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;
72 unsigned Align; // CP alignment.
73 unsigned char SymbolFlags; // X86II::MO_*
76 : BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
77 Segment(), GV(nullptr), CP(nullptr), BlockAddr(nullptr), ES(nullptr),
78 MCSym(nullptr), JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {}
80 bool hasSymbolicDisplacement() const {
81 return GV != nullptr || CP != nullptr || ES != nullptr ||
82 MCSym != nullptr || 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'
142 dbgs() << " JT" << JT << " Align" << Align << '\n';
149 //===--------------------------------------------------------------------===//
150 /// ISel - X86 specific code to select X86 machine instructions for
151 /// SelectionDAG operations.
153 class X86DAGToDAGISel final : public SelectionDAGISel {
154 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
155 /// make the right decision when generating code for different targets.
156 const X86Subtarget *Subtarget;
158 /// OptForSize - If true, selector should try to optimize for code size
159 /// instead of performance.
163 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
164 : SelectionDAGISel(tm, OptLevel), OptForSize(false) {}
166 const char *getPassName() const override {
167 return "X86 DAG->DAG Instruction Selection";
170 bool runOnMachineFunction(MachineFunction &MF) override {
171 // Reset the subtarget each time through.
172 Subtarget = &MF.getSubtarget<X86Subtarget>();
173 SelectionDAGISel::runOnMachineFunction(MF);
177 void EmitFunctionEntryCode() override;
179 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
181 void PreprocessISelDAG() override;
183 inline bool immSext8(SDNode *N) const {
184 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
187 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
188 // sign extended field.
189 inline bool i64immSExt32(SDNode *N) const {
190 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
191 return (int64_t)v == (int32_t)v;
194 // Include the pieces autogenerated from the target description.
195 #include "X86GenDAGISel.inc"
198 SDNode *Select(SDNode *N) override;
199 SDNode *SelectGather(SDNode *N, unsigned Opc);
200 SDNode *SelectAtomicLoadArith(SDNode *Node, MVT NVT);
202 bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
203 bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
204 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
205 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
206 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
208 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
209 bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
210 SDValue &Scale, SDValue &Index, SDValue &Disp,
212 bool SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
213 SDValue &Scale, SDValue &Index, SDValue &Disp,
215 bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
216 bool SelectLEAAddr(SDValue N, SDValue &Base,
217 SDValue &Scale, SDValue &Index, SDValue &Disp,
219 bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
220 SDValue &Scale, SDValue &Index, SDValue &Disp,
222 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
223 SDValue &Scale, SDValue &Index, SDValue &Disp,
225 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
226 SDValue &Base, SDValue &Scale,
227 SDValue &Index, SDValue &Disp,
229 SDValue &NodeWithChain);
231 bool TryFoldLoad(SDNode *P, SDValue N,
232 SDValue &Base, SDValue &Scale,
233 SDValue &Index, SDValue &Disp,
236 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
237 /// inline asm expressions.
238 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
239 unsigned ConstraintID,
240 std::vector<SDValue> &OutOps) override;
242 void EmitSpecialCodeForMain();
244 inline void getAddressOperands(X86ISelAddressMode &AM, SDLoc DL,
245 SDValue &Base, SDValue &Scale,
246 SDValue &Index, SDValue &Disp,
248 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
249 ? CurDAG->getTargetFrameIndex(
251 TLI->getPointerTy(CurDAG->getDataLayout()))
253 Scale = getI8Imm(AM.Scale, DL);
255 // These are 32-bit even in 64-bit mode since RIP relative offset
258 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
262 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
263 AM.Align, AM.Disp, AM.SymbolFlags);
265 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
266 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
267 } else if (AM.MCSym) {
268 assert(!AM.Disp && "Non-zero displacement is ignored with MCSym.");
269 assert(AM.SymbolFlags == 0 && "oo");
270 Disp = CurDAG->getMCSymbol(AM.MCSym, MVT::i32);
271 } else if (AM.JT != -1) {
272 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
273 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
274 } else if (AM.BlockAddr)
275 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
278 Disp = CurDAG->getTargetConstant(AM.Disp, DL, MVT::i32);
280 if (AM.Segment.getNode())
281 Segment = AM.Segment;
283 Segment = CurDAG->getRegister(0, MVT::i32);
286 /// getI8Imm - Return a target constant with the specified value, of type
288 inline SDValue getI8Imm(unsigned Imm, SDLoc DL) {
289 return CurDAG->getTargetConstant(Imm, DL, MVT::i8);
292 /// getI32Imm - Return a target constant with the specified value, of type
294 inline SDValue getI32Imm(unsigned Imm, SDLoc DL) {
295 return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
298 /// getGlobalBaseReg - Return an SDNode that returns the value of
299 /// the global base register. Output instructions required to
300 /// initialize the global base register, if necessary.
302 SDNode *getGlobalBaseReg();
304 /// getTargetMachine - Return a reference to the TargetMachine, casted
305 /// to the target-specific type.
306 const X86TargetMachine &getTargetMachine() const {
307 return static_cast<const X86TargetMachine &>(TM);
310 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
311 /// to the target-specific type.
312 const X86InstrInfo *getInstrInfo() const {
313 return Subtarget->getInstrInfo();
316 /// \brief Address-mode matching performs shift-of-and to and-of-shift
317 /// reassociation in order to expose more scaled addressing
319 bool ComplexPatternFuncMutatesDAG() const override {
327 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
328 if (OptLevel == CodeGenOpt::None) return false;
333 if (N.getOpcode() != ISD::LOAD)
336 // If N is a load, do additional profitability checks.
338 switch (U->getOpcode()) {
351 SDValue Op1 = U->getOperand(1);
353 // If the other operand is a 8-bit immediate we should fold the immediate
354 // instead. This reduces code size.
356 // movl 4(%esp), %eax
360 // addl 4(%esp), %eax
361 // The former is 2 bytes shorter. In case where the increment is 1, then
362 // the saving can be 4 bytes (by using incl %eax).
363 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
364 if (Imm->getAPIntValue().isSignedIntN(8))
367 // If the other operand is a TLS address, we should fold it instead.
370 // leal i@NTPOFF(%eax), %eax
372 // movl $i@NTPOFF, %eax
374 // if the block also has an access to a second TLS address this will save
376 // FIXME: This is probably also true for non-TLS addresses.
377 if (Op1.getOpcode() == X86ISD::Wrapper) {
378 SDValue Val = Op1.getOperand(0);
379 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
389 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
390 /// load's chain operand and move load below the call's chain operand.
391 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
392 SDValue Call, SDValue OrigChain) {
393 SmallVector<SDValue, 8> Ops;
394 SDValue Chain = OrigChain.getOperand(0);
395 if (Chain.getNode() == Load.getNode())
396 Ops.push_back(Load.getOperand(0));
398 assert(Chain.getOpcode() == ISD::TokenFactor &&
399 "Unexpected chain operand");
400 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
401 if (Chain.getOperand(i).getNode() == Load.getNode())
402 Ops.push_back(Load.getOperand(0));
404 Ops.push_back(Chain.getOperand(i));
406 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
408 Ops.push_back(NewChain);
410 Ops.append(OrigChain->op_begin() + 1, OrigChain->op_end());
411 CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
412 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
413 Load.getOperand(1), Load.getOperand(2));
416 Ops.push_back(SDValue(Load.getNode(), 1));
417 Ops.append(Call->op_begin() + 1, Call->op_end());
418 CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
421 /// isCalleeLoad - Return true if call address is a load and it can be
422 /// moved below CALLSEQ_START and the chains leading up to the call.
423 /// Return the CALLSEQ_START by reference as a second output.
424 /// In the case of a tail call, there isn't a callseq node between the call
425 /// chain and the load.
426 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
427 // The transformation is somewhat dangerous if the call's chain was glued to
428 // the call. After MoveBelowOrigChain the load is moved between the call and
429 // the chain, this can create a cycle if the load is not folded. So it is
430 // *really* important that we are sure the load will be folded.
431 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
433 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
436 LD->getAddressingMode() != ISD::UNINDEXED ||
437 LD->getExtensionType() != ISD::NON_EXTLOAD)
440 // Now let's find the callseq_start.
441 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
442 if (!Chain.hasOneUse())
444 Chain = Chain.getOperand(0);
447 if (!Chain.getNumOperands())
449 // Since we are not checking for AA here, conservatively abort if the chain
450 // writes to memory. It's not safe to move the callee (a load) across a store.
451 if (isa<MemSDNode>(Chain.getNode()) &&
452 cast<MemSDNode>(Chain.getNode())->writeMem())
454 if (Chain.getOperand(0).getNode() == Callee.getNode())
456 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
457 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
458 Callee.getValue(1).hasOneUse())
463 void X86DAGToDAGISel::PreprocessISelDAG() {
464 // OptForSize is used in pattern predicates that isel is matching.
465 // FIXME: Use Function::optForSize().
