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 "X86MachineFunctionInfo.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/Type.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CFG.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/ADT/Statistic.h"
42 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
44 //===----------------------------------------------------------------------===//
45 // Pattern Matcher Implementation
46 //===----------------------------------------------------------------------===//
49 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
50 /// SDValue's instead of register numbers for the leaves of the matched
52 struct X86ISelAddressMode {
58 // This is really a union, discriminated by BaseType!
66 const GlobalValue *GV;
68 const BlockAddress *BlockAddr;
71 unsigned Align; // CP alignment.
72 unsigned char SymbolFlags; // X86II::MO_*
75 : BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
76 Segment(), GV(0), CP(0), BlockAddr(0), ES(0), JT(-1), Align(0),
77 SymbolFlags(X86II::MO_NO_FLAG) {
80 bool hasSymbolicDisplacement() const {
81 return GV != 0 || CP != 0 || ES != 0 || JT != -1 || BlockAddr != 0;
84 bool hasBaseOrIndexReg() const {
85 return IndexReg.getNode() != 0 || Base_Reg.getNode() != 0;
88 /// isRIPRelative - Return true if this addressing mode is already RIP
90 bool isRIPRelative() const {
91 if (BaseType != RegBase) return false;
92 if (RegisterSDNode *RegNode =
93 dyn_cast_or_null<RegisterSDNode>(Base_Reg.getNode()))
94 return RegNode->getReg() == X86::RIP;
98 void setBaseReg(SDValue Reg) {
103 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
105 dbgs() << "X86ISelAddressMode " << this << '\n';
106 dbgs() << "Base_Reg ";
107 if (Base_Reg.getNode() != 0)
108 Base_Reg.getNode()->dump();
111 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
112 << " Scale" << Scale << '\n'
114 if (IndexReg.getNode() != 0)
115 IndexReg.getNode()->dump();
118 dbgs() << " Disp " << Disp << '\n'
135 dbgs() << " JT" << JT << " Align" << Align << '\n';
142 //===--------------------------------------------------------------------===//
143 /// ISel - X86 specific code to select X86 machine instructions for
144 /// SelectionDAG operations.
146 class X86DAGToDAGISel : public SelectionDAGISel {
147 /// X86Lowering - This object fully describes how to lower LLVM code to an
148 /// X86-specific SelectionDAG.
149 const X86TargetLowering &X86Lowering;
151 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
152 /// make the right decision when generating code for different targets.
153 const X86Subtarget *Subtarget;
155 /// OptForSize - If true, selector should try to optimize for code size
156 /// instead of performance.
160 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
161 : SelectionDAGISel(tm, OptLevel),
162 X86Lowering(*tm.getTargetLowering()),
163 Subtarget(&tm.getSubtarget<X86Subtarget>()),
166 virtual const char *getPassName() const {
167 return "X86 DAG->DAG Instruction Selection";
170 virtual void EmitFunctionEntryCode();
172 virtual bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const;
174 virtual void PreprocessISelDAG();
176 inline bool immSext8(SDNode *N) const {
177 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
180 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
181 // sign extended field.
182 inline bool i64immSExt32(SDNode *N) const {
183 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
184 return (int64_t)v == (int32_t)v;
187 // Include the pieces autogenerated from the target description.
188 #include "X86GenDAGISel.inc"
191 SDNode *Select(SDNode *N);
192 SDNode *SelectGather(SDNode *N, unsigned Opc);
193 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
194 SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
195 SDNode *SelectAtomicLoadArith(SDNode *Node, EVT NVT);
197 bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
198 bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
199 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
200 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
201 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
203 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
204 bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
205 SDValue &Scale, SDValue &Index, SDValue &Disp,
207 bool SelectSingleRegAddr(SDNode *Parent, SDValue N, SDValue &Base,
208 SDValue &Scale, SDValue &Index, SDValue &Disp,
210 bool SelectLEAAddr(SDValue N, SDValue &Base,
211 SDValue &Scale, SDValue &Index, SDValue &Disp,
213 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
214 SDValue &Scale, SDValue &Index, SDValue &Disp,
216 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
217 SDValue &Base, SDValue &Scale,
218 SDValue &Index, SDValue &Disp,
220 SDValue &NodeWithChain);
222 bool TryFoldLoad(SDNode *P, SDValue N,
223 SDValue &Base, SDValue &Scale,
224 SDValue &Index, SDValue &Disp,
227 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
228 /// inline asm expressions.
229 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
231 std::vector<SDValue> &OutOps);
233 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
235 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
236 SDValue &Scale, SDValue &Index,
237 SDValue &Disp, SDValue &Segment) {
238 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
239 CurDAG->getTargetFrameIndex(AM.Base_FrameIndex, TLI.getPointerTy()) :
241 Scale = getI8Imm(AM.Scale);
243 // These are 32-bit even in 64-bit mode since RIP relative offset
246 Disp = CurDAG->getTargetGlobalAddress(AM.GV, DebugLoc(),
250 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
251 AM.Align, AM.Disp, AM.SymbolFlags);
253 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
254 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
255 } else if (AM.JT != -1) {
256 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
257 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
258 } else if (AM.BlockAddr)
259 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
262 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
264 if (AM.Segment.getNode())
265 Segment = AM.Segment;
267 Segment = CurDAG->getRegister(0, MVT::i32);
270 /// getI8Imm - Return a target constant with the specified value, of type
272 inline SDValue getI8Imm(unsigned Imm) {
273 return CurDAG->getTargetConstant(Imm, MVT::i8);
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();
304 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
305 if (OptLevel == CodeGenOpt::None) return false;
310 if (N.getOpcode() != ISD::LOAD)
313 // If N is a load, do additional profitability checks.
315 switch (U->getOpcode()) {
328 SDValue Op1 = U->getOperand(1);
330 // If the other operand is a 8-bit immediate we should fold the immediate
331 // instead. This reduces code size.
333 // movl 4(%esp), %eax
337 // addl 4(%esp), %eax
338 // The former is 2 bytes shorter. In case where the increment is 1, then
339 // the saving can be 4 bytes (by using incl %eax).
340 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
341 if (Imm->getAPIntValue().isSignedIntN(8))
344 // If the other operand is a TLS address, we should fold it instead.
347 // leal i@NTPOFF(%eax), %eax
349 // movl $i@NTPOFF, %eax
351 // if the block also has an access to a second TLS address this will save
353 // FIXME: This is probably also true for non TLS addresses.
354 if (Op1.getOpcode() == X86ISD::Wrapper) {
355 SDValue Val = Op1.getOperand(0);
356 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
366 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
367 /// load's chain operand and move load below the call's chain operand.
368 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
369 SDValue Call, SDValue OrigChain) {
370 SmallVector<SDValue, 8> Ops;
371 SDValue Chain = OrigChain.getOperand(0);
372 if (Chain.getNode() == Load.getNode())
373 Ops.push_back(Load.getOperand(0));
375 assert(Chain.getOpcode() == ISD::TokenFactor &&
376 "Unexpected chain operand");
377 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
378 if (Chain.getOperand(i).getNode() == Load.getNode())
379 Ops.push_back(Load.getOperand(0));
381 Ops.push_back(Chain.getOperand(i));
383 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
384 MVT::Other, &Ops[0], Ops.size());
386 Ops.push_back(NewChain);
388 for (unsigned i = 1, e = OrigChain.getNumOperands(); i != e; ++i)
389 Ops.push_back(OrigChain.getOperand(i));
390 CurDAG->UpdateNodeOperands(OrigChain.getNode(), &Ops[0], Ops.size());
391 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
392 Load.getOperand(1), Load.getOperand(2));
394 Ops.push_back(SDValue(Load.getNode(), 1));
395 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
396 Ops.push_back(Call.getOperand(i));
397 CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], Ops.size());
400 /// isCalleeLoad - Return true if call address is a load and it can be
401 /// moved below CALLSEQ_START and the chains leading up to the call.
402 /// Return the CALLSEQ_START by reference as a second output.
403 /// In the case of a tail call, there isn't a callseq node between the call
404 /// chain and the load.
405 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
406 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
408 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
411 LD->getAddressingMode() != ISD::UNINDEXED ||
412 LD->getExtensionType() != ISD::NON_EXTLOAD)
415 // Now let's find the callseq_start.
416 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
417 if (!Chain.hasOneUse())
419 Chain = Chain.getOperand(0);
422 if (!Chain.getNumOperands())
424 if (Chain.getOperand(0).getNode() == Callee.getNode())
426 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
427 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
428 Callee.getValue(1).hasOneUse())
433 void X86DAGToDAGISel::PreprocessISelDAG() {
434 // OptForSize is used in pattern predicates that isel is matching.
