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 "x86-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/Compiler.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/ADT/Statistic.h"
42 //===----------------------------------------------------------------------===//
43 // Pattern Matcher Implementation
44 //===----------------------------------------------------------------------===//
47 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
48 /// SDOperand's instead of register numbers for the leaves of the matched
50 struct X86ISelAddressMode {
56 struct { // This is really a union, discriminated by BaseType!
61 bool isRIPRel; // RIP relative?
69 unsigned Align; // CP alignment.
72 : BaseType(RegBase), isRIPRel(false), Scale(1), IndexReg(), Disp(0),
73 GV(0), CP(0), ES(0), JT(-1), Align(0) {
80 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
83 NumLoadMoved("x86-codegen", "Number of loads moved below TokenFactor");
85 //===--------------------------------------------------------------------===//
86 /// ISel - X86 specific code to select X86 machine instructions for
87 /// SelectionDAG operations.
89 class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
90 /// ContainsFPCode - Every instruction we select that uses or defines a FP
91 /// register should set this to true.
94 /// FastISel - Enable fast(er) instruction selection.
98 /// TM - Keep a reference to X86TargetMachine.
100 X86TargetMachine &TM;
102 /// X86Lowering - This object fully describes how to lower LLVM code to an
103 /// X86-specific SelectionDAG.
104 X86TargetLowering X86Lowering;
106 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
107 /// make the right decision when generating code for different targets.
108 const X86Subtarget *Subtarget;
110 /// GlobalBaseReg - keeps track of the virtual register mapped onto global
112 unsigned GlobalBaseReg;
115 X86DAGToDAGISel(X86TargetMachine &tm, bool fast)
116 : SelectionDAGISel(X86Lowering),
117 ContainsFPCode(false), FastISel(fast), TM(tm),
118 X86Lowering(*TM.getTargetLowering()),
119 Subtarget(&TM.getSubtarget<X86Subtarget>()) {}
121 virtual bool runOnFunction(Function &Fn) {
122 // Make sure we re-emit a set of the global base reg if necessary
124 return SelectionDAGISel::runOnFunction(Fn);
127 virtual const char *getPassName() const {
128 return "X86 DAG->DAG Instruction Selection";
131 /// InstructionSelectBasicBlock - This callback is invoked by
132 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
133 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
135 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
137 virtual bool CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root);
139 // Include the pieces autogenerated from the target description.
140 #include "X86GenDAGISel.inc"
143 SDNode *Select(SDOperand N);
145 bool MatchAddress(SDOperand N, X86ISelAddressMode &AM, bool isRoot = true);
146 bool SelectAddr(SDOperand Op, SDOperand N, SDOperand &Base,
147 SDOperand &Scale, SDOperand &Index, SDOperand &Disp);
148 bool SelectLEAAddr(SDOperand Op, SDOperand N, SDOperand &Base,
149 SDOperand &Scale, SDOperand &Index, SDOperand &Disp);
150 bool SelectScalarSSELoad(SDOperand Op, SDOperand Pred,
151 SDOperand N, SDOperand &Base, SDOperand &Scale,
152 SDOperand &Index, SDOperand &Disp,
153 SDOperand &InChain, SDOperand &OutChain);
154 bool TryFoldLoad(SDOperand P, SDOperand N,
155 SDOperand &Base, SDOperand &Scale,
156 SDOperand &Index, SDOperand &Disp);
157 void InstructionSelectPreprocess(SelectionDAG &DAG);
159 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
160 /// inline asm expressions.
161 virtual bool SelectInlineAsmMemoryOperand(const SDOperand &Op,
163 std::vector<SDOperand> &OutOps,
166 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
168 inline void getAddressOperands(X86ISelAddressMode &AM, SDOperand &Base,
169 SDOperand &Scale, SDOperand &Index,
171 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
172 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
174 Scale = getI8Imm(AM.Scale);
176 // These are 32-bit even in 64-bit mode since RIP relative offset
179 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp);
181 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32, AM.Align, AM.Disp);
183 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32);
184 else if (AM.JT != -1)
185 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32);
187 Disp = getI32Imm(AM.Disp);
190 /// getI8Imm - Return a target constant with the specified value, of type
192 inline SDOperand getI8Imm(unsigned Imm) {
193 return CurDAG->getTargetConstant(Imm, MVT::i8);
196 /// getI16Imm - Return a target constant with the specified value, of type
198 inline SDOperand getI16Imm(unsigned Imm) {
199 return CurDAG->getTargetConstant(Imm, MVT::i16);
202 /// getI32Imm - Return a target constant with the specified value, of type
204 inline SDOperand getI32Imm(unsigned Imm) {
205 return CurDAG->getTargetConstant(Imm, MVT::i32);
208 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
209 /// base register. Return the virtual register that holds this value.
