#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/StringExtras.h"
+#include "llvm/ParameterAttributes.h"
using namespace llvm;
X86TargetLowering::X86TargetLowering(TargetMachine &TM)
: TargetLowering(TM) {
Subtarget = &TM.getSubtarget<X86Subtarget>();
- X86ScalarSSE = Subtarget->hasSSE2();
+ X86ScalarSSEf64 = Subtarget->hasSSE2();
+ X86ScalarSSEf32 = Subtarget->hasSSE1();
X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
+ RegInfo = TM.getRegisterInfo();
+
// Set up the TargetLowering object.
// X86 is weird, it always uses i8 for shift amounts and setcc results.
setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
} else {
- if (X86ScalarSSE)
+ if (X86ScalarSSEf64)
// If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
else
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 (X86ScalarSSE)
+ if (X86ScalarSSEf32) {
setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
- else {
+ // 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);
}
- if (!Subtarget->is64Bit()) {
- // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
- setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
- setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
- }
+ // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
+ // are Legal, f80 is custom lowered.
+ setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
+ setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
// Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
// this operation.
setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
- if (X86ScalarSSE) {
+ if (X86ScalarSSEf32) {
setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
+ // f32 and f64 cases are Legal, f80 case is not
+ setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
} else {
setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
} else {
- if (X86ScalarSSE && !Subtarget->hasSSE3())
+ 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.
}
// TODO: when we have SSE, these could be more efficient, by using movd/movq.
- if (!X86ScalarSSE) {
+ if (!X86ScalarSSEf64) {
setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
}
+ // Divide and remainder are lowered to use div or idiv in legalize in
+ // order to expose the intermediate computations to trivial CSE. This is
+ // most noticeable when both x/y and x%y are being computed; they can be
+ // done with a single div or idiv.
+ setOperationAction(ISD::SDIV , MVT::i8 , Custom);
+ setOperationAction(ISD::UDIV , MVT::i8 , Custom);
+ setOperationAction(ISD::SREM , MVT::i8 , Custom);
+ setOperationAction(ISD::UREM , MVT::i8 , Custom);
+ setOperationAction(ISD::SDIV , MVT::i16 , Custom);
+ setOperationAction(ISD::UDIV , MVT::i16 , Custom);
+ setOperationAction(ISD::SREM , MVT::i16 , Custom);
+ setOperationAction(ISD::UREM , MVT::i16 , Custom);
+ setOperationAction(ISD::SDIV , MVT::i32 , Custom);
+ setOperationAction(ISD::UDIV , MVT::i32 , Custom);
+ setOperationAction(ISD::SREM , MVT::i32 , Custom);
+ setOperationAction(ISD::UREM , MVT::i32 , Custom);
+ setOperationAction(ISD::SDIV , MVT::i64 , Custom);
+ setOperationAction(ISD::UDIV , MVT::i64 , Custom);
+ setOperationAction(ISD::SREM , MVT::i64 , Custom);
+ setOperationAction(ISD::UREM , MVT::i64 , Custom);
+
setOperationAction(ISD::BR_JT , MVT::Other, Expand);
setOperationAction(ISD::BRCOND , MVT::Other, Custom);
setOperationAction(ISD::BR_CC , MVT::Other, Expand);
setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
if (Subtarget->is64Bit())
- setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
- setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
- setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
setOperationAction(ISD::FREM , MVT::f64 , Expand);
setOperationAction(ISD::SELECT , MVT::i32 , Custom);
setOperationAction(ISD::SELECT , MVT::f32 , Custom);
setOperationAction(ISD::SELECT , MVT::f64 , Custom);
+ setOperationAction(ISD::SELECT , MVT::f80 , Custom);
setOperationAction(ISD::SETCC , MVT::i8 , Custom);
setOperationAction(ISD::SETCC , MVT::i16 , Custom);
setOperationAction(ISD::SETCC , MVT::i32 , Custom);
setOperationAction(ISD::SETCC , MVT::f32 , Custom);
setOperationAction(ISD::SETCC , MVT::f64 , Custom);
+ setOperationAction(ISD::SETCC , MVT::f80 , Custom);
if (Subtarget->is64Bit()) {
setOperationAction(ISD::SELECT , MVT::i64 , Custom);
setOperationAction(ISD::SETCC , MVT::i64 , Custom);
}
// X86 ret instruction may pop stack.
setOperationAction(ISD::RET , MVT::Other, Custom);
+ if (!Subtarget->is64Bit())
+ setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
+
// Darwin ABI issue.
setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
setOperationAction(ISD::MEMSET , MVT::Other, Custom);
setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
- // We don't have line number support yet.
+ // Use the default ISD::LOCATION expansion.
setOperationAction(ISD::LOCATION, MVT::Other, Expand);
- setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
// FIXME - use subtarget debug flags
if (!Subtarget->isTargetDarwin() &&
!Subtarget->isTargetELF() &&
setExceptionPointerRegister(X86::EAX);
setExceptionSelectorRegister(X86::EDX);
}
+ setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
+ setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
+
// VASTART needs to be custom lowered to use the VarArgsFrameIndex
setOperationAction(ISD::VASTART , MVT::Other, Custom);
setOperationAction(ISD::VAARG , MVT::Other, Expand);
else
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
- if (X86ScalarSSE) {
+ if (X86ScalarSSEf64) {
+ // f32 and f64 use SSE.
// Set up the FP register classes.
addRegisterClass(MVT::f32, X86::FR32RegisterClass);
addRegisterClass(MVT::f64, X86::FR64RegisterClass);
// cases we handle.
setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
- addLegalFPImmediate(+0.0); // xorps / xorpd
+ addLegalFPImmediate(APFloat(+0.0)); // xorpd
+ addLegalFPImmediate(APFloat(+0.0f)); // xorps
+
+ // Conversions to long double (in X87) go through memory.
+ setConvertAction(MVT::f32, MVT::f80, Expand);
+ setConvertAction(MVT::f64, MVT::f80, Expand);
+
+ // Conversions from long double (in X87) go through memory.
+ setConvertAction(MVT::f80, MVT::f32, Expand);
+ setConvertAction(MVT::f80, MVT::f64, Expand);
+ } else if (X86ScalarSSEf32) {
+ // Use SSE for f32, x87 for f64.
+ // Set up the FP register classes.
+ addRegisterClass(MVT::f32, X86::FR32RegisterClass);
+ addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
+
+ // Use ANDPS to simulate FABS.
+ setOperationAction(ISD::FABS , MVT::f32, Custom);
+
+ // Use XORP to simulate FNEG.
+ setOperationAction(ISD::FNEG , MVT::f32, Custom);
+
+ setOperationAction(ISD::UNDEF, MVT::f64, Expand);
+
+ // Use ANDPS and ORPS to simulate FCOPYSIGN.
+ setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
+
+ // We don't support sin/cos/fmod
+ setOperationAction(ISD::FSIN , MVT::f32, Expand);
+ setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FREM , MVT::f32, Expand);
+
+ // Expand FP immediates into loads from the stack, except for the special
+ // cases we handle.
+ setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
+ setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
+ addLegalFPImmediate(APFloat(+0.0f)); // xorps
+ addLegalFPImmediate(APFloat(+0.0)); // FLD0
+ addLegalFPImmediate(APFloat(+1.0)); // FLD1
+ addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
+ addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
+
+ // SSE->x87 conversions go through memory.
+ setConvertAction(MVT::f32, MVT::f64, Expand);
+ setConvertAction(MVT::f32, MVT::f80, Expand);
+
+ // x87->SSE truncations need to go through memory.
+ 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);
+ }
} else {
+ // f32 and f64 in x87.
// Set up the FP register classes.
- addRegisterClass(MVT::f64, X86::RFPRegisterClass);
+ addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
+ addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
setOperationAction(ISD::UNDEF, MVT::f64, Expand);
+ setOperationAction(ISD::UNDEF, MVT::f32, Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
+ // Floating truncations need to go through memory.
+ 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);
}
setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
- addLegalFPImmediate(+0.0); // FLD0
- addLegalFPImmediate(+1.0); // FLD1
- addLegalFPImmediate(-0.0); // FLD0/FCHS
- addLegalFPImmediate(-1.0); // FLD1/FCHS
+ setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
+ addLegalFPImmediate(APFloat(+0.0)); // FLD0
+ addLegalFPImmediate(APFloat(+1.0)); // FLD1
+ addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
+ addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
+ addLegalFPImmediate(APFloat(+0.0f)); // FLD0
+ addLegalFPImmediate(APFloat(+1.0f)); // FLD1
+ addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
+ addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
+ }
+
+ // Long double always uses X87.
+ addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
+ setOperationAction(ISD::UNDEF, MVT::f80, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
+ setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
+ if (!UnsafeFPMath) {
+ setOperationAction(ISD::FSIN , MVT::f80 , Expand);
+ setOperationAction(ISD::FCOS , MVT::f80 , Expand);
}
// First set operation action for all vector types to expand. Then we
setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FABS, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FSIN, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FCOS, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FREM, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FPOWI, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FSQRT, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FCOPYSIGN, (MVT::ValueType)VT, Expand);
}
if (Subtarget->hasMMX()) {
setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
+ setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
+ setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
+ setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
+ setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
SmallVector<CCValAssign, 16> RVLocs;
unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
- CCState CCInfo(CC, getTargetMachine(), RVLocs);
+ bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
+ CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
// If this is an FP return with ScalarSSE, we need to move the value from
// an XMM register onto the fp-stack.
- if (X86ScalarSSE) {
+ if ((X86ScalarSSEf32 && RVLocs[0].getValVT()==MVT::f32) ||
+ (X86ScalarSSEf64 && RVLocs[0].getValVT()==MVT::f64)) {
SDOperand MemLoc;
// If this is a load into a scalarsse value, don't store the loaded value
MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
}
- SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
+ SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other);
SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
Chain = Value.getValue(1);
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
+ bool isVarArg = cast<ConstantSDNode>(TheCall->getOperand(2))->getValue() != 0;
+ CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
// before the fp stackifier runs.
// Copy ST0 into an RFP register with FP_GET_RESULT.
- SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(RVLocs[0].getValVT(), MVT::Other, MVT::Flag);
SDOperand GROps[] = { Chain, InFlag };
SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
Chain = RetVal.getValue(1);
// If we are using ScalarSSE, store ST(0) to the stack and reload it into
// an XMM register.
- if (X86ScalarSSE) {
+ if ((X86ScalarSSEf32 && RVLocs[0].getValVT() == MVT::f32) ||
+ (X86ScalarSSEf64 && RVLocs[0].getValVT() == MVT::f64)) {
// FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
// shouldn't be necessary except that RFP cannot be live across
// multiple blocks. When stackifier is fixed, they can be uncoupled.
RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
Chain = RetVal.getValue(1);
}
-
- if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
- // FIXME: we would really like to remember that this FP_ROUND
- // operation is okay to eliminate if we allow excess FP precision.
- RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
ResultVals.push_back(RetVal);
}
return VReg;
}
+SDOperand X86TargetLowering::LowerMemArgument(SDOperand Op, SelectionDAG &DAG,
+ const CCValAssign &VA,
+ MachineFrameInfo *MFI,
+ SDOperand Root, unsigned i) {
+ // Create the nodes corresponding to a load from this parameter slot.
+ int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
+ VA.getLocMemOffset());
+ SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
+
+ unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3 + i))->getValue();
+
+ if (Flags & ISD::ParamFlags::ByVal)
+ return FIN;
+ else
+ return DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0);
+}
+
SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
bool isStdCall) {
unsigned NumArgs = Op.Val->getNumValues() - 1;
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
- ArgLocs);
+ CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
+ getTargetMachine(), ArgLocs);
CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
SmallVector<SDOperand, 8> ArgValues;
ArgValues.push_back(ArgValue);
} else {
assert(VA.isMemLoc());
-
- // Create the nodes corresponding to a load from this parameter slot.
- int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
- VA.getLocMemOffset());
- SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
- ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
+ ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
}
}
BytesCallerReserves = StackSize;
}
-
+
RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
- ReturnAddrIndex = 0; // No return address slot generated yet.
- MF.getInfo<X86MachineFunctionInfo>()
- ->setBytesToPopOnReturn(BytesToPopOnReturn);
+ X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+ FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
// Return the new list of results.
return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, getTargetMachine(), ArgLocs);
+ CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
// Get a count of how many bytes are to be pushed on the stack.
assert(VA.isMemLoc());
if (StackPtr.Val == 0)
StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
- SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
- MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
+
+ MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
+ Arg));
}
}
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
SDOperand Root = Op.getOperand(0);
+ bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
- ArgLocs);
+ CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
+ getTargetMachine(), ArgLocs);
CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
SmallVector<SDOperand, 8> ArgValues;
ArgValues.push_back(ArgValue);
} else {
assert(VA.isMemLoc());
-
- // Create the nodes corresponding to a load from this parameter slot.
- int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
- VA.getLocMemOffset());
- SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
- ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
+ ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
}
}
VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
- ReturnAddrIndex = 0; // No return address slot generated yet.
BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
BytesCallerReserves = 0;
- MF.getInfo<X86MachineFunctionInfo>()
- ->setBytesToPopOnReturn(BytesToPopOnReturn);
+ X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+ FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
// Return the new list of results.
return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
&ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
}
+SDOperand
+X86TargetLowering::LowerMemOpCallTo(SDOperand Op, SelectionDAG &DAG,
+ const SDOperand &StackPtr,
+ const CCValAssign &VA,
+ SDOperand Chain,
+ SDOperand Arg) {
+ SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
+ PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
+ SDOperand FlagsOp = Op.getOperand(6+2*VA.getValNo());
+ unsigned Flags = cast<ConstantSDNode>(FlagsOp)->getValue();
+ if (Flags & ISD::ParamFlags::ByVal) {
+ unsigned Align = 1 << ((Flags & ISD::ParamFlags::ByValAlign) >>
+ ISD::ParamFlags::ByValAlignOffs);
+
+ unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
+ ISD::ParamFlags::ByValSizeOffs;
+
+ SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
+ SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
+
+ return DAG.getNode(ISD::MEMCPY, MVT::Other, Chain, PtrOff, Arg, SizeNode,
+ AlignNode);
+ } else {
+ return DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
+ }
+}
+
SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
unsigned CC) {
SDOperand Chain = Op.getOperand(0);
bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
+ bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
SDOperand Callee = Op.getOperand(4);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, getTargetMachine(), ArgLocs);
+ CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
// Get a count of how many bytes are to be pushed on the stack.
assert(VA.isMemLoc());
if (StackPtr.Val == 0)
StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
- SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
- MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
+
+ MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
+ Arg));
}
}
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
- ArgLocs);
+ CCState CCInfo(MF.getFunction()->getCallingConv(), isVarArg,
+ getTargetMachine(), ArgLocs);
CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
SmallVector<SDOperand, 8> ArgValues;
ArgValues.push_back(ArgValue);
} else {
assert(VA.isMemLoc());
-
- // Create the nodes corresponding to a load from this parameter slot.
- int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
- VA.getLocMemOffset());
- SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
- ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
+ ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, Root, i));
}
}
ArgValues.push_back(Root);
- ReturnAddrIndex = 0; // No return address slot generated yet.
BytesToPopOnReturn = 0; // Callee pops nothing.
BytesCallerReserves = StackSize;
+ X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+ FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
+
// Return the new list of results.
return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
&ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, getTargetMachine(), ArgLocs);
+ CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
// Get a count of how many bytes are to be pushed on the stack.
assert(VA.isMemLoc());
if (StackPtr.Val == 0)
StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
- SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
- MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
+
+ MemOpChains.push_back(LowerMemOpCallTo(Op, DAG, StackPtr, VA, Chain,
+ Arg));
}
}
SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+ int ReturnAddrIndex = FuncInfo->getRAIndex();
+
if (ReturnAddrIndex == 0) {
// Set up a frame object for the return address.
- MachineFunction &MF = DAG.getMachineFunction();
if (Subtarget->is64Bit())
ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
else
ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
+
+ FuncInfo->setRAIndex(ReturnAddrIndex);
}
return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
// X < 0 -> X == 0, jump on sign.
X86CC = X86::COND_S;
return true;
+ } else if (SetCCOpcode == ISD::SETLT && RHSC->getValue() == 1) {
+ // X < 1 -> X <= 0
+ RHS = DAG.getConstant(0, RHS.getValueType());
+ X86CC = X86::COND_LE;
+ return true;
}
}
bool X86::isPSHUFDMask(SDNode *N) {
assert(N->getOpcode() == ISD::BUILD_VECTOR);
- if (N->getNumOperands() != 4)
+ if (N->getNumOperands() != 2 && N->getNumOperands() != 4)
return false;
// Check if the value doesn't reference the second vector.
SDOperand Arg = N->getOperand(i);
if (Arg.getOpcode() == ISD::UNDEF) continue;
assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
+ if (cast<ConstantSDNode>(Arg)->getValue() >= e)
return false;
}
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)))
+ 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) {
return ((isa<ConstantSDNode>(Elt) &&
cast<ConstantSDNode>(Elt)->getValue() == 0) ||
(isa<ConstantFPSDNode>(Elt) &&
- cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
+ cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
}
/// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
unsigned NumZero = 0;
unsigned NumNonZero = 0;
unsigned NonZeros = 0;
+ unsigned NumNonZeroImms = 0;
std::set<SDOperand> Values;
for (unsigned i = 0; i < NumElems; ++i) {
SDOperand Elt = Op.getOperand(i);
else {
NonZeros |= (1 << i);
NumNonZero++;
+ if (Elt.getOpcode() == ISD::Constant ||
+ Elt.getOpcode() == ISD::ConstantFP)
+ NumNonZeroImms++;
}
}
}
- if (NumNonZero == 0)
- // Must be a mix of zero and undef. Return a zero vector.
- return getZeroVector(VT, DAG);
+ if (NumNonZero == 0) {
+ if (NumZero == 0)
+ // All undef vector. Return an UNDEF.
+ return DAG.getNode(ISD::UNDEF, VT);
+ else
+ // A mix of zero and undef. Return a zero vector.
+ return getZeroVector(VT, DAG);
+ }
// Splat is obviously ok. Let legalizer expand it to a shuffle.
if (Values.size() == 1)
}
}
+ // A vector full of immediates; various special cases are already
+ // handled, so this is best done with a single constant-pool load.
+ if (NumNonZero == NumNonZeroImms)
+ return SDOperand();
+
// Let legalizer expand 2-wide build_vectors.
if (EVTBits == 64)
return SDOperand();
if (isZeroShuffle(Op.Val))
return getZeroVector(VT, DAG);
+ if (isIdentityMask(PermMask.Val))
+ return V1;
+ else if (isIdentityMask(PermMask.Val, true))
+ return V2;
+
if (isSplatMask(PermMask.Val)) {
if (NumElems <= 4) return Op;
// Promote it to a v4i32 splat.
// If VT is integer, try PSHUF* first, then SHUFP*.
if (MVT::isInteger(VT)) {
- if (X86::isPSHUFDMask(PermMask.Val) ||
+ // MMX doesn't have PSHUFD; it does have PSHUFW. While it's theoretically
+ // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
+ if (((MVT::getSizeInBits(VT) != 64 || NumElems == 4) &&
+ X86::isPSHUFDMask(PermMask.Val)) ||
X86::isPSHUFHWMask(PermMask.Val) ||
X86::isPSHUFLWMask(PermMask.Val)) {
if (V2.getOpcode() != ISD::UNDEF)
if (N1.getValueType() != MVT::i32)
N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
if (N2.getValueType() != MVT::i32)
- N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
+ N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
} else if (MVT::getSizeInBits(BaseVT) == 32) {
unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
// Use two pinsrw instructions to insert a 32 bit value.
Idx <<= 1;
if (MVT::isFloatingPoint(N1.getValueType())) {
- if (ISD::isNON_EXTLoad(N1.Val)) {
- // Just load directly from f32mem to GR32.
