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 virtual bool IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const;
190 // Include the pieces autogenerated from the target description.
191 #include "X86GenDAGISel.inc"
194 SDNode *Select(SDNode *N);
195 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
196 SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
198 bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
199 bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
200 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
201 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
202 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
204 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
205 bool SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
206 SDValue &Scale, SDValue &Index, SDValue &Disp,
208 bool SelectLEAAddr(SDNode *Op, SDValue N, SDValue &Base,
209 SDValue &Scale, SDValue &Index, SDValue &Disp);
210 bool SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
211 SDValue &Scale, SDValue &Index, SDValue &Disp);
212 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
213 SDValue &Base, SDValue &Scale,
214 SDValue &Index, SDValue &Disp,
216 SDValue &InChain, SDValue &OutChain);
217 bool TryFoldLoad(SDNode *P, SDValue N,
218 SDValue &Base, SDValue &Scale,
219 SDValue &Index, SDValue &Disp,
221 void PreprocessForRMW();
222 void PreprocessForFPConvert();
224 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
225 /// inline asm expressions.
226 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
228 std::vector<SDValue> &OutOps);
230 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
232 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
233 SDValue &Scale, SDValue &Index,
234 SDValue &Disp, SDValue &Segment) {
235 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
236 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
238 Scale = getI8Imm(AM.Scale);
240 // These are 32-bit even in 64-bit mode since RIP relative offset
243 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp,
246 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
247 AM.Align, AM.Disp, AM.SymbolFlags);
249 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
250 else if (AM.JT != -1)
251 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
252 else if (AM.BlockAddr)
253 Disp = CurDAG->getBlockAddress(AM.BlockAddr, MVT::i32,
254 true, AM.SymbolFlags);
256 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
258 if (AM.Segment.getNode())
259 Segment = AM.Segment;
261 Segment = CurDAG->getRegister(0, MVT::i32);
264 /// getI8Imm - Return a target constant with the specified value, of type
266 inline SDValue getI8Imm(unsigned Imm) {
267 return CurDAG->getTargetConstant(Imm, MVT::i8);
270 /// getI16Imm - Return a target constant with the specified value, of type
272 inline SDValue getI16Imm(unsigned Imm) {
273 return CurDAG->getTargetConstant(Imm, MVT::i16);
276 /// getI32Imm - Return a target constant with the specified value, of type
278 inline SDValue getI32Imm(unsigned Imm) {
279 return CurDAG->getTargetConstant(Imm, MVT::i32);
282 /// getGlobalBaseReg - Return an SDNode that returns the value of
283 /// the global base register. Output instructions required to
284 /// initialize the global base register, if necessary.
286 SDNode *getGlobalBaseReg();
288 /// getTargetMachine - Return a reference to the TargetMachine, casted
289 /// to the target-specific type.
290 const X86TargetMachine &getTargetMachine() {
291 return static_cast<const X86TargetMachine &>(TM);
294 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
295 /// to the target-specific type.
296 const X86InstrInfo *getInstrInfo() {
297 return getTargetMachine().getInstrInfo();
308 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
309 if (OptLevel == CodeGenOpt::None) return false;
314 if (N.getOpcode() != ISD::LOAD)
317 // If N is a load, do additional profitability checks.
319 switch (U->getOpcode()) {
332 SDValue Op1 = U->getOperand(1);
334 // If the other operand is a 8-bit immediate we should fold the immediate
335 // instead. This reduces code size.
337 // movl 4(%esp), %eax
341 // addl 4(%esp), %eax
342 // The former is 2 bytes shorter. In case where the increment is 1, then
343 // the saving can be 4 bytes (by using incl %eax).
344 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
345 if (Imm->getAPIntValue().isSignedIntN(8))
348 // If the other operand is a TLS address, we should fold it instead.
351 // leal i@NTPOFF(%eax), %eax
353 // movl $i@NTPOFF, %eax
355 // if the block also has an access to a second TLS address this will save
357 // FIXME: This is probably also true for non TLS addresses.
358 if (Op1.getOpcode() == X86ISD::Wrapper) {
359 SDValue Val = Op1.getOperand(0);
360 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
371 bool X86DAGToDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const {
372 if (OptLevel == CodeGenOpt::None) return false;
374 // Proceed to 'generic' cycle finder code
375 return SelectionDAGISel::IsLegalToFold(N, U, Root);
378 /// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
379 /// and move load below the TokenFactor. Replace store's chain operand with
380 /// load's chain result.
381 static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
382 SDValue Store, SDValue TF) {
383 SmallVector<SDValue, 4> Ops;
384 for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i)
385 if (Load.getNode() == TF.getOperand(i).getNode())
386 Ops.push_back(Load.getOperand(0));
388 Ops.push_back(TF.getOperand(i));
389 SDValue NewTF = CurDAG->UpdateNodeOperands(TF, &Ops[0], Ops.size());
390 SDValue NewLoad = CurDAG->UpdateNodeOperands(Load, NewTF,
393 CurDAG->UpdateNodeOperands(Store, NewLoad.getValue(1), Store.getOperand(1),
394 Store.getOperand(2), Store.getOperand(3));
397 /// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG. The
398 /// chain produced by the load must only be used by the store's chain operand,
399 /// otherwise this may produce a cycle in the DAG.
401 static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
403 if (N.getOpcode() == ISD::BIT_CONVERT) {
409 LoadSDNode *LD = dyn_cast<LoadSDNode>(N);
410 if (!LD || LD->isVolatile())
412 if (LD->getAddressingMode() != ISD::UNINDEXED)
415 ISD::LoadExtType ExtType = LD->getExtensionType();
416 if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
420 LD->hasNUsesOfValue(1, 1) &&
421 N.getOperand(1) == Address &&
422 LD->isOperandOf(Chain.getNode())) {
429 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
430 /// operand and move load below the call's chain operand.
