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), ReachabilityMatrix(NULL), ReachMatrixRange(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 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 unsigned char *ReachabilityMatrix;
194 /// ReachMatrixRange - The range of reachability information available for
195 /// the particular source node.
196 unsigned *ReachMatrixRange;
198 inline void setReachable(SDNode *f, SDNode *t) {
199 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
200 ReachabilityMatrix[Idx / 8] |= 1 << (Idx % 8);
203 inline bool isReachable(SDNode *f, SDNode *t) {
204 unsigned Idx = f->getNodeId() * DAGSize + t->getNodeId();
205 return ReachabilityMatrix[Idx / 8] & (1 << (Idx % 8));
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 unsigned char *UnfoldableSet;
213 inline void setUnfoldable(SDNode *N) {
214 unsigned Id = N->getNodeId();
215 UnfoldableSet[Id / 8] |= 1 << (Id % 8);
218 inline bool isUnfoldable(SDNode *N) {
219 unsigned Id = N->getNodeId();
220 return UnfoldableSet[Id / 8] & (1 << (Id % 8));
229 bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U) {
230 // Is it already folded by SelectAddr / SelectLEAAddr?
234 // If U use can somehow reach N through another path then U can't fold N or
235 // it will create a cycle. e.g. In the following diagram, U can reach N
236 // through X. If N is folded into into U, then X is both a predecessor and
246 DetermineReachability(U, N);
247 assert(isReachable(U, N) && "Attempting to fold a non-operand node?");
248 for (SDNode::op_iterator I = U->op_begin(), E = U->op_end(); I != E; ++I) {
250 if (P != N && isReachable(P, N))
256 /// DetermineReachability - Determine reachability between all pairs of nodes
257 /// between f and t in topological order.
258 void X86DAGToDAGISel::DetermineReachability(SDNode *f, SDNode *t) {
259 unsigned Orderf = f->getNodeId();
260 unsigned Ordert = t->getNodeId();
261 unsigned Range = ReachMatrixRange[Orderf];
267 for (unsigned i = Range; i < Ordert; ++i) {
268 SDNode *N = TopOrder[i];
270 // If N is a leaf node, there is nothing more to do.
271 if (N->getNumOperands() == 0)
274 for (unsigned i2 = Range; ; ++i2) {
275 SDNode *M = TopOrder[i2];
276 if (isReachable(M, N)) {
277 // Update reachability from M to N's operands.
278 for (SDNode::op_iterator I = N->op_begin(),E = N->op_end(); I != E;++I){
280 if (P->getNodeId() >= 0)
288 ReachMatrixRange[Orderf] = Ordert;
291 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
292 /// when it has created a SelectionDAG for us to codegen.
293 void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
295 MachineFunction::iterator FirstMBB = BB;
297 DAGSize = DAG.AssignTopologicalOrder(TopOrder);
298 unsigned RMSize = (DAGSize * DAGSize + 7) / 8;
299 ReachabilityMatrix = new unsigned char[RMSize];
300 memset(ReachabilityMatrix, 0, RMSize);
301 ReachMatrixRange = new unsigned[DAGSize];
302 memset(ReachMatrixRange, 0, DAGSize * sizeof(unsigned));
303 unsigned NumBytes = (DAGSize + 7) / 8;
304 UnfoldableSet = new unsigned char[NumBytes];
305 memset(UnfoldableSet, 0, NumBytes);
307 // Codegen the basic block.
309 DEBUG(std::cerr << "===== Instruction selection begins:\n");
312 DAG.setRoot(SelectRoot(DAG.getRoot()));
314 DEBUG(std::cerr << "===== Instruction selection ends:\n");
317 delete[] ReachabilityMatrix;
318 delete[] ReachMatrixRange;
319 delete[] UnfoldableSet;
320 ReachabilityMatrix = NULL;
321 ReachMatrixRange = NULL;
322 UnfoldableSet = NULL;
327 DAG.RemoveDeadNodes();
329 // Emit machine code to BB.
330 ScheduleAndEmitDAG(DAG);
332 // If we are emitting FP stack code, scan the basic block to determine if this
333 // block defines any FP values. If so, put an FP_REG_KILL instruction before
334 // the terminator of the block.
