1 //===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 "X86MachineFunctionInfo.h"
20 #include "X86RegisterInfo.h"
21 #include "X86Subtarget.h"
22 #include "X86TargetMachine.h"
23 #include "llvm/GlobalValue.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/Support/CFG.h"
27 #include "llvm/Type.h"
28 #include "llvm/CodeGen/MachineConstantPool.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAGISel.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetOptions.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/Streams.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/Statistic.h"
46 STATISTIC(NumFPKill , "Number of FP_REG_KILL instructions added");
47 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
49 //===----------------------------------------------------------------------===//
50 // Pattern Matcher Implementation
51 //===----------------------------------------------------------------------===//
54 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
55 /// SDValue's instead of register numbers for the leaves of the matched
57 struct X86ISelAddressMode {
63 struct { // This is really a union, discriminated by BaseType!
68 bool isRIPRel; // RIP as base?
76 unsigned Align; // CP alignment.
79 : BaseType(RegBase), isRIPRel(false), Scale(1), IndexReg(), Disp(0),
80 GV(0), CP(0), ES(0), JT(-1), Align(0) {
83 cerr << "X86ISelAddressMode " << this << "\n";
85 if (Base.Reg.getNode() != 0) Base.Reg.getNode()->dump();
87 cerr << " Base.FrameIndex " << Base.FrameIndex << "\n";
88 cerr << "isRIPRel " << isRIPRel << " Scale" << Scale << "\n";
90 if (IndexReg.getNode() != 0) IndexReg.getNode()->dump();
92 cerr << " Disp " << Disp << "\n";
93 cerr << "GV "; if (GV) GV->dump();
95 cerr << " CP "; if (CP) CP->dump();
98 cerr << "ES "; if (ES) cerr << ES; else cerr << "nul";
99 cerr << " JT" << JT << " Align" << Align << "\n";
105 //===--------------------------------------------------------------------===//
106 /// ISel - X86 specific code to select X86 machine instructions for
107 /// SelectionDAG operations.
109 class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
110 /// ContainsFPCode - Every instruction we select that uses or defines a FP
111 /// register should set this to true.
114 /// TM - Keep a reference to X86TargetMachine.
116 X86TargetMachine &TM;
118 /// X86Lowering - This object fully describes how to lower LLVM code to an
119 /// X86-specific SelectionDAG.
120 X86TargetLowering X86Lowering;
122 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
123 /// make the right decision when generating code for different targets.
124 const X86Subtarget *Subtarget;
126 /// GlobalBaseReg - keeps track of the virtual register mapped onto global
128 unsigned GlobalBaseReg;
130 /// CurBB - Current BB being isel'd.
132 MachineBasicBlock *CurBB;
134 /// OptForSize - If true, selector should try to optimize for code size
135 /// instead of performance.
139 X86DAGToDAGISel(X86TargetMachine &tm, bool fast)
140 : SelectionDAGISel(X86Lowering, fast),
141 ContainsFPCode(false), TM(tm),
142 X86Lowering(*TM.getTargetLowering()),
143 Subtarget(&TM.getSubtarget<X86Subtarget>()),
144 OptForSize(OptimizeForSize) {}
146 virtual bool runOnFunction(Function &Fn) {
147 // Make sure we re-emit a set of the global base reg if necessary
149 return SelectionDAGISel::runOnFunction(Fn);
152 virtual const char *getPassName() const {
153 return "X86 DAG->DAG Instruction Selection";
156 /// InstructionSelect - This callback is invoked by
157 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
158 virtual void InstructionSelect();
160 /// InstructionSelectPostProcessing - Post processing of selected and
161 /// scheduled basic blocks.
162 virtual void InstructionSelectPostProcessing();
164 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
166 virtual bool CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) const;
168 // Include the pieces autogenerated from the target description.
169 #include "X86GenDAGISel.inc"
172 SDNode *Select(SDValue N);
174 bool MatchAddress(SDValue N, X86ISelAddressMode &AM,
175 bool isRoot = true, unsigned Depth = 0);
176 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM,
177 bool isRoot, unsigned Depth);
178 bool SelectAddr(SDValue Op, SDValue N, SDValue &Base,
179 SDValue &Scale, SDValue &Index, SDValue &Disp);
180 bool SelectLEAAddr(SDValue Op, SDValue N, SDValue &Base,
181 SDValue &Scale, SDValue &Index, SDValue &Disp);
182 bool SelectScalarSSELoad(SDValue Op, SDValue Pred,
183 SDValue N, SDValue &Base, SDValue &Scale,
184 SDValue &Index, SDValue &Disp,
185 SDValue &InChain, SDValue &OutChain);
186 bool TryFoldLoad(SDValue P, SDValue N,
187 SDValue &Base, SDValue &Scale,
188 SDValue &Index, SDValue &Disp);
189 void PreprocessForRMW();
190 void PreprocessForFPConvert();
192 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
193 /// inline asm expressions.
194 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
196 std::vector<SDValue> &OutOps);
198 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
200 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
201 SDValue &Scale, SDValue &Index,
203 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
204 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
206 Scale = getI8Imm(AM.Scale);
208 // These are 32-bit even in 64-bit mode since RIP relative offset
211 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp);
213 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
216 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32);
217 else if (AM.JT != -1)
218 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32);
220 Disp = getI32Imm(AM.Disp);
223 /// getI8Imm - Return a target constant with the specified value, of type
225 inline SDValue getI8Imm(unsigned Imm) {
226 return CurDAG->getTargetConstant(Imm, MVT::i8);
229 /// getI16Imm - Return a target constant with the specified value, of type
231 inline SDValue getI16Imm(unsigned Imm) {
232 return CurDAG->getTargetConstant(Imm, MVT::i16);
235 /// getI32Imm - Return a target constant with the specified value, of type
237 inline SDValue getI32Imm(unsigned Imm) {
238 return CurDAG->getTargetConstant(Imm, MVT::i32);
241 /// getGlobalBaseReg - Return an SDNode that returns the value of
242 /// the global base register. Output instructions required to
243 /// initialize the global base register, if necessary.