431 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
432 SDValue Call, SDValue CallSeqStart) {
433 SmallVector<SDValue, 8> Ops;
434 SDValue Chain = CallSeqStart.getOperand(0);
435 if (Chain.getNode() == Load.getNode())
436 Ops.push_back(Load.getOperand(0));
438 assert(Chain.getOpcode() == ISD::TokenFactor &&
439 "Unexpected CallSeqStart chain operand");
440 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
441 if (Chain.getOperand(i).getNode() == Load.getNode())
442 Ops.push_back(Load.getOperand(0));
444 Ops.push_back(Chain.getOperand(i));
446 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
447 MVT::Other, &Ops[0], Ops.size());
449 Ops.push_back(NewChain);
451 for (unsigned i = 1, e = CallSeqStart.getNumOperands(); i != e; ++i)
452 Ops.push_back(CallSeqStart.getOperand(i));
453 CurDAG->UpdateNodeOperands(CallSeqStart, &Ops[0], Ops.size());
454 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
455 Load.getOperand(1), Load.getOperand(2));
457 Ops.push_back(SDValue(Load.getNode(), 1));
458 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
459 Ops.push_back(Call.getOperand(i));
460 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
463 /// isCalleeLoad - Return true if call address is a load and it can be
464 /// moved below CALLSEQ_START and the chains leading up to the call.
465 /// Return the CALLSEQ_START by reference as a second output.
466 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
467 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
469 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
472 LD->getAddressingMode() != ISD::UNINDEXED ||
473 LD->getExtensionType() != ISD::NON_EXTLOAD)
476 // Now let's find the callseq_start.
477 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
478 if (!Chain.hasOneUse())
480 Chain = Chain.getOperand(0);
483 if (Chain.getOperand(0).getNode() == Callee.getNode())
485 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
486 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
487 Callee.getValue(1).hasOneUse())
493 /// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
494 /// This is only run if not in -O0 mode.
495 /// This allows the instruction selector to pick more read-modify-write
496 /// instructions. This is a common case:
506 /// [TokenFactor] [Op]
513 /// The fact the store's chain operand != load's chain will prevent the
514 /// (store (op (load))) instruction from being selected. We can transform it to:
533 void X86DAGToDAGISel::PreprocessForRMW() {
534 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
535 E = CurDAG->allnodes_end(); I != E; ++I) {
536 if (I->getOpcode() == X86ISD::CALL) {
537 /// Also try moving call address load from outside callseq_start to just
538 /// before the call to allow it to be folded.
556 SDValue Chain = I->getOperand(0);
557 SDValue Load = I->getOperand(1);
558 if (!isCalleeLoad(Load, Chain))
560 MoveBelowCallSeqStart(CurDAG, Load, SDValue(I, 0), Chain);
565 if (!ISD::isNON_TRUNCStore(I))
567 SDValue Chain = I->getOperand(0);
569 if (Chain.getNode()->getOpcode() != ISD::TokenFactor)
572 SDValue N1 = I->getOperand(1);
573 SDValue N2 = I->getOperand(2);
574 if ((N1.getValueType().isFloatingPoint() &&
575 !N1.getValueType().isVector()) ||
581 unsigned Opcode = N1.getNode()->getOpcode();
590 case ISD::VECTOR_SHUFFLE: {
591 SDValue N10 = N1.getOperand(0);
592 SDValue N11 = N1.getOperand(1);
593 RModW = isRMWLoad(N10, Chain, N2, Load);
595 RModW = isRMWLoad(N11, Chain, N2, Load);
608 SDValue N10 = N1.getOperand(0);
609 RModW = isRMWLoad(N10, Chain, N2, Load);
615 MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
623 /// PreprocessForFPConvert - Walk over the dag lowering fpround and fpextend
624 /// nodes that target the FP stack to be store and load to the stack. This is a
625 /// gross hack. We would like to simply mark these as being illegal, but when
626 /// we do that, legalize produces these when it expands calls, then expands
627 /// these in the same legalize pass. We would like dag combine to be able to
628 /// hack on these between the call expansion and the node legalization. As such
629 /// this pass basically does "really late" legalization of these inline with the
631 void X86DAGToDAGISel::PreprocessForFPConvert() {
632 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
633 E = CurDAG->allnodes_end(); I != E; ) {
634 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
635 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
638 // If the source and destination are SSE registers, then this is a legal
639 // conversion that should not be lowered.
640 EVT SrcVT = N->getOperand(0).getValueType();
641 EVT DstVT = N->getValueType(0);
642 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
643 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
644 if (SrcIsSSE && DstIsSSE)
647 if (!SrcIsSSE && !DstIsSSE) {
648 // If this is an FPStack extension, it is a noop.
649 if (N->getOpcode() == ISD::FP_EXTEND)
651 // If this is a value-preserving FPStack truncation, it is a noop.
652 if (N->getConstantOperandVal(1))
656 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
657 // FPStack has extload and truncstore. SSE can fold direct loads into other
658 // operations. Based on this, decide what we want to do.
660 if (N->getOpcode() == ISD::FP_ROUND)
661 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
663 MemVT = SrcIsSSE ? SrcVT : DstVT;
665 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
666 DebugLoc dl = N->getDebugLoc();
668 // FIXME: optimize the case where the src/dest is a load or store?
669 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
671 MemTmp, NULL, 0, MemVT,
673 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
674 NULL, 0, MemVT, false, false, 0);
676 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
677 // extload we created. This will cause general havok on the dag because
678 // anything below the conversion could be folded into other existing nodes.
679 // To avoid invalidating 'I', back it up to the convert node.
681 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
683 // Now that we did that, the node is dead. Increment the iterator to the
684 // next node to process, then delete N.
686 CurDAG->DeleteNode(N);
690 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
691 /// when it has created a SelectionDAG for us to codegen.
692 void X86DAGToDAGISel::InstructionSelect() {
693 const Function *F = MF->getFunction();
694 OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
696 if (OptLevel != CodeGenOpt::None)
699 // FIXME: This should only happen when not compiled with -O0.
700 PreprocessForFPConvert();
702 // Codegen the basic block.
