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 #include "llvm/Support/CommandLine.h"
45 static cl::opt<bool> AvoidDupAddrCompute("x86-avoid-dup-address", cl::Hidden);
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?
77 unsigned Align; // CP alignment.
80 : BaseType(RegBase), isRIPRel(false), Scale(1), IndexReg(), Disp(0),
81 Segment(), GV(0), CP(0), ES(0), JT(-1), Align(0) {
84 bool hasSymbolicDisplacement() const {
85 return GV != 0 || CP != 0 || ES != 0 || JT != -1;
89 cerr << "X86ISelAddressMode " << this << "\n";
91 if (Base.Reg.getNode() != 0) Base.Reg.getNode()->dump();
93 cerr << " Base.FrameIndex " << Base.FrameIndex << "\n";
94 cerr << "isRIPRel " << isRIPRel << " Scale" << Scale << "\n";
96 if (IndexReg.getNode() != 0) IndexReg.getNode()->dump();
98 cerr << " Disp " << Disp << "\n";
99 cerr << "GV "; if (GV) GV->dump();
101 cerr << " CP "; if (CP) CP->dump();
104 cerr << "ES "; if (ES) cerr << ES; else cerr << "nul";
105 cerr << " JT" << JT << " Align" << Align << "\n";
111 //===--------------------------------------------------------------------===//
112 /// ISel - X86 specific code to select X86 machine instructions for
113 /// SelectionDAG operations.
115 class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
116 /// TM - Keep a reference to X86TargetMachine.
118 X86TargetMachine &TM;
120 /// X86Lowering - This object fully describes how to lower LLVM code to an
121 /// X86-specific SelectionDAG.
122 X86TargetLowering &X86Lowering;
124 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
125 /// make the right decision when generating code for different targets.
126 const X86Subtarget *Subtarget;
128 /// CurBB - Current BB being isel'd.
130 MachineBasicBlock *CurBB;
132 /// OptForSize - If true, selector should try to optimize for code size
133 /// instead of performance.
137 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
138 : SelectionDAGISel(tm, OptLevel),
139 TM(tm), X86Lowering(*TM.getTargetLowering()),
140 Subtarget(&TM.getSubtarget<X86Subtarget>()),
143 virtual const char *getPassName() const {
144 return "X86 DAG->DAG Instruction Selection";
147 /// InstructionSelect - This callback is invoked by
148 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
149 virtual void InstructionSelect();
151 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
154 bool IsLegalAndProfitableToFold(SDNode *N, SDNode *U, SDNode *Root) const;
156 // Include the pieces autogenerated from the target description.
157 #include "X86GenDAGISel.inc"
160 SDNode *Select(SDValue N);
161 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
163 bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
164 bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
165 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
166 bool MatchAddress(SDValue N, X86ISelAddressMode &AM,
168 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
169 bool SelectAddr(SDValue Op, SDValue N, SDValue &Base,
170 SDValue &Scale, SDValue &Index, SDValue &Disp,
172 bool SelectLEAAddr(SDValue Op, SDValue N, SDValue &Base,
173 SDValue &Scale, SDValue &Index, SDValue &Disp);
174 bool SelectScalarSSELoad(SDValue Op, SDValue Pred,
175 SDValue N, SDValue &Base, SDValue &Scale,
176 SDValue &Index, SDValue &Disp,
178 SDValue &InChain, SDValue &OutChain);
179 bool TryFoldLoad(SDValue P, SDValue N,
180 SDValue &Base, SDValue &Scale,
181 SDValue &Index, SDValue &Disp,
183 void PreprocessForRMW();
184 void PreprocessForFPConvert();
186 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
187 /// inline asm expressions.
188 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
190 std::vector<SDValue> &OutOps);
192 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
194 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
195 SDValue &Scale, SDValue &Index,
196 SDValue &Disp, SDValue &Segment) {
197 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
198 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
200 Scale = getI8Imm(AM.Scale);
202 // These are 32-bit even in 64-bit mode since RIP relative offset
205 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp);
207 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
210 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32);
211 else if (AM.JT != -1)
212 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32);
214 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
216 if (AM.Segment.getNode())
217 Segment = AM.Segment;
219 Segment = CurDAG->getRegister(0, MVT::i32);
222 /// getI8Imm - Return a target constant with the specified value, of type
224 inline SDValue getI8Imm(unsigned Imm) {
225 return CurDAG->getTargetConstant(Imm, MVT::i8);
228 /// getI16Imm - Return a target constant with the specified value, of type
230 inline SDValue getI16Imm(unsigned Imm) {
231 return CurDAG->getTargetConstant(Imm, MVT::i16);
234 /// getI32Imm - Return a target constant with the specified value, of type
236 inline SDValue getI32Imm(unsigned Imm) {
237 return CurDAG->getTargetConstant(Imm, MVT::i32);
240 /// getGlobalBaseReg - Return an SDNode that returns the value of
241 /// the global base register. Output instructions required to
242 /// initialize the global base register, if necessary.
244 SDNode *getGlobalBaseReg();
253 bool X86DAGToDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U,
254 SDNode *Root) const {
255 if (OptLevel == CodeGenOpt::None) return false;
258 switch (U->getOpcode()) {
266 SDValue Op1 = U->getOperand(1);
268 // If the other operand is a 8-bit immediate we should fold the immediate
269 // instead. This reduces code size.
