
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
SmallVector<SDValue, 8> Ops;
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
DAG.getIntPtrConstant(0, true), InFlag);
InFlag = Chain.getValue(1);
-
+
// Returns a chain & a flag for retval copy to use.
NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
Ops.clear();
}
-
+
Ops.push_back(Chain);
Ops.push_back(Callee);
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
Ops.push_back(DAG.getRegister(RegsToPass[i].first,
RegsToPass[i].second.getValueType()));
-
+
// Add an implicit use GOT pointer in EBX.
if (!IsTailCall && !Is64Bit &&
getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
Ops.push_back(InFlag);
if (IsTailCall) {
- assert(InFlag.getNode() &&
+ assert(InFlag.getNode() &&
"Flag must be set. Depend on flag being set in LowerRET");
Chain = DAG.getNode(X86ISD::TAILCALL, dl,
TheCall->getVTList(), &Ops[0], Ops.size());
-
+
return SDValue(Chain.getNode(), Op.getResNo());
}
NumBytesForCalleeToPush = 4;
else
NumBytesForCalleeToPush = 0; // Callee pops nothing.
-
+
// Returns a flag for retval copy to use.
Chain = DAG.getCALLSEQ_END(Chain,
DAG.getIntPtrConstant(NumBytes, true),
// arg1
// arg2
// RETADDR
-// [ new RETADDR
+// [ new RETADDR
// move area ]
// (possible EBP)
// ESI
/// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
/// for a 16 byte align requirement.
-unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
+unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
SelectionDAG& DAG) {
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
const TargetFrameInfo &TFI = *TM.getFrameInfo();
unsigned StackAlignment = TFI.getStackAlignment();
- uint64_t AlignMask = StackAlignment - 1;
+ uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
uint64_t SlotSize = TD->getPointerSize();
if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
} else {
// Mask out lower bits, add stackalignment once plus the 12 bytes.
- Offset = ((~AlignMask) & Offset) + StackAlignment +
+ Offset = ((~AlignMask) & Offset) + StackAlignment +
(StackAlignment-SlotSize);
}
return Offset;
case ISD::SETUGE: return X86::COND_AE;
}
}
-
+
// First determine if it is required or is profitable to flip the operands.
// If LHS is a foldable load, but RHS is not, flip the condition.
}
}
-/// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
-/// true if Op is undef or if its value falls within the specified range (L, H].
-static bool isUndefOrInRange(SDValue Op, unsigned Low, unsigned Hi) {
- if (Op.getOpcode() == ISD::UNDEF)
- return true;
-
- unsigned Val = cast<ConstantSDNode>(Op)->getZExtValue();
- return (Val >= Low && Val < Hi);
+/// isUndefOrInRange - Return true if Val is undef or if its value falls within
+/// the specified range (L, H].
+static bool isUndefOrInRange(int Val, int Low, int Hi) {
+ return (Val < 0) || (Val >= Low && Val < Hi);
}
-/// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
-/// true if Op is undef or if its value equal to the specified value.
-static bool isUndefOrEqual(SDValue Op, unsigned Val) {
- if (Op.getOpcode() == ISD::UNDEF)
+/// isUndefOrEqual - Val is either less than zero (undef) or equal to the
+/// specified value.
+static bool isUndefOrEqual(int Val, int CmpVal) {
+ if (Val < 0 || Val == CmpVal)
return true;
- return cast<ConstantSDNode>(Op)->getZExtValue() == Val;
+ return false;
}
-/// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to PSHUFD.
-bool X86::isPSHUFDMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
- return false;
-
- // Check if the value doesn't reference the second vector.
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Arg)->getZExtValue() >= e)
- return false;
- }
-
- return true;
+/// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
+/// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
+/// the second operand.
+static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
+ return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
+ if (VT == MVT::v2f64 || VT == MVT::v2i64)
+ return (Mask[0] < 2 && Mask[1] < 2);
+ return false;
}
-/// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to PSHUFHW.
-bool X86::isPSHUFHWMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
+bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isPSHUFDMask(M, N->getValueType(0));
+}
- if (N->getNumOperands() != 8)
+/// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
+/// is suitable for input to PSHUFHW.
+static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ if (VT != MVT::v8i16)
return false;
-
- // Lower quadword copied in order.
- for (unsigned i = 0; i != 4; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Arg)->getZExtValue() != i)
+
+ // 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 (unsigned i = 4; i != 8; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val < 4 || Val > 7)
+ for (int i = 4; i != 8; ++i)
+ if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
return false;
- }
-
+
return true;
}
-/// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to PSHUFLW.
-bool X86::isPSHUFLWMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
+bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isPSHUFHWMask(M, N->getValueType(0));
+}
- if (N->getNumOperands() != 8)
+/// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
+/// is suitable for input to PSHUFLW.
+static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ if (VT != MVT::v8i16)
return false;
-
+
// Upper quadword copied in order.
- for (unsigned i = 4; i != 8; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i))
+ for (int i = 4; i != 8; ++i)
+ if (Mask[i] >= 0 && Mask[i] != i)
return false;
-
+
// Lower quadword shuffled.
- for (unsigned i = 0; i != 4; ++i)
- if (!isUndefOrInRange(N->getOperand(i), 0, 4))
+ for (int i = 0; i != 4; ++i)
+ if (Mask[i] >= 4)
return false;
-
+
return true;
}
+bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isPSHUFLWMask(M, N->getValueType(0));
+}
+
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to SHUFP*.
-template<class SDOperand>
-static bool isSHUFPMask(SDOperand *Elems, unsigned NumElems) {
- if (NumElems != 2 && NumElems != 4) return false;
-
- unsigned Half = NumElems / 2;
- for (unsigned i = 0; i < Half; ++i)
- if (!isUndefOrInRange(Elems[i], 0, NumElems))
+static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ 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))
return false;
- for (unsigned i = Half; i < NumElems; ++i)
- if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
+ for (int i = Half; i < NumElems; ++i)
+ if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
return false;
-
+
return true;
}
-bool X86::isSHUFPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
+bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isSHUFPMask(M, N->getValueType(0));
}
/// isCommutedSHUFP - Returns true if the shuffle mask is exactly
/// the reverse of what x86 shuffles want. x86 shuffles requires the lower
/// half elements to come from vector 1 (which would equal the dest.) and
/// the upper half to come from vector 2.
-template<class SDOperand>
-static bool isCommutedSHUFP(SDOperand *Ops, unsigned NumOps) {
- if (NumOps != 2 && NumOps != 4) return false;
-
- unsigned Half = NumOps / 2;
- for (unsigned i = 0; i < Half; ++i)
- if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
+static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ 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], NumElems, NumElems*2))
return false;
- for (unsigned i = Half; i < NumOps; ++i)
- if (!isUndefOrInRange(Ops[i], 0, NumOps))
+ for (int i = Half; i < NumElems; ++i)
+ if (!isUndefOrInRange(Mask[i], 0, NumElems))
return false;
return true;
}
-static bool isCommutedSHUFP(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
+static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return isCommutedSHUFPMask(M, N->getValueType(0));
}
/// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVHLPS.
-bool X86::isMOVHLPSMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 4)
+bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
+ if (N->getValueType(0).getVectorNumElements() != 4)
return false;
// Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
- return isUndefOrEqual(N->getOperand(0), 6) &&
- isUndefOrEqual(N->getOperand(1), 7) &&
- isUndefOrEqual(N->getOperand(2), 2) &&
- isUndefOrEqual(N->getOperand(3), 3);
-}
-
-/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
-/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
-/// <2, 3, 2, 3>
-bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 4)
- return false;
-
- // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
- return isUndefOrEqual(N->getOperand(0), 2) &&
- isUndefOrEqual(N->getOperand(1), 3) &&
- isUndefOrEqual(N->getOperand(2), 2) &&
- isUndefOrEqual(N->getOperand(3), 3);
+ return isUndefOrEqual(N->getMaskElt(0), 6) &&
+ isUndefOrEqual(N->getMaskElt(1), 7) &&
+ isUndefOrEqual(N->getMaskElt(2), 2) &&
+ isUndefOrEqual(N->getMaskElt(3), 3);
}
/// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
-bool X86::isMOVLPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
+bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
+ unsigned NumElems = N->getValueType(0).getVectorNumElements();
- unsigned NumElems = N->getNumOperands();
if (NumElems != 2 && NumElems != 4)
return false;
for (unsigned i = 0; i < NumElems/2; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
+ if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
return false;
for (unsigned i = NumElems/2; i < NumElems; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i))
+ if (!isUndefOrEqual(N->getMaskElt(i), i))
return false;
return true;
/// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
/// and MOVLHPS.
-bool X86::isMOVHPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
+bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
+ unsigned NumElems = N->getValueType(0).getVectorNumElements();
- unsigned NumElems = N->getNumOperands();
if (NumElems != 2 && NumElems != 4)
return false;
for (unsigned i = 0; i < NumElems/2; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i))
+ if (!isUndefOrEqual(N->getMaskElt(i), i))
return false;
- for (unsigned i = 0; i < NumElems/2; ++i) {
- SDValue Arg = N->getOperand(i + NumElems/2);
- if (!isUndefOrEqual(Arg, i + NumElems))
+ for (unsigned i = 0; i < NumElems/2; ++i)
+ if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
return false;
- }
return true;
}
+/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
+/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
+/// <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) &&
+ isUndefOrEqual(N->getMaskElt(1), 3) &&
+ isUndefOrEqual(N->getMaskElt(2), 2) &&
+ isUndefOrEqual(N->getMaskElt(3), 3);
+}
+
/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
-template<class SDOperand>
-bool static isUNPCKLMask(SDOperand *Elts, unsigned NumElts,
+static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
bool V2IsSplat = false) {
+ int NumElts = VT.getVectorNumElements();
if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
return false;
-
- for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
- SDValue BitI = Elts[i];
- SDValue BitI1 = Elts[i+1];
+
+ for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j))
return false;
if (V2IsSplat) {
return false;
}
}
-
return true;
}
-bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
+bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
}
/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKH.
-template<class SDOperand>
-bool static isUNPCKHMask(SDOperand *Elts, unsigned NumElts,
+static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
bool V2IsSplat = false) {
+ int NumElts = VT.getVectorNumElements();
if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
return false;
-
- for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
- SDValue BitI = Elts[i];
- SDValue BitI1 = Elts[i+1];
+
+ for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j + NumElts/2))
return false;
if (V2IsSplat) {
return false;
}
}
-
return true;
}
-bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
+bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
}
/// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
/// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
/// <0, 0, 1, 1>
-bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- unsigned NumElems = N->getNumOperands();
+static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ int NumElems = VT.getVectorNumElements();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
-
- for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
- SDValue BitI = N->getOperand(i);
- SDValue BitI1 = N->getOperand(i+1);
-
+
+ for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j))
return false;
if (!isUndefOrEqual(BitI1, j))
return false;
}
-
return true;
}
+bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
+}
+
/// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
/// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
/// <2, 2, 3, 3>
-bool X86::isUNPCKH_v_undef_Mask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- unsigned NumElems = N->getNumOperands();
+static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ int NumElems = VT.getVectorNumElements();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
-
- for (unsigned i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
- SDValue BitI = N->getOperand(i);
- SDValue BitI1 = N->getOperand(i + 1);
-
+
+ for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j))
return false;
if (!isUndefOrEqual(BitI1, j))
return false;
}
-
return true;
}
+bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
+}
+
/// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVSS,
/// MOVSD, and MOVD, i.e. setting the lowest element.
-template<class SDOperand>
-static bool isMOVLMask(SDOperand *Elts, unsigned NumElts) {
- if (NumElts != 2 && NumElts != 4)
+static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
+ if (VT.getVectorElementType().getSizeInBits() < 32)
return false;
- if (!isUndefOrEqual(Elts[0], NumElts))
+ int NumElts = VT.getVectorNumElements();
+
+ if (!isUndefOrEqual(Mask[0], NumElts))
return false;
-
- for (unsigned i = 1; i < NumElts; ++i) {
- if (!isUndefOrEqual(Elts[i], i))
+
+ for (int i = 1; i < NumElts; ++i)
+ if (!isUndefOrEqual(Mask[i], i))
return false;
- }
-
+
return true;
}
-bool X86::isMOVLMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return ::isMOVLMask(N->op_begin(), N->getNumOperands());
+bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isMOVLMask(M, N->getValueType(0));
}
/// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
/// of what x86 movss want. X86 movs requires the lowest element to be lowest
/// element of vector 2 and the other elements to come from vector 1 in order.
-template<class SDOperand>
-static bool isCommutedMOVL(SDOperand *Ops, unsigned NumOps,
- bool V2IsSplat = false,
- bool V2IsUndef = false) {
+static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
+ bool V2IsSplat = false, bool V2IsUndef = false) {
+ int NumOps = VT.getVectorNumElements();
if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
return false;
-
- if (!isUndefOrEqual(Ops[0], 0))
+
+ if (!isUndefOrEqual(Mask[0], 0))
return false;
-
- for (unsigned i = 1; i < NumOps; ++i) {
- SDValue Arg = Ops[i];
- if (!(isUndefOrEqual(Arg, i+NumOps) ||
- (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
- (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
+
+ 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;
}
-static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
+static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
bool V2IsUndef = false) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
- return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
- V2IsSplat, V2IsUndef);
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
}
/// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
-bool X86::isMOVSHDUPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 4)
+bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
+ if (N->getValueType(0).getVectorNumElements() != 4)
return false;
// Expect 1, 1, 3, 3
for (unsigned i = 0; i < 2; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val != 1) return false;
+ int Elt = N->getMaskElt(i);
+ if (Elt >= 0 && Elt != 1)
+ return false;
}
bool HasHi = false;
for (unsigned i = 2; i < 4; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val != 3) return false;
- HasHi = true;
+ int Elt = N->getMaskElt(i);
+ if (Elt >= 0 && Elt != 3)
+ return false;
+ if (Elt == 3)
+ HasHi = true;
}
-
// Don't use movshdup if it can be done with a shufps.
+ // FIXME: verify that matching u, u, 3, 3 is what we want.
return HasHi;
}
/// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
-bool X86::isMOVSLDUPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 4)
+bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
+ if (N->getValueType(0).getVectorNumElements() != 4)
return false;
// Expect 0, 0, 2, 2
- for (unsigned i = 0; i < 2; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val != 0) return false;
- }
+ for (unsigned i = 0; i < 2; ++i)
+ if (N->getMaskElt(i) > 0)
+ return false;
bool HasHi = false;
for (unsigned i = 2; i < 4; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val != 2) return false;
- HasHi = true;
+ int Elt = N->getMaskElt(i);
+ if (Elt >= 0 && Elt != 2)
+ return false;
+ if (Elt == 2)
+ HasHi = true;
}
-
- // Don't use movshdup if it can be done with a shufps.
+ // Don't use movsldup if it can be done with a shufps.
return HasHi;
}
-/// isIdentityMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a identity operation on the LHS or RHS.
-static bool isIdentityMask(SDNode *N, bool RHS = false) {
- unsigned NumElems = N->getNumOperands();
- for (unsigned i = 0; i < NumElems; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i + (RHS ? NumElems : 0)))
+/// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
+/// 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;
+ for (int i = 0; i < e; ++i)
+ if (!isUndefOrEqual(N->getMaskElt(e+i), i))
return false;
return true;
}
-/// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
-/// a splat of a single element.
-static bool isSplatMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- // This is a splat operation if each element of the permute is the same, and
- // if the value doesn't reference the second vector.
- unsigned NumElems = N->getNumOperands();
- SDValue ElementBase;
- unsigned i = 0;
- for (; i != NumElems; ++i) {
- SDValue Elt = N->getOperand(i);
- if (isa<ConstantSDNode>(Elt)) {
- ElementBase = Elt;
- break;
- }
- }
-
- if (!ElementBase.getNode())
- return false;
+/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
+/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
+/// instructions.
+unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+ int NumOperands = SVOp->getValueType(0).getVectorNumElements();
- for (; i != NumElems; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (Arg != ElementBase) return false;
+ unsigned Shift = (NumOperands == 4) ? 2 : 1;
+ unsigned Mask = 0;
+ for (int i = 0; i < NumOperands; ++i) {
+ int Val = SVOp->getMaskElt(NumOperands-i-1);
+ if (Val < 0) Val = 0;
+ if (Val >= NumOperands) Val -= NumOperands;
+ Mask |= Val;
+ if (i != NumOperands - 1)
+ Mask <<= Shift;
}
-
- // Make sure it is a splat of the first vector operand.
- return cast<ConstantSDNode>(ElementBase)->getZExtValue() < NumElems;
-}
-
-/// getSplatMaskEltNo - Given a splat mask, return the index to the element
-/// we want to splat.
-static SDValue getSplatMaskEltNo(SDNode *N) {
- assert(isSplatMask(N) && "Not a splat mask");
- unsigned NumElems = N->getNumOperands();
- SDValue ElementBase;
- unsigned i = 0;
- for (; i != NumElems; ++i) {
- SDValue Elt = N->getOperand(i);
- if (isa<ConstantSDNode>(Elt))
- return Elt;
- }
- assert(0 && " No splat value found!");
- return SDValue();
-}
-
-
-/// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
-/// a splat of a single element and it's a 2 or 4 element mask.
-bool X86::isSplatMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- // We can only splat 64-bit, and 32-bit quantities with a single instruction.
- if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
- return false;
- return ::isSplatMask(N);
-}
-
-/// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a splat of zero element.
-bool X86::isSplatLoMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
- if (!isUndefOrEqual(N->getOperand(i), 0))
- return false;
- return true;
-}
-
-/// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVDDUP.
-bool X86::isMOVDDUPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- unsigned e = N->getNumOperands() / 2;
- for (unsigned i = 0; i < e; ++i)
- if (!isUndefOrEqual(N->getOperand(i), i))
- return false;
- for (unsigned i = 0; i < e; ++i)
- if (!isUndefOrEqual(N->getOperand(e+i), i))
- return false;
- return true;
-}
-
-/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
-/// instructions.
-unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
- unsigned NumOperands = N->getNumOperands();
- unsigned Shift = (NumOperands == 4) ? 2 : 1;
- unsigned Mask = 0;
- for (unsigned i = 0; i < NumOperands; ++i) {
- unsigned Val = 0;
- SDValue Arg = N->getOperand(NumOperands-i-1);
- if (Arg.getOpcode() != ISD::UNDEF)
- Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val >= NumOperands) Val -= NumOperands;
- Mask |= Val;
- if (i != NumOperands - 1)
- Mask <<= Shift;
- }
-
- return Mask;
+ return Mask;
}
/// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
/// instructions.
unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
// 8 nodes, but we only care about the last 4.
for (unsigned i = 7; i >= 4; --i) {
- unsigned Val = 0;
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() != ISD::UNDEF) {
- Val = cast<ConstantSDNode>(Arg)->getZExtValue();
+ int Val = SVOp->getMaskElt(i);
+ if (Val >= 0)
Mask |= (Val - 4);
- }
if (i != 4)
Mask <<= 2;
}
-
return Mask;
}
/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
/// instructions.
unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
// 8 nodes, but we only care about the first 4.
for (int i = 3; i >= 0; --i) {
- unsigned Val = 0;
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() != ISD::UNDEF)
- Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- Mask |= Val;
+ int Val = SVOp->getMaskElt(i);
+ if (Val >= 0)
+ Mask |= Val;
if (i != 0)
Mask <<= 2;
}
-
return Mask;
}
-/// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
-/// specifies a 8 element shuffle that can be broken into a pair of
-/// PSHUFHW and PSHUFLW.
-static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- if (N->getNumOperands() != 8)
- return false;
-
- // Lower quadword shuffled.
- for (unsigned i = 0; i != 4; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val >= 4)
- return false;
- }
-
- // Upper quadword shuffled.
- for (unsigned i = 4; i != 8; ++i) {
- SDValue Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val < 4 || Val > 7)
- return false;
- }
-
- return true;
-}
-
-/// CommuteVectorShuffle - Swap vector_shuffle operands as well as
-/// values in ther permute mask.
-static SDValue CommuteVectorShuffle(SDValue Op, SDValue &V1,
- SDValue &V2, SDValue &Mask,
- SelectionDAG &DAG) {
- MVT VT = Op.getValueType();
- MVT MaskVT = Mask.getValueType();
- MVT EltVT = MaskVT.getVectorElementType();
- unsigned NumElems = Mask.getNumOperands();
- SmallVector<SDValue, 8> MaskVec;
- DebugLoc dl = Op.getDebugLoc();
-
+/// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
+/// their permute mask.
+static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
+ SelectionDAG &DAG) {
+ MVT VT = SVOp->getValueType(0);
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> MaskVec;
+
for (unsigned i = 0; i != NumElems; ++i) {
- SDValue Arg = Mask.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) {
- MaskVec.push_back(DAG.getUNDEF(EltVT));
- continue;
- }
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val < NumElems)
- MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
+ int idx = SVOp->getMaskElt(i);
+ if (idx < 0)
+ MaskVec.push_back(idx);
+ else if (idx < (int)NumElems)
+ MaskVec.push_back(idx + NumElems);
else
- MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
+ MaskVec.push_back(idx - NumElems);
}
-
- std::swap(V1, V2);
- Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
+ return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
+ SVOp->getOperand(0), &MaskVec[0]);
}
/// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
/// the two vector operands have swapped position.
-static
-SDValue CommuteVectorShuffleMask(SDValue Mask, SelectionDAG &DAG, DebugLoc dl) {
- MVT MaskVT = Mask.getValueType();
- MVT EltVT = MaskVT.getVectorElementType();
- unsigned NumElems = Mask.getNumOperands();
- SmallVector<SDValue, 8> MaskVec;
+static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
+ unsigned NumElems = VT.getVectorNumElements();
for (unsigned i = 0; i != NumElems; ++i) {
- SDValue Arg = Mask.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) {
- MaskVec.push_back(DAG.getUNDEF(EltVT));
+ int idx = Mask[i];
+ if (idx < 0)
continue;
- }
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val < NumElems)
- MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
+ else if (idx < (int)NumElems)
+ Mask[i] = idx + NumElems;
else
- MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
+ Mask[i] = idx - NumElems;
}
- return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], NumElems);
}
-
/// ShouldXformToMOVHLPS - Return true if the node should be transformed to
/// match movhlps. The lower half elements should come from upper half of
/// V1 (and in order), and the upper half elements should come from the upper
/// half of V2 (and in order).
-static bool ShouldXformToMOVHLPS(SDNode *Mask) {
- unsigned NumElems = Mask->getNumOperands();
- if (NumElems != 4)
+static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
+ if (Op->getValueType(0).getVectorNumElements() != 4)
return false;
for (unsigned i = 0, e = 2; i != e; ++i)
- if (!isUndefOrEqual(Mask->getOperand(i), i+2))
+ if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
return false;
for (unsigned i = 2; i != 4; ++i)
- if (!isUndefOrEqual(Mask->getOperand(i), i+4))
+ if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
return false;
return true;
}
/// V1 (and in order), and the upper half elements should come from the upper
/// half of V2 (and in order). And since V1 will become the source of the
/// MOVLP, it must be either a vector load or a scalar load to vector.
-static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
+static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
+ ShuffleVectorSDNode *Op) {
if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
return false;
// Is V2 is a vector load, don't do this transformation. We will try to use
if (ISD::isNON_EXTLoad(V2))
return false;
- unsigned NumElems = Mask->getNumOperands();
+ unsigned NumElems = Op->getValueType(0).getVectorNumElements();
+
if (NumElems != 2 && NumElems != 4)
return false;
for (unsigned i = 0, e = NumElems/2; i != e; ++i)
- if (!isUndefOrEqual(Mask->getOperand(i), i))
+ if (!isUndefOrEqual(Op->getMaskElt(i), i))
return false;
for (unsigned i = NumElems/2; i != NumElems; ++i)
- if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
+ if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
return false;
return true;
}
return true;
}
-/// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
-/// to an undef.
-static bool isUndefShuffle(SDNode *N) {
- if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
- return false;
-
- SDValue V1 = N->getOperand(0);
- SDValue V2 = N->getOperand(1);
- SDValue Mask = N->getOperand(2);
- unsigned NumElems = Mask.getNumOperands();
- for (unsigned i = 0; i != NumElems; ++i) {
- SDValue Arg = Mask.getOperand(i);
- if (Arg.getOpcode() != ISD::UNDEF) {
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
- return false;
- else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
- return false;
- }
- }
- return true;
-}
-
/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
static inline bool isZeroNode(SDValue Elt) {
}
/// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
-/// to an zero vector.
-static bool isZeroShuffle(SDNode *N) {
- if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
- return false;
-
+/// to an zero vector.
+/// FIXME: move to dag combiner / method on ShuffleVectorSDNode
+static bool isZeroShuffle(ShuffleVectorSDNode *N) {
SDValue V1 = N->getOperand(0);
SDValue V2 = N->getOperand(1);
- SDValue Mask = N->getOperand(2);
- unsigned NumElems = Mask.getNumOperands();
+ unsigned NumElems = N->getValueType(0).getVectorNumElements();
for (unsigned i = 0; i != NumElems; ++i) {
- SDValue Arg = Mask.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF)
- continue;
-
- unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Idx < NumElems) {
- unsigned Opc = V1.getNode()->getOpcode();
- if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
+ int Idx = N->getMaskElt(i);
+ if (Idx >= (int)NumElems) {
+ unsigned Opc = V2.getOpcode();
+ if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
continue;
- if (Opc != ISD::BUILD_VECTOR ||
- !isZeroNode(V1.getNode()->getOperand(Idx)))
+ if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
return false;
- } else if (Idx >= NumElems) {
- unsigned Opc = V2.getNode()->getOpcode();
- if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
+ } else if (Idx >= 0) {
+ unsigned Opc = V1.getOpcode();
+ if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
continue;
- if (Opc != ISD::BUILD_VECTOR ||
- !isZeroNode(V2.getNode()->getOperand(Idx - NumElems)))
+ if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
return false;
}
}
static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
DebugLoc dl) {
assert(VT.isVector() && "Expected a vector type");
-
+
// Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
// type. This ensures they get CSE'd.
SDValue Vec;
///
static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
assert(VT.isVector() && "Expected a vector type");
-
+
// Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
// type. This ensures they get CSE'd.
SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
/// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
/// that point to V2 points to its first element.
-static SDValue NormalizeMask(SDValue Mask, SelectionDAG &DAG) {
- assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
-
+static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
+ MVT VT = SVOp->getValueType(0);
+ unsigned NumElems = VT.getVectorNumElements();
+
bool Changed = false;
- SmallVector<SDValue, 8> MaskVec;
- unsigned NumElems = Mask.getNumOperands();
+ SmallVector<int, 8> MaskVec;
+ SVOp->getMask(MaskVec);
+
for (unsigned i = 0; i != NumElems; ++i) {
- SDValue Arg = Mask.getOperand(i);
- if (Arg.getOpcode() != ISD::UNDEF) {
- unsigned Val = cast<ConstantSDNode>(Arg)->getZExtValue();
- if (Val > NumElems) {
- Arg = DAG.getConstant(NumElems, Arg.getValueType());
- Changed = true;
- }
+ if (MaskVec[i] > (int)NumElems) {
+ MaskVec[i] = NumElems;
+ Changed = true;
}
- MaskVec.push_back(Arg);
}
-
if (Changed)
- Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getDebugLoc(),
- Mask.getValueType(),
- &MaskVec[0], MaskVec.size());
- return Mask;
+ return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
+ SVOp->getOperand(1), &MaskVec[0]);
+ return SDValue(SVOp, 0);
}
/// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
/// operation of specified width.
-static SDValue getMOVLMask(unsigned NumElems, SelectionDAG &DAG, DebugLoc dl) {
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT BaseVT = MaskVT.getVectorElementType();
-
- SmallVector<SDValue, 8> MaskVec;
- MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
+static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> Mask;
+ Mask.push_back(NumElems);
for (unsigned i = 1; i != NumElems; ++i)
- MaskVec.push_back(DAG.getConstant(i, BaseVT));
- return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
-}
-
-/// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
-/// of specified width.
-static SDValue getUnpacklMask(unsigned NumElems, SelectionDAG &DAG,
- DebugLoc dl) {
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT BaseVT = MaskVT.getVectorElementType();
- SmallVector<SDValue, 8> MaskVec;
+ Mask.push_back(i);
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
+}
+
+/// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
+static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> Mask;
for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
- MaskVec.push_back(DAG.getConstant(i, BaseVT));
- MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
+ Mask.push_back(i);
+ Mask.push_back(i + NumElems);
}
- return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
}
-/// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
-/// of specified width.
-static SDValue getUnpackhMask(unsigned NumElems, SelectionDAG &DAG,
- DebugLoc dl) {
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT BaseVT = MaskVT.getVectorElementType();
+/// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
+static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
unsigned Half = NumElems/2;
- SmallVector<SDValue, 8> MaskVec;
+ SmallVector<int, 8> Mask;
for (unsigned i = 0; i != Half; ++i) {
- MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
- MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
- }
- return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
-}
-
-/// getSwapEltZeroMask - Returns a vector_shuffle mask for a shuffle that swaps
-/// element #0 of a vector with the specified index, leaving the rest of the
-/// elements in place.
-static SDValue getSwapEltZeroMask(unsigned NumElems, unsigned DestElt,
- SelectionDAG &DAG, DebugLoc dl) {
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT BaseVT = MaskVT.getVectorElementType();
- SmallVector<SDValue, 8> MaskVec;
- // Element #0 of the result gets the elt we are replacing.
- MaskVec.push_back(DAG.getConstant(DestElt, BaseVT));
- for (unsigned i = 1; i != NumElems; ++i)
- MaskVec.push_back(DAG.getConstant(i == DestElt ? 0 : i, BaseVT));
- return DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
+ Mask.push_back(i + Half);
+ Mask.push_back(i + NumElems + Half);
+ }
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
}
/// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
-static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
- MVT PVT = HasSSE2 ? MVT::v4i32 : MVT::v4f32;
- MVT VT = Op.getValueType();
- if (PVT == VT)
- return Op;
- SDValue V1 = Op.getOperand(0);
- SDValue Mask = Op.getOperand(2);
- unsigned MaskNumElems = Mask.getNumOperands();
- unsigned NumElems = MaskNumElems;
- DebugLoc dl = Op.getDebugLoc();
- // Special handling of v4f32 -> v4i32.
- if (VT != MVT::v4f32) {
- // Find which element we want to splat.
- SDNode* EltNoNode = getSplatMaskEltNo(Mask.getNode()).getNode();
- unsigned EltNo = cast<ConstantSDNode>(EltNoNode)->getZExtValue();
- // unpack elements to the correct location
- while (NumElems > 4) {
- if (EltNo < NumElems/2) {
- Mask = getUnpacklMask(MaskNumElems, DAG, dl);
- } else {
- Mask = getUnpackhMask(MaskNumElems, DAG, dl);
- EltNo -= NumElems/2;
- }
- V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1, Mask);
- NumElems >>= 1;
+static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
+ bool HasSSE2) {
+ if (SV->getValueType(0).getVectorNumElements() <= 4)
+ return SDValue(SV, 0);
+
+ MVT PVT = MVT::v4f32;
+ MVT VT = SV->getValueType(0);
+ DebugLoc dl = SV->getDebugLoc();
+ SDValue V1 = SV->getOperand(0);
+ int NumElems = VT.getVectorNumElements();
+ int EltNo = SV->getSplatIndex();
+
+ // unpack elements to the correct location
+ while (NumElems > 4) {
+ if (EltNo < NumElems/2) {
+ V1 = getUnpackl(DAG, dl, VT, V1, V1);
+ } else {
+ V1 = getUnpackh(DAG, dl, VT, V1, V1);
+ EltNo -= NumElems/2;
}
- SDValue Cst = DAG.getConstant(EltNo, MVT::i32);
- Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
- }
-
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
- SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
- DAG.getUNDEF(PVT), Mask);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
-}
-
-/// isVectorLoad - Returns true if the node is a vector load, a scalar
-/// load that's promoted to vector, or a load bitcasted.
-static bool isVectorLoad(SDValue Op) {
- assert(Op.getValueType().isVector() && "Expected a vector type");
- if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
- Op.getOpcode() == ISD::BIT_CONVERT) {
- return isa<LoadSDNode>(Op.getOperand(0));
- }
- return isa<LoadSDNode>(Op);
-}
-
-
-/// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
-///
-static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
- SelectionDAG &DAG, bool HasSSE3) {
- // If we have sse3 and shuffle has more than one use or input is a load, then
- // use movddup. Otherwise, use movlhps.
- bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
- MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
- MVT VT = Op.getValueType();
- if (VT == PVT)
- return Op;
- DebugLoc dl = Op.getDebugLoc();
- unsigned NumElems = PVT.getVectorNumElements();
- if (NumElems == 2) {
- SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
- Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
- } else {
- assert(NumElems == 4);
- SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
- SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
- Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
- Cst0, Cst1, Cst0, Cst1);
+ NumElems >>= 1;
}
-
+
+ // Perform the splat.
+ int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
- SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, PVT, V1,
- DAG.getUNDEF(PVT), Mask);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuffle);
+ V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
}
/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
bool isZero, bool HasSSE2,
SelectionDAG &DAG) {
- DebugLoc dl = V2.getDebugLoc();
MVT VT = V2.getValueType();
SDValue V1 = isZero
- ? getZeroVector(VT, HasSSE2, DAG, dl) : DAG.getUNDEF(VT);
- unsigned NumElems = V2.getValueType().getVectorNumElements();
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT EVT = MaskVT.getVectorElementType();
- SmallVector<SDValue, 16> MaskVec;
+ ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 16> MaskVec;
for (unsigned i = 0; i != NumElems; ++i)
- if (i == Idx) // If this is the insertion idx, put the low elt of V2 here.
- MaskVec.push_back(DAG.getConstant(NumElems, EVT));
- else
- MaskVec.push_back(DAG.getConstant(i, EVT));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, Mask);
+ // If this is the insertion idx, put the low elt of V2 here.
+ MaskVec.push_back(i == Idx ? NumElems : i);
+ return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
}
/// getNumOfConsecutiveZeros - Return the number of elements in a result of
/// a shuffle that is zero.
static
-unsigned getNumOfConsecutiveZeros(SDValue Op, SDValue Mask,
- unsigned NumElems, bool Low,
- SelectionDAG &DAG) {
+unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
+ bool Low, SelectionDAG &DAG) {
unsigned NumZeros = 0;
- for (unsigned i = 0; i < NumElems; ++i) {
+ for (int i = 0; i < NumElems; ++i) {
unsigned Index = Low ? i : NumElems-i-1;
- SDValue Idx = Mask.getOperand(Index);
- if (Idx.getOpcode() == ISD::UNDEF) {
+ int Idx = SVOp->getMaskElt(Index);
+ if (Idx < 0) {
++NumZeros;
continue;
}
- SDValue Elt = DAG.getShuffleScalarElt(Op.getNode(), Index);
+ SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
if (Elt.getNode() && isZeroNode(Elt))
++NumZeros;
else
/// isVectorShift - Returns true if the shuffle can be implemented as a
/// logical left or right shift of a vector.
-static bool isVectorShift(SDValue Op, SDValue Mask, SelectionDAG &DAG,
+/// FIXME: split into pslldqi, psrldqi, palignr variants.
+static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
- unsigned NumElems = Mask.getNumOperands();
+ int NumElems = SVOp->getValueType(0).getVectorNumElements();
isLeft = true;
- unsigned NumZeros= getNumOfConsecutiveZeros(Op, Mask, NumElems, true, DAG);
+ unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
if (!NumZeros) {
isLeft = false;
- NumZeros = getNumOfConsecutiveZeros(Op, Mask, NumElems, false, DAG);
+ NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
if (!NumZeros)
return false;
}
-
bool SeenV1 = false;
bool SeenV2 = false;
- for (unsigned i = NumZeros; i < NumElems; ++i) {
- unsigned Val = isLeft ? (i - NumZeros) : i;
- SDValue Idx = Mask.getOperand(isLeft ? i : (i - NumZeros));
- if (Idx.getOpcode() == ISD::UNDEF)
+ for (int i = NumZeros; i < NumElems; ++i) {
+ int Val = isLeft ? (i - NumZeros) : i;
+ int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
+ if (Idx < 0)
continue;
- unsigned Index = cast<ConstantSDNode>(Idx)->getZExtValue();
- if (Index < NumElems)
+ if (Idx < NumElems)
SeenV1 = true;
else {
- Index -= NumElems;
+ Idx -= NumElems;
SeenV2 = true;
}
- if (Index != Val)
+ if (Idx != Val)
return false;
}
if (SeenV1 && SeenV2)
return false;
- ShVal = SeenV1 ? Op.getOperand(0) : Op.getOperand(1);
+ ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
ShAmt = NumZeros;
return true;
}
SDValue ThisElt(0, 0), LastElt(0, 0);
bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
if (LastIsNonZero) {
- LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
+ LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
MVT::i16, Op.getOperand(i-1));
}
if (ThisIsNonZero) {
V = DAG.getUNDEF(MVT::v8i16);
First = false;
}
- V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
+ V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
MVT::v8i16, V, Op.getOperand(i),
DAG.getIntPtrConstant(i));
}
/// getVShift - Return a vector logical shift node.