435 OptForSize = MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize);
437 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
438 E = CurDAG->allnodes_end(); I != E; ) {
439 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
441 if (OptLevel != CodeGenOpt::None &&
442 (N->getOpcode() == X86ISD::CALL ||
443 N->getOpcode() == X86ISD::TC_RETURN)) {
444 /// Also try moving call address load from outside callseq_start to just
445 /// before the call to allow it to be folded.
463 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
464 SDValue Chain = N->getOperand(0);
465 SDValue Load = N->getOperand(1);
466 if (!isCalleeLoad(Load, Chain, HasCallSeq))
468 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
473 // Lower fpround and fpextend nodes that target the FP stack to be store and
474 // load to the stack. This is a gross hack. We would like to simply mark
475 // these as being illegal, but when we do that, legalize produces these when
476 // it expands calls, then expands these in the same legalize pass. We would
477 // like dag combine to be able to hack on these between the call expansion
478 // and the node legalization. As such this pass basically does "really
479 // late" legalization of these inline with the X86 isel pass.
480 // FIXME: This should only happen when not compiled with -O0.
481 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
484 EVT SrcVT = N->getOperand(0).getValueType();
485 EVT DstVT = N->getValueType(0);
487 // If any of the sources are vectors, no fp stack involved.
488 if (SrcVT.isVector() || DstVT.isVector())
491 // If the source and destination are SSE registers, then this is a legal
492 // conversion that should not be lowered.
493 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
494 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
495 if (SrcIsSSE && DstIsSSE)
498 if (!SrcIsSSE && !DstIsSSE) {
499 // If this is an FPStack extension, it is a noop.
500 if (N->getOpcode() == ISD::FP_EXTEND)
502 // If this is a value-preserving FPStack truncation, it is a noop.
503 if (N->getConstantOperandVal(1))
507 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
508 // FPStack has extload and truncstore. SSE can fold direct loads into other
509 // operations. Based on this, decide what we want to do.
511 if (N->getOpcode() == ISD::FP_ROUND)
512 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
514 MemVT = SrcIsSSE ? SrcVT : DstVT;
516 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
517 DebugLoc dl = N->getDebugLoc();
519 // FIXME: optimize the case where the src/dest is a load or store?
520 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
522 MemTmp, MachinePointerInfo(), MemVT,
524 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
525 MachinePointerInfo(),
526 MemVT, false, false, 0);
528 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
529 // extload we created. This will cause general havok on the dag because
530 // anything below the conversion could be folded into other existing nodes.
531 // To avoid invalidating 'I', back it up to the convert node.
533 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
535 // Now that we did that, the node is dead. Increment the iterator to the
536 // next node to process, then delete N.
538 CurDAG->DeleteNode(N);
543 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
544 /// the main function.
545 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
546 MachineFrameInfo *MFI) {
547 const TargetInstrInfo *TII = TM.getInstrInfo();
548 if (Subtarget->isTargetCygMing()) {
550 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
551 BuildMI(BB, DebugLoc(),
552 TII->get(CallOp)).addExternalSymbol("__main");
556 void X86DAGToDAGISel::EmitFunctionEntryCode() {
557 // If this is main, emit special code for main.
558 if (const Function *Fn = MF->getFunction())
559 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
560 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
563 static bool isDispSafeForFrameIndex(int64_t Val) {
564 // On 64-bit platforms, we can run into an issue where a frame index
565 // includes a displacement that, when added to the explicit displacement,
566 // will overflow the displacement field. Assuming that the frame index
567 // displacement fits into a 31-bit integer (which is only slightly more
568 // aggressive than the current fundamental assumption that it fits into
569 // a 32-bit integer), a 31-bit disp should always be safe.
570 return isInt<31>(Val);
573 bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
574 X86ISelAddressMode &AM) {
575 int64_t Val = AM.Disp + Offset;
576 CodeModel::Model M = TM.getCodeModel();
577 if (Subtarget->is64Bit()) {
578 if (!X86::isOffsetSuitableForCodeModel(Val, M,
579 AM.hasSymbolicDisplacement()))
581 // In addition to the checks required for a register base, check that
582 // we do not try to use an unsafe Disp with a frame index.
583 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
584 !isDispSafeForFrameIndex(Val))
592 bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
593 SDValue Address = N->getOperand(1);
595 // load gs:0 -> GS segment register.
596 // load fs:0 -> FS segment register.
598 // This optimization is valid because the GNU TLS model defines that
599 // gs:0 (or fs:0 on X86-64) contains its own address.
600 // For more information see http://people.redhat.com/drepper/tls.pdf
601 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
602 if (C->getSExtValue() == 0 && AM.Segment.getNode() == 0 &&
603 Subtarget->isTargetLinux())
604 switch (N->getPointerInfo().getAddrSpace()) {
606 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
609 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
616 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
617 /// into an addressing mode. These wrap things that will resolve down into a
618 /// symbol reference. If no match is possible, this returns true, otherwise it
620 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
621 // If the addressing mode already has a symbol as the displacement, we can
622 // never match another symbol.
623 if (AM.hasSymbolicDisplacement())
626 SDValue N0 = N.getOperand(0);
627 CodeModel::Model M = TM.getCodeModel();
629 // Handle X86-64 rip-relative addresses. We check this before checking direct
630 // folding because RIP is preferable to non-RIP accesses.
631 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
632 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
633 // they cannot be folded into immediate fields.
634 // FIXME: This can be improved for kernel and other models?
635 (M == CodeModel::Small || M == CodeModel::Kernel)) {
636 // Base and index reg must be 0 in order to use %rip as base.
637 if (AM.hasBaseOrIndexReg())
639 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
640 X86ISelAddressMode Backup = AM;
641 AM.GV = G->getGlobal();
642 AM.SymbolFlags = G->getTargetFlags();
643 if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
647 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
648 X86ISelAddressMode Backup = AM;
649 AM.CP = CP->getConstVal();
650 AM.Align = CP->getAlignment();
651 AM.SymbolFlags = CP->getTargetFlags();
652 if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
656 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
657 AM.ES = S->getSymbol();
658 AM.SymbolFlags = S->getTargetFlags();
659 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
660 AM.JT = J->getIndex();
661 AM.SymbolFlags = J->getTargetFlags();
662 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
663 X86ISelAddressMode Backup = AM;
664 AM.BlockAddr = BA->getBlockAddress();
665 AM.SymbolFlags = BA->getTargetFlags();
666 if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
671 llvm_unreachable("Unhandled symbol reference node.");
673 if (N.getOpcode() == X86ISD::WrapperRIP)
674 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
678 // Handle the case when globals fit in our immediate field: This is true for
679 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
680 // mode, this only applies to a non-RIP-relative computation.
681 if (!Subtarget->is64Bit() ||
682 M == CodeModel::Small || M == CodeModel::Kernel) {
683 assert(N.getOpcode() != X86ISD::WrapperRIP &&
684 "RIP-relative addressing already handled");
685 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
686 AM.GV = G->getGlobal();
687 AM.Disp += G->getOffset();
688 AM.SymbolFlags = G->getTargetFlags();
689 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
690 AM.CP = CP->getConstVal();
691 AM.Align = CP->getAlignment();
692 AM.Disp += CP->getOffset();
693 AM.SymbolFlags = CP->getTargetFlags();
694 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
695 AM.ES = S->getSymbol();
696 AM.SymbolFlags = S->getTargetFlags();
697 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
698 AM.JT = J->getIndex();
699 AM.SymbolFlags = J->getTargetFlags();
700 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
701 AM.BlockAddr = BA->getBlockAddress();
702 AM.Disp += BA->getOffset();
703 AM.SymbolFlags = BA->getTargetFlags();
705 llvm_unreachable("Unhandled symbol reference node.");
712 /// MatchAddress - Add the specified node to the specified addressing mode,
713 /// returning true if it cannot be done. This just pattern matches for the
715 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
716 if (MatchAddressRecursively(N, AM, 0))
719 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
720 // a smaller encoding and avoids a scaled-index.
722 AM.BaseType == X86ISelAddressMode::RegBase &&
723 AM.Base_Reg.getNode() == 0) {
724 AM.Base_Reg = AM.IndexReg;
728 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
729 // because it has a smaller encoding.
730 // TODO: Which other code models can use this?
731 if (TM.getCodeModel() == CodeModel::Small &&
732 Subtarget->is64Bit() &&
734 AM.BaseType == X86ISelAddressMode::RegBase &&
735 AM.Base_Reg.getNode() == 0 &&
736 AM.IndexReg.getNode() == 0 &&
737 AM.SymbolFlags == X86II::MO_NO_FLAG &&
738 AM.hasSymbolicDisplacement())
739 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
744 // Insert a node into the DAG at least before the Pos node's position. This
745 // will reposition the node as needed, and will assign it a node ID that is <=
746 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
747 // IDs! The selection DAG must no longer depend on their uniqueness when this
749 static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
750 if (N.getNode()->getNodeId() == -1 ||
751 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
752 DAG.RepositionNode(Pos.getNode(), N.getNode());
753 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
757 // Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if safe. This
758 // allows us to convert the shift and and into an h-register extract and
759 // a scaled index. Returns false if the simplification is performed.