466 OptForSize = MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
468 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
469 E = CurDAG->allnodes_end(); I != E; ) {
470 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
472 if (OptLevel != CodeGenOpt::None &&
473 // Only does this when target favors doesn't favor register indirect
475 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
476 (N->getOpcode() == X86ISD::TC_RETURN &&
477 // Only does this if load can be folded into TC_RETURN.
478 (Subtarget->is64Bit() ||
479 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
480 /// Also try moving call address load from outside callseq_start to just
481 /// before the call to allow it to be folded.
499 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
500 SDValue Chain = N->getOperand(0);
501 SDValue Load = N->getOperand(1);
502 if (!isCalleeLoad(Load, Chain, HasCallSeq))
504 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
509 // Lower fpround and fpextend nodes that target the FP stack to be store and
510 // load to the stack. This is a gross hack. We would like to simply mark
511 // these as being illegal, but when we do that, legalize produces these when
512 // it expands calls, then expands these in the same legalize pass. We would
513 // like dag combine to be able to hack on these between the call expansion
514 // and the node legalization. As such this pass basically does "really
515 // late" legalization of these inline with the X86 isel pass.
516 // FIXME: This should only happen when not compiled with -O0.
517 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
520 MVT SrcVT = N->getOperand(0).getSimpleValueType();
521 MVT DstVT = N->getSimpleValueType(0);
523 // If any of the sources are vectors, no fp stack involved.
524 if (SrcVT.isVector() || DstVT.isVector())
527 // If the source and destination are SSE registers, then this is a legal
528 // conversion that should not be lowered.
529 const X86TargetLowering *X86Lowering =
530 static_cast<const X86TargetLowering *>(TLI);
531 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
532 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
533 if (SrcIsSSE && DstIsSSE)
536 if (!SrcIsSSE && !DstIsSSE) {
537 // If this is an FPStack extension, it is a noop.
538 if (N->getOpcode() == ISD::FP_EXTEND)
540 // If this is a value-preserving FPStack truncation, it is a noop.
541 if (N->getConstantOperandVal(1))
545 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
546 // FPStack has extload and truncstore. SSE can fold direct loads into other
547 // operations. Based on this, decide what we want to do.
549 if (N->getOpcode() == ISD::FP_ROUND)
550 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
552 MemVT = SrcIsSSE ? SrcVT : DstVT;
554 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
557 // FIXME: optimize the case where the src/dest is a load or store?
558 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
560 MemTmp, MachinePointerInfo(), MemVT,
562 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
563 MachinePointerInfo(),
564 MemVT, false, false, false, 0);
566 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
567 // extload we created. This will cause general havok on the dag because
568 // anything below the conversion could be folded into other existing nodes.
569 // To avoid invalidating 'I', back it up to the convert node.
571 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
573 // Now that we did that, the node is dead. Increment the iterator to the
574 // next node to process, then delete N.
576 CurDAG->DeleteNode(N);
581 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
582 /// the main function.
583 void X86DAGToDAGISel::EmitSpecialCodeForMain() {
584 if (Subtarget->isTargetCygMing()) {
585 TargetLowering::ArgListTy Args;
586 auto &DL = CurDAG->getDataLayout();
588 TargetLowering::CallLoweringInfo CLI(*CurDAG);
589 CLI.setChain(CurDAG->getRoot())
590 .setCallee(CallingConv::C, Type::getVoidTy(*CurDAG->getContext()),
591 CurDAG->getExternalSymbol("__main", TLI->getPointerTy(DL)),
593 const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
594 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
595 CurDAG->setRoot(Result.second);
599 void X86DAGToDAGISel::EmitFunctionEntryCode() {
600 // If this is main, emit special code for main.
601 if (const Function *Fn = MF->getFunction())
602 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
603 EmitSpecialCodeForMain();
606 static bool isDispSafeForFrameIndex(int64_t Val) {
607 // On 64-bit platforms, we can run into an issue where a frame index
608 // includes a displacement that, when added to the explicit displacement,
609 // will overflow the displacement field. Assuming that the frame index
610 // displacement fits into a 31-bit integer (which is only slightly more
611 // aggressive than the current fundamental assumption that it fits into
612 // a 32-bit integer), a 31-bit disp should always be safe.
613 return isInt<31>(Val);
616 bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
617 X86ISelAddressMode &AM) {
618 // Cannot combine ExternalSymbol displacements with integer offsets.
619 if (Offset != 0 && (AM.ES || AM.MCSym))
621 int64_t Val = AM.Disp + Offset;
622 CodeModel::Model M = TM.getCodeModel();
623 if (Subtarget->is64Bit()) {
624 if (!X86::isOffsetSuitableForCodeModel(Val, M,
625 AM.hasSymbolicDisplacement()))
627 // In addition to the checks required for a register base, check that
628 // we do not try to use an unsafe Disp with a frame index.
629 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
630 !isDispSafeForFrameIndex(Val))
638 bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
639 SDValue Address = N->getOperand(1);
641 // load gs:0 -> GS segment register.
642 // load fs:0 -> FS segment register.
644 // This optimization is valid because the GNU TLS model defines that
645 // gs:0 (or fs:0 on X86-64) contains its own address.
646 // For more information see http://people.redhat.com/drepper/tls.pdf
647 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
648 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
649 Subtarget->isTargetLinux())
650 switch (N->getPointerInfo().getAddrSpace()) {
652 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
655 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
662 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
663 /// into an addressing mode. These wrap things that will resolve down into a
664 /// symbol reference. If no match is possible, this returns true, otherwise it
666 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
667 // If the addressing mode already has a symbol as the displacement, we can
668 // never match another symbol.
669 if (AM.hasSymbolicDisplacement())
672 SDValue N0 = N.getOperand(0);
673 CodeModel::Model M = TM.getCodeModel();
675 // Handle X86-64 rip-relative addresses. We check this before checking direct
676 // folding because RIP is preferable to non-RIP accesses.
677 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
678 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
679 // they cannot be folded into immediate fields.
680 // FIXME: This can be improved for kernel and other models?
681 (M == CodeModel::Small || M == CodeModel::Kernel)) {
682 // Base and index reg must be 0 in order to use %rip as base.
683 if (AM.hasBaseOrIndexReg())
685 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
686 X86ISelAddressMode Backup = AM;
687 AM.GV = G->getGlobal();
688 AM.SymbolFlags = G->getTargetFlags();
689 if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
693 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
694 X86ISelAddressMode Backup = AM;
695 AM.CP = CP->getConstVal();
696 AM.Align = CP->getAlignment();
697 AM.SymbolFlags = CP->getTargetFlags();
698 if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
702 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
703 AM.ES = S->getSymbol();
704 AM.SymbolFlags = S->getTargetFlags();
705 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
706 AM.MCSym = S->getMCSymbol();
707 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
708 AM.JT = J->getIndex();
709 AM.SymbolFlags = J->getTargetFlags();
710 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
711 X86ISelAddressMode Backup = AM;
712 AM.BlockAddr = BA->getBlockAddress();
713 AM.SymbolFlags = BA->getTargetFlags();
714 if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
719 llvm_unreachable("Unhandled symbol reference node.");
721 if (N.getOpcode() == X86ISD::WrapperRIP)
722 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
726 // Handle the case when globals fit in our immediate field: This is true for
727 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
728 // mode, this only applies to a non-RIP-relative computation.
729 if (!Subtarget->is64Bit() ||
730 M == CodeModel::Small || M == CodeModel::Kernel) {
731 assert(N.getOpcode() != X86ISD::WrapperRIP &&
732 "RIP-relative addressing already handled");
733 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
734 AM.GV = G->getGlobal();
735 AM.Disp += G->getOffset();
736 AM.SymbolFlags = G->getTargetFlags();
737 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
738 AM.CP = CP->getConstVal();
739 AM.Align = CP->getAlignment();
740 AM.Disp += CP->getOffset();
741 AM.SymbolFlags = CP->getTargetFlags();
742 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
743 AM.ES = S->getSymbol();
744 AM.SymbolFlags = S->getTargetFlags();
745 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
746 AM.MCSym = S->getMCSymbol();
747 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
748 AM.JT = J->getIndex();
749 AM.SymbolFlags = J->getTargetFlags();
750 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
751 AM.BlockAddr = BA->getBlockAddress();
752 AM.Disp += BA->getOffset();
753 AM.SymbolFlags = BA->getTargetFlags();
755 llvm_unreachable("Unhandled symbol reference node.");
762 /// MatchAddress - Add the specified node to the specified addressing mode,
763 /// returning true if it cannot be done. This just pattern matches for the
765 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
766 if (MatchAddressRecursively(N, AM, 0))
769 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
770 // a smaller encoding and avoids a scaled-index.
772 AM.BaseType == X86ISelAddressMode::RegBase &&
773 AM.Base_Reg.getNode() == nullptr) {
774 AM.Base_Reg = AM.IndexReg;
778 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
779 // because it has a smaller encoding.
