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"
44 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
46 //===----------------------------------------------------------------------===//
47 // Pattern Matcher Implementation
48 //===----------------------------------------------------------------------===//
51 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
52 /// SDValue's instead of register numbers for the leaves of the matched
54 struct X86ISelAddressMode {
60 struct { // This is really a union, discriminated by BaseType!
65 bool isRIPRel; // RIP as base?
73 unsigned Align; // CP alignment.
76 : BaseType(RegBase), isRIPRel(false), Scale(1), IndexReg(), Disp(0),
77 GV(0), CP(0), ES(0), JT(-1), Align(0) {
80 cerr << "X86ISelAddressMode " << this << "\n";
82 if (Base.Reg.getNode() != 0) Base.Reg.getNode()->dump();
84 cerr << " Base.FrameIndex " << Base.FrameIndex << "\n";
85 cerr << "isRIPRel " << isRIPRel << " Scale" << Scale << "\n";
87 if (IndexReg.getNode() != 0) IndexReg.getNode()->dump();
89 cerr << " Disp " << Disp << "\n";
90 cerr << "GV "; if (GV) GV->dump();
92 cerr << " CP "; if (CP) CP->dump();
95 cerr << "ES "; if (ES) cerr << ES; else cerr << "nul";
96 cerr << " JT" << JT << " Align" << Align << "\n";
102 //===--------------------------------------------------------------------===//
103 /// ISel - X86 specific code to select X86 machine instructions for
104 /// SelectionDAG operations.
106 class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
107 /// TM - Keep a reference to X86TargetMachine.
109 X86TargetMachine &TM;
111 /// X86Lowering - This object fully describes how to lower LLVM code to an
112 /// X86-specific SelectionDAG.
113 X86TargetLowering &X86Lowering;
115 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
116 /// make the right decision when generating code for different targets.
117 const X86Subtarget *Subtarget;
119 /// CurBB - Current BB being isel'd.
121 MachineBasicBlock *CurBB;
123 /// OptForSize - If true, selector should try to optimize for code size
124 /// instead of performance.
128 X86DAGToDAGISel(X86TargetMachine &tm, bool fast)
129 : SelectionDAGISel(*tm.getTargetLowering(), fast),
130 TM(tm), X86Lowering(*TM.getTargetLowering()),
131 Subtarget(&TM.getSubtarget<X86Subtarget>()),
134 virtual const char *getPassName() const {
135 return "X86 DAG->DAG Instruction Selection";
138 /// InstructionSelect - This callback is invoked by
139 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
140 virtual void InstructionSelect();
142 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
144 virtual bool CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) const;
146 // Include the pieces autogenerated from the target description.
147 #include "X86GenDAGISel.inc"
150 SDNode *Select(SDValue N);
151 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
153 bool MatchAddress(SDValue N, X86ISelAddressMode &AM,
154 bool isRoot = true, unsigned Depth = 0);
155 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM,
156 bool isRoot, unsigned Depth);
157 bool SelectAddr(SDValue Op, SDValue N, SDValue &Base,
158 SDValue &Scale, SDValue &Index, SDValue &Disp);
159 bool SelectLEAAddr(SDValue Op, SDValue N, SDValue &Base,
160 SDValue &Scale, SDValue &Index, SDValue &Disp);
161 bool SelectScalarSSELoad(SDValue Op, SDValue Pred,
162 SDValue N, SDValue &Base, SDValue &Scale,
163 SDValue &Index, SDValue &Disp,
164 SDValue &InChain, SDValue &OutChain);
165 bool TryFoldLoad(SDValue P, SDValue N,
166 SDValue &Base, SDValue &Scale,
167 SDValue &Index, SDValue &Disp);
168 void PreprocessForRMW();
169 void PreprocessForFPConvert();
171 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
172 /// inline asm expressions.
173 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
175 std::vector<SDValue> &OutOps);
177 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
179 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
180 SDValue &Scale, SDValue &Index,
182 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
183 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
185 Scale = getI8Imm(AM.Scale);
187 // These are 32-bit even in 64-bit mode since RIP relative offset
190 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp);
192 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
195 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32);
196 else if (AM.JT != -1)
197 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32);
199 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
202 /// getI8Imm - Return a target constant with the specified value, of type
204 inline SDValue getI8Imm(unsigned Imm) {
205 return CurDAG->getTargetConstant(Imm, MVT::i8);
208 /// getI16Imm - Return a target constant with the specified value, of type
210 inline SDValue getI16Imm(unsigned Imm) {
211 return CurDAG->getTargetConstant(Imm, MVT::i16);
214 /// getI32Imm - Return a target constant with the specified value, of type
216 inline SDValue getI32Imm(unsigned Imm) {
217 return CurDAG->getTargetConstant(Imm, MVT::i32);
220 /// getGlobalBaseReg - Return an SDNode that returns the value of
221 /// the global base register. Output instructions required to
222 /// initialize the global base register, if necessary.
224 SDNode *getGlobalBaseReg();
226 /// getTruncateTo8Bit - return an SDNode that implements a subreg based
227 /// truncate of the specified operand to i8. This can be done with tablegen,
228 /// except that this code uses MVT::Flag in a tricky way that happens to
229 /// improve scheduling in some cases.