760 static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
762 SDValue Shift, SDValue X,
763 X86ISelAddressMode &AM) {
764 if (Shift.getOpcode() != ISD::SRL ||
765 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
769 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
770 if (ScaleLog <= 0 || ScaleLog >= 4 ||
771 Mask != (0xffu << ScaleLog))
774 EVT VT = N.getValueType();
775 DebugLoc DL = N.getDebugLoc();
776 SDValue Eight = DAG.getConstant(8, MVT::i8);
777 SDValue NewMask = DAG.getConstant(0xff, VT);
778 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
779 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
780 SDValue ShlCount = DAG.getConstant(ScaleLog, MVT::i8);
781 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
783 // Insert the new nodes into the topological ordering. We must do this in
784 // a valid topological ordering as nothing is going to go back and re-sort
785 // these nodes. We continually insert before 'N' in sequence as this is
786 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
787 // hierarchy left to express.
788 InsertDAGNode(DAG, N, Eight);
789 InsertDAGNode(DAG, N, Srl);
790 InsertDAGNode(DAG, N, NewMask);
791 InsertDAGNode(DAG, N, And);
792 InsertDAGNode(DAG, N, ShlCount);
793 InsertDAGNode(DAG, N, Shl);
794 DAG.ReplaceAllUsesWith(N, Shl);
796 AM.Scale = (1 << ScaleLog);
800 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
801 // allows us to fold the shift into this addressing mode. Returns false if the
802 // transform succeeded.
803 static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
805 SDValue Shift, SDValue X,
806 X86ISelAddressMode &AM) {
807 if (Shift.getOpcode() != ISD::SHL ||
808 !isa<ConstantSDNode>(Shift.getOperand(1)))
811 // Not likely to be profitable if either the AND or SHIFT node has more
812 // than one use (unless all uses are for address computation). Besides,
813 // isel mechanism requires their node ids to be reused.
814 if (!N.hasOneUse() || !Shift.hasOneUse())
817 // Verify that the shift amount is something we can fold.
818 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
819 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
822 EVT VT = N.getValueType();
823 DebugLoc DL = N.getDebugLoc();
824 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
825 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
826 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
828 // Insert the new nodes into the topological ordering. We must do this in
829 // a valid topological ordering as nothing is going to go back and re-sort
830 // these nodes. We continually insert before 'N' in sequence as this is
831 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
832 // hierarchy left to express.
833 InsertDAGNode(DAG, N, NewMask);
834 InsertDAGNode(DAG, N, NewAnd);
835 InsertDAGNode(DAG, N, NewShift);
836 DAG.ReplaceAllUsesWith(N, NewShift);
838 AM.Scale = 1 << ShiftAmt;
839 AM.IndexReg = NewAnd;
843 // Implement some heroics to detect shifts of masked values where the mask can
844 // be replaced by extending the shift and undoing that in the addressing mode
845 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
846 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
847 // the addressing mode. This results in code such as:
849 // int f(short *y, int *lookup_table) {
851 // return *y + lookup_table[*y >> 11];
855 // movzwl (%rdi), %eax
858 // addl (%rsi,%rcx,4), %eax
861 // movzwl (%rdi), %eax
865 // addl (%rsi,%rcx), %eax
867 // Note that this function assumes the mask is provided as a mask *after* the
868 // value is shifted. The input chain may or may not match that, but computing
869 // such a mask is trivial.
870 static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
872 SDValue Shift, SDValue X,
873 X86ISelAddressMode &AM) {
874 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
875 !isa<ConstantSDNode>(Shift.getOperand(1)))
878 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
879 unsigned MaskLZ = CountLeadingZeros_64(Mask);
880 unsigned MaskTZ = CountTrailingZeros_64(Mask);
882 // The amount of shift we're trying to fit into the addressing mode is taken
883 // from the trailing zeros of the mask.
884 unsigned AMShiftAmt = MaskTZ;
886 // There is nothing we can do here unless the mask is removing some bits.
887 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
888 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
890 // We also need to ensure that mask is a continuous run of bits.
891 if (CountTrailingOnes_64(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
893 // Scale the leading zero count down based on the actual size of the value.
894 // Also scale it down based on the size of the shift.
895 MaskLZ -= (64 - X.getValueSizeInBits()) + ShiftAmt;
897 // The final check is to ensure that any masked out high bits of X are
898 // already known to be zero. Otherwise, the mask has a semantic impact
899 // other than masking out a couple of low bits. Unfortunately, because of
900 // the mask, zero extensions will be removed from operands in some cases.
901 // This code works extra hard to look through extensions because we can
902 // replace them with zero extensions cheaply if necessary.
903 bool ReplacingAnyExtend = false;
904 if (X.getOpcode() == ISD::ANY_EXTEND) {
905 unsigned ExtendBits =
906 X.getValueSizeInBits() - X.getOperand(0).getValueSizeInBits();
907 // Assume that we'll replace the any-extend with a zero-extend, and
908 // narrow the search to the extended value.
910 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
911 ReplacingAnyExtend = true;
913 APInt MaskedHighBits = APInt::getHighBitsSet(X.getValueSizeInBits(),
915 APInt KnownZero, KnownOne;
916 DAG.ComputeMaskedBits(X, KnownZero, KnownOne);
917 if (MaskedHighBits != KnownZero) return true;
919 // We've identified a pattern that can be transformed into a single shift
920 // and an addressing mode. Make it so.
921 EVT VT = N.getValueType();
922 if (ReplacingAnyExtend) {
923 assert(X.getValueType() != VT);
924 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
925 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, X.getDebugLoc(), VT, X);
926 InsertDAGNode(DAG, N, NewX);
929 DebugLoc DL = N.getDebugLoc();
930 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
931 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
932 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
933 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
935 // Insert the new nodes into the topological ordering. We must do this in
936 // a valid topological ordering as nothing is going to go back and re-sort
937 // these nodes. We continually insert before 'N' in sequence as this is
938 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
939 // hierarchy left to express.
940 InsertDAGNode(DAG, N, NewSRLAmt);
941 InsertDAGNode(DAG, N, NewSRL);
942 InsertDAGNode(DAG, N, NewSHLAmt);
943 InsertDAGNode(DAG, N, NewSHL);
944 DAG.ReplaceAllUsesWith(N, NewSHL);
946 AM.Scale = 1 << AMShiftAmt;
947 AM.IndexReg = NewSRL;
951 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
953 DebugLoc dl = N.getDebugLoc();
955 dbgs() << "MatchAddress: ";
960 return MatchAddressBase(N, AM);
962 // If this is already a %rip relative address, we can only merge immediates
963 // into it. Instead of handling this in every case, we handle it here.
964 // RIP relative addressing: %rip + 32-bit displacement!
965 if (AM.isRIPRelative()) {
966 // FIXME: JumpTable and ExternalSymbol address currently don't like
967 // displacements. It isn't very important, but this should be fixed for
969 if (!AM.ES && AM.JT != -1) return true;
971 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
972 if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
977 switch (N.getOpcode()) {
979 case ISD::Constant: {
980 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
981 if (!FoldOffsetIntoAddress(Val, AM))
986 case X86ISD::Wrapper:
987 case X86ISD::WrapperRIP:
988 if (!MatchWrapper(N, AM))
993 if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
997 case ISD::FrameIndex:
998 if (AM.BaseType == X86ISelAddressMode::RegBase &&
999 AM.Base_Reg.getNode() == 0 &&
1000 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1001 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1002 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1008 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
1012 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1013 unsigned Val = CN->getZExtValue();
1014 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1015 // that the base operand remains free for further matching. If
1016 // the base doesn't end up getting used, a post-processing step
1017 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1018 if (Val == 1 || Val == 2 || Val == 3) {
1019 AM.Scale = 1 << Val;
1020 SDValue ShVal = N.getNode()->getOperand(0);
1022 // Okay, we know that we have a scale by now. However, if the scaled
1023 // value is an add of something and a constant, we can fold the
1024 // constant into the disp field here.
1025 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1026 AM.IndexReg = ShVal.getNode()->getOperand(0);
1027 ConstantSDNode *AddVal =
1028 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1029 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1030 if (!FoldOffsetIntoAddress(Disp, AM))
1034 AM.IndexReg = ShVal;
1041 // Scale must not be used already.
