1 //===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a DAG pattern matching instruction selector for X86,
11 // converting from a legalized dag to a X86 dag.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86RegisterInfo.h"
21 #include "X86Subtarget.h"
22 #include "X86TargetMachine.h"
23 #include "llvm/GlobalValue.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/Support/CFG.h"
27 #include "llvm/Type.h"
28 #include "llvm/CodeGen/MachineConstantPool.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAGISel.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetOptions.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/Statistic.h"
44 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
46 //===----------------------------------------------------------------------===//
47 // Pattern Matcher Implementation
48 //===----------------------------------------------------------------------===//
51 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
52 /// SDValue's instead of register numbers for the leaves of the matched
54 struct X86ISelAddressMode {
60 struct { // This is really a union, discriminated by BaseType!
71 BlockAddress *BlockAddr;
74 unsigned Align; // CP alignment.
75 unsigned char SymbolFlags; // X86II::MO_*
78 : BaseType(RegBase), Scale(1), IndexReg(), Disp(0),
79 Segment(), GV(0), CP(0), BlockAddr(0), ES(0), JT(-1), Align(0),
80 SymbolFlags(X86II::MO_NO_FLAG) {
83 bool hasSymbolicDisplacement() const {
84 return GV != 0 || CP != 0 || ES != 0 || JT != -1 || BlockAddr != 0;
87 bool hasBaseOrIndexReg() const {
88 return IndexReg.getNode() != 0 || Base.Reg.getNode() != 0;
91 /// isRIPRelative - Return true if this addressing mode is already RIP
93 bool isRIPRelative() const {
94 if (BaseType != RegBase) return false;
95 if (RegisterSDNode *RegNode =
96 dyn_cast_or_null<RegisterSDNode>(Base.Reg.getNode()))
97 return RegNode->getReg() == X86::RIP;
101 void setBaseReg(SDValue Reg) {
107 dbgs() << "X86ISelAddressMode " << this << '\n';
108 dbgs() << "Base.Reg ";
109 if (Base.Reg.getNode() != 0)
110 Base.Reg.getNode()->dump();
113 dbgs() << " Base.FrameIndex " << Base.FrameIndex << '\n'
114 << " Scale" << Scale << '\n'
116 if (IndexReg.getNode() != 0)
117 IndexReg.getNode()->dump();
120 dbgs() << " Disp " << Disp << '\n'
137 dbgs() << " JT" << JT << " Align" << Align << '\n';
143 //===--------------------------------------------------------------------===//
144 /// ISel - X86 specific code to select X86 machine instructions for
145 /// SelectionDAG operations.
147 class X86DAGToDAGISel : public SelectionDAGISel {
148 /// X86Lowering - This object fully describes how to lower LLVM code to an
149 /// X86-specific SelectionDAG.
150 X86TargetLowering &X86Lowering;
152 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
153 /// make the right decision when generating code for different targets.
154 const X86Subtarget *Subtarget;
156 /// OptForSize - If true, selector should try to optimize for code size
157 /// instead of performance.
161 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
162 : SelectionDAGISel(tm, OptLevel),
163 X86Lowering(*tm.getTargetLowering()),
164 Subtarget(&tm.getSubtarget<X86Subtarget>()),
167 virtual const char *getPassName() const {
168 return "X86 DAG->DAG Instruction Selection";
171 virtual void EmitFunctionEntryCode(Function &Fn, MachineFunction &MF);
173 virtual bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const;
175 virtual void PreprocessISelDAG();
177 // Include the pieces autogenerated from the target description.
178 #include "X86GenDAGISel.inc"
181 SDNode *Select(SDNode *N);
182 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
183 SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
185 bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
186 bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
187 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
188 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
189 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
191 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
192 bool SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
193 SDValue &Scale, SDValue &Index, SDValue &Disp,
195 bool SelectLEAAddr(SDNode *Op, SDValue N, SDValue &Base,
196 SDValue &Scale, SDValue &Index, SDValue &Disp);
197 bool SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
198 SDValue &Scale, SDValue &Index, SDValue &Disp);
199 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
200 SDValue &Base, SDValue &Scale,
201 SDValue &Index, SDValue &Disp,
203 SDValue &NodeWithChain);
205 bool TryFoldLoad(SDNode *P, SDValue N,
206 SDValue &Base, SDValue &Scale,
207 SDValue &Index, SDValue &Disp,
210 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
211 /// inline asm expressions.
212 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
214 std::vector<SDValue> &OutOps);
216 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
218 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
219 SDValue &Scale, SDValue &Index,
220 SDValue &Disp, SDValue &Segment) {
221 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
222 CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy()) :
224 Scale = getI8Imm(AM.Scale);
226 // These are 32-bit even in 64-bit mode since RIP relative offset
229 Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp,
232 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
233 AM.Align, AM.Disp, AM.SymbolFlags);
235 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
236 else if (AM.JT != -1)
237 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
238 else if (AM.BlockAddr)
239 Disp = CurDAG->getBlockAddress(AM.BlockAddr, MVT::i32,
240 true, AM.SymbolFlags);
242 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
244 if (AM.Segment.getNode())
245 Segment = AM.Segment;
247 Segment = CurDAG->getRegister(0, MVT::i32);
250 /// getI8Imm - Return a target constant with the specified value, of type
252 inline SDValue getI8Imm(unsigned Imm) {
253 return CurDAG->getTargetConstant(Imm, MVT::i8);
256 /// getI16Imm - Return a target constant with the specified value, of type
258 inline SDValue getI16Imm(unsigned Imm) {
259 return CurDAG->getTargetConstant(Imm, MVT::i16);
262 /// getI32Imm - Return a target constant with the specified value, of type
264 inline SDValue getI32Imm(unsigned Imm) {
265 return CurDAG->getTargetConstant(Imm, MVT::i32);
268 /// getGlobalBaseReg - Return an SDNode that returns the value of
269 /// the global base register. Output instructions required to
270 /// initialize the global base register, if necessary.
