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();
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.
189 std::vector<SDNode*> TopOrder;
191 /// ReachibilityMatrix - A N x N matrix representing all pairs reachibility
192 /// information. One bit per potential edge.
193 unsigned char *ReachibilityMatrix;
195 inline void setReachable(SDNode *f, SDNode *t) {
196 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
197 ReachibilityMatrix[Idx / 8] |= 1 << (Idx % 8);
200 inline bool isReachable(SDNode *f, SDNode *t) {
201 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
202 return ReachibilityMatrix[Idx / 8] & (1 << (Idx % 8));
211 bool X86DAGToDAGISel::IsFoldableBy(SDNode *N, SDNode *U) {
212 // If U use can somehow reach N through another path then U can't fold N or
213 // it will create a cycle. e.g. In the following diagram, U can reach N
214 // through X. If N is foled into into U, then X is both a predecessor and
224 if (!ReachibilityMatrix)
225 DeterminReachibility();
226 assert(isReachable(U, N) && "Attempting to fold a non-operand node?");
227 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end(); I != E; ++I) {
229 if (P != N && isReachable(P, N))
235 /// DetermineTopologicalOrdering - Determine topological ordering of the nodes
237 void X86DAGToDAGISel::DetermineTopologicalOrdering() {
238 DAGSize = CurDAG->AssignNodeIds();
239 TopOrder.reserve(DAGSize);
241 std::vector<unsigned> InDegree(DAGSize);
242 std::list<SDNode*> Sources;
243 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
244 E = CurDAG->allnodes_end(); I != E; ++I) {
246 unsigned Degree = N->use_size();
247 InDegree[N->getNodeId()] = Degree;
249 Sources.push_back(I);
253 while (!Sources.empty()) {
254 SDNode *N = Sources.front();
258 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) {
260 int PId = P->getNodeId();
261 unsigned Degree = InDegree[PId] - 1;
263 Sources.push_back(P);
264 InDegree[PId] = Degree;
269 void X86DAGToDAGISel::DeterminReachibility() {
270 DetermineTopologicalOrdering();
271 ReachibilityMatrix = new unsigned char[DAGSize * DAGSize];
272 memset(ReachibilityMatrix, 0, DAGSize * DAGSize * sizeof(unsigned char));
274 for (unsigned i = 0; i < DAGSize; ++i) {
275 SDNode *N = TopOrder[i];
277 // If N is a leaf node, there is nothing more to do.
278 if (N->getNumOperands() == 0)
281 for (unsigned i2 = 0; ; ++i2) {
282 SDNode *M = TopOrder[i2];
283 if (isReachable(M, N)) {
284 // Update reachibility from M to N's operands.
285 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E;++I)
286 setReachable(M, I->Val);
293 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
294 /// when it has created a SelectionDAG for us to codegen.
295 void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
297 MachineFunction::iterator FirstMBB = BB;
299 // Codegen the basic block.
301 DEBUG(std::cerr << "===== Instruction selection begins:\n");
304 DAG.setRoot(SelectRoot(DAG.getRoot()));
305 assert(InFlightSet.empty() && "ISel InFlightSet has not been emptied!");
307 DEBUG(std::cerr << "===== Instruction selection ends:\n");
309 if (ReachibilityMatrix) {
310 delete[] ReachibilityMatrix;
311 ReachibilityMatrix = NULL;
316 DAG.RemoveDeadNodes();
318 // Emit machine code to BB.
319 ScheduleAndEmitDAG(DAG);
321 // If we are emitting FP stack code, scan the basic block to determine if this
322 // block defines any FP values. If so, put an FP_REG_KILL instruction before
323 // the terminator of the block.
324 if (!Subtarget->hasSSE2()) {
325 // Note that FP stack instructions *are* used in SSE code when returning
326 // values, but these are not live out of the basic block, so we don't need
327 // an FP_REG_KILL in this case either.