1042 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1044 SDValue And = N.getOperand(0);
1045 if (And.getOpcode() != ISD::AND) break;
1046 SDValue X = And.getOperand(0);
1048 // We only handle up to 64-bit values here as those are what matter for
1049 // addressing mode optimizations.
1050 if (X.getValueSizeInBits() > 64) break;
1052 // The mask used for the transform is expected to be post-shift, but we
1053 // found the shift first so just apply the shift to the mask before passing
1055 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1056 !isa<ConstantSDNode>(And.getOperand(1)))
1058 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1060 // Try to fold the mask and shift into the scale, and return false if we
1062 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1067 case ISD::SMUL_LOHI:
1068 case ISD::UMUL_LOHI:
1069 // A mul_lohi where we need the low part can be folded as a plain multiply.
1070 if (N.getResNo() != 0) break;
1073 case X86ISD::MUL_IMM:
1074 // X*[3,5,9] -> X+X*[2,4,8]
1075 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1076 AM.Base_Reg.getNode() == 0 &&
1077 AM.IndexReg.getNode() == 0) {
1079 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1080 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1081 CN->getZExtValue() == 9) {
1082 AM.Scale = unsigned(CN->getZExtValue())-1;
1084 SDValue MulVal = N.getNode()->getOperand(0);
1087 // Okay, we know that we have a scale by now. However, if the scaled
1088 // value is an add of something and a constant, we can fold the
1089 // constant into the disp field here.
1090 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1091 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1092 Reg = MulVal.getNode()->getOperand(0);
1093 ConstantSDNode *AddVal =
1094 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1095 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1096 if (FoldOffsetIntoAddress(Disp, AM))
1097 Reg = N.getNode()->getOperand(0);
1099 Reg = N.getNode()->getOperand(0);
1102 AM.IndexReg = AM.Base_Reg = Reg;
1109 // Given A-B, if A can be completely folded into the address and
1110 // the index field with the index field unused, use -B as the index.
1111 // This is a win if a has multiple parts that can be folded into
1112 // the address. Also, this saves a mov if the base register has
1113 // other uses, since it avoids a two-address sub instruction, however
1114 // it costs an additional mov if the index register has other uses.
1116 // Add an artificial use to this node so that we can keep track of
1117 // it if it gets CSE'd with a different node.
1118 HandleSDNode Handle(N);
1120 // Test if the LHS of the sub can be folded.
1121 X86ISelAddressMode Backup = AM;
1122 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1126 // Test if the index field is free for use.
1127 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1133 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1134 // If the RHS involves a register with multiple uses, this
1135 // transformation incurs an extra mov, due to the neg instruction
1136 // clobbering its operand.
1137 if (!RHS.getNode()->hasOneUse() ||
1138 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1139 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1140 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1141 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1142 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1144 // If the base is a register with multiple uses, this
1145 // transformation may save a mov.
1146 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1147 AM.Base_Reg.getNode() &&
1148 !AM.Base_Reg.getNode()->hasOneUse()) ||
1149 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1151 // If the folded LHS was interesting, this transformation saves
1152 // address arithmetic.
1153 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1154 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1155 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1157 // If it doesn't look like it may be an overall win, don't do it.
1163 // Ok, the transformation is legal and appears profitable. Go for it.
1164 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
1165 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1169 // Insert the new nodes into the topological ordering.
1170 InsertDAGNode(*CurDAG, N, Zero);
1171 InsertDAGNode(*CurDAG, N, Neg);
1176 // Add an artificial use to this node so that we can keep track of
1177 // it if it gets CSE'd with a different node.
1178 HandleSDNode Handle(N);
1180 X86ISelAddressMode Backup = AM;
1181 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1182 !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1186 // Try again after commuting the operands.
1187 if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1188 !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1192 // If we couldn't fold both operands into the address at the same time,
1193 // see if we can just put each operand into a register and fold at least
1195 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1196 !AM.Base_Reg.getNode() &&
1197 !AM.IndexReg.getNode()) {
1198 N = Handle.getValue();
1199 AM.Base_Reg = N.getOperand(0);
1200 AM.IndexReg = N.getOperand(1);
1204 N = Handle.getValue();
1209 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1210 if (CurDAG->isBaseWithConstantOffset(N)) {
1211 X86ISelAddressMode Backup = AM;
1212 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1214 // Start with the LHS as an addr mode.
1215 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1216 !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
1223 // Perform some heroic transforms on an and of a constant-count shift
1224 // with a constant to enable use of the scaled offset field.
1226 // Scale must not be used already.
1227 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1229 SDValue Shift = N.getOperand(0);
1230 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1231 SDValue X = Shift.getOperand(0);
1233 // We only handle up to 64-bit values here as those are what matter for
1234 // addressing mode optimizations.
1235 if (X.getValueSizeInBits() > 64) break;
1237 if (!isa<ConstantSDNode>(N.getOperand(1)))
1239 uint64_t Mask = N.getConstantOperandVal(1);
1241 // Try to fold the mask and shift into an extract and scale.
1242 if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1245 // Try to fold the mask and shift directly into the scale.
1246 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1249 // Try to swap the mask and shift to place shifts which can be done as
1250 // a scale on the outside of the mask.
1251 if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1257 return MatchAddressBase(N, AM);
1260 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1261 /// specified addressing mode without any further recursion.
1262 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1263 // Is the base register already occupied?
1264 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1265 // If so, check to see if the scale index register is set.
1266 if (AM.IndexReg.getNode() == 0) {
1272 // Otherwise, we cannot select it.
1276 // Default, generate it as a register.
1277 AM.BaseType = X86ISelAddressMode::RegBase;
1282 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1283 /// It returns the operands which make up the maximal addressing mode it can
1284 /// match by reference.
1286 /// Parent is the parent node of the addr operand that is being matched. It
1287 /// is always a load, store, atomic node, or null. It is only null when
1288 /// checking memory operands for inline asm nodes.
1289 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1290 SDValue &Scale, SDValue &Index,
1291 SDValue &Disp, SDValue &Segment) {
1292 X86ISelAddressMode AM;
1295 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1296 // that are not a MemSDNode, and thus don't have proper addrspace info.
1297 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1298 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1299 Parent->getOpcode() != X86ISD::TLSCALL) { // Fixme
1300 unsigned AddrSpace =
1301 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1302 // AddrSpace 256 -> GS, 257 -> FS.
1303 if (AddrSpace == 256)
1304 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1305 if (AddrSpace == 257)
1306 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1309 if (MatchAddress(N, AM))
1312 EVT VT = N.getValueType();
1313 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1314 if (!AM.Base_Reg.getNode())
1315 AM.Base_Reg = CurDAG->getRegister(0, VT);
1318 if (!AM.IndexReg.getNode())
1319 AM.IndexReg = CurDAG->getRegister(0, VT);
1321 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1325 /// SelectSingleRegAddr - Like SelectAddr, but reject any address that would
1326 /// require more than one allocatable register.
1328 /// This is used for a TCRETURNmi64 instruction when used to tail call a
1329 /// variadic function with 6 arguments: Only %r11 is available from GR64_TC.
1330 /// The other scratch register, %rax, is needed to pass in the number of vector
1331 /// registers used in the variadic arguments.
1333 bool X86DAGToDAGISel::SelectSingleRegAddr(SDNode *Parent, SDValue N,
1335 SDValue &Scale, SDValue &Index,
1336 SDValue &Disp, SDValue &Segment) {
1337 if (!SelectAddr(Parent, N, Base, Scale, Index, Disp, Segment))
1339 // Anything %RIP relative is fine.
1340 if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Base))
1341 if (Reg->getReg() == X86::RIP)
1343 // Check that the index register is 0.
1344 if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Index))
1345 if (Reg->getReg() == 0)
1350 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1351 /// match a load whose top elements are either undef or zeros. The load flavor
1352 /// is derived from the type of N, which is either v4f32 or v2f64.
1355 /// PatternChainNode: this is the matched node that has a chain input and
1357 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1358 SDValue N, SDValue &Base,
1359 SDValue &Scale, SDValue &Index,
1360 SDValue &Disp, SDValue &Segment,
1361 SDValue &PatternNodeWithChain) {
1362 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1363 PatternNodeWithChain = N.getOperand(0);
1364 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1365 PatternNodeWithChain.hasOneUse() &&
1366 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1367 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1368 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1369 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1375 // Also handle the case where we explicitly require zeros in the top
1376 // elements. This is a vector shuffle from the zero vector.
1377 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1378 // Check to see if the top elements are all zeros (or bitcast of zeros).
1379 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1380 N.getOperand(0).getNode()->hasOneUse() &&
1381 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1382 N.getOperand(0).getOperand(0).hasOneUse() &&
1383 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1384 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1385 // Okay, this is a zero extending load. Fold it.