272 SDNode *getGlobalBaseReg();
274 /// getTargetMachine - Return a reference to the TargetMachine, casted
275 /// to the target-specific type.
276 const X86TargetMachine &getTargetMachine() {
277 return static_cast<const X86TargetMachine &>(TM);
280 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
281 /// to the target-specific type.
282 const X86InstrInfo *getInstrInfo() {
283 return getTargetMachine().getInstrInfo();
290 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
291 if (OptLevel == CodeGenOpt::None) return false;
296 if (N.getOpcode() != ISD::LOAD)
299 // If N is a load, do additional profitability checks.
301 switch (U->getOpcode()) {
314 SDValue Op1 = U->getOperand(1);
316 // If the other operand is a 8-bit immediate we should fold the immediate
317 // instead. This reduces code size.
319 // movl 4(%esp), %eax
323 // addl 4(%esp), %eax
324 // The former is 2 bytes shorter. In case where the increment is 1, then
325 // the saving can be 4 bytes (by using incl %eax).
326 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
327 if (Imm->getAPIntValue().isSignedIntN(8))
330 // If the other operand is a TLS address, we should fold it instead.
333 // leal i@NTPOFF(%eax), %eax
335 // movl $i@NTPOFF, %eax
337 // if the block also has an access to a second TLS address this will save
339 // FIXME: This is probably also true for non TLS addresses.
340 if (Op1.getOpcode() == X86ISD::Wrapper) {
341 SDValue Val = Op1.getOperand(0);
342 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
352 /// MoveBelowCallSeqStart - Replace CALLSEQ_START operand with load's chain
353 /// operand and move load below the call's chain operand.
354 static void MoveBelowCallSeqStart(SelectionDAG *CurDAG, SDValue Load,
355 SDValue Call, SDValue CallSeqStart) {
356 SmallVector<SDValue, 8> Ops;
357 SDValue Chain = CallSeqStart.getOperand(0);
358 if (Chain.getNode() == Load.getNode())
359 Ops.push_back(Load.getOperand(0));
361 assert(Chain.getOpcode() == ISD::TokenFactor &&
362 "Unexpected CallSeqStart chain operand");
363 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
364 if (Chain.getOperand(i).getNode() == Load.getNode())
365 Ops.push_back(Load.getOperand(0));
367 Ops.push_back(Chain.getOperand(i));
369 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
370 MVT::Other, &Ops[0], Ops.size());
372 Ops.push_back(NewChain);
374 for (unsigned i = 1, e = CallSeqStart.getNumOperands(); i != e; ++i)
375 Ops.push_back(CallSeqStart.getOperand(i));
376 CurDAG->UpdateNodeOperands(CallSeqStart, &Ops[0], Ops.size());
377 CurDAG->UpdateNodeOperands(Load, Call.getOperand(0),
378 Load.getOperand(1), Load.getOperand(2));
380 Ops.push_back(SDValue(Load.getNode(), 1));
381 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
382 Ops.push_back(Call.getOperand(i));
383 CurDAG->UpdateNodeOperands(Call, &Ops[0], Ops.size());
386 /// isCalleeLoad - Return true if call address is a load and it can be
387 /// moved below CALLSEQ_START and the chains leading up to the call.
388 /// Return the CALLSEQ_START by reference as a second output.
389 static bool isCalleeLoad(SDValue Callee, SDValue &Chain) {
390 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
392 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
395 LD->getAddressingMode() != ISD::UNINDEXED ||
396 LD->getExtensionType() != ISD::NON_EXTLOAD)
399 // Now let's find the callseq_start.
400 while (Chain.getOpcode() != ISD::CALLSEQ_START) {
401 if (!Chain.hasOneUse())
403 Chain = Chain.getOperand(0);
406 if (Chain.getOperand(0).getNode() == Callee.getNode())
408 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
409 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
410 Callee.getValue(1).hasOneUse())
415 void X86DAGToDAGISel::PreprocessISelDAG() {
416 OptForSize = MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize);
418 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
419 E = CurDAG->allnodes_end(); I != E; ) {
420 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
422 if (OptLevel != CodeGenOpt::None && N->getOpcode() == X86ISD::CALL) {
423 /// Also try moving call address load from outside callseq_start to just
424 /// before the call to allow it to be folded.
442 SDValue Chain = N->getOperand(0);
443 SDValue Load = N->getOperand(1);
444 if (!isCalleeLoad(Load, Chain))
446 MoveBelowCallSeqStart(CurDAG, Load, SDValue(N, 0), Chain);
451 // Lower fpround and fpextend nodes that target the FP stack to be store and
452 // load to the stack. This is a gross hack. We would like to simply mark
453 // these as being illegal, but when we do that, legalize produces these when
454 // it expands calls, then expands these in the same legalize pass. We would
455 // like dag combine to be able to hack on these between the call expansion
456 // and the node legalization. As such this pass basically does "really
457 // late" legalization of these inline with the X86 isel pass.
458 // FIXME: This should only happen when not compiled with -O0.
459 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
462 // If the source and destination are SSE registers, then this is a legal
463 // conversion that should not be lowered.
464 EVT SrcVT = N->getOperand(0).getValueType();
465 EVT DstVT = N->getValueType(0);
466 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
467 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
468 if (SrcIsSSE && DstIsSSE)
471 if (!SrcIsSSE && !DstIsSSE) {
472 // If this is an FPStack extension, it is a noop.
473 if (N->getOpcode() == ISD::FP_EXTEND)
475 // If this is a value-preserving FPStack truncation, it is a noop.
476 if (N->getConstantOperandVal(1))
480 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
481 // FPStack has extload and truncstore. SSE can fold direct loads into other
482 // operations. Based on this, decide what we want to do.
484 if (N->getOpcode() == ISD::FP_ROUND)
485 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
487 MemVT = SrcIsSSE ? SrcVT : DstVT;
489 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
490 DebugLoc dl = N->getDebugLoc();
492 // FIXME: optimize the case where the src/dest is a load or store?
