1 //===-- X86ISelPattern.cpp - A pattern matching inst selector for X86 -----===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for X86.
12 //===----------------------------------------------------------------------===//
15 #include "X86InstrBuilder.h"
16 #include "X86RegisterInfo.h"
17 #include "llvm/Constants.h" // FIXME: REMOVE
18 #include "llvm/Function.h"
19 #include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SelectionDAGISel.h"
24 #include "llvm/CodeGen/SSARegMap.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetLowering.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/ADT/Statistic.h"
33 //===----------------------------------------------------------------------===//
34 // X86TargetLowering - X86 Implementation of the TargetLowering interface
36 class X86TargetLowering : public TargetLowering {
37 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
38 int ReturnAddrIndex; // FrameIndex for return slot.
40 X86TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
41 // Set up the TargetLowering object.
43 // X86 is wierd, it always uses i8 for shift amounts and setcc results.
44 setShiftAmountType(MVT::i8);
45 setSetCCResultType(MVT::i8);
46 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
48 // Set up the register classes.
49 addRegisterClass(MVT::i8, X86::R8RegisterClass);
50 addRegisterClass(MVT::i16, X86::R16RegisterClass);
51 addRegisterClass(MVT::i32, X86::R32RegisterClass);
52 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
54 // FIXME: Eliminate these two classes when legalize can handle promotions
56 /**/ addRegisterClass(MVT::i1, X86::R8RegisterClass);
58 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
59 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
60 setOperationAction(ISD::ZERO_EXTEND_INREG, MVT::i16 , Expand);
61 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
62 setOperationAction(ISD::ZERO_EXTEND_INREG, MVT::i1 , Expand);
63 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
64 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
65 setOperationAction(ISD::SREM , MVT::f64 , Expand);
67 // These should be promoted to a larger select which is supported.
68 /**/ setOperationAction(ISD::SELECT , MVT::i1 , Promote);
69 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
71 computeRegisterProperties();
73 addLegalFPImmediate(+0.0); // FLD0
74 addLegalFPImmediate(+1.0); // FLD1
75 addLegalFPImmediate(-0.0); // FLD0/FCHS
76 addLegalFPImmediate(-1.0); // FLD1/FCHS
79 /// LowerArguments - This hook must be implemented to indicate how we should
80 /// lower the arguments for the specified function, into the specified DAG.
81 virtual std::vector<SDOperand>
82 LowerArguments(Function &F, SelectionDAG &DAG);
84 /// LowerCallTo - This hook lowers an abstract call to a function into an
86 virtual std::pair<SDOperand, SDOperand>
87 LowerCallTo(SDOperand Chain, const Type *RetTy, SDOperand Callee,
88 ArgListTy &Args, SelectionDAG &DAG);
90 virtual std::pair<SDOperand, SDOperand>
91 LowerVAStart(SDOperand Chain, SelectionDAG &DAG);
93 virtual std::pair<SDOperand,SDOperand>
94 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
95 const Type *ArgTy, SelectionDAG &DAG);
97 virtual std::pair<SDOperand, SDOperand>
98 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
104 std::vector<SDOperand>
105 X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
106 std::vector<SDOperand> ArgValues;
108 // Add DAG nodes to load the arguments... On entry to a function on the X86,
109 // the stack frame looks like this:
111 // [ESP] -- return address
112 // [ESP + 4] -- first argument (leftmost lexically)
113 // [ESP + 8] -- second argument, if first argument is four bytes in size
116 MachineFunction &MF = DAG.getMachineFunction();
117 MachineFrameInfo *MFI = MF.getFrameInfo();
119 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
120 for (Function::aiterator I = F.abegin(), E = F.aend(); I != E; ++I) {
121 MVT::ValueType ObjectVT = getValueType(I->getType());
122 unsigned ArgIncrement = 4;
125 default: assert(0 && "Unhandled argument type!");
127 case MVT::i8: ObjSize = 1; break;
128 case MVT::i16: ObjSize = 2; break;
129 case MVT::i32: ObjSize = 4; break;
130 case MVT::i64: ObjSize = ArgIncrement = 8; break;
131 case MVT::f32: ObjSize = 4; break;
132 case MVT::f64: ObjSize = ArgIncrement = 8; break;
134 // Create the frame index object for this incoming parameter...
135 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
137 // Create the SelectionDAG nodes corresponding to a load from this parameter
138 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
140 // Don't codegen dead arguments. FIXME: remove this check when we can nuke
144 ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN);
146 if (MVT::isInteger(ObjectVT))
147 ArgValue = DAG.getConstant(0, ObjectVT);
149 ArgValue = DAG.getConstantFP(0, ObjectVT);
151 ArgValues.push_back(ArgValue);
153 ArgOffset += ArgIncrement; // Move on to the next argument...
156 // If the function takes variable number of arguments, make a frame index for
157 // the start of the first vararg value... for expansion of llvm.va_start.
159 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
160 ReturnAddrIndex = 0; // No return address slot generated yet.
164 std::pair<SDOperand, SDOperand>
165 X86TargetLowering::LowerCallTo(SDOperand Chain,
166 const Type *RetTy, SDOperand Callee,
167 ArgListTy &Args, SelectionDAG &DAG) {
168 // Count how many bytes are to be pushed on the stack.
169 unsigned NumBytes = 0;
173 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
174 DAG.getConstant(0, getPointerTy()));
176 for (unsigned i = 0, e = Args.size(); i != e; ++i)
177 switch (getValueType(Args[i].second)) {
178 default: assert(0 && "Unknown value type!");
192 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
193 DAG.getConstant(NumBytes, getPointerTy()));
195 // Arguments go on the stack in reverse order, as specified by the ABI.
196 unsigned ArgOffset = 0;
197 SDOperand StackPtr = DAG.getCopyFromReg(X86::ESP, MVT::i32,
199 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
201 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
202 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
204 switch (getValueType(Args[i].second)) {
205 default: assert(0 && "Unexpected ValueType for argument!");
209 // Promote the integer to 32 bits. If the input type is signed use a
210 // sign extend, otherwise use a zero extend.
211 if (Args[i].second->isSigned())
212 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
214 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
219 // FIXME: Note that all of these stores are independent of each other.
220 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
221 Args[i].first, PtrOff);
226 // FIXME: Note that all of these stores are independent of each other.
227 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
228 Args[i].first, PtrOff);
235 std::vector<MVT::ValueType> RetVals;
236 MVT::ValueType RetTyVT = getValueType(RetTy);
237 if (RetTyVT != MVT::isVoid)
238 RetVals.push_back(RetTyVT);
239 RetVals.push_back(MVT::Other);
241 SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain, Callee), 0);
242 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
243 Chain = DAG.getNode(ISD::ADJCALLSTACKUP, MVT::Other, Chain,
244 DAG.getConstant(NumBytes, getPointerTy()));
245 return std::make_pair(TheCall, Chain);
248 std::pair<SDOperand, SDOperand>
249 X86TargetLowering::LowerVAStart(SDOperand Chain, SelectionDAG &DAG) {
250 // vastart just returns the address of the VarArgsFrameIndex slot.
251 return std::make_pair(DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32), Chain);
254 std::pair<SDOperand,SDOperand> X86TargetLowering::
255 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
256 const Type *ArgTy, SelectionDAG &DAG) {
257 MVT::ValueType ArgVT = getValueType(ArgTy);
260 Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), VAList);
263 if (ArgVT == MVT::i32)
266 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
267 "Other types should have been promoted for varargs!");
270 Result = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
271 DAG.getConstant(Amt, VAList.getValueType()));
273 return std::make_pair(Result, Chain);
277 std::pair<SDOperand, SDOperand> X86TargetLowering::
278 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
281 if (Depth) // Depths > 0 not supported yet!
282 Result = DAG.getConstant(0, getPointerTy());
284 if (ReturnAddrIndex == 0) {
285 // Set up a frame object for the return address.
286 MachineFunction &MF = DAG.getMachineFunction();
287 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
290 SDOperand RetAddrFI = DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
293 // Just load the return address
294 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI);
296 Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
297 DAG.getConstant(4, MVT::i32));
299 return std::make_pair(Result, Chain);
304 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
305 /// SDOperand's instead of register numbers for the leaves of the matched
307 struct X86ISelAddressMode {
313 struct { // This is really a union, discriminated by BaseType!
324 : BaseType(RegBase), Scale(1), IndexReg(), Disp(), GV(0) {
332 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
334 //===--------------------------------------------------------------------===//
335 /// ISel - X86 specific code to select X86 machine instructions for
336 /// SelectionDAG operations.
338 class ISel : public SelectionDAGISel {
339 /// ContainsFPCode - Every instruction we select that uses or defines a FP
340 /// register should set this to true.
343 /// X86Lowering - This object fully describes how to lower LLVM code to an
344 /// X86-specific SelectionDAG.
345 X86TargetLowering X86Lowering;
347 /// RegPressureMap - This keeps an approximate count of the number of
348 /// registers required to evaluate each node in the graph.
349 std::map<SDNode*, unsigned> RegPressureMap;
351 /// ExprMap - As shared expressions are codegen'd, we keep track of which
352 /// vreg the value is produced in, so we only emit one copy of each compiled
354 std::map<SDOperand, unsigned> ExprMap;
357 ISel(TargetMachine &TM) : SelectionDAGISel(X86Lowering), X86Lowering(TM) {
360 unsigned getRegPressure(SDOperand O) {
361 return RegPressureMap[O.Val];
363 unsigned ComputeRegPressure(SDOperand O);
365 /// InstructionSelectBasicBlock - This callback is invoked by
366 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
367 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
369 bool isFoldableLoad(SDOperand Op, SDOperand OtherOp);
370 void EmitFoldedLoad(SDOperand Op, X86AddressMode &AM);
371 bool TryToFoldLoadOpStore(SDNode *Node);
373 void EmitCMP(SDOperand LHS, SDOperand RHS, bool isOnlyUse);
374 bool EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain, SDOperand Cond);
375 void EmitSelectCC(SDOperand Cond, MVT::ValueType SVT,
376 unsigned RTrue, unsigned RFalse, unsigned RDest);
377 unsigned SelectExpr(SDOperand N);
379 X86AddressMode SelectAddrExprs(const X86ISelAddressMode &IAM);
380 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM);
381 void SelectAddress(SDOperand N, X86AddressMode &AM);
382 void Select(SDOperand N);
386 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
387 /// when it has created a SelectionDAG for us to codegen.
388 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
389 // While we're doing this, keep track of whether we see any FP code for
390 // FP_REG_KILL insertion.
391 ContainsFPCode = false;
393 // Scan the PHI nodes that already are inserted into this basic block. If any
394 // of them is a PHI of a floating point value, we need to insert an
396 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
397 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
399 assert(I->getOpcode() == X86::PHI &&
400 "Isn't just PHI nodes?");
401 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
402 X86::RFPRegisterClass) {
403 ContainsFPCode = true;
408 // Compute the RegPressureMap, which is an approximation for the number of
409 // registers required to compute each node.
410 ComputeRegPressure(DAG.getRoot());
412 // Codegen the basic block.
413 Select(DAG.getRoot());
415 // Finally, look at all of the successors of this block. If any contain a PHI
416 // node of FP type, we need to insert an FP_REG_KILL in this block.
417 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
418 E = BB->succ_end(); SI != E && !ContainsFPCode; ++SI)
419 for (MachineBasicBlock::iterator I = (*SI)->begin(), E = (*SI)->end();
420 I != E && I->getOpcode() == X86::PHI; ++I) {
421 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
422 X86::RFPRegisterClass) {
423 ContainsFPCode = true;
428 // Insert FP_REG_KILL instructions into basic blocks that need them. This
429 // only occurs due to the floating point stackifier not being aggressive
430 // enough to handle arbitrary global stackification.
432 // Currently we insert an FP_REG_KILL instruction into each block that uses or
433 // defines a floating point virtual register.
435 // When the global register allocators (like linear scan) finally update live
436 // variable analysis, we can keep floating point values in registers across
437 // basic blocks. This will be a huge win, but we are waiting on the global
438 // allocators before we can do this.
440 if (ContainsFPCode && BB->succ_size()) {
441 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
445 // Clear state used for selection.
