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 "X86Subtarget.h"
18 #include "X86ISelLowering.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/Function.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/CodeGen/SSARegMap.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/ADT/Statistic.h"
40 //===----------------------------------------------------------------------===//
41 // Pattern Matcher Implementation
42 //===----------------------------------------------------------------------===//
45 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
46 /// SDOperand's instead of register numbers for the leaves of the matched
48 struct X86ISelAddressMode {
54 struct { // This is really a union, discriminated by BaseType!
65 : BaseType(RegBase), Scale(1), IndexReg(), Disp(), GV(0) {
73 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
75 //===--------------------------------------------------------------------===//
76 /// ISel - X86 specific code to select X86 machine instructions for
77 /// SelectionDAG operations.
79 class ISel : public SelectionDAGISel {
80 /// ContainsFPCode - Every instruction we select that uses or defines a FP
81 /// register should set this to true.
84 /// X86Lowering - This object fully describes how to lower LLVM code to an
85 /// X86-specific SelectionDAG.
86 X86TargetLowering X86Lowering;
88 /// RegPressureMap - This keeps an approximate count of the number of
89 /// registers required to evaluate each node in the graph.
90 std::map<SDNode*, unsigned> RegPressureMap;
92 /// ExprMap - As shared expressions are codegen'd, we keep track of which
93 /// vreg the value is produced in, so we only emit one copy of each compiled
95 std::map<SDOperand, unsigned> ExprMap;
97 /// TheDAG - The DAG being selected during Select* operations.
100 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
101 /// make the right decision when generating code for different targets.
102 const X86Subtarget *Subtarget;
104 ISel(TargetMachine &TM) : SelectionDAGISel(X86Lowering), X86Lowering(TM) {
105 Subtarget = &TM.getSubtarget<X86Subtarget>();
108 virtual const char *getPassName() const {
109 return "X86 Pattern Instruction Selection";
112 unsigned getRegPressure(SDOperand O) {
113 return RegPressureMap[O.Val];
115 unsigned ComputeRegPressure(SDOperand O);
117 /// InstructionSelectBasicBlock - This callback is invoked by
118 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
119 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
121 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
123 bool isFoldableLoad(SDOperand Op, SDOperand OtherOp,
124 bool FloatPromoteOk = false);
125 void EmitFoldedLoad(SDOperand Op, X86AddressMode &AM);
126 bool TryToFoldLoadOpStore(SDNode *Node);
127 bool EmitOrOpOp(SDOperand Op1, SDOperand Op2, unsigned DestReg);
128 void EmitCMP(SDOperand LHS, SDOperand RHS, bool isOnlyUse);
129 bool EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain, SDOperand Cond);
130 void EmitSelectCC(SDOperand Cond, SDOperand True, SDOperand False,
131 MVT::ValueType SVT, unsigned RDest);
132 unsigned SelectExpr(SDOperand N);
134 X86AddressMode SelectAddrExprs(const X86ISelAddressMode &IAM);
135 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM);
136 void SelectAddress(SDOperand N, X86AddressMode &AM);
137 bool EmitPotentialTailCall(SDNode *Node);
138 void EmitFastCCToFastCCTailCall(SDNode *TailCallNode);
139 void Select(SDOperand N);
143 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
144 /// the main function.
145 static void EmitSpecialCodeForMain(MachineBasicBlock *BB,
146 MachineFrameInfo *MFI) {
147 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
148 int CWFrameIdx = MFI->CreateStackObject(2, 2);
149 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
151 // Set the high part to be 64-bit precision.
152 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
153 CWFrameIdx, 1).addImm(2);
155 // Reload the modified control word now.
156 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
159 void ISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
160 // If this is main, emit special code for main.
161 MachineBasicBlock *BB = MF.begin();
162 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
163 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
167 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
168 /// when it has created a SelectionDAG for us to codegen.
169 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
170 // While we're doing this, keep track of whether we see any FP code for
171 // FP_REG_KILL insertion.
172 ContainsFPCode = false;
173 MachineFunction *MF = BB->getParent();
175 // Scan the PHI nodes that already are inserted into this basic block. If any
176 // of them is a PHI of a floating point value, we need to insert an
178 SSARegMap *RegMap = MF->getSSARegMap();
179 if (BB != MF->begin())
180 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
182 assert(I->getOpcode() == X86::PHI &&
183 "Isn't just PHI nodes?");
184 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
185 X86::RFPRegisterClass) {
186 ContainsFPCode = true;
191 // Compute the RegPressureMap, which is an approximation for the number of
192 // registers required to compute each node.
193 ComputeRegPressure(DAG.getRoot());
197 // Codegen the basic block.
198 Select(DAG.getRoot());
202 // Finally, look at all of the successors of this block. If any contain a PHI
203 // node of FP type, we need to insert an FP_REG_KILL in this block.
204 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
205 E = BB->succ_end(); SI != E && !ContainsFPCode; ++SI)
206 for (MachineBasicBlock::iterator I = (*SI)->begin(), E = (*SI)->end();
207 I != E && I->getOpcode() == X86::PHI; ++I) {
208 if (RegMap->getRegClass(I->getOperand(0).getReg()) ==
209 X86::RFPRegisterClass) {
210 ContainsFPCode = true;
215 // Final check, check LLVM BB's that are successors to the LLVM BB
216 // corresponding to BB for FP PHI nodes.
217 const BasicBlock *LLVMBB = BB->getBasicBlock();
220 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
221 SI != E && !ContainsFPCode; ++SI)
222 for (BasicBlock::const_iterator II = SI->begin();
223 (PN = dyn_cast<PHINode>(II)); ++II)
224 if (PN->getType()->isFloatingPoint()) {
225 ContainsFPCode = true;
229 // Insert FP_REG_KILL instructions into basic blocks that need them. This
230 // only occurs due to the floating point stackifier not being aggressive
231 // enough to handle arbitrary global stackification.
233 // Currently we insert an FP_REG_KILL instruction into each block that uses or
234 // defines a floating point virtual register.
236 // When the global register allocators (like linear scan) finally update live
237 // variable analysis, we can keep floating point values in registers across
238 // basic blocks. This will be a huge win, but we are waiting on the global
239 // allocators before we can do this.
241 if (ContainsFPCode) {
242 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
246 // Clear state used for selection.
248 RegPressureMap.clear();
252 // ComputeRegPressure - Compute the RegPressureMap, which is an approximation
253 // for the number of registers required to compute each node. This is basically
254 // computing a generalized form of the Sethi-Ullman number for each node.
255 unsigned ISel::ComputeRegPressure(SDOperand O) {
257 unsigned &Result = RegPressureMap[N];
258 if (Result) return Result;
260 // FIXME: Should operations like CALL (which clobber lots o regs) have a
261 // higher fixed cost??
263 if (N->getNumOperands() == 0) {
266 unsigned MaxRegUse = 0;
267 unsigned NumExtraMaxRegUsers = 0;
268 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
270 if (N->getOperand(i).getOpcode() == ISD::Constant)
273 Regs = ComputeRegPressure(N->getOperand(i));
274 if (Regs > MaxRegUse) {
276 NumExtraMaxRegUsers = 0;
277 } else if (Regs == MaxRegUse &&
278 N->getOperand(i).getValueType() != MVT::Other) {
279 ++NumExtraMaxRegUsers;
283 if (O.getOpcode() != ISD::TokenFactor)
284 Result = MaxRegUse+NumExtraMaxRegUsers;
286 Result = MaxRegUse == 1 ? 0 : MaxRegUse-1;
289 //std::cerr << " WEIGHT: " << Result << " "; N->dump(); std::cerr << "\n";
293 /// NodeTransitivelyUsesValue - Return true if N or any of its uses uses Op.
294 /// The DAG cannot have cycles in it, by definition, so the visited set is not
295 /// needed to prevent infinite loops. The DAG CAN, however, have unbounded
296 /// reuse, so it prevents exponential cases.
298 static bool NodeTransitivelyUsesValue(SDOperand N, SDOperand Op,
299 std::set<SDNode*> &Visited) {
300 if (N == Op) return true; // Found it.
301 SDNode *Node = N.Val;
302 if (Node->getNumOperands() == 0 || // Leaf?
303 Node->getNodeDepth() <= Op.getNodeDepth()) return false; // Can't find it?
304 if (!Visited.insert(Node).second) return false; // Already visited?
306 // Recurse for the first N-1 operands.
307 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
308 if (NodeTransitivelyUsesValue(Node->getOperand(i), Op, Visited))
311 // Tail recurse for the last operand.
312 return NodeTransitivelyUsesValue(Node->getOperand(0), Op, Visited);
315 X86AddressMode ISel::SelectAddrExprs(const X86ISelAddressMode &IAM) {
316 X86AddressMode Result;
318 // If we need to emit two register operands, emit the one with the highest
319 // register pressure first.
320 if (IAM.BaseType == X86ISelAddressMode::RegBase &&
321 IAM.Base.Reg.Val && IAM.IndexReg.Val) {
322 bool EmitBaseThenIndex;
323 if (getRegPressure(IAM.Base.Reg) > getRegPressure(IAM.IndexReg)) {
324 std::set<SDNode*> Visited;
325 EmitBaseThenIndex = true;
326 // If Base ends up pointing to Index, we must emit index first. This is
327 // because of the way we fold loads, we may end up doing bad things with
329 if (NodeTransitivelyUsesValue(IAM.Base.Reg, IAM.IndexReg, Visited))
330 EmitBaseThenIndex = false;
332 std::set<SDNode*> Visited;
333 EmitBaseThenIndex = false;
334 // If Base ends up pointing to Index, we must emit index first. This is
335 // because of the way we fold loads, we may end up doing bad things with
337 if (NodeTransitivelyUsesValue(IAM.IndexReg, IAM.Base.Reg, Visited))
338 EmitBaseThenIndex = true;
341 if (EmitBaseThenIndex) {
342 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
343 Result.IndexReg = SelectExpr(IAM.IndexReg);
345 Result.IndexReg = SelectExpr(IAM.IndexReg);
346 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
349 } else if (IAM.BaseType == X86ISelAddressMode::RegBase && IAM.Base.Reg.Val) {
350 Result.Base.Reg = SelectExpr(IAM.Base.Reg);
351 } else if (IAM.IndexReg.Val) {
352 Result.IndexReg = SelectExpr(IAM.IndexReg);
355 switch (IAM.BaseType) {
356 case X86ISelAddressMode::RegBase:
357 Result.BaseType = X86AddressMode::RegBase;
359 case X86ISelAddressMode::FrameIndexBase:
360 Result.BaseType = X86AddressMode::FrameIndexBase;
361 Result.Base.FrameIndex = IAM.Base.FrameIndex;
364 assert(0 && "Unknown base type!");
367 Result.Scale = IAM.Scale;
368 Result.Disp = IAM.Disp;
373 /// SelectAddress - Pattern match the maximal addressing mode for this node and
374 /// emit all of the leaf registers.
375 void ISel::SelectAddress(SDOperand N, X86AddressMode &AM) {
376 X86ISelAddressMode IAM;
377 MatchAddress(N, IAM);
378 AM = SelectAddrExprs(IAM);
381 /// MatchAddress - Add the specified node to the specified addressing mode,
382 /// returning true if it cannot be done. This just pattern matches for the
383 /// addressing mode, it does not cause any code to be emitted. For that, use
385 bool ISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM) {
386 switch (N.getOpcode()) {
388 case ISD::FrameIndex:
389 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
390 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
391 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
395 case ISD::GlobalAddress:
397 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
398 // For Darwin, external and weak symbols are indirect, so we want to load
399 // the value at address GV, not the value of GV itself. This means that
400 // the GlobalAddress must be in the base or index register of the address,
401 // not the GV offset field.
402 if (Subtarget->getIndirectExternAndWeakGlobals() &&
403 (GV->hasWeakLinkage() || GV->isExternal())) {
412 AM.Disp += cast<ConstantSDNode>(N)->getValue();
415 // We might have folded the load into this shift, so don't regen the value
417 if (ExprMap.count(N)) break;
419 if (AM.IndexReg.Val == 0 && AM.Scale == 1)
420 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
421 unsigned Val = CN->getValue();
422 if (Val == 1 || Val == 2 || Val == 3) {
424 SDOperand ShVal = N.Val->getOperand(0);
426 // Okay, we know that we have a scale by now. However, if the scaled
427 // value is an add of something and a constant, we can fold the
428 // constant into the disp field here.
429 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
430 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
431 AM.IndexReg = ShVal.Val->getOperand(0);
432 ConstantSDNode *AddVal =
433 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
434 AM.Disp += AddVal->getValue() << Val;
443 // We might have folded the load into this mul, so don't regen the value if
445 if (ExprMap.count(N)) break;
447 // X*[3,5,9] -> X+X*[2,4,8]
448 if (AM.IndexReg.Val == 0 && AM.BaseType == X86ISelAddressMode::RegBase &&
449 AM.Base.Reg.Val == 0)
450 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
451 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
452 AM.Scale = unsigned(CN->getValue())-1;
454 SDOperand MulVal = N.Val->getOperand(0);
457 // Okay, we know that we have a scale by now. However, if the scaled
458 // value is an add of something and a constant, we can fold the
459 // constant into the disp field here.
460 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
461 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
462 Reg = MulVal.Val->getOperand(0);
463 ConstantSDNode *AddVal =
464 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
465 AM.Disp += AddVal->getValue() * CN->getValue();
467 Reg = N.Val->getOperand(0);
470 AM.IndexReg = AM.Base.Reg = Reg;
476 // We might have folded the load into this mul, so don't regen the value if
478 if (ExprMap.count(N)) break;
480 X86ISelAddressMode Backup = AM;
481 if (!MatchAddress(N.Val->getOperand(0), AM) &&
482 !MatchAddress(N.Val->getOperand(1), AM))
485 if (!MatchAddress(N.Val->getOperand(1), AM) &&
486 !MatchAddress(N.Val->getOperand(0), AM))
493 // Is the base register already occupied?
494 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
495 // If so, check to see if the scale index register is set.
496 if (AM.IndexReg.Val == 0) {
502 // Otherwise, we cannot select it.
506 // Default, generate it as a register.
507 AM.BaseType = X86ISelAddressMode::RegBase;
512 /// Emit2SetCCsAndLogical - Emit the following sequence of instructions,
513 /// assuming that the temporary registers are in the 8-bit register class.
517 /// DestReg = logicalop Tmp1, Tmp2
519 static void Emit2SetCCsAndLogical(MachineBasicBlock *BB, unsigned SetCC1,
520 unsigned SetCC2, unsigned LogicalOp,
522 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
523 unsigned Tmp1 = RegMap->createVirtualRegister(X86::R8RegisterClass);
524 unsigned Tmp2 = RegMap->createVirtualRegister(X86::R8RegisterClass);
525 BuildMI(BB, SetCC1, 0, Tmp1);
526 BuildMI(BB, SetCC2, 0, Tmp2);
527 BuildMI(BB, LogicalOp, 2, DestReg).addReg(Tmp1).addReg(Tmp2);
530 /// EmitSetCC - Emit the code to set the specified 8-bit register to 1 if the
531 /// condition codes match the specified SetCCOpcode. Note that some conditions
532 /// require multiple instructions to generate the correct value.