245 SDNode *getGlobalBaseReg();
247 /// getTruncateTo8Bit - return an SDNode that implements a subreg based
248 /// truncate of the specified operand to i8. This can be done with tablegen,
249 /// except that this code uses MVT::Flag in a tricky way that happens to
250 /// improve scheduling in some cases.
251 SDNode *getTruncateTo8Bit(SDValue N0);
259 /// findFlagUse - Return use of MVT::Flag value produced by the specified
262 static SDNode *findFlagUse(SDNode *N) {
263 unsigned FlagResNo = N->getNumValues()-1;
264 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
266 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
267 SDValue Op = User->getOperand(i);
268 if (Op.getNode() == N && Op.getResNo() == FlagResNo)
275 /// findNonImmUse - Return true by reference in "found" if "Use" is an
276 /// non-immediate use of "Def". This function recursively traversing
277 /// up the operand chain ignoring certain nodes.
278 static void findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
279 SDNode *Root, bool &found,
280 SmallPtrSet<SDNode*, 16> &Visited) {
282 Use->getNodeId() > Def->getNodeId() ||
283 !Visited.insert(Use))
286 for (unsigned i = 0, e = Use->getNumOperands(); !found && i != e; ++i) {
287 SDNode *N = Use->getOperand(i).getNode();
289 if (Use == ImmedUse || Use == Root)
290 continue; // We are not looking for immediate use.
296 // Traverse up the operand chain.
297 findNonImmUse(N, Def, ImmedUse, Root, found, Visited);
301 /// isNonImmUse - Start searching from Root up the DAG to check is Def can
302 /// be reached. Return true if that's the case. However, ignore direct uses
303 /// by ImmedUse (which would be U in the example illustrated in
304 /// CanBeFoldedBy) and by Root (which can happen in the store case).
305 /// FIXME: to be really generic, we should allow direct use by any node
306 /// that is being folded. But realisticly since we only fold loads which
307 /// have one non-chain use, we only need to watch out for load/op/store
308 /// and load/op/cmp case where the root (store / cmp) may reach the load via
309 /// its chain operand.
310 static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
311 SmallPtrSet<SDNode*, 16> Visited;
313 findNonImmUse(Root, Def, ImmedUse, Root, found, Visited);
318 bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) const {
319 if (Fast) return false;
321 // If Root use can somehow reach N through a path that that doesn't contain
322 // U then folding N would create a cycle. e.g. In the following
323 // diagram, Root can reach N through X. If N is folded into into Root, then
324 // X is both a predecessor and a successor of U.
335 // * indicates nodes to be folded together.
337 // If Root produces a flag, then it gets (even more) interesting. Since it
338 // will be "glued" together with its flag use in the scheduler, we need to
339 // check if it might reach N.
358 // If FU (flag use) indirectly reaches N (the load), and Root folds N
359 // (call it Fold), then X is a predecessor of FU and a successor of
360 // Fold. But since Fold and FU are flagged together, this will create
361 // a cycle in the scheduling graph.
363 MVT VT = Root->getValueType(Root->getNumValues()-1);
364 while (VT == MVT::Flag) {
365 SDNode *FU = findFlagUse(Root);
369 VT = Root->getValueType(Root->getNumValues()-1);
372 return !isNonImmUse(Root, N, U);
375 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
376 /// and move load below the TokenFactor. Replace store's chain operand with
377 /// load's chain result.
378 static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
379 SDValue Store, SDValue TF) {
380 SmallVector<SDValue, 4> Ops;
381 for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i)
382 if (Load.getNode() == TF.getOperand(i).getNode())
383 Ops.push_back(Load.getOperand(0));
385 Ops.push_back(TF.getOperand(i));
386 CurDAG->UpdateNodeOperands(TF, &Ops[0], Ops.size());
387 CurDAG->UpdateNodeOperands(Load, TF, Load.getOperand(1), Load.getOperand(2));
388 CurDAG->UpdateNodeOperands(Store, Load.getValue(1), Store.getOperand(1),
389 Store.getOperand(2), Store.getOperand(3));
392 /// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
394 static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
396 if (N.getOpcode() == ISD::BIT_CONVERT)
399 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
400 if (!LD || LD->isVolatile())
402 if (LD->getAddressingMode() != ISD::UNINDEXED)
405 ISD::LoadExtType ExtType = LD->getExtensionType();
406 if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
410 N.getOperand(1) == Address &&
411 N.getNode()->isOperandOf(Chain.getNode())) {
418 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
419 /// operand and move load below the call's chain operand.
420 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
421 SDValue Call, SDValue Chain) {
422 SmallVector<SDValue, 8> Ops;
423 for (unsigned i = 0, e = Chain.getNode()->getNumOperands(); i != e; ++i)
424 if (Load.getNode() == Chain.getOperand(i).getNode())
425 Ops.push_back(Load.getOperand(0));
427 Ops.push_back(Chain.getOperand(i));
428 CurDAG->UpdateNodeOperands(Chain, &Ops[0], Ops.size());
429 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
430 Load.getOperand(1), Load.getOperand(2));
432 Ops.push_back(SDValue(Load.getNode(), 1));
433 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
434 Ops.push_back(Call.getOperand(i));
435 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
438 /// isCalleeLoad - Return true if call address is a load and it can be
439 /// moved below CALLSEQ_START and the chains leading up to the call.
440 /// Return the CALLSEQ_START by reference as a second output.