210 SDNode *getGlobalBaseReg();
218 static SDNode *findFlagUse(SDNode *N) {
219 unsigned FlagResNo = N->getNumValues()-1;
220 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
222 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
223 SDOperand Op = User->getOperand(i);
224 if (Op.Val == N && Op.ResNo == FlagResNo)
231 static void findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
232 SDNode *Root, SDNode *Skip, bool &found,
233 std::set<SDNode *> &Visited) {
235 Use->getNodeId() > Def->getNodeId() ||
236 !Visited.insert(Use).second)
239 for (unsigned i = 0, e = Use->getNumOperands(); !found && i != e; ++i) {
240 SDNode *N = Use->getOperand(i).Val;
245 continue; // Immediate use is ok.
247 assert(Use->getOpcode() == ISD::STORE ||
248 Use->getOpcode() == X86ISD::CMP);
254 findNonImmUse(N, Def, ImmedUse, Root, Skip, found, Visited);
258 /// isNonImmUse - Start searching from Root up the DAG to check is Def can
259 /// be reached. Return true if that's the case. However, ignore direct uses
260 /// by ImmedUse (which would be U in the example illustrated in
261 /// CanBeFoldedBy) and by Root (which can happen in the store case).
262 /// FIXME: to be really generic, we should allow direct use by any node
263 /// that is being folded. But realisticly since we only fold loads which
264 /// have one non-chain use, we only need to watch out for load/op/store
265 /// and load/op/cmp case where the root (store / cmp) may reach the load via
266 /// its chain operand.
267 static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse,
268 SDNode *Skip = NULL) {
269 std::set<SDNode *> Visited;
271 findNonImmUse(Root, Def, ImmedUse, Root, Skip, found, Visited);
276 bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) {
277 if (FastISel) return false;
279 // If U use can somehow reach N through another path then U can't fold N or
280 // it will create a cycle. e.g. In the following diagram, U can reach N
281 // through X. If N is folded into into U, then X is both a predecessor and
292 if (isNonImmUse(Root, N, U))
295 // If U produces a flag, then it gets (even more) interesting. Since it
296 // would have been "glued" together with its flag use, we need to check if
309 // If FU (flag use) indirectly reach N (the load), and U fold N (call it
310 // NU), then TF is a predecessor of FU and a successor of NU. But since
311 // NU and FU are flagged together, this effectively creates a cycle.
312 bool HasFlagUse = false;
313 MVT::ValueType VT = Root->getValueType(Root->getNumValues()-1);
314 while ((VT == MVT::Flag && !Root->use_empty())) {
315 SDNode *FU = findFlagUse(Root);
322 VT = Root->getValueType(Root->getNumValues()-1);
326 return !isNonImmUse(Root, N, Root, U);
330 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
331 /// and move load below the TokenFactor. Replace store's chain operand with
332 /// load's chain result.