447 RegPressureMap.clear();
451 // ComputeRegPressure - Compute the RegPressureMap, which is an approximation
452 // for the number of registers required to compute each node. This is basically
453 // computing a generalized form of the Sethi-Ullman number for each node.
454 unsigned ISel::ComputeRegPressure(SDOperand O) {
456 unsigned &Result = RegPressureMap[N];
457 if (Result) return Result;
459 // FIXME: Should operations like CALL (which clobber lots o regs) have a
460 // higher fixed cost??
462 if (N->getNumOperands() == 0) {
465 unsigned MaxRegUse = 0;
466 unsigned NumExtraMaxRegUsers = 0;
467 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
469 if (N->getOperand(i).getOpcode() == ISD::Constant)
472 Regs = ComputeRegPressure(N->getOperand(i));
473 if (Regs > MaxRegUse) {
475 NumExtraMaxRegUsers = 0;
476 } else if (Regs == MaxRegUse &&
477 N->getOperand(i).getValueType() != MVT::Other) {
478 ++NumExtraMaxRegUsers;
482 if (O.getOpcode() != ISD::TokenFactor)
483 Result = MaxRegUse+NumExtraMaxRegUsers;
485 Result = MaxRegUse == 1 ? 0 : MaxRegUse-1;
488 //std::cerr << " WEIGHT: " << Result << " "; N->dump(); std::cerr << "\n";
492 X86AddressMode ISel::SelectAddrExprs(const X86ISelAddressMode &IAM) {
493 X86AddressMode Result;
495 // If we need to emit two register operands, emit the one with the highest
496 // register pressure first.
497 if (IAM.BaseType == X86ISelAddressMode::RegBase &&
498 IAM.Base.Reg.Val && IAM.IndexReg.Val) {
499 if (getRegPressure(IAM.Base.Reg) > getRegPressure(IAM.IndexReg)) {
500 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
501 Result.IndexReg = SelectExpr(IAM.IndexReg);
503 Result.IndexReg = SelectExpr(IAM.IndexReg);
504 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
506 } else if (IAM.BaseType == X86ISelAddressMode::RegBase && IAM.Base.Reg.Val) {
507 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
508 } else if (IAM.IndexReg.Val) {
509 Result.IndexReg = SelectExpr(IAM.IndexReg);
512 switch (IAM.BaseType) {
513 case X86ISelAddressMode::RegBase:
514 Result.BaseType = X86AddressMode::RegBase;
516 case X86ISelAddressMode::FrameIndexBase:
517 Result.BaseType = X86AddressMode::FrameIndexBase;
518 Result.Base.FrameIndex = IAM.Base.FrameIndex;
521 assert(0 && "Unknown base type!");
524 Result.Scale = IAM.Scale;
525 Result.Disp = IAM.Disp;
530 /// SelectAddress - Pattern match the maximal addressing mode for this node and
531 /// emit all of the leaf registers.
532 void ISel::SelectAddress(SDOperand N, X86AddressMode &AM) {
533 X86ISelAddressMode IAM;
534 MatchAddress(N, IAM);
535 AM = SelectAddrExprs(IAM);
538 /// MatchAddress - Add the specified node to the specified addressing mode,
539 /// returning true if it cannot be done. This just pattern matches for the
540 /// addressing mode, it does not cause any code to be emitted. For that, use
542 bool ISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM) {
543 switch (N.getOpcode()) {
545 case ISD::FrameIndex:
546 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
547 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
548 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
552 case ISD::GlobalAddress:
554 AM.GV = cast<GlobalAddressSDNode>(N)->getGlobal();
559 AM.Disp += cast<ConstantSDNode>(N)->getValue();
562 // We might have folded the load into this shift, so don't regen the value
564 if (ExprMap.count(N)) break;
566 if (AM.IndexReg.Val == 0 && AM.Scale == 1)
567 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
568 unsigned Val = CN->getValue();
569 if (Val == 1 || Val == 2 || Val == 3) {
571 SDOperand ShVal = N.Val->getOperand(0);
573 // Okay, we know that we have a scale by now. However, if the scaled
574 // value is an add of something and a constant, we can fold the
575 // constant into the disp field here.
576 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
577 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
578 AM.IndexReg = ShVal.Val->getOperand(0);
579 ConstantSDNode *AddVal =
580 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
581 AM.Disp += AddVal->getValue() << Val;
590 // We might have folded the load into this mul, so don't regen the value if
592 if (ExprMap.count(N)) break;
594 // X*[3,5,9] -> X+X*[2,4,8]
595 if (AM.IndexReg.Val == 0 && AM.BaseType == X86ISelAddressMode::RegBase &&
596 AM.Base.Reg.Val == 0)
597 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
598 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
599 AM.Scale = unsigned(CN->getValue())-1;
601 SDOperand MulVal = N.Val->getOperand(0);
604 // Okay, we know that we have a scale by now. However, if the scaled
605 // value is an add of something and a constant, we can fold the
606 // constant into the disp field here.
607 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
608 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
609 Reg = MulVal.Val->getOperand(0);
610 ConstantSDNode *AddVal =
611 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
612 AM.Disp += AddVal->getValue() * CN->getValue();
614 Reg = N.Val->getOperand(0);
617 AM.IndexReg = AM.Base.Reg = Reg;
623 // We might have folded the load into this mul, so don't regen the value if
625 if (ExprMap.count(N)) break;
627 X86ISelAddressMode Backup = AM;
628 if (!MatchAddress(N.Val->getOperand(0), AM) &&
629 !MatchAddress(N.Val->getOperand(1), AM))
632 if (!MatchAddress(N.Val->getOperand(1), AM) &&
633 !MatchAddress(N.Val->getOperand(0), AM))
640 // Is the base register already occupied?
641 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
642 // If so, check to see if the scale index register is set.
643 if (AM.IndexReg.Val == 0) {
649 // Otherwise, we cannot select it.
653 // Default, generate it as a register.
654 AM.BaseType = X86ISelAddressMode::RegBase;
659 /// Emit2SetCCsAndLogical - Emit the following sequence of instructions,
660 /// assuming that the temporary registers are in the 8-bit register class.
664 /// DestReg = logicalop Tmp1, Tmp2
666 static void Emit2SetCCsAndLogical(MachineBasicBlock *BB, unsigned SetCC1,
667 unsigned SetCC2, unsigned LogicalOp,
669 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
670 unsigned Tmp1 = RegMap->createVirtualRegister(X86::R8RegisterClass);
671 unsigned Tmp2 = RegMap->createVirtualRegister(X86::R8RegisterClass);
672 BuildMI(BB, SetCC1, 0, Tmp1);
673 BuildMI(BB, SetCC2, 0, Tmp2);
674 BuildMI(BB, LogicalOp, 2, DestReg).addReg(Tmp1).addReg(Tmp2);
677 /// EmitSetCC - Emit the code to set the specified 8-bit register to 1 if the
678 /// condition codes match the specified SetCCOpcode. Note that some conditions
679 /// require multiple instructions to generate the correct value.
680 static void EmitSetCC(MachineBasicBlock *BB, unsigned DestReg,
681 ISD::CondCode SetCCOpcode, bool isFP) {
684 switch (SetCCOpcode) {
685 default: assert(0 && "Illegal integer SetCC!");
686 case ISD::SETEQ: Opc = X86::SETEr; break;
687 case ISD::SETGT: Opc = X86::SETGr; break;
688 case ISD::SETGE: Opc = X86::SETGEr; break;
689 case ISD::SETLT: Opc = X86::SETLr; break;
690 case ISD::SETLE: Opc = X86::SETLEr; break;
691 case ISD::SETNE: Opc = X86::SETNEr; break;
692 case ISD::SETULT: Opc = X86::SETBr; break;
693 case ISD::SETUGT: Opc = X86::SETAr; break;
694 case ISD::SETULE: Opc = X86::SETBEr; break;
695 case ISD::SETUGE: Opc = X86::SETAEr; break;
698 // On a floating point condition, the flags are set as follows:
702 // 1 | 0 | 0 | X == Y
703 // 1 | 1 | 1 | unordered
705 switch (SetCCOpcode) {
706 default: assert(0 && "Invalid FP setcc!");
709 Opc = X86::SETEr; // True if ZF = 1
713 Opc = X86::SETAr; // True if CF = 0 and ZF = 0
717 Opc = X86::SETAEr; // True if CF = 0
721 Opc = X86::SETBr; // True if CF = 1
725 Opc = X86::SETBEr; // True if CF = 1 or ZF = 1
729 Opc = X86::SETNEr; // True if ZF = 0
732 Opc = X86::SETPr; // True if PF = 1
735 Opc = X86::SETNPr; // True if PF = 0
737 case ISD::SETOEQ: // !PF & ZF
738 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETEr, X86::AND8rr, DestReg);
740 case ISD::SETOLT: // !PF & CF
741 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBr, X86::AND8rr, DestReg);
743 case ISD::SETOLE: // !PF & (CF || ZF)
744 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBEr, X86::AND8rr, DestReg);
746 case ISD::SETUGT: // PF | (!ZF & !CF)
747 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAr, X86::OR8rr, DestReg);
749 case ISD::SETUGE: // PF | !CF
750 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAEr, X86::OR8rr, DestReg);
752 case ISD::SETUNE: // PF | !ZF
753 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETNEr, X86::OR8rr, DestReg);
757 BuildMI(BB, Opc, 0, DestReg);
761 /// EmitBranchCC - Emit code into BB that arranges for control to transfer to
762 /// the Dest block if the Cond condition is true. If we cannot fold this
763 /// condition into the branch, return true.
765 bool ISel::EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain,
767 // FIXME: Evaluate whether it would be good to emit code like (X < Y) | (A >
768 // B) using two conditional branches instead of one condbr, two setcc's, and
770 if ((Cond.getOpcode() == ISD::OR ||
771 Cond.getOpcode() == ISD::AND) && Cond.Val->hasOneUse()) {
772 // And and or set the flags for us, so there is no need to emit a TST of the
773 // result. It is only safe to do this if there is only a single use of the
774 // AND/OR though, otherwise we don't know it will be emitted here.
777 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
781 // Codegen br not C -> JE.
782 if (Cond.getOpcode() == ISD::XOR)
783 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(Cond.Val->getOperand(1)))
784 if (NC->isAllOnesValue()) {
786 if (getRegPressure(Chain) > getRegPressure(Cond)) {
788 CondR = SelectExpr(Cond.Val->getOperand(0));
790 CondR = SelectExpr(Cond.Val->getOperand(0));
793 BuildMI(BB, X86::TEST8rr, 2).addReg(CondR).addReg(CondR);
794 BuildMI(BB, X86::JE, 1).addMBB(Dest);
798 SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Cond);
800 return true; // Can only handle simple setcc's so far.
804 // Handle integer conditions first.
805 if (MVT::isInteger(SetCC->getOperand(0).getValueType())) {
806 switch (SetCC->getCondition()) {
807 default: assert(0 && "Illegal integer SetCC!");
808 case ISD::SETEQ: Opc = X86::JE; break;
809 case ISD::SETGT: Opc = X86::JG; break;
810 case ISD::SETGE: Opc = X86::JGE; break;
811 case ISD::SETLT: Opc = X86::JL; break;
812 case ISD::SETLE: Opc = X86::JLE; break;
813 case ISD::SETNE: Opc = X86::JNE; break;
814 case ISD::SETULT: Opc = X86::JB; break;
815 case ISD::SETUGT: Opc = X86::JA; break;
816 case ISD::SETULE: Opc = X86::JBE; break;
817 case ISD::SETUGE: Opc = X86::JAE; break;
820 EmitCMP(SetCC->getOperand(0), SetCC->getOperand(1), SetCC->hasOneUse());
821 BuildMI(BB, Opc, 1).addMBB(Dest);
825 unsigned Opc2 = 0; // Second branch if needed.