- LoadSDNode *LD = cast<LoadSDNode>(N1);
- N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset());
- } else {
- N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
- N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
- N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
- DAG.getConstant(0, getPointerTy()));
- }
+ N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
+ N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
+ N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
+ DAG.getConstant(0, getPointerTy()));
}
N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
DAG.getConstant(32, MVT::i8));
SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
- SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
+ SDOperand Cond = NewCCModeling
+ ? DAG.getNode(X86ISD::CMP_NEW, MVT::i32,
+ AndNode, DAG.getConstant(0, MVT::i8))
+ : DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
SDOperand Hi, Lo;
SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
-
+ unsigned Opc = NewCCModeling ? X86ISD::CMOV_NEW : X86ISD::CMOV;
VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
SmallVector<SDOperand, 4> Ops;
if (Op.getOpcode() == ISD::SHL_PARTS) {
Ops.push_back(Tmp2);
Ops.push_back(Tmp3);
Ops.push_back(CC);
- Ops.push_back(InFlag);
- Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
- InFlag = Hi.getValue(1);
+ Ops.push_back(Cond);
+ if (NewCCModeling)
+ Hi = DAG.getNode(Opc, MVT::i32, &Ops[0], Ops.size());
+ else {
+ Hi = DAG.getNode(Opc, VTs, 2, &Ops[0], Ops.size());
+ Cond = Hi.getValue(1);
+ }
Ops.clear();
Ops.push_back(Tmp3);
Ops.push_back(Tmp1);
Ops.push_back(CC);
- Ops.push_back(InFlag);
- Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
+ Ops.push_back(Cond);
+ if (NewCCModeling)
+ Lo = DAG.getNode(Opc, MVT::i32, &Ops[0], Ops.size());
+ else
+ Lo = DAG.getNode(Opc, VTs, 2, &Ops[0], Ops.size());
} else {
Ops.push_back(Tmp2);
Ops.push_back(Tmp3);
Ops.push_back(CC);
- Ops.push_back(InFlag);
- Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
- InFlag = Lo.getValue(1);
+ Ops.push_back(Cond);
+ if (NewCCModeling)
+ Lo = DAG.getNode(Opc, MVT::i32, &Ops[0], Ops.size());
+ else {
+ Lo = DAG.getNode(Opc, VTs, 2, &Ops[0], Ops.size());
+ Cond = Lo.getValue(1);
+ }
Ops.clear();
Ops.push_back(Tmp3);
Ops.push_back(Tmp1);
Ops.push_back(CC);
- Ops.push_back(InFlag);
- Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
+ Ops.push_back(Cond);
+ if (NewCCModeling)
+ Hi = DAG.getNode(Opc, MVT::i32, &Ops[0], Ops.size());
+ else
+ Hi = DAG.getNode(Opc, VTs, 2, &Ops[0], Ops.size());
}
VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
}
+SDOperand X86TargetLowering::LowerIntegerDivOrRem(SDOperand Op, SelectionDAG &DAG) {
+ unsigned Opcode = Op.getOpcode();
+ MVT::ValueType NVT = Op.getValueType();
+ bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
+ bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
+ unsigned Opc = isSigned ? X86ISD::IDIV : X86ISD::DIV;
+
+ SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
+ SDOperand DR = DAG.getNode(Opc, DAG.getVTList(NVT, NVT), Ops, 2);
+
+ if (isDiv)
+ return DR;
+
+ return SDOperand(DR.Val, 1);
+}
+
SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
Op.getOperand(0).getValueType() >= MVT::i16 &&
SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
StackSlot, NULL, 0);
+ // These are really Legal; caller falls through into that case.
+ if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f32 && X86ScalarSSEf32)
+ return Result;
+ if (SrcVT==MVT::i32 && Op.getValueType() == MVT::f64 && X86ScalarSSEf64)
+ return Result;
+ if (SrcVT==MVT::i64 && Op.getValueType() != MVT::f80 &&
+ Subtarget->is64Bit())
+ return Result;
+
// Build the FILD
SDVTList Tys;
- if (X86ScalarSSE)
+ bool useSSE = (X86ScalarSSEf32 && Op.getValueType() == MVT::f32) ||
+ (X86ScalarSSEf64 && Op.getValueType() == MVT::f64);
+ if (useSSE)
Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
else
- Tys = DAG.getVTList(MVT::f64, MVT::Other);
+ Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
SmallVector<SDOperand, 8> Ops;
Ops.push_back(Chain);
Ops.push_back(StackSlot);
Ops.push_back(DAG.getValueType(SrcVT));
- Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
+ Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
Tys, &Ops[0], Ops.size());
- if (X86ScalarSSE) {
+ if (useSSE) {
Chain = Result.getValue(1);
SDOperand InFlag = Result.getValue(2);
"Unknown FP_TO_SINT to lower!");
// We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
// stack slot.
+ SDOperand Result;
MachineFunction &MF = DAG.getMachineFunction();
unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ // These are really Legal.
+ if (Op.getValueType() == MVT::i32 &&
+ X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32)
+ return Result;
+ if (Op.getValueType() == MVT::i32 &&
+ X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)
+ return Result;
+ if (Subtarget->is64Bit() &&
+ Op.getValueType() == MVT::i64 &&
+ Op.getOperand(0).getValueType() != MVT::f80)
+ return Result;
+
unsigned Opc;
switch (Op.getValueType()) {
default: assert(0 && "Invalid FP_TO_SINT to lower!");
SDOperand Chain = DAG.getEntryNode();
SDOperand Value = Op.getOperand(0);
- if (X86ScalarSSE) {
+ if ((X86ScalarSSEf32 && Op.getOperand(0).getValueType() == MVT::f32) ||
+ (X86ScalarSSEf64 && Op.getOperand(0).getValueType() == MVT::f64)) {
assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
- SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
+ SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
SDOperand Ops[] = {
Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
};
SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
MVT::ValueType VT = Op.getValueType();
- const Type *OpNTy = MVT::getTypeForValueType(VT);
+ MVT::ValueType EltVT = VT;
+ if (MVT::isVector(VT))
+ EltVT = MVT::getVectorElementType(VT);
+ const Type *OpNTy = MVT::getTypeForValueType(EltVT);
std::vector<Constant*> CV;
- if (VT == MVT::f64) {
- CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
+ if (EltVT == MVT::f64) {
+ Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, ~(1ULL << 63))));
+ CV.push_back(C);
+ CV.push_back(C);
} else {
- CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- }
- Constant *CS = ConstantStruct::get(CV);
- SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- SDVTList Tys = DAG.getVTList(VT, MVT::Other);
- SmallVector<SDOperand, 3> Ops;
- Ops.push_back(DAG.getEntryNode());
- Ops.push_back(CPIdx);
- Ops.push_back(DAG.getSrcValue(NULL));
- SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
+ Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, ~(1U << 31))));
+ CV.push_back(C);
+ CV.push_back(C);
+ CV.push_back(C);
+ CV.push_back(C);
+ }
+ Constant *C = ConstantVector::get(CV);
+ SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
+ false, 16);
return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
}
SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
MVT::ValueType VT = Op.getValueType();
- const Type *OpNTy = MVT::getTypeForValueType(VT);
+ MVT::ValueType EltVT = VT;
+ unsigned EltNum = 1;
+ if (MVT::isVector(VT)) {
+ EltVT = MVT::getVectorElementType(VT);
+ EltNum = MVT::getVectorNumElements(VT);
+ }
+ const Type *OpNTy = MVT::getTypeForValueType(EltVT);
std::vector<Constant*> CV;
- if (VT == MVT::f64) {
- CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
+ if (EltVT == MVT::f64) {
+ Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(64, 1ULL << 63)));
+ CV.push_back(C);
+ CV.push_back(C);
+ } else {
+ Constant *C = ConstantFP::get(OpNTy, APFloat(APInt(32, 1U << 31)));
+ CV.push_back(C);
+ CV.push_back(C);
+ CV.push_back(C);
+ CV.push_back(C);
+ }
+ Constant *C = ConstantVector::get(CV);
+ SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDOperand Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
+ false, 16);
+ if (MVT::isVector(VT)) {
+ return DAG.getNode(ISD::BIT_CONVERT, VT,
+ DAG.getNode(ISD::XOR, MVT::v2i64,
+ DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
+ DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
} else {
- CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- CV.push_back(ConstantFP::get(OpNTy, 0.0));
- }
- Constant *CS = ConstantStruct::get(CV);
- SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- SDVTList Tys = DAG.getVTList(VT, MVT::Other);
- SmallVector<SDOperand, 3> Ops;
- Ops.push_back(DAG.getEntryNode());
- Ops.push_back(CPIdx);
- Ops.push_back(DAG.getSrcValue(NULL));
- SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
- return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
+ return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
+ }
}
SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
SrcVT = VT;
+ SrcTy = MVT::getTypeForValueType(SrcVT);
}
// First get the sign bit of second operand.