333 static void MoveBelowTokenFactor(SelectionDAG &DAG, SDOperand Load,
334 SDOperand Store, SDOperand TF) {
335 std::vector<SDOperand> Ops;
336 for (unsigned i = 0, e = TF.Val->getNumOperands(); i != e; ++i)
337 if (Load.Val == TF.Val->getOperand(i).Val)
338 Ops.push_back(Load.Val->getOperand(0));
340 Ops.push_back(TF.Val->getOperand(i));
341 DAG.UpdateNodeOperands(TF, &Ops[0], Ops.size());
342 DAG.UpdateNodeOperands(Load, TF, Load.getOperand(1), Load.getOperand(2));
343 DAG.UpdateNodeOperands(Store, Load.getValue(1), Store.getOperand(1),
344 Store.getOperand(2), Store.getOperand(3));
347 /// InstructionSelectPreprocess - Preprocess the DAG to allow the instruction
348 /// selector to pick more load-modify-store instructions. This is a common
359 /// [TokenFactor] [Op]
366 /// The fact the store's chain operand != load's chain will prevent the
367 /// (store (op (load))) instruction from being selected. We can transform it to:
386 void X86DAGToDAGISel::InstructionSelectPreprocess(SelectionDAG &DAG) {
387 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
388 E = DAG.allnodes_end(); I != E; ++I) {
389 if (!ISD::isNON_TRUNCStore(I))
391 SDOperand Chain = I->getOperand(0);
392 if (Chain.Val->getOpcode() != ISD::TokenFactor)
395 SDOperand N1 = I->getOperand(1);
396 SDOperand N2 = I->getOperand(2);
397 if (MVT::isFloatingPoint(N1.getValueType()) ||
398 MVT::isVector(N1.getValueType()) ||
404 unsigned Opcode = N1.Val->getOpcode();
413 SDOperand N10 = N1.getOperand(0);
414 SDOperand N11 = N1.getOperand(1);
415 if (ISD::isNON_EXTLoad(N10.Val))
417 else if (ISD::isNON_EXTLoad(N11.Val)) {
421 RModW = RModW && N10.Val->isOperand(Chain.Val) && N10.hasOneUse() &&
422 (N10.getOperand(1) == N2) &&
423 (N10.Val->getValueType(0) == N1.getValueType());
438 SDOperand N10 = N1.getOperand(0);
439 if (ISD::isNON_EXTLoad(N10.Val))
440 RModW = N10.Val->isOperand(Chain.Val) && N10.hasOneUse() &&
441 (N10.getOperand(1) == N2) &&
442 (N10.Val->getValueType(0) == N1.getValueType());
450 MoveBelowTokenFactor(DAG, Load, SDOperand(I, 0), Chain);
456 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
457 /// when it has created a SelectionDAG for us to codegen.
458 void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
460 MachineFunction::iterator FirstMBB = BB;
463 InstructionSelectPreprocess(DAG);
465 // Codegen the basic block.
467 DEBUG(std::cerr << "===== Instruction selection begins:\n");
470 DAG.setRoot(SelectRoot(DAG.getRoot()));
472 DEBUG(std::cerr << "===== Instruction selection ends:\n");
475 DAG.RemoveDeadNodes();
477 // Emit machine code to BB.
478 ScheduleAndEmitDAG(DAG);
480 // If we are emitting FP stack code, scan the basic block to determine if this
481 // block defines any FP values. If so, put an FP_REG_KILL instruction before
482 // the terminator of the block.
483 if (!Subtarget->hasSSE2()) {
484 // Note that FP stack instructions *are* used in SSE code when returning
485 // values, but these are not live out of the basic block, so we don't need
486 // an FP_REG_KILL in this case either.
487 bool ContainsFPCode = false;
489 // Scan all of the machine instructions in these MBBs, checking for FP
491 MachineFunction::iterator MBBI = FirstMBB;
493 for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end();
494 !ContainsFPCode && I != E; ++I) {
495 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
496 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
497 MRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
498 RegMap->getRegClass(I->getOperand(0).getReg()) ==
499 X86::RFPRegisterClass) {
500 ContainsFPCode = true;
505 } while (!ContainsFPCode && &*(MBBI++) != BB);
507 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
508 // a copy of the input value in this block.
509 if (!ContainsFPCode) {
510 // Final check, check LLVM BB's that are successors to the LLVM BB
511 // corresponding to BB for FP PHI nodes.
512 const BasicBlock *LLVMBB = BB->getBasicBlock();
514 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
515 !ContainsFPCode && SI != E; ++SI) {
516 for (BasicBlock::const_iterator II = SI->begin();
517 (PN = dyn_cast<PHINode>(II)); ++II) {
518 if (PN->getType()->isFloatingPoint()) {
519 ContainsFPCode = true;
526 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
527 if (ContainsFPCode) {
528 const TargetInstrDescriptor &II= TM.getInstrInfo()->get(X86::FP_REG_KILL);
529 MachineInstrBuilder MIB =
530 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
531 for (const unsigned *ImplicitDefs = II.ImplicitDefs;
532 *ImplicitDefs; ++ImplicitDefs)
533 MIB = MIB.addReg(*ImplicitDefs, true, true);
539 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
540 /// the main function.
541 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
542 MachineFrameInfo *MFI) {
543 if (Subtarget->isTargetCygwin())
544 BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("__main");
546 // Switch the FPU to 64-bit precision mode for better compatibility and speed.
547 int CWFrameIdx = MFI->CreateStackObject(2, 2);
548 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
550 // Set the high part to be 64-bit precision.
551 addFrameReference(BuildMI(BB, X86::MOV8mi, 5),
552 CWFrameIdx, 1).addImm(2);
554 // Reload the modified control word now.
