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 // Force NDEBUG on in any optimized build on Darwin.
17 // FIXME: This is a huge hack, to work around ridiculously awful compile times
18 // on this file with gcc-4.2 on Darwin, in Release mode.
19 #if (!defined(__llvm__) && defined(__APPLE__) && \
20 defined(__OPTIMIZE__) && !defined(NDEBUG))
24 #define DEBUG_TYPE "x86-isel"
26 #include "X86InstrBuilder.h"
27 #include "X86ISelLowering.h"
28 #include "X86MachineFunctionInfo.h"
29 #include "X86RegisterInfo.h"
30 #include "X86Subtarget.h"
31 #include "X86TargetMachine.h"
32 #include "llvm/GlobalValue.h"
33 #include "llvm/Instructions.h"
34 #include "llvm/Intrinsics.h"
35 #include "llvm/Support/CFG.h"
36 #include "llvm/Type.h"
37 #include "llvm/CodeGen/MachineConstantPool.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineFrameInfo.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAGISel.h"
43 #include "llvm/Target/TargetMachine.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/ADT/SmallPtrSet.h"
50 #include "llvm/ADT/Statistic.h"
53 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
55 //===----------------------------------------------------------------------===//
56 // Pattern Matcher Implementation
57 //===----------------------------------------------------------------------===//
60 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
61 /// SDValue's instead of register numbers for the leaves of the matched
63 struct X86ISelAddressMode {
69 struct { // This is really a union, discriminated by BaseType!
80 BlockAddress *BlockAddr;
83 unsigned Align; // CP alignment.
84 unsigned char SymbolFlags; // X86II::MO_*
87 : BaseType(RegBase), Scale(1), IndexReg(), Disp(0),
88 Segment(), GV(0), CP(0), BlockAddr(0), ES(0), JT(-1), Align(0),
89 SymbolFlags(X86II::MO_NO_FLAG) {
92 bool hasSymbolicDisplacement() const {
93 return GV != 0 || CP != 0 || ES != 0 || JT != -1 || BlockAddr != 0;
96 bool hasBaseOrIndexReg() const {
97 return IndexReg.getNode() != 0 || Base.Reg.getNode() != 0;
100 /// isRIPRelative - Return true if this addressing mode is already RIP
102 bool isRIPRelative() const {
103 if (BaseType != RegBase) return false;
104 if (RegisterSDNode *RegNode =
105 dyn_cast_or_null<RegisterSDNode>(Base.Reg.getNode()))
106 return RegNode->getReg() == X86::RIP;
110 void setBaseReg(SDValue Reg) {
116 dbgs() << "X86ISelAddressMode " << this << '\n';
117 dbgs() << "Base.Reg ";
118 if (Base.Reg.getNode() != 0)
119 Base.Reg.getNode()->dump();
122 dbgs() << " Base.FrameIndex " << Base.FrameIndex << '\n'
123 << " Scale" << Scale << '\n'
125 if (IndexReg.getNode() != 0)
126 IndexReg.getNode()->dump();
129 dbgs() << " Disp " << Disp << '\n'
146 dbgs() << " JT" << JT << " Align" << Align << '\n';
152 //===--------------------------------------------------------------------===//
153 /// ISel - X86 specific code to select X86 machine instructions for
154 /// SelectionDAG operations.
156 class X86DAGToDAGISel : public SelectionDAGISel {
157 /// X86Lowering - This object fully describes how to lower LLVM code to an
158 /// X86-specific SelectionDAG.
159 X86TargetLowering &X86Lowering;
161 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
162 /// make the right decision when generating code for different targets.
163 const X86Subtarget *Subtarget;
165 /// OptForSize - If true, selector should try to optimize for code size
166 /// instead of performance.
170 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
171 : SelectionDAGISel(tm, OptLevel),
172 X86Lowering(*tm.getTargetLowering()),
173 Subtarget(&tm.getSubtarget<X86Subtarget>()),
176 virtual const char *getPassName() const {
177 return "X86 DAG->DAG Instruction Selection";
180 /// InstructionSelect - This callback is invoked by
181 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
182 virtual void InstructionSelect();
184 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
186 virtual bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const;
188 // Include the pieces autogenerated from the target description.
189 #include "X86GenDAGISel.inc"
192 SDNode *Select(SDNode *N);
193 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
194 SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
196 bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
197 bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
198 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
199 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
200 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
202 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
203 bool SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
204 SDValue &Scale, SDValue &Index, SDValue &Disp,
206 bool SelectLEAAddr(SDNode *Op, SDValue N, SDValue &Base,
207 SDValue &Scale, SDValue &Index, SDValue &Disp);
208 bool SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
209 SDValue &Scale, SDValue &Index, SDValue &Disp);
210 bool SelectScalarSSELoadXXX(SDNode *Root, SDValue N,
211 SDValue &Base, SDValue &Scale,
212 SDValue &Index, SDValue &Disp,
214 SDValue &NodeWithChain);
216 // FIXME: Remove this hacky wrapper.
217 bool SelectScalarSSELoad(SDNode *Root, SDValue N, SDValue &Base,
218 SDValue &Scale, SDValue &Index,
219 SDValue &Disp, SDValue &Segment,
220 SDValue &PatternChainResult,
221 SDValue &PatternInputChain) {
223 if (!SelectScalarSSELoadXXX(Root, N, Base, Scale, Index, Disp, Segment,
226 PatternInputChain = Tmp.getOperand(0);
227 PatternChainResult = Tmp.getValue(1);
230 bool TryFoldLoad(SDNode *P, SDValue N,
231 SDValue &Base, SDValue &Scale,
232 SDValue &Index, SDValue &Disp,
234 void PreprocessForRMW();
235 void PreprocessForFPConvert();
237 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
238 /// inline asm expressions.