328 bool ContainsFPCode = false;
330 // Scan all of the machine instructions in these MBBs, checking for FP
332 MachineFunction::iterator MBBI = FirstMBB;
334 for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
335 !ContainsFPCode && I != E; ++I) {
336 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
337 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
338 MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
339 RegMap->getRegClass(I->getOperand(0).getReg()) ==
340 X86::RFPRegisterClass) {
341 ContainsFPCode = true;
346 } while (!ContainsFPCode && &*(MBBI++) != BB);
348 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
349 // a copy of the input value in this block.
350 if (!ContainsFPCode) {
351 // Final check, check LLVM BB's that are successors to the LLVM BB
352 // corresponding to BB for FP PHI nodes.
353 const BasicBlock *LLVMBB = BB->getBasicBlock();
355 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
356 !ContainsFPCode && SI != E; ++SI) {
357 for (BasicBlock::const_iterator II = SI->begin();
358 (PN = dyn_cast<PHINode>(II)); ++II) {
359 if (PN->getType()->isFloatingPoint()) {
360 ContainsFPCode = true;
367 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
368 if (ContainsFPCode) {
369 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
375 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
376 /// the main function.
377 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
378 MachineFrameInfo *MFI) {
379 if (Subtarget->TargetType == X86Subtarget::isCygwin)
380 BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("__main");
382 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
383 int CWFrameIdx = MFI->CreateStackObject(2, 2);
384 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
386 // Set the high part to be 64-bit precision.
387 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
388 CWFrameIdx, 1).addImm(2);
390 // Reload the modified control word now.
391 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
394 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
395 // If this is main, emit special code for main.
396 MachineBasicBlock *BB = MF.begin();
397 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
398 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
401 /// MatchAddress - Add the specified node to the specified addressing mode,
402 /// returning true if it cannot be done. This just pattern matches for the
404 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
406 bool Available = false;
407 // If N has already been selected, reuse the result unless in some very
409 std::map<SDOperand, SDOperand>::iterator CGMI= CodeGenMap.find(N.getValue(0));
410 if (CGMI != CodeGenMap.end()) {
414 switch (N.getOpcode()) {
417 AM.Disp += cast<ConstantSDNode>(N)->getValue();
420 case X86ISD::Wrapper:
421 // If both base and index components have been picked, we can't fit
422 // the result available in the register in the addressing mode. Duplicate
423 // GlobalAddress or ConstantPool as displacement.
424 if (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val)) {
425 if (ConstantPoolSDNode *CP =
426 dyn_cast<ConstantPoolSDNode>(N.getOperand(0))) {
429 AM.Align = CP->getAlignment();
430 AM.Disp += CP->getOffset();
433 } else if (GlobalAddressSDNode *G =
434 dyn_cast<GlobalAddressSDNode>(N.getOperand(0))) {
436 AM.GV = G->getGlobal();
437 AM.Disp += G->getOffset();
444 case ISD::FrameIndex:
445 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
446 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
447 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
453 if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
454 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
455 unsigned Val = CN->getValue();
456 if (Val == 1 || Val == 2 || Val == 3) {
458 SDOperand ShVal = N.Val->getOperand(0);
460 // Okay, we know that we have a scale by now. However, if the scaled
461 // value is an add of something and a constant, we can fold the
462 // constant into the disp field here.
463 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
464 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
465 AM.IndexReg = ShVal.Val->getOperand(0);
466 ConstantSDNode *AddVal =
467 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
468 AM.Disp += AddVal->getValue() << Val;
478 // X*[3,5,9] -> X+X*[2,4,8]
480 AM.BaseType == X86ISelAddressMode::RegBase &&
481 AM.Base.Reg.Val == 0 &&
482 AM.IndexReg.Val == 0)
483 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
484 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
485 AM.Scale = unsigned(CN->getValue())-1;
487 SDOperand MulVal = N.Val->getOperand(0);
490 // Okay, we know that we have a scale by now. However, if the scaled
491 // value is an add of something and a constant, we can fold the
492 // constant into the disp field here.