555 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
558 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
559 // If this is main, emit special code for main.
560 MachineBasicBlock *BB = MF.begin();
561 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
562 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
565 /// MatchAddress - Add the specified node to the specified addressing mode,
566 /// returning true if it cannot be done. This just pattern matches for the
568 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
570 // RIP relative addressing: %rip + 32-bit displacement!
572 if (!AM.ES && AM.JT != -1 && N.getOpcode() == ISD::Constant) {
573 int64_t Val = cast<ConstantSDNode>(N)->getSignExtended();
574 if (isInt32(AM.Disp + Val)) {
582 int id = N.Val->getNodeId();
583 bool Available = isSelected(id);
585 switch (N.getOpcode()) {
587 case ISD::Constant: {
588 int64_t Val = cast<ConstantSDNode>(N)->getSignExtended();
589 if (isInt32(AM.Disp + Val)) {
596 case X86ISD::Wrapper:
597 // If value is available in a register both base and index components have
598 // been picked, we can't fit the result available in the register in the
599 // addressing mode. Duplicate GlobalAddress or ConstantPool as displacement.
601 // Can't fit GV or CP in addressing mode for X86-64 medium or large code
602 // model since the displacement field is 32-bit. Ok for small code model.
604 // For X86-64 PIC code, only allow GV / CP + displacement so we can use RIP
605 // relative addressing mode.
606 if ((!Subtarget->is64Bit() || TM.getCodeModel() == CodeModel::Small) &&
607 (!Available || (AM.Base.Reg.Val && AM.IndexReg.Val))) {
608 bool isRIP = Subtarget->is64Bit();
609 if (isRIP && (AM.Base.Reg.Val || AM.Scale > 1 || AM.IndexReg.Val ||
610 AM.BaseType == X86ISelAddressMode::FrameIndexBase))
612 if (ConstantPoolSDNode *CP =
613 dyn_cast<ConstantPoolSDNode>(N.getOperand(0))) {
615 AM.CP = CP->getConstVal();
616 AM.Align = CP->getAlignment();
617 AM.Disp += CP->getOffset();
622 } else if (GlobalAddressSDNode *G =
623 dyn_cast<GlobalAddressSDNode>(N.getOperand(0))) {
625 AM.GV = G->getGlobal();
626 AM.Disp += G->getOffset();
631 } else if (isRoot && isRIP) {
632 if (ExternalSymbolSDNode *S =
633 dyn_cast<ExternalSymbolSDNode>(N.getOperand(0))) {
634 AM.ES = S->getSymbol();
637 } else if (JumpTableSDNode *J =
638 dyn_cast<JumpTableSDNode>(N.getOperand(0))) {
639 AM.JT = J->getIndex();
647 case ISD::FrameIndex:
648 if (AM.BaseType == X86ISelAddressMode::RegBase && AM.Base.Reg.Val == 0) {
649 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
650 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
656 if (!Available && AM.IndexReg.Val == 0 && AM.Scale == 1)
657 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1))) {
658 unsigned Val = CN->getValue();
659 if (Val == 1 || Val == 2 || Val == 3) {
661 SDOperand ShVal = N.Val->getOperand(0);
663 // Okay, we know that we have a scale by now. However, if the scaled
664 // value is an add of something and a constant, we can fold the
665 // constant into the disp field here.
666 if (ShVal.Val->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
667 isa<ConstantSDNode>(ShVal.Val->getOperand(1))) {
668 AM.IndexReg = ShVal.Val->getOperand(0);
669 ConstantSDNode *AddVal =
670 cast<ConstantSDNode>(ShVal.Val->getOperand(1));
671 uint64_t Disp = AM.Disp + (AddVal->getValue() << Val);
685 // X*[3,5,9] -> X+X*[2,4,8]
687 AM.BaseType == X86ISelAddressMode::RegBase &&
688 AM.Base.Reg.Val == 0 &&
689 AM.IndexReg.Val == 0)
690 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1)))
691 if (CN->getValue() == 3 || CN->getValue() == 5 || CN->getValue() == 9) {
692 AM.Scale = unsigned(CN->getValue())-1;
694 SDOperand MulVal = N.Val->getOperand(0);
697 // Okay, we know that we have a scale by now. However, if the scaled
698 // value is an add of something and a constant, we can fold the
699 // constant into the disp field here.