441 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
442 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
444 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
447 LD->getAddressingMode() != ISD::UNINDEXED ||
448 LD->getExtensionType() != ISD::NON_EXTLOAD)
451 // Now let's find the callseq_start.
452 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
453 if (!Chain.hasOneUse())
455 Chain = Chain.getOperand(0);
457 return Chain.getOperand(0).getNode() == Callee.getNode();
461 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
462 /// This is only run if not in -fast mode (aka -O0).
463 /// This allows the instruction selector to pick more read-modify-write
464 /// instructions. This is a common case:
474 /// [TokenFactor] [Op]
481 /// The fact the store's chain operand != load's chain will prevent the
482 /// (store (op (load))) instruction from being selected. We can transform it to:
501 void X86DAGToDAGISel::PreprocessForRMW() {
502 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
503 E = CurDAG->allnodes_end(); I != E; ++I) {
504 if (I->getOpcode() == X86ISD::CALL) {
505 /// Also try moving call address load from outside callseq_start to just
506 /// before the call to allow it to be folded.
524 SDValue Chain = I->getOperand(0);
525 SDValue Load = I->getOperand(1);
526 if (!isCalleeLoad(Load, Chain))
528 MoveBelowCallSeqStart(CurDAG, Load, SDValue(I, 0), Chain);
533 if (!ISD::isNON_TRUNCStore(I))
535 SDValue Chain = I->getOperand(0);
537 if (Chain.getNode()->getOpcode() != ISD::TokenFactor)
540 SDValue N1 = I->getOperand(1);
541 SDValue N2 = I->getOperand(2);
542 if ((N1.getValueType().isFloatingPoint() &&
543 !N1.getValueType().isVector()) ||
549 unsigned Opcode = N1.getNode()->getOpcode();
558 case ISD::VECTOR_SHUFFLE: {
559 SDValue N10 = N1.getOperand(0);
560 SDValue N11 = N1.getOperand(1);
561 RModW = isRMWLoad(N10, Chain, N2, Load);
563 RModW = isRMWLoad(N11, Chain, N2, Load);
576 SDValue N10 = N1.getOperand(0);
577 RModW = isRMWLoad(N10, Chain, N2, Load);
583 MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
590 /// PreprocessForFPConvert - Walk over the dag lowering fpround and fpextend
591 /// nodes that target the FP stack to be store and load to the stack. This is a
592 /// gross hack. We would like to simply mark these as being illegal, but when
593 /// we do that, legalize produces these when it expands calls, then expands
594 /// these in the same legalize pass. We would like dag combine to be able to
595 /// hack on these between the call expansion and the node legalization. As such
596 /// this pass basically does "really late" legalization of these inline with the
598 void X86DAGToDAGISel::PreprocessForFPConvert() {
599 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
600 E = CurDAG->allnodes_end(); I != E; ) {
601 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
602 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
605 // If the source and destination are SSE registers, then this is a legal
606 // conversion that should not be lowered.
607 MVT SrcVT = N->getOperand(0).getValueType();
608 MVT DstVT = N->getValueType(0);
609 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
610 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
611 if (SrcIsSSE && DstIsSSE)
614 if (!SrcIsSSE && !DstIsSSE) {
615 // If this is an FPStack extension, it is a noop.
616 if (N->getOpcode() == ISD::FP_EXTEND)
618 // If this is a value-preserving FPStack truncation, it is a noop.
619 if (N->getConstantOperandVal(1))
623 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
624 // FPStack has extload and truncstore. SSE can fold direct loads into other
625 // operations. Based on this, decide what we want to do.
627 if (N->getOpcode() == ISD::FP_ROUND)
628 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
630 MemVT = SrcIsSSE ? SrcVT : DstVT;
632 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
634 // FIXME: optimize the case where the src/dest is a load or store?
635 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(),
637 MemTmp, NULL, 0, MemVT);
638 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, DstVT, Store, MemTmp,
641 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
642 // extload we created. This will cause general havok on the dag because
643 // anything below the conversion could be folded into other existing nodes.
644 // To avoid invalidating 'I', back it up to the convert node.
646 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
648 // Now that we did that, the node is dead. Increment the iterator to the
649 // next node to process, then delete N.
651 CurDAG->DeleteNode(N);
655 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
656 /// when it has created a SelectionDAG for us to codegen.
657 void X86DAGToDAGISel::InstructionSelect() {
658 CurBB = BB; // BB can change as result of isel.
660 const Function *F = CurDAG->getMachineFunction().getFunction();
661 OptForSize = !F->isDeclaration() && F->hasNote(Attribute::OptimizeForSize);
668 // FIXME: This should only happen when not -fast.
669 PreprocessForFPConvert();
671 // Codegen the basic block.
673 DOUT << "===== Instruction selection begins:\n";
678 DOUT << "===== Instruction selection ends:\n";
681 CurDAG->RemoveDeadNodes();
684 void X86DAGToDAGISel::InstructionSelectPostProcessing() {
685 // If we are emitting FP stack code, scan the basic block to determine if this
686 // block defines any FP values. If so, put an FP_REG_KILL instruction before
687 // the terminator of the block.
689 // Note that FP stack instructions are used in all modes for long double,
690 // so we always need to do this check.
691 // Also note that it's possible for an FP stack register to be live across
692 // an instruction that produces multiple basic blocks (SSE CMOV) so we
693 // must check all the generated basic blocks.
695 // Scan all of the machine instructions in these MBBs, checking for FP
696 // stores. (RFP32 and RFP64 will not exist in SSE mode, but RFP80 might.)
697 MachineFunction::iterator MBBI = CurBB;
698 MachineFunction::iterator EndMBB = BB; ++EndMBB;
699 for (; MBBI != EndMBB; ++MBBI) {
700 MachineBasicBlock *MBB = MBBI;
702 // If this block returns, ignore it. We don't want to insert an FP_REG_KILL
703 // before the return.
