extern X86VectorEnum X86Vector;
extern bool X86ScalarSSE;
-/// createX86PatternInstructionSelector - This pass converts an LLVM function
-/// into a machine code representation in a more aggressive way.
+/// createX86ISelPattern - This pass converts an LLVM function into a
+/// machine code representation using pattern matching and a machine
+/// description file.
///
-FunctionPass *createX86PatternInstructionSelector(TargetMachine &TM);
+FunctionPass *createX86ISelPattern(TargetMachine &TM);
+
+/// createX86ISelDag - This pass converts a legalized DAG into a
+/// X86-specific DAG, ready for instruction scheduling.
+///
+FunctionPass *createX86ISelDag(TargetMachine &TM);
/// createX86SSAPeepholeOptimizerPass - Create a pass to perform SSA-based X86
/// specific peephole optimizations.
--- /dev/null
+//===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file was developed by Chris Lattner and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the interfaces that X86 uses to lower LLVM code into a
+// selection DAG.
+//
+//===----------------------------------------------------------------------===//
+
+#include "X86.h"
+#include "X86ISelLowering.h"
+#include "X86TargetMachine.h"
+#include "llvm/CallingConv.h"
+#include "llvm/Function.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/Target/TargetOptions.h"
+using namespace llvm;
+
+// FIXME: temporary.
+#include "llvm/Support/CommandLine.h"
+static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
+ cl::desc("Enable fastcc on X86"));
+
+X86TargetLowering::X86TargetLowering(TargetMachine &TM)
+ : TargetLowering(TM) {
+
+ // 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);
+ setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
+
+ // Set up the register classes.
+ // FIXME: Eliminate these two classes when legalize can handle promotions
+ // well.
+ addRegisterClass(MVT::i1, X86::R8RegisterClass);
+ addRegisterClass(MVT::i8, X86::R8RegisterClass);
+ addRegisterClass(MVT::i16, X86::R16RegisterClass);
+ addRegisterClass(MVT::i32, X86::R32RegisterClass);
+
+ // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
+ // operation.
+ 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);
+
+ // 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);
+
+ 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);
+ setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
+ }
+
+ // Handle FP_TO_UINT by promoting the destination to a larger signed
+ // conversion.
+ setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
+ setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
+ setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
+
+ if (!X86ScalarSSE)
+ 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::BRCONDTWOWAY , MVT::Other, Expand);
+ setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
+ setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
+ setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
+ setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
+ setOperationAction(ISD::FREM , MVT::f64 , Expand);
+ setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
+ setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
+ setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
+ setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
+ setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
+ setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
+ setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
+ setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
+ setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
+
+ setOperationAction(ISD::READIO , MVT::i1 , Expand);
+ setOperationAction(ISD::READIO , MVT::i8 , Expand);
+ setOperationAction(ISD::READIO , MVT::i16 , Expand);
+ setOperationAction(ISD::READIO , MVT::i32 , Expand);
+ setOperationAction(ISD::WRITEIO , MVT::i1 , Expand);
+ setOperationAction(ISD::WRITEIO , MVT::i8 , Expand);
+ setOperationAction(ISD::WRITEIO , MVT::i16 , Expand);
+ setOperationAction(ISD::WRITEIO , MVT::i32 , Expand);
+
+ // These should be promoted to a larger select which is supported.
+ setOperationAction(ISD::SELECT , MVT::i1 , Promote);
+ setOperationAction(ISD::SELECT , MVT::i8 , Promote);
+
+ if (X86ScalarSSE) {
+ // Set up the FP register classes.
+ addRegisterClass(MVT::f32, X86::V4F4RegisterClass);
+ addRegisterClass(MVT::f64, X86::V2F8RegisterClass);
+
+ // 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);
+
+ // 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
+ 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);
+
+ addLegalFPImmediate(+0.0); // xorps / xorpd
+ } else {
+ // Set up the FP register classes.
+ addRegisterClass(MVT::f64, X86::RFPRegisterClass);
+
+ if (!UnsafeFPMath) {
+ setOperationAction(ISD::FSIN , MVT::f64 , Expand);
+ setOperationAction(ISD::FCOS , MVT::f64 , Expand);
+ }
+
+ addLegalFPImmediate(+0.0); // FLD0
+ addLegalFPImmediate(+1.0); // FLD1
+ addLegalFPImmediate(-0.0); // FLD0/FCHS
+ addLegalFPImmediate(-1.0); // FLD1/FCHS
+ }
+ 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
+ allowUnalignedMemoryAccesses = true; // x86 supports it!