1386 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1387 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1389 PatternNodeWithChain = SDValue(LD, 0);
1396 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1397 /// mode it matches can be cost effectively emitted as an LEA instruction.
1398 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1399 SDValue &Base, SDValue &Scale,
1400 SDValue &Index, SDValue &Disp,
1402 X86ISelAddressMode AM;
1404 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1406 SDValue Copy = AM.Segment;
1407 SDValue T = CurDAG->getRegister(0, MVT::i32);
1409 if (MatchAddress(N, AM))
1411 assert (T == AM.Segment);
1414 EVT VT = N.getValueType();
1415 unsigned Complexity = 0;
1416 if (AM.BaseType == X86ISelAddressMode::RegBase)
1417 if (AM.Base_Reg.getNode())
1420 AM.Base_Reg = CurDAG->getRegister(0, VT);
1421 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1424 if (AM.IndexReg.getNode())
1427 AM.IndexReg = CurDAG->getRegister(0, VT);
1429 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1434 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1435 // to a LEA. This is determined with some expermentation but is by no means
1436 // optimal (especially for code size consideration). LEA is nice because of
1437 // its three-address nature. Tweak the cost function again when we can run
1438 // convertToThreeAddress() at register allocation time.
1439 if (AM.hasSymbolicDisplacement()) {
1440 // For X86-64, we should always use lea to materialize RIP relative
1442 if (Subtarget->is64Bit())
1448 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1451 // If it isn't worth using an LEA, reject it.
1452 if (Complexity <= 2)
1455 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1459 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1460 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1461 SDValue &Scale, SDValue &Index,
1462 SDValue &Disp, SDValue &Segment) {
1463 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1464 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1466 X86ISelAddressMode AM;
1467 AM.GV = GA->getGlobal();
1468 AM.Disp += GA->getOffset();
1469 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1470 AM.SymbolFlags = GA->getTargetFlags();
1472 if (N.getValueType() == MVT::i32) {
1474 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1476 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1479 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1484 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1485 SDValue &Base, SDValue &Scale,
1486 SDValue &Index, SDValue &Disp,
1488 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1489 !IsProfitableToFold(N, P, P) ||
1490 !IsLegalToFold(N, P, P, OptLevel))
1493 return SelectAddr(N.getNode(),
1494 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1497 /// getGlobalBaseReg - Return an SDNode that returns the value of
1498 /// the global base register. Output instructions required to
1499 /// initialize the global base register, if necessary.
1501 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1502 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1503 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1506 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1507 SDValue Chain = Node->getOperand(0);
1508 SDValue In1 = Node->getOperand(1);
1509 SDValue In2L = Node->getOperand(2);
1510 SDValue In2H = Node->getOperand(3);
1511 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1512 if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1514 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1515 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1516 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1517 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1518 MVT::i32, MVT::i32, MVT::Other, Ops,
1519 array_lengthof(Ops));
1520 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1524 // FIXME: Figure out some way to unify this with the 'or' and other code
1526 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1527 if (Node->hasAnyUseOfValue(0))
1530 // Optimize common patterns for __sync_add_and_fetch and
1531 // __sync_sub_and_fetch where the result is not used. This allows us
1532 // to use "lock" version of add, sub, inc, dec instructions.
1533 // FIXME: Do not use special instructions but instead add the "lock"
1534 // prefix to the target node somehow. The extra information will then be
1535 // transferred to machine instruction and it denotes the prefix.
1536 SDValue Chain = Node->getOperand(0);
1537 SDValue Ptr = Node->getOperand(1);
1538 SDValue Val = Node->getOperand(2);
1539 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1540 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1543 bool isInc = false, isDec = false, isSub = false, isCN = false;
1544 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1545 if (CN && CN->getSExtValue() == (int32_t)CN->getSExtValue()) {
1547 int64_t CNVal = CN->getSExtValue();
1550 else if (CNVal == -1)
1552 else if (CNVal >= 0)
1553 Val = CurDAG->getTargetConstant(CNVal, NVT);
1556 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1558 } else if (Val.hasOneUse() &&
1559 Val.getOpcode() == ISD::SUB &&
1560 X86::isZeroNode(Val.getOperand(0))) {
1562 Val = Val.getOperand(1);
1565 DebugLoc dl = Node->getDebugLoc();
1567 switch (NVT.getSimpleVT().SimpleTy) {
1571 Opc = X86::LOCK_INC8m;
1573 Opc = X86::LOCK_DEC8m;
1576 Opc = X86::LOCK_SUB8mi;
1578 Opc = X86::LOCK_SUB8mr;
1581 Opc = X86::LOCK_ADD8mi;
1583 Opc = X86::LOCK_ADD8mr;
1588 Opc = X86::LOCK_INC16m;
1590 Opc = X86::LOCK_DEC16m;
1593 if (immSext8(Val.getNode()))
1594 Opc = X86::LOCK_SUB16mi8;
1596 Opc = X86::LOCK_SUB16mi;
1598 Opc = X86::LOCK_SUB16mr;
1601 if (immSext8(Val.getNode()))
1602 Opc = X86::LOCK_ADD16mi8;
1604 Opc = X86::LOCK_ADD16mi;
1606 Opc = X86::LOCK_ADD16mr;
1611 Opc = X86::LOCK_INC32m;
1613 Opc = X86::LOCK_DEC32m;
1616 if (immSext8(Val.getNode()))
1617 Opc = X86::LOCK_SUB32mi8;
1619 Opc = X86::LOCK_SUB32mi;
1621 Opc = X86::LOCK_SUB32mr;
1624 if (immSext8(Val.getNode()))
1625 Opc = X86::LOCK_ADD32mi8;
1627 Opc = X86::LOCK_ADD32mi;
1629 Opc = X86::LOCK_ADD32mr;
1634 Opc = X86::LOCK_INC64m;
1636 Opc = X86::LOCK_DEC64m;
1638 Opc = X86::LOCK_SUB64mr;
1640 if (immSext8(Val.getNode()))
1641 Opc = X86::LOCK_SUB64mi8;
1642 else if (i64immSExt32(Val.getNode()))
1643 Opc = X86::LOCK_SUB64mi32;
1646 Opc = X86::LOCK_ADD64mr;
1648 if (immSext8(Val.getNode()))
1649 Opc = X86::LOCK_ADD64mi8;
1650 else if (i64immSExt32(Val.getNode()))
1651 Opc = X86::LOCK_ADD64mi32;
1657 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1659 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1660 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1661 if (isInc || isDec) {
1662 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1663 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1664 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1665 SDValue RetVals[] = { Undef, Ret };
1666 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1668 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1669 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1670 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1671 SDValue RetVals[] = { Undef, Ret };
1672 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1698 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1722 X86::LOCK_AND64mi32,
1735 X86::LOCK_XOR64mi32,
1740 SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, EVT NVT) {
1741 if (Node->hasAnyUseOfValue(0))
1744 // Optimize common patterns for __sync_or_and_fetch and similar arith
1745 // operations where the result is not used. This allows us to use the "lock"
1746 // version of the arithmetic instruction.
1747 // FIXME: Same as for 'add' and 'sub', try to merge those down here.
1748 SDValue Chain = Node->getOperand(0);
1749 SDValue Ptr = Node->getOperand(1);
1750 SDValue Val = Node->getOperand(2);
1751 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1752 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1755 // Which index into the table.
1757 switch (Node->getOpcode()) {
1758 case ISD::ATOMIC_LOAD_OR:
1761 case ISD::ATOMIC_LOAD_AND:
1764 case ISD::ATOMIC_LOAD_XOR:
1772 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1773 if (CN && (int32_t)CN->getSExtValue() == CN->getSExtValue()) {
1775 Val = CurDAG->getTargetConstant(CN->getSExtValue(), NVT);
1779 switch (NVT.getSimpleVT().SimpleTy) {
1783 Opc = AtomicOpcTbl[Op][ConstantI8];
1785 Opc = AtomicOpcTbl[Op][I8];
1789 if (immSext8(Val.getNode()))
1790 Opc = AtomicOpcTbl[Op][SextConstantI16];
1792 Opc = AtomicOpcTbl[Op][ConstantI16];
1794 Opc = AtomicOpcTbl[Op][I16];
1798 if (immSext8(Val.getNode()))
1799 Opc = AtomicOpcTbl[Op][SextConstantI32];
1801 Opc = AtomicOpcTbl[Op][ConstantI32];
1803 Opc = AtomicOpcTbl[Op][I32];
1806 Opc = AtomicOpcTbl[Op][I64];
1808 if (immSext8(Val.getNode()))
1809 Opc = AtomicOpcTbl[Op][SextConstantI64];
1810 else if (i64immSExt32(Val.getNode()))
1811 Opc = AtomicOpcTbl[Op][ConstantI64];
1816 assert(Opc != 0 && "Invalid arith lock transform!");
1818 DebugLoc dl = Node->getDebugLoc();
1819 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1821 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1822 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1823 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1824 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1825 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1826 SDValue RetVals[] = { Undef, Ret };
1827 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1830 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1831 /// any uses which require the SF or OF bits to be accurate.
