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 /// This corresponds to X86AddressMode, but uses SDValue's instead of register
50 /// numbers for the leaves of the matched tree.
51 struct X86ISelAddressMode {
57 // This is really a union, discriminated by BaseType!
65 const GlobalValue *GV;
67 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 MCSym(nullptr), JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {}
79 bool hasSymbolicDisplacement() const {
80 return GV != nullptr || CP != nullptr || ES != nullptr ||
81 MCSym != nullptr || JT != -1 || BlockAddr != nullptr;
84 bool hasBaseOrIndexReg() const {
85 return BaseType == FrameIndexBase ||
86 IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
89 /// Return true if this addressing mode is already RIP-relative.
90 bool isRIPRelative() const {
91 if (BaseType != RegBase) return false;
92 if (RegisterSDNode *RegNode =
93 dyn_cast_or_null<RegisterSDNode>(Base_Reg.getNode()))
94 return RegNode->getReg() == X86::RIP;
98 void setBaseReg(SDValue Reg) {
103 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
105 dbgs() << "X86ISelAddressMode " << this << '\n';
106 dbgs() << "Base_Reg ";
107 if (Base_Reg.getNode())
108 Base_Reg.getNode()->dump();
111 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
112 << " Scale" << Scale << '\n'
114 if (IndexReg.getNode())
115 IndexReg.getNode()->dump();
118 dbgs() << " Disp " << Disp << '\n'
140 dbgs() << " JT" << JT << " Align" << Align << '\n';
147 //===--------------------------------------------------------------------===//
148 /// ISel - X86-specific code to select X86 machine instructions for
149 /// SelectionDAG operations.
151 class X86DAGToDAGISel final : public SelectionDAGISel {
152 /// Keep a pointer to the X86Subtarget around so that we can
153 /// make the right decision when generating code for different targets.
154 const X86Subtarget *Subtarget;
156 /// If true, selector should try to optimize for code size instead of
160 /// If true, selector should try to optimize for minimum code size.
164 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
165 : SelectionDAGISel(tm, OptLevel), OptForSize(false),
166 OptForMinSize(false) {}
168 const char *getPassName() const override {
169 return "X86 DAG->DAG Instruction Selection";
172 bool runOnMachineFunction(MachineFunction &MF) override {
173 // Reset the subtarget each time through.
174 Subtarget = &MF.getSubtarget<X86Subtarget>();
175 SelectionDAGISel::runOnMachineFunction(MF);
179 void EmitFunctionEntryCode() override;
181 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
183 void PreprocessISelDAG() override;
185 inline bool immSext8(SDNode *N) const {
186 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
189 // True if the 64-bit immediate fits in a 32-bit sign-extended field.
190 inline bool i64immSExt32(SDNode *N) const {
191 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
192 return (int64_t)v == (int32_t)v;
195 // Include the pieces autogenerated from the target description.
196 #include "X86GenDAGISel.inc"
199 SDNode *Select(SDNode *N) override;
200 SDNode *selectGather(SDNode *N, unsigned Opc);
201 SDNode *selectAtomicLoadArith(SDNode *Node, MVT NVT);
203 bool foldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
204 bool matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
205 bool matchWrapper(SDValue N, X86ISelAddressMode &AM);
206 bool matchAddress(SDValue N, X86ISelAddressMode &AM);
207 bool matchAdd(SDValue N, X86ISelAddressMode &AM, unsigned Depth);
208 bool matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
210 bool matchAddressBase(SDValue N, X86ISelAddressMode &AM);
211 bool selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
212 SDValue &Scale, SDValue &Index, SDValue &Disp,
214 bool selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
215 SDValue &Scale, SDValue &Index, SDValue &Disp,
217 bool selectMOV64Imm32(SDValue N, SDValue &Imm);
218 bool selectLEAAddr(SDValue N, SDValue &Base,
219 SDValue &Scale, SDValue &Index, SDValue &Disp,
221 bool selectLEA64_32Addr(SDValue N, SDValue &Base,
222 SDValue &Scale, SDValue &Index, SDValue &Disp,
224 bool selectTLSADDRAddr(SDValue N, SDValue &Base,
225 SDValue &Scale, SDValue &Index, SDValue &Disp,
227 bool selectScalarSSELoad(SDNode *Root, SDValue N,
228 SDValue &Base, SDValue &Scale,
229 SDValue &Index, SDValue &Disp,
231 SDValue &NodeWithChain);
233 bool tryFoldLoad(SDNode *P, SDValue N,
234 SDValue &Base, SDValue &Scale,
235 SDValue &Index, SDValue &Disp,
238 /// Implement addressing mode selection for inline asm expressions.
239 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
240 unsigned ConstraintID,
241 std::vector<SDValue> &OutOps) override;
243 void emitSpecialCodeForMain();
245 inline void getAddressOperands(X86ISelAddressMode &AM, SDLoc DL,
246 SDValue &Base, SDValue &Scale,
247 SDValue &Index, SDValue &Disp,
249 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
250 ? CurDAG->getTargetFrameIndex(
252 TLI->getPointerTy(CurDAG->getDataLayout()))
254 Scale = getI8Imm(AM.Scale, DL);
256 // These are 32-bit even in 64-bit mode since RIP-relative offset
259 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
263 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
264 AM.Align, AM.Disp, AM.SymbolFlags);
266 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
267 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
268 } else if (AM.MCSym) {
269 assert(!AM.Disp && "Non-zero displacement is ignored with MCSym.");
270 assert(AM.SymbolFlags == 0 && "oo");
271 Disp = CurDAG->getMCSymbol(AM.MCSym, MVT::i32);
272 } else if (AM.JT != -1) {
273 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
274 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
275 } else if (AM.BlockAddr)
276 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
279 Disp = CurDAG->getTargetConstant(AM.Disp, DL, MVT::i32);
281 if (AM.Segment.getNode())
282 Segment = AM.Segment;
284 Segment = CurDAG->getRegister(0, MVT::i32);
287 // Utility function to determine whether we should avoid selecting
288 // immediate forms of instructions for better code size or not.
289 // At a high level, we'd like to avoid such instructions when
290 // we have similar constants used within the same basic block
291 // that can be kept in a register.
293 bool shouldAvoidImmediateInstFormsForSize(SDNode *N) const {
294 uint32_t UseCount = 0;
296 // Do not want to hoist if we're not optimizing for size.
