//===----------------------------------------------------------------------===//
#include "X86.h"
+#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
+#include "llvm/Constants.h"
#include "llvm/Function.h"
-#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/ADT/VectorExtras.h"
using namespace llvm;
// FIXME: temporary.
X86TargetLowering::X86TargetLowering(TargetMachine &TM)
: TargetLowering(TM) {
+ Subtarget = &TM.getSubtarget<X86Subtarget>();
+ X86ScalarSSE = Subtarget->hasSSE2();
+
// Set up the TargetLowering object.
// X86 is weird, it always uses i8 for shift amounts and setcc results.
setShiftAmountType(MVT::i8);
setSetCCResultType(MVT::i8);
setSetCCResultContents(ZeroOrOneSetCCResult);
+ setSchedulingPreference(SchedulingForRegPressure);
setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
-
+ setStackPointerRegisterToSaveRestore(X86::ESP);
+
// Set up the register classes.
addRegisterClass(MVT::i8, X86::R8RegisterClass);
addRegisterClass(MVT::i16, X86::R16RegisterClass);
setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
- setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
+
+ if (X86ScalarSSE)
+ // No SSE i64 SINT_TO_FP, so expand i32 UINT_TO_FP instead.
+ setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
+ else
+ setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
// Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
// this operation.
setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
+ // SSE has no i16 to fp conversion, only i32
+ if (X86ScalarSSE)
+ setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
+ else {
+ setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
+ setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
+ }
- if (!X86ScalarSSE) {
- // We can handle SINT_TO_FP and FP_TO_SINT from/TO i64 even though i64
- // isn't legal.
- setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
- setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
- setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
+ // We can handle SINT_TO_FP and FP_TO_SINT from/to i64 even though i64
+ // isn't legal.
+ setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
+ setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
+
+ // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
+ // this operation.
+ setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
+ setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
+
+ if (X86ScalarSSE) {
+ setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
+ } else {
setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
+ setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
}
// Handle FP_TO_UINT by promoting the destination to a larger signed
setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
- if (!X86ScalarSSE)
+ if (X86ScalarSSE && !Subtarget->hasSSE3())
+ // Expand FP_TO_UINT into a select.
+ // FIXME: We would like to use a Custom expander here eventually to do
+ // the optimal thing for SSE vs. the default expansion in the legalizer.
+ setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
+ else
+ // With SSE3 we can use fisttpll to convert to a signed i64.
setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
- // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
- // this operation.
- setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
- setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
- setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
+ setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
+ setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
- if (X86DAGIsel) {
- setOperationAction(ISD::BRCOND , MVT::Other, Custom);
- }
+ setOperationAction(ISD::BRCOND , MVT::Other, Custom);
setOperationAction(ISD::BRCONDTWOWAY , MVT::Other, Expand);
setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
+ setOperationAction(ISD::BR_CC , MVT::Other, Expand);
+ setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
+ setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
setOperationAction(ISD::READIO , MVT::i1 , Expand);
setOperationAction(ISD::READIO , MVT::i8 , Expand);
// These should be promoted to a larger select which is supported.
setOperationAction(ISD::SELECT , MVT::i1 , Promote);
setOperationAction(ISD::SELECT , MVT::i8 , Promote);
+
// X86 wants to expand cmov itself.
- if (X86DAGIsel) {
- setOperationAction(ISD::SELECT , MVT::i16 , Custom);
- setOperationAction(ISD::SELECT , MVT::i32 , Custom);
- setOperationAction(ISD::SETCC , MVT::i8 , Custom);
- setOperationAction(ISD::SETCC , MVT::i16 , Custom);
- setOperationAction(ISD::SETCC , MVT::i32 , Custom);
- setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
- }
+ setOperationAction(ISD::SELECT , MVT::i16 , Custom);
+ setOperationAction(ISD::SELECT , MVT::i32 , Custom);
+ setOperationAction(ISD::SELECT , MVT::f32 , Custom);
+ setOperationAction(ISD::SELECT , MVT::f64 , Custom);
+ setOperationAction(ISD::SETCC , MVT::i8 , Custom);
+ setOperationAction(ISD::SETCC , MVT::i16 , Custom);
+ setOperationAction(ISD::SETCC , MVT::i32 , Custom);
+ setOperationAction(ISD::SETCC , MVT::f32 , Custom);
+ setOperationAction(ISD::SETCC , MVT::f64 , Custom);
+ // X86 ret instruction may pop stack.
+ setOperationAction(ISD::RET , MVT::Other, Custom);
+ // Darwin ABI issue.
+ setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
+ setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
+ // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
+ setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
+ setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
+ setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
+ // X86 wants to expand memset / memcpy itself.
+ setOperationAction(ISD::MEMSET , MVT::Other, Custom);
+ setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
// We don't have line number support yet.
setOperationAction(ISD::LOCATION, MVT::Other, Expand);
setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
+ setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
+
+ // VASTART needs to be custom lowered to use the VarArgsFrameIndex
+ setOperationAction(ISD::VASTART , MVT::Other, Custom);
+
+ // Use the default implementation.