271 // movl 4(%esp), %eax
275 // addl 4(%esp), %eax
276 // The former is 2 bytes shorter. In case where the increment is 1, then
277 // the saving can be 4 bytes (by using incl %eax).
278 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
279 if (Imm->getAPIntValue().isSignedIntN(8))
282 // If the other operand is a TLS address, we should fold it instead.
285 // leal i@NTPOFF(%eax), %eax
287 // movl $i@NTPOFF, %eax
289 // if the block also has an access to a second TLS address this will save
291 // FIXME: This is probably also true for non TLS addresses.
292 if (Op1.getOpcode() == X86ISD::Wrapper) {
293 SDValue Val = Op1.getOperand(0);
294 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
300 // Proceed to 'generic' cycle finder code
301 return SelectionDAGISel::IsLegalAndProfitableToFold(N, U, Root);
304 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
305 /// and move load below the TokenFactor. Replace store's chain operand with
306 /// load's chain result.
307 static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
308 SDValue Store, SDValue TF) {
309 SmallVector<SDValue, 4> Ops;
310 for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i)
311 if (Load.getNode() == TF.getOperand(i).getNode())
312 Ops.push_back(Load.getOperand(0));
314 Ops.push_back(TF.getOperand(i));
315 CurDAG->UpdateNodeOperands(TF, &Ops[0], Ops.size());
316 CurDAG->UpdateNodeOperands(Load, TF, Load.getOperand(1), Load.getOperand(2));
317 CurDAG->UpdateNodeOperands(Store, Load.getValue(1), Store.getOperand(1),
318 Store.getOperand(2), Store.getOperand(3));
321 /// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
323 static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
325 if (N.getOpcode() == ISD::BIT_CONVERT)
328 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
329 if (!LD || LD->isVolatile())
331 if (LD->getAddressingMode() != ISD::UNINDEXED)
334 ISD::LoadExtType ExtType = LD->getExtensionType();
335 if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
339 N.getOperand(1) == Address &&
340 N.getNode()->isOperandOf(Chain.getNode())) {
347 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
348 /// operand and move load below the call's chain operand.
349 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
350 SDValue Call, SDValue CallSeqStart) {
351 SmallVector<SDValue, 8> Ops;
352 SDValue Chain = CallSeqStart.getOperand(0);
353 if (Chain.getNode() == Load.getNode())
354 Ops.push_back(Load.getOperand(0));
356 assert(Chain.getOpcode() == ISD::TokenFactor &&
357 "Unexpected CallSeqStart chain operand");
358 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
359 if (Chain.getOperand(i).getNode() == Load.getNode())
360 Ops.push_back(Load.getOperand(0));
362 Ops.push_back(Chain.getOperand(i));
364 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
365 MVT::Other, &Ops[0], Ops.size());
367 Ops.push_back(NewChain);
369 for (unsigned i = 1, e = CallSeqStart.getNumOperands(); i != e; ++i)
370 Ops.push_back(CallSeqStart.getOperand(i));
371 CurDAG->UpdateNodeOperands(CallSeqStart, &Ops[0], Ops.size());
372 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
373 Load.getOperand(1), Load.getOperand(2));
375 Ops.push_back(SDValue(Load.getNode(), 1));
376 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
377 Ops.push_back(Call.getOperand(i));
378 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
381 /// isCalleeLoad - Return true if call address is a load and it can be
382 /// moved below CALLSEQ_START and the chains leading up to the call.
383 /// Return the CALLSEQ_START by reference as a second output.
384 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
385 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
387 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
390 LD->getAddressingMode() != ISD::UNINDEXED ||
391 LD->getExtensionType() != ISD::NON_EXTLOAD)
394 // Now let's find the callseq_start.
395 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
396 if (!Chain.hasOneUse())
398 Chain = Chain.getOperand(0);
401 if (Chain.getOperand(0).getNode() == Callee.getNode())
403 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
404 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()))
410 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
411 /// This is only run if not in -O0 mode.
412 /// This allows the instruction selector to pick more read-modify-write
413 /// instructions. This is a common case:
423 /// [TokenFactor] [Op]
430 /// The fact the store's chain operand != load's chain will prevent the
431 /// (store (op (load))) instruction from being selected. We can transform it to:
450 void X86DAGToDAGISel::PreprocessForRMW() {
451 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
452 E = CurDAG->allnodes_end(); I != E; ++I) {
453 if (I->getOpcode() == X86ISD::CALL) {
454 /// Also try moving call address load from outside callseq_start to just
455 /// before the call to allow it to be folded.