705 MachineBasicBlock::iterator EndI = MBB->end();
707 if (EndI->getDesc().isReturn())
711 bool ContainsFPCode = false;
712 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
713 !ContainsFPCode && I != E; ++I) {
714 if (I->getNumOperands() != 0 && I->getOperand(0).isRegister()) {
715 const TargetRegisterClass *clas;
716 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
717 if (I->getOperand(op).isRegister() && I->getOperand(op).isDef() &&
718 TargetRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
719 ((clas = RegInfo->getRegClass(I->getOperand(0).getReg())) ==
720 X86::RFP32RegisterClass ||
721 clas == X86::RFP64RegisterClass ||
722 clas == X86::RFP80RegisterClass)) {
723 ContainsFPCode = true;
729 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
730 // a copy of the input value in this block. In SSE mode, we only care about
732 if (!ContainsFPCode) {
733 // Final check, check LLVM BB's that are successors to the LLVM BB
734 // corresponding to BB for FP PHI nodes.
735 const BasicBlock *LLVMBB = BB->getBasicBlock();
737 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
738 !ContainsFPCode && SI != E; ++SI) {
739 for (BasicBlock::const_iterator II = SI->begin();
740 (PN = dyn_cast<PHINode>(II)); ++II) {
741 if (PN->getType()==Type::X86_FP80Ty ||
742 (!Subtarget->hasSSE1() && PN->getType()->isFloatingPoint()) ||
743 (!Subtarget->hasSSE2() && PN->getType()==Type::DoubleTy)) {
744 ContainsFPCode = true;
750 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
751 if (ContainsFPCode) {
752 BuildMI(*MBB, MBBI->getFirstTerminator(),
753 TM.getInstrInfo()->get(X86::FP_REG_KILL));
759 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
760 /// the main function.
761 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
762 MachineFrameInfo *MFI) {
763 const TargetInstrInfo *TII = TM.getInstrInfo();
764 if (Subtarget->isTargetCygMing())
765 BuildMI(BB, TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
768 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
769 // If this is main, emit special code for main.
770 MachineBasicBlock *BB = MF.begin();
771 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
772 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
775 /// MatchAddress - Add the specified node to the specified addressing mode,
776 /// returning true if it cannot be done. This just pattern matches for the
778 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM,
779 bool isRoot, unsigned Depth) {
780 DOUT << "MatchAddress: "; DEBUG(AM.dump());
783 return MatchAddressBase(N, AM, isRoot, Depth);
785 // RIP relative addressing: %rip + 32-bit displacement!
787 if (!AM.ES && AM.JT != -1 && N.getOpcode() == ISD::Constant) {
788 int64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
789 if (isInt32(AM.Disp + Val)) {
797 int id = N.getNode()->getNodeId();
798 bool AlreadySelected = isSelected(id); // Already selected, not yet replaced.
800 switch (N.getOpcode()) {
802 case ISD::Constant: {
803 int64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
804 if (isInt32(AM.Disp + Val)) {
811 case X86ISD::Wrapper: {
812 DOUT << "Wrapper: 64bit " << Subtarget->is64Bit();
813 DOUT << " AM "; DEBUG(AM.dump()); DOUT << "\n";
814 DOUT << "AlreadySelected " << AlreadySelected << "\n";
815 bool is64Bit = Subtarget->is64Bit();
816 // Under X86-64 non-small code model, GV (and friends) are 64-bits.
817 // Also, base and index reg must be 0 in order to use rip as base.
818 if (is64Bit && (TM.getCodeModel() != CodeModel::Small ||
819 AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
821 if (AM.GV != 0 || AM.CP != 0 || AM.ES != 0 || AM.JT != -1)
823 // If value is available in a register both base and index components have
824 // been picked, we can't fit the result available in the register in the
825 // addressing mode. Duplicate GlobalAddress or ConstantPool as displacement.
826 if (!AlreadySelected || (AM.Base.Reg.getNode() && AM.IndexReg.getNode())) {
827 SDValue N0 = N.getOperand(0);
828 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
829 GlobalValue *GV = G->getGlobal();
831 AM.Disp += G->getOffset();
832 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
834 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
835 AM.CP = CP->getConstVal();
836 AM.Align = CP->getAlignment();
837 AM.Disp += CP->getOffset();
838 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
840 } else if (ExternalSymbolSDNode *S =dyn_cast<ExternalSymbolSDNode>(N0)) {
841 AM.ES = S->getSymbol();
842 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
844 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
845 AM.JT = J->getIndex();
846 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
853 case ISD::FrameIndex:
854 if (AM.BaseType == X86ISelAddressMode::RegBase
855 && AM.Base.Reg.getNode() == 0) {
856 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
857 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
863 if (AlreadySelected || AM.IndexReg.getNode() != 0
864 || AM.Scale != 1 || AM.isRIPRel)
868 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
869 unsigned Val = CN->getZExtValue();
870 if (Val == 1 || Val == 2 || Val == 3) {
872 SDValue ShVal = N.getNode()->getOperand(0);
874 // Okay, we know that we have a scale by now. However, if the scaled
875 // value is an add of something and a constant, we can fold the
876 // constant into the disp field here.
877 if (ShVal.getNode()->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
878 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
879 AM.IndexReg = ShVal.getNode()->getOperand(0);
880 ConstantSDNode *AddVal =
881 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
882 uint64_t Disp = AM.Disp + (AddVal->getZExtValue() << Val);
897 // A mul_lohi where we need the low part can be folded as a plain multiply.