533 static void EmitSetCC(MachineBasicBlock *BB, unsigned DestReg,
534 ISD::CondCode SetCCOpcode, bool isFP) {
537 switch (SetCCOpcode) {
538 default: assert(0 && "Illegal integer SetCC!");
539 case ISD::SETEQ: Opc = X86::SETEr; break;
540 case ISD::SETGT: Opc = X86::SETGr; break;
541 case ISD::SETGE: Opc = X86::SETGEr; break;
542 case ISD::SETLT: Opc = X86::SETLr; break;
543 case ISD::SETLE: Opc = X86::SETLEr; break;
544 case ISD::SETNE: Opc = X86::SETNEr; break;
545 case ISD::SETULT: Opc = X86::SETBr; break;
546 case ISD::SETUGT: Opc = X86::SETAr; break;
547 case ISD::SETULE: Opc = X86::SETBEr; break;
548 case ISD::SETUGE: Opc = X86::SETAEr; break;
551 // On a floating point condition, the flags are set as follows:
555 // 1 | 0 | 0 | X == Y
556 // 1 | 1 | 1 | unordered
558 switch (SetCCOpcode) {
559 default: assert(0 && "Invalid FP setcc!");
562 Opc = X86::SETEr; // True if ZF = 1
566 Opc = X86::SETAr; // True if CF = 0 and ZF = 0
570 Opc = X86::SETAEr; // True if CF = 0
574 Opc = X86::SETBr; // True if CF = 1
578 Opc = X86::SETBEr; // True if CF = 1 or ZF = 1
582 Opc = X86::SETNEr; // True if ZF = 0
585 Opc = X86::SETPr; // True if PF = 1
588 Opc = X86::SETNPr; // True if PF = 0
590 case ISD::SETOEQ: // !PF & ZF
591 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETEr, X86::AND8rr, DestReg);
593 case ISD::SETOLT: // !PF & CF
594 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBr, X86::AND8rr, DestReg);
596 case ISD::SETOLE: // !PF & (CF || ZF)
597 Emit2SetCCsAndLogical(BB, X86::SETNPr, X86::SETBEr, X86::AND8rr, DestReg);
599 case ISD::SETUGT: // PF | (!ZF & !CF)
600 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAr, X86::OR8rr, DestReg);
602 case ISD::SETUGE: // PF | !CF
603 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETAEr, X86::OR8rr, DestReg);
605 case ISD::SETUNE: // PF | !ZF
606 Emit2SetCCsAndLogical(BB, X86::SETPr, X86::SETNEr, X86::OR8rr, DestReg);
610 BuildMI(BB, Opc, 0, DestReg);
614 /// EmitBranchCC - Emit code into BB that arranges for control to transfer to
615 /// the Dest block if the Cond condition is true. If we cannot fold this
616 /// condition into the branch, return true.
618 bool ISel::EmitBranchCC(MachineBasicBlock *Dest, SDOperand Chain,
620 // FIXME: Evaluate whether it would be good to emit code like (X < Y) | (A >
621 // B) using two conditional branches instead of one condbr, two setcc's, and
623 if ((Cond.getOpcode() == ISD::OR ||
624 Cond.getOpcode() == ISD::AND) && Cond.Val->hasOneUse()) {
625 // And and or set the flags for us, so there is no need to emit a TST of the
626 // result. It is only safe to do this if there is only a single use of the
627 // AND/OR though, otherwise we don't know it will be emitted here.
630 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
634 // Codegen br not C -> JE.
635 if (Cond.getOpcode() == ISD::XOR)
636 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(Cond.Val->getOperand(1)))
637 if (NC->isAllOnesValue()) {
639 if (getRegPressure(Chain) > getRegPressure(Cond)) {
641 CondR = SelectExpr(Cond.Val->getOperand(0));
643 CondR = SelectExpr(Cond.Val->getOperand(0));
646 BuildMI(BB, X86::TEST8rr, 2).addReg(CondR).addReg(CondR);
647 BuildMI(BB, X86::JE, 1).addMBB(Dest);
651 if (Cond.getOpcode() != ISD::SETCC)
652 return true; // Can only handle simple setcc's so far.
653 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
657 // Handle integer conditions first.
658 if (MVT::isInteger(Cond.getOperand(0).getValueType())) {
660 default: assert(0 && "Illegal integer SetCC!");
661 case ISD::SETEQ: Opc = X86::JE; break;
662 case ISD::SETGT: Opc = X86::JG; break;
663 case ISD::SETGE: Opc = X86::JGE; break;
664 case ISD::SETLT: Opc = X86::JL; break;
665 case ISD::SETLE: Opc = X86::JLE; break;
666 case ISD::SETNE: Opc = X86::JNE; break;
667 case ISD::SETULT: Opc = X86::JB; break;
668 case ISD::SETUGT: Opc = X86::JA; break;
669 case ISD::SETULE: Opc = X86::JBE; break;
670 case ISD::SETUGE: Opc = X86::JAE; break;
673 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.hasOneUse());
674 BuildMI(BB, Opc, 1).addMBB(Dest);
678 unsigned Opc2 = 0; // Second branch if needed.
680 // On a floating point condition, the flags are set as follows:
684 // 1 | 0 | 0 | X == Y
685 // 1 | 1 | 1 | unordered
688 default: assert(0 && "Invalid FP setcc!");
690 case ISD::SETEQ: Opc = X86::JE; break; // True if ZF = 1
692 case ISD::SETGT: Opc = X86::JA; break; // True if CF = 0 and ZF = 0
694 case ISD::SETGE: Opc = X86::JAE; break; // True if CF = 0
696 case ISD::SETLT: Opc = X86::JB; break; // True if CF = 1
698 case ISD::SETLE: Opc = X86::JBE; break; // True if CF = 1 or ZF = 1
700 case ISD::SETNE: Opc = X86::JNE; break; // True if ZF = 0
701 case ISD::SETUO: Opc = X86::JP; break; // True if PF = 1
702 case ISD::SETO: Opc = X86::JNP; break; // True if PF = 0
703 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
704 Opc = X86::JA; // ZF = 0 & CF = 0
705 Opc2 = X86::JP; // PF = 1
707 case ISD::SETUGE: // PF = 1 | CF = 0
708 Opc = X86::JAE; // CF = 0
709 Opc2 = X86::JP; // PF = 1
711 case ISD::SETUNE: // PF = 1 | ZF = 0
712 Opc = X86::JNE; // ZF = 0
713 Opc2 = X86::JP; // PF = 1
715 case ISD::SETOEQ: // PF = 0 & ZF = 1
718 return true; // FIXME: Emit more efficient code for this branch.
719 case ISD::SETOLT: // PF = 0 & CF = 1
722 return true; // FIXME: Emit more efficient code for this branch.
723 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
726 return true; // FIXME: Emit more efficient code for this branch.
730 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.hasOneUse());
731 BuildMI(BB, Opc, 1).addMBB(Dest);
733 BuildMI(BB, Opc2, 1).addMBB(Dest);
737 /// EmitSelectCC - Emit code into BB that performs a select operation between
738 /// the two registers RTrue and RFalse, generating a result into RDest.
740 void ISel::EmitSelectCC(SDOperand Cond, SDOperand True, SDOperand False,
741 MVT::ValueType SVT, unsigned RDest) {
742 unsigned RTrue, RFalse;
744 EQ, NE, LT, LE, GT, GE, B, BE, A, AE, P, NP,
746 } CondCode = NOT_SET;
748 static const unsigned CMOVTAB16[] = {
749 X86::CMOVE16rr, X86::CMOVNE16rr, X86::CMOVL16rr, X86::CMOVLE16rr,
750 X86::CMOVG16rr, X86::CMOVGE16rr, X86::CMOVB16rr, X86::CMOVBE16rr,
751 X86::CMOVA16rr, X86::CMOVAE16rr, X86::CMOVP16rr, X86::CMOVNP16rr,
753 static const unsigned CMOVTAB32[] = {
754 X86::CMOVE32rr, X86::CMOVNE32rr, X86::CMOVL32rr, X86::CMOVLE32rr,
755 X86::CMOVG32rr, X86::CMOVGE32rr, X86::CMOVB32rr, X86::CMOVBE32rr,
756 X86::CMOVA32rr, X86::CMOVAE32rr, X86::CMOVP32rr, X86::CMOVNP32rr,
758 static const unsigned CMOVTABFP[] = {
759 X86::FpCMOVE, X86::FpCMOVNE, /*missing*/0, /*missing*/0,
760 /*missing*/0, /*missing*/ 0, X86::FpCMOVB, X86::FpCMOVBE,
761 X86::FpCMOVA, X86::FpCMOVAE, X86::FpCMOVP, X86::FpCMOVNP
763 static const int SSE_CMOVTAB[] = {
764 /*CMPEQ*/ 0, /*CMPNEQ*/ 4, /*missing*/ 0, /*missing*/ 0,
765 /*missing*/ 0, /*missing*/ 0, /*CMPLT*/ 1, /*CMPLE*/ 2,
766 /*CMPNLE*/ 6, /*CMPNLT*/ 5, /*CMPUNORD*/ 3, /*CMPORD*/ 7
769 if (Cond.getOpcode() == ISD::SETCC) {
770 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
771 if (MVT::isInteger(Cond.getOperand(0).getValueType())) {
773 default: assert(0 && "Unknown integer comparison!");
774 case ISD::SETEQ: CondCode = EQ; break;
775 case ISD::SETGT: CondCode = GT; break;
776 case ISD::SETGE: CondCode = GE; break;
777 case ISD::SETLT: CondCode = LT; break;
778 case ISD::SETLE: CondCode = LE; break;
779 case ISD::SETNE: CondCode = NE; break;
780 case ISD::SETULT: CondCode = B; break;
781 case ISD::SETUGT: CondCode = A; break;
782 case ISD::SETULE: CondCode = BE; break;
783 case ISD::SETUGE: CondCode = AE; break;
786 // On a floating point condition, the flags are set as follows:
790 // 1 | 0 | 0 | X == Y
791 // 1 | 1 | 1 | unordered
794 default: assert(0 && "Unknown FP comparison!");
796 case ISD::SETEQ: CondCode = EQ; break; // True if ZF = 1
798 case ISD::SETGT: CondCode = A; break; // True if CF = 0 and ZF = 0
800 case ISD::SETGE: CondCode = AE; break; // True if CF = 0
802 case ISD::SETLT: CondCode = B; break; // True if CF = 1
804 case ISD::SETLE: CondCode = BE; break; // True if CF = 1 or ZF = 1
806 case ISD::SETNE: CondCode = NE; break; // True if ZF = 0
807 case ISD::SETUO: CondCode = P; break; // True if PF = 1
808 case ISD::SETO: CondCode = NP; break; // True if PF = 0
809 case ISD::SETUGT: // PF = 1 | (ZF = 0 & CF = 0)
810 case ISD::SETUGE: // PF = 1 | CF = 0
811 case ISD::SETUNE: // PF = 1 | ZF = 0
812 case ISD::SETOEQ: // PF = 0 & ZF = 1
813 case ISD::SETOLT: // PF = 0 & CF = 1
814 case ISD::SETOLE: // PF = 0 & (CF = 1 || ZF = 1)
815 // We cannot emit this comparison as a single cmov.
821 // There's no SSE equivalent of FCMOVE. For cases where we set a condition
822 // code above and one of the results of the select is +0.0, then we can fake
823 // it up through a clever AND with mask. Otherwise, we will fall through to
824 // the code below that will use a PHI node to select the right value.
825 if (X86ScalarSSE && (SVT == MVT::f32 || SVT == MVT::f64)) {
826 if (Cond.getOperand(0).getValueType() == SVT &&
827 NOT_SET != CondCode) {
828 ConstantFPSDNode *CT = dyn_cast<ConstantFPSDNode>(True);
829 ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(False);
830 bool TrueZero = CT && CT->isExactlyValue(0.0);
831 bool FalseZero = CF && CF->isExactlyValue(0.0);
832 if (TrueZero || FalseZero) {
833 SDOperand LHS = Cond.getOperand(0);
834 SDOperand RHS = Cond.getOperand(1);
836 // Select the two halves of the condition
838 if (getRegPressure(LHS) > getRegPressure(RHS)) {
839 RLHS = SelectExpr(LHS);
840 RRHS = SelectExpr(RHS);
842 RRHS = SelectExpr(RHS);
843 RLHS = SelectExpr(LHS);
846 // Emit the comparison and generate a mask from it
847 unsigned MaskReg = MakeReg(SVT);
848 unsigned Opc = (SVT == MVT::f32) ? X86::CMPSSrr : X86::CMPSDrr;
849 BuildMI(BB, Opc, 3, MaskReg).addReg(RLHS).addReg(RRHS)
850 .addImm(SSE_CMOVTAB[CondCode]);
853 RFalse = SelectExpr(False);
854 Opc = (SVT == MVT::f32) ? X86::ANDNPSrr : X86::ANDNPDrr;
855 BuildMI(BB, Opc, 2, RDest).addReg(MaskReg).addReg(RFalse);
857 RTrue = SelectExpr(True);
858 Opc = (SVT == MVT::f32) ? X86::ANDPSrr : X86::ANDPDrr;
859 BuildMI(BB, Opc, 2, RDest).addReg(MaskReg).addReg(RTrue);
867 // Select the true and false values for use in both the SSE PHI case, and the
868 // integer or x87 cmov cases below.
869 if (getRegPressure(True) > getRegPressure(False)) {
870 RTrue = SelectExpr(True);
871 RFalse = SelectExpr(False);
873 RFalse = SelectExpr(False);
874 RTrue = SelectExpr(True);
877 // Since there's no SSE equivalent of FCMOVE, and we couldn't generate an
878 // AND with mask, we'll have to do the normal RISC thing and generate a PHI
879 // node to select between the true and false values.
880 if (X86ScalarSSE && (SVT == MVT::f32 || SVT == MVT::f64)) {
881 // FIXME: emit a direct compare and branch rather than setting a cond reg
883 unsigned CondReg = SelectExpr(Cond);
884 BuildMI(BB, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
886 // Create an iterator with which to insert the MBB for copying the false
887 // value and the MBB to hold the PHI instruction for this SetCC.
888 MachineBasicBlock *thisMBB = BB;
889 const BasicBlock *LLVM_BB = BB->getBasicBlock();
890 ilist<MachineBasicBlock>::iterator It = BB;
898 // fallthrough --> copy0MBB
899 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
900 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
901 BuildMI(BB, X86::JNE, 1).addMBB(sinkMBB);
902 MachineFunction *F = BB->getParent();
903 F->getBasicBlockList().insert(It, copy0MBB);
904 F->getBasicBlockList().insert(It, sinkMBB);
905 // Update machine-CFG edges
906 BB->addSuccessor(copy0MBB);
907 BB->addSuccessor(sinkMBB);
911 // # fallthrough to sinkMBB
913 // Update machine-CFG edges
914 BB->addSuccessor(sinkMBB);
917 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
920 BuildMI(BB, X86::PHI, 4, RDest).addReg(RFalse)
921 .addMBB(copy0MBB).addReg(RTrue).addMBB(thisMBB);
926 if (CondCode != NOT_SET) {
928 default: assert(0 && "Cannot select this type!");
929 case MVT::i16: Opc = CMOVTAB16[CondCode]; break;
930 case MVT::i32: Opc = CMOVTAB32[CondCode]; break;
931 case MVT::f64: Opc = CMOVTABFP[CondCode]; break;
935 // Finally, if we weren't able to fold this, just emit the condition and test
937 if (CondCode == NOT_SET || Opc == 0) {
938 // Get the condition into the zero flag.