230 SDNode *getTruncateTo8Bit(SDValue N0);
238 /// findFlagUse - Return use of MVT::Flag value produced by the specified
241 static SDNode *findFlagUse(SDNode *N) {
242 unsigned FlagResNo = N->getNumValues()-1;
243 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
245 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
246 SDValue Op = User->getOperand(i);
247 if (Op.getNode() == N && Op.getResNo() == FlagResNo)
254 /// findNonImmUse - Return true by reference in "found" if "Use" is an
255 /// non-immediate use of "Def". This function recursively traversing
256 /// up the operand chain ignoring certain nodes.
257 static void findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
258 SDNode *Root, bool &found,
259 SmallPtrSet<SDNode*, 16> &Visited) {
261 Use->getNodeId() < Def->getNodeId() ||
262 !Visited.insert(Use))
265 for (unsigned i = 0, e = Use->getNumOperands(); !found && i != e; ++i) {
266 SDNode *N = Use->getOperand(i).getNode();
268 if (Use == ImmedUse || Use == Root)
269 continue; // We are not looking for immediate use.
275 // Traverse up the operand chain.
276 findNonImmUse(N, Def, ImmedUse, Root, found, Visited);
280 /// isNonImmUse - Start searching from Root up the DAG to check is Def can
281 /// be reached. Return true if that's the case. However, ignore direct uses
282 /// by ImmedUse (which would be U in the example illustrated in
283 /// CanBeFoldedBy) and by Root (which can happen in the store case).
284 /// FIXME: to be really generic, we should allow direct use by any node
285 /// that is being folded. But realisticly since we only fold loads which
286 /// have one non-chain use, we only need to watch out for load/op/store
287 /// and load/op/cmp case where the root (store / cmp) may reach the load via
288 /// its chain operand.
289 static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
290 SmallPtrSet<SDNode*, 16> Visited;
292 findNonImmUse(Root, Def, ImmedUse, Root, found, Visited);
297 bool X86DAGToDAGISel::CanBeFoldedBy(SDNode *N, SDNode *U, SDNode *Root) const {
298 if (Fast) return false;
300 // If Root use can somehow reach N through a path that that doesn't contain
301 // U then folding N would create a cycle. e.g. In the following
302 // diagram, Root can reach N through X. If N is folded into into Root, then
303 // X is both a predecessor and a successor of U.
314 // * indicates nodes to be folded together.
316 // If Root produces a flag, then it gets (even more) interesting. Since it
317 // will be "glued" together with its flag use in the scheduler, we need to
318 // check if it might reach N.
337 // If FU (flag use) indirectly reaches N (the load), and Root folds N
338 // (call it Fold), then X is a predecessor of FU and a successor of
339 // Fold. But since Fold and FU are flagged together, this will create
340 // a cycle in the scheduling graph.
342 MVT VT = Root->getValueType(Root->getNumValues()-1);
343 while (VT == MVT::Flag) {
344 SDNode *FU = findFlagUse(Root);
348 VT = Root->getValueType(Root->getNumValues()-1);
351 return !isNonImmUse(Root, N, U);
354 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
355 /// and move load below the TokenFactor. Replace store's chain operand with
356 /// load's chain result.
357 static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
358 SDValue Store, SDValue TF) {
359 SmallVector<SDValue, 4> Ops;
360 for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i)
361 if (Load.getNode() == TF.getOperand(i).getNode())
362 Ops.push_back(Load.getOperand(0));
364 Ops.push_back(TF.getOperand(i));
365 CurDAG->UpdateNodeOperands(TF, &Ops[0], Ops.size());
366 CurDAG->UpdateNodeOperands(Load, TF, Load.getOperand(1), Load.getOperand(2));
367 CurDAG->UpdateNodeOperands(Store, Load.getValue(1), Store.getOperand(1),
368 Store.getOperand(2), Store.getOperand(3));
371 /// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
373 static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
375 if (N.getOpcode() == ISD::BIT_CONVERT)
378 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
379 if (!LD || LD->isVolatile())
381 if (LD->getAddressingMode() != ISD::UNINDEXED)
384 ISD::LoadExtType ExtType = LD->getExtensionType();
385 if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
389 N.getOperand(1) == Address &&
390 N.getNode()->isOperandOf(Chain.getNode())) {
397 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
398 /// operand and move load below the call's chain operand.
399 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
400 SDValue Call, SDValue Chain) {
401 SmallVector<SDValue, 8> Ops;
402 for (unsigned i = 0, e = Chain.getNode()->getNumOperands(); i != e; ++i)
403 if (Load.getNode() == Chain.getOperand(i).getNode())
404 Ops.push_back(Load.getOperand(0));
406 Ops.push_back(Chain.getOperand(i));
407 CurDAG->UpdateNodeOperands(Chain, &Ops[0], Ops.size());
408 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
409 Load.getOperand(1), Load.getOperand(2));
411 Ops.push_back(SDValue(Load.getNode(), 1));
412 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
413 Ops.push_back(Call.getOperand(i));
414 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
417 /// isCalleeLoad - Return true if call address is a load and it can be
418 /// moved below CALLSEQ_START and the chains leading up to the call.