700 if (MulVal.Val->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
701 isa<ConstantSDNode>(MulVal.Val->getOperand(1))) {
702 Reg = MulVal.Val->getOperand(0);
703 ConstantSDNode *AddVal =
704 cast<ConstantSDNode>(MulVal.Val->getOperand(1));
705 uint64_t Disp = AM.Disp + AddVal->getValue() * CN->getValue();
709 Reg = N.Val->getOperand(0);
711 Reg = N.Val->getOperand(0);
714 AM.IndexReg = AM.Base.Reg = Reg;
721 X86ISelAddressMode Backup = AM;
722 if (!MatchAddress(N.Val->getOperand(0), AM, false) &&
723 !MatchAddress(N.Val->getOperand(1), AM, false))
726 if (!MatchAddress(N.Val->getOperand(1), AM, false) &&
727 !MatchAddress(N.Val->getOperand(0), AM, false))
736 X86ISelAddressMode Backup = AM;
737 // Look for (x << c1) | c2 where (c2 < c1)
738 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(0));
739 if (CN && !MatchAddress(N.Val->getOperand(1), AM, false)) {
740 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
741 AM.Disp = CN->getValue();
746 CN = dyn_cast<ConstantSDNode>(N.Val->getOperand(1));
747 if (CN && !MatchAddress(N.Val->getOperand(0), AM, false)) {
748 if (AM.GV == NULL && AM.Disp == 0 && CN->getValue() < AM.Scale) {
749 AM.Disp = CN->getValue();
759 // Is the base register already occupied?
760 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
761 // If so, check to see if the scale index register is set.
762 if (AM.IndexReg.Val == 0) {
768 // Otherwise, we cannot select it.
772 // Default, generate it as a register.
773 AM.BaseType = X86ISelAddressMode::RegBase;
778 /// SelectAddr - returns true if it is able pattern match an addressing mode.
779 /// It returns the operands which make up the maximal addressing mode it can
780 /// match by reference.
781 bool X86DAGToDAGISel::SelectAddr(SDOperand Op, SDOperand N, SDOperand &Base,
782 SDOperand &Scale, SDOperand &Index,
784 X86ISelAddressMode AM;
785 if (MatchAddress(N, AM))
788 MVT::ValueType VT = N.getValueType();
789 if (AM.BaseType == X86ISelAddressMode::RegBase) {
790 if (!AM.Base.Reg.Val)
791 AM.Base.Reg = CurDAG->getRegister(0, VT);
794 if (!AM.IndexReg.Val)
795 AM.IndexReg = CurDAG->getRegister(0, VT);
797 getAddressOperands(AM, Base, Scale, Index, Disp);
801 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
803 static inline bool isZeroNode(SDOperand Elt) {
804 return ((isa<ConstantSDNode>(Elt) &&
805 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
806 (isa<ConstantFPSDNode>(Elt) &&
807 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
811 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
812 /// match a load whose top elements are either undef or zeros. The load flavor
813 /// is derived from the type of N, which is either v4f32 or v2f64.
814 bool X86DAGToDAGISel::SelectScalarSSELoad(SDOperand Op, SDOperand Pred,
815 SDOperand N, SDOperand &Base,
816 SDOperand &Scale, SDOperand &Index,
817 SDOperand &Disp, SDOperand &InChain,
818 SDOperand &OutChain) {
819 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
820 InChain = N.getOperand(0).getValue(1);
821 if (ISD::isNON_EXTLoad(InChain.Val) &&
822 InChain.getValue(0).hasOneUse() &&
824 CanBeFoldedBy(N.Val, Pred.Val, Op.Val)) {
825 LoadSDNode *LD = cast<LoadSDNode>(InChain);
826 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
828 OutChain = LD->getChain();
833 // Also handle the case where we explicitly require zeros in the top
834 // elements. This is a vector shuffle from the zero vector.
835 if (N.getOpcode() == ISD::VECTOR_SHUFFLE && N.Val->hasOneUse() &&
836 N.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
837 N.getOperand(1).getOpcode() == ISD::SCALAR_TO_VECTOR &&
838 N.getOperand(1).Val->hasOneUse() &&
839 ISD::isNON_EXTLoad(N.getOperand(1).getOperand(0).Val) &&
840 N.getOperand(1).getOperand(0).hasOneUse()) {
841 // Check to see if the BUILD_VECTOR is building a zero vector.