939 unsigned CondReg = SelectExpr(Cond);
940 BuildMI(BB, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
943 default: assert(0 && "Cannot select this type!");
944 case MVT::i16: Opc = X86::CMOVE16rr; break;
945 case MVT::i32: Opc = X86::CMOVE32rr; break;
946 case MVT::f64: Opc = X86::FpCMOVE; break;
949 // FIXME: CMP R, 0 -> TEST R, R
950 EmitCMP(Cond.getOperand(0), Cond.getOperand(1), Cond.Val->hasOneUse());
951 std::swap(RTrue, RFalse);
953 BuildMI(BB, Opc, 2, RDest).addReg(RTrue).addReg(RFalse);
956 void ISel::EmitCMP(SDOperand LHS, SDOperand RHS, bool HasOneUse) {
958 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(RHS)) {
960 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
961 switch (RHS.getValueType()) {
964 case MVT::i8: Opc = X86::CMP8mi; break;
965 case MVT::i16: Opc = X86::CMP16mi; break;
966 case MVT::i32: Opc = X86::CMP32mi; break;
970 EmitFoldedLoad(LHS, AM);
971 addFullAddress(BuildMI(BB, Opc, 5), AM).addImm(CN->getValue());
976 switch (RHS.getValueType()) {
979 case MVT::i8: Opc = X86::CMP8ri; break;
980 case MVT::i16: Opc = X86::CMP16ri; break;
981 case MVT::i32: Opc = X86::CMP32ri; break;
984 unsigned Tmp1 = SelectExpr(LHS);
985 BuildMI(BB, Opc, 2).addReg(Tmp1).addImm(CN->getValue());
988 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(RHS)) {
989 if (!X86ScalarSSE && (CN->isExactlyValue(+0.0) ||
990 CN->isExactlyValue(-0.0))) {
991 unsigned Reg = SelectExpr(LHS);
992 BuildMI(BB, X86::FpTST, 1).addReg(Reg);
993 BuildMI(BB, X86::FNSTSW8r, 0);
994 BuildMI(BB, X86::SAHF, 1);
1000 if (HasOneUse && isFoldableLoad(LHS, RHS)) {
1001 switch (RHS.getValueType()) {
1004 case MVT::i8: Opc = X86::CMP8mr; break;
1005 case MVT::i16: Opc = X86::CMP16mr; break;
1006 case MVT::i32: Opc = X86::CMP32mr; break;
1010 EmitFoldedLoad(LHS, AM);
1011 unsigned Reg = SelectExpr(RHS);
1012 addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(Reg);
1017 switch (LHS.getValueType()) {
1018 default: assert(0 && "Cannot compare this value!");
1020 case MVT::i8: Opc = X86::CMP8rr; break;
1021 case MVT::i16: Opc = X86::CMP16rr; break;
1022 case MVT::i32: Opc = X86::CMP32rr; break;
1023 case MVT::f32: Opc = X86::UCOMISSrr; break;
1024 case MVT::f64: Opc = X86ScalarSSE ? X86::UCOMISDrr : X86::FpUCOMIr; break;
1026 unsigned Tmp1, Tmp2;
1027 if (getRegPressure(LHS) > getRegPressure(RHS)) {
1028 Tmp1 = SelectExpr(LHS);
1029 Tmp2 = SelectExpr(RHS);
1031 Tmp2 = SelectExpr(RHS);
1032 Tmp1 = SelectExpr(LHS);
1034 BuildMI(BB, Opc, 2).addReg(Tmp1).addReg(Tmp2);
1037 /// isFoldableLoad - Return true if this is a load instruction that can safely
1038 /// be folded into an operation that uses it.
1039 bool ISel::isFoldableLoad(SDOperand Op, SDOperand OtherOp, bool FloatPromoteOk){
1040 if (Op.getOpcode() == ISD::LOAD) {
1041 // FIXME: currently can't fold constant pool indexes.
1042 if (isa<ConstantPoolSDNode>(Op.getOperand(1)))
1044 } else if (FloatPromoteOk && Op.getOpcode() == ISD::EXTLOAD &&
1045 cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::f32) {
1046 // FIXME: currently can't fold constant pool indexes.
1047 if (isa<ConstantPoolSDNode>(Op.getOperand(1)))
1053 // If this load has already been emitted, we clearly can't fold it.
1054 assert(Op.ResNo == 0 && "Not a use of the value of the load?");
1055 if (ExprMap.count(Op.getValue(1))) return false;
1056 assert(!ExprMap.count(Op.getValue(0)) && "Value in map but not token chain?");
1057 assert(!ExprMap.count(Op.getValue(1))&&"Token lowered but value not in map?");
1059 // If there is not just one use of its value, we cannot fold.
1060 if (!Op.Val->hasNUsesOfValue(1, 0)) return false;
1062 // Finally, we cannot fold the load into the operation if this would induce a
1063 // cycle into the resultant dag. To check for this, see if OtherOp (the other
1064 // operand of the operation we are folding the load into) can possible use the
1065 // chain node defined by the load.
1066 if (OtherOp.Val && !Op.Val->hasNUsesOfValue(0, 1)) { // Has uses of chain?
1067 std::set<SDNode*> Visited;
1068 if (NodeTransitivelyUsesValue(OtherOp, Op.getValue(1), Visited))
1075 /// EmitFoldedLoad - Ensure that the arguments of the load are code generated,
1076 /// and compute the address being loaded into AM.
1077 void ISel::EmitFoldedLoad(SDOperand Op, X86AddressMode &AM) {
1078 SDOperand Chain = Op.getOperand(0);
1079 SDOperand Address = Op.getOperand(1);
1081 if (getRegPressure(Chain) > getRegPressure(Address)) {
1083 SelectAddress(Address, AM);
1085 SelectAddress(Address, AM);
1089 // The chain for this load is now lowered.
1090 assert(ExprMap.count(SDOperand(Op.Val, 1)) == 0 &&
1091 "Load emitted more than once?");
1092 if (!ExprMap.insert(std::make_pair(Op.getValue(1), 1)).second)
1093 assert(0 && "Load emitted more than once!");
1096 // EmitOrOpOp - Pattern match the expression (Op1|Op2), where we know that op1
1097 // and op2 are i8/i16/i32 values with one use each (the or). If we can form a
1098 // SHLD or SHRD, emit the instruction (generating the value into DestReg) and
1100 bool ISel::EmitOrOpOp(SDOperand Op1, SDOperand Op2, unsigned DestReg) {
1101 if (Op1.getOpcode() == ISD::SHL && Op2.getOpcode() == ISD::SRL) {
1103 } else if (Op2.getOpcode() == ISD::SHL && Op1.getOpcode() == ISD::SRL) {
1104 std::swap(Op1, Op2); // Op1 is the SHL now.
1106 return false; // No match
1109 SDOperand ShlVal = Op1.getOperand(0);
1110 SDOperand ShlAmt = Op1.getOperand(1);
1111 SDOperand ShrVal = Op2.getOperand(0);
1112 SDOperand ShrAmt = Op2.getOperand(1);
1114 unsigned RegSize = MVT::getSizeInBits(Op1.getValueType());
1116 // Find out if ShrAmt = 32-ShlAmt or ShlAmt = 32-ShrAmt.
1117 if (ShlAmt.getOpcode() == ISD::SUB && ShlAmt.getOperand(1) == ShrAmt)
1118 if (ConstantSDNode *SubCST = dyn_cast<ConstantSDNode>(ShlAmt.getOperand(0)))
1119 if (SubCST->getValue() == RegSize) {
1120 // (A >> ShrAmt) | (A << (32-ShrAmt)) ==> ROR A, ShrAmt
1121 // (A >> ShrAmt) | (B << (32-ShrAmt)) ==> SHRD A, B, ShrAmt
1122 if (ShrVal == ShlVal) {
1123 unsigned Reg, ShAmt;
1124 if (getRegPressure(ShrVal) > getRegPressure(ShrAmt)) {
1125 Reg = SelectExpr(ShrVal);
1126 ShAmt = SelectExpr(ShrAmt);
1128 ShAmt = SelectExpr(ShrAmt);
1129 Reg = SelectExpr(ShrVal);
1131 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
1132 unsigned Opc = RegSize == 8 ? X86::ROR8rCL :
1133 (RegSize == 16 ? X86::ROR16rCL : X86::ROR32rCL);
1134 BuildMI(BB, Opc, 1, DestReg).addReg(Reg);
1136 } else if (RegSize != 8) {
1137 unsigned AReg, BReg;
1138 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
1139 BReg = SelectExpr(ShlVal);
1140 AReg = SelectExpr(ShrVal);
1142 AReg = SelectExpr(ShrVal);
1143 BReg = SelectExpr(ShlVal);
1145 unsigned ShAmt = SelectExpr(ShrAmt);
1146 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
1147 unsigned Opc = RegSize == 16 ? X86::SHRD16rrCL : X86::SHRD32rrCL;
1148 BuildMI(BB, Opc, 2, DestReg).addReg(AReg).addReg(BReg);
1153 if (ShrAmt.getOpcode() == ISD::SUB && ShrAmt.getOperand(1) == ShlAmt)
1154 if (ConstantSDNode *SubCST = dyn_cast<ConstantSDNode>(ShrAmt.getOperand(0)))
1155 if (SubCST->getValue() == RegSize) {
1156 // (A << ShlAmt) | (A >> (32-ShlAmt)) ==> ROL A, ShrAmt
1157 // (A << ShlAmt) | (B >> (32-ShlAmt)) ==> SHLD A, B, ShrAmt
1158 if (ShrVal == ShlVal) {
1159 unsigned Reg, ShAmt;
1160 if (getRegPressure(ShrVal) > getRegPressure(ShlAmt)) {
1161 Reg = SelectExpr(ShrVal);
1162 ShAmt = SelectExpr(ShlAmt);
1164 ShAmt = SelectExpr(ShlAmt);
1165 Reg = SelectExpr(ShrVal);
1167 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
1168 unsigned Opc = RegSize == 8 ? X86::ROL8rCL :
1169 (RegSize == 16 ? X86::ROL16rCL : X86::ROL32rCL);
1170 BuildMI(BB, Opc, 1, DestReg).addReg(Reg);
1172 } else if (RegSize != 8) {
1173 unsigned AReg, BReg;
1174 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
1175 AReg = SelectExpr(ShlVal);
1176 BReg = SelectExpr(ShrVal);
1178 BReg = SelectExpr(ShrVal);
1179 AReg = SelectExpr(ShlVal);
1181 unsigned ShAmt = SelectExpr(ShlAmt);
1182 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShAmt);
1183 unsigned Opc = RegSize == 16 ? X86::SHLD16rrCL : X86::SHLD32rrCL;
1184 BuildMI(BB, Opc, 2, DestReg).addReg(AReg).addReg(BReg);
1189 if (ConstantSDNode *ShrCst = dyn_cast<ConstantSDNode>(ShrAmt))
1190 if (ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(ShlAmt))
1191 if (ShrCst->getValue() < RegSize && ShlCst->getValue() < RegSize)
1192 if (ShrCst->getValue() == RegSize-ShlCst->getValue()) {
1193 // (A >> 5) | (A << 27) --> ROR A, 5
1194 // (A >> 5) | (B << 27) --> SHRD A, B, 5
1195 if (ShrVal == ShlVal) {
1196 unsigned Reg = SelectExpr(ShrVal);
1197 unsigned Opc = RegSize == 8 ? X86::ROR8ri :
1198 (RegSize == 16 ? X86::ROR16ri : X86::ROR32ri);
1199 BuildMI(BB, Opc, 2, DestReg).addReg(Reg).addImm(ShrCst->getValue());
1201 } else if (RegSize != 8) {
1202 unsigned AReg, BReg;
1203 if (getRegPressure(ShlVal) > getRegPressure(ShrVal)) {
1204 BReg = SelectExpr(ShlVal);
1205 AReg = SelectExpr(ShrVal);
1207 AReg = SelectExpr(ShrVal);
1208 BReg = SelectExpr(ShlVal);
1210 unsigned Opc = RegSize == 16 ? X86::SHRD16rri8 : X86::SHRD32rri8;
1211 BuildMI(BB, Opc, 3, DestReg).addReg(AReg).addReg(BReg)
1212 .addImm(ShrCst->getValue());
1220 unsigned ISel::SelectExpr(SDOperand N) {
1222 unsigned Tmp1 = 0, Tmp2 = 0, Tmp3 = 0, Opc = 0;
1223 SDNode *Node = N.Val;
1226 if (Node->getOpcode() == ISD::CopyFromReg ||
1227 Node->getOpcode() == ISD::Register) {
1228 unsigned Reg = (Node->getOpcode() == ISD::CopyFromReg) ?
1229 cast<RegisterSDNode>(Node->getOperand(1))->getReg() :
1230 cast<RegisterSDNode>(Node)->getReg();
1231 // Just use the specified register as our input if we can.
1232 if (MRegisterInfo::isVirtualRegister(Reg) || Reg == X86::ESP)
1236 unsigned &Reg = ExprMap[N];
1237 if (Reg) return Reg;
1239 switch (N.getOpcode()) {
1241 Reg = Result = (N.getValueType() != MVT::Other) ?
1242 MakeReg(N.getValueType()) : 1;
1244 case X86ISD::TAILCALL:
1246 // If this is a call instruction, make sure to prepare ALL of the result
1247 // values as well as the chain.
1248 ExprMap[N.getValue(0)] = 1;
1249 if (Node->getNumValues() > 1) {
1250 Result = MakeReg(Node->getValueType(1));
1251 ExprMap[N.getValue(1)] = Result;
1252 for (unsigned i = 2, e = Node->getNumValues(); i != e; ++i)
1253 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1258 case ISD::ADD_PARTS:
1259 case ISD::SUB_PARTS:
1260 case ISD::SHL_PARTS:
1261 case ISD::SRL_PARTS:
1262 case ISD::SRA_PARTS:
1263 Result = MakeReg(Node->getValueType(0));
1264 ExprMap[N.getValue(0)] = Result;
1265 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
1266 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
1270 switch (N.getOpcode()) {
1273 assert(0 && "Node not handled!\n");
1274 case ISD::FP_EXTEND:
1275 assert(X86ScalarSSE && "Scalar SSE FP must be enabled to use f32");
1276 Tmp1 = SelectExpr(N.getOperand(0));
1277 BuildMI(BB, X86::CVTSS2SDrr, 1, Result).addReg(Tmp1);
1280 assert(X86ScalarSSE && "Scalar SSE FP must be enabled to use f32");
1281 Tmp1 = SelectExpr(N.getOperand(0));
1282 BuildMI(BB, X86::CVTSD2SSrr, 1, Result).addReg(Tmp1);
1284 case ISD::CopyFromReg:
1285 Select(N.getOperand(0));
1287 Reg = Result = ExprMap[N.getValue(0)] =
1288 MakeReg(N.getValue(0).getValueType());
1290 Tmp1 = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
1291 switch (Node->getValueType(0)) {
1292 default: assert(0 && "Cannot CopyFromReg this!");
1295 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(Tmp1);
1298 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(Tmp1);
1301 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(Tmp1);
1305 case ISD::FrameIndex:
1306 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
1307 addFrameReference(BuildMI(BB, X86::LEA32r, 4, Result), (int)Tmp1);
1309 case ISD::ConstantPool:
1310 Tmp1 = BB->getParent()->getConstantPool()->
1311 getConstantPoolIndex(cast<ConstantPoolSDNode>(N)->get());
1312 addConstantPoolReference(BuildMI(BB, X86::LEA32r, 4, Result), Tmp1);
1314 case ISD::ConstantFP:
1316 assert(cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) &&
1317 "SSE only supports +0.0");
1318 Opc = (N.getValueType() == MVT::f32) ? X86::FLD0SS : X86::FLD0SD;
1319 BuildMI(BB, Opc, 0, Result);
1322 ContainsFPCode = true;
1323 Tmp1 = Result; // Intermediate Register
1324 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
1325 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
1326 Tmp1 = MakeReg(MVT::f64);
1328 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
1329 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
1330 BuildMI(BB, X86::FpLD0, 0, Tmp1);
1331 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
1332 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
1333 BuildMI(BB, X86::FpLD1, 0, Tmp1);
1335 assert(0 && "Unexpected constant!");
1337 BuildMI(BB, X86::FpCHS, 1, Result).addReg(Tmp1);
1340 switch (N.getValueType()) {
1341 default: assert(0 && "Cannot use constants of this type!");
1343 case MVT::i8: Opc = X86::MOV8ri; break;
1344 case MVT::i16: Opc = X86::MOV16ri; break;
1345 case MVT::i32: Opc = X86::MOV32ri; break;
1347 BuildMI(BB, Opc, 1,Result).addImm(cast<ConstantSDNode>(N)->getValue());
1350 if (Node->getValueType(0) == MVT::f64) {
1351 // FIXME: SHOULD TEACH STACKIFIER ABOUT UNDEF VALUES!
1352 BuildMI(BB, X86::FpLD0, 0, Result);
1354 BuildMI(BB, X86::IMPLICIT_DEF, 0, Result);
1357 case ISD::GlobalAddress: {
1358 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1359 // For Darwin, external and weak symbols are indirect, so we want to load
1360 // the value at address GV, not the value of GV itself.
1361 if (Subtarget->getIndirectExternAndWeakGlobals() &&
1362 (GV->hasWeakLinkage() || GV->isExternal())) {
1363 BuildMI(BB, X86::MOV32rm, 4, Result).addReg(0).addZImm(1).addReg(0)
1364 .addGlobalAddress(GV, false, 0);
1366 BuildMI(BB, X86::MOV32ri, 1, Result).addGlobalAddress(GV);
1370 case ISD::ExternalSymbol: {
1371 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1372 BuildMI(BB, X86::MOV32ri, 1, Result).addExternalSymbol(Sym);
1375 case ISD::ANY_EXTEND: // treat any extend like zext
1376 case ISD::ZERO_EXTEND: {
1377 int DestIs16 = N.getValueType() == MVT::i16;
1378 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
1380 // FIXME: This hack is here for zero extension casts from bool to i8. This
1381 // would not be needed if bools were promoted by Legalize.
