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>()),
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 CanBeFoldedBy(SDNode *N, SDNode *U);
122 // Include the pieces autogenerated from the target description.
123 #include "X86GenDAGISel.inc"
126 void DetermineReachability(SDNode *f, SDNode *t);
128 void Select(SDOperand &Result, SDOperand N);
130 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM, bool isRoot = true);
131 bool SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
132 SDOperand &Index, SDOperand &Disp);
133 bool SelectLEAAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
134 SDOperand &Index, SDOperand &Disp);
135 bool TryFoldLoad(SDOperand P, SDOperand N,
136 SDOperand &Base, SDOperand &Scale,
137 SDOperand &Index, SDOperand &Disp);
138 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
139 /// inline asm expressions.
140 virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
142 std::vector<SDOperand> &OutOps,
145 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
147 inline void getAddressOperands(X86ISelAddressMode &AM, SDOperand &Base,
148 SDOperand &Scale, SDOperand &Index,
150 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
151 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, MVT::i32) : AM.Base.Reg;
152 Scale = getI8Imm(AM.Scale);
154 Disp = AM.GV ? CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp)
156 CurDAG->getTargetConstantPool(AM.CP, MVT::i32, AM.Align, AM.Disp)
157 : getI32Imm(AM.Disp));
160 /// getI8Imm - Return a target constant with the specified value, of type
162 inline SDOperand getI8Imm(unsigned Imm) {
163 return CurDAG->getTargetConstant(Imm, MVT::i8);
166 /// getI16Imm - Return a target constant with the specified value, of type
168 inline SDOperand getI16Imm(unsigned Imm) {
169 return CurDAG->getTargetConstant(Imm, MVT::i16);
172 /// getI32Imm - Return a target constant with the specified value, of type
174 inline SDOperand getI32Imm(unsigned Imm) {
175 return CurDAG->getTargetConstant(Imm, MVT::i32);
178 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
179 /// base register. Return the virtual register that holds this value.
180 SDOperand getGlobalBaseReg();
182 /// DAGSize - Number of nodes in the DAG.
186 /// TopOrder - Topological ordering of all nodes in the DAG.
188 std::vector<SDNode*> TopOrder;
190 /// ReachabilityMatrix - A N x N matrix representing all pairs reachability
191 /// information. One bit per potential edge.
192 std::vector<bool> ReachabilityMatrix;
194 /// RMRange - The range of reachability information available for the
195 /// particular source node.
196 std::vector<unsigned> ReachMatrixRange;
198 inline void setReachable(SDNode *f, SDNode *t) {
199 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
200 ReachabilityMatrix[Idx] = true;
203 inline bool isReachable(SDNode *f, SDNode *t) {
204 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
205 return ReachabilityMatrix[Idx];
208 /// UnfoldableSet - An boolean array representing nodes which have been
209 /// folded into addressing modes and therefore should not be folded in
210 /// another operation.
211 std::vector<bool> UnfoldableSet;
213 inline void setUnfoldable(SDNode *N) {
214 UnfoldableSet[N->getNodeId()] = true;
217 inline bool isUnfoldable(SDNode *N) {
218 return UnfoldableSet[N->getNodeId()];
227 bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U) {
228 // Is it already folded by SelectAddr / SelectLEAAddr?
232 // If U use can somehow reach N through another path then U can't fold N or
233 // it will create a cycle. e.g. In the following diagram, U can reach N
234 // through X. If N is folded into into U, then X is both a predecessor and
244 DetermineReachability(U, N);
245 assert(isReachable(U, N) && "Attempting to fold a non-operand node?");
246 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end(); I != E; ++I) {
248 if (P != N && isReachable(P, N))
254 /// DetermineReachability - Determine reachability between all pairs of nodes
255 /// between f and t in topological order.
256 void X86DAGToDAGISel::DetermineReachability(SDNode *f, SDNode *t) {
257 unsigned Orderf = f->getNodeId();
258 unsigned Ordert = t->getNodeId();
259 unsigned Range = ReachMatrixRange[Orderf];
265 for (unsigned i = Range; i < Ordert; ++i) {
266 SDNode *N = TopOrder[i];
268 // If N is a leaf node, there is nothing more to do.
269 if (N->getNumOperands() == 0)
272 for (unsigned i2 = Orderf; ; ++i2) {
273 SDNode *M = TopOrder[i2];
274 if (isReachable(M, N)) {
275 // Update reachability from M to N's operands.