297 // TODO: We'd like to remove this restriction.
298 // See the comment in X86InstrInfo.td for more info.
302 // Walk all the users of the immediate.
303 for (SDNode::use_iterator UI = N->use_begin(),
304 UE = N->use_end(); (UI != UE) && (UseCount < 2); ++UI) {
308 // This user is already selected. Count it as a legitimate use and
310 if (User->isMachineOpcode()) {
315 // We want to count stores of immediates as real uses.
316 if (User->getOpcode() == ISD::STORE &&
317 User->getOperand(1).getNode() == N) {
322 // We don't currently match users that have > 2 operands (except
323 // for stores, which are handled above)
324 // Those instruction won't match in ISEL, for now, and would
325 // be counted incorrectly.
326 // This may change in the future as we add additional instruction
328 if (User->getNumOperands() != 2)
331 // Immediates that are used for offsets as part of stack
332 // manipulation should be left alone. These are typically
333 // used to indicate SP offsets for argument passing and
334 // will get pulled into stores/pushes (implicitly).
335 if (User->getOpcode() == X86ISD::ADD ||
336 User->getOpcode() == ISD::ADD ||
337 User->getOpcode() == X86ISD::SUB ||
338 User->getOpcode() == ISD::SUB) {
340 // Find the other operand of the add/sub.
341 SDValue OtherOp = User->getOperand(0);
342 if (OtherOp.getNode() == N)
343 OtherOp = User->getOperand(1);
345 // Don't count if the other operand is SP.
346 RegisterSDNode *RegNode;
347 if (OtherOp->getOpcode() == ISD::CopyFromReg &&
348 (RegNode = dyn_cast_or_null<RegisterSDNode>(
349 OtherOp->getOperand(1).getNode())))
350 if ((RegNode->getReg() == X86::ESP) ||
351 (RegNode->getReg() == X86::RSP))
355 // ... otherwise, count this and move on.
359 // If we have more than 1 use, then recommend for hoisting.
360 return (UseCount > 1);
363 /// Return a target constant with the specified value of type i8.
364 inline SDValue getI8Imm(unsigned Imm, SDLoc DL) {
365 return CurDAG->getTargetConstant(Imm, DL, MVT::i8);
368 /// Return a target constant with the specified value, of type i32.
369 inline SDValue getI32Imm(unsigned Imm, SDLoc DL) {
370 return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
373 /// Return an SDNode that returns the value of the global base register.
374 /// Output instructions required to initialize the global base register,
376 SDNode *getGlobalBaseReg();
378 /// Return a reference to the TargetMachine, casted to the target-specific
380 const X86TargetMachine &getTargetMachine() const {
381 return static_cast<const X86TargetMachine &>(TM);
384 /// Return a reference to the TargetInstrInfo, casted to the target-specific
386 const X86InstrInfo *getInstrInfo() const {
387 return Subtarget->getInstrInfo();
390 /// \brief Address-mode matching performs shift-of-and to and-of-shift
391 /// reassociation in order to expose more scaled addressing
393 bool ComplexPatternFuncMutatesDAG() const override {
401 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
402 if (OptLevel == CodeGenOpt::None) return false;
407 if (N.getOpcode() != ISD::LOAD)
410 // If N is a load, do additional profitability checks.
412 switch (U->getOpcode()) {
425 SDValue Op1 = U->getOperand(1);
427 // If the other operand is a 8-bit immediate we should fold the immediate
428 // instead. This reduces code size.
430 // movl 4(%esp), %eax
434 // addl 4(%esp), %eax
435 // The former is 2 bytes shorter. In case where the increment is 1, then
436 // the saving can be 4 bytes (by using incl %eax).
437 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
438 if (Imm->getAPIntValue().isSignedIntN(8))
441 // If the other operand is a TLS address, we should fold it instead.
444 // leal i@NTPOFF(%eax), %eax
446 // movl $i@NTPOFF, %eax
448 // if the block also has an access to a second TLS address this will save
450 // FIXME: This is probably also true for non-TLS addresses.
451 if (Op1.getOpcode() == X86ISD::Wrapper) {
452 SDValue Val = Op1.getOperand(0);
453 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
463 /// Replace the original chain operand of the call with
464 /// load's chain operand and move load below the call's chain operand.
465 static void moveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
466 SDValue Call, SDValue OrigChain) {
467 SmallVector<SDValue, 8> Ops;
468 SDValue Chain = OrigChain.getOperand(0);
469 if (Chain.getNode() == Load.getNode())
470 Ops.push_back(Load.getOperand(0));
472 assert(Chain.getOpcode() == ISD::TokenFactor &&
473 "Unexpected chain operand");
474 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
475 if (Chain.getOperand(i).getNode() == Load.getNode())
476 Ops.push_back(Load.getOperand(0));
478 Ops.push_back(Chain.getOperand(i));
480 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
482 Ops.push_back(NewChain);
484 Ops.append(OrigChain->op_begin() + 1, OrigChain->op_end());
485 CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
486 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
487 Load.getOperand(1), Load.getOperand(2));
490 Ops.push_back(SDValue(Load.getNode(), 1));
491 Ops.append(Call->op_begin() + 1, Call->op_end());
492 CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
495 /// Return true if call address is a load and it can be
496 /// moved below CALLSEQ_START and the chains leading up to the call.
497 /// Return the CALLSEQ_START by reference as a second output.
498 /// In the case of a tail call, there isn't a callseq node between the call
499 /// chain and the load.
500 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
501 // The transformation is somewhat dangerous if the call's chain was glued to
502 // the call. After MoveBelowOrigChain the load is moved between the call and
503 // the chain, this can create a cycle if the load is not folded. So it is
504 // *really* important that we are sure the load will be folded.
505 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
507 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
510 LD->getAddressingMode() != ISD::UNINDEXED ||
511 LD->getExtensionType() != ISD::NON_EXTLOAD)
514 // Now let's find the callseq_start.
515 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
516 if (!Chain.hasOneUse())
518 Chain = Chain.getOperand(0);
521 if (!Chain.getNumOperands())
523 // Since we are not checking for AA here, conservatively abort if the chain
524 // writes to memory. It's not safe to move the callee (a load) across a store.