493 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
494 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
495 Reg = MulVal.Val->getOperand(0);
496 ConstantSDNode *AddVal =
497 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
498 AM.Disp += AddVal->getValue() * CN->getValue();
500 Reg = N.Val->getOperand(0);
503 AM.IndexReg = AM.Base.Reg = Reg;
510 X86ISelAddressMode Backup = AM;
511 if (!MatchAddress(N.Val->getOperand(0), AM, false) &&
512 !MatchAddress(N.Val->getOperand(1), AM, false))
515 if (!MatchAddress(N.Val->getOperand(1), AM, false) &&
516 !MatchAddress(N.Val->getOperand(0), AM, false))
525 X86ISelAddressMode Backup = AM;
526 // Look for (x << c1) | c2 where (c2 < c1)
527 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(0));
528 if (CN && !MatchAddress(N.Val->getOperand(1), AM, false)) {
529 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
530 AM.Disp = CN->getValue();
535 CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1));
536 if (CN && !MatchAddress(N.Val->getOperand(0), AM, false)) {
537 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
538 AM.Disp = CN->getValue();
548 // Is the base register already occupied?
549 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
550 // If so, check to see if the scale index register is set.
551 if (AM.IndexReg.Val == 0) {
557 // Otherwise, we cannot select it.
561 // Default, generate it as a register.
562 AM.BaseType = X86ISelAddressMode::RegBase;
567 /// SelectAddr - returns true if it is able pattern match an addressing mode.
568 /// It returns the operands which make up the maximal addressing mode it can
569 /// match by reference.
570 bool X86DAGToDAGISel::SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
571 SDOperand &Index, SDOperand &Disp) {
572 X86ISelAddressMode AM;
573 if (MatchAddress(N, AM))
576 if (AM.BaseType == X86ISelAddressMode::RegBase) {
577 if (!AM.Base.Reg.Val)
578 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
581 if (!AM.IndexReg.Val)
582 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
584 getAddressOperands(AM, Base, Scale, Index, Disp);
589 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
590 /// mode it matches can be cost effectively emitted as an LEA instruction.
591 bool X86DAGToDAGISel::SelectLEAAddr(SDOperand N, SDOperand &Base,
593 SDOperand &Index, SDOperand &Disp) {
594 X86ISelAddressMode AM;
595 if (MatchAddress(N, AM))
598 unsigned Complexity = 0;
599 if (AM.BaseType == X86ISelAddressMode::RegBase)
603 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
604 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
610 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
614 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg
615 else if (AM.Scale > 1)
618 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
619 // to a LEA. This is determined with some expermentation but is by no means
620 // optimal (especially for code size consideration). LEA is nice because of
621 // its three-address nature. Tweak the cost function again when we can run
622 // convertToThreeAddress() at register allocation time.
626 if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
629 if (Complexity > 2) {
630 getAddressOperands(AM, Base, Scale, Index, Disp);
637 bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
638 SDOperand &Base, SDOperand &Scale,
639 SDOperand &Index, SDOperand &Disp) {
640 if (N.getOpcode() == ISD::LOAD &&
642 !CodeGenMap.count(N.getValue(0)) &&
643 !IsFoldableBy(N.Val, P.Val))
644 return SelectAddr(N.getOperand(1), Base, Scale, Index, Disp);
648 static bool isRegister0(SDOperand Op) {
649 if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op))
650 return (R->getReg() == 0);
654 /// getGlobalBaseReg - Output the instructions required to put the
655 /// base address to use for accessing globals into a register.
657 SDOperand X86DAGToDAGISel::getGlobalBaseReg() {
658 if (!GlobalBaseReg) {
659 // Insert the set of GlobalBaseReg into the first MBB of the function
660 MachineBasicBlock &FirstMBB = BB->getParent()->front();
661 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
662 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
663 // FIXME: when we get to LP64, we will need to create the appropriate
664 // type of register here.