276 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E;++I)
277 setReachable(M, I->Val);
283 ReachMatrixRange[Orderf] = Ordert;
286 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
287 /// when it has created a SelectionDAG for us to codegen.
288 void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
290 MachineFunction::iterator FirstMBB = BB;
292 TopOrder = DAG.AssignTopologicalOrder();
293 DAGSize = TopOrder.size();
294 ReachabilityMatrix.assign(DAGSize*DAGSize, false);
295 ReachMatrixRange.assign(DAGSize, 0);
296 UnfoldableSet.assign(DAGSize, false);
298 // Codegen the basic block.
300 DEBUG(std::cerr << "===== Instruction selection begins:\n");
303 DAG.setRoot(SelectRoot(DAG.getRoot()));
305 DEBUG(std::cerr << "===== Instruction selection ends:\n");
311 DAG.RemoveDeadNodes();
313 // Emit machine code to BB.
314 ScheduleAndEmitDAG(DAG);
316 // If we are emitting FP stack code, scan the basic block to determine if this
317 // block defines any FP values. If so, put an FP_REG_KILL instruction before
318 // the terminator of the block.
319 if (!Subtarget->hasSSE2()) {
320 // Note that FP stack instructions *are* used in SSE code when returning
321 // values, but these are not live out of the basic block, so we don't need
322 // an FP_REG_KILL in this case either.
323 bool ContainsFPCode = false;
325 // Scan all of the machine instructions in these MBBs, checking for FP
327 MachineFunction::iterator MBBI = FirstMBB;
329 for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
330 !ContainsFPCode && I != E; ++I) {
331 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
332 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
333 MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
334 RegMap->getRegClass(I->getOperand(0).getReg()) ==
335 X86::RFPRegisterClass) {
336 ContainsFPCode = true;
341 } while (!ContainsFPCode && &*(MBBI++) != BB);
343 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
344 // a copy of the input value in this block.
345 if (!ContainsFPCode) {
346 // Final check, check LLVM BB's that are successors to the LLVM BB
347 // corresponding to BB for FP PHI nodes.
348 const BasicBlock *LLVMBB = BB->getBasicBlock();
350 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
351 !ContainsFPCode && SI != E; ++SI) {
352 for (BasicBlock::const_iterator II = SI->begin();
353 (PN = dyn_cast<PHINode>(II)); ++II) {
354 if (PN->getType()->isFloatingPoint()) {
355 ContainsFPCode = true;
362 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
363 if (ContainsFPCode) {
364 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
370 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
371 /// the main function.
372 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
373 MachineFrameInfo *MFI) {
374 if (Subtarget->TargetType == X86Subtarget::isCygwin)
375 BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("__main");
377 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
378 int CWFrameIdx = MFI->CreateStackObject(2, 2);
379 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
381 // Set the high part to be 64-bit precision.
382 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
383 CWFrameIdx, 1).addImm(2);
385 // Reload the modified control word now.
386 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
389 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
390 // If this is main, emit special code for main.
391 MachineBasicBlock *BB = MF.begin();
392 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
393 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
396 /// MatchAddress - Add the specified node to the specified addressing mode,
397 /// returning true if it cannot be done. This just pattern matches for the
399 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
401 bool Available = false;
402 // If N has already been selected, reuse the result unless in some very
404 std::map<SDOperand, SDOperand>::iterator CGMI= CodeGenMap.find(N.getValue(0));
405 if (CGMI != CodeGenMap.end()) {
409 switch (N.getOpcode()) {
412 AM.Disp += cast<ConstantSDNode>(N)->getValue();
415 case X86ISD::Wrapper:
416 // If both base and index components have been picked, we can't fit
417 // the result available in the register in the addressing mode. Duplicate
418 // GlobalAddress or ConstantPool as displacement.
419 if (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val)) {
420 if (ConstantPoolSDNode *CP =
421 dyn_cast<ConstantPoolSDNode>(N.getOperand(0))) {
424 AM.Align = CP->getAlignment();
425 AM.Disp += CP->getOffset();
428 } else if (GlobalAddressSDNode *G =
429 dyn_cast<GlobalAddressSDNode>(N.getOperand(0))) {
431 AM.GV = G->getGlobal();
432 AM.Disp += G->getOffset();
439 case ISD::FrameIndex:
440 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
441 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
442 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
448 if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
449 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
450 unsigned Val = CN->getValue();
451 if (Val == 1 || Val == 2 || Val == 3) {
453 SDOperand ShVal = N.Val->getOperand(0);
455 // Okay, we know that we have a scale by now. However, if the scaled
456 // value is an add of something and a constant, we can fold the
457 // constant into the disp field here.