+}
+
+std::vector<SDOperand>
+X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
+ if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
+ return LowerFastCCArguments(F, DAG);
+ return LowerCCCArguments(F, DAG);
+}
+
+std::pair<SDOperand, SDOperand>
+X86TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
+ bool isVarArg, unsigned CallingConv,
+ bool isTailCall,
+ SDOperand Callee, ArgListTy &Args,
+ SelectionDAG &DAG) {
+ assert((!isVarArg || CallingConv == CallingConv::C) &&
+ "Only C takes varargs!");
+ if (CallingConv == CallingConv::Fast && EnableFastCC)
+ return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
+ return LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
+}
+
+//===----------------------------------------------------------------------===//
+// C Calling Convention implementation
+//===----------------------------------------------------------------------===//
+
+std::vector<SDOperand>
+X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
+ std::vector<SDOperand> ArgValues;
+
+ MachineFunction &MF = DAG.getMachineFunction();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+
+ // Add DAG nodes to load the arguments... On entry to a function on the X86,
+ // the stack frame looks like this:
+ //
+ // [ESP] -- return address
+ // [ESP + 4] -- first argument (leftmost lexically)
+ // [ESP + 8] -- second argument, if first argument is four bytes in size
+ // ...
+ //
+ 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);
+ }
+ 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
+ // the start of the first vararg value... for expansion of llvm.va_start.
+ if (F.isVarArg())
+ VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
+ 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;
+ }
+ return ArgValues;
+}
+
+std::pair<SDOperand, SDOperand>
+X86TargetLowering::LowerCCCCallTo(SDOperand Chain, const Type *RetTy,
+ bool isVarArg, bool isTailCall,
+ SDOperand Callee, ArgListTy &Args,
+ SelectionDAG &DAG) {
+ // Count how many bytes are to be pushed on the stack.
+ unsigned NumBytes = 0;
+
+ if (Args.empty()) {
+ // Save zero bytes.
+ Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
+ DAG.getConstant(0, getPointerTy()));
+ } else {
+ for (unsigned i = 0, e = Args.size(); i != e; ++i)
+ switch (getValueType(Args[i].second)) {
+ default: assert(0 && "Unknown value type!");
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ case MVT::f32:
+ NumBytes += 4;
+ break;
+ case MVT::i64:
+ case MVT::f64:
+ NumBytes += 8;
+ break;
+ }
+
+ Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, 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);
+ 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:
+ case MVT::i8:
+ case MVT::i16:
+ // Promote the integer to 32 bits. If the input type is signed use a
+ // sign extend, otherwise use a zero extend.
+ if (Args[i].second->isSigned())
+ Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
+ else
+ Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
+
+ // FALL THROUGH
+ case MVT::i32:
+ case MVT::f32:
+ 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:
+ Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
+ Args[i].first, PtrOff,
+ DAG.getSrcValue(NULL)));
+ ArgOffset += 8;
+ break;
+ }
+ }
+ Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
+ }
+
+ std::vector<MVT::ValueType> RetVals;
+ MVT::ValueType RetTyVT = getValueType(RetTy);
+ RetVals.push_back(MVT::Other);
+
+ // The result values produced have to be legal. Promote the result.
+ switch (RetTyVT) {
+ case MVT::isVoid: break;
+ default:
+ RetVals.push_back(RetTyVT);
+ break;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ RetVals.push_back(MVT::i32);
+ break;
+ case MVT::f32:
+ if (X86ScalarSSE)
+ RetVals.push_back(MVT::f32);
+ else
+ RetVals.push_back(MVT::f64);
+ break;
+ case MVT::i64:
+ RetVals.push_back(MVT::i32);
+ RetVals.push_back(MVT::i32);
+ break;
+ }
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(Callee);
+ 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);
+
+ 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;
+ }
+
+ 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);
+}
+
+//===----------------------------------------------------------------------===//
+// Fast Calling Convention implementation
+//===----------------------------------------------------------------------===//
+//
+// The X86 'fast' calling convention passes up to two integer arguments in
+// registers (an appropriate portion of EAX/EDX), passes arguments in C order,
+// and requires that the callee pop its arguments off the stack (allowing proper
+// tail calls), and has the same return value conventions as C calling convs.
+//
+// This calling convention always arranges for the callee pop value to be 8n+4
+// bytes, which is needed for tail recursion elimination and stack alignment
+// reasons.
+//
+// Note that this can be enhanced in the future to pass fp vals in registers
+// (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;
+}
+
+
+std::vector<SDOperand>
+X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
+ std::vector<SDOperand> ArgValues;
+
+ MachineFunction &MF = DAG.getMachineFunction();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+
+ // Add DAG nodes to load the arguments... On entry to a function the stack
+ // frame looks like this:
+ //
+ // [ESP] -- return address
+ // [ESP + 4] -- first nonreg argument (leftmost lexically)
+ // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
+ // ...
+ unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
+
+ // Keep track of the number of integer regs passed so far. This can be either
+ // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
+ // used).
+ unsigned NumIntRegs = 0;
+
+ 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 < 2) {
+ 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));
+ }
+ ++NumIntRegs;
+ break;
+ }
+
+ ObjSize = 1;
+ break;
+ case MVT::i16:
+ if (NumIntRegs < 2) {
+ 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 < 2) {
+ 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));
+ }
+ ++NumIntRegs;
+ break;
+ }
+ ObjSize = 4;
+ break;
+ case MVT::i64:
+ if (NumIntRegs == 0) {
+ 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) {
+ 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);
+ }
+ ArgOffset += 4;
+ NumIntRegs = 2;
+ break;
+ }
+ 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);
+ }
+ 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
+ // arguments and the arguments after the retaddr has been pushed are aligned.