419 /// Return the CALLSEQ_START by reference as a second output.
420 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
421 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
423 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
426 LD->getAddressingMode() != ISD::UNINDEXED ||
427 LD->getExtensionType() != ISD::NON_EXTLOAD)
430 // Now let's find the callseq_start.
431 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
432 if (!Chain.hasOneUse())
434 Chain = Chain.getOperand(0);
436 return Chain.getOperand(0).getNode() == Callee.getNode();
440 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
441 /// This is only run if not in -fast mode (aka -O0).
442 /// This allows the instruction selector to pick more read-modify-write
443 /// instructions. This is a common case:
453 /// [TokenFactor] [Op]
460 /// The fact the store's chain operand != load's chain will prevent the
461 /// (store (op (load))) instruction from being selected. We can transform it to:
480 void X86DAGToDAGISel::PreprocessForRMW() {
481 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
482 E = CurDAG->allnodes_end(); I != E; ++I) {
483 if (I->getOpcode() == X86ISD::CALL) {
484 /// Also try moving call address load from outside callseq_start to just
485 /// before the call to allow it to be folded.
503 SDValue Chain = I->getOperand(0);
504 SDValue Load = I->getOperand(1);
505 if (!isCalleeLoad(Load, Chain))
507 MoveBelowCallSeqStart(CurDAG, Load, SDValue(I, 0), Chain);
512 if (!ISD::isNON_TRUNCStore(I))
514 SDValue Chain = I->getOperand(0);
516 if (Chain.getNode()->getOpcode() != ISD::TokenFactor)
519 SDValue N1 = I->getOperand(1);
520 SDValue N2 = I->getOperand(2);
521 if ((N1.getValueType().isFloatingPoint() &&
522 !N1.getValueType().isVector()) ||
528 unsigned Opcode = N1.getNode()->getOpcode();
537 case ISD::VECTOR_SHUFFLE: {
538 SDValue N10 = N1.getOperand(0);
539 SDValue N11 = N1.getOperand(1);
540 RModW = isRMWLoad(N10, Chain, N2, Load);
542 RModW = isRMWLoad(N11, Chain, N2, Load);
555 SDValue N10 = N1.getOperand(0);
556 RModW = isRMWLoad(N10, Chain, N2, Load);
562 MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
569 /// PreprocessForFPConvert - Walk over the dag lowering fpround and fpextend
570 /// nodes that target the FP stack to be store and load to the stack. This is a
571 /// gross hack. We would like to simply mark these as being illegal, but when
572 /// we do that, legalize produces these when it expands calls, then expands
573 /// these in the same legalize pass. We would like dag combine to be able to
574 /// hack on these between the call expansion and the node legalization. As such
575 /// this pass basically does "really late" legalization of these inline with the
577 void X86DAGToDAGISel::PreprocessForFPConvert() {
578 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
579 E = CurDAG->allnodes_end(); I != E; ) {
580 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
581 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
584 // If the source and destination are SSE registers, then this is a legal
585 // conversion that should not be lowered.
586 MVT SrcVT = N->getOperand(0).getValueType();
587 MVT DstVT = N->getValueType(0);
588 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
589 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
590 if (SrcIsSSE && DstIsSSE)
593 if (!SrcIsSSE && !DstIsSSE) {
594 // If this is an FPStack extension, it is a noop.
595 if (N->getOpcode() == ISD::FP_EXTEND)
597 // If this is a value-preserving FPStack truncation, it is a noop.
598 if (N->getConstantOperandVal(1))
602 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
603 // FPStack has extload and truncstore. SSE can fold direct loads into other
604 // operations. Based on this, decide what we want to do.
606 if (N->getOpcode() == ISD::FP_ROUND)
607 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
609 MemVT = SrcIsSSE ? SrcVT : DstVT;
611 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
613 // FIXME: optimize the case where the src/dest is a load or store?
614 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(),
616 MemTmp, NULL, 0, MemVT);
617 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, DstVT, Store, MemTmp,
620 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
621 // extload we created. This will cause general havok on the dag because
622 // anything below the conversion could be folded into other existing nodes.
623 // To avoid invalidating 'I', back it up to the convert node.
625 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
627 // Now that we did that, the node is dead. Increment the iterator to the
628 // next node to process, then delete N.
630 CurDAG->DeleteNode(N);
634 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
635 /// when it has created a SelectionDAG for us to codegen.
636 void X86DAGToDAGISel::InstructionSelect() {
637 CurBB = BB; // BB can change as result of isel.
638 const Function *F = CurDAG->getMachineFunction().getFunction();
639 OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
645 // FIXME: This should only happen when not -fast.
646 PreprocessForFPConvert();
648 // Codegen the basic block.
650 DOUT << "===== Instruction selection begins:\n";
655 DOUT << "===== Instruction selection ends:\n";
658 CurDAG->RemoveDeadNodes();
661 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
662 /// the main function.
663 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
664 MachineFrameInfo *MFI) {
665 const TargetInstrInfo *TII = TM.getInstrInfo();
666 if (Subtarget->isTargetCygMing())
667 BuildMI(BB, TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
670 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
671 // If this is main, emit special code for main.