458 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
459 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
460 AM.IndexReg = ShVal.Val->getOperand(0);
461 ConstantSDNode *AddVal =
462 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
463 AM.Disp += AddVal->getValue() << Val;
473 // X*[3,5,9] -> X+X*[2,4,8]
475 AM.BaseType == X86ISelAddressMode::RegBase &&
476 AM.Base.Reg.Val == 0 &&
477 AM.IndexReg.Val == 0)
478 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
479 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
480 AM.Scale = unsigned(CN->getValue())-1;
482 SDOperand MulVal = N.Val->getOperand(0);
485 // Okay, we know that we have a scale by now. However, if the scaled
486 // value is an add of something and a constant, we can fold the
487 // constant into the disp field here.
488 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
489 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
490 Reg = MulVal.Val->getOperand(0);
491 ConstantSDNode *AddVal =
492 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
493 AM.Disp += AddVal->getValue() * CN->getValue();
495 Reg = N.Val->getOperand(0);
498 AM.IndexReg = AM.Base.Reg = Reg;
505 X86ISelAddressMode Backup = AM;
506 if (!MatchAddress(N.Val->getOperand(0), AM, false) &&
507 !MatchAddress(N.Val->getOperand(1), AM, false))
510 if (!MatchAddress(N.Val->getOperand(1), AM, false) &&
511 !MatchAddress(N.Val->getOperand(0), AM, false))
520 X86ISelAddressMode Backup = AM;
521 // Look for (x << c1) | c2 where (c2 < c1)
522 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(0));
523 if (CN && !MatchAddress(N.Val->getOperand(1), AM, false)) {
524 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
525 AM.Disp = CN->getValue();
530 CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1));
531 if (CN && !MatchAddress(N.Val->getOperand(0), AM, false)) {
532 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
533 AM.Disp = CN->getValue();
543 // Is the base register already occupied?
544 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
545 // If so, check to see if the scale index register is set.
546 if (AM.IndexReg.Val == 0) {
552 // Otherwise, we cannot select it.
556 // Default, generate it as a register.
557 AM.BaseType = X86ISelAddressMode::RegBase;
562 /// SelectAddr - returns true if it is able pattern match an addressing mode.
563 /// It returns the operands which make up the maximal addressing mode it can
564 /// match by reference.
565 bool X86DAGToDAGISel::SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
566 SDOperand &Index, SDOperand &Disp) {
567 X86ISelAddressMode AM;
568 if (MatchAddress(N, AM))
571 if (AM.BaseType == X86ISelAddressMode::RegBase) {
572 if (!AM.Base.Reg.Val)
573 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
576 if (!AM.IndexReg.Val)
577 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
579 getAddressOperands(AM, Base, Scale, Index, Disp);
581 int Id = Base.Val ? Base.Val->getNodeId() : -1;
583 setUnfoldable(Base.Val);
584 Id = Index.Val ? Index.Val->getNodeId() : -1;
586 setUnfoldable(Index.Val);
591 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
592 /// mode it matches can be cost effectively emitted as an LEA instruction.
593 bool X86DAGToDAGISel::SelectLEAAddr(SDOperand N, SDOperand &Base,
595 SDOperand &Index, SDOperand &Disp) {
596 X86ISelAddressMode AM;
597 if (MatchAddress(N, AM))
600 unsigned Complexity = 0;
601 if (AM.BaseType == X86ISelAddressMode::RegBase)
605 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
606 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
612 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
616 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg
617 else if (AM.Scale > 1)
620 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
621 // to a LEA. This is determined with some expermentation but is by no means
622 // optimal (especially for code size consideration). LEA is nice because of
623 // its three-address nature. Tweak the cost function again when we can run
624 // convertToThreeAddress() at register allocation time.
628 if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
631 if (Complexity > 2) {
632 getAddressOperands(AM, Base, Scale, Index, Disp);
636 int Id = Base.Val ? Base.Val->getNodeId() : -1;
638 setUnfoldable(Base.Val);
639 Id = Index.Val ? Index.Val->getNodeId() : -1;
641 setUnfoldable(Index.Val);
646 bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
647 SDOperand &Base, SDOperand &Scale,
648 SDOperand &Index, SDOperand &Disp) {
649 if (N.getOpcode() == ISD::LOAD &&
651 !CodeGenMap.count(N.getValue(0)) &&
652 !CanBeFoldedBy(N.Val, P.Val))
653 return SelectAddr(N.getOperand(1), Base, Scale, Index, Disp);
657 static bool isRegister0(SDOperand Op) {
658 if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op))
659 return (R->getReg() == 0);
663 /// getGlobalBaseReg - Output the instructions required to put the
664 /// base address to use for accessing globals into a register.