335 if (!Subtarget->hasSSE2()) {
336 // Note that FP stack instructions *are* used in SSE code when returning
337 // values, but these are not live out of the basic block, so we don't need
338 // an FP_REG_KILL in this case either.
339 bool ContainsFPCode = false;
341 // Scan all of the machine instructions in these MBBs, checking for FP
343 MachineFunction::iterator MBBI = FirstMBB;
345 for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
346 !ContainsFPCode && I != E; ++I) {
347 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
348 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
349 MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
350 RegMap->getRegClass(I->getOperand(0).getReg()) ==
351 X86::RFPRegisterClass) {
352 ContainsFPCode = true;
357 } while (!ContainsFPCode && &*(MBBI++) != BB);
359 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
360 // a copy of the input value in this block.
361 if (!ContainsFPCode) {
362 // Final check, check LLVM BB's that are successors to the LLVM BB
363 // corresponding to BB for FP PHI nodes.
364 const BasicBlock *LLVMBB = BB->getBasicBlock();
366 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
367 !ContainsFPCode && SI != E; ++SI) {
368 for (BasicBlock::const_iterator II = SI->begin();
369 (PN = dyn_cast<PHINode>(II)); ++II) {
370 if (PN->getType()->isFloatingPoint()) {
371 ContainsFPCode = true;
378 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
379 if (ContainsFPCode) {
380 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
386 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
387 /// the main function.
388 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
389 MachineFrameInfo *MFI) {
390 if (Subtarget->TargetType == X86Subtarget::isCygwin)
391 BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("__main");
393 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
394 int CWFrameIdx = MFI->CreateStackObject(2, 2);
395 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
397 // Set the high part to be 64-bit precision.
398 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
399 CWFrameIdx, 1).addImm(2);
401 // Reload the modified control word now.
402 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
405 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
406 // If this is main, emit special code for main.
407 MachineBasicBlock *BB = MF.begin();
408 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
409 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
412 /// MatchAddress - Add the specified node to the specified addressing mode,
413 /// returning true if it cannot be done. This just pattern matches for the
415 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
417 bool Available = false;
418 // If N has already been selected, reuse the result unless in some very
420 std::map<SDOperand, SDOperand>::iterator CGMI= CodeGenMap.find(N.getValue(0));
421 if (CGMI != CodeGenMap.end()) {
425 switch (N.getOpcode()) {
428 AM.Disp += cast<ConstantSDNode>(N)->getValue();
431 case X86ISD::Wrapper:
432 // If both base and index components have been picked, we can't fit
433 // the result available in the register in the addressing mode. Duplicate
434 // GlobalAddress or ConstantPool as displacement.
435 if (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val)) {
436 if (ConstantPoolSDNode *CP =
437 dyn_cast<ConstantPoolSDNode>(N.getOperand(0))) {
440 AM.Align = CP->getAlignment();
441 AM.Disp += CP->getOffset();
444 } else if (GlobalAddressSDNode *G =
445 dyn_cast<GlobalAddressSDNode>(N.getOperand(0))) {
447 AM.GV = G->getGlobal();
448 AM.Disp += G->getOffset();
455 case ISD::FrameIndex:
456 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
457 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
458 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
464 if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
465 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
466 unsigned Val = CN->getValue();
467 if (Val == 1 || Val == 2 || Val == 3) {
469 SDOperand ShVal = N.Val->getOperand(0);
471 // Okay, we know that we have a scale by now. However, if the scaled
472 // value is an add of something and a constant, we can fold the
473 // constant into the disp field here.
474 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
475 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
476 AM.IndexReg = ShVal.Val->getOperand(0);
477 ConstantSDNode *AddVal =
478 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
479 AM.Disp += AddVal->getValue() << Val;
489 // X*[3,5,9] -> X+X*[2,4,8]
491 AM.BaseType == X86ISelAddressMode::RegBase &&
492 AM.Base.Reg.Val == 0 &&
493 AM.IndexReg.Val == 0)
494 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
495 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
496 AM.Scale = unsigned(CN->getValue())-1;
498 SDOperand MulVal = N.Val->getOperand(0);
501 // Okay, we know that we have a scale by now. However, if the scaled
502 // value is an add of something and a constant, we can fold the
503 // constant into the disp field here.