525 if (isa<MemSDNode>(Chain.getNode()) &&
526 cast<MemSDNode>(Chain.getNode())->writeMem())
528 if (Chain.getOperand(0).getNode() == Callee.getNode())
530 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
531 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
532 Callee.getValue(1).hasOneUse())
537 void X86DAGToDAGISel::PreprocessISelDAG() {
538 // OptFor[Min]Size are used in pattern predicates that isel is matching.
539 OptForSize = MF->getFunction()->optForSize();
540 OptForMinSize = MF->getFunction()->optForMinSize();
541 assert((!OptForMinSize || OptForSize) && "OptForMinSize implies OptForSize");
543 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
544 E = CurDAG->allnodes_end(); I != E; ) {
545 SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
547 if (OptLevel != CodeGenOpt::None &&
548 // Only does this when target favors doesn't favor register indirect
550 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
551 (N->getOpcode() == X86ISD::TC_RETURN &&
552 // Only does this if load can be folded into TC_RETURN.
553 (Subtarget->is64Bit() ||
554 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
555 /// Also try moving call address load from outside callseq_start to just
556 /// before the call to allow it to be folded.
574 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
575 SDValue Chain = N->getOperand(0);
576 SDValue Load = N->getOperand(1);
577 if (!isCalleeLoad(Load, Chain, HasCallSeq))
579 moveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
584 // Lower fpround and fpextend nodes that target the FP stack to be store and
585 // load to the stack. This is a gross hack. We would like to simply mark
586 // these as being illegal, but when we do that, legalize produces these when
587 // it expands calls, then expands these in the same legalize pass. We would
588 // like dag combine to be able to hack on these between the call expansion
589 // and the node legalization. As such this pass basically does "really
590 // late" legalization of these inline with the X86 isel pass.
591 // FIXME: This should only happen when not compiled with -O0.
592 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
595 MVT SrcVT = N->getOperand(0).getSimpleValueType();
596 MVT DstVT = N->getSimpleValueType(0);
598 // If any of the sources are vectors, no fp stack involved.
599 if (SrcVT.isVector() || DstVT.isVector())
602 // If the source and destination are SSE registers, then this is a legal
603 // conversion that should not be lowered.
604 const X86TargetLowering *X86Lowering =
605 static_cast<const X86TargetLowering *>(TLI);
606 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
607 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
608 if (SrcIsSSE && DstIsSSE)
611 if (!SrcIsSSE && !DstIsSSE) {
612 // If this is an FPStack extension, it is a noop.
613 if (N->getOpcode() == ISD::FP_EXTEND)
615 // If this is a value-preserving FPStack truncation, it is a noop.
616 if (N->getConstantOperandVal(1))
620 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
621 // FPStack has extload and truncstore. SSE can fold direct loads into other
622 // operations. Based on this, decide what we want to do.
624 if (N->getOpcode() == ISD::FP_ROUND)
625 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
627 MemVT = SrcIsSSE ? SrcVT : DstVT;
629 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
632 // FIXME: optimize the case where the src/dest is a load or store?
633 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
635 MemTmp, MachinePointerInfo(), MemVT,
637 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
638 MachinePointerInfo(),
639 MemVT, false, false, false, 0);
641 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
642 // extload we created. This will cause general havok on the dag because
643 // anything below the conversion could be folded into other existing nodes.
644 // To avoid invalidating 'I', back it up to the convert node.
646 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
648 // Now that we did that, the node is dead. Increment the iterator to the
649 // next node to process, then delete N.
651 CurDAG->DeleteNode(N);
656 /// Emit any code that needs to be executed only in the main function.
657 void X86DAGToDAGISel::emitSpecialCodeForMain() {
658 if (Subtarget->isTargetCygMing()) {
659 TargetLowering::ArgListTy Args;
660 auto &DL = CurDAG->getDataLayout();
662 TargetLowering::CallLoweringInfo CLI(*CurDAG);
663 CLI.setChain(CurDAG->getRoot())
664 .setCallee(CallingConv::C, Type::getVoidTy(*CurDAG->getContext()),
665 CurDAG->getExternalSymbol("__main", TLI->getPointerTy(DL)),
667 const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
668 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
669 CurDAG->setRoot(Result.second);
673 void X86DAGToDAGISel::EmitFunctionEntryCode() {
674 // If this is main, emit special code for main.
675 if (const Function *Fn = MF->getFunction())
676 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
677 emitSpecialCodeForMain();
680 static bool isDispSafeForFrameIndex(int64_t Val) {
681 // On 64-bit platforms, we can run into an issue where a frame index
682 // includes a displacement that, when added to the explicit displacement,
683 // will overflow the displacement field. Assuming that the frame index
684 // displacement fits into a 31-bit integer (which is only slightly more
685 // aggressive than the current fundamental assumption that it fits into
686 // a 32-bit integer), a 31-bit disp should always be safe.
687 return isInt<31>(Val);
690 bool X86DAGToDAGISel::foldOffsetIntoAddress(uint64_t Offset,
691 X86ISelAddressMode &AM) {
692 // Cannot combine ExternalSymbol displacements with integer offsets.
693 if (Offset != 0 && (AM.ES || AM.MCSym))
695 int64_t Val = AM.Disp + Offset;
696 CodeModel::Model M = TM.getCodeModel();
697 if (Subtarget->is64Bit()) {
698 if (!X86::isOffsetSuitableForCodeModel(Val, M,
699 AM.hasSymbolicDisplacement()))
701 // In addition to the checks required for a register base, check that
702 // we do not try to use an unsafe Disp with a frame index.
703 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
704 !isDispSafeForFrameIndex(Val))
712 bool X86DAGToDAGISel::matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
713 SDValue Address = N->getOperand(1);
715 // load gs:0 -> GS segment register.
716 // load fs:0 -> FS segment register.
718 // This optimization is valid because the GNU TLS model defines that
719 // gs:0 (or fs:0 on X86-64) contains its own address.
720 // For more information see http://people.redhat.com/drepper/tls.pdf
721 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
722 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
723 Subtarget->isTargetLinux())
724 switch (N->getPointerInfo().getAddrSpace()) {
726 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
729 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
736 /// Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes into an addressing
737 /// mode. These wrap things that will resolve down into a symbol reference.
738 /// If no match is possible, this returns true, otherwise it returns false.
739 bool X86DAGToDAGISel::matchWrapper(SDValue N, X86ISelAddressMode &AM) {
740 // If the addressing mode already has a symbol as the displacement, we can
741 // never match another symbol.