+ setOperationAction(ISD::VAARG , MVT::Other, Expand);
+ setOperationAction(ISD::VACOPY , MVT::Other, Expand);
+ setOperationAction(ISD::VAEND , MVT::Other, Expand);
+ setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
+ setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
+ setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
if (X86ScalarSSE) {
// Set up the FP register classes.
- addRegisterClass(MVT::f32, X86::V4F4RegisterClass);
- addRegisterClass(MVT::f64, X86::V2F8RegisterClass);
+ 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);
- // SSE has no i16 to fp conversion, only i32
- setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
- setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
+ // Use ANDPD to simulate FABS.
+ setOperationAction(ISD::FABS , MVT::f64, Custom);
+ setOperationAction(ISD::FABS , MVT::f32, Custom);
- // Expand FP_TO_UINT into a select.
- // FIXME: We would like to use a Custom expander here eventually to do
- // the optimal thing for SSE vs. the default expansion in the legalizer.
- setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
-
- // We don't support sin/cos/sqrt/fmod
+ // Use XORP to simulate FNEG.
+ setOperationAction(ISD::FNEG , MVT::f64, Custom);
+ setOperationAction(ISD::FNEG , MVT::f32, Custom);
+
+ // We don't support sin/cos/fmod
setOperationAction(ISD::FSIN , MVT::f64, Expand);
setOperationAction(ISD::FCOS , MVT::f64, Expand);
- setOperationAction(ISD::FABS , MVT::f64, Expand);
- setOperationAction(ISD::FNEG , MVT::f64, Expand);
setOperationAction(ISD::FREM , MVT::f64, Expand);
setOperationAction(ISD::FSIN , MVT::f32, Expand);
setOperationAction(ISD::FCOS , MVT::f32, Expand);
- setOperationAction(ISD::FABS , MVT::f32, Expand);
- setOperationAction(ISD::FNEG , MVT::f32, Expand);
setOperationAction(ISD::FREM , MVT::f32, Expand);
+ // Expand FP immediates into loads from the stack, except for the special
+ // cases we handle.
+ setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
+ setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
addLegalFPImmediate(+0.0); // xorps / xorpd
} else {
// Set up the FP register classes.
addRegisterClass(MVT::f64, X86::RFPRegisterClass);
-
+
+ setOperationAction(ISD::UNDEF, MVT::f64, Expand);
+
if (!UnsafeFPMath) {
setOperationAction(ISD::FSIN , MVT::f64 , Expand);
setOperationAction(ISD::FCOS , MVT::f64 , Expand);
}
+ setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
addLegalFPImmediate(+0.0); // FLD0
addLegalFPImmediate(+1.0); // FLD1
addLegalFPImmediate(-0.0); // FLD0/FCHS
addLegalFPImmediate(-1.0); // FLD1/FCHS
}
+
+ if (TM.getSubtarget<X86Subtarget>().hasMMX()) {
+ addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
+ addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
+ addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
+
+ // FIXME: We don't support any ConstantVec's yet. We should custom expand
+ // the ones we do!
+ setOperationAction(ISD::ConstantVec, MVT::v8i8, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v4i16, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v2i32, Expand);
+ }
+
+ if (TM.getSubtarget<X86Subtarget>().hasSSE1()) {
+ addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
+
+ // FIXME: We don't support any ConstantVec's yet. We should custom expand
+ // the ones we do!
+ setOperationAction(ISD::ConstantVec, MVT::v4f32, Expand);
+ }
+
+ if (TM.getSubtarget<X86Subtarget>().hasSSE2()) {
+ addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
+ addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
+ addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
+ addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
+ addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
+
+
+ // FIXME: We don't support any ConstantVec's yet. We should custom expand
+ // the ones we do!
+ setOperationAction(ISD::ConstantVec, MVT::v2f64, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v16i8, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v8i16, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v4i32, Expand);
+ setOperationAction(ISD::ConstantVec, MVT::v2i64, Expand);
+ }
+
computeRegisterProperties();
- maxStoresPerMemSet = 8; // For %llvm.memset -> sequence of stores
- maxStoresPerMemCpy = 8; // For %llvm.memcpy -> sequence of stores
- maxStoresPerMemMove = 8; // For %llvm.memmove -> sequence of stores
+ // FIXME: These should be based on subtarget info. Plus, the values should
+ // be smaller when we are in optimizing for size mode.
+ maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
+ maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
+ maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
allowUnalignedMemoryAccesses = true; // x86 supports it!