///
static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
- unsigned NumBits, SelectionDAG &DAG,
- const TargetLowering &TLI, DebugLoc dl) {
+ unsigned NumBits, SelectionDAG &DAG,
+ const TargetLowering &TLI, DebugLoc dl) {
bool isMMX = VT.getSizeInBits() == 64;
MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
}
// Special case for single non-zero, non-undef, element.
- if (NumNonZero == 1 && NumElems <= 4) {
+ if (NumNonZero == 1) {
unsigned Idx = CountTrailingZeros_32(NonZeros);
SDValue Item = Op.getOperand(Idx);
-
+
// If this is an insertion of an i64 value on x86-32, and if the top bits of
// the value are obviously zero, truncate the value to i32 and do the
// insertion that way. Only do this if the value is non-constant or if the
// Handle MMX and SSE both.
MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
-
+
// Truncate the value (which may itself be a constant) to i32, and
// convert it to a vector with movd (S2V+shuffle to zero extend).
Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
Item = getShuffleVectorZeroOrUndef(Item, 0, true,
Subtarget->hasSSE2(), DAG);
-
+
// Now we have our 32-bit value zero extended in the low element of
// a vector. If Idx != 0, swizzle it into place.
if (Idx != 0) {
- SDValue Ops[] = {
- Item, DAG.getUNDEF(Item.getValueType()),
- getSwapEltZeroMask(VecElts, Idx, DAG, dl)
- };
- Item = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VecVT, Ops, 3);
+ SmallVector<int, 4> Mask;
+ Mask.push_back(Idx);
+ for (unsigned i = 1; i != VecElts; ++i)
+ Mask.push_back(i);
+ Item = DAG.getVectorShuffle(VecVT, dl, Item,
+ DAG.getUNDEF(Item.getValueType()),
+ &Mask[0]);
}
return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
}
}
-
+
// If we have a constant or non-constant insertion into the low element of
// a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
// the rest of the elements. This will be matched as movd/movq/movss/movsd
- // depending on what the source datatype is. Because we can only get here
- // when NumElems <= 4, this only needs to handle i32/f32/i64/f64.
- if (Idx == 0 &&
- // Don't do this for i64 values on x86-32.
- (EVT != MVT::i64 || Subtarget->is64Bit())) {
- Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
- // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
- return getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
- Subtarget->hasSSE2(), DAG);
+ // depending on what the source datatype is.
+ if (Idx == 0) {
+ if (NumZero == 0) {
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
+ (EVT == MVT::i64 && Subtarget->is64Bit())) {
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
+ return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
+ DAG);
+ } else if (EVT == MVT::i16 || EVT == MVT::i8) {
+ Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
+ MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
+ Item = getShuffleVectorZeroOrUndef(Item, 0, true,
+ Subtarget->hasSSE2(), DAG);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
+ }
}
// Is it a vector logical left shift?
isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
unsigned NumBits = VT.getSizeInBits();
return getVShift(true, VT,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
+ DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
VT, Op.getOperand(1)),
NumBits/2, DAG, *this, dl);
}
-
+
if (IsAllConstants) // Otherwise, it's better to do a constpool load.
return SDValue();
// place.
if (EVTBits == 32) {
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
-
+
// Turn it into a shuffle of zero and zero-extended scalar to vector.
Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
Subtarget->hasSSE2(), DAG);
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT MaskEVT = MaskVT.getVectorElementType();
- SmallVector<SDValue, 8> MaskVec;
+ SmallVector<int, 8> MaskVec;
for (unsigned i = 0; i < NumElems; i++)
- MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, Item,
- DAG.getUNDEF(VT), Mask);
+ MaskVec.push_back(i == Idx ? 0 : 1);
+ return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
}
}
// Splat is obviously ok. Let legalizer expand it to a shuffle.
if (Values.size() == 1)
return SDValue();
-
+
// A vector full of immediates; various special cases are already
// handled, so this is best done with a single constant-pool load.
if (IsAllConstants)
V[i] = V[i*2]; // Must be a zero vector.
break;
case 1:
- V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2+1], V[i*2],
- getMOVLMask(NumElems, DAG, dl));
+ V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
break;
case 2:
- V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
- getMOVLMask(NumElems, DAG, dl));
+ V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
break;
case 3:
- V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i*2], V[i*2+1],
- getUnpacklMask(NumElems, DAG, dl));
+ V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
break;
}
}
- MVT MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT EVT = MaskVT.getVectorElementType();
- SmallVector<SDValue, 8> MaskVec;
+ SmallVector<int, 8> MaskVec;
bool Reverse = (NonZeros & 0x3) == 2;
for (unsigned i = 0; i < 2; ++i)
- if (Reverse)
- MaskVec.push_back(DAG.getConstant(1-i, EVT));
- else
- MaskVec.push_back(DAG.getConstant(i, EVT));
+ MaskVec.push_back(Reverse ? 1-i : i);
Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
for (unsigned i = 0; i < 2; ++i)
- if (Reverse)
- MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
- else
- MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
- SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size());
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[0], V[1], ShufMask);
+ MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
+ return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
}
if (Values.size() > 2) {
+ // 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.
+ // FIXME: We could probably just check that here directly.
+ if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
+ getSubtarget()->hasSSE41()) {
+ V[0] = DAG.getUNDEF(VT);
+ for (unsigned i = 0; i < NumElems; ++i)
+ if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
+ V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
+ Op.getOperand(i), DAG.getIntPtrConstant(i));
+ return V[0];
+ }
// Expand into a number of unpckl*.
// e.g. for v4f32
// Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
// : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
// Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
- SDValue UnpckMask = getUnpacklMask(NumElems, DAG, dl);
for (unsigned i = 0; i < NumElems; ++i)
V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
NumElems >>= 1;
while (NumElems != 0) {
for (unsigned i = 0; i < NumElems; ++i)
- V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V[i], V[i + NumElems],
- UnpckMask);
+ V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
NumElems >>= 1;
}
return V[0];
return SDValue();
}
+// v8i16 shuffles - Prefer shuffles in the following order:
+// 1. [all] pshuflw, pshufhw, optional move
+// 2. [ssse3] 1 x pshufb
+// 3. [ssse3] 2 x pshufb + 1 x por
+// 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
static
-SDValue LowerVECTOR_SHUFFLEv8i16(SDValue V1, SDValue V2,
- SDValue PermMask, SelectionDAG &DAG,
- TargetLowering &TLI, DebugLoc dl) {
- SDValue NewV;
- MVT MaskVT = MVT::getIntVectorWithNumElements(8);
- MVT MaskEVT = MaskVT.getVectorElementType();
- MVT PtrVT = TLI.getPointerTy();
- SmallVector<SDValue, 8> MaskElts(PermMask.getNode()->op_begin(),
- PermMask.getNode()->op_end());
-
- // First record which half of which vector the low elements come from.
- SmallVector<unsigned, 4> LowQuad(4);
- for (unsigned i = 0; i < 4; ++i) {
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF)
+SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
+ SelectionDAG &DAG, X86TargetLowering &TLI) {
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ DebugLoc dl = SVOp->getDebugLoc();
+ SmallVector<int, 8> MaskVals;
+
+ // Determine if more than 1 of the words in each of the low and high quadwords
+ // of the result come from the same quadword of one of the two inputs. Undef
+ // mask values count as coming from any quadword, for better codegen.
+ SmallVector<unsigned, 4> LoQuad(4);
+ SmallVector<unsigned, 4> HiQuad(4);
+ BitVector InputQuads(4);
+ for (unsigned i = 0; i < 8; ++i) {
+ SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
+ int EltIdx = SVOp->getMaskElt(i);
+ MaskVals.push_back(EltIdx);
+ if (EltIdx < 0) {
+ ++Quad[0];
+ ++Quad[1];
+ ++Quad[2];
+ ++Quad[3];
continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- int QuadIdx = EltIdx / 4;
- ++LowQuad[QuadIdx];
+ }
+ ++Quad[EltIdx / 4];
+ InputQuads.set(EltIdx / 4);
}
- int BestLowQuad = -1;
+ int BestLoQuad = -1;
unsigned MaxQuad = 1;
for (unsigned i = 0; i < 4; ++i) {
- if (LowQuad[i] > MaxQuad) {
- BestLowQuad = i;
- MaxQuad = LowQuad[i];
+ if (LoQuad[i] > MaxQuad) {
+ BestLoQuad = i;
+ MaxQuad = LoQuad[i];
}
}
- // Record which half of which vector the high elements come from.
- SmallVector<unsigned, 4> HighQuad(4);
- for (unsigned i = 4; i < 8; ++i) {
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF)
- continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- int QuadIdx = EltIdx / 4;
- ++HighQuad[QuadIdx];
- }
-
- int BestHighQuad = -1;
+ int BestHiQuad = -1;
MaxQuad = 1;
for (unsigned i = 0; i < 4; ++i) {
- if (HighQuad[i] > MaxQuad) {
- BestHighQuad = i;
- MaxQuad = HighQuad[i];
+ if (HiQuad[i] > MaxQuad) {
+ BestHiQuad = i;
+ MaxQuad = HiQuad[i];
}
}
- // If it's possible to sort parts of either half with PSHUF{H|L}W, then do it.
- if (BestLowQuad != -1 || BestHighQuad != -1) {
- // First sort the 4 chunks in order using shufpd.
- SmallVector<SDValue, 8> MaskVec;
-
- if (BestLowQuad != -1)
- MaskVec.push_back(DAG.getConstant(BestLowQuad, MVT::i32));
- else
- MaskVec.push_back(DAG.getConstant(0, MVT::i32));
-
- if (BestHighQuad != -1)
- MaskVec.push_back(DAG.getConstant(BestHighQuad, MVT::i32));
- else
- MaskVec.push_back(DAG.getConstant(1, MVT::i32));
+ // 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
+ // 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];
+ bool V2Used = InputQuads[2] || InputQuads[3];
+ if (TLI.getSubtarget()->hasSSSE3()) {
+ if (InputQuads.count() == 2 && V1Used && V2Used) {
+ BestLoQuad = InputQuads.find_first();
+ BestHiQuad = InputQuads.find_next(BestLoQuad);
+ }
+ if (InputQuads.count() > 2) {
+ BestLoQuad = -1;
+ BestHiQuad = -1;
+ }
+ }
- SDValue Mask= DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, &MaskVec[0],2);
- NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2i64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), Mask);
+ // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
+ // the shuffle mask. If a quad is scored as -1, that means that it contains
+ // words from all 4 input quadwords.
+ SDValue NewV;
+ if (BestLoQuad >= 0 || BestHiQuad >= 0) {
+ SmallVector<int, 8> MaskV;
+ MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
+ MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
+ 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);
- // Now sort high and low parts separately.
- BitVector InOrder(8);
- if (BestLowQuad != -1) {
- // Sort lower half in order using PSHUFLW.
- MaskVec.clear();
- bool AnyOutOrder = false;
-
- for (unsigned i = 0; i != 4; ++i) {
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF) {
- MaskVec.push_back(Elt);
- InOrder.set(i);
- } else {
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (EltIdx != i)
- AnyOutOrder = true;
-
- MaskVec.push_back(DAG.getConstant(EltIdx % 4, MaskEVT));
-
- // If this element is in the right place after this shuffle, then
- // remember it.
- if ((int)(EltIdx / 4) == BestLowQuad)
- InOrder.set(i);
- }
- }
- if (AnyOutOrder) {
- for (unsigned i = 4; i != 8; ++i)
- MaskVec.push_back(DAG.getConstant(i, MaskEVT));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], 8);
- NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16,
- NewV, NewV, Mask);
- }
+ // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
+ // source words for the shuffle, to aid later transformations.
+ bool AllWordsInNewV = true;
+ bool InOrder[2] = { true, true };
+ for (unsigned i = 0; i != 8; ++i) {
+ int idx = MaskVals[i];
+ if (idx != (int)i)
+ InOrder[i/4] = false;
+ if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
+ continue;
+ AllWordsInNewV = false;
+ break;
}
- if (BestHighQuad != -1) {
- // Sort high half in order using PSHUFHW if possible.
- MaskVec.clear();
-
- for (unsigned i = 0; i != 4; ++i)
- MaskVec.push_back(DAG.getConstant(i, MaskEVT));
-
- bool AnyOutOrder = false;
- for (unsigned i = 4; i != 8; ++i) {
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF) {
- MaskVec.push_back(Elt);
- InOrder.set(i);
- } else {
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (EltIdx != i)
- AnyOutOrder = true;
-
- MaskVec.push_back(DAG.getConstant((EltIdx % 4) + 4, MaskEVT));
-
- // If this element is in the right place after this shuffle, then
- // remember it.
- if ((int)(EltIdx / 4) == BestHighQuad)
- InOrder.set(i);
- }
- }
-
- if (AnyOutOrder) {
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl,
- MaskVT, &MaskVec[0], 8);
- NewV = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16,
- NewV, NewV, Mask);
+ bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
+ if (AllWordsInNewV) {
+ for (int i = 0; i != 8; ++i) {
+ int idx = MaskVals[i];
+ if (idx < 0)
+ continue;
+ idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
+ if ((idx != i) && idx < 4)
+ pshufhw = false;
+ if ((idx != i) && idx > 3)
+ pshuflw = false;
}
+ V1 = NewV;
+ V2Used = false;
+ BestLoQuad = 0;
+ BestHiQuad = 1;
}
- // The other elements are put in the right place using pextrw and pinsrw.
+ // 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,
+ 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
+ // 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;
for (unsigned i = 0; i != 8; ++i) {
- if (InOrder[i])
+ int EltIdx = MaskVals[i] * 2;
+ if (TwoInputs && (EltIdx >= 16)) {
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
continue;
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF)
+ }
+ pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
+ 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,
+ 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();
+ for (unsigned i = 0; i != 8; ++i) {
+ int EltIdx = MaskVals[i] * 2;
+ if (EltIdx < 16) {
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- SDValue ExtOp = (EltIdx < 8)
- ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
- DAG.getConstant(EltIdx, PtrVT))
- : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
- DAG.getConstant(EltIdx - 8, PtrVT));
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
- DAG.getConstant(i, PtrVT));
- }
-
- return NewV;
- }
-
- // PSHUF{H|L}W are not used. Lower into extracts and inserts but try to use as
- // few as possible. First, let's find out how many elements are already in the
- // right order.
- unsigned V1InOrder = 0;
- unsigned V1FromV1 = 0;
- unsigned V2InOrder = 0;
- unsigned V2FromV2 = 0;
- SmallVector<SDValue, 8> V1Elts;
- SmallVector<SDValue, 8> V2Elts;
- for (unsigned i = 0; i < 8; ++i) {
- SDValue Elt = MaskElts[i];
- if (Elt.getOpcode() == ISD::UNDEF) {
- V1Elts.push_back(Elt);
- V2Elts.push_back(Elt);
- ++V1InOrder;
- ++V2InOrder;
- continue;
+ }
+ pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
}
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (EltIdx == i) {
- V1Elts.push_back(Elt);
- V2Elts.push_back(DAG.getConstant(i+8, MaskEVT));
- ++V1InOrder;
- } else if (EltIdx == i+8) {
- V1Elts.push_back(Elt);
- V2Elts.push_back(DAG.getConstant(i, MaskEVT));
- ++V2InOrder;
- } else if (EltIdx < 8) {
- V1Elts.push_back(Elt);
- V2Elts.push_back(DAG.getConstant(EltIdx+8, MaskEVT));
- ++V1FromV1;
- } else {
- V1Elts.push_back(Elt);
- V2Elts.push_back(DAG.getConstant(EltIdx-8, MaskEVT));
- ++V2FromV2;
+ V2 = DAG.getNode(ISD::BIT_CONVERT, 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);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
+ }
+
+ // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
+ // and update MaskVals with new element order.
+ BitVector InOrder(8);
+ if (BestLoQuad >= 0) {
+ SmallVector<int, 8> MaskV;
+ for (int i = 0; i != 4; ++i) {
+ int idx = MaskVals[i];
+ if (idx < 0) {
+ MaskV.push_back(-1);
+ InOrder.set(i);
+ } else if ((idx / 4) == BestLoQuad) {
+ MaskV.push_back(idx & 3);
+ InOrder.set(i);
+ } else {
+ MaskV.push_back(-1);
+ }
}
+ for (unsigned i = 4; i != 8; ++i)
+ MaskV.push_back(i);
+ NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
+ &MaskV[0]);
}
-
- if (V2InOrder > V1InOrder) {
- PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
- std::swap(V1, V2);
- std::swap(V1Elts, V2Elts);
- std::swap(V1FromV1, V2FromV2);
- }
-
- if ((V1FromV1 + V1InOrder) != 8) {
- // Some elements are from V2.
- if (V1FromV1) {
- // If there are elements that are from V1 but out of place,
- // then first sort them in place
- SmallVector<SDValue, 8> MaskVec;
- for (unsigned i = 0; i < 8; ++i) {
- SDValue Elt = V1Elts[i];
- if (Elt.getOpcode() == ISD::UNDEF) {
- MaskVec.push_back(DAG.getUNDEF(MaskEVT));
- continue;
- }
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (EltIdx >= 8)
- MaskVec.push_back(DAG.getUNDEF(MaskEVT));
- else
- MaskVec.push_back(DAG.getConstant(EltIdx, MaskEVT));
+
+ // If BestHi >= 0, generate a pshufhw to put the high elements in order,
+ // and update MaskVals with the new element order.
+ if (BestHiQuad >= 0) {
+ SmallVector<int, 8> MaskV;
+ for (unsigned i = 0; i != 4; ++i)
+ MaskV.push_back(i);
+ for (unsigned i = 4; i != 8; ++i) {
+ int idx = MaskVals[i];
+ if (idx < 0) {
+ MaskV.push_back(-1);
+ InOrder.set(i);
+ } else if ((idx / 4) == BestHiQuad) {
+ MaskV.push_back((idx & 3) + 4);
+ InOrder.set(i);
+ } else {
+ MaskV.push_back(-1);
}
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &MaskVec[0], 8);
- V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v8i16, V1, V1, Mask);
}
-
+ 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 (BestLoQuad == -1 && BestHiQuad == -1) {
NewV = V1;
- for (unsigned i = 0; i < 8; ++i) {
- SDValue Elt = V1Elts[i];
- if (Elt.getOpcode() == ISD::UNDEF)
- continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (EltIdx < 8)
+ for (int i = 0; i != 8; ++i)
+ 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])
+ continue;
+ int EltIdx = MaskVals[i];
+ if (EltIdx < 0)
+ continue;
+ SDValue ExtOp = (EltIdx < 8)
+ ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
+ DAG.getIntPtrConstant(EltIdx))
+ : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
+ DAG.getIntPtrConstant(EltIdx - 8));
+ NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
+ DAG.getIntPtrConstant(i));
+ }
+ return NewV;
+}
+
+// v16i8 shuffles - Prefer shuffles in the following order:
+// 1. [ssse3] 1 x pshufb
+// 2. [ssse3] 2 x pshufb + 1 x por
+// 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
+static
+SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
+ SelectionDAG &DAG, X86TargetLowering &TLI) {
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ 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
+ // present, fall back to case 3.