504 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
505 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
506 Reg = MulVal.Val->getOperand(0);
507 ConstantSDNode *AddVal =
508 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
509 AM.Disp += AddVal->getValue() * CN->getValue();
511 Reg = N.Val->getOperand(0);
514 AM.IndexReg = AM.Base.Reg = Reg;
521 X86ISelAddressMode Backup = AM;
522 if (!MatchAddress(N.Val->getOperand(0), AM, false) &&
523 !MatchAddress(N.Val->getOperand(1), AM, false))
526 if (!MatchAddress(N.Val->getOperand(1), AM, false) &&
527 !MatchAddress(N.Val->getOperand(0), AM, false))
536 X86ISelAddressMode Backup = AM;
537 // Look for (x << c1) | c2 where (c2 < c1)
538 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(0));
539 if (CN && !MatchAddress(N.Val->getOperand(1), AM, false)) {
540 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
541 AM.Disp = CN->getValue();
546 CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1));
547 if (CN && !MatchAddress(N.Val->getOperand(0), AM, false)) {
548 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
549 AM.Disp = CN->getValue();
559 // Is the base register already occupied?
560 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
561 // If so, check to see if the scale index register is set.
562 if (AM.IndexReg.Val == 0) {
568 // Otherwise, we cannot select it.
572 // Default, generate it as a register.
573 AM.BaseType = X86ISelAddressMode::RegBase;
578 /// SelectAddr - returns true if it is able pattern match an addressing mode.
579 /// It returns the operands which make up the maximal addressing mode it can
580 /// match by reference.
581 bool X86DAGToDAGISel::SelectAddr(SDOperand N, SDOperand &Base, SDOperand &Scale,
582 SDOperand &Index, SDOperand &Disp) {
583 X86ISelAddressMode AM;
584 if (MatchAddress(N, AM))
587 if (AM.BaseType == X86ISelAddressMode::RegBase) {
588 if (!AM.Base.Reg.Val)
589 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
592 if (!AM.IndexReg.Val)
593 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
595 getAddressOperands(AM, Base, Scale, Index, Disp);
597 int Id = Base.Val ? Base.Val->getNodeId() : -1;
599 setUnfoldable(Base.Val);
600 Id = Index.Val ? Index.Val->getNodeId() : -1;
602 setUnfoldable(Index.Val);
607 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
608 /// mode it matches can be cost effectively emitted as an LEA instruction.
609 bool X86DAGToDAGISel::SelectLEAAddr(SDOperand N, SDOperand &Base,
611 SDOperand &Index, SDOperand &Disp) {
612 X86ISelAddressMode AM;
613 if (MatchAddress(N, AM))
616 unsigned Complexity = 0;
617 if (AM.BaseType == X86ISelAddressMode::RegBase)
621 AM.Base.Reg = CurDAG->getRegister(0, MVT::i32);
622 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
628 AM.IndexReg = CurDAG->getRegister(0, MVT::i32);
632 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg
633 else if (AM.Scale > 1)
636 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
637 // to a LEA. This is determined with some expermentation but is by no means
638 // optimal (especially for code size consideration). LEA is nice because of
639 // its three-address nature. Tweak the cost function again when we can run
640 // convertToThreeAddress() at register allocation time.
644 if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
647 if (Complexity > 2) {
648 getAddressOperands(AM, Base, Scale, Index, Disp);
652 int Id = Base.Val ? Base.Val->getNodeId() : -1;
654 setUnfoldable(Base.Val);
655 Id = Index.Val ? Index.Val->getNodeId() : -1;
657 setUnfoldable(Index.Val);
662 bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
663 SDOperand &Base, SDOperand &Scale,
664 SDOperand &Index, SDOperand &Disp) {
665 if (N.getOpcode() == ISD::LOAD &&
667 !CodeGenMap.count(N.getValue(0)) &&
668 !CanBeFoldedBy(N.Val, P.Val))
669 return SelectAddr(N.getOperand(1), Base, Scale, Index, Disp);
673 static bool isRegister0(SDOperand Op) {
674 if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op))
675 return (R->getReg() == 0);
679 /// getGlobalBaseReg - Output the instructions required to put the
680 /// base address to use for accessing globals into a register.