742 if (AM.hasSymbolicDisplacement())
745 SDValue N0 = N.getOperand(0);
746 CodeModel::Model M = TM.getCodeModel();
748 // Handle X86-64 rip-relative addresses. We check this before checking direct
749 // folding because RIP is preferable to non-RIP accesses.
750 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
751 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
752 // they cannot be folded into immediate fields.
753 // FIXME: This can be improved for kernel and other models?
754 (M == CodeModel::Small || M == CodeModel::Kernel)) {
755 // Base and index reg must be 0 in order to use %rip as base.
756 if (AM.hasBaseOrIndexReg())
758 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
759 X86ISelAddressMode Backup = AM;
760 AM.GV = G->getGlobal();
761 AM.SymbolFlags = G->getTargetFlags();
762 if (foldOffsetIntoAddress(G->getOffset(), AM)) {
766 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
767 X86ISelAddressMode Backup = AM;
768 AM.CP = CP->getConstVal();
769 AM.Align = CP->getAlignment();
770 AM.SymbolFlags = CP->getTargetFlags();
771 if (foldOffsetIntoAddress(CP->getOffset(), AM)) {
775 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
776 AM.ES = S->getSymbol();
777 AM.SymbolFlags = S->getTargetFlags();
778 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
779 AM.MCSym = S->getMCSymbol();
780 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
781 AM.JT = J->getIndex();
782 AM.SymbolFlags = J->getTargetFlags();
783 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
784 X86ISelAddressMode Backup = AM;
785 AM.BlockAddr = BA->getBlockAddress();
786 AM.SymbolFlags = BA->getTargetFlags();
787 if (foldOffsetIntoAddress(BA->getOffset(), AM)) {
792 llvm_unreachable("Unhandled symbol reference node.");
794 if (N.getOpcode() == X86ISD::WrapperRIP)
795 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
799 // Handle the case when globals fit in our immediate field: This is true for
800 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
801 // mode, this only applies to a non-RIP-relative computation.
802 if (!Subtarget->is64Bit() ||
803 M == CodeModel::Small || M == CodeModel::Kernel) {
804 assert(N.getOpcode() != X86ISD::WrapperRIP &&
805 "RIP-relative addressing already handled");
806 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
807 AM.GV = G->getGlobal();
808 AM.Disp += G->getOffset();
809 AM.SymbolFlags = G->getTargetFlags();
810 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
811 AM.CP = CP->getConstVal();
812 AM.Align = CP->getAlignment();
813 AM.Disp += CP->getOffset();
814 AM.SymbolFlags = CP->getTargetFlags();
815 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
816 AM.ES = S->getSymbol();
817 AM.SymbolFlags = S->getTargetFlags();
818 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
819 AM.MCSym = S->getMCSymbol();
820 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
821 AM.JT = J->getIndex();
822 AM.SymbolFlags = J->getTargetFlags();
823 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
824 AM.BlockAddr = BA->getBlockAddress();
825 AM.Disp += BA->getOffset();
826 AM.SymbolFlags = BA->getTargetFlags();
828 llvm_unreachable("Unhandled symbol reference node.");
835 /// Add the specified node to the specified addressing mode, returning true if
836 /// it cannot be done. This just pattern matches for the addressing mode.
837 bool X86DAGToDAGISel::matchAddress(SDValue N, X86ISelAddressMode &AM) {
838 if (matchAddressRecursively(N, AM, 0))
841 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
842 // a smaller encoding and avoids a scaled-index.
844 AM.BaseType == X86ISelAddressMode::RegBase &&
845 AM.Base_Reg.getNode() == nullptr) {
846 AM.Base_Reg = AM.IndexReg;
850 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
851 // because it has a smaller encoding.
852 // TODO: Which other code models can use this?
853 if (TM.getCodeModel() == CodeModel::Small &&
854 Subtarget->is64Bit() &&
856 AM.BaseType == X86ISelAddressMode::RegBase &&
857 AM.Base_Reg.getNode() == nullptr &&
858 AM.IndexReg.getNode() == nullptr &&
859 AM.SymbolFlags == X86II::MO_NO_FLAG &&
860 AM.hasSymbolicDisplacement())
861 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
866 bool X86DAGToDAGISel::matchAdd(SDValue N, X86ISelAddressMode &AM,
868 // Add an artificial use to this node so that we can keep track of
869 // it if it gets CSE'd with a different node.
870 HandleSDNode Handle(N);
872 X86ISelAddressMode Backup = AM;
873 if (!matchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
874 !matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
878 // Try again after commuting the operands.
879 if (!matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1) &&
880 !matchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
884 // If we couldn't fold both operands into the address at the same time,
885 // see if we can just put each operand into a register and fold at least
887 if (AM.BaseType == X86ISelAddressMode::RegBase &&
888 !AM.Base_Reg.getNode() &&
889 !AM.IndexReg.getNode()) {
890 N = Handle.getValue();
891 AM.Base_Reg = N.getOperand(0);
892 AM.IndexReg = N.getOperand(1);
896 N = Handle.getValue();
900 // Insert a node into the DAG at least before the Pos node's position. This
901 // will reposition the node as needed, and will assign it a node ID that is <=
902 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
903 // IDs! The selection DAG must no longer depend on their uniqueness when this
905 static void insertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
906 if (N.getNode()->getNodeId() == -1 ||
907 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
908 DAG.RepositionNode(Pos.getNode()->getIterator(), N.getNode());
909 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
913 // Transform "(X >> (8-C1)) & (0xff << C1)" to "((X >> 8) & 0xff) << C1" if
914 // safe. This allows us to convert the shift and and into an h-register
915 // extract and a scaled index. Returns false if the simplification is
917 static bool foldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
919 SDValue Shift, SDValue X,
920 X86ISelAddressMode &AM) {
921 if (Shift.getOpcode() != ISD::SRL ||
922 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
926 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
927 if (ScaleLog <= 0 || ScaleLog >= 4 ||
928 Mask != (0xffu << ScaleLog))
931 MVT VT = N.getSimpleValueType();
933 SDValue Eight = DAG.getConstant(8, DL, MVT::i8);
934 SDValue NewMask = DAG.getConstant(0xff, DL, VT);
935 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
936 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
937 SDValue ShlCount = DAG.getConstant(ScaleLog, DL, MVT::i8);
938 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
940 // Insert the new nodes into the topological ordering. We must do this in
941 // a valid topological ordering as nothing is going to go back and re-sort
942 // these nodes. We continually insert before 'N' in sequence as this is
943 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
944 // hierarchy left to express.