672 MachineBasicBlock *BB = MF.begin();
673 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
674 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
677 /// MatchAddress - Add the specified node to the specified addressing mode,
678 /// returning true if it cannot be done. This just pattern matches for the
680 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM,
681 bool isRoot, unsigned Depth) {
682 bool is64Bit = Subtarget->is64Bit();
683 DOUT << "MatchAddress: "; DEBUG(AM.dump());
686 return MatchAddressBase(N, AM, isRoot, Depth);
688 // RIP relative addressing: %rip + 32-bit displacement!
690 if (!AM.ES && AM.JT != -1 && N.getOpcode() == ISD::Constant) {
691 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
692 if (!is64Bit || isInt32(AM.Disp + Val)) {
700 switch (N.getOpcode()) {
702 case ISD::Constant: {
703 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
704 if (!is64Bit || isInt32(AM.Disp + Val)) {
711 case X86ISD::Wrapper: {
712 DOUT << "Wrapper: 64bit " << is64Bit;
713 DOUT << " AM "; DEBUG(AM.dump()); DOUT << "\n";
714 // Under X86-64 non-small code model, GV (and friends) are 64-bits.
715 // Also, base and index reg must be 0 in order to use rip as base.
716 if (is64Bit && (TM.getCodeModel() != CodeModel::Small ||
717 AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
719 if (AM.GV != 0 || AM.CP != 0 || AM.ES != 0 || AM.JT != -1)
721 // If value is available in a register both base and index components have
722 // been picked, we can't fit the result available in the register in the
723 // addressing mode. Duplicate GlobalAddress or ConstantPool as displacement.
725 SDValue N0 = N.getOperand(0);
726 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
727 uint64_t Offset = G->getOffset();
728 if (!is64Bit || isInt32(AM.Disp + Offset)) {
729 GlobalValue *GV = G->getGlobal();
732 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
735 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
736 uint64_t Offset = CP->getOffset();
737 if (!is64Bit || isInt32(AM.Disp + Offset)) {
738 AM.CP = CP->getConstVal();
739 AM.Align = CP->getAlignment();
741 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
744 } else if (ExternalSymbolSDNode *S =dyn_cast<ExternalSymbolSDNode>(N0)) {
745 AM.ES = S->getSymbol();
746 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
748 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
749 AM.JT = J->getIndex();
750 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
757 case ISD::FrameIndex:
758 if (AM.BaseType == X86ISelAddressMode::RegBase
759 && AM.Base.Reg.getNode() == 0) {
760 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
761 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
767 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1 || AM.isRIPRel)
771 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
772 unsigned Val = CN->getZExtValue();
773 if (Val == 1 || Val == 2 || Val == 3) {
775 SDValue ShVal = N.getNode()->getOperand(0);
777 // Okay, we know that we have a scale by now. However, if the scaled
778 // value is an add of something and a constant, we can fold the
779 // constant into the disp field here.
780 if (ShVal.getNode()->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
781 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
782 AM.IndexReg = ShVal.getNode()->getOperand(0);
783 ConstantSDNode *AddVal =
784 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
785 uint64_t Disp = AM.Disp + (AddVal->getZExtValue() << Val);
786 if (!is64Bit || isInt32(Disp))
800 // A mul_lohi where we need the low part can be folded as a plain multiply.
801 if (N.getResNo() != 0) break;
804 // X*[3,5,9] -> X+X*[2,4,8]
805 if (AM.BaseType == X86ISelAddressMode::RegBase &&
806 AM.Base.Reg.getNode() == 0 &&
807 AM.IndexReg.getNode() == 0 &&
810 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
811 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
812 CN->getZExtValue() == 9) {
813 AM.Scale = unsigned(CN->getZExtValue())-1;
815 SDValue MulVal = N.getNode()->getOperand(0);
818 // Okay, we know that we have a scale by now. However, if the scaled
819 // value is an add of something and a constant, we can fold the
820 // constant into the disp field here.
821 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
822 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
823 Reg = MulVal.getNode()->getOperand(0);
824 ConstantSDNode *AddVal =
825 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
826 uint64_t Disp = AM.Disp + AddVal->getZExtValue() *
828 if (!is64Bit || isInt32(Disp))
831 Reg = N.getNode()->getOperand(0);
833 Reg = N.getNode()->getOperand(0);
836 AM.IndexReg = AM.Base.Reg = Reg;
844 X86ISelAddressMode Backup = AM;
845 if (!MatchAddress(N.getNode()->getOperand(0), AM, false, Depth+1) &&
846 !MatchAddress(N.getNode()->getOperand(1), AM, false, Depth+1))
849 if (!MatchAddress(N.getNode()->getOperand(1), AM, false, Depth+1) &&
850 !MatchAddress(N.getNode()->getOperand(0), AM, false, Depth+1))
857 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
858 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
859 X86ISelAddressMode Backup = AM;
860 uint64_t Offset = CN->getSExtValue();
861 // Start with the LHS as an addr mode.
862 if (!MatchAddress(N.getOperand(0), AM, false) &&
863 // Address could not have picked a GV address for the displacement.
865 // On x86-64, the resultant disp must fit in 32-bits.