898 if (N.getResNo() != 0) break;
901 // X*[3,5,9] -> X+X*[2,4,8]
902 if (!AlreadySelected &&
903 AM.BaseType == X86ISelAddressMode::RegBase &&
904 AM.Base.Reg.getNode() == 0 &&
905 AM.IndexReg.getNode() == 0 &&
908 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
909 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
910 CN->getZExtValue() == 9) {
911 AM.Scale = unsigned(CN->getZExtValue())-1;
913 SDValue MulVal = N.getNode()->getOperand(0);
916 // Okay, we know that we have a scale by now. However, if the scaled
917 // value is an add of something and a constant, we can fold the
918 // constant into the disp field here.
919 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
920 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
921 Reg = MulVal.getNode()->getOperand(0);
922 ConstantSDNode *AddVal =
923 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
924 uint64_t Disp = AM.Disp + AddVal->getZExtValue() *
929 Reg = N.getNode()->getOperand(0);
931 Reg = N.getNode()->getOperand(0);
934 AM.IndexReg = AM.Base.Reg = Reg;
941 if (!AlreadySelected) {
942 X86ISelAddressMode Backup = AM;
943 if (!MatchAddress(N.getNode()->getOperand(0), AM, false, Depth+1) &&
944 !MatchAddress(N.getNode()->getOperand(1), AM, false, Depth+1))
947 if (!MatchAddress(N.getNode()->getOperand(1), AM, false, Depth+1) &&
948 !MatchAddress(N.getNode()->getOperand(0), AM, false, Depth+1))
955 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
956 if (AlreadySelected) break;
958 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
959 X86ISelAddressMode Backup = AM;
960 // Start with the LHS as an addr mode.
961 if (!MatchAddress(N.getOperand(0), AM, false) &&
962 // Address could not have picked a GV address for the displacement.
964 // On x86-64, the resultant disp must fit in 32-bits.
965 isInt32(AM.Disp + CN->getSExtValue()) &&
966 // Check to see if the LHS & C is zero.
967 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
968 AM.Disp += CN->getZExtValue();
976 // Handle "(x << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
977 // allows us to fold the shift into this addressing mode.
978 if (AlreadySelected) break;
979 SDValue Shift = N.getOperand(0);
980 if (Shift.getOpcode() != ISD::SHL) break;
982 // Scale must not be used already.
983 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
985 // Not when RIP is used as the base.
986 if (AM.isRIPRel) break;
988 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
989 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
990 if (!C1 || !C2) break;
992 // Not likely to be profitable if either the AND or SHIFT node has more
993 // than one use (unless all uses are for address computation). Besides,
994 // isel mechanism requires their node ids to be reused.
995 if (!N.hasOneUse() || !Shift.hasOneUse())
998 // Verify that the shift amount is something we can fold.
999 unsigned ShiftCst = C1->getZExtValue();
1000 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1003 // Get the new AND mask, this folds to a constant.
1004 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, N.getValueType(),
1005 SDValue(C2, 0), SDValue(C1, 0));
1006 SDValue NewAND = CurDAG->getNode(ISD::AND, N.getValueType(),
1007 Shift.getOperand(0), NewANDMask);
1008 NewANDMask.getNode()->setNodeId(Shift.getNode()->getNodeId());
1009 NewAND.getNode()->setNodeId(N.getNode()->getNodeId());
1011 AM.Scale = 1 << ShiftCst;
1012 AM.IndexReg = NewAND;
1017 return MatchAddressBase(N, AM, isRoot, Depth);
1020 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1021 /// specified addressing mode without any further recursion.
1022 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM,
1023 bool isRoot, unsigned Depth) {
1024 // Is the base register already occupied?
1025 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
1026 // If so, check to see if the scale index register is set.
1027 if (AM.IndexReg.getNode() == 0 && !AM.isRIPRel) {
1033 // Otherwise, we cannot select it.
1037 // Default, generate it as a register.
1038 AM.BaseType = X86ISelAddressMode::RegBase;
1043 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1044 /// It returns the operands which make up the maximal addressing mode it can
1045 /// match by reference.
1046 bool X86DAGToDAGISel::SelectAddr(SDValue Op, SDValue N, SDValue &Base,
1047 SDValue &Scale, SDValue &Index,
1049 X86ISelAddressMode AM;
1050 if (MatchAddress(N, AM))
1053 MVT VT = N.getValueType();
1054 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1055 if (!AM.Base.Reg.getNode())
1056 AM.Base.Reg = CurDAG->getRegister(0, VT);
1059 if (!AM.IndexReg.getNode())
1060 AM.IndexReg = CurDAG->getRegister(0, VT);
1062 getAddressOperands(AM, Base, Scale, Index, Disp);
1066 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1067 /// match a load whose top elements are either undef or zeros. The load flavor
1068 /// is derived from the type of N, which is either v4f32 or v2f64.
1069 bool X86DAGToDAGISel::SelectScalarSSELoad(SDValue Op, SDValue Pred,
1070 SDValue N, SDValue &Base,
1071 SDValue &Scale, SDValue &Index,
1072 SDValue &Disp, SDValue &InChain,
1073 SDValue &OutChain) {
1074 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1075 InChain = N.getOperand(0).getValue(1);
1076 if (ISD::isNON_EXTLoad(InChain.getNode()) &&
1077 InChain.getValue(0).hasOneUse() &&
1079 CanBeFoldedBy(N.getNode(), Pred.getNode(), Op.getNode())) {
1080 LoadSDNode *LD = cast<LoadSDNode>(InChain);
1081 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
1083 OutChain = LD->getChain();
1088 // Also handle the case where we explicitly require zeros in the top
1089 // elements. This is a vector shuffle from the zero vector.
1090 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1091 // Check to see if the top elements are all zeros (or bitcast of zeros).
1092 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1093 N.getOperand(0).getNode()->hasOneUse() &&
1094 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1095 N.getOperand(0).getOperand(0).hasOneUse()) {
1096 // Okay, this is a zero extending load. Fold it.
