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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86RegisterInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetOptions.h"
40 #define DEBUG_TYPE "x86-isel"
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(nullptr), CP(nullptr), BlockAddr(nullptr), ES(nullptr),
77 JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {
80 bool hasSymbolicDisplacement() const {
81 return GV != nullptr || CP != nullptr || ES != nullptr ||
82 JT != -1 || BlockAddr != nullptr;
85 bool hasBaseOrIndexReg() const {
86 return BaseType == FrameIndexBase ||
87 IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
90 /// isRIPRelative - Return true if this addressing mode is already RIP
92 bool isRIPRelative() const {
93 if (BaseType != RegBase) return false;
94 if (RegisterSDNode *RegNode =
95 dyn_cast_or_null<RegisterSDNode>(Base_Reg.getNode()))
96 return RegNode->getReg() == X86::RIP;
100 void setBaseReg(SDValue Reg) {
105 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
107 dbgs() << "X86ISelAddressMode " << this << '\n';
108 dbgs() << "Base_Reg ";
109 if (Base_Reg.getNode())
110 Base_Reg.getNode()->dump();
113 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
114 << " Scale" << Scale << '\n'
116 if (IndexReg.getNode())
117 IndexReg.getNode()->dump();
120 dbgs() << " Disp " << Disp << '\n'
137 dbgs() << " JT" << JT << " Align" << Align << '\n';
144 //===--------------------------------------------------------------------===//
145 /// ISel - X86 specific code to select X86 machine instructions for
146 /// SelectionDAG operations.
148 class X86DAGToDAGISel final : public SelectionDAGISel {
149 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
150 /// make the right decision when generating code for different targets.
151 const X86Subtarget *Subtarget;
153 /// OptForSize - If true, selector should try to optimize for code size
154 /// instead of performance.
158 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
159 : SelectionDAGISel(tm, OptLevel), OptForSize(false) {}
161 const char *getPassName() const override {
162 return "X86 DAG->DAG Instruction Selection";
165 bool runOnMachineFunction(MachineFunction &MF) override {
166 // Reset the subtarget each time through.
167 Subtarget = &MF.getSubtarget<X86Subtarget>();
168 SelectionDAGISel::runOnMachineFunction(MF);
172 void EmitFunctionEntryCode() override;
174 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
176 void PreprocessISelDAG() override;
178 inline bool immSext8(SDNode *N) const {
179 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
182 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
183 // sign extended field.
184 inline bool i64immSExt32(SDNode *N) const {
185 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
186 return (int64_t)v == (int32_t)v;
189 // Include the pieces autogenerated from the target description.
190 #include "X86GenDAGISel.inc"
193 SDNode *Select(SDNode *N) override;
194 SDNode *SelectGather(SDNode *N, unsigned Opc);
195 SDNode *SelectAtomicLoadArith(SDNode *Node, MVT 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 SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
208 SDValue &Scale, SDValue &Index, SDValue &Disp,
210 bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
211 bool SelectLEAAddr(SDValue N, SDValue &Base,
212 SDValue &Scale, SDValue &Index, SDValue &Disp,
214 bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
215 SDValue &Scale, SDValue &Index, SDValue &Disp,
217 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
218 SDValue &Scale, SDValue &Index, SDValue &Disp,
220 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
221 SDValue &Base, SDValue &Scale,
222 SDValue &Index, SDValue &Disp,
224 SDValue &NodeWithChain);
226 bool TryFoldLoad(SDNode *P, SDValue N,
227 SDValue &Base, SDValue &Scale,
228 SDValue &Index, SDValue &Disp,
231 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
232 /// inline asm expressions.
233 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
234 unsigned ConstraintID,
235 std::vector<SDValue> &OutOps) override;
237 void EmitSpecialCodeForMain();
239 inline void getAddressOperands(X86ISelAddressMode &AM, SDLoc DL,
240 SDValue &Base, SDValue &Scale,
241 SDValue &Index, SDValue &Disp,
243 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
244 ? CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
247 Scale = getI8Imm(AM.Scale, DL);
249 // These are 32-bit even in 64-bit mode since RIP relative offset
252 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
256 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
257 AM.Align, AM.Disp, AM.SymbolFlags);
259 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
260 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
261 } else if (AM.JT != -1) {
262 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
263 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
264 } else if (AM.BlockAddr)
265 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
268 Disp = CurDAG->getTargetConstant(AM.Disp, DL, MVT::i32);
270 if (AM.Segment.getNode())
271 Segment = AM.Segment;
273 Segment = CurDAG->getRegister(0, MVT::i32);
276 /// getI8Imm - Return a target constant with the specified value, of type
278 inline SDValue getI8Imm(unsigned Imm, SDLoc DL) {
279 return CurDAG->getTargetConstant(Imm, DL, MVT::i8);
282 /// getI32Imm - Return a target constant with the specified value, of type
284 inline SDValue getI32Imm(unsigned Imm, SDLoc DL) {
285 return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
288 /// getGlobalBaseReg - Return an SDNode that returns the value of
289 /// the global base register. Output instructions required to
290 /// initialize the global base register, if necessary.
292 SDNode *getGlobalBaseReg();
294 /// getTargetMachine - Return a reference to the TargetMachine, casted
295 /// to the target-specific type.
296 const X86TargetMachine &getTargetMachine() const {
297 return static_cast<const X86TargetMachine &>(TM);
300 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
301 /// to the target-specific type.
302 const X86InstrInfo *getInstrInfo() const {
303 return Subtarget->getInstrInfo();
306 /// \brief Address-mode matching performs shift-of-and to and-of-shift
307 /// reassociation in order to expose more scaled addressing
309 bool ComplexPatternFuncMutatesDAG() const override {
317 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
318 if (OptLevel == CodeGenOpt::None) return false;
323 if (N.getOpcode() != ISD::LOAD)
326 // If N is a load, do additional profitability checks.
328 switch (U->getOpcode()) {
341 SDValue Op1 = U->getOperand(1);
343 // If the other operand is a 8-bit immediate we should fold the immediate
344 // instead. This reduces code size.
346 // movl 4(%esp), %eax
350 // addl 4(%esp), %eax
351 // The former is 2 bytes shorter. In case where the increment is 1, then
352 // the saving can be 4 bytes (by using incl %eax).
353 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
354 if (Imm->getAPIntValue().isSignedIntN(8))
357 // If the other operand is a TLS address, we should fold it instead.
360 // leal i@NTPOFF(%eax), %eax
362 // movl $i@NTPOFF, %eax
364 // if the block also has an access to a second TLS address this will save
366 // FIXME: This is probably also true for non-TLS addresses.
367 if (Op1.getOpcode() == X86ISD::Wrapper) {
368 SDValue Val = Op1.getOperand(0);
369 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
379 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
380 /// load's chain operand and move load below the call's chain operand.
381 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
382 SDValue Call, SDValue OrigChain) {
383 SmallVector<SDValue, 8> Ops;
384 SDValue Chain = OrigChain.getOperand(0);
385 if (Chain.getNode() == Load.getNode())
386 Ops.push_back(Load.getOperand(0));
388 assert(Chain.getOpcode() == ISD::TokenFactor &&
389 "Unexpected chain operand");
390 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
391 if (Chain.getOperand(i).getNode() == Load.getNode())
392 Ops.push_back(Load.getOperand(0));
394 Ops.push_back(Chain.getOperand(i));
396 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
398 Ops.push_back(NewChain);
400 Ops.append(OrigChain->op_begin() + 1, OrigChain->op_end());
401 CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
402 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
403 Load.getOperand(1), Load.getOperand(2));
406 Ops.push_back(SDValue(Load.getNode(), 1));
407 Ops.append(Call->op_begin() + 1, Call->op_end());
408 CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
411 /// isCalleeLoad - Return true if call address is a load and it can be
412 /// moved below CALLSEQ_START and the chains leading up to the call.
413 /// Return the CALLSEQ_START by reference as a second output.
414 /// In the case of a tail call, there isn't a callseq node between the call
415 /// chain and the load.
416 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
417 // The transformation is somewhat dangerous if the call's chain was glued to
418 // the call. After MoveBelowOrigChain the load is moved between the call and
419 // the chain, this can create a cycle if the load is not folded. So it is
420 // *really* important that we are sure the load will be folded.
421 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
423 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
426 LD->getAddressingMode() != ISD::UNINDEXED ||
427 LD->getExtensionType() != ISD::NON_EXTLOAD)
430 // Now let's find the callseq_start.
431 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
432 if (!Chain.hasOneUse())
434 Chain = Chain.getOperand(0);
437 if (!Chain.getNumOperands())
439 // Since we are not checking for AA here, conservatively abort if the chain
440 // writes to memory. It's not safe to move the callee (a load) across a store.