}
SelectionDAG &DAG) {
assert((!isVarArg || CallingConv == CallingConv::C) &&
"Only C takes varargs!");
+
+ // If the callee is a GlobalAddress node (quite common, every direct call is)
+ // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
+ if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
+ Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
+ else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
+ Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
+
if (CallingConv == CallingConv::Fast && EnableFastCC)
return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
return LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
}
-SDOperand X86TargetLowering::LowerReturnTo(SDOperand Chain, SDOperand Op,
- SelectionDAG &DAG) {
- if (!X86DAGIsel)
- return DAG.getNode(ISD::RET, MVT::Other, Chain, Op);
-
- SDOperand Copy;
- MVT::ValueType OpVT = Op.getValueType();
- switch (OpVT) {
- default: assert(0 && "Unknown type to return!");
- case MVT::i32:
- Copy = DAG.getCopyToReg(Chain, X86::EAX, Op, SDOperand());
- break;
- case MVT::i64: {
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op,
- DAG.getConstant(1, MVT::i32));
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op,
- DAG.getConstant(0, MVT::i32));
- Copy = DAG.getCopyToReg(Chain, X86::EAX, Hi, SDOperand());
- Copy = DAG.getCopyToReg(Copy, X86::EDX, Lo, Copy.getValue(1));
- break;
- }
- case MVT::f32:
- assert(X86ScalarSSE && "MVT::f32 only legal with scalar sse fp");
- // Fallthrough intended
- case MVT::f64:
- if (!X86ScalarSSE) {
- 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(Op);
- Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
- } else {
- // Spill the value to memory and reload it into top of stack.
- unsigned Size = MVT::getSizeInBits(OpVT)/8;
- MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
- SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Op,
- StackSlot, DAG.getSrcValue(NULL));
- 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(OpVT));
- Copy = DAG.getNode(X86ISD::FLD, Tys, Ops);
- Tys.clear();
- Tys.push_back(MVT::Other);
- Tys.push_back(MVT::Flag);
- Ops.clear();
- Ops.push_back(Copy.getValue(1));
- Ops.push_back(Copy);
- Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
- }
- break;
- }
-
- return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
- Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
- Copy.getValue(1));
-}
-
//===----------------------------------------------------------------------===//
// C Calling Convention implementation
//===----------------------------------------------------------------------===//
if (Args.empty()) {
// Save zero bytes.
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(0, getPointerTy()));
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(0, getPointerTy()));
} else {
for (unsigned i = 0, e = Args.size(); i != e; ++i)
switch (getValueType(Args[i].second)) {
break;
}
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
+ Chain = DAG.getCALLSEQ_START(Chain,
+ DAG.getConstant(NumBytes, getPointerTy()));
// Arguments go on the stack in reverse order, as specified by the ABI.
unsigned ArgOffset = 0;
- SDOperand StackPtr = DAG.getCopyFromReg(DAG.getEntryNode(),
- X86::ESP, MVT::i32);
+ SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
std::vector<SDOperand> Stores;
for (unsigned i = 0, e = Args.size(); i != e; ++i) {
RetVals.push_back(MVT::i32);
break;
}
+
+ 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);
+
+ // FIXME: Do not generate X86ISD::TAILCALL for now.
+ Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
+ SDOperand InFlag = Chain.getValue(1);
+
+ NodeTys.clear();
+ NodeTys.push_back(MVT::Other); // Returns a chain
+ NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
+ Ops.clear();
+ Ops.push_back(Chain);
Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
Ops.push_back(DAG.getConstant(0, getPointerTy()));
- SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
- RetVals, Ops);
- Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
+ InFlag = Chain.getValue(1);
+
+ SDOperand RetVal;
+ if (RetTyVT != MVT::isVoid) {
+ switch (RetTyVT) {
+ default: assert(0 && "Unknown value type to return!");
+ case MVT::i1:
+ case MVT::i8:
+ RetVal = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag);
+ Chain = RetVal.getValue(1);
+ if (RetTyVT == MVT::i1)
+ RetVal = DAG.getNode(ISD::TRUNCATE, MVT::i1, RetVal);
+ break;
+ case MVT::i16:
+ RetVal = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
+ Chain = RetVal.getValue(1);
+ break;
+ case MVT::i32:
+ RetVal = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
+ Chain = RetVal.getValue(1);
+ break;
+ case MVT::i64: {
+ SDOperand Lo = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
+ SDOperand Hi = DAG.getCopyFromReg(Lo.getValue(1), X86::EDX, MVT::i32,
+ Lo.getValue(2));
+ RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
+ Chain = Hi.getValue(1);
+ break;
+ }
+ case MVT::f32:
+ case MVT::f64: {
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::f64);
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(InFlag);
+ RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
+ Chain = RetVal.getValue(1);
+ InFlag = RetVal.getValue(2);
+ if (X86ScalarSSE) {
+ // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
+ // shouldn't be necessary except that RFP cannot be live across
+ // multiple blocks. When stackifier is fixed, they can be uncoupled.
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ Tys.clear();
+ Tys.push_back(MVT::Other);
+ Ops.clear();
+ Ops.push_back(Chain);
+ Ops.push_back(RetVal);
+ Ops.push_back(StackSlot);
+ Ops.push_back(DAG.getValueType(RetTyVT));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
+ RetVal = DAG.getLoad(RetTyVT, Chain, StackSlot,
+ DAG.getSrcValue(NULL));
+ Chain = RetVal.getValue(1);
+ }
- SDOperand ResultVal;
- switch (RetTyVT) {
- case MVT::isVoid: break;
- default:
- ResultVal = TheCall.getValue(1);
- break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- ResultVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, TheCall.getValue(1));
- break;
- case MVT::f32:
- // FIXME: we would really like to remember that this FP_ROUND operation is
- // okay to eliminate if we allow excess FP precision.
- ResultVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, TheCall.getValue(1));
- break;
- case MVT::i64:
- ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, TheCall.getValue(1),
- TheCall.getValue(2));
- break;
+ if (RetTyVT == MVT::f32 && !X86ScalarSSE)
+ // FIXME: we would really like to remember that this FP_ROUND
+ // operation is okay to eliminate if we allow excess FP precision.
+ RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
+ break;
+ }
+ }
}
- return std::make_pair(ResultVal, Chain);
-}
-
-SDOperand
-X86TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG) {
- // vastart just stores the address of the VarArgsFrameIndex slot.
- SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
- return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR, VAListP,
- DAG.getSrcValue(VAListV));
-}
-
-
-std::pair<SDOperand,SDOperand>
-X86TargetLowering::LowerVAArg(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, const Type *ArgTy,
- SelectionDAG &DAG) {
- MVT::ValueType ArgVT = getValueType(ArgTy);
- SDOperand Val = DAG.getLoad(MVT::i32, Chain,
- VAListP, DAG.getSrcValue(VAListV));
- SDOperand Result = DAG.getLoad(ArgVT, Chain, Val,
- DAG.getSrcValue(NULL));
- unsigned Amt;
- if (ArgVT == MVT::i32)
- Amt = 4;
- else {
- assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
- "Other types should have been promoted for varargs!");
- Amt = 8;
- }
- Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
- DAG.getConstant(Amt, Val.getValueType()));
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Val, VAListP, DAG.getSrcValue(VAListV));
- return std::make_pair(Result, Chain);
+ return std::make_pair(RetVal, Chain);
}
//===----------------------------------------------------------------------===//
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;
if ((NumBytes & 7) == 0)
NumBytes += 4;
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
+ Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
// Arguments go on the stack in reverse order, as specified by the ABI.
unsigned ArgOffset = 0;
- SDOperand StackPtr = DAG.getCopyFromReg(DAG.getEntryNode(),
- X86::ESP, MVT::i32);
+ SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
NumIntRegs = 0;
std::vector<SDOperand> Stores;
std::vector<SDOperand> RegValuesToPass;
switch (getValueType(Args[i].second)) {
default: assert(0 && "Unexpected ValueType for argument!");
case MVT::i1:
+ Args[i].first = DAG.getNode(ISD::ANY_EXTEND, MVT::i8, Args[i].first);
+ // Fall through.
case MVT::i8:
case MVT::i16:
case MVT::i32:
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;
+ SDOperand RegToPass = RegValuesToPass[i];
+ switch (RegToPass.getValueType()) {
+ default: assert(0 && "Bad thing to pass in regs");
+ case MVT::i8:
+ CCReg = (i == 0) ? X86::AL : X86::DL;
+ break;
+ case MVT::i16:
+ CCReg = (i == 0) ? X86::AX : X86::DX;
+ break;
+ case MVT::i32:
+ CCReg = (i == 0) ? X86::EAX : X86::EDX;
+ break;
+ }
+
+ Chain = DAG.getCopyToReg(Chain, CCReg, 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);
- Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
- // Callee pops all arg values on the stack.
- Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
+ if (InFlag.Val)
+ Ops.push_back(InFlag);
- // Pass register arguments as needed.
- Ops.insert(Ops.end(), RegValuesToPass.begin(), RegValuesToPass.end());
+ // FIXME: Do not generate X86ISD::TAILCALL for now.
+ Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
+ InFlag = Chain.getValue(1);
- SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
- RetVals, Ops);
- Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
+ NodeTys.clear();
+ NodeTys.push_back(MVT::Other); // Returns a chain
+ NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
+ Ops.clear();
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
+ Ops.push_back(DAG.getConstant(ArgOffset, getPointerTy()));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
+ InFlag = Chain.getValue(1);
+
+ SDOperand RetVal;
+ if (RetTyVT != MVT::isVoid) {
+ switch (RetTyVT) {
+ default: assert(0 && "Unknown value type to return!");
+ case MVT::i1:
+ case MVT::i8:
+ RetVal = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag);
+ Chain = RetVal.getValue(1);
+ if (RetTyVT == MVT::i1)
+ RetVal = DAG.getNode(ISD::TRUNCATE, MVT::i1, RetVal);
+ break;
+ case MVT::i16:
+ RetVal = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
+ Chain = RetVal.getValue(1);
+ break;
+ case MVT::i32:
+ RetVal = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
+ Chain = RetVal.getValue(1);
+ break;
+ case MVT::i64: {
+ SDOperand Lo = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
+ SDOperand Hi = DAG.getCopyFromReg(Lo.getValue(1), X86::EDX, MVT::i32,
+ Lo.getValue(2));
+ RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
+ Chain = Hi.getValue(1);
+ break;
+ }
+ case MVT::f32:
+ case MVT::f64: {
+ std::vector<MVT::ValueType> Tys;
+ Tys.push_back(MVT::f64);
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(InFlag);
+ RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
+ Chain = RetVal.getValue(1);
+ InFlag = RetVal.getValue(2);
+ if (X86ScalarSSE) {
+ // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
+ // shouldn't be necessary except that RFP cannot be live across
+ // multiple blocks. When stackifier is fixed, they can be uncoupled.