866 (!is64Bit || isInt32(AM.Disp + Offset)) &&
867 // Check to see if the LHS & C is zero.
868 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
877 // Handle "(x << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
878 // allows us to fold the shift into this addressing mode.
879 SDValue Shift = N.getOperand(0);
880 if (Shift.getOpcode() != ISD::SHL) break;
882 // Scale must not be used already.
883 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
885 // Not when RIP is used as the base.
886 if (AM.isRIPRel) break;
888 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
889 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
890 if (!C1 || !C2) break;
892 // Not likely to be profitable if either the AND or SHIFT node has more
893 // than one use (unless all uses are for address computation). Besides,
894 // isel mechanism requires their node ids to be reused.
895 if (!N.hasOneUse() || !Shift.hasOneUse())
898 // Verify that the shift amount is something we can fold.
899 unsigned ShiftCst = C1->getZExtValue();
900 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
903 // Get the new AND mask, this folds to a constant.
904 SDValue X = Shift.getOperand(0);
905 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, N.getValueType(),
906 SDValue(C2, 0), SDValue(C1, 0));
907 SDValue NewAND = CurDAG->getNode(ISD::AND, N.getValueType(), X, NewANDMask);
908 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, N.getValueType(),
909 NewAND, SDValue(C1, 0));
911 // Insert the new nodes into the topological ordering.
912 if (C1->getNodeId() > X.getNode()->getNodeId()) {
913 CurDAG->RepositionNode(X.getNode(), C1);
914 C1->setNodeId(X.getNode()->getNodeId());
916 if (NewANDMask.getNode()->getNodeId() == -1 ||
917 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
918 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
919 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
921 if (NewAND.getNode()->getNodeId() == -1 ||
922 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
923 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
924 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
926 if (NewSHIFT.getNode()->getNodeId() == -1 ||
927 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
928 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
929 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
932 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
934 AM.Scale = 1 << ShiftCst;
935 AM.IndexReg = NewAND;
940 return MatchAddressBase(N, AM, isRoot, Depth);
943 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
944 /// specified addressing mode without any further recursion.
945 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM,
946 bool isRoot, unsigned Depth) {
947 // Is the base register already occupied?
948 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
949 // If so, check to see if the scale index register is set.
950 if (AM.IndexReg.getNode() == 0 && !AM.isRIPRel) {
956 // Otherwise, we cannot select it.
960 // Default, generate it as a register.
961 AM.BaseType = X86ISelAddressMode::RegBase;
966 /// SelectAddr - returns true if it is able pattern match an addressing mode.
967 /// It returns the operands which make up the maximal addressing mode it can
968 /// match by reference.
969 bool X86DAGToDAGISel::SelectAddr(SDValue Op, SDValue N, SDValue &Base,
970 SDValue &Scale, SDValue &Index,
972 X86ISelAddressMode AM;
973 if (MatchAddress(N, AM))
976 MVT VT = N.getValueType();
977 if (AM.BaseType == X86ISelAddressMode::RegBase) {
978 if (!AM.Base.Reg.getNode())
979 AM.Base.Reg = CurDAG->getRegister(0, VT);
982 if (!AM.IndexReg.getNode())
983 AM.IndexReg = CurDAG->getRegister(0, VT);
985 getAddressOperands(AM, Base, Scale, Index, Disp);
989 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
990 /// match a load whose top elements are either undef or zeros. The load flavor
991 /// is derived from the type of N, which is either v4f32 or v2f64.
992 bool X86DAGToDAGISel::SelectScalarSSELoad(SDValue Op, SDValue Pred,
993 SDValue N, SDValue &Base,
994 SDValue &Scale, SDValue &Index,
995 SDValue &Disp, SDValue &InChain,
997 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
998 InChain = N.getOperand(0).getValue(1);
999 if (ISD::isNON_EXTLoad(InChain.getNode()) &&
1000 InChain.getValue(0).hasOneUse() &&
1002 CanBeFoldedBy(N.getNode(), Pred.getNode(), Op.getNode())) {
1003 LoadSDNode *LD = cast<LoadSDNode>(InChain);
1004 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
1006 OutChain = LD->getChain();
1011 // Also handle the case where we explicitly require zeros in the top
1012 // elements. This is a vector shuffle from the zero vector.
1013 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1014 // Check to see if the top elements are all zeros (or bitcast of zeros).
1015 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1016 N.getOperand(0).getNode()->hasOneUse() &&
1017 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1018 N.getOperand(0).getOperand(0).hasOneUse()) {
1019 // Okay, this is a zero extending load. Fold it.
1020 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1021 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp))
1023 OutChain = LD->getChain();
1024 InChain = SDValue(LD, 1);
1031 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1032 /// mode it matches can be cost effectively emitted as an LEA instruction.