239 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
241 std::vector<SDValue> &OutOps);
243 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
245 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
246 SDValue &Scale, SDValue &Index,
247 SDValue &Disp, SDValue &Segment) {
248 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
249 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
251 Scale = getI8Imm(AM.Scale);
253 // These are 32-bit even in 64-bit mode since RIP relative offset
256 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp,
259 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
260 AM.Align, AM.Disp, AM.SymbolFlags);
262 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
263 else if (AM.JT != -1)
264 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
265 else if (AM.BlockAddr)
266 Disp = CurDAG->getBlockAddress(AM.BlockAddr, MVT::i32,
267 true, AM.SymbolFlags);
269 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
271 if (AM.Segment.getNode())
272 Segment = AM.Segment;
274 Segment = CurDAG->getRegister(0, MVT::i32);
277 /// getI8Imm - Return a target constant with the specified value, of type
279 inline SDValue getI8Imm(unsigned Imm) {
280 return CurDAG->getTargetConstant(Imm, MVT::i8);
283 /// getI16Imm - Return a target constant with the specified value, of type
285 inline SDValue getI16Imm(unsigned Imm) {
286 return CurDAG->getTargetConstant(Imm, MVT::i16);
289 /// getI32Imm - Return a target constant with the specified value, of type
291 inline SDValue getI32Imm(unsigned Imm) {
292 return CurDAG->getTargetConstant(Imm, MVT::i32);
295 /// getGlobalBaseReg - Return an SDNode that returns the value of
296 /// the global base register. Output instructions required to
297 /// initialize the global base register, if necessary.
299 SDNode *getGlobalBaseReg();
301 /// getTargetMachine - Return a reference to the TargetMachine, casted
302 /// to the target-specific type.
303 const X86TargetMachine &getTargetMachine() {
304 return static_cast<const X86TargetMachine &>(TM);
307 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
308 /// to the target-specific type.
309 const X86InstrInfo *getInstrInfo() {
310 return getTargetMachine().getInstrInfo();
321 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
322 if (OptLevel == CodeGenOpt::None) return false;
327 if (N.getOpcode() != ISD::LOAD)
330 // If N is a load, do additional profitability checks.
332 switch (U->getOpcode()) {
345 SDValue Op1 = U->getOperand(1);
347 // If the other operand is a 8-bit immediate we should fold the immediate
348 // instead. This reduces code size.
350 // movl 4(%esp), %eax
354 // addl 4(%esp), %eax
355 // The former is 2 bytes shorter. In case where the increment is 1, then
356 // the saving can be 4 bytes (by using incl %eax).
357 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
358 if (Imm->getAPIntValue().isSignedIntN(8))
361 // If the other operand is a TLS address, we should fold it instead.
364 // leal i@NTPOFF(%eax), %eax
366 // movl $i@NTPOFF, %eax
368 // if the block also has an access to a second TLS address this will save
370 // FIXME: This is probably also true for non TLS addresses.
371 if (Op1.getOpcode() == X86ISD::Wrapper) {
372 SDValue Val = Op1.getOperand(0);
373 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
383 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
384 /// and move load below the TokenFactor. Replace store's chain operand with
385 /// load's chain result.
386 static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
387 SDValue Store, SDValue TF) {
388 SmallVector<SDValue, 4> Ops;
389 for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i)
390 if (Load.getNode() == TF.getOperand(i).getNode())
391 Ops.push_back(Load.getOperand(0));
393 Ops.push_back(TF.getOperand(i));
394 SDValue NewTF = CurDAG->UpdateNodeOperands(TF, &Ops[0], Ops.size());
395 SDValue NewLoad = CurDAG->UpdateNodeOperands(Load, NewTF,
398 CurDAG->UpdateNodeOperands(Store, NewLoad.getValue(1), Store.getOperand(1),
399 Store.getOperand(2), Store.getOperand(3));
402 /// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG. The
403 /// chain produced by the load must only be used by the store's chain operand,
404 /// otherwise this may produce a cycle in the DAG.
406 static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
408 if (N.getOpcode() == ISD::BIT_CONVERT) {
414 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
415 if (!LD || LD->isVolatile())
417 if (LD->getAddressingMode() != ISD::UNINDEXED)
420 ISD::LoadExtType ExtType = LD->getExtensionType();
421 if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
425 LD->hasNUsesOfValue(1, 1) &&
426 N.getOperand(1) == Address &&
427 LD->isOperandOf(Chain.getNode())) {
434 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
435 /// operand and move load below the call's chain operand.
436 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
437 SDValue Call, SDValue CallSeqStart) {
438 SmallVector<SDValue, 8> Ops;
439 SDValue Chain = CallSeqStart.getOperand(0);
440 if (Chain.getNode() == Load.getNode())
441 Ops.push_back(Load.getOperand(0));
443 assert(Chain.getOpcode() == ISD::TokenFactor &&
444 "Unexpected CallSeqStart chain operand");
445 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
446 if (Chain.getOperand(i).getNode() == Load.getNode())
447 Ops.push_back(Load.getOperand(0));
449 Ops.push_back(Chain.getOperand(i));
451 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
452 MVT::Other, &Ops[0], Ops.size());
454 Ops.push_back(NewChain);
456 for (unsigned i = 1, e = CallSeqStart.getNumOperands(); i != e; ++i)
457 Ops.push_back(CallSeqStart.getOperand(i));
458 CurDAG->UpdateNodeOperands(CallSeqStart, &Ops[0], Ops.size());
459 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
460 Load.getOperand(1), Load.getOperand(2));
462 Ops.push_back(SDValue(Load.getNode(), 1));
463 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
464 Ops.push_back(Call.getOperand(i));
465 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
468 /// isCalleeLoad - Return true if call address is a load and it can be
469 /// moved below CALLSEQ_START and the chains leading up to the call.
470 /// Return the CALLSEQ_START by reference as a second output.
471 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
472 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
474 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
477 LD->getAddressingMode() != ISD::UNINDEXED ||
478 LD->getExtensionType() != ISD::NON_EXTLOAD)
481 // Now let's find the callseq_start.
482 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
483 if (!Chain.hasOneUse())
485 Chain = Chain.getOperand(0);
488 if (Chain.getOperand(0).getNode() == Callee.getNode())
490 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
491 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
492 Callee.getValue(1).hasOneUse())
498 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
499 /// This is only run if not in -O0 mode.
500 /// This allows the instruction selector to pick more read-modify-write
501 /// instructions. This is a common case:
511 /// [TokenFactor] [Op]
518 /// The fact the store's chain operand != load's chain will prevent the
519 /// (store (op (load))) instruction from being selected. We can transform it to:
538 void X86DAGToDAGISel::PreprocessForRMW() {
539 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
540 E = CurDAG->allnodes_end(); I != E; ++I) {
541 if (I->getOpcode() == X86ISD::CALL) {
542 /// Also try moving call address load from outside callseq_start to just
543 /// before the call to allow it to be folded.