1097 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1098 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
1100 OutChain = LD->getChain();
1101 InChain = SDValue(LD, 1);
1108 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1109 /// mode it matches can be cost effectively emitted as an LEA instruction.
1110 bool X86DAGToDAGISel::SelectLEAAddr(SDValue Op, SDValue N,
1111 SDValue &Base, SDValue &Scale,
1112 SDValue &Index, SDValue &Disp) {
1113 X86ISelAddressMode AM;
1114 if (MatchAddress(N, AM))
1117 MVT VT = N.getValueType();
1118 unsigned Complexity = 0;
1119 if (AM.BaseType == X86ISelAddressMode::RegBase)
1120 if (AM.Base.Reg.getNode())
1123 AM.Base.Reg = CurDAG->getRegister(0, VT);
1124 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1127 if (AM.IndexReg.getNode())
1130 AM.IndexReg = CurDAG->getRegister(0, VT);
1132 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1137 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1138 // to a LEA. This is determined with some expermentation but is by no means
1139 // optimal (especially for code size consideration). LEA is nice because of
1140 // its three-address nature. Tweak the cost function again when we can run
1141 // convertToThreeAddress() at register allocation time.
1142 if (AM.GV || AM.CP || AM.ES || AM.JT != -1) {
1143 // For X86-64, we should always use lea to materialize RIP relative
1145 if (Subtarget->is64Bit())
1151 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1154 if (Complexity > 2) {
1155 getAddressOperands(AM, Base, Scale, Index, Disp);
1161 bool X86DAGToDAGISel::TryFoldLoad(SDValue P, SDValue N,
1162 SDValue &Base, SDValue &Scale,
1163 SDValue &Index, SDValue &Disp) {
1164 if (ISD::isNON_EXTLoad(N.getNode()) &&
1166 CanBeFoldedBy(N.getNode(), P.getNode(), P.getNode()))
1167 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp);
1171 /// getGlobalBaseReg - Return an SDNode that returns the value of
1172 /// the global base register. Output instructions required to
1173 /// initialize the global base register, if necessary.
1175 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1176 assert(!Subtarget->is64Bit() && "X86-64 PIC uses RIP relative addressing");
1178 GlobalBaseReg = TM.getInstrInfo()->initializeGlobalBaseReg(BB->getParent());
1179 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1182 static SDNode *FindCallStartFromCall(SDNode *Node) {
1183 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1184 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1185 "Node doesn't have a token chain argument!");
1186 return FindCallStartFromCall(Node->getOperand(0).getNode());
1189 /// getTruncateTo8Bit - return an SDNode that implements a subreg based
1190 /// truncate of the specified operand to i8. This can be done with tablegen,
1191 /// except that this code uses MVT::Flag in a tricky way that happens to
1192 /// improve scheduling in some cases.
1193 SDNode *X86DAGToDAGISel::getTruncateTo8Bit(SDValue N0) {
1194 assert(!Subtarget->is64Bit() &&
1195 "getTruncateTo8Bit is only needed on x86-32!");
1196 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1198 // Ensure that the source register has an 8-bit subreg on 32-bit targets
1200 MVT N0VT = N0.getValueType();
1201 switch (N0VT.getSimpleVT()) {
1202 default: assert(0 && "Unknown truncate!");
1204 Opc = X86::MOV16to16_;
1207 Opc = X86::MOV32to32_;
1211 // The use of MVT::Flag here is not strictly accurate, but it helps
1212 // scheduling in some cases.
1213 N0 = SDValue(CurDAG->getTargetNode(Opc, N0VT, MVT::Flag, N0), 0);
1214 return CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1215 MVT::i8, N0, SRIdx, N0.getValue(1));
1219 SDNode *X86DAGToDAGISel::Select(SDValue N) {
1220 SDNode *Node = N.getNode();
1221 MVT NVT = Node->getValueType(0);
1223 unsigned Opcode = Node->getOpcode();
1226 DOUT << std::string(Indent, ' ') << "Selecting: ";
1227 DEBUG(Node->dump(CurDAG));
1232 if (Node->isMachineOpcode()) {
1234 DOUT << std::string(Indent-2, ' ') << "== ";
1235 DEBUG(Node->dump(CurDAG));
1239 return NULL; // Already selected.
1244 case X86ISD::GlobalBaseReg:
1245 return getGlobalBaseReg();
1248 // Turn ADD X, c to MOV32ri X+c. This cannot be done with tblgen'd
1249 // code and is matched first so to prevent it from being turned into
1251 // In 64-bit small code size mode, use LEA to take advantage of
1252 // RIP-relative addressing.
1253 if (TM.getCodeModel() != CodeModel::Small)
1255 MVT PtrVT = TLI.getPointerTy();
1256 SDValue N0 = N.getOperand(0);
1257 SDValue N1 = N.getOperand(1);
1258 if (N.getNode()->getValueType(0) == PtrVT &&
1259 N0.getOpcode() == X86ISD::Wrapper &&
1260 N1.getOpcode() == ISD::Constant) {
1261 unsigned Offset = (unsigned)cast<ConstantSDNode>(N1)->getZExtValue();
1263 // TODO: handle ExternalSymbolSDNode.
1264 if (GlobalAddressSDNode *G =
1265 dyn_cast<GlobalAddressSDNode>(N0.getOperand(0))) {
1266 C = CurDAG->getTargetGlobalAddress(G->getGlobal(), PtrVT,
1267 G->getOffset() + Offset);
1268 } else if (ConstantPoolSDNode *CP =
1269 dyn_cast<ConstantPoolSDNode>(N0.getOperand(0))) {
1270 C = CurDAG->getTargetConstantPool(CP->getConstVal(), PtrVT,
1272 CP->getOffset()+Offset);
1276 if (Subtarget->is64Bit()) {
1277 SDValue Ops[] = { CurDAG->getRegister(0, PtrVT), getI8Imm(1),
1278 CurDAG->getRegister(0, PtrVT), C };
1279 return CurDAG->SelectNodeTo(N.getNode(), X86::LEA64r,
1282 return CurDAG->SelectNodeTo(N.getNode(), X86::MOV32ri, PtrVT, C);
1286 // Other cases are handled by auto-generated code.