473 SDValue Chain = I->getOperand(0);
474 SDValue Load = I->getOperand(1);
475 if (!isCalleeLoad(Load, Chain))
477 MoveBelowCallSeqStart(CurDAG, Load, SDValue(I, 0), Chain);
482 if (!ISD::isNON_TRUNCStore(I))
484 SDValue Chain = I->getOperand(0);
486 if (Chain.getNode()->getOpcode() != ISD::TokenFactor)
489 SDValue N1 = I->getOperand(1);
490 SDValue N2 = I->getOperand(2);
491 if ((N1.getValueType().isFloatingPoint() &&
492 !N1.getValueType().isVector()) ||
498 unsigned Opcode = N1.getNode()->getOpcode();
507 case ISD::VECTOR_SHUFFLE: {
508 SDValue N10 = N1.getOperand(0);
509 SDValue N11 = N1.getOperand(1);
510 RModW = isRMWLoad(N10, Chain, N2, Load);
512 RModW = isRMWLoad(N11, Chain, N2, Load);
525 SDValue N10 = N1.getOperand(0);
526 RModW = isRMWLoad(N10, Chain, N2, Load);
532 MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
539 /// PreprocessForFPConvert - Walk over the dag lowering fpround and fpextend
540 /// nodes that target the FP stack to be store and load to the stack. This is a
541 /// gross hack. We would like to simply mark these as being illegal, but when
542 /// we do that, legalize produces these when it expands calls, then expands
543 /// these in the same legalize pass. We would like dag combine to be able to
544 /// hack on these between the call expansion and the node legalization. As such
545 /// this pass basically does "really late" legalization of these inline with the
547 void X86DAGToDAGISel::PreprocessForFPConvert() {
548 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
549 E = CurDAG->allnodes_end(); I != E; ) {
550 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
551 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
554 // If the source and destination are SSE registers, then this is a legal
555 // conversion that should not be lowered.
556 MVT SrcVT = N->getOperand(0).getValueType();
557 MVT DstVT = N->getValueType(0);
558 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
559 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
560 if (SrcIsSSE && DstIsSSE)
563 if (!SrcIsSSE && !DstIsSSE) {
564 // If this is an FPStack extension, it is a noop.
565 if (N->getOpcode() == ISD::FP_EXTEND)
567 // If this is a value-preserving FPStack truncation, it is a noop.
568 if (N->getConstantOperandVal(1))
572 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
573 // FPStack has extload and truncstore. SSE can fold direct loads into other
574 // operations. Based on this, decide what we want to do.
576 if (N->getOpcode() == ISD::FP_ROUND)
577 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
579 MemVT = SrcIsSSE ? SrcVT : DstVT;
581 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
582 DebugLoc dl = N->getDebugLoc();
584 // FIXME: optimize the case where the src/dest is a load or store?
585 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
587 MemTmp, NULL, 0, MemVT);
588 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
591 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
592 // extload we created. This will cause general havok on the dag because
593 // anything below the conversion could be folded into other existing nodes.
594 // To avoid invalidating 'I', back it up to the convert node.
596 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
598 // Now that we did that, the node is dead. Increment the iterator to the
599 // next node to process, then delete N.
601 CurDAG->DeleteNode(N);
605 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
606 /// when it has created a SelectionDAG for us to codegen.
607 void X86DAGToDAGISel::InstructionSelect() {
608 CurBB = BB; // BB can change as result of isel.
609 const Function *F = CurDAG->getMachineFunction().getFunction();
610 OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
613 if (OptLevel != CodeGenOpt::None)
616 // FIXME: This should only happen when not compiled with -O0.
617 PreprocessForFPConvert();
619 // Codegen the basic block.
621 DOUT << "===== Instruction selection begins:\n";
626 DOUT << "===== Instruction selection ends:\n";
629 CurDAG->RemoveDeadNodes();
632 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
633 /// the main function.
634 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
635 MachineFrameInfo *MFI) {
636 const TargetInstrInfo *TII = TM.getInstrInfo();
637 if (Subtarget->isTargetCygMing())
638 BuildMI(BB, DebugLoc::getUnknownLoc(),
639 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
642 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
643 // If this is main, emit special code for main.
644 MachineBasicBlock *BB = MF.begin();
645 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
646 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
650 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
651 X86ISelAddressMode &AM) {
652 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
653 SDValue Segment = N.getOperand(0);
655 if (AM.Segment.getNode() == 0) {
656 AM.Segment = Segment;
663 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
664 // This optimization is valid because the GNU TLS model defines that
665 // gs:0 (or fs:0 on X86-64) contains its own address.
666 // For more information see http://people.redhat.com/drepper/tls.pdf
668 SDValue Address = N.getOperand(1);
669 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
670 !MatchSegmentBaseAddress (Address, AM))
676 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
677 bool is64Bit = Subtarget->is64Bit();
678 DOUT << "Wrapper: 64bit " << is64Bit;
679 DOUT << " AM "; DEBUG(AM.dump()); DOUT << "\n";
681 // Under X86-64 non-small code model, GV (and friends) are 64-bits.
682 if (is64Bit && (TM.getCodeModel() != CodeModel::Small))
685 // Base and index reg must be 0 in order to use rip as base.
686 bool canUsePICRel = !AM.Base.Reg.getNode() && !AM.IndexReg.getNode();
687 if (is64Bit && !canUsePICRel && TM.symbolicAddressesAreRIPRel())
690 if (AM.hasSymbolicDisplacement())
692 // If value is available in a register both base and index components have
693 // been picked, we can't fit the result available in the register in the
694 // addressing mode. Duplicate GlobalAddress or ConstantPool as displacement.