+ if ((ArgOffset & 7) == 0)
+ ArgOffset += 4;
+
+ VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
+ ReturnAddrIndex = 0; // No return address slot generated yet.
+ BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
+ BytesCallerReserves = 0;
+
+ // 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;
+ }
+ return ArgValues;
+}
+
+std::pair<SDOperand, SDOperand>
+X86TargetLowering::LowerFastCCCallTo(SDOperand Chain, const Type *RetTy,
+ bool isTailCall, SDOperand Callee,
+ ArgListTy &Args, SelectionDAG &DAG) {
+ // Count how many bytes are to be pushed on the stack.
+ unsigned NumBytes = 0;
+
+ // Keep track of the number of integer regs passed so far. This can be either
+ // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
+ // used).
+ unsigned NumIntRegs = 0;
+
+ for (unsigned i = 0, e = Args.size(); i != e; ++i)
+ switch (getValueType(Args[i].second)) {
+ default: assert(0 && "Unknown value type!");
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ if (NumIntRegs < 2) {
+ ++NumIntRegs;
+ break;
+ }
+ // fall through
+ case MVT::f32:
+ NumBytes += 4;
+ break;
+ case MVT::i64:
+ if (NumIntRegs == 0) {
+ NumIntRegs = 2;
+ break;
+ } else if (NumIntRegs == 1) {
+ NumIntRegs = 2;
+ NumBytes += 4;
+ break;
+ }
+
+ // fall through
+ case MVT::f64:
+ NumBytes += 8;
+ break;
+ }
+
+ // Make sure the instruction takes 8n+4 bytes to make sure the start of the
+ // arguments and the arguments after the retaddr has been pushed are aligned.
+ if ((NumBytes & 7) == 0)
+ NumBytes += 4;
+
+ Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, 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);
+ NumIntRegs = 0;
+ std::vector<SDOperand> Stores;
+ std::vector<SDOperand> RegValuesToPass;
+ for (unsigned i = 0, e = Args.size(); i != e; ++i) {
+ switch (getValueType(Args[i].second)) {
+ default: assert(0 && "Unexpected ValueType for argument!");
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ if (NumIntRegs < 2) {
+ RegValuesToPass.push_back(Args[i].first);
+ ++NumIntRegs;
+ break;
+ }
+ // Fall through
+ 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:
+ if (NumIntRegs < 2) { // Can pass part of it in regs?
+ SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
+ Args[i].first, DAG.getConstant(1, MVT::i32));
+ SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
+ Args[i].first, DAG.getConstant(0, MVT::i32));
+ RegValuesToPass.push_back(Lo);
+ ++NumIntRegs;
+ if (NumIntRegs < 2) { // Pass both parts in regs?
+ RegValuesToPass.push_back(Hi);
+ ++NumIntRegs;
+ } else {
+ // Pass the high part in memory.
+ 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,
+ Hi, PtrOff, DAG.getSrcValue(NULL)));
+ ArgOffset += 4;
+ }
+ break;
+ }
+ // Fall through
+ 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;
+ }
+ }
+ if (!Stores.empty())
+ Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
+
+ // Make sure the instruction takes 8n+4 bytes to make sure the start of the
+ // arguments and the arguments after the retaddr has been pushed are aligned.
+ if ((ArgOffset & 7) == 0)
+ ArgOffset += 4;
+
+ std::vector<MVT::ValueType> RetVals;
+ MVT::ValueType RetTyVT = getValueType(RetTy);
+
+ RetVals.push_back(MVT::Other);
+
+ // The result values produced have to be legal. Promote the result.
+ switch (RetTyVT) {
+ case MVT::isVoid: break;
+ default:
+ RetVals.push_back(RetTyVT);
+ break;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ RetVals.push_back(MVT::i32);
+ break;
+ case MVT::f32:
+ if (X86ScalarSSE)
+ RetVals.push_back(MVT::f32);
+ else
+ RetVals.push_back(MVT::f64);
+ break;
+ case MVT::i64:
+ RetVals.push_back(MVT::i32);
+ RetVals.push_back(MVT::i32);
+ break;
+ }
+
+ 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()));
+
+ // Pass register arguments as needed.
+ Ops.insert(Ops.end(), RegValuesToPass.begin(), RegValuesToPass.end());
+
+ SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
+ RetVals, Ops);
+ Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
+
+ 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;
+ }
+
+ return std::make_pair(ResultVal, Chain);
+}
+
+SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
+ if (ReturnAddrIndex == 0) {
+ // Set up a frame object for the return address.
+ MachineFunction &MF = DAG.getMachineFunction();
+ ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
+ }
+
+ return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
+}
+
+
+
+std::pair<SDOperand, SDOperand> X86TargetLowering::
+LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
+ SelectionDAG &DAG) {
+ SDOperand Result;
+ if (Depth) // Depths > 0 not supported yet!