1382 if (N.getValueType() == MVT::i8) {
1383 Tmp1 = SelectExpr(N.getOperand(0));
1384 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(Tmp1);
1388 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
1389 static const unsigned Opc[3] = {
1390 X86::MOVZX32rm8, X86::MOVZX32rm16, X86::MOVZX16rm8
1394 EmitFoldedLoad(N.getOperand(0), AM);
1395 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
1400 static const unsigned Opc[3] = {
1401 X86::MOVZX32rr8, X86::MOVZX32rr16, X86::MOVZX16rr8
1403 Tmp1 = SelectExpr(N.getOperand(0));
1404 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
1407 case ISD::SIGN_EXTEND: {
1408 int DestIs16 = N.getValueType() == MVT::i16;
1409 int SrcIs16 = N.getOperand(0).getValueType() == MVT::i16;
1411 // FIXME: Legalize should promote bools to i8!
1412 assert(N.getOperand(0).getValueType() != MVT::i1 &&
1413 "Sign extend from bool not implemented!");
1415 if (isFoldableLoad(N.getOperand(0), SDOperand())) {
1416 static const unsigned Opc[3] = {
1417 X86::MOVSX32rm8, X86::MOVSX32rm16, X86::MOVSX16rm8
1421 EmitFoldedLoad(N.getOperand(0), AM);
1422 addFullAddress(BuildMI(BB, Opc[SrcIs16+DestIs16*2], 4, Result), AM);
1426 static const unsigned Opc[3] = {
1427 X86::MOVSX32rr8, X86::MOVSX32rr16, X86::MOVSX16rr8
1429 Tmp1 = SelectExpr(N.getOperand(0));
1430 BuildMI(BB, Opc[SrcIs16+DestIs16*2], 1, Result).addReg(Tmp1);
1434 // Handle cast of LARGER int to SMALLER int using a move to EAX followed by
1435 // a move out of AX or AL.
1436 switch (N.getOperand(0).getValueType()) {
1437 default: assert(0 && "Unknown truncate!");
1438 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
1439 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
1440 case MVT::i32: Tmp2 = X86::EAX; Opc = X86::MOV32rr; break;
1442 Tmp1 = SelectExpr(N.getOperand(0));
1443 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
1445 switch (N.getValueType()) {
1446 default: assert(0 && "Unknown truncate!");
1448 case MVT::i8: Tmp2 = X86::AL; Opc = X86::MOV8rr; break;
1449 case MVT::i16: Tmp2 = X86::AX; Opc = X86::MOV16rr; break;
1451 BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
1454 case ISD::SINT_TO_FP: {
1455 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
1456 unsigned PromoteOpcode = 0;
1458 // We can handle any sint to fp with the direct sse conversion instructions.
1460 Opc = (N.getValueType() == MVT::f64) ? X86::CVTSI2SDrr : X86::CVTSI2SSrr;
1461 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1465 ContainsFPCode = true;
1467 // Spill the integer to memory and reload it from there.
1468 MVT::ValueType SrcTy = N.getOperand(0).getValueType();
1469 unsigned Size = MVT::getSizeInBits(SrcTy)/8;
1470 MachineFunction *F = BB->getParent();
1471 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
1475 addFrameReference(BuildMI(BB, X86::MOV32mr, 5), FrameIdx).addReg(Tmp1);
1476 addFrameReference(BuildMI(BB, X86::FpILD32m, 5, Result), FrameIdx);
1479 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), FrameIdx).addReg(Tmp1);
1480 addFrameReference(BuildMI(BB, X86::FpILD16m, 5, Result), FrameIdx);
1482 default: break; // No promotion required.
1486 case ISD::FP_TO_SINT:
1487 Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
1489 // If the target supports SSE2 and is performing FP operations in SSE regs
1490 // instead of the FP stack, then we can use the efficient CVTSS2SI and
1491 // CVTSD2SI instructions.
1492 assert(X86ScalarSSE);
1493 if (MVT::f32 == N.getOperand(0).getValueType()) {
1494 BuildMI(BB, X86::CVTTSS2SIrr, 1, Result).addReg(Tmp1);
1495 } else if (MVT::f64 == N.getOperand(0).getValueType()) {
1496 BuildMI(BB, X86::CVTTSD2SIrr, 1, Result).addReg(Tmp1);
1498 assert(0 && "Not an f32 or f64?");
1505 Op0 = N.getOperand(0);
1506 Op1 = N.getOperand(1);
1508 if (isFoldableLoad(Op0, Op1, true)) {
1509 std::swap(Op0, Op1);
1513 if (isFoldableLoad(Op1, Op0, true)) {
1515 switch (N.getValueType()) {
1516 default: assert(0 && "Cannot add this type!");
1518 case MVT::i8: Opc = X86::ADD8rm; break;
1519 case MVT::i16: Opc = X86::ADD16rm; break;
1520 case MVT::i32: Opc = X86::ADD32rm; break;
1521 case MVT::f32: Opc = X86::ADDSSrm; break;
1523 // For F64, handle promoted load operations (from F32) as well!
1525 assert(Op1.getOpcode() == ISD::LOAD && "SSE load not promoted");
1528 Opc = Op1.getOpcode() == ISD::LOAD ? X86::FpADD64m : X86::FpADD32m;
1533 EmitFoldedLoad(Op1, AM);
1534 Tmp1 = SelectExpr(Op0);
1535 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1539 // See if we can codegen this as an LEA to fold operations together.
1540 if (N.getValueType() == MVT::i32) {
1542 X86ISelAddressMode AM;
1543 MatchAddress(N, AM);
1544 ExprMap[N] = Result;
1546 // If this is not just an add, emit the LEA. For a simple add (like
1547 // reg+reg or reg+imm), we just emit an add. It might be a good idea to
1548 // leave this as LEA, then peephole it to 'ADD' after two address elim
1550 if (AM.Scale != 1 || AM.BaseType == X86ISelAddressMode::FrameIndexBase||
1551 AM.GV || (AM.Base.Reg.Val && AM.IndexReg.Val && AM.Disp)) {
1552 X86AddressMode XAM = SelectAddrExprs(AM);
1553 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), XAM);
1558 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
1560 if (CN->getValue() == 1) { // add X, 1 -> inc X
1561 switch (N.getValueType()) {
1562 default: assert(0 && "Cannot integer add this type!");
1563 case MVT::i8: Opc = X86::INC8r; break;
1564 case MVT::i16: Opc = X86::INC16r; break;
1565 case MVT::i32: Opc = X86::INC32r; break;
1567 } else if (CN->isAllOnesValue()) { // add X, -1 -> dec X
1568 switch (N.getValueType()) {
1569 default: assert(0 && "Cannot integer add this type!");
1570 case MVT::i8: Opc = X86::DEC8r; break;
1571 case MVT::i16: Opc = X86::DEC16r; break;
1572 case MVT::i32: Opc = X86::DEC32r; break;
1577 Tmp1 = SelectExpr(Op0);
1578 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1582 switch (N.getValueType()) {
1583 default: assert(0 && "Cannot add this type!");
1584 case MVT::i8: Opc = X86::ADD8ri; break;
1585 case MVT::i16: Opc = X86::ADD16ri; break;
1586 case MVT::i32: Opc = X86::ADD32ri; break;
1589 Tmp1 = SelectExpr(Op0);
1590 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
1595 switch (N.getValueType()) {
1596 default: assert(0 && "Cannot add this type!");
1597 case MVT::i8: Opc = X86::ADD8rr; break;
1598 case MVT::i16: Opc = X86::ADD16rr; break;
1599 case MVT::i32: Opc = X86::ADD32rr; break;
1600 case MVT::f32: Opc = X86::ADDSSrr; break;
1601 case MVT::f64: Opc = X86ScalarSSE ? X86::ADDSDrr : X86::FpADD; break;
1604 if (getRegPressure(Op0) > getRegPressure(Op1)) {
1605 Tmp1 = SelectExpr(Op0);
1606 Tmp2 = SelectExpr(Op1);
1608 Tmp2 = SelectExpr(Op1);
1609 Tmp1 = SelectExpr(Op0);
1612 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1616 Tmp1 = SelectExpr(Node->getOperand(0));
1618 Opc = (N.getValueType() == MVT::f32) ? X86::SQRTSSrr : X86::SQRTSDrr;
1619 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1621 BuildMI(BB, X86::FpSQRT, 1, Result).addReg(Tmp1);
1626 // Once we can spill 16 byte constants into the constant pool, we can
1627 // implement SSE equivalents of FABS and FCHS.
1632 assert(N.getValueType()==MVT::f64 && "Illegal type for this operation");
1633 Tmp1 = SelectExpr(Node->getOperand(0));
1634 switch (N.getOpcode()) {
1635 default: assert(0 && "Unreachable!");
1636 case ISD::FABS: BuildMI(BB, X86::FpABS, 1, Result).addReg(Tmp1); break;
1637 case ISD::FNEG: BuildMI(BB, X86::FpCHS, 1, Result).addReg(Tmp1); break;
1638 case ISD::FSIN: BuildMI(BB, X86::FpSIN, 1, Result).addReg(Tmp1); break;
1639 case ISD::FCOS: BuildMI(BB, X86::FpCOS, 1, Result).addReg(Tmp1); break;
1644 switch (N.getValueType()) {
1645 default: assert(0 && "Unsupported VT!");
1646 case MVT::i8: Tmp2 = X86::MUL8r; break;
1647 case MVT::i16: Tmp2 = X86::MUL16r; break;
1648 case MVT::i32: Tmp2 = X86::MUL32r; break;
1652 unsigned MovOpc, LowReg, HiReg;
1653 switch (N.getValueType()) {
1654 default: assert(0 && "Unsupported VT!");
1656 MovOpc = X86::MOV8rr;
1662 MovOpc = X86::MOV16rr;
1668 MovOpc = X86::MOV32rr;
1674 if (Node->getOpcode() != ISD::MULHS)
1675 Opc = Tmp2; // Get the MULHU opcode.
1677 Op0 = Node->getOperand(0);
1678 Op1 = Node->getOperand(1);
1679 if (getRegPressure(Op0) > getRegPressure(Op1)) {
1680 Tmp1 = SelectExpr(Op0);
1681 Tmp2 = SelectExpr(Op1);
1683 Tmp2 = SelectExpr(Op1);
1684 Tmp1 = SelectExpr(Op0);
1687 // FIXME: Implement folding of loads into the memory operands here!
1688 BuildMI(BB, MovOpc, 1, LowReg).addReg(Tmp1);
1689 BuildMI(BB, Opc, 1).addReg(Tmp2);
1690 BuildMI(BB, MovOpc, 1, Result).addReg(HiReg);
1701 static const unsigned SUBTab[] = {
1702 X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, 0,
1703 X86::SUB8rm, X86::SUB16rm, X86::SUB32rm, X86::FpSUB32m, X86::FpSUB64m,
1704 X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::FpSUB , X86::FpSUB,
1706 static const unsigned SSE_SUBTab[] = {
1707 X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, 0,
1708 X86::SUB8rm, X86::SUB16rm, X86::SUB32rm, X86::SUBSSrm, X86::SUBSDrm,
1709 X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::SUBSSrr, X86::SUBSDrr,
1711 static const unsigned MULTab[] = {
1712 0, X86::IMUL16rri, X86::IMUL32rri, 0, 0,
1713 0, X86::IMUL16rm , X86::IMUL32rm, X86::FpMUL32m, X86::FpMUL64m,
1714 0, X86::IMUL16rr , X86::IMUL32rr, X86::FpMUL , X86::FpMUL,
1716 static const unsigned SSE_MULTab[] = {
1717 0, X86::IMUL16rri, X86::IMUL32rri, 0, 0,
1718 0, X86::IMUL16rm , X86::IMUL32rm, X86::MULSSrm, X86::MULSDrm,
1719 0, X86::IMUL16rr , X86::IMUL32rr, X86::MULSSrr, X86::MULSDrr,
1721 static const unsigned ANDTab[] = {
1722 X86::AND8ri, X86::AND16ri, X86::AND32ri, 0, 0,
1723 X86::AND8rm, X86::AND16rm, X86::AND32rm, 0, 0,
1724 X86::AND8rr, X86::AND16rr, X86::AND32rr, 0, 0,
1726 static const unsigned ORTab[] = {
1727 X86::OR8ri, X86::OR16ri, X86::OR32ri, 0, 0,
1728 X86::OR8rm, X86::OR16rm, X86::OR32rm, 0, 0,
1729 X86::OR8rr, X86::OR16rr, X86::OR32rr, 0, 0,
1731 static const unsigned XORTab[] = {
1732 X86::XOR8ri, X86::XOR16ri, X86::XOR32ri, 0, 0,
1733 X86::XOR8rm, X86::XOR16rm, X86::XOR32rm, 0, 0,
1734 X86::XOR8rr, X86::XOR16rr, X86::XOR32rr, 0, 0,
1737 Op0 = Node->getOperand(0);
1738 Op1 = Node->getOperand(1);
1740 if (Node->getOpcode() == ISD::OR && Op0.hasOneUse() && Op1.hasOneUse())
1741 if (EmitOrOpOp(Op0, Op1, Result)) // Match SHLD, SHRD, and rotates.
1744 if (Node->getOpcode() == ISD::SUB)
1745 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(0)))
1746 if (CN->isNullValue()) { // 0 - N -> neg N
1747 switch (N.getValueType()) {
1748 default: assert(0 && "Cannot sub this type!");
1750 case MVT::i8: Opc = X86::NEG8r; break;
1751 case MVT::i16: Opc = X86::NEG16r; break;
1752 case MVT::i32: Opc = X86::NEG32r; break;
1754 Tmp1 = SelectExpr(N.getOperand(1));
1755 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1759 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
1760 if (CN->isAllOnesValue() && Node->getOpcode() == ISD::XOR) {
1762 switch (N.getValueType()) {
1763 default: assert(0 && "Cannot add this type!");
1764 case MVT::i1: break; // Not supported, don't invert upper bits!
1765 case MVT::i8: Opc = X86::NOT8r; break;
1766 case MVT::i16: Opc = X86::NOT16r; break;
1767 case MVT::i32: Opc = X86::NOT32r; break;
1770 Tmp1 = SelectExpr(Op0);
1771 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1776 // Fold common multiplies into LEA instructions.
1777 if (Node->getOpcode() == ISD::MUL && N.getValueType() == MVT::i32) {
1778 switch ((int)CN->getValue()) {
1783 // Remove N from exprmap so SelectAddress doesn't get confused.
1786 SelectAddress(N, AM);
1787 // Restore it to the map.
1788 ExprMap[N] = Result;
1789 addFullAddress(BuildMI(BB, X86::LEA32r, 4, Result), AM);
1794 switch (N.getValueType()) {
1795 default: assert(0 && "Cannot xor this type!");
1797 case MVT::i8: Opc = 0; break;
1798 case MVT::i16: Opc = 1; break;
1799 case MVT::i32: Opc = 2; break;
1801 switch (Node->getOpcode()) {
1802 default: assert(0 && "Unreachable!");
1804 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
1806 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
1807 case ISD::AND: Opc = ANDTab[Opc]; break;
1808 case ISD::OR: Opc = ORTab[Opc]; break;
1809 case ISD::XOR: Opc = XORTab[Opc]; break;
1811 if (Opc) { // Can't fold MUL:i8 R, imm
1812 Tmp1 = SelectExpr(Op0);
1813 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
1818 if (isFoldableLoad(Op0, Op1, true))
1819 if (Node->getOpcode() != ISD::SUB && Node->getOpcode() != ISD::FSUB) {
1820 std::swap(Op0, Op1);
1823 // For FP, emit 'reverse' subract, with a memory operand.
1824 if (N.getValueType() == MVT::f64 && !X86ScalarSSE) {
1825 if (Op0.getOpcode() == ISD::EXTLOAD)
1826 Opc = X86::FpSUBR32m;
1828 Opc = X86::FpSUBR64m;
1831 EmitFoldedLoad(Op0, AM);
1832 Tmp1 = SelectExpr(Op1);
1833 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1838 if (isFoldableLoad(Op1, Op0, true)) {
1840 switch (N.getValueType()) {
1841 default: assert(0 && "Cannot operate on this type!");
1843 case MVT::i8: Opc = 5; break;
1844 case MVT::i16: Opc = 6; break;
1845 case MVT::i32: Opc = 7; break;
1846 case MVT::f32: Opc = 8; break;
1847 // For F64, handle promoted load operations (from F32) as well!