+ // FIXME: kill V2Only once shuffles are canonizalized by getNode.
+ bool V1Only = true;
+ bool V2Only = true;
+ for (unsigned i = 0; i < 16; ++i) {
+ int EltIdx = MaskVals[i];
+ if (EltIdx < 0)
+ continue;
+ if (EltIdx < 16)
+ V2Only = false;
+ 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.
+ //
+ // 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
+ // so that we can OR them together.
+ bool TwoInputs = !(V1Only || V2Only);
+ for (unsigned i = 0; i != 16; ++i) {
+ int EltIdx = MaskVals[i];
+ if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
continue;
- SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
- DAG.getConstant(EltIdx - 8, PtrVT));
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
- DAG.getConstant(i, PtrVT));
+ }
+ pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
}
- return NewV;
- } else {
- // All elements are from V1.
- NewV = V1;
- for (unsigned i = 0; i < 8; ++i) {
- SDValue Elt = V1Elts[i];
- if (Elt.getOpcode() == ISD::UNDEF)
+ // If all the elements are from V2, assign it to V1 and return after
+ // building the first pshufb.
+ if (V2Only)
+ V1 = V2;
+ V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, dl,
+ 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();
+ for (unsigned i = 0; i != 16; ++i) {
+ int EltIdx = MaskVals[i];
+ if (EltIdx < 16) {
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- SDValue ExtOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
- DAG.getConstant(EltIdx, PtrVT));
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
- DAG.getConstant(i, PtrVT));
+ }
+ pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
}
- return NewV;
+ V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, dl,
+ 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.
+ V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
+ V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
+ SDValue NewV = V2Only ? V2 : V1;
+ 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;
+
+ // If Elt0 and Elt1 are defined, are consecutive, and can be load
+ // using a single extract together, load it and store it.
+ if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
+ InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
+ DAG.getIntPtrConstant(Elt1 / 2));
+ NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
+ DAG.getIntPtrConstant(i));
+ continue;
+ }
+
+ // If Elt1 is defined, extract it from the appropriate source. If the
+ // source byte is not also odd, shift the extracted word left 8 bits
+ // otherwise clear the bottom 8 bits if we need to do an or.
+ if (Elt1 >= 0) {
+ InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
+ DAG.getIntPtrConstant(Elt1 / 2));
+ if ((Elt1 & 1) == 0)
+ InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
+ DAG.getConstant(8, TLI.getShiftAmountTy()));
+ else if (Elt0 >= 0)
+ InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
+ DAG.getConstant(0xFF00, MVT::i16));
+ }
+ // If Elt0 is defined, extract it from the appropriate source. If the
+ // source byte is not also even, shift the extracted word right 8 bits. If
+ // Elt1 was also defined, OR the extracted values together before
+ // inserting them in the result.
+ if (Elt0 >= 0) {
+ SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
+ Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
+ if ((Elt0 & 1) != 0)
+ InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
+ DAG.getConstant(8, TLI.getShiftAmountTy()));
+ else if (Elt1 >= 0)
+ InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
+ DAG.getConstant(0x00FF, MVT::i16));
+ InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
+ : InsElt0;
+ }
+ NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
+ DAG.getIntPtrConstant(i));
+ }
+ return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
}
/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
/// the right sequence. e.g.
/// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
static
-SDValue RewriteAsNarrowerShuffle(SDValue V1, SDValue V2,
- MVT VT,
- SDValue PermMask, SelectionDAG &DAG,
- TargetLowering &TLI, DebugLoc dl) {
- unsigned NumElems = PermMask.getNumOperands();
+SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
+ SelectionDAG &DAG,
+ TargetLowering &TLI, DebugLoc dl) {
+ MVT VT = SVOp->getValueType(0);
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ unsigned NumElems = VT.getVectorNumElements();
unsigned NewWidth = (NumElems == 4) ? 2 : 4;
MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
MVT MaskEltVT = MaskVT.getVectorElementType();
else
NewVT = MVT::v2f64;
}
- unsigned Scale = NumElems / NewWidth;
- SmallVector<SDValue, 8> MaskVec;
+ int Scale = NumElems / NewWidth;
+ SmallVector<int, 8> MaskVec;
for (unsigned i = 0; i < NumElems; i += Scale) {
- unsigned StartIdx = ~0U;
- for (unsigned j = 0; j < Scale; ++j) {
- SDValue Elt = PermMask.getOperand(i+j);
- if (Elt.getOpcode() == ISD::UNDEF)
+ int StartIdx = -1;
+ for (int j = 0; j < Scale; ++j) {
+ int EltIdx = SVOp->getMaskElt(i+j);
+ if (EltIdx < 0)
continue;
- unsigned EltIdx = cast<ConstantSDNode>(Elt)->getZExtValue();
- if (StartIdx == ~0U)
+ if (StartIdx == -1)
StartIdx = EltIdx - (EltIdx % Scale);
if (EltIdx != StartIdx + j)
return SDValue();
}
- if (StartIdx == ~0U)
- MaskVec.push_back(DAG.getUNDEF(MaskEltVT));
+ if (StartIdx == -1)
+ MaskVec.push_back(-1);
else
- MaskVec.push_back(DAG.getConstant(StartIdx / Scale, MaskEltVT));
+ MaskVec.push_back(StartIdx / Scale);
}
V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, NewVT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskVec[0], MaskVec.size()));
+ return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
}
/// getVZextMovL - Return a zero-extending vector move low node.
///
static SDValue getVZextMovL(MVT VT, MVT OpVT,
- SDValue SrcOp, SelectionDAG &DAG,
- const X86Subtarget *Subtarget, DebugLoc dl) {
+ SDValue SrcOp, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget, DebugLoc dl) {
if (VT == MVT::v2f64 || VT == MVT::v4f32) {
LoadSDNode *LD = NULL;
if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
- DAG.getNode(ISD::BIT_CONVERT, dl,
+ DAG.getNode(ISD::BIT_CONVERT, dl,
OpVT, SrcOp)));
}
/// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
/// shuffles.
static SDValue
-LowerVECTOR_SHUFFLE_4wide(SDValue V1, SDValue V2,
- SDValue PermMask, MVT VT, SelectionDAG &DAG,
- DebugLoc dl) {
- MVT MaskVT = PermMask.getValueType();
- MVT MaskEVT = MaskVT.getVectorElementType();
+LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ DebugLoc dl = SVOp->getDebugLoc();
+ MVT VT = SVOp->getValueType(0);
+
SmallVector<std::pair<int, int>, 8> Locs;
Locs.resize(4);
- SmallVector<SDValue, 8> Mask1(4, DAG.getUNDEF(MaskEVT));
+ SmallVector<int, 8> Mask1(4U, -1);
+ SmallVector<int, 8> PermMask;
+ SVOp->getMask(PermMask);
+
unsigned NumHi = 0;
unsigned NumLo = 0;
for (unsigned i = 0; i != 4; ++i) {
- SDValue Elt = PermMask.getOperand(i);
- if (Elt.getOpcode() == ISD::UNDEF) {
+ int Idx = PermMask[i];
+ if (Idx < 0) {
Locs[i] = std::make_pair(-1, -1);
} else {
- unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
- assert(Val < 8 && "Invalid VECTOR_SHUFFLE index!");
- if (Val < 4) {
+ assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
+ if (Idx < 4) {
Locs[i] = std::make_pair(0, NumLo);
- Mask1[NumLo] = Elt;
+ Mask1[NumLo] = Idx;
NumLo++;
} else {
Locs[i] = std::make_pair(1, NumHi);
if (2+NumHi < 4)
- Mask1[2+NumHi] = Elt;
+ Mask1[2+NumHi] = Idx;
NumHi++;
}
}
// implemented with two shuffles. First shuffle gather the elements.
// The second shuffle, which takes the first shuffle as both of its
// vector operands, put the elements into the right order.
- V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &Mask1[0], Mask1.size()));
+ V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
- SmallVector<SDValue, 8> Mask2(4, DAG.getUNDEF(MaskEVT));
+ SmallVector<int, 8> Mask2(4U, -1);
+
for (unsigned i = 0; i != 4; ++i) {
if (Locs[i].first == -1)
continue;
else {
unsigned Idx = (i < 2) ? 0 : 4;
Idx += Locs[i].first * 2 + Locs[i].second;
- Mask2[i] = DAG.getConstant(Idx, MaskEVT);
+ Mask2[i] = Idx;
}
}
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &Mask2[0], Mask2.size()));
+ return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
} else if (NumLo == 3 || NumHi == 3) {
// Otherwise, we must have three elements from one vector, call it X, and
// one element from the other, call it Y. First, use a shufps to build an
// from X.
if (NumHi == 3) {
// Normalize it so the 3 elements come from V1.
- PermMask = CommuteVectorShuffleMask(PermMask, DAG, dl);
+ CommuteVectorShuffleMask(PermMask, VT);
std::swap(V1, V2);
}
// Find the element from V2.
unsigned HiIndex;
for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
- SDValue Elt = PermMask.getOperand(HiIndex);
- if (Elt.getOpcode() == ISD::UNDEF)
+ int Val = PermMask[HiIndex];
+ if (Val < 0)
continue;
- unsigned Val = cast<ConstantSDNode>(Elt)->getZExtValue();
if (Val >= 4)
break;
}
- Mask1[0] = PermMask.getOperand(HiIndex);
- Mask1[1] = DAG.getUNDEF(MaskEVT);
- Mask1[2] = PermMask.getOperand(HiIndex^1);
- Mask1[3] = DAG.getUNDEF(MaskEVT);
- V2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT, &Mask1[0], 4));
+ Mask1[0] = PermMask[HiIndex];
+ Mask1[1] = -1;
+ Mask1[2] = PermMask[HiIndex^1];
+ Mask1[3] = -1;
+ V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
if (HiIndex >= 2) {
- Mask1[0] = PermMask.getOperand(0);
- Mask1[1] = PermMask.getOperand(1);
- Mask1[2] = DAG.getConstant(HiIndex & 1 ? 6 : 4, MaskEVT);
- Mask1[3] = DAG.getConstant(HiIndex & 1 ? 4 : 6, MaskEVT);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MaskVT, &Mask1[0], 4));
+ Mask1[0] = PermMask[0];
+ Mask1[1] = PermMask[1];
+ Mask1[2] = HiIndex & 1 ? 6 : 4;
+ Mask1[3] = HiIndex & 1 ? 4 : 6;
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
} else {
- Mask1[0] = DAG.getConstant(HiIndex & 1 ? 2 : 0, MaskEVT);
- Mask1[1] = DAG.getConstant(HiIndex & 1 ? 0 : 2, MaskEVT);
- Mask1[2] = PermMask.getOperand(2);
- Mask1[3] = PermMask.getOperand(3);
- if (Mask1[2].getOpcode() != ISD::UNDEF)
- Mask1[2] =
- DAG.getConstant(cast<ConstantSDNode>(Mask1[2])->getZExtValue()+4,
- MaskEVT);
- if (Mask1[3].getOpcode() != ISD::UNDEF)
- Mask1[3] =
- DAG.getConstant(cast<ConstantSDNode>(Mask1[3])->getZExtValue()+4,
- MaskEVT);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V2, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MaskVT, &Mask1[0], 4));
+ Mask1[0] = HiIndex & 1 ? 2 : 0;
+ Mask1[1] = HiIndex & 1 ? 0 : 2;
+ Mask1[2] = PermMask[2];
+ Mask1[3] = PermMask[3];
+ if (Mask1[2] >= 0)
+ Mask1[2] += 4;
+ if (Mask1[3] >= 0)
+ Mask1[3] += 4;
+ return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
}
}
// Break it into (shuffle shuffle_hi, shuffle_lo).
Locs.clear();
- SmallVector<SDValue,8> LoMask(4, DAG.getUNDEF(MaskEVT));
- SmallVector<SDValue,8> HiMask(4, DAG.getUNDEF(MaskEVT));
- SmallVector<SDValue,8> *MaskPtr = &LoMask;
+ SmallVector<int,8> LoMask(4U, -1);
+ SmallVector<int,8> HiMask(4U, -1);
+
+ SmallVector<int,8> *MaskPtr = &LoMask;
unsigned MaskIdx = 0;
unsigned LoIdx = 0;
unsigned HiIdx = 2;
LoIdx = 0;
HiIdx = 2;
}
- SDValue Elt = PermMask.getOperand(i);
- if (Elt.getOpcode() == ISD::UNDEF) {
+ int Idx = PermMask[i];
+ if (Idx < 0) {
Locs[i] = std::make_pair(-1, -1);
- } else if (cast<ConstantSDNode>(Elt)->getZExtValue() < 4) {
+ } else if (Idx < 4) {
Locs[i] = std::make_pair(MaskIdx, LoIdx);
- (*MaskPtr)[LoIdx] = Elt;
+ (*MaskPtr)[LoIdx] = Idx;
LoIdx++;
} else {
Locs[i] = std::make_pair(MaskIdx, HiIdx);
- (*MaskPtr)[HiIdx] = Elt;
+ (*MaskPtr)[HiIdx] = Idx;
HiIdx++;
}
}
- SDValue LoShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &LoMask[0], LoMask.size()));
- SDValue HiShuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &HiMask[0], HiMask.size()));
- SmallVector<SDValue, 8> MaskOps;
+ SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
+ SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
+ SmallVector<int, 8> MaskOps;
for (unsigned i = 0; i != 4; ++i) {
if (Locs[i].first == -1) {
- MaskOps.push_back(DAG.getUNDEF(MaskEVT));
+ MaskOps.push_back(-1);
} else {
unsigned Idx = Locs[i].first * 4 + Locs[i].second;
- MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
+ MaskOps.push_back(Idx);
}
}
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, LoShuffle, HiShuffle,
- DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &MaskOps[0], MaskOps.size()));
+ return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
}
SDValue
X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
- SDValue PermMask = Op.getOperand(2);
MVT VT = Op.getValueType();
DebugLoc dl = Op.getDebugLoc();
- unsigned NumElems = PermMask.getNumOperands();
+ unsigned NumElems = VT.getVectorNumElements();
bool isMMX = VT.getSizeInBits() == 64;
bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
bool V1IsSplat = false;
bool V2IsSplat = false;
- if (isUndefShuffle(Op.getNode()))
- return DAG.getUNDEF(VT);
-
- if (isZeroShuffle(Op.getNode()))
+ if (isZeroShuffle(SVOp))
return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
- if (isIdentityMask(PermMask.getNode()))
- return V1;
- else if (isIdentityMask(PermMask.getNode(), true))
- return V2;
-
- // Canonicalize movddup shuffles.
- if (V2IsUndef && Subtarget->hasSSE2() &&
- VT.getSizeInBits() == 128 &&
- X86::isMOVDDUPMask(PermMask.getNode()))
- return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
-
- if (isSplatMask(PermMask.getNode())) {
- if (isMMX || NumElems < 4) return Op;
- // Promote it to a v4{if}32 splat.
- return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
+ // Promote splats to v4f32.
+ if (SVOp->isSplat()) {
+ if (isMMX || NumElems < 4)
+ return Op;
+ return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
}
// If the shuffle can be profitably rewritten as a narrower shuffle, then
// do it!
if (VT == MVT::v8i16 || VT == MVT::v16i8) {
- SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask, DAG,
- *this, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
if (NewOp.getNode())
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
LowerVECTOR_SHUFFLE(NewOp, DAG));
} else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
// FIXME: Figure out a cleaner way to do this.
// Try to make use of movq to zero out the top part.
if (ISD::isBuildVectorAllZeros(V2.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
- DAG, *this, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
if (NewOp.getNode()) {
- SDValue NewV1 = NewOp.getOperand(0);
- SDValue NewV2 = NewOp.getOperand(1);
- SDValue NewMask = NewOp.getOperand(2);
- if (isCommutedMOVL(NewMask.getNode(), true, false)) {
- NewOp = CommuteVectorShuffle(NewOp, NewV1, NewV2, NewMask, DAG);
- return getVZextMovL(VT, NewOp.getValueType(), NewV2, DAG, Subtarget,
- dl);
- }
+ if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
+ return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
+ DAG, Subtarget, dl);
}
} else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
- SDValue NewOp= RewriteAsNarrowerShuffle(V1, V2, VT, PermMask,
- DAG, *this, dl);
- if (NewOp.getNode() && X86::isMOVLMask(NewOp.getOperand(2).getNode()))
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
+ if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
- DAG, Subtarget, dl);
+ 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;
SDValue ShVal;
- bool isShift = isVectorShift(Op, PermMask, DAG, isLeft, ShVal, ShAmt);
+ bool isShift = getSubtarget()->hasSSE2() &&
+ isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
if (isShift && ShVal.hasOneUse()) {
- // If the shifted value has multiple uses, it may be cheaper to use
+ // If the shifted value has multiple uses, it may be cheaper to use
// v_set0 + movlhps or movhlps, etc.
MVT EVT = VT.getVectorElementType();
ShAmt *= EVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
-
- if (X86::isMOVLMask(PermMask.getNode())) {
+
+ if (X86::isMOVLMask(SVOp)) {
if (V1IsUndef)
return V2;
if (ISD::isBuildVectorAllZeros(V1.getNode()))
if (!isMMX)
return Op;
}
-
- if (!isMMX && (X86::isMOVSHDUPMask(PermMask.getNode()) ||
- X86::isMOVSLDUPMask(PermMask.getNode()) ||
- X86::isMOVHLPSMask(PermMask.getNode()) ||
- X86::isMOVHPMask(PermMask.getNode()) ||
- X86::isMOVLPMask(PermMask.getNode())))
+
+ // FIXME: fold these into legal mask.