945 insertDAGNode(DAG, N, Eight);
946 insertDAGNode(DAG, N, Srl);
947 insertDAGNode(DAG, N, NewMask);
948 insertDAGNode(DAG, N, And);
949 insertDAGNode(DAG, N, ShlCount);
950 insertDAGNode(DAG, N, Shl);
951 DAG.ReplaceAllUsesWith(N, Shl);
953 AM.Scale = (1 << ScaleLog);
957 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
958 // allows us to fold the shift into this addressing mode. Returns false if the
959 // transform succeeded.
960 static bool foldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
962 SDValue Shift, SDValue X,
963 X86ISelAddressMode &AM) {
964 if (Shift.getOpcode() != ISD::SHL ||
965 !isa<ConstantSDNode>(Shift.getOperand(1)))
968 // Not likely to be profitable if either the AND or SHIFT node has more
969 // than one use (unless all uses are for address computation). Besides,
970 // isel mechanism requires their node ids to be reused.
971 if (!N.hasOneUse() || !Shift.hasOneUse())
974 // Verify that the shift amount is something we can fold.
975 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
976 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
979 MVT VT = N.getSimpleValueType();
981 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, DL, VT);
982 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
983 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
985 // Insert the new nodes into the topological ordering. We must do this in
986 // a valid topological ordering as nothing is going to go back and re-sort
987 // these nodes. We continually insert before 'N' in sequence as this is
988 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
989 // hierarchy left to express.
990 insertDAGNode(DAG, N, NewMask);
991 insertDAGNode(DAG, N, NewAnd);
992 insertDAGNode(DAG, N, NewShift);
993 DAG.ReplaceAllUsesWith(N, NewShift);
995 AM.Scale = 1 << ShiftAmt;
996 AM.IndexReg = NewAnd;
1000 // Implement some heroics to detect shifts of masked values where the mask can
1001 // be replaced by extending the shift and undoing that in the addressing mode
1002 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
1003 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
1004 // the addressing mode. This results in code such as:
1006 // int f(short *y, int *lookup_table) {
1008 // return *y + lookup_table[*y >> 11];
1012 // movzwl (%rdi), %eax
1015 // addl (%rsi,%rcx,4), %eax
1018 // movzwl (%rdi), %eax
1022 // addl (%rsi,%rcx), %eax
1024 // Note that this function assumes the mask is provided as a mask *after* the
1025 // value is shifted. The input chain may or may not match that, but computing
1026 // such a mask is trivial.
1027 static bool foldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
1029 SDValue Shift, SDValue X,
1030 X86ISelAddressMode &AM) {
1031 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
1032 !isa<ConstantSDNode>(Shift.getOperand(1)))
1035 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
1036 unsigned MaskLZ = countLeadingZeros(Mask);
1037 unsigned MaskTZ = countTrailingZeros(Mask);
1039 // The amount of shift we're trying to fit into the addressing mode is taken
1040 // from the trailing zeros of the mask.
1041 unsigned AMShiftAmt = MaskTZ;
1043 // There is nothing we can do here unless the mask is removing some bits.
1044 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
1045 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
1047 // We also need to ensure that mask is a continuous run of bits.
1048 if (countTrailingOnes(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
1050 // Scale the leading zero count down based on the actual size of the value.
1051 // Also scale it down based on the size of the shift.
1052 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
1054 // The final check is to ensure that any masked out high bits of X are
1055 // already known to be zero. Otherwise, the mask has a semantic impact
1056 // other than masking out a couple of low bits. Unfortunately, because of
1057 // the mask, zero extensions will be removed from operands in some cases.
1058 // This code works extra hard to look through extensions because we can
1059 // replace them with zero extensions cheaply if necessary.
1060 bool ReplacingAnyExtend = false;
1061 if (X.getOpcode() == ISD::ANY_EXTEND) {
1062 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
1063 X.getOperand(0).getSimpleValueType().getSizeInBits();
1064 // Assume that we'll replace the any-extend with a zero-extend, and
1065 // narrow the search to the extended value.
1066 X = X.getOperand(0);
1067 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
1068 ReplacingAnyExtend = true;
1070 APInt MaskedHighBits =
1071 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
1072 APInt KnownZero, KnownOne;
1073 DAG.computeKnownBits(X, KnownZero, KnownOne);
1074 if (MaskedHighBits != KnownZero) return true;
1076 // We've identified a pattern that can be transformed into a single shift
1077 // and an addressing mode. Make it so.
1078 MVT VT = N.getSimpleValueType();
1079 if (ReplacingAnyExtend) {
1080 assert(X.getValueType() != VT);
1081 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
1082 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
1083 insertDAGNode(DAG, N, NewX);
1087 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, DL, MVT::i8);
1088 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
1089 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, DL, MVT::i8);
1090 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
1092 // Insert the new nodes into the topological ordering. We must do this in
1093 // a valid topological ordering as nothing is going to go back and re-sort
1094 // these nodes. We continually insert before 'N' in sequence as this is
1095 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
1096 // hierarchy left to express.
1097 insertDAGNode(DAG, N, NewSRLAmt);
1098 insertDAGNode(DAG, N, NewSRL);
1099 insertDAGNode(DAG, N, NewSHLAmt);
1100 insertDAGNode(DAG, N, NewSHL);
1101 DAG.ReplaceAllUsesWith(N, NewSHL);
1103 AM.Scale = 1 << AMShiftAmt;
1104 AM.IndexReg = NewSRL;
1108 bool X86DAGToDAGISel::matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
1112 dbgs() << "MatchAddress: ";
1117 return matchAddressBase(N, AM);
1119 // If this is already a %rip relative address, we can only merge immediates
1120 // into it. Instead of handling this in every case, we handle it here.
1121 // RIP relative addressing: %rip + 32-bit displacement!