1290 case ISD::SMUL_LOHI:
1291 case ISD::UMUL_LOHI: {
1292 SDValue N0 = Node->getOperand(0);
1293 SDValue N1 = Node->getOperand(1);
1295 bool isSigned = Opcode == ISD::SMUL_LOHI;
1297 switch (NVT.getSimpleVT()) {
1298 default: assert(0 && "Unsupported VT!");
1299 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1300 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1301 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1302 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1305 switch (NVT.getSimpleVT()) {
1306 default: assert(0 && "Unsupported VT!");
1307 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1308 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1309 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1310 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1313 unsigned LoReg, HiReg;
1314 switch (NVT.getSimpleVT()) {
1315 default: assert(0 && "Unsupported VT!");
1316 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1317 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1318 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1319 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1322 SDValue Tmp0, Tmp1, Tmp2, Tmp3;
1323 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1324 // multiplty is commmutative
1326 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
1332 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), LoReg,
1333 N0, SDValue()).getValue(1);
1336 AddToISelQueue(N1.getOperand(0));
1337 AddToISelQueue(Tmp0);
1338 AddToISelQueue(Tmp1);
1339 AddToISelQueue(Tmp2);
1340 AddToISelQueue(Tmp3);
1341 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N1.getOperand(0), InFlag };
1343 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1344 InFlag = SDValue(CNode, 1);
1345 // Update the chain.
1346 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1350 SDValue(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1353 // Copy the low half of the result, if it is needed.
1354 if (!N.getValue(0).use_empty()) {
1355 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1356 LoReg, NVT, InFlag);
1357 InFlag = Result.getValue(2);
1358 ReplaceUses(N.getValue(0), Result);
1360 DOUT << std::string(Indent-2, ' ') << "=> ";
1361 DEBUG(Result.getNode()->dump(CurDAG));
1365 // Copy the high half of the result, if it is needed.
1366 if (!N.getValue(1).use_empty()) {
1368 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1369 // Prevent use of AH in a REX instruction by referencing AX instead.
1370 // Shift it down 8 bits.
1371 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1372 X86::AX, MVT::i16, InFlag);
1373 InFlag = Result.getValue(2);
1374 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
1375 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1376 // Then truncate it down to i8.
1377 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1378 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1379 MVT::i8, Result, SRIdx), 0);
1381 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1382 HiReg, NVT, InFlag);
1383 InFlag = Result.getValue(2);
1385 ReplaceUses(N.getValue(1), Result);
1387 DOUT << std::string(Indent-2, ' ') << "=> ";
1388 DEBUG(Result.getNode()->dump(CurDAG));
1401 case ISD::UDIVREM: {
1402 SDValue N0 = Node->getOperand(0);
1403 SDValue N1 = Node->getOperand(1);
1405 bool isSigned = Opcode == ISD::SDIVREM;
1407 switch (NVT.getSimpleVT()) {
1408 default: assert(0 && "Unsupported VT!");
1409 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1410 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1411 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1412 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1415 switch (NVT.getSimpleVT()) {
1416 default: assert(0 && "Unsupported VT!");
1417 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1418 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1419 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1420 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1423 unsigned LoReg, HiReg;
1424 unsigned ClrOpcode, SExtOpcode;
1425 switch (NVT.getSimpleVT()) {
1426 default: assert(0 && "Unsupported VT!");
1428 LoReg = X86::AL; HiReg = X86::AH;
1430 SExtOpcode = X86::CBW;
1433 LoReg = X86::AX; HiReg = X86::DX;
1434 ClrOpcode = X86::MOV16r0;
1435 SExtOpcode = X86::CWD;
1438 LoReg = X86::EAX; HiReg = X86::EDX;
1439 ClrOpcode = X86::MOV32r0;
1440 SExtOpcode = X86::CDQ;
1443 LoReg = X86::RAX; HiReg = X86::RDX;
1444 ClrOpcode = X86::MOV64r0;
1445 SExtOpcode = X86::CQO;
1449 SDValue Tmp0, Tmp1, Tmp2, Tmp3;
1450 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1453 if (NVT == MVT::i8 && !isSigned) {
1454 // Special case for div8, just use a move with zero extension to AX to
1455 // clear the upper 8 bits (AH).
1456 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Move, Chain;
1457 if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3)) {
1458 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N0.getOperand(0) };
1459 AddToISelQueue(N0.getOperand(0));
1460 AddToISelQueue(Tmp0);
1461 AddToISelQueue(Tmp1);
1462 AddToISelQueue(Tmp2);
1463 AddToISelQueue(Tmp3);
1465 SDValue(CurDAG->getTargetNode(X86::MOVZX16rm8, MVT::i16, MVT::Other,
1467 Chain = Move.getValue(1);
1468 ReplaceUses(N0.getValue(1), Chain);
1472 SDValue(CurDAG->getTargetNode(X86::MOVZX16rr8, MVT::i16, N0), 0);
1473 Chain = CurDAG->getEntryNode();
1475 Chain = CurDAG->getCopyToReg(Chain, X86::AX, Move, SDValue());
1476 InFlag = Chain.getValue(1);
1480 CurDAG->getCopyToReg(CurDAG->getEntryNode(),
1481 LoReg, N0, SDValue()).getValue(1);
1483 // Sign extend the low part into the high part.