493 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
495 MemTmp, NULL, 0, MemVT,
497 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
498 NULL, 0, MemVT, false, false, 0);
500 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
501 // extload we created. This will cause general havok on the dag because
502 // anything below the conversion could be folded into other existing nodes.
503 // To avoid invalidating 'I', back it up to the convert node.
505 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
507 // Now that we did that, the node is dead. Increment the iterator to the
508 // next node to process, then delete N.
510 CurDAG->DeleteNode(N);
515 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
516 /// the main function.
517 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
518 MachineFrameInfo *MFI) {
519 const TargetInstrInfo *TII = TM.getInstrInfo();
520 if (Subtarget->isTargetCygMing())
521 BuildMI(BB, DebugLoc::getUnknownLoc(),
522 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
525 void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
526 // If this is main, emit special code for main.
527 MachineBasicBlock *BB = MF.begin();
528 if (Fn.hasExternalLinkage() && Fn.getName() == "main")
529 EmitSpecialCodeForMain(BB, MF.getFrameInfo());
533 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
534 X86ISelAddressMode &AM) {
535 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
536 SDValue Segment = N.getOperand(0);
538 if (AM.Segment.getNode() == 0) {
539 AM.Segment = Segment;
546 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
547 // This optimization is valid because the GNU TLS model defines that
548 // gs:0 (or fs:0 on X86-64) contains its own address.
549 // For more information see http://people.redhat.com/drepper/tls.pdf
551 SDValue Address = N.getOperand(1);
552 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
553 !MatchSegmentBaseAddress (Address, AM))
559 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
560 /// into an addressing mode. These wrap things that will resolve down into a
561 /// symbol reference. If no match is possible, this returns true, otherwise it
563 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
564 // If the addressing mode already has a symbol as the displacement, we can
565 // never match another symbol.
566 if (AM.hasSymbolicDisplacement())
569 SDValue N0 = N.getOperand(0);
570 CodeModel::Model M = TM.getCodeModel();
572 // Handle X86-64 rip-relative addresses. We check this before checking direct
573 // folding because RIP is preferable to non-RIP accesses.
574 if (Subtarget->is64Bit() &&
575 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
576 // they cannot be folded into immediate fields.
577 // FIXME: This can be improved for kernel and other models?
578 (M == CodeModel::Small || M == CodeModel::Kernel) &&
579 // Base and index reg must be 0 in order to use %rip as base and lowering
581 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
582 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
583 int64_t Offset = AM.Disp + G->getOffset();
584 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
585 AM.GV = G->getGlobal();
587 AM.SymbolFlags = G->getTargetFlags();
588 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
589 int64_t Offset = AM.Disp + CP->getOffset();
590 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
591 AM.CP = CP->getConstVal();
592 AM.Align = CP->getAlignment();
594 AM.SymbolFlags = CP->getTargetFlags();
595 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
596 AM.ES = S->getSymbol();
597 AM.SymbolFlags = S->getTargetFlags();
598 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
599 AM.JT = J->getIndex();
600 AM.SymbolFlags = J->getTargetFlags();
602 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
603 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
606 if (N.getOpcode() == X86ISD::WrapperRIP)
607 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
611 // Handle the case when globals fit in our immediate field: This is true for
612 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
613 // mode, this results in a non-RIP-relative computation.
614 if (!Subtarget->is64Bit() ||
615 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
616 TM.getRelocationModel() == Reloc::Static)) {
617 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
618 AM.GV = G->getGlobal();
619 AM.Disp += G->getOffset();
620 AM.SymbolFlags = G->getTargetFlags();
621 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
622 AM.CP = CP->getConstVal();
623 AM.Align = CP->getAlignment();
624 AM.Disp += CP->getOffset();
625 AM.SymbolFlags = CP->getTargetFlags();
626 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
627 AM.ES = S->getSymbol();
628 AM.SymbolFlags = S->getTargetFlags();
629 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
630 AM.JT = J->getIndex();
631 AM.SymbolFlags = J->getTargetFlags();
633 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
634 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
642 /// MatchAddress - Add the specified node to the specified addressing mode,
643 /// returning true if it cannot be done. This just pattern matches for the
645 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
646 if (MatchAddressRecursively(N, AM, 0))
649 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
650 // a smaller encoding and avoids a scaled-index.
652 AM.BaseType == X86ISelAddressMode::RegBase &&
653 AM.Base.Reg.getNode() == 0) {
654 AM.Base.Reg = AM.IndexReg;
658 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
659 // because it has a smaller encoding.
660 // TODO: Which other code models can use this?
661 if (TM.getCodeModel() == CodeModel::Small &&
662 Subtarget->is64Bit() &&
664 AM.BaseType == X86ISelAddressMode::RegBase &&
665 AM.Base.Reg.getNode() == 0 &&
666 AM.IndexReg.getNode() == 0 &&
667 AM.SymbolFlags == X86II::MO_NO_FLAG &&
668 AM.hasSymbolicDisplacement())
669 AM.Base.Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
674 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
676 bool is64Bit = Subtarget->is64Bit();
677 DebugLoc dl = N.getDebugLoc();
679 dbgs() << "MatchAddress: ";
684 return MatchAddressBase(N, AM);
686 CodeModel::Model M = TM.getCodeModel();
688 // If this is already a %rip relative address, we can only merge immediates
689 // into it. Instead of handling this in every case, we handle it here.
690 // RIP relative addressing: %rip + 32-bit displacement!
