#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
+#include "llvm/Intrinsics.h"
#include "llvm/ADT/VectorExtras.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
addLegalAddressScale(3);
// Set up the register classes.
- addRegisterClass(MVT::i8, X86::R8RegisterClass);
- addRegisterClass(MVT::i16, X86::R16RegisterClass);
- addRegisterClass(MVT::i32, X86::R32RegisterClass);
+ addRegisterClass(MVT::i8, X86::GR8RegisterClass);
+ addRegisterClass(MVT::i16, X86::GR16RegisterClass);
+ addRegisterClass(MVT::i32, X86::GR32RegisterClass);
// Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
// operation.
setOperationAction(ISD::RET , MVT::Other, Custom);
// Darwin ABI issue.
setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
+ setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
// 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
addRegisterClass(MVT::f32, X86::FR32RegisterClass);
addRegisterClass(MVT::f64, X86::FR64RegisterClass);
- // SSE has no load+extend ops
- setOperationAction(ISD::EXTLOAD, MVT::f32, Expand);
- setOperationAction(ISD::ZEXTLOAD, MVT::f32, Expand);
-
// Use ANDPD to simulate FABS.
setOperationAction(ISD::FABS , MVT::f64, Custom);
setOperationAction(ISD::FABS , MVT::f32, Custom);
if (Subtarget->hasSSE1()) {
addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
- setOperationAction(ISD::ADD, MVT::v4f32, Legal);
- setOperationAction(ISD::SUB, MVT::v4f32, Legal);
- setOperationAction(ISD::MUL, MVT::v4f32, Legal);
- setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
+ setOperationAction(ISD::AND, MVT::v4f32, Legal);
+ setOperationAction(ISD::OR, MVT::v4f32, Legal);
+ setOperationAction(ISD::XOR, MVT::v4f32, Legal);
+ setOperationAction(ISD::ADD, MVT::v4f32, Legal);
+ setOperationAction(ISD::SUB, MVT::v4f32, Legal);
+ setOperationAction(ISD::MUL, MVT::v4f32, Legal);
+ setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
+ setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
+ setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
}
if (Subtarget->hasSSE2()) {
addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
-
- setOperationAction(ISD::ADD, MVT::v2f64, Legal);
- setOperationAction(ISD::ADD, MVT::v16i8, Legal);
- setOperationAction(ISD::ADD, MVT::v8i16, Legal);
- setOperationAction(ISD::ADD, MVT::v4i32, Legal);
- setOperationAction(ISD::SUB, MVT::v2f64, Legal);
- setOperationAction(ISD::SUB, MVT::v16i8, Legal);
- setOperationAction(ISD::SUB, MVT::v8i16, Legal);
- setOperationAction(ISD::SUB, MVT::v4i32, Legal);
- setOperationAction(ISD::MUL, MVT::v2f64, Legal);
- setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
- setOperationAction(ISD::LOAD, MVT::v16i8, Legal);
- setOperationAction(ISD::LOAD, MVT::v8i16, Legal);
- setOperationAction(ISD::LOAD, MVT::v4i32, Legal);
- setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
- setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
- setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i16, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i32, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
+ setOperationAction(ISD::ADD, MVT::v2f64, Legal);
+ setOperationAction(ISD::ADD, MVT::v16i8, Legal);
+ setOperationAction(ISD::ADD, MVT::v8i16, Legal);
+ setOperationAction(ISD::ADD, MVT::v4i32, Legal);
+ setOperationAction(ISD::SUB, MVT::v2f64, Legal);
+ setOperationAction(ISD::SUB, MVT::v16i8, Legal);
+ setOperationAction(ISD::SUB, MVT::v8i16, Legal);
+ setOperationAction(ISD::SUB, MVT::v4i32, Legal);
+ setOperationAction(ISD::MUL, MVT::v8i16, Legal);
+ setOperationAction(ISD::MUL, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
+ // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
+
+ // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
+ for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
+ setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)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::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
+
+ // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
+ for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
+ setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
+ AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
+ setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
+ AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
+ setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
+ AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
+ setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
+ AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
+ setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
+ AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
+ }
+
+ // Custom lower v2i64 and v2f64 selects.
+ setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
+ setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
+ setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
+ setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
}
+ // We want to custom lower some of our intrinsics.
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+
computeRegisterProperties();
// FIXME: These should be based on subtarget info. Plus, the values should
std::vector<SDOperand>
X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
+ std::vector<SDOperand> Args = TargetLowering::LowerArguments(F, DAG);
+
+ FormalArgs.clear();
+ FormalArgLocs.clear();
+
+ // This sets BytesToPopOnReturn, BytesCallerReserves, etc. which have to be
+ // set before the rest of the function can be lowered.
if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
- return LowerFastCCArguments(F, DAG);
- return LowerCCCArguments(F, DAG);
+ PreprocessFastCCArguments(Args, F, DAG);
+ else
+ PreprocessCCCArguments(Args, F, DAG);
+ return Args;
}
std::pair<SDOperand, SDOperand>
// C Calling Convention implementation
//===----------------------------------------------------------------------===//
-std::vector<SDOperand>
-X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
- std::vector<SDOperand> ArgValues;
+/// 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,
+ TargetRegisterClass *RC) {
+ assert(RC->contains(PReg) && "Not the correct regclass!");
+ unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
+ MF.addLiveIn(PReg, VReg);
+ return VReg;
+}
+
+/// HowToPassCCCArgument - Returns how an formal argument of the specified type
+/// should be passed. If it is through stack, returns the size of the stack
+/// frame; if it is through XMM register, returns the number of XMM registers
+/// are needed.
+static void
+HowToPassCCCArgument(MVT::ValueType ObjectVT, unsigned NumXMMRegs,
+ unsigned &ObjSize, unsigned &ObjXMMRegs) {
+ switch (ObjectVT) {
+ default: assert(0 && "Unhandled argument type!");
+ case MVT::i1:
+ case MVT::i8: ObjSize = 1; break;
+ case MVT::i16: ObjSize = 2; break;
+ case MVT::i32: ObjSize = 4; break;
+ case MVT::i64: ObjSize = 8; break;
+ case MVT::f32: ObjSize = 4; break;
+ case MVT::f64: ObjSize = 8; break;
+ case MVT::v16i8:
+ case MVT::v8i16:
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v4f32:
+ case MVT::v2f64:
+ if (NumXMMRegs < 3)
+ ObjXMMRegs = 1;
+ else
+ ObjSize = 16;
+ break;
+ }
+}
+
+/// getFormalArgObjects - Returns itself if Op is a FORMAL_ARGUMENTS, otherwise
+/// returns the FORMAL_ARGUMENTS node(s) that made up parts of the node.
+static std::vector<SDOperand> getFormalArgObjects(SDOperand Op) {
+ unsigned Opc = Op.getOpcode();
+ std::vector<SDOperand> Objs;
+ if (Opc == ISD::TRUNCATE) {
+ Op = Op.getOperand(0);
+ assert(Op.getOpcode() == ISD::AssertSext ||
+ Op.getOpcode() == ISD::AssertZext);
+ Objs.push_back(Op.getOperand(0));
+ } else if (Opc == ISD::FP_ROUND || Opc == ISD::VBIT_CONVERT) {
+ Objs.push_back(Op.getOperand(0));
+ } else if (Opc == ISD::BUILD_PAIR) {
+ Objs.push_back(Op.getOperand(0));
+ Objs.push_back(Op.getOperand(1));
+ } else {
+ Objs.push_back(Op);
+ }
+ return Objs;
+}
+void X86TargetLowering::PreprocessCCCArguments(std::vector<SDOperand>Args,
+ Function &F, SelectionDAG &DAG) {
+ unsigned NumArgs = Args.size();
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
// ...
//
unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
- for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
- MVT::ValueType ObjectVT = getValueType(I->getType());
- unsigned ArgIncrement = 4;
- unsigned ObjSize;
- switch (ObjectVT) {
- default: assert(0 && "Unhandled argument type!");
- case MVT::i1:
- case MVT::i8: ObjSize = 1; break;
- case MVT::i16: ObjSize = 2; break;
- case MVT::i32: ObjSize = 4; break;
- case MVT::i64: ObjSize = ArgIncrement = 8; break;
- case MVT::f32: ObjSize = 4; break;
- case MVT::f64: ObjSize = ArgIncrement = 8; break;
- }
- // Create the frame index object for this incoming parameter...
- int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
-
- // Create the SelectionDAG nodes corresponding to a load from this parameter
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
-
- // Don't codegen dead arguments. FIXME: remove this check when we can nuke
- // dead loads.
- SDOperand ArgValue;
- if (!I->use_empty())
- ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
- DAG.getSrcValue(NULL));
- else {
- if (MVT::isInteger(ObjectVT))
- ArgValue = DAG.getConstant(0, ObjectVT);
- else
- ArgValue = DAG.getConstantFP(0, ObjectVT);
+ unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
+ unsigned XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2 };
+ for (unsigned i = 0; i < NumArgs; ++i) {
+ SDOperand Op = Args[i];
+ std::vector<SDOperand> Objs = getFormalArgObjects(Op);
+ for (std::vector<SDOperand>::iterator I = Objs.begin(), E = Objs.end();
+ I != E; ++I) {
+ SDOperand Obj = *I;
+ MVT::ValueType ObjectVT = Obj.getValueType();
+ unsigned ArgIncrement = 4;
+ unsigned ObjSize = 0;
+ unsigned ObjXMMRegs = 0;
+ HowToPassCCCArgument(ObjectVT, NumXMMRegs, ObjSize, ObjXMMRegs);
+ if (ObjSize >= 8)
+ ArgIncrement = ObjSize;
+
+ if (ObjXMMRegs) {
+ // Passed in a XMM register.
+ unsigned Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
+ X86::VR128RegisterClass);
+ std::pair<FALocInfo, FALocInfo> Loc =
+ std::make_pair(FALocInfo(FALocInfo::LiveInRegLoc, Reg, ObjectVT),
+ FALocInfo());
+ FormalArgLocs.push_back(Loc);
+ NumXMMRegs += ObjXMMRegs;
+ } else {
+ // Create the frame index object for this incoming parameter...
+ int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
+ std::pair<FALocInfo, FALocInfo> Loc =
+ std::make_pair(FALocInfo(FALocInfo::StackFrameLoc, FI), FALocInfo());
+ FormalArgLocs.push_back(Loc);
+ ArgOffset += ArgIncrement; // Move on to the next argument...
+ }
}
- ArgValues.push_back(ArgValue);
-
- ArgOffset += ArgIncrement; // Move on to the next argument...
}
// If the function takes variable number of arguments, make a frame index for
ReturnAddrIndex = 0; // No return address slot generated yet.
BytesToPopOnReturn = 0; // Callee pops nothing.
BytesCallerReserves = ArgOffset;
+}
- // Finally, inform the code generator which regs we return values in.
- switch (getValueType(F.getReturnType())) {
- default: assert(0 && "Unknown type!");
- case MVT::isVoid: break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- MF.addLiveOut(X86::EAX);
- break;
- case MVT::i64:
- MF.addLiveOut(X86::EAX);
- MF.addLiveOut(X86::EDX);
- break;
- case MVT::f32:
- case MVT::f64:
- MF.addLiveOut(X86::ST0);
- break;
+void X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG) {
+ unsigned NumArgs = Op.Val->getNumValues() - 1;
+ MachineFunction &MF = DAG.getMachineFunction();
+
+ for (unsigned i = 0; i < NumArgs; ++i) {
+ std::pair<FALocInfo, FALocInfo> Loc = FormalArgLocs[i];
+ SDOperand ArgValue;
+ if (Loc.first.Kind == FALocInfo::StackFrameLoc) {
+ // Create the SelectionDAG nodes corresponding to a load from this
+ // parameter.
+ unsigned FI = FormalArgLocs[i].first.Loc;
+ SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
+ ArgValue = DAG.getLoad(Op.Val->getValueType(i),
+ DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
+ } else {
+ // Must be a CopyFromReg
+ ArgValue= DAG.getCopyFromReg(DAG.getEntryNode(), Loc.first.Loc,
+ Loc.first.Typ);
+ }
+ FormalArgs.push_back(ArgValue);
}
- return ArgValues;
+ // Provide a chain. Note that this isn't the right one, but it works as well
+ // as before.
+ FormalArgs.push_back(DAG.getEntryNode());
}
std::pair<SDOperand, SDOperand>
// Count how many bytes are to be pushed on the stack.
unsigned NumBytes = 0;
+ // Keep track of the number of XMM regs passed so far.
+ unsigned NumXMMRegs = 0;
+ unsigned XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2 };
+
+ std::vector<SDOperand> RegValuesToPass;
if (Args.empty()) {
// Save zero bytes.
Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(0, getPointerTy()));
case MVT::f64:
NumBytes += 8;
break;
+ case MVT::Vector:
+ if (NumXMMRegs < 3)
+ ++NumXMMRegs;
+ else
+ NumBytes += 16;
+ break;
}
Chain = DAG.getCALLSEQ_START(Chain,
// Arguments go on the stack in reverse order, as specified by the ABI.
unsigned ArgOffset = 0;
+ NumXMMRegs = 0;
SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
std::vector<SDOperand> Stores;
-
for (unsigned i = 0, e = Args.size(); i != e; ++i) {
- SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
-
switch (getValueType(Args[i].second)) {
default: assert(0 && "Unexpected ValueType for argument!");
case MVT::i1:
// FALL THROUGH
case MVT::i32:
- case MVT::f32:
+ case MVT::f32: {
+ SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
+ PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
Args[i].first, PtrOff,
DAG.getSrcValue(NULL)));
ArgOffset += 4;
break;
+ }
case MVT::i64:
- case MVT::f64:
+ case MVT::f64: {
+ SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
+ PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
Args[i].first, PtrOff,
DAG.getSrcValue(NULL)));
ArgOffset += 8;
break;
}
+ case MVT::Vector:
+ if (NumXMMRegs < 3) {
+ RegValuesToPass.push_back(Args[i].first);
+ NumXMMRegs++;
+ } else {
+ SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
+ PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
+ Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
+ Args[i].first, PtrOff,
+ DAG.getSrcValue(NULL)));
+ ArgOffset += 16;
+ }
+ }
}
+ if (!Stores.empty())
Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
}
break;
}
+ // Build a sequence of copy-to-reg nodes chained together with token chain
+ // and flag operands which copy the outgoing args into registers.
+ SDOperand InFlag;
+ for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
+ unsigned CCReg = XMMArgRegs[i];
+ SDOperand RegToPass = RegValuesToPass[i];
+ assert(RegToPass.getValueType() == MVT::Vector);
+ unsigned NumElems =
+ cast<ConstantSDNode>(*(RegToPass.Val->op_end()-2))->getValue();
+ MVT::ValueType EVT = cast<VTSDNode>(*(RegToPass.Val->op_end()-1))->getVT();
+ MVT::ValueType PVT = getVectorType(EVT, NumElems);
+ SDOperand CCRegNode = DAG.getRegister(CCReg, PVT);
+ RegToPass = DAG.getNode(ISD::VBIT_CONVERT, PVT, RegToPass);
+ Chain = DAG.getCopyToReg(Chain, CCRegNode, RegToPass, InFlag);
+ InFlag = Chain.getValue(1);
+ }
+
std::vector<MVT::ValueType> NodeTys;
NodeTys.push_back(MVT::Other); // Returns a chain
NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
std::vector<SDOperand> Ops;
Ops.push_back(Chain);
Ops.push_back(Callee);
+ if (InFlag.Val)
+ Ops.push_back(InFlag);
// FIXME: Do not generate X86ISD::TAILCALL for now.
Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
- SDOperand InFlag = Chain.getValue(1);
+ InFlag = Chain.getValue(1);
NodeTys.clear();
NodeTys.push_back(MVT::Other); // Returns a chain
RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
break;
}
+ case MVT::Vector: {
+ const PackedType *PTy = cast<PackedType>(RetTy);
+ MVT::ValueType EVT;
+ MVT::ValueType LVT;
+ unsigned NumRegs = getPackedTypeBreakdown(PTy, EVT, LVT);
+ assert(NumRegs == 1 && "Unsupported type!");
+ RetVal = DAG.getCopyFromReg(Chain, X86::XMM0, EVT, InFlag);
+ Chain = RetVal.getValue(1);
+ break;
+ }
}
}
// (when we have a global fp allocator) and do other tricks.
//
-/// 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,
- TargetRegisterClass *RC) {
- assert(RC->contains(PReg) && "Not the correct regclass!");
- unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
- MF.addLiveIn(PReg, VReg);
- return VReg;
-}
-
// FASTCC_NUM_INT_ARGS_INREGS - This is the max number of integer arguments
// to pass in registers. 0 is none, 1 is is "use EAX", 2 is "use EAX and
// EDX". Anything more is illegal.
static unsigned FASTCC_NUM_INT_ARGS_INREGS = 0;
-std::vector<SDOperand>
-X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
- std::vector<SDOperand> ArgValues;
+/// HowToPassFastCCArgument - Returns how an formal argument of the specified
+/// type should be passed. If it is through stack, returns the size of the stack
+/// frame; if it is through integer or XMM register, returns the number of
+/// integer or XMM registers are needed.
+static void
+HowToPassFastCCArgument(MVT::ValueType ObjectVT,
+ unsigned NumIntRegs, unsigned NumXMMRegs,
+ unsigned &ObjSize, unsigned &ObjIntRegs,
+ unsigned &ObjXMMRegs) {
+ ObjSize = 0;
+ NumIntRegs = 0;
+
+ switch (ObjectVT) {
+ default: assert(0 && "Unhandled argument type!");
+ case MVT::i1:
+ case MVT::i8:
+ if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
+ ObjIntRegs = 1;
+ else
+ ObjSize = 1;
+ break;
+ case MVT::i16:
+ if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
+ ObjIntRegs = 1;
+ else
+ ObjSize = 2;
+ break;
+ case MVT::i32:
+ if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
+ ObjIntRegs = 1;
+ else
+ ObjSize = 4;
+ break;
+ case MVT::i64:
+ if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
+ ObjIntRegs = 2;
+ } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
+ ObjIntRegs = 1;
+ ObjSize = 4;
+ } else
+ ObjSize = 8;
+ case MVT::f32:
+ ObjSize = 4;
+ break;
+ case MVT::f64:
+ ObjSize = 8;
+ break;
+ case MVT::v16i8:
+ case MVT::v8i16:
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v4f32:
+ case MVT::v2f64:
+ if (NumXMMRegs < 3)
+ ObjXMMRegs = 1;
+ else
+ ObjSize = 16;
+ break;
+ }
+}
+void
+X86TargetLowering::PreprocessFastCCArguments(std::vector<SDOperand>Args,
+ Function &F, SelectionDAG &DAG) {
+ unsigned NumArgs = Args.size();
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
// 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
// used).
unsigned NumIntRegs = 0;
+ unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
+ unsigned XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2 };
- for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
- MVT::ValueType ObjectVT = getValueType(I->getType());
- unsigned ArgIncrement = 4;
- unsigned ObjSize = 0;
- SDOperand ArgValue;
-
- switch (ObjectVT) {
- default: assert(0 && "Unhandled argument type!");
- case MVT::i1:
- case MVT::i8:
- if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
- if (!I->use_empty()) {
- unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
- X86::R8RegisterClass);
- ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i8);
- DAG.setRoot(ArgValue.getValue(1));
- if (ObjectVT == MVT::i1)
- // FIXME: Should insert a assertzext here.
- ArgValue = DAG.getNode(ISD::TRUNCATE, MVT::i1, ArgValue);
- }
- ++NumIntRegs;
- break;
- }
-
- ObjSize = 1;
- break;
- case MVT::i16:
- if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
- if (!I->use_empty()) {
- unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
- X86::R16RegisterClass);
- ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i16);
- DAG.setRoot(ArgValue.getValue(1));
- }
- ++NumIntRegs;
- break;
- }
- ObjSize = 2;
- break;
- case MVT::i32:
- if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
- if (!I->use_empty()) {
- unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
- X86::R32RegisterClass);
- ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
- DAG.setRoot(ArgValue.getValue(1));
+ for (unsigned i = 0; i < NumArgs; ++i) {
+ SDOperand Op = Args[i];
+ std::vector<SDOperand> Objs = getFormalArgObjects(Op);
+ for (std::vector<SDOperand>::iterator I = Objs.begin(), E = Objs.end();
+ I != E; ++I) {
+ SDOperand Obj = *I;
+ MVT::ValueType ObjectVT = Obj.getValueType();
+ unsigned ArgIncrement = 4;
+ unsigned ObjSize = 0;
+ unsigned ObjIntRegs = 0;
+ unsigned ObjXMMRegs = 0;
+
+ HowToPassFastCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
+ ObjSize, ObjIntRegs, ObjXMMRegs);
+ if (ObjSize >= 8)
+ ArgIncrement = ObjSize;
+
+ unsigned Reg;
+ std::pair<FALocInfo,FALocInfo> Loc = std::make_pair(FALocInfo(),
+ FALocInfo());
+ if (ObjIntRegs) {
+ switch (ObjectVT) {
+ default: assert(0 && "Unhandled argument type!");
+ case MVT::i1:
+ case MVT::i8:
+ Reg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
+ X86::GR8RegisterClass);
+ Loc.first.Kind = FALocInfo::LiveInRegLoc;
+ Loc.first.Loc = Reg;
+ Loc.first.Typ = MVT::i8;
+ break;
+ case MVT::i16:
+ Reg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
+ X86::GR16RegisterClass);
+ Loc.first.Kind = FALocInfo::LiveInRegLoc;
+ Loc.first.Loc = Reg;
+ Loc.first.Typ = MVT::i16;
+ break;
+ case MVT::i32:
+ Reg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
+ X86::GR32RegisterClass);
+ Loc.first.Kind = FALocInfo::LiveInRegLoc;
+ Loc.first.Loc = Reg;
+ Loc.first.Typ = MVT::i32;
+ break;
+ case MVT::i64:
+ Reg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
+ X86::GR32RegisterClass);
+ Loc.first.Kind = FALocInfo::LiveInRegLoc;
+ Loc.first.Loc = Reg;
+ Loc.first.Typ = MVT::i32;
+ if (ObjIntRegs == 2) {
+ Reg = AddLiveIn(MF, X86::EDX, X86::GR32RegisterClass);
+ Loc.second.Kind = FALocInfo::LiveInRegLoc;
+ Loc.second.Loc = Reg;
+ Loc.second.Typ = MVT::i32;
+ }
+ break;
+ case MVT::v16i8:
+ case MVT::v8i16:
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v4f32:
+ case MVT::v2f64:
+ Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
+ Loc.first.Kind = FALocInfo::LiveInRegLoc;
+ Loc.first.Loc = Reg;
+ Loc.first.Typ = ObjectVT;
+ break;
}
- ++NumIntRegs;
- break;
+ NumIntRegs += ObjIntRegs;
+ NumXMMRegs += ObjXMMRegs;
}
- ObjSize = 4;
- break;
- case MVT::i64:
- if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
- if (!I->use_empty()) {
- unsigned BotReg = AddLiveIn(MF, X86::EAX, X86::R32RegisterClass);
- unsigned TopReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
-
- SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
- SDOperand Hi = DAG.getCopyFromReg(Low.getValue(1), TopReg, MVT::i32);
- DAG.setRoot(Hi.getValue(1));
-
- ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
- }
- NumIntRegs += 2;
- break;
- } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
- if (!I->use_empty()) {
- unsigned BotReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
- SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
- DAG.setRoot(Low.getValue(1));
-
- // Load the high part from memory.
- // Create the frame index object for this incoming parameter...
- int FI = MFI->CreateFixedObject(4, ArgOffset);
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
- SDOperand Hi = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
- DAG.getSrcValue(NULL));
- ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
+ if (ObjSize) {
+ int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
+ if (ObjectVT == MVT::i64 && ObjIntRegs) {
+ Loc.second.Kind = FALocInfo::StackFrameLoc;
+ Loc.second.Loc = FI;
+ } else {
+ Loc.first.Kind = FALocInfo::StackFrameLoc;
+ Loc.first.Loc = FI;
}
- ArgOffset += 4;
- NumIntRegs = FASTCC_NUM_INT_ARGS_INREGS;
- break;
+ ArgOffset += ArgIncrement; // Move on to the next argument.
}
- ObjSize = ArgIncrement = 8;
- break;
- case MVT::f32: ObjSize = 4; break;
- case MVT::f64: ObjSize = ArgIncrement = 8; break;
- }
-
- // Don't codegen dead arguments. FIXME: remove this check when we can nuke
- // dead loads.
- if (ObjSize && !I->use_empty()) {
- // Create the frame index object for this incoming parameter...
- int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
-
- // Create the SelectionDAG nodes corresponding to a load from this
- // parameter.
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
- ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
- DAG.getSrcValue(NULL));
- } else if (ArgValue.Val == 0) {
- if (MVT::isInteger(ObjectVT))
- ArgValue = DAG.getConstant(0, ObjectVT);
- else
- ArgValue = DAG.getConstantFP(0, ObjectVT);
+ FormalArgLocs.push_back(Loc);
}
- ArgValues.push_back(ArgValue);
-
- if (ObjSize)
- ArgOffset += ArgIncrement; // Move on to the next argument.
}
// Make sure the instruction takes 8n+4 bytes to make sure the start of the
case MVT::f64:
MF.addLiveOut(X86::ST0);
break;
+ case MVT::Vector: {
+ const PackedType *PTy = cast<PackedType>(F.getReturnType());
+ MVT::ValueType EVT;
+ MVT::ValueType LVT;
+ unsigned NumRegs = getPackedTypeBreakdown(PTy, EVT, LVT);
+ assert(NumRegs == 1 && "Unsupported type!");
+ MF.addLiveOut(X86::XMM0);
+ break;
+ }
+ }
+}
+
+void
+X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
+ unsigned NumArgs = Op.Val->getNumValues()-1;
+ MachineFunction &MF = DAG.getMachineFunction();
+
+ for (unsigned i = 0; i < NumArgs; ++i) {
+ MVT::ValueType VT = Op.Val->getValueType(i);
+ std::pair<FALocInfo, FALocInfo> Loc = FormalArgLocs[i];
+ SDOperand ArgValue;
+ if (Loc.first.Kind == FALocInfo::StackFrameLoc) {
+ // Create the SelectionDAG nodes corresponding to a load from this
+ // parameter.
+ SDOperand FIN = DAG.getFrameIndex(Loc.first.Loc, MVT::i32);
+ ArgValue = DAG.getLoad(Op.Val->getValueType(i), DAG.getEntryNode(), FIN,
+ DAG.getSrcValue(NULL));
+ } else {
+ // Must be a CopyFromReg
+ ArgValue= DAG.getCopyFromReg(DAG.getEntryNode(), Loc.first.Loc,
+ Loc.first.Typ);
+ }
+
+ if (Loc.second.Kind != FALocInfo::None) {
+ SDOperand ArgValue2;
+ if (Loc.second.Kind == FALocInfo::StackFrameLoc) {
+ // Create the SelectionDAG nodes corresponding to a load from this
+ // parameter.
+ SDOperand FIN = DAG.getFrameIndex(Loc.second.Loc, MVT::i32);
+ ArgValue2 = DAG.getLoad(Op.Val->getValueType(i), DAG.getEntryNode(),
+ FIN, DAG.getSrcValue(NULL));
+ } else {
+ // Must be a CopyFromReg
+ ArgValue2 = DAG.getCopyFromReg(DAG.getEntryNode(),
+ Loc.second.Loc, Loc.second.Typ);
+ }
+ ArgValue = DAG.getNode(ISD::BUILD_PAIR, VT, ArgValue, ArgValue2);
+ }
+ FormalArgs.push_back(ArgValue);
}
- return ArgValues;
+
+ // Provide a chain. Note that this isn't the right one, but it works as well
+ // as before.
+ FormalArgs.push_back(DAG.getEntryNode());
}
std::pair<SDOperand, SDOperand>
Ops.push_back(InFlag);
// FIXME: Do not generate X86ISD::TAILCALL for now.
- Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
+ Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
+ NodeTys, Ops);
InFlag = Chain.getValue(1);
NodeTys.clear();
/// specific condition code. It returns a false if it cannot do a direct
/// translation. X86CC is the translated CondCode. Flip is set to true if the
/// the order of comparison operands should be flipped.
-static bool translateX86CC(SDOperand CC, bool isFP, unsigned &X86CC,
- bool &Flip) {
- ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
+static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
+ unsigned &X86CC, bool &Flip) {
Flip = false;
X86CC = X86ISD::COND_INVALID;
if (!isFP) {
default: break;
case ISD::SETUEQ:
case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
- case ISD::SETOLE: Flip = true; // Fallthrough
+ case ISD::SETOLT: Flip = true; // Fallthrough
case ISD::SETOGT:
case ISD::SETGT: X86CC = X86ISD::COND_A; break;
- case ISD::SETOLT: Flip = true; // Fallthrough
+ case ISD::SETOLE: Flip = true; // Fallthrough
case ISD::SETOGE:
case ISD::SETGE: X86CC = X86ISD::COND_AE; break;
- case ISD::SETUGE: Flip = true; // Fallthrough
+ case ISD::SETUGT: Flip = true; // Fallthrough
case ISD::SETULT:
case ISD::SETLT: X86CC = X86ISD::COND_B; break;
- case ISD::SETUGT: Flip = true; // Fallthrough
+ case ISD::SETUGE: Flip = true; // Fallthrough
case ISD::SETULE:
case ISD::SETLE: X86CC = X86ISD::COND_BE; break;
case ISD::SETONE:
return X86CC != X86ISD::COND_INVALID;
}
+static bool translateX86CC(SDOperand CC, bool isFP, unsigned &X86CC,
+ bool &Flip) {
+ return translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC, Flip);
+}
+
/// hasFPCMov - is there a floating point cmov for the specific X86 condition
/// code. Current x86 isa includes the following FP cmov instructions:
/// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
case X86::CMOV_FR32:
- case X86::CMOV_FR64: {
+ case X86::CMOV_FR64:
+ case X86::CMOV_V4F32:
+ case X86::CMOV_V2F64:
+ case X86::CMOV_V2I64: {
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
// destination vreg to set, the condition code register to branch on, the
// Load the old value of the high byte of the control word...
unsigned OldCW =
- F->getSSARegMap()->createVirtualRegister(X86::R16RegisterClass);
+ F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
// Set the high part to be round to zero...