+ Result = DAG.getConstant(0, getPointerTy());
+ else {
+ SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
+ if (!isFrameAddress)
+ // Just load the return address
+ Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
+ DAG.getSrcValue(NULL));
+ else
+ Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
+ DAG.getConstant(4, MVT::i32));
+ }
+ return std::make_pair(Result, Chain);
+}
+
+//===----------------------------------------------------------------------===//
+// X86 Custom Lowering Hooks
+//===----------------------------------------------------------------------===//
+
+/// 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::SINT_TO_FP: {
+ assert(Op.getValueType() == MVT::f64 &&
+ Op.getOperand(0).getValueType() == MVT::i64 &&
+ "Unknown SINT_TO_FP to lower!");
+ // We lower sint64->FP into a store to a temporary stack slot, followed by a
+ // FILD64m node.
+ MachineFunction &MF = DAG.getMachineFunction();
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
+ 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);
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Store);
+ Ops.push_back(StackSlot);
+ return DAG.getNode(X86ISD::FILD64m, RTs, Ops);
+ }
+ 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.
+ 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;
+ }
+
+ // 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(StackSlot);
+ SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
+
+ // Load the result.
+ return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
+ DAG.getSrcValue(NULL));
+ }
+ }
+}
--- /dev/null
+//===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file was developed by Chris Lattner and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the interfaces that X86 uses to lower LLVM code into a
+// selection DAG.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef X86ISELLOWERING_H
+#define X86ISELLOWERING_H
+
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+
+namespace llvm {
+ // X86 Specific DAG Nodes
+ namespace X86ISD {
+ enum NodeType {
+ // Start the numbering where the builtin ops leave off.
+ FIRST_NUMBER = ISD::BUILTIN_OP_END,
+
+ /// FILD64m - This instruction implements SINT_TO_FP with a
+ /// 64-bit source in memory and a FP reg result. This corresponds to
+ /// the X86::FILD64m instruction. It has two inputs (token chain and
+ /// address) and two outputs (FP value and token chain).
+ FILD64m,
+
+ /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
+ /// integer destination in memory and a FP reg source. This corresponds
+ /// to the X86::FIST*m instructions and the rounding mode change stuff. It
+ /// has two inputs (token chain and address) and two outputs (FP value and
+ /// token chain).
+ FP_TO_INT16_IN_MEM,
+ FP_TO_INT32_IN_MEM,
+ FP_TO_INT64_IN_MEM,
+
+ /// CALL/TAILCALL - These operations represent an abstract X86 call
+ /// instruction, which includes a bunch of information. In particular the
+ /// operands of these node are:
+ ///
+ /// #0 - The incoming token chain
+ /// #1 - The callee
+ /// #2 - The number of arg bytes the caller pushes on the stack.
+ /// #3 - The number of arg bytes the callee pops off the stack.
+ /// #4 - The value to pass in AL/AX/EAX (optional)
+ /// #5 - The value to pass in DL/DX/EDX (optional)
+ ///
+ /// The result values of these nodes are:
+ ///
+ /// #0 - The outgoing token chain
+ /// #1 - The first register result value (optional)
+ /// #2 - The second register result value (optional)
+ ///
+ /// The CALL vs TAILCALL distinction boils down to whether the callee is
+ /// known not to modify the caller's stack frame, as is standard with
+ /// LLVM.
+ CALL,
+ TAILCALL,
+ };
+ }
+
+ //===----------------------------------------------------------------------===//
+ // X86TargetLowering - X86 Implementation of the TargetLowering interface
+ class X86TargetLowering : public TargetLowering {
+ int VarArgsFrameIndex; // FrameIndex for start of varargs area.
+ int ReturnAddrIndex; // FrameIndex for return slot.
+ int BytesToPopOnReturn; // Number of arg bytes ret should pop.
+ int BytesCallerReserves; // Number of arg bytes caller makes.
+ public:
+ X86TargetLowering(TargetMachine &TM);
+
+ // Return the number of bytes that a function should pop when it returns (in
+ // addition to the space used by the return address).
+ //
+ unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }
+
+ // Return the number of bytes that the caller reserves for arguments passed
+ // to this function.
+ unsigned getBytesCallerReserves() const { return BytesCallerReserves; }
+
+ /// LowerOperation - Provide custom lowering hooks for some operations.
+ ///
+ virtual SDOperand LowerOperation(SDOperand Op, SelectionDAG &DAG);
+
+ /// LowerArguments - This hook must be implemented to indicate how we should
+ /// lower the arguments for the specified function, into the specified DAG.
+ virtual std::vector<SDOperand>
+ LowerArguments(Function &F, SelectionDAG &DAG);
+
+ /// LowerCallTo - This hook lowers an abstract call to a function into an
+ /// actual call.
+ virtual std::pair<SDOperand, SDOperand>
+ LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
+ bool isTailCall, SDOperand Callee, ArgListTy &Args,
+ SelectionDAG &DAG);
+
+ virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
+ Value *VAListV, SelectionDAG &DAG);
+ virtual std::pair<SDOperand,SDOperand>
+ LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
+ const Type *ArgTy, SelectionDAG &DAG);
+
+ virtual std::pair<SDOperand, SDOperand>
+ LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
+ SelectionDAG &DAG);
+
+ SDOperand getReturnAddressFrameIndex(SelectionDAG &DAG);
+
+ private:
+ // C Calling Convention implementation.