1033 bool X86DAGToDAGISel::SelectLEAAddr(SDValue Op, SDValue N,
1034 SDValue &Base, SDValue &Scale,
1035 SDValue &Index, SDValue &Disp) {
1036 X86ISelAddressMode AM;
1037 if (MatchAddress(N, AM))
1040 MVT VT = N.getValueType();
1041 unsigned Complexity = 0;
1042 if (AM.BaseType == X86ISelAddressMode::RegBase)
1043 if (AM.Base.Reg.getNode())
1046 AM.Base.Reg = CurDAG->getRegister(0, VT);
1047 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1050 if (AM.IndexReg.getNode())
1053 AM.IndexReg = CurDAG->getRegister(0, VT);
1055 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1060 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1061 // to a LEA. This is determined with some expermentation but is by no means
1062 // optimal (especially for code size consideration). LEA is nice because of
1063 // its three-address nature. Tweak the cost function again when we can run
1064 // convertToThreeAddress() at register allocation time.
1065 if (AM.GV || AM.CP || AM.ES || AM.JT != -1) {
1066 // For X86-64, we should always use lea to materialize RIP relative
1068 if (Subtarget->is64Bit())
1074 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1077 if (Complexity > 2) {
1078 getAddressOperands(AM, Base, Scale, Index, Disp);
1084 bool X86DAGToDAGISel::TryFoldLoad(SDValue P, SDValue N,
1085 SDValue &Base, SDValue &Scale,
1086 SDValue &Index, SDValue &Disp) {
1087 if (ISD::isNON_EXTLoad(N.getNode()) &&
1089 CanBeFoldedBy(N.getNode(), P.getNode(), P.getNode()))
1090 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp);
1094 /// getGlobalBaseReg - Return an SDNode that returns the value of
1095 /// the global base register. Output instructions required to
1096 /// initialize the global base register, if necessary.
1098 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1099 MachineFunction *MF = CurBB->getParent();
1100 unsigned GlobalBaseReg = TM.getInstrInfo()->getGlobalBaseReg(MF);
1101 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1104 static SDNode *FindCallStartFromCall(SDNode *Node) {
1105 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1106 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1107 "Node doesn't have a token chain argument!");
1108 return FindCallStartFromCall(Node->getOperand(0).getNode());
1111 /// getTruncateTo8Bit - return an SDNode that implements a subreg based
1112 /// truncate of the specified operand to i8. This can be done with tablegen,
1113 /// except that this code uses MVT::Flag in a tricky way that happens to
1114 /// improve scheduling in some cases.
1115 SDNode *X86DAGToDAGISel::getTruncateTo8Bit(SDValue N0) {
1116 assert(!Subtarget->is64Bit() &&
1117 "getTruncateTo8Bit is only needed on x86-32!");
1118 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1120 // Ensure that the source register has an 8-bit subreg on 32-bit targets
1122 MVT N0VT = N0.getValueType();
1123 switch (N0VT.getSimpleVT()) {
1124 default: assert(0 && "Unknown truncate!");
1126 Opc = X86::MOV16to16_;
1129 Opc = X86::MOV32to32_;
1133 // The use of MVT::Flag here is not strictly accurate, but it helps
1134 // scheduling in some cases.
1135 N0 = SDValue(CurDAG->getTargetNode(Opc, N0VT, MVT::Flag, N0), 0);
1136 return CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1137 MVT::i8, N0, SRIdx, N0.getValue(1));
1140 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1141 SDValue Chain = Node->getOperand(0);
1142 SDValue In1 = Node->getOperand(1);
1143 SDValue In2L = Node->getOperand(2);
1144 SDValue In2H = Node->getOperand(3);
1145 SDValue Tmp0, Tmp1, Tmp2, Tmp3;
1146 if (!SelectAddr(In1, In1, Tmp0, Tmp1, Tmp2, Tmp3))
1148 SDValue LSI = Node->getOperand(4); // MemOperand
1149 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, In2L, In2H, LSI, Chain };
1150 return CurDAG->getTargetNode(Opc, MVT::i32, MVT::i32, MVT::Other, Ops, 8);
1153 SDNode *X86DAGToDAGISel::Select(SDValue N) {
1154 SDNode *Node = N.getNode();
1155 MVT NVT = Node->getValueType(0);
1157 unsigned Opcode = Node->getOpcode();
1160 DOUT << std::string(Indent, ' ') << "Selecting: ";
1161 DEBUG(Node->dump(CurDAG));
1166 if (Node->isMachineOpcode()) {
1168 DOUT << std::string(Indent-2, ' ') << "== ";
1169 DEBUG(Node->dump(CurDAG));
1173 return NULL; // Already selected.