696 SDValue N0 = N.getOperand(0);
697 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
698 uint64_t Offset = G->getOffset();
699 if (!is64Bit || isInt32(AM.Disp + Offset)) {
700 GlobalValue *GV = G->getGlobal();
701 bool isRIPRel = TM.symbolicAddressesAreRIPRel();
702 if (N0.getOpcode() == llvm::ISD::TargetGlobalTLSAddress) {
703 TLSModel::Model model =
704 getTLSModel (GV, TM.getRelocationModel());
705 if (is64Bit && model == TLSModel::InitialExec)
710 AM.isRIPRel = isRIPRel;
713 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
714 uint64_t Offset = CP->getOffset();
715 if (!is64Bit || isInt32(AM.Disp + Offset)) {
716 AM.CP = CP->getConstVal();
717 AM.Align = CP->getAlignment();
719 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
722 } else if (ExternalSymbolSDNode *S =dyn_cast<ExternalSymbolSDNode>(N0)) {
723 AM.ES = S->getSymbol();
724 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
726 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
727 AM.JT = J->getIndex();
728 AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
735 /// MatchAddress - Add the specified node to the specified addressing mode,
736 /// returning true if it cannot be done. This just pattern matches for the
738 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM,
740 bool is64Bit = Subtarget->is64Bit();
741 DebugLoc dl = N.getDebugLoc();
742 DOUT << "MatchAddress: "; DEBUG(AM.dump());
745 return MatchAddressBase(N, AM);
747 // RIP relative addressing: %rip + 32-bit displacement!
749 if (!AM.ES && AM.JT != -1 && N.getOpcode() == ISD::Constant) {
750 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
751 if (!is64Bit || isInt32(AM.Disp + Val)) {
759 switch (N.getOpcode()) {
761 case ISD::Constant: {
762 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
763 if (!is64Bit || isInt32(AM.Disp + Val)) {
770 case X86ISD::SegmentBaseAddress:
771 if (!MatchSegmentBaseAddress(N, AM))
775 case X86ISD::Wrapper:
776 if (!MatchWrapper(N, AM))
781 if (!MatchLoad(N, AM))
785 case ISD::FrameIndex:
786 if (AM.BaseType == X86ISelAddressMode::RegBase
787 && AM.Base.Reg.getNode() == 0) {
788 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
789 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
795 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1 || AM.isRIPRel)
799 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
800 unsigned Val = CN->getZExtValue();
801 if (Val == 1 || Val == 2 || Val == 3) {
803 SDValue ShVal = N.getNode()->getOperand(0);
805 // Okay, we know that we have a scale by now. However, if the scaled
806 // value is an add of something and a constant, we can fold the
807 // constant into the disp field here.
808 if (ShVal.getNode()->getOpcode() == ISD::ADD && ShVal.hasOneUse() &&
809 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
810 AM.IndexReg = ShVal.getNode()->getOperand(0);
811 ConstantSDNode *AddVal =
812 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
813 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
814 if (!is64Bit || isInt32(Disp))
828 // A mul_lohi where we need the low part can be folded as a plain multiply.
829 if (N.getResNo() != 0) break;
832 case X86ISD::MUL_IMM:
833 // X*[3,5,9] -> X+X*[2,4,8]
834 if (AM.BaseType == X86ISelAddressMode::RegBase &&
835 AM.Base.Reg.getNode() == 0 &&
836 AM.IndexReg.getNode() == 0 &&
839 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
840 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
841 CN->getZExtValue() == 9) {
842 AM.Scale = unsigned(CN->getZExtValue())-1;
844 SDValue MulVal = N.getNode()->getOperand(0);
847 // Okay, we know that we have a scale by now. However, if the scaled
848 // value is an add of something and a constant, we can fold the
849 // constant into the disp field here.
850 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
851 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
852 Reg = MulVal.getNode()->getOperand(0);
853 ConstantSDNode *AddVal =
854 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
855 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
857 if (!is64Bit || isInt32(Disp))
860 Reg = N.getNode()->getOperand(0);
862 Reg = N.getNode()->getOperand(0);
865 AM.IndexReg = AM.Base.Reg = Reg;
872 // Given A-B, if A can be completely folded into the address and
873 // the index field with the index field unused, use -B as the index.
874 // This is a win if a has multiple parts that can be folded into
875 // the address. Also, this saves a mov if the base register has
876 // other uses, since it avoids a two-address sub instruction, however
877 // it costs an additional mov if the index register has other uses.
879 // Test if the LHS of the sub can be folded.
880 X86ISelAddressMode Backup = AM;
881 if (MatchAddress(N.getNode()->getOperand(0), AM, Depth+1)) {
885 // Test if the index field is free for use.
886 if (AM.IndexReg.getNode() || AM.isRIPRel) {
891 SDValue RHS = N.getNode()->getOperand(1);
892 // If the RHS involves a register with multiple uses, this
893 // transformation incurs an extra mov, due to the neg instruction
894 // clobbering its operand.
895 if (!RHS.getNode()->hasOneUse() ||
896 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
897 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
898 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
899 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
900 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
902 // If the base is a register with multiple uses, this
903 // transformation may save a mov.
904 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
905 AM.Base.Reg.getNode() &&
906 !AM.Base.Reg.getNode()->hasOneUse()) ||
907 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
909 // If the folded LHS was interesting, this transformation saves
910 // address arithmetic.
911 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
912 ((AM.Disp != 0) && (Backup.Disp == 0)) +
913 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
915 // If it doesn't look like it may be an overall win, don't do it.
921 // Ok, the transformation is legal and appears profitable. Go for it.
922 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
923 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
927 // Insert the new nodes into the topological ordering.