std::vector<Constant*> CV;
if (SrcVT == MVT::f64) {
- CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 1ULL << 63))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
} else {
- CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- }
- Constant *CS = ConstantStruct::get(CV);
- SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
- SmallVector<SDOperand, 3> Ops;
- Ops.push_back(DAG.getEntryNode());
- Ops.push_back(CPIdx);
- Ops.push_back(DAG.getSrcValue(NULL));
- SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 1U << 31))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ }
+ Constant *C = ConstantVector::get(CV);
+ SDOperand CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDOperand Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx, NULL, 0,
+ false, 16);
SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
// Shift sign bit right or left if the two operands have different types.
// Clear first operand sign bit.
CV.clear();
if (VT == MVT::f64) {
- CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, ~(1ULL << 63)))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(64, 0))));
} else {
- CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- CV.push_back(ConstantFP::get(SrcTy, 0.0));
- }
- CS = ConstantStruct::get(CV);
- CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- Tys = DAG.getVTList(VT, MVT::Other);
- Ops.clear();
- Ops.push_back(DAG.getEntryNode());
- Ops.push_back(CPIdx);
- Ops.push_back(DAG.getSrcValue(NULL));
- SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, ~(1U << 31)))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ CV.push_back(ConstantFP::get(SrcTy, APFloat(APInt(32, 0))));
+ }
+ C = ConstantVector::get(CV);
+ CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
+ SDOperand Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx, NULL, 0,
+ false, 16);
SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
// Or the value with the sign bit.
}
}
+SDOperand X86TargetLowering::LowerSETCC_New(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
+ SDOperand Cond;
+ SDOperand Op0 = Op.getOperand(0);
+ SDOperand Op1 = Op.getOperand(1);
+ SDOperand CC = Op.getOperand(2);
+ ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
+ bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
+ unsigned X86CC;
+
+ if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
+ Op0, Op1, DAG)) {
+ Cond = DAG.getNode(X86ISD::CMP_NEW, MVT::i32, Op0, Op1);
+ return DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond);
+ }
+
+ assert(isFP && "Illegal integer SetCC!");
+
+ Cond = DAG.getNode(X86ISD::CMP_NEW, MVT::i32, Op0, Op1);
+ switch (SetCCOpcode) {
+ default: assert(false && "Illegal floating point SetCC!");
+ case ISD::SETOEQ: { // !PF & ZF
+ SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
+ SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86::COND_E, MVT::i8), Cond);
+ return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
+ }
+ case ISD::SETUNE: { // PF | !ZF
+ SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86::COND_P, MVT::i8), Cond);
+ SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
+ return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
+ }
+ }
+}
+
+
SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
bool addTest = true;
SDOperand Chain = DAG.getEntryNode();
// pressure reason)?
SDOperand Cmp = Cond.getOperand(1);
unsigned Opc = Cmp.getOpcode();
- bool IllegalFPCMov = !X86ScalarSSE &&
- MVT::isFloatingPoint(Op.getValueType()) &&
+ bool IllegalFPCMov =
+ ! ((X86ScalarSSEf32 && Op.getValueType()==MVT::f32) ||
+ (X86ScalarSSEf64 && Op.getValueType()==MVT::f64)) &&
!hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
!IllegalFPCMov) {
return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
}
+SDOperand X86TargetLowering::LowerSELECT_New(SDOperand Op, SelectionDAG &DAG) {
+ bool addTest = true;
+ SDOperand Cond = Op.getOperand(0);
+ SDOperand CC;
+
+ if (Cond.getOpcode() == ISD::SETCC)
+ Cond = LowerSETCC_New(Cond, DAG);
+
+ if (Cond.getOpcode() == X86ISD::SETCC_NEW) {
+ CC = Cond.getOperand(0);
+
+ // If condition flag is set by a X86ISD::CMP, then make a copy of it
+ // (since flag operand cannot be shared). Use it as the condition setting
+ // operand in place of the X86ISD::SETCC.
+ // If the X86ISD::SETCC has more than one use, then perhaps it's better
+ // to use a test instead of duplicating the X86ISD::CMP (for register
+ // pressure reason)?
+ SDOperand Cmp = Cond.getOperand(1);
+ unsigned Opc = Cmp.getOpcode();
+ bool IllegalFPCMov =
+ ! ((X86ScalarSSEf32 && Op.getValueType()==MVT::f32) ||
+ (X86ScalarSSEf64 && Op.getValueType()==MVT::f64)) &&
+ !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
+ if ((Opc == X86ISD::CMP_NEW ||
+ Opc == X86ISD::COMI_NEW ||
+ Opc == X86ISD::UCOMI_NEW) &&
+ !IllegalFPCMov) {
+ Cond = DAG.getNode(Opc, MVT::i32, Cmp.getOperand(0), Cmp.getOperand(1));
+ addTest = false;
+ }
+ }
+
+ if (addTest) {
+ CC = DAG.getConstant(X86::COND_NE, MVT::i8);
+ Cond = DAG.getNode(X86ISD::CMP_NEW, MVT::i32, Cond,
+ DAG.getConstant(0, MVT::i8));
+ }
+
+ const MVT::ValueType *VTs = DAG.getNodeValueTypes(Op.getValueType(),
+ MVT::Flag);
+ SmallVector<SDOperand, 4> Ops;
+ // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
+ // condition is true.