(GV->isExternal() && !GV->hasNotBeenReadFromBytecode()));
}
+/// 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(SDOperand Op, unsigned Low, unsigned Hi) {
+ if (Op.getOpcode() == ISD::UNDEF)
+ return true;
+
+ unsigned Val = cast<ConstantSDNode>(Op)->getValue();
+ return (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(SDOperand Op, unsigned Val) {
+ if (Op.getOpcode() == ISD::UNDEF)
+ return true;
+ return cast<ConstantSDNode>(Op)->getValue() == Val;
+}
+
/// 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) {
return false;
// Upper quadword copied in order.
- for (unsigned i = 4; i != 8; ++i) {
- SDOperand Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Arg)->getValue() != i)
+ for (unsigned i = 4; i != 8; ++i)
+ if (!isUndefOrEqual(N->getOperand(i), i))
return false;
- }
// Lower quadword shuffled.
- for (unsigned i = 0; i != 4; ++i) {
- SDOperand Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
- if (Val > 4)
+ for (unsigned i = 0; i != 4; ++i)
+ if (!isUndefOrInRange(N->getOperand(i), 0, 4))
return false;
- }
return true;
}
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to SHUFP*.
-bool X86::isSHUFPMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- unsigned NumElems = N->getNumOperands();
- if (NumElems == 2) {
- // The only case that ought be handled by SHUFPD is
- // Dest { 2, 1 } <= shuffle( Dest { 1, 0 }, Src { 3, 2 }
- // Expect bit 0 == 1, bit1 == 2
- SDOperand Bit0 = N->getOperand(0);
- if (Bit0.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit0) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit0)->getValue() != 1)
- return false;
- }
+static bool isSHUFPMask(std::vector<SDOperand> &N) {
+ unsigned NumElems = N.size();
+ if (NumElems != 2 && NumElems != 4) return false;
- SDOperand Bit1 = N->getOperand(1);
- if (Bit1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit1)->getValue() != 2)
- return false;
- }
+ unsigned Half = NumElems / 2;
+ for (unsigned i = 0; i < Half; ++i)
+ if (!isUndefOrInRange(N[i], 0, NumElems))
+ return false;
+ for (unsigned i = Half; i < NumElems; ++i)
+ if (!isUndefOrInRange(N[i], NumElems, NumElems*2))
+ return false;
- return true;
- }
+ return true;
+}
- if (NumElems != 4) return false;
-
- // Each half must refer to only one of the vector.
- for (unsigned i = 0; i < 2; ++i) {
- SDOperand Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
- if (Val >= 4) return false;
- }
- for (unsigned i = 2; i < 4; ++i) {
- SDOperand Arg = N->getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
- if (Val < 4) return false;
- }
+bool X86::isSHUFPMask(SDNode *N) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return ::isSHUFPMask(Ops);
+}
+/// isCommutedSHUFP - Returns true if the shuffle mask is except
+/// 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.
+static bool isCommutedSHUFP(std::vector<SDOperand> &Ops) {
+ unsigned NumElems = Ops.size();
+ if (NumElems != 2 && NumElems != 4) return false;
+
+ unsigned Half = NumElems / 2;
+ for (unsigned i = 0; i < Half; ++i)
+ if (!isUndefOrInRange(Ops[i], NumElems, NumElems*2))
+ return false;
+ for (unsigned i = Half; i < NumElems; ++i)
+ if (!isUndefOrInRange(Ops[i], 0, NumElems))
+ return false;
return true;
}
+static bool isCommutedSHUFP(SDNode *N) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return isCommutedSHUFP(Ops);
+}
+
/// 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) {
return false;
// Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
- SDOperand Bit0 = N->getOperand(0);
- SDOperand Bit1 = N->getOperand(1);
- SDOperand Bit2 = N->getOperand(2);
- SDOperand Bit3 = N->getOperand(3);
-
- if (Bit0.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit0) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit0)->getValue() != 6)
- return false;
- }
+ return isUndefOrEqual(N->getOperand(0), 6) &&
+ isUndefOrEqual(N->getOperand(1), 7) &&
+ isUndefOrEqual(N->getOperand(2), 2) &&
+ isUndefOrEqual(N->getOperand(3), 3);
+}
- if (Bit1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit1)->getValue() != 7)
- return false;
- }
+/// 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);
+
+ unsigned NumElems = N->getNumOperands();
+ if (NumElems != 2 && NumElems != 4)
+ return false;
- if (Bit2.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit2) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit2)->getValue() != 2)
+ for (unsigned i = 0; i < NumElems/2; ++i)
+ if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
return false;
- }
- if (Bit3.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit3) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit3)->getValue() != 3)
+ for (unsigned i = NumElems/2; i < NumElems; ++i)
+ if (!isUndefOrEqual(N->getOperand(i), i))
return false;
- }
return true;
}
-/// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVHLPS.
-bool X86::isMOVLHPSMask(SDNode *N) {
+/// 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);
- if (N->getNumOperands() != 4)
+ unsigned NumElems = N->getNumOperands();
+ if (NumElems != 2 && NumElems != 4)
return false;
- // Expect bit0 == 0, bit1 == 1, bit2 == 4, bit3 == 5
- SDOperand Bit0 = N->getOperand(0);
- SDOperand Bit1 = N->getOperand(1);
- SDOperand Bit2 = N->getOperand(2);
- SDOperand Bit3 = N->getOperand(3);
-
- if (Bit0.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit0) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit0)->getValue() != 0)
- return false;
- }
-
- if (Bit1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit1)->getValue() != 1)
- return false;
- }
-
- if (Bit2.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit2) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit2)->getValue() != 4)
+ for (unsigned i = 0; i < NumElems/2; ++i)
+ if (!isUndefOrEqual(N->getOperand(i), i))
return false;
- }
- if (Bit3.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(Bit3) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(Bit3)->getValue() != 5)
+ for (unsigned i = 0; i < NumElems/2; ++i) {
+ SDOperand Arg = N->getOperand(i + NumElems/2);
+ if (!isUndefOrEqual(Arg, i + NumElems))
return false;
}
/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
-bool X86::isUNPCKLMask(SDNode *N) {
- assert(N->getOpcode() == ISD::BUILD_VECTOR);
-
- unsigned NumElems = N->getNumOperands();
+bool static isUNPCKLMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
+ unsigned NumElems = N.size();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
- SDOperand BitI = N->getOperand(i);
- SDOperand BitI1 = N->getOperand(i+1);
-
- if (BitI.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI)->getValue() != j)
+ SDOperand BitI = N[i];
+ SDOperand BitI1 = N[i+1];
+ if (!isUndefOrEqual(BitI, j))
+ return false;
+ if (V2IsSplat) {
+ if (isUndefOrEqual(BitI1, NumElems))
return false;
- }
-
- if (BitI1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI1)->getValue() != j + NumElems)
+ } else {
+ if (!isUndefOrEqual(BitI1, j + NumElems))
return false;
}
}
return true;
}
-/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to UNPCKH.
-bool X86::isUNPCKHMask(SDNode *N) {
+bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return ::isUNPCKLMask(Ops, V2IsSplat);
+}
- unsigned NumElems = N->getNumOperands();
+/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to UNPCKH.
+bool static isUNPCKHMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
+ unsigned NumElems = N.size();
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
- SDOperand BitI = N->getOperand(i);
- SDOperand BitI1 = N->getOperand(i+1);
-
- if (BitI.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI)->getValue() != j + NumElems/2)
+ SDOperand BitI = N[i];
+ SDOperand BitI1 = N[i+1];
+ if (!isUndefOrEqual(BitI, j + NumElems/2))
+ return false;
+ if (V2IsSplat) {
+ if (isUndefOrEqual(BitI1, NumElems))
return false;
- }
-
- if (BitI1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI1)->getValue() != j + NumElems/2 + NumElems)
+ } else {
+ if (!isUndefOrEqual(BitI1, j + NumElems/2 + NumElems))
return false;
}
}
return true;
}
+bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return ::isUNPCKHMask(Ops, 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>
SDOperand BitI = N->getOperand(i);
SDOperand BitI1 = N->getOperand(i+1);
- if (BitI.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI)->getValue() != j)
- return false;
- }
+ if (!isUndefOrEqual(BitI, j))
+ return false;
+ if (!isUndefOrEqual(BitI1, j))
+ return false;
+ }
+
+ return true;
+}
+
+/// 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.
+static bool isMOVLMask(std::vector<SDOperand> &N) {
+ unsigned NumElems = N.size();
+ if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+ return false;
+
+ if (!isUndefOrEqual(N[0], NumElems))
+ return false;
+
+ for (unsigned i = 1; i < NumElems; ++i) {
+ SDOperand Arg = N[i];
+ if (!isUndefOrEqual(Arg, i))
+ return false;
+ }
+
+ return true;
+}
+
+bool X86::isMOVLMask(SDNode *N) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return ::isMOVLMask(Ops);
+}
+
+/// 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.
+static bool isCommutedMOVL(std::vector<SDOperand> &Ops, bool V2IsSplat = false) {
+ unsigned NumElems = Ops.size();
+ if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+ return false;
+
+ if (!isUndefOrEqual(Ops[0], 0))
+ return false;
- if (BitI1.getOpcode() != ISD::UNDEF) {
- assert(isa<ConstantSDNode>(BitI1) && "Invalid VECTOR_SHUFFLE mask!");
- if (cast<ConstantSDNode>(BitI1)->getValue() != j)
+ for (unsigned i = 1; i < NumElems; ++i) {
+ SDOperand Arg = Ops[i];
+ if (V2IsSplat) {
+ if (!isUndefOrEqual(Arg, NumElems))
+ return false;
+ } else {
+ if (!isUndefOrEqual(Arg, i+NumElems))
return false;
}
}
return true;
}
+static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+ std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
+ return isCommutedMOVL(Ops, V2IsSplat);
+}
-/// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
-/// a splat of a single element.
-bool X86::isSplatMask(SDNode *N) {
+/// 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);
- // We can only splat 64-bit, and 32-bit quantities.
- if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
+ if (N->getNumOperands() != 4)
+ return false;
+
+ // Expect 1, 1, 3, 3
+ for (unsigned i = 0; i < 2; ++i) {
+ SDOperand Arg = N->getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val != 1) return false;
+ }
+
+ bool HasHi = false;
+ for (unsigned i = 2; i < 4; ++i) {
+ SDOperand Arg = N->getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val != 3) return false;
+ HasHi = true;
+ }
+
+ // Don't use movshdup if it can be done with a shufps.
+ 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)
return false;
+ // Expect 0, 0, 2, 2
+ for (unsigned i = 0; i < 2; ++i) {
+ SDOperand Arg = N->getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val != 0) return false;
+ }
+
+ bool HasHi = false;
+ for (unsigned i = 2; i < 4; ++i) {
+ SDOperand Arg = N->getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val != 2) return false;
+ HasHi = true;
+ }
+
+ // Don't use movshdup if it can be done with a shufps.
+ return HasHi;
+}
+
+/// 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.
- SDOperand Elt = N->getOperand(0);
- assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
- for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i) {
+ unsigned NumElems = N->getNumOperands();
+ SDOperand ElementBase;
+ unsigned i = 0;
+ for (; i != NumElems; ++i) {
+ SDOperand Elt = N->getOperand(i);
+ if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt)) {
+ ElementBase = Elt;
+ break;
+ }
+ }
+
+ if (!ElementBase.Val)
+ return false;
+
+ for (; i != NumElems; ++i) {
SDOperand Arg = N->getOperand(i);
if (Arg.getOpcode() == ISD::UNDEF) continue;
assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
- if (Arg != Elt) return false;
+ if (Arg != ElementBase) return false;
}
// Make sure it is a splat of the first vector operand.
- return cast<ConstantSDNode>(Elt)->getValue() < N->getNumOperands();
+ return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
+}
+
+/// 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);
}
/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
return Mask;
}
-/// NormalizeVectorShuffle - Swap vector_shuffle operands (as well as
-/// values in ther permute mask if needed. Use V1 as second vector if it is
-/// undef. Return an empty SDOperand is it is already well formed.
-static SDOperand NormalizeVectorShuffle(SDOperand Op, SelectionDAG &DAG) {
- SDOperand V1 = Op.getOperand(0);
- SDOperand V2 = Op.getOperand(1);
- SDOperand Mask = Op.getOperand(2);
- MVT::ValueType VT = Op.getValueType();
- unsigned NumElems = Mask.getNumOperands();
- SDOperand Half1 = Mask.getOperand(0);
- SDOperand Half2 = Mask.getOperand(NumElems/2);
- bool V2Undef = false;
- if (V2.getOpcode() == ISD::UNDEF) {
- V2Undef = true;
- V2 = V1;
- }
-
- if (cast<ConstantSDNode>(Half1)->getValue() >= NumElems &&
- cast<ConstantSDNode>(Half2)->getValue() < NumElems) {
- // Swap the operands and change mask.
- std::vector<SDOperand> MaskVec;
- for (unsigned i = NumElems / 2; i != NumElems; ++i)
- MaskVec.push_back(Mask.getOperand(i));
- for (unsigned i = 0; i != NumElems / 2; ++i)
- MaskVec.push_back(Mask.getOperand(i));
- Mask =
- DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(), MaskVec);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1, Mask);
- }
-
- if (V2Undef)
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
-
- return Op;
-}
-
/// 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.
return true;
}
-/// LowerOperation - Provide custom lowering hooks for some operations.
-///
-SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
- switch (Op.getOpcode()) {
- default: assert(0 && "Should not custom lower this!");
- case ISD::SHL_PARTS:
- case ISD::SRA_PARTS:
- case ISD::SRL_PARTS: {
- assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
- "Not an i64 shift!");
- bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
- SDOperand ShOpLo = Op.getOperand(0);
- SDOperand ShOpHi = Op.getOperand(1);
- SDOperand ShAmt = Op.getOperand(2);
- SDOperand Tmp1 = isSRA ? DAG.getNode(ISD::SRA, MVT::i32, ShOpHi,
- DAG.getConstant(31, MVT::i8))
- : DAG.getConstant(0, MVT::i32);
+/// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
+/// values in ther permute mask.
+static SDOperand CommuteVectorShuffle(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand V1 = Op.getOperand(0);
+ SDOperand V2 = Op.getOperand(1);
+ SDOperand Mask = Op.getOperand(2);
+ MVT::ValueType VT = Op.getValueType();
+ MVT::ValueType MaskVT = Mask.getValueType();
+ MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
+ unsigned NumElems = Mask.getNumOperands();
+ std::vector<SDOperand> MaskVec;
- SDOperand Tmp2, Tmp3;
- if (Op.getOpcode() == ISD::SHL_PARTS) {
- Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
- Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
- } else {
- Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
- Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
+ for (unsigned i = 0; i != NumElems; ++i) {
+ SDOperand Arg = Mask.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) {
+ MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
+ continue;
}
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val < NumElems)
+ MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
+ else
+ MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
+ }
- SDOperand InFlag = DAG.getNode(X86ISD::TEST, MVT::Flag,
- ShAmt, DAG.getConstant(32, MVT::i8));
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1, Mask);
+}
- SDOperand Hi, Lo;
- SDOperand CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
+/// 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)
+ return false;
+ for (unsigned i = 0, e = 2; i != e; ++i)
+ if (!isUndefOrEqual(Mask->getOperand(i), i+2))
+ return false;
+ for (unsigned i = 2; i != 4; ++i)
+ if (!isUndefOrEqual(Mask->getOperand(i), i+4))
+ return false;
+ return true;
+}
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::i32);
- Tys.push_back(MVT::Flag);
- std::vector<SDOperand> 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, Tys, Ops);
- InFlag = Hi.getValue(1);
+/// isScalarLoadToVector - Returns true if the node is a scalar load that
+/// is promoted to a vector.