780 // TODO: Which other code models can use this?
781 if (TM.getCodeModel() == CodeModel::Small &&
782 Subtarget->is64Bit() &&
784 AM.BaseType == X86ISelAddressMode::RegBase &&
785 AM.Base_Reg.getNode() == nullptr &&
786 AM.IndexReg.getNode() == nullptr &&
787 AM.SymbolFlags == X86II::MO_NO_FLAG &&
788 AM.hasSymbolicDisplacement())
789 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
794 // Insert a node into the DAG at least before the Pos node's position. This
795 // will reposition the node as needed, and will assign it a node ID that is <=
796 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
797 // IDs! The selection DAG must no longer depend on their uniqueness when this
799 static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
800 if (N.getNode()->getNodeId() == -1 ||
801 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
802 DAG.RepositionNode(Pos.getNode(), N.getNode());
803 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
807 // Transform "(X >> (8-C1)) & (0xff << C1)" to "((X >> 8) & 0xff) << C1" if
808 // safe. This allows us to convert the shift and and into an h-register
809 // extract and a scaled index. Returns false if the simplification is
811 static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
813 SDValue Shift, SDValue X,
814 X86ISelAddressMode &AM) {
815 if (Shift.getOpcode() != ISD::SRL ||
816 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
820 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
821 if (ScaleLog <= 0 || ScaleLog >= 4 ||
822 Mask != (0xffu << ScaleLog))
825 MVT VT = N.getSimpleValueType();
827 SDValue Eight = DAG.getConstant(8, DL, MVT::i8);
828 SDValue NewMask = DAG.getConstant(0xff, DL, VT);
829 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
830 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
831 SDValue ShlCount = DAG.getConstant(ScaleLog, DL, MVT::i8);
832 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
834 // Insert the new nodes into the topological ordering. We must do this in
835 // a valid topological ordering as nothing is going to go back and re-sort
836 // these nodes. We continually insert before 'N' in sequence as this is
837 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
838 // hierarchy left to express.
839 InsertDAGNode(DAG, N, Eight);
840 InsertDAGNode(DAG, N, Srl);
841 InsertDAGNode(DAG, N, NewMask);
842 InsertDAGNode(DAG, N, And);
843 InsertDAGNode(DAG, N, ShlCount);
844 InsertDAGNode(DAG, N, Shl);
845 DAG.ReplaceAllUsesWith(N, Shl);
847 AM.Scale = (1 << ScaleLog);
851 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
852 // allows us to fold the shift into this addressing mode. Returns false if the
853 // transform succeeded.
854 static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
856 SDValue Shift, SDValue X,
857 X86ISelAddressMode &AM) {
858 if (Shift.getOpcode() != ISD::SHL ||
859 !isa<ConstantSDNode>(Shift.getOperand(1)))
862 // Not likely to be profitable if either the AND or SHIFT node has more
863 // than one use (unless all uses are for address computation). Besides,
864 // isel mechanism requires their node ids to be reused.
865 if (!N.hasOneUse() || !Shift.hasOneUse())
868 // Verify that the shift amount is something we can fold.
869 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
870 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
873 MVT VT = N.getSimpleValueType();
875 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, DL, VT);
876 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
877 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
879 // Insert the new nodes into the topological ordering. We must do this in
880 // a valid topological ordering as nothing is going to go back and re-sort
881 // these nodes. We continually insert before 'N' in sequence as this is
882 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
883 // hierarchy left to express.
884 InsertDAGNode(DAG, N, NewMask);
885 InsertDAGNode(DAG, N, NewAnd);
886 InsertDAGNode(DAG, N, NewShift);
887 DAG.ReplaceAllUsesWith(N, NewShift);
889 AM.Scale = 1 << ShiftAmt;
890 AM.IndexReg = NewAnd;
894 // Implement some heroics to detect shifts of masked values where the mask can
895 // be replaced by extending the shift and undoing that in the addressing mode
896 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
897 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
898 // the addressing mode. This results in code such as:
900 // int f(short *y, int *lookup_table) {
902 // return *y + lookup_table[*y >> 11];
906 // movzwl (%rdi), %eax
909 // addl (%rsi,%rcx,4), %eax
912 // movzwl (%rdi), %eax
916 // addl (%rsi,%rcx), %eax
918 // Note that this function assumes the mask is provided as a mask *after* the
919 // value is shifted. The input chain may or may not match that, but computing
920 // such a mask is trivial.
921 static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
923 SDValue Shift, SDValue X,
924 X86ISelAddressMode &AM) {
925 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
926 !isa<ConstantSDNode>(Shift.getOperand(1)))
929 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
930 unsigned MaskLZ = countLeadingZeros(Mask);
931 unsigned MaskTZ = countTrailingZeros(Mask);
933 // The amount of shift we're trying to fit into the addressing mode is taken
934 // from the trailing zeros of the mask.
935 unsigned AMShiftAmt = MaskTZ;
937 // There is nothing we can do here unless the mask is removing some bits.
938 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
939 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
941 // We also need to ensure that mask is a continuous run of bits.
942 if (countTrailingOnes(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
944 // Scale the leading zero count down based on the actual size of the value.
945 // Also scale it down based on the size of the shift.
946 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
948 // The final check is to ensure that any masked out high bits of X are
949 // already known to be zero. Otherwise, the mask has a semantic impact
950 // other than masking out a couple of low bits. Unfortunately, because of
951 // the mask, zero extensions will be removed from operands in some cases.
952 // This code works extra hard to look through extensions because we can
953 // replace them with zero extensions cheaply if necessary.
954 bool ReplacingAnyExtend = false;
955 if (X.getOpcode() == ISD::ANY_EXTEND) {
956 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
957 X.getOperand(0).getSimpleValueType().getSizeInBits();
958 // Assume that we'll replace the any-extend with a zero-extend, and
959 // narrow the search to the extended value.
961 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
962 ReplacingAnyExtend = true;
964 APInt MaskedHighBits =
965 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
966 APInt KnownZero, KnownOne;
967 DAG.computeKnownBits(X, KnownZero, KnownOne);
968 if (MaskedHighBits != KnownZero) return true;
970 // We've identified a pattern that can be transformed into a single shift
971 // and an addressing mode. Make it so.
972 MVT VT = N.getSimpleValueType();
973 if (ReplacingAnyExtend) {
974 assert(X.getValueType() != VT);
975 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
976 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
977 InsertDAGNode(DAG, N, NewX);
981 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, DL, MVT::i8);
982 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
983 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, DL, MVT::i8);
984 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
986 // Insert the new nodes into the topological ordering. We must do this in
987 // a valid topological ordering as nothing is going to go back and re-sort
988 // these nodes. We continually insert before 'N' in sequence as this is
989 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
990 // hierarchy left to express.
991 InsertDAGNode(DAG, N, NewSRLAmt);
992 InsertDAGNode(DAG, N, NewSRL);
993 InsertDAGNode(DAG, N, NewSHLAmt);
994 InsertDAGNode(DAG, N, NewSHL);
995 DAG.ReplaceAllUsesWith(N, NewSHL);
997 AM.Scale = 1 << AMShiftAmt;
998 AM.IndexReg = NewSRL;
1002 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
1006 dbgs() << "MatchAddress: ";
1011 return MatchAddressBase(N, AM);
1013 // If this is already a %rip relative address, we can only merge immediates
1014 // into it. Instead of handling this in every case, we handle it here.
1015 // RIP relative addressing: %rip + 32-bit displacement!
1016 if (AM.isRIPRelative()) {
1017 // FIXME: JumpTable and ExternalSymbol address currently don't like
1018 // displacements. It isn't very important, but this should be fixed for
1020 if (!(AM.ES || AM.MCSym) && AM.JT != -1)
1023 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
1024 if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
1029 switch (N.getOpcode()) {
1031 case ISD::LOCAL_RECOVER: {
1032 if (!AM.hasSymbolicDisplacement() && AM.Disp == 0)
1033 if (const auto *ESNode = dyn_cast<MCSymbolSDNode>(N.getOperand(0))) {
1034 // Use the symbol and don't prefix it.
1035 AM.MCSym = ESNode->getMCSymbol();
1040 case ISD::Constant: {
1041 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1042 if (!FoldOffsetIntoAddress(Val, AM))
1047 case X86ISD::Wrapper:
1048 case X86ISD::WrapperRIP:
1049 if (!MatchWrapper(N, AM))
1054 if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
1058 case ISD::FrameIndex:
1059 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1060 AM.Base_Reg.getNode() == nullptr &&
1061 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1062 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1063 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1069 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1073 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1074 unsigned Val = CN->getZExtValue();
1075 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1076 // that the base operand remains free for further matching. If
1077 // the base doesn't end up getting used, a post-processing step
1078 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1079 if (Val == 1 || Val == 2 || Val == 3) {
1080 AM.Scale = 1 << Val;
1081 SDValue ShVal = N.getNode()->getOperand(0);
1083 // Okay, we know that we have a scale by now. However, if the scaled
1084 // value is an add of something and a constant, we can fold the
1085 // constant into the disp field here.