+ if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
+ X86::isMOVSLDUPMask(SVOp) ||
+ X86::isMOVHLPSMask(SVOp) ||
+ X86::isMOVHPMask(SVOp) ||
+ X86::isMOVLPMask(SVOp)))
return Op;
- if (ShouldXformToMOVHLPS(PermMask.getNode()) ||
- ShouldXformToMOVLP(V1.getNode(), V2.getNode(), PermMask.getNode()))
- return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
+ if (ShouldXformToMOVHLPS(SVOp) ||
+ ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
+ return CommuteVectorShuffle(SVOp, DAG);
if (isShift) {
// No better options. Use a vshl / vsrl.
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.
V1IsSplat = isSplatVector(V1.getNode());
V2IsSplat = isSplatVector(V2.getNode());
-
+
// Canonicalize the splat or undef, if present, to be on the RHS.
if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
- Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
+ Op = CommuteVectorShuffle(SVOp, DAG);
+ SVOp = cast<ShuffleVectorSDNode>(Op);
+ V1 = SVOp->getOperand(0);
+ V2 = SVOp->getOperand(1);
std::swap(V1IsSplat, V2IsSplat);
std::swap(V1IsUndef, V2IsUndef);
Commuted = true;
}
- // FIXME: Figure out a cleaner way to do this.
- if (isCommutedMOVL(PermMask.getNode(), V2IsSplat, V2IsUndef)) {
- if (V2IsUndef) return V1;
- Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
- if (V2IsSplat) {
- // V2 is a splat, so the mask may be malformed. That is, it may point
- // to any V2 element. The instruction selectior won't like this. Get
- // a corrected mask and commute to form a proper MOVS{S|D}.
- SDValue NewMask = getMOVLMask(NumElems, DAG, dl);
- if (NewMask.getNode() != PermMask.getNode())
- Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
- }
- return Op;
+ if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
+ // Shuffling low element of v1 into undef, just return v1.
+ 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
+ // swapped operands to undo the commute.
+ return getMOVL(DAG, dl, VT, V2, V1);
}
- if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
- X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
- X86::isUNPCKLMask(PermMask.getNode()) ||
- X86::isUNPCKHMask(PermMask.getNode()))
+ if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
+ X86::isUNPCKH_v_undef_Mask(SVOp) ||
+ X86::isUNPCKLMask(SVOp) ||
+ X86::isUNPCKHMask(SVOp))
return Op;
if (V2IsSplat) {
// Normalize mask so all entries that point to V2 points to its first
// element then try to match unpck{h|l} again. If match, return a
// new vector_shuffle with the corrected mask.
- SDValue NewMask = NormalizeMask(PermMask, DAG);
- if (NewMask.getNode() != PermMask.getNode()) {
- if (X86::isUNPCKLMask(NewMask.getNode(), true)) {
- SDValue NewMask = getUnpacklMask(NumElems, DAG, dl);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
- } else if (X86::isUNPCKHMask(NewMask.getNode(), true)) {
- SDValue NewMask = getUnpackhMask(NumElems, DAG, dl);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1, V2, NewMask);
+ SDValue NewMask = NormalizeMask(SVOp, DAG);
+ ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
+ if (NSVOp != SVOp) {
+ if (X86::isUNPCKLMask(NSVOp, true)) {
+ return NewMask;
+ } else if (X86::isUNPCKHMask(NSVOp, true)) {
+ return NewMask;
}
}
}
- // Normalize the node to match x86 shuffle ops if needed
- if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.getNode()))
- Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
-
if (Commuted) {
// Commute is back and try unpck* again.
- Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
- if (X86::isUNPCKL_v_undef_Mask(PermMask.getNode()) ||
- X86::isUNPCKH_v_undef_Mask(PermMask.getNode()) ||
- X86::isUNPCKLMask(PermMask.getNode()) ||
- X86::isUNPCKHMask(PermMask.getNode()))
- return Op;
+ // FIXME: this seems wrong.
+ SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
+ ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
+ if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
+ X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
+ X86::isUNPCKLMask(NewSVOp) ||
+ X86::isUNPCKHMask(NewSVOp))
+ return NewOp;
}
- // Try PSHUF* first, then SHUFP*.
- // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
- // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
- if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.getNode())) {
- if (V2.getOpcode() != ISD::UNDEF)
- return DAG.getNode(ISD::VECTOR_SHUFFLE, dl, VT, V1,
- DAG.getUNDEF(VT), PermMask);
- return Op;
- }
+ // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
- if (!isMMX) {
- if (Subtarget->hasSSE2() &&
- (X86::isPSHUFDMask(PermMask.getNode()) ||
- X86::isPSHUFHWMask(PermMask.getNode()) ||
- X86::isPSHUFLWMask(PermMask.getNode()))) {
- MVT RVT = VT;
- if (VT == MVT::v4f32) {
- RVT = MVT::v4i32;
- Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT,
- DAG.getNode(ISD::BIT_CONVERT, dl, RVT, V1),
- DAG.getUNDEF(RVT), PermMask);
- } else if (V2.getOpcode() != ISD::UNDEF)
- Op = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, RVT, V1,
- DAG.getUNDEF(RVT), PermMask);
- if (RVT != VT)
- Op = DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
- return Op;
- }
-
- // Binary or unary shufps.
- if (X86::isSHUFPMask(PermMask.getNode()) ||
- (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.getNode())))
- return Op;
- }
+ // Normalize the node to match x86 shuffle ops if needed
+ if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
+ return CommuteVectorShuffle(SVOp, DAG);
+ // Check for legal shuffle and return?
+ SmallVector<int, 16> PermMask;
+ 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(V1, V2, PermMask, DAG, *this, dl);
+ SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
if (NewOp.getNode())
return NewOp;
}
+ if (VT == MVT::v16i8) {
+ SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(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(V1, V2, PermMask, VT, DAG, dl);
+ return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
return SDValue();
}
User->getValueType(0) != MVT::i32))
return SDValue();
SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
+ DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
Op.getOperand(0)),
Op.getOperand(1));
return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
if (Idx == 0)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
- DAG.getNode(ISD::BIT_CONVERT, dl,
+ DAG.getNode(ISD::BIT_CONVERT, dl,
MVT::v4i32, Vec),
Op.getOperand(1)));
// Transform it so it match pextrw which produces a 32-bit result.
unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
if (Idx == 0)
return Op;
+
// SHUFPS the element to the lowest double word, then movss.
- MVT MaskVT = MVT::getIntVectorWithNumElements(4);
- SmallVector<SDValue, 8> IdxVec;
- IdxVec.
- push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
- IdxVec.
- push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
- IdxVec.
- push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
- IdxVec.
- push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &IdxVec[0], IdxVec.size());
- SDValue Vec = Op.getOperand(0);
- Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
- Vec, DAG.getUNDEF(Vec.getValueType()), Mask);
+ int Mask[4] = { Idx, -1, -1, -1 };
+ MVT VVT = Op.getOperand(0).getValueType();
+ 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));
} else if (VT.getSizeInBits() == 64) {
// UNPCKHPD the element to the lowest double word, then movsd.
// Note if the lower 64 bits of the result of the UNPCKHPD is then stored
// to a f64mem, the whole operation is folded into a single MOVHPDmr.
- MVT MaskVT = MVT::getIntVectorWithNumElements(2);
- SmallVector<SDValue, 8> IdxVec;
- IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
- IdxVec.
- push_back(DAG.getUNDEF(MaskVT.getVectorElementType()));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVT,
- &IdxVec[0], IdxVec.size());
- SDValue Vec = Op.getOperand(0);
- Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, Vec.getValueType(),
- Vec, DAG.getUNDEF(Vec.getValueType()),
- Mask);
+ int Mask[2] = { 1, -1 };
+ MVT VVT = Op.getOperand(0).getValueType();
+ 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));
}
if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
isa<ConstantSDNode>(N2)) {
unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
- : X86ISD::PINSRW;
+ : X86ISD::PINSRW;
// Transform it so it match pinsr{b,w} which expects a GR32 as its second
// argument.
if (N1.getValueType() != MVT::i32)
// zero here. The DAG Combiner may combine an extract_elt index into these
// bits. For example (insert (extract, 3), 2) could be matched by putting
// the '3' into bits [7:6] of X86ISD::INSERTPS.
- // Bits [5:4] of the constant are the destination select. This is the
+ // Bits [5:4] of the constant are the destination select. This is the
// value of the incoming immediate.
- // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
+ // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
// combine either bitwise AND or insert of float 0.0 to set these bits.
N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
- if (EVT.getSizeInBits() == 16) {
+ if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
// Transform it so it match pinsrw which expects a 16-bit value in a GR32
// as its second argument.
if (N1.getValueType() != MVT::i32)
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
// FIXME there isn't really any debug info here, should come from the parent
DebugLoc dl = CP->getDebugLoc();
- SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
- getPointerTy(),
- CP->getAlignment());
+ SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
+ CP->getAlignment());
Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
// With PIC, the address is actually $g + Offset.
if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
Result);
}
-
+
// For Darwin & Mingw32, external and weak symbols are indirect, so we want to
// load the value at address GV, not the value of GV itself. This means that
// the GlobalAddress must be in the base or index register of the address, not
return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
}
+static SDValue
+GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
+ SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg) {
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ DebugLoc dl = GA->getDebugLoc();
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
+ GA->getValueType(0),
+ GA->getOffset());
+ if (InFlag) {
+ SDValue Ops[] = { Chain, TGA, *InFlag };
+ Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
+ } else {
+ SDValue Ops[] = { Chain, TGA };
+ Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
+ }
+ SDValue Flag = Chain.getValue(1);
+ return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
+}
+
// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
static SDValue
LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
PtrVT), InFlag);
InFlag = Chain.getValue(1);
- // emit leal symbol@TLSGD(,%ebx,1), %eax
- SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
- GA->getValueType(0),
- GA->getOffset());
- SDValue Ops[] = { Chain, TGA, InFlag };
- SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
- InFlag = Result.getValue(2);
- Chain = Result.getValue(1);
-
- // call ___tls_get_addr. This function receives its argument in
- // the register EAX.
- Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Result, InFlag);
- InFlag = Chain.getValue(1);
-
- NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
- SDValue Ops1[] = { Chain,
- DAG.getTargetExternalSymbol("___tls_get_addr",
- PtrVT),
- DAG.getRegister(X86::EAX, PtrVT),
- DAG.getRegister(X86::EBX, PtrVT),
- InFlag };
- Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 5);
- InFlag = Chain.getValue(1);
-
- return DAG.getCopyFromReg(Chain, dl, X86::EAX, PtrVT, InFlag);
+ return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX);
}
// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
static SDValue
LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
const MVT PtrVT) {
- SDValue InFlag, Chain;
- DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
-
- // emit leaq symbol@TLSGD(%rip), %rdi
- SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
- GA->getValueType(0),
- GA->getOffset());
- SDValue Ops[] = { DAG.getEntryNode(), TGA};
- SDValue Result = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
- Chain = Result.getValue(1);
- InFlag = Result.getValue(2);
-
- // call __tls_get_addr. This function receives its argument in
- // the register RDI.
- Chain = DAG.getCopyToReg(Chain, dl, X86::RDI, Result, InFlag);
- InFlag = Chain.getValue(1);
-
- NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
- SDValue Ops1[] = { Chain,
- DAG.getTargetExternalSymbol("__tls_get_addr",
- PtrVT),
- DAG.getRegister(X86::RDI, PtrVT),
- InFlag };
- Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops1, 4);
- InFlag = Chain.getValue(1);
-
- return DAG.getCopyFromReg(Chain, dl, X86::RAX, PtrVT, InFlag);
+ return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX);
}
// Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
// "local exec" model.
static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
- const MVT PtrVT) {
+ const MVT PtrVT, TLSModel::Model model,
+ bool is64Bit) {
DebugLoc dl = GA->getDebugLoc();
// Get the Thread Pointer
- SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER,
- DebugLoc::getUnknownLoc(), PtrVT);
+ SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
+ DebugLoc::getUnknownLoc(), PtrVT,
+ DAG.getRegister(is64Bit? X86::FS : X86::GS,
+ MVT::i32));
+
+ SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
+ NULL, 0);
+
// emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
// exec)
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
- GA->getValueType(0),
- GA->getOffset());
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
+ GA->getOffset());
SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
- if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
+ if (model == TLSModel::InitialExec)
Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
PseudoSourceValue::getGOT(), 0);
assert(Subtarget->isTargetELF() &&
"TLS not implemented for non-ELF targets");
GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
- // If the relocation model is PIC, use the "General Dynamic" TLS Model,
- // otherwise use the "Local Exec"TLS Model
+ GlobalValue *GV = GA->getGlobal();
+ TLSModel::Model model =
+ getTLSModel (GV, getTargetMachine().getRelocationModel());
if (Subtarget->is64Bit()) {
- return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
+ switch (model) {
+ case TLSModel::GeneralDynamic:
+ case TLSModel::LocalDynamic: // not implemented
+ return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
+
+ case TLSModel::InitialExec:
+ case TLSModel::LocalExec:
+ return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, true);
+ }
} else {
- if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
+ switch (model) {
+ case TLSModel::GeneralDynamic:
+ case TLSModel::LocalDynamic: // not implemented
return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
- else
- return LowerToTLSExecModel(GA, DAG, getPointerTy());
+
+ case TLSModel::InitialExec:
+ case TLSModel::LocalExec:
+ return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, false);
+ }
}
+ assert(0 && "Unreachable");
+ return SDValue();
}
SDValue
if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
!Subtarget->isPICStyleRIPRel()) {
Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
- DAG.getNode(X86ISD::GlobalBaseReg,
+ DAG.getNode(X86ISD::GlobalBaseReg,
DebugLoc::getUnknownLoc(),
getPointerTy()),
Result);
}
/// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
-/// take a 2 x i32 value to shift plus a shift amount.
+/// take a 2 x i32 value to shift plus a shift amount.
SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
assert(Op.getNumOperands() == 3 && "Not a double-shift!");
MVT VT = Op.getValueType();
SDValue ShOpHi = Op.getOperand(1);
SDValue ShAmt = Op.getOperand(2);
SDValue Tmp1 = isSRA ?
- DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
+ DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
DAG.getConstant(VTBits - 1, MVT::i8)) :
DAG.getConstant(0, VT);
SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
MVT SrcVT = Op.getOperand(0).getValueType();
+
+ if (SrcVT.isVector()) {
+ if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
+ return Op;
+ }
+ return SDValue();
+ }
+
assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
"Unknown SINT_TO_FP to lower!");
-
- // These are really Legal; caller falls through into that case.
+
+ // These are really Legal; return the operand so the caller accepts it as
+ // Legal.
if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
- return SDValue();
- if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 &&
- Subtarget->is64Bit())
- return SDValue();
-
+ return Op;
+ if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
+ Subtarget->is64Bit()) {
+ return Op;
+ }
+
DebugLoc dl = Op.getDebugLoc();
unsigned Size = SrcVT.getSizeInBits()/8;
MachineFunction &MF = DAG.getMachineFunction();
int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
- StackSlot,
- PseudoSourceValue::getFixedStack(SSFI), 0);
+ StackSlot,
+ PseudoSourceValue::getFixedStack(SSFI), 0);
+ return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
+}
+SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
+ SDValue StackSlot,
+ SelectionDAG &DAG) {
// Build the FILD
+ DebugLoc dl = Op.getDebugLoc();
SDVTList Tys;
bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
if (useSSE)
CV0.push_back(ConstantInt::get(APInt(32, 0)));
CV0.push_back(ConstantInt::get(APInt(32, 0)));
Constant *C0 = ConstantVector::get(CV0);
- SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 4);
+ SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
std::vector<Constant*> CV1;
CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
Constant *C1 = ConstantVector::get(CV1);
- SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 4);
-
- SmallVector<SDValue, 4> MaskVec;
- MaskVec.push_back(DAG.getConstant(0, MVT::i32));
- MaskVec.push_back(DAG.getConstant(4, MVT::i32));
- MaskVec.push_back(DAG.getConstant(1, MVT::i32));
- MaskVec.push_back(DAG.getConstant(5, MVT::i32));
- SDValue UnpcklMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
- &MaskVec[0], MaskVec.size());
- SmallVector<SDValue, 4> MaskVec2;
- MaskVec2.push_back(DAG.getConstant(1, MVT::i32));
- MaskVec2.push_back(DAG.getConstant(0, MVT::i32));
- SDValue ShufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32,
- &MaskVec2[0], MaskVec2.size());
+ SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
Op.getOperand(0),
DAG.getIntPtrConstant(0)));
- SDValue Unpck1 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
- XR1, XR2, UnpcklMask);
+ SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
PseudoSourceValue::getConstantPool(), 0,
false, 16);
- SDValue Unpck2 = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v4i32,
- Unpck1, CLod0, UnpcklMask);
+ SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
PseudoSourceValue::getConstantPool(), 0,
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
// Add the halves; easiest way is to swap them into another reg first.
- SDValue Shuf = DAG.getNode(ISD::VECTOR_SHUFFLE, dl, MVT::v2f64,
- Sub, Sub, ShufMask);
+ int ShufMask[2] = { 1, -1 };
+ SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
+ DAG.getUNDEF(MVT::v2f64), ShufMask);
SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
DAG.getIntPtrConstant(0));
MVT SrcVT = N0.getValueType();
if (SrcVT == MVT::i64) {
- // We only handle SSE2 f64 target here; caller can handle the rest.
+ // We only handle SSE2 f64 target here; caller can expand the rest.
if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
return SDValue();
return LowerUINT_TO_FP_i64(Op, DAG);
- } else if (SrcVT == MVT::i32) {
+ } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
return LowerUINT_TO_FP_i32(Op, DAG);
}
- assert(0 && "Unknown UINT_TO_FP to lower!");
- return SDValue();
+ assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
+
+ // Make a 64-bit buffer, and use it to build an FILD.
+ SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
+ SDValue WordOff = DAG.getConstant(4, getPointerTy());
+ SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
+ getPointerTy(), StackSlot, WordOff);
+ SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
+ StackSlot, NULL, 0);
+ SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
+ OffsetSlot, NULL, 0);
+ return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
}
std::pair<SDValue,SDValue> X86TargetLowering::
-FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
+FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
DebugLoc dl = Op.getDebugLoc();
- assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
- Op.getValueType().getSimpleVT() >= MVT::i16 &&
+
+ MVT DstTy = Op.getValueType();
+
+ if (!IsSigned) {
+ assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
+ DstTy = MVT::i64;
+ }
+
+ assert(DstTy.getSimpleVT() <= MVT::i64 &&
+ DstTy.getSimpleVT() >= MVT::i16 &&
"Unknown FP_TO_SINT to lower!");
// These are really Legal.