+static inline bool isScalarLoadToVector(SDNode *N) {
+ if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ N = N->getOperand(0).Val;
+ return (N->getOpcode() == ISD::LOAD);
+ }
+ return false;
+}
- Ops.clear();
- Ops.push_back(Tmp3);
- Ops.push_back(Tmp1);
- Ops.push_back(CC);
- Ops.push_back(InFlag);
- Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
- } else {
- Ops.push_back(Tmp2);
- Ops.push_back(Tmp3);
- Ops.push_back(CC);
- Ops.push_back(InFlag);
- Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
- InFlag = Lo.getValue(1);
+/// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
+/// match movlp{s|d}. The lower half elements should come from lower half of
+/// 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 *Mask) {
+ if (V1->getOpcode() != ISD::LOAD && !isScalarLoadToVector(V1))
+ return false;
- Ops.clear();
- Ops.push_back(Tmp3);
- Ops.push_back(Tmp1);
- Ops.push_back(CC);
- Ops.push_back(InFlag);
- Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
- }
+ unsigned NumElems = Mask->getNumOperands();
+ if (NumElems != 2 && NumElems != 4)
+ return false;
+ for (unsigned i = 0, e = NumElems/2; i != e; ++i)
+ if (!isUndefOrEqual(Mask->getOperand(i), i))
+ return false;
+ for (unsigned i = NumElems/2; i != NumElems; ++i)
+ if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
+ return false;
+ return true;
+}
- Tys.clear();
- Tys.push_back(MVT::i32);
- Tys.push_back(MVT::i32);
- Ops.clear();
- Ops.push_back(Lo);
- Ops.push_back(Hi);
- return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
+/// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
+/// all the same.
+static bool isSplatVector(SDNode *N) {
+ if (N->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+
+ SDOperand SplatValue = N->getOperand(0);
+ for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
+ if (N->getOperand(i) != SplatValue)
+ return false;
+ return true;
+}
+
+/// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
+/// that point to V2 points to its first element.
+static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
+ assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
+
+ bool Changed = false;
+ std::vector<SDOperand> MaskVec;
+ unsigned NumElems = Mask.getNumOperands();
+ for (unsigned i = 0; i != NumElems; ++i) {
+ SDOperand Arg = Mask.getOperand(i);
+ if (Arg.getOpcode() != ISD::UNDEF) {
+ unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
+ if (Val > NumElems) {
+ Arg = DAG.getConstant(NumElems, Arg.getValueType());
+ Changed = true;
+ }
+ }
+ MaskVec.push_back(Arg);
}
- case ISD::SINT_TO_FP: {
- assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
- Op.getOperand(0).getValueType() >= MVT::i16 &&
- "Unknown SINT_TO_FP to lower!");
- SDOperand Result;
- MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
- unsigned Size = MVT::getSizeInBits(SrcVT)/8;
- MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
- SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
- SDOperand Chain = DAG.getNode(ISD::STORE, MVT::Other,
- DAG.getEntryNode(), Op.getOperand(0),
- StackSlot, DAG.getSrcValue(NULL));
+ if (Changed)
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(), MaskVec);
+ return Mask;
+}
- // Build the FILD
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::f64);
- Tys.push_back(MVT::Other);
- if (X86ScalarSSE) Tys.push_back(MVT::Flag);
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(StackSlot);
- Ops.push_back(DAG.getValueType(SrcVT));
- Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
- Tys, Ops);
-
- if (X86ScalarSSE) {
- Chain = Result.getValue(1);
- SDOperand InFlag = Result.getValue(2);
-
- // FIXME: Currently the FST is flagged to the FILD_FLAG. This
- // shouldn't be necessary except that RFP cannot be live across
- // multiple blocks. When stackifier is fixed, they can be uncoupled.
- MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
- SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::Other);
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(Result);
- Ops.push_back(StackSlot);
- Ops.push_back(DAG.getValueType(Op.getValueType()));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
- Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
- DAG.getSrcValue(NULL));
+/// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
+/// operation of specified width.
+static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
+
+ std::vector<SDOperand> MaskVec;
+ MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
+ for (unsigned i = 1; i != NumElems; ++i)
+ MaskVec.push_back(DAG.getConstant(i, BaseVT));
+ return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+}
+
+/// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
+/// of specified width.
+static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<SDOperand> MaskVec;
+ 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));
+ }
+ return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+}
+
+/// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
+/// of specified width.
+static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
+ unsigned Half = NumElems/2;
+ std::vector<SDOperand> MaskVec;
+ 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, MaskVT, MaskVec);
+}
+
+/// getZeroVector - Returns a vector of specified type with all zero elements.
+///
+static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
+ assert(MVT::isVector(VT) && "Expected a vector type");
+ unsigned NumElems = getVectorNumElements(VT);
+ MVT::ValueType EVT = MVT::getVectorBaseType(VT);
+ bool isFP = MVT::isFloatingPoint(EVT);
+ SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
+ std::vector<SDOperand> ZeroVec(NumElems, Zero);
+ return DAG.getNode(ISD::BUILD_VECTOR, VT, ZeroVec);
+}
+
+/// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
+///
+static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand V1 = Op.getOperand(0);
+ SDOperand Mask = Op.getOperand(2);
+ MVT::ValueType VT = Op.getValueType();
+ unsigned NumElems = Mask.getNumOperands();
+ Mask = getUnpacklMask(NumElems, DAG);
+ while (NumElems != 4) {
+ V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
+ NumElems >>= 1;
+ }
+ V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
+
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
+ Mask = getZeroVector(MaskVT, DAG);
+ SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
+ DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
+ return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
+}
+
+/// isZeroNode - Returns true if Elt is a constant zero or a floating point
+/// constant +0.0.
+static inline bool isZeroNode(SDOperand Elt) {
+ return ((isa<ConstantSDNode>(Elt) &&
+ cast<ConstantSDNode>(Elt)->getValue() == 0) ||
+ (isa<ConstantFPSDNode>(Elt) &&
+ cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
+}
+
+/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
+/// vector and zero or undef vector.
+static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
+ unsigned NumElems, unsigned Idx,
+ bool isZero, SelectionDAG &DAG) {
+ SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
+ SDOperand Zero = DAG.getConstant(0, EVT);
+ std::vector<SDOperand> MaskVec(NumElems, Zero);
+ MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
+ SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
+}
+
+/// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
+///
+static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
+ unsigned NumNonZero, unsigned NumZero,
+ SelectionDAG &DAG) {
+ if (NumNonZero > 8)
+ return SDOperand();
+
+ SDOperand V(0, 0);
+ bool First = true;
+ for (unsigned i = 0; i < 16; ++i) {
+ bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
+ if (ThisIsNonZero && First) {
+ if (NumZero)
+ V = getZeroVector(MVT::v8i16, DAG);
+ else
+ V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
+ First = false;
}
- return Result;
+ if ((i & 1) != 0) {
+ SDOperand ThisElt(0, 0), LastElt(0, 0);
+ bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
+ if (LastIsNonZero) {
+ LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
+ }
+ if (ThisIsNonZero) {
+ ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
+ ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
+ ThisElt, DAG.getConstant(8, MVT::i8));
+ if (LastIsNonZero)
+ ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
+ } else
+ ThisElt = LastElt;
+
+ if (ThisElt.Val)
+ V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
+ DAG.getConstant(i/2, MVT::i32));
+ }
}
- case ISD::FP_TO_SINT: {
- assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
- "Unknown FP_TO_SINT to lower!");
- // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
- // stack slot.
- MachineFunction &MF = DAG.getMachineFunction();
- unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
- int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
- SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
- unsigned Opc;
- switch (Op.getValueType()) {
- 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;
- case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
+ return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
+}
+
+/// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
+///
+static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
+ unsigned NumNonZero, unsigned NumZero,
+ SelectionDAG &DAG) {
+ if (NumNonZero > 4)
+ return SDOperand();
+
+ SDOperand V(0, 0);
+ bool First = true;
+ for (unsigned i = 0; i < 8; ++i) {
+ bool isNonZero = (NonZeros & (1 << i)) != 0;
+ if (isNonZero) {
+ if (First) {
+ if (NumZero)
+ V = getZeroVector(MVT::v8i16, DAG);
+ else
+ V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
+ First = false;
+ }
+ V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
+ DAG.getConstant(i, MVT::i32));
}
+ }
- SDOperand Chain = DAG.getEntryNode();
- SDOperand Value = Op.getOperand(0);
- if (X86ScalarSSE) {
- assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value, StackSlot,
- DAG.getSrcValue(0));
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::f64);
- Tys.push_back(MVT::Other);
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(StackSlot);
- Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
- Value = DAG.getNode(X86ISD::FLD, Tys, Ops);
- Chain = Value.getValue(1);
- SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
- StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ return V;
+}
+
+SDOperand
+X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
+ // All zero's are handled with pxor.
+ if (ISD::isBuildVectorAllZeros(Op.Val))
+ return Op;
+
+ // All one's are handled with pcmpeqd.
+ if (ISD::isBuildVectorAllOnes(Op.Val))
+ return Op;
+
+ MVT::ValueType VT = Op.getValueType();
+ MVT::ValueType EVT = MVT::getVectorBaseType(VT);
+ unsigned EVTBits = MVT::getSizeInBits(EVT);
+
+ unsigned NumElems = Op.getNumOperands();
+ unsigned NumZero = 0;
+ unsigned NumNonZero = 0;
+ unsigned NonZeros = 0;
+ std::set<SDOperand> Values;
+ for (unsigned i = 0; i < NumElems; ++i) {
+ SDOperand Elt = Op.getOperand(i);
+ if (Elt.getOpcode() != ISD::UNDEF) {
+ Values.insert(Elt);
+ if (isZeroNode(Elt))
+ NumZero++;
+ else {
+ NonZeros |= (1 << i);
+ NumNonZero++;
+ }
}
+ }
- // Build the FP_TO_INT*_IN_MEM
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(Value);
- Ops.push_back(StackSlot);
- SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
+ if (NumNonZero == 0)
+ // Must be a mix of zero and undef. Return a zero vector.
+ return getZeroVector(VT, DAG);
- // Load the result.
- return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
- DAG.getSrcValue(NULL));
+ // Splat is obviously ok. Let legalizer expand it to a shuffle.
+ if (Values.size() == 1)
+ return SDOperand();
+
+ // Special case for single non-zero element.
+ if (NumNonZero == 1) {
+ unsigned Idx = CountTrailingZeros_32(NonZeros);
+ SDOperand Item = Op.getOperand(Idx);
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
+ if (Idx == 0)
+ // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
+ return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
+ NumZero > 0, DAG);
+
+ if (EVTBits == 32) {
+ // Turn it into a shuffle of zero and zero-extended scalar to vector.
+ Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
+ DAG);
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<SDOperand> MaskVec;
+ for (unsigned i = 0; i < NumElems; i++)
+ MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
+ SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
+ DAG.getNode(ISD::UNDEF, VT), Mask);
+ }
}
- case ISD::READCYCLECOUNTER: {
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::Other);
- Tys.push_back(MVT::Flag);
- std::vector<SDOperand> Ops;
- Ops.push_back(Op.getOperand(0));
- SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, Ops);
- Ops.clear();
- Ops.push_back(DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)));
- Ops.push_back(DAG.getCopyFromReg(Ops[0].getValue(1), X86::EDX,
- MVT::i32, Ops[0].getValue(2)));
- Ops.push_back(Ops[1].getValue(1));
- Tys[0] = Tys[1] = MVT::i32;
- Tys.push_back(MVT::Other);
- return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
+
+ // Let legalizer expand 2-widde build_vector's.
+ if (EVTBits == 64)
+ return SDOperand();
+
+ // If element VT is < 32 bits, convert it to inserts into a zero vector.
+ if (EVTBits == 8) {
+ SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG);
+ if (V.Val) return V;
}
- case ISD::FABS: {
- MVT::ValueType VT = Op.getValueType();
- const Type *OpNTy = MVT::getTypeForValueType(VT);
- 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));
- } 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));
+
+ if (EVTBits == 16) {
+ SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG);
+ if (V.Val) return V;
+ }
+
+ // If element VT is == 32 bits, turn it into a number of shuffles.
+ std::vector<SDOperand> V(NumElems);
+ if (NumElems == 4 && NumZero > 0) {
+ for (unsigned i = 0; i < 4; ++i) {
+ bool isZero = !(NonZeros & (1 << i));
+ if (isZero)
+ V[i] = getZeroVector(VT, DAG);
+ else
+ V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
}
- Constant *CS = ConstantStruct::get(CV);
- SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- SDOperand Mask
- = DAG.getNode(X86ISD::LOAD_PACK,
- VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
- return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
- }
- case ISD::FNEG: {
- MVT::ValueType VT = Op.getValueType();
- const Type *OpNTy = MVT::getTypeForValueType(VT);
- 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));
- } 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));
+
+ for (unsigned i = 0; i < 2; ++i) {
+ switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
+ default: break;
+ case 0:
+ V[i] = V[i*2]; // Must be a zero vector.
+ break;
+ case 1:
+ V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
+ getMOVLMask(NumElems, DAG));
+ break;
+ case 2:
+ V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
+ getMOVLMask(NumElems, DAG));
+ break;
+ case 3:
+ V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
+ getUnpacklMask(NumElems, DAG));
+ break;
+ }
}
- Constant *CS = ConstantStruct::get(CV);
- SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
- SDOperand Mask
- = DAG.getNode(X86ISD::LOAD_PACK,
- VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
- return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
- }
- case ISD::SETCC: {
- assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
- SDOperand Cond;
- SDOperand CC = Op.getOperand(2);
- ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
- bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
- bool Flip;
- unsigned X86CC;
- if (translateX86CC(CC, isFP, X86CC, Flip)) {
- if (Flip)
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Op.getOperand(1), Op.getOperand(0));
+
+ // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
+ // clears the upper bits.
+ // FIXME: we can do the same for v4f32 case when we know both parts of
+ // the lower half come from scalar_to_vector (loadf32). We should do
+ // that in post legalizer dag combiner with target specific hooks.