1178 case X86ISD::GlobalBaseReg:
1179 return getGlobalBaseReg();
1181 case X86ISD::ATOMOR64_DAG:
1182 return SelectAtomic64(Node, X86::ATOMOR6432);
1183 case X86ISD::ATOMXOR64_DAG:
1184 return SelectAtomic64(Node, X86::ATOMXOR6432);
1185 case X86ISD::ATOMADD64_DAG:
1186 return SelectAtomic64(Node, X86::ATOMADD6432);
1187 case X86ISD::ATOMSUB64_DAG:
1188 return SelectAtomic64(Node, X86::ATOMSUB6432);
1189 case X86ISD::ATOMNAND64_DAG:
1190 return SelectAtomic64(Node, X86::ATOMNAND6432);
1191 case X86ISD::ATOMAND64_DAG:
1192 return SelectAtomic64(Node, X86::ATOMAND6432);
1193 case X86ISD::ATOMSWAP64_DAG:
1194 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1196 case ISD::SMUL_LOHI:
1197 case ISD::UMUL_LOHI: {
1198 SDValue N0 = Node->getOperand(0);
1199 SDValue N1 = Node->getOperand(1);
1201 bool isSigned = Opcode == ISD::SMUL_LOHI;
1203 switch (NVT.getSimpleVT()) {
1204 default: assert(0 && "Unsupported VT!");
1205 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1206 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1207 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1208 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1211 switch (NVT.getSimpleVT()) {
1212 default: assert(0 && "Unsupported VT!");
1213 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1214 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1215 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1216 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1219 unsigned LoReg, HiReg;
1220 switch (NVT.getSimpleVT()) {
1221 default: assert(0 && "Unsupported VT!");
1222 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1223 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1224 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1225 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1228 SDValue Tmp0, Tmp1, Tmp2, Tmp3;
1229 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1230 // multiplty is commmutative
1232 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3);
1237 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), LoReg,
1238 N0, SDValue()).getValue(1);
1241 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N1.getOperand(0), InFlag };
1243 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1244 InFlag = SDValue(CNode, 1);
1245 // Update the chain.
1246 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1249 SDValue(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1252 // Copy the low half of the result, if it is needed.
1253 if (!N.getValue(0).use_empty()) {
1254 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1255 LoReg, NVT, InFlag);
1256 InFlag = Result.getValue(2);
1257 ReplaceUses(N.getValue(0), Result);
1259 DOUT << std::string(Indent-2, ' ') << "=> ";
1260 DEBUG(Result.getNode()->dump(CurDAG));
1264 // Copy the high half of the result, if it is needed.
1265 if (!N.getValue(1).use_empty()) {
1267 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1268 // Prevent use of AH in a REX instruction by referencing AX instead.
1269 // Shift it down 8 bits.
1270 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1271 X86::AX, MVT::i16, InFlag);
1272 InFlag = Result.getValue(2);
1273 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
1274 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1275 // Then truncate it down to i8.
1276 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1277 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1278 MVT::i8, Result, SRIdx), 0);
1280 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1281 HiReg, NVT, InFlag);
1282 InFlag = Result.getValue(2);
1284 ReplaceUses(N.getValue(1), Result);
1286 DOUT << std::string(Indent-2, ' ') << "=> ";
1287 DEBUG(Result.getNode()->dump(CurDAG));
1300 case ISD::UDIVREM: {
1301 SDValue N0 = Node->getOperand(0);
1302 SDValue N1 = Node->getOperand(1);
1304 bool isSigned = Opcode == ISD::SDIVREM;
1306 switch (NVT.getSimpleVT()) {
1307 default: assert(0 && "Unsupported VT!");
1308 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1309 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1310 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1311 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1314 switch (NVT.getSimpleVT()) {
1315 default: assert(0 && "Unsupported VT!");
1316 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1317 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1318 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1319 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1322 unsigned LoReg, HiReg;
1323 unsigned ClrOpcode, SExtOpcode;
1324 switch (NVT.getSimpleVT()) {
1325 default: assert(0 && "Unsupported VT!");
1327 LoReg = X86::AL; HiReg = X86::AH;
1329 SExtOpcode = X86::CBW;
1332 LoReg = X86::AX; HiReg = X86::DX;
1333 ClrOpcode = X86::MOV16r0;
1334 SExtOpcode = X86::CWD;
1337 LoReg = X86::EAX; HiReg = X86::EDX;
1338 ClrOpcode = X86::MOV32r0;
1339 SExtOpcode = X86::CDQ;
1342 LoReg = X86::RAX; HiReg = X86::RDX;
1343 ClrOpcode = X86::MOV64r0;
1344 SExtOpcode = X86::CQO;
1348 SDValue Tmp0, Tmp1, Tmp2, Tmp3;
1349 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3);
1352 if (NVT == MVT::i8 && !isSigned) {
1353 // Special case for div8, just use a move with zero extension to AX to
1354 // clear the upper 8 bits (AH).
1355 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Move, Chain;
1356 if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3)) {
1357 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N0.getOperand(0) };
1359 SDValue(CurDAG->getTargetNode(X86::MOVZX16rm8, MVT::i16, MVT::Other,
1361 Chain = Move.getValue(1);
1362 ReplaceUses(N0.getValue(1), Chain);
1365 SDValue(CurDAG->getTargetNode(X86::MOVZX16rr8, MVT::i16, N0), 0);
1366 Chain = CurDAG->getEntryNode();
1368 Chain = CurDAG->getCopyToReg(Chain, X86::AX, Move, SDValue());
1369 InFlag = Chain.getValue(1);
1372 CurDAG->getCopyToReg(CurDAG->getEntryNode(),
1373 LoReg, N0, SDValue()).getValue(1);
1375 // Sign extend the low part into the high part.
1377 SDValue(CurDAG->getTargetNode(SExtOpcode, MVT::Flag, InFlag), 0);
1379 // Zero out the high part, effectively zero extending the input.