928 if (Zero.getNode()->getNodeId() == -1 ||
929 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
930 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
931 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
933 if (Neg.getNode()->getNodeId() == -1 ||
934 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
935 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
936 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
942 X86ISelAddressMode Backup = AM;
943 if (!MatchAddress(N.getNode()->getOperand(0), AM, Depth+1) &&
944 !MatchAddress(N.getNode()->getOperand(1), AM, Depth+1))
947 if (!MatchAddress(N.getNode()->getOperand(1), AM, Depth+1) &&
948 !MatchAddress(N.getNode()->getOperand(0), AM, Depth+1))
952 // If we couldn't fold both operands into the address at the same time,
953 // see if we can just put each operand into a register and fold at least
955 if (AM.BaseType == X86ISelAddressMode::RegBase &&
956 !AM.Base.Reg.getNode() &&
957 !AM.IndexReg.getNode() &&
959 AM.Base.Reg = N.getNode()->getOperand(0);
960 AM.IndexReg = N.getNode()->getOperand(1);
968 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
969 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
970 X86ISelAddressMode Backup = AM;
971 uint64_t Offset = CN->getSExtValue();
972 // Start with the LHS as an addr mode.
973 if (!MatchAddress(N.getOperand(0), AM, Depth+1) &&
974 // Address could not have picked a GV address for the displacement.
976 // On x86-64, the resultant disp must fit in 32-bits.
977 (!is64Bit || isInt32(AM.Disp + Offset)) &&
978 // Check to see if the LHS & C is zero.
979 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
988 // Perform some heroic transforms on an and of a constant-count shift
989 // with a constant to enable use of the scaled offset field.
991 SDValue Shift = N.getOperand(0);
992 if (Shift.getNumOperands() != 2) break;
994 // Scale must not be used already.
995 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
997 // Not when RIP is used as the base.
998 if (AM.isRIPRel) break;
1000 SDValue X = Shift.getOperand(0);
1001 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1002 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1003 if (!C1 || !C2) break;
1005 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1006 // allows us to convert the shift and and into an h-register extract and
1008 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1009 unsigned ScaleLog = 8 - C1->getZExtValue();
1010 if (ScaleLog > 0 && ScaleLog < 4 &&
1011 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1012 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1013 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1014 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1016 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1018 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1019 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1022 // Insert the new nodes into the topological ordering.
1023 if (Eight.getNode()->getNodeId() == -1 ||
1024 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1025 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1026 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1028 if (Mask.getNode()->getNodeId() == -1 ||
1029 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1030 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1031 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1033 if (Srl.getNode()->getNodeId() == -1 ||
1034 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1035 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1036 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1038 if (And.getNode()->getNodeId() == -1 ||
1039 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1040 CurDAG->RepositionNode(N.getNode(), And.getNode());
1041 And.getNode()->setNodeId(N.getNode()->getNodeId());
1043 if (ShlCount.getNode()->getNodeId() == -1 ||
1044 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1045 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1046 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1048 if (Shl.getNode()->getNodeId() == -1 ||
1049 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1050 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1051 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1053 CurDAG->ReplaceAllUsesWith(N, Shl);
1055 AM.Scale = (1 << ScaleLog);
1060 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1061 // allows us to fold the shift into this addressing mode.
1062 if (Shift.getOpcode() != ISD::SHL) break;
1064 // Not likely to be profitable if either the AND or SHIFT node has more
1065 // than one use (unless all uses are for address computation). Besides,
1066 // isel mechanism requires their node ids to be reused.
1067 if (!N.hasOneUse() || !Shift.hasOneUse())
1070 // Verify that the shift amount is something we can fold.
1071 unsigned ShiftCst = C1->getZExtValue();
1072 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1075 // Get the new AND mask, this folds to a constant.
1076 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1077 SDValue(C2, 0), SDValue(C1, 0));
1078 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1080 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1081 NewAND, SDValue(C1, 0));
1083 // Insert the new nodes into the topological ordering.
1084 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1085 CurDAG->RepositionNode(X.getNode(), C1);
1086 C1->setNodeId(X.getNode()->getNodeId());
1088 if (NewANDMask.getNode()->getNodeId() == -1 ||
1089 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1090 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1091 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1093 if (NewAND.getNode()->getNodeId() == -1 ||
1094 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1095 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1096 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1098 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1099 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1100 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1101 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1104 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1106 AM.Scale = 1 << ShiftCst;
1107 AM.IndexReg = NewAND;
1112 return MatchAddressBase(N, AM);
1115 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1116 /// specified addressing mode without any further recursion.
1117 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1118 // Is the base register already occupied?
1119 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
1120 // If so, check to see if the scale index register is set.
1121 if (AM.IndexReg.getNode() == 0 && !AM.isRIPRel) {
1127 // Otherwise, we cannot select it.
1131 // Default, generate it as a register.
1132 AM.BaseType = X86ISelAddressMode::RegBase;
1137 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1138 /// It returns the operands which make up the maximal addressing mode it can
1139 /// match by reference.
1140 bool X86DAGToDAGISel::SelectAddr(SDValue Op, SDValue N, SDValue &Base,
1141 SDValue &Scale, SDValue &Index,
1142 SDValue &Disp, SDValue &Segment) {
1143 X86ISelAddressMode AM;
1145 if (AvoidDupAddrCompute && !N.hasOneUse()) {
1146 unsigned Opcode = N.getOpcode();
1147 if (Opcode != ISD::Constant && Opcode != ISD::FrameIndex &&
1148 Opcode != X86ISD::Wrapper) {
1149 // If we are able to fold N into addressing mode, then we'll allow it even
1150 // if N has multiple uses. In general, addressing computation is used as
1151 // addresses by all of its uses. But watch out for CopyToReg uses, that
1152 // means the address computation is liveout. It will be computed by a LEA
1153 // so we want to avoid computing the address twice.