704 DEBUG(dbgs() << "===== Instruction selection begins:\n");
709 DEBUG(dbgs() << "===== Instruction selection ends:\n");
712 CurDAG->RemoveDeadNodes();
715 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
716 /// the main function.
717 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
718 MachineFrameInfo *MFI) {
719 const TargetInstrInfo *TII = TM.getInstrInfo();
720 if (Subtarget->isTargetCygMing())
721 BuildMI(BB, DebugLoc::getUnknownLoc(),
722 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
725 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
726 // If this is main, emit special code for main.
727 MachineBasicBlock *BB = MF.begin();
728 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
729 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
733 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
734 X86ISelAddressMode &AM) {
735 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
736 SDValue Segment = N.getOperand(0);
738 if (AM.Segment.getNode() == 0) {
739 AM.Segment = Segment;
746 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
747 // This optimization is valid because the GNU TLS model defines that
748 // gs:0 (or fs:0 on X86-64) contains its own address.
749 // For more information see http://people.redhat.com/drepper/tls.pdf
751 SDValue Address = N.getOperand(1);
752 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
753 !MatchSegmentBaseAddress (Address, AM))
759 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
760 /// into an addressing mode. These wrap things that will resolve down into a
761 /// symbol reference. If no match is possible, this returns true, otherwise it
763 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
764 // If the addressing mode already has a symbol as the displacement, we can
765 // never match another symbol.
766 if (AM.hasSymbolicDisplacement())
769 SDValue N0 = N.getOperand(0);
770 CodeModel::Model M = TM.getCodeModel();
772 // Handle X86-64 rip-relative addresses. We check this before checking direct
773 // folding because RIP is preferable to non-RIP accesses.
774 if (Subtarget->is64Bit() &&
775 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
776 // they cannot be folded into immediate fields.
777 // FIXME: This can be improved for kernel and other models?
778 (M == CodeModel::Small || M == CodeModel::Kernel) &&
779 // Base and index reg must be 0 in order to use %rip as base and lowering
781 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
782 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
783 int64_t Offset = AM.Disp + G->getOffset();
784 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
785 AM.GV = G->getGlobal();
787 AM.SymbolFlags = G->getTargetFlags();
788 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
789 int64_t Offset = AM.Disp + CP->getOffset();
790 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
791 AM.CP = CP->getConstVal();
792 AM.Align = CP->getAlignment();
794 AM.SymbolFlags = CP->getTargetFlags();
795 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
796 AM.ES = S->getSymbol();
797 AM.SymbolFlags = S->getTargetFlags();
798 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
799 AM.JT = J->getIndex();
800 AM.SymbolFlags = J->getTargetFlags();
802 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
803 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
806 if (N.getOpcode() == X86ISD::WrapperRIP)
807 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
811 // Handle the case when globals fit in our immediate field: This is true for
812 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
813 // mode, this results in a non-RIP-relative computation.
814 if (!Subtarget->is64Bit() ||
815 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
816 TM.getRelocationModel() == Reloc::Static)) {
817 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
818 AM.GV = G->getGlobal();
819 AM.Disp += G->getOffset();
820 AM.SymbolFlags = G->getTargetFlags();
821 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
822 AM.CP = CP->getConstVal();
823 AM.Align = CP->getAlignment();
824 AM.Disp += CP->getOffset();
825 AM.SymbolFlags = CP->getTargetFlags();
826 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
827 AM.ES = S->getSymbol();
828 AM.SymbolFlags = S->getTargetFlags();
829 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
830 AM.JT = J->getIndex();
831 AM.SymbolFlags = J->getTargetFlags();
833 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
834 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
842 /// MatchAddress - Add the specified node to the specified addressing mode,
843 /// returning true if it cannot be done. This just pattern matches for the
845 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
846 if (MatchAddressRecursively(N, AM, 0))
849 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
850 // a smaller encoding and avoids a scaled-index.
852 AM.BaseType == X86ISelAddressMode::RegBase &&
853 AM.Base.Reg.getNode() == 0) {
854 AM.Base.Reg = AM.IndexReg;
858 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
859 // because it has a smaller encoding.
860 // TODO: Which other code models can use this?
861 if (TM.getCodeModel() == CodeModel::Small &&
862 Subtarget->is64Bit() &&
864 AM.BaseType == X86ISelAddressMode::RegBase &&
865 AM.Base.Reg.getNode() == 0 &&
866 AM.IndexReg.getNode() == 0 &&
867 AM.SymbolFlags == X86II::MO_NO_FLAG &&
868 AM.hasSymbolicDisplacement())
869 AM.Base.Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
874 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
876 bool is64Bit = Subtarget->is64Bit();
877 DebugLoc dl = N.getDebugLoc();
879 dbgs() << "MatchAddress: ";
884 return MatchAddressBase(N, AM);
886 CodeModel::Model M = TM.getCodeModel();
888 // If this is already a %rip relative address, we can only merge immediates
889 // into it. Instead of handling this in every case, we handle it here.
890 // RIP relative addressing: %rip + 32-bit displacement!
891 if (AM.isRIPRelative()) {
892 // FIXME: JumpTable and ExternalSymbol address currently don't like
893 // displacements. It isn't very important, but this should be fixed for
895 if (!AM.ES && AM.JT != -1) return true;
897 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
898 int64_t Val = AM.Disp + Cst->getSExtValue();
899 if (X86::isOffsetSuitableForCodeModel(Val, M,
900 AM.hasSymbolicDisplacement())) {
908 switch (N.getOpcode()) {
910 case ISD::Constant: {
911 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
913 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
914 AM.hasSymbolicDisplacement())) {
921 case X86ISD::SegmentBaseAddress:
922 if (!MatchSegmentBaseAddress(N, AM))
926 case X86ISD::Wrapper:
927 case X86ISD::WrapperRIP:
928 if (!MatchWrapper(N, AM))
933 if (!MatchLoad(N, AM))
937 case ISD::FrameIndex:
938 if (AM.BaseType == X86ISelAddressMode::RegBase
939 && AM.Base.Reg.getNode() == 0) {
940 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
941 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
947 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
951 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
952 unsigned Val = CN->getZExtValue();
953 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
954 // that the base operand remains free for further matching. If
955 // the base doesn't end up getting used, a post-processing step
956 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
957 if (Val == 1 || Val == 2 || Val == 3) {
959 SDValue ShVal = N.getNode()->getOperand(0);
961 // Okay, we know that we have a scale by now. However, if the scaled
962 // value is an add of something and a constant, we can fold the
963 // constant into the disp field here.