+ Ops.push_back(Op.getOperand(2));
+ Ops.push_back(Op.getOperand(1));
+ Ops.push_back(CC);
+ Ops.push_back(Cond);
+ return DAG.getNode(X86ISD::CMOV_NEW, VTs, 2, &Ops[0], Ops.size());
+}
+
SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
bool addTest = true;
SDOperand Chain = Op.getOperand(0);
Cond, Op.getOperand(2), CC, Cond.getValue(1));
}
+SDOperand X86TargetLowering::LowerBRCOND_New(SDOperand Op, SelectionDAG &DAG) {
+ bool addTest = true;
+ SDOperand Chain = Op.getOperand(0);
+ SDOperand Cond = Op.getOperand(1);
+ SDOperand Dest = Op.getOperand(2);
+ SDOperand CC;
+
+ if (Cond.getOpcode() == ISD::SETCC)
+ Cond = LowerSETCC_New(Cond, DAG);
+
+ if (Cond.getOpcode() == X86ISD::SETCC_NEW) {
+ CC = Cond.getOperand(0);
+
+ // If condition flag is set by a X86ISD::CMP, then make a copy of it
+ // (since flag operand cannot be shared). Use it as the condition setting
+ // operand in place of the X86ISD::SETCC.
+ // If the X86ISD::SETCC has more than one use, then perhaps it's better
+ // to use a test instead of duplicating the X86ISD::CMP (for register
+ // pressure reason)?
+ SDOperand Cmp = Cond.getOperand(1);
+ unsigned Opc = Cmp.getOpcode();
+ if (Opc == X86ISD::CMP_NEW ||
+ Opc == X86ISD::COMI_NEW ||
+ Opc == X86ISD::UCOMI_NEW) {
+ Cond = DAG.getNode(Opc, MVT::i32, Cmp.getOperand(0), Cmp.getOperand(1));
+ addTest = false;
+ }
+ }
+
+ if (addTest) {
+ CC = DAG.getConstant(X86::COND_NE, MVT::i8);
+ Cond= DAG.getNode(X86ISD::CMP_NEW, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
+ }
+ return DAG.getNode(X86ISD::BRCOND_NEW, Op.getValueType(),
+ Chain, Op.getOperand(2), CC, Cond);
+}
+
SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
// bytes in one go. Touching the stack at 4K increments is necessary to ensure
// that the guard pages used by the OS virtual memory manager are allocated in
// correct sequence.
-SDOperand X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
- SelectionDAG &DAG) {
+SDOperand
+X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
+ SelectionDAG &DAG) {
assert(Subtarget->isTargetCygMing() &&
"This should be used only on Cygwin/Mingw targets");
SDOperand Size = Op.getOperand(1);
// FIXME: Ensure alignment here
- TargetLowering::ArgListTy Args;
- TargetLowering::ArgListEntry Entry;
+ SDOperand Flag;
+
MVT::ValueType IntPtr = getPointerTy();
MVT::ValueType SPTy = (Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
- const Type *IntPtrTy = getTargetData()->getIntPtrType();
-
- Entry.Node = Size;
- Entry.Ty = IntPtrTy;
- Entry.isInReg = true; // Should pass in EAX
- Args.push_back(Entry);
- std::pair<SDOperand, SDOperand> CallResult =
- LowerCallTo(Chain, IntPtrTy, false, false, CallingConv::C, false,
- DAG.getExternalSymbol("_alloca", IntPtr), Args, DAG);
-
- SDOperand SP = DAG.getCopyFromReg(CallResult.second, X86StackPtr, SPTy);
+
+ Chain = DAG.getCopyToReg(Chain, X86::EAX, Size, Flag);
+ Flag = Chain.getValue(1);
+
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDOperand Ops[] = { Chain,
+ DAG.getTargetExternalSymbol("_alloca", IntPtr),
+ DAG.getRegister(X86::EAX, IntPtr),
+ Flag };
+ Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops, 4);
+ Flag = Chain.getValue(1);
+
+ Chain = DAG.getCopyFromReg(Chain, X86StackPtr, SPTy).getValue(1);
std::vector<MVT::ValueType> Tys;
Tys.push_back(SPTy);
Tys.push_back(MVT::Other);
- SDOperand Ops[2] = { SP, CallResult.second };
- return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
+ SDOperand Ops1[2] = { Chain.getValue(0), Chain };
+ return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops1, 2);
}
SDOperand
if (Align == 0) Align = 1;
ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
- // If not DWORD aligned, call memset if size is less than the threshold.
- // It knows how to align to the right boundary first.
+ // If not DWORD aligned or size is more than the threshold, call memset.
+ // The libc version is likely to be faster for these cases. It can use the
+ // address value and run time information about the CPU.
if ((Align & 3) != 0 ||
- (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
+ (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
MVT::ValueType IntPtr = getPointerTy();
const Type *IntPtrTy = getTargetData()->getIntPtrType();
TargetLowering::ArgListTy Args;
if (Align == 0) Align = 1;
ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
- // If not DWORD aligned, call memcpy if size is less than the threshold.
- // It knows how to align to the right boundary first.
+ // If not DWORD aligned or size is more than the threshold, call memcpy.
+ // The libc version is likely to be faster for these cases. It can use the
+ // address value and run time information about the CPU.
+ // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
if ((Align & 3) != 0 ||
- (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
+ (I && I->getValue() > Subtarget->getMinRepStrSizeThreshold())) {
MVT::ValueType IntPtr = getPointerTy();
TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
SDOperand RHS = Op.getOperand(2);
translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
- const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
- SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
- SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
- VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
- SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
- SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
- return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
+ if (NewCCModeling) {
+ Opc = (Opc == X86ISD::UCOMI) ? X86ISD::UCOMI_NEW : X86ISD::COMI_NEW;
+ SDOperand Cond = DAG.getNode(Opc, MVT::i32, LHS, RHS);
+ SDOperand SetCC = DAG.getNode(X86ISD::SETCC_NEW, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond);
+ return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
+ } else {
+ const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
+ SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
+ SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
+ VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
+ SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
+ SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
+ return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
+ }
}
}
}
DAG.getConstant(4, getPointerTy()));
}
+SDOperand X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDOperand Op,
+ SelectionDAG &DAG) {
+ // Is not yet supported on x86-64
+ if (Subtarget->is64Bit())
+ return SDOperand();
+
+ return DAG.getConstant(8, getPointerTy());
+}
+
+SDOperand X86TargetLowering::LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG)
+{
+ assert(!Subtarget->is64Bit() &&
+ "Lowering of eh_return builtin is not supported yet on x86-64");
+
+ MachineFunction &MF = DAG.getMachineFunction();
+ SDOperand Chain = Op.getOperand(0);
+ SDOperand Offset = Op.getOperand(1);
+ SDOperand Handler = Op.getOperand(2);
+
+ SDOperand Frame = DAG.getRegister(RegInfo->getFrameRegister(MF),
+ getPointerTy());
+
+ SDOperand StoreAddr = DAG.getNode(ISD::SUB, getPointerTy(), Frame,
+ DAG.getConstant(-4UL, getPointerTy()));
+ StoreAddr = DAG.getNode(ISD::ADD, getPointerTy(), StoreAddr, Offset);
+ Chain = DAG.getStore(Chain, Handler, StoreAddr, NULL, 0);
+ Chain = DAG.getCopyToReg(Chain, X86::ECX, StoreAddr);
+ MF.addLiveOut(X86::ECX);
+
+ return DAG.getNode(X86ISD::EH_RETURN, MVT::Other,
+ Chain, DAG.getRegister(X86::ECX, getPointerTy()));
+}
+
+SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
+ SelectionDAG &DAG) {
+ SDOperand Root = Op.getOperand(0);
+ SDOperand Trmp = Op.getOperand(1); // trampoline
+ SDOperand FPtr = Op.getOperand(2); // nested function
+ SDOperand Nest = Op.getOperand(3); // 'nest' parameter value
+
+ SrcValueSDNode *TrmpSV = cast<SrcValueSDNode>(Op.getOperand(4));
+
+ if (Subtarget->is64Bit()) {
+ return SDOperand(); // not yet supported
+ } else {
+ Function *Func = (Function *)
+ cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
+ unsigned CC = Func->getCallingConv();
+ unsigned NestReg;
+
+ switch (CC) {
+ default:
+ assert(0 && "Unsupported calling convention");
+ case CallingConv::C:
+ case CallingConv::Fast:
+ case CallingConv::X86_StdCall: {
+ // Pass 'nest' parameter in ECX.