666 SDOperand X86DAGToDAGISel::getGlobalBaseReg() {
667 if (!GlobalBaseReg) {
668 // Insert the set of GlobalBaseReg into the first MBB of the function
669 MachineBasicBlock &FirstMBB = BB->getParent()->front();
670 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
671 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
672 // FIXME: when we get to LP64, we will need to create the appropriate
673 // type of register here.
674 GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
675 BuildMI(FirstMBB, MBBI, X86::MovePCtoStack, 0);
676 BuildMI(FirstMBB, MBBI, X86::POP32r, 1, GlobalBaseReg);
678 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
681 static SDNode *FindCallStartFromCall(SDNode *Node) {
682 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
683 assert(Node->getOperand(0).getValueType() == MVT::Other &&
684 "Node doesn't have a token chain argument!");
685 return FindCallStartFromCall(Node->getOperand(0).Val);
688 void X86DAGToDAGISel::Select(SDOperand &Result, SDOperand N) {
689 SDNode *Node = N.Val;
690 MVT::ValueType NVT = Node->getValueType(0);
692 unsigned Opcode = Node->getOpcode();
695 DEBUG(std::cerr << std::string(Indent, ' '));
696 DEBUG(std::cerr << "Selecting: ");
697 DEBUG(Node->dump(CurDAG));
698 DEBUG(std::cerr << "\n");
702 if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
705 DEBUG(std::cerr << std::string(Indent-2, ' '));
706 DEBUG(std::cerr << "== ");
707 DEBUG(Node->dump(CurDAG));
708 DEBUG(std::cerr << "\n");
711 return; // Already selected.
714 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(N);
715 if (CGMI != CodeGenMap.end()) {
716 Result = CGMI->second;
718 DEBUG(std::cerr << std::string(Indent-2, ' '));
719 DEBUG(std::cerr << "== ");
720 DEBUG(Result.Val->dump(CurDAG));
721 DEBUG(std::cerr << "\n");
729 case X86ISD::GlobalBaseReg:
730 Result = getGlobalBaseReg();
734 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
735 // code and is matched first so to prevent it from being turned into
737 SDOperand N0 = N.getOperand(0);
738 SDOperand N1 = N.getOperand(1);
739 if (N.Val->getValueType(0) == MVT::i32 &&
740 N0.getOpcode() == X86ISD::Wrapper &&
741 N1.getOpcode() == ISD::Constant) {
742 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
744 // TODO: handle ExternalSymbolSDNode.
745 if (GlobalAddressSDNode *G =
746 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
747 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), MVT::i32,
748 G->getOffset() + Offset);
749 } else if (ConstantPoolSDNode *CP =
750 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
751 C = CurDAG->getTargetConstantPool(CP->get(), MVT::i32,
753 CP->getOffset()+Offset);
757 if (N.Val->hasOneUse()) {
758 Result = CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, MVT::i32, C);
760 SDNode *ResNode = CurDAG->getTargetNode(X86::MOV32ri, MVT::i32, C);
761 Result = CodeGenMap[N] = SDOperand(ResNode, 0);
767 // Other cases are handled by auto-generated code.