- if (Op.getValueType() == MVT::i32 &&
+ if (DstTy == MVT::i32 &&
isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
return std::make_pair(SDValue(), SDValue());
if (Subtarget->is64Bit() &&
- Op.getValueType() == MVT::i64 &&
- Op.getOperand(0).getValueType() != MVT::f80)
+ DstTy == MVT::i64 &&
+ isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
return std::make_pair(SDValue(), SDValue());
// We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
// stack slot.
MachineFunction &MF = DAG.getMachineFunction();
- unsigned MemSize = Op.getValueType().getSizeInBits()/8;
+ unsigned MemSize = DstTy.getSizeInBits()/8;
int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+
unsigned Opc;
- switch (Op.getValueType().getSimpleVT()) {
+ switch (DstTy.getSimpleVT()) {
default: assert(0 && "Invalid FP_TO_SINT to lower!");
case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
SDValue Chain = DAG.getEntryNode();
SDValue Value = Op.getOperand(0);
if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
- assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
+ assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
Chain = DAG.getStore(Chain, dl, Value, StackSlot,
PseudoSourceValue::getFixedStack(SSFI), 0);
SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
}
SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
- std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
+ if (Op.getValueType().isVector()) {
+ if (Op.getValueType() == MVT::v2i32 &&
+ Op.getOperand(0).getValueType() == MVT::v2f64) {
+ return Op;
+ }
+ return SDValue();
+ }
+
+ std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
SDValue FIST = Vals.first, StackSlot = Vals.second;
- if (FIST.getNode() == 0) return SDValue();
-
+ // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
+ if (FIST.getNode() == 0) return Op;
+
+ // Load the result.
+ return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
+ FIST, StackSlot, NULL, 0);
+}
+
+SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
+ std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
+ SDValue FIST = Vals.first, StackSlot = Vals.second;
+ assert(FIST.getNode() && "Unexpected failure");
+
// Load the result.
return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
FIST, StackSlot, NULL, 0);
CV.push_back(C);
}
Constant *C = ConstantVector::get(CV);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
PseudoSourceValue::getConstantPool(), 0,
false, 16);
CV.push_back(C);
}
Constant *C = ConstantVector::get(CV);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
PseudoSourceValue::getConstantPool(), 0,
false, 16);
if (VT.isVector()) {
return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
DAG.getNode(ISD::XOR, dl, MVT::v2i64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
+ DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
Op.getOperand(0)),
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
} else {
CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
}
Constant *C = ConstantVector::get(CV);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
PseudoSourceValue::getConstantPool(), 0,
false, 16);
CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
}
C = ConstantVector::get(CV);
- CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
PseudoSourceValue::getConstantPool(), 0,
false, 16);
return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
}
+/// Emit nodes that will be selected as "test Op0,Op0", or something
+/// equivalent.
+SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
+ SelectionDAG &DAG) {
+ DebugLoc dl = Op.getDebugLoc();
+
+ // CF and OF aren't always set the way we want. Determine which
+ // of these we need.
+ bool NeedCF = false;
+ bool NeedOF = false;
+ switch (X86CC) {
+ case X86::COND_A: case X86::COND_AE:
+ case X86::COND_B: case X86::COND_BE:
+ NeedCF = true;
+ break;
+ case X86::COND_G: case X86::COND_GE:
+ case X86::COND_L: case X86::COND_LE:
+ case X86::COND_O: case X86::COND_NO:
+ NeedOF = true;
+ break;
+ default: break;
+ }
+
+ // See if we can use the EFLAGS value from the operand instead of
+ // doing a separate TEST. TEST always sets OF and CF to 0, so unless
+ // we prove that the arithmetic won't overflow, we can't use OF or CF.
+ if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
+ unsigned Opcode = 0;
+ unsigned NumOperands = 0;
+ switch (Op.getNode()->getOpcode()) {
+ case ISD::ADD:
+ // Due to an isel shortcoming, be conservative if this add is likely to
+ // be selected as part of a load-modify-store instruction. When the root
+ // node in a match is a store, isel doesn't know how to remap non-chain
+ // non-flag uses of other nodes in the match, such as the ADD in this
+ // case. This leads to the ADD being left around and reselected, with
+ // the result being two adds in the output.
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ UE = Op.getNode()->use_end(); UI != UE; ++UI)
+ if (UI->getOpcode() == ISD::STORE)
+ goto default_case;
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
+ // An add of one will be selected as an INC.
+ if (C->getAPIntValue() == 1) {
+ Opcode = X86ISD::INC;
+ NumOperands = 1;
+ break;
+ }
+ // An add of negative one (subtract of one) will be selected as a DEC.
+ if (C->getAPIntValue().isAllOnesValue()) {
+ Opcode = X86ISD::DEC;
+ NumOperands = 1;
+ break;
+ }
+ }
+ // Otherwise use a regular EFLAGS-setting add.
+ Opcode = X86ISD::ADD;
+ NumOperands = 2;
+ break;
+ case ISD::SUB:
+ // Due to the ISEL shortcoming noted above, be conservative if this sub is
+ // likely to be selected as part of a load-modify-store instruction.
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ UE = Op.getNode()->use_end(); UI != UE; ++UI)
+ if (UI->getOpcode() == ISD::STORE)
+ goto default_case;
+ // Otherwise use a regular EFLAGS-setting sub.
+ Opcode = X86ISD::SUB;
+ NumOperands = 2;
+ break;
+ case X86ISD::ADD:
+ case X86ISD::SUB:
+ case X86ISD::INC:
+ case X86ISD::DEC:
+ return SDValue(Op.getNode(), 1);
+ default:
+ default_case:
+ break;
+ }
+ if (Opcode != 0) {
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
+ SmallVector<SDValue, 4> Ops;
+ for (unsigned i = 0; i != NumOperands; ++i)
+ Ops.push_back(Op.getOperand(i));
+ SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
+ DAG.ReplaceAllUsesWith(Op, New);
+ return SDValue(New.getNode(), 1);
+ }
+ }
+
+ // Otherwise just emit a CMP with 0, which is the TEST pattern.
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
+ DAG.getConstant(0, Op.getValueType()));
+}
+
+/// Emit nodes that will be selected as "cmp Op0,Op1", or something
+/// equivalent.
+SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
+ SelectionDAG &DAG) {
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
+ if (C->getAPIntValue() == 0)
+ return EmitTest(Op0, X86CC, DAG);
+
+ DebugLoc dl = Op0.getDebugLoc();
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
+}
+
SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
DebugLoc dl = Op.getDebugLoc();
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
-
+
// Lower (X & (1 << N)) == 0 to BT(X, N).
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
-
- SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
+
+ SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86CC, MVT::i8), Cond);
}
default: break;
case ISD::SETOEQ:
case ISD::SETEQ: SSECC = 0; break;
- case ISD::SETOGT:
+ case ISD::SETOGT:
case ISD::SETGT: Swap = true; // Fallthrough
case ISD::SETLT:
case ISD::SETOLT: SSECC = 1; break;
// Handle all other FP comparisons here.
return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, 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.
unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
bool Swap = false, Invert = false, FlipSigns = false;
-
+
switch (VT.getSimpleVT()) {
default: break;
case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
}
-
+
switch (SetCCOpcode) {
default: break;
case ISD::SETNE: Invert = true;
}
if (Swap)
std::swap(Op0, Op1);
-
+
// Since SSE has no unsigned integer comparisons, we need to flip the sign
// bits of the inputs before performing those operations.
if (FlipSigns) {
Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
}
-
+
SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
// If the logical-not of the result is required, perform that now.
}
// isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
-static bool isX86LogicalCmp(unsigned Opc) {
- return Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI;
+static bool isX86LogicalCmp(SDValue Op) {
+ unsigned Opc = Op.getNode()->getOpcode();
+ if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
+ return true;
+ if (Op.getResNo() == 1 &&
+ (Opc == X86ISD::ADD ||
+ Opc == X86ISD::SUB ||
+ Opc == X86ISD::SMUL ||
+ Opc == X86ISD::UMUL ||
+ Opc == X86ISD::INC ||
+ Opc == X86ISD::DEC))
+ return true;
+
+ return false;
}
SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
SDValue Cmp = Cond.getOperand(1);
unsigned Opc = Cmp.getOpcode();
MVT VT = Op.getValueType();
-
+
bool IllegalFPCMov = false;
if (VT.isFloatingPoint() && !VT.isVector() &&
!isScalarFPTypeInSSEReg(VT)) // FPStack?
IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
-
- if ((isX86LogicalCmp(Opc) && !IllegalFPCMov) || Opc == X86ISD::BT) { // FIXME
+
+ if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
+ Opc == X86ISD::BT) { // FIXME
Cond = Cmp;
addTest = false;
}
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
- Cond= DAG.getNode(X86ISD::CMP, dl, MVT::i32, Cond,
- DAG.getConstant(0, MVT::i8));
+ Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
- const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
- MVT::Flag);
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
SmallVector<SDValue, 4> Ops;
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
Ops.push_back(Op.getOperand(1));
Ops.push_back(CC);
Ops.push_back(Cond);
- return DAG.getNode(X86ISD::CMOV, dl, VTs, 2, &Ops[0], Ops.size());
+ return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
}
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
Cond.getOpcode() == X86ISD::UMUL)
Cond = LowerXALUO(Cond, DAG);
#endif
-
+
// If condition flag is set by a X86ISD::CMP, then use it as the condition
// setting operand in place of the X86ISD::SETCC.
if (Cond.getOpcode() == X86ISD::SETCC) {
SDValue Cmp = Cond.getOperand(1);
unsigned Opc = Cmp.getOpcode();
// FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
- if (isX86LogicalCmp(Opc) || Opc == X86ISD::BT) {
+ if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
Cond = Cmp;
addTest = false;
} else {
unsigned CondOpc;
if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
SDValue Cmp = Cond.getOperand(0).getOperand(1);
- unsigned Opc = Cmp.getOpcode();
if (CondOpc == ISD::OR) {
// Also, recognize the pattern generated by an FCMP_UNE. We can emit
// two branches instead of an explicit OR instruction with a
// separate test.
if (Cmp == Cond.getOperand(1).getOperand(1) &&
- isX86LogicalCmp(Opc)) {
+ isX86LogicalCmp(Cmp)) {
CC = Cond.getOperand(0).getOperand(0);
Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
Chain, Dest, CC, Cmp);
// have a fall-through edge, because this requires an explicit
// jmp when the condition is false.
if (Cmp == Cond.getOperand(1).getOperand(1) &&
- isX86LogicalCmp(Opc) &&
+ isX86LogicalCmp(Cmp) &&
Op.getNode()->hasOneUse()) {
X86::CondCode CCode =
(X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
- Cond= DAG.getNode(X86ISD::CMP, dl, MVT::i32, Cond,
- DAG.getConstant(0, MVT::i8));
+ Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
Chain, Dest, CC, Cond);
V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
MVT IntPtr = getPointerTy();
const Type *IntPtrTy = TD->getIntPtrType();
- TargetLowering::ArgListTy Args;
+ TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
Entry.Node = Dst;
Entry.Ty = IntPtrTy;
Entry.Node = Size;
Args.push_back(Entry);
std::pair<SDValue,SDValue> CallResult =
- LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
- CallingConv::C, false,
+ LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
+ CallingConv::C, false,
DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
return CallResult.second;
}
InFlag = Chain.getValue(1);
}
- Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
+ Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
+ Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
MVT CVT = Count.getValueType();
SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
- Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
+ Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
X86::ECX,
Left, InFlag);
InFlag = Chain.getValue(1);
SDValue Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstSVOff,
- const Value *SrcSV, uint64_t SrcSVOff) {
+ const Value *SrcSV, uint64_t SrcSVOff) {
// This requires the copy size to be a constant, preferrably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
unsigned BytesLeft = SizeVal % UBytes;
SDValue InFlag(0, 0);
- Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
+ Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
+ Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
+ Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
X86::ESI,
Src, InFlag);
InFlag = Chain.getValue(1);
MVT DstVT = Dst.getValueType();
MVT SrcVT = Src.getValueType();
MVT SizeVT = Size.getValueType();
- Results.push_back(DAG.getMemcpy(Chain, dl,
+ Results.push_back(DAG.getMemcpy(Chain, dl,
DAG.getNode(ISD::ADD, dl, DstVT, Dst,
DAG.getConstant(Offset, DstVT)),
DAG.getNode(ISD::ADD, dl, SrcVT, Src,
SrcSV, SrcSVOff + Offset));
}
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&Results[0], Results.size());
}
MemOps.push_back(Store);
// Store fp_offset
- FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
FIN, DAG.getIntPtrConstant(4));
Store = DAG.getStore(Op.getOperand(0), dl,
DAG.getConstant(VarArgsFPOffset, MVT::i32),
MemOps.push_back(Store);
// Store ptr to overflow_arg_area
- FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
FIN, DAG.getIntPtrConstant(4));
SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
MemOps.push_back(Store);
// Store ptr to reg_save_area.
- FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
FIN, DAG.getIntPtrConstant(8));
SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
MemOps.push_back(Store);
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&MemOps[0], MemOps.size());
}
DAG.getConstant(TD->getPointerSize(),
Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
- DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ DAG.getNode(ISD::ADD, dl, getPointerTy(),
FrameAddr, Offset),
NULL, 0);
}
// Just load the return address.
SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
- return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
+ return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
RetAddrFI, NULL, 0);
}
OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
Addr, TrmpAddr, 0);
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(2, MVT::i64));
OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
// Load the 'nest' parameter value into R10.
// R10 is specified in X86CallingConv.td
OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(10, MVT::i64));
OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
Addr, TrmpAddr, 10);
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(12, MVT::i64));
OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
// Jump to the nested function.
OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(20, MVT::i64));
OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
Addr, TrmpAddr, 20);
unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
DAG.getConstant(22, MVT::i64));
OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
TrmpAddr, 22);
SDValue OutChains[4];
SDValue Addr, Disp;
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(10, MVT::i32));
Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
- OutChains[0] = DAG.getStore(Root, dl,
+ OutChains[0] = DAG.getStore(Root, dl,
DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
Trmp, TrmpAddr, 0);
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(1, MVT::i32));
OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(5, MVT::i32));
OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
TrmpAddr, 5, false, 1);
- Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
+ Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
DAG.getConstant(6, MVT::i32));
OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
MVT VT = Op.getValueType();
assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
DebugLoc dl = Op.getDebugLoc();
-
+
// ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
// ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
// ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
SDValue A = Op.getOperand(0);
SDValue B = Op.getOperand(1);
-
+
SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
A, DAG.getConstant(32, MVT::i32));
switch (Op.getOpcode()) {
default: assert(0 && "Unknown ovf instruction!");
case ISD::SADDO:
+ // A subtract of one will be selected as a INC. Note that INC doesn't
+ // set CF, so we can't do this for UADDO.
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (C->getAPIntValue() == 1) {
+ BaseOp = X86ISD::INC;
+ Cond = X86::COND_O;
+ break;
+ }
BaseOp = X86ISD::ADD;
Cond = X86::COND_O;
break;
Cond = X86::COND_B;
break;
case ISD::SSUBO:
+ // A subtract of one will be selected as a DEC. Note that DEC doesn't
+ // set CF, so we can't do this for USUBO.
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (C->getAPIntValue() == 1) {
+ BaseOp = X86ISD::DEC;
+ Cond = X86::COND_O;
+ break;
+ }
BaseOp = X86ISD::SUB;
Cond = X86::COND_O;
break;
case MVT::i8: Reg = X86::AL; size = 1; break;
case MVT::i16: Reg = X86::AX; size = 2; break;
case MVT::i32: Reg = X86::EAX; size = 4; break;
- case MVT::i64:
+ case MVT::i64:
assert(Subtarget->is64Bit() && "Node not type legal!");
Reg = X86::RAX; size = 8;
break;
cpIn.getValue(1) };
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
- SDValue cpOut =
+ SDValue cpOut =
DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
return cpOut;
}
DebugLoc dl = Node->getDebugLoc();
MVT T = Node->getValueType(0);
SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
- DAG.getConstant(0, T), Node->getOperand(2));
+ DAG.getConstant(0, T), Node->getOperand(2));
return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
cast<AtomicSDNode>(Node)->getMemoryVT(),
Node->getOperand(0),
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
+ case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
case ISD::FABS: return LowerFABS(Op, DAG);
case ISD::FNEG: return LowerFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
assert(false && "Do not know how to custom type legalize this operation!");
return;
case ISD::FP_TO_SINT: {
- std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
+ std::pair<SDValue,SDValue> Vals =
+ FP_TO_INTHelper(SDValue(N, 0), DAG, true);
SDValue FIST = Vals.first, StackSlot = Vals.second;
if (FIST.getNode() != 0) {
MVT VT = N->getValueType(0);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
SDValue TheChain = N->getOperand(0);
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
- SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
+ SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
rd.getValue(1));
SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
eax.getValue(2));
case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
case X86ISD::PINSRB: return "X86ISD::PINSRB";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
+ case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::FRCP: return "X86ISD::FRCP";
case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
- case X86ISD::THREAD_POINTER: return "X86ISD::THREAD_POINTER";
+ case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
case X86ISD::SUB: return "X86ISD::SUB";
case X86ISD::SMUL: return "X86ISD::SMUL";
case X86ISD::UMUL: return "X86ISD::UMUL";
+ case X86ISD::INC: return "X86ISD::INC";
+ case X86ISD::DEC: return "X86ISD::DEC";
+ case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
}
}
// isLegalAddressingMode - Return true if the addressing mode represented
// by AM is legal for this target, for a load/store of the specified type.
-bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
+bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
const Type *Ty) const {
// X86 supports extremely general addressing modes.
-
+
// X86 allows a sign-extended 32-bit immediate field as a displacement.
if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
return false;
-
+
if (AM.BaseGV) {
// We can only fold this if we don't need an extra load.
if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
return false;
}
}
-
+
switch (AM.Scale) {
case 0:
case 1:
default: // Other stuff never works.
return false;
}
-
+
return true;
}
return Subtarget->is64Bit() || NumBits1 < 64;
}
+bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
+ // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
+ return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
+}
+
+bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
+ // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
+ return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
+}
+
+bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
+ // i16 instructions are longer (0x66 prefix) and potentially slower.
+ return !(VT1 == MVT::i32 && VT2 == MVT::i16);
+}
+
/// isShuffleMaskLegal - Targets can use this to indicate that they only
/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
/// are assumed to be legal.
bool
-X86TargetLowering::isShuffleMaskLegal(SDValue Mask, MVT VT) const {
+X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
+ MVT VT) const {
// Only do shuffles on 128-bit vector types for now.