1849 assert((!X86ScalarSSE || Op1.getOpcode() == ISD::LOAD) &&
1850 "SSE load should have been promoted");
1851 Opc = Op1.getOpcode() == ISD::LOAD ? 9 : 8; break;
1853 switch (Node->getOpcode()) {
1854 default: assert(0 && "Unreachable!");
1856 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
1858 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
1859 case ISD::AND: Opc = ANDTab[Opc]; break;
1860 case ISD::OR: Opc = ORTab[Opc]; break;
1861 case ISD::XOR: Opc = XORTab[Opc]; break;
1865 EmitFoldedLoad(Op1, AM);
1866 Tmp1 = SelectExpr(Op0);
1868 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
1870 assert(Node->getOpcode() == ISD::MUL &&
1871 N.getValueType() == MVT::i8 && "Unexpected situation!");
1872 // Must use the MUL instruction, which forces use of AL.
1873 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
1874 addFullAddress(BuildMI(BB, X86::MUL8m, 1), AM);
1875 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
1880 if (getRegPressure(Op0) > getRegPressure(Op1)) {
1881 Tmp1 = SelectExpr(Op0);
1882 Tmp2 = SelectExpr(Op1);
1884 Tmp2 = SelectExpr(Op1);
1885 Tmp1 = SelectExpr(Op0);
1888 switch (N.getValueType()) {
1889 default: assert(0 && "Cannot add this type!");
1891 case MVT::i8: Opc = 10; break;
1892 case MVT::i16: Opc = 11; break;
1893 case MVT::i32: Opc = 12; break;
1894 case MVT::f32: Opc = 13; break;
1895 case MVT::f64: Opc = 14; break;
1897 switch (Node->getOpcode()) {
1898 default: assert(0 && "Unreachable!");
1900 case ISD::SUB: Opc = X86ScalarSSE ? SSE_SUBTab[Opc] : SUBTab[Opc]; break;
1902 case ISD::MUL: Opc = X86ScalarSSE ? SSE_MULTab[Opc] : MULTab[Opc]; break;
1903 case ISD::AND: Opc = ANDTab[Opc]; break;
1904 case ISD::OR: Opc = ORTab[Opc]; break;
1905 case ISD::XOR: Opc = XORTab[Opc]; break;
1908 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1910 assert(Node->getOpcode() == ISD::MUL &&
1911 N.getValueType() == MVT::i8 && "Unexpected situation!");
1912 // Must use the MUL instruction, which forces use of AL.
1913 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
1914 BuildMI(BB, X86::MUL8r, 1).addReg(Tmp2);
1915 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
1919 case ISD::ADD_PARTS:
1920 case ISD::SUB_PARTS: {
1921 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1922 "Not an i64 add/sub!");
1923 // Emit all of the operands.
1924 std::vector<unsigned> InVals;
1925 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1926 InVals.push_back(SelectExpr(N.getOperand(i)));
1927 if (N.getOpcode() == ISD::ADD_PARTS) {
1928 BuildMI(BB, X86::ADD32rr, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1929 BuildMI(BB, X86::ADC32rr,2,Result+1).addReg(InVals[1]).addReg(InVals[3]);
1931 BuildMI(BB, X86::SUB32rr, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1932 BuildMI(BB, X86::SBB32rr, 2,Result+1).addReg(InVals[1]).addReg(InVals[3]);
1934 return Result+N.ResNo;
1937 case ISD::SHL_PARTS:
1938 case ISD::SRA_PARTS:
1939 case ISD::SRL_PARTS: {
1940 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1941 "Not an i64 shift!");
1942 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1943 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1944 unsigned TmpReg = MakeReg(MVT::i32);
1945 if (N.getOpcode() == ISD::SRA_PARTS) {
1946 // If this is a SHR of a Long, then we need to do funny sign extension
1947 // stuff. TmpReg gets the value to use as the high-part if we are
1948 // shifting more than 32 bits.
1949 BuildMI(BB, X86::SAR32ri, 2, TmpReg).addReg(ShiftOpHi).addImm(31);
1951 // Other shifts use a fixed zero value if the shift is more than 32 bits.
1952 BuildMI(BB, X86::MOV32ri, 1, TmpReg).addImm(0);
1955 // Initialize CL with the shift amount.
1956 unsigned ShiftAmountReg = SelectExpr(N.getOperand(2));
1957 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(ShiftAmountReg);
1959 unsigned TmpReg2 = MakeReg(MVT::i32);
1960 unsigned TmpReg3 = MakeReg(MVT::i32);
1961 if (N.getOpcode() == ISD::SHL_PARTS) {
1962 // TmpReg2 = shld inHi, inLo
1963 BuildMI(BB, X86::SHLD32rrCL, 2,TmpReg2).addReg(ShiftOpHi)
1965 // TmpReg3 = shl inLo, CL
1966 BuildMI(BB, X86::SHL32rCL, 1, TmpReg3).addReg(ShiftOpLo);
1968 // Set the flags to indicate whether the shift was by more than 32 bits.
1969 BuildMI(BB, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
1971 // DestHi = (>32) ? TmpReg3 : TmpReg2;
1972 BuildMI(BB, X86::CMOVNE32rr, 2,
1973 Result+1).addReg(TmpReg2).addReg(TmpReg3);
1974 // DestLo = (>32) ? TmpReg : TmpReg3;
1975 BuildMI(BB, X86::CMOVNE32rr, 2,
1976 Result).addReg(TmpReg3).addReg(TmpReg);
1978 // TmpReg2 = shrd inLo, inHi
1979 BuildMI(BB, X86::SHRD32rrCL,2,TmpReg2).addReg(ShiftOpLo)
1981 // TmpReg3 = s[ah]r inHi, CL
1982 BuildMI(BB, N.getOpcode() == ISD::SRA_PARTS ? X86::SAR32rCL
1983 : X86::SHR32rCL, 1, TmpReg3)
1986 // Set the flags to indicate whether the shift was by more than 32 bits.
1987 BuildMI(BB, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
1989 // DestLo = (>32) ? TmpReg3 : TmpReg2;
1990 BuildMI(BB, X86::CMOVNE32rr, 2,
1991 Result).addReg(TmpReg2).addReg(TmpReg3);
1993 // DestHi = (>32) ? TmpReg : TmpReg3;
1994 BuildMI(BB, X86::CMOVNE32rr, 2,
1995 Result+1).addReg(TmpReg3).addReg(TmpReg);
1997 return Result+N.ResNo;
2001 EmitSelectCC(N.getOperand(0), N.getOperand(1), N.getOperand(2),
2002 N.getValueType(), Result);
2011 assert((N.getOpcode() != ISD::SREM || MVT::isInteger(N.getValueType())) &&
2012 "We don't support this operator!");
2014 if (N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::FDIV) {
2015 // We can fold loads into FpDIVs, but not really into any others.
2016 if (N.getValueType() == MVT::f64 && !X86ScalarSSE) {
2017 // Check for reversed and unreversed DIV.
2018 if (isFoldableLoad(N.getOperand(0), N.getOperand(1), true)) {
2019 if (N.getOperand(0).getOpcode() == ISD::EXTLOAD)
2020 Opc = X86::FpDIVR32m;
2022 Opc = X86::FpDIVR64m;
2024 EmitFoldedLoad(N.getOperand(0), AM);
2025 Tmp1 = SelectExpr(N.getOperand(1));
2026 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
2028 } else if (isFoldableLoad(N.getOperand(1), N.getOperand(0), true) &&
2029 N.getOperand(1).getOpcode() == ISD::LOAD) {
2030 if (N.getOperand(1).getOpcode() == ISD::EXTLOAD)
2031 Opc = X86::FpDIV32m;
2033 Opc = X86::FpDIV64m;
2035 EmitFoldedLoad(N.getOperand(1), AM);
2036 Tmp1 = SelectExpr(N.getOperand(0));
2037 addFullAddress(BuildMI(BB, Opc, 5, Result).addReg(Tmp1), AM);
2043 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2044 Tmp1 = SelectExpr(N.getOperand(0));
2045 Tmp2 = SelectExpr(N.getOperand(1));
2047 Tmp2 = SelectExpr(N.getOperand(1));
2048 Tmp1 = SelectExpr(N.getOperand(0));
2051 bool isSigned = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::SREM;
2052 bool isDiv = N.getOpcode() == ISD::SDIV || N.getOpcode() == ISD::UDIV;
2053 unsigned LoReg, HiReg, DivOpcode, MovOpcode, ClrOpcode, SExtOpcode;
2054 switch (N.getValueType()) {
2055 default: assert(0 && "Cannot sdiv this type!");
2057 DivOpcode = isSigned ? X86::IDIV8r : X86::DIV8r;
2060 MovOpcode = X86::MOV8rr;
2061 ClrOpcode = X86::MOV8ri;
2062 SExtOpcode = X86::CBW;
2065 DivOpcode = isSigned ? X86::IDIV16r : X86::DIV16r;
2068 MovOpcode = X86::MOV16rr;
2069 ClrOpcode = X86::MOV16ri;
2070 SExtOpcode = X86::CWD;
2073 DivOpcode = isSigned ? X86::IDIV32r : X86::DIV32r;
2076 MovOpcode = X86::MOV32rr;
2077 ClrOpcode = X86::MOV32ri;
2078 SExtOpcode = X86::CDQ;
2081 BuildMI(BB, X86::DIVSSrr, 2, Result).addReg(Tmp1).addReg(Tmp2);
2084 Opc = X86ScalarSSE ? X86::DIVSDrr : X86::FpDIV;
2085 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
2089 // Set up the low part.
2090 BuildMI(BB, MovOpcode, 1, LoReg).addReg(Tmp1);
2093 // Sign extend the low part into the high part.
2094 BuildMI(BB, SExtOpcode, 0);
2096 // Zero out the high part, effectively zero extending the input.
2097 BuildMI(BB, ClrOpcode, 1, HiReg).addImm(0);
2100 // Emit the DIV/IDIV instruction.
2101 BuildMI(BB, DivOpcode, 1).addReg(Tmp2);
2103 // Get the result of the divide or rem.
2104 BuildMI(BB, MovOpcode, 1, Result).addReg(isDiv ? LoReg : HiReg);
2109 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2110 if (CN->getValue() == 1) { // X = SHL Y, 1 -> X = ADD Y, Y
2111 switch (N.getValueType()) {
2112 default: assert(0 && "Cannot shift this type!");
2113 case MVT::i8: Opc = X86::ADD8rr; break;
2114 case MVT::i16: Opc = X86::ADD16rr; break;
2115 case MVT::i32: Opc = X86::ADD32rr; break;
2117 Tmp1 = SelectExpr(N.getOperand(0));
2118 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp1);
2122 switch (N.getValueType()) {
2123 default: assert(0 && "Cannot shift this type!");
2124 case MVT::i8: Opc = X86::SHL8ri; break;
2125 case MVT::i16: Opc = X86::SHL16ri; break;
2126 case MVT::i32: Opc = X86::SHL32ri; break;
2128 Tmp1 = SelectExpr(N.getOperand(0));
2129 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2133 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2134 Tmp1 = SelectExpr(N.getOperand(0));
2135 Tmp2 = SelectExpr(N.getOperand(1));
2137 Tmp2 = SelectExpr(N.getOperand(1));
2138 Tmp1 = SelectExpr(N.getOperand(0));
2141 switch (N.getValueType()) {
2142 default: assert(0 && "Cannot shift this type!");
2143 case MVT::i8 : Opc = X86::SHL8rCL; break;
2144 case MVT::i16: Opc = X86::SHL16rCL; break;
2145 case MVT::i32: Opc = X86::SHL32rCL; break;
2147 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
2148 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2151 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2152 switch (N.getValueType()) {
2153 default: assert(0 && "Cannot shift this type!");
2154 case MVT::i8: Opc = X86::SHR8ri; break;
2155 case MVT::i16: Opc = X86::SHR16ri; break;
2156 case MVT::i32: Opc = X86::SHR32ri; break;
2158 Tmp1 = SelectExpr(N.getOperand(0));
2159 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2163 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2164 Tmp1 = SelectExpr(N.getOperand(0));
2165 Tmp2 = SelectExpr(N.getOperand(1));
2167 Tmp2 = SelectExpr(N.getOperand(1));
2168 Tmp1 = SelectExpr(N.getOperand(0));
2171 switch (N.getValueType()) {
2172 default: assert(0 && "Cannot shift this type!");
2173 case MVT::i8 : Opc = X86::SHR8rCL; break;
2174 case MVT::i16: Opc = X86::SHR16rCL; break;
2175 case MVT::i32: Opc = X86::SHR32rCL; break;
2177 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
2178 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2181 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2182 switch (N.getValueType()) {
2183 default: assert(0 && "Cannot shift this type!");
2184 case MVT::i8: Opc = X86::SAR8ri; break;
2185 case MVT::i16: Opc = X86::SAR16ri; break;
2186 case MVT::i32: Opc = X86::SAR32ri; break;
2188 Tmp1 = SelectExpr(N.getOperand(0));
2189 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(CN->getValue());
2193 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2194 Tmp1 = SelectExpr(N.getOperand(0));
2195 Tmp2 = SelectExpr(N.getOperand(1));
2197 Tmp2 = SelectExpr(N.getOperand(1));
2198 Tmp1 = SelectExpr(N.getOperand(0));
2201 switch (N.getValueType()) {
2202 default: assert(0 && "Cannot shift this type!");
2203 case MVT::i8 : Opc = X86::SAR8rCL; break;
2204 case MVT::i16: Opc = X86::SAR16rCL; break;
2205 case MVT::i32: Opc = X86::SAR32rCL; break;
2207 BuildMI(BB, X86::MOV8rr, 1, X86::CL).addReg(Tmp2);
2208 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
2212 EmitCMP(N.getOperand(0), N.getOperand(1), Node->hasOneUse());
2213 EmitSetCC(BB, Result, cast<CondCodeSDNode>(N.getOperand(2))->get(),
2214 MVT::isFloatingPoint(N.getOperand(1).getValueType()));
2217 // Make sure we generate both values.
2218 if (Result != 1) { // Generate the token
2219 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
2220 assert(0 && "Load already emitted!?");
2222 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2224 switch (Node->getValueType(0)) {
2225 default: assert(0 && "Cannot load this type!");
2227 case MVT::i8: Opc = X86::MOV8rm; break;
2228 case MVT::i16: Opc = X86::MOV16rm; break;
2229 case MVT::i32: Opc = X86::MOV32rm; break;
2230 case MVT::f32: Opc = X86::MOVSSrm; break;
2236 ContainsFPCode = true;
2241 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1))){
2242 unsigned CPIdx = BB->getParent()->getConstantPool()->
2243 getConstantPoolIndex(CP->get());
2244 Select(N.getOperand(0));
2245 addConstantPoolReference(BuildMI(BB, Opc, 4, Result), CPIdx);
2249 SDOperand Chain = N.getOperand(0);
2250 SDOperand Address = N.getOperand(1);
2251 if (getRegPressure(Chain) > getRegPressure(Address)) {
2253 SelectAddress(Address, AM);
2255 SelectAddress(Address, AM);
2259 addFullAddress(BuildMI(BB, Opc, 4, Result), AM);
2262 case X86ISD::FILD64m:
2263 // Make sure we generate both values.
2264 assert(Result != 1 && N.getValueType() == MVT::f64);
2265 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
2266 assert(0 && "Load already emitted!?");
2271 SDOperand Chain = N.getOperand(0);
2272 SDOperand Address = N.getOperand(1);
2273 if (getRegPressure(Chain) > getRegPressure(Address)) {
2275 SelectAddress(Address, AM);
2277 SelectAddress(Address, AM);
2281 addFullAddress(BuildMI(BB, X86::FpILD64m, 4, Result), AM);
2285 case ISD::EXTLOAD: // Arbitrarily codegen extloads as MOVZX*
2286 case ISD::ZEXTLOAD: {
2287 // Make sure we generate both values.