665 GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
666 BuildMI(FirstMBB, MBBI, X86::MovePCtoStack, 0);
667 BuildMI(FirstMBB, MBBI, X86::POP32r, 1, GlobalBaseReg);
669 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
672 static SDNode *FindCallStartFromCall(SDNode *Node) {
673 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
674 assert(Node->getOperand(0).getValueType() == MVT::Other &&
675 "Node doesn't have a token chain argument!");
676 return FindCallStartFromCall(Node->getOperand(0).Val);
679 void X86DAGToDAGISel::Select(SDOperand &Result, SDOperand N) {
680 SDNode *Node = N.Val;
681 MVT::ValueType NVT = Node->getValueType(0);
683 unsigned Opcode = Node->getOpcode();
686 DEBUG(std::cerr << std::string(Indent, ' '));
687 DEBUG(std::cerr << "Selecting: ");
688 DEBUG(Node->dump(CurDAG));
689 DEBUG(std::cerr << "\n");
693 if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
696 DEBUG(std::cerr << std::string(Indent-2, ' '));
697 DEBUG(std::cerr << "== ");
698 DEBUG(Node->dump(CurDAG));
699 DEBUG(std::cerr << "\n");
702 return; // Already selected.
705 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(N);
706 if (CGMI != CodeGenMap.end()) {
707 Result = CGMI->second;
709 DEBUG(std::cerr << std::string(Indent-2, ' '));
710 DEBUG(std::cerr << "== ");
711 DEBUG(Result.Val->dump(CurDAG));
712 DEBUG(std::cerr << "\n");
720 case X86ISD::GlobalBaseReg:
721 Result = getGlobalBaseReg();
725 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
726 // code and is matched first so to prevent it from being turned into
728 SDOperand N0 = N.getOperand(0);
729 SDOperand N1 = N.getOperand(1);
730 if (N.Val->getValueType(0) == MVT::i32 &&
731 N0.getOpcode() == X86ISD::Wrapper &&
732 N1.getOpcode() == ISD::Constant) {
733 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
735 // TODO: handle ExternalSymbolSDNode.
736 if (GlobalAddressSDNode *G =
737 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
738 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), MVT::i32,
739 G->getOffset() + Offset);
740 } else if (ConstantPoolSDNode *CP =
741 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
742 C = CurDAG->getTargetConstantPool(CP->get(), MVT::i32,
744 CP->getOffset()+Offset);
748 if (N.Val->hasOneUse()) {
749 Result = CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, MVT::i32, C);
751 SDNode *ResNode = CurDAG->getTargetNode(X86::MOV32ri, MVT::i32, C);
752 Result = CodeGenMap[N] = SDOperand(ResNode, 0);
758 // Other cases are handled by auto-generated code.
764 if (Opcode == ISD::MULHU)
766 default: assert(0 && "Unsupported VT!");
767 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
768 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
769 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
773 default: assert(0 && "Unsupported VT!");
774 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
775 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
776 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
779 unsigned LoReg, HiReg;
781 default: assert(0 && "Unsupported VT!");
782 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
783 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
784 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
787 SDOperand N0 = Node->getOperand(0);
788 SDOperand N1 = Node->getOperand(1);
790 bool foldedLoad = false;
791 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
792 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
793 // MULHU and MULHS are commmutative
795 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
797 N0 = Node->getOperand(1);
798 N1 = Node->getOperand(0);
804 Select(Chain, N1.getOperand(0));
806 Chain = CurDAG->getEntryNode();
808 SDOperand InFlag(0, 0);
810 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
812 InFlag = Chain.getValue(1);
820 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
821 Tmp2, Tmp3, Chain, InFlag);
822 Chain = SDOperand(CNode, 0);
823 InFlag = SDOperand(CNode, 1);
827 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
830 Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