773 if (Opcode == ISD::MULHU)
775 default: assert(0 && "Unsupported VT!");
776 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
777 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
778 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
782 default: assert(0 && "Unsupported VT!");
783 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
784 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
785 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
788 unsigned LoReg, HiReg;
790 default: assert(0 && "Unsupported VT!");
791 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
792 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
793 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
796 SDOperand N0 = Node->getOperand(0);
797 SDOperand N1 = Node->getOperand(1);
799 bool foldedLoad = false;
800 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
801 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
802 // MULHU and MULHS are commmutative
804 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
806 N0 = Node->getOperand(1);
807 N1 = Node->getOperand(0);
813 Select(Chain, N1.getOperand(0));
815 Chain = CurDAG->getEntryNode();
817 SDOperand InFlag(0, 0);
819 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
821 InFlag = Chain.getValue(1);
829 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
830 Tmp2, Tmp3, Chain, InFlag);
831 Chain = SDOperand(CNode, 0);
832 InFlag = SDOperand(CNode, 1);
836 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
839 Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
840 CodeGenMap[N.getValue(0)] = Result;
842 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
843 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
847 DEBUG(std::cerr << std::string(Indent-2, ' '));
848 DEBUG(std::cerr << "== ");
849 DEBUG(Result.Val->dump(CurDAG));
850 DEBUG(std::cerr << "\n");
860 bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
861 bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
864 default: assert(0 && "Unsupported VT!");
865 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
866 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
867 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
871 default: assert(0 && "Unsupported VT!");
872 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
873 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
874 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
877 unsigned LoReg, HiReg;
878 unsigned ClrOpcode, SExtOpcode;
880 default: assert(0 && "Unsupported VT!");
882 LoReg = X86::AL; HiReg = X86::AH;
883 ClrOpcode = X86::MOV8r0;
884 SExtOpcode = X86::CBW;
887 LoReg = X86::AX; HiReg = X86::DX;
888 ClrOpcode = X86::MOV16r0;
889 SExtOpcode = X86::CWD;
892 LoReg = X86::EAX; HiReg = X86::EDX;
893 ClrOpcode = X86::MOV32r0;
894 SExtOpcode = X86::CDQ;
898 SDOperand N0 = Node->getOperand(0);
899 SDOperand N1 = Node->getOperand(1);
901 bool foldedLoad = false;
902 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
903 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
906 Select(Chain, N1.getOperand(0));
908 Chain = CurDAG->getEntryNode();
910 SDOperand InFlag(0, 0);
912 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
914 InFlag = Chain.getValue(1);
917 // Sign extend the low part into the high part.
919 SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
921 // Zero out the high part, effectively zero extending the input.
922 SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
923 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(HiReg, NVT),
925 InFlag = Chain.getValue(1);
934 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
935 Tmp2, Tmp3, Chain, InFlag);
936 Chain = SDOperand(CNode, 0);
937 InFlag = SDOperand(CNode, 1);
941 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
944 Result = CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg,
946 CodeGenMap[N.getValue(0)] = Result;
948 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
949 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
953 DEBUG(std::cerr << std::string(Indent-2, ' '));
954 DEBUG(std::cerr << "== ");
955 DEBUG(Result.Val->dump(CurDAG));
956 DEBUG(std::cerr << "\n");
962 case ISD::TRUNCATE: {
963 if (NVT == MVT::i8) {
966 switch (Node->getOperand(0).getValueType()) {
967 default: assert(0 && "Unknown truncate!");
969 Opc = X86::MOV16to16_;
971 Opc2 = X86::TRUNC_GR16_GR8;
974 Opc = X86::MOV32to32_;
976 Opc2 = X86::TRUNC_GR32_GR8;
980 SDOperand Tmp0, Tmp1;
981 Select(Tmp0, Node->getOperand(0));
982 Tmp1 = SDOperand(CurDAG->getTargetNode(Opc, VT, Tmp0), 0);
983 Result = CodeGenMap[N] =
984 SDOperand(CurDAG->getTargetNode(Opc2, NVT, Tmp1), 0);
987 DEBUG(std::cerr << std::string(Indent-2, ' '));
988 DEBUG(std::cerr << "== ");
989 DEBUG(Result.Val->dump(CurDAG));
990 DEBUG(std::cerr << "\n");
1000 SelectCode(Result, N);
1002 DEBUG(std::cerr << std::string(Indent-2, ' '));
1003 DEBUG(std::cerr << "=> ");
1004 DEBUG(Result.Val->dump(CurDAG));
1005 DEBUG(std::cerr << "\n");
1010 bool X86DAGToDAGISel::
1011 SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
1012 std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
1013 SDOperand Op0, Op1, Op2, Op3;
1014 switch (ConstraintCode) {
1015 case 'o': // offsetable ??
1016 case 'v': // not offsetable ??
1017 default: return true;
1019 if (!SelectAddr(Op, Op0, Op1, Op2, Op3))
1025 Select(OutOps[0], Op0);
1026 Select(OutOps[1], Op1);
1027 Select(OutOps[2], Op2);
1028 Select(OutOps[3], Op3);
1032 /// createX86ISelDag - This pass converts a legalized DAG into a
1033 /// X86-specific DAG, ready for instruction scheduling.
1035 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM) {
1036 return new X86DAGToDAGISel(TM);