1 //===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the Evan Cheng and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a DAG pattern matching instruction selector for X86,
11 // converting from a legalized dag to a X86 dag.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/GlobalValue.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/SSARegMap.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/Visibility.h"
35 #include "llvm/ADT/Statistic.h"
41 //===----------------------------------------------------------------------===//
42 // Pattern Matcher Implementation
43 //===----------------------------------------------------------------------===//
46 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
47 /// SDOperand's instead of register numbers for the leaves of the matched
49 struct X86ISelAddressMode {
55 struct { // This is really a union, discriminated by BaseType!
65 unsigned Align; // CP alignment.
68 : BaseType(RegBase), Scale(1), IndexReg(), Disp(0), GV(0),
76 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
78 //===--------------------------------------------------------------------===//
79 /// ISel - X86 specific code to select X86 machine instructions for
80 /// SelectionDAG operations.
82 class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
83 /// ContainsFPCode - Every instruction we select that uses or defines a FP
84 /// register should set this to true.
87 /// X86Lowering - This object fully describes how to lower LLVM code to an
88 /// X86-specific SelectionDAG.
89 X86TargetLowering X86Lowering;
91 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
92 /// make the right decision when generating code for different targets.
93 const X86Subtarget *Subtarget;
95 unsigned GlobalBaseReg;
98 X86DAGToDAGISel(X86TargetMachine &TM)
99 : SelectionDAGISel(X86Lowering),
100 X86Lowering(*TM.getTargetLowering()),
101 Subtarget(&TM.getSubtarget<X86Subtarget>()),
102 DAGSize(0), ReachibilityMatrix(NULL) {}
104 virtual bool runOnFunction(Function &Fn) {
105 // Make sure we re-emit a set of the global base reg if necessary
107 return SelectionDAGISel::runOnFunction(Fn);
110 virtual const char *getPassName() const {
111 return "X86 DAG->DAG Instruction Selection";
114 /// InstructionSelectBasicBlock - This callback is invoked by
115 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
116 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
118 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
120 virtual bool IsFoldableBy(SDNode *N, SDNode *U);
122 // Include the pieces autogenerated from the target description.
123 #include "X86GenDAGISel.inc"
126 void DetermineTopologicalOrdering();
127 void DeterminReachibility(SDNode *f, SDNode *t);
129 void Select(SDOperand &Result, SDOperand N);
131 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM, bool isRoot = true);
132 bool SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
133 SDOperand &Index, SDOperand &Disp);
134 bool SelectLEAAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
135 SDOperand &Index, SDOperand &Disp);
136 bool TryFoldLoad(SDOperand P, SDOperand N,
137 SDOperand &Base, SDOperand &Scale,
138 SDOperand &Index, SDOperand &Disp);
139 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
140 /// inline asm expressions.
141 virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
143 std::vector<SDOperand> &OutOps,
146 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
148 inline void getAddressOperands(X86ISelAddressMode &AM, SDOperand &Base,
149 SDOperand &Scale, SDOperand &Index,
151 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
152 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, MVT::i32) : AM.Base.Reg;
153 Scale = getI8Imm(AM.Scale);
155 Disp = AM.GV ? CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp)
157 CurDAG->getTargetConstantPool(AM.CP, MVT::i32, AM.Align, AM.Disp)
158 : getI32Imm(AM.Disp));
161 /// getI8Imm - Return a target constant with the specified value, of type
163 inline SDOperand getI8Imm(unsigned Imm) {
164 return CurDAG->getTargetConstant(Imm, MVT::i8);
167 /// getI16Imm - Return a target constant with the specified value, of type
169 inline SDOperand getI16Imm(unsigned Imm) {
170 return CurDAG->getTargetConstant(Imm, MVT::i16);
173 /// getI32Imm - Return a target constant with the specified value, of type
175 inline SDOperand getI32Imm(unsigned Imm) {
176 return CurDAG->getTargetConstant(Imm, MVT::i32);
179 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
180 /// base register. Return the virtual register that holds this value.
181 SDOperand getGlobalBaseReg();
183 /// DAGSize - Number of nodes in the DAG.
187 /// TopOrder - Topological ordering of all nodes in the DAG.