441 if (isa<MemSDNode>(Chain.getNode()) &&
442 cast<MemSDNode>(Chain.getNode())->writeMem())
444 if (Chain.getOperand(0).getNode() == Callee.getNode())
446 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
447 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
448 Callee.getValue(1).hasOneUse())
453 void X86DAGToDAGISel::PreprocessISelDAG() {
454 // OptForSize is used in pattern predicates that isel is matching.
455 OptForSize = MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
457 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
458 E = CurDAG->allnodes_end(); I != E; ) {
459 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
461 if (OptLevel != CodeGenOpt::None &&
462 // Only does this when target favors doesn't favor register indirect
464 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
465 (N->getOpcode() == X86ISD::TC_RETURN &&
466 // Only does this if load can be folded into TC_RETURN.
467 (Subtarget->is64Bit() ||
468 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
469 /// Also try moving call address load from outside callseq_start to just
470 /// before the call to allow it to be folded.
488 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
489 SDValue Chain = N->getOperand(0);
490 SDValue Load = N->getOperand(1);
491 if (!isCalleeLoad(Load, Chain, HasCallSeq))
493 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
498 // Lower fpround and fpextend nodes that target the FP stack to be store and
499 // load to the stack. This is a gross hack. We would like to simply mark
500 // these as being illegal, but when we do that, legalize produces these when
501 // it expands calls, then expands these in the same legalize pass. We would
502 // like dag combine to be able to hack on these between the call expansion
503 // and the node legalization. As such this pass basically does "really
504 // late" legalization of these inline with the X86 isel pass.
505 // FIXME: This should only happen when not compiled with -O0.
506 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
509 MVT SrcVT = N->getOperand(0).getSimpleValueType();
510 MVT DstVT = N->getSimpleValueType(0);
512 // If any of the sources are vectors, no fp stack involved.
513 if (SrcVT.isVector() || DstVT.isVector())
516 // If the source and destination are SSE registers, then this is a legal
517 // conversion that should not be lowered.
518 const X86TargetLowering *X86Lowering =
519 static_cast<const X86TargetLowering *>(TLI);
520 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
521 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
522 if (SrcIsSSE && DstIsSSE)
525 if (!SrcIsSSE && !DstIsSSE) {
526 // If this is an FPStack extension, it is a noop.
527 if (N->getOpcode() == ISD::FP_EXTEND)
529 // If this is a value-preserving FPStack truncation, it is a noop.
530 if (N->getConstantOperandVal(1))
534 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
535 // FPStack has extload and truncstore. SSE can fold direct loads into other
536 // operations. Based on this, decide what we want to do.
538 if (N->getOpcode() == ISD::FP_ROUND)
539 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
541 MemVT = SrcIsSSE ? SrcVT : DstVT;
543 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
546 // FIXME: optimize the case where the src/dest is a load or store?
547 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
549 MemTmp, MachinePointerInfo(), MemVT,
551 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
552 MachinePointerInfo(),
553 MemVT, false, false, false, 0);
555 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
556 // extload we created. This will cause general havok on the dag because
557 // anything below the conversion could be folded into other existing nodes.
558 // To avoid invalidating 'I', back it up to the convert node.
560 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
562 // Now that we did that, the node is dead. Increment the iterator to the
563 // next node to process, then delete N.
565 CurDAG->DeleteNode(N);
570 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
571 /// the main function.
572 void X86DAGToDAGISel::EmitSpecialCodeForMain() {
573 if (Subtarget->isTargetCygMing()) {
574 TargetLowering::ArgListTy Args;
576 TargetLowering::CallLoweringInfo CLI(*CurDAG);
577 CLI.setChain(CurDAG->getRoot())
578 .setCallee(CallingConv::C, Type::getVoidTy(*CurDAG->getContext()),
579 CurDAG->getExternalSymbol("__main", TLI->getPointerTy()),
581 const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
582 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
583 CurDAG->setRoot(Result.second);
587 void X86DAGToDAGISel::EmitFunctionEntryCode() {
588 // If this is main, emit special code for main.
589 if (const Function *Fn = MF->getFunction())
590 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
591 EmitSpecialCodeForMain();
594 static bool isDispSafeForFrameIndex(int64_t Val) {
595 // On 64-bit platforms, we can run into an issue where a frame index
596 // includes a displacement that, when added to the explicit displacement,
597 // will overflow the displacement field. Assuming that the frame index
598 // displacement fits into a 31-bit integer (which is only slightly more
599 // aggressive than the current fundamental assumption that it fits into
600 // a 32-bit integer), a 31-bit disp should always be safe.
601 return isInt<31>(Val);
604 bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
605 X86ISelAddressMode &AM) {
606 int64_t Val = AM.Disp + Offset;
607 CodeModel::Model M = TM.getCodeModel();
608 if (Subtarget->is64Bit()) {
609 if (!X86::isOffsetSuitableForCodeModel(Val, M,
610 AM.hasSymbolicDisplacement()))
612 // In addition to the checks required for a register base, check that
613 // we do not try to use an unsafe Disp with a frame index.
614 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
615 !isDispSafeForFrameIndex(Val))
623 bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
624 SDValue Address = N->getOperand(1);
626 // load gs:0 -> GS segment register.
627 // load fs:0 -> FS segment register.
629 // This optimization is valid because the GNU TLS model defines that
630 // gs:0 (or fs:0 on X86-64) contains its own address.
631 // For more information see http://people.redhat.com/drepper/tls.pdf
632 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
633 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
634 Subtarget->isTargetLinux())
635 switch (N->getPointerInfo().getAddrSpace()) {
637 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
640 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
647 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
648 /// into an addressing mode. These wrap things that will resolve down into a
649 /// symbol reference. If no match is possible, this returns true, otherwise it
651 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
652 // If the addressing mode already has a symbol as the displacement, we can
653 // never match another symbol.
654 if (AM.hasSymbolicDisplacement())
657 SDValue N0 = N.getOperand(0);
658 CodeModel::Model M = TM.getCodeModel();
660 // Handle X86-64 rip-relative addresses. We check this before checking direct
661 // folding because RIP is preferable to non-RIP accesses.
662 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
663 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
664 // they cannot be folded into immediate fields.
665 // FIXME: This can be improved for kernel and other models?
666 (M == CodeModel::Small || M == CodeModel::Kernel)) {
667 // Base and index reg must be 0 in order to use %rip as base.
668 if (AM.hasBaseOrIndexReg())
670 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
671 X86ISelAddressMode Backup = AM;
672 AM.GV = G->getGlobal();
673 AM.SymbolFlags = G->getTargetFlags();
674 if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
678 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
679 X86ISelAddressMode Backup = AM;
680 AM.CP = CP->getConstVal();
681 AM.Align = CP->getAlignment();
682 AM.SymbolFlags = CP->getTargetFlags();
683 if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
687 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
688 AM.ES = S->getSymbol();
689 AM.SymbolFlags = S->getTargetFlags();
690 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
691 AM.JT = J->getIndex();
692 AM.SymbolFlags = J->getTargetFlags();
693 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
694 X86ISelAddressMode Backup = AM;
695 AM.BlockAddr = BA->getBlockAddress();
696 AM.SymbolFlags = BA->getTargetFlags();
697 if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
702 llvm_unreachable("Unhandled symbol reference node.");
704 if (N.getOpcode() == X86ISD::WrapperRIP)
705 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
709 // Handle the case when globals fit in our immediate field: This is true for
710 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
711 // mode, this only applies to a non-RIP-relative computation.
712 if (!Subtarget->is64Bit() ||
713 M == CodeModel::Small || M == CodeModel::Kernel) {
714 assert(N.getOpcode() != X86ISD::WrapperRIP &&
715 "RIP-relative addressing already handled");
716 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
717 AM.GV = G->getGlobal();
718 AM.Disp += G->getOffset();
719 AM.SymbolFlags = G->getTargetFlags();
720 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
721 AM.CP = CP->getConstVal();
722 AM.Align = CP->getAlignment();
723 AM.Disp += CP->getOffset();
724 AM.SymbolFlags = CP->getTargetFlags();
725 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
726 AM.ES = S->getSymbol();
727 AM.SymbolFlags = S->getTargetFlags();
728 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
729 AM.JT = J->getIndex();
730 AM.SymbolFlags = J->getTargetFlags();
731 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
732 AM.BlockAddr = BA->getBlockAddress();
733 AM.Disp += BA->getOffset();
734 AM.SymbolFlags = BA->getTargetFlags();
736 llvm_unreachable("Unhandled symbol reference node.");
743 /// MatchAddress - Add the specified node to the specified addressing mode,
744 /// returning true if it cannot be done. This just pattern matches for the
746 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
747 if (MatchAddressRecursively(N, AM, 0))
750 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
751 // a smaller encoding and avoids a scaled-index.
753 AM.BaseType == X86ISelAddressMode::RegBase &&
754 AM.Base_Reg.getNode() == nullptr) {
755 AM.Base_Reg = AM.IndexReg;
759 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
760 // because it has a smaller encoding.