682 SDOperand X86DAGToDAGISel::getGlobalBaseReg() {
683 if (!GlobalBaseReg) {
684 // Insert the set of GlobalBaseReg into the first MBB of the function
685 MachineBasicBlock &FirstMBB = BB->getParent()->front();
686 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
687 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
688 // FIXME: when we get to LP64, we will need to create the appropriate
689 // type of register here.
690 GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
691 BuildMI(FirstMBB, MBBI, X86::MovePCtoStack, 0);
692 BuildMI(FirstMBB, MBBI, X86::POP32r, 1, GlobalBaseReg);
694 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
697 static SDNode *FindCallStartFromCall(SDNode *Node) {
698 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
699 assert(Node->getOperand(0).getValueType() == MVT::Other &&
700 "Node doesn't have a token chain argument!");
701 return FindCallStartFromCall(Node->getOperand(0).Val);
704 void X86DAGToDAGISel::Select(SDOperand &Result, SDOperand N) {
705 SDNode *Node = N.Val;
706 MVT::ValueType NVT = Node->getValueType(0);
708 unsigned Opcode = Node->getOpcode();
711 DEBUG(std::cerr << std::string(Indent, ' '));
712 DEBUG(std::cerr << "Selecting: ");
713 DEBUG(Node->dump(CurDAG));
714 DEBUG(std::cerr << "\n");
718 if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
721 DEBUG(std::cerr << std::string(Indent-2, ' '));
722 DEBUG(std::cerr << "== ");
723 DEBUG(Node->dump(CurDAG));
724 DEBUG(std::cerr << "\n");
727 return; // Already selected.
730 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(N);
731 if (CGMI != CodeGenMap.end()) {
732 Result = CGMI->second;
734 DEBUG(std::cerr << std::string(Indent-2, ' '));
735 DEBUG(std::cerr << "== ");
736 DEBUG(Result.Val->dump(CurDAG));
737 DEBUG(std::cerr << "\n");
745 case X86ISD::GlobalBaseReg:
746 Result = getGlobalBaseReg();
750 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
751 // code and is matched first so to prevent it from being turned into
753 SDOperand N0 = N.getOperand(0);
754 SDOperand N1 = N.getOperand(1);
755 if (N.Val->getValueType(0) == MVT::i32 &&
756 N0.getOpcode() == X86ISD::Wrapper &&
757 N1.getOpcode() == ISD::Constant) {
758 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
760 // TODO: handle ExternalSymbolSDNode.
761 if (GlobalAddressSDNode *G =
762 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
763 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), MVT::i32,
764 G->getOffset() + Offset);
765 } else if (ConstantPoolSDNode *CP =
766 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
767 C = CurDAG->getTargetConstantPool(CP->get(), MVT::i32,
769 CP->getOffset()+Offset);
773 if (N.Val->hasOneUse()) {
774 Result = CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, MVT::i32, C);
776 SDNode *ResNode = CurDAG->getTargetNode(X86::MOV32ri, MVT::i32, C);
777 Result = CodeGenMap[N] = SDOperand(ResNode, 0);
783 // Other cases are handled by auto-generated code.