1086 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1087 AM.IndexReg = ShVal.getNode()->getOperand(0);
1088 ConstantSDNode *AddVal =
1089 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1090 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1091 if (!FoldOffsetIntoAddress(Disp, AM))
1095 AM.IndexReg = ShVal;
1102 // Scale must not be used already.
1103 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1105 SDValue And = N.getOperand(0);
1106 if (And.getOpcode() != ISD::AND) break;
1107 SDValue X = And.getOperand(0);
1109 // We only handle up to 64-bit values here as those are what matter for
1110 // addressing mode optimizations.
1111 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1113 // The mask used for the transform is expected to be post-shift, but we
1114 // found the shift first so just apply the shift to the mask before passing
1116 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1117 !isa<ConstantSDNode>(And.getOperand(1)))
1119 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1121 // Try to fold the mask and shift into the scale, and return false if we
1123 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1128 case ISD::SMUL_LOHI:
1129 case ISD::UMUL_LOHI:
1130 // A mul_lohi where we need the low part can be folded as a plain multiply.
1131 if (N.getResNo() != 0) break;
1134 case X86ISD::MUL_IMM:
1135 // X*[3,5,9] -> X+X*[2,4,8]
1136 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1137 AM.Base_Reg.getNode() == nullptr &&
1138 AM.IndexReg.getNode() == nullptr) {
1140 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1141 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1142 CN->getZExtValue() == 9) {
1143 AM.Scale = unsigned(CN->getZExtValue())-1;
1145 SDValue MulVal = N.getNode()->getOperand(0);
1148 // Okay, we know that we have a scale by now. However, if the scaled
1149 // value is an add of something and a constant, we can fold the
1150 // constant into the disp field here.
1151 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1152 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1153 Reg = MulVal.getNode()->getOperand(0);
1154 ConstantSDNode *AddVal =
1155 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1156 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1157 if (FoldOffsetIntoAddress(Disp, AM))
1158 Reg = N.getNode()->getOperand(0);
1160 Reg = N.getNode()->getOperand(0);
1163 AM.IndexReg = AM.Base_Reg = Reg;
1170 // Given A-B, if A can be completely folded into the address and
1171 // the index field with the index field unused, use -B as the index.
1172 // This is a win if a has multiple parts that can be folded into
1173 // the address. Also, this saves a mov if the base register has
1174 // other uses, since it avoids a two-address sub instruction, however
1175 // it costs an additional mov if the index register has other uses.
1177 // Add an artificial use to this node so that we can keep track of
1178 // it if it gets CSE'd with a different node.
1179 HandleSDNode Handle(N);
1181 // Test if the LHS of the sub can be folded.
1182 X86ISelAddressMode Backup = AM;
1183 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1187 // Test if the index field is free for use.
1188 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1194 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1195 // If the RHS involves a register with multiple uses, this
1196 // transformation incurs an extra mov, due to the neg instruction
1197 // clobbering its operand.
1198 if (!RHS.getNode()->hasOneUse() ||
1199 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1200 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1201 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1202 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1203 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1205 // If the base is a register with multiple uses, this
1206 // transformation may save a mov.
1207 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1208 AM.Base_Reg.getNode() &&
1209 !AM.Base_Reg.getNode()->hasOneUse()) ||
1210 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1212 // If the folded LHS was interesting, this transformation saves
1213 // address arithmetic.
1214 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1215 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1216 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1218 // If it doesn't look like it may be an overall win, don't do it.
1224 // Ok, the transformation is legal and appears profitable. Go for it.
1225 SDValue Zero = CurDAG->getConstant(0, dl, N.getValueType());
1226 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1230 // Insert the new nodes into the topological ordering.
1231 InsertDAGNode(*CurDAG, N, Zero);
1232 InsertDAGNode(*CurDAG, N, Neg);
1237 // Add an artificial use to this node so that we can keep track of
1238 // it if it gets CSE'd with a different node.
1239 HandleSDNode Handle(N);
1241 X86ISelAddressMode Backup = AM;
1242 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1243 !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1247 // Try again after commuting the operands.
1248 if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1249 !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1253 // If we couldn't fold both operands into the address at the same time,
1254 // see if we can just put each operand into a register and fold at least
1256 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1257 !AM.Base_Reg.getNode() &&
1258 !AM.IndexReg.getNode()) {
1259 N = Handle.getValue();
1260 AM.Base_Reg = N.getOperand(0);
1261 AM.IndexReg = N.getOperand(1);
1265 N = Handle.getValue();
1270 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1271 if (CurDAG->isBaseWithConstantOffset(N)) {
1272 X86ISelAddressMode Backup = AM;
1273 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1275 // Start with the LHS as an addr mode.
1276 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1277 !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
1284 // Perform some heroic transforms on an and of a constant-count shift
1285 // with a constant to enable use of the scaled offset field.
1287 // Scale must not be used already.
1288 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1290 SDValue Shift = N.getOperand(0);
1291 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1292 SDValue X = Shift.getOperand(0);
1294 // We only handle up to 64-bit values here as those are what matter for
1295 // addressing mode optimizations.
1296 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1298 if (!isa<ConstantSDNode>(N.getOperand(1)))
1300 uint64_t Mask = N.getConstantOperandVal(1);
1302 // Try to fold the mask and shift into an extract and scale.
1303 if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1306 // Try to fold the mask and shift directly into the scale.
1307 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1310 // Try to swap the mask and shift to place shifts which can be done as
1311 // a scale on the outside of the mask.
1312 if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1318 return MatchAddressBase(N, AM);
1321 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1322 /// specified addressing mode without any further recursion.
1323 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1324 // Is the base register already occupied?
1325 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1326 // If so, check to see if the scale index register is set.
1327 if (!AM.IndexReg.getNode()) {
1333 // Otherwise, we cannot select it.
1337 // Default, generate it as a register.
1338 AM.BaseType = X86ISelAddressMode::RegBase;
1343 bool X86DAGToDAGISel::SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
1344 SDValue &Scale, SDValue &Index,
1345 SDValue &Disp, SDValue &Segment) {
1347 MaskedGatherScatterSDNode *Mgs = dyn_cast<MaskedGatherScatterSDNode>(Parent);
1350 X86ISelAddressMode AM;
1351 unsigned AddrSpace = Mgs->getPointerInfo().getAddrSpace();
1352 // AddrSpace 256 -> GS, 257 -> FS.
1353 if (AddrSpace == 256)
1354 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1355 if (AddrSpace == 257)
1356 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1359 Base = Mgs->getBasePtr();
1360 Index = Mgs->getIndex();
1361 unsigned ScalarSize = Mgs->getValue().getValueType().getScalarSizeInBits();
1362 Scale = getI8Imm(ScalarSize/8, DL);
1364 // If Base is 0, the whole address is in index and the Scale is 1
1365 if (isa<ConstantSDNode>(Base)) {
1366 assert(dyn_cast<ConstantSDNode>(Base)->isNullValue() &&
1367 "Unexpected base in gather/scatter");
1368 Scale = getI8Imm(1, DL);
1369 Base = CurDAG->getRegister(0, MVT::i32);
1371 if (AM.Segment.getNode())
1372 Segment = AM.Segment;
1374 Segment = CurDAG->getRegister(0, MVT::i32);
1375 Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
1379 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1380 /// It returns the operands which make up the maximal addressing mode it can
1381 /// match by reference.
1383 /// Parent is the parent node of the addr operand that is being matched. It
1384 /// is always a load, store, atomic node, or null. It is only null when
1385 /// checking memory operands for inline asm nodes.
1386 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1387 SDValue &Scale, SDValue &Index,
1388 SDValue &Disp, SDValue &Segment) {
1389 X86ISelAddressMode AM;
1392 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1393 // that are not a MemSDNode, and thus don't have proper addrspace info.
1394 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1395 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1396 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1397 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1398 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1399 unsigned AddrSpace =
1400 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1401 // AddrSpace 256 -> GS, 257 -> FS.
1402 if (AddrSpace == 256)
1403 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1404 if (AddrSpace == 257)
1405 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1408 if (MatchAddress(N, AM))
1411 MVT VT = N.getSimpleValueType();
1412 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1413 if (!AM.Base_Reg.getNode())
1414 AM.Base_Reg = CurDAG->getRegister(0, VT);
1417 if (!AM.IndexReg.getNode())
1418 AM.IndexReg = CurDAG->getRegister(0, VT);
1420 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1424 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1425 /// match a load whose top elements are either undef or zeros. The load flavor
1426 /// is derived from the type of N, which is either v4f32 or v2f64.