691 if (AM.isRIPRelative()) {
692 // FIXME: JumpTable and ExternalSymbol address currently don't like
693 // displacements. It isn't very important, but this should be fixed for
695 if (!AM.ES && AM.JT != -1) return true;
697 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
698 int64_t Val = AM.Disp + Cst->getSExtValue();
699 if (X86::isOffsetSuitableForCodeModel(Val, M,
700 AM.hasSymbolicDisplacement())) {
708 switch (N.getOpcode()) {
710 case ISD::Constant: {
711 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
713 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
714 AM.hasSymbolicDisplacement())) {
721 case X86ISD::SegmentBaseAddress:
722 if (!MatchSegmentBaseAddress(N, AM))
726 case X86ISD::Wrapper:
727 case X86ISD::WrapperRIP:
728 if (!MatchWrapper(N, AM))
733 if (!MatchLoad(N, AM))
737 case ISD::FrameIndex:
738 if (AM.BaseType == X86ISelAddressMode::RegBase
739 && AM.Base.Reg.getNode() == 0) {
740 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
741 AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
747 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
751 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
752 unsigned Val = CN->getZExtValue();
753 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
754 // that the base operand remains free for further matching. If
755 // the base doesn't end up getting used, a post-processing step
756 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
757 if (Val == 1 || Val == 2 || Val == 3) {
759 SDValue ShVal = N.getNode()->getOperand(0);
761 // Okay, we know that we have a scale by now. However, if the scaled
762 // value is an add of something and a constant, we can fold the
763 // constant into the disp field here.
764 if (ShVal.getNode()->getOpcode() == ISD::ADD &&
765 isa<ConstantSDNode>(ShVal.getNode()->getOperand(1))) {
766 AM.IndexReg = ShVal.getNode()->getOperand(0);
767 ConstantSDNode *AddVal =
768 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
769 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
771 X86::isOffsetSuitableForCodeModel(Disp, M,
772 AM.hasSymbolicDisplacement()))
786 // A mul_lohi where we need the low part can be folded as a plain multiply.
787 if (N.getResNo() != 0) break;
790 case X86ISD::MUL_IMM:
791 // X*[3,5,9] -> X+X*[2,4,8]
792 if (AM.BaseType == X86ISelAddressMode::RegBase &&
793 AM.Base.Reg.getNode() == 0 &&
794 AM.IndexReg.getNode() == 0) {
796 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
797 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
798 CN->getZExtValue() == 9) {
799 AM.Scale = unsigned(CN->getZExtValue())-1;
801 SDValue MulVal = N.getNode()->getOperand(0);
804 // Okay, we know that we have a scale by now. However, if the scaled
805 // value is an add of something and a constant, we can fold the
806 // constant into the disp field here.
807 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
808 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
809 Reg = MulVal.getNode()->getOperand(0);
810 ConstantSDNode *AddVal =
811 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
812 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
815 X86::isOffsetSuitableForCodeModel(Disp, M,
816 AM.hasSymbolicDisplacement()))
819 Reg = N.getNode()->getOperand(0);
821 Reg = N.getNode()->getOperand(0);
824 AM.IndexReg = AM.Base.Reg = Reg;
831 // Given A-B, if A can be completely folded into the address and
832 // the index field with the index field unused, use -B as the index.
833 // This is a win if a has multiple parts that can be folded into
834 // the address. Also, this saves a mov if the base register has
835 // other uses, since it avoids a two-address sub instruction, however
836 // it costs an additional mov if the index register has other uses.
838 // Test if the LHS of the sub can be folded.
839 X86ISelAddressMode Backup = AM;
840 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
844 // Test if the index field is free for use.
845 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
850 SDValue RHS = N.getNode()->getOperand(1);
851 // If the RHS involves a register with multiple uses, this
852 // transformation incurs an extra mov, due to the neg instruction
853 // clobbering its operand.
854 if (!RHS.getNode()->hasOneUse() ||
855 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
856 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
857 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
858 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
859 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
861 // If the base is a register with multiple uses, this
862 // transformation may save a mov.
863 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
864 AM.Base.Reg.getNode() &&
865 !AM.Base.Reg.getNode()->hasOneUse()) ||
866 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
868 // If the folded LHS was interesting, this transformation saves
869 // address arithmetic.
870 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
871 ((AM.Disp != 0) && (Backup.Disp == 0)) +
872 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
874 // If it doesn't look like it may be an overall win, don't do it.
880 // Ok, the transformation is legal and appears profitable. Go for it.
881 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
882 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
886 // Insert the new nodes into the topological ordering.
887 if (Zero.getNode()->getNodeId() == -1 ||
888 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
889 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
890 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
892 if (Neg.getNode()->getNodeId() == -1 ||
893 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
894 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
895 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
901 X86ISelAddressMode Backup = AM;
902 if (!MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1) &&
903 !MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1))
906 if (!MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1) &&
907 !MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1))
911 // If we couldn't fold both operands into the address at the same time,
912 // see if we can just put each operand into a register and fold at least
914 if (AM.BaseType == X86ISelAddressMode::RegBase &&
915 !AM.Base.Reg.getNode() &&
916 !AM.IndexReg.getNode()) {
917 AM.Base.Reg = N.getNode()->getOperand(0);
918 AM.IndexReg = N.getNode()->getOperand(1);
926 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
927 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
928 X86ISelAddressMode Backup = AM;
929 uint64_t Offset = CN->getSExtValue();
930 // Start with the LHS as an addr mode.
931 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
932 // Address could not have picked a GV address for the displacement.
934 // On x86-64, the resultant disp must fit in 32-bits.
936 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
937 AM.hasSymbolicDisplacement())) &&
938 // Check to see if the LHS & C is zero.
939 CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
948 // Perform some heroic transforms on an and of a constant-count shift
949 // with a constant to enable use of the scaled offset field.
951 SDValue Shift = N.getOperand(0);
952 if (Shift.getNumOperands() != 2) break;
954 // Scale must not be used already.
955 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
957 SDValue X = Shift.getOperand(0);
958 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
959 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
960 if (!C1 || !C2) break;
962 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
963 // allows us to convert the shift and and into an h-register extract and
965 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
966 unsigned ScaleLog = 8 - C1->getZExtValue();
967 if (ScaleLog > 0 && ScaleLog < 4 &&
968 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
969 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
970 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
971 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
973 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
975 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
976 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
979 // Insert the new nodes into the topological ordering.