2289 ExprMap[N.getValue(1)] = 1; // Generate the token
2291 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2293 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N.getOperand(1)))
2294 if (Node->getValueType(0) == MVT::f64) {
2295 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::f32 &&
2297 unsigned CPIdx = BB->getParent()->getConstantPool()->
2298 getConstantPoolIndex(CP->get());
2300 addConstantPoolReference(BuildMI(BB, X86::FpLD32m, 4, Result), CPIdx);
2305 if (getRegPressure(Node->getOperand(0)) >
2306 getRegPressure(Node->getOperand(1))) {
2307 Select(Node->getOperand(0)); // chain
2308 SelectAddress(Node->getOperand(1), AM);
2310 SelectAddress(Node->getOperand(1), AM);
2311 Select(Node->getOperand(0)); // chain
2314 switch (Node->getValueType(0)) {
2315 default: assert(0 && "Unknown type to sign extend to.");
2317 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::f32 &&
2319 addFullAddress(BuildMI(BB, X86::FpLD32m, 5, Result), AM);
2322 switch (cast<VTSDNode>(Node->getOperand(3))->getVT()) {
2324 assert(0 && "Bad zero extend!");
2327 addFullAddress(BuildMI(BB, X86::MOVZX32rm8, 5, Result), AM);
2330 addFullAddress(BuildMI(BB, X86::MOVZX32rm16, 5, Result), AM);
2335 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() <= MVT::i8 &&
2336 "Bad zero extend!");
2337 addFullAddress(BuildMI(BB, X86::MOVZX16rm8, 5, Result), AM);
2340 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::i1 &&
2341 "Bad zero extend!");
2342 addFullAddress(BuildMI(BB, X86::MOV8rm, 5, Result), AM);
2347 case ISD::SEXTLOAD: {
2348 // Make sure we generate both values.
2350 ExprMap[N.getValue(1)] = 1; // Generate the token
2352 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2355 if (getRegPressure(Node->getOperand(0)) >
2356 getRegPressure(Node->getOperand(1))) {
2357 Select(Node->getOperand(0)); // chain
2358 SelectAddress(Node->getOperand(1), AM);
2360 SelectAddress(Node->getOperand(1), AM);
2361 Select(Node->getOperand(0)); // chain
2364 switch (Node->getValueType(0)) {
2365 case MVT::i8: assert(0 && "Cannot sign extend from bool!");
2366 default: assert(0 && "Unknown type to sign extend to.");
2368 switch (cast<VTSDNode>(Node->getOperand(3))->getVT()) {
2370 case MVT::i1: assert(0 && "Cannot sign extend from bool!");
2372 addFullAddress(BuildMI(BB, X86::MOVSX32rm8, 5, Result), AM);
2375 addFullAddress(BuildMI(BB, X86::MOVSX32rm16, 5, Result), AM);
2380 assert(cast<VTSDNode>(Node->getOperand(3))->getVT() == MVT::i8 &&
2381 "Cannot sign extend from bool!");
2382 addFullAddress(BuildMI(BB, X86::MOVSX16rm8, 5, Result), AM);
2388 case ISD::DYNAMIC_STACKALLOC:
2389 // Generate both result values.
2391 ExprMap[N.getValue(1)] = 1; // Generate the token
2393 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2395 // FIXME: We are currently ignoring the requested alignment for handling
2396 // greater than the stack alignment. This will need to be revisited at some
2397 // point. Align = N.getOperand(2);
2399 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
2400 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
2401 std::cerr << "Cannot allocate stack object with greater alignment than"
2402 << " the stack alignment yet!";
2406 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
2407 Select(N.getOperand(0));
2408 BuildMI(BB, X86::SUB32ri, 2, X86::ESP).addReg(X86::ESP)
2409 .addImm(CN->getValue());
2411 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2412 Select(N.getOperand(0));
2413 Tmp1 = SelectExpr(N.getOperand(1));
2415 Tmp1 = SelectExpr(N.getOperand(1));
2416 Select(N.getOperand(0));
2419 // Subtract size from stack pointer, thereby allocating some space.
2420 BuildMI(BB, X86::SUB32rr, 2, X86::ESP).addReg(X86::ESP).addReg(Tmp1);
2423 // Put a pointer to the space into the result register, by copying the stack
2425 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::ESP);
2428 case X86ISD::TAILCALL:
2429 case X86ISD::CALL: {
2430 // The chain for this call is now lowered.
2431 ExprMap.insert(std::make_pair(N.getValue(0), 1));
2433 bool isDirect = isa<GlobalAddressSDNode>(N.getOperand(1)) ||
2434 isa<ExternalSymbolSDNode>(N.getOperand(1));
2435 unsigned Callee = 0;
2437 Select(N.getOperand(0));
2439 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2440 Select(N.getOperand(0));
2441 Callee = SelectExpr(N.getOperand(1));
2443 Callee = SelectExpr(N.getOperand(1));
2444 Select(N.getOperand(0));
2448 // If this call has values to pass in registers, do so now.
2449 if (Node->getNumOperands() > 4) {
2450 // The first value is passed in (a part of) EAX, the second in EDX.
2451 unsigned RegOp1 = SelectExpr(N.getOperand(4));
2453 Node->getNumOperands() > 5 ? SelectExpr(N.getOperand(5)) : 0;
2455 switch (N.getOperand(4).getValueType()) {
2456 default: assert(0 && "Bad thing to pass in regs");
2458 case MVT::i8: BuildMI(BB, X86::MOV8rr , 1,X86::AL).addReg(RegOp1); break;
2459 case MVT::i16: BuildMI(BB, X86::MOV16rr, 1,X86::AX).addReg(RegOp1); break;
2460 case MVT::i32: BuildMI(BB, X86::MOV32rr, 1,X86::EAX).addReg(RegOp1);break;
2463 switch (N.getOperand(5).getValueType()) {
2464 default: assert(0 && "Bad thing to pass in regs");
2467 BuildMI(BB, X86::MOV8rr , 1, X86::DL).addReg(RegOp2);
2470 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(RegOp2);
2473 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(RegOp2);
2478 if (GlobalAddressSDNode *GASD =
2479 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
2480 BuildMI(BB, X86::CALLpcrel32, 1).addGlobalAddress(GASD->getGlobal(),true);
2481 } else if (ExternalSymbolSDNode *ESSDN =
2482 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
2483 BuildMI(BB, X86::CALLpcrel32,
2484 1).addExternalSymbol(ESSDN->getSymbol(), true);
2486 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
2487 Select(N.getOperand(0));
2488 Tmp1 = SelectExpr(N.getOperand(1));
2490 Tmp1 = SelectExpr(N.getOperand(1));
2491 Select(N.getOperand(0));
2494 BuildMI(BB, X86::CALL32r, 1).addReg(Tmp1);
2497 // Get caller stack amount and amount the callee added to the stack pointer.
2498 Tmp1 = cast<ConstantSDNode>(N.getOperand(2))->getValue();
2499 Tmp2 = cast<ConstantSDNode>(N.getOperand(3))->getValue();
2500 BuildMI(BB, X86::ADJCALLSTACKUP, 2).addImm(Tmp1).addImm(Tmp2);
2502 if (Node->getNumValues() != 1)
2503 switch (Node->getValueType(1)) {
2504 default: assert(0 && "Unknown value type for call result!");
2505 case MVT::Other: return 1;
2508 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2511 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
2514 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
2515 if (Node->getNumValues() == 3 && Node->getValueType(2) == MVT::i32)
2516 BuildMI(BB, X86::MOV32rr, 1, Result+1).addReg(X86::EDX);
2518 case MVT::f64: // Floating-point return values live in %ST(0)
2520 ContainsFPCode = true;
2521 BuildMI(BB, X86::FpGETRESULT, 1, X86::FP0);
2523 unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
2524 MachineFunction *F = BB->getParent();
2525 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
2526 addFrameReference(BuildMI(BB, X86::FpST64m, 5), FrameIdx).addReg(X86::FP0);
2527 addFrameReference(BuildMI(BB, X86::MOVSDrm, 4, Result), FrameIdx);
2530 ContainsFPCode = true;
2531 BuildMI(BB, X86::FpGETRESULT, 1, Result);
2535 return Result+N.ResNo-1;
2538 // First, determine that the size of the operand falls within the acceptable
2539 // range for this architecture.
2541 if (Node->getOperand(1).getValueType() != MVT::i16) {
2542 std::cerr << "llvm.readport: Address size is not 16 bits\n";
2546 // Make sure we generate both values.
2547 if (Result != 1) { // Generate the token
2548 if (!ExprMap.insert(std::make_pair(N.getValue(1), 1)).second)
2549 assert(0 && "readport already emitted!?");
2551 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
2553 Select(Node->getOperand(0)); // Select the chain.
2555 // If the port is a single-byte constant, use the immediate form.
2556 if (ConstantSDNode *Port = dyn_cast<ConstantSDNode>(Node->getOperand(1)))
2557 if ((Port->getValue() & 255) == Port->getValue()) {
2558 switch (Node->getValueType(0)) {
2560 BuildMI(BB, X86::IN8ri, 1).addImm(Port->getValue());
2561 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2564 BuildMI(BB, X86::IN16ri, 1).addImm(Port->getValue());
2565 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
2568 BuildMI(BB, X86::IN32ri, 1).addImm(Port->getValue());
2569 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
2575 // Now, move the I/O port address into the DX register and use the IN
2576 // instruction to get the input data.
2578 Tmp1 = SelectExpr(Node->getOperand(1));
2579 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Tmp1);
2580 switch (Node->getValueType(0)) {
2582 BuildMI(BB, X86::IN8rr, 0);
2583 BuildMI(BB, X86::MOV8rr, 1, Result).addReg(X86::AL);
2586 BuildMI(BB, X86::IN16rr, 0);
2587 BuildMI(BB, X86::MOV16rr, 1, Result).addReg(X86::AX);
2590 BuildMI(BB, X86::IN32rr, 0);
2591 BuildMI(BB, X86::MOV32rr, 1, Result).addReg(X86::EAX);
2594 std::cerr << "Cannot do input on this data type";
2603 /// TryToFoldLoadOpStore - Given a store node, try to fold together a
2604 /// load/op/store instruction. If successful return true.
2605 bool ISel::TryToFoldLoadOpStore(SDNode *Node) {
2606 assert(Node->getOpcode() == ISD::STORE && "Can only do this for stores!");
2607 SDOperand Chain = Node->getOperand(0);
2608 SDOperand StVal = Node->getOperand(1);
2609 SDOperand StPtr = Node->getOperand(2);
2611 // The chain has to be a load, the stored value must be an integer binary
2612 // operation with one use.
2613 if (!StVal.Val->hasOneUse() || StVal.Val->getNumOperands() != 2 ||
2614 MVT::isFloatingPoint(StVal.getValueType()))
2617 // Token chain must either be a factor node or the load to fold.
2618 if (Chain.getOpcode() != ISD::LOAD && Chain.getOpcode() != ISD::TokenFactor)
2623 // Check to see if there is a load from the same pointer that we're storing
2624 // to in either operand of the binop.
2625 if (StVal.getOperand(0).getOpcode() == ISD::LOAD &&
2626 StVal.getOperand(0).getOperand(1) == StPtr)
2627 TheLoad = StVal.getOperand(0);
2628 else if (StVal.getOperand(1).getOpcode() == ISD::LOAD &&
2629 StVal.getOperand(1).getOperand(1) == StPtr)
2630 TheLoad = StVal.getOperand(1);
2632 return false; // No matching load operand.
2634 // We can only fold the load if there are no intervening side-effecting
2635 // operations. This means that the store uses the load as its token chain, or
2636 // there are only token factor nodes in between the store and load.
2637 if (Chain != TheLoad.getValue(1)) {
2638 // Okay, the other option is that we have a store referring to (possibly
2639 // nested) token factor nodes. For now, just try peeking through one level
2640 // of token factors to see if this is the case.
2641 bool ChainOk = false;
2642 if (Chain.getOpcode() == ISD::TokenFactor) {
2643 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
2644 if (Chain.getOperand(i) == TheLoad.getValue(1)) {
2650 if (!ChainOk) return false;
2653 if (TheLoad.getOperand(1) != StPtr)
2656 // Make sure that one of the operands of the binop is the load, and that the
2657 // load folds into the binop.
2658 if (((StVal.getOperand(0) != TheLoad ||
2659 !isFoldableLoad(TheLoad, StVal.getOperand(1))) &&
2660 (StVal.getOperand(1) != TheLoad ||
2661 !isFoldableLoad(TheLoad, StVal.getOperand(0)))))
2664 // Finally, check to see if this is one of the ops we can handle!
2665 static const unsigned ADDTAB[] = {
2666 X86::ADD8mi, X86::ADD16mi, X86::ADD32mi,
2667 X86::ADD8mr, X86::ADD16mr, X86::ADD32mr,
2669 static const unsigned SUBTAB[] = {
2670 X86::SUB8mi, X86::SUB16mi, X86::SUB32mi,
2671 X86::SUB8mr, X86::SUB16mr, X86::SUB32mr,
2673 static const unsigned ANDTAB[] = {
2674 X86::AND8mi, X86::AND16mi, X86::AND32mi,
2675 X86::AND8mr, X86::AND16mr, X86::AND32mr,
2677 static const unsigned ORTAB[] = {
2678 X86::OR8mi, X86::OR16mi, X86::OR32mi,
2679 X86::OR8mr, X86::OR16mr, X86::OR32mr,
2681 static const unsigned XORTAB[] = {
2682 X86::XOR8mi, X86::XOR16mi, X86::XOR32mi,
2683 X86::XOR8mr, X86::XOR16mr, X86::XOR32mr,
2685 static const unsigned SHLTAB[] = {
2686 X86::SHL8mi, X86::SHL16mi, X86::SHL32mi,
2687 /*Have to put the reg in CL*/0, 0, 0,
2689 static const unsigned SARTAB[] = {
2690 X86::SAR8mi, X86::SAR16mi, X86::SAR32mi,
2691 /*Have to put the reg in CL*/0, 0, 0,
2693 static const unsigned SHRTAB[] = {
2694 X86::SHR8mi, X86::SHR16mi, X86::SHR32mi,
2695 /*Have to put the reg in CL*/0, 0, 0,
2698 const unsigned *TabPtr = 0;
2699 switch (StVal.getOpcode()) {
2701 std::cerr << "CANNOT [mem] op= val: ";
2702 StVal.Val->dump(); std::cerr << "\n";
2710 case ISD::UREM: return false;
2712 case ISD::ADD: TabPtr = ADDTAB; break;
2713 case ISD::SUB: TabPtr = SUBTAB; break;
2714 case ISD::AND: TabPtr = ANDTAB; break;
2715 case ISD:: OR: TabPtr = ORTAB; break;
2716 case ISD::XOR: TabPtr = XORTAB; break;
2717 case ISD::SHL: TabPtr = SHLTAB; break;
2718 case ISD::SRA: TabPtr = SARTAB; break;
2719 case ISD::SRL: TabPtr = SHRTAB; break;
2722 // Handle: [mem] op= CST
2723 SDOperand Op0 = StVal.getOperand(0);
2724 SDOperand Op1 = StVal.getOperand(1);
2726 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
2727 switch (Op0.getValueType()) { // Use Op0's type because of shifts.
2730 case MVT::i8: Opc = TabPtr[0]; break;
2731 case MVT::i16: Opc = TabPtr[1]; break;
2732 case MVT::i32: Opc = TabPtr[2]; break;
2736 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
2737 assert(0 && "Already emitted?");
2741 if (getRegPressure(TheLoad.getOperand(0)) >
2742 getRegPressure(TheLoad.getOperand(1))) {
2743 Select(TheLoad.getOperand(0));
2744 SelectAddress(TheLoad.getOperand(1), AM);
2746 SelectAddress(TheLoad.getOperand(1), AM);
2747 Select(TheLoad.getOperand(0));
2750 if (StVal.getOpcode() == ISD::ADD) {
2751 if (CN->getValue() == 1) {
2752 switch (Op0.getValueType()) {
2755 addFullAddress(BuildMI(BB, X86::INC8m, 4), AM);
2757 case MVT::i16: Opc = TabPtr[1];
2758 addFullAddress(BuildMI(BB, X86::INC16m, 4), AM);
2760 case MVT::i32: Opc = TabPtr[2];
2761 addFullAddress(BuildMI(BB, X86::INC32m, 4), AM);
2764 } else if (CN->getValue()+1 == 0) { // [X] += -1 -> DEC [X]
2765 switch (Op0.getValueType()) {
2768 addFullAddress(BuildMI(BB, X86::DEC8m, 4), AM);
2770 case MVT::i16: Opc = TabPtr[1];
2771 addFullAddress(BuildMI(BB, X86::DEC16m, 4), AM);
2773 case MVT::i32: Opc = TabPtr[2];
2774 addFullAddress(BuildMI(BB, X86::DEC32m, 4), AM);
2780 addFullAddress(BuildMI(BB, Opc, 4+1),AM).addImm(CN->getValue());
2785 // If we have [mem] = V op [mem], try to turn it into:
2786 // [mem] = [mem] op V.