561 SDValue Chain = I->getOperand(0);
562 SDValue Load = I->getOperand(1);
563 if (!isCalleeLoad(Load, Chain))
565 MoveBelowCallSeqStart(CurDAG, Load, SDValue(I, 0), Chain);
570 if (!ISD::isNON_TRUNCStore(I))
572 SDValue Chain = I->getOperand(0);
574 if (Chain.getNode()->getOpcode() != ISD::TokenFactor)
577 SDValue N1 = I->getOperand(1);
578 SDValue N2 = I->getOperand(2);
579 if ((N1.getValueType().isFloatingPoint() &&
580 !N1.getValueType().isVector()) ||
586 unsigned Opcode = N1.getNode()->getOpcode();
595 case ISD::VECTOR_SHUFFLE: {
596 SDValue N10 = N1.getOperand(0);
597 SDValue N11 = N1.getOperand(1);
598 RModW = isRMWLoad(N10, Chain, N2, Load);
600 RModW = isRMWLoad(N11, Chain, N2, Load);
613 SDValue N10 = N1.getOperand(0);
614 RModW = isRMWLoad(N10, Chain, N2, Load);
620 MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
628 /// PreprocessForFPConvert - Walk over the dag lowering fpround and fpextend
629 /// nodes that target the FP stack to be store and load to the stack. This is a
630 /// gross hack. We would like to simply mark these as being illegal, but when
631 /// we do that, legalize produces these when it expands calls, then expands
632 /// these in the same legalize pass. We would like dag combine to be able to
633 /// hack on these between the call expansion and the node legalization. As such
634 /// this pass basically does "really late" legalization of these inline with the
636 void X86DAGToDAGISel::PreprocessForFPConvert() {
637 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
638 E = CurDAG->allnodes_end(); I != E; ) {
639 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
640 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
643 // If the source and destination are SSE registers, then this is a legal
644 // conversion that should not be lowered.
645 EVT SrcVT = N->getOperand(0).getValueType();
646 EVT DstVT = N->getValueType(0);
647 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
648 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
649 if (SrcIsSSE && DstIsSSE)
652 if (!SrcIsSSE && !DstIsSSE) {
653 // If this is an FPStack extension, it is a noop.
654 if (N->getOpcode() == ISD::FP_EXTEND)
656 // If this is a value-preserving FPStack truncation, it is a noop.
657 if (N->getConstantOperandVal(1))
661 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
662 // FPStack has extload and truncstore. SSE can fold direct loads into other
663 // operations. Based on this, decide what we want to do.
665 if (N->getOpcode() == ISD::FP_ROUND)
666 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
668 MemVT = SrcIsSSE ? SrcVT : DstVT;
670 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
671 DebugLoc dl = N->getDebugLoc();
673 // FIXME: optimize the case where the src/dest is a load or store?
674 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
676 MemTmp, NULL, 0, MemVT,
678 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
679 NULL, 0, MemVT, false, false, 0);
681 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
682 // extload we created. This will cause general havok on the dag because
683 // anything below the conversion could be folded into other existing nodes.
684 // To avoid invalidating 'I', back it up to the convert node.
686 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
688 // Now that we did that, the node is dead. Increment the iterator to the
689 // next node to process, then delete N.
691 CurDAG->DeleteNode(N);
695 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
696 /// when it has created a SelectionDAG for us to codegen.
697 void X86DAGToDAGISel::InstructionSelect() {
698 const Function *F = MF->getFunction();
699 OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
701 if (OptLevel != CodeGenOpt::None)
704 // FIXME: This should only happen when not compiled with -O0.
705 PreprocessForFPConvert();
707 // Codegen the basic block.
709 DEBUG(dbgs() << "===== Instruction selection begins:\n");
714 DEBUG(dbgs() << "===== Instruction selection ends:\n");
717 CurDAG->RemoveDeadNodes();
720 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
721 /// the main function.
722 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
723 MachineFrameInfo *MFI) {
724 const TargetInstrInfo *TII = TM.getInstrInfo();
725 if (Subtarget->isTargetCygMing())
726 BuildMI(BB, DebugLoc::getUnknownLoc(),
727 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
730 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
731 // If this is main, emit special code for main.
732 MachineBasicBlock *BB = MF.begin();
733 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
734 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
738 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
739 X86ISelAddressMode &AM) {
740 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
741 SDValue Segment = N.getOperand(0);
743 if (AM.Segment.getNode() == 0) {
744 AM.Segment = Segment;
751 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
752 // This optimization is valid because the GNU TLS model defines that
753 // gs:0 (or fs:0 on X86-64) contains its own address.
754 // For more information see http://people.redhat.com/drepper/tls.pdf
756 SDValue Address = N.getOperand(1);
757 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
758 !MatchSegmentBaseAddress (Address, AM))
764 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
765 /// into an addressing mode. These wrap things that will resolve down into a
766 /// symbol reference. If no match is possible, this returns true, otherwise it
768 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
769 // If the addressing mode already has a symbol as the displacement, we can
770 // never match another symbol.
771 if (AM.hasSymbolicDisplacement())
774 SDValue N0 = N.getOperand(0);
775 CodeModel::Model M = TM.getCodeModel();
777 // Handle X86-64 rip-relative addresses. We check this before checking direct
778 // folding because RIP is preferable to non-RIP accesses.
779 if (Subtarget->is64Bit() &&
780 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
781 // they cannot be folded into immediate fields.
782 // FIXME: This can be improved for kernel and other models?
783 (M == CodeModel::Small || M == CodeModel::Kernel) &&
784 // Base and index reg must be 0 in order to use %rip as base and lowering
786 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
787 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
788 int64_t Offset = AM.Disp + G->getOffset();
789 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
790 AM.GV = G->getGlobal();
792 AM.SymbolFlags = G->getTargetFlags();
793 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
794 int64_t Offset = AM.Disp + CP->getOffset();
795 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
796 AM.CP = CP->getConstVal();
797 AM.Align = CP->getAlignment();
799 AM.SymbolFlags = CP->getTargetFlags();
800 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
801 AM.ES = S->getSymbol();
802 AM.SymbolFlags = S->getTargetFlags();
803 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
804 AM.JT = J->getIndex();
805 AM.SymbolFlags = J->getTargetFlags();
807 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
808 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
811 if (N.getOpcode() == X86ISD::WrapperRIP)
812 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
816 // Handle the case when globals fit in our immediate field: This is true for
817 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
818 // mode, this results in a non-RIP-relative computation.