1429 /// PatternChainNode: this is the matched node that has a chain input and
1431 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1432 SDValue N, SDValue &Base,
1433 SDValue &Scale, SDValue &Index,
1434 SDValue &Disp, SDValue &Segment,
1435 SDValue &PatternNodeWithChain) {
1436 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1437 PatternNodeWithChain = N.getOperand(0);
1438 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1439 PatternNodeWithChain.hasOneUse() &&
1440 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1441 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1442 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1443 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1449 // Also handle the case where we explicitly require zeros in the top
1450 // elements. This is a vector shuffle from the zero vector.
1451 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1452 // Check to see if the top elements are all zeros (or bitcast of zeros).
1453 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1454 N.getOperand(0).getNode()->hasOneUse() &&
1455 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1456 N.getOperand(0).getOperand(0).hasOneUse() &&
1457 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1458 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1459 // Okay, this is a zero extending load. Fold it.
1460 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1461 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1463 PatternNodeWithChain = SDValue(LD, 0);
1470 bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
1471 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1472 uint64_t ImmVal = CN->getZExtValue();
1473 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1476 Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i64);
1480 // In static codegen with small code model, we can get the address of a label
1481 // into a register with 'movl'. TableGen has already made sure we're looking
1482 // at a label of some kind.
1483 assert(N->getOpcode() == X86ISD::Wrapper &&
1484 "Unexpected node type for MOV32ri64");
1485 N = N.getOperand(0);
1487 if (N->getOpcode() != ISD::TargetConstantPool &&
1488 N->getOpcode() != ISD::TargetJumpTable &&
1489 N->getOpcode() != ISD::TargetGlobalAddress &&
1490 N->getOpcode() != ISD::TargetExternalSymbol &&
1491 N->getOpcode() != ISD::MCSymbol &&
1492 N->getOpcode() != ISD::TargetBlockAddress)
1496 return TM.getCodeModel() == CodeModel::Small;
1499 bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
1500 SDValue &Scale, SDValue &Index,
1501 SDValue &Disp, SDValue &Segment) {
1502 if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1506 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1507 if (RN && RN->getReg() == 0)
1508 Base = CurDAG->getRegister(0, MVT::i64);
1509 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
1510 // Base could already be %rip, particularly in the x32 ABI.
1511 Base = SDValue(CurDAG->getMachineNode(
1512 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1513 CurDAG->getTargetConstant(0, DL, MVT::i64),
1515 CurDAG->getTargetConstant(X86::sub_32bit, DL, MVT::i32)),
1519 RN = dyn_cast<RegisterSDNode>(Index);
1520 if (RN && RN->getReg() == 0)
1521 Index = CurDAG->getRegister(0, MVT::i64);
1523 assert(Index.getValueType() == MVT::i32 &&
1524 "Expect to be extending 32-bit registers for use in LEA");
1525 Index = SDValue(CurDAG->getMachineNode(
1526 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1527 CurDAG->getTargetConstant(0, DL, MVT::i64),
1529 CurDAG->getTargetConstant(X86::sub_32bit, DL,
1537 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1538 /// mode it matches can be cost effectively emitted as an LEA instruction.
1539 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1540 SDValue &Base, SDValue &Scale,
1541 SDValue &Index, SDValue &Disp,
1543 X86ISelAddressMode AM;
1545 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1547 SDValue Copy = AM.Segment;
1548 SDValue T = CurDAG->getRegister(0, MVT::i32);
1550 if (MatchAddress(N, AM))
1552 assert (T == AM.Segment);
1555 MVT VT = N.getSimpleValueType();
1556 unsigned Complexity = 0;
1557 if (AM.BaseType == X86ISelAddressMode::RegBase)
1558 if (AM.Base_Reg.getNode())
1561 AM.Base_Reg = CurDAG->getRegister(0, VT);
1562 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1565 if (AM.IndexReg.getNode())
1568 AM.IndexReg = CurDAG->getRegister(0, VT);
1570 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1575 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1576 // to a LEA. This is determined with some expermentation but is by no means
1577 // optimal (especially for code size consideration). LEA is nice because of
1578 // its three-address nature. Tweak the cost function again when we can run
1579 // convertToThreeAddress() at register allocation time.
1580 if (AM.hasSymbolicDisplacement()) {
1581 // For X86-64, we should always use lea to materialize RIP relative
1583 if (Subtarget->is64Bit())
1589 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1592 // If it isn't worth using an LEA, reject it.
1593 if (Complexity <= 2)
1596 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1600 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1601 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1602 SDValue &Scale, SDValue &Index,
1603 SDValue &Disp, SDValue &Segment) {
1604 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1605 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1607 X86ISelAddressMode AM;
1608 AM.GV = GA->getGlobal();
1609 AM.Disp += GA->getOffset();
1610 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1611 AM.SymbolFlags = GA->getTargetFlags();
1613 if (N.getValueType() == MVT::i32) {
1615 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1617 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1620 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1625 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1626 SDValue &Base, SDValue &Scale,
1627 SDValue &Index, SDValue &Disp,
1629 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1630 !IsProfitableToFold(N, P, P) ||
1631 !IsLegalToFold(N, P, P, OptLevel))
1634 return SelectAddr(N.getNode(),
1635 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1638 /// getGlobalBaseReg - Return an SDNode that returns the value of
1639 /// the global base register. Output instructions required to
1640 /// initialize the global base register, if necessary.
1642 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1643 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1644 auto &DL = MF->getDataLayout();
1645 return CurDAG->getRegister(GlobalBaseReg, TLI->getPointerTy(DL)).getNode();
1648 /// Atomic opcode table
1676 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1687 X86::LOCK_ADD64mi32,
1700 X86::LOCK_SUB64mi32,
1752 X86::LOCK_AND64mi32,
1765 X86::LOCK_XOR64mi32,
1770 // Return the target constant operand for atomic-load-op and do simple
1771 // translations, such as from atomic-load-add to lock-sub. The return value is
1772 // one of the following 3 cases:
1773 // + target-constant, the operand could be supported as a target constant.
1774 // + empty, the operand is not needed any more with the new op selected.
1775 // + non-empty, otherwise.
1776 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1778 enum AtomicOpc &Op, MVT NVT,
1780 const X86Subtarget *Subtarget) {
1781 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1782 int64_t CNVal = CN->getSExtValue();
1783 // Quit if not 32-bit imm.
1784 if ((int32_t)CNVal != CNVal)
1786 // Quit if INT32_MIN: it would be negated as it is negative and overflow,
1787 // producing an immediate that does not fit in the 32 bits available for
1788 // an immediate operand to sub. However, it still fits in 32 bits for the
1789 // add (since it is not negated) so we can return target-constant.
1790 if (CNVal == INT32_MIN)
1791 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1792 // For atomic-load-add, we could do some optimizations.
1794 // Translate to INC/DEC if ADD by 1 or -1.
1795 if (((CNVal == 1) || (CNVal == -1)) && !Subtarget->slowIncDec()) {
1796 Op = (CNVal == 1) ? INC : DEC;
1797 // No more constant operand after being translated into INC/DEC.
1800 // Translate to SUB if ADD by negative value.
1806 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1809 // If the value operand is single-used, try to optimize it.
1810 if (Op == ADD && Val.hasOneUse()) {
1811 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1812 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1814 return Val.getOperand(1);
1816 // A special case for i16, which needs truncating as, in most cases, it's
1817 // promoted to i32. We will translate
1818 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1819 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1820 Val.getOperand(0).getOpcode() == ISD::SUB &&
1821 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1823 Val = Val.getOperand(0);
1824 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1832 SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
1833 if (Node->hasAnyUseOfValue(0))
1838 // Optimize common patterns for __sync_or_and_fetch and similar arith
1839 // operations where the result is not used. This allows us to use the "lock"
1840 // version of the arithmetic instruction.
1841 SDValue Chain = Node->getOperand(0);
1842 SDValue Ptr = Node->getOperand(1);
1843 SDValue Val = Node->getOperand(2);
1844 SDValue Base, Scale, Index, Disp, Segment;
1845 if (!SelectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
1848 // Which index into the table.
1850 switch (Node->getOpcode()) {
1853 case ISD::ATOMIC_LOAD_OR:
1856 case ISD::ATOMIC_LOAD_AND:
1859 case ISD::ATOMIC_LOAD_XOR:
1862 case ISD::ATOMIC_LOAD_ADD:
1867 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val, Subtarget);
1868 bool isUnOp = !Val.getNode();
1869 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1872 switch (NVT.SimpleTy) {
1873 default: return nullptr;
1876 Opc = AtomicOpcTbl[Op][ConstantI8];
1878 Opc = AtomicOpcTbl[Op][I8];
1882 if (immSext8(Val.getNode()))
1883 Opc = AtomicOpcTbl[Op][SextConstantI16];
1885 Opc = AtomicOpcTbl[Op][ConstantI16];
1887 Opc = AtomicOpcTbl[Op][I16];
1891 if (immSext8(Val.getNode()))
1892 Opc = AtomicOpcTbl[Op][SextConstantI32];
1894 Opc = AtomicOpcTbl[Op][ConstantI32];
1896 Opc = AtomicOpcTbl[Op][I32];
1900 if (immSext8(Val.getNode()))
1901 Opc = AtomicOpcTbl[Op][SextConstantI64];
1902 else if (i64immSExt32(Val.getNode()))
1903 Opc = AtomicOpcTbl[Op][ConstantI64];
1905 llvm_unreachable("True 64 bits constant in SelectAtomicLoadArith");
1907 Opc = AtomicOpcTbl[Op][I64];
1911 assert(Opc != 0 && "Invalid arith lock transform!");
1913 // Building the new node.