+ std::vector<SDOperand> LowerCCCArguments(Function &F, SelectionDAG &DAG);
+ std::pair<SDOperand, SDOperand>
+ LowerCCCCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
+ bool isTailCall,
+ SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
+
+ // Fast Calling Convention implementation.
+ std::vector<SDOperand> LowerFastCCArguments(Function &F, SelectionDAG &DAG);
+ std::pair<SDOperand, SDOperand>
+ LowerFastCCCallTo(SDOperand Chain, const Type *RetTy, bool isTailCall,
+ SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
+ };
+}
+
+#endif // X86ISELLOWERING_H
#include "X86InstrBuilder.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
+#include "X86ISelLowering.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include <algorithm>
using namespace llvm;
-// FIXME: temporary.
-#include "llvm/Support/CommandLine.h"
-static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
- cl::desc("Enable fastcc on X86"));
-
-namespace {
- // X86 Specific DAG Nodes
- namespace X86ISD {
- enum NodeType {
- // Start the numbering where the builtin ops leave off.
- FIRST_NUMBER = ISD::BUILTIN_OP_END,
-
- /// FILD64m - This instruction implements SINT_TO_FP with a
- /// 64-bit source in memory and a FP reg result. This corresponds to
- /// the X86::FILD64m instruction. It has two inputs (token chain and
- /// address) and two outputs (FP value and token chain).
- FILD64m,
-
- /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
- /// integer destination in memory and a FP reg source. This corresponds
- /// to the X86::FIST*m instructions and the rounding mode change stuff. It
- /// has two inputs (token chain and address) and two outputs (FP value and
- /// token chain).
- FP_TO_INT16_IN_MEM,
- FP_TO_INT32_IN_MEM,
- FP_TO_INT64_IN_MEM,
-
- /// CALL/TAILCALL - These operations represent an abstract X86 call
- /// instruction, which includes a bunch of information. In particular the
- /// operands of these node are:
- ///
- /// #0 - The incoming token chain
- /// #1 - The callee
- /// #2 - The number of arg bytes the caller pushes on the stack.
- /// #3 - The number of arg bytes the callee pops off the stack.
- /// #4 - The value to pass in AL/AX/EAX (optional)
- /// #5 - The value to pass in DL/DX/EDX (optional)
- ///
- /// The result values of these nodes are:
- ///
- /// #0 - The outgoing token chain
- /// #1 - The first register result value (optional)
- /// #2 - The second register result value (optional)
- ///
- /// The CALL vs TAILCALL distinction boils down to whether the callee is
- /// known not to modify the caller's stack frame, as is standard with
- /// LLVM.
- CALL,
- TAILCALL,
- };
- }
-}
-
-//===----------------------------------------------------------------------===//
-// X86TargetLowering - X86 Implementation of the TargetLowering interface
-namespace {
- class X86TargetLowering : public TargetLowering {
- int VarArgsFrameIndex; // FrameIndex for start of varargs area.
- int ReturnAddrIndex; // FrameIndex for return slot.
- int BytesToPopOnReturn; // Number of arg bytes ret should pop.
- int BytesCallerReserves; // Number of arg bytes caller makes.
- public:
- X86TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
- // 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);
- setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
-
- // Set up the register classes.
- // FIXME: Eliminate these two classes when legalize can handle promotions
- // well.
- addRegisterClass(MVT::i1, X86::R8RegisterClass);
- addRegisterClass(MVT::i8, X86::R8RegisterClass);
- addRegisterClass(MVT::i16, X86::R16RegisterClass);
- addRegisterClass(MVT::i32, X86::R32RegisterClass);
-
- // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
- // operation.
- 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);
-
- // 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);
-
- 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);
- setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
- }
-
- // Handle FP_TO_UINT by promoting the destination to a larger signed
- // conversion.
- setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
- setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
- setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
-
- if (!X86ScalarSSE)
- 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::BRCONDTWOWAY , MVT::Other, Expand);
- setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
- setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
- setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
- setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
- setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
- setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
- setOperationAction(ISD::FREM , MVT::f64 , Expand);
- setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
- setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
- setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
- setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
- setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
- setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
- setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
- setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
- setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
-
- setOperationAction(ISD::READIO , MVT::i1 , Expand);
- setOperationAction(ISD::READIO , MVT::i8 , Expand);
- setOperationAction(ISD::READIO , MVT::i16 , Expand);
- setOperationAction(ISD::READIO , MVT::i32 , Expand);
- setOperationAction(ISD::WRITEIO , MVT::i1 , Expand);
- setOperationAction(ISD::WRITEIO , MVT::i8 , Expand);
- setOperationAction(ISD::WRITEIO , MVT::i16 , Expand);
- setOperationAction(ISD::WRITEIO , MVT::i32 , Expand);
-
- // These should be promoted to a larger select which is supported.