980 if (Eight.getNode()->getNodeId() == -1 ||
981 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
982 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
983 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
985 if (Mask.getNode()->getNodeId() == -1 ||
986 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
987 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
988 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
990 if (Srl.getNode()->getNodeId() == -1 ||
991 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
992 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
993 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
995 if (And.getNode()->getNodeId() == -1 ||
996 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
997 CurDAG->RepositionNode(N.getNode(), And.getNode());
998 And.getNode()->setNodeId(N.getNode()->getNodeId());
1000 if (ShlCount.getNode()->getNodeId() == -1 ||
1001 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1002 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1003 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1005 if (Shl.getNode()->getNodeId() == -1 ||
1006 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1007 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1008 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1010 CurDAG->ReplaceAllUsesWith(N, Shl);
1012 AM.Scale = (1 << ScaleLog);
1017 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1018 // allows us to fold the shift into this addressing mode.
1019 if (Shift.getOpcode() != ISD::SHL) break;
1021 // Not likely to be profitable if either the AND or SHIFT node has more
1022 // than one use (unless all uses are for address computation). Besides,
1023 // isel mechanism requires their node ids to be reused.
1024 if (!N.hasOneUse() || !Shift.hasOneUse())
1027 // Verify that the shift amount is something we can fold.
1028 unsigned ShiftCst = C1->getZExtValue();
1029 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1032 // Get the new AND mask, this folds to a constant.
1033 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1034 SDValue(C2, 0), SDValue(C1, 0));
1035 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1037 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1038 NewAND, SDValue(C1, 0));
1040 // Insert the new nodes into the topological ordering.
1041 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1042 CurDAG->RepositionNode(X.getNode(), C1);
1043 C1->setNodeId(X.getNode()->getNodeId());
1045 if (NewANDMask.getNode()->getNodeId() == -1 ||
1046 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1047 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1048 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1050 if (NewAND.getNode()->getNodeId() == -1 ||
1051 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1052 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1053 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1055 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1056 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1057 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1058 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1061 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1063 AM.Scale = 1 << ShiftCst;
1064 AM.IndexReg = NewAND;
1069 return MatchAddressBase(N, AM);
1072 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1073 /// specified addressing mode without any further recursion.
1074 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1075 // Is the base register already occupied?
1076 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.getNode()) {
1077 // If so, check to see if the scale index register is set.
1078 if (AM.IndexReg.getNode() == 0) {
1084 // Otherwise, we cannot select it.
1088 // Default, generate it as a register.
1089 AM.BaseType = X86ISelAddressMode::RegBase;
1094 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1095 /// It returns the operands which make up the maximal addressing mode it can
1096 /// match by reference.
1097 bool X86DAGToDAGISel::SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
1098 SDValue &Scale, SDValue &Index,
1099 SDValue &Disp, SDValue &Segment) {
1100 X86ISelAddressMode AM;
1101 if (MatchAddress(N, AM))
1104 EVT VT = N.getValueType();
1105 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1106 if (!AM.Base.Reg.getNode())
1107 AM.Base.Reg = CurDAG->getRegister(0, VT);
1110 if (!AM.IndexReg.getNode())
1111 AM.IndexReg = CurDAG->getRegister(0, VT);
1113 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1117 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1118 /// match a load whose top elements are either undef or zeros. The load flavor
1119 /// is derived from the type of N, which is either v4f32 or v2f64.
1122 /// PatternChainNode: this is the matched node that has a chain input and
1124 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1125 SDValue N, SDValue &Base,
1126 SDValue &Scale, SDValue &Index,
1127 SDValue &Disp, SDValue &Segment,
1128 SDValue &PatternNodeWithChain) {
1129 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1130 PatternNodeWithChain = N.getOperand(0);
1131 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1132 PatternNodeWithChain.hasOneUse() &&
1133 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1134 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1135 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1136 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp,Segment))
1142 // Also handle the case where we explicitly require zeros in the top
1143 // elements. This is a vector shuffle from the zero vector.
1144 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1145 // Check to see if the top elements are all zeros (or bitcast of zeros).
1146 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1147 N.getOperand(0).getNode()->hasOneUse() &&
1148 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1149 N.getOperand(0).getOperand(0).hasOneUse() &&
1150 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1151 IsLegalToFold(N.getOperand(0), N.getNode(), Root)) {
1152 // Okay, this is a zero extending load. Fold it.
1153 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1154 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1156 PatternNodeWithChain = SDValue(LD, 0);
1163 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1164 /// mode it matches can be cost effectively emitted as an LEA instruction.
1165 bool X86DAGToDAGISel::SelectLEAAddr(SDNode *Op, SDValue N,
1166 SDValue &Base, SDValue &Scale,
1167 SDValue &Index, SDValue &Disp) {
1168 X86ISelAddressMode AM;
1170 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1172 SDValue Copy = AM.Segment;
1173 SDValue T = CurDAG->getRegister(0, MVT::i32);
1175 if (MatchAddress(N, AM))
1177 assert (T == AM.Segment);
1180 EVT VT = N.getValueType();
1181 unsigned Complexity = 0;
1182 if (AM.BaseType == X86ISelAddressMode::RegBase)
1183 if (AM.Base.Reg.getNode())
1186 AM.Base.Reg = CurDAG->getRegister(0, VT);
1187 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1190 if (AM.IndexReg.getNode())
1193 AM.IndexReg = CurDAG->getRegister(0, VT);
1195 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1200 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1201 // to a LEA. This is determined with some expermentation but is by no means
1202 // optimal (especially for code size consideration). LEA is nice because of
1203 // its three-address nature. Tweak the cost function again when we can run
1204 // convertToThreeAddress() at register allocation time.
1205 if (AM.hasSymbolicDisplacement()) {
1206 // For X86-64, we should always use lea to materialize RIP relative
1208 if (Subtarget->is64Bit())
1214 if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
1217 // If it isn't worth using an LEA, reject it.