+ // Must be kept in sync with X86CallingConv.td
+ NestReg = X86::ECX;
+
+ // Check that ECX wasn't needed by an 'inreg' parameter.
+ const FunctionType *FTy = Func->getFunctionType();
+ const ParamAttrsList *Attrs = FTy->getParamAttrs();
+
+ if (Attrs && !Func->isVarArg()) {
+ unsigned InRegCount = 0;
+ unsigned Idx = 1;
+
+ for (FunctionType::param_iterator I = FTy->param_begin(),
+ E = FTy->param_end(); I != E; ++I, ++Idx)
+ if (Attrs->paramHasAttr(Idx, ParamAttr::InReg))
+ // FIXME: should only count parameters that are lowered to integers.
+ InRegCount += (getTargetData()->getTypeSizeInBits(*I) + 31) / 32;
+
+ if (InRegCount > 2) {
+ cerr << "Nest register in use - reduce number of inreg parameters!\n";
+ abort();
+ }
+ }
+ break;
+ }
+ case CallingConv::X86_FastCall:
+ // Pass 'nest' parameter in EAX.
+ // Must be kept in sync with X86CallingConv.td
+ NestReg = X86::EAX;
+ break;
+ }
+
+ const X86InstrInfo *TII =
+ ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
+
+ SDOperand OutChains[4];
+ SDOperand Addr, Disp;
+
+ Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(10, MVT::i32));
+ Disp = DAG.getNode(ISD::SUB, MVT::i32, FPtr, Addr);
+
+ unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
+ unsigned char N86Reg = ((X86RegisterInfo&)RegInfo).getX86RegNum(NestReg);
+ OutChains[0] = DAG.getStore(Root, DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
+ Trmp, TrmpSV->getValue(), TrmpSV->getOffset());
+
+ Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(1, MVT::i32));
+ OutChains[1] = DAG.getStore(Root, Nest, Addr, TrmpSV->getValue(),
+ TrmpSV->getOffset() + 1, false, 1);
+
+ unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
+ Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(5, MVT::i32));
+ OutChains[2] = DAG.getStore(Root, DAG.getConstant(JMP, MVT::i8), Addr,
+ TrmpSV->getValue() + 5, TrmpSV->getOffset());
+
+ Addr = DAG.getNode(ISD::ADD, MVT::i32, Trmp, DAG.getConstant(6, MVT::i32));
+ OutChains[3] = DAG.getStore(Root, Disp, Addr, TrmpSV->getValue(),
+ TrmpSV->getOffset() + 6, false, 1);
+
+ SDOperand Ops[] =
+ { Trmp, DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains, 4) };
+ return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), Ops, 2);
+ }
+}
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
case ISD::SHL_PARTS:
case ISD::SRA_PARTS:
case ISD::SRL_PARTS: return LowerShift(Op, DAG);
+ case ISD::SDIV:
+ case ISD::UDIV:
+ case ISD::SREM:
+ case ISD::UREM: return LowerIntegerDivOrRem(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::FABS: return LowerFABS(Op, DAG);
case ISD::FNEG: return LowerFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
- case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
- case ISD::SELECT: return LowerSELECT(Op, DAG);
- case ISD::BRCOND: return LowerBRCOND(Op, DAG);
+ case ISD::SETCC: return NewCCModeling
+ ? LowerSETCC_New(Op, DAG)
+ : LowerSETCC(Op, DAG, DAG.getEntryNode());
+ case ISD::SELECT: return NewCCModeling
+ ? LowerSELECT_New(Op, DAG)
+ : LowerSELECT(Op, DAG);
+ case ISD::BRCOND: return NewCCModeling
+ ? LowerBRCOND_New(Op, DAG)
+ : LowerBRCOND(Op, DAG);
case ISD::JumpTable: return LowerJumpTable(Op, DAG);
case ISD::CALL: return LowerCALL(Op, DAG);
case ISD::RET: return LowerRET(Op, DAG);
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
+ case ISD::FRAME_TO_ARGS_OFFSET:
+ return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
+ case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
+ case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
}
return SDOperand();
}
case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
case X86ISD::CMP: return "X86ISD::CMP";
+ case X86ISD::CMP_NEW: return "X86ISD::CMP_NEW";
case X86ISD::COMI: return "X86ISD::COMI";
+ case X86ISD::COMI_NEW: return "X86ISD::COMI_NEW";
case X86ISD::UCOMI: return "X86ISD::UCOMI";
+ case X86ISD::UCOMI_NEW: return "X86ISD::UCOMI_NEW";
case X86ISD::SETCC: return "X86ISD::SETCC";
+ case X86ISD::SETCC_NEW: return "X86ISD::SETCC_NEW";
case X86ISD::CMOV: return "X86ISD::CMOV";
+ case X86ISD::CMOV_NEW: return "X86ISD::CMOV_NEW";
case X86ISD::BRCOND: return "X86ISD::BRCOND";
+ case X86ISD::BRCOND_NEW: return "X86ISD::BRCOND_NEW";
case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
- case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
- case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
case X86ISD::Wrapper: return "X86ISD::Wrapper";
case X86ISD::S2VEC: return "X86ISD::S2VEC";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
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::EH_RETURN: return "X86ISD::EH_RETURN";
+ case X86ISD::DIV: return "X86ISD::DIV";
+ case X86ISD::IDIV: return "X86ISD::IDIV";
}
}
return false;
if (AM.BaseGV) {
- // X86-64 only supports addr of globals in small code model.
- if (Subtarget->is64Bit() &&
- getTargetMachine().getCodeModel() != CodeModel::Small)
- return false;
-
- // We can only fold this if we don't need a load either.
+ // We can only fold this if we don't need an extra load.
if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
return false;
+
+ // X86-64 only supports addr of globals in small code model.
+ if (Subtarget->is64Bit()) {
+ if (getTargetMachine().getCodeModel() != CodeModel::Small)
+ return false;
+ // If lower 4G is not available, then we must use rip-relative addressing.