842 SDOperand BV = N.getOperand(0);
843 for (unsigned i = 0, e = BV.getNumOperands(); i != e; ++i)
844 if (!isZeroNode(BV.getOperand(i)) &&
845 BV.getOperand(i).getOpcode() != ISD::UNDEF)
846 return false; // Not a zero/undef vector.
847 // Check to see if the shuffle mask is 4/L/L/L or 2/L, where L is something
849 unsigned VecWidth = BV.getNumOperands();
850 SDOperand ShufMask = N.getOperand(2);
851 assert(ShufMask.getOpcode() == ISD::BUILD_VECTOR && "Invalid shuf mask!");
852 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(ShufMask.getOperand(0))) {
853 if (C->getValue() == VecWidth) {
854 for (unsigned i = 1; i != VecWidth; ++i) {
855 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF) {
858 ConstantSDNode *C = cast<ConstantSDNode>(ShufMask.getOperand(i));
859 if (C->getValue() >= VecWidth) return false;
864 // Okay, this is a zero extending load. Fold it.
865 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(1).getOperand(0));
866 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
868 OutChain = LD->getChain();
869 InChain = SDOperand(LD, 1);
877 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
878 /// mode it matches can be cost effectively emitted as an LEA instruction.
879 bool X86DAGToDAGISel::SelectLEAAddr(SDOperand Op, SDOperand N,
880 SDOperand &Base, SDOperand &Scale,
881 SDOperand &Index, SDOperand &Disp) {
882 X86ISelAddressMode AM;
883 if (MatchAddress(N, AM))
886 MVT::ValueType VT = N.getValueType();
887 unsigned Complexity = 0;
888 if (AM.BaseType == X86ISelAddressMode::RegBase)
892 AM.Base.Reg = CurDAG->getRegister(0, VT);
893 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
899 AM.IndexReg = CurDAG->getRegister(0, VT);
903 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg
904 else if (AM.Scale > 1)
907 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
908 // to a LEA. This is determined with some expermentation but is by no means
909 // optimal (especially for code size consideration). LEA is nice because of
910 // its three-address nature. Tweak the cost function again when we can run
911 // convertToThreeAddress() at register allocation time.
912 if (AM.GV || AM.CP || AM.ES || AM.JT != -1) {
913 // For X86-64, we should always use lea to materialize RIP relative
915 if (Subtarget->is64Bit())
921 if (AM.Disp && (AM.Base.Reg.Val || AM.IndexReg.Val))
924 if (Complexity > 2) {
925 getAddressOperands(AM, Base, Scale, Index, Disp);
931 bool X86DAGToDAGISel::TryFoldLoad(SDOperand P, SDOperand N,
932 SDOperand &Base, SDOperand &Scale,
933 SDOperand &Index, SDOperand &Disp) {
934 if (ISD::isNON_EXTLoad(N.Val) &&
936 CanBeFoldedBy(N.Val, P.Val, P.Val))
937 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp);
941 /// getGlobalBaseReg - Output the instructions required to put the
942 /// base address to use for accessing globals into a register.
944 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
945 assert(!Subtarget->is64Bit() && "X86-64 PIC uses RIP relative addressing");
946 if (!GlobalBaseReg) {
947 // Insert the set of GlobalBaseReg into the first MBB of the function
948 MachineBasicBlock &FirstMBB = BB->getParent()->front();
949 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
950 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
951 // FIXME: when we get to LP64, we will need to create the appropriate
952 // type of register here.
953 GlobalBaseReg = RegMap->createVirtualRegister(X86::GR32RegisterClass);
954 BuildMI(FirstMBB, MBBI, X86::MovePCtoStack, 0);
955 BuildMI(FirstMBB, MBBI, X86::POP32r, 1, GlobalBaseReg);
957 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).Val;
960 static SDNode *FindCallStartFromCall(SDNode *Node) {
961 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
962 assert(Node->getOperand(0).getValueType() == MVT::Other &&
963 "Node doesn't have a token chain argument!");
964 return FindCallStartFromCall(Node->getOperand(0).Val);
967 SDNode *X86DAGToDAGISel::Select(SDOperand N) {
968 SDNode *Node = N.Val;
969 MVT::ValueType NVT = Node->getValueType(0);
971 unsigned Opcode = Node->getOpcode();
974 DEBUG(std::cerr << std::string(Indent, ' '));
975 DEBUG(std::cerr << "Selecting: ");
976 DEBUG(Node->dump(CurDAG));
977 DEBUG(std::cerr << "\n");
981 if (Opcode >= ISD::BUILTIN_OP_END && Opcode < X86ISD::FIRST_NUMBER) {
983 DEBUG(std::cerr << std::string(Indent-2, ' '));
984 DEBUG(std::cerr << "== ");
985 DEBUG(Node->dump(CurDAG));
986 DEBUG(std::cerr << "\n");
989 return NULL; // Already selected.