1154 for (SDNode::use_iterator UI = N.getNode()->use_begin(),
1155 UE = N.getNode()->use_end(); UI != UE; ++UI) {
1156 if (UI->getOpcode() == ISD::CopyToReg) {
1157 MatchAddressBase(N, AM);
1165 if (!Done && MatchAddress(N, AM))
1168 MVT VT = N.getValueType();
1169 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1170 if (!AM.Base.Reg.getNode())
1171 AM.Base.Reg = CurDAG->getRegister(0, VT);
1174 if (!AM.IndexReg.getNode())
1175 AM.IndexReg = CurDAG->getRegister(0, VT);
1177 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1181 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1182 /// match a load whose top elements are either undef or zeros. The load flavor
1183 /// is derived from the type of N, which is either v4f32 or v2f64.
1184 bool X86DAGToDAGISel::SelectScalarSSELoad(SDValue Op, SDValue Pred,
1185 SDValue N, SDValue &Base,
1186 SDValue &Scale, SDValue &Index,
1187 SDValue &Disp, SDValue &Segment,
1189 SDValue &OutChain) {
1190 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1191 InChain = N.getOperand(0).getValue(1);
1192 if (ISD::isNON_EXTLoad(InChain.getNode()) &&
1193 InChain.getValue(0).hasOneUse() &&
1195 IsLegalAndProfitableToFold(N.getNode(), Pred.getNode(), Op.getNode())) {
1196 LoadSDNode *LD = cast<LoadSDNode>(InChain);
1197 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1199 OutChain = LD->getChain();
1204 // Also handle the case where we explicitly require zeros in the top
1205 // elements. This is a vector shuffle from the zero vector.
1206 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1207 // Check to see if the top elements are all zeros (or bitcast of zeros).
1208 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1209 N.getOperand(0).getNode()->hasOneUse() &&
1210 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1211 N.getOperand(0).getOperand(0).hasOneUse()) {
1212 // Okay, this is a zero extending load. Fold it.
1213 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1214 if (!SelectAddr(Op, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1216 OutChain = LD->getChain();
1217 InChain = SDValue(LD, 1);
1224 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1225 /// mode it matches can be cost effectively emitted as an LEA instruction.
1226 bool X86DAGToDAGISel::SelectLEAAddr(SDValue Op, SDValue N,
1227 SDValue &Base, SDValue &Scale,
1228 SDValue &Index, SDValue &Disp) {
1229 X86ISelAddressMode AM;
1231 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1233 SDValue Copy = AM.Segment;
1234 SDValue T = CurDAG->getRegister(0, MVT::i32);
1236 if (MatchAddress(N, AM))
1238 assert (T == AM.Segment);
1241 MVT VT = N.getValueType();
1242 unsigned Complexity = 0;
1243 if (AM.BaseType == X86ISelAddressMode::RegBase)
1244 if (AM.Base.Reg.getNode())
1247 AM.Base.Reg = CurDAG->getRegister(0, VT);
1248 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1251 if (AM.IndexReg.getNode())
1254 AM.IndexReg = CurDAG->getRegister(0, VT);
1256 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1261 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1262 // to a LEA. This is determined with some expermentation but is by no means
1263 // optimal (especially for code size consideration). LEA is nice because of
1264 // its three-address nature. Tweak the cost function again when we can run
1265 // convertToThreeAddress() at register allocation time.
1266 if (AM.hasSymbolicDisplacement()) {
1267 // For X86-64, we should always use lea to materialize RIP relative
1269 if (Subtarget->is64Bit())
1275 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1278 if (Complexity > 2) {
1280 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1286 bool X86DAGToDAGISel::TryFoldLoad(SDValue P, SDValue N,
1287 SDValue &Base, SDValue &Scale,
1288 SDValue &Index, SDValue &Disp,
1290 if (ISD::isNON_EXTLoad(N.getNode()) &&
1292 IsLegalAndProfitableToFold(N.getNode(), P.getNode(), P.getNode()))
1293 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
1297 /// getGlobalBaseReg - Return an SDNode that returns the value of
1298 /// the global base register. Output instructions required to
1299 /// initialize the global base register, if necessary.
1301 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1302 MachineFunction *MF = CurBB->getParent();
1303 unsigned GlobalBaseReg = TM.getInstrInfo()->getGlobalBaseReg(MF);
1304 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1307 static SDNode *FindCallStartFromCall(SDNode *Node) {
1308 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1309 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1310 "Node doesn't have a token chain argument!");
1311 return FindCallStartFromCall(Node->getOperand(0).getNode());
1314 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1315 SDValue Chain = Node->getOperand(0);
1316 SDValue In1 = Node->getOperand(1);
1317 SDValue In2L = Node->getOperand(2);
1318 SDValue In2H = Node->getOperand(3);
1319 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1320 if (!SelectAddr(In1, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1322 SDValue LSI = Node->getOperand(4); // MemOperand
1323 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, LSI, Chain};
1324 return CurDAG->getTargetNode(Opc, Node->getDebugLoc(),
1325 MVT::i32, MVT::i32, MVT::Other, Ops,
1326 array_lengthof(Ops));
1329 SDNode *X86DAGToDAGISel::Select(SDValue N) {
1330 SDNode *Node = N.getNode();
1331 MVT NVT = Node->getValueType(0);
1333 unsigned Opcode = Node->getOpcode();
1334 DebugLoc dl = Node->getDebugLoc();
1337 DOUT << std::string(Indent, ' ') << "Selecting: ";
1338 DEBUG(Node->dump(CurDAG));
1343 if (Node->isMachineOpcode()) {
1345 DOUT << std::string(Indent-2, ' ') << "== ";
1346 DEBUG(Node->dump(CurDAG));
1350 return NULL; // Already selected.