191 /// IdToOrder - Node id to topological order map.
195 /// RMRange - The range of reachibility information available for the
196 /// particular source node.
199 /// ReachibilityMatrix - A N x N matrix representing all pairs reachibility
200 /// information. One bit per potential edge.
201 unsigned char *ReachibilityMatrix;
203 inline void setReachable(SDNode *f, SDNode *t) {
204 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
205 ReachibilityMatrix[Idx / 8] |= 1 << (Idx % 8);
208 inline bool isReachable(SDNode *f, SDNode *t) {
209 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
210 return ReachibilityMatrix[Idx / 8] & (1 << (Idx % 8));
219 bool X86DAGToDAGISel::IsFoldableBy(SDNode *N, SDNode *U) {
220 // If U use can somehow reach N through another path then U can't fold N or
221 // it will create a cycle. e.g. In the following diagram, U can reach N
222 // through X. If N is foled into into U, then X is both a predecessor and
232 DeterminReachibility(U, N);
233 assert(isReachable(U, N) && "Attempting to fold a non-operand node?");
234 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end(); I != E; ++I) {
236 if (P != N && isReachable(P, N))
242 /// DetermineTopologicalOrdering - Determine topological ordering of the nodes
244 void X86DAGToDAGISel::DetermineTopologicalOrdering() {
245 DAGSize = CurDAG->AssignNodeIds();
246 TopOrder = new SDNode*[DAGSize];
247 IdToOrder = new unsigned[DAGSize];
248 memset(IdToOrder, 0, DAGSize * sizeof(unsigned));
249 RMRange = new unsigned[DAGSize];
250 memset(RMRange, 0, DAGSize * sizeof(unsigned));
252 std::vector<unsigned> InDegree(DAGSize);
253 std::list<SDNode*> Sources;
254 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
255 E = CurDAG->allnodes_end(); I != E; ++I) {
257 unsigned Degree = N->use_size();
258 InDegree[N->getNodeId()] = Degree;
260 Sources.push_back(I);
264 while (!Sources.empty()) {
265 SDNode *N = Sources.front();
268 IdToOrder[N->getNodeId()] = Order;
270 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) {
272 int PId = P->getNodeId();
273 unsigned Degree = InDegree[PId] - 1;
275 Sources.push_back(P);
276 InDegree[PId] = Degree;
281 void X86DAGToDAGISel::DeterminReachibility(SDNode *f, SDNode *t) {
282 if (!ReachibilityMatrix) {
283 DetermineTopologicalOrdering();
284 unsigned RMSize = (DAGSize * DAGSize + 7) / 8;
285 ReachibilityMatrix = new unsigned char[RMSize];
286 memset(ReachibilityMatrix, 0, RMSize);
289 int Idf = f->getNodeId();
290 int Idt = t->getNodeId();
291 unsigned Orderf = IdToOrder[Idf];
292 unsigned Ordert = IdToOrder[Idt];
293 unsigned Range = RMRange[Idf];
299 for (unsigned i = Range; i < Ordert; ++i) {
300 SDNode *N = TopOrder[i];
302 // If N is a leaf node, there is nothing more to do.
303 if (N->getNumOperands() == 0)
306 for (unsigned i2 = Orderf; ; ++i2) {
307 SDNode *M = TopOrder[i2];
308 if (isReachable(M, N)) {
309 // Update reachibility from M to N's operands.
310 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E;++I)
311 setReachable(M, I->Val);
317 RMRange[Idf] = Ordert;
320 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
321 /// when it has created a SelectionDAG for us to codegen.
322 void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
324 MachineFunction::iterator FirstMBB = BB;
326 // Codegen the basic block.
328 DEBUG(std::cerr << "===== Instruction selection begins:\n");
331 DAG.setRoot(SelectRoot(DAG.getRoot()));
332 assert(InFlightSet.empty() && "ISel InFlightSet has not been emptied!");
334 DEBUG(std::cerr << "===== Instruction selection ends:\n");
336 if (ReachibilityMatrix) {
337 delete[] ReachibilityMatrix;
341 ReachibilityMatrix = NULL;
343 IdToOrder = RMRange = NULL;
348 DAG.RemoveDeadNodes();
350 // Emit machine code to BB.