827 // On a floating point condition, the flags are set as follows:
831 // 1 | 0 | 0 | X == Y
832 // 1 | 1 | 1 | unordered
834 switch (SetCC->getCondition()) {
835 default: assert(0 && "Invalid FP setcc!");
837 case ISD::SETEQ: Opc = X86::JE; break; // True if ZF = 1
839 case ISD::SETGT: Opc = X86::JA; break; // True if CF = 0 and ZF = 0
841 case ISD::SETGE: Opc = X86::JAE; break; // True if CF = 0
843 case ISD::SETLT: Opc = X86::JB; break; // True if CF = 1
845 case ISD::SETLE: Opc = X86::JBE; break; // True if CF = 1 or ZF = 1
847 case ISD::SETNE: Opc = X86::JNE; break; // True if ZF = 0
848 case ISD::SETUO: Opc = X86::JP; break; // True if PF = 1
849 case ISD::SETO: Opc = X86::JNP; break; // True if PF = 0
850 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
851 Opc = X86::JA; // ZF = 0 & CF = 0
852 Opc2 = X86::JP; // PF = 1
854 case ISD::SETUGE: // PF = 1 | CF = 0
855 Opc = X86::JAE; // CF = 0
856 Opc2 = X86::JP; // PF = 1
858 case ISD::SETUNE: // PF = 1 | ZF = 0
859 Opc = X86::JNE; // ZF = 0
860 Opc2 = X86::JP; // PF = 1
862 case ISD::SETOEQ: // PF = 0 & ZF = 1
865 return true; // FIXME: Emit more efficient code for this branch.
866 case ISD::SETOLT: // PF = 0 & CF = 1
869 return true; // FIXME: Emit more efficient code for this branch.
870 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
873 return true; // FIXME: Emit more efficient code for this branch.
877 EmitCMP(SetCC->getOperand(0), SetCC->getOperand(1), SetCC->hasOneUse());
878 BuildMI(BB, Opc, 1).addMBB(Dest);
880 BuildMI(BB, Opc2, 1).addMBB(Dest);
884 /// EmitSelectCC - Emit code into BB that performs a select operation between
885 /// the two registers RTrue and RFalse, generating a result into RDest. Return
886 /// true if the fold cannot be performed.
888 void ISel::EmitSelectCC(SDOperand Cond, MVT::ValueType SVT,
889 unsigned RTrue, unsigned RFalse, unsigned RDest) {
891 EQ, NE, LT, LE, GT, GE, B, BE, A, AE, P, NP,
893 } CondCode = NOT_SET;
895 static const unsigned CMOVTAB16[] = {
896 X86::CMOVE16rr, X86::CMOVNE16rr, X86::CMOVL16rr, X86::CMOVLE16rr,
897 X86::CMOVG16rr, X86::CMOVGE16rr, X86::CMOVB16rr, X86::CMOVBE16rr,
898 X86::CMOVA16rr, X86::CMOVAE16rr, X86::CMOVP16rr, X86::CMOVNP16rr,
900 static const unsigned CMOVTAB32[] = {
901 X86::CMOVE32rr, X86::CMOVNE32rr, X86::CMOVL32rr, X86::CMOVLE32rr,
902 X86::CMOVG32rr, X86::CMOVGE32rr, X86::CMOVB32rr, X86::CMOVBE32rr,
903 X86::CMOVA32rr, X86::CMOVAE32rr, X86::CMOVP32rr, X86::CMOVNP32rr,
905 static const unsigned CMOVTABFP[] = {
906 X86::FCMOVE , X86::FCMOVNE, /*missing*/0, /*missing*/0,
907 /*missing*/0, /*missing*/0, X86::FCMOVB , X86::FCMOVBE,
908 X86::FCMOVA , X86::FCMOVAE, X86::FCMOVP , X86::FCMOVNP
911 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Cond)) {
912 if (MVT::isInteger(SetCC->getOperand(0).getValueType())) {
913 switch (SetCC->getCondition()) {
914 default: assert(0 && "Unknown integer comparison!");
915 case ISD::SETEQ: CondCode = EQ; break;
916 case ISD::SETGT: CondCode = GT; break;
917 case ISD::SETGE: CondCode = GE; break;
918 case ISD::SETLT: CondCode = LT; break;
919 case ISD::SETLE: CondCode = LE; break;
920 case ISD::SETNE: CondCode = NE; break;
921 case ISD::SETULT: CondCode = B; break;
922 case ISD::SETUGT: CondCode = A; break;
923 case ISD::SETULE: CondCode = BE; break;
924 case ISD::SETUGE: CondCode = AE; break;
927 // On a floating point condition, the flags are set as follows:
931 // 1 | 0 | 0 | X == Y
932 // 1 | 1 | 1 | unordered
934 switch (SetCC->getCondition()) {
935 default: assert(0 && "Unknown FP comparison!");
937 case ISD::SETEQ: CondCode = EQ; break; // True if ZF = 1
939 case ISD::SETGT: CondCode = A; break; // True if CF = 0 and ZF = 0
941 case ISD::SETGE: CondCode = AE; break; // True if CF = 0
943 case ISD::SETLT: CondCode = B; break; // True if CF = 1
945 case ISD::SETLE: CondCode = BE; break; // True if CF = 1 or ZF = 1
947 case ISD::SETNE: CondCode = NE; break; // True if ZF = 0
948 case ISD::SETUO: CondCode = P; break; // True if PF = 1
949 case ISD::SETO: CondCode = NP; break; // True if PF = 0
950 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
951 case ISD::SETUGE: // PF = 1 | CF = 0
952 case ISD::SETUNE: // PF = 1 | ZF = 0
953 case ISD::SETOEQ: // PF = 0 & ZF = 1
954 case ISD::SETOLT: // PF = 0 & CF = 1
955 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
956 // We cannot emit this comparison as a single cmov.
963 if (CondCode != NOT_SET) {
965 default: assert(0 && "Cannot select this type!");
966 case MVT::i16: Opc = CMOVTAB16[CondCode]; break;
967 case MVT::i32: Opc = CMOVTAB32[CondCode]; break;
968 case MVT::f64: Opc = CMOVTABFP[CondCode]; break;
972 // Finally, if we weren't able to fold this, just emit the condition and test
974 if (CondCode == NOT_SET || Opc == 0) {
975 // Get the condition into the zero flag.
976 unsigned CondReg = SelectExpr(Cond);
977 BuildMI(BB, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
980 default: assert(0 && "Cannot select this type!");
981 case MVT::i16: Opc = X86::CMOVE16rr; break;
982 case MVT::i32: Opc = X86::CMOVE32rr; break;
983 case MVT::f64: Opc = X86::FCMOVE; break;
986 // FIXME: CMP R, 0 -> TEST R, R
987 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.Val->hasOneUse());
988 std::swap(RTrue, RFalse);
990 BuildMI(BB, Opc, 2, RDest).addReg(RTrue).addReg(RFalse);
993 void ISel::EmitCMP(SDOperand LHS, SDOperand RHS, bool HasOneUse) {
995 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(RHS)) {
997 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
998 switch (RHS.getValueType()) {
1001 case MVT::i8: Opc = X86::CMP8mi; break;
1002 case MVT::i16: Opc = X86::CMP16mi; break;
1003 case MVT::i32: Opc = X86::CMP32mi; break;
1007 EmitFoldedLoad(LHS, AM);
1008 addFullAddress(BuildMI(BB, Opc, 5), AM).addImm(CN->getValue());
1013 switch (RHS.getValueType()) {
1016 case MVT::i8: Opc = X86::CMP8ri; break;
1017 case MVT::i16: Opc = X86::CMP16ri; break;
1018 case MVT::i32: Opc = X86::CMP32ri; break;
1021 unsigned Tmp1 = SelectExpr(LHS);
1022 BuildMI(BB, Opc, 2).addReg(Tmp1).addImm(CN->getValue());
1025 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(RHS)) {
1026 if (CN->isExactlyValue(+0.0) ||
1027 CN->isExactlyValue(-0.0)) {
1028 unsigned Reg = SelectExpr(LHS);
1029 BuildMI(BB, X86::FTST, 1).addReg(Reg);
1030 BuildMI(BB, X86::FNSTSW8r, 0);
1031 BuildMI(BB, X86::SAHF, 1);
1036 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
1037 switch (RHS.getValueType()) {
1040 case MVT::i8: Opc = X86::CMP8mr; break;
1041 case MVT::i16: Opc = X86::CMP16mr; break;
1042 case MVT::i32: Opc = X86::CMP32mr; break;
1046 EmitFoldedLoad(LHS, AM);
1047 unsigned Reg = SelectExpr(RHS);
1048 addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(Reg);
1053 switch (LHS.getValueType()) {
1054 default: assert(0 && "Cannot compare this value!");
1056 case MVT::i8: Opc = X86::CMP8rr; break;
1057 case MVT::i16: Opc = X86::CMP16rr; break;
1058 case MVT::i32: Opc = X86::CMP32rr; break;
1059 case MVT::f64: Opc = X86::FUCOMIr; break;
1061 unsigned Tmp1, Tmp2;
1062 if (getRegPressure(LHS) > getRegPressure(RHS)) {
1063 Tmp1 = SelectExpr(LHS);
1064 Tmp2 = SelectExpr(RHS);
1066 Tmp2 = SelectExpr(RHS);
1067 Tmp1 = SelectExpr(LHS);
1069 BuildMI(BB, Opc, 2).addReg(Tmp1).addReg(Tmp2);
1072 /// NodeTransitivelyUsesValue - Return true if N or any of its uses uses Op.
1073 /// The DAG cannot have cycles in it, by definition, so the visited set is not
1074 /// needed to prevent infinite loops. The DAG CAN, however, have unbounded
1075 /// reuse, so it prevents exponential cases.
1077 static bool NodeTransitivelyUsesValue(SDOperand N, SDOperand Op,
1078 std::set<SDNode*> &Visited) {
1079 if (N == Op) return true; // Found it.
1080 SDNode *Node = N.Val;
1081 if (Node->getNumOperands() == 0) return false; // Leaf?
1082 if (!Visited.insert(Node).second) return false; // Already visited?
1084 // Recurse for the first N-1 operands.
1085 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
1086 if (NodeTransitivelyUsesValue(Node->getOperand(i), Op, Visited))
1089 // Tail recurse for the last operand.
1090 return NodeTransitivelyUsesValue(Node->getOperand(0), Op, Visited);
1093 /// isFoldableLoad - Return true if this is a load instruction that can safely
1094 /// be folded into an operation that uses it.
1095 bool ISel::isFoldableLoad(SDOperand Op, SDOperand OtherOp) {
1096 if (Op.getOpcode() != ISD::LOAD ||
1097 // FIXME: currently can't fold constant pool indexes.
1098 isa<ConstantPoolSDNode>(Op.getOperand(1)))
1101 // If this load has already been emitted, we clearly can't fold it.
1102 assert(Op.ResNo == 0 && "Not a use of the value of the load?");
1103 if (ExprMap.count(Op.getValue(1))) return false;
1104 assert(!ExprMap.count(Op.getValue(0)) && "Value in map but not token chain?");
1105 assert(!ExprMap.count(Op.getValue(1))&&"Token lowered but value not in map?");
1107 // If there is not just one use of its value, we cannot fold.
1108 if (!Op.Val->hasNUsesOfValue(1, 0)) return false;
1110 // Finally, we cannot fold the load into the operation if this would induce a
1111 // cycle into the resultant dag. To check for this, see if OtherOp (the other
1112 // operand of the operation we are folding the load into) can possible use the
1113 // chain node defined by the load.
1114 if (OtherOp.Val && !Op.Val->hasNUsesOfValue(0, 1)) { // Has uses of chain?
1115 std::set<SDNode*> Visited;
1116 if (NodeTransitivelyUsesValue(OtherOp, Op.getValue(1), Visited))
1123 /// EmitFoldedLoad - Ensure that the arguments of the load are code generated,
1124 /// and compute the address being loaded into AM.
1125 void ISel::EmitFoldedLoad(SDOperand Op, X86AddressMode &AM) {
1126 SDOperand Chain = Op.getOperand(0);
1127 SDOperand Address = Op.getOperand(1);
1129 if (getRegPressure(Chain) > getRegPressure(Address)) {
1131 SelectAddress(Address, AM);
1133 SelectAddress(Address, AM);
1137 // The chain for this load is now lowered.