1916 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Chain };
1917 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1919 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Val, Chain };
1920 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1923 // Copying the MachineMemOperand.
1924 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1925 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1926 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1928 // We need to have two outputs as that is what the original instruction had.
1929 // So we add a dummy, undefined output. This is safe as we checked first
1930 // that no-one uses our output anyway.
1931 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1933 SDValue RetVals[] = { Undef, Ret };
1934 return CurDAG->getMergeValues(RetVals, dl).getNode();
1937 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1938 /// any uses which require the SF or OF bits to be accurate.
1939 static bool HasNoSignedComparisonUses(SDNode *N) {
1940 // Examine each user of the node.
1941 for (SDNode::use_iterator UI = N->use_begin(),
1942 UE = N->use_end(); UI != UE; ++UI) {
1943 // Only examine CopyToReg uses.
1944 if (UI->getOpcode() != ISD::CopyToReg)
1946 // Only examine CopyToReg uses that copy to EFLAGS.
1947 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1950 // Examine each user of the CopyToReg use.
1951 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1952 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1953 // Only examine the Flag result.
1954 if (FlagUI.getUse().getResNo() != 1) continue;
1955 // Anything unusual: assume conservatively.
1956 if (!FlagUI->isMachineOpcode()) return false;
1957 // Examine the opcode of the user.
1958 switch (FlagUI->getMachineOpcode()) {
1959 // These comparisons don't treat the most significant bit specially.
1960 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1961 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1962 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1963 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1964 case X86::JA_1: case X86::JAE_1: case X86::JB_1: case X86::JBE_1:
1965 case X86::JE_1: case X86::JNE_1: case X86::JP_1: case X86::JNP_1:
1966 case X86::CMOVA16rr: case X86::CMOVA16rm:
1967 case X86::CMOVA32rr: case X86::CMOVA32rm:
1968 case X86::CMOVA64rr: case X86::CMOVA64rm:
1969 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1970 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1971 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1972 case X86::CMOVB16rr: case X86::CMOVB16rm:
1973 case X86::CMOVB32rr: case X86::CMOVB32rm:
1974 case X86::CMOVB64rr: case X86::CMOVB64rm:
1975 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1976 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1977 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1978 case X86::CMOVE16rr: case X86::CMOVE16rm:
1979 case X86::CMOVE32rr: case X86::CMOVE32rm:
1980 case X86::CMOVE64rr: case X86::CMOVE64rm:
1981 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1982 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1983 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1984 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1985 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1986 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1987 case X86::CMOVP16rr: case X86::CMOVP16rm:
1988 case X86::CMOVP32rr: case X86::CMOVP32rm:
1989 case X86::CMOVP64rr: case X86::CMOVP64rm:
1991 // Anything else: assume conservatively.
1992 default: return false;
1999 /// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
2000 /// is suitable for doing the {load; increment or decrement; store} to modify
2002 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
2003 SDValue StoredVal, SelectionDAG *CurDAG,
2004 LoadSDNode* &LoadNode, SDValue &InputChain) {
2006 // is the value stored the result of a DEC or INC?
2007 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
2009 // is the stored value result 0 of the load?
2010 if (StoredVal.getResNo() != 0) return false;
2012 // are there other uses of the loaded value than the inc or dec?
2013 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
2015 // is the store non-extending and non-indexed?
2016 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
2019 SDValue Load = StoredVal->getOperand(0);
2020 // Is the stored value a non-extending and non-indexed load?
2021 if (!ISD::isNormalLoad(Load.getNode())) return false;
2023 // Return LoadNode by reference.
2024 LoadNode = cast<LoadSDNode>(Load);
2025 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
2026 EVT LdVT = LoadNode->getMemoryVT();
2027 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
2031 // Is store the only read of the loaded value?
2032 if (!Load.hasOneUse())
2035 // Is the address of the store the same as the load?
2036 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
2037 LoadNode->getOffset() != StoreNode->getOffset())
2040 // Check if the chain is produced by the load or is a TokenFactor with
2041 // the load output chain as an operand. Return InputChain by reference.
2042 SDValue Chain = StoreNode->getChain();
2044 bool ChainCheck = false;
2045 if (Chain == Load.getValue(1)) {
2047 InputChain = LoadNode->getChain();
2048 } else if (Chain.getOpcode() == ISD::TokenFactor) {
2049 SmallVector<SDValue, 4> ChainOps;
2050 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
2051 SDValue Op = Chain.getOperand(i);
2052 if (Op == Load.getValue(1)) {
2057 // Make sure using Op as part of the chain would not cause a cycle here.
2058 // In theory, we could check whether the chain node is a predecessor of
2059 // the load. But that can be very expensive. Instead visit the uses and
2060 // make sure they all have smaller node id than the load.
2061 int LoadId = LoadNode->getNodeId();
2062 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
2063 UE = UI->use_end(); UI != UE; ++UI) {
2064 if (UI.getUse().getResNo() != 0)
2066 if (UI->getNodeId() > LoadId)
2070 ChainOps.push_back(Op);
2074 // Make a new TokenFactor with all the other input chains except
2076 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
2077 MVT::Other, ChainOps);
2085 /// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
2086 /// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
2087 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2088 if (Opc == X86ISD::DEC) {
2089 if (LdVT == MVT::i64) return X86::DEC64m;
2090 if (LdVT == MVT::i32) return X86::DEC32m;
2091 if (LdVT == MVT::i16) return X86::DEC16m;
2092 if (LdVT == MVT::i8) return X86::DEC8m;
2094 assert(Opc == X86ISD::INC && "unrecognized opcode");
2095 if (LdVT == MVT::i64) return X86::INC64m;
2096 if (LdVT == MVT::i32) return X86::INC32m;
2097 if (LdVT == MVT::i16) return X86::INC16m;
2098 if (LdVT == MVT::i8) return X86::INC8m;
2100 llvm_unreachable("unrecognized size for LdVT");
2103 /// SelectGather - Customized ISel for GATHER operations.
2105 SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
2106 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2107 SDValue Chain = Node->getOperand(0);
2108 SDValue VSrc = Node->getOperand(2);
2109 SDValue Base = Node->getOperand(3);
2110 SDValue VIdx = Node->getOperand(4);
2111 SDValue VMask = Node->getOperand(5);
2112 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2116 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2121 // Memory Operands: Base, Scale, Index, Disp, Segment
2122 SDValue Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
2123 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2124 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue(), DL), VIdx,
2125 Disp, Segment, VMask, Chain};
2126 SDNode *ResNode = CurDAG->getMachineNode(Opc, DL, VTs, Ops);
2127 // Node has 2 outputs: VDst and MVT::Other.
2128 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2129 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2131 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2132 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2136 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2137 MVT NVT = Node->getSimpleValueType(0);
2139 unsigned Opcode = Node->getOpcode();
2142 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2144 if (Node->isMachineOpcode()) {
2145 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2146 Node->setNodeId(-1);
2147 return nullptr; // Already selected.
2152 case ISD::INTRINSIC_W_CHAIN: {
2153 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2156 case Intrinsic::x86_avx2_gather_d_pd:
2157 case Intrinsic::x86_avx2_gather_d_pd_256:
2158 case Intrinsic::x86_avx2_gather_q_pd:
2159 case Intrinsic::x86_avx2_gather_q_pd_256:
2160 case Intrinsic::x86_avx2_gather_d_ps:
2161 case Intrinsic::x86_avx2_gather_d_ps_256:
2162 case Intrinsic::x86_avx2_gather_q_ps:
2163 case Intrinsic::x86_avx2_gather_q_ps_256:
2164 case Intrinsic::x86_avx2_gather_d_q:
2165 case Intrinsic::x86_avx2_gather_d_q_256:
2166 case Intrinsic::x86_avx2_gather_q_q:
2167 case Intrinsic::x86_avx2_gather_q_q_256:
2168 case Intrinsic::x86_avx2_gather_d_d:
2169 case Intrinsic::x86_avx2_gather_d_d_256:
2170 case Intrinsic::x86_avx2_gather_q_d:
2171 case Intrinsic::x86_avx2_gather_q_d_256: {
2172 if (!Subtarget->hasAVX2())
2176 default: llvm_unreachable("Impossible intrinsic");
2177 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2178 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2179 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2180 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2181 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2182 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2183 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2184 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2185 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2186 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2187 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2188 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2189 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2190 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2191 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2192 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2194 SDNode *RetVal = SelectGather(Node, Opc);
2196 // We already called ReplaceUses inside SelectGather.