994 case X86ISD::GlobalBaseReg:
995 return getGlobalBaseReg();
998 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
999 // code and is matched first so to prevent it from being turned into
1001 // In 64-bit mode, use LEA to take advantage of RIP-relative addressing.
1002 MVT::ValueType PtrVT = TLI.getPointerTy();
1003 SDOperand N0 = N.getOperand(0);
1004 SDOperand N1 = N.getOperand(1);
1005 if (N.Val->getValueType(0) == PtrVT &&
1006 N0.getOpcode() == X86ISD::Wrapper &&
1007 N1.getOpcode() == ISD::Constant) {
1008 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getValue();
1010 // TODO: handle ExternalSymbolSDNode.
1011 if (GlobalAddressSDNode *G =
1012 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
1013 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), PtrVT,
1014 G->getOffset() + Offset);
1015 } else if (ConstantPoolSDNode *CP =
1016 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
1017 C = CurDAG->getTargetConstantPool(CP->getConstVal(), PtrVT,
1019 CP->getOffset()+Offset);
1023 if (Subtarget->is64Bit()) {
1024 SDOperand Ops[] = { CurDAG->getRegister(0, PtrVT), getI8Imm(1),
1025 CurDAG->getRegister(0, PtrVT), C };
1026 return CurDAG->SelectNodeTo(N.Val, X86::LEA64r, MVT::i64, Ops, 4);
1028 return CurDAG->SelectNodeTo(N.Val, X86::MOV32ri, PtrVT, C);
1032 // Other cases are handled by auto-generated code.
1038 if (Opcode == ISD::MULHU)
1040 default: assert(0 && "Unsupported VT!");
1041 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1042 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1043 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1044 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1048 default: assert(0 && "Unsupported VT!");
1049 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1050 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1051 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1052 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1055 unsigned LoReg, HiReg;
1057 default: assert(0 && "Unsupported VT!");
1058 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1059 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1060 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1061 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1064 SDOperand N0 = Node->getOperand(0);
1065 SDOperand N1 = Node->getOperand(1);
1067 bool foldedLoad = false;
1068 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
1069 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1070 // MULHU and MULHS are commmutative
1072 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
1074 N0 = Node->getOperand(1);
1075 N1 = Node->getOperand(0);
1081 Chain = N1.getOperand(0);
1082 AddToISelQueue(Chain);
1084 Chain = CurDAG->getEntryNode();
1086 SDOperand InFlag(0, 0);
1088 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
1090 InFlag = Chain.getValue(1);
1093 AddToISelQueue(Tmp0);
1094 AddToISelQueue(Tmp1);
1095 AddToISelQueue(Tmp2);
1096 AddToISelQueue(Tmp3);
1097 SDOperand Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Chain, InFlag };
1099 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1100 Chain = SDOperand(CNode, 0);
1101 InFlag = SDOperand(CNode, 1);
1105 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1108 SDOperand Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
1109 ReplaceUses(N.getValue(0), Result);
1111 ReplaceUses(N1.getValue(1), Result.getValue(1));
1114 DEBUG(std::cerr << std::string(Indent-2, ' '));
1115 DEBUG(std::cerr << "=> ");
1116 DEBUG(Result.Val->dump(CurDAG));
1117 DEBUG(std::cerr << "\n");
1127 bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
1128 bool isDiv = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
1131 default: assert(0 && "Unsupported VT!");
1132 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1133 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1134 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1135 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1139 default: assert(0 && "Unsupported VT!");
1140 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1141 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1142 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1143 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1146 unsigned LoReg, HiReg;
1147 unsigned ClrOpcode, SExtOpcode;
1149 default: assert(0 && "Unsupported VT!");
1151 LoReg = X86::AL; HiReg = X86::AH;
1152 ClrOpcode = X86::MOV8r0;
1153 SExtOpcode = X86::CBW;
1156 LoReg = X86::AX; HiReg = X86::DX;
1157 ClrOpcode = X86::MOV16r0;
1158 SExtOpcode = X86::CWD;
1161 LoReg = X86::EAX; HiReg = X86::EDX;
1162 ClrOpcode = X86::MOV32r0;
1163 SExtOpcode = X86::CDQ;
1166 LoReg = X86::RAX; HiReg = X86::RDX;
1167 ClrOpcode = X86::MOV64r0;
1168 SExtOpcode = X86::CQO;
1172 SDOperand N0 = Node->getOperand(0);
1173 SDOperand N1 = Node->getOperand(1);
1175 bool foldedLoad = false;
1176 SDOperand Tmp0, Tmp1, Tmp2, Tmp3;
1177 foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1180 Chain = N1.getOperand(0);
1181 AddToISelQueue(Chain);
1183 Chain = CurDAG->getEntryNode();
1185 SDOperand InFlag(0, 0);
1187 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(LoReg, NVT),
1189 InFlag = Chain.getValue(1);
1192 // Sign extend the low part into the high part.