819 if (!Subtarget->is64Bit() ||
820 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
821 TM.getRelocationModel() == Reloc::Static)) {
822 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
823 AM.GV = G->getGlobal();
824 AM.Disp += G->getOffset();
825 AM.SymbolFlags = G->getTargetFlags();
826 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
827 AM.CP = CP->getConstVal();
828 AM.Align = CP->getAlignment();
829 AM.Disp += CP->getOffset();
830 AM.SymbolFlags = CP->getTargetFlags();
831 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
832 AM.ES = S->getSymbol();
833 AM.SymbolFlags = S->getTargetFlags();
834 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
835 AM.JT = J->getIndex();
836 AM.SymbolFlags = J->getTargetFlags();
838 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
839 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
847 /// MatchAddress - Add the specified node to the specified addressing mode,
848 /// returning true if it cannot be done. This just pattern matches for the
850 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
851 if (MatchAddressRecursively(N, AM, 0))
854 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
855 // a smaller encoding and avoids a scaled-index.
857 AM.BaseType == X86ISelAddressMode::RegBase &&
858 AM.Base.Reg.getNode() == 0) {
859 AM.Base.Reg = AM.IndexReg;
863 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
864 // because it has a smaller encoding.
865 // TODO: Which other code models can use this?
866 if (TM.getCodeModel() == CodeModel::Small &&
867 Subtarget->is64Bit() &&
869 AM.BaseType == X86ISelAddressMode::RegBase &&
870 AM.Base.Reg.getNode() == 0 &&
871 AM.IndexReg.getNode() == 0 &&
872 AM.SymbolFlags == X86II::MO_NO_FLAG &&
873 AM.hasSymbolicDisplacement())
874 AM.Base.Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
879 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
881 bool is64Bit = Subtarget->is64Bit();
882 DebugLoc dl = N.getDebugLoc();
884 dbgs() << "MatchAddress: ";
889 return MatchAddressBase(N, AM);
891 CodeModel::Model M = TM.getCodeModel();
893 // If this is already a %rip relative address, we can only merge immediates
894 // into it. Instead of handling this in every case, we handle it here.
895 // RIP relative addressing: %rip + 32-bit displacement!
896 if (AM.isRIPRelative()) {
897 // FIXME: JumpTable and ExternalSymbol address currently don't like
898 // displacements. It isn't very important, but this should be fixed for
900 if (!AM.ES && AM.JT != -1) return true;
902 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
903 int64_t Val = AM.Disp + Cst->getSExtValue();
904 if (X86::isOffsetSuitableForCodeModel(Val, M,
905 AM.hasSymbolicDisplacement())) {
913 switch (N.getOpcode()) {
915 case ISD::Constant: {
916 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
918 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
919 AM.hasSymbolicDisplacement())) {
926 case X86ISD::SegmentBaseAddress:
927 if (!MatchSegmentBaseAddress(N, AM))
931 case X86ISD::Wrapper:
932 case X86ISD::WrapperRIP:
933 if (!MatchWrapper(N, AM))
938 if (!MatchLoad(N, AM))
942 case ISD::FrameIndex:
943 if (AM.BaseType == X86ISelAddressMode::RegBase
944 && AM.Base.Reg.getNode() == 0) {
945 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
946 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
952 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
956 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
957 unsigned Val = CN->getZExtValue();
958 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
959 // that the base operand remains free for further matching. If
960 // the base doesn't end up getting used, a post-processing step
961 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
962 if (Val == 1 || Val == 2 || Val == 3) {
964 SDValue ShVal = N.getNode()->getOperand(0);
966 // Okay, we know that we have a scale by now. However, if the scaled
967 // value is an add of something and a constant, we can fold the
968 // constant into the disp field here.
969 if (ShVal.getNode()->getOpcode() == ISD::ADD &&
970 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
971 AM.IndexReg = ShVal.getNode()->getOperand(0);
972 ConstantSDNode *AddVal =
973 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
974 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
976 X86::isOffsetSuitableForCodeModel(Disp, M,
977 AM.hasSymbolicDisplacement()))
991 // A mul_lohi where we need the low part can be folded as a plain multiply.
992 if (N.getResNo() != 0) break;
995 case X86ISD::MUL_IMM:
996 // X*[3,5,9] -> X+X*[2,4,8]
997 if (AM.BaseType == X86ISelAddressMode::RegBase &&
998 AM.Base.Reg.getNode() == 0 &&
999 AM.IndexReg.getNode() == 0) {
1001 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1002 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1003 CN->getZExtValue() == 9) {
1004 AM.Scale = unsigned(CN->getZExtValue())-1;
1006 SDValue MulVal = N.getNode()->getOperand(0);
1009 // Okay, we know that we have a scale by now. However, if the scaled
1010 // value is an add of something and a constant, we can fold the
1011 // constant into the disp field here.
1012 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1013 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1014 Reg = MulVal.getNode()->getOperand(0);
1015 ConstantSDNode *AddVal =
1016 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1017 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
1020 X86::isOffsetSuitableForCodeModel(Disp, M,
1021 AM.hasSymbolicDisplacement()))
1024 Reg = N.getNode()->getOperand(0);
1026 Reg = N.getNode()->getOperand(0);
1029 AM.IndexReg = AM.Base.Reg = Reg;
1036 // Given A-B, if A can be completely folded into the address and
1037 // the index field with the index field unused, use -B as the index.
1038 // This is a win if a has multiple parts that can be folded into
1039 // the address. Also, this saves a mov if the base register has
1040 // other uses, since it avoids a two-address sub instruction, however
1041 // it costs an additional mov if the index register has other uses.
1043 // Test if the LHS of the sub can be folded.
1044 X86ISelAddressMode Backup = AM;
1045 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1049 // Test if the index field is free for use.
1050 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1055 SDValue RHS = N.getNode()->getOperand(1);
1056 // If the RHS involves a register with multiple uses, this
1057 // transformation incurs an extra mov, due to the neg instruction
1058 // clobbering its operand.
1059 if (!RHS.getNode()->hasOneUse() ||
1060 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1061 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1062 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1063 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1064 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1066 // If the base is a register with multiple uses, this
1067 // transformation may save a mov.
1068 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1069 AM.Base.Reg.getNode() &&
1070 !AM.Base.Reg.getNode()->hasOneUse()) ||
1071 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1073 // If the folded LHS was interesting, this transformation saves
1074 // address arithmetic.
1075 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1076 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1077 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1079 // If it doesn't look like it may be an overall win, don't do it.
1085 // Ok, the transformation is legal and appears profitable. Go for it.
1086 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
1087 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1091 // Insert the new nodes into the topological ordering.