964 if (ShVal.getNode()->getOpcode() == ISD::ADD &&
965 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
966 AM.IndexReg = ShVal.getNode()->getOperand(0);
967 ConstantSDNode *AddVal =
968 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
969 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
971 X86::isOffsetSuitableForCodeModel(Disp, M,
972 AM.hasSymbolicDisplacement()))
986 // A mul_lohi where we need the low part can be folded as a plain multiply.
987 if (N.getResNo() != 0) break;
990 case X86ISD::MUL_IMM:
991 // X*[3,5,9] -> X+X*[2,4,8]
992 if (AM.BaseType == X86ISelAddressMode::RegBase &&
993 AM.Base.Reg.getNode() == 0 &&
994 AM.IndexReg.getNode() == 0) {
996 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
997 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
998 CN->getZExtValue() == 9) {
999 AM.Scale = unsigned(CN->getZExtValue())-1;
1001 SDValue MulVal = N.getNode()->getOperand(0);
1004 // Okay, we know that we have a scale by now. However, if the scaled
1005 // value is an add of something and a constant, we can fold the
1006 // constant into the disp field here.
1007 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1008 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1009 Reg = MulVal.getNode()->getOperand(0);
1010 ConstantSDNode *AddVal =
1011 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1012 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
1015 X86::isOffsetSuitableForCodeModel(Disp, M,
1016 AM.hasSymbolicDisplacement()))
1019 Reg = N.getNode()->getOperand(0);
1021 Reg = N.getNode()->getOperand(0);
1024 AM.IndexReg = AM.Base.Reg = Reg;
1031 // Given A-B, if A can be completely folded into the address and
1032 // the index field with the index field unused, use -B as the index.
1033 // This is a win if a has multiple parts that can be folded into
1034 // the address. Also, this saves a mov if the base register has
1035 // other uses, since it avoids a two-address sub instruction, however
1036 // it costs an additional mov if the index register has other uses.
1038 // Test if the LHS of the sub can be folded.
1039 X86ISelAddressMode Backup = AM;
1040 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1044 // Test if the index field is free for use.
1045 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1050 SDValue RHS = N.getNode()->getOperand(1);
1051 // If the RHS involves a register with multiple uses, this
1052 // transformation incurs an extra mov, due to the neg instruction
1053 // clobbering its operand.
1054 if (!RHS.getNode()->hasOneUse() ||
1055 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1056 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1057 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1058 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1059 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1061 // If the base is a register with multiple uses, this
1062 // transformation may save a mov.
1063 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1064 AM.Base.Reg.getNode() &&
1065 !AM.Base.Reg.getNode()->hasOneUse()) ||
1066 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1068 // If the folded LHS was interesting, this transformation saves
1069 // address arithmetic.
1070 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1071 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1072 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1074 // If it doesn't look like it may be an overall win, don't do it.
1080 // Ok, the transformation is legal and appears profitable. Go for it.
1081 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
1082 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1086 // Insert the new nodes into the topological ordering.
1087 if (Zero.getNode()->getNodeId() == -1 ||
1088 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1089 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
1090 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
1092 if (Neg.getNode()->getNodeId() == -1 ||
1093 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1094 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
1095 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
1101 X86ISelAddressMode Backup = AM;
1102 if (!MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1) &&
1103 !MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1))
1106 if (!MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1) &&
1107 !MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1))
1111 // If we couldn't fold both operands into the address at the same time,
1112 // see if we can just put each operand into a register and fold at least
1114 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1115 !AM.Base.Reg.getNode() &&
1116 !AM.IndexReg.getNode()) {
1117 AM.Base.Reg = N.getNode()->getOperand(0);
1118 AM.IndexReg = N.getNode()->getOperand(1);
1126 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1127 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1128 X86ISelAddressMode Backup = AM;
1129 uint64_t Offset = CN->getSExtValue();
1130 // Start with the LHS as an addr mode.
1131 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1132 // Address could not have picked a GV address for the displacement.
1134 // On x86-64, the resultant disp must fit in 32-bits.
1136 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
1137 AM.hasSymbolicDisplacement())) &&
1138 // Check to see if the LHS & C is zero.
1139 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
1148 // Perform some heroic transforms on an and of a constant-count shift
1149 // with a constant to enable use of the scaled offset field.
1151 SDValue Shift = N.getOperand(0);
1152 if (Shift.getNumOperands() != 2) break;
1154 // Scale must not be used already.
1155 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1157 SDValue X = Shift.getOperand(0);
1158 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1159 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1160 if (!C1 || !C2) break;
1162 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1163 // allows us to convert the shift and and into an h-register extract and
1165 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1166 unsigned ScaleLog = 8 - C1->getZExtValue();
1167 if (ScaleLog > 0 && ScaleLog < 4 &&
1168 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1169 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1170 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1171 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1173 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1175 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1176 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1179 // Insert the new nodes into the topological ordering.
1180 if (Eight.getNode()->getNodeId() == -1 ||
1181 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1182 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1183 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1185 if (Mask.getNode()->getNodeId() == -1 ||
1186 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1187 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1188 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1190 if (Srl.getNode()->getNodeId() == -1 ||
1191 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1192 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1193 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1195 if (And.getNode()->getNodeId() == -1 ||
1196 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1197 CurDAG->RepositionNode(N.getNode(), And.getNode());
1198 And.getNode()->setNodeId(N.getNode()->getNodeId());
1200 if (ShlCount.getNode()->getNodeId() == -1 ||
1201 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1202 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1203 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1205 if (Shl.getNode()->getNodeId() == -1 ||
1206 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1207 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1208 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1210 CurDAG->ReplaceAllUsesWith(N, Shl);
1212 AM.Scale = (1 << ScaleLog);
1217 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1218 // allows us to fold the shift into this addressing mode.