761 // TODO: Which other code models can use this?
762 if (TM.getCodeModel() == CodeModel::Small &&
763 Subtarget->is64Bit() &&
765 AM.BaseType == X86ISelAddressMode::RegBase &&
766 AM.Base_Reg.getNode() == nullptr &&
767 AM.IndexReg.getNode() == nullptr &&
768 AM.SymbolFlags == X86II::MO_NO_FLAG &&
769 AM.hasSymbolicDisplacement())
770 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
775 // Insert a node into the DAG at least before the Pos node's position. This
776 // will reposition the node as needed, and will assign it a node ID that is <=
777 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
778 // IDs! The selection DAG must no longer depend on their uniqueness when this
780 static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
781 if (N.getNode()->getNodeId() == -1 ||
782 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
783 DAG.RepositionNode(Pos.getNode(), N.getNode());
784 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
788 // Transform "(X >> (8-C1)) & (0xff << C1)" to "((X >> 8) & 0xff) << C1" if
789 // safe. This allows us to convert the shift and and into an h-register
790 // extract and a scaled index. Returns false if the simplification is
792 static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
794 SDValue Shift, SDValue X,
795 X86ISelAddressMode &AM) {
796 if (Shift.getOpcode() != ISD::SRL ||
797 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
801 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
802 if (ScaleLog <= 0 || ScaleLog >= 4 ||
803 Mask != (0xffu << ScaleLog))
806 MVT VT = N.getSimpleValueType();
808 SDValue Eight = DAG.getConstant(8, DL, MVT::i8);
809 SDValue NewMask = DAG.getConstant(0xff, DL, VT);
810 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
811 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
812 SDValue ShlCount = DAG.getConstant(ScaleLog, DL, MVT::i8);
813 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
815 // Insert the new nodes into the topological ordering. We must do this in
816 // a valid topological ordering as nothing is going to go back and re-sort
817 // these nodes. We continually insert before 'N' in sequence as this is
818 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
819 // hierarchy left to express.
820 InsertDAGNode(DAG, N, Eight);
821 InsertDAGNode(DAG, N, Srl);
822 InsertDAGNode(DAG, N, NewMask);
823 InsertDAGNode(DAG, N, And);
824 InsertDAGNode(DAG, N, ShlCount);
825 InsertDAGNode(DAG, N, Shl);
826 DAG.ReplaceAllUsesWith(N, Shl);
828 AM.Scale = (1 << ScaleLog);
832 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
833 // allows us to fold the shift into this addressing mode. Returns false if the
834 // transform succeeded.
835 static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
837 SDValue Shift, SDValue X,
838 X86ISelAddressMode &AM) {
839 if (Shift.getOpcode() != ISD::SHL ||
840 !isa<ConstantSDNode>(Shift.getOperand(1)))
843 // Not likely to be profitable if either the AND or SHIFT node has more
844 // than one use (unless all uses are for address computation). Besides,
845 // isel mechanism requires their node ids to be reused.
846 if (!N.hasOneUse() || !Shift.hasOneUse())
849 // Verify that the shift amount is something we can fold.
850 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
851 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
854 MVT VT = N.getSimpleValueType();
856 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, DL, VT);
857 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
858 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
860 // Insert the new nodes into the topological ordering. We must do this in
861 // a valid topological ordering as nothing is going to go back and re-sort
862 // these nodes. We continually insert before 'N' in sequence as this is
863 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
864 // hierarchy left to express.
865 InsertDAGNode(DAG, N, NewMask);
866 InsertDAGNode(DAG, N, NewAnd);
867 InsertDAGNode(DAG, N, NewShift);
868 DAG.ReplaceAllUsesWith(N, NewShift);
870 AM.Scale = 1 << ShiftAmt;
871 AM.IndexReg = NewAnd;
875 // Implement some heroics to detect shifts of masked values where the mask can
876 // be replaced by extending the shift and undoing that in the addressing mode
877 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
878 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
879 // the addressing mode. This results in code such as:
881 // int f(short *y, int *lookup_table) {
883 // return *y + lookup_table[*y >> 11];
887 // movzwl (%rdi), %eax
890 // addl (%rsi,%rcx,4), %eax
893 // movzwl (%rdi), %eax
897 // addl (%rsi,%rcx), %eax
899 // Note that this function assumes the mask is provided as a mask *after* the
900 // value is shifted. The input chain may or may not match that, but computing
901 // such a mask is trivial.
902 static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
904 SDValue Shift, SDValue X,
905 X86ISelAddressMode &AM) {
906 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
907 !isa<ConstantSDNode>(Shift.getOperand(1)))
910 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
911 unsigned MaskLZ = countLeadingZeros(Mask);
912 unsigned MaskTZ = countTrailingZeros(Mask);
914 // The amount of shift we're trying to fit into the addressing mode is taken
915 // from the trailing zeros of the mask.
916 unsigned AMShiftAmt = MaskTZ;
918 // There is nothing we can do here unless the mask is removing some bits.
919 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
920 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
922 // We also need to ensure that mask is a continuous run of bits.
923 if (countTrailingOnes(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
925 // Scale the leading zero count down based on the actual size of the value.
926 // Also scale it down based on the size of the shift.
927 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
929 // The final check is to ensure that any masked out high bits of X are
930 // already known to be zero. Otherwise, the mask has a semantic impact
931 // other than masking out a couple of low bits. Unfortunately, because of
932 // the mask, zero extensions will be removed from operands in some cases.
933 // This code works extra hard to look through extensions because we can
934 // replace them with zero extensions cheaply if necessary.
935 bool ReplacingAnyExtend = false;
936 if (X.getOpcode() == ISD::ANY_EXTEND) {
937 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
938 X.getOperand(0).getSimpleValueType().getSizeInBits();
939 // Assume that we'll replace the any-extend with a zero-extend, and
940 // narrow the search to the extended value.
942 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
943 ReplacingAnyExtend = true;
945 APInt MaskedHighBits =
946 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
947 APInt KnownZero, KnownOne;
948 DAG.computeKnownBits(X, KnownZero, KnownOne);
949 if (MaskedHighBits != KnownZero) return true;
951 // We've identified a pattern that can be transformed into a single shift
952 // and an addressing mode. Make it so.
953 MVT VT = N.getSimpleValueType();
954 if (ReplacingAnyExtend) {
955 assert(X.getValueType() != VT);
956 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
957 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
958 InsertDAGNode(DAG, N, NewX);
962 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, DL, MVT::i8);
963 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
964 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, DL, MVT::i8);
965 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
967 // Insert the new nodes into the topological ordering. We must do this in
968 // a valid topological ordering as nothing is going to go back and re-sort
969 // these nodes. We continually insert before 'N' in sequence as this is
970 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
971 // hierarchy left to express.
972 InsertDAGNode(DAG, N, NewSRLAmt);
973 InsertDAGNode(DAG, N, NewSRL);
974 InsertDAGNode(DAG, N, NewSHLAmt);
975 InsertDAGNode(DAG, N, NewSHL);
976 DAG.ReplaceAllUsesWith(N, NewSHL);
978 AM.Scale = 1 << AMShiftAmt;
979 AM.IndexReg = NewSRL;
983 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
987 dbgs() << "MatchAddress: ";
992 return MatchAddressBase(N, AM);
994 // If this is already a %rip relative address, we can only merge immediates
995 // into it. Instead of handling this in every case, we handle it here.
996 // RIP relative addressing: %rip + 32-bit displacement!
997 if (AM.isRIPRelative()) {
998 // FIXME: JumpTable and ExternalSymbol address currently don't like
999 // displacements. It isn't very important, but this should be fixed for
1001 if (!AM.ES && AM.JT != -1) return true;
1003 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
1004 if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
1009 switch (N.getOpcode()) {
1011 case ISD::FRAME_ALLOC_RECOVER: {
1012 if (!AM.hasSymbolicDisplacement())
1013 if (const auto *ESNode = dyn_cast<ExternalSymbolSDNode>(N.getOperand(0)))
1014 if (ESNode->getOpcode() == ISD::TargetExternalSymbol) {
1015 // Use the symbol and don't prefix it.
1016 AM.ES = ESNode->getSymbol();
1017 AM.SymbolFlags = X86II::MO_NOPREFIX;
1022 case ISD::Constant: {
1023 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1024 if (!FoldOffsetIntoAddress(Val, AM))
1029 case X86ISD::Wrapper:
1030 case X86ISD::WrapperRIP:
1031 if (!MatchWrapper(N, AM))
1036 if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
1040 case ISD::FrameIndex:
1041 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1042 AM.Base_Reg.getNode() == nullptr &&
1043 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1044 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1045 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1051 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1055 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1056 unsigned Val = CN->getZExtValue();
1057 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1058 // that the base operand remains free for further matching. If
1059 // the base doesn't end up getting used, a post-processing step
1060 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1061 if (Val == 1 || Val == 2 || Val == 3) {
1062 AM.Scale = 1 << Val;
1063 SDValue ShVal = N.getNode()->getOperand(0);
1065 // Okay, we know that we have a scale by now. However, if the scaled
1066 // value is an add of something and a constant, we can fold the
1067 // constant into the disp field here.