1092 if (Zero.getNode()->getNodeId() == -1 ||
1093 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1094 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
1095 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
1097 if (Neg.getNode()->getNodeId() == -1 ||
1098 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1099 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
1100 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
1106 X86ISelAddressMode Backup = AM;
1107 if (!MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1) &&
1108 !MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1))
1111 if (!MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1) &&
1112 !MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1))
1116 // If we couldn't fold both operands into the address at the same time,
1117 // see if we can just put each operand into a register and fold at least
1119 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1120 !AM.Base.Reg.getNode() &&
1121 !AM.IndexReg.getNode()) {
1122 AM.Base.Reg = N.getNode()->getOperand(0);
1123 AM.IndexReg = N.getNode()->getOperand(1);
1131 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1132 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1133 X86ISelAddressMode Backup = AM;
1134 uint64_t Offset = CN->getSExtValue();
1135 // Start with the LHS as an addr mode.
1136 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1137 // Address could not have picked a GV address for the displacement.
1139 // On x86-64, the resultant disp must fit in 32-bits.
1141 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
1142 AM.hasSymbolicDisplacement())) &&
1143 // Check to see if the LHS & C is zero.
1144 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
1153 // Perform some heroic transforms on an and of a constant-count shift
1154 // with a constant to enable use of the scaled offset field.
1156 SDValue Shift = N.getOperand(0);
1157 if (Shift.getNumOperands() != 2) break;
1159 // Scale must not be used already.
1160 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1162 SDValue X = Shift.getOperand(0);
1163 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1164 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1165 if (!C1 || !C2) break;
1167 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1168 // allows us to convert the shift and and into an h-register extract and
1170 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1171 unsigned ScaleLog = 8 - C1->getZExtValue();
1172 if (ScaleLog > 0 && ScaleLog < 4 &&
1173 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1174 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1175 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1176 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1178 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1180 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1181 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1184 // Insert the new nodes into the topological ordering.
1185 if (Eight.getNode()->getNodeId() == -1 ||
1186 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1187 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1188 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1190 if (Mask.getNode()->getNodeId() == -1 ||
1191 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1192 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1193 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1195 if (Srl.getNode()->getNodeId() == -1 ||
1196 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1197 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1198 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1200 if (And.getNode()->getNodeId() == -1 ||
1201 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1202 CurDAG->RepositionNode(N.getNode(), And.getNode());
1203 And.getNode()->setNodeId(N.getNode()->getNodeId());
1205 if (ShlCount.getNode()->getNodeId() == -1 ||
1206 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1207 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1208 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1210 if (Shl.getNode()->getNodeId() == -1 ||
1211 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1212 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1213 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1215 CurDAG->ReplaceAllUsesWith(N, Shl);
1217 AM.Scale = (1 << ScaleLog);
1222 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1223 // allows us to fold the shift into this addressing mode.
1224 if (Shift.getOpcode() != ISD::SHL) break;
1226 // Not likely to be profitable if either the AND or SHIFT node has more
1227 // than one use (unless all uses are for address computation). Besides,
1228 // isel mechanism requires their node ids to be reused.
1229 if (!N.hasOneUse() || !Shift.hasOneUse())
1232 // Verify that the shift amount is something we can fold.
1233 unsigned ShiftCst = C1->getZExtValue();
1234 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1237 // Get the new AND mask, this folds to a constant.
1238 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1239 SDValue(C2, 0), SDValue(C1, 0));
1240 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1242 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1243 NewAND, SDValue(C1, 0));
1245 // Insert the new nodes into the topological ordering.
1246 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1247 CurDAG->RepositionNode(X.getNode(), C1);
1248 C1->setNodeId(X.getNode()->getNodeId());
1250 if (NewANDMask.getNode()->getNodeId() == -1 ||
1251 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1252 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1253 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1255 if (NewAND.getNode()->getNodeId() == -1 ||
1256 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1257 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1258 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1260 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1261 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1262 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1263 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1266 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1268 AM.Scale = 1 << ShiftCst;
1269 AM.IndexReg = NewAND;
1274 return MatchAddressBase(N, AM);
1277 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1278 /// specified addressing mode without any further recursion.
1279 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1280 // Is the base register already occupied?
1281 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
1282 // If so, check to see if the scale index register is set.
1283 if (AM.IndexReg.getNode() == 0) {
1289 // Otherwise, we cannot select it.
1293 // Default, generate it as a register.
1294 AM.BaseType = X86ISelAddressMode::RegBase;
1299 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1300 /// It returns the operands which make up the maximal addressing mode it can
1301 /// match by reference.
1302 bool X86DAGToDAGISel::SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
1303 SDValue &Scale, SDValue &Index,
1304 SDValue &Disp, SDValue &Segment) {
1305 X86ISelAddressMode AM;
1306 if (MatchAddress(N, AM))
1309 EVT VT = N.getValueType();
1310 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1311 if (!AM.Base.Reg.getNode())
1312 AM.Base.Reg = CurDAG->getRegister(0, VT);
1315 if (!AM.IndexReg.getNode())
1316 AM.IndexReg = CurDAG->getRegister(0, VT);
1318 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1322 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1323 /// match a load whose top elements are either undef or zeros. The load flavor
1324 /// is derived from the type of N, which is either v4f32 or v2f64.
1327 /// PatternChainNode: this is the matched node that has a chain input and
1329 bool X86DAGToDAGISel::SelectScalarSSELoadXXX(SDNode *Root,
1330 SDValue N, SDValue &Base,
1331 SDValue &Scale, SDValue &Index,
1332 SDValue &Disp, SDValue &Segment,
1333 SDValue &PatternNodeWithChain) {
1334 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1335 PatternNodeWithChain = N.getOperand(0);
1336 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1337 PatternNodeWithChain.hasOneUse() &&
1338 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1339 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1340 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1341 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp,Segment))
1347 // Also handle the case where we explicitly require zeros in the top
1348 // elements. This is a vector shuffle from the zero vector.
1349 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1350 // Check to see if the top elements are all zeros (or bitcast of zeros).
1351 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1352 N.getOperand(0).getNode()->hasOneUse() &&
1353 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1354 N.getOperand(0).getOperand(0).hasOneUse() &&
1355 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1356 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1357 // Okay, this is a zero extending load. Fold it.
1358 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1359 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1361 PatternNodeWithChain = SDValue(LD, 0);
1368 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1369 /// mode it matches can be cost effectively emitted as an LEA instruction.