1138 assert(ExprMap.count(SDOperand(Op.Val, 1)) == 0 &&
1139 "Load emitted more than once?");
1140 if (!ExprMap.insert(std::make_pair(Op.getValue(1), 1)).second)
1141 assert(0 && "Load emitted more than once!");
1144 unsigned ISel::SelectExpr(SDOperand N) {
1146 unsigned Tmp1, Tmp2, Tmp3;
1148 SDNode *Node = N.Val;
1151 if (Node->getOpcode() == ISD::CopyFromReg) {
1152 // FIXME: Handle copy from physregs!
1154 // Just use the specified register as our input.
1155 return dyn_cast<RegSDNode>(Node)->getReg();
1158 unsigned &Reg = ExprMap[N];
1159 if (Reg) return Reg;
1161 if (N.getOpcode() != ISD::CALL)
1162 Reg = Result = (N.getValueType() != MVT::Other) ?
1163 MakeReg(N.getValueType()) : 1;
1165 // If this is a call instruction, make sure to prepare ALL of the result
1166 // values as well as the chain.
1167 if (Node->getNumValues() == 1)
1168 Reg = Result = 1; // Void call, just a chain.
1170 Result = MakeReg(Node->getValueType(0));
1171 ExprMap[N.getValue(0)] = Result;
1172 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
1173 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1174 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
1178 switch (N.getOpcode()) {
1181 assert(0 && "Node not handled!\n");
1182 case ISD::FrameIndex:
1183 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1184 addFrameReference(BuildMI(BB, X86::LEA32r, 4, Result), (int)Tmp1);
1186 case ISD::ConstantPool:
1187 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
1188 addConstantPoolReference(BuildMI(BB, X86::LEA32r, 4, Result), Tmp1);
1190 case ISD::ConstantFP:
1191 ContainsFPCode = true;
1192 Tmp1 = Result; // Intermediate Register
1193 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
1194 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
1195 Tmp1 = MakeReg(MVT::f64);
1197 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
1198 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
1199 BuildMI(BB, X86::FLD0, 0, Tmp1);
1200 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
1201 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
1202 BuildMI(BB, X86::FLD1, 0, Tmp1);
1204 assert(0 && "Unexpected constant!");
1206 BuildMI(BB, X86::FCHS, 1, Result).addReg(Tmp1);
1209 switch (N.getValueType()) {
1210 default: assert(0 && "Cannot use constants of this type!");
1212 case MVT::i8: Opc = X86::MOV8ri; break;
1213 case MVT::i16: Opc = X86::MOV16ri; break;
1214 case MVT::i32: Opc = X86::MOV32ri; break;
1216 BuildMI(BB, Opc, 1,Result).addImm(cast<ConstantSDNode>(N)->getValue());
1218 case ISD::GlobalAddress: {
1219 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1220 BuildMI(BB, X86::MOV32ri, 1, Result).addGlobalAddress(GV);
1223 case ISD::ExternalSymbol: {
1224 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1225 BuildMI(BB, X86::MOV32ri, 1, Result).addExternalSymbol(Sym);
1228 case ISD::ZERO_EXTEND: {
1229 int DestIs16 = N.getValueType() == MVT::i16;
1230 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
1232 // FIXME: This hack is here for zero extension casts from bool to i8. This
1233 // would not be needed if bools were promoted by Legalize.
1234 if (N.getValueType() == MVT::i8) {
1235 Tmp1 = SelectExpr(N.getOperand(0));
1236 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(Tmp1);
1240 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
1241 static const unsigned Opc[3] = {
1242 X86::MOVZX32rm8, X86::MOVZX32rm16, X86::MOVZX16rm8
1246 EmitFoldedLoad(N.getOperand(0), AM);
1247 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
1252 static const unsigned Opc[3] = {
1253 X86::MOVZX32rr8, X86::MOVZX32rr16, X86::MOVZX16rr8
1255 Tmp1 = SelectExpr(N.getOperand(0));
1256 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
1259 case ISD::SIGN_EXTEND: {
1260 int DestIs16 = N.getValueType() == MVT::i16;
1261 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
1263 // FIXME: Legalize should promote bools to i8!
1264 assert(N.getOperand(0).getValueType() != MVT::i1 &&
1265 "Sign extend from bool not implemented!");
1267 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
1268 static const unsigned Opc[3] = {
1269 X86::MOVSX32rm8, X86::MOVSX32rm16, X86::MOVSX16rm8
1273 EmitFoldedLoad(N.getOperand(0), AM);
1274 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
1278 static const unsigned Opc[3] = {
1279 X86::MOVSX32rr8, X86::MOVSX32rr16, X86::MOVSX16rr8
1281 Tmp1 = SelectExpr(N.getOperand(0));
1282 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
1286 // Fold TRUNCATE (LOAD P) into a smaller load from P.
1287 // FIXME: This should be performed by the DAGCombiner.
1288 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
1289 switch (N.getValueType()) {
1290 default: assert(0 && "Unknown truncate!");
1292 case MVT::i8: Opc = X86::MOV8rm; break;
1293 case MVT::i16: Opc = X86::MOV16rm; break;
1296 EmitFoldedLoad(N.getOperand(0), AM);
1297 addFullAddress(BuildMI(BB, Opc, 4, Result), AM);
1301 // Handle cast of LARGER int to SMALLER int using a move to EAX followed by
1302 // a move out of AX or AL.
1303 switch (N.getOperand(0).getValueType()) {
1304 default: assert(0 && "Unknown truncate!");
1305 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
1306 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
1307 case MVT::i32: Tmp2 = X86::EAX; Opc = X86::MOV32rr; break;
1309 Tmp1 = SelectExpr(N.getOperand(0));
1310 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
1312 switch (N.getValueType()) {
1313 default: assert(0 && "Unknown truncate!");
1315 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
1316 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
1318 BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
1321 case ISD::SINT_TO_FP:
1322 case ISD::UINT_TO_FP: {
1323 // FIXME: Most of this grunt work should be done by legalize!
1324 ContainsFPCode = true;
1326 // Promote the integer to a type supported by FLD. We do this because there
1327 // are no unsigned FLD instructions, so we must promote an unsigned value to
1328 // a larger signed value, then use FLD on the larger value.
1330 MVT::ValueType PromoteType = MVT::Other;
1331 MVT::ValueType SrcTy = N.getOperand(0).getValueType();
1332 unsigned PromoteOpcode = 0;
1333 unsigned RealDestReg = Result;
1337 // We don't have the facilities for directly loading byte sized data from
1338 // memory (even signed). Promote it to 16 bits.
1339 PromoteType = MVT::i16;
1340 PromoteOpcode = Node->getOpcode() == ISD::SINT_TO_FP ?
1341 X86::MOVSX16rr8 : X86::MOVZX16rr8;
1344 if (Node->getOpcode() == ISD::UINT_TO_FP) {
1345 PromoteType = MVT::i32;
1346 PromoteOpcode = X86::MOVZX32rr16;
1350 // Don't fild into the real destination.
1351 if (Node->getOpcode() == ISD::UINT_TO_FP)
1352 Result = MakeReg(Node->getValueType(0));
1356 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
1358 if (PromoteType != MVT::Other) {
1359 Tmp2 = MakeReg(PromoteType);
1360 BuildMI(BB, PromoteOpcode, 1, Tmp2).addReg(Tmp1);
1361 SrcTy = PromoteType;
1365 // Spill the integer to memory and reload it from there.
1366 unsigned Size = MVT::getSizeInBits(SrcTy)/8;
1367 MachineFunction *F = BB->getParent();
1368 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
1372 assert(0 && "Cast ulong to FP not implemented yet!");
1373 // FIXME: this won't work for cast [u]long to FP
1374 addFrameReference(BuildMI(BB, X86::MOV32mr, 5),
1375 FrameIdx).addReg(Tmp1);
1376 addFrameReference(BuildMI(BB, X86::MOV32mr, 5),
1377 FrameIdx, 4).addReg(Tmp1+1);
1378 addFrameReference(BuildMI(BB, X86::FILD64m, 5, Result), FrameIdx);
1381 addFrameReference(BuildMI(BB, X86::MOV32mr, 5),
1382 FrameIdx).addReg(Tmp1);
1383 addFrameReference(BuildMI(BB, X86::FILD32m, 5, Result), FrameIdx);
1386 addFrameReference(BuildMI(BB, X86::MOV16mr, 5),
1387 FrameIdx).addReg(Tmp1);
1388 addFrameReference(BuildMI(BB, X86::FILD16m, 5, Result), FrameIdx);
1390 default: break; // No promotion required.
1393 if (Node->getOpcode() == ISD::UINT_TO_FP && Result != RealDestReg) {
1394 // If this is a cast from uint -> double, we need to be careful when if
1395 // the "sign" bit is set. If so, we don't want to make a negative number,
1396 // we want to make a positive number. Emit code to add an offset if the
1399 // Compute whether the sign bit is set by shifting the reg right 31 bits.
1400 unsigned IsNeg = MakeReg(MVT::i32);
1401 BuildMI(BB, X86::SHR32ri, 2, IsNeg).addReg(Tmp1).addImm(31);
1403 // Create a CP value that has the offset in one word and 0 in the other.
1404 static ConstantInt *TheOffset = ConstantUInt::get(Type::ULongTy,
1405 0x4f80000000000000ULL);
1406 unsigned CPI = F->getConstantPool()->getConstantPoolIndex(TheOffset);
1407 BuildMI(BB, X86::FADD32m, 5, RealDestReg).addReg(Result)
1408 .addConstantPoolIndex(CPI).addZImm(4).addReg(IsNeg).addSImm(0);
1410 } else if (Node->getOpcode() == ISD::UINT_TO_FP && SrcTy == MVT::i64) {
1411 // We need special handling for unsigned 64-bit integer sources. If the
1412 // input number has the "sign bit" set, then we loaded it incorrectly as a
1413 // negative 64-bit number. In this case, add an offset value.
1415 // Emit a test instruction to see if the dynamic input value was signed.
1416 BuildMI(BB, X86::TEST32rr, 2).addReg(Tmp1+1).addReg(Tmp1+1);
1418 // If the sign bit is set, get a pointer to an offset, otherwise get a
1419 // pointer to a zero.
1420 MachineConstantPool *CP = F->getConstantPool();
1421 unsigned Zero = MakeReg(MVT::i32);
1422 Constant *Null = Constant::getNullValue(Type::UIntTy);
1423 addConstantPoolReference(BuildMI(BB, X86::LEA32r, 5, Zero),
1424 CP->getConstantPoolIndex(Null));
1425 unsigned Offset = MakeReg(MVT::i32);
1426 Constant *OffsetCst = ConstantUInt::get(Type::UIntTy, 0x5f800000);
1428 addConstantPoolReference(BuildMI(BB, X86::LEA32r, 5, Offset),
1429 CP->getConstantPoolIndex(OffsetCst));
1430 unsigned Addr = MakeReg(MVT::i32);
1431 BuildMI(BB, X86::CMOVS32rr, 2, Addr).addReg(Zero).addReg(Offset);
1433 // Load the constant for an add. FIXME: this could make an 'fadd' that
1434 // reads directly from memory, but we don't support these yet.
1435 unsigned ConstReg = MakeReg(MVT::f64);
1436 addDirectMem(BuildMI(BB, X86::FLD32m, 4, ConstReg), Addr);
1438 BuildMI(BB, X86::FpADD, 2, RealDestReg).addReg(ConstReg).addReg(Result);
1442 case ISD::FP_TO_SINT:
1443 case ISD::FP_TO_UINT: {
1444 // FIXME: Most of this grunt work should be done by legalize!
1445 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
1447 // Change the floating point control register to use "round towards zero"
1448 // mode when truncating to an integer value.
1450 MachineFunction *F = BB->getParent();
1451 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
1452 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
1454 // Load the old value of the high byte of the control word...
1455 unsigned HighPartOfCW = MakeReg(MVT::i8);
1456 addFrameReference(BuildMI(BB, X86::MOV8rm, 4, HighPartOfCW),
1459 // Set the high part to be round to zero...
1460 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
1461 CWFrameIdx, 1).addImm(12);
1463 // Reload the modified control word now...