1832 static bool HasNoSignedComparisonUses(SDNode *N) {
1833 // Examine each user of the node.
1834 for (SDNode::use_iterator UI = N->use_begin(),
1835 UE = N->use_end(); UI != UE; ++UI) {
1836 // Only examine CopyToReg uses.
1837 if (UI->getOpcode() != ISD::CopyToReg)
1839 // Only examine CopyToReg uses that copy to EFLAGS.
1840 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1843 // Examine each user of the CopyToReg use.
1844 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1845 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1846 // Only examine the Flag result.
1847 if (FlagUI.getUse().getResNo() != 1) continue;
1848 // Anything unusual: assume conservatively.
1849 if (!FlagUI->isMachineOpcode()) return false;
1850 // Examine the opcode of the user.
1851 switch (FlagUI->getMachineOpcode()) {
1852 // These comparisons don't treat the most significant bit specially.
1853 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1854 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1855 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1856 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1857 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1858 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1859 case X86::CMOVA16rr: case X86::CMOVA16rm:
1860 case X86::CMOVA32rr: case X86::CMOVA32rm:
1861 case X86::CMOVA64rr: case X86::CMOVA64rm:
1862 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1863 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1864 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1865 case X86::CMOVB16rr: case X86::CMOVB16rm:
1866 case X86::CMOVB32rr: case X86::CMOVB32rm:
1867 case X86::CMOVB64rr: case X86::CMOVB64rm:
1868 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1869 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1870 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1871 case X86::CMOVE16rr: case X86::CMOVE16rm:
1872 case X86::CMOVE32rr: case X86::CMOVE32rm:
1873 case X86::CMOVE64rr: case X86::CMOVE64rm:
1874 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1875 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1876 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1877 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1878 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1879 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1880 case X86::CMOVP16rr: case X86::CMOVP16rm:
1881 case X86::CMOVP32rr: case X86::CMOVP32rm:
1882 case X86::CMOVP64rr: case X86::CMOVP64rm:
1884 // Anything else: assume conservatively.
1885 default: return false;
1892 /// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
1893 /// is suitable for doing the {load; increment or decrement; store} to modify
1895 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
1896 SDValue StoredVal, SelectionDAG *CurDAG,
1897 LoadSDNode* &LoadNode, SDValue &InputChain) {
1899 // is the value stored the result of a DEC or INC?
1900 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
1902 // is the stored value result 0 of the load?
1903 if (StoredVal.getResNo() != 0) return false;
1905 // are there other uses of the loaded value than the inc or dec?
1906 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
1908 // is the store non-extending and non-indexed?
1909 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
1912 SDValue Load = StoredVal->getOperand(0);
1913 // Is the stored value a non-extending and non-indexed load?
1914 if (!ISD::isNormalLoad(Load.getNode())) return false;
1916 // Return LoadNode by reference.
1917 LoadNode = cast<LoadSDNode>(Load);
1918 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
1919 EVT LdVT = LoadNode->getMemoryVT();
1920 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
1924 // Is store the only read of the loaded value?
1925 if (!Load.hasOneUse())
1928 // Is the address of the store the same as the load?
1929 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
1930 LoadNode->getOffset() != StoreNode->getOffset())
1933 // Check if the chain is produced by the load or is a TokenFactor with
1934 // the load output chain as an operand. Return InputChain by reference.
1935 SDValue Chain = StoreNode->getChain();
1937 bool ChainCheck = false;
1938 if (Chain == Load.getValue(1)) {
1940 InputChain = LoadNode->getChain();
1941 } else if (Chain.getOpcode() == ISD::TokenFactor) {
1942 SmallVector<SDValue, 4> ChainOps;
1943 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
1944 SDValue Op = Chain.getOperand(i);
1945 if (Op == Load.getValue(1)) {
1950 // Make sure using Op as part of the chain would not cause a cycle here.
1951 // In theory, we could check whether the chain node is a predecessor of
1952 // the load. But that can be very expensive. Instead visit the uses and
1953 // make sure they all have smaller node id than the load.
1954 int LoadId = LoadNode->getNodeId();
1955 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
1956 UE = UI->use_end(); UI != UE; ++UI) {
1957 if (UI.getUse().getResNo() != 0)
1959 if (UI->getNodeId() > LoadId)
1963 ChainOps.push_back(Op);
1967 // Make a new TokenFactor with all the other input chains except
1969 InputChain = CurDAG->getNode(ISD::TokenFactor, Chain.getDebugLoc(),
1970 MVT::Other, &ChainOps[0], ChainOps.size());
1978 /// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
1979 /// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
1980 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
1981 if (Opc == X86ISD::DEC) {
1982 if (LdVT == MVT::i64) return X86::DEC64m;
1983 if (LdVT == MVT::i32) return X86::DEC32m;
1984 if (LdVT == MVT::i16) return X86::DEC16m;
1985 if (LdVT == MVT::i8) return X86::DEC8m;
1987 assert(Opc == X86ISD::INC && "unrecognized opcode");
1988 if (LdVT == MVT::i64) return X86::INC64m;
1989 if (LdVT == MVT::i32) return X86::INC32m;
1990 if (LdVT == MVT::i16) return X86::INC16m;
1991 if (LdVT == MVT::i8) return X86::INC8m;
1993 llvm_unreachable("unrecognized size for LdVT");
1996 /// SelectGather - Customized ISel for GATHER operations.
1998 SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
1999 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2000 SDValue Chain = Node->getOperand(0);
2001 SDValue VSrc = Node->getOperand(2);
2002 SDValue Base = Node->getOperand(3);
2003 SDValue VIdx = Node->getOperand(4);
2004 SDValue VMask = Node->getOperand(5);
2005 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2009 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2012 // Memory Operands: Base, Scale, Index, Disp, Segment
2013 SDValue Disp = CurDAG->getTargetConstant(0, MVT::i32);
2014 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2015 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue()), VIdx,
2016 Disp, Segment, VMask, Chain};
2017 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
2018 VTs, Ops, array_lengthof(Ops));
2019 // Node has 2 outputs: VDst and MVT::Other.
2020 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2021 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2023 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2024 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2028 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2029 EVT NVT = Node->getValueType(0);
2031 unsigned Opcode = Node->getOpcode();
2032 DebugLoc dl = Node->getDebugLoc();
2034 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2036 if (Node->isMachineOpcode()) {
2037 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2038 return NULL; // Already selected.
2043 case ISD::INTRINSIC_W_CHAIN: {
2044 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2047 case Intrinsic::x86_avx2_gather_d_pd:
2048 case Intrinsic::x86_avx2_gather_d_pd_256:
2049 case Intrinsic::x86_avx2_gather_q_pd:
2050 case Intrinsic::x86_avx2_gather_q_pd_256:
2051 case Intrinsic::x86_avx2_gather_d_ps:
2052 case Intrinsic::x86_avx2_gather_d_ps_256:
2053 case Intrinsic::x86_avx2_gather_q_ps:
2054 case Intrinsic::x86_avx2_gather_q_ps_256:
2055 case Intrinsic::x86_avx2_gather_d_q:
2056 case Intrinsic::x86_avx2_gather_d_q_256:
2057 case Intrinsic::x86_avx2_gather_q_q:
2058 case Intrinsic::x86_avx2_gather_q_q_256:
2059 case Intrinsic::x86_avx2_gather_d_d:
2060 case Intrinsic::x86_avx2_gather_d_d_256:
2061 case Intrinsic::x86_avx2_gather_q_d:
2062 case Intrinsic::x86_avx2_gather_q_d_256: {
2065 default: llvm_unreachable("Impossible intrinsic");
2066 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2067 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2068 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2069 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2070 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2071 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2072 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2073 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2074 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2075 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2076 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2077 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2078 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2079 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2080 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2081 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2083 SDNode *RetVal = SelectGather(Node, Opc);
2085 // We already called ReplaceUses inside SelectGather.