351 ScheduleAndEmitDAG(DAG);
353 // If we are emitting FP stack code, scan the basic block to determine if this
354 // block defines any FP values. If so, put an FP_REG_KILL instruction before
355 // the terminator of the block.
356 if (!Subtarget->hasSSE2()) {
357 // Note that FP stack instructions *are* used in SSE code when returning
358 // values, but these are not live out of the basic block, so we don't need
359 // an FP_REG_KILL in this case either.
360 bool ContainsFPCode = false;
362 // Scan all of the machine instructions in these MBBs, checking for FP
364 MachineFunction::iterator MBBI = FirstMBB;
366 for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
367 !ContainsFPCode && I != E; ++I) {
368 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
369 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
370 MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
371 RegMap->getRegClass(I->getOperand(0).getReg()) ==
372 X86::RFPRegisterClass) {
373 ContainsFPCode = true;
378 } while (!ContainsFPCode && &*(MBBI++) != BB);
380 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
381 // a copy of the input value in this block.
382 if (!ContainsFPCode) {
383 // Final check, check LLVM BB's that are successors to the LLVM BB
384 // corresponding to BB for FP PHI nodes.
385 const BasicBlock *LLVMBB = BB->getBasicBlock();
387 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
388 !ContainsFPCode && SI != E; ++SI) {
389 for (BasicBlock::const_iterator II = SI->begin();
390 (PN = dyn_cast<PHINode>(II)); ++II) {
391 if (PN->getType()->isFloatingPoint()) {
392 ContainsFPCode = true;
399 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
400 if (ContainsFPCode) {
401 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
407 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
408 /// the main function.
409 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
410 MachineFrameInfo *MFI) {
411 if (Subtarget->TargetType == X86Subtarget::isCygwin)
412 BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("__main");
414 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
415 int CWFrameIdx = MFI->CreateStackObject(2, 2);
416 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
418 // Set the high part to be 64-bit precision.
419 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
420 CWFrameIdx, 1).addImm(2);
422 // Reload the modified control word now.
423 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
426 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
427 // If this is main, emit special code for main.
428 MachineBasicBlock *BB = MF.begin();
429 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
430 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
433 /// MatchAddress - Add the specified node to the specified addressing mode,
434 /// returning true if it cannot be done. This just pattern matches for the
436 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
438 bool Available = false;
439 // If N has already been selected, reuse the result unless in some very
441 std::map<SDOperand, SDOperand>::iterator CGMI= CodeGenMap.find(N.getValue(0));
442 if (CGMI != CodeGenMap.end()) {
446 switch (N.getOpcode()) {
449 AM.Disp += cast<ConstantSDNode>(N)->getValue();
452 case X86ISD::Wrapper:
453 // If both base and index components have been picked, we can't fit
454 // the result available in the register in the addressing mode. Duplicate
455 // GlobalAddress or ConstantPool as displacement.
456 if (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val)) {
457 if (ConstantPoolSDNode *CP =
458 dyn_cast<ConstantPoolSDNode>(N.getOperand(0))) {
461 AM.Align = CP->getAlignment();
462 AM.Disp += CP->getOffset();
465 } else if (GlobalAddressSDNode *G =
466 dyn_cast<GlobalAddressSDNode>(N.getOperand(0))) {
468 AM.GV = G->getGlobal();
469 AM.Disp += G->getOffset();
476 case ISD::FrameIndex:
477 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
478 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
479 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
485 if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
486 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
487 unsigned Val = CN->getValue();
488 if (Val == 1 || Val == 2 || Val == 3) {
490 SDOperand ShVal = N.Val->getOperand(0);
492 // Okay, we know that we have a scale by now. However, if the scaled
493 // value is an add of something and a constant, we can fold the
494 // constant into the disp field here.