1194 SDOperand(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
1196 // Zero out the high part, effectively zero extending the input.
1197 SDOperand ClrNode = SDOperand(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
1198 Chain = CurDAG->getCopyToReg(Chain, CurDAG->getRegister(HiReg, NVT),
1200 InFlag = Chain.getValue(1);
1204 AddToISelQueue(Tmp0);
1205 AddToISelQueue(Tmp1);
1206 AddToISelQueue(Tmp2);
1207 AddToISelQueue(Tmp3);
1208 SDOperand Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Chain, InFlag };
1210 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1211 Chain = SDOperand(CNode, 0);
1212 InFlag = SDOperand(CNode, 1);
1216 SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1219 SDOperand Result = CurDAG->getCopyFromReg(Chain, isDiv ? LoReg : HiReg,
1221 ReplaceUses(N.getValue(0), Result);
1223 ReplaceUses(N1.getValue(1), Result.getValue(1));
1226 DEBUG(std::cerr << std::string(Indent-2, ' '));
1227 DEBUG(std::cerr << "=> ");
1228 DEBUG(Result.Val->dump(CurDAG));
1229 DEBUG(std::cerr << "\n");
1236 case ISD::TRUNCATE: {
1237 if (!Subtarget->is64Bit() && NVT == MVT::i8) {
1240 switch (Node->getOperand(0).getValueType()) {
1241 default: assert(0 && "Unknown truncate!");
1243 Opc = X86::MOV16to16_;
1245 Opc2 = X86::TRUNC_16_to8;
1248 Opc = X86::MOV32to32_;
1250 Opc2 = X86::TRUNC_32_to8;
1254 AddToISelQueue(Node->getOperand(0));
1256 SDOperand(CurDAG->getTargetNode(Opc, VT, Node->getOperand(0)), 0);
1257 SDNode *ResNode = CurDAG->getTargetNode(Opc2, NVT, Tmp);
1260 DEBUG(std::cerr << std::string(Indent-2, ' '));
1261 DEBUG(std::cerr << "=> ");
1262 DEBUG(ResNode->dump(CurDAG));
1263 DEBUG(std::cerr << "\n");
1273 SDNode *ResNode = SelectCode(N);
1276 DEBUG(std::cerr << std::string(Indent-2, ' '));
1277 DEBUG(std::cerr << "=> ");
1278 if (ResNode == NULL || ResNode == N.Val)
1279 DEBUG(N.Val->dump(CurDAG));
1281 DEBUG(ResNode->dump(CurDAG));
1282 DEBUG(std::cerr << "\n");
1289 bool X86DAGToDAGISel::
1290 SelectInlineAsmMemoryOperand(const SDOperand &Op, char ConstraintCode,
1291 std::vector<SDOperand> &OutOps, SelectionDAG &DAG){
1292 SDOperand Op0, Op1, Op2, Op3;
1293 switch (ConstraintCode) {
1294 case 'o': // offsetable ??
1295 case 'v': // not offsetable ??
1296 default: return true;
1298 if (!SelectAddr(Op, Op, Op0, Op1, Op2, Op3))
1303 OutOps.push_back(Op0);
1304 OutOps.push_back(Op1);
1305 OutOps.push_back(Op2);
1306 OutOps.push_back(Op3);
1307 AddToISelQueue(Op0);
1308 AddToISelQueue(Op1);
1309 AddToISelQueue(Op2);
1310 AddToISelQueue(Op3);
1314 /// createX86ISelDag - This pass converts a legalized DAG into a
1315 /// X86-specific DAG, ready for instruction scheduling.
1317 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM, bool Fast) {
1318 return new X86DAGToDAGISel(TM, Fast);