1218 if (Complexity <= 2)
1222 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1226 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1227 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
1228 SDValue &Scale, SDValue &Index,
1230 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1231 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1233 X86ISelAddressMode AM;
1234 AM.GV = GA->getGlobal();
1235 AM.Disp += GA->getOffset();
1236 AM.Base.Reg = CurDAG->getRegister(0, N.getValueType());
1237 AM.SymbolFlags = GA->getTargetFlags();
1239 if (N.getValueType() == MVT::i32) {
1241 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1243 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1247 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1252 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1253 SDValue &Base, SDValue &Scale,
1254 SDValue &Index, SDValue &Disp,
1256 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1257 !IsProfitableToFold(N, P, P) ||
1258 !IsLegalToFold(N, P, P))
1261 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
1264 /// getGlobalBaseReg - Return an SDNode that returns the value of
1265 /// the global base register. Output instructions required to
1266 /// initialize the global base register, if necessary.
1268 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1269 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1270 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1273 static SDNode *FindCallStartFromCall(SDNode *Node) {
1274 if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
1275 assert(Node->getOperand(0).getValueType() == MVT::Other &&
1276 "Node doesn't have a token chain argument!");
1277 return FindCallStartFromCall(Node->getOperand(0).getNode());
1280 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1281 SDValue Chain = Node->getOperand(0);
1282 SDValue In1 = Node->getOperand(1);
1283 SDValue In2L = Node->getOperand(2);
1284 SDValue In2H = Node->getOperand(3);
1285 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1286 if (!SelectAddr(In1.getNode(), In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1288 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1289 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1290 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1291 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1292 MVT::i32, MVT::i32, MVT::Other, Ops,
1293 array_lengthof(Ops));
1294 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1298 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1299 if (Node->hasAnyUseOfValue(0))
1302 // Optimize common patterns for __sync_add_and_fetch and
1303 // __sync_sub_and_fetch where the result is not used. This allows us
1304 // to use "lock" version of add, sub, inc, dec instructions.
1305 // FIXME: Do not use special instructions but instead add the "lock"
1306 // prefix to the target node somehow. The extra information will then be
1307 // transferred to machine instruction and it denotes the prefix.
1308 SDValue Chain = Node->getOperand(0);
1309 SDValue Ptr = Node->getOperand(1);
1310 SDValue Val = Node->getOperand(2);
1311 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1312 if (!SelectAddr(Ptr.getNode(), Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1315 bool isInc = false, isDec = false, isSub = false, isCN = false;
1316 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1319 int64_t CNVal = CN->getSExtValue();
1322 else if (CNVal == -1)
1324 else if (CNVal >= 0)
1325 Val = CurDAG->getTargetConstant(CNVal, NVT);
1328 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1330 } else if (Val.hasOneUse() &&
1331 Val.getOpcode() == ISD::SUB &&
1332 X86::isZeroNode(Val.getOperand(0))) {
1334 Val = Val.getOperand(1);
1338 switch (NVT.getSimpleVT().SimpleTy) {
1342 Opc = X86::LOCK_INC8m;
1344 Opc = X86::LOCK_DEC8m;
1347 Opc = X86::LOCK_SUB8mi;
1349 Opc = X86::LOCK_SUB8mr;
1352 Opc = X86::LOCK_ADD8mi;
1354 Opc = X86::LOCK_ADD8mr;
1359 Opc = X86::LOCK_INC16m;
1361 Opc = X86::LOCK_DEC16m;
1364 if (Predicate_i16immSExt8(Val.getNode()))
1365 Opc = X86::LOCK_SUB16mi8;
1367 Opc = X86::LOCK_SUB16mi;
1369 Opc = X86::LOCK_SUB16mr;
1372 if (Predicate_i16immSExt8(Val.getNode()))
1373 Opc = X86::LOCK_ADD16mi8;
1375 Opc = X86::LOCK_ADD16mi;
1377 Opc = X86::LOCK_ADD16mr;
1382 Opc = X86::LOCK_INC32m;
1384 Opc = X86::LOCK_DEC32m;
1387 if (Predicate_i32immSExt8(Val.getNode()))
1388 Opc = X86::LOCK_SUB32mi8;
1390 Opc = X86::LOCK_SUB32mi;
1392 Opc = X86::LOCK_SUB32mr;
1395 if (Predicate_i32immSExt8(Val.getNode()))
1396 Opc = X86::LOCK_ADD32mi8;
1398 Opc = X86::LOCK_ADD32mi;
1400 Opc = X86::LOCK_ADD32mr;
1405 Opc = X86::LOCK_INC64m;
1407 Opc = X86::LOCK_DEC64m;
1409 Opc = X86::LOCK_SUB64mr;
1411 if (Predicate_i64immSExt8(Val.getNode()))
1412 Opc = X86::LOCK_SUB64mi8;
1413 else if (Predicate_i64immSExt32(Val.getNode()))
1414 Opc = X86::LOCK_SUB64mi32;
1417 Opc = X86::LOCK_ADD64mr;
1419 if (Predicate_i64immSExt8(Val.getNode()))
1420 Opc = X86::LOCK_ADD64mi8;
1421 else if (Predicate_i64immSExt32(Val.getNode()))
1422 Opc = X86::LOCK_ADD64mi32;
1428 DebugLoc dl = Node->getDebugLoc();
1429 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1431 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1432 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1433 if (isInc || isDec) {
1434 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1435 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1436 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1437 SDValue RetVals[] = { Undef, Ret };
1438 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1440 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1441 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1442 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1443 SDValue RetVals[] = { Undef, Ret };
1444 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1448 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1449 /// any uses which require the SF or OF bits to be accurate.
1450 static bool HasNoSignedComparisonUses(SDNode *N) {
1451 // Examine each user of the node.
1452 for (SDNode::use_iterator UI = N->use_begin(),
1453 UE = N->use_end(); UI != UE; ++UI) {
1454 // Only examine CopyToReg uses.
1455 if (UI->getOpcode() != ISD::CopyToReg)
1457 // Only examine CopyToReg uses that copy to EFLAGS.