+ if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
+ return V[0];
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<SDOperand> MaskVec;
+ bool Reverse = (NonZeros & 0x3) == 2;
+ for (unsigned i = 0; i < 2; ++i)
+ if (Reverse)
+ MaskVec.push_back(DAG.getConstant(1-i, EVT));
else
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Op.getOperand(0), Op.getOperand(1));
- return DAG.getNode(X86ISD::SETCC, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond);
- } else {
- assert(isFP && "Illegal integer SetCC!");
+ MaskVec.push_back(DAG.getConstant(i, EVT));
+ 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));
+ SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
+ }
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Op.getOperand(0), Op.getOperand(1));
- std::vector<MVT::ValueType> Tys;
- std::vector<SDOperand> Ops;
- switch (SetCCOpcode) {
- default: assert(false && "Illegal floating point SetCC!");
- case ISD::SETOEQ: { // !PF & ZF
- Tys.push_back(MVT::i8);
- Tys.push_back(MVT::Flag);
- Ops.push_back(DAG.getConstant(X86ISD::COND_NP, MVT::i8));
- Ops.push_back(Cond);
- SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
- SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
- DAG.getConstant(X86ISD::COND_E, MVT::i8),
- Tmp1.getValue(1));
- return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
- }
- case ISD::SETUNE: { // PF | !ZF
- Tys.push_back(MVT::i8);
- Tys.push_back(MVT::Flag);
- Ops.push_back(DAG.getConstant(X86ISD::COND_P, MVT::i8));
- Ops.push_back(Cond);
- SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
- SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
- DAG.getConstant(X86ISD::COND_NE, MVT::i8),
- Tmp1.getValue(1));
- return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
- }
+ if (Values.size() > 2) {
+ // 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>
+ SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
+ for (unsigned i = 0; i < NumElems; ++i)
+ V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
+ NumElems >>= 1;
+ while (NumElems != 0) {
+ for (unsigned i = 0; i < NumElems; ++i)
+ V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
+ UnpckMask);
+ NumElems >>= 1;
+ }
+ return V[0];
+ }
+
+ return SDOperand();
+}
+
+SDOperand
+X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand V1 = Op.getOperand(0);
+ SDOperand V2 = Op.getOperand(1);
+ SDOperand PermMask = Op.getOperand(2);
+ MVT::ValueType VT = Op.getValueType();
+ unsigned NumElems = PermMask.getNumOperands();
+ bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
+ bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
+
+ if (isSplatMask(PermMask.Val)) {
+ if (NumElems <= 4) return Op;
+ // Promote it to a v4i32 splat.
+ return PromoteSplat(Op, DAG);
+ }
+
+ if (X86::isMOVLMask(PermMask.Val))
+ return (V1IsUndef) ? V2 : Op;
+
+ if (X86::isMOVSHDUPMask(PermMask.Val) ||
+ X86::isMOVSLDUPMask(PermMask.Val) ||
+ X86::isMOVHLPSMask(PermMask.Val) ||
+ X86::isMOVHPMask(PermMask.Val) ||
+ X86::isMOVLPMask(PermMask.Val))
+ return Op;
+
+ if (ShouldXformToMOVHLPS(PermMask.Val) ||
+ ShouldXformToMOVLP(V1.Val, PermMask.Val))
+ return CommuteVectorShuffle(Op, DAG);
+
+ bool V1IsSplat = isSplatVector(V1.Val) || V1.getOpcode() == ISD::UNDEF;
+ bool V2IsSplat = isSplatVector(V2.Val) || V2.getOpcode() == ISD::UNDEF;
+ if (V1IsSplat && !V2IsSplat) {
+ Op = CommuteVectorShuffle(Op, DAG);
+ V1 = Op.getOperand(0);
+ V2 = Op.getOperand(1);
+ PermMask = Op.getOperand(2);
+ V2IsSplat = true;
+ }
+
+ if (isCommutedMOVL(PermMask.Val, V2IsSplat)) {
+ if (V2IsUndef) return V1;
+ Op = CommuteVectorShuffle(Op, DAG);
+ V1 = Op.getOperand(0);
+ V2 = Op.getOperand(1);
+ PermMask = Op.getOperand(2);
+ 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}.
+ SDOperand NewMask = getMOVLMask(NumElems, DAG);
+ if (NewMask.Val != PermMask.Val)
+ Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
+ }
+ return Op;
+ }
+
+ if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
+ X86::isUNPCKLMask(PermMask.Val) ||
+ X86::isUNPCKHMask(PermMask.Val))
+ 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.
+ SDOperand NewMask = NormalizeMask(PermMask, DAG);
+ if (NewMask.Val != PermMask.Val) {
+ if (X86::isUNPCKLMask(PermMask.Val, true)) {
+ SDOperand NewMask = getUnpacklMask(NumElems, DAG);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
+ } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
+ SDOperand NewMask = getUnpackhMask(NumElems, DAG);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
}
}
}
- case ISD::SELECT: {
- MVT::ValueType VT = Op.getValueType();
- bool isFP = MVT::isFloatingPoint(VT);
- bool isFPStack = isFP && !X86ScalarSSE;
- bool isFPSSE = isFP && X86ScalarSSE;
- bool addTest = false;
- SDOperand Op0 = Op.getOperand(0);
- SDOperand Cond, CC;
- if (Op0.getOpcode() == ISD::SETCC)
- Op0 = LowerOperation(Op0, DAG);
-
- if (Op0.getOpcode() == X86ISD::SETCC) {
- // If condition flag is set by a X86ISD::CMP, then make a copy of it
- // (since flag operand cannot be shared). If the X86ISD::SETCC does not
- // have another use it will be eliminated.
- // If the X86ISD::SETCC has more than one use, then it's probably better
- // to use a test instead of duplicating the X86ISD::CMP (for register
- // pressure reason).
- if (Op0.getOperand(1).getOpcode() == X86ISD::CMP) {
- if (!Op0.hasOneUse()) {
- std::vector<MVT::ValueType> Tys;
- for (unsigned i = 0; i < Op0.Val->getNumValues(); ++i)
- Tys.push_back(Op0.Val->getValueType(i));
- std::vector<SDOperand> Ops;
- for (unsigned i = 0; i < Op0.getNumOperands(); ++i)
- Ops.push_back(Op0.getOperand(i));
- Op0 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
- }
- CC = Op0.getOperand(0);
- Cond = Op0.getOperand(1);
- // Make a copy as flag result cannot be used by more than one.
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Cond.getOperand(0), Cond.getOperand(1));
- addTest =
- isFPStack && !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
- } else
- addTest = true;
- } else
- addTest = true;
+ // Normalize the node to match x86 shuffle ops if needed
+ if (V2.getOpcode() != ISD::UNDEF)
+ if (isCommutedSHUFP(PermMask.Val)) {
+ Op = CommuteVectorShuffle(Op, DAG);
+ V1 = Op.getOperand(0);
+ V2 = Op.getOperand(1);
+ PermMask = Op.getOperand(2);
+ }
- if (addTest) {
- CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
- Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
+ // If VT is integer, try PSHUF* first, then SHUFP*.
+ if (MVT::isInteger(VT)) {
+ if (X86::isPSHUFDMask(PermMask.Val) ||
+ X86::isPSHUFHWMask(PermMask.Val) ||
+ X86::isPSHUFLWMask(PermMask.Val)) {
+ if (V2.getOpcode() != ISD::UNDEF)
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
+ DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
+ return Op;
}
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(Op.getValueType());
- Tys.push_back(MVT::Flag);
- std::vector<SDOperand> 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, Tys, Ops);
- }
- case ISD::BRCOND: {
- bool addTest = false;
- SDOperand Cond = Op.getOperand(1);
- SDOperand Dest = Op.getOperand(2);
- SDOperand CC;
- if (Cond.getOpcode() == ISD::SETCC)
- Cond = LowerOperation(Cond, DAG);
-
- if (Cond.getOpcode() == X86ISD::SETCC) {
- // If condition flag is set by a X86ISD::CMP, then make a copy of it
- // (since flag operand cannot be shared). If the X86ISD::SETCC does not
- // have another use it will be eliminated.
- // If the X86ISD::SETCC has more than one use, then it's probably better
- // to use a test instead of duplicating the X86ISD::CMP (for register
- // pressure reason).
- if (Cond.getOperand(1).getOpcode() == X86ISD::CMP) {
- if (!Cond.hasOneUse()) {
- std::vector<MVT::ValueType> Tys;
- for (unsigned i = 0; i < Cond.Val->getNumValues(); ++i)
- Tys.push_back(Cond.Val->getValueType(i));
- std::vector<SDOperand> Ops;
- for (unsigned i = 0; i < Cond.getNumOperands(); ++i)
- Ops.push_back(Cond.getOperand(i));
- Cond = DAG.getNode(X86ISD::SETCC, Tys, Ops);
+ if (X86::isSHUFPMask(PermMask.Val))
+ return Op;
+
+ // Handle v8i16 shuffle high / low shuffle node pair.
+ if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<SDOperand> MaskVec;
+ for (unsigned i = 0; i != 4; ++i)
+ MaskVec.push_back(PermMask.getOperand(i));
+ for (unsigned i = 4; i != 8; ++i)
+ MaskVec.push_back(DAG.getConstant(i, BaseVT));
+ SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
+ MaskVec.clear();
+ for (unsigned i = 0; i != 4; ++i)
+ MaskVec.push_back(DAG.getConstant(i, BaseVT));
+ for (unsigned i = 4; i != 8; ++i)
+ MaskVec.push_back(PermMask.getOperand(i));
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
+ }
+ } else {
+ // Floating point cases in the other order.
+ if (X86::isSHUFPMask(PermMask.Val))
+ return Op;
+ if (X86::isPSHUFDMask(PermMask.Val) ||
+ X86::isPSHUFHWMask(PermMask.Val) ||
+ X86::isPSHUFLWMask(PermMask.Val)) {
+ if (V2.getOpcode() != ISD::UNDEF)
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
+ DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
+ return Op;
+ }
+ }
+
+ if (NumElems == 4) {
+ MVT::ValueType MaskVT = PermMask.getValueType();
+ MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<std::pair<int, int> > Locs;
+ Locs.reserve(NumElems);
+ std::vector<SDOperand> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
+ std::vector<SDOperand> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
+ unsigned NumHi = 0;
+ unsigned NumLo = 0;
+ // If no more than two elements come from either vector. This can be
+ // 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.
+ for (unsigned i = 0; i != NumElems; ++i) {
+ SDOperand Elt = PermMask.getOperand(i);
+ if (Elt.getOpcode() == ISD::UNDEF) {
+ Locs[i] = std::make_pair(-1, -1);
+ } else {
+ unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
+ if (Val < NumElems) {
+ Locs[i] = std::make_pair(0, NumLo);
+ Mask1[NumLo] = Elt;
+ NumLo++;
+ } else {
+ Locs[i] = std::make_pair(1, NumHi);
+ if (2+NumHi < NumElems)
+ Mask1[2+NumHi] = Elt;
+ NumHi++;
}
+ }
+ }
+ if (NumLo <= 2 && NumHi <= 2) {
+ V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, Mask1));
+ for (unsigned i = 0; i != NumElems; ++i) {
+ if (Locs[i].first == -1)
+ continue;
+ else {
+ unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
+ Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
+ Mask2[i] = DAG.getConstant(Idx, MaskEVT);
+ }
+ }
- CC = Cond.getOperand(0);
- Cond = Cond.getOperand(1);
- // Make a copy as flag result cannot be used by more than one.
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Cond.getOperand(0), Cond.getOperand(1));
- } else
- addTest = true;
- } else
- addTest = true;
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, Mask2));
+ }
- if (addTest) {
- CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
- Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
+ // Break it into (shuffle shuffle_hi, shuffle_lo).
+ Locs.clear();
+ std::vector<SDOperand> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
+ std::vector<SDOperand> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
+ std::vector<SDOperand> *MaskPtr = &LoMask;
+ unsigned MaskIdx = 0;
+ unsigned LoIdx = 0;
+ unsigned HiIdx = NumElems/2;
+ for (unsigned i = 0; i != NumElems; ++i) {
+ if (i == NumElems/2) {
+ MaskPtr = &HiMask;
+ MaskIdx = 1;
+ LoIdx = 0;
+ HiIdx = NumElems/2;
+ }
+ SDOperand Elt = PermMask.getOperand(i);
+ if (Elt.getOpcode() == ISD::UNDEF) {
+ Locs[i] = std::make_pair(-1, -1);
+ } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
+ Locs[i] = std::make_pair(MaskIdx, LoIdx);
+ (*MaskPtr)[LoIdx] = Elt;
+ LoIdx++;
+ } else {
+ Locs[i] = std::make_pair(MaskIdx, HiIdx);
+ (*MaskPtr)[HiIdx] = Elt;
+ HiIdx++;
+ }
}
- return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
- Op.getOperand(0), Op.getOperand(2), CC, Cond);
- }
- case ISD::MEMSET: {
- SDOperand InFlag(0, 0);
- SDOperand Chain = Op.getOperand(0);
- unsigned Align =
- (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
- 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 ((Align & 3) != 0 ||
- (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
- MVT::ValueType IntPtr = getPointerTy();
- const Type *IntPtrTy = getTargetData().getIntPtrType();
- std::vector<std::pair<SDOperand, const Type*> > Args;
- Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
- // Extend the ubyte argument to be an int value for the call.
- SDOperand Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
- Args.push_back(std::make_pair(Val, IntPtrTy));
- Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
- std::pair<SDOperand,SDOperand> CallResult =
- LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
- DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
- return CallResult.second;
+
+ SDOperand LoShuffle =
+ DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, LoMask));
+ SDOperand HiShuffle =
+ DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, HiMask));
+ std::vector<SDOperand> MaskOps;
+ for (unsigned i = 0; i != NumElems; ++i) {
+ if (Locs[i].first == -1) {
+ MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
+ } else {
+ unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
+ MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
+ }
}
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskOps));
+ }
- MVT::ValueType AVT;
- SDOperand Count;
- ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
- unsigned BytesLeft = 0;
- bool TwoRepStos = false;
- if (ValC) {
- unsigned ValReg;
- unsigned Val = ValC->getValue() & 255;
-
- // If the value is a constant, then we can potentially use larger sets.
- switch (Align & 3) {
- case 2: // WORD aligned
- AVT = MVT::i16;
- Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
- BytesLeft = I->getValue() % 2;
- Val = (Val << 8) | Val;
- ValReg = X86::AX;
- break;
- case 0: // DWORD aligned
- AVT = MVT::i32;
- if (I) {
- Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
- BytesLeft = I->getValue() % 4;
+ return SDOperand();
+}
+
+SDOperand
+X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
+ if (!isa<ConstantSDNode>(Op.getOperand(1)))
+ return SDOperand();
+
+ MVT::ValueType VT = Op.getValueType();
+ // TODO: handle v16i8.
+ if (MVT::getSizeInBits(VT) == 16) {
+ // Transform it so it match pextrw which produces a 32-bit result.
+ MVT::ValueType EVT = (MVT::ValueType)(VT+1);
+ SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
+ Op.getOperand(0), Op.getOperand(1));
+ SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
+ DAG.getValueType(VT));
+ return DAG.getNode(ISD::TRUNCATE, VT, Assert);
+ } else if (MVT::getSizeInBits(VT) == 32) {
+ SDOperand Vec = Op.getOperand(0);
+ unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
+ if (Idx == 0)
+ return Op;
+
+ // SHUFPS the element to the lowest double word, then movss.
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
+ SDOperand IdxNode = DAG.getConstant((Idx < 2) ? Idx : Idx+4,
+ MVT::getVectorBaseType(MaskVT));
+ std::vector<SDOperand> IdxVec;
+ IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
+ IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
+ IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
+ IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
+ SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
+ Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
+ Vec, Vec, Mask);
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
+ DAG.getConstant(0, MVT::i32));
+ } else if (MVT::getSizeInBits(VT) == 64) {
+ SDOperand Vec = Op.getOperand(0);
+ unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
+ if (Idx == 0)
+ return Op;
+
+ // 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::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
+ std::vector<SDOperand> IdxVec;
+ IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
+ IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
+ SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
+ Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
+ Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
+ DAG.getConstant(0, MVT::i32));
+ }
+
+ return SDOperand();
+}
+
+SDOperand
+X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
+ // Transform it so it match pinsrw which expects a 16-bit value in a GR32
+ // as its second argument.