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+ Tys.clear();
+ Tys.push_back(MVT::Other);
+ Ops.clear();
+ Ops.push_back(Chain);
+ Ops.push_back(RetVal);
+ Ops.push_back(StackSlot);
+ Ops.push_back(DAG.getValueType(RetTyVT));
+ Ops.push_back(InFlag);
+ Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
+ RetVal = DAG.getLoad(RetTyVT, Chain, StackSlot,
+ DAG.getSrcValue(NULL));
+ Chain = RetVal.getValue(1);
+ }
- SDOperand ResultVal;
- switch (RetTyVT) {
- case MVT::isVoid: break;
- default:
- ResultVal = TheCall.getValue(1);
- break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- ResultVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, TheCall.getValue(1));
- break;
- case MVT::f32:
- // FIXME: we would really like to remember that this FP_ROUND operation is
- // okay to eliminate if we allow excess FP precision.
- ResultVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, TheCall.getValue(1));
- break;
- case MVT::i64:
- ResultVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, TheCall.getValue(1),
- TheCall.getValue(2));
- break;
+ if (RetTyVT == MVT::f32 && !X86ScalarSSE)
+ // FIXME: we would really like to remember that this FP_ROUND
+ // operation is okay to eliminate if we allow excess FP precision.
+ RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
+ break;
+ }
+ }
}
- return std::make_pair(ResultVal, Chain);
+ return std::make_pair(RetVal, Chain);
}
SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
return std::make_pair(Result, Chain);
}
+/// getCondBrOpcodeForX86CC - Returns the X86 conditional branch opcode
+/// which corresponds to the condition code.
+static unsigned getCondBrOpcodeForX86CC(unsigned X86CC) {
+ switch (X86CC) {
+ default: assert(0 && "Unknown X86 conditional code!");
+ case X86ISD::COND_A: return X86::JA;
+ case X86ISD::COND_AE: return X86::JAE;
+ case X86ISD::COND_B: return X86::JB;
+ case X86ISD::COND_BE: return X86::JBE;
+ case X86ISD::COND_E: return X86::JE;
+ case X86ISD::COND_G: return X86::JG;
+ case X86ISD::COND_GE: return X86::JGE;
+ case X86ISD::COND_L: return X86::JL;
+ case X86ISD::COND_LE: return X86::JLE;
+ case X86ISD::COND_NE: return X86::JNE;
+ case X86ISD::COND_NO: return X86::JNO;
+ case X86ISD::COND_NP: return X86::JNP;
+ case X86ISD::COND_NS: return X86::JNS;
+ case X86ISD::COND_O: return X86::JO;
+ case X86ISD::COND_P: return X86::JP;
+ case X86ISD::COND_S: return X86::JS;
+ }
+}
+
+/// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
+/// 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();
+ Flip = false;
+ X86CC = X86ISD::COND_INVALID;
+ if (!isFP) {
+ switch (SetCCOpcode) {
+ default: break;
+ case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
+ case ISD::SETGT: X86CC = X86ISD::COND_G; break;
+ case ISD::SETGE: X86CC = X86ISD::COND_GE; break;
+ case ISD::SETLT: X86CC = X86ISD::COND_L; break;
+ case ISD::SETLE: X86CC = X86ISD::COND_LE; break;
+ case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
+ case ISD::SETULT: X86CC = X86ISD::COND_B; break;
+ case ISD::SETUGT: X86CC = X86ISD::COND_A; break;
+ case ISD::SETULE: X86CC = X86ISD::COND_BE; break;
+ case ISD::SETUGE: X86CC = X86ISD::COND_AE; break;
+ }
+ } else {
+ // On a floating point condition, the flags are set as follows:
+ // ZF PF CF op
+ // 0 | 0 | 0 | X > Y
+ // 0 | 0 | 1 | X < Y
+ // 1 | 0 | 0 | X == Y
+ // 1 | 1 | 1 | unordered
+ switch (SetCCOpcode) {
+ default: break;
+ case ISD::SETUEQ:
+ case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
+ case ISD::SETOLE: Flip = true; // Fallthrough
+ case ISD::SETOGT:
+ case ISD::SETGT: X86CC = X86ISD::COND_A; break;
+ case ISD::SETOLT: Flip = true; // Fallthrough
+ case ISD::SETOGE:
+ case ISD::SETGE: X86CC = X86ISD::COND_AE; break;
+ case ISD::SETUGE: Flip = true; // Fallthrough
+ case ISD::SETULT:
+ case ISD::SETLT: X86CC = X86ISD::COND_B; break;
+ case ISD::SETUGT: Flip = true; // Fallthrough
+ case ISD::SETULE:
+ case ISD::SETLE: X86CC = X86ISD::COND_BE; break;
+ case ISD::SETONE:
+ case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
+ case ISD::SETUO: X86CC = X86ISD::COND_P; break;
+ case ISD::SETO: X86CC = X86ISD::COND_NP; break;
+ }
+ }
+
+ return X86CC != X86ISD::COND_INVALID;
+}
+
+/// 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.