1068 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1069 AM.IndexReg = ShVal.getNode()->getOperand(0);
1070 ConstantSDNode *AddVal =
1071 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1072 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1073 if (!FoldOffsetIntoAddress(Disp, AM))
1077 AM.IndexReg = ShVal;
1084 // Scale must not be used already.
1085 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1087 SDValue And = N.getOperand(0);
1088 if (And.getOpcode() != ISD::AND) break;
1089 SDValue X = And.getOperand(0);
1091 // We only handle up to 64-bit values here as those are what matter for
1092 // addressing mode optimizations.
1093 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1095 // The mask used for the transform is expected to be post-shift, but we
1096 // found the shift first so just apply the shift to the mask before passing
1098 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1099 !isa<ConstantSDNode>(And.getOperand(1)))
1101 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1103 // Try to fold the mask and shift into the scale, and return false if we
1105 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1110 case ISD::SMUL_LOHI:
1111 case ISD::UMUL_LOHI:
1112 // A mul_lohi where we need the low part can be folded as a plain multiply.
1113 if (N.getResNo() != 0) break;
1116 case X86ISD::MUL_IMM:
1117 // X*[3,5,9] -> X+X*[2,4,8]
1118 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1119 AM.Base_Reg.getNode() == nullptr &&
1120 AM.IndexReg.getNode() == nullptr) {
1122 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1123 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1124 CN->getZExtValue() == 9) {
1125 AM.Scale = unsigned(CN->getZExtValue())-1;
1127 SDValue MulVal = N.getNode()->getOperand(0);
1130 // Okay, we know that we have a scale by now. However, if the scaled
1131 // value is an add of something and a constant, we can fold the
1132 // constant into the disp field here.
1133 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1134 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1135 Reg = MulVal.getNode()->getOperand(0);
1136 ConstantSDNode *AddVal =
1137 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1138 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1139 if (FoldOffsetIntoAddress(Disp, AM))
1140 Reg = N.getNode()->getOperand(0);
1142 Reg = N.getNode()->getOperand(0);
1145 AM.IndexReg = AM.Base_Reg = Reg;
1152 // Given A-B, if A can be completely folded into the address and
1153 // the index field with the index field unused, use -B as the index.
1154 // This is a win if a has multiple parts that can be folded into
1155 // the address. Also, this saves a mov if the base register has
1156 // other uses, since it avoids a two-address sub instruction, however
1157 // it costs an additional mov if the index register has other uses.
1159 // Add an artificial use to this node so that we can keep track of
1160 // it if it gets CSE'd with a different node.
1161 HandleSDNode Handle(N);
1163 // Test if the LHS of the sub can be folded.
1164 X86ISelAddressMode Backup = AM;
1165 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1169 // Test if the index field is free for use.
1170 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1176 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1177 // If the RHS involves a register with multiple uses, this
1178 // transformation incurs an extra mov, due to the neg instruction
1179 // clobbering its operand.
1180 if (!RHS.getNode()->hasOneUse() ||
1181 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1182 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1183 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1184 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1185 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1187 // If the base is a register with multiple uses, this
1188 // transformation may save a mov.
1189 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1190 AM.Base_Reg.getNode() &&
1191 !AM.Base_Reg.getNode()->hasOneUse()) ||
1192 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1194 // If the folded LHS was interesting, this transformation saves
1195 // address arithmetic.
1196 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1197 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1198 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1200 // If it doesn't look like it may be an overall win, don't do it.
1206 // Ok, the transformation is legal and appears profitable. Go for it.
1207 SDValue Zero = CurDAG->getConstant(0, dl, N.getValueType());
1208 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1212 // Insert the new nodes into the topological ordering.
1213 InsertDAGNode(*CurDAG, N, Zero);
1214 InsertDAGNode(*CurDAG, N, Neg);
1219 // Add an artificial use to this node so that we can keep track of
1220 // it if it gets CSE'd with a different node.
1221 HandleSDNode Handle(N);
1223 X86ISelAddressMode Backup = AM;
1224 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1225 !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1229 // Try again after commuting the operands.
1230 if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1231 !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1235 // If we couldn't fold both operands into the address at the same time,
1236 // see if we can just put each operand into a register and fold at least
1238 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1239 !AM.Base_Reg.getNode() &&
1240 !AM.IndexReg.getNode()) {
1241 N = Handle.getValue();
1242 AM.Base_Reg = N.getOperand(0);
1243 AM.IndexReg = N.getOperand(1);
1247 N = Handle.getValue();
1252 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1253 if (CurDAG->isBaseWithConstantOffset(N)) {
1254 X86ISelAddressMode Backup = AM;
1255 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1257 // Start with the LHS as an addr mode.
1258 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1259 !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
1266 // Perform some heroic transforms on an and of a constant-count shift
1267 // with a constant to enable use of the scaled offset field.
1269 // Scale must not be used already.
1270 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1272 SDValue Shift = N.getOperand(0);
1273 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1274 SDValue X = Shift.getOperand(0);
1276 // We only handle up to 64-bit values here as those are what matter for
1277 // addressing mode optimizations.
1278 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1280 if (!isa<ConstantSDNode>(N.getOperand(1)))
1282 uint64_t Mask = N.getConstantOperandVal(1);
1284 // Try to fold the mask and shift into an extract and scale.
1285 if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1288 // Try to fold the mask and shift directly into the scale.
1289 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1292 // Try to swap the mask and shift to place shifts which can be done as
1293 // a scale on the outside of the mask.
1294 if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1300 return MatchAddressBase(N, AM);
1303 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1304 /// specified addressing mode without any further recursion.
1305 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1306 // Is the base register already occupied?
1307 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1308 // If so, check to see if the scale index register is set.
1309 if (!AM.IndexReg.getNode()) {
1315 // Otherwise, we cannot select it.
1319 // Default, generate it as a register.
1320 AM.BaseType = X86ISelAddressMode::RegBase;
1325 bool X86DAGToDAGISel::SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
1326 SDValue &Scale, SDValue &Index,
1327 SDValue &Disp, SDValue &Segment) {
1329 MaskedGatherScatterSDNode *Mgs = dyn_cast<MaskedGatherScatterSDNode>(Parent);
1332 X86ISelAddressMode AM;
1333 unsigned AddrSpace = Mgs->getPointerInfo().getAddrSpace();
1334 // AddrSpace 256 -> GS, 257 -> FS.
1335 if (AddrSpace == 256)
1336 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1337 if (AddrSpace == 257)
1338 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1341 Base = Mgs->getBasePtr();
1342 Index = Mgs->getIndex();
1343 unsigned ScalarSize = Mgs->getValue().getValueType().getScalarSizeInBits();
1344 Scale = getI8Imm(ScalarSize/8, DL);
1346 // If Base is 0, the whole address is in index and the Scale is 1
1347 if (isa<ConstantSDNode>(Base)) {
1348 assert(dyn_cast<ConstantSDNode>(Base)->isNullValue() &&
1349 "Unexpected base in gather/scatter");
1350 Scale = getI8Imm(1, DL);
1351 Base = CurDAG->getRegister(0, MVT::i32);
1353 if (AM.Segment.getNode())
1354 Segment = AM.Segment;
1356 Segment = CurDAG->getRegister(0, MVT::i32);
1357 Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
1361 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1362 /// It returns the operands which make up the maximal addressing mode it can
1363 /// match by reference.
1365 /// Parent is the parent node of the addr operand that is being matched. It
1366 /// is always a load, store, atomic node, or null. It is only null when
1367 /// checking memory operands for inline asm nodes.
1368 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1369 SDValue &Scale, SDValue &Index,
1370 SDValue &Disp, SDValue &Segment) {
1371 X86ISelAddressMode AM;
1374 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1375 // that are not a MemSDNode, and thus don't have proper addrspace info.
1376 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1377 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1378 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1379 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1380 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1381 unsigned AddrSpace =
1382 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1383 // AddrSpace 256 -> GS, 257 -> FS.
1384 if (AddrSpace == 256)
1385 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1386 if (AddrSpace == 257)
1387 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1390 if (MatchAddress(N, AM))
1393 MVT VT = N.getSimpleValueType();
1394 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1395 if (!AM.Base_Reg.getNode())
1396 AM.Base_Reg = CurDAG->getRegister(0, VT);
1399 if (!AM.IndexReg.getNode())
1400 AM.IndexReg = CurDAG->getRegister(0, VT);
1402 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1406 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1407 /// match a load whose top elements are either undef or zeros. The load flavor
1408 /// is derived from the type of N, which is either v4f32 or v2f64.