1219 if (Shift.getOpcode() != ISD::SHL) break;
1221 // Not likely to be profitable if either the AND or SHIFT node has more
1222 // than one use (unless all uses are for address computation). Besides,
1223 // isel mechanism requires their node ids to be reused.
1224 if (!N.hasOneUse() || !Shift.hasOneUse())
1227 // Verify that the shift amount is something we can fold.
1228 unsigned ShiftCst = C1->getZExtValue();
1229 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1232 // Get the new AND mask, this folds to a constant.
1233 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1234 SDValue(C2, 0), SDValue(C1, 0));
1235 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1237 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1238 NewAND, SDValue(C1, 0));
1240 // Insert the new nodes into the topological ordering.
1241 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1242 CurDAG->RepositionNode(X.getNode(), C1);
1243 C1->setNodeId(X.getNode()->getNodeId());
1245 if (NewANDMask.getNode()->getNodeId() == -1 ||
1246 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1247 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1248 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1250 if (NewAND.getNode()->getNodeId() == -1 ||
1251 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1252 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1253 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1255 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1256 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1257 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1258 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1261 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1263 AM.Scale = 1 << ShiftCst;
1264 AM.IndexReg = NewAND;
1269 return MatchAddressBase(N, AM);
1272 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1273 /// specified addressing mode without any further recursion.
1274 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1275 // Is the base register already occupied?
1276 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
1277 // If so, check to see if the scale index register is set.
1278 if (AM.IndexReg.getNode() == 0) {
1284 // Otherwise, we cannot select it.
1288 // Default, generate it as a register.
1289 AM.BaseType = X86ISelAddressMode::RegBase;
1294 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1295 /// It returns the operands which make up the maximal addressing mode it can
1296 /// match by reference.
1297 bool X86DAGToDAGISel::SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
1298 SDValue &Scale, SDValue &Index,
1299 SDValue &Disp, SDValue &Segment) {
1300 X86ISelAddressMode AM;
1301 if (MatchAddress(N, AM))
1304 EVT VT = N.getValueType();
1305 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1306 if (!AM.Base.Reg.getNode())
1307 AM.Base.Reg = CurDAG->getRegister(0, VT);
1310 if (!AM.IndexReg.getNode())
1311 AM.IndexReg = CurDAG->getRegister(0, VT);
1313 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1317 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1318 /// match a load whose top elements are either undef or zeros. The load flavor
1319 /// is derived from the type of N, which is either v4f32 or v2f64.
1320 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1321 SDValue N, SDValue &Base,
1322 SDValue &Scale, SDValue &Index,
1323 SDValue &Disp, SDValue &Segment,
1325 SDValue &OutChain) {
1326 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1327 InChain = N.getOperand(0).getValue(1);
1328 if (ISD::isNON_EXTLoad(InChain.getNode()) &&
1329 InChain.getValue(0).hasOneUse() &&
1330 IsProfitableToFold(N.getOperand(0), InChain.getNode(), Root) &&
1331 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1332 LoadSDNode *LD = cast<LoadSDNode>(InChain);
1333 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp,Segment))
1335 OutChain = LD->getChain();
1340 // Also handle the case where we explicitly require zeros in the top
1341 // elements. This is a vector shuffle from the zero vector.
1342 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1343 // Check to see if the top elements are all zeros (or bitcast of zeros).
1344 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1345 N.getOperand(0).getNode()->hasOneUse() &&
1346 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1347 N.getOperand(0).getOperand(0).hasOneUse() &&
1348 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1349 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1350 // Okay, this is a zero extending load. Fold it.
1351 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1352 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1354 OutChain = LD->getChain();
1355 InChain = SDValue(LD, 1);
1362 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1363 /// mode it matches can be cost effectively emitted as an LEA instruction.
1364 bool X86DAGToDAGISel::SelectLEAAddr(SDNode *Op, SDValue N,
1365 SDValue &Base, SDValue &Scale,
1366 SDValue &Index, SDValue &Disp) {
1367 X86ISelAddressMode AM;
1369 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1371 SDValue Copy = AM.Segment;
1372 SDValue T = CurDAG->getRegister(0, MVT::i32);
1374 if (MatchAddress(N, AM))
1376 assert (T == AM.Segment);
1379 EVT VT = N.getValueType();
1380 unsigned Complexity = 0;
1381 if (AM.BaseType == X86ISelAddressMode::RegBase)
1382 if (AM.Base.Reg.getNode())
1385 AM.Base.Reg = CurDAG->getRegister(0, VT);
1386 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1389 if (AM.IndexReg.getNode())
1392 AM.IndexReg = CurDAG->getRegister(0, VT);
1394 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1399 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1400 // to a LEA. This is determined with some expermentation but is by no means
1401 // optimal (especially for code size consideration). LEA is nice because of
1402 // its three-address nature. Tweak the cost function again when we can run
1403 // convertToThreeAddress() at register allocation time.
1404 if (AM.hasSymbolicDisplacement()) {
1405 // For X86-64, we should always use lea to materialize RIP relative
1407 if (Subtarget->is64Bit())
1413 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1416 // If it isn't worth using an LEA, reject it.