+static bool hasFPCMov(unsigned X86CC) {
+ switch (X86CC) {
+ default:
+ return false;
+ case X86ISD::COND_B:
+ case X86ISD::COND_BE:
+ case X86ISD::COND_E:
+ case X86ISD::COND_P:
+ case X86ISD::COND_A:
+ case X86ISD::COND_AE:
+ case X86ISD::COND_NE:
+ case X86ISD::COND_NP:
+ return true;
+ }
+}
+
+MachineBasicBlock *
+X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
+ MachineBasicBlock *BB) {
+ switch (MI->getOpcode()) {
+ default: assert(false && "Unexpected instr type to insert");
+ case X86::CMOV_FR32:
+ case X86::CMOV_FR64: {
+ // To "insert" a SELECT_CC instruction, we actually have to insert the
+ // diamond control-flow pattern. The incoming instruction knows the
+ // destination vreg to set, the condition code register to branch on, the
+ // true/false values to select between, and a branch opcode to use.
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ ilist<MachineBasicBlock>::iterator It = BB;
+ ++It;
+
+ // thisMBB:
+ // ...
+ // TrueVal = ...
+ // cmpTY ccX, r1, r2
+ // bCC copy1MBB
+ // fallthrough --> copy0MBB
+ MachineBasicBlock *thisMBB = BB;
+ MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
+ unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
+ BuildMI(BB, Opc, 1).addMBB(sinkMBB);
+ MachineFunction *F = BB->getParent();
+ F->getBasicBlockList().insert(It, copy0MBB);
+ F->getBasicBlockList().insert(It, sinkMBB);
+ // Update machine-CFG edges
+ BB->addSuccessor(copy0MBB);
+ BB->addSuccessor(sinkMBB);
+
+ // copy0MBB:
+ // %FalseValue = ...
+ // # fallthrough to sinkMBB
+ BB = copy0MBB;
+
+ // Update machine-CFG edges
+ BB->addSuccessor(sinkMBB);
+
+ // sinkMBB:
+ // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
+ // ...
+ BB = sinkMBB;
+ BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
+ .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
+ .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
+
+ delete MI; // The pseudo instruction is gone now.
+ return BB;
+ }
+
+ case X86::FP_TO_INT16_IN_MEM:
+ case X86::FP_TO_INT32_IN_MEM:
+ case X86::FP_TO_INT64_IN_MEM: {
+ // Change the floating point control register to use "round towards zero"
+ // mode when truncating to an integer value.
+ MachineFunction *F = BB->getParent();
+ int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
+ addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
+
+ // Load the old value of the high byte of the control word...
+ unsigned OldCW =
+ F->getSSARegMap()->createVirtualRegister(X86::R16RegisterClass);
+ addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
+
+ // Set the high part to be round to zero...
+ addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
+
+ // Reload the modified control word now...
+ addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
+
+ // Restore the memory image of control word to original value
+ addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
+
+ // Get the X86 opcode to use.
+ unsigned Opc;
+ switch (MI->getOpcode()) {
+ default: assert(0 && "illegal opcode!");
+ case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
+ case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
+ case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
+ }
+
+ X86AddressMode AM;
+ MachineOperand &Op = MI->getOperand(0);
+ if (Op.isRegister()) {
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = Op.getReg();
+ } else {
+ AM.BaseType = X86AddressMode::FrameIndexBase;
+ AM.Base.FrameIndex = Op.getFrameIndex();
+ }
+ Op = MI->getOperand(1);
+ if (Op.isImmediate())
+ AM.Scale = Op.getImmedValue();
+ Op = MI->getOperand(2);
+ if (Op.isImmediate())
+ AM.IndexReg = Op.getImmedValue();
+ Op = MI->getOperand(3);
+ if (Op.isGlobalAddress()) {
+ AM.GV = Op.getGlobal();
+ } else {
+ AM.Disp = Op.getImmedValue();
+ }
+ addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(MI->getOperand(4).getReg());
+
+ // Reload the original control word now.
+ addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
+
+ delete MI; // The pseudo instruction is gone now.
+ return BB;
+ }
+ }
+}
+
+
//===----------------------------------------------------------------------===//
// X86 Custom Lowering Hooks
//===----------------------------------------------------------------------===//
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);
+
+ 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);
+ }
+
+ 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::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);
+
+ 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);
+
+ 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);
+ }
case ISD::SINT_TO_FP: {
- assert(Op.getValueType() == MVT::f64 &&
- Op.getOperand(0).getValueType() == MVT::i64 &&
+ assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
+ Op.getOperand(0).getValueType() >= MVT::i16 &&
"Unknown SINT_TO_FP to lower!");
- // We lower sint64->FP into a store to a temporary stack slot, followed by a
- // FILD64m node.