1458 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1461 // Examine each user of the CopyToReg use.
1462 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1463 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1464 // Only examine the Flag result.
1465 if (FlagUI.getUse().getResNo() != 1) continue;
1466 // Anything unusual: assume conservatively.
1467 if (!FlagUI->isMachineOpcode()) return false;
1468 // Examine the opcode of the user.
1469 switch (FlagUI->getMachineOpcode()) {
1470 // These comparisons don't treat the most significant bit specially.
1471 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1472 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1473 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1474 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1475 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1476 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1477 case X86::CMOVA16rr: case X86::CMOVA16rm:
1478 case X86::CMOVA32rr: case X86::CMOVA32rm:
1479 case X86::CMOVA64rr: case X86::CMOVA64rm:
1480 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1481 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1482 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1483 case X86::CMOVB16rr: case X86::CMOVB16rm:
1484 case X86::CMOVB32rr: case X86::CMOVB32rm:
1485 case X86::CMOVB64rr: case X86::CMOVB64rm:
1486 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1487 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1488 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1489 case X86::CMOVE16rr: case X86::CMOVE16rm:
1490 case X86::CMOVE32rr: case X86::CMOVE32rm:
1491 case X86::CMOVE64rr: case X86::CMOVE64rm:
1492 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1493 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1494 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1495 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1496 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1497 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1498 case X86::CMOVP16rr: case X86::CMOVP16rm:
1499 case X86::CMOVP32rr: case X86::CMOVP32rm:
1500 case X86::CMOVP64rr: case X86::CMOVP64rm:
1502 // Anything else: assume conservatively.
1503 default: return false;
1510 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1511 EVT NVT = Node->getValueType(0);
1513 unsigned Opcode = Node->getOpcode();
1514 DebugLoc dl = Node->getDebugLoc();
1516 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
1518 if (Node->isMachineOpcode()) {
1519 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
1520 return NULL; // Already selected.
1525 case X86ISD::GlobalBaseReg:
1526 return getGlobalBaseReg();
1528 case X86ISD::ATOMOR64_DAG:
1529 return SelectAtomic64(Node, X86::ATOMOR6432);
1530 case X86ISD::ATOMXOR64_DAG:
1531 return SelectAtomic64(Node, X86::ATOMXOR6432);
1532 case X86ISD::ATOMADD64_DAG:
1533 return SelectAtomic64(Node, X86::ATOMADD6432);
1534 case X86ISD::ATOMSUB64_DAG:
1535 return SelectAtomic64(Node, X86::ATOMSUB6432);
1536 case X86ISD::ATOMNAND64_DAG:
1537 return SelectAtomic64(Node, X86::ATOMNAND6432);
1538 case X86ISD::ATOMAND64_DAG:
1539 return SelectAtomic64(Node, X86::ATOMAND6432);
1540 case X86ISD::ATOMSWAP64_DAG:
1541 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1543 case ISD::ATOMIC_LOAD_ADD: {
1544 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1550 case ISD::SMUL_LOHI:
1551 case ISD::UMUL_LOHI: {
1552 SDValue N0 = Node->getOperand(0);
1553 SDValue N1 = Node->getOperand(1);
1555 bool isSigned = Opcode == ISD::SMUL_LOHI;
1557 switch (NVT.getSimpleVT().SimpleTy) {
1558 default: llvm_unreachable("Unsupported VT!");
1559 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1560 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1561 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1562 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1565 switch (NVT.getSimpleVT().SimpleTy) {
1566 default: llvm_unreachable("Unsupported VT!");
1567 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1568 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1569 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1570 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1574 unsigned LoReg, HiReg;
1575 switch (NVT.getSimpleVT().SimpleTy) {
1576 default: llvm_unreachable("Unsupported VT!");
1577 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1578 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1579 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1580 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1583 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1584 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1585 // Multiply is commmutative.
1587 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1592 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1593 N0, SDValue()).getValue(1);
1596 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1599 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1600 array_lengthof(Ops));
1601 InFlag = SDValue(CNode, 1);
1602 // Update the chain.
1603 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1606 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1609 // Copy the low half of the result, if it is needed.
1610 if (!SDValue(Node, 0).use_empty()) {
1611 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1612 LoReg, NVT, InFlag);
1613 InFlag = Result.getValue(2);
1614 ReplaceUses(SDValue(Node, 0), Result);
1615 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1617 // Copy the high half of the result, if it is needed.
1618 if (!SDValue(Node, 1).use_empty()) {
1620 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1621 // Prevent use of AH in a REX instruction by referencing AX instead.
1622 // Shift it down 8 bits.
1623 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1624 X86::AX, MVT::i16, InFlag);
1625 InFlag = Result.getValue(2);
1626 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1628 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1629 // Then truncate it down to i8.
1630 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
1633 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1634 HiReg, NVT, InFlag);
1635 InFlag = Result.getValue(2);
1637 ReplaceUses(SDValue(Node, 1), Result);
1638 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1645 case ISD::UDIVREM: {
1646 SDValue N0 = Node->getOperand(0);
1647 SDValue N1 = Node->getOperand(1);
1649 bool isSigned = Opcode == ISD::SDIVREM;
1651 switch (NVT.getSimpleVT().SimpleTy) {
1652 default: llvm_unreachable("Unsupported VT!");
1653 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1654 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1655 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1656 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1659 switch (NVT.getSimpleVT().SimpleTy) {
1660 default: llvm_unreachable("Unsupported VT!");
1661 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1662 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1663 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1664 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1668 unsigned LoReg, HiReg, ClrReg;
1669 unsigned ClrOpcode, SExtOpcode;
1670 switch (NVT.getSimpleVT().SimpleTy) {
1671 default: llvm_unreachable("Unsupported VT!");
1673 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1675 SExtOpcode = X86::CBW;
1678 LoReg = X86::AX; HiReg = X86::DX;
1679 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1680 SExtOpcode = X86::CWD;
1683 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1684 ClrOpcode = X86::MOV32r0;
1685 SExtOpcode = X86::CDQ;
1688 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1689 ClrOpcode = X86::MOV64r0;
1690 SExtOpcode = X86::CQO;
1694 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1695 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1696 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1699 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1700 // Special case for div8, just use a move with zero extension to AX to
1701 // clear the upper 8 bits (AH).