1485 SDValue(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
1487 // Zero out the high part, effectively zero extending the input.
1488 SDValue ClrNode = SDValue(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
1489 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), HiReg,
1490 ClrNode, InFlag).getValue(1);
1495 AddToISelQueue(N1.getOperand(0));
1496 AddToISelQueue(Tmp0);
1497 AddToISelQueue(Tmp1);
1498 AddToISelQueue(Tmp2);
1499 AddToISelQueue(Tmp3);
1500 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N1.getOperand(0), InFlag };
1502 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1503 InFlag = SDValue(CNode, 1);
1504 // Update the chain.
1505 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1509 SDValue(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1512 // Copy the division (low) result, if it is needed.
1513 if (!N.getValue(0).use_empty()) {
1514 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1515 LoReg, NVT, InFlag);
1516 InFlag = Result.getValue(2);
1517 ReplaceUses(N.getValue(0), Result);
1519 DOUT << std::string(Indent-2, ' ') << "=> ";
1520 DEBUG(Result.getNode()->dump(CurDAG));
1524 // Copy the remainder (high) result, if it is needed.
1525 if (!N.getValue(1).use_empty()) {
1527 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1528 // Prevent use of AH in a REX instruction by referencing AX instead.
1529 // Shift it down 8 bits.
1530 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1531 X86::AX, MVT::i16, InFlag);
1532 InFlag = Result.getValue(2);
1533 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
1534 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1535 // Then truncate it down to i8.
1536 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1537 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1538 MVT::i8, Result, SRIdx), 0);
1540 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1541 HiReg, NVT, InFlag);
1542 InFlag = Result.getValue(2);
1544 ReplaceUses(N.getValue(1), Result);
1546 DOUT << std::string(Indent-2, ' ') << "=> ";
1547 DEBUG(Result.getNode()->dump(CurDAG));
1559 case ISD::SIGN_EXTEND_INREG: {
1560 MVT SVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
1561 if (SVT == MVT::i8 && !Subtarget->is64Bit()) {
1562 SDValue N0 = Node->getOperand(0);
1565 SDValue TruncOp = SDValue(getTruncateTo8Bit(N0), 0);
1567 switch (NVT.getSimpleVT()) {
1568 default: assert(0 && "Unknown sign_extend_inreg!");
1570 Opc = X86::MOVSX16rr8;
1573 Opc = X86::MOVSX32rr8;
1577 SDNode *ResNode = CurDAG->getTargetNode(Opc, NVT, TruncOp);
1580 DOUT << std::string(Indent-2, ' ') << "=> ";
1581 DEBUG(TruncOp.getNode()->dump(CurDAG));
1583 DOUT << std::string(Indent-2, ' ') << "=> ";
1584 DEBUG(ResNode->dump(CurDAG));
1593 case ISD::TRUNCATE: {
1594 if (NVT == MVT::i8 && !Subtarget->is64Bit()) {
1595 SDValue Input = Node->getOperand(0);
1596 AddToISelQueue(Node->getOperand(0));
1597 SDNode *ResNode = getTruncateTo8Bit(Input);
1600 DOUT << std::string(Indent-2, ' ') << "=> ";
1601 DEBUG(ResNode->dump(CurDAG));
1610 case ISD::DECLARE: {
1611 // Handle DECLARE nodes here because the second operand may have been
1612 // wrapped in X86ISD::Wrapper.
1613 SDValue Chain = Node->getOperand(0);
1614 SDValue N1 = Node->getOperand(1);
1615 SDValue N2 = Node->getOperand(2);
1616 if (!isa<FrameIndexSDNode>(N1))
1618 int FI = cast<FrameIndexSDNode>(N1)->getIndex();
1619 if (N2.getOpcode() == ISD::ADD &&
1620 N2.getOperand(0).getOpcode() == X86ISD::GlobalBaseReg)
1621 N2 = N2.getOperand(1);
1622 if (N2.getOpcode() == X86ISD::Wrapper &&
1623 isa<GlobalAddressSDNode>(N2.getOperand(0))) {
1625 cast<GlobalAddressSDNode>(N2.getOperand(0))->getGlobal();
1626 SDValue Tmp1 = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1627 SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GV, TLI.getPointerTy());
1628 AddToISelQueue(Chain);
1629 SDValue Ops[] = { Tmp1, Tmp2, Chain };
1630 return CurDAG->getTargetNode(TargetInstrInfo::DECLARE,
1631 MVT::Other, Ops, 3);
1637 SDNode *ResNode = SelectCode(N);
1640 DOUT << std::string(Indent-2, ' ') << "=> ";
1641 if (ResNode == NULL || ResNode == N.getNode())
1642 DEBUG(N.getNode()->dump(CurDAG));
1644 DEBUG(ResNode->dump(CurDAG));
1652 bool X86DAGToDAGISel::
1653 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1654 std::vector<SDValue> &OutOps) {
1655 SDValue Op0, Op1, Op2, Op3;
1656 switch (ConstraintCode) {
1657 case 'o': // offsetable ??
1658 case 'v': // not offsetable ??
1659 default: return true;
1661 if (!SelectAddr(Op, Op, Op0, Op1, Op2, Op3))
1666 OutOps.push_back(Op0);
1667 OutOps.push_back(Op1);
1668 OutOps.push_back(Op2);
1669 OutOps.push_back(Op3);
1670 AddToISelQueue(Op0);
1671 AddToISelQueue(Op1);
1672 AddToISelQueue(Op2);
1673 AddToISelQueue(Op3);
1677 /// createX86ISelDag - This pass converts a legalized DAG into a
1678 /// X86-specific DAG, ready for instruction scheduling.
1680 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM, bool Fast) {
1681 return new X86DAGToDAGISel(TM, Fast);