+ if (AM.BaseOffs || AM.Scale > 1)
+ return false;
+ }
}
switch (AM.Scale) {
// Only do shuffles on 128-bit vector types for now.
if (MVT::getSizeInBits(VT) == 64) return false;
return (Mask.Val->getNumOperands() <= 4 ||
+ isIdentityMask(Mask.Val) ||
+ isIdentityMask(Mask.Val, true) ||
isSplatMask(Mask.Val) ||
isPSHUFHW_PSHUFLWMask(Mask.Val) ||
X86::isUNPCKLMask(Mask.Val) ||
+ X86::isUNPCKHMask(Mask.Val) ||
X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
- X86::isUNPCKH_v_undef_Mask(Mask.Val) ||
- X86::isUNPCKHMask(Mask.Val));
+ X86::isUNPCKH_v_undef_Mask(Mask.Val));
}
bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
case X86::CMOV_FR64:
case X86::CMOV_V4F32:
case X86::CMOV_V2F64:
- case X86::CMOV_V2I64: {
+ case X86::CMOV_V2I64:
+
+ case X86::NEW_CMOV_FR32:
+ case X86::NEW_CMOV_FR64:
+ case X86::NEW_CMOV_V4F32:
+ case X86::NEW_CMOV_V2F64:
+ case X86::NEW_CMOV_V2I64: {
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
// destination vreg to set, the condition code register to branch on, the
return BB;
}
- case X86::FP_TO_INT16_IN_MEM:
- case X86::FP_TO_INT32_IN_MEM:
- case X86::FP_TO_INT64_IN_MEM: {
+ case X86::FP32_TO_INT16_IN_MEM:
+ case X86::FP32_TO_INT32_IN_MEM:
+ case X86::FP32_TO_INT64_IN_MEM:
+ case X86::FP64_TO_INT16_IN_MEM:
+ case X86::FP64_TO_INT32_IN_MEM:
+ case X86::FP64_TO_INT64_IN_MEM:
+ case X86::FP80_TO_INT16_IN_MEM:
+ case X86::FP80_TO_INT32_IN_MEM:
+ case X86::FP80_TO_INT64_IN_MEM: {
// Change the floating point control register to use "round towards zero"
// mode when truncating to an integer value.
MachineFunction *F = BB->getParent();
unsigned Opc;
switch (MI->getOpcode()) {
default: assert(0 && "illegal opcode!");
- case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
- case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
- case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
+ case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
+ case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
+ case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
+ case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
+ case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
+ case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
+ case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
+ case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
+ case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
}
X86AddressMode AM;
uint64_t Mask,
uint64_t &KnownZero,
uint64_t &KnownOne,
+ const SelectionDAG &DAG,
unsigned Depth) const {
unsigned Opc = Op.getOpcode();
assert((Opc >= ISD::BUILTIN_OP_END ||
switch (Opc) {
default: break;
case X86ISD::SETCC:
+ case X86ISD::SETCC_NEW:
KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
break;
}
if (Loc.getOpcode() == ISD::FrameIndex) {
if (BaseLoc.getOpcode() != ISD::FrameIndex)
return false;
- int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
- int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
+ int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
+ int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
int FS = MFI->getObjectSize(FI);
int BFS = MFI->getObjectSize(BFI);
if (FS != BFS || FS != Size) return false;
return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
else {
assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
- int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
+ int BFI = cast<FrameIndexSDNode>(Base)->getIndex();
if (BFI < 0)
// Fixed objects do not specify alignment, however the offsets are known.
return ((Subtarget->getStackAlignment() % 16) == 0 &&
}
bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
+ LoadSDNode *LD = cast<LoadSDNode>(Base);
if (isAlign16) {
- LoadSDNode *LD = cast<LoadSDNode>(Base);
return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
- LD->getSrcValueOffset());
+ LD->getSrcValueOffset(), LD->isVolatile());
} else {
- // Just use movups, it's shorter.
- SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
- SmallVector<SDOperand, 3> Ops;
- Ops.push_back(Base->getOperand(0));
- Ops.push_back(Base->getOperand(1));
- Ops.push_back(Base->getOperand(2));
- return DAG.getNode(ISD::BIT_CONVERT, VT,
- DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
+ return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
+ LD->getSrcValueOffset(), LD->isVolatile(),
+ LD->getAlignment());
}
}
return TargetLowering::getConstraintType(Constraint);
}
-/// isOperandValidForConstraint - Return the specified operand (possibly
-/// modified) if the specified SDOperand is valid for the specified target
-/// constraint letter, otherwise return null.
-SDOperand X86TargetLowering::
-isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
+/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
+/// vector. If it is invalid, don't add anything to Ops.
+void X86TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
+ char Constraint,
+ std::vector<SDOperand>&Ops,
+ SelectionDAG &DAG) {
+ SDOperand Result(0, 0);
+
switch (Constraint) {
default: break;
case 'I':
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
- if (C->getValue() <= 31)
- return DAG.getTargetConstant(C->getValue(), Op.getValueType());
+ if (C->getValue() <= 31) {
+ Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
+ break;
+ }
}
- return SDOperand(0,0);
+ return;
case 'N':
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
- if (C->getValue() <= 255)
- return DAG.getTargetConstant(C->getValue(), Op.getValueType());
+ if (C->getValue() <= 255) {
+ Result = DAG.getTargetConstant(C->getValue(), Op.getValueType());
+ break;
+ }
}
- return SDOperand(0,0);
+ return;
case 'i': {
// Literal immediates are always ok.
- if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op))
- return DAG.getTargetConstant(CST->getValue(), Op.getValueType());
+ if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
+ Result = DAG.getTargetConstant(CST->getValue(), Op.getValueType());
+ break;
+ }
// If we are in non-pic codegen mode, we allow the address of a global (with
// an optional displacement) to be used with 'i'.
// match.
if (Subtarget->GVRequiresExtraLoad(GA->getGlobal(), getTargetMachine(),
false))
- return SDOperand(0, 0);
+ return;
Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
Offset);
- return Op;
+ Result = Op;
+ break;
}
// Otherwise, not valid for this mode.
- return SDOperand(0, 0);
+ return;
+ }
}
+
+ if (Result.Val) {
+ Ops.push_back(Result);
+ return;
}
- return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
+ return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}
std::vector<unsigned> X86TargetLowering::
else if (VT == MVT::i16)
return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
else if (VT == MVT::i8)
- return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
+ return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
break;
}
}
case MVT::i64:
return std::make_pair(0U, X86::FR64RegisterClass);
// Vector types.
- case MVT::Vector:
case MVT::v16i8:
case MVT::v8i16:
case MVT::v4i32:
// GCC calls "st(0)" just plain "st".
if (StringsEqualNoCase("{st}", Constraint)) {
Res.first = X86::ST0;
- Res.second = X86::RSTRegisterClass;
+ Res.second = X86::RFP80RegisterClass;
}
return Res;
projects / oota-llvm.git / blobdiff