831 CodeGenMap[N.getValue(0)] = Result;
833 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
834 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
838 DEBUG(std::cerr << std::string(Indent-2, ' '));
839 DEBUG(std::cerr << "== ");
840 DEBUG(Result.Val->dump(CurDAG));
841 DEBUG(std::cerr << "\n");
851 bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
852 bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
855 default: assert(0 && "Unsupported VT!");
856 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
857 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
858 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
862 default: assert(0 && "Unsupported VT!");
863 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
864 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
865 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
868 unsigned LoReg, HiReg;
869 unsigned ClrOpcode, SExtOpcode;
871 default: assert(0 && "Unsupported VT!");
873 LoReg = X86::AL; HiReg = X86::AH;
874 ClrOpcode = X86::MOV8r0;
875 SExtOpcode = X86::CBW;
878 LoReg = X86::AX; HiReg = X86::DX;
879 ClrOpcode = X86::MOV16r0;
880 SExtOpcode = X86::CWD;
883 LoReg = X86::EAX; HiReg = X86::EDX;
884 ClrOpcode = X86::MOV32r0;
885 SExtOpcode = X86::CDQ;
889 SDOperand N0 = Node->getOperand(0);
890 SDOperand N1 = Node->getOperand(1);
892 bool foldedLoad = false;
893 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
894 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
897 Select(Chain, N1.getOperand(0));
899 Chain = CurDAG->getEntryNode();
901 SDOperand InFlag(0, 0);
903 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
905 InFlag = Chain.getValue(1);
908 // Sign extend the low part into the high part.
910 SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
912 // Zero out the high part, effectively zero extending the input.
913 SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
914 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(HiReg, NVT),
916 InFlag = Chain.getValue(1);
925 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
926 Tmp2, Tmp3, Chain, InFlag);
927 Chain = SDOperand(CNode, 0);
928 InFlag = SDOperand(CNode, 1);
932 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
935 Result = CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg,
937 CodeGenMap[N.getValue(0)] = Result;
939 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
940 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
944 DEBUG(std::cerr << std::string(Indent-2, ' '));
945 DEBUG(std::cerr << "== ");
946 DEBUG(Result.Val->dump(CurDAG));
947 DEBUG(std::cerr << "\n");
953 case ISD::TRUNCATE: {
954 if (NVT == MVT::i8) {
957 switch (Node->getOperand(0).getValueType()) {
958 default: assert(0 && "Unknown truncate!");
960 Opc = X86::MOV16to16_;
962 Opc2 = X86::TRUNC_GR16_GR8;
965 Opc = X86::MOV32to32_;
967 Opc2 = X86::TRUNC_GR32_GR8;
971 SDOperand Tmp0, Tmp1;
972 Select(Tmp0, Node->getOperand(0));
973 Tmp1 = SDOperand(CurDAG->getTargetNode(Opc, VT, Tmp0), 0);
974 Result = CodeGenMap[N] =
975 SDOperand(CurDAG->getTargetNode(Opc2, NVT, Tmp1), 0);
978 DEBUG(std::cerr << std::string(Indent-2, ' '));
979 DEBUG(std::cerr << "== ");
980 DEBUG(Result.Val->dump(CurDAG));
981 DEBUG(std::cerr << "\n");
991 SelectCode(Result, N);
993 DEBUG(std::cerr << std::string(Indent-2, ' '));
994 DEBUG(std::cerr << "=> ");
995 DEBUG(Result.Val->dump(CurDAG));
996 DEBUG(std::cerr << "\n");
1001 bool X86DAGToDAGISel::
1002 SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
1003 std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
1004 SDOperand Op0, Op1, Op2, Op3;
1005 switch (ConstraintCode) {
1006 case 'o': // offsetable ??
1007 case 'v': // not offsetable ??
1008 default: return true;
1010 if (!SelectAddr(Op, Op0, Op1, Op2, Op3))
1016 Select(OutOps[0], Op0);
1017 Select(OutOps[1], Op1);
1018 Select(OutOps[2], Op2);
1019 Select(OutOps[3], Op3);
1023 /// createX86ISelDag - This pass converts a legalized DAG into a
1024 /// X86-specific DAG, ready for instruction scheduling.
1026 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM) {
1027 return new X86DAGToDAGISel(TM);