+
+ SDOperand Result;
+ MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
+ unsigned Size = MVT::getSizeInBits(SrcVT)/8;
MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
- Op.getOperand(0), StackSlot, DAG.getSrcValue(NULL));
- std::vector<MVT::ValueType> RTs;
- RTs.push_back(MVT::f64);
- RTs.push_back(MVT::Other);
+ 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(Store);
+ Ops.push_back(Chain);
Ops.push_back(StackSlot);
- return DAG.getNode(X86ISD::FILD64m, RTs, Ops);
+ 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;
}
case ISD::FP_TO_SINT: {
assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
- Op.getOperand(0).getValueType() == MVT::f64 &&
"Unknown FP_TO_SINT to lower!");
// We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
// stack slot.
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(DAG.getEntryNode());
- Ops.push_back(Op.getOperand(0));
+ Ops.push_back(Chain);
+ Ops.push_back(Value);
Ops.push_back(StackSlot);
SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
Tys.push_back(MVT::Other);
return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
}
+ 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));
+ }
+ 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));
+ }
+ 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 CC = Op.getOperand(2);
- SDOperand Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Op.getOperand(0), Op.getOperand(1));
- return DAG.getNode(X86ISD::SETCC, MVT::i8, CC, Cond);
+ 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);
+ }
+ 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);
+ }
+ }
+ }
}
case ISD::SELECT: {
- SDOperand Cond = Op.getOperand(0);
- SDOperand CC;
- if (Cond.getOpcode() == X86ISD::SETCC) {
- CC = Cond.getOperand(0);
- Cond = Cond.getOperand(1);
- } else if (Cond.getOpcode() == ISD::SETCC) {
- CC = Cond.getOperand(2);
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Cond.getOperand(0), Cond.getOperand(1));
- } else {
- CC = DAG.getCondCode(ISD::SETEQ);
- Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
+ 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;
+
+ if (addTest) {
+ CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
+ Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
}
- return DAG.getNode(X86ISD::CMOV, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2), CC, Cond);
+
+ 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;
- // TODO: handle Cond == OR / AND / XOR
+ if (Cond.getOpcode() == ISD::SETCC)
+ Cond = LowerOperation(Cond, DAG);
+
if (Cond.getOpcode() == X86ISD::SETCC) {
- CC = Cond.getOperand(0);
- Cond = Cond.getOperand(1);
- } else if (Cond.getOpcode() == ISD::SETCC) {
- CC = Cond.getOperand(2);
- Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
- Cond.getOperand(0), Cond.getOperand(1));
- } else {
- CC = DAG.getCondCode(ISD::SETNE);
+ // 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);
+ }
+
+ 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;
+
+ if (addTest) {
+ CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
}
return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
Op.getOperand(0), Op.getOperand(2), CC, Cond);
}
- case ISD::GlobalAddress:
- GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
- SDOperand GVOp = DAG.getTargetGlobalAddress(GV, getPointerTy());
- // 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.
+ case ISD::MEMSET: {
+ SDOperand InFlag;
+ 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;
+ }
+
+ MVT::ValueType AVT;
+ SDOperand Count;
+ if (ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2))) {
+ 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;
+ if (I)
+ Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
+ else
+ Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
+ DAG.getConstant(1, MVT::i8));
+ 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);
+ else
+ Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
+ DAG.getConstant(2, MVT::i8));
+ 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;
+ }
+
+ 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);
+ }
+
+ 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);
+
+ return DAG.getNode(X86ISD::REP_STOS, MVT::Other, Chain,
+ DAG.getValueType(AVT), InFlag);
+ }
+ 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;
+ switch (Align & 3) {
+ case 2: // WORD aligned
+ AVT = MVT::i16;
+ if (I)
+ Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
+ else
+ Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
+ break;
+ case 0: // DWORD aligned
+ AVT = MVT::i32;
+ if (I)
+ Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
+ else
+ Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
+ DAG.getConstant(2, MVT::i8));
+ break;
+ default: // Byte aligned
+ AVT = MVT::i8;
+ Count = Op.getOperand(3);
+ break;
+ }
+
+ SDOperand InFlag;
+ 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);
+
+ return DAG.getNode(X86ISD::REP_MOVS, MVT::Other, Chain,
+ DAG.getValueType(AVT), InFlag);
+ }
+ case ISD::ConstantPool: {
+ ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
+ SDOperand Result =
+ DAG.getTargetConstantPool(CP->get(), getPointerTy(), CP->getAlignment());
+ // Only lower ConstantPool on Darwin.
+ if (getTargetMachine().getSubtarget<X86Subtarget>().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::GlobalAddress: {
+ SDOperand Result;
+ // Only lower GlobalAddress on Darwin.
if (getTargetMachine().