1417 if (Complexity <= 2)
1421 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1425 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1426 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
1427 SDValue &Scale, SDValue &Index,
1429 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1430 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1432 X86ISelAddressMode AM;
1433 AM.GV = GA->getGlobal();
1434 AM.Disp += GA->getOffset();
1435 AM.Base.Reg = CurDAG->getRegister(0, N.getValueType());
1436 AM.SymbolFlags = GA->getTargetFlags();
1438 if (N.getValueType() == MVT::i32) {
1440 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1442 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1446 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1451 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1452 SDValue &Base, SDValue &Scale,
1453 SDValue &Index, SDValue &Disp,
1455 if (ISD::isNON_EXTLoad(N.getNode()) &&
1456 IsProfitableToFold(N, P, P) &&
1457 IsLegalToFold(N, P, P))
1458 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
1462 /// getGlobalBaseReg - Return an SDNode that returns the value of
1463 /// the global base register. Output instructions required to
1464 /// initialize the global base register, if necessary.
1466 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1467 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1468 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1471 static SDNode *FindCallStartFromCall(SDNode *Node) {
1472 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1473 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1474 "Node doesn't have a token chain argument!");
1475 return FindCallStartFromCall(Node->getOperand(0).getNode());
1478 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1479 SDValue Chain = Node->getOperand(0);
1480 SDValue In1 = Node->getOperand(1);
1481 SDValue In2L = Node->getOperand(2);
1482 SDValue In2H = Node->getOperand(3);
1483 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1484 if (!SelectAddr(In1.getNode(), In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1486 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1487 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1488 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1489 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1490 MVT::i32, MVT::i32, MVT::Other, Ops,
1491 array_lengthof(Ops));
1492 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1496 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1497 if (Node->hasAnyUseOfValue(0))
1500 // Optimize common patterns for __sync_add_and_fetch and
1501 // __sync_sub_and_fetch where the result is not used. This allows us
1502 // to use "lock" version of add, sub, inc, dec instructions.
1503 // FIXME: Do not use special instructions but instead add the "lock"
1504 // prefix to the target node somehow. The extra information will then be
1505 // transferred to machine instruction and it denotes the prefix.
1506 SDValue Chain = Node->getOperand(0);
1507 SDValue Ptr = Node->getOperand(1);
1508 SDValue Val = Node->getOperand(2);
1509 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1510 if (!SelectAddr(Ptr.getNode(), Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1513 bool isInc = false, isDec = false, isSub = false, isCN = false;
1514 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1517 int64_t CNVal = CN->getSExtValue();
1520 else if (CNVal == -1)
1522 else if (CNVal >= 0)
1523 Val = CurDAG->getTargetConstant(CNVal, NVT);
1526 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1528 } else if (Val.hasOneUse() &&
1529 Val.getOpcode() == ISD::SUB &&
1530 X86::isZeroNode(Val.getOperand(0))) {
1532 Val = Val.getOperand(1);
1536 switch (NVT.getSimpleVT().SimpleTy) {
1540 Opc = X86::LOCK_INC8m;
1542 Opc = X86::LOCK_DEC8m;
1545 Opc = X86::LOCK_SUB8mi;
1547 Opc = X86::LOCK_SUB8mr;
1550 Opc = X86::LOCK_ADD8mi;
1552 Opc = X86::LOCK_ADD8mr;
1557 Opc = X86::LOCK_INC16m;
1559 Opc = X86::LOCK_DEC16m;
1562 if (Predicate_i16immSExt8(Val.getNode()))
1563 Opc = X86::LOCK_SUB16mi8;
1565 Opc = X86::LOCK_SUB16mi;
1567 Opc = X86::LOCK_SUB16mr;
1570 if (Predicate_i16immSExt8(Val.getNode()))
1571 Opc = X86::LOCK_ADD16mi8;
1573 Opc = X86::LOCK_ADD16mi;
1575 Opc = X86::LOCK_ADD16mr;
1580 Opc = X86::LOCK_INC32m;
1582 Opc = X86::LOCK_DEC32m;
1585 if (Predicate_i32immSExt8(Val.getNode()))
1586 Opc = X86::LOCK_SUB32mi8;
1588 Opc = X86::LOCK_SUB32mi;
1590 Opc = X86::LOCK_SUB32mr;
1593 if (Predicate_i32immSExt8(Val.getNode()))
1594 Opc = X86::LOCK_ADD32mi8;
1596 Opc = X86::LOCK_ADD32mi;
1598 Opc = X86::LOCK_ADD32mr;
1603 Opc = X86::LOCK_INC64m;
1605 Opc = X86::LOCK_DEC64m;
1607 Opc = X86::LOCK_SUB64mr;
1609 if (Predicate_i64immSExt8(Val.getNode()))
1610 Opc = X86::LOCK_SUB64mi8;
1611 else if (Predicate_i64immSExt32(Val.getNode()))
1612 Opc = X86::LOCK_SUB64mi32;
1615 Opc = X86::LOCK_ADD64mr;
1617 if (Predicate_i64immSExt8(Val.getNode()))
1618 Opc = X86::LOCK_ADD64mi8;
1619 else if (Predicate_i64immSExt32(Val.getNode()))
1620 Opc = X86::LOCK_ADD64mi32;
1626 DebugLoc dl = Node->getDebugLoc();
1627 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1629 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1630 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1631 if (isInc || isDec) {
1632 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1633 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1634 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1635 SDValue RetVals[] = { Undef, Ret };
1636 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1638 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1639 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1640 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1641 SDValue RetVals[] = { Undef, Ret };
1642 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1646 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1647 /// any uses which require the SF or OF bits to be accurate.
1648 static bool HasNoSignedComparisonUses(SDNode *N) {
1649 // Examine each user of the node.
1650 for (SDNode::use_iterator UI = N->use_begin(),
1651 UE = N->use_end(); UI != UE; ++UI) {
1652 // Only examine CopyToReg uses.
1653 if (UI->getOpcode() != ISD::CopyToReg)
1655 // Only examine CopyToReg uses that copy to EFLAGS.
1656 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1659 // Examine each user of the CopyToReg use.
1660 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1661 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1662 // Only examine the Flag result.
1663 if (FlagUI.getUse().getResNo() != 1) continue;
1664 // Anything unusual: assume conservatively.
1665 if (!FlagUI->isMachineOpcode()) return false;
1666 // Examine the opcode of the user.