2092 case X86ISD::GlobalBaseReg:
2093 return getGlobalBaseReg();
2096 case X86ISD::ATOMOR64_DAG:
2097 case X86ISD::ATOMXOR64_DAG:
2098 case X86ISD::ATOMADD64_DAG:
2099 case X86ISD::ATOMSUB64_DAG:
2100 case X86ISD::ATOMNAND64_DAG:
2101 case X86ISD::ATOMAND64_DAG:
2102 case X86ISD::ATOMSWAP64_DAG: {
2105 default: llvm_unreachable("Impossible opcode");
2106 case X86ISD::ATOMOR64_DAG: Opc = X86::ATOMOR6432; break;
2107 case X86ISD::ATOMXOR64_DAG: Opc = X86::ATOMXOR6432; break;
2108 case X86ISD::ATOMADD64_DAG: Opc = X86::ATOMADD6432; break;
2109 case X86ISD::ATOMSUB64_DAG: Opc = X86::ATOMSUB6432; break;
2110 case X86ISD::ATOMNAND64_DAG: Opc = X86::ATOMNAND6432; break;
2111 case X86ISD::ATOMAND64_DAG: Opc = X86::ATOMAND6432; break;
2112 case X86ISD::ATOMSWAP64_DAG: Opc = X86::ATOMSWAP6432; break;
2114 SDNode *RetVal = SelectAtomic64(Node, Opc);
2120 case ISD::ATOMIC_LOAD_ADD: {
2121 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
2126 case ISD::ATOMIC_LOAD_XOR:
2127 case ISD::ATOMIC_LOAD_AND:
2128 case ISD::ATOMIC_LOAD_OR: {
2129 SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
2137 // For operations of the form (x << C1) op C2, check if we can use a smaller
2138 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2139 SDValue N0 = Node->getOperand(0);
2140 SDValue N1 = Node->getOperand(1);
2142 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2145 // i8 is unshrinkable, i16 should be promoted to i32.
2146 if (NVT != MVT::i32 && NVT != MVT::i64)
2149 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2150 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2151 if (!Cst || !ShlCst)
2154 int64_t Val = Cst->getSExtValue();
2155 uint64_t ShlVal = ShlCst->getZExtValue();
2157 // Make sure that we don't change the operation by removing bits.
2158 // This only matters for OR and XOR, AND is unaffected.
2159 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2160 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2166 // Check the minimum bitwidth for the new constant.
2167 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2168 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2169 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2170 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2172 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2175 // Bail if there is no smaller encoding.
2179 switch (NVT.getSimpleVT().SimpleTy) {
2180 default: llvm_unreachable("Unsupported VT!");
2182 assert(CstVT == MVT::i8);
2183 ShlOp = X86::SHL32ri;
2186 default: llvm_unreachable("Impossible opcode");
2187 case ISD::AND: Op = X86::AND32ri8; break;
2188 case ISD::OR: Op = X86::OR32ri8; break;
2189 case ISD::XOR: Op = X86::XOR32ri8; break;
2193 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2194 ShlOp = X86::SHL64ri;
2197 default: llvm_unreachable("Impossible opcode");
2198 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2199 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2200 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2205 // Emit the smaller op and the shift.
2206 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, CstVT);
2207 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2208 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2211 case X86ISD::UMUL: {
2212 SDValue N0 = Node->getOperand(0);
2213 SDValue N1 = Node->getOperand(1);
2216 switch (NVT.getSimpleVT().SimpleTy) {
2217 default: llvm_unreachable("Unsupported VT!");
2218 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2219 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2220 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2221 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2224 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2225 N0, SDValue()).getValue(1);
2227 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2228 SDValue Ops[] = {N1, InFlag};
2229 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops, 2);
2231 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2232 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2233 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2237 case ISD::SMUL_LOHI:
2238 case ISD::UMUL_LOHI: {
2239 SDValue N0 = Node->getOperand(0);
2240 SDValue N1 = Node->getOperand(1);
2242 bool isSigned = Opcode == ISD::SMUL_LOHI;
2244 switch (NVT.getSimpleVT().SimpleTy) {
2245 default: llvm_unreachable("Unsupported VT!");
2246 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2247 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2248 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
2249 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
2252 switch (NVT.getSimpleVT().SimpleTy) {
2253 default: llvm_unreachable("Unsupported VT!");
2254 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2255 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2256 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2257 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2261 unsigned LoReg, HiReg;
2262 switch (NVT.getSimpleVT().SimpleTy) {
2263 default: llvm_unreachable("Unsupported VT!");
2264 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
2265 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
2266 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
2267 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
2270 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2271 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2272 // Multiply is commmutative.
2274 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2279 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2280 N0, SDValue()).getValue(1);
2283 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2286 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops,
2287 array_lengthof(Ops));
2288 InFlag = SDValue(CNode, 1);
2290 // Update the chain.
2291 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2293 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag);
2294 InFlag = SDValue(CNode, 0);
2297 // Prevent use of AH in a REX instruction by referencing AX instead.
2298 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2299 !SDValue(Node, 1).use_empty()) {
2300 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2301 X86::AX, MVT::i16, InFlag);
2302 InFlag = Result.getValue(2);
2303 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2305 if (!SDValue(Node, 0).use_empty())
2306 ReplaceUses(SDValue(Node, 1),
2307 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2309 // Shift AX down 8 bits.
2310 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2312 CurDAG->getTargetConstant(8, MVT::i8)), 0);
2313 // Then truncate it down to i8.
2314 ReplaceUses(SDValue(Node, 1),
2315 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2317 // Copy the low half of the result, if it is needed.
2318 if (!SDValue(Node, 0).use_empty()) {
2319 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2320 LoReg, NVT, InFlag);
2321 InFlag = Result.getValue(2);
2322 ReplaceUses(SDValue(Node, 0), Result);
2323 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2325 // Copy the high half of the result, if it is needed.
2326 if (!SDValue(Node, 1).use_empty()) {
2327 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2328 HiReg, NVT, InFlag);
2329 InFlag = Result.getValue(2);
2330 ReplaceUses(SDValue(Node, 1), Result);
2331 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2338 case ISD::UDIVREM: {
2339 SDValue N0 = Node->getOperand(0);
2340 SDValue N1 = Node->getOperand(1);
2342 bool isSigned = Opcode == ISD::SDIVREM;
2344 switch (NVT.getSimpleVT().SimpleTy) {
2345 default: llvm_unreachable("Unsupported VT!");
2346 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2347 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2348 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2349 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2352 switch (NVT.getSimpleVT().SimpleTy) {
2353 default: llvm_unreachable("Unsupported VT!");
2354 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2355 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2356 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2357 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2361 unsigned LoReg, HiReg, ClrReg;
2362 unsigned ClrOpcode, SExtOpcode;
2363 switch (NVT.getSimpleVT().SimpleTy) {
2364 default: llvm_unreachable("Unsupported VT!");
2366 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2368 SExtOpcode = X86::CBW;
2371 LoReg = X86::AX; HiReg = X86::DX;
2372 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
2373 SExtOpcode = X86::CWD;
2376 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2377 ClrOpcode = X86::MOV32r0;
2378 SExtOpcode = X86::CDQ;
2381 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2382 ClrOpcode = X86::MOV64r0;
2383 SExtOpcode = X86::CQO;
2387 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2388 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2389 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2392 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2393 // Special case for div8, just use a move with zero extension to AX to
2394 // clear the upper 8 bits (AH).
2395 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2396 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2397 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2399 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2401 array_lengthof(Ops)), 0);
2402 Chain = Move.getValue(1);
2403 ReplaceUses(N0.getValue(1), Chain);
2406 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2407 Chain = CurDAG->getEntryNode();
2409 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2410 InFlag = Chain.getValue(1);
2413 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2414 LoReg, N0, SDValue()).getValue(1);
2415 if (isSigned && !signBitIsZero) {
2416 // Sign extend the low part into the high part.
2418 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2420 // Zero out the high part, effectively zero extending the input.
2422 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
2423 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2424 ClrNode, InFlag).getValue(1);
2429 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2432 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops,
2433 array_lengthof(Ops));
2434 InFlag = SDValue(CNode, 1);
2435 // Update the chain.
2436 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2439 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2442 // Prevent use of AH in a REX instruction by referencing AX instead.
2443 // Shift it down 8 bits.
2444 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2445 !SDValue(Node, 1).use_empty()) {
2446 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2447 X86::AX, MVT::i16, InFlag);
2448 InFlag = Result.getValue(2);
2450 // If we also need AL (the quotient), get it by extracting a subreg from
2451 // Result. The fast register allocator does not like multiple CopyFromReg
2452 // nodes using aliasing registers.
2453 if (!SDValue(Node, 0).use_empty())
2454 ReplaceUses(SDValue(Node, 0),
2455 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2457 // Shift AX right by 8 bits instead of using AH.
2458 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2460 CurDAG->getTargetConstant(8, MVT::i8)),
2462 ReplaceUses(SDValue(Node, 1),
2463 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2465 // Copy the division (low) result, if it is needed.