2203 case X86ISD::GlobalBaseReg:
2204 return getGlobalBaseReg();
2206 case X86ISD::SHRUNKBLEND: {
2207 // SHRUNKBLEND selects like a regular VSELECT.
2208 SDValue VSelect = CurDAG->getNode(
2209 ISD::VSELECT, SDLoc(Node), Node->getValueType(0), Node->getOperand(0),
2210 Node->getOperand(1), Node->getOperand(2));
2211 ReplaceUses(SDValue(Node, 0), VSelect);
2212 SelectCode(VSelect.getNode());
2213 // We already called ReplaceUses.
2217 case ISD::ATOMIC_LOAD_XOR:
2218 case ISD::ATOMIC_LOAD_AND:
2219 case ISD::ATOMIC_LOAD_OR:
2220 case ISD::ATOMIC_LOAD_ADD: {
2221 SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
2229 // For operations of the form (x << C1) op C2, check if we can use a smaller
2230 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2231 SDValue N0 = Node->getOperand(0);
2232 SDValue N1 = Node->getOperand(1);
2234 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2237 // i8 is unshrinkable, i16 should be promoted to i32.
2238 if (NVT != MVT::i32 && NVT != MVT::i64)
2241 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2242 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2243 if (!Cst || !ShlCst)
2246 int64_t Val = Cst->getSExtValue();
2247 uint64_t ShlVal = ShlCst->getZExtValue();
2249 // Make sure that we don't change the operation by removing bits.
2250 // This only matters for OR and XOR, AND is unaffected.
2251 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2252 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2255 unsigned ShlOp, AddOp, Op;
2258 // Check the minimum bitwidth for the new constant.
2259 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2260 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2261 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2262 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2264 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2267 // Bail if there is no smaller encoding.
2271 switch (NVT.SimpleTy) {
2272 default: llvm_unreachable("Unsupported VT!");
2274 assert(CstVT == MVT::i8);
2275 ShlOp = X86::SHL32ri;
2276 AddOp = X86::ADD32rr;
2279 default: llvm_unreachable("Impossible opcode");
2280 case ISD::AND: Op = X86::AND32ri8; break;
2281 case ISD::OR: Op = X86::OR32ri8; break;
2282 case ISD::XOR: Op = X86::XOR32ri8; break;
2286 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2287 ShlOp = X86::SHL64ri;
2288 AddOp = X86::ADD64rr;
2291 default: llvm_unreachable("Impossible opcode");
2292 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2293 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2294 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2299 // Emit the smaller op and the shift.
2300 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, dl, CstVT);
2301 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2303 return CurDAG->SelectNodeTo(Node, AddOp, NVT, SDValue(New, 0),
2305 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2306 getI8Imm(ShlVal, dl));
2309 case X86ISD::SMUL8: {
2310 SDValue N0 = Node->getOperand(0);
2311 SDValue N1 = Node->getOperand(1);
2313 Opc = (Opcode == X86ISD::SMUL8 ? X86::IMUL8r : X86::MUL8r);
2315 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::AL,
2316 N0, SDValue()).getValue(1);
2318 SDVTList VTs = CurDAG->getVTList(NVT, MVT::i32);
2319 SDValue Ops[] = {N1, InFlag};
2320 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2322 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2323 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2327 case X86ISD::UMUL: {
2328 SDValue N0 = Node->getOperand(0);
2329 SDValue N1 = Node->getOperand(1);
2332 switch (NVT.SimpleTy) {
2333 default: llvm_unreachable("Unsupported VT!");
2334 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2335 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2336 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2337 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2340 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2341 N0, SDValue()).getValue(1);
2343 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2344 SDValue Ops[] = {N1, InFlag};
2345 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2347 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2348 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2349 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2353 case ISD::SMUL_LOHI:
2354 case ISD::UMUL_LOHI: {
2355 SDValue N0 = Node->getOperand(0);
2356 SDValue N1 = Node->getOperand(1);
2358 bool isSigned = Opcode == ISD::SMUL_LOHI;
2359 bool hasBMI2 = Subtarget->hasBMI2();
2361 switch (NVT.SimpleTy) {
2362 default: llvm_unreachable("Unsupported VT!");
2363 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2364 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2365 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2366 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2367 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2368 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2371 switch (NVT.SimpleTy) {
2372 default: llvm_unreachable("Unsupported VT!");
2373 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2374 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2375 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2376 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2380 unsigned SrcReg, LoReg, HiReg;
2382 default: llvm_unreachable("Unknown MUL opcode!");
2385 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2389 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2393 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2397 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2400 SrcReg = X86::EDX; LoReg = HiReg = 0;
2403 SrcReg = X86::RDX; LoReg = HiReg = 0;
2407 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2408 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2409 // Multiply is commmutative.
2411 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2416 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2417 N0, SDValue()).getValue(1);
2418 SDValue ResHi, ResLo;
2422 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2424 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2425 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2426 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2427 ResHi = SDValue(CNode, 0);
2428 ResLo = SDValue(CNode, 1);
2429 Chain = SDValue(CNode, 2);
2430 InFlag = SDValue(CNode, 3);
2432 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2433 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2434 Chain = SDValue(CNode, 0);
2435 InFlag = SDValue(CNode, 1);
2438 // Update the chain.
2439 ReplaceUses(N1.getValue(1), Chain);
2441 SDValue Ops[] = { N1, InFlag };
2442 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2443 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2444 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2445 ResHi = SDValue(CNode, 0);
2446 ResLo = SDValue(CNode, 1);
2447 InFlag = SDValue(CNode, 2);
2449 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2450 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2451 InFlag = SDValue(CNode, 0);
2455 // Prevent use of AH in a REX instruction by referencing AX instead.
2456 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2457 !SDValue(Node, 1).use_empty()) {
2458 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2459 X86::AX, MVT::i16, InFlag);
2460 InFlag = Result.getValue(2);
2461 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2463 if (!SDValue(Node, 0).use_empty())
2464 ReplaceUses(SDValue(Node, 1),
2465 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2467 // Shift AX down 8 bits.
2468 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2470 CurDAG->getTargetConstant(8, dl, MVT::i8)),
2472 // Then truncate it down to i8.
2473 ReplaceUses(SDValue(Node, 1),
2474 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2476 // Copy the low half of the result, if it is needed.
2477 if (!SDValue(Node, 0).use_empty()) {
2478 if (!ResLo.getNode()) {
2479 assert(LoReg && "Register for low half is not defined!");
2480 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2482 InFlag = ResLo.getValue(2);
2484 ReplaceUses(SDValue(Node, 0), ResLo);
2485 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2487 // Copy the high half of the result, if it is needed.
2488 if (!SDValue(Node, 1).use_empty()) {
2489 if (!ResHi.getNode()) {
2490 assert(HiReg && "Register for high half is not defined!");
2491 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2493 InFlag = ResHi.getValue(2);
2495 ReplaceUses(SDValue(Node, 1), ResHi);
2496 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2504 case X86ISD::SDIVREM8_SEXT_HREG:
2505 case X86ISD::UDIVREM8_ZEXT_HREG: {
2506 SDValue N0 = Node->getOperand(0);
2507 SDValue N1 = Node->getOperand(1);
2509 bool isSigned = (Opcode == ISD::SDIVREM ||
2510 Opcode == X86ISD::SDIVREM8_SEXT_HREG);
2512 switch (NVT.SimpleTy) {
2513 default: llvm_unreachable("Unsupported VT!");
2514 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2515 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2516 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2517 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2520 switch (NVT.SimpleTy) {
2521 default: llvm_unreachable("Unsupported VT!");
2522 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2523 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2524 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2525 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2529 unsigned LoReg, HiReg, ClrReg;
2530 unsigned SExtOpcode;
2531 switch (NVT.SimpleTy) {
2532 default: llvm_unreachable("Unsupported VT!");
2534 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2535 SExtOpcode = X86::CBW;
2538 LoReg = X86::AX; HiReg = X86::DX;
2540 SExtOpcode = X86::CWD;
2543 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2544 SExtOpcode = X86::CDQ;
2547 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2548 SExtOpcode = X86::CQO;
2552 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2553 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2554 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2557 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2558 // Special case for div8, just use a move with zero extension to AX to
2559 // clear the upper 8 bits (AH).
2560 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2561 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2562 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2564 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2565 MVT::Other, Ops), 0);
2566 Chain = Move.getValue(1);
2567 ReplaceUses(N0.getValue(1), Chain);
2570 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2571 Chain = CurDAG->getEntryNode();
2573 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2574 InFlag = Chain.getValue(1);
2577 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2578 LoReg, N0, SDValue()).getValue(1);
2579 if (isSigned && !signBitIsZero) {
2580 // Sign extend the low part into the high part.