1667 switch (FlagUI->getMachineOpcode()) {
1668 // These comparisons don't treat the most significant bit specially.
1669 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1670 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1671 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1672 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1673 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1674 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1675 case X86::CMOVA16rr: case X86::CMOVA16rm:
1676 case X86::CMOVA32rr: case X86::CMOVA32rm:
1677 case X86::CMOVA64rr: case X86::CMOVA64rm:
1678 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1679 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1680 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1681 case X86::CMOVB16rr: case X86::CMOVB16rm:
1682 case X86::CMOVB32rr: case X86::CMOVB32rm:
1683 case X86::CMOVB64rr: case X86::CMOVB64rm:
1684 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1685 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1686 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1687 case X86::CMOVE16rr: case X86::CMOVE16rm:
1688 case X86::CMOVE32rr: case X86::CMOVE32rm:
1689 case X86::CMOVE64rr: case X86::CMOVE64rm:
1690 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1691 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1692 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1693 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1694 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1695 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1696 case X86::CMOVP16rr: case X86::CMOVP16rm:
1697 case X86::CMOVP32rr: case X86::CMOVP32rm:
1698 case X86::CMOVP64rr: case X86::CMOVP64rm:
1700 // Anything else: assume conservatively.
1701 default: return false;
1708 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1709 EVT NVT = Node->getValueType(0);
1711 unsigned Opcode = Node->getOpcode();
1712 DebugLoc dl = Node->getDebugLoc();
1716 dbgs() << std::string(Indent, ' ') << "Selecting: ";
1723 if (Node->isMachineOpcode()) {
1726 dbgs() << std::string(Indent-2, ' ') << "== ";
1732 return NULL; // Already selected.
1737 case X86ISD::GlobalBaseReg:
1738 return getGlobalBaseReg();
1740 case X86ISD::ATOMOR64_DAG:
1741 return SelectAtomic64(Node, X86::ATOMOR6432);
1742 case X86ISD::ATOMXOR64_DAG:
1743 return SelectAtomic64(Node, X86::ATOMXOR6432);
1744 case X86ISD::ATOMADD64_DAG:
1745 return SelectAtomic64(Node, X86::ATOMADD6432);
1746 case X86ISD::ATOMSUB64_DAG:
1747 return SelectAtomic64(Node, X86::ATOMSUB6432);
1748 case X86ISD::ATOMNAND64_DAG:
1749 return SelectAtomic64(Node, X86::ATOMNAND6432);
1750 case X86ISD::ATOMAND64_DAG:
1751 return SelectAtomic64(Node, X86::ATOMAND6432);
1752 case X86ISD::ATOMSWAP64_DAG:
1753 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1755 case ISD::ATOMIC_LOAD_ADD: {
1756 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1762 case ISD::SMUL_LOHI:
1763 case ISD::UMUL_LOHI: {
1764 SDValue N0 = Node->getOperand(0);
1765 SDValue N1 = Node->getOperand(1);
1767 bool isSigned = Opcode == ISD::SMUL_LOHI;
1769 switch (NVT.getSimpleVT().SimpleTy) {
1770 default: llvm_unreachable("Unsupported VT!");
1771 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1772 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1773 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1774 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1777 switch (NVT.getSimpleVT().SimpleTy) {
1778 default: llvm_unreachable("Unsupported VT!");
1779 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1780 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1781 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1782 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1786 unsigned LoReg, HiReg;
1787 switch (NVT.getSimpleVT().SimpleTy) {
1788 default: llvm_unreachable("Unsupported VT!");
1789 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1790 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1791 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1792 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1795 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1796 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1797 // Multiply is commmutative.
1799 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1804 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1805 N0, SDValue()).getValue(1);
1808 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1811 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1812 array_lengthof(Ops));
1813 InFlag = SDValue(CNode, 1);
1814 // Update the chain.
1815 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1818 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1821 // Copy the low half of the result, if it is needed.
1822 if (!SDValue(Node, 0).use_empty()) {
1823 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1824 LoReg, NVT, InFlag);
1825 InFlag = Result.getValue(2);
1826 ReplaceUses(SDValue(Node, 0), Result);
1829 dbgs() << std::string(Indent-2, ' ') << "=> ";
1830 Result.getNode()->dump(CurDAG);
1835 // Copy the high half of the result, if it is needed.
1836 if (!SDValue(Node, 1).use_empty()) {
1838 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1839 // Prevent use of AH in a REX instruction by referencing AX instead.
1840 // Shift it down 8 bits.
1841 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1842 X86::AX, MVT::i16, InFlag);
1843 InFlag = Result.getValue(2);
1844 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1846 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1847 // Then truncate it down to i8.
1848 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
1851 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1852 HiReg, NVT, InFlag);
1853 InFlag = Result.getValue(2);
1855 ReplaceUses(SDValue(Node, 1), Result);
1858 dbgs() << std::string(Indent-2, ' ') << "=> ";
1859 Result.getNode()->dump(CurDAG);
1873 case ISD::UDIVREM: {
1874 SDValue N0 = Node->getOperand(0);
1875 SDValue N1 = Node->getOperand(1);
1877 bool isSigned = Opcode == ISD::SDIVREM;
1879 switch (NVT.getSimpleVT().SimpleTy) {
1880 default: llvm_unreachable("Unsupported VT!");
1881 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1882 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1883 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1884 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1887 switch (NVT.getSimpleVT().SimpleTy) {
1888 default: llvm_unreachable("Unsupported VT!");
1889 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1890 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1891 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1892 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1896 unsigned LoReg, HiReg, ClrReg;
1897 unsigned ClrOpcode, SExtOpcode;
1898 switch (NVT.getSimpleVT().SimpleTy) {
1899 default: llvm_unreachable("Unsupported VT!");
1901 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1903 SExtOpcode = X86::CBW;
1906 LoReg = X86::AX; HiReg = X86::DX;
1907 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1908 SExtOpcode = X86::CWD;
1911 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1912 ClrOpcode = X86::MOV32r0;
1913 SExtOpcode = X86::CDQ;
1916 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1917 ClrOpcode = X86::MOV64r0;
1918 SExtOpcode = X86::CQO;
1922 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1923 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1924 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1927 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1928 // Special case for div8, just use a move with zero extension to AX to
1929 // clear the upper 8 bits (AH).