- setOperationAction(ISD::SELECT , MVT::i1 , Promote);
- setOperationAction(ISD::SELECT , MVT::i8 , Promote);
-
- if (X86ScalarSSE) {
- // Set up the FP register classes.
- addRegisterClass(MVT::f32, X86::V4F4RegisterClass);
- addRegisterClass(MVT::f64, X86::V2F8RegisterClass);
-
- // 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);
-
- // 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
- 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);
-
- addLegalFPImmediate(+0.0); // xorps / xorpd
- } else {
- // Set up the FP register classes.
- addRegisterClass(MVT::f64, X86::RFPRegisterClass);
-
- if (!UnsafeFPMath) {
- setOperationAction(ISD::FSIN , MVT::f64 , Expand);
- setOperationAction(ISD::FCOS , MVT::f64 , Expand);
- }
-
- addLegalFPImmediate(+0.0); // FLD0
- addLegalFPImmediate(+1.0); // FLD1
- addLegalFPImmediate(-0.0); // FLD0/FCHS
- addLegalFPImmediate(-1.0); // FLD1/FCHS
- }
- 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
- allowUnalignedMemoryAccesses = true; // x86 supports it!
- }
-
- // Return the number of bytes that a function should pop when it returns (in
- // addition to the space used by the return address).
- //
- unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }
-
- // Return the number of bytes that the caller reserves for arguments passed
- // to this function.
- unsigned getBytesCallerReserves() const { return BytesCallerReserves; }
-
- /// LowerOperation - Provide custom lowering hooks for some operations.
- ///
- virtual SDOperand LowerOperation(SDOperand Op, SelectionDAG &DAG);
-
- /// LowerArguments - This hook must be implemented to indicate how we should
- /// lower the arguments for the specified function, into the specified DAG.
- virtual std::vector<SDOperand>
- LowerArguments(Function &F, SelectionDAG &DAG);
-
- /// LowerCallTo - This hook lowers an abstract call to a function into an
- /// actual call.
- virtual std::pair<SDOperand, SDOperand>
- LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
- bool isTailCall, SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG);
-
- virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG);
- virtual std::pair<SDOperand,SDOperand>
- LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
- const Type *ArgTy, SelectionDAG &DAG);
-
- virtual std::pair<SDOperand, SDOperand>
- LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
- SelectionDAG &DAG);
-
- SDOperand getReturnAddressFrameIndex(SelectionDAG &DAG);
-
- private:
- // C Calling Convention implementation.
- std::vector<SDOperand> LowerCCCArguments(Function &F, SelectionDAG &DAG);
- std::pair<SDOperand, SDOperand>
- LowerCCCCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
- bool isTailCall,
- SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
-
- // Fast Calling Convention implementation.
- std::vector<SDOperand> LowerFastCCArguments(Function &F, SelectionDAG &DAG);
- std::pair<SDOperand, SDOperand>
- LowerFastCCCallTo(SDOperand Chain, const Type *RetTy, bool isTailCall,
- SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
- };
-}
-
-std::vector<SDOperand>
-X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
- if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
- return LowerFastCCArguments(F, DAG);
- return LowerCCCArguments(F, DAG);
-}
-
-std::pair<SDOperand, SDOperand>
-X86TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
- bool isVarArg, unsigned CallingConv,
- bool isTailCall,
- SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG) {
- assert((!isVarArg || CallingConv == CallingConv::C) &&
- "Only C takes varargs!");
- if (CallingConv == CallingConv::Fast && EnableFastCC)
- return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
- return LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
-}
-
-//===----------------------------------------------------------------------===//
-// C Calling Convention implementation
-//===----------------------------------------------------------------------===//
-
-std::vector<SDOperand>
-X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
- std::vector<SDOperand> ArgValues;
-
- MachineFunction &MF = DAG.getMachineFunction();
- MachineFrameInfo *MFI = MF.getFrameInfo();
-
- // Add DAG nodes to load the arguments... On entry to a function on the X86,
- // the stack frame looks like this:
- //
- // [ESP] -- return address
- // [ESP + 4] -- first argument (leftmost lexically)
- // [ESP + 8] -- second argument, if first argument is four bytes in size
- // ...
- //
- 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);
- }
- 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
- // the start of the first vararg value... for expansion of llvm.va_start.
- if (F.isVarArg())
- VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
- 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;
- }
- return ArgValues;
-}
-
-std::pair<SDOperand, SDOperand>
-X86TargetLowering::LowerCCCCallTo(SDOperand Chain, const Type *RetTy,
- bool isVarArg, bool isTailCall,
- SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG) {
- // Count how many bytes are to be pushed on the stack.
- unsigned NumBytes = 0;
-
- if (Args.empty()) {
- // Save zero bytes.
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(0, getPointerTy()));
- } else {
- for (unsigned i = 0, e = Args.size(); i != e; ++i)
- switch (getValueType(Args[i].second)) {
- default: assert(0 && "Unknown value type!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- case MVT::f32:
- NumBytes += 4;
- break;
- case MVT::i64:
- case MVT::f64:
- NumBytes += 8;
- break;
- }
-
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, 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);
- 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:
- case MVT::i8:
- case MVT::i16:
- // Promote the integer to 32 bits. If the input type is signed use a
- // sign extend, otherwise use a zero extend.