1370 bool X86DAGToDAGISel::SelectLEAAddr(SDNode *Op, SDValue N,
1371 SDValue &Base, SDValue &Scale,
1372 SDValue &Index, SDValue &Disp) {
1373 X86ISelAddressMode AM;
1375 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1377 SDValue Copy = AM.Segment;
1378 SDValue T = CurDAG->getRegister(0, MVT::i32);
1380 if (MatchAddress(N, AM))
1382 assert (T == AM.Segment);
1385 EVT VT = N.getValueType();
1386 unsigned Complexity = 0;
1387 if (AM.BaseType == X86ISelAddressMode::RegBase)
1388 if (AM.Base.Reg.getNode())
1391 AM.Base.Reg = CurDAG->getRegister(0, VT);
1392 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1395 if (AM.IndexReg.getNode())
1398 AM.IndexReg = CurDAG->getRegister(0, VT);
1400 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1405 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1406 // to a LEA. This is determined with some expermentation but is by no means
1407 // optimal (especially for code size consideration). LEA is nice because of
1408 // its three-address nature. Tweak the cost function again when we can run
1409 // convertToThreeAddress() at register allocation time.
1410 if (AM.hasSymbolicDisplacement()) {
1411 // For X86-64, we should always use lea to materialize RIP relative
1413 if (Subtarget->is64Bit())
1419 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1422 // If it isn't worth using an LEA, reject it.
1423 if (Complexity <= 2)
1427 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1431 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1432 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
1433 SDValue &Scale, SDValue &Index,
1435 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1436 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1438 X86ISelAddressMode AM;
1439 AM.GV = GA->getGlobal();
1440 AM.Disp += GA->getOffset();
1441 AM.Base.Reg = CurDAG->getRegister(0, N.getValueType());
1442 AM.SymbolFlags = GA->getTargetFlags();
1444 if (N.getValueType() == MVT::i32) {
1446 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1448 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1452 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1457 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1458 SDValue &Base, SDValue &Scale,
1459 SDValue &Index, SDValue &Disp,
1461 if (ISD::isNON_EXTLoad(N.getNode()) &&
1462 IsProfitableToFold(N, P, P) &&
1463 IsLegalToFold(N, P, P))
1464 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
1468 /// getGlobalBaseReg - Return an SDNode that returns the value of
1469 /// the global base register. Output instructions required to
1470 /// initialize the global base register, if necessary.
1472 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1473 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1474 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1477 static SDNode *FindCallStartFromCall(SDNode *Node) {
1478 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1479 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1480 "Node doesn't have a token chain argument!");
1481 return FindCallStartFromCall(Node->getOperand(0).getNode());
1484 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1485 SDValue Chain = Node->getOperand(0);
1486 SDValue In1 = Node->getOperand(1);
1487 SDValue In2L = Node->getOperand(2);
1488 SDValue In2H = Node->getOperand(3);
1489 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1490 if (!SelectAddr(In1.getNode(), In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1492 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1493 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1494 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1495 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1496 MVT::i32, MVT::i32, MVT::Other, Ops,
1497 array_lengthof(Ops));
1498 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1502 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1503 if (Node->hasAnyUseOfValue(0))
1506 // Optimize common patterns for __sync_add_and_fetch and
1507 // __sync_sub_and_fetch where the result is not used. This allows us
1508 // to use "lock" version of add, sub, inc, dec instructions.
1509 // FIXME: Do not use special instructions but instead add the "lock"
1510 // prefix to the target node somehow. The extra information will then be
1511 // transferred to machine instruction and it denotes the prefix.
1512 SDValue Chain = Node->getOperand(0);
1513 SDValue Ptr = Node->getOperand(1);
1514 SDValue Val = Node->getOperand(2);
1515 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1516 if (!SelectAddr(Ptr.getNode(), Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1519 bool isInc = false, isDec = false, isSub = false, isCN = false;
1520 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1523 int64_t CNVal = CN->getSExtValue();
1526 else if (CNVal == -1)
1528 else if (CNVal >= 0)
1529 Val = CurDAG->getTargetConstant(CNVal, NVT);
1532 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1534 } else if (Val.hasOneUse() &&
1535 Val.getOpcode() == ISD::SUB &&
1536 X86::isZeroNode(Val.getOperand(0))) {
1538 Val = Val.getOperand(1);
1542 switch (NVT.getSimpleVT().SimpleTy) {
1546 Opc = X86::LOCK_INC8m;
1548 Opc = X86::LOCK_DEC8m;
1551 Opc = X86::LOCK_SUB8mi;
1553 Opc = X86::LOCK_SUB8mr;
1556 Opc = X86::LOCK_ADD8mi;
1558 Opc = X86::LOCK_ADD8mr;
1563 Opc = X86::LOCK_INC16m;
1565 Opc = X86::LOCK_DEC16m;
1568 if (Predicate_i16immSExt8(Val.getNode()))
1569 Opc = X86::LOCK_SUB16mi8;
1571 Opc = X86::LOCK_SUB16mi;
1573 Opc = X86::LOCK_SUB16mr;
1576 if (Predicate_i16immSExt8(Val.getNode()))
1577 Opc = X86::LOCK_ADD16mi8;
1579 Opc = X86::LOCK_ADD16mi;
1581 Opc = X86::LOCK_ADD16mr;
1586 Opc = X86::LOCK_INC32m;
1588 Opc = X86::LOCK_DEC32m;
1591 if (Predicate_i32immSExt8(Val.getNode()))
1592 Opc = X86::LOCK_SUB32mi8;
1594 Opc = X86::LOCK_SUB32mi;
1596 Opc = X86::LOCK_SUB32mr;
1599 if (Predicate_i32immSExt8(Val.getNode()))
1600 Opc = X86::LOCK_ADD32mi8;
1602 Opc = X86::LOCK_ADD32mi;
1604 Opc = X86::LOCK_ADD32mr;
1609 Opc = X86::LOCK_INC64m;
1611 Opc = X86::LOCK_DEC64m;
1613 Opc = X86::LOCK_SUB64mr;
1615 if (Predicate_i64immSExt8(Val.getNode()))
1616 Opc = X86::LOCK_SUB64mi8;
1617 else if (Predicate_i64immSExt32(Val.getNode()))
1618 Opc = X86::LOCK_SUB64mi32;
1621 Opc = X86::LOCK_ADD64mr;
1623 if (Predicate_i64immSExt8(Val.getNode()))
1624 Opc = X86::LOCK_ADD64mi8;
1625 else if (Predicate_i64immSExt32(Val.getNode()))
1626 Opc = X86::LOCK_ADD64mi32;
1632 DebugLoc dl = Node->getDebugLoc();
1633 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1635 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1636 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1637 if (isInc || isDec) {
1638 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1639 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1640 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1641 SDValue RetVals[] = { Undef, Ret };
1642 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1644 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1645 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1646 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1647 SDValue RetVals[] = { Undef, Ret };
1648 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1652 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1653 /// any uses which require the SF or OF bits to be accurate.