1411 /// PatternChainNode: this is the matched node that has a chain input and
1413 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1414 SDValue N, SDValue &Base,
1415 SDValue &Scale, SDValue &Index,
1416 SDValue &Disp, SDValue &Segment,
1417 SDValue &PatternNodeWithChain) {
1418 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1419 PatternNodeWithChain = N.getOperand(0);
1420 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1421 PatternNodeWithChain.hasOneUse() &&
1422 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1423 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1424 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1425 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1431 // Also handle the case where we explicitly require zeros in the top
1432 // elements. This is a vector shuffle from the zero vector.
1433 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1434 // Check to see if the top elements are all zeros (or bitcast of zeros).
1435 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1436 N.getOperand(0).getNode()->hasOneUse() &&
1437 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1438 N.getOperand(0).getOperand(0).hasOneUse() &&
1439 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1440 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1441 // Okay, this is a zero extending load. Fold it.
1442 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1443 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1445 PatternNodeWithChain = SDValue(LD, 0);
1452 bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
1453 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1454 uint64_t ImmVal = CN->getZExtValue();
1455 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1458 Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i64);
1462 // In static codegen with small code model, we can get the address of a label
1463 // into a register with 'movl'. TableGen has already made sure we're looking
1464 // at a label of some kind.
1465 assert(N->getOpcode() == X86ISD::Wrapper &&
1466 "Unexpected node type for MOV32ri64");
1467 N = N.getOperand(0);
1469 if (N->getOpcode() != ISD::TargetConstantPool &&
1470 N->getOpcode() != ISD::TargetJumpTable &&
1471 N->getOpcode() != ISD::TargetGlobalAddress &&
1472 N->getOpcode() != ISD::TargetExternalSymbol &&
1473 N->getOpcode() != ISD::TargetBlockAddress)
1477 return TM.getCodeModel() == CodeModel::Small;
1480 bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
1481 SDValue &Scale, SDValue &Index,
1482 SDValue &Disp, SDValue &Segment) {
1483 if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1487 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1488 if (RN && RN->getReg() == 0)
1489 Base = CurDAG->getRegister(0, MVT::i64);
1490 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
1491 // Base could already be %rip, particularly in the x32 ABI.
1492 Base = SDValue(CurDAG->getMachineNode(
1493 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1494 CurDAG->getTargetConstant(0, DL, MVT::i64),
1496 CurDAG->getTargetConstant(X86::sub_32bit, DL, MVT::i32)),
1500 RN = dyn_cast<RegisterSDNode>(Index);
1501 if (RN && RN->getReg() == 0)
1502 Index = CurDAG->getRegister(0, MVT::i64);
1504 assert(Index.getValueType() == MVT::i32 &&
1505 "Expect to be extending 32-bit registers for use in LEA");
1506 Index = SDValue(CurDAG->getMachineNode(
1507 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1508 CurDAG->getTargetConstant(0, DL, MVT::i64),
1510 CurDAG->getTargetConstant(X86::sub_32bit, DL,
1518 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1519 /// mode it matches can be cost effectively emitted as an LEA instruction.
1520 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1521 SDValue &Base, SDValue &Scale,
1522 SDValue &Index, SDValue &Disp,
1524 X86ISelAddressMode AM;
1526 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1528 SDValue Copy = AM.Segment;
1529 SDValue T = CurDAG->getRegister(0, MVT::i32);
1531 if (MatchAddress(N, AM))
1533 assert (T == AM.Segment);
1536 MVT VT = N.getSimpleValueType();
1537 unsigned Complexity = 0;
1538 if (AM.BaseType == X86ISelAddressMode::RegBase)
1539 if (AM.Base_Reg.getNode())
1542 AM.Base_Reg = CurDAG->getRegister(0, VT);
1543 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1546 if (AM.IndexReg.getNode())
1549 AM.IndexReg = CurDAG->getRegister(0, VT);
1551 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1556 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1557 // to a LEA. This is determined with some expermentation but is by no means
1558 // optimal (especially for code size consideration). LEA is nice because of
1559 // its three-address nature. Tweak the cost function again when we can run
1560 // convertToThreeAddress() at register allocation time.
1561 if (AM.hasSymbolicDisplacement()) {
1562 // For X86-64, we should always use lea to materialize RIP relative
1564 if (Subtarget->is64Bit())
1570 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1573 // If it isn't worth using an LEA, reject it.
1574 if (Complexity <= 2)
1577 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1581 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1582 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1583 SDValue &Scale, SDValue &Index,
1584 SDValue &Disp, SDValue &Segment) {
1585 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1586 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1588 X86ISelAddressMode AM;
1589 AM.GV = GA->getGlobal();
1590 AM.Disp += GA->getOffset();
1591 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1592 AM.SymbolFlags = GA->getTargetFlags();
1594 if (N.getValueType() == MVT::i32) {
1596 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1598 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1601 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1606 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1607 SDValue &Base, SDValue &Scale,
1608 SDValue &Index, SDValue &Disp,
1610 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1611 !IsProfitableToFold(N, P, P) ||
1612 !IsLegalToFold(N, P, P, OptLevel))
1615 return SelectAddr(N.getNode(),
1616 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1619 /// getGlobalBaseReg - Return an SDNode that returns the value of
1620 /// the global base register. Output instructions required to
1621 /// initialize the global base register, if necessary.
1623 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1624 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1625 return CurDAG->getRegister(GlobalBaseReg, TLI->getPointerTy()).getNode();
1628 /// Atomic opcode table
1656 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1667 X86::LOCK_ADD64mi32,
1680 X86::LOCK_SUB64mi32,
1732 X86::LOCK_AND64mi32,
1745 X86::LOCK_XOR64mi32,
1750 // Return the target constant operand for atomic-load-op and do simple
1751 // translations, such as from atomic-load-add to lock-sub. The return value is
1752 // one of the following 3 cases:
1753 // + target-constant, the operand could be supported as a target constant.
1754 // + empty, the operand is not needed any more with the new op selected.
1755 // + non-empty, otherwise.
1756 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1758 enum AtomicOpc &Op, MVT NVT,
1760 const X86Subtarget *Subtarget) {
1761 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1762 int64_t CNVal = CN->getSExtValue();
1763 // Quit if not 32-bit imm.
1764 if ((int32_t)CNVal != CNVal)
1766 // Quit if INT32_MIN: it would be negated as it is negative and overflow,
1767 // producing an immediate that does not fit in the 32 bits available for
1768 // an immediate operand to sub. However, it still fits in 32 bits for the
1769 // add (since it is not negated) so we can return target-constant.
1770 if (CNVal == INT32_MIN)
1771 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1772 // For atomic-load-add, we could do some optimizations.
1774 // Translate to INC/DEC if ADD by 1 or -1.
1775 if (((CNVal == 1) || (CNVal == -1)) && !Subtarget->slowIncDec()) {
1776 Op = (CNVal == 1) ? INC : DEC;
1777 // No more constant operand after being translated into INC/DEC.
1780 // Translate to SUB if ADD by negative value.
1786 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1789 // If the value operand is single-used, try to optimize it.
1790 if (Op == ADD && Val.hasOneUse()) {
1791 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1792 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1794 return Val.getOperand(1);
1796 // A special case for i16, which needs truncating as, in most cases, it's
1797 // promoted to i32. We will translate
1798 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1799 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1800 Val.getOperand(0).getOpcode() == ISD::SUB &&
1801 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1803 Val = Val.getOperand(0);
1804 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1812 SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
1813 if (Node->hasAnyUseOfValue(0))
1818 // Optimize common patterns for __sync_or_and_fetch and similar arith
1819 // operations where the result is not used. This allows us to use the "lock"
1820 // version of the arithmetic instruction.
1821 SDValue Chain = Node->getOperand(0);
1822 SDValue Ptr = Node->getOperand(1);
1823 SDValue Val = Node->getOperand(2);
1824 SDValue Base, Scale, Index, Disp, Segment;
1825 if (!SelectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
1828 // Which index into the table.
1830 switch (Node->getOpcode()) {
1833 case ISD::ATOMIC_LOAD_OR:
1836 case ISD::ATOMIC_LOAD_AND:
1839 case ISD::ATOMIC_LOAD_XOR:
1842 case ISD::ATOMIC_LOAD_ADD:
1847 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val, Subtarget);
1848 bool isUnOp = !Val.getNode();
1849 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1852 switch (NVT.SimpleTy) {
1853 default: return nullptr;
1856 Opc = AtomicOpcTbl[Op][ConstantI8];
1858 Opc = AtomicOpcTbl[Op][I8];
1862 if (immSext8(Val.getNode()))
1863 Opc = AtomicOpcTbl[Op][SextConstantI16];
1865 Opc = AtomicOpcTbl[Op][ConstantI16];
1867 Opc = AtomicOpcTbl[Op][I16];
1871 if (immSext8(Val.getNode()))
1872 Opc = AtomicOpcTbl[Op][SextConstantI32];
1874 Opc = AtomicOpcTbl[Op][ConstantI32];
1876 Opc = AtomicOpcTbl[Op][I32];
1880 if (immSext8(Val.getNode()))
1881 Opc = AtomicOpcTbl[Op][SextConstantI64];
1882 else if (i64immSExt32(Val.getNode()))
1883 Opc = AtomicOpcTbl[Op][ConstantI64];
1885 llvm_unreachable("True 64 bits constant in SelectAtomicLoadArith");
1887 Opc = AtomicOpcTbl[Op][I64];
1891 assert(Opc != 0 && "Invalid arith lock transform!");
1893 // Building the new node.