1355 case X86ISD::GlobalBaseReg:
1356 return getGlobalBaseReg();
1358 case X86ISD::ATOMOR64_DAG:
1359 return SelectAtomic64(Node, X86::ATOMOR6432);
1360 case X86ISD::ATOMXOR64_DAG:
1361 return SelectAtomic64(Node, X86::ATOMXOR6432);
1362 case X86ISD::ATOMADD64_DAG:
1363 return SelectAtomic64(Node, X86::ATOMADD6432);
1364 case X86ISD::ATOMSUB64_DAG:
1365 return SelectAtomic64(Node, X86::ATOMSUB6432);
1366 case X86ISD::ATOMNAND64_DAG:
1367 return SelectAtomic64(Node, X86::ATOMNAND6432);
1368 case X86ISD::ATOMAND64_DAG:
1369 return SelectAtomic64(Node, X86::ATOMAND6432);
1370 case X86ISD::ATOMSWAP64_DAG:
1371 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1373 case ISD::SMUL_LOHI:
1374 case ISD::UMUL_LOHI: {
1375 SDValue N0 = Node->getOperand(0);
1376 SDValue N1 = Node->getOperand(1);
1378 bool isSigned = Opcode == ISD::SMUL_LOHI;
1380 switch (NVT.getSimpleVT()) {
1381 default: assert(0 && "Unsupported VT!");
1382 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1383 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1384 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1385 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1388 switch (NVT.getSimpleVT()) {
1389 default: assert(0 && "Unsupported VT!");
1390 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1391 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1392 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1393 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1396 unsigned LoReg, HiReg;
1397 switch (NVT.getSimpleVT()) {
1398 default: assert(0 && "Unsupported VT!");
1399 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1400 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1401 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1402 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1405 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1406 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1407 // multiplty is commmutative
1409 foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1414 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1415 N0, SDValue()).getValue(1);
1418 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1421 CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1422 array_lengthof(Ops));
1423 InFlag = SDValue(CNode, 1);
1424 // Update the chain.
1425 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1428 SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1431 // Copy the low half of the result, if it is needed.
1432 if (!N.getValue(0).use_empty()) {
1433 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1434 LoReg, NVT, InFlag);
1435 InFlag = Result.getValue(2);
1436 ReplaceUses(N.getValue(0), Result);
1438 DOUT << std::string(Indent-2, ' ') << "=> ";
1439 DEBUG(Result.getNode()->dump(CurDAG));
1443 // Copy the high half of the result, if it is needed.
1444 if (!N.getValue(1).use_empty()) {
1446 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1447 // Prevent use of AH in a REX instruction by referencing AX instead.
1448 // Shift it down 8 bits.
1449 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1450 X86::AX, MVT::i16, InFlag);
1451 InFlag = Result.getValue(2);
1452 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
1454 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1455 // Then truncate it down to i8.
1456 SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
1457 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
1458 MVT::i8, Result, SRIdx), 0);
1460 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1461 HiReg, NVT, InFlag);
1462 InFlag = Result.getValue(2);
1464 ReplaceUses(N.getValue(1), Result);
1466 DOUT << std::string(Indent-2, ' ') << "=> ";
1467 DEBUG(Result.getNode()->dump(CurDAG));
1480 case ISD::UDIVREM: {
1481 SDValue N0 = Node->getOperand(0);
1482 SDValue N1 = Node->getOperand(1);
1484 bool isSigned = Opcode == ISD::SDIVREM;
1486 switch (NVT.getSimpleVT()) {
1487 default: assert(0 && "Unsupported VT!");
1488 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1489 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1490 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1491 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1494 switch (NVT.getSimpleVT()) {
1495 default: assert(0 && "Unsupported VT!");
1496 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1497 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1498 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1499 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1502 unsigned LoReg, HiReg;
1503 unsigned ClrOpcode, SExtOpcode;
1504 switch (NVT.getSimpleVT()) {
1505 default: assert(0 && "Unsupported VT!");
1507 LoReg = X86::AL; HiReg = X86::AH;
1509 SExtOpcode = X86::CBW;
1512 LoReg = X86::AX; HiReg = X86::DX;
1513 ClrOpcode = X86::MOV16r0;
1514 SExtOpcode = X86::CWD;
1517 LoReg = X86::EAX; HiReg = X86::EDX;
1518 ClrOpcode = X86::MOV32r0;
1519 SExtOpcode = X86::CDQ;
1522 LoReg = X86::RAX; HiReg = X86::RDX;
1523 ClrOpcode = X86::MOV64r0;
1524 SExtOpcode = X86::CQO;
1528 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1529 bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1530 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1533 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1534 // Special case for div8, just use a move with zero extension to AX to
1535 // clear the upper 8 bits (AH).