1464 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1466 // Restore the memory image of control word to original value
1467 addFrameReference(BuildMI(BB, X86::MOV8mr, 5),
1468 CWFrameIdx, 1).addReg(HighPartOfCW);
1470 // We don't have the facilities for directly storing byte sized data to
1471 // memory. Promote it to 16 bits. We also must promote unsigned values to
1472 // larger classes because we only have signed FP stores.
1473 MVT::ValueType StoreClass = Node->getValueType(0);
1474 if (StoreClass == MVT::i8 || Node->getOpcode() == ISD::FP_TO_UINT)
1475 switch (StoreClass) {
1476 case MVT::i8: StoreClass = MVT::i16; break;
1477 case MVT::i16: StoreClass = MVT::i32; break;
1478 case MVT::i32: StoreClass = MVT::i64; break;
1479 // The following treatment of cLong may not be perfectly right,
1480 // but it survives chains of casts of the form
1481 // double->ulong->double.
1482 case MVT::i64: StoreClass = MVT::i64; break;
1483 default: assert(0 && "Unknown store class!");
1486 // Spill the integer to memory and reload it from there.
1487 unsigned Size = MVT::getSizeInBits(StoreClass)/8;
1488 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
1490 switch (StoreClass) {
1491 default: assert(0 && "Unknown store class!");
1493 addFrameReference(BuildMI(BB, X86::FIST16m, 5), FrameIdx).addReg(Tmp1);
1496 addFrameReference(BuildMI(BB, X86::FIST32m, 5), FrameIdx).addReg(Tmp1);
1499 addFrameReference(BuildMI(BB, X86::FISTP64m, 5), FrameIdx).addReg(Tmp1);
1503 switch (Node->getValueType(0)) {
1505 assert(0 && "Unknown integer type!");
1507 // FIXME: this isn't gunna work.
1508 assert(0 && "Cast FP to long not implemented yet!");
1509 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Result), FrameIdx);
1510 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Result+1), FrameIdx, 4);
1512 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Result), FrameIdx);
1515 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, Result), FrameIdx);
1518 addFrameReference(BuildMI(BB, X86::MOV8rm, 4, Result), FrameIdx);
1522 // Reload the original control word now.
1523 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1527 Op0 = N.getOperand(0);
1528 Op1 = N.getOperand(1);
1530 if (isFoldableLoad(Op0, Op1)) {
1531 std::swap(Op0, Op1);
1535 if (isFoldableLoad(Op1, Op0)) {
1537 switch (N.getValueType()) {
1538 default: assert(0 && "Cannot add this type!");
1540 case MVT::i8: Opc = X86::ADD8rm; break;
1541 case MVT::i16: Opc = X86::ADD16rm; break;
1542 case MVT::i32: Opc = X86::ADD32rm; break;
1543 case MVT::f32: Opc = X86::FADD32m; break;
1544 case MVT::f64: Opc = X86::FADD64m; break;
1547 EmitFoldedLoad(Op1, AM);
1548 Tmp1 = SelectExpr(Op0);
1549 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1553 // See if we can codegen this as an LEA to fold operations together.
1554 if (N.getValueType() == MVT::i32) {
1556 X86ISelAddressMode AM;
1557 MatchAddress(N, AM);
1558 ExprMap[N] = Result;
1560 // If this is not just an add, emit the LEA. For a simple add (like
1561 // reg+reg or reg+imm), we just emit an add. It might be a good idea to
1562 // leave this as LEA, then peephole it to 'ADD' after two address elim
1564 if (AM.Scale != 1 || AM.BaseType == X86ISelAddressMode::FrameIndexBase||
1565 AM.GV || (AM.Base.Reg.Val && AM.IndexReg.Val && AM.Disp)) {
1566 X86AddressMode XAM = SelectAddrExprs(AM);
1567 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), XAM);
1572 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
1574 if (CN->getValue() == 1) { // add X, 1 -> inc X
1575 switch (N.getValueType()) {
1576 default: assert(0 && "Cannot integer add this type!");
1577 case MVT::i8: Opc = X86::INC8r; break;
1578 case MVT::i16: Opc = X86::INC16r; break;
1579 case MVT::i32: Opc = X86::INC32r; break;
1581 } else if (CN->isAllOnesValue()) { // add X, -1 -> dec X
1582 switch (N.getValueType()) {
1583 default: assert(0 && "Cannot integer add this type!");
1584 case MVT::i8: Opc = X86::DEC8r; break;
1585 case MVT::i16: Opc = X86::DEC16r; break;
1586 case MVT::i32: Opc = X86::DEC32r; break;
1591 Tmp1 = SelectExpr(Op0);
1592 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1596 switch (N.getValueType()) {
1597 default: assert(0 && "Cannot add this type!");
1598 case MVT::i8: Opc = X86::ADD8ri; break;
1599 case MVT::i16: Opc = X86::ADD16ri; break;
1600 case MVT::i32: Opc = X86::ADD32ri; break;
1603 Tmp1 = SelectExpr(Op0);
1604 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
1609 switch (N.getValueType()) {
1610 default: assert(0 && "Cannot add this type!");
1611 case MVT::i8: Opc = X86::ADD8rr; break;
1612 case MVT::i16: Opc = X86::ADD16rr; break;
1613 case MVT::i32: Opc = X86::ADD32rr; break;
1614 case MVT::f64: Opc = X86::FpADD; break;
1617 if (getRegPressure(Op0) > getRegPressure(Op1)) {
1618 Tmp1 = SelectExpr(Op0);
1619 Tmp2 = SelectExpr(Op1);
1621 Tmp2 = SelectExpr(Op1);
1622 Tmp1 = SelectExpr(Op0);
1625 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1632 static const unsigned SUBTab[] = {
1633 X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, 0,
1634 X86::SUB8rm, X86::SUB16rm, X86::SUB32rm, X86::FSUB32m, X86::FSUB64m,
1635 X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::FpSUB , X86::FpSUB,
1637 static const unsigned MULTab[] = {
1638 0, X86::IMUL16rri, X86::IMUL32rri, 0, 0,
1639 0, X86::IMUL16rm , X86::IMUL32rm, X86::FMUL32m, X86::FMUL64m,
1640 0, X86::IMUL16rr , X86::IMUL32rr, X86::FpMUL , X86::FpMUL,
1642 static const unsigned ANDTab[] = {
1643 X86::AND8ri, X86::AND16ri, X86::AND32ri, 0, 0,
1644 X86::AND8rm, X86::AND16rm, X86::AND32rm, 0, 0,
1645 X86::AND8rr, X86::AND16rr, X86::AND32rr, 0, 0,
1647 static const unsigned ORTab[] = {
1648 X86::OR8ri, X86::OR16ri, X86::OR32ri, 0, 0,
1649 X86::OR8rm, X86::OR16rm, X86::OR32rm, 0, 0,
1650 X86::OR8rr, X86::OR16rr, X86::OR32rr, 0, 0,
1652 static const unsigned XORTab[] = {
1653 X86::XOR8ri, X86::XOR16ri, X86::XOR32ri, 0, 0,
1654 X86::XOR8rm, X86::XOR16rm, X86::XOR32rm, 0, 0,
1655 X86::XOR8rr, X86::XOR16rr, X86::XOR32rr, 0, 0,
1658 Op0 = Node->getOperand(0);
1659 Op1 = Node->getOperand(1);
1661 if (Node->getOpcode() == ISD::SUB && MVT::isInteger(N.getValueType()))
1662 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(0)))
1663 if (CN->isNullValue()) { // 0 - N -> neg N
1664 switch (N.getValueType()) {
1665 default: assert(0 && "Cannot sub this type!");
1667 case MVT::i8: Opc = X86::NEG8r; break;
1668 case MVT::i16: Opc = X86::NEG16r; break;
1669 case MVT::i32: Opc = X86::NEG32r; break;
1671 Tmp1 = SelectExpr(N.getOperand(1));
1672 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1676 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
1677 if (CN->isAllOnesValue() && Node->getOpcode() == ISD::XOR) {
1679 switch (N.getValueType()) {
1680 default: assert(0 && "Cannot add this type!");
1681 case MVT::i1: break; // Not supported, don't invert upper bits!
1682 case MVT::i8: Opc = X86::NOT8r; break;
1683 case MVT::i16: Opc = X86::NOT16r; break;
1684 case MVT::i32: Opc = X86::NOT32r; break;
1687 Tmp1 = SelectExpr(Op0);
1688 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1693 // Fold common multiplies into LEA instructions.
1694 if (Node->getOpcode() == ISD::MUL && N.getValueType() == MVT::i32) {
1695 switch ((int)CN->getValue()) {
1700 // Remove N from exprmap so SelectAddress doesn't get confused.
1703 SelectAddress(N, AM);
1704 // Restore it to the map.
1705 ExprMap[N] = Result;
1706 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), AM);
1711 switch (N.getValueType()) {
1712 default: assert(0 && "Cannot xor this type!");
1714 case MVT::i8: Opc = 0; break;
1715 case MVT::i16: Opc = 1; break;
1716 case MVT::i32: Opc = 2; break;
1718 switch (Node->getOpcode()) {
1719 default: assert(0 && "Unreachable!");
1720 case ISD::SUB: Opc = SUBTab[Opc]; break;
1721 case ISD::MUL: Opc = MULTab[Opc]; break;
1722 case ISD::AND: Opc = ANDTab[Opc]; break;
1723 case ISD::OR: Opc = ORTab[Opc]; break;
1724 case ISD::XOR: Opc = XORTab[Opc]; break;
1726 if (Opc) { // Can't fold MUL:i8 R, imm
1727 Tmp1 = SelectExpr(Op0);
1728 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
1733 if (isFoldableLoad(Op0, Op1))
1734 if (Node->getOpcode() != ISD::SUB) {
1735 std::swap(Op0, Op1);
1738 // Emit 'reverse' subract, with a memory operand.
1739 switch (N.getValueType()) {
1740 default: Opc = 0; break;
1741 case MVT::f32: Opc = X86::FSUBR32m; break;
1742 case MVT::f64: Opc = X86::FSUBR64m; break;
1746 EmitFoldedLoad(Op0, AM);
1747 Tmp1 = SelectExpr(Op1);
1748 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1753 if (isFoldableLoad(Op1, Op0)) {
1755 switch (N.getValueType()) {
1756 default: assert(0 && "Cannot operate on this type!");
1758 case MVT::i8: Opc = 5; break;
1759 case MVT::i16: Opc = 6; break;
1760 case MVT::i32: Opc = 7; break;
1761 case MVT::f32: Opc = 8; break;
1762 case MVT::f64: Opc = 9; break;
1764 switch (Node->getOpcode()) {
1765 default: assert(0 && "Unreachable!");
1766 case ISD::SUB: Opc = SUBTab[Opc]; break;
1767 case ISD::MUL: Opc = MULTab[Opc]; break;
1768 case ISD::AND: Opc = ANDTab[Opc]; break;
1769 case ISD::OR: Opc = ORTab[Opc]; break;
1770 case ISD::XOR: Opc = XORTab[Opc]; break;
1774 EmitFoldedLoad(Op1, AM);
1775 Tmp1 = SelectExpr(Op0);
1777 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1779 assert(Node->getOpcode() == ISD::MUL &&
1780 N.getValueType() == MVT::i8 && "Unexpected situation!");
1781 // Must use the MUL instruction, which forces use of AL.
1782 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
1783 addFullAddress(BuildMI(BB, X86::MUL8m, 1), AM);
1784 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
1789 if (getRegPressure(Op0) > getRegPressure(Op1)) {
1790 Tmp1 = SelectExpr(Op0);
1791 Tmp2 = SelectExpr(Op1);
1793 Tmp2 = SelectExpr(Op1);
1794 Tmp1 = SelectExpr(Op0);
1797 switch (N.getValueType()) {
1798 default: assert(0 && "Cannot add this type!");
1800 case MVT::i8: Opc = 10; break;
1801 case MVT::i16: Opc = 11; break;
1802 case MVT::i32: Opc = 12; break;
1803 case MVT::f32: Opc = 13; break;
1804 case MVT::f64: Opc = 14; break;
1806 switch (Node->getOpcode()) {
1807 default: assert(0 && "Unreachable!");
1808 case ISD::SUB: Opc = SUBTab[Opc]; break;
1809 case ISD::MUL: Opc = MULTab[Opc]; break;
1810 case ISD::AND: Opc = ANDTab[Opc]; break;
1811 case ISD::OR: Opc = ORTab[Opc]; break;
1812 case ISD::XOR: Opc = XORTab[Opc]; break;
1815 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1817 assert(Node->getOpcode() == ISD::MUL &&
1818 N.getValueType() == MVT::i8 && "Unexpected situation!");
1819 // Must use the MUL instruction, which forces use of AL.