+ MVT::ValueType VT = Op.getValueType();
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
+ SDOperand N0 = Op.getOperand(0);
+ SDOperand N1 = Op.getOperand(1);
+ SDOperand N2 = Op.getOperand(2);
+ if (MVT::getSizeInBits(BaseVT) == 16) {
+ 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);
+ return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
+ } else if (MVT::getSizeInBits(BaseVT) == 32) {
+ unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
+ if (Idx == 0) {
+ // Use a movss.
+ N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
+ MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
+ MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
+ std::vector<SDOperand> MaskVec;
+ MaskVec.push_back(DAG.getConstant(4, BaseVT));
+ for (unsigned i = 1; i <= 3; ++i)
+ MaskVec.push_back(DAG.getConstant(i, BaseVT));
+ return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
+ DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec));
+ } else {
+ // Use two pinsrw instructions to insert a 32 bit value.
+ Idx <<= 1;
+ if (MVT::isFloatingPoint(N1.getValueType())) {
+ if (N1.getOpcode() == ISD::LOAD) {
+ // Just load directly from f32mem to GR32.
+ N1 = DAG.getLoad(MVT::i32, N1.getOperand(0), N1.getOperand(1),
+ N1.getOperand(2));
} else {
- Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
- DAG.getConstant(2, MVT::i8));
- TwoRepStos = true;
+ 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, MVT::i32));
}
- Val = (Val << 8) | Val;
- Val = (Val << 16) | Val;
- ValReg = X86::EAX;
- break;
- default: // Byte aligned
- AVT = MVT::i8;
- Count = Op.getOperand(3);
- ValReg = X86::AL;
- break;
}
+ N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
+ N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
+ DAG.getConstant(Idx, MVT::i32));
+ N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
+ N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
+ DAG.getConstant(Idx+1, MVT::i32));
+ return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
+ }
+ }
+
+ return SDOperand();
+}
- Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
- InFlag);
- InFlag = Chain.getValue(1);
+SDOperand
+X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
+ return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
+}
+
+// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
+// their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
+// one of the above mentioned nodes. It has to be wrapped because otherwise
+// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
+// be used to form addressing mode. These wrapped nodes will be selected
+// into MOV32ri.
+SDOperand
+X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
+ ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
+ SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
+ DAG.getTargetConstantPool(CP->get(), getPointerTy(),
+ CP->getAlignment()));
+ if (Subtarget->isTargetDarwin()) {
+ // With PIC, the address is actually $g + Offset.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC)
+ Result = DAG.getNode(ISD::ADD, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
+ }
+
+ return Result;
+}
+
+SDOperand
+X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
+ GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
+ SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
+ DAG.getTargetGlobalAddress(GV,
+ getPointerTy()));
+ if (Subtarget->isTargetDarwin()) {
+ // With PIC, the address is actually $g + Offset.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC)
+ Result = DAG.getNode(ISD::ADD, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
+ Result);
+
+ // For Darwin, 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 the GV offset field.
+ if (getTargetMachine().getRelocationModel() != Reloc::Static &&
+ DarwinGVRequiresExtraLoad(GV))
+ Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
+ Result, DAG.getSrcValue(NULL));
+ }
+
+ return Result;
+}
+
+SDOperand
+X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
+ const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
+ SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
+ DAG.getTargetExternalSymbol(Sym,
+ getPointerTy()));
+ if (Subtarget->isTargetDarwin()) {
+ // With PIC, the address is actually $g + Offset.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC)
+ Result = DAG.getNode(ISD::ADD, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
+ Result);
+ }
+
+ return Result;
+}
+
+SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
+ "Not an i64 shift!");
+ bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
+ SDOperand ShOpLo = Op.getOperand(0);
+ SDOperand ShOpHi = Op.getOperand(1);
+ SDOperand ShAmt = Op.getOperand(2);
+ SDOperand Tmp1 = isSRA ? DAG.getNode(ISD::SRA, MVT::i32, ShOpHi,
+ DAG.getConstant(31, MVT::i8))
+ : DAG.getConstant(0, MVT::i32);
+
+ SDOperand Tmp2, Tmp3;
+ if (Op.getOpcode() == ISD::SHL_PARTS) {
+ Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
+ Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
} else {
- AVT = MVT::i8;
- Count = Op.getOperand(3);
- Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
- InFlag = Chain.getValue(1);
+ Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
+ Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
}
- Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
- InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
- InFlag = Chain.getValue(1);
+ SDOperand InFlag = DAG.getNode(X86ISD::TEST, MVT::Flag,
+ ShAmt, DAG.getConstant(32, MVT::i8));
+
+ SDOperand Hi, Lo;
+ SDOperand CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::i32);
Tys.push_back(MVT::Flag);
std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(AVT));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_STOS, Tys, 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, Tys, Ops);
+ InFlag = Hi.getValue(1);
- if (TwoRepStos) {
- InFlag = Chain.getValue(1);
- Count = Op.getOperand(3);
- MVT::ValueType CVT = Count.getValueType();
- SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
- DAG.getConstant(3, CVT));
- Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
- InFlag = Chain.getValue(1);
- Tys.clear();
- Tys.push_back(MVT::Other);
- Tys.push_back(MVT::Flag);
Ops.clear();
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(MVT::i8));
+ Ops.push_back(Tmp3);
+ Ops.push_back(Tmp1);
+ Ops.push_back(CC);
+ Ops.push_back(InFlag);
+ Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
+ } else {
+ Ops.push_back(Tmp2);
+ Ops.push_back(Tmp3);
+ Ops.push_back(CC);
Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
- } else if (BytesLeft) {
- // Issue stores for the last 1 - 3 bytes.
- SDOperand Value;
- unsigned Val = ValC->getValue() & 255;
- unsigned Offset = I->getValue() - BytesLeft;
- SDOperand DstAddr = Op.getOperand(1);
- MVT::ValueType AddrVT = DstAddr.getValueType();
- if (BytesLeft >= 2) {
- Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
- DAG.getNode(ISD::ADD, AddrVT, DstAddr,
- DAG.getConstant(Offset, AddrVT)),
- DAG.getSrcValue(NULL));
- BytesLeft -= 2;
- Offset += 2;
+ Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
+ InFlag = 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, Tys, Ops);
+ }
+
+ Tys.clear();
+ Tys.push_back(MVT::i32);
+ Tys.push_back(MVT::i32);
+ Ops.clear();
+ Ops.push_back(Lo);
+ Ops.push_back(Hi);
+ return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
+}
+
+SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
+ Op.getOperand(0).getValueType() >= MVT::i16 &&
+ "Unknown SINT_TO_FP to lower!");
+
+ SDOperand Result;
+ MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
+ unsigned Size = MVT::getSizeInBits(SrcVT)/8;
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
+ SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ SDOperand Chain = DAG.getNode(ISD::STORE, MVT::Other,
+ DAG.getEntryNode(), Op.getOperand(0),
+ StackSlot, DAG.getSrcValue(NULL));
+
+ // Build the FILD
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::f64);
+ Tys.push_back(MVT::Other);
+ if (X86ScalarSSE) Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(StackSlot);
+ Ops.push_back(DAG.getValueType(SrcVT));
+ Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
+ Tys, Ops);
+
+ if (X86ScalarSSE) {
+ Chain = Result.getValue(1);
+ SDOperand InFlag = Result.getValue(2);
+
+ // FIXME: Currently the FST is flagged to the FILD_FLAG. This
+ // shouldn't be necessary except that RFP cannot be live across
+ // multiple blocks. When stackifier is fixed, they can be uncoupled.
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::Other);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(Result);
+ Ops.push_back(StackSlot);
+ Ops.push_back(DAG.getValueType(Op.getValueType()));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
+ Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
+ DAG.getSrcValue(NULL));
+ }
+
+ return Result;
+}
+
+SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
+ "Unknown FP_TO_SINT to lower!");
+ // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
+ // stack slot.
+ MachineFunction &MF = DAG.getMachineFunction();
+ unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
+ int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+
+ unsigned Opc;
+ switch (Op.getValueType()) {
+ 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;
+ case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
+ }
+
+ SDOperand Chain = DAG.getEntryNode();
+ SDOperand Value = Op.getOperand(0);
+ if (X86ScalarSSE) {
+ assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value, StackSlot,
+ DAG.getSrcValue(0));
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::f64);
+ Tys.push_back(MVT::Other);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(StackSlot);
+ Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
+ Value = DAG.getNode(X86ISD::FLD, Tys, Ops);
+ Chain = Value.getValue(1);
+ SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ }
+
+ // Build the FP_TO_INT*_IN_MEM
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(Value);
+ Ops.push_back(StackSlot);
+ SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
+
+ // Load the result.
+ return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
+ DAG.getSrcValue(NULL));
+}
+
+SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
+ MVT::ValueType VT = Op.getValueType();
+ const Type *OpNTy = MVT::getTypeForValueType(VT);
+ 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));
+ } 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);
+ SDOperand Mask
+ = DAG.getNode(X86ISD::LOAD_PACK,
+ VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
+ 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);
+ 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));
+ } 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);
+ SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK,
+ VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
+ return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
+}
+
+SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
+ SDOperand Cond;
+ SDOperand CC = Op.getOperand(2);
+ ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
+ bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
+ bool Flip;
+ unsigned X86CC;
+ if (translateX86CC(CC, isFP, X86CC, Flip)) {
+ if (Flip)
+ Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
+ Op.getOperand(1), Op.getOperand(0));
+ else
+ Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
+ Op.getOperand(0), Op.getOperand(1));
+ return DAG.getNode(X86ISD::SETCC, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond);
+ } else {
+ assert(isFP && "Illegal integer SetCC!");
+
+ Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
+ Op.getOperand(0), Op.getOperand(1));
+ std::vector<MVT::ValueType> Tys;
+ std::vector<SDOperand> Ops;
+ switch (SetCCOpcode) {
+ default: assert(false && "Illegal floating point SetCC!");
+ case ISD::SETOEQ: { // !PF & ZF
+ Tys.push_back(MVT::i8);
+ Tys.push_back(MVT::Flag);
+ Ops.push_back(DAG.getConstant(X86ISD::COND_NP, MVT::i8));
+ Ops.push_back(Cond);
+ SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
+ SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
+ DAG.getConstant(X86ISD::COND_E, MVT::i8),
+ Tmp1.getValue(1));
+ return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
}
-
- if (BytesLeft == 1) {
- Value = DAG.getConstant(Val, MVT::i8);
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
- DAG.getNode(ISD::ADD, AddrVT, DstAddr,
- DAG.getConstant(Offset, AddrVT)),
- DAG.getSrcValue(NULL));
+ case ISD::SETUNE: { // PF | !ZF
+ Tys.push_back(MVT::i8);
+ Tys.push_back(MVT::Flag);
+ Ops.push_back(DAG.getConstant(X86ISD::COND_P, MVT::i8));
+ Ops.push_back(Cond);
+ SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
+ SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
+ DAG.getConstant(X86ISD::COND_NE, MVT::i8),
+ Tmp1.getValue(1));
+ return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
}
}
+ }
+}
- return Chain;
- }
- case ISD::MEMCPY: {
- SDOperand Chain = Op.getOperand(0);
- unsigned Align =
- (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
- 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 ((Align & 3) != 0 ||
- (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
- MVT::ValueType IntPtr = getPointerTy();
- const Type *IntPtrTy = getTargetData().getIntPtrType();
- std::vector<std::pair<SDOperand, const Type*> > Args;
- Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
- Args.push_back(std::make_pair(Op.getOperand(2), IntPtrTy));
- Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
- std::pair<SDOperand,SDOperand> CallResult =
- LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
- DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
- return CallResult.second;
- }
-
- MVT::ValueType AVT;
- SDOperand Count;
- unsigned BytesLeft = 0;
- bool TwoRepMovs = false;
- switch (Align & 3) {
- case 2: // WORD aligned
- AVT = MVT::i16;
- Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
- BytesLeft = I->getValue() % 2;
- break;
- case 0: // DWORD aligned
- AVT = MVT::i32;
- if (I) {
- Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
- BytesLeft = I->getValue() % 4;
- } else {
- Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
- DAG.getConstant(2, MVT::i8));
- TwoRepMovs = true;
+SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
+ MVT::ValueType VT = Op.getValueType();
+ bool isFPStack = MVT::isFloatingPoint(VT) && !X86ScalarSSE;
+ bool addTest = false;
+ SDOperand Op0 = Op.getOperand(0);
+ SDOperand Cond, CC;
+ if (Op0.getOpcode() == ISD::SETCC)
+ Op0 = LowerOperation(Op0, DAG);
+
+ if (Op0.getOpcode() == X86ISD::SETCC) {
+ // If condition flag is set by a X86ISD::CMP, then make a copy of it
+ // (since flag operand cannot be shared). If the X86ISD::SETCC does not
+ // have another use it will be eliminated.
+ // If the X86ISD::SETCC has more than one use, then it's probably better
+ // to use a test instead of duplicating the X86ISD::CMP (for register
+ // pressure reason).
+ unsigned CmpOpc = Op0.getOperand(1).getOpcode();
+ if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
+ CmpOpc == X86ISD::UCOMI) {
+ if (!Op0.hasOneUse()) {
+ std::vector<MVT::ValueType> Tys;
+ for (unsigned i = 0; i < Op0.Val->getNumValues(); ++i)
+ Tys.push_back(Op0.Val->getValueType(i));
+ std::vector<SDOperand> Ops;
+ for (unsigned i = 0; i < Op0.getNumOperands(); ++i)
+ Ops.push_back(Op0.getOperand(i));
+ Op0 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
}
- break;
- default: // Byte aligned
- AVT = MVT::i8;
- Count = Op.getOperand(3);
- break;
- }
- SDOperand InFlag(0, 0);
- Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
- InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
- InFlag = Chain.getValue(1);
- Chain = DAG.getCopyToReg(Chain, X86::ESI, Op.getOperand(2), InFlag);
- InFlag = Chain.getValue(1);
+ CC = Op0.getOperand(0);
+ Cond = Op0.getOperand(1);
+ // Make a copy as flag result cannot be used by more than one.
+ Cond = DAG.getNode(CmpOpc, MVT::Flag,
+ Cond.getOperand(0), Cond.getOperand(1));
+ addTest =
+ isFPStack && !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
+ } else
+ addTest = true;
+ } else
+ addTest = true;
- std::vector<MVT::ValueType> Tys;
- Tys.push_back(MVT::Other);
- Tys.push_back(MVT::Flag);
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(AVT));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
+ if (addTest) {
+ CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
+ Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
+ }
- if (TwoRepMovs) {
- InFlag = Chain.getValue(1);
- Count = Op.getOperand(3);
- MVT::ValueType CVT = Count.getValueType();
- SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
- DAG.getConstant(3, CVT));
- Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
- InFlag = Chain.getValue(1);
- Tys.clear();
- Tys.push_back(MVT::Other);
- Tys.push_back(MVT::Flag);
- Ops.clear();
- Ops.push_back(Chain);
- Ops.push_back(DAG.getValueType(MVT::i8));
- Ops.push_back(InFlag);
- Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
- } else if (BytesLeft) {
- // Issue loads and stores for the last 1 - 3 bytes.