495 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
496 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
497 AM.IndexReg = ShVal.Val->getOperand(0);
498 ConstantSDNode *AddVal =
499 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
500 AM.Disp += AddVal->getValue() << Val;
510 // X*[3,5,9] -> X+X*[2,4,8]
512 AM.BaseType == X86ISelAddressMode::RegBase &&
513 AM.Base.Reg.Val == 0 &&
514 AM.IndexReg.Val == 0)
515 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
516 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
517 AM.Scale = unsigned(CN->getValue())-1;
519 SDOperand MulVal = N.Val->getOperand(0);
522 // Okay, we know that we have a scale by now. However, if the scaled
523 // value is an add of something and a constant, we can fold the
524 // constant into the disp field here.
525 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
526 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
527 Reg = MulVal.Val->getOperand(0);
528 ConstantSDNode *AddVal =
529 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
530 AM.Disp += AddVal->getValue() * CN->getValue();
532 Reg = N.Val->getOperand(0);
535 AM.IndexReg = AM.Base.Reg = Reg;
542 X86ISelAddressMode Backup = AM;
543 if (!MatchAddress(N.Val->getOperand(0), AM, false) &&
544 !MatchAddress(N.Val->getOperand(1), AM, false))
547 if (!MatchAddress(N.Val->getOperand(1), AM, false) &&
548 !MatchAddress(N.Val->getOperand(0), AM, false))
557 X86ISelAddressMode Backup = AM;
558 // Look for (x << c1) | c2 where (c2 < c1)
559 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(0));
560 if (CN && !MatchAddress(N.Val->getOperand(1), AM, false)) {
561 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
562 AM.Disp = CN->getValue();
567 CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1));
568 if (CN && !MatchAddress(N.Val->getOperand(0), AM, false)) {
569 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
570 AM.Disp = CN->getValue();
580 // Is the base register already occupied?
581 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
582 // If so, check to see if the scale index register is set.
583 if (AM.IndexReg.Val == 0) {
589 // Otherwise, we cannot select it.
593 // Default, generate it as a register.
594 AM.BaseType = X86ISelAddressMode::RegBase;
599 /// SelectAddr - returns true if it is able pattern match an addressing mode.
600 /// It returns the operands which make up the maximal addressing mode it can
601 /// match by reference.
602 bool X86DAGToDAGISel::SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
603 SDOperand &Index, SDOperand &Disp) {
604 X86ISelAddressMode AM;
605 if (MatchAddress(N, AM))
608 if (AM.BaseType == X86ISelAddressMode::RegBase) {
609 if (!AM.Base.Reg.Val)
610 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
613 if (!AM.IndexReg.Val)
614 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
616 getAddressOperands(AM, Base, Scale, Index, Disp);
621 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
622 /// mode it matches can be cost effectively emitted as an LEA instruction.
623 bool X86DAGToDAGISel::SelectLEAAddr(SDOperand N, SDOperand &Base,
625 SDOperand &Index, SDOperand &Disp) {
626 X86ISelAddressMode AM;
627 if (MatchAddress(N, AM))
630 unsigned Complexity = 0;
631 if (AM.BaseType == X86ISelAddressMode::RegBase)
635 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
636 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
642 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
646 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg
647 else if (AM.Scale > 1)
650 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
651 // to a LEA. This is determined with some expermentation but is by no means
652 // optimal (especially for code size consideration). LEA is nice because of
653 // its three-address nature. Tweak the cost function again when we can run
654 // convertToThreeAddress() at register allocation time.
658 if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
661 if (Complexity > 2) {
662 getAddressOperands(AM, Base, Scale, Index, Disp);
669 bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
670 SDOperand &Base, SDOperand &Scale,
671 SDOperand &Index, SDOperand &Disp) {
672 if (N.getOpcode() == ISD::LOAD &&
674 !CodeGenMap.count(N.getValue(0)) &&
675 !IsFoldableBy(N.Val, P.Val))
676 return SelectAddr(N.getOperand(1), Base, Scale, Index, Disp);
680 static bool isRegister0(SDOperand Op) {
681 if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op))
682 return (R->getReg() == 0);
686 /// getGlobalBaseReg - Output the instructions required to put the
687 /// base address to use for accessing globals into a register.