2582 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2584 // Zero out the high part, effectively zero extending the input.
2585 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2586 switch (NVT.SimpleTy) {
2589 SDValue(CurDAG->getMachineNode(
2590 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2591 CurDAG->getTargetConstant(X86::sub_16bit, dl,
2599 SDValue(CurDAG->getMachineNode(
2600 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2601 CurDAG->getTargetConstant(0, dl, MVT::i64), ClrNode,
2602 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2607 llvm_unreachable("Unexpected division source");
2610 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2611 ClrNode, InFlag).getValue(1);
2616 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2619 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2620 InFlag = SDValue(CNode, 1);
2621 // Update the chain.
2622 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2625 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2628 // Prevent use of AH in a REX instruction by explicitly copying it to
2629 // an ABCD_L register.
2631 // The current assumption of the register allocator is that isel
2632 // won't generate explicit references to the GR8_ABCD_H registers. If
2633 // the allocator and/or the backend get enhanced to be more robust in
2634 // that regard, this can be, and should be, removed.
2635 if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
2636 SDValue AHCopy = CurDAG->getRegister(X86::AH, MVT::i8);
2637 unsigned AHExtOpcode =
2638 isSigned ? X86::MOVSX32_NOREXrr8 : X86::MOVZX32_NOREXrr8;
2640 SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
2641 MVT::Glue, AHCopy, InFlag);
2642 SDValue Result(RNode, 0);
2643 InFlag = SDValue(RNode, 1);
2645 if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
2646 Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
2647 if (Node->getValueType(1) == MVT::i64) {
2648 // It's not possible to directly movsx AH to a 64bit register, because
2649 // the latter needs the REX prefix, but the former can't have it.
2650 assert(Opcode != X86ISD::SDIVREM8_SEXT_HREG &&
2651 "Unexpected i64 sext of h-register");
2653 SDValue(CurDAG->getMachineNode(
2654 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2655 CurDAG->getTargetConstant(0, dl, MVT::i64), Result,
2656 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2662 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
2664 ReplaceUses(SDValue(Node, 1), Result);
2665 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2667 // Copy the division (low) result, if it is needed.
2668 if (!SDValue(Node, 0).use_empty()) {
2669 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2670 LoReg, NVT, InFlag);
2671 InFlag = Result.getValue(2);
2672 ReplaceUses(SDValue(Node, 0), Result);
2673 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2675 // Copy the remainder (high) result, if it is needed.
2676 if (!SDValue(Node, 1).use_empty()) {
2677 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2678 HiReg, NVT, InFlag);
2679 InFlag = Result.getValue(2);
2680 ReplaceUses(SDValue(Node, 1), Result);
2681 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2688 // Sometimes a SUB is used to perform comparison.
2689 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2690 // This node is not a CMP.
2692 SDValue N0 = Node->getOperand(0);
2693 SDValue N1 = Node->getOperand(1);
2695 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2696 HasNoSignedComparisonUses(Node))
2697 N0 = N0.getOperand(0);
2699 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2700 // use a smaller encoding.
2701 // Look past the truncate if CMP is the only use of it.
2702 if ((N0.getNode()->getOpcode() == ISD::AND ||
2703 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2704 N0.getNode()->hasOneUse() &&
2705 N0.getValueType() != MVT::i8 &&
2706 X86::isZeroNode(N1)) {
2707 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2710 // For example, convert "testl %eax, $8" to "testb %al, $8"
2711 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2712 (!(C->getZExtValue() & 0x80) ||
2713 HasNoSignedComparisonUses(Node))) {
2714 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl, MVT::i8);
2715 SDValue Reg = N0.getNode()->getOperand(0);
2717 // On x86-32, only the ABCD registers have 8-bit subregisters.
2718 if (!Subtarget->is64Bit()) {
2719 const TargetRegisterClass *TRC;
2720 switch (N0.getSimpleValueType().SimpleTy) {
2721 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2722 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2723 default: llvm_unreachable("Unsupported TEST operand type!");
2725 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2726 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2727 Reg.getValueType(), Reg, RC), 0);
2730 // Extract the l-register.
2731 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2735 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2737 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2738 // one, do not call ReplaceAllUsesWith.
2739 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2740 SDValue(NewNode, 0));
2744 // For example, "testl %eax, $2048" to "testb %ah, $8".
2745 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2746 (!(C->getZExtValue() & 0x8000) ||
2747 HasNoSignedComparisonUses(Node))) {
2748 // Shift the immediate right by 8 bits.
2749 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2751 SDValue Reg = N0.getNode()->getOperand(0);
2753 // Put the value in an ABCD register.
2754 const TargetRegisterClass *TRC;
2755 switch (N0.getSimpleValueType().SimpleTy) {
2756 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2757 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2758 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2759 default: llvm_unreachable("Unsupported TEST operand type!");
2761 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2762 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2763 Reg.getValueType(), Reg, RC), 0);
2765 // Extract the h-register.
2766 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2769 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2770 // target GR8_NOREX registers, so make sure the register class is
2772 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2773 MVT::i32, Subreg, ShiftedImm);
2774 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2775 // one, do not call ReplaceAllUsesWith.
2776 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2777 SDValue(NewNode, 0));
2781 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2782 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2783 N0.getValueType() != MVT::i16 &&
2784 (!(C->getZExtValue() & 0x8000) ||
2785 HasNoSignedComparisonUses(Node))) {
2786 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2788 SDValue Reg = N0.getNode()->getOperand(0);
2790 // Extract the 16-bit subregister.
2791 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2795 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2797 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2798 // one, do not call ReplaceAllUsesWith.
2799 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2800 SDValue(NewNode, 0));
2804 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2805 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2806 N0.getValueType() == MVT::i64 &&
2807 (!(C->getZExtValue() & 0x80000000) ||
2808 HasNoSignedComparisonUses(Node))) {
2809 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2811 SDValue Reg = N0.getNode()->getOperand(0);
2813 // Extract the 32-bit subregister.
2814 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2818 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2820 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2821 // one, do not call ReplaceAllUsesWith.
2822 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2823 SDValue(NewNode, 0));
2830 // Change a chain of {load; incr or dec; store} of the same value into
2831 // a simple increment or decrement through memory of that value, if the
2832 // uses of the modified value and its address are suitable.
2833 // The DEC64m tablegen pattern is currently not able to match the case where
2834 // the EFLAGS on the original DEC are used. (This also applies to
2835 // {INC,DEC}X{64,32,16,8}.)
2836 // We'll need to improve tablegen to allow flags to be transferred from a
2837 // node in the pattern to the result node. probably with a new keyword
2838 // for example, we have this
2839 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2840 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2841 // (implicit EFLAGS)]>;
2842 // but maybe need something like this
2843 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2844 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2845 // (transferrable EFLAGS)]>;
2847 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2848 SDValue StoredVal = StoreNode->getOperand(1);
2849 unsigned Opc = StoredVal->getOpcode();
2851 LoadSDNode *LoadNode = nullptr;
2853 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2854 LoadNode, InputChain))
2857 SDValue Base, Scale, Index, Disp, Segment;
2858 if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
2859 Base, Scale, Index, Disp, Segment))
2862 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2863 MemOp[0] = StoreNode->getMemOperand();
2864 MemOp[1] = LoadNode->getMemOperand();
2865 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2866 EVT LdVT = LoadNode->getMemoryVT();
2867 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2868 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2870 MVT::i32, MVT::Other, Ops);
2871 Result->setMemRefs(MemOp, MemOp + 2);
2873 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2874 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2880 SDNode *ResNode = SelectCode(Node);
2882 DEBUG(dbgs() << "=> ";
2883 if (ResNode == nullptr || ResNode == Node)
2886 ResNode->dump(CurDAG);
2892 bool X86DAGToDAGISel::
2893 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
2894 std::vector<SDValue> &OutOps) {
2895 SDValue Op0, Op1, Op2, Op3, Op4;
2896 switch (ConstraintID) {
2898 llvm_unreachable("Unexpected asm memory constraint");
2899 case InlineAsm::Constraint_i:
2900 // FIXME: It seems strange that 'i' is needed here since it's supposed to
2901 // be an immediate and not a memory constraint.
2903 case InlineAsm::Constraint_o: // offsetable ??
2904 case InlineAsm::Constraint_v: // not offsetable ??
2905 case InlineAsm::Constraint_m: // memory
2906 case InlineAsm::Constraint_X:
2907 if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
2912 OutOps.push_back(Op0);
2913 OutOps.push_back(Op1);
2914 OutOps.push_back(Op2);
2915 OutOps.push_back(Op3);
2916 OutOps.push_back(Op4);
2920 /// createX86ISelDag - This pass converts a legalized DAG into a
2921 /// X86-specific DAG, ready for instruction scheduling.
2923 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2924 CodeGenOpt::Level OptLevel) {
2925 return new X86DAGToDAGISel(TM, OptLevel);