- if (Args[i].second->isSigned())
- Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
- else
- Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
-
- // FALL THROUGH
- case MVT::i32:
- case MVT::f32:
- 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:
- Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Args[i].first, PtrOff,
- DAG.getSrcValue(NULL)));
- ArgOffset += 8;
- break;
- }
- }
- Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
- }
-
- std::vector<MVT::ValueType> RetVals;
- MVT::ValueType RetTyVT = getValueType(RetTy);
- RetVals.push_back(MVT::Other);
-
- // The result values produced have to be legal. Promote the result.
- switch (RetTyVT) {
- case MVT::isVoid: break;
- default:
- RetVals.push_back(RetTyVT);
- break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- RetVals.push_back(MVT::i32);
- break;
- case MVT::f32:
- if (X86ScalarSSE)
- RetVals.push_back(MVT::f32);
- else
- RetVals.push_back(MVT::f64);
- break;
- case MVT::i64:
- RetVals.push_back(MVT::i32);
- RetVals.push_back(MVT::i32);
- break;
- }
- std::vector<SDOperand> Ops;
- Ops.push_back(Chain);
- Ops.push_back(Callee);
- 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);
-
- 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;
- }
-
- 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);
-}
-
-//===----------------------------------------------------------------------===//
-// Fast Calling Convention implementation
-//===----------------------------------------------------------------------===//
-//
-// The X86 'fast' calling convention passes up to two integer arguments in
-// registers (an appropriate portion of EAX/EDX), passes arguments in C order,
-// and requires that the callee pop its arguments off the stack (allowing proper
-// tail calls), and has the same return value conventions as C calling convs.
-//
-// This calling convention always arranges for the callee pop value to be 8n+4
-// bytes, which is needed for tail recursion elimination and stack alignment
-// reasons.
-//
-// Note that this can be enhanced in the future to pass fp vals in registers
-// (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;
-}
-
-
-std::vector<SDOperand>
-X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
- std::vector<SDOperand> ArgValues;
-
- MachineFunction &MF = DAG.getMachineFunction();
- MachineFrameInfo *MFI = MF.getFrameInfo();
-
- // Add DAG nodes to load the arguments... On entry to a function the stack
- // frame looks like this:
- //
- // [ESP] -- return address
- // [ESP + 4] -- first nonreg argument (leftmost lexically)
- // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
- // ...
- unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
-
- // Keep track of the number of integer regs passed so far. This can be either
- // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
- // used).
- unsigned NumIntRegs = 0;
-
- 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 < 2) {
- 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));
- }
- ++NumIntRegs;
- break;
- }
-
- ObjSize = 1;
- break;
- case MVT::i16:
- if (NumIntRegs < 2) {
- 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 < 2) {
- 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));
- }
- ++NumIntRegs;
- break;
- }
- ObjSize = 4;
- break;
- case MVT::i64:
- if (NumIntRegs == 0) {
- 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) {
- 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);
- }
- ArgOffset += 4;
- NumIntRegs = 2;
- break;
- }
- 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);
- }
- 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
- // arguments and the arguments after the retaddr has been pushed are aligned.
- if ((ArgOffset & 7) == 0)
- ArgOffset += 4;
-
- VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
- ReturnAddrIndex = 0; // No return address slot generated yet.
- BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
- BytesCallerReserves = 0;
-
- // 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;
- }
- return ArgValues;
-}
-
-std::pair<SDOperand, SDOperand>
-X86TargetLowering::LowerFastCCCallTo(SDOperand Chain, const Type *RetTy,
- bool isTailCall, SDOperand Callee,
- ArgListTy &Args, SelectionDAG &DAG) {
- // Count how many bytes are to be pushed on the stack.
- unsigned NumBytes = 0;
-
- // Keep track of the number of integer regs passed so far. This can be either
- // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
- // used).
- unsigned NumIntRegs = 0;
-
- for (unsigned i = 0, e = Args.size(); i != e; ++i)
- switch (getValueType(Args[i].second)) {
- default: assert(0 && "Unknown value type!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- if (NumIntRegs < 2) {
- ++NumIntRegs;
- break;
- }
- // fall through
- case MVT::f32:
- NumBytes += 4;
- break;
- case MVT::i64:
- if (NumIntRegs == 0) {
- NumIntRegs = 2;
- break;
- } else if (NumIntRegs == 1) {
- NumIntRegs = 2;
- NumBytes += 4;
- break;
- }
-
- // fall through
- case MVT::f64:
- NumBytes += 8;
- break;
- }
-
- // Make sure the instruction takes 8n+4 bytes to make sure the start of the
- // arguments and the arguments after the retaddr has been pushed are aligned.