- if (VT.getSizeInBits() == 64) return false;
- return (Mask.getNode()->getNumOperands() <= 4 ||
- isIdentityMask(Mask.getNode()) ||
- isIdentityMask(Mask.getNode(), true) ||
- isSplatMask(Mask.getNode()) ||
- isPSHUFHW_PSHUFLWMask(Mask.getNode()) ||
- X86::isUNPCKLMask(Mask.getNode()) ||
- X86::isUNPCKHMask(Mask.getNode()) ||
- X86::isUNPCKL_v_undef_Mask(Mask.getNode()) ||
- X86::isUNPCKH_v_undef_Mask(Mask.getNode()));
+ if (VT.getSizeInBits() == 64)
+ return false;
+
+ // FIXME: pshufb, blends, palignr, shifts.
+ return (VT.getVectorNumElements() == 2 ||
+ ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
+ isMOVLMask(M, VT) ||
+ isSHUFPMask(M, VT) ||
+ isPSHUFDMask(M, VT) ||
+ isPSHUFHWMask(M, VT) ||
+ isPSHUFLWMask(M, VT) ||
+ isUNPCKLMask(M, VT) ||
+ isUNPCKHMask(M, VT) ||
+ isUNPCKL_v_undef_Mask(M, VT) ||
+ isUNPCKH_v_undef_Mask(M, VT));
}
bool
-X86TargetLowering::isVectorClearMaskLegal(const std::vector<SDValue> &BVOps,
- MVT EVT, SelectionDAG &DAG) const {
- unsigned NumElts = BVOps.size();
- // Only do shuffles on 128-bit vector types for now.
- if (EVT.getSizeInBits() * NumElts == 64) return false;
- if (NumElts == 2) return true;
- if (NumElts == 4) {
- return (isMOVLMask(&BVOps[0], 4) ||
- isCommutedMOVL(&BVOps[0], 4, true) ||
- isSHUFPMask(&BVOps[0], 4) ||
- isCommutedSHUFP(&BVOps[0], 4));
+X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
+ MVT VT) const {
+ unsigned NumElts = VT.getVectorNumElements();
+ // FIXME: This collection of masks seems suspect.
+ if (NumElts == 2)
+ return true;
+ if (NumElts == 4 && VT.getSizeInBits() == 128) {
+ return (isMOVLMask(Mask, VT) ||
+ isCommutedMOVLMask(Mask, VT, true) ||
+ isSHUFPMask(Mask, VT) ||
+ isCommutedSHUFPMask(Mask, VT));
}
return false;
}
const BasicBlock *LLVM_BB = MBB->getBasicBlock();
MachineFunction::iterator MBBIter = MBB;
++MBBIter;
-
+
/// First build the CFG
MachineFunction *F = MBB->getParent();
MachineBasicBlock *thisMBB = MBB;
MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
-
+
// Move all successors to thisMBB to nextMBB
nextMBB->transferSuccessors(thisMBB);
-
+
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
-
+
// newMBB jumps to itself and fall through to nextMBB
newMBB->addSuccessor(nextMBB);
newMBB->addSuccessor(newMBB);
-
+
// Insert instructions into newMBB based on incoming instruction
- assert(bInstr->getNumOperands() < 8 && "unexpected number of operands");
+ assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
+ "unexpected number of operands");
DebugLoc dl = bInstr->getDebugLoc();
MachineOperand& destOper = bInstr->getOperand(0);
- MachineOperand* argOpers[6];
+ MachineOperand* argOpers[2 + X86AddrNumOperands];
int numArgs = bInstr->getNumOperands() - 1;
for (int i=0; i < numArgs; ++i)
argOpers[i] = &bInstr->getOperand(i+1);
// x86 address has 4 operands: base, index, scale, and displacement
- int lastAddrIndx = 3; // [0,3]
- int valArgIndx = 4;
-
+ int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
+ int valArgIndx = lastAddrIndx + 1;
+
unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
for (int i=0; i <= lastAddrIndx; ++i)
if (invSrc) {
MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
}
- else
+ else
tt = t1;
unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
MIB.addReg(t1);
-
+
MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
for (int i=0; i <= lastAddrIndx; ++i)
(*MIB).addOperand(*argOpers[i]);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
MIB.addReg(EAXreg);
-
+
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
const BasicBlock *LLVM_BB = MBB->getBasicBlock();
MachineFunction::iterator MBBIter = MBB;
++MBBIter;
-
+
/// First build the CFG
MachineFunction *F = MBB->getParent();
MachineBasicBlock *thisMBB = MBB;
MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
-
+
// Move all successors to thisMBB to nextMBB
nextMBB->transferSuccessors(thisMBB);
-
+
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
-
+
// newMBB jumps to itself and fall through to nextMBB
newMBB->addSuccessor(nextMBB);
newMBB->addSuccessor(newMBB);
-
+
DebugLoc dl = bInstr->getDebugLoc();
// Insert instructions into newMBB based on incoming instruction
// There are 8 "real" operands plus 9 implicit def/uses, ignored here.
- assert(bInstr->getNumOperands() < 18 && "unexpected number of operands");
+ assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
+ "unexpected number of operands");
MachineOperand& dest1Oper = bInstr->getOperand(0);
MachineOperand& dest2Oper = bInstr->getOperand(1);
- MachineOperand* argOpers[6];
- for (int i=0; i < 6; ++i)
+ MachineOperand* argOpers[2 + X86AddrNumOperands];
+ for (int i=0; i < 2 + X86AddrNumOperands; ++i)
argOpers[i] = &bInstr->getOperand(i+2);
// x86 address has 4 operands: base, index, scale, and displacement
- int lastAddrIndx = 3; // [0,3]
-
+ int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
+
unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
for (int i=0; i <= lastAddrIndx; ++i)
unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
// add 4 to displacement.
- for (int i=0; i <= lastAddrIndx-1; ++i)
+ for (int i=0; i <= lastAddrIndx-2; ++i)
(*MIB).addOperand(*argOpers[i]);
MachineOperand newOp3 = *(argOpers[3]);
if (newOp3.isImm())
else
newOp3.setOffset(newOp3.getOffset()+4);
(*MIB).addOperand(newOp3);
+ (*MIB).addOperand(*argOpers[lastAddrIndx]);
// t3/4 are defined later, at the bottom of the loop
unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
- if (invSrc) {
+ if (invSrc) {
MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
} else {
tt2 = t2;
}
- assert((argOpers[4]->isReg() || argOpers[4]->isImm()) &&
+ int valArgIndx = lastAddrIndx + 1;
+ assert((argOpers[valArgIndx]->isReg() ||
+ argOpers[valArgIndx]->isImm()) &&
"invalid operand");
unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
- if (argOpers[4]->isReg())
+ if (argOpers[valArgIndx]->isReg())
MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
else
MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
if (regOpcL != X86::MOV32rr)
MIB.addReg(tt1);
- (*MIB).addOperand(*argOpers[4]);
- assert(argOpers[5]->isReg() == argOpers[4]->isReg());
- assert(argOpers[5]->isImm() == argOpers[4]->isImm());
- if (argOpers[5]->isReg())
+ (*MIB).addOperand(*argOpers[valArgIndx]);
+ assert(argOpers[valArgIndx + 1]->isReg() ==
+ argOpers[valArgIndx]->isReg());
+ assert(argOpers[valArgIndx + 1]->isImm() ==
+ argOpers[valArgIndx]->isImm());
+ if (argOpers[valArgIndx + 1]->isReg())
MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
else
MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
if (regOpcH != X86::MOV32rr)
MIB.addReg(tt2);
- (*MIB).addOperand(*argOpers[5]);
+ (*MIB).addOperand(*argOpers[valArgIndx + 1]);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
MIB.addReg(t1);
MIB.addReg(t5);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
MIB.addReg(t6);
-
+
MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
for (int i=0; i <= lastAddrIndx; ++i)
(*MIB).addOperand(*argOpers[i]);
MIB.addReg(X86::EAX);
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
MIB.addReg(X86::EDX);
-
+
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
// thisMBB:
// newMBB:
// ld t1 = [min/max.addr]
- // mov t2 = [min/max.val]
+ // mov t2 = [min/max.val]
// cmp t1, t2
// cmov[cond] t2 = t1
// mov EAX = t1
const BasicBlock *LLVM_BB = MBB->getBasicBlock();
MachineFunction::iterator MBBIter = MBB;
++MBBIter;
-
+
/// First build the CFG
MachineFunction *F = MBB->getParent();
MachineBasicBlock *thisMBB = MBB;
MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
-
+
// Move all successors to thisMBB to nextMBB
nextMBB->transferSuccessors(thisMBB);
-
+
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
-
+
// newMBB jumps to newMBB and fall through to nextMBB
newMBB->addSuccessor(nextMBB);
newMBB->addSuccessor(newMBB);
-
+
DebugLoc dl = mInstr->getDebugLoc();
// Insert instructions into newMBB based on incoming instruction
- assert(mInstr->getNumOperands() < 8 && "unexpected number of operands");
+ assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
+ "unexpected number of operands");
MachineOperand& destOper = mInstr->getOperand(0);
- MachineOperand* argOpers[6];
+ MachineOperand* argOpers[2 + X86AddrNumOperands];
int numArgs = mInstr->getNumOperands() - 1;
for (int i=0; i < numArgs; ++i)
argOpers[i] = &mInstr->getOperand(i+1);
-
+
// x86 address has 4 operands: base, index, scale, and displacement
- int lastAddrIndx = 3; // [0,3]
- int valArgIndx = 4;
-
+ int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
+ int valArgIndx = lastAddrIndx + 1;
+
unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
for (int i=0; i <= lastAddrIndx; ++i)
assert((argOpers[valArgIndx]->isReg() ||
argOpers[valArgIndx]->isImm()) &&
"invalid operand");
-
- unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
+
+ unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
if (argOpers[valArgIndx]->isReg())
MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
- else
+ else
MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
(*MIB).addOperand(*argOpers[valArgIndx]);
MIB.addReg(t3);
assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
(*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
-
+
MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
MIB.addReg(X86::EAX);
-
+
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
// 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...
AM.Disp = Op.getImm();
}
addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
- .addReg(MI->getOperand(4).getReg());
+ .addReg(MI->getOperand(X86AddrNumOperands).getReg());
// Reload the original control word now.
addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
}
case X86::ATOMAND32:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
- X86::AND32ri, X86::MOV32rm,
+ X86::AND32ri, X86::MOV32rm,
X86::LCMPXCHG32, X86::MOV32rr,
X86::NOT32r, X86::EAX,
X86::GR32RegisterClass);
case X86::ATOMOR32:
- return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
- X86::OR32ri, X86::MOV32rm,
+ return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
+ X86::OR32ri, X86::MOV32rm,
X86::LCMPXCHG32, X86::MOV32rr,
X86::NOT32r, X86::EAX,
X86::GR32RegisterClass);
case X86::ATOMXOR32:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
- X86::XOR32ri, X86::MOV32rm,
+ X86::XOR32ri, X86::MOV32rm,
X86::LCMPXCHG32, X86::MOV32rr,
X86::NOT32r, X86::EAX,
X86::GR32RegisterClass);
X86::NOT16r, X86::AX,
X86::GR16RegisterClass);
case X86::ATOMOR16:
- return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
+ return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
X86::OR16ri, X86::MOV16rm,
X86::LCMPXCHG16, X86::MOV16rr,
X86::NOT16r, X86::AX,
X86::NOT8r, X86::AL,
X86::GR8RegisterClass);
case X86::ATOMOR8:
- return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
+ return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
X86::OR8ri, X86::MOV8rm,
X86::LCMPXCHG8, X86::MOV8rr,
X86::NOT8r, X86::AL,
// This group is for 64-bit host.
case X86::ATOMAND64:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
- X86::AND64ri32, X86::MOV64rm,
+ X86::AND64ri32, X86::MOV64rm,
X86::LCMPXCHG64, X86::MOV64rr,
X86::NOT64r, X86::RAX,
X86::GR64RegisterClass);
case X86::ATOMOR64:
- return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
- X86::OR64ri32, X86::MOV64rm,
+ return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
+ X86::OR64ri32, X86::MOV64rm,
X86::LCMPXCHG64, X86::MOV64rr,
X86::NOT64r, X86::RAX,
X86::GR64RegisterClass);
case X86::ATOMXOR64:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
- X86::XOR64ri32, X86::MOV64rm,
+ X86::XOR64ri32, X86::MOV64rm,
X86::LCMPXCHG64, X86::MOV64rr,
X86::NOT64r, X86::RAX,
X86::GR64RegisterClass);
// This group does 64-bit operations on a 32-bit host.
case X86::ATOMAND6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::AND32rr, X86::AND32rr,
X86::AND32ri, X86::AND32ri,
false);
case X86::ATOMOR6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::OR32rr, X86::OR32rr,
X86::OR32ri, X86::OR32ri,
false);
case X86::ATOMXOR6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::XOR32rr, X86::XOR32rr,
X86::XOR32ri, X86::XOR32ri,
false);
case X86::ATOMNAND6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::AND32rr, X86::AND32rr,
X86::AND32ri, X86::AND32ri,
true);
case X86::ATOMADD6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::ADD32rr, X86::ADC32rr,
X86::ADD32ri, X86::ADC32ri,
false);
case X86::ATOMSUB6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::SUB32rr, X86::SBB32rr,
X86::SUB32ri, X86::SBB32ri,
false);
case X86::ATOMSWAP6432:
- return EmitAtomicBit6432WithCustomInserter(MI, BB,
+ return EmitAtomicBit6432WithCustomInserter(MI, BB,
X86::MOV32rr, X86::MOV32rr,
X86::MOV32ri, X86::MOV32ri,
false);
case X86ISD::SUB:
case X86ISD::SMUL:
case X86ISD::UMUL:
+ case X86ISD::INC:
+ case X86ISD::DEC:
// These nodes' second result is a boolean.
if (Op.getResNo() == 0)
break;
return false;
}
-static bool EltsFromConsecutiveLoads(SDNode *N, SDValue PermMask,
- unsigned NumElems, MVT EVT,
- SDNode *&Base,
+static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
+ MVT EVT, LoadSDNode *&LDBase,
+ unsigned &LastLoadedElt,
SelectionDAG &DAG, MachineFrameInfo *MFI,
const TargetLowering &TLI) {
- Base = NULL;
+ LDBase = NULL;
+ LastLoadedElt = -1U;
for (unsigned i = 0; i < NumElems; ++i) {
- SDValue Idx = PermMask.getOperand(i);
- if (Idx.getOpcode() == ISD::UNDEF) {
- if (!Base)
+ if (N->getMaskElt(i) < 0) {
+ if (!LDBase)
return false;
continue;
}
if (!Elt.getNode() ||
(Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
return false;
- if (!Base) {
- Base = Elt.getNode();
- if (Base->getOpcode() == ISD::UNDEF)
+ if (!LDBase) {
+ if (Elt.getNode()->getOpcode() == ISD::UNDEF)
return false;
+ LDBase = cast<LoadSDNode>(Elt.getNode());
+ LastLoadedElt = i;
continue;
}
if (Elt.getOpcode() == ISD::UNDEF)
continue;
- if (!TLI.isConsecutiveLoad(Elt.getNode(), Base,
- EVT.getSizeInBits()/8, i, MFI))
+ LoadSDNode *LD = cast<LoadSDNode>(Elt);
+ if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
return false;
+ LastLoadedElt = i;
}
return true;
}
/// PerformShuffleCombine - Combine a vector_shuffle that is equal to
/// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
/// if the load addresses are consecutive, non-overlapping, and in the right
-/// order.
+/// order. In the case of v2i64, it will see if it can rewrite the
+/// shuffle to be an appropriate build vector so it can take advantage of
+// performBuildVectorCombine.
static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
- const TargetLowering &TLI) {
- MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ const TargetLowering &TLI) {
DebugLoc dl = N->getDebugLoc();
MVT VT = N->getValueType(0);
MVT EVT = VT.getVectorElementType();
- SDValue PermMask = N->getOperand(2);
- unsigned NumElems = PermMask.getNumOperands();
- SDNode *Base = NULL;
- if (!EltsFromConsecutiveLoads(N, PermMask, NumElems, EVT, Base,
- DAG, MFI, TLI))
- return SDValue();
+ ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+ unsigned NumElems = VT.getVectorNumElements();
- LoadSDNode *LD = cast<LoadSDNode>(Base);
- if (isBaseAlignmentOfN(16, Base->getOperand(1).getNode(), TLI))
- return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile());
- return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), LD->getAlignment());
-}
-
-/// PerformBuildVectorCombine - build_vector 0,(load i64 / f64) -> movq / movsd.
-static SDValue PerformBuildVectorCombine(SDNode *N, SelectionDAG &DAG,
- TargetLowering::DAGCombinerInfo &DCI,
- const X86Subtarget *Subtarget,
- const TargetLowering &TLI) {
- unsigned NumOps = N->getNumOperands();
- DebugLoc dl = N->getDebugLoc();
-
- // Ignore single operand BUILD_VECTOR.
- if (NumOps == 1)
+ if (VT.getSizeInBits() != 128)
return SDValue();
- MVT VT = N->getValueType(0);
- MVT EVT = VT.getVectorElementType();
- if ((EVT != MVT::i64 && EVT != MVT::f64) || Subtarget->is64Bit())
- // We are looking for load i64 and zero extend. We want to transform
- // it before legalizer has a chance to expand it. Also look for i64
- // BUILD_PAIR bit casted to f64.
- return SDValue();
- // This must be an insertion into a zero vector.
- SDValue HighElt = N->getOperand(1);
- if (!isZeroNode(HighElt))
+ // Try to combine a vector_shuffle into a 128-bit load.
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ LoadSDNode *LD = NULL;
+ unsigned LastLoadedElt;
+ if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
+ MFI, TLI))
return SDValue();
- // Value must be a load.
- SDNode *Base = N->getOperand(0).getNode();
- if (!isa<LoadSDNode>(Base)) {
- if (Base->getOpcode() != ISD::BIT_CONVERT)
- return SDValue();
- Base = Base->getOperand(0).getNode();
- if (!isa<LoadSDNode>(Base))
- return SDValue();
+ if (LastLoadedElt == NumElems - 1) {
+ if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
+ return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->isVolatile());
+ return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->isVolatile(), LD->getAlignment());
+ } else if (NumElems == 4 && LastLoadedElt == 1) {
+ SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+ SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
+ SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
}
-
- // Transform it into VZEXT_LOAD addr.
- LoadSDNode *LD = cast<LoadSDNode>(Base);
-
- // Load must not be an extload.
- if (LD->getExtensionType() != ISD::NON_EXTLOAD)
- return SDValue();
-
- // Load type should legal type so we don't have to legalize it.