1930 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1931 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1932 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1934 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1936 array_lengthof(Ops)), 0);
1937 Chain = Move.getValue(1);
1938 ReplaceUses(N0.getValue(1), Chain);
1941 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1942 Chain = CurDAG->getEntryNode();
1944 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1945 InFlag = Chain.getValue(1);
1948 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1949 LoReg, N0, SDValue()).getValue(1);
1950 if (isSigned && !signBitIsZero) {
1951 // Sign extend the low part into the high part.
1953 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1955 // Zero out the high part, effectively zero extending the input.
1957 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1958 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1959 ClrNode, InFlag).getValue(1);
1964 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1967 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1968 array_lengthof(Ops));
1969 InFlag = SDValue(CNode, 1);
1970 // Update the chain.
1971 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1974 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1977 // Copy the division (low) result, if it is needed.
1978 if (!SDValue(Node, 0).use_empty()) {
1979 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1980 LoReg, NVT, InFlag);
1981 InFlag = Result.getValue(2);
1982 ReplaceUses(SDValue(Node, 0), Result);
1985 dbgs() << std::string(Indent-2, ' ') << "=> ";
1986 Result.getNode()->dump(CurDAG);
1991 // Copy the remainder (high) result, if it is needed.
1992 if (!SDValue(Node, 1).use_empty()) {
1994 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1995 // Prevent use of AH in a REX instruction by referencing AX instead.
1996 // Shift it down 8 bits.
1997 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1998 X86::AX, MVT::i16, InFlag);
1999 InFlag = Result.getValue(2);
2000 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2002 CurDAG->getTargetConstant(8, MVT::i8)),
2004 // Then truncate it down to i8.
2005 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
2008 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2009 HiReg, NVT, InFlag);
2010 InFlag = Result.getValue(2);
2012 ReplaceUses(SDValue(Node, 1), Result);
2015 dbgs() << std::string(Indent-2, ' ') << "=> ";
2016 Result.getNode()->dump(CurDAG);
2030 SDValue N0 = Node->getOperand(0);
2031 SDValue N1 = Node->getOperand(1);
2033 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2034 // use a smaller encoding.
2035 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
2036 N0.getValueType() != MVT::i8 &&
2037 X86::isZeroNode(N1)) {
2038 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2041 // For example, convert "testl %eax, $8" to "testb %al, $8"
2042 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2043 (!(C->getZExtValue() & 0x80) ||
2044 HasNoSignedComparisonUses(Node))) {
2045 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
2046 SDValue Reg = N0.getNode()->getOperand(0);
2048 // On x86-32, only the ABCD registers have 8-bit subregisters.
2049 if (!Subtarget->is64Bit()) {
2050 TargetRegisterClass *TRC = 0;
2051 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2052 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2053 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2054 default: llvm_unreachable("Unsupported TEST operand type!");
2056 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2057 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2058 Reg.getValueType(), Reg, RC), 0);
2061 // Extract the l-register.
2062 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
2066 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
2069 // For example, "testl %eax, $2048" to "testb %ah, $8".
2070 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2071 (!(C->getZExtValue() & 0x8000) ||
2072 HasNoSignedComparisonUses(Node))) {
2073 // Shift the immediate right by 8 bits.
2074 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2076 SDValue Reg = N0.getNode()->getOperand(0);
2078 // Put the value in an ABCD register.
2079 TargetRegisterClass *TRC = 0;
2080 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2081 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2082 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2083 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2084 default: llvm_unreachable("Unsupported TEST operand type!");
2086 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2087 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2088 Reg.getValueType(), Reg, RC), 0);
2090 // Extract the h-register.
2091 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT_HI, dl,
2094 // Emit a testb. No special NOREX tricks are needed since there's
2095 // only one GPR operand!
2096 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2097 Subreg, ShiftedImm);
2100 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2101 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2102 N0.getValueType() != MVT::i16 &&
2103 (!(C->getZExtValue() & 0x8000) ||
2104 HasNoSignedComparisonUses(Node))) {
2105 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
2106 SDValue Reg = N0.getNode()->getOperand(0);
2108 // Extract the 16-bit subregister.
2109 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_16BIT, dl,
2113 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
2116 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2117 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2118 N0.getValueType() == MVT::i64 &&
2119 (!(C->getZExtValue() & 0x80000000) ||
2120 HasNoSignedComparisonUses(Node))) {
2121 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
2122 SDValue Reg = N0.getNode()->getOperand(0);
2124 // Extract the 32-bit subregister.
2125 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_32BIT, dl,
2129 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
2136 SDNode *ResNode = SelectCode(Node);
2140 dbgs() << std::string(Indent-2, ' ') << "=> ";
2141 if (ResNode == NULL || ResNode == Node)
2144 ResNode->dump(CurDAG);
2153 bool X86DAGToDAGISel::
2154 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2155 std::vector<SDValue> &OutOps) {
2156 SDValue Op0, Op1, Op2, Op3, Op4;
2157 switch (ConstraintCode) {
2158 case 'o': // offsetable ??
2159 case 'v': // not offsetable ??
2160 default: return true;
2162 if (!SelectAddr(Op.getNode(), Op, Op0, Op1, Op2, Op3, Op4))
2167 OutOps.push_back(Op0);
2168 OutOps.push_back(Op1);
2169 OutOps.push_back(Op2);
2170 OutOps.push_back(Op3);
2171 OutOps.push_back(Op4);
2175 /// createX86ISelDag - This pass converts a legalized DAG into a
2176 /// X86-specific DAG, ready for instruction scheduling.
2178 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2179 llvm::CodeGenOpt::Level OptLevel) {
2180 return new X86DAGToDAGISel(TM, OptLevel);