- unsigned Offset = I->getValue() - BytesLeft;
- SDOperand DstAddr = Op.getOperand(1);
- MVT::ValueType DstVT = DstAddr.getValueType();
- SDOperand SrcAddr = Op.getOperand(2);
- MVT::ValueType SrcVT = SrcAddr.getValueType();
- SDOperand Value;
- if (BytesLeft >= 2) {
- Value = DAG.getLoad(MVT::i16, Chain,
- DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
- DAG.getConstant(Offset, SrcVT)),
- DAG.getSrcValue(NULL));
- Chain = Value.getValue(1);
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
- DAG.getNode(ISD::ADD, DstVT, DstAddr,
- DAG.getConstant(Offset, DstVT)),
- DAG.getSrcValue(NULL));
- BytesLeft -= 2;
- Offset += 2;
- }
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(Op.getValueType());
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> 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, Tys, Ops);
+}
- if (BytesLeft == 1) {
- Value = DAG.getLoad(MVT::i8, Chain,
- DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
- DAG.getConstant(Offset, SrcVT)),
- DAG.getSrcValue(NULL));
- Chain = Value.getValue(1);
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
- DAG.getNode(ISD::ADD, DstVT, DstAddr,
- DAG.getConstant(Offset, DstVT)),
- DAG.getSrcValue(NULL));
+SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
+ bool addTest = false;
+ SDOperand Cond = Op.getOperand(1);
+ SDOperand Dest = Op.getOperand(2);
+ SDOperand CC;
+ if (Cond.getOpcode() == ISD::SETCC)
+ Cond = LowerOperation(Cond, DAG);
+
+ if (Cond.getOpcode() == X86ISD::SETCC) {
+ // If condition flag is set by a X86ISD::CMP, then make a copy of it
+ // (since flag operand cannot be shared). If the X86ISD::SETCC does not
+ // have another use it will be eliminated.
+ // If the X86ISD::SETCC has more than one use, then it's probably better
+ // to use a test instead of duplicating the X86ISD::CMP (for register
+ // pressure reason).
+ unsigned CmpOpc = Cond.getOperand(1).getOpcode();
+ if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
+ CmpOpc == X86ISD::UCOMI) {
+ if (!Cond.hasOneUse()) {
+ std::vector<MVT::ValueType> Tys;
+ for (unsigned i = 0; i < Cond.Val->getNumValues(); ++i)
+ Tys.push_back(Cond.Val->getValueType(i));
+ std::vector<SDOperand> Ops;
+ for (unsigned i = 0; i < Cond.getNumOperands(); ++i)
+ Ops.push_back(Cond.getOperand(i));
+ Cond = DAG.getNode(X86ISD::SETCC, Tys, Ops);
}
- }
- return Chain;
- }
-
- // ConstantPool, GlobalAddress, and ExternalSymbol are lowered as their
- // target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
- // one of the above mentioned nodes. It has to be wrapped because otherwise
- // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
- // be used to form addressing mode. These wrapped nodes will be selected
- // into MOV32ri.
- case ISD::ConstantPool: {
- ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
- SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
- DAG.getTargetConstantPool(CP->get(), getPointerTy(),
- CP->getAlignment()));
- if (Subtarget->isTargetDarwin()) {
- // With PIC, the address is actually $g + Offset.
- if (getTargetMachine().getRelocationModel() == Reloc::PIC)
- Result = DAG.getNode(ISD::ADD, getPointerTy(),
- DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
- }
+ CC = Cond.getOperand(0);
+ Cond = Cond.getOperand(1);
+ // Make a copy as flag result cannot be used by more than one.
+ Cond = DAG.getNode(CmpOpc, MVT::Flag,
+ Cond.getOperand(0), Cond.getOperand(1));
+ } else
+ addTest = true;
+ } else
+ addTest = true;
- return Result;
+ if (addTest) {
+ CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
+ Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
}
- case ISD::GlobalAddress: {
- GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
- SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
- DAG.getTargetGlobalAddress(GV, getPointerTy()));
- if (Subtarget->isTargetDarwin()) {
- // With PIC, the address is actually $g + Offset.
- if (getTargetMachine().getRelocationModel() == Reloc::PIC)
- Result = DAG.getNode(ISD::ADD, getPointerTy(),
- DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
-
- // For Darwin, 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 the GV offset field.
- if (getTargetMachine().getRelocationModel() != Reloc::Static &&
- DarwinGVRequiresExtraLoad(GV))
- Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
- Result, DAG.getSrcValue(NULL));
- }
+ return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
+ Op.getOperand(0), Op.getOperand(2), CC, Cond);
+}
- return Result;
+SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
+ JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
+ SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
+ DAG.getTargetJumpTable(JT->getIndex(),
+ getPointerTy()));
+ if (Subtarget->isTargetDarwin()) {
+ // With PIC, the address is actually $g + Offset.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC)
+ Result = DAG.getNode(ISD::ADD, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
+ Result);
}
- case ISD::ExternalSymbol: {
- const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
- SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
- DAG.getTargetExternalSymbol(Sym, getPointerTy()));
- if (Subtarget->isTargetDarwin()) {
- // With PIC, the address is actually $g + Offset.
- if (getTargetMachine().getRelocationModel() == Reloc::PIC)
- Result = DAG.getNode(ISD::ADD, getPointerTy(),
- DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
- }
- return Result;
- }
- case ISD::VASTART: {
- // vastart just stores the address of the VarArgsFrameIndex slot into the
- // memory location argument.
- // FIXME: Replace MVT::i32 with PointerTy
- SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
- return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
- Op.getOperand(1), Op.getOperand(2));
- }
- case ISD::RET: {
- SDOperand Copy;
+ return Result;
+}
+
+SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand Copy;
- switch(Op.getNumOperands()) {
+ switch(Op.getNumOperands()) {
default:
assert(0 && "Do not know how to return this many arguments!");
abort();
- case 1:
+ case 1: // ret void.
return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
- DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
+ DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
case 2: {
MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
- if (MVT::isInteger(ArgVT))
+
+ if (MVT::isVector(ArgVT)) {
+ // Integer or FP vector result -> XMM0.
+ if (DAG.getMachineFunction().liveout_empty())
+ DAG.getMachineFunction().addLiveOut(X86::XMM0);
+ Copy = DAG.getCopyToReg(Op.getOperand(0), X86::XMM0, Op.getOperand(1),
+ SDOperand());
+ } else if (MVT::isInteger(ArgVT)) {
+ // Integer result -> EAX
+ if (DAG.getMachineFunction().liveout_empty())
+ DAG.getMachineFunction().addLiveOut(X86::EAX);
+
Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EAX, Op.getOperand(1),
SDOperand());
- else if (!X86ScalarSSE) {
+ } else if (!X86ScalarSSE) {
+ // FP return with fp-stack value.
+ if (DAG.getMachineFunction().liveout_empty())
+ DAG.getMachineFunction().addLiveOut(X86::ST0);
+
std::vector<MVT::ValueType> Tys;
Tys.push_back(MVT::Other);
Tys.push_back(MVT::Flag);
Ops.push_back(Op.getOperand(1));
Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
} else {
+ // FP return with ScalarSSE (return on fp-stack).
+ if (DAG.getMachineFunction().liveout_empty())
+ DAG.getMachineFunction().addLiveOut(X86::ST0);
+
SDOperand MemLoc;
SDOperand Chain = Op.getOperand(0);
SDOperand Value = Op.getOperand(1);
break;
}
case 3:
+ if (DAG.getMachineFunction().liveout_empty()) {
+ DAG.getMachineFunction().addLiveOut(X86::EAX);
+ DAG.getMachineFunction().addLiveOut(X86::EDX);
+ }
+
Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EDX, Op.getOperand(2),
SDOperand());
Copy = DAG.getCopyToReg(Copy, X86::EAX,Op.getOperand(1),Copy.getValue(1));
break;
- }
- return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
- Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
- Copy.getValue(1));
- }
- case ISD::SCALAR_TO_VECTOR: {
- SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
- return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
- }
- case ISD::VECTOR_SHUFFLE: {
- SDOperand V1 = Op.getOperand(0);
- SDOperand V2 = Op.getOperand(1);
- SDOperand PermMask = Op.getOperand(2);
- MVT::ValueType VT = Op.getValueType();
- unsigned NumElems = PermMask.getNumOperands();
-
- // Splat && PSHUFD's 2nd vector must be undef.
- if (X86::isSplatMask(PermMask.Val)) {
- if (V2.getOpcode() != ISD::UNDEF)
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
- return Op;
- }
+ }
+ return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
+ Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
+ Copy.getValue(1));
+}
- if (X86::isUNPCKLMask(PermMask.Val) ||
- X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
- X86::isUNPCKHMask(PermMask.Val))
- // Leave the VECTOR_SHUFFLE alone. It matches {P}UNPCKL*.
- return Op;
+SDOperand
+X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
+ if (FormalArgs.size() == 0) {
+ unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
+ if (CC == CallingConv::Fast && EnableFastCC)
+ LowerFastCCArguments(Op, DAG);
+ else
+ LowerCCCArguments(Op, DAG);
+ }
+
+ // Return the new list of results.
+ std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
+ Op.Val->value_end());
+ return DAG.getNode(ISD::MERGE_VALUES, RetVTs, FormalArgs);
+}
- if (NumElems == 2)
- return NormalizeVectorShuffle(Op, DAG);
-
- // If VT is integer, try PSHUF* first, then SHUFP*.
- if (MVT::isInteger(VT)) {
- if (X86::isPSHUFDMask(PermMask.Val) ||
- X86::isPSHUFHWMask(PermMask.Val) ||
- X86::isPSHUFLWMask(PermMask.Val)) {
- if (V2.getOpcode() != ISD::UNDEF)
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
- return Op;
- }
+SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand InFlag(0, 0);
+ SDOperand Chain = Op.getOperand(0);
+ unsigned Align =
+ (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
+ 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 ((Align & 3) != 0 ||
+ (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
+ MVT::ValueType IntPtr = getPointerTy();
+ const Type *IntPtrTy = getTargetData()->getIntPtrType();
+ std::vector<std::pair<SDOperand, const Type*> > Args;
+ Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
+ // Extend the ubyte argument to be an int value for the call.
+ SDOperand Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
+ Args.push_back(std::make_pair(Val, IntPtrTy));
+ Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
+ std::pair<SDOperand,SDOperand> CallResult =
+ LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
+ DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
+ return CallResult.second;
+ }
- if (X86::isSHUFPMask(PermMask.Val))
- return NormalizeVectorShuffle(Op, DAG);
-
- // Handle v8i16 shuffle high / low shuffle node pair.
- if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
- MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
- std::vector<SDOperand> MaskVec;
- for (unsigned i = 0; i != 4; ++i)
- MaskVec.push_back(PermMask.getOperand(i));
- for (unsigned i = 4; i != 8; ++i)
- MaskVec.push_back(DAG.getConstant(i, BaseVT));
- SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
- V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
- MaskVec.clear();
- for (unsigned i = 0; i != 4; ++i)
- MaskVec.push_back(DAG.getConstant(i, BaseVT));
- for (unsigned i = 4; i != 8; ++i)
- MaskVec.push_back(PermMask.getOperand(i));
- Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
- }
- } else {
- // Floating point cases in the other order.
- if (X86::isSHUFPMask(PermMask.Val))
- return NormalizeVectorShuffle(Op, DAG);
- if (X86::isPSHUFDMask(PermMask.Val) ||
- X86::isPSHUFHWMask(PermMask.Val) ||
- X86::isPSHUFLWMask(PermMask.Val)) {
- if (V2.getOpcode() != ISD::UNDEF)
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
- return Op;
- }
+ MVT::ValueType AVT;
+ SDOperand Count;
+ ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
+ unsigned BytesLeft = 0;
+ bool TwoRepStos = false;
+ if (ValC) {
+ unsigned ValReg;
+ unsigned Val = ValC->getValue() & 255;
+
+ // If the value is a constant, then we can potentially use larger sets.
+ switch (Align & 3) {
+ case 2: // WORD aligned
+ AVT = MVT::i16;
+ Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
+ BytesLeft = I->getValue() % 2;
+ Val = (Val << 8) | Val;
+ ValReg = X86::AX;
+ break;
+ case 0: // DWORD aligned
+ AVT = MVT::i32;
+ if (I) {
+ Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
+ BytesLeft = I->getValue() % 4;
+ } else {
+ Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
+ DAG.getConstant(2, MVT::i8));
+ TwoRepStos = true;
+ }
+ Val = (Val << 8) | Val;
+ Val = (Val << 16) | Val;
+ ValReg = X86::EAX;
+ break;
+ default: // Byte aligned
+ AVT = MVT::i8;
+ Count = Op.getOperand(3);
+ ValReg = X86::AL;
+ break;
}
- return SDOperand();
+ Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
+ InFlag);
+ InFlag = Chain.getValue(1);
+ } else {
+ AVT = MVT::i8;
+ Count = Op.getOperand(3);
+ Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
+ InFlag = Chain.getValue(1);
}
- case ISD::BUILD_VECTOR: {
- // All one's are handled with pcmpeqd.
- if (ISD::isBuildVectorAllOnes(Op.Val))
- return Op;
- std::set<SDOperand> Values;
- SDOperand Elt0 = Op.getOperand(0);
- Values.insert(Elt0);
- bool Elt0IsZero = (isa<ConstantSDNode>(Elt0) &&
- cast<ConstantSDNode>(Elt0)->getValue() == 0) ||
- (isa<ConstantFPSDNode>(Elt0) &&
- cast<ConstantFPSDNode>(Elt0)->isExactlyValue(0.0));
- bool RestAreZero = true;
- unsigned NumElems = Op.getNumOperands();
- for (unsigned i = 1; i < NumElems; ++i) {
- SDOperand Elt = Op.getOperand(i);
- if (ConstantFPSDNode *FPC = dyn_cast<ConstantFPSDNode>(Elt)) {
- if (!FPC->isExactlyValue(+0.0))
- RestAreZero = false;
- } else if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
- if (!C->isNullValue())
- RestAreZero = false;
- } else
- RestAreZero = false;
- Values.insert(Elt);
- }
+ Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
+ InFlag = Chain.getValue(1);
+ Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
+ InFlag = Chain.getValue(1);
- if (RestAreZero) {
- if (Elt0IsZero) return Op;
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getValueType(AVT));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
- // Zero extend a scalar to a vector.
- return DAG.getNode(X86ISD::ZEXT_S2VEC, Op.getValueType(), Elt0);
+ if (TwoRepStos) {
+ InFlag = Chain.getValue(1);
+ Count = Op.getOperand(3);
+ MVT::ValueType CVT = Count.getValueType();
+ SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
+ DAG.getConstant(3, CVT));
+ Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
+ InFlag = Chain.getValue(1);
+ Tys.clear();
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ Ops.clear();
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getValueType(MVT::i8));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
+ } else if (BytesLeft) {
+ // Issue stores for the last 1 - 3 bytes.