1122 if (AM.isRIPRelative()) {
1123 // FIXME: JumpTable and ExternalSymbol address currently don't like
1124 // displacements. It isn't very important, but this should be fixed for
1126 if (!(AM.ES || AM.MCSym) && AM.JT != -1)
1129 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
1130 if (!foldOffsetIntoAddress(Cst->getSExtValue(), AM))
1135 switch (N.getOpcode()) {
1137 case ISD::LOCAL_RECOVER: {
1138 if (!AM.hasSymbolicDisplacement() && AM.Disp == 0)
1139 if (const auto *ESNode = dyn_cast<MCSymbolSDNode>(N.getOperand(0))) {
1140 // Use the symbol and don't prefix it.
1141 AM.MCSym = ESNode->getMCSymbol();
1146 case ISD::Constant: {
1147 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1148 if (!foldOffsetIntoAddress(Val, AM))
1153 case X86ISD::Wrapper:
1154 case X86ISD::WrapperRIP:
1155 if (!matchWrapper(N, AM))
1160 if (!matchLoadInAddress(cast<LoadSDNode>(N), AM))
1164 case ISD::FrameIndex:
1165 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1166 AM.Base_Reg.getNode() == nullptr &&
1167 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1168 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1169 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1175 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1179 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1180 unsigned Val = CN->getZExtValue();
1181 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1182 // that the base operand remains free for further matching. If
1183 // the base doesn't end up getting used, a post-processing step
1184 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1185 if (Val == 1 || Val == 2 || Val == 3) {
1186 AM.Scale = 1 << Val;
1187 SDValue ShVal = N.getNode()->getOperand(0);
1189 // Okay, we know that we have a scale by now. However, if the scaled
1190 // value is an add of something and a constant, we can fold the
1191 // constant into the disp field here.
1192 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1193 AM.IndexReg = ShVal.getNode()->getOperand(0);
1194 ConstantSDNode *AddVal =
1195 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1196 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1197 if (!foldOffsetIntoAddress(Disp, AM))
1201 AM.IndexReg = ShVal;
1208 // Scale must not be used already.
1209 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1211 SDValue And = N.getOperand(0);
1212 if (And.getOpcode() != ISD::AND) break;
1213 SDValue X = And.getOperand(0);
1215 // We only handle up to 64-bit values here as those are what matter for
1216 // addressing mode optimizations.
1217 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1219 // The mask used for the transform is expected to be post-shift, but we
1220 // found the shift first so just apply the shift to the mask before passing
1222 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1223 !isa<ConstantSDNode>(And.getOperand(1)))
1225 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1227 // Try to fold the mask and shift into the scale, and return false if we
1229 if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1234 case ISD::SMUL_LOHI:
1235 case ISD::UMUL_LOHI:
1236 // A mul_lohi where we need the low part can be folded as a plain multiply.
1237 if (N.getResNo() != 0) break;
1240 case X86ISD::MUL_IMM:
1241 // X*[3,5,9] -> X+X*[2,4,8]
1242 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1243 AM.Base_Reg.getNode() == nullptr &&
1244 AM.IndexReg.getNode() == nullptr) {
1246 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1247 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1248 CN->getZExtValue() == 9) {
1249 AM.Scale = unsigned(CN->getZExtValue())-1;
1251 SDValue MulVal = N.getNode()->getOperand(0);
1254 // Okay, we know that we have a scale by now. However, if the scaled
1255 // value is an add of something and a constant, we can fold the
1256 // constant into the disp field here.
1257 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1258 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1259 Reg = MulVal.getNode()->getOperand(0);
1260 ConstantSDNode *AddVal =
1261 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1262 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1263 if (foldOffsetIntoAddress(Disp, AM))
1264 Reg = N.getNode()->getOperand(0);
1266 Reg = N.getNode()->getOperand(0);
1269 AM.IndexReg = AM.Base_Reg = Reg;
1276 // Given A-B, if A can be completely folded into the address and
1277 // the index field with the index field unused, use -B as the index.
1278 // This is a win if a has multiple parts that can be folded into
1279 // the address. Also, this saves a mov if the base register has
1280 // other uses, since it avoids a two-address sub instruction, however
1281 // it costs an additional mov if the index register has other uses.
1283 // Add an artificial use to this node so that we can keep track of
1284 // it if it gets CSE'd with a different node.
1285 HandleSDNode Handle(N);
1287 // Test if the LHS of the sub can be folded.
1288 X86ISelAddressMode Backup = AM;
1289 if (matchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1293 // Test if the index field is free for use.
1294 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1300 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1301 // If the RHS involves a register with multiple uses, this
1302 // transformation incurs an extra mov, due to the neg instruction
1303 // clobbering its operand.
1304 if (!RHS.getNode()->hasOneUse() ||
1305 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1306 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1307 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1308 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1309 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1311 // If the base is a register with multiple uses, this
1312 // transformation may save a mov.
1313 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1314 AM.Base_Reg.getNode() &&
1315 !AM.Base_Reg.getNode()->hasOneUse()) ||
1316 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1318 // If the folded LHS was interesting, this transformation saves
1319 // address arithmetic.
1320 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1321 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1322 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1324 // If it doesn't look like it may be an overall win, don't do it.
1330 // Ok, the transformation is legal and appears profitable. Go for it.
1331 SDValue Zero = CurDAG->getConstant(0, dl, N.getValueType());
1332 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1336 // Insert the new nodes into the topological ordering.
1337 insertDAGNode(*CurDAG, N, Zero);
1338 insertDAGNode(*CurDAG, N, Neg);
1343 if (!matchAdd(N, AM, Depth))
1348 // We want to look through a transform in InstCombine and DAGCombiner that
1349 // turns 'add' into 'or', so we can treat this 'or' exactly like an 'add'.
1350 // Example: (or (and x, 1), (shl y, 3)) --> (add (and x, 1), (shl y, 3))
1351 // An 'lea' can then be used to match the shift (multiply) and add:
1353 // lea (%rsi, %rdi, 8), %rax
1354 if (CurDAG->haveNoCommonBitsSet(N.getOperand(0), N.getOperand(1)) &&
1355 !matchAdd(N, AM, Depth))
1360 // Perform some heroic transforms on an and of a constant-count shift
1361 // with a constant to enable use of the scaled offset field.
1363 // Scale must not be used already.
1364 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1366 SDValue Shift = N.getOperand(0);
1367 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1368 SDValue X = Shift.getOperand(0);
1370 // We only handle up to 64-bit values here as those are what matter for
1371 // addressing mode optimizations.