1380 SDValue ClrNode = SDValue(CurDAG->getTargetNode(ClrOpcode, NVT), 0);
1381 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), HiReg,
1382 ClrNode, InFlag).getValue(1);
1387 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, N1.getOperand(0), InFlag };
1389 CurDAG->getTargetNode(MOpc, MVT::Other, MVT::Flag, Ops, 6);
1390 InFlag = SDValue(CNode, 1);
1391 // Update the chain.
1392 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1395 SDValue(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
1398 // Copy the division (low) result, if it is needed.
1399 if (!N.getValue(0).use_empty()) {
1400 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1401 LoReg, NVT, InFlag);
1402 InFlag = Result.getValue(2);
1403 ReplaceUses(N.getValue(0), Result);
1405 DOUT << std::string(Indent-2, ' ') << "=> ";
1406 DEBUG(Result.getNode()->dump(CurDAG));
1410 // Copy the remainder (high) result, if it is needed.
1411 if (!N.getValue(1).use_empty()) {
1413 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1414 // Prevent use of AH in a REX instruction by referencing AX instead.
1415 // Shift it down 8 bits.
1416 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1417 X86::AX, MVT::i16, InFlag);
1418 InFlag = Result.getValue(2);
1419 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
1420 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1421 // Then truncate it down to i8.
1422 SDValue SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
1423 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
1424 MVT::i8, Result, SRIdx), 0);
1426 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
1427 HiReg, NVT, InFlag);
1428 InFlag = Result.getValue(2);
1430 ReplaceUses(N.getValue(1), Result);
1432 DOUT << std::string(Indent-2, ' ') << "=> ";
1433 DEBUG(Result.getNode()->dump(CurDAG));
1445 case ISD::SIGN_EXTEND_INREG: {
1446 MVT SVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
1447 if (SVT == MVT::i8 && !Subtarget->is64Bit()) {
1448 SDValue N0 = Node->getOperand(0);
1450 SDValue TruncOp = SDValue(getTruncateTo8Bit(N0), 0);
1452 switch (NVT.getSimpleVT()) {
1453 default: assert(0 && "Unknown sign_extend_inreg!");
1455 Opc = X86::MOVSX16rr8;
1458 Opc = X86::MOVSX32rr8;
1462 SDNode *ResNode = CurDAG->getTargetNode(Opc, NVT, TruncOp);
1465 DOUT << std::string(Indent-2, ' ') << "=> ";
1466 DEBUG(TruncOp.getNode()->dump(CurDAG));
1468 DOUT << std::string(Indent-2, ' ') << "=> ";
1469 DEBUG(ResNode->dump(CurDAG));
1478 case ISD::TRUNCATE: {
1479 if (NVT == MVT::i8 && !Subtarget->is64Bit()) {
1480 SDValue Input = Node->getOperand(0);
1481 SDNode *ResNode = getTruncateTo8Bit(Input);
1484 DOUT << std::string(Indent-2, ' ') << "=> ";
1485 DEBUG(ResNode->dump(CurDAG));
1494 case ISD::DECLARE: {
1495 // Handle DECLARE nodes here because the second operand may have been
1496 // wrapped in X86ISD::Wrapper.
1497 SDValue Chain = Node->getOperand(0);
1498 SDValue N1 = Node->getOperand(1);
1499 SDValue N2 = Node->getOperand(2);
1500 if (!isa<FrameIndexSDNode>(N1))
1502 int FI = cast<FrameIndexSDNode>(N1)->getIndex();
1503 if (N2.getOpcode() == ISD::ADD &&
1504 N2.getOperand(0).getOpcode() == X86ISD::GlobalBaseReg)
1505 N2 = N2.getOperand(1);
1506 if (N2.getOpcode() == X86ISD::Wrapper &&
1507 isa<GlobalAddressSDNode>(N2.getOperand(0))) {
1509 cast<GlobalAddressSDNode>(N2.getOperand(0))->getGlobal();
1510 SDValue Tmp1 = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1511 SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GV, TLI.getPointerTy());
1512 SDValue Ops[] = { Tmp1, Tmp2, Chain };
1513 return CurDAG->getTargetNode(TargetInstrInfo::DECLARE,
1514 MVT::Other, Ops, 3);
1520 SDNode *ResNode = SelectCode(N);
1523 DOUT << std::string(Indent-2, ' ') << "=> ";
1524 if (ResNode == NULL || ResNode == N.getNode())
1525 DEBUG(N.getNode()->dump(CurDAG));
1527 DEBUG(ResNode->dump(CurDAG));
1535 bool X86DAGToDAGISel::
1536 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1537 std::vector<SDValue> &OutOps) {
1538 SDValue Op0, Op1, Op2, Op3;
1539 switch (ConstraintCode) {
1540 case 'o': // offsetable ??
1541 case 'v': // not offsetable ??
1542 default: return true;
1544 if (!SelectAddr(Op, Op, Op0, Op1, Op2, Op3))
1549 OutOps.push_back(Op0);
1550 OutOps.push_back(Op1);
1551 OutOps.push_back(Op2);
1552 OutOps.push_back(Op3);
1556 /// createX86ISelDag - This pass converts a legalized DAG into a
1557 /// X86-specific DAG, ready for instruction scheduling.
1559 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM, bool Fast) {
1560 return new X86DAGToDAGISel(TM, Fast);