1536 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1537 if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1538 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1540 SDValue(CurDAG->getTargetNode(X86::MOVZX16rm8, dl, MVT::i16,
1542 array_lengthof(Ops)), 0);
1543 Chain = Move.getValue(1);
1544 ReplaceUses(N0.getValue(1), Chain);
1547 SDValue(CurDAG->getTargetNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1548 Chain = CurDAG->getEntryNode();
1550 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1551 InFlag = Chain.getValue(1);
1554 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1555 LoReg, N0, SDValue()).getValue(1);
1556 if (isSigned && !signBitIsZero) {
1557 // Sign extend the low part into the high part.
1559 SDValue(CurDAG->getTargetNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1561 // Zero out the high part, effectively zero extending the input.
1562 SDValue ClrNode = SDValue(CurDAG->getTargetNode(ClrOpcode, dl, NVT),
1564 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, HiReg,
1565 ClrNode, InFlag).getValue(1);
1570 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1573 CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1574 array_lengthof(Ops));
1575 InFlag = SDValue(CNode, 1);
1576 // Update the chain.
1577 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1580 SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1583 // Copy the division (low) result, if it is needed.
1584 if (!N.getValue(0).use_empty()) {
1585 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1586 LoReg, NVT, InFlag);
1587 InFlag = Result.getValue(2);
1588 ReplaceUses(N.getValue(0), Result);
1590 DOUT << std::string(Indent-2, ' ') << "=> ";
1591 DEBUG(Result.getNode()->dump(CurDAG));
1595 // Copy the remainder (high) result, if it is needed.
1596 if (!N.getValue(1).use_empty()) {
1598 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1599 // Prevent use of AH in a REX instruction by referencing AX instead.
1600 // Shift it down 8 bits.
1601 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1602 X86::AX, MVT::i16, InFlag);
1603 InFlag = Result.getValue(2);
1604 Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
1606 CurDAG->getTargetConstant(8, MVT::i8)),
1608 // Then truncate it down to i8.
1609 SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
1610 Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
1611 MVT::i8, Result, SRIdx), 0);
1613 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1614 HiReg, NVT, InFlag);
1615 InFlag = Result.getValue(2);
1617 ReplaceUses(N.getValue(1), Result);
1619 DOUT << std::string(Indent-2, ' ') << "=> ";
1620 DEBUG(Result.getNode()->dump(CurDAG));
1632 case ISD::DECLARE: {
1633 // Handle DECLARE nodes here because the second operand may have been
1634 // wrapped in X86ISD::Wrapper.
1635 SDValue Chain = Node->getOperand(0);
1636 SDValue N1 = Node->getOperand(1);
1637 SDValue N2 = Node->getOperand(2);
1638 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N1);
1640 // FIXME: We need to handle this for VLAs.
1642 ReplaceUses(N.getValue(0), Chain);
1646 if (N2.getOpcode() == ISD::ADD &&
1647 N2.getOperand(0).getOpcode() == X86ISD::GlobalBaseReg)
1648 N2 = N2.getOperand(1);
1650 // If N2 is not Wrapper(decriptor) then the llvm.declare is mangled
1651 // somehow, just ignore it.
1652 if (N2.getOpcode() != X86ISD::Wrapper) {
1653 ReplaceUses(N.getValue(0), Chain);
1656 GlobalAddressSDNode *GVNode =
1657 dyn_cast<GlobalAddressSDNode>(N2.getOperand(0));
1659 ReplaceUses(N.getValue(0), Chain);
1662 SDValue Tmp1 = CurDAG->getTargetFrameIndex(FINode->getIndex(),
1663 TLI.getPointerTy());
1664 SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GVNode->getGlobal(),
1665 TLI.getPointerTy());
1666 SDValue Ops[] = { Tmp1, Tmp2, Chain };
1667 return CurDAG->getTargetNode(TargetInstrInfo::DECLARE, dl,
1669 array_lengthof(Ops));
1673 SDNode *ResNode = SelectCode(N);
1676 DOUT << std::string(Indent-2, ' ') << "=> ";
1677 if (ResNode == NULL || ResNode == N.getNode())
1678 DEBUG(N.getNode()->dump(CurDAG));
1680 DEBUG(ResNode->dump(CurDAG));
1688 bool X86DAGToDAGISel::
1689 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1690 std::vector<SDValue> &OutOps) {
1691 SDValue Op0, Op1, Op2, Op3, Op4;
1692 switch (ConstraintCode) {
1693 case 'o': // offsetable ??
1694 case 'v': // not offsetable ??
1695 default: return true;
1697 if (!SelectAddr(Op, Op, Op0, Op1, Op2, Op3, Op4))
1702 OutOps.push_back(Op0);
1703 OutOps.push_back(Op1);
1704 OutOps.push_back(Op2);
1705 OutOps.push_back(Op3);
1706 OutOps.push_back(Op4);
1710 /// createX86ISelDag - This pass converts a legalized DAG into a
1711 /// X86-specific DAG, ready for instruction scheduling.
1713 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
1714 llvm::CodeGenOpt::Level OptLevel) {
1715 return new X86DAGToDAGISel(TM, OptLevel);