1654 static bool HasNoSignedComparisonUses(SDNode *N) {
1655 // Examine each user of the node.
1656 for (SDNode::use_iterator UI = N->use_begin(),
1657 UE = N->use_end(); UI != UE; ++UI) {
1658 // Only examine CopyToReg uses.
1659 if (UI->getOpcode() != ISD::CopyToReg)
1661 // Only examine CopyToReg uses that copy to EFLAGS.
1662 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1665 // Examine each user of the CopyToReg use.
1666 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1667 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1668 // Only examine the Flag result.
1669 if (FlagUI.getUse().getResNo() != 1) continue;
1670 // Anything unusual: assume conservatively.
1671 if (!FlagUI->isMachineOpcode()) return false;
1672 // Examine the opcode of the user.
1673 switch (FlagUI->getMachineOpcode()) {
1674 // These comparisons don't treat the most significant bit specially.
1675 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1676 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1677 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1678 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1679 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1680 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1681 case X86::CMOVA16rr: case X86::CMOVA16rm:
1682 case X86::CMOVA32rr: case X86::CMOVA32rm:
1683 case X86::CMOVA64rr: case X86::CMOVA64rm:
1684 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1685 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1686 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1687 case X86::CMOVB16rr: case X86::CMOVB16rm:
1688 case X86::CMOVB32rr: case X86::CMOVB32rm:
1689 case X86::CMOVB64rr: case X86::CMOVB64rm:
1690 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1691 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1692 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1693 case X86::CMOVE16rr: case X86::CMOVE16rm:
1694 case X86::CMOVE32rr: case X86::CMOVE32rm:
1695 case X86::CMOVE64rr: case X86::CMOVE64rm:
1696 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1697 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1698 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1699 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1700 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1701 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1702 case X86::CMOVP16rr: case X86::CMOVP16rm:
1703 case X86::CMOVP32rr: case X86::CMOVP32rm:
1704 case X86::CMOVP64rr: case X86::CMOVP64rm:
1706 // Anything else: assume conservatively.
1707 default: return false;
1714 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1715 EVT NVT = Node->getValueType(0);
1717 unsigned Opcode = Node->getOpcode();
1718 DebugLoc dl = Node->getDebugLoc();
1722 dbgs() << std::string(Indent, ' ') << "Selecting: ";
1729 if (Node->isMachineOpcode()) {
1732 dbgs() << std::string(Indent-2, ' ') << "== ";
1738 return NULL; // Already selected.
1743 case X86ISD::GlobalBaseReg:
1744 return getGlobalBaseReg();
1746 case X86ISD::ATOMOR64_DAG:
1747 return SelectAtomic64(Node, X86::ATOMOR6432);
1748 case X86ISD::ATOMXOR64_DAG:
1749 return SelectAtomic64(Node, X86::ATOMXOR6432);
1750 case X86ISD::ATOMADD64_DAG:
1751 return SelectAtomic64(Node, X86::ATOMADD6432);
1752 case X86ISD::ATOMSUB64_DAG:
1753 return SelectAtomic64(Node, X86::ATOMSUB6432);
1754 case X86ISD::ATOMNAND64_DAG:
1755 return SelectAtomic64(Node, X86::ATOMNAND6432);
1756 case X86ISD::ATOMAND64_DAG:
1757 return SelectAtomic64(Node, X86::ATOMAND6432);
1758 case X86ISD::ATOMSWAP64_DAG:
1759 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1761 case ISD::ATOMIC_LOAD_ADD: {
1762 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1768 case ISD::SMUL_LOHI:
1769 case ISD::UMUL_LOHI: {
1770 SDValue N0 = Node->getOperand(0);
1771 SDValue N1 = Node->getOperand(1);
1773 bool isSigned = Opcode == ISD::SMUL_LOHI;
1775 switch (NVT.getSimpleVT().SimpleTy) {
1776 default: llvm_unreachable("Unsupported VT!");
1777 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1778 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1779 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1780 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1783 switch (NVT.getSimpleVT().SimpleTy) {
1784 default: llvm_unreachable("Unsupported VT!");
1785 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1786 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1787 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1788 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1792 unsigned LoReg, HiReg;
1793 switch (NVT.getSimpleVT().SimpleTy) {
1794 default: llvm_unreachable("Unsupported VT!");
1795 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1796 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1797 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1798 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1801 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1802 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1803 // Multiply is commmutative.
1805 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1810 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1811 N0, SDValue()).getValue(1);
1814 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1817 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1818 array_lengthof(Ops));
1819 InFlag = SDValue(CNode, 1);
1820 // Update the chain.
1821 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1824 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1827 // Copy the low half of the result, if it is needed.
1828 if (!SDValue(Node, 0).use_empty()) {
1829 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1830 LoReg, NVT, InFlag);
1831 InFlag = Result.getValue(2);
1832 ReplaceUses(SDValue(Node, 0), Result);
1835 dbgs() << std::string(Indent-2, ' ') << "=> ";
1836 Result.getNode()->dump(CurDAG);
1841 // Copy the high half of the result, if it is needed.
1842 if (!SDValue(Node, 1).use_empty()) {
1844 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1845 // Prevent use of AH in a REX instruction by referencing AX instead.
1846 // Shift it down 8 bits.
1847 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1848 X86::AX, MVT::i16, InFlag);
1849 InFlag = Result.getValue(2);
1850 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1852 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1853 // Then truncate it down to i8.
1854 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
1857 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1858 HiReg, NVT, InFlag);
1859 InFlag = Result.getValue(2);
1861 ReplaceUses(SDValue(Node, 1), Result);
1864 dbgs() << std::string(Indent-2, ' ') << "=> ";
1865 Result.getNode()->dump(CurDAG);
1879 case ISD::UDIVREM: {
1880 SDValue N0 = Node->getOperand(0);
1881 SDValue N1 = Node->getOperand(1);
1883 bool isSigned = Opcode == ISD::SDIVREM;
1885 switch (NVT.getSimpleVT().SimpleTy) {
1886 default: llvm_unreachable("Unsupported VT!");
1887 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1888 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1889 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1890 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1893 switch (NVT.getSimpleVT().SimpleTy) {
1894 default: llvm_unreachable("Unsupported VT!");
1895 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1896 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1897 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1898 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1902 unsigned LoReg, HiReg, ClrReg;
1903 unsigned ClrOpcode, SExtOpcode;
1904 switch (NVT.getSimpleVT().SimpleTy) {
1905 default: llvm_unreachable("Unsupported VT!");
1907 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1909 SExtOpcode = X86::CBW;
1912 LoReg = X86::AX; HiReg = X86::DX;
1913 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1914 SExtOpcode = X86::CWD;
1917 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1918 ClrOpcode = X86::MOV32r0;
1919 SExtOpcode = X86::CDQ;
1922 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1923 ClrOpcode = X86::MOV64r0;
1924 SExtOpcode = X86::CQO;
1928 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1929 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1930 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1933 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1934 // Special case for div8, just use a move with zero extension to AX to
1935 // clear the upper 8 bits (AH).