1820 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
1821 BuildMI(BB, X86::MUL8r, 1).addReg(Tmp2);
1822 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
1827 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
1828 Tmp2 = SelectExpr(N.getOperand(1));
1829 Tmp3 = SelectExpr(N.getOperand(2));
1831 Tmp3 = SelectExpr(N.getOperand(2));
1832 Tmp2 = SelectExpr(N.getOperand(1));
1834 EmitSelectCC(N.getOperand(0), N.getValueType(), Tmp2, Tmp3, Result);
1841 assert((N.getOpcode() != ISD::SREM || MVT::isInteger(N.getValueType())) &&
1842 "We don't support this operator!");
1844 if (N.getOpcode() == ISD::SDIV)
1845 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1846 // FIXME: These special cases should be handled by the lowering impl!
1847 unsigned RHS = CN->getValue();
1853 if (RHS && (RHS & (RHS-1)) == 0) { // Signed division by power of 2?
1854 unsigned Log = log2(RHS);
1855 unsigned TmpReg = MakeReg(N.getValueType());
1856 unsigned SAROpc, SHROpc, ADDOpc, NEGOpc;
1857 switch (N.getValueType()) {
1858 default: assert("Unknown type to signed divide!");
1860 SAROpc = X86::SAR8ri;
1861 SHROpc = X86::SHR8ri;
1862 ADDOpc = X86::ADD8rr;
1863 NEGOpc = X86::NEG8r;
1866 SAROpc = X86::SAR16ri;
1867 SHROpc = X86::SHR16ri;
1868 ADDOpc = X86::ADD16rr;
1869 NEGOpc = X86::NEG16r;
1872 SAROpc = X86::SAR32ri;
1873 SHROpc = X86::SHR32ri;
1874 ADDOpc = X86::ADD32rr;
1875 NEGOpc = X86::NEG32r;
1878 Tmp1 = SelectExpr(N.getOperand(0));
1879 BuildMI(BB, SAROpc, 2, TmpReg).addReg(Tmp1).addImm(Log-1);
1880 unsigned TmpReg2 = MakeReg(N.getValueType());
1881 BuildMI(BB, SHROpc, 2, TmpReg2).addReg(TmpReg).addImm(32-Log);
1882 unsigned TmpReg3 = MakeReg(N.getValueType());
1883 BuildMI(BB, ADDOpc, 2, TmpReg3).addReg(Tmp1).addReg(TmpReg2);
1885 unsigned TmpReg4 = isNeg ? MakeReg(N.getValueType()) : Result;
1886 BuildMI(BB, SAROpc, 2, TmpReg4).addReg(TmpReg3).addImm(Log);
1888 BuildMI(BB, NEGOpc, 1, Result).addReg(TmpReg4);
1893 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
1894 Tmp1 = SelectExpr(N.getOperand(0));
1895 Tmp2 = SelectExpr(N.getOperand(1));
1897 Tmp2 = SelectExpr(N.getOperand(1));
1898 Tmp1 = SelectExpr(N.getOperand(0));
1901 bool isSigned = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::SREM;
1902 bool isDiv = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::UDIV;
1903 unsigned LoReg, HiReg, DivOpcode, MovOpcode, ClrOpcode, SExtOpcode;
1904 switch (N.getValueType()) {
1905 default: assert(0 && "Cannot sdiv this type!");
1907 DivOpcode = isSigned ? X86::IDIV8r : X86::DIV8r;
1910 MovOpcode = X86::MOV8rr;
1911 ClrOpcode = X86::MOV8ri;
1912 SExtOpcode = X86::CBW;
1915 DivOpcode = isSigned ? X86::IDIV16r : X86::DIV16r;
1918 MovOpcode = X86::MOV16rr;
1919 ClrOpcode = X86::MOV16ri;
1920 SExtOpcode = X86::CWD;
1923 DivOpcode = isSigned ? X86::IDIV32r : X86::DIV32r;
1926 MovOpcode = X86::MOV32rr;
1927 ClrOpcode = X86::MOV32ri;
1928 SExtOpcode = X86::CDQ;
1931 BuildMI(BB, X86::FpDIV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1935 // Set up the low part.
1936 BuildMI(BB, MovOpcode, 1, LoReg).addReg(Tmp1);
1939 // Sign extend the low part into the high part.
1940 BuildMI(BB, SExtOpcode, 0);
1942 // Zero out the high part, effectively zero extending the input.
1943 BuildMI(BB, ClrOpcode, 1, HiReg).addImm(0);
1946 // Emit the DIV/IDIV instruction.
1947 BuildMI(BB, DivOpcode, 1).addReg(Tmp2);
1949 // Get the result of the divide or rem.
1950 BuildMI(BB, MovOpcode, 1, Result).addReg(isDiv ? LoReg : HiReg);
1955 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1956 if (CN->getValue() == 1) { // X = SHL Y, 1 -> X = ADD Y, Y
1957 switch (N.getValueType()) {
1958 default: assert(0 && "Cannot shift this type!");
1959 case MVT::i8: Opc = X86::ADD8rr; break;
1960 case MVT::i16: Opc = X86::ADD16rr; break;
1961 case MVT::i32: Opc = X86::ADD32rr; break;
1963 Tmp1 = SelectExpr(N.getOperand(0));
1964 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp1);
1968 switch (N.getValueType()) {
1969 default: assert(0 && "Cannot shift this type!");
1970 case MVT::i8: Opc = X86::SHL8ri; break;
1971 case MVT::i16: Opc = X86::SHL16ri; break;
1972 case MVT::i32: Opc = X86::SHL32ri; break;
1974 Tmp1 = SelectExpr(N.getOperand(0));
1975 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
1979 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
1980 Tmp1 = SelectExpr(N.getOperand(0));
1981 Tmp2 = SelectExpr(N.getOperand(1));
1983 Tmp2 = SelectExpr(N.getOperand(1));
1984 Tmp1 = SelectExpr(N.getOperand(0));
1987 switch (N.getValueType()) {
1988 default: assert(0 && "Cannot shift this type!");
1989 case MVT::i8 : Opc = X86::SHL8rCL; break;
1990 case MVT::i16: Opc = X86::SHL16rCL; break;
1991 case MVT::i32: Opc = X86::SHL32rCL; break;
1993 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
1994 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1997 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1998 switch (N.getValueType()) {
1999 default: assert(0 && "Cannot shift this type!");
2000 case MVT::i8: Opc = X86::SHR8ri; break;
2001 case MVT::i16: Opc = X86::SHR16ri; break;
2002 case MVT::i32: Opc = X86::SHR32ri; break;
2004 Tmp1 = SelectExpr(N.getOperand(0));
2005 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2009 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2010 Tmp1 = SelectExpr(N.getOperand(0));
2011 Tmp2 = SelectExpr(N.getOperand(1));
2013 Tmp2 = SelectExpr(N.getOperand(1));
2014 Tmp1 = SelectExpr(N.getOperand(0));
2017 switch (N.getValueType()) {
2018 default: assert(0 && "Cannot shift this type!");
2019 case MVT::i8 : Opc = X86::SHR8rCL; break;
2020 case MVT::i16: Opc = X86::SHR16rCL; break;
2021 case MVT::i32: Opc = X86::SHR32rCL; break;
2023 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
2024 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
2027 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2028 switch (N.getValueType()) {
2029 default: assert(0 && "Cannot shift this type!");
2030 case MVT::i8: Opc = X86::SAR8ri; break;
2031 case MVT::i16: Opc = X86::SAR16ri; break;
2032 case MVT::i32: Opc = X86::SAR32ri; break;
2034 Tmp1 = SelectExpr(N.getOperand(0));
2035 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2039 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2040 Tmp1 = SelectExpr(N.getOperand(0));
2041 Tmp2 = SelectExpr(N.getOperand(1));
2043 Tmp2 = SelectExpr(N.getOperand(1));
2044 Tmp1 = SelectExpr(N.getOperand(0));
2047 switch (N.getValueType()) {
2048 default: assert(0 && "Cannot shift this type!");
2049 case MVT::i8 : Opc = X86::SAR8rCL; break;
2050 case MVT::i16: Opc = X86::SAR16rCL; break;
2051 case MVT::i32: Opc = X86::SAR32rCL; break;
2053 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
2054 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
2058 EmitCMP(N.getOperand(0), N.getOperand(1), Node->hasOneUse());
2059 EmitSetCC(BB, Result, cast<SetCCSDNode>(N)->getCondition(),
2060 MVT::isFloatingPoint(N.getOperand(1).getValueType()));
2063 // Make sure we generate both values.
2064 if (Result != 1) { // Generate the token
2065 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
2066 assert(0 && "Load already emitted!?");
2068 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2070 switch (Node->getValueType(0)) {
2071 default: assert(0 && "Cannot load this type!");
2073 case MVT::i8: Opc = X86::MOV8rm; break;
2074 case MVT::i16: Opc = X86::MOV16rm; break;
2075 case MVT::i32: Opc = X86::MOV32rm; break;
2076 case MVT::f64: Opc = X86::FLD64m; ContainsFPCode = true; break;
2079 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1))){
2080 Select(N.getOperand(0));
2081 addConstantPoolReference(BuildMI(BB, Opc, 4, Result), CP->getIndex());
2085 SDOperand Chain = N.getOperand(0);
2086 SDOperand Address = N.getOperand(1);
2087 if (getRegPressure(Chain) > getRegPressure(Address)) {
2089 SelectAddress(Address, AM);
2091 SelectAddress(Address, AM);
2095 addFullAddress(BuildMI(BB, Opc, 4, Result), AM);
2099 case ISD::EXTLOAD: // Arbitrarily codegen extloads as MOVZX*
2100 case ISD::ZEXTLOAD: {
2101 // Make sure we generate both values.
2103 ExprMap[N.getValue(1)] = 1; // Generate the token
2105 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2107 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1)))
2108 if (Node->getValueType(0) == MVT::f64) {
2109 assert(cast<MVTSDNode>(Node)->getExtraValueType() == MVT::f32 &&
2111 addConstantPoolReference(BuildMI(BB, X86::FLD32m, 4, Result),
2117 if (getRegPressure(Node->getOperand(0)) >
2118 getRegPressure(Node->getOperand(1))) {
2119 Select(Node->getOperand(0)); // chain
2120 SelectAddress(Node->getOperand(1), AM);
2122 SelectAddress(Node->getOperand(1), AM);
2123 Select(Node->getOperand(0)); // chain
2126 switch (Node->getValueType(0)) {
2127 default: assert(0 && "Unknown type to sign extend to.");
2129 assert(cast<MVTSDNode>(Node)->getExtraValueType() == MVT::f32 &&
2131 addFullAddress(BuildMI(BB, X86::FLD32m, 5, Result), AM);
2134 switch (cast<MVTSDNode>(Node)->getExtraValueType()) {
2136 assert(0 && "Bad zero extend!");
2139 addFullAddress(BuildMI(BB, X86::MOVZX32rm8, 5, Result), AM);
2142 addFullAddress(BuildMI(BB, X86::MOVZX32rm16, 5, Result), AM);
2147 assert(cast<MVTSDNode>(Node)->getExtraValueType() <= MVT::i8 &&
2148 "Bad zero extend!");
2149 addFullAddress(BuildMI(BB, X86::MOVSX16rm8, 5, Result), AM);
2152 assert(cast<MVTSDNode>(Node)->getExtraValueType() == MVT::i1 &&
2153 "Bad zero extend!");
2154 addFullAddress(BuildMI(BB, X86::MOV8rm, 5, Result), AM);
2159 case ISD::SEXTLOAD: {
2160 // Make sure we generate both values.