1896 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Chain };
1897 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1899 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Val, Chain };
1900 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1903 // Copying the MachineMemOperand.
1904 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1905 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1906 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1908 // We need to have two outputs as that is what the original instruction had.
1909 // So we add a dummy, undefined output. This is safe as we checked first
1910 // that no-one uses our output anyway.
1911 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1913 SDValue RetVals[] = { Undef, Ret };
1914 return CurDAG->getMergeValues(RetVals, dl).getNode();
1917 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1918 /// any uses which require the SF or OF bits to be accurate.
1919 static bool HasNoSignedComparisonUses(SDNode *N) {
1920 // Examine each user of the node.
1921 for (SDNode::use_iterator UI = N->use_begin(),
1922 UE = N->use_end(); UI != UE; ++UI) {
1923 // Only examine CopyToReg uses.
1924 if (UI->getOpcode() != ISD::CopyToReg)
1926 // Only examine CopyToReg uses that copy to EFLAGS.
1927 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1930 // Examine each user of the CopyToReg use.
1931 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1932 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1933 // Only examine the Flag result.
1934 if (FlagUI.getUse().getResNo() != 1) continue;
1935 // Anything unusual: assume conservatively.
1936 if (!FlagUI->isMachineOpcode()) return false;
1937 // Examine the opcode of the user.
1938 switch (FlagUI->getMachineOpcode()) {
1939 // These comparisons don't treat the most significant bit specially.
1940 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1941 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1942 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1943 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1944 case X86::JA_1: case X86::JAE_1: case X86::JB_1: case X86::JBE_1:
1945 case X86::JE_1: case X86::JNE_1: case X86::JP_1: case X86::JNP_1:
1946 case X86::CMOVA16rr: case X86::CMOVA16rm:
1947 case X86::CMOVA32rr: case X86::CMOVA32rm:
1948 case X86::CMOVA64rr: case X86::CMOVA64rm:
1949 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1950 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1951 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1952 case X86::CMOVB16rr: case X86::CMOVB16rm:
1953 case X86::CMOVB32rr: case X86::CMOVB32rm:
1954 case X86::CMOVB64rr: case X86::CMOVB64rm:
1955 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1956 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1957 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1958 case X86::CMOVE16rr: case X86::CMOVE16rm:
1959 case X86::CMOVE32rr: case X86::CMOVE32rm:
1960 case X86::CMOVE64rr: case X86::CMOVE64rm:
1961 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1962 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1963 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1964 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1965 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1966 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1967 case X86::CMOVP16rr: case X86::CMOVP16rm:
1968 case X86::CMOVP32rr: case X86::CMOVP32rm:
1969 case X86::CMOVP64rr: case X86::CMOVP64rm:
1971 // Anything else: assume conservatively.
1972 default: return false;
1979 /// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
1980 /// is suitable for doing the {load; increment or decrement; store} to modify
1982 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
1983 SDValue StoredVal, SelectionDAG *CurDAG,
1984 LoadSDNode* &LoadNode, SDValue &InputChain) {
1986 // is the value stored the result of a DEC or INC?
1987 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
1989 // is the stored value result 0 of the load?
1990 if (StoredVal.getResNo() != 0) return false;
1992 // are there other uses of the loaded value than the inc or dec?
1993 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
1995 // is the store non-extending and non-indexed?
1996 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
1999 SDValue Load = StoredVal->getOperand(0);
2000 // Is the stored value a non-extending and non-indexed load?
2001 if (!ISD::isNormalLoad(Load.getNode())) return false;
2003 // Return LoadNode by reference.
2004 LoadNode = cast<LoadSDNode>(Load);
2005 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
2006 EVT LdVT = LoadNode->getMemoryVT();
2007 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
2011 // Is store the only read of the loaded value?
2012 if (!Load.hasOneUse())
2015 // Is the address of the store the same as the load?
2016 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
2017 LoadNode->getOffset() != StoreNode->getOffset())
2020 // Check if the chain is produced by the load or is a TokenFactor with
2021 // the load output chain as an operand. Return InputChain by reference.
2022 SDValue Chain = StoreNode->getChain();
2024 bool ChainCheck = false;
2025 if (Chain == Load.getValue(1)) {
2027 InputChain = LoadNode->getChain();
2028 } else if (Chain.getOpcode() == ISD::TokenFactor) {
2029 SmallVector<SDValue, 4> ChainOps;
2030 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
2031 SDValue Op = Chain.getOperand(i);
2032 if (Op == Load.getValue(1)) {
2037 // Make sure using Op as part of the chain would not cause a cycle here.
2038 // In theory, we could check whether the chain node is a predecessor of
2039 // the load. But that can be very expensive. Instead visit the uses and
2040 // make sure they all have smaller node id than the load.
2041 int LoadId = LoadNode->getNodeId();
2042 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
2043 UE = UI->use_end(); UI != UE; ++UI) {
2044 if (UI.getUse().getResNo() != 0)
2046 if (UI->getNodeId() > LoadId)
2050 ChainOps.push_back(Op);
2054 // Make a new TokenFactor with all the other input chains except
2056 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
2057 MVT::Other, ChainOps);
2065 /// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
2066 /// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
2067 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2068 if (Opc == X86ISD::DEC) {
2069 if (LdVT == MVT::i64) return X86::DEC64m;
2070 if (LdVT == MVT::i32) return X86::DEC32m;
2071 if (LdVT == MVT::i16) return X86::DEC16m;
2072 if (LdVT == MVT::i8) return X86::DEC8m;
2074 assert(Opc == X86ISD::INC && "unrecognized opcode");
2075 if (LdVT == MVT::i64) return X86::INC64m;
2076 if (LdVT == MVT::i32) return X86::INC32m;
2077 if (LdVT == MVT::i16) return X86::INC16m;
2078 if (LdVT == MVT::i8) return X86::INC8m;
2080 llvm_unreachable("unrecognized size for LdVT");
2083 /// SelectGather - Customized ISel for GATHER operations.
2085 SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
2086 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2087 SDValue Chain = Node->getOperand(0);
2088 SDValue VSrc = Node->getOperand(2);
2089 SDValue Base = Node->getOperand(3);
2090 SDValue VIdx = Node->getOperand(4);
2091 SDValue VMask = Node->getOperand(5);
2092 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2096 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2101 // Memory Operands: Base, Scale, Index, Disp, Segment
2102 SDValue Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
2103 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2104 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue(), DL), VIdx,
2105 Disp, Segment, VMask, Chain};
2106 SDNode *ResNode = CurDAG->getMachineNode(Opc, DL, VTs, Ops);
2107 // Node has 2 outputs: VDst and MVT::Other.
2108 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2109 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2111 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2112 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2116 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2117 MVT NVT = Node->getSimpleValueType(0);
2119 unsigned Opcode = Node->getOpcode();
2122 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2124 if (Node->isMachineOpcode()) {
2125 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2126 Node->setNodeId(-1);
2127 return nullptr; // Already selected.
2132 case ISD::INTRINSIC_W_CHAIN: {
2133 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2136 case Intrinsic::x86_avx2_gather_d_pd:
2137 case Intrinsic::x86_avx2_gather_d_pd_256:
2138 case Intrinsic::x86_avx2_gather_q_pd:
2139 case Intrinsic::x86_avx2_gather_q_pd_256:
2140 case Intrinsic::x86_avx2_gather_d_ps:
2141 case Intrinsic::x86_avx2_gather_d_ps_256:
2142 case Intrinsic::x86_avx2_gather_q_ps:
2143 case Intrinsic::x86_avx2_gather_q_ps_256:
2144 case Intrinsic::x86_avx2_gather_d_q:
2145 case Intrinsic::x86_avx2_gather_d_q_256:
2146 case Intrinsic::x86_avx2_gather_q_q:
2147 case Intrinsic::x86_avx2_gather_q_q_256:
2148 case Intrinsic::x86_avx2_gather_d_d:
2149 case Intrinsic::x86_avx2_gather_d_d_256:
2150 case Intrinsic::x86_avx2_gather_q_d:
2151 case Intrinsic::x86_avx2_gather_q_d_256: {
2152 if (!Subtarget->hasAVX2())
2156 default: llvm_unreachable("Impossible intrinsic");
2157 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2158 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2159 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2160 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2161 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2162 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2163 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2164 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2165 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2166 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2167 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2168 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2169 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2170 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2171 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2172 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2174 SDNode *RetVal = SelectGather(Node, Opc);
2176 // We already called ReplaceUses inside SelectGather.