1702 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1703 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1704 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1706 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1708 array_lengthof(Ops)), 0);
1709 Chain = Move.getValue(1);
1710 ReplaceUses(N0.getValue(1), Chain);
1713 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1714 Chain = CurDAG->getEntryNode();
1716 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1717 InFlag = Chain.getValue(1);
1720 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1721 LoReg, N0, SDValue()).getValue(1);
1722 if (isSigned && !signBitIsZero) {
1723 // Sign extend the low part into the high part.
1725 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1727 // Zero out the high part, effectively zero extending the input.
1729 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1730 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1731 ClrNode, InFlag).getValue(1);
1736 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1739 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1740 array_lengthof(Ops));
1741 InFlag = SDValue(CNode, 1);
1742 // Update the chain.
1743 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1746 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1749 // Copy the division (low) result, if it is needed.
1750 if (!SDValue(Node, 0).use_empty()) {
1751 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1752 LoReg, NVT, InFlag);
1753 InFlag = Result.getValue(2);
1754 ReplaceUses(SDValue(Node, 0), Result);
1755 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1757 // Copy the remainder (high) result, if it is needed.
1758 if (!SDValue(Node, 1).use_empty()) {
1760 if (HiReg == X86::AH && Subtarget->is64Bit()) {
1761 // Prevent use of AH in a REX instruction by referencing AX instead.
1762 // Shift it down 8 bits.
1763 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1764 X86::AX, MVT::i16, InFlag);
1765 InFlag = Result.getValue(2);
1766 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1768 CurDAG->getTargetConstant(8, MVT::i8)),
1770 // Then truncate it down to i8.
1771 Result = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
1774 Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1775 HiReg, NVT, InFlag);
1776 InFlag = Result.getValue(2);
1778 ReplaceUses(SDValue(Node, 1), Result);
1779 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1785 SDValue N0 = Node->getOperand(0);
1786 SDValue N1 = Node->getOperand(1);
1788 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
1789 // use a smaller encoding.
1790 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
1791 N0.getValueType() != MVT::i8 &&
1792 X86::isZeroNode(N1)) {
1793 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
1796 // For example, convert "testl %eax, $8" to "testb %al, $8"
1797 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
1798 (!(C->getZExtValue() & 0x80) ||
1799 HasNoSignedComparisonUses(Node))) {
1800 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
1801 SDValue Reg = N0.getNode()->getOperand(0);
1803 // On x86-32, only the ABCD registers have 8-bit subregisters.
1804 if (!Subtarget->is64Bit()) {
1805 TargetRegisterClass *TRC = 0;
1806 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1807 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1808 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1809 default: llvm_unreachable("Unsupported TEST operand type!");
1811 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1812 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1813 Reg.getValueType(), Reg, RC), 0);
1816 // Extract the l-register.
1817 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT, dl,
1821 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
1824 // For example, "testl %eax, $2048" to "testb %ah, $8".
1825 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
1826 (!(C->getZExtValue() & 0x8000) ||
1827 HasNoSignedComparisonUses(Node))) {
1828 // Shift the immediate right by 8 bits.
1829 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
1831 SDValue Reg = N0.getNode()->getOperand(0);
1833 // Put the value in an ABCD register.
1834 TargetRegisterClass *TRC = 0;
1835 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1836 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
1837 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1838 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1839 default: llvm_unreachable("Unsupported TEST operand type!");
1841 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1842 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1843 Reg.getValueType(), Reg, RC), 0);
1845 // Extract the h-register.
1846 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_8BIT_HI, dl,
1849 // Emit a testb. No special NOREX tricks are needed since there's
1850 // only one GPR operand!
1851 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
1852 Subreg, ShiftedImm);
1855 // For example, "testl %eax, $32776" to "testw %ax, $32776".
1856 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
1857 N0.getValueType() != MVT::i16 &&
1858 (!(C->getZExtValue() & 0x8000) ||
1859 HasNoSignedComparisonUses(Node))) {
1860 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
1861 SDValue Reg = N0.getNode()->getOperand(0);
1863 // Extract the 16-bit subregister.
1864 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_16BIT, dl,
1868 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
1871 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
1872 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
1873 N0.getValueType() == MVT::i64 &&
1874 (!(C->getZExtValue() & 0x80000000) ||
1875 HasNoSignedComparisonUses(Node))) {
1876 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
1877 SDValue Reg = N0.getNode()->getOperand(0);
1879 // Extract the 32-bit subregister.
1880 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::SUBREG_32BIT, dl,
1884 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
1891 SDNode *ResNode = SelectCode(Node);
1893 DEBUG(dbgs() << "=> ";
1894 if (ResNode == NULL || ResNode == Node)
1897 ResNode->dump(CurDAG);
1903 bool X86DAGToDAGISel::
1904 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1905 std::vector<SDValue> &OutOps) {
1906 SDValue Op0, Op1, Op2, Op3, Op4;
1907 switch (ConstraintCode) {
1908 case 'o': // offsetable ??
1909 case 'v': // not offsetable ??
1910 default: return true;
1912 if (!SelectAddr(Op.getNode(), Op, Op0, Op1, Op2, Op3, Op4))
1917 OutOps.push_back(Op0);
1918 OutOps.push_back(Op1);
1919 OutOps.push_back(Op2);
1920 OutOps.push_back(Op3);
1921 OutOps.push_back(Op4);
1925 /// createX86ISelDag - This pass converts a legalized DAG into a
1926 /// X86-specific DAG, ready for instruction scheduling.
1928 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
1929 llvm::CodeGenOpt::Level OptLevel) {
1930 return new X86DAGToDAGISel(TM, OptLevel);