2466 if (!SDValue(Node, 0).use_empty()) {
2467 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2468 LoReg, NVT, InFlag);
2469 InFlag = Result.getValue(2);
2470 ReplaceUses(SDValue(Node, 0), Result);
2471 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2473 // Copy the remainder (high) result, if it is needed.
2474 if (!SDValue(Node, 1).use_empty()) {
2475 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2476 HiReg, NVT, InFlag);
2477 InFlag = Result.getValue(2);
2478 ReplaceUses(SDValue(Node, 1), Result);
2479 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2486 // Sometimes a SUB is used to perform comparison.
2487 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2488 // This node is not a CMP.
2490 SDValue N0 = Node->getOperand(0);
2491 SDValue N1 = Node->getOperand(1);
2493 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2494 // use a smaller encoding.
2495 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2496 HasNoSignedComparisonUses(Node))
2497 // Look past the truncate if CMP is the only use of it.
2498 N0 = N0.getOperand(0);
2499 if ((N0.getNode()->getOpcode() == ISD::AND ||
2500 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2501 N0.getNode()->hasOneUse() &&
2502 N0.getValueType() != MVT::i8 &&
2503 X86::isZeroNode(N1)) {
2504 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2507 // For example, convert "testl %eax, $8" to "testb %al, $8"
2508 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2509 (!(C->getZExtValue() & 0x80) ||
2510 HasNoSignedComparisonUses(Node))) {
2511 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
2512 SDValue Reg = N0.getNode()->getOperand(0);
2514 // On x86-32, only the ABCD registers have 8-bit subregisters.
2515 if (!Subtarget->is64Bit()) {
2516 const TargetRegisterClass *TRC;
2517 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2518 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2519 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2520 default: llvm_unreachable("Unsupported TEST operand type!");
2522 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2523 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2524 Reg.getValueType(), Reg, RC), 0);
2527 // Extract the l-register.
2528 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2532 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
2535 // For example, "testl %eax, $2048" to "testb %ah, $8".
2536 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2537 (!(C->getZExtValue() & 0x8000) ||
2538 HasNoSignedComparisonUses(Node))) {
2539 // Shift the immediate right by 8 bits.
2540 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2542 SDValue Reg = N0.getNode()->getOperand(0);
2544 // Put the value in an ABCD register.
2545 const TargetRegisterClass *TRC;
2546 switch (N0.getValueType().getSimpleVT().SimpleTy) {
2547 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2548 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2549 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2550 default: llvm_unreachable("Unsupported TEST operand type!");
2552 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2553 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2554 Reg.getValueType(), Reg, RC), 0);
2556 // Extract the h-register.
2557 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2560 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2561 // target GR8_NOREX registers, so make sure the register class is
2563 return CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl, MVT::i32,
2564 Subreg, ShiftedImm);
2567 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2568 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2569 N0.getValueType() != MVT::i16 &&
2570 (!(C->getZExtValue() & 0x8000) ||
2571 HasNoSignedComparisonUses(Node))) {
2572 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
2573 SDValue Reg = N0.getNode()->getOperand(0);
2575 // Extract the 16-bit subregister.
2576 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2580 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
2583 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2584 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2585 N0.getValueType() == MVT::i64 &&
2586 (!(C->getZExtValue() & 0x80000000) ||
2587 HasNoSignedComparisonUses(Node))) {
2588 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
2589 SDValue Reg = N0.getNode()->getOperand(0);
2591 // Extract the 32-bit subregister.
2592 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2596 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
2602 // Change a chain of {load; incr or dec; store} of the same value into
2603 // a simple increment or decrement through memory of that value, if the
2604 // uses of the modified value and its address are suitable.
2605 // The DEC64m tablegen pattern is currently not able to match the case where
2606 // the EFLAGS on the original DEC are used. (This also applies to
2607 // {INC,DEC}X{64,32,16,8}.)
2608 // We'll need to improve tablegen to allow flags to be transferred from a
2609 // node in the pattern to the result node. probably with a new keyword
2610 // for example, we have this
2611 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2612 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2613 // (implicit EFLAGS)]>;
2614 // but maybe need something like this
2615 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2616 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2617 // (transferrable EFLAGS)]>;
2619 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2620 SDValue StoredVal = StoreNode->getOperand(1);
2621 unsigned Opc = StoredVal->getOpcode();
2623 LoadSDNode *LoadNode = 0;
2625 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2626 LoadNode, InputChain))
2629 SDValue Base, Scale, Index, Disp, Segment;
2630 if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
2631 Base, Scale, Index, Disp, Segment))
2634 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2635 MemOp[0] = StoreNode->getMemOperand();
2636 MemOp[1] = LoadNode->getMemOperand();
2637 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2638 EVT LdVT = LoadNode->getMemoryVT();
2639 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2640 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2641 Node->getDebugLoc(),
2642 MVT::i32, MVT::Other, Ops,
2643 array_lengthof(Ops));
2644 Result->setMemRefs(MemOp, MemOp + 2);
2646 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2647 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2652 // FIXME: Custom handling because TableGen doesn't support multiple implicit
2653 // defs in an instruction pattern
2654 case X86ISD::PCMPESTRI: {
2655 SDValue N0 = Node->getOperand(0);
2656 SDValue N1 = Node->getOperand(1);
2657 SDValue N2 = Node->getOperand(2);
2658 SDValue N3 = Node->getOperand(3);
2659 SDValue N4 = Node->getOperand(4);
2661 // Make sure last argument is a constant
2662 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N4);
2666 uint64_t Imm = Cst->getZExtValue();
2668 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2669 X86::EAX, N1, SDValue()).getValue(1);
2670 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::EDX,
2671 N3, InFlag).getValue(1);
2673 SDValue Ops[] = { N0, N2, getI8Imm(Imm), InFlag };
2674 unsigned Opc = Subtarget->hasAVX() ? X86::VPCMPESTRIrr :
2676 InFlag = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Ops,
2677 array_lengthof(Ops)), 0);
2679 if (!SDValue(Node, 0).use_empty()) {
2680 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2681 X86::ECX, NVT, InFlag);
2682 InFlag = Result.getValue(2);
2683 ReplaceUses(SDValue(Node, 0), Result);
2685 if (!SDValue(Node, 1).use_empty()) {
2686 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2687 X86::EFLAGS, NVT, InFlag);
2688 InFlag = Result.getValue(2);
2689 ReplaceUses(SDValue(Node, 1), Result);
2695 // FIXME: Custom handling because TableGen doesn't support multiple implicit
2696 // defs in an instruction pattern
2697 case X86ISD::PCMPISTRI: {
2698 SDValue N0 = Node->getOperand(0);
2699 SDValue N1 = Node->getOperand(1);
2700 SDValue N2 = Node->getOperand(2);
2702 // Make sure last argument is a constant
2703 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N2);
2707 uint64_t Imm = Cst->getZExtValue();
2709 SDValue Ops[] = { N0, N1, getI8Imm(Imm) };
2710 unsigned Opc = Subtarget->hasAVX() ? X86::VPCMPISTRIrr :
2712 SDValue InFlag = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Ops,
2713 array_lengthof(Ops)), 0);
2715 if (!SDValue(Node, 0).use_empty()) {
2716 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2717 X86::ECX, NVT, InFlag);
2718 InFlag = Result.getValue(2);
2719 ReplaceUses(SDValue(Node, 0), Result);
2721 if (!SDValue(Node, 1).use_empty()) {
2722 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2723 X86::EFLAGS, NVT, InFlag);
2724 InFlag = Result.getValue(2);
2725 ReplaceUses(SDValue(Node, 1), Result);
2732 SDNode *ResNode = SelectCode(Node);
2734 DEBUG(dbgs() << "=> ";
2735 if (ResNode == NULL || ResNode == Node)
2738 ResNode->dump(CurDAG);
2744 bool X86DAGToDAGISel::
2745 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2746 std::vector<SDValue> &OutOps) {
2747 SDValue Op0, Op1, Op2, Op3, Op4;
2748 switch (ConstraintCode) {
2749 case 'o': // offsetable ??
2750 case 'v': // not offsetable ??
2751 default: return true;
2753 if (!SelectAddr(0, Op, Op0, Op1, Op2, Op3, Op4))
2758 OutOps.push_back(Op0);
2759 OutOps.push_back(Op1);
2760 OutOps.push_back(Op2);
2761 OutOps.push_back(Op3);
2762 OutOps.push_back(Op4);
2766 /// createX86ISelDag - This pass converts a legalized DAG into a
2767 /// X86-specific DAG, ready for instruction scheduling.
2769 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2770 CodeGenOpt::Level OptLevel) {
2771 return new X86DAGToDAGISel(TM, OptLevel);