2183 case X86ISD::GlobalBaseReg:
2184 return getGlobalBaseReg();
2186 case X86ISD::SHRUNKBLEND: {
2187 // SHRUNKBLEND selects like a regular VSELECT.
2188 SDValue VSelect = CurDAG->getNode(
2189 ISD::VSELECT, SDLoc(Node), Node->getValueType(0), Node->getOperand(0),
2190 Node->getOperand(1), Node->getOperand(2));
2191 ReplaceUses(SDValue(Node, 0), VSelect);
2192 SelectCode(VSelect.getNode());
2193 // We already called ReplaceUses.
2197 case ISD::ATOMIC_LOAD_XOR:
2198 case ISD::ATOMIC_LOAD_AND:
2199 case ISD::ATOMIC_LOAD_OR:
2200 case ISD::ATOMIC_LOAD_ADD: {
2201 SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
2209 // For operations of the form (x << C1) op C2, check if we can use a smaller
2210 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2211 SDValue N0 = Node->getOperand(0);
2212 SDValue N1 = Node->getOperand(1);
2214 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2217 // i8 is unshrinkable, i16 should be promoted to i32.
2218 if (NVT != MVT::i32 && NVT != MVT::i64)
2221 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2222 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2223 if (!Cst || !ShlCst)
2226 int64_t Val = Cst->getSExtValue();
2227 uint64_t ShlVal = ShlCst->getZExtValue();
2229 // Make sure that we don't change the operation by removing bits.
2230 // This only matters for OR and XOR, AND is unaffected.
2231 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2232 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2235 unsigned ShlOp, AddOp, Op;
2238 // Check the minimum bitwidth for the new constant.
2239 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2240 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2241 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2242 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2244 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2247 // Bail if there is no smaller encoding.
2251 switch (NVT.SimpleTy) {
2252 default: llvm_unreachable("Unsupported VT!");
2254 assert(CstVT == MVT::i8);
2255 ShlOp = X86::SHL32ri;
2256 AddOp = X86::ADD32rr;
2259 default: llvm_unreachable("Impossible opcode");
2260 case ISD::AND: Op = X86::AND32ri8; break;
2261 case ISD::OR: Op = X86::OR32ri8; break;
2262 case ISD::XOR: Op = X86::XOR32ri8; break;
2266 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2267 ShlOp = X86::SHL64ri;
2268 AddOp = X86::ADD64rr;
2271 default: llvm_unreachable("Impossible opcode");
2272 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2273 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2274 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2279 // Emit the smaller op and the shift.
2280 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, dl, CstVT);
2281 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2283 return CurDAG->SelectNodeTo(Node, AddOp, NVT, SDValue(New, 0),
2285 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2286 getI8Imm(ShlVal, dl));
2289 case X86ISD::SMUL8: {
2290 SDValue N0 = Node->getOperand(0);
2291 SDValue N1 = Node->getOperand(1);
2293 Opc = (Opcode == X86ISD::SMUL8 ? X86::IMUL8r : X86::MUL8r);
2295 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::AL,
2296 N0, SDValue()).getValue(1);
2298 SDVTList VTs = CurDAG->getVTList(NVT, MVT::i32);
2299 SDValue Ops[] = {N1, InFlag};
2300 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2302 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2303 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2307 case X86ISD::UMUL: {
2308 SDValue N0 = Node->getOperand(0);
2309 SDValue N1 = Node->getOperand(1);
2312 switch (NVT.SimpleTy) {
2313 default: llvm_unreachable("Unsupported VT!");
2314 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2315 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2316 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2317 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2320 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2321 N0, SDValue()).getValue(1);
2323 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2324 SDValue Ops[] = {N1, InFlag};
2325 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2327 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2328 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2329 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2333 case ISD::SMUL_LOHI:
2334 case ISD::UMUL_LOHI: {
2335 SDValue N0 = Node->getOperand(0);
2336 SDValue N1 = Node->getOperand(1);
2338 bool isSigned = Opcode == ISD::SMUL_LOHI;
2339 bool hasBMI2 = Subtarget->hasBMI2();
2341 switch (NVT.SimpleTy) {
2342 default: llvm_unreachable("Unsupported VT!");
2343 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2344 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2345 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2346 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2347 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2348 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2351 switch (NVT.SimpleTy) {
2352 default: llvm_unreachable("Unsupported VT!");
2353 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2354 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2355 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2356 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2360 unsigned SrcReg, LoReg, HiReg;
2362 default: llvm_unreachable("Unknown MUL opcode!");
2365 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2369 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2373 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2377 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2380 SrcReg = X86::EDX; LoReg = HiReg = 0;
2383 SrcReg = X86::RDX; LoReg = HiReg = 0;
2387 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2388 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2389 // Multiply is commmutative.
2391 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2396 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2397 N0, SDValue()).getValue(1);
2398 SDValue ResHi, ResLo;
2402 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2404 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2405 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2406 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2407 ResHi = SDValue(CNode, 0);
2408 ResLo = SDValue(CNode, 1);
2409 Chain = SDValue(CNode, 2);
2410 InFlag = SDValue(CNode, 3);
2412 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2413 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2414 Chain = SDValue(CNode, 0);
2415 InFlag = SDValue(CNode, 1);
2418 // Update the chain.
2419 ReplaceUses(N1.getValue(1), Chain);
2421 SDValue Ops[] = { N1, InFlag };
2422 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2423 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2424 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2425 ResHi = SDValue(CNode, 0);
2426 ResLo = SDValue(CNode, 1);
2427 InFlag = SDValue(CNode, 2);
2429 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2430 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2431 InFlag = SDValue(CNode, 0);
2435 // Prevent use of AH in a REX instruction by referencing AX instead.
2436 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2437 !SDValue(Node, 1).use_empty()) {
2438 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2439 X86::AX, MVT::i16, InFlag);
2440 InFlag = Result.getValue(2);
2441 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2443 if (!SDValue(Node, 0).use_empty())
2444 ReplaceUses(SDValue(Node, 1),
2445 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2447 // Shift AX down 8 bits.
2448 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2450 CurDAG->getTargetConstant(8, dl, MVT::i8)),
2452 // Then truncate it down to i8.
2453 ReplaceUses(SDValue(Node, 1),
2454 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2456 // Copy the low half of the result, if it is needed.
2457 if (!SDValue(Node, 0).use_empty()) {
2458 if (!ResLo.getNode()) {
2459 assert(LoReg && "Register for low half is not defined!");
2460 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2462 InFlag = ResLo.getValue(2);
2464 ReplaceUses(SDValue(Node, 0), ResLo);
2465 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2467 // Copy the high half of the result, if it is needed.
2468 if (!SDValue(Node, 1).use_empty()) {
2469 if (!ResHi.getNode()) {
2470 assert(HiReg && "Register for high half is not defined!");
2471 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2473 InFlag = ResHi.getValue(2);
2475 ReplaceUses(SDValue(Node, 1), ResHi);
2476 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2484 case X86ISD::SDIVREM8_SEXT_HREG:
2485 case X86ISD::UDIVREM8_ZEXT_HREG: {
2486 SDValue N0 = Node->getOperand(0);
2487 SDValue N1 = Node->getOperand(1);
2489 bool isSigned = (Opcode == ISD::SDIVREM ||
2490 Opcode == X86ISD::SDIVREM8_SEXT_HREG);
2492 switch (NVT.SimpleTy) {
2493 default: llvm_unreachable("Unsupported VT!");
2494 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2495 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2496 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2497 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2500 switch (NVT.SimpleTy) {
2501 default: llvm_unreachable("Unsupported VT!");
2502 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2503 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2504 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2505 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2509 unsigned LoReg, HiReg, ClrReg;
2510 unsigned SExtOpcode;
2511 switch (NVT.SimpleTy) {
2512 default: llvm_unreachable("Unsupported VT!");
2514 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2515 SExtOpcode = X86::CBW;
2518 LoReg = X86::AX; HiReg = X86::DX;
2520 SExtOpcode = X86::CWD;
2523 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2524 SExtOpcode = X86::CDQ;
2527 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2528 SExtOpcode = X86::CQO;
2532 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2533 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2534 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2537 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2538 // Special case for div8, just use a move with zero extension to AX to
2539 // clear the upper 8 bits (AH).