- getSubtarget<X86Subtarget>().getIndirectExternAndWeakGlobals() &&
- (GV->hasWeakLinkage() || GV->isExternal()))
- return DAG.getLoad(MVT::i32, DAG.getEntryNode(),
- GVOp, DAG.getSrcValue(NULL));
- else
- return GVOp;
- break;
+ getSubtarget<X86Subtarget>().isTargetDarwin()) {
+ GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
+ Result = DAG.getNode(X86ISD::TGAWrapper, getPointerTy(),
+ DAG.getTargetGlobalAddress(GV, getPointerTy()));
+ // 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 &&
+ (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
+ (GV->isExternal() && !GV->hasNotBeenReadFromBytecode())))
+ Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
+ Result, DAG.getSrcValue(NULL));
+ }
+
+ 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;
+
+ switch(Op.getNumOperands()) {
+ default:
+ assert(0 && "Do not know how to return this many arguments!");
+ abort();
+ case 1:
+ return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
+ DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
+ case 2: {
+ MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
+ if (MVT::isInteger(ArgVT))
+ Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EAX, Op.getOperand(1),
+ SDOperand());
+ else if (!X86ScalarSSE) {
+ 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));
+ Ops.push_back(Op.getOperand(1));
+ Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
+ } else {
+ SDOperand MemLoc;
+ SDOperand Chain = Op.getOperand(0);
+ SDOperand Value = Op.getOperand(1);
+
+ if (Value.getOpcode() == ISD::LOAD &&
+ (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
+ Chain = Value.getOperand(0);
+ MemLoc = Value.getOperand(1);
+ } else {
+ // Spill the value to memory and reload it into top of stack.
+ unsigned Size = MVT::getSizeInBits(ArgVT)/8;
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
+ MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
+ Chain = DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0),
+ Value, MemLoc, 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(MemLoc);
+ Ops.push_back(DAG.getValueType(ArgVT));
+ Copy = DAG.getNode(X86ISD::FLD, Tys, Ops);
+ Tys.clear();
+ Tys.push_back(MVT::Other);
+ Tys.push_back(MVT::Flag);
+ Ops.clear();
+ Ops.push_back(Copy.getValue(1));
+ Ops.push_back(Copy);
+ Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
+ }
+ break;
+ }
+ case 3:
+ 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));
+ }
}
}
const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
switch (Opcode) {
default: return NULL;
- case X86ISD::FILD64m: return "X86ISD::FILD64m";
+ case X86ISD::SHLD: return "X86ISD::SHLD";
+ case X86ISD::SHRD: return "X86ISD::SHRD";
+ case X86ISD::FAND: return "X86ISD::FAND";
+ case X86ISD::FXOR: return "X86ISD::FXOR";
+ case X86ISD::FILD: return "X86ISD::FILD";
+ case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
case X86ISD::FLD: return "X86ISD::FLD";
+ case X86ISD::FST: return "X86ISD::FST";
+ case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
case X86ISD::CALL: return "X86ISD::CALL";
case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
case X86ISD::CMOV: return "X86ISD::CMOV";
case X86ISD::BRCOND: return "X86ISD::BRCOND";
case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
+ case X86ISD::REP_STOS: return "X86ISD::RET_STOS";
+ case X86ISD::REP_MOVS: return "X86ISD::RET_MOVS";
+ case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
+ case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
+ case X86ISD::TGAWrapper: return "X86ISD::TGAWrapper";
}
}
-bool X86TargetLowering::isMaskedValueZeroForTargetNode(const SDOperand &Op,
- uint64_t Mask) const {
+void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
+ uint64_t Mask,
+ uint64_t &KnownZero,
+ uint64_t &KnownOne,
+ unsigned Depth) const {
unsigned Opc = Op.getOpcode();
+ KnownZero = KnownOne = 0; // Don't know anything.
switch (Opc) {
default:
assert(Opc >= ISD::BUILTIN_OP_END && "Expected a target specific node");
break;
- case X86ISD::SETCC: return (Mask & 1) == 0;
+ case X86ISD::SETCC:
+ KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
+ break;
}
+}
- return false;
+std::vector<unsigned> X86TargetLowering::
+getRegClassForInlineAsmConstraint(const std::string &Constraint,
+ MVT::ValueType VT) const {
+ if (Constraint.size() == 1) {
+ // FIXME: not handling fp-stack yet!
+ // FIXME: not handling MMX registers yet ('y' constraint).
+ switch (Constraint[0]) { // GCC X86 Constraint Letters
+ 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);
+ case 'l': // INDEX_REGS
+ return make_vector<unsigned>(X86::EAX, X86::EBX, X86::ECX, X86::EDX,
+ X86::ESI, X86::EDI, X86::EBP, 0);
+ 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);
+ case 'x': // SSE_REGS if SSE1 allowed
+ if (Subtarget->hasSSE1())
+ return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
+ X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
+ 0);
+ return std::vector<unsigned>();
+ case 'Y': // SSE_REGS if SSE2 allowed
+ if (Subtarget->hasSSE2())
+ return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
+ X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
+ 0);
+ return std::vector<unsigned>();
+ }
+ }
+
+ return std::vector<unsigned>();
}