- if ((NumBytes & 7) == 0)
- NumBytes += 4;
-
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, 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);
- NumIntRegs = 0;
- std::vector<SDOperand> Stores;
- std::vector<SDOperand> RegValuesToPass;
- for (unsigned i = 0, e = Args.size(); i != e; ++i) {
- switch (getValueType(Args[i].second)) {
- default: assert(0 && "Unexpected ValueType for argument!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- if (NumIntRegs < 2) {
- RegValuesToPass.push_back(Args[i].first);
- ++NumIntRegs;
- break;
- }
- // Fall through
- 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:
- if (NumIntRegs < 2) { // Can pass part of it in regs?
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
- Args[i].first, DAG.getConstant(1, MVT::i32));
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
- Args[i].first, DAG.getConstant(0, MVT::i32));
- RegValuesToPass.push_back(Lo);
- ++NumIntRegs;
- if (NumIntRegs < 2) { // Pass both parts in regs?
- RegValuesToPass.push_back(Hi);
- ++NumIntRegs;
- } else {
- // Pass the high part in memory.
- 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,
- Hi, PtrOff, DAG.getSrcValue(NULL)));
- ArgOffset += 4;
- }
- break;
- }
- // Fall through
- 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;
- }
- }
- if (!Stores.empty())
- Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
-
- // Make sure the instruction takes 8n+4 bytes to make sure the start of the
- // arguments and the arguments after the retaddr has been pushed are aligned.
- if ((ArgOffset & 7) == 0)
- ArgOffset += 4;
-
- std::vector<MVT::ValueType> RetVals;
- MVT::ValueType RetTyVT = getValueType(RetTy);
-
- RetVals.push_back(MVT::Other);
-
- // The result values produced have to be legal. Promote the result.
- switch (RetTyVT) {
- case MVT::isVoid: break;
- default:
- RetVals.push_back(RetTyVT);
- break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- RetVals.push_back(MVT::i32);
- break;
- case MVT::f32:
- if (X86ScalarSSE)
- RetVals.push_back(MVT::f32);
- else
- RetVals.push_back(MVT::f64);
- break;
- case MVT::i64:
- RetVals.push_back(MVT::i32);
- RetVals.push_back(MVT::i32);
- break;
- }
-
- 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()));
-
- // Pass register arguments as needed.
- Ops.insert(Ops.end(), RegValuesToPass.begin(), RegValuesToPass.end());
-
- SDOperand TheCall = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
- RetVals, Ops);
- Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, TheCall);
-
- 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;
- }
-
- return std::make_pair(ResultVal, Chain);
-}
-
-SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
- if (ReturnAddrIndex == 0) {
- // Set up a frame object for the return address.
- MachineFunction &MF = DAG.getMachineFunction();
- ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
- }
-
- return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
-}
-
-
-
-std::pair<SDOperand, SDOperand> X86TargetLowering::
-LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
- SelectionDAG &DAG) {
- SDOperand Result;
- if (Depth) // Depths > 0 not supported yet!
- Result = DAG.getConstant(0, getPointerTy());
- else {
- SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
- if (!isFrameAddress)
- // Just load the return address
- Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
- DAG.getSrcValue(NULL));
- else
- Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
- DAG.getConstant(4, MVT::i32));
- }
- return std::make_pair(Result, Chain);
-}
-
-//===----------------------------------------------------------------------===//
-// X86 Custom Lowering Hooks
-//===----------------------------------------------------------------------===//
-
-/// 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::SINT_TO_FP: {
- assert(Op.getValueType() == MVT::f64 &&
- Op.getOperand(0).getValueType() == MVT::i64 &&
- "Unknown SINT_TO_FP to lower!");
- // We lower sint64->FP into a store to a temporary stack slot, followed by a
- // FILD64m node.
- MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
- 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);
- std::vector<SDOperand> Ops;
- Ops.push_back(Store);
- Ops.push_back(StackSlot);
- return DAG.getNode(X86ISD::FILD64m, RTs, Ops);
- }
- 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.
- 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;
- }
-
- // 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(StackSlot);
- SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
-
- // Load the result.
- return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
- DAG.getSrcValue(NULL));
- }
- }
-}
-
-
//===----------------------------------------------------------------------===//
// Pattern Matcher Implementation
//===----------------------------------------------------------------------===//
}
-/// createX86PatternInstructionSelector - This pass converts an LLVM function
+/// createX86ISelPattern - This pass converts an LLVM function
/// into a machine code representation using pattern matching and a machine
/// description file.
///
-FunctionPass *llvm::createX86PatternInstructionSelector(TargetMachine &TM) {
+FunctionPass *llvm::createX86ISelPattern(TargetMachine &TM) {
return new ISel(TM);
}
PM.add(createUnreachableBlockEliminationPass());
// Install an instruction selector.
- PM.add(createX86PatternInstructionSelector(*this));
+ PM.add(createX86ISelPattern(*this));
// Run optional SSA-based machine code optimizations next...
if (!NoSSAPeephole)
PM.add(createUnreachableBlockEliminationPass());
// Install an instruction selector.
- PM.add(createX86PatternInstructionSelector(TM));
+ PM.add(createX86ISelPattern(TM));
// Run optional SSA-based machine code optimizations next...
if (!NoSSAPeephole)
projects / oota-llvm.git / commitdiff