2787 if (Op1 == TheLoad &&
2788 StVal.getOpcode() != ISD::SUB && StVal.getOpcode() != ISD::FSUB &&
2789 StVal.getOpcode() != ISD::SHL && StVal.getOpcode() != ISD::SRA &&
2790 StVal.getOpcode() != ISD::SRL)
2791 std::swap(Op0, Op1);
2793 if (Op0 != TheLoad) return false;
2795 switch (Op0.getValueType()) {
2796 default: return false;
2798 case MVT::i8: Opc = TabPtr[3]; break;
2799 case MVT::i16: Opc = TabPtr[4]; break;
2800 case MVT::i32: Opc = TabPtr[5]; break;
2803 // Table entry doesn't exist?
2804 if (Opc == 0) return false;
2806 if (!ExprMap.insert(std::make_pair(TheLoad.getValue(1), 1)).second)
2807 assert(0 && "Already emitted?");
2809 Select(TheLoad.getOperand(0));
2812 SelectAddress(TheLoad.getOperand(1), AM);
2813 unsigned Reg = SelectExpr(Op1);
2814 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Reg);
2818 /// If node is a ret(tailcall) node, emit the specified tail call and return
2819 /// true, otherwise return false.
2821 /// FIXME: This whole thing should be a post-legalize optimization pass which
2822 /// recognizes and transforms the dag. We don't want the selection phase doing
2825 bool ISel::EmitPotentialTailCall(SDNode *RetNode) {
2826 assert(RetNode->getOpcode() == ISD::RET && "Not a return");
2828 SDOperand Chain = RetNode->getOperand(0);
2830 // If this is a token factor node where one operand is a call, dig into it.
2831 SDOperand TokFactor;
2832 unsigned TokFactorOperand = 0;
2833 if (Chain.getOpcode() == ISD::TokenFactor) {
2834 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
2835 if (Chain.getOperand(i).getOpcode() == ISD::CALLSEQ_END ||
2836 Chain.getOperand(i).getOpcode() == X86ISD::TAILCALL) {
2837 TokFactorOperand = i;
2839 Chain = Chain.getOperand(i);
2842 if (TokFactor.Val == 0) return false; // No call operand.
2845 // Skip the CALLSEQ_END node if present.
2846 if (Chain.getOpcode() == ISD::CALLSEQ_END)
2847 Chain = Chain.getOperand(0);
2849 // Is a tailcall the last control operation that occurs before the return?
2850 if (Chain.getOpcode() != X86ISD::TAILCALL)
2853 // If we return a value, is it the value produced by the call?
2854 if (RetNode->getNumOperands() > 1) {
2855 // Not returning the ret val of the call?
2856 if (Chain.Val->getNumValues() == 1 ||
2857 RetNode->getOperand(1) != Chain.getValue(1))
2860 if (RetNode->getNumOperands() > 2) {
2861 if (Chain.Val->getNumValues() == 2 ||
2862 RetNode->getOperand(2) != Chain.getValue(2))
2865 assert(RetNode->getNumOperands() <= 3);
2868 // CalleeCallArgAmt - The total number of bytes used for the callee arg area.
2869 // For FastCC, this will always be > 0.
2870 unsigned CalleeCallArgAmt =
2871 cast<ConstantSDNode>(Chain.getOperand(2))->getValue();
2873 // CalleeCallArgPopAmt - The number of bytes in the call area popped by the
2874 // callee. For FastCC this will always be > 0, for CCC this is always 0.
2875 unsigned CalleeCallArgPopAmt =
2876 cast<ConstantSDNode>(Chain.getOperand(3))->getValue();
2878 // There are several cases we can handle here. First, if the caller and
2879 // callee are both CCC functions, we can tailcall if the callee takes <= the
2880 // number of argument bytes that the caller does.
2881 if (CalleeCallArgPopAmt == 0 && // Callee is C CallingConv?
2882 X86Lowering.getBytesToPopOnReturn() == 0) { // Caller is C CallingConv?
2883 // Check to see if caller arg area size >= callee arg area size.
2884 if (X86Lowering.getBytesCallerReserves() >= CalleeCallArgAmt) {
2885 //std::cerr << "CCC TAILCALL UNIMP!\n";
2886 // If TokFactor is non-null, emit all operands.
2888 //EmitCCCToCCCTailCall(Chain.Val);
2894 // Second, if both are FastCC functions, we can always perform the tail call.
2895 if (CalleeCallArgPopAmt && X86Lowering.getBytesToPopOnReturn()) {
2896 // If TokFactor is non-null, emit all operands before the call.
2897 if (TokFactor.Val) {
2898 for (unsigned i = 0, e = TokFactor.getNumOperands(); i != e; ++i)
2899 if (i != TokFactorOperand)
2900 Select(TokFactor.getOperand(i));
2903 EmitFastCCToFastCCTailCall(Chain.Val);
2907 // We don't support mixed calls, due to issues with alignment. We could in
2908 // theory handle some mixed calls from CCC -> FastCC if the stack is properly
2909 // aligned (which depends on the number of arguments to the callee). TODO.
2913 static SDOperand GetAdjustedArgumentStores(SDOperand Chain, int Offset,
2914 SelectionDAG &DAG) {
2915 MVT::ValueType StoreVT;
2916 switch (Chain.getOpcode()) {
2917 default: assert(0 && "Unexpected node!");
2918 case ISD::CALLSEQ_START:
2919 // If we found the start of the call sequence, we're done. We actually
2920 // strip off the CALLSEQ_START node, to avoid generating the
2921 // ADJCALLSTACKDOWN marker for the tail call.
2922 return Chain.getOperand(0);
2923 case ISD::TokenFactor: {
2924 std::vector<SDOperand> Ops;
2925 Ops.reserve(Chain.getNumOperands());
2926 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
2927 Ops.push_back(GetAdjustedArgumentStores(Chain.getOperand(i), Offset,DAG));
2928 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops);
2930 case ISD::STORE: // Normal store
2931 StoreVT = Chain.getOperand(1).getValueType();
2933 case ISD::TRUNCSTORE: // FLOAT store
2934 StoreVT = cast<VTSDNode>(Chain.getOperand(4))->getVT();
2938 SDOperand OrigDest = Chain.getOperand(2);
2939 unsigned OrigOffset;
2941 if (OrigDest.getOpcode() == ISD::CopyFromReg) {
2943 assert(cast<RegisterSDNode>(OrigDest.getOperand(1))->getReg() == X86::ESP);
2945 // We expect only (ESP+C)
2946 assert(OrigDest.getOpcode() == ISD::ADD &&
2947 isa<ConstantSDNode>(OrigDest.getOperand(1)) &&
2948 OrigDest.getOperand(0).getOpcode() == ISD::CopyFromReg &&
2949 cast<RegisterSDNode>(OrigDest.getOperand(0).getOperand(1))->getReg()
2951 OrigOffset = cast<ConstantSDNode>(OrigDest.getOperand(1))->getValue();
2954 // Compute the new offset from the incoming ESP value we wish to use.
2955 unsigned NewOffset = OrigOffset + Offset;
2957 unsigned OpSize = (MVT::getSizeInBits(StoreVT)+7)/8; // Bits -> Bytes
2958 MachineFunction &MF = DAG.getMachineFunction();
2959 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, NewOffset);
2960 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
2962 SDOperand InChain = GetAdjustedArgumentStores(Chain.getOperand(0), Offset,
2964 if (Chain.getOpcode() == ISD::STORE)
2965 return DAG.getNode(ISD::STORE, MVT::Other, InChain, Chain.getOperand(1),
2967 assert(Chain.getOpcode() == ISD::TRUNCSTORE);
2968 return DAG.getNode(ISD::TRUNCSTORE, MVT::Other, InChain, Chain.getOperand(1),
2969 FIN, DAG.getSrcValue(NULL), DAG.getValueType(StoreVT));
2973 /// EmitFastCCToFastCCTailCall - Given a tailcall in the tail position to a
2974 /// fastcc function from a fastcc function, emit the code to emit a 'proper'
2976 void ISel::EmitFastCCToFastCCTailCall(SDNode *TailCallNode) {
2977 unsigned CalleeCallArgSize =
2978 cast<ConstantSDNode>(TailCallNode->getOperand(2))->getValue();
2979 unsigned CallerArgSize = X86Lowering.getBytesToPopOnReturn();
2981 //std::cerr << "****\n*** EMITTING TAIL CALL!\n****\n";
2983 // Adjust argument stores. Instead of storing to [ESP], f.e., store to frame
2984 // indexes that are relative to the incoming ESP. If the incoming and
2985 // outgoing arg sizes are the same we will store to [InESP] instead of
2986 // [CurESP] and the ESP referenced will be relative to the incoming function
2988 int ESPOffset = CallerArgSize-CalleeCallArgSize;
2989 SDOperand AdjustedArgStores =
2990 GetAdjustedArgumentStores(TailCallNode->getOperand(0), ESPOffset, *TheDAG);
2992 // Copy the return address of the caller into a virtual register so we don't
2994 SDOperand RetVal(0, 0);
2996 SDOperand RetValAddr = X86Lowering.getReturnAddressFrameIndex(*TheDAG);
2997 RetVal = TheDAG->getLoad(MVT::i32, TheDAG->getEntryNode(),
2998 RetValAddr, TheDAG->getSrcValue(NULL));
3002 // Codegen all of the argument stores.
3003 Select(AdjustedArgStores);
3006 // Emit a store of the saved ret value to the new location.
3007 MachineFunction &MF = TheDAG->getMachineFunction();
3008 int ReturnAddrFI = MF.getFrameInfo()->CreateFixedObject(4, ESPOffset-4);
3009 SDOperand RetValAddr = TheDAG->getFrameIndex(ReturnAddrFI, MVT::i32);
3010 Select(TheDAG->getNode(ISD::STORE, MVT::Other, TheDAG->getEntryNode(),
3011 RetVal, RetValAddr));
3014 // Get the destination value.
3015 SDOperand Callee = TailCallNode->getOperand(1);
3016 bool isDirect = isa<GlobalAddressSDNode>(Callee) ||
3017 isa<ExternalSymbolSDNode>(Callee);
3018 unsigned CalleeReg = 0;
3020 // If this is not a direct tail call, evaluate the callee's address.
3021 CalleeReg = SelectExpr(Callee);
3024 unsigned RegOp1 = 0;
3025 unsigned RegOp2 = 0;
3027 if (TailCallNode->getNumOperands() > 4) {
3028 // The first value is passed in (a part of) EAX, the second in EDX.
3029 RegOp1 = SelectExpr(TailCallNode->getOperand(4));
3030 if (TailCallNode->getNumOperands() > 5)
3031 RegOp2 = SelectExpr(TailCallNode->getOperand(5));
3033 switch (TailCallNode->getOperand(4).getValueType()) {
3034 default: assert(0 && "Bad thing to pass in regs");
3037 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(RegOp1);
3041 BuildMI(BB, X86::MOV16rr, 1,X86::AX).addReg(RegOp1);
3045 BuildMI(BB, X86::MOV32rr, 1,X86::EAX).addReg(RegOp1);
3050 switch (TailCallNode->getOperand(5).getValueType()) {
3051 default: assert(0 && "Bad thing to pass in regs");
3054 BuildMI(BB, X86::MOV8rr, 1, X86::DL).addReg(RegOp2);
3058 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(RegOp2);
3062 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(RegOp2);
3068 // If this is not a direct tail call, put the callee's address into ECX.
3069 // The address has to be evaluated into a non-callee save register that is
3070 // not used for arguments. This means either ECX, as EAX and EDX may be
3071 // used for argument passing. We do this here to make sure that the
3072 // expressions for arguments and callee are all evaluated before the copies
3073 // into physical registers.
3075 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CalleeReg);
3079 BuildMI(BB, X86::ADJSTACKPTRri, 2,
3080 X86::ESP).addReg(X86::ESP).addImm(ESPOffset);
3082 // TODO: handle jmp [mem]
3084 BuildMI(BB, X86::TAILJMPr, 1).addReg(X86::ECX);
3085 } else if (GlobalAddressSDNode *GASD = dyn_cast<GlobalAddressSDNode>(Callee)){
3086 BuildMI(BB, X86::TAILJMPd, 1).addGlobalAddress(GASD->getGlobal(), true);
3088 ExternalSymbolSDNode *ESSDN = cast<ExternalSymbolSDNode>(Callee);
3089 BuildMI(BB, X86::TAILJMPd, 1).addExternalSymbol(ESSDN->getSymbol(), true);
3091 // ADD IMPLICIT USE RegOp1/RegOp2's
3095 void ISel::Select(SDOperand N) {
3096 unsigned Tmp1 = 0, Tmp2 = 0, Opc = 0;
3098 if (!ExprMap.insert(std::make_pair(N, 1)).second)
3099 return; // Already selected.
3101 SDNode *Node = N.Val;
3103 switch (Node->getOpcode()) {
3105 Node->dump(); std::cerr << "\n";
3106 assert(0 && "Node not handled yet!");
3107 case X86ISD::RDTSC_DAG:
3108 Select(Node->getOperand(0)); //Chain
3109 BuildMI(BB, X86::RDTSC, 0);
3112 case ISD::EntryToken: return; // Noop
3113 case ISD::TokenFactor:
3114 if (Node->getNumOperands() == 2) {
3116 getRegPressure(Node->getOperand(1))>getRegPressure(Node->getOperand(0));
3117 Select(Node->getOperand(OneFirst));
3118 Select(Node->getOperand(!OneFirst));
3120 std::vector<std::pair<unsigned, unsigned> > OpsP;
3121 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
3122 OpsP.push_back(std::make_pair(getRegPressure(Node->getOperand(i)), i));
3123 std::sort(OpsP.begin(), OpsP.end());
3124 std::reverse(OpsP.begin(), OpsP.end());
3125 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
3126 Select(Node->getOperand(OpsP[i].second));
3129 case ISD::CopyToReg:
3130 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
3131 Select(N.getOperand(0));
3132 Tmp1 = SelectExpr(N.getOperand(2));
3134 Tmp1 = SelectExpr(N.getOperand(2));
3135 Select(N.getOperand(0));
3137 Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3140 switch (N.getOperand(2).getValueType()) {
3141 default: assert(0 && "Invalid type for operation!");
3143 case MVT::i8: Opc = X86::MOV8rr; break;
3144 case MVT::i16: Opc = X86::MOV16rr; break;
3145 case MVT::i32: Opc = X86::MOV32rr; break;
3146 case MVT::f32: Opc = X86::MOVSSrr; break;
3152 ContainsFPCode = true;
3156 BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
3160 if (N.getOperand(0).getOpcode() == ISD::CALLSEQ_END ||
3161 N.getOperand(0).getOpcode() == X86ISD::TAILCALL ||
3162 N.getOperand(0).getOpcode() == ISD::TokenFactor)
3163 if (EmitPotentialTailCall(Node))
3166 switch (N.getNumOperands()) {
3168 assert(0 && "Unknown return instruction!");
3170 assert(N.getOperand(1).getValueType() == MVT::i32 &&
3171 N.getOperand(2).getValueType() == MVT::i32 &&
3172 "Unknown two-register value!");
3173 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
3174 Tmp1 = SelectExpr(N.getOperand(1));
3175 Tmp2 = SelectExpr(N.getOperand(2));
3177 Tmp2 = SelectExpr(N.getOperand(2));
3178 Tmp1 = SelectExpr(N.getOperand(1));
3180 Select(N.getOperand(0));
3182 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
3183 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(Tmp2);
3186 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3187 Select(N.getOperand(0));
3188 Tmp1 = SelectExpr(N.getOperand(1));
3190 Tmp1 = SelectExpr(N.getOperand(1));
3191 Select(N.getOperand(0));
3193 switch (N.getOperand(1).getValueType()) {
3194 default: assert(0 && "All other types should have been promoted!!");
3197 // Spill the value to memory and reload it into top of stack.