689 SDOperand X86DAGToDAGISel::getGlobalBaseReg() {
690 if (!GlobalBaseReg) {
691 // Insert the set of GlobalBaseReg into the first MBB of the function
692 MachineBasicBlock &FirstMBB = BB->getParent()->front();
693 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
694 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
695 // FIXME: when we get to LP64, we will need to create the appropriate
696 // type of register here.
697 GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
698 BuildMI(FirstMBB, MBBI, X86::MovePCtoStack, 0);
699 BuildMI(FirstMBB, MBBI, X86::POP32r, 1, GlobalBaseReg);
701 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
704 static SDNode *FindCallStartFromCall(SDNode *Node) {
705 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
706 assert(Node->getOperand(0).getValueType() == MVT::Other &&
707 "Node doesn't have a token chain argument!");
708 return FindCallStartFromCall(Node->getOperand(0).Val);
711 void X86DAGToDAGISel::Select(SDOperand &Result, SDOperand N) {
712 SDNode *Node = N.Val;
713 MVT::ValueType NVT = Node->getValueType(0);
715 unsigned Opcode = Node->getOpcode();
718 DEBUG(std::cerr << std::string(Indent, ' '));
719 DEBUG(std::cerr << "Selecting: ");
720 DEBUG(Node->dump(CurDAG));
721 DEBUG(std::cerr << "\n");
725 if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
728 DEBUG(std::cerr << std::string(Indent-2, ' '));
729 DEBUG(std::cerr << "== ");
730 DEBUG(Node->dump(CurDAG));
731 DEBUG(std::cerr << "\n");
734 return; // Already selected.
737 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(N);
738 if (CGMI != CodeGenMap.end()) {
739 Result = CGMI->second;
741 DEBUG(std::cerr << std::string(Indent-2, ' '));
742 DEBUG(std::cerr << "== ");
743 DEBUG(Result.Val->dump(CurDAG));
744 DEBUG(std::cerr << "\n");
752 case X86ISD::GlobalBaseReg:
753 Result = getGlobalBaseReg();
757 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
758 // code and is matched first so to prevent it from being turned into
760 SDOperand N0 = N.getOperand(0);
761 SDOperand N1 = N.getOperand(1);
762 if (N.Val->getValueType(0) == MVT::i32 &&
763 N0.getOpcode() == X86ISD::Wrapper &&
764 N1.getOpcode() == ISD::Constant) {
765 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
767 // TODO: handle ExternalSymbolSDNode.
768 if (GlobalAddressSDNode *G =
769 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
770 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), MVT::i32,
771 G->getOffset() + Offset);
772 } else if (ConstantPoolSDNode *CP =
773 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
774 C = CurDAG->getTargetConstantPool(CP->get(), MVT::i32,
776 CP->getOffset()+Offset);
780 if (N.Val->hasOneUse()) {
781 Result = CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, MVT::i32, C);
783 SDNode *ResNode = CurDAG->getTargetNode(X86::MOV32ri, MVT::i32, C);
784 Result = CodeGenMap[N] = SDOperand(ResNode, 0);
790 // Other cases are handled by auto-generated code.