- if (!TLI.isTypeLegal(VT))
- return SDValue();
-
- SDVTList Tys = DAG.getVTList(VT, MVT::Other);
- SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
- SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
- TargetLowering::TargetLoweringOpt TLO(DAG);
- TLO.CombineTo(SDValue(Base, 1), ResNode.getValue(1));
- DCI.CommitTargetLoweringOpt(TLO);
- return ResNode;
-}
+ return SDValue();
+}
/// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
- DebugLoc dl = N->getDebugLoc();
+ const X86Subtarget *Subtarget) {
+ DebugLoc DL = N->getDebugLoc();
SDValue Cond = N->getOperand(0);
-
+ // 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() &&
- (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
- if (Cond.getOpcode() == ISD::SETCC) {
- // Get the LHS/RHS of the select.
- SDValue LHS = N->getOperand(1);
- SDValue RHS = N->getOperand(2);
- ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
-
- unsigned Opcode = 0;
- if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
- switch (CC) {
- default: break;
- case ISD::SETOLE: // (X <= Y) ? X : Y -> min
- case ISD::SETULE:
- case ISD::SETLE:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
- case ISD::SETLT:
- Opcode = X86ISD::FMIN;
- break;
+ (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
+ Cond.getOpcode() == ISD::SETCC) {
+ ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
- case ISD::SETOGT: // (X > Y) ? X : Y -> max
- case ISD::SETUGT:
- case ISD::SETGT:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
- case ISD::SETGE:
- Opcode = X86ISD::FMAX;
- break;
- }
- } 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::SETGT:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
- case ISD::SETGE:
- Opcode = X86ISD::FMIN;
- break;
+ unsigned Opcode = 0;
+ if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
+ switch (CC) {
+ default: break;
+ case ISD::SETOLE: // (X <= Y) ? X : Y -> min
+ case ISD::SETULE:
+ case ISD::SETLE:
+ if (!UnsafeFPMath) break;
+ // FALL THROUGH.
+ case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
+ case ISD::SETLT:
+ Opcode = X86ISD::FMIN;
+ break;
- case ISD::SETOLE: // (X <= Y) ? Y : X -> max
- case ISD::SETULE:
- case ISD::SETLE:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
- case ISD::SETLT:
- Opcode = X86ISD::FMAX;
- break;
- }
+ case ISD::SETOGT: // (X > Y) ? X : Y -> max
+ case ISD::SETUGT:
+ case ISD::SETGT:
+ if (!UnsafeFPMath) break;
+ // FALL THROUGH.
+ case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
+ case ISD::SETGE:
+ Opcode = X86ISD::FMAX;
+ break;
}
+ } 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::SETGT:
+ if (!UnsafeFPMath) break;
+ // FALL THROUGH.
+ case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
+ case ISD::SETGE:
+ Opcode = X86ISD::FMIN;
+ break;
- if (Opcode)
- return DAG.getNode(Opcode, dl, N->getValueType(0), LHS, RHS);
+ case ISD::SETOLE: // (X <= Y) ? Y : X -> max
+ case ISD::SETULE:
+ case ISD::SETLE:
+ if (!UnsafeFPMath) break;
+ // FALL THROUGH.
+ case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
+ case ISD::SETLT:
+ Opcode = X86ISD::FMAX;
+ break;
+ }
}
+ 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 (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
+ // Don't do this for crazy integer types.
+ if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
+ // 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.
+ (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
+ isa<ConstantSDNode>(Cond.getOperand(1))))) {
+ 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) {
+ default: break;
+ case 1: // result = add base, cond
+ case 2: // result = lea base( , cond*2)
+ case 3: // result = lea base(cond, cond*2)
+ case 4: // result = lea base( , cond*4)
+ case 5: // result = lea base(cond, cond*4)
+ case 8: // result = lea base( , cond*8)
+ case 9: // result = lea base(cond, cond*8)
+ isFastMultiplier = true;
+ 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);
+ // Scale the condition by the difference.
+ 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();
+}
+
+/// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
+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.
+ if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
+ if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
+ // 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.
+ if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
+ 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));
+ if (N->getNumValues() == 2) // Dead flag value?
+ 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) {
+ default: break;
+ case 1: // result = add base, cond
+ case 2: // result = lea base( , cond*2)
+ case 3: // result = lea base(cond, cond*2)
+ case 4: // result = lea base( , cond*4)
+ case 5: // result = lea base(cond, cond*4)
+ case 8: // result = lea base( , cond*8)
+ case 9: // result = lea base(cond, cond*8)
+ isFastMultiplier = true;
+ break;
+ }
+ }
+
+ if (isFastMultiplier) {
+ APInt Diff = TrueC->getAPIntValue()-FalseC->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);
+ // Scale the condition by the difference.
+ 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));
+ if (N->getNumValues() == 2) // Dead flag value?
+ return DCI.CombineTo(N, Cond, SDValue());
+ return Cond;
+ }
+ }
+ }
}
+ return SDValue();
+}
+
+
+/// PerformMulCombine - Optimize a single multiply with constant into two
+/// in order to implement it with two cheaper instructions, e.g.
+/// LEA + SHL, LEA + LEA.
+static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI) {
+ if (DAG.getMachineFunction().
+ getFunction()->hasFnAttr(Attribute::OptimizeForSize))
+ return SDValue();
+
+ if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
+ return SDValue();
+
+ MVT VT = N->getValueType(0);
+ if (VT != MVT::i64)
+ return SDValue();
+
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ if (!C)
+ return SDValue();
+ uint64_t MulAmt = C->getZExtValue();
+ if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
+ return SDValue();
+
+ uint64_t MulAmt1 = 0;
+ uint64_t MulAmt2 = 0;
+ if ((MulAmt % 9) == 0) {
+ MulAmt1 = 9;
+ MulAmt2 = MulAmt / 9;
+ } else if ((MulAmt % 5) == 0) {
+ MulAmt1 = 5;
+ MulAmt2 = MulAmt / 5;
+ } else if ((MulAmt % 3) == 0) {
+ MulAmt1 = 3;
+ MulAmt2 = MulAmt / 3;
+ }
+ if (MulAmt2 &&
+ (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
+ DebugLoc DL = N->getDebugLoc();
+
+ if (isPowerOf2_64(MulAmt2) &&
+ !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
+ // If second multiplifer is pow2, issue it first. We want the multiply by
+ // 3, 5, or 9 to be folded into the addressing mode unless the lone use
+ // is an add.
+ std::swap(MulAmt1, MulAmt2);
+
+ SDValue NewMul;
+ 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))
+ NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
+ DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
+ else
+ NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
+ DAG.getConstant(MulAmt2, VT));
+ // Do not add new nodes to DAG combiner worklist.
+ DCI.CombineTo(N, NewMul, false);
+ }
return SDValue();
}
+
/// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
/// when possible.
static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
// so we have no knowledge of the shift amount.
if (!Subtarget->hasSSE2())
return SDValue();
-
+
MVT VT = N->getValueType(0);
if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
return SDValue();
-
+
SDValue ShAmtOp = N->getOperand(1);
MVT EltVT = VT.getVectorElementType();
- DebugLoc dl = N->getDebugLoc();
+ DebugLoc DL = N->getDebugLoc();
SDValue BaseShAmt;
if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
unsigned NumElts = VT.getVectorNumElements();
}
}
} else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
- isSplatMask(ShAmtOp.getOperand(2).getNode())) {
- BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ShAmtOp,
- DAG.getIntPtrConstant(0));
+ cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
+ BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
+ DAG.getIntPtrConstant(0));
} else
return SDValue();
if (EltVT.bitsGT(MVT::i32))
- BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
+ 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::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
// The shift amount is identical so we can do a vector shift.
SDValue ValOp = N->getOperand(0);
break;
case ISD::SHL:
if (VT == MVT::v2i64)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
ValOp, BaseShAmt);
if (VT == MVT::v4i32)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
ValOp, BaseShAmt);
if (VT == MVT::v8i16)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
ValOp, BaseShAmt);
break;
case ISD::SRA:
if (VT == MVT::v4i32)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
ValOp, BaseShAmt);
if (VT == MVT::v8i16)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
ValOp, BaseShAmt);
break;
case ISD::SRL:
if (VT == MVT::v2i64)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
ValOp, BaseShAmt);
if (VT == MVT::v4i32)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
ValOp, BaseShAmt);
if (VT == MVT::v8i16)
- return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
ValOp, BaseShAmt);
break;
/// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
+ const X86Subtarget *Subtarget) {
// Turn load->store of MMX types into GPR load/stores. This avoids clobbering
// the FP state in cases where an emms may be missing.
// A preferable solution to the general problem is to figure out the right
// places to insert EMMS. This qualifies as a quick hack.
+
+ // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
StoreSDNode *St = cast<StoreSDNode>(N);
- if (St->getValue().getValueType().isVector() &&
- St->getValue().getValueType().getSizeInBits() == 64 &&
+ MVT VT = St->getValue().getValueType();
+ if (VT.getSizeInBits() != 64)
+ return SDValue();
+
+ const Function *F = DAG.getMachineFunction().getFunction();
+ bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
+ bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
+ && Subtarget->hasSSE2();
+ if ((VT.isVector() ||
+ (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
isa<LoadSDNode>(St->getValue()) &&
!cast<LoadSDNode>(St->getValue())->isVolatile() &&
St->getChain().hasOneUse() && !St->isVolatile()) {
Ops.push_back(ChainVal->getOperand(i));
}
}
- if (Ld) {
- DebugLoc dl = N->getDebugLoc();
- // If we are a 64-bit capable x86, lower to a single movq load/store pair.
- if (Subtarget->is64Bit()) {
- SDValue NewLd = DAG.getLoad(MVT::i64, dl, Ld->getChain(),
- Ld->getBasePtr(), Ld->getSrcValue(),
- Ld->getSrcValueOffset(), Ld->isVolatile(),
- Ld->getAlignment());
- SDValue NewChain = NewLd.getValue(1);
- if (TokenFactorIndex != -1) {
- Ops.push_back(NewChain);
- NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Ops[0],
- Ops.size());
- }
- return DAG.getStore(NewChain, dl, NewLd, St->getBasePtr(),
- St->getSrcValue(), St->getSrcValueOffset(),
- St->isVolatile(), St->getAlignment());
- }
- // Otherwise, lower to two 32-bit copies.
- SDValue LoAddr = Ld->getBasePtr();
- SDValue HiAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, LoAddr,
- DAG.getConstant(4, MVT::i32));
+ if (!Ld || !ISD::isNormalLoad(Ld))
+ return SDValue();
- SDValue LoLd = DAG.getLoad(MVT::i32, dl, Ld->getChain(), LoAddr,
- Ld->getSrcValue(), Ld->getSrcValueOffset(),
- Ld->isVolatile(), Ld->getAlignment());
- SDValue HiLd = DAG.getLoad(MVT::i32, dl, Ld->getChain(), HiAddr,
- Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
- Ld->isVolatile(),
- MinAlign(Ld->getAlignment(), 4));
+ // If this is not the MMX case, i.e. we are just turning i64 load/store
+ // into f64 load/store, avoid the transformation if there are multiple
+ // uses of the loaded value.
+ if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
+ return SDValue();
- SDValue NewChain = LoLd.getValue(1);
+ DebugLoc LdDL = Ld->getDebugLoc();
+ DebugLoc StDL = N->getDebugLoc();
+ // If we are a 64-bit capable x86, lower to a single movq load/store pair.
+ // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
+ // pair instead.
+ if (Subtarget->is64Bit() || F64IsLegal) {
+ MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
+ SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
+ Ld->getBasePtr(), Ld->getSrcValue(),
+ Ld->getSrcValueOffset(), Ld->isVolatile(),
+ Ld->getAlignment());
+ SDValue NewChain = NewLd.getValue(1);
if (TokenFactorIndex != -1) {
- Ops.push_back(LoLd);
- Ops.push_back(HiLd);
- NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Ops[0],
+ Ops.push_back(NewChain);
+ NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
Ops.size());
}
-
- LoAddr = St->getBasePtr();
- HiAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, LoAddr,
- DAG.getConstant(4, MVT::i32));
-
- SDValue LoSt = DAG.getStore(NewChain, dl, LoLd, LoAddr,
+ return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
St->getSrcValue(), St->getSrcValueOffset(),
St->isVolatile(), St->getAlignment());
- SDValue HiSt = DAG.getStore(NewChain, dl, HiLd, HiAddr,
- St->getSrcValue(),
- St->getSrcValueOffset() + 4,
- St->isVolatile(),
- MinAlign(St->getAlignment(), 4));
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoSt, HiSt);
}
+
+ // Otherwise, lower to two pairs of 32-bit loads / stores.
+ SDValue LoAddr = Ld->getBasePtr();
+ SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
+ DAG.getConstant(4, MVT::i32));
+
+ SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
+ Ld->getSrcValue(), Ld->getSrcValueOffset(),
+ Ld->isVolatile(), Ld->getAlignment());
+ SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
+ Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
+ Ld->isVolatile(),
+ MinAlign(Ld->getAlignment(), 4));
+
+ SDValue NewChain = LoLd.getValue(1);
+ if (TokenFactorIndex != -1) {
+ Ops.push_back(LoLd);
+ Ops.push_back(HiLd);
+ NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
+ Ops.size());
+ }
+
+ LoAddr = St->getBasePtr();
+ HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
+ DAG.getConstant(4, MVT::i32));
+
+ SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
+ St->getSrcValue(), St->getSrcValueOffset(),
+ St->isVolatile(), St->getAlignment());
+ SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
+ St->getSrcValue(),
+ St->getSrcValueOffset() + 4,
+ St->isVolatile(),
+ MinAlign(St->getAlignment(), 4));
+ return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
}
return SDValue();
}
return SDValue();
}
+static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue Op = N->getOperand(0);
+ if (Op.getOpcode() == ISD::BIT_CONVERT)
+ Op = Op.getOperand(0);
+ MVT VT = N->getValueType(0), OpVT = Op.getValueType();
+ if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
+ VT.getVectorElementType().getSizeInBits() ==
+ OpVT.getVectorElementType().getSizeInBits()) {
+ return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
+ }
+ return SDValue();
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
switch (N->getOpcode()) {
default: break;
case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
- case ISD::BUILD_VECTOR:
- return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
+ case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
+ case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
case X86ISD::FOR: return PerformFORCombine(N, DAG);
case X86ISD::FAND: return PerformFANDCombine(N, DAG);
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
+ case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
}
return SDValue();
if (Subtarget->hasSSE1())
return "x";
}
-
+
return TargetLowering::LowerXConstraint(ConstraintVT);
}
std::vector<SDValue>&Ops,
SelectionDAG &DAG) const {
SDValue Result(0, 0);
-
+
switch (Constraint) {
default: break;
case 'I':
}
}
return;
+ case 'K':
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+ if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
+ Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
+ break;
+ }
+ }
+ return;
case 'N':
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
if (C->getZExtValue() <= 255) {
// If we are in non-pic codegen mode, we allow the address of a global (with
// an optional displacement) to be used with 'i'.
- GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
+ GlobalAddressSDNode *GA = 0;
int64_t Offset = 0;
-
- // Match either (GA) or (GA+C)
- if (GA) {
- Offset = GA->getOffset();
- } else if (Op.getOpcode() == ISD::ADD) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
- GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
- if (C && GA) {
- Offset = GA->getOffset()+C->getZExtValue();
- } else {
- C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
- GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
- if (C && GA)
- Offset = GA->getOffset()+C->getZExtValue();
- else
- C = 0, GA = 0;
+
+ // Match either (GA), (GA+C), (GA+C1+C2), etc.
+ while (1) {
+ if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
+ Offset += GA->getOffset();
+ break;
+ } else if (Op.getOpcode() == ISD::ADD) {
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ Offset += C->getZExtValue();
+ Op = Op.getOperand(0);
+ continue;
+ }
+ } else if (Op.getOpcode() == ISD::SUB) {
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ Offset += -C->getZExtValue();
+ Op = Op.getOperand(0);
+ continue;
+ }
}
- }
-
- if (GA) {
- if (hasMemory)
- Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
- Offset, DAG);
- else
- Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
- Offset);
- Result = Op;
- break;
+
+ // Otherwise, this isn't something we can handle, reject it.
+ return;
}
- // Otherwise, not valid for this mode.
- return;
+ if (hasMemory)
+ Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
+ else
+ Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
+ Offset);
+ Result = Op;
+ break;
}
}
-
+
if (Result.getNode()) {
Ops.push_back(Result);
return;
if (VT == MVT::i16)
return std::make_pair(0U, X86::GR16RegisterClass);
if (VT == MVT::i32 || !Subtarget->is64Bit())
- return std::make_pair(0U, X86::GR32RegisterClass);
+ return std::make_pair(0U, X86::GR32RegisterClass);
return std::make_pair(0U, X86::GR64RegisterClass);
case 'f': // FP Stack registers.
// If SSE is enabled for this VT, use f80 to ensure the isel moves the
break;
}
}
-
+
// Use the default implementation in TargetLowering to convert the register
// constraint into a member of a register class.
std::pair<unsigned, const TargetRegisterClass*> Res;
}
if (DestReg) {
Res.first = DestReg;
- Res.second = Res.second = X86::GR8RegisterClass;
+ Res.second = X86::GR8RegisterClass;
}
} else if (VT == MVT::i32) {
unsigned DestReg = 0;
}
if (DestReg) {
Res.first = DestReg;
- Res.second = Res.second = X86::GR32RegisterClass;
+ Res.second = X86::GR32RegisterClass;
}
} else if (VT == MVT::i64) {
unsigned DestReg = 0;
}
if (DestReg) {
Res.first = DestReg;
- Res.second = Res.second = X86::GR64RegisterClass;
+ Res.second = X86::GR64RegisterClass;
}
}
} else if (Res.second == X86::FR32RegisterClass ||
assert(VT.isVector());
if (isTypeLegal(VT))
return VT;
-
+
// TODO: In computeRegisterProperty, we can compute the list of legal vector
// type based on element type. This would speed up our search (though
// it may not be worth it since the size of the list is relatively
// small).
MVT EltVT = VT.getVectorElementType();
unsigned NElts = VT.getVectorNumElements();
-
+
// On X86, it make sense to widen any vector wider than 1
if (NElts <= 1)
return MVT::Other;
-
- for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
+
+ for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
MVT SVT = (MVT::SimpleValueType)nVT;
-
- if (isTypeLegal(SVT) &&
- SVT.getVectorElementType() == EltVT &&
+
+ if (isTypeLegal(SVT) &&
+ SVT.getVectorElementType() == EltVT &&
SVT.getVectorNumElements() > NElts)
return SVT;
}