3198 unsigned Size = MVT::getSizeInBits(MVT::f32)/8;
3199 MachineFunction *F = BB->getParent();
3200 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
3201 addFrameReference(BuildMI(BB, X86::MOVSSmr, 5), FrameIdx).addReg(Tmp1);
3202 addFrameReference(BuildMI(BB, X86::FpLD32m, 4, X86::FP0), FrameIdx);
3203 BuildMI(BB, X86::FpSETRESULT, 1).addReg(X86::FP0);
3204 ContainsFPCode = true;
3206 assert(0 && "MVT::f32 only legal with scalar sse fp");
3212 // Spill the value to memory and reload it into top of stack.
3213 unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
3214 MachineFunction *F = BB->getParent();
3215 int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
3216 addFrameReference(BuildMI(BB, X86::MOVSDmr, 5), FrameIdx).addReg(Tmp1);
3217 addFrameReference(BuildMI(BB, X86::FpLD64m, 4, X86::FP0), FrameIdx);
3218 BuildMI(BB, X86::FpSETRESULT, 1).addReg(X86::FP0);
3219 ContainsFPCode = true;
3221 BuildMI(BB, X86::FpSETRESULT, 1).addReg(Tmp1);
3225 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
3230 Select(N.getOperand(0));
3233 if (X86Lowering.getBytesToPopOnReturn() == 0)
3234 BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
3236 BuildMI(BB, X86::RETI, 1).addImm(X86Lowering.getBytesToPopOnReturn());
3239 Select(N.getOperand(0));
3240 MachineBasicBlock *Dest =
3241 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
3242 BuildMI(BB, X86::JMP, 1).addMBB(Dest);
3247 MachineBasicBlock *Dest =
3248 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
3250 // Try to fold a setcc into the branch. If this fails, emit a test/jne
3252 if (EmitBranchCC(Dest, N.getOperand(0), N.getOperand(1))) {
3253 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(1))) {
3254 Select(N.getOperand(0));
3255 Tmp1 = SelectExpr(N.getOperand(1));
3257 Tmp1 = SelectExpr(N.getOperand(1));
3258 Select(N.getOperand(0));
3260 BuildMI(BB, X86::TEST8rr, 2).addReg(Tmp1).addReg(Tmp1);
3261 BuildMI(BB, X86::JNE, 1).addMBB(Dest);
3268 // If this load could be folded into the only using instruction, and if it
3269 // is safe to emit the instruction here, try to do so now.
3270 if (Node->hasNUsesOfValue(1, 0)) {
3271 SDOperand TheVal = N.getValue(0);
3273 for (SDNode::use_iterator UI = Node->use_begin(); ; ++UI) {
3274 assert(UI != Node->use_end() && "Didn't find use!");
3276 for (unsigned i = 0, e = UN->getNumOperands(); i != e; ++i)
3277 if (UN->getOperand(i) == TheVal) {
3283 // Only handle unary operators right now.
3284 if (User->getNumOperands() == 1) {
3286 SelectExpr(SDOperand(User, 0));
3297 case ISD::DYNAMIC_STACKALLOC:
3298 case X86ISD::TAILCALL:
3303 case ISD::CopyFromReg:
3304 case X86ISD::FILD64m:
3306 SelectExpr(N.getValue(0));
3309 case X86ISD::FP_TO_INT16_IN_MEM:
3310 case X86ISD::FP_TO_INT32_IN_MEM:
3311 case X86ISD::FP_TO_INT64_IN_MEM: {
3312 assert(N.getOperand(1).getValueType() == MVT::f64);
3314 Select(N.getOperand(0)); // Select the token chain
3317 if (getRegPressure(N.getOperand(1)) > getRegPressure(N.getOperand(2))) {
3318 ValReg = SelectExpr(N.getOperand(1));
3319 SelectAddress(N.getOperand(2), AM);
3321 SelectAddress(N.getOperand(2), AM);
3322 ValReg = SelectExpr(N.getOperand(1));
3325 // Change the floating point control register to use "round towards zero"
3326 // mode when truncating to an integer value.
3328 MachineFunction *F = BB->getParent();
3329 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
3330 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
3332 // Load the old value of the high byte of the control word...
3333 unsigned OldCW = MakeReg(MVT::i16);
3334 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
3336 // Set the high part to be round to zero...
3337 addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
3339 // Reload the modified control word now...
3340 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
3342 // Restore the memory image of control word to original value
3343 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
3345 // Get the X86 opcode to use.
3346 switch (N.getOpcode()) {
3347 case X86ISD::FP_TO_INT16_IN_MEM: Tmp1 = X86::FpIST16m; break;
3348 case X86ISD::FP_TO_INT32_IN_MEM: Tmp1 = X86::FpIST32m; break;
3349 case X86ISD::FP_TO_INT64_IN_MEM: Tmp1 = X86::FpIST64m; break;
3352 addFullAddress(BuildMI(BB, Tmp1, 5), AM).addReg(ValReg);
3354 // Reload the original control word now.
3355 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
3359 case ISD::TRUNCSTORE: { // truncstore chain, val, ptr, SRCVALUE, storety
3361 MVT::ValueType StoredTy = cast<VTSDNode>(N.getOperand(4))->getVT();
3362 assert((StoredTy == MVT::i1 || StoredTy == MVT::f32 ||
3363 StoredTy == MVT::i16 /*FIXME: THIS IS JUST FOR TESTING!*/)
3364 && "Unsupported TRUNCSTORE for this target!");
3366 if (StoredTy == MVT::i16) {
3367 // FIXME: This is here just to allow testing. X86 doesn't really have a
3368 // TRUNCSTORE i16 operation, but this is required for targets that do not
3369 // have 16-bit integer registers. We occasionally disable 16-bit integer
3370 // registers to test the promotion code.
3371 Select(N.getOperand(0));
3372 Tmp1 = SelectExpr(N.getOperand(1));
3373 SelectAddress(N.getOperand(2), AM);
3375 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
3376 addFullAddress(BuildMI(BB, X86::MOV16mr, 5), AM).addReg(X86::AX);
3380 // Store of constant bool?
3381 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3382 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
3383 Select(N.getOperand(0));
3384 SelectAddress(N.getOperand(2), AM);
3386 SelectAddress(N.getOperand(2), AM);
3387 Select(N.getOperand(0));
3389 addFullAddress(BuildMI(BB, X86::MOV8mi, 5), AM).addImm(CN->getValue());
3394 default: assert(0 && "Cannot truncstore this type!");
3395 case MVT::i1: Opc = X86::MOV8mr; break;
3397 assert(!X86ScalarSSE && "Cannot truncstore scalar SSE regs");
3398 Opc = X86::FpST32m; break;
3401 std::vector<std::pair<unsigned, unsigned> > RP;
3402 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
3403 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
3404 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
3405 std::sort(RP.begin(), RP.end());
3407 Tmp1 = 0; // Silence a warning.
3408 for (unsigned i = 0; i != 3; ++i)
3409 switch (RP[2-i].second) {
3410 default: assert(0 && "Unknown operand number!");
3411 case 0: Select(N.getOperand(0)); break;
3412 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
3413 case 2: SelectAddress(N.getOperand(2), AM); break;
3416 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
3422 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
3424 switch (CN->getValueType(0)) {
3425 default: assert(0 && "Invalid type for operation!");
3427 case MVT::i8: Opc = X86::MOV8mi; break;
3428 case MVT::i16: Opc = X86::MOV16mi; break;
3429 case MVT::i32: Opc = X86::MOV32mi; break;
3432 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
3433 Select(N.getOperand(0));
3434 SelectAddress(N.getOperand(2), AM);
3436 SelectAddress(N.getOperand(2), AM);
3437 Select(N.getOperand(0));
3439 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addImm(CN->getValue());
3442 } else if (GlobalAddressSDNode *GA =
3443 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
3444 assert(GA->getValueType(0) == MVT::i32 && "Bad pointer operand");
3446 if (getRegPressure(N.getOperand(0)) > getRegPressure(N.getOperand(2))) {
3447 Select(N.getOperand(0));
3448 SelectAddress(N.getOperand(2), AM);
3450 SelectAddress(N.getOperand(2), AM);
3451 Select(N.getOperand(0));
3453 GlobalValue *GV = GA->getGlobal();
3454 // For Darwin, external and weak symbols are indirect, so we want to load
3455 // the value at address GV, not the value of GV itself.
3456 if (Subtarget->getIndirectExternAndWeakGlobals() &&
3457 (GV->hasWeakLinkage() || GV->isExternal())) {
3458 Tmp1 = MakeReg(MVT::i32);
3459 BuildMI(BB, X86::MOV32rm, 4, Tmp1).addReg(0).addZImm(1).addReg(0)
3460 .addGlobalAddress(GV, false, 0);
3461 addFullAddress(BuildMI(BB, X86::MOV32mr, 4+1),AM).addReg(Tmp1);
3463 addFullAddress(BuildMI(BB, X86::MOV32mi, 4+1),AM).addGlobalAddress(GV);
3468 // Check to see if this is a load/op/store combination.
3469 if (TryToFoldLoadOpStore(Node))
3472 switch (N.getOperand(1).getValueType()) {
3473 default: assert(0 && "Cannot store this type!");
3475 case MVT::i8: Opc = X86::MOV8mr; break;
3476 case MVT::i16: Opc = X86::MOV16mr; break;
3477 case MVT::i32: Opc = X86::MOV32mr; break;
3478 case MVT::f32: Opc = X86::MOVSSmr; break;
3479 case MVT::f64: Opc = X86ScalarSSE ? X86::MOVSDmr : X86::FpST64m; break;
3482 std::vector<std::pair<unsigned, unsigned> > RP;
3483 RP.push_back(std::make_pair(getRegPressure(N.getOperand(0)), 0));
3484 RP.push_back(std::make_pair(getRegPressure(N.getOperand(1)), 1));
3485 RP.push_back(std::make_pair(getRegPressure(N.getOperand(2)), 2));
3486 std::sort(RP.begin(), RP.end());
3488 Tmp1 = 0; // Silence a warning.
3489 for (unsigned i = 0; i != 3; ++i)
3490 switch (RP[2-i].second) {
3491 default: assert(0 && "Unknown operand number!");
3492 case 0: Select(N.getOperand(0)); break;
3493 case 1: Tmp1 = SelectExpr(N.getOperand(1)); break;
3494 case 2: SelectAddress(N.getOperand(2), AM); break;
3497 addFullAddress(BuildMI(BB, Opc, 4+1), AM).addReg(Tmp1);
3500 case ISD::CALLSEQ_START:
3501 Select(N.getOperand(0));
3503 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
3504 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addImm(Tmp1);
3506 case ISD::CALLSEQ_END:
3507 Select(N.getOperand(0));
3510 Select(N.getOperand(0)); // Select the chain.
3512 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
3513 if (Align == 0) Align = 1;
3515 // Turn the byte code into # iterations
3518 if (ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Node->getOperand(2))) {
3519 unsigned Val = ValC->getValue() & 255;
3521 // If the value is a constant, then we can potentially use larger sets.
3522 switch (Align & 3) {
3523 case 2: // WORD aligned
3524 CountReg = MakeReg(MVT::i32);
3525 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
3526 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
3528 unsigned ByteReg = SelectExpr(Node->getOperand(3));
3529 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
3531 BuildMI(BB, X86::MOV16ri, 1, X86::AX).addImm((Val << 8) | Val);
3532 Opcode = X86::REP_STOSW;
3534 case 0: // DWORD aligned
3535 CountReg = MakeReg(MVT::i32);
3536 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
3537 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
3539 unsigned ByteReg = SelectExpr(Node->getOperand(3));
3540 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
3542 Val = (Val << 8) | Val;
3543 BuildMI(BB, X86::MOV32ri, 1, X86::EAX).addImm((Val << 16) | Val);
3544 Opcode = X86::REP_STOSD;
3546 default: // BYTE aligned
3547 CountReg = SelectExpr(Node->getOperand(3));
3548 BuildMI(BB, X86::MOV8ri, 1, X86::AL).addImm(Val);
3549 Opcode = X86::REP_STOSB;
3553 // If it's not a constant value we are storing, just fall back. We could
3554 // try to be clever to form 16 bit and 32 bit values, but we don't yet.
3555 unsigned ValReg = SelectExpr(Node->getOperand(2));
3556 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(ValReg);
3557 CountReg = SelectExpr(Node->getOperand(3));
3558 Opcode = X86::REP_STOSB;
3561 // No matter what the alignment is, we put the source in ESI, the
3562 // destination in EDI, and the count in ECX.
3563 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
3564 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
3565 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
3566 BuildMI(BB, Opcode, 0);
3570 Select(N.getOperand(0)); // Select the chain.
3572 (unsigned)cast<ConstantSDNode>(Node->getOperand(4))->getValue();
3573 if (Align == 0) Align = 1;
3575 // Turn the byte code into # iterations
3578 switch (Align & 3) {
3579 case 2: // WORD aligned
3580 CountReg = MakeReg(MVT::i32);
3581 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
3582 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/2);
3584 unsigned ByteReg = SelectExpr(Node->getOperand(3));
3585 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
3587 Opcode = X86::REP_MOVSW;
3589 case 0: // DWORD aligned
3590 CountReg = MakeReg(MVT::i32);
3591 if (ConstantSDNode *I = dyn_cast<ConstantSDNode>(Node->getOperand(3))) {
3592 BuildMI(BB, X86::MOV32ri, 1, CountReg).addImm(I->getValue()/4);
3594 unsigned ByteReg = SelectExpr(Node->getOperand(3));
3595 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
3597 Opcode = X86::REP_MOVSD;
3599 default: // BYTE aligned
3600 CountReg = SelectExpr(Node->getOperand(3));
3601 Opcode = X86::REP_MOVSB;
3605 // No matter what the alignment is, we put the source in ESI, the
3606 // destination in EDI, and the count in ECX.
3607 unsigned TmpReg1 = SelectExpr(Node->getOperand(1));
3608 unsigned TmpReg2 = SelectExpr(Node->getOperand(2));
3609 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
3610 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
3611 BuildMI(BB, X86::MOV32rr, 1, X86::ESI).addReg(TmpReg2);
3612 BuildMI(BB, Opcode, 0);
3615 case ISD::WRITEPORT:
3616 if (Node->getOperand(2).getValueType() != MVT::i16) {
3617 std::cerr << "llvm.writeport: Address size is not 16 bits\n";
3620 Select(Node->getOperand(0)); // Emit the chain.
3622 Tmp1 = SelectExpr(Node->getOperand(1));
3623 switch (Node->getOperand(1).getValueType()) {
3625 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(Tmp1);
3626 Tmp2 = X86::OUT8ir; Opc = X86::OUT8rr;
3629 BuildMI(BB, X86::MOV16rr, 1, X86::AX).addReg(Tmp1);
3630 Tmp2 = X86::OUT16ir; Opc = X86::OUT16rr;
3633 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Tmp1);
3634 Tmp2 = X86::OUT32ir; Opc = X86::OUT32rr;
3637 std::cerr << "llvm.writeport: invalid data type for X86 target";
3641 // If the port is a single-byte constant, use the immediate form.
3642 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Node->getOperand(2)))
3643 if ((CN->getValue() & 255) == CN->getValue()) {
3644 BuildMI(BB, Tmp2, 1).addImm(CN->getValue());
3648 // Otherwise, move the I/O port address into the DX register.
3649 unsigned Reg = SelectExpr(Node->getOperand(2));
3650 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Reg);
3651 BuildMI(BB, Opc, 0);
3654 assert(0 && "Should not be reached!");
3658 /// createX86ISelPattern - This pass converts an LLVM function
3659 /// into a machine code representation using pattern matching and a machine
3660 /// description file.
3662 FunctionPass *llvm::createX86ISelPattern(TargetMachine &TM) {
3663 return new ISel(TM);