796 if (Opcode == ISD::MULHU)
798 default: assert(0 && "Unsupported VT!");
799 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
800 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
801 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
805 default: assert(0 && "Unsupported VT!");
806 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
807 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
808 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
811 unsigned LoReg, HiReg;
813 default: assert(0 && "Unsupported VT!");
814 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
815 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
816 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
819 SDOperand N0 = Node->getOperand(0);
820 SDOperand N1 = Node->getOperand(1);
822 bool foldedLoad = false;
823 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
824 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
825 // MULHU and MULHS are commmutative
827 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
829 N0 = Node->getOperand(1);
830 N1 = Node->getOperand(0);
836 Select(Chain, N1.getOperand(0));
838 Chain = CurDAG->getEntryNode();
840 SDOperand InFlag(0, 0);
842 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
844 InFlag = Chain.getValue(1);
852 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
853 Tmp2, Tmp3, Chain, InFlag);
854 Chain = SDOperand(CNode, 0);
855 InFlag = SDOperand(CNode, 1);
859 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
862 Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
863 CodeGenMap[N.getValue(0)] = Result;
865 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
866 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
870 DEBUG(std::cerr << std::string(Indent-2, ' '));
871 DEBUG(std::cerr << "== ");
872 DEBUG(Result.Val->dump(CurDAG));
873 DEBUG(std::cerr << "\n");
883 bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
884 bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
887 default: assert(0 && "Unsupported VT!");
888 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
889 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
890 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
894 default: assert(0 && "Unsupported VT!");
895 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
896 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
897 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
900 unsigned LoReg, HiReg;
901 unsigned ClrOpcode, SExtOpcode;
903 default: assert(0 && "Unsupported VT!");
905 LoReg = X86::AL; HiReg = X86::AH;
906 ClrOpcode = X86::MOV8r0;
907 SExtOpcode = X86::CBW;
910 LoReg = X86::AX; HiReg = X86::DX;
911 ClrOpcode = X86::MOV16r0;
912 SExtOpcode = X86::CWD;
915 LoReg = X86::EAX; HiReg = X86::EDX;
916 ClrOpcode = X86::MOV32r0;
917 SExtOpcode = X86::CDQ;
921 SDOperand N0 = Node->getOperand(0);
922 SDOperand N1 = Node->getOperand(1);
924 bool foldedLoad = false;
925 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
926 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
929 Select(Chain, N1.getOperand(0));
931 Chain = CurDAG->getEntryNode();
933 SDOperand InFlag(0, 0);
935 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
937 InFlag = Chain.getValue(1);
940 // Sign extend the low part into the high part.
942 SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
944 // Zero out the high part, effectively zero extending the input.
945 SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
946 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(HiReg, NVT),
948 InFlag = Chain.getValue(1);
957 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
958 Tmp2, Tmp3, Chain, InFlag);
959 Chain = SDOperand(CNode, 0);
960 InFlag = SDOperand(CNode, 1);
964 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
967 Result = CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg,
969 CodeGenMap[N.getValue(0)] = Result;
971 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
972 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
976 DEBUG(std::cerr << std::string(Indent-2, ' '));
977 DEBUG(std::cerr << "== ");
978 DEBUG(Result.Val->dump(CurDAG));
979 DEBUG(std::cerr << "\n");
985 case ISD::TRUNCATE: {
986 if (NVT == MVT::i8) {
989 switch (Node->getOperand(0).getValueType()) {
990 default: assert(0 && "Unknown truncate!");
992 Opc = X86::MOV16to16_;
994 Opc2 = X86::TRUNC_GR16_GR8;
997 Opc = X86::MOV32to32_;
999 Opc2 = X86::TRUNC_GR32_GR8;
1003 SDOperand Tmp0, Tmp1;
1004 Select(Tmp0, Node->getOperand(0));
1005 Tmp1 = SDOperand(CurDAG->getTargetNode(Opc, VT, Tmp0), 0);
1006 Result = CodeGenMap[N] =
1007 SDOperand(CurDAG->getTargetNode(Opc2, NVT, Tmp1), 0);
1010 DEBUG(std::cerr << std::string(Indent-2, ' '));
1011 DEBUG(std::cerr << "== ");
1012 DEBUG(Result.Val->dump(CurDAG));
1013 DEBUG(std::cerr << "\n");
1023 SelectCode(Result, N);
1025 DEBUG(std::cerr << std::string(Indent-2, ' '));
1026 DEBUG(std::cerr << "=> ");
1027 DEBUG(Result.Val->dump(CurDAG));
1028 DEBUG(std::cerr << "\n");
1033 bool X86DAGToDAGISel::
1034 SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
1035 std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
1036 SDOperand Op0, Op1, Op2, Op3;
1037 switch (ConstraintCode) {
1038 case 'o': // offsetable ??
1039 case 'v': // not offsetable ??
1040 default: return true;
1042 if (!SelectAddr(Op, Op0, Op1, Op2, Op3))
1048 Select(OutOps[0], Op0);
1049 Select(OutOps[1], Op1);
1050 Select(OutOps[2], Op2);
1051 Select(OutOps[3], Op3);
1055 /// createX86ISelDag - This pass converts a legalized DAG into a
1056 /// X86-specific DAG, ready for instruction scheduling.
1058 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM) {
1059 return new X86DAGToDAGISel(TM);