2540 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2541 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2542 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2544 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2545 MVT::Other, Ops), 0);
2546 Chain = Move.getValue(1);
2547 ReplaceUses(N0.getValue(1), Chain);
2550 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2551 Chain = CurDAG->getEntryNode();
2553 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2554 InFlag = Chain.getValue(1);
2557 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2558 LoReg, N0, SDValue()).getValue(1);
2559 if (isSigned && !signBitIsZero) {
2560 // Sign extend the low part into the high part.
2562 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2564 // Zero out the high part, effectively zero extending the input.
2565 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2566 switch (NVT.SimpleTy) {
2569 SDValue(CurDAG->getMachineNode(
2570 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2571 CurDAG->getTargetConstant(X86::sub_16bit, dl,
2579 SDValue(CurDAG->getMachineNode(
2580 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2581 CurDAG->getTargetConstant(0, dl, MVT::i64), ClrNode,
2582 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2587 llvm_unreachable("Unexpected division source");
2590 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2591 ClrNode, InFlag).getValue(1);
2596 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2599 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2600 InFlag = SDValue(CNode, 1);
2601 // Update the chain.
2602 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2605 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2608 // Prevent use of AH in a REX instruction by explicitly copying it to
2609 // an ABCD_L register.
2611 // The current assumption of the register allocator is that isel
2612 // won't generate explicit references to the GR8_ABCD_H registers. If
2613 // the allocator and/or the backend get enhanced to be more robust in
2614 // that regard, this can be, and should be, removed.
2615 if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
2616 SDValue AHCopy = CurDAG->getRegister(X86::AH, MVT::i8);
2617 unsigned AHExtOpcode =
2618 isSigned ? X86::MOVSX32_NOREXrr8 : X86::MOVZX32_NOREXrr8;
2620 SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
2621 MVT::Glue, AHCopy, InFlag);
2622 SDValue Result(RNode, 0);
2623 InFlag = SDValue(RNode, 1);
2625 if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
2626 Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
2627 if (Node->getValueType(1) == MVT::i64) {
2628 // It's not possible to directly movsx AH to a 64bit register, because
2629 // the latter needs the REX prefix, but the former can't have it.
2630 assert(Opcode != X86ISD::SDIVREM8_SEXT_HREG &&
2631 "Unexpected i64 sext of h-register");
2633 SDValue(CurDAG->getMachineNode(
2634 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2635 CurDAG->getTargetConstant(0, dl, MVT::i64), Result,
2636 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2642 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
2644 ReplaceUses(SDValue(Node, 1), Result);
2645 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2647 // Copy the division (low) result, if it is needed.
2648 if (!SDValue(Node, 0).use_empty()) {
2649 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2650 LoReg, NVT, InFlag);
2651 InFlag = Result.getValue(2);
2652 ReplaceUses(SDValue(Node, 0), Result);
2653 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2655 // Copy the remainder (high) result, if it is needed.
2656 if (!SDValue(Node, 1).use_empty()) {
2657 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2658 HiReg, NVT, InFlag);
2659 InFlag = Result.getValue(2);
2660 ReplaceUses(SDValue(Node, 1), Result);
2661 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2668 // Sometimes a SUB is used to perform comparison.
2669 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2670 // This node is not a CMP.
2672 SDValue N0 = Node->getOperand(0);
2673 SDValue N1 = Node->getOperand(1);
2675 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2676 HasNoSignedComparisonUses(Node))
2677 N0 = N0.getOperand(0);
2679 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2680 // use a smaller encoding.
2681 // Look past the truncate if CMP is the only use of it.
2682 if ((N0.getNode()->getOpcode() == ISD::AND ||
2683 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2684 N0.getNode()->hasOneUse() &&
2685 N0.getValueType() != MVT::i8 &&
2686 X86::isZeroNode(N1)) {
2687 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2690 // For example, convert "testl %eax, $8" to "testb %al, $8"
2691 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2692 (!(C->getZExtValue() & 0x80) ||
2693 HasNoSignedComparisonUses(Node))) {
2694 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl, MVT::i8);
2695 SDValue Reg = N0.getNode()->getOperand(0);
2697 // On x86-32, only the ABCD registers have 8-bit subregisters.
2698 if (!Subtarget->is64Bit()) {
2699 const TargetRegisterClass *TRC;
2700 switch (N0.getSimpleValueType().SimpleTy) {
2701 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2702 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2703 default: llvm_unreachable("Unsupported TEST operand type!");
2705 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2706 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2707 Reg.getValueType(), Reg, RC), 0);
2710 // Extract the l-register.
2711 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2715 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2717 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2718 // one, do not call ReplaceAllUsesWith.
2719 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2720 SDValue(NewNode, 0));
2724 // For example, "testl %eax, $2048" to "testb %ah, $8".
2725 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2726 (!(C->getZExtValue() & 0x8000) ||
2727 HasNoSignedComparisonUses(Node))) {
2728 // Shift the immediate right by 8 bits.
2729 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2731 SDValue Reg = N0.getNode()->getOperand(0);
2733 // Put the value in an ABCD register.
2734 const TargetRegisterClass *TRC;
2735 switch (N0.getSimpleValueType().SimpleTy) {
2736 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2737 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2738 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2739 default: llvm_unreachable("Unsupported TEST operand type!");
2741 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2742 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2743 Reg.getValueType(), Reg, RC), 0);
2745 // Extract the h-register.
2746 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2749 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2750 // target GR8_NOREX registers, so make sure the register class is
2752 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2753 MVT::i32, Subreg, ShiftedImm);
2754 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2755 // one, do not call ReplaceAllUsesWith.
2756 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2757 SDValue(NewNode, 0));
2761 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2762 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2763 N0.getValueType() != MVT::i16 &&
2764 (!(C->getZExtValue() & 0x8000) ||
2765 HasNoSignedComparisonUses(Node))) {
2766 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2768 SDValue Reg = N0.getNode()->getOperand(0);
2770 // Extract the 16-bit subregister.
2771 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2775 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2777 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2778 // one, do not call ReplaceAllUsesWith.
2779 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2780 SDValue(NewNode, 0));
2784 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2785 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2786 N0.getValueType() == MVT::i64 &&
2787 (!(C->getZExtValue() & 0x80000000) ||
2788 HasNoSignedComparisonUses(Node))) {
2789 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2791 SDValue Reg = N0.getNode()->getOperand(0);
2793 // Extract the 32-bit subregister.
2794 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2798 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2800 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2801 // one, do not call ReplaceAllUsesWith.
2802 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2803 SDValue(NewNode, 0));
2810 // Change a chain of {load; incr or dec; store} of the same value into
2811 // a simple increment or decrement through memory of that value, if the
2812 // uses of the modified value and its address are suitable.
2813 // The DEC64m tablegen pattern is currently not able to match the case where
2814 // the EFLAGS on the original DEC are used. (This also applies to
2815 // {INC,DEC}X{64,32,16,8}.)
2816 // We'll need to improve tablegen to allow flags to be transferred from a
2817 // node in the pattern to the result node. probably with a new keyword
2818 // for example, we have this
2819 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2820 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2821 // (implicit EFLAGS)]>;
2822 // but maybe need something like this
2823 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2824 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2825 // (transferrable EFLAGS)]>;
2827 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2828 SDValue StoredVal = StoreNode->getOperand(1);
2829 unsigned Opc = StoredVal->getOpcode();
2831 LoadSDNode *LoadNode = nullptr;
2833 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2834 LoadNode, InputChain))
2837 SDValue Base, Scale, Index, Disp, Segment;
2838 if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
2839 Base, Scale, Index, Disp, Segment))
2842 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2843 MemOp[0] = StoreNode->getMemOperand();
2844 MemOp[1] = LoadNode->getMemOperand();
2845 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2846 EVT LdVT = LoadNode->getMemoryVT();
2847 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2848 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2850 MVT::i32, MVT::Other, Ops);
2851 Result->setMemRefs(MemOp, MemOp + 2);
2853 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2854 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2860 SDNode *ResNode = SelectCode(Node);
2862 DEBUG(dbgs() << "=> ";
2863 if (ResNode == nullptr || ResNode == Node)
2866 ResNode->dump(CurDAG);
2872 bool X86DAGToDAGISel::
2873 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
2874 std::vector<SDValue> &OutOps) {
2875 SDValue Op0, Op1, Op2, Op3, Op4;
2876 switch (ConstraintID) {
2877 case InlineAsm::Constraint_o: // offsetable ??
2878 case InlineAsm::Constraint_v: // not offsetable ??
2879 default: return true;
2880 case InlineAsm::Constraint_m: // memory
2881 if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
2886 OutOps.push_back(Op0);
2887 OutOps.push_back(Op1);
2888 OutOps.push_back(Op2);
2889 OutOps.push_back(Op3);
2890 OutOps.push_back(Op4);
2894 /// createX86ISelDag - This pass converts a legalized DAG into a
2895 /// X86-specific DAG, ready for instruction scheduling.
2897 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2898 CodeGenOpt::Level OptLevel) {
2899 return new X86DAGToDAGISel(TM, OptLevel);