789 if (Opcode == ISD::MULHU)
791 default: assert(0 && "Unsupported VT!");
792 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
793 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
794 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
798 default: assert(0 && "Unsupported VT!");
799 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
800 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
801 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
804 unsigned LoReg, HiReg;
806 default: assert(0 && "Unsupported VT!");
807 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
808 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
809 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
812 SDOperand N0 = Node->getOperand(0);
813 SDOperand N1 = Node->getOperand(1);
815 bool foldedLoad = false;
816 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
817 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
818 // MULHU and MULHS are commmutative
820 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
822 N0 = Node->getOperand(1);
823 N1 = Node->getOperand(0);
829 Select(Chain, N1.getOperand(0));
831 Chain = CurDAG->getEntryNode();
833 SDOperand InFlag(0, 0);
835 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
837 InFlag = Chain.getValue(1);
845 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
846 Tmp2, Tmp3, Chain, InFlag);
847 Chain = SDOperand(CNode, 0);
848 InFlag = SDOperand(CNode, 1);
852 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
855 Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
856 CodeGenMap[N.getValue(0)] = Result;
858 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
859 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
863 DEBUG(std::cerr << std::string(Indent-2, ' '));
864 DEBUG(std::cerr << "== ");
865 DEBUG(Result.Val->dump(CurDAG));
866 DEBUG(std::cerr << "\n");
876 bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
877 bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
880 default: assert(0 && "Unsupported VT!");
881 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
882 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
883 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
887 default: assert(0 && "Unsupported VT!");
888 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
889 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
890 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
893 unsigned LoReg, HiReg;
894 unsigned ClrOpcode, SExtOpcode;
896 default: assert(0 && "Unsupported VT!");
898 LoReg = X86::AL; HiReg = X86::AH;
899 ClrOpcode = X86::MOV8r0;
900 SExtOpcode = X86::CBW;
903 LoReg = X86::AX; HiReg = X86::DX;
904 ClrOpcode = X86::MOV16r0;
905 SExtOpcode = X86::CWD;
908 LoReg = X86::EAX; HiReg = X86::EDX;
909 ClrOpcode = X86::MOV32r0;
910 SExtOpcode = X86::CDQ;
914 SDOperand N0 = Node->getOperand(0);
915 SDOperand N1 = Node->getOperand(1);
917 bool foldedLoad = false;
918 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
919 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
922 Select(Chain, N1.getOperand(0));
924 Chain = CurDAG->getEntryNode();
926 SDOperand InFlag(0, 0);
928 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
930 InFlag = Chain.getValue(1);
933 // Sign extend the low part into the high part.
935 SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
937 // Zero out the high part, effectively zero extending the input.
938 SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
939 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(HiReg, NVT),
941 InFlag = Chain.getValue(1);
950 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Tmp0, Tmp1,
951 Tmp2, Tmp3, Chain, InFlag);
952 Chain = SDOperand(CNode, 0);
953 InFlag = SDOperand(CNode, 1);
957 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
960 Result = CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg,
962 CodeGenMap[N.getValue(0)] = Result;
964 CodeGenMap[N1.getValue(1)] = Result.getValue(1);
965 AddHandleReplacement(N1.Val, 1, Result.Val, 1);
969 DEBUG(std::cerr << std::string(Indent-2, ' '));
970 DEBUG(std::cerr << "== ");
971 DEBUG(Result.Val->dump(CurDAG));
972 DEBUG(std::cerr << "\n");
978 case ISD::TRUNCATE: {
979 if (NVT == MVT::i8) {
982 switch (Node->getOperand(0).getValueType()) {
983 default: assert(0 && "Unknown truncate!");
985 Opc = X86::MOV16to16_;
987 Opc2 = X86::TRUNC_GR16_GR8;
990 Opc = X86::MOV32to32_;
992 Opc2 = X86::TRUNC_GR32_GR8;
996 SDOperand Tmp0, Tmp1;
997 Select(Tmp0, Node->getOperand(0));
998 Tmp1 = SDOperand(CurDAG->getTargetNode(Opc, VT, Tmp0), 0);
999 Result = CodeGenMap[N] =
1000 SDOperand(CurDAG->getTargetNode(Opc2, NVT, Tmp1), 0);
1003 DEBUG(std::cerr << std::string(Indent-2, ' '));
1004 DEBUG(std::cerr << "== ");
1005 DEBUG(Result.Val->dump(CurDAG));
1006 DEBUG(std::cerr << "\n");
1016 SelectCode(Result, N);
1018 DEBUG(std::cerr << std::string(Indent-2, ' '));
1019 DEBUG(std::cerr << "=> ");
1020 DEBUG(Result.Val->dump(CurDAG));
1021 DEBUG(std::cerr << "\n");
1026 bool X86DAGToDAGISel::
1027 SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
1028 std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
1029 SDOperand Op0, Op1, Op2, Op3;
1030 switch (ConstraintCode) {
1031 case 'o': // offsetable ??
1032 case 'v': // not offsetable ??
1033 default: return true;
1035 if (!SelectAddr(Op, Op0, Op1, Op2, Op3))
1041 Select(OutOps[0], Op0);
1042 Select(OutOps[1], Op1);
1043 Select(OutOps[2], Op2);
1044 Select(OutOps[3], Op3);
1048 /// createX86ISelDag - This pass converts a legalized DAG into a
1049 /// X86-specific DAG, ready for instruction scheduling.
1051 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM) {
1052 return new X86DAGToDAGISel(TM);