1936 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1937 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1938 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1940 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1942 array_lengthof(Ops)), 0);
1943 Chain = Move.getValue(1);
1944 ReplaceUses(N0.getValue(1), Chain);
1947 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1948 Chain = CurDAG->getEntryNode();
1950 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1951 InFlag = Chain.getValue(1);
1954 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1955 LoReg, N0, SDValue()).getValue(1);
1956 if (isSigned && !signBitIsZero) {
1957 // Sign extend the low part into the high part.
1959 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1961 // Zero out the high part, effectively zero extending the input.
1963 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1964 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1965 ClrNode, InFlag).getValue(1);
1970 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1973 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1974 array_lengthof(Ops));
1975 InFlag = SDValue(CNode, 1);
1976 // Update the chain.
1977 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1980 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1983 // Copy the division (low) result, if it is needed.
1984 if (!SDValue(Node, 0).use_empty()) {
1985 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1986 LoReg, NVT, InFlag);
1987 InFlag = Result.getValue(2);
1988 ReplaceUses(SDValue(Node, 0), Result);
1991 dbgs() << std::string(Indent-2, ' ') << "=> ";
1992 Result.getNode()->dump(CurDAG);
1997 // Copy the remainder (high) result, if it is needed.
1998 if (!SDValue(Node, 1).use_empty()) {
2000 if (HiReg == X86::AH && Subtarget->is64Bit()) {
2001 // Prevent use of AH in a REX instruction by referencing AX instead.
2002 // Shift it down 8 bits.
2003 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2004 X86::AX, MVT::i16, InFlag);
2005 InFlag = Result.getValue(2);
2006 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2008 CurDAG->getTargetConstant(8, MVT::i8)),
2010 // Then truncate it down to i8.
2011 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
2014 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2015 HiReg, NVT, InFlag);
2016 InFlag = Result.getValue(2);
2018 ReplaceUses(SDValue(Node, 1), Result);
2021 dbgs() << std::string(Indent-2, ' ') << "=> ";
2022 Result.getNode()->dump(CurDAG);
2036 SDValue N0 = Node->getOperand(0);
2037 SDValue N1 = Node->getOperand(1);
2039 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2040 // use a smaller encoding.
2041 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
2042 N0.getValueType() != MVT::i8 &&
2043 X86::isZeroNode(N1)) {
2044 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2047 // For example, convert "testl %eax, $8" to "testb %al, $8"
2048 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2049 (!(C->getZExtValue() & 0x80) ||
2050 HasNoSignedComparisonUses(Node))) {
2051 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
2052 SDValue Reg = N0.getNode()->getOperand(0);
2054 // On x86-32, only the ABCD registers have 8-bit subregisters.
2055 if (!Subtarget->is64Bit()) {
2056 TargetRegisterClass *TRC = 0;
2057 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2058 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2059 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2060 default: llvm_unreachable("Unsupported TEST operand type!");
2062 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2063 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2064 Reg.getValueType(), Reg, RC), 0);
2067 // Extract the l-register.
2068 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
2072 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
2075 // For example, "testl %eax, $2048" to "testb %ah, $8".
2076 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2077 (!(C->getZExtValue() & 0x8000) ||
2078 HasNoSignedComparisonUses(Node))) {
2079 // Shift the immediate right by 8 bits.
2080 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2082 SDValue Reg = N0.getNode()->getOperand(0);
2084 // Put the value in an ABCD register.
2085 TargetRegisterClass *TRC = 0;
2086 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2087 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2088 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2089 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2090 default: llvm_unreachable("Unsupported TEST operand type!");
2092 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2093 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2094 Reg.getValueType(), Reg, RC), 0);
2096 // Extract the h-register.
2097 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT_HI, dl,
2100 // Emit a testb. No special NOREX tricks are needed since there's
2101 // only one GPR operand!
2102 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2103 Subreg, ShiftedImm);
2106 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2107 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2108 N0.getValueType() != MVT::i16 &&
2109 (!(C->getZExtValue() & 0x8000) ||
2110 HasNoSignedComparisonUses(Node))) {
2111 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
2112 SDValue Reg = N0.getNode()->getOperand(0);
2114 // Extract the 16-bit subregister.
2115 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_16BIT, dl,
2119 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
2122 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2123 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2124 N0.getValueType() == MVT::i64 &&
2125 (!(C->getZExtValue() & 0x80000000) ||
2126 HasNoSignedComparisonUses(Node))) {
2127 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
2128 SDValue Reg = N0.getNode()->getOperand(0);
2130 // Extract the 32-bit subregister.
2131 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_32BIT, dl,
2135 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
2142 SDNode *ResNode = SelectCode(Node);
2146 dbgs() << std::string(Indent-2, ' ') << "=> ";
2147 if (ResNode == NULL || ResNode == Node)
2150 ResNode->dump(CurDAG);
2159 bool X86DAGToDAGISel::
2160 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2161 std::vector<SDValue> &OutOps) {
2162 SDValue Op0, Op1, Op2, Op3, Op4;
2163 switch (ConstraintCode) {
2164 case 'o': // offsetable ??
2165 case 'v': // not offsetable ??
2166 default: return true;
2168 if (!SelectAddr(Op.getNode(), Op, Op0, Op1, Op2, Op3, Op4))
2173 OutOps.push_back(Op0);
2174 OutOps.push_back(Op1);
2175 OutOps.push_back(Op2);
2176 OutOps.push_back(Op3);
2177 OutOps.push_back(Op4);
2181 /// createX86ISelDag - This pass converts a legalized DAG into a
2182 /// X86-specific DAG, ready for instruction scheduling.
2184 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2185 llvm::CodeGenOpt::Level OptLevel) {
2186 return new X86DAGToDAGISel(TM, OptLevel);