2162 ExprMap[N.getValue(1)] = 1; // Generate the token
2164 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2167 if (getRegPressure(Node->getOperand(0)) >
2168 getRegPressure(Node->getOperand(1))) {
2169 Select(Node->getOperand(0)); // chain
2170 SelectAddress(Node->getOperand(1), AM);
2172 SelectAddress(Node->getOperand(1), AM);
2173 Select(Node->getOperand(0)); // chain
2176 switch (Node->getValueType(0)) {
2177 case MVT::i8: assert(0 && "Cannot sign extend from bool!");
2178 default: assert(0 && "Unknown type to sign extend to.");
2180 switch (cast<MVTSDNode>(Node)->getExtraValueType()) {
2182 case MVT::i1: assert(0 && "Cannot sign extend from bool!");
2184 addFullAddress(BuildMI(BB, X86::MOVSX32rm8, 5, Result), AM);
2187 addFullAddress(BuildMI(BB, X86::MOVSX32rm16, 5, Result), AM);
2192 assert(cast<MVTSDNode>(Node)->getExtraValueType() == MVT::i8 &&
2193 "Cannot sign extend from bool!");
2194 addFullAddress(BuildMI(BB, X86::MOVSX16rm8, 5, Result), AM);
2200 case ISD::DYNAMIC_STACKALLOC:
2201 // Generate both result values.
2203 ExprMap[N.getValue(1)] = 1; // Generate the token
2205 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2207 // FIXME: We are currently ignoring the requested alignment for handling
2208 // greater than the stack alignment. This will need to be revisited at some
2209 // point. Align = N.getOperand(2);
2211 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
2212 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
2213 std::cerr << "Cannot allocate stack object with greater alignment than"
2214 << " the stack alignment yet!";
2218 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2219 Select(N.getOperand(0));
2220 BuildMI(BB, X86::SUB32ri, 2, X86::ESP).addReg(X86::ESP)
2221 .addImm(CN->getValue());
2223 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2224 Select(N.getOperand(0));
2225 Tmp1 = SelectExpr(N.getOperand(1));
2227 Tmp1 = SelectExpr(N.getOperand(1));
2228 Select(N.getOperand(0));
2231 // Subtract size from stack pointer, thereby allocating some space.
2232 BuildMI(BB, X86::SUB32rr, 2, X86::ESP).addReg(X86::ESP).addReg(Tmp1);
2235 // Put a pointer to the space into the result register, by copying the stack
2237 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::ESP);
2241 // The chain for this call is now lowered.
2242 ExprMap.insert(std::make_pair(N.getValue(Node->getNumValues()-1), 1));
2244 if (GlobalAddressSDNode *GASD =
2245 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
2246 Select(N.getOperand(0));
2247 BuildMI(BB, X86::CALLpcrel32, 1).addGlobalAddress(GASD->getGlobal(),true);
2248 } else if (ExternalSymbolSDNode *ESSDN =
2249 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
2250 Select(N.getOperand(0));
2251 BuildMI(BB, X86::CALLpcrel32,
2252 1).addExternalSymbol(ESSDN->getSymbol(), true);
2254 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2255 Select(N.getOperand(0));
2256 Tmp1 = SelectExpr(N.getOperand(1));
2258 Tmp1 = SelectExpr(N.getOperand(1));
2259 Select(N.getOperand(0));
2262 BuildMI(BB, X86::CALL32r, 1).addReg(Tmp1);
2264 switch (Node->getValueType(0)) {
2265 default: assert(0 && "Unknown value type for call result!");
2266 case MVT::Other: return 1;
2269 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2272 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
2275 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
2276 if (Node->getValueType(1) == MVT::i32)
2277 BuildMI(BB, X86::MOV32rr, 1, Result+1).addReg(X86::EDX);
2279 case MVT::f64: // Floating-point return values live in %ST(0)
2280 ContainsFPCode = true;
2281 BuildMI(BB, X86::FpGETRESULT, 1, Result);
2284 return Result+N.ResNo;
2290 /// TryToFoldLoadOpStore - Given a store node, try to fold together a
2291 /// load/op/store instruction. If successful return true.
2292 bool ISel::TryToFoldLoadOpStore(SDNode *Node) {
2293 assert(Node->getOpcode() == ISD::STORE && "Can only do this for stores!");
2294 SDOperand Chain = Node->getOperand(0);
2295 SDOperand StVal = Node->getOperand(1);
2296 SDOperand StPtr = Node->getOperand(2);
2298 // The chain has to be a load, the stored value must be an integer binary
2299 // operation with one use.
2300 if (!StVal.Val->hasOneUse() || StVal.Val->getNumOperands() != 2 ||
2301 MVT::isFloatingPoint(StVal.getValueType()))
2304 // Token chain must either be a factor node or the load to fold.
2305 if (Chain.getOpcode() != ISD::LOAD && Chain.getOpcode() != ISD::TokenFactor)
2310 // Check to see if there is a load from the same pointer that we're storing
2311 // to in either operand of the binop.
2312 if (StVal.getOperand(0).getOpcode() == ISD::LOAD &&
2313 StVal.getOperand(0).getOperand(1) == StPtr)
2314 TheLoad = StVal.getOperand(0);
2315 else if (StVal.getOperand(1).getOpcode() == ISD::LOAD &&
2316 StVal.getOperand(1).getOperand(1) == StPtr)
2317 TheLoad = StVal.getOperand(1);
2319 return false; // No matching load operand.
2321 // We can only fold the load if there are no intervening side-effecting
2322 // operations. This means that the store uses the load as its token chain, or
2323 // there are only token factor nodes in between the store and load.
2324 if (Chain != TheLoad.getValue(1)) {
2325 // Okay, the other option is that we have a store referring to (possibly
2326 // nested) token factor nodes. For now, just try peeking through one level
2327 // of token factors to see if this is the case.
2328 bool ChainOk = false;
2329 if (Chain.getOpcode() == ISD::TokenFactor) {
2330 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
2331 if (Chain.getOperand(i) == TheLoad.getValue(1)) {
2337 if (!ChainOk) return false;
2340 if (TheLoad.getOperand(1) != StPtr)
2343 // Make sure that one of the operands of the binop is the load, and that the
2344 // load folds into the binop.
2345 if (((StVal.getOperand(0) != TheLoad ||
2346 !isFoldableLoad(TheLoad, StVal.getOperand(1))) &&
2347 (StVal.getOperand(1) != TheLoad ||
2348 !isFoldableLoad(TheLoad, StVal.getOperand(0)))))
2351 // Finally, check to see if this is one of the ops we can handle!
2352 static const unsigned ADDTAB[] = {
2353 X86::ADD8mi, X86::ADD16mi, X86::ADD32mi,
2354 X86::ADD8mr, X86::ADD16mr, X86::ADD32mr,
2356 static const unsigned SUBTAB[] = {
2357 X86::SUB8mi, X86::SUB16mi, X86::SUB32mi,
2358 X86::SUB8mr, X86::SUB16mr, X86::SUB32mr,
2360 static const unsigned ANDTAB[] = {
2361 X86::AND8mi, X86::AND16mi, X86::AND32mi,
2362 X86::AND8mr, X86::AND16mr, X86::AND32mr,
2364 static const unsigned ORTAB[] = {
2365 X86::OR8mi, X86::OR16mi, X86::OR32mi,
2366 X86::OR8mr, X86::OR16mr, X86::OR32mr,
2368 static const unsigned XORTAB[] = {
2369 X86::XOR8mi, X86::XOR16mi, X86::XOR32mi,
2370 X86::XOR8mr, X86::XOR16mr, X86::XOR32mr,
2372 static const unsigned SHLTAB[] = {
2373 X86::SHL8mi, X86::SHL16mi, X86::SHL32mi,
2374 /*Have to put the reg in CL*/0, 0, 0,
2376 static const unsigned SARTAB[] = {
2377 X86::SAR8mi, X86::SAR16mi, X86::SAR32mi,
2378 /*Have to put the reg in CL*/0, 0, 0,
2380 static const unsigned SHRTAB[] = {
2381 X86::SHR8mi, X86::SHR16mi, X86::SHR32mi,
2382 /*Have to put the reg in CL*/0, 0, 0,
2385 const unsigned *TabPtr = 0;
2386 switch (StVal.getOpcode()) {
2388 std::cerr << "CANNOT [mem] op= val: ";
2389 StVal.Val->dump(); std::cerr << "\n";
2394 case ISD::UREM: return false;
2396 case ISD::ADD: TabPtr = ADDTAB; break;
2397 case ISD::SUB: TabPtr = SUBTAB; break;
2398 case ISD::AND: TabPtr = ANDTAB; break;
2399 case ISD:: OR: TabPtr = ORTAB; break;
2400 case ISD::XOR: TabPtr = XORTAB; break;
2401 case ISD::SHL: TabPtr = SHLTAB; break;
2402 case ISD::SRA: TabPtr = SARTAB; break;
2403 case ISD::SRL: TabPtr = SHRTAB; break;
2406 // Handle: [mem] op= CST
2407 SDOperand Op0 = StVal.getOperand(0);
2408 SDOperand Op1 = StVal.getOperand(1);
2410 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
2411 switch (Op0.getValueType()) { // Use Op0's type because of shifts.
2414 case MVT::i8: Opc = TabPtr[0]; break;
2415 case MVT::i16: Opc = TabPtr[1]; break;
2416 case MVT::i32: Opc = TabPtr[2]; break;
2420 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
2421 assert(0 && "Already emitted?");
2425 if (getRegPressure(TheLoad.getOperand(0)) >
2426 getRegPressure(TheLoad.getOperand(1))) {
2427 Select(TheLoad.getOperand(0));
2428 SelectAddress(TheLoad.getOperand(1), AM);
2430 SelectAddress(TheLoad.getOperand(1), AM);
2431 Select(TheLoad.getOperand(0));
2434 if (StVal.getOpcode() == ISD::ADD) {
2435 if (CN->getValue() == 1) {
2436 switch (Op0.getValueType()) {
2439 addFullAddress(BuildMI(BB, X86::INC8m, 4), AM);
2441 case MVT::i16: Opc = TabPtr[1];
2442 addFullAddress(BuildMI(BB, X86::INC16m, 4), AM);
2444 case MVT::i32: Opc = TabPtr[2];
2445 addFullAddress(BuildMI(BB, X86::INC32m, 4), AM);
2448 } else if (CN->getValue()+1 == 0) { // [X] += -1 -> DEC [X]
2449 switch (Op0.getValueType()) {
2452 addFullAddress(BuildMI(BB, X86::DEC8m, 4), AM);
2454 case MVT::i16: Opc = TabPtr[1];
2455 addFullAddress(BuildMI(BB, X86::DEC16m, 4), AM);
2457 case MVT::i32: Opc = TabPtr[2];
2458 addFullAddress(BuildMI(BB, X86::DEC32m, 4), AM);
2464 addFullAddress(BuildMI(BB, Opc, 4+1),AM).addImm(CN->getValue());
2469 // If we have [mem] = V op [mem], try to turn it into:
2470 // [mem] = [mem] op V.
2471 if (Op1 == TheLoad && StVal.getOpcode() != ISD::SUB &&
2472 StVal.getOpcode() != ISD::SHL && StVal.getOpcode() != ISD::SRA &&
2473 StVal.getOpcode() != ISD::SRL)
2474 std::swap(Op0, Op1);
2476 if (Op0 != TheLoad) return false;
2478 switch (Op0.getValueType()) {
2479 default: return false;
2481 case MVT::i8: Opc = TabPtr[3]; break;
2482 case MVT::i16: Opc = TabPtr[4]; break;
2483 case MVT::i32: Opc = TabPtr[5]; break;
2486 // Table entry doesn't exist?
2487 if (Opc == 0) return false;
2489 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
2490 assert(0 && "Already emitted?");
2492 Select(TheLoad.getOperand(0));
2495 SelectAddress(TheLoad.getOperand(1), AM);
2496 unsigned Reg = SelectExpr(Op1);
2497 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Reg);
2502 void ISel::Select(SDOperand N) {
2503 unsigned Tmp1, Tmp2, Opc;
2505 // FIXME: Disable for our current expansion model!
2506 if (/*!N->hasOneUse() &&*/ !ExprMap.insert(std::make_pair(N, 1)).second)
2507 return; // Already selected.