+ SDOperand Value;
+ unsigned Val = ValC->getValue() & 255;
+ unsigned Offset = I->getValue() - BytesLeft;
+ SDOperand DstAddr = Op.getOperand(1);
+ MVT::ValueType AddrVT = DstAddr.getValueType();
+ if (BytesLeft >= 2) {
+ Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
+ DAG.getNode(ISD::ADD, AddrVT, DstAddr,
+ DAG.getConstant(Offset, AddrVT)),
+ DAG.getSrcValue(NULL));
+ BytesLeft -= 2;
+ Offset += 2;
}
- if (Values.size() > 2) {
- // 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>
- MVT::ValueType VT = Op.getValueType();
- MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
- MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
- std::vector<SDOperand> MaskVec;
- 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));
- }
- SDOperand PermMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
- std::vector<SDOperand> V(NumElems);
- for (unsigned i = 0; i < NumElems; ++i)
- V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
- NumElems >>= 1;
- while (NumElems != 0) {
- for (unsigned i = 0; i < NumElems; ++i)
- V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
- PermMask);
- NumElems >>= 1;
- }
- return V[0];
+ if (BytesLeft == 1) {
+ Value = DAG.getConstant(Val, MVT::i8);
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
+ DAG.getNode(ISD::ADD, AddrVT, DstAddr,
+ DAG.getConstant(Offset, AddrVT)),
+ DAG.getSrcValue(NULL));
}
+ }
- return SDOperand();
+ return Chain;
+}
+
+SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
+ SDOperand Chain = Op.getOperand(0);
+ unsigned Align =
+ (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
+ 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 ((Align & 3) != 0 ||
+ (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
+ MVT::ValueType IntPtr = getPointerTy();
+ const Type *IntPtrTy = getTargetData()->getIntPtrType();
+ std::vector<std::pair<SDOperand, const Type*> > Args;
+ Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
+ Args.push_back(std::make_pair(Op.getOperand(2), IntPtrTy));
+ Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
+ std::pair<SDOperand,SDOperand> CallResult =
+ LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
+ DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
+ return CallResult.second;
}
- case ISD::EXTRACT_VECTOR_ELT: {
- if (!isa<ConstantSDNode>(Op.getOperand(1)))
- return SDOperand();
-
- MVT::ValueType VT = Op.getValueType();
- if (MVT::getSizeInBits(VT) == 16) {
- // Transform it so it match pextrw which produces a 32-bit result.
- MVT::ValueType EVT = (MVT::ValueType)(VT+1);
- SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
- Op.getOperand(0), Op.getOperand(1));
- SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
- DAG.getValueType(VT));
- return DAG.getNode(ISD::TRUNCATE, VT, Assert);
- } else if (MVT::getSizeInBits(VT) == 32) {
- SDOperand Vec = Op.getOperand(0);
- unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
- if (Idx == 0)
- return Op;
-
- // TODO: if Idex == 2, we can use unpckhps
- // SHUFPS the element to the lowest double word, then movss.
- MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
- SDOperand IdxNode = DAG.getConstant((Idx < 2) ? Idx : Idx+4,
- MVT::getVectorBaseType(MaskVT));
- std::vector<SDOperand> IdxVec;
- IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
- IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
- IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
- IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
- SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
- Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
- Vec, Vec, Mask);
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
- DAG.getConstant(0, MVT::i32));
- } else if (MVT::getSizeInBits(VT) == 64) {
- SDOperand Vec = Op.getOperand(0);
- unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
- if (Idx == 0)
- return Op;
-
- // 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::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
- std::vector<SDOperand> IdxVec;
- IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
- IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
- SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
- Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
- Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
- DAG.getConstant(0, MVT::i32));
- }
- return SDOperand();
+ MVT::ValueType AVT;
+ SDOperand Count;
+ unsigned BytesLeft = 0;
+ bool TwoRepMovs = false;
+ switch (Align & 3) {
+ case 2: // WORD aligned
+ AVT = MVT::i16;
+ Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
+ BytesLeft = I->getValue() % 2;
+ break;
+ case 0: // DWORD aligned
+ AVT = MVT::i32;
+ if (I) {
+ Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
+ BytesLeft = I->getValue() % 4;
+ } else {
+ Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
+ DAG.getConstant(2, MVT::i8));
+ TwoRepMovs = true;
+ }
+ break;
+ default: // Byte aligned
+ AVT = MVT::i8;
+ Count = Op.getOperand(3);
+ break;
}
- case ISD::INSERT_VECTOR_ELT: {
- // Transform it so it match pinsrw which expects a 16-bit value in a R32
- // as its second argument.
- MVT::ValueType VT = Op.getValueType();
- MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
- if (MVT::getSizeInBits(BaseVT) == 16) {
- SDOperand N1 = Op.getOperand(1);
- SDOperand N2 = Op.getOperand(2);
- 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);
- return DAG.getNode(X86ISD::PINSRW, VT, Op.getOperand(0), N1, N2);
+
+ SDOperand InFlag(0, 0);
+ Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
+ InFlag = Chain.getValue(1);
+ Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
+ InFlag = Chain.getValue(1);
+ Chain = DAG.getCopyToReg(Chain, X86::ESI, Op.getOperand(2), InFlag);
+ InFlag = Chain.getValue(1);
+
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getValueType(AVT));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
+
+ if (TwoRepMovs) {
+ InFlag = Chain.getValue(1);
+ Count = Op.getOperand(3);
+ MVT::ValueType CVT = Count.getValueType();
+ SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
+ DAG.getConstant(3, CVT));
+ Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
+ InFlag = Chain.getValue(1);
+ Tys.clear();
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ Ops.clear();
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getValueType(MVT::i8));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
+ } else if (BytesLeft) {
+ // Issue loads and stores for the last 1 - 3 bytes.
+ unsigned Offset = I->getValue() - BytesLeft;
+ SDOperand DstAddr = Op.getOperand(1);
+ MVT::ValueType DstVT = DstAddr.getValueType();
+ SDOperand SrcAddr = Op.getOperand(2);
+ MVT::ValueType SrcVT = SrcAddr.getValueType();
+ SDOperand Value;
+ if (BytesLeft >= 2) {
+ Value = DAG.getLoad(MVT::i16, Chain,
+ DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
+ DAG.getConstant(Offset, SrcVT)),
+ DAG.getSrcValue(NULL));
+ Chain = Value.getValue(1);
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
+ DAG.getNode(ISD::ADD, DstVT, DstAddr,
+ DAG.getConstant(Offset, DstVT)),
+ DAG.getSrcValue(NULL));
+ BytesLeft -= 2;
+ Offset += 2;
}
- return SDOperand();
+ if (BytesLeft == 1) {
+ Value = DAG.getLoad(MVT::i8, Chain,
+ DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
+ DAG.getConstant(Offset, SrcVT)),
+ DAG.getSrcValue(NULL));
+ Chain = Value.getValue(1);
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
+ DAG.getNode(ISD::ADD, DstVT, DstAddr,
+ DAG.getConstant(Offset, DstVT)),
+ DAG.getSrcValue(NULL));
+ }
+ }
+
+ return Chain;
+}
+
+SDOperand
+X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Op.getOperand(0));
+ SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, Ops);
+ Ops.clear();
+ Ops.push_back(DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)));
+ Ops.push_back(DAG.getCopyFromReg(Ops[0].getValue(1), X86::EDX,
+ MVT::i32, Ops[0].getValue(2)));
+ Ops.push_back(Ops[1].getValue(1));
+ Tys[0] = Tys[1] = MVT::i32;
+ Tys.push_back(MVT::Other);
+ return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
+}
+
+SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
+ // vastart just stores the address of the VarArgsFrameIndex slot into the
+ // memory location argument.
+ // FIXME: Replace MVT::i32 with PointerTy
+ SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
+ return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
+ Op.getOperand(1), Op.getOperand(2));
+}
+
+SDOperand
+X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
+ unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
+ switch (IntNo) {
+ default: return SDOperand(); // Don't custom lower most intrinsics.
+ // Comparison intrinsics.
+ case Intrinsic::x86_sse_comieq_ss:
+ case Intrinsic::x86_sse_comilt_ss:
+ case Intrinsic::x86_sse_comile_ss:
+ case Intrinsic::x86_sse_comigt_ss:
+ case Intrinsic::x86_sse_comige_ss:
+ case Intrinsic::x86_sse_comineq_ss:
+ case Intrinsic::x86_sse_ucomieq_ss:
+ case Intrinsic::x86_sse_ucomilt_ss:
+ case Intrinsic::x86_sse_ucomile_ss:
+ case Intrinsic::x86_sse_ucomigt_ss:
+ case Intrinsic::x86_sse_ucomige_ss:
+ case Intrinsic::x86_sse_ucomineq_ss:
+ case Intrinsic::x86_sse2_comieq_sd:
+ case Intrinsic::x86_sse2_comilt_sd:
+ case Intrinsic::x86_sse2_comile_sd:
+ case Intrinsic::x86_sse2_comigt_sd:
+ case Intrinsic::x86_sse2_comige_sd:
+ case Intrinsic::x86_sse2_comineq_sd:
+ case Intrinsic::x86_sse2_ucomieq_sd:
+ case Intrinsic::x86_sse2_ucomilt_sd:
+ case Intrinsic::x86_sse2_ucomile_sd:
+ case Intrinsic::x86_sse2_ucomigt_sd:
+ case Intrinsic::x86_sse2_ucomige_sd:
+ case Intrinsic::x86_sse2_ucomineq_sd: {
+ unsigned Opc = 0;
+ ISD::CondCode CC = ISD::SETCC_INVALID;
+ switch (IntNo) {
+ default: break;
+ case Intrinsic::x86_sse_comieq_ss:
+ case Intrinsic::x86_sse2_comieq_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETEQ;
+ break;
+ case Intrinsic::x86_sse_comilt_ss:
+ case Intrinsic::x86_sse2_comilt_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETLT;
+ break;
+ case Intrinsic::x86_sse_comile_ss:
+ case Intrinsic::x86_sse2_comile_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETLE;
+ break;
+ case Intrinsic::x86_sse_comigt_ss:
+ case Intrinsic::x86_sse2_comigt_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETGT;
+ break;
+ case Intrinsic::x86_sse_comige_ss:
+ case Intrinsic::x86_sse2_comige_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETGE;
+ break;
+ case Intrinsic::x86_sse_comineq_ss:
+ case Intrinsic::x86_sse2_comineq_sd:
+ Opc = X86ISD::COMI;
+ CC = ISD::SETNE;
+ break;
+ case Intrinsic::x86_sse_ucomieq_ss:
+ case Intrinsic::x86_sse2_ucomieq_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETEQ;
+ break;
+ case Intrinsic::x86_sse_ucomilt_ss:
+ case Intrinsic::x86_sse2_ucomilt_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETLT;
+ break;
+ case Intrinsic::x86_sse_ucomile_ss:
+ case Intrinsic::x86_sse2_ucomile_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETLE;
+ break;
+ case Intrinsic::x86_sse_ucomigt_ss:
+ case Intrinsic::x86_sse2_ucomigt_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETGT;
+ break;
+ case Intrinsic::x86_sse_ucomige_ss:
+ case Intrinsic::x86_sse2_ucomige_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETGE;
+ break;
+ case Intrinsic::x86_sse_ucomineq_ss:
+ case Intrinsic::x86_sse2_ucomineq_sd:
+ Opc = X86ISD::UCOMI;
+ CC = ISD::SETNE;
+ break;
+ }
+ bool Flip;
+ unsigned X86CC;
+ translateX86CC(CC, true, X86CC, Flip);
+ SDOperand Cond = DAG.getNode(Opc, MVT::Flag, Op.getOperand(Flip?2:1),
+ Op.getOperand(Flip?1:2));
+ SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond);
+ return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
+ }
}
+}
+
+/// LowerOperation - Provide custom lowering hooks for some operations.
+///
+SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
+ switch (Op.getOpcode()) {
+ default: assert(0 && "Should not custom lower this!");
+ case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
+ case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
+ case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
+ case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
+ case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
+ case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
+ case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
+ case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
+ case ISD::SHL_PARTS:
+ case ISD::SRA_PARTS:
+ case ISD::SRL_PARTS: return LowerShift(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::SETCC: return LowerSETCC(Op, DAG);
+ case ISD::SELECT: return LowerSELECT(Op, DAG);
+ case ISD::BRCOND: return LowerBRCOND(Op, DAG);
+ case ISD::JumpTable: return LowerJumpTable(Op, DAG);
+ case ISD::RET: return LowerRET(Op, DAG);
+ case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
+ case ISD::MEMSET: return LowerMEMSET(Op, DAG);
+ case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
+ case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
+ case ISD::VASTART: return LowerVASTART(Op, DAG);
+ case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
}
}
case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
case X86ISD::CMP: return "X86ISD::CMP";
case X86ISD::TEST: return "X86ISD::TEST";
+ case X86ISD::COMI: return "X86ISD::COMI";
+ case X86ISD::UCOMI: return "X86ISD::UCOMI";
case X86ISD::SETCC: return "X86ISD::SETCC";
case X86ISD::CMOV: return "X86ISD::CMOV";
case X86ISD::BRCOND: return "X86ISD::BRCOND";
case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
case X86ISD::Wrapper: return "X86ISD::Wrapper";
case X86ISD::S2VEC: return "X86ISD::S2VEC";
- case X86ISD::ZEXT_S2VEC: return "X86ISD::ZEXT_S2VEC";
case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
}
default: break; // Unknown constriant letter
case 'r': // GENERAL_REGS
case 'R': // LEGACY_REGS
- return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX,
- X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
+ if (VT == MVT::i32)
+ return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
+ X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
+ else if (VT == MVT::i16)
+ return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
+ X86::SI, X86::DI, X86::BP, X86::SP, 0);
+ else if (VT == MVT::i8)
+ return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
+ break;
case 'l': // INDEX_REGS
- return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX,
- X86::ESI, X86::EDI, X86::EBP, 0);
+ if (VT == MVT::i32)
+ return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
+ X86::ESI, X86::EDI, X86::EBP, 0);
+ else if (VT == MVT::i16)
+ return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
+ X86::SI, X86::DI, X86::BP, 0);
+ else if (VT == MVT::i8)
+ return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
+ break;
case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
case 'Q': // Q_REGS
- return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX, 0);
+ if (VT == MVT::i32)
+ return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
+ 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);
+ break;
case 'x': // SSE_REGS if SSE1 allowed
if (Subtarget->hasSSE1())
return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
// Only do shuffles on 128-bit vector types for now.
if (MVT::getSizeInBits(VT) == 64) return false;
- return (Mask.Val->getNumOperands() == 2 ||
- X86::isSplatMask(Mask.Val) ||
- X86::isPSHUFDMask(Mask.Val) ||
+ return (Mask.Val->getNumOperands() <= 4 ||
+ isSplatMask(Mask.Val) ||
isPSHUFHW_PSHUFLWMask(Mask.Val) ||
- X86::isSHUFPMask(Mask.Val) ||
X86::isUNPCKLMask(Mask.Val) ||
X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
X86::isUNPCKHMask(Mask.Val));
}
+
+bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
+ MVT::ValueType EVT,
+ SelectionDAG &DAG) const {
+ unsigned NumElts = BVOps.size();
+ // Only do shuffles on 128-bit vector types for now.
+ if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
+ if (NumElts == 2) return true;
+ if (NumElts == 4) {
+ return (isMOVLMask(BVOps) || isCommutedMOVL(BVOps, true) ||
+ isSHUFPMask(BVOps) || isCommutedSHUFP(BVOps));
+ }
+ return false;
+}