1372 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1374 if (!isa<ConstantSDNode>(N.getOperand(1)))
1376 uint64_t Mask = N.getConstantOperandVal(1);
1378 // Try to fold the mask and shift into an extract and scale.
1379 if (!foldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1382 // Try to fold the mask and shift directly into the scale.
1383 if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1386 // Try to swap the mask and shift to place shifts which can be done as
1387 // a scale on the outside of the mask.
1388 if (!foldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1394 return matchAddressBase(N, AM);
1397 /// Helper for MatchAddress. Add the specified node to the
1398 /// specified addressing mode without any further recursion.
1399 bool X86DAGToDAGISel::matchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1400 // Is the base register already occupied?
1401 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1402 // If so, check to see if the scale index register is set.
1403 if (!AM.IndexReg.getNode()) {
1409 // Otherwise, we cannot select it.
1413 // Default, generate it as a register.
1414 AM.BaseType = X86ISelAddressMode::RegBase;
1419 bool X86DAGToDAGISel::selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
1420 SDValue &Scale, SDValue &Index,
1421 SDValue &Disp, SDValue &Segment) {
1423 MaskedGatherScatterSDNode *Mgs = dyn_cast<MaskedGatherScatterSDNode>(Parent);
1426 X86ISelAddressMode AM;
1427 unsigned AddrSpace = Mgs->getPointerInfo().getAddrSpace();
1428 // AddrSpace 256 -> GS, 257 -> FS.
1429 if (AddrSpace == 256)
1430 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1431 if (AddrSpace == 257)
1432 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1435 Base = Mgs->getBasePtr();
1436 Index = Mgs->getIndex();
1437 unsigned ScalarSize = Mgs->getValue().getValueType().getScalarSizeInBits();
1438 Scale = getI8Imm(ScalarSize/8, DL);
1440 // If Base is 0, the whole address is in index and the Scale is 1
1441 if (isa<ConstantSDNode>(Base)) {
1442 assert(cast<ConstantSDNode>(Base)->isNullValue() &&
1443 "Unexpected base in gather/scatter");
1444 Scale = getI8Imm(1, DL);
1445 Base = CurDAG->getRegister(0, MVT::i32);
1447 if (AM.Segment.getNode())
1448 Segment = AM.Segment;
1450 Segment = CurDAG->getRegister(0, MVT::i32);
1451 Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
1455 /// Returns true if it is able to pattern match an addressing mode.
1456 /// It returns the operands which make up the maximal addressing mode it can
1457 /// match by reference.
1459 /// Parent is the parent node of the addr operand that is being matched. It
1460 /// is always a load, store, atomic node, or null. It is only null when
1461 /// checking memory operands for inline asm nodes.
1462 bool X86DAGToDAGISel::selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1463 SDValue &Scale, SDValue &Index,
1464 SDValue &Disp, SDValue &Segment) {
1465 X86ISelAddressMode AM;
1468 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1469 // that are not a MemSDNode, and thus don't have proper addrspace info.
1470 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1471 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1472 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1473 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1474 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1475 unsigned AddrSpace =
1476 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1477 // AddrSpace 256 -> GS, 257 -> FS.
1478 if (AddrSpace == 256)
1479 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1480 if (AddrSpace == 257)
1481 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1484 if (matchAddress(N, AM))
1487 MVT VT = N.getSimpleValueType();
1488 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1489 if (!AM.Base_Reg.getNode())
1490 AM.Base_Reg = CurDAG->getRegister(0, VT);
1493 if (!AM.IndexReg.getNode())
1494 AM.IndexReg = CurDAG->getRegister(0, VT);
1496 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1500 /// Match a scalar SSE load. In particular, we want to match a load whose top
1501 /// elements are either undef or zeros. The load flavor is derived from the
1502 /// type of N, which is either v4f32 or v2f64.
1505 /// PatternChainNode: this is the matched node that has a chain input and
1507 bool X86DAGToDAGISel::selectScalarSSELoad(SDNode *Root,
1508 SDValue N, SDValue &Base,
1509 SDValue &Scale, SDValue &Index,
1510 SDValue &Disp, SDValue &Segment,
1511 SDValue &PatternNodeWithChain) {
1512 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1513 PatternNodeWithChain = N.getOperand(0);
1514 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1515 PatternNodeWithChain.hasOneUse() &&
1516 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1517 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1518 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1519 if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1525 // Also handle the case where we explicitly require zeros in the top
1526 // elements. This is a vector shuffle from the zero vector.
1527 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1528 // Check to see if the top elements are all zeros (or bitcast of zeros).
1529 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1530 N.getOperand(0).getNode()->hasOneUse() &&
1531 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1532 N.getOperand(0).getOperand(0).hasOneUse() &&
1533 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1534 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1535 // Okay, this is a zero extending load. Fold it.
1536 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1537 if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1539 PatternNodeWithChain = SDValue(LD, 0);
1546 bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
1547 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1548 uint64_t ImmVal = CN->getZExtValue();
1549 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1552 Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i64);
1556 // In static codegen with small code model, we can get the address of a label
1557 // into a register with 'movl'. TableGen has already made sure we're looking
1558 // at a label of some kind.
1559 assert(N->getOpcode() == X86ISD::Wrapper &&
1560 "Unexpected node type for MOV32ri64");
1561 N = N.getOperand(0);
1563 if (N->getOpcode() != ISD::TargetConstantPool &&
1564 N->getOpcode() != ISD::TargetJumpTable &&
1565 N->getOpcode() != ISD::TargetGlobalAddress &&
1566 N->getOpcode() != ISD::TargetExternalSymbol &&
1567 N->getOpcode() != ISD::MCSymbol &&
1568 N->getOpcode() != ISD::TargetBlockAddress)
1572 return TM.getCodeModel() == CodeModel::Small;
1575 bool X86DAGToDAGISel::selectLEA64_32Addr(SDValue N, SDValue &Base,
1576 SDValue &Scale, SDValue &Index,
1577 SDValue &Disp, SDValue &Segment) {
1578 if (!selectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1582 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1583 if (RN && RN->getReg() == 0)
1584 Base = CurDAG->getRegister(0, MVT::i64);
1585 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
1586 // Base could already be %rip, particularly in the x32 ABI.