2509 SDNode *Node = N.Val;
2511 switch (Node->getOpcode()) {
2513 Node->dump(); std::cerr << "\n";
2514 assert(0 && "Node not handled yet!");
2515 case ISD::EntryToken: return; // Noop
2516 case ISD::TokenFactor:
2517 if (Node->getNumOperands() == 2) {
2519 getRegPressure(Node->getOperand(1))>getRegPressure(Node->getOperand(0));
2520 Select(Node->getOperand(OneFirst));
2521 Select(Node->getOperand(!OneFirst));
2523 std::vector<std::pair<unsigned, unsigned> > OpsP;
2524 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2525 OpsP.push_back(std::make_pair(getRegPressure(Node->getOperand(i)), i));
2526 std::sort(OpsP.begin(), OpsP.end());
2527 std::reverse(OpsP.begin(), OpsP.end());
2528 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2529 Select(Node->getOperand(OpsP[i].second));
2532 case ISD::CopyToReg:
2533 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2534 Select(N.getOperand(0));
2535 Tmp1 = SelectExpr(N.getOperand(1));
2537 Tmp1 = SelectExpr(N.getOperand(1));
2538 Select(N.getOperand(0));
2540 Tmp2 = cast<RegSDNode>(N)->getReg();
2543 switch (N.getOperand(1).getValueType()) {
2544 default: assert(0 && "Invalid type for operation!");
2546 case MVT::i8: Opc = X86::MOV8rr; break;
2547 case MVT::i16: Opc = X86::MOV16rr; break;
2548 case MVT::i32: Opc = X86::MOV32rr; break;
2549 case MVT::f64: Opc = X86::FpMOV; ContainsFPCode = true; break;
2551 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
2555 switch (N.getNumOperands()) {
2557 assert(0 && "Unknown return instruction!");
2559 assert(N.getOperand(1).getValueType() == MVT::i32 &&
2560 N.getOperand(2).getValueType() == MVT::i32 &&
2561 "Unknown two-register value!");
2562 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
2563 Tmp1 = SelectExpr(N.getOperand(1));
2564 Tmp2 = SelectExpr(N.getOperand(2));
2566 Tmp2 = SelectExpr(N.getOperand(2));
2567 Tmp1 = SelectExpr(N.getOperand(1));
2569 Select(N.getOperand(0));
2571 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
2572 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(Tmp2);
2573 // Declare that EAX & EDX are live on exit.
2574 BuildMI(BB, X86::IMPLICIT_USE, 3).addReg(X86::EAX).addReg(X86::EDX)
2578 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2579 Select(N.getOperand(0));
2580 Tmp1 = SelectExpr(N.getOperand(1));
2582 Tmp1 = SelectExpr(N.getOperand(1));
2583 Select(N.getOperand(0));
2585 switch (N.getOperand(1).getValueType()) {
2586 default: assert(0 && "All other types should have been promoted!!");
2588 BuildMI(BB, X86::FpSETRESULT, 1).addReg(Tmp1);
2589 // Declare that top-of-stack is live on exit
2590 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::ST0).addReg(X86::ESP);
2593 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
2594 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::EAX).addReg(X86::ESP);
2599 Select(N.getOperand(0));
2602 BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
2605 Select(N.getOperand(0));
2606 MachineBasicBlock *Dest =
2607 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2608 BuildMI(BB, X86::JMP, 1).addMBB(Dest);
2613 MachineBasicBlock *Dest =
2614 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
2616 // Try to fold a setcc into the branch. If this fails, emit a test/jne
2618 if (EmitBranchCC(Dest, N.getOperand(0), N.getOperand(1))) {
2619 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2620 Select(N.getOperand(0));
2621 Tmp1 = SelectExpr(N.getOperand(1));
2623 Tmp1 = SelectExpr(N.getOperand(1));
2624 Select(N.getOperand(0));
2626 BuildMI(BB, X86::TEST8rr, 2).addReg(Tmp1).addReg(Tmp1);
2627 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
2634 // If this load could be folded into the only using instruction, and if it
2635 // is safe to emit the instruction here, try to do so now.
2636 if (Node->hasNUsesOfValue(1, 0)) {
2637 SDOperand TheVal = N.getValue(0);
2639 for (SDNode::use_iterator UI = Node->use_begin(); ; ++UI) {
2640 assert(UI != Node->use_end() && "Didn't find use!");
2642 for (unsigned i = 0, e = UN->getNumOperands(); i != e; ++i)
2643 if (UN->getOperand(i) == TheVal) {
2649 // Only handle unary operators right now.
2650 if (User->getNumOperands() == 1) {
2652 SelectExpr(SDOperand(User, 0));
2664 case ISD::DYNAMIC_STACKALLOC:
2669 case ISD::TRUNCSTORE: { // truncstore chain, val, ptr :storety
2670 // On X86, we can represent all types except for Bool and Float natively.
2672 MVT::ValueType StoredTy = cast<MVTSDNode>(Node)->getExtraValueType();
2673 assert((StoredTy == MVT::i1 || StoredTy == MVT::f32 ||
2674 StoredTy == MVT::i16 /*FIXME: THIS IS JUST FOR TESTING!*/)
2675 && "Unsupported TRUNCSTORE for this target!");
2677 if (StoredTy == MVT::i16) {
2678 // FIXME: This is here just to allow testing. X86 doesn't really have a
2679 // TRUNCSTORE i16 operation, but this is required for targets that do not
2680 // have 16-bit integer registers. We occasionally disable 16-bit integer
2681 // registers to test the promotion code.
2682 Select(N.getOperand(0));
2683 Tmp1 = SelectExpr(N.getOperand(1));
2684 SelectAddress(N.getOperand(2), AM);
2686 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
2687 addFullAddress(BuildMI(BB, X86::MOV16mr, 5), AM).addReg(X86::AX);
2691 // Store of constant bool?
2692 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2693 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
2694 Select(N.getOperand(0));
2695 SelectAddress(N.getOperand(2), AM);
2697 SelectAddress(N.getOperand(2), AM);
2698 Select(N.getOperand(0));
2700 addFullAddress(BuildMI(BB, X86::MOV8mi, 5), AM).addImm(CN->getValue());
2705 default: assert(0 && "Cannot truncstore this type!");
2706 case MVT::i1: Opc = X86::MOV8mr; break;
2707 case MVT::f32: Opc = X86::FST32m; break;
2710 std::vector<std::pair<unsigned, unsigned> > RP;
2711 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
2712 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
2713 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
2714 std::sort(RP.begin(), RP.end());
2716 for (unsigned i = 0; i != 3; ++i)
2717 switch (RP[2-i].second) {
2718 default: assert(0 && "Unknown operand number!");
2719 case 0: Select(N.getOperand(0)); break;
2720 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
2721 case 2: SelectAddress(N.getOperand(2), AM); break;
2724 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
2730 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2732 switch (CN->getValueType(0)) {
2733 default: assert(0 && "Invalid type for operation!");
2735 case MVT::i8: Opc = X86::MOV8mi; break;
2736 case MVT::i16: Opc = X86::MOV16mi; break;
2737 case MVT::i32: Opc = X86::MOV32mi; break;
2738 case MVT::f64: break;
2741 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
2742 Select(N.getOperand(0));
2743 SelectAddress(N.getOperand(2), AM);
2745 SelectAddress(N.getOperand(2), AM);
2746 Select(N.getOperand(0));
2748 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addImm(CN->getValue());
2753 // Check to see if this is a load/op/store combination.
2754 if (TryToFoldLoadOpStore(Node))
2757 switch (N.getOperand(1).getValueType()) {
2758 default: assert(0 && "Cannot store this type!");
2760 case MVT::i8: Opc = X86::MOV8mr; break;
2761 case MVT::i16: Opc = X86::MOV16mr; break;
2762 case MVT::i32: Opc = X86::MOV32mr; break;
2763 case MVT::f64: Opc = X86::FST64m; break;
2766 std::vector<std::pair<unsigned, unsigned> > RP;
2767 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
2768 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
2769 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
2770 std::sort(RP.begin(), RP.end());
2772 for (unsigned i = 0; i != 3; ++i)
2773 switch (RP[2-i].second) {
2774 default: assert(0 && "Unknown operand number!");
2775 case 0: Select(N.getOperand(0)); break;
2776 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
2777 case 2: SelectAddress(N.getOperand(2), AM); break;
2780 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
2783 case ISD::ADJCALLSTACKDOWN:
2784 case ISD::ADJCALLSTACKUP:
2785 Select(N.getOperand(0));
2786 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2788 Opc = N.getOpcode() == ISD::ADJCALLSTACKDOWN ? X86::ADJCALLSTACKDOWN :
2789 X86::ADJCALLSTACKUP;
2790 BuildMI(BB, Opc, 1).addImm(Tmp1);
2793 Select(N.getOperand(0)); // Select the chain.
2795 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
2796 if (Align == 0) Align = 1;
2798 // Turn the byte code into # iterations
2801 if (ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Node->getOperand(2))) {
2802 unsigned Val = ValC->getValue() & 255;
2804 // If the value is a constant, then we can potentially use larger sets.
2805 switch (Align & 3) {
2806 case 2: // WORD aligned
2807 CountReg = MakeReg(MVT::i32);
2808 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
2809 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
2811 unsigned ByteReg = SelectExpr(Node->getOperand(3));
2812 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
2814 BuildMI(BB, X86::MOV16ri, 1, X86::AX).addImm((Val << 8) | Val);
2815 Opcode = X86::REP_STOSW;
2817 case 0: // DWORD aligned
2818 CountReg = MakeReg(MVT::i32);
2819 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
2820 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
2822 unsigned ByteReg = SelectExpr(Node->getOperand(3));
2823 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
2825 Val = (Val << 8) | Val;
2826 BuildMI(BB, X86::MOV32ri, 1, X86::EAX).addImm((Val << 16) | Val);
2827 Opcode = X86::REP_STOSD;
2829 default: // BYTE aligned
2830 CountReg = SelectExpr(Node->getOperand(3));
2831 BuildMI(BB, X86::MOV8ri, 1, X86::AL).addImm(Val);
2832 Opcode = X86::REP_STOSB;
2836 // If it's not a constant value we are storing, just fall back. We could
2837 // try to be clever to form 16 bit and 32 bit values, but we don't yet.
2838 unsigned ValReg = SelectExpr(Node->getOperand(2));
2839 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(ValReg);
2840 CountReg = SelectExpr(Node->getOperand(3));
2841 Opcode = X86::REP_STOSB;
2844 // No matter what the alignment is, we put the source in ESI, the
2845 // destination in EDI, and the count in ECX.
2846 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
2847 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
2848 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
2849 BuildMI(BB, Opcode, 0);
2853 Select(N.getOperand(0)); // Select the chain.
2855 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
2856 if (Align == 0) Align = 1;
2858 // Turn the byte code into # iterations
2861 switch (Align & 3) {
2862 case 2: // WORD aligned
2863 CountReg = MakeReg(MVT::i32);
2864 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
2865 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
2867 unsigned ByteReg = SelectExpr(Node->getOperand(3));
2868 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
2870 Opcode = X86::REP_MOVSW;
2872 case 0: // DWORD aligned
2873 CountReg = MakeReg(MVT::i32);
2874 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
2875 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
2877 unsigned ByteReg = SelectExpr(Node->getOperand(3));
2878 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
2880 Opcode = X86::REP_MOVSD;
2882 default: // BYTE aligned
2883 CountReg = SelectExpr(Node->getOperand(3));
2884 Opcode = X86::REP_MOVSB;
2888 // No matter what the alignment is, we put the source in ESI, the
2889 // destination in EDI, and the count in ECX.
2890 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
2891 unsigned TmpReg2 = SelectExpr(Node->getOperand(2));
2892 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
2893 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
2894 BuildMI(BB, X86::MOV32rr, 1, X86::ESI).addReg(TmpReg2);
2895 BuildMI(BB, Opcode, 0);
2898 assert(0 && "Should not be reached!");
2902 /// createX86PatternInstructionSelector - This pass converts an LLVM function
2903 /// into a machine code representation using pattern matching and a machine
2904 /// description file.
2906 FunctionPass *llvm::createX86PatternInstructionSelector(TargetMachine &TM) {
2907 return new ISel(TM);