1587 Base = SDValue(CurDAG->getMachineNode(
1588 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1589 CurDAG->getTargetConstant(0, DL, MVT::i64),
1591 CurDAG->getTargetConstant(X86::sub_32bit, DL, MVT::i32)),
1595 RN = dyn_cast<RegisterSDNode>(Index);
1596 if (RN && RN->getReg() == 0)
1597 Index = CurDAG->getRegister(0, MVT::i64);
1599 assert(Index.getValueType() == MVT::i32 &&
1600 "Expect to be extending 32-bit registers for use in LEA");
1601 Index = SDValue(CurDAG->getMachineNode(
1602 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1603 CurDAG->getTargetConstant(0, DL, MVT::i64),
1605 CurDAG->getTargetConstant(X86::sub_32bit, DL,
1613 /// Calls SelectAddr and determines if the maximal addressing
1614 /// mode it matches can be cost effectively emitted as an LEA instruction.
1615 bool X86DAGToDAGISel::selectLEAAddr(SDValue N,
1616 SDValue &Base, SDValue &Scale,
1617 SDValue &Index, SDValue &Disp,
1619 X86ISelAddressMode AM;
1621 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1623 SDValue Copy = AM.Segment;
1624 SDValue T = CurDAG->getRegister(0, MVT::i32);
1626 if (matchAddress(N, AM))
1628 assert (T == AM.Segment);
1631 MVT VT = N.getSimpleValueType();
1632 unsigned Complexity = 0;
1633 if (AM.BaseType == X86ISelAddressMode::RegBase)
1634 if (AM.Base_Reg.getNode())
1637 AM.Base_Reg = CurDAG->getRegister(0, VT);
1638 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1641 if (AM.IndexReg.getNode())
1644 AM.IndexReg = CurDAG->getRegister(0, VT);
1646 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1651 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1652 // to a LEA. This is determined with some experimentation but is by no means
1653 // optimal (especially for code size consideration). LEA is nice because of
1654 // its three-address nature. Tweak the cost function again when we can run
1655 // convertToThreeAddress() at register allocation time.
1656 if (AM.hasSymbolicDisplacement()) {
1657 // For X86-64, always use LEA to materialize RIP-relative addresses.
1658 if (Subtarget->is64Bit())
1664 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1667 // If it isn't worth using an LEA, reject it.
1668 if (Complexity <= 2)
1671 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1675 /// This is only run on TargetGlobalTLSAddress nodes.
1676 bool X86DAGToDAGISel::selectTLSADDRAddr(SDValue N, SDValue &Base,
1677 SDValue &Scale, SDValue &Index,
1678 SDValue &Disp, SDValue &Segment) {
1679 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1680 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1682 X86ISelAddressMode AM;
1683 AM.GV = GA->getGlobal();
1684 AM.Disp += GA->getOffset();
1685 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1686 AM.SymbolFlags = GA->getTargetFlags();
1688 if (N.getValueType() == MVT::i32) {
1690 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1692 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1695 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1700 bool X86DAGToDAGISel::tryFoldLoad(SDNode *P, SDValue N,
1701 SDValue &Base, SDValue &Scale,
1702 SDValue &Index, SDValue &Disp,
1704 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1705 !IsProfitableToFold(N, P, P) ||
1706 !IsLegalToFold(N, P, P, OptLevel))
1709 return selectAddr(N.getNode(),
1710 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1713 /// Return an SDNode that returns the value of the global base register.
1714 /// Output instructions required to initialize the global base register,
1716 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1717 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1718 auto &DL = MF->getDataLayout();
1719 return CurDAG->getRegister(GlobalBaseReg, TLI->getPointerTy(DL)).getNode();
1722 /// Atomic opcode table
1750 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1761 X86::LOCK_ADD64mi32,
1774 X86::LOCK_SUB64mi32,
1826 X86::LOCK_AND64mi32,
1839 X86::LOCK_XOR64mi32,
1844 // Return the target constant operand for atomic-load-op and do simple
1845 // translations, such as from atomic-load-add to lock-sub. The return value is
1846 // one of the following 3 cases:
1847 // + target-constant, the operand could be supported as a target constant.
1848 // + empty, the operand is not needed any more with the new op selected.
1849 // + non-empty, otherwise.
1850 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1852 enum AtomicOpc &Op, MVT NVT,
1854 const X86Subtarget *Subtarget) {
1855 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1856 int64_t CNVal = CN->getSExtValue();
1857 // Quit if not 32-bit imm.
1858 if ((int32_t)CNVal != CNVal)
1860 // Quit if INT32_MIN: it would be negated as it is negative and overflow,
1861 // producing an immediate that does not fit in the 32 bits available for
1862 // an immediate operand to sub. However, it still fits in 32 bits for the
1863 // add (since it is not negated) so we can return target-constant.
1864 if (CNVal == INT32_MIN)
1865 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1866 // For atomic-load-add, we could do some optimizations.
1868 // Translate to INC/DEC if ADD by 1 or -1.
1869 if (((CNVal == 1) || (CNVal == -1)) && !Subtarget->slowIncDec()) {
1870 Op = (CNVal == 1) ? INC : DEC;
1871 // No more constant operand after being translated into INC/DEC.
1874 // Translate to SUB if ADD by negative value.
1880 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1883 // If the value operand is single-used, try to optimize it.
1884 if (Op == ADD && Val.hasOneUse()) {
1885 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1886 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1888 return Val.getOperand(1);
1890 // A special case for i16, which needs truncating as, in most cases, it's
1891 // promoted to i32. We will translate
1892 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1893 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1894 Val.getOperand(0).getOpcode() == ISD::SUB &&
1895 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1897 Val = Val.getOperand(0);
1898 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1906 SDNode *X86DAGToDAGISel::selectAtomicLoadArith(SDNode *Node, MVT NVT) {
1907 if (Node->hasAnyUseOfValue(0))
1912 // Optimize common patterns for __sync_or_and_fetch and similar arith
1913 // operations where the result is not used. This allows us to use the "lock"
1914 // version of the arithmetic instruction.
1915 SDValue Chain = Node->getOperand(0);
1916 SDValue Ptr = Node->getOperand(1);
1917 SDValue Val = Node->getOperand(2);
1918 SDValue Base, Scale, Index, Disp, Segment;
1919 if (!selectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
1922 // Which index into the table.
1924 switch (Node->getOpcode()) {
1927 case ISD::ATOMIC_LOAD_OR:
1930 case ISD::ATOMIC_LOAD_AND:
1933 case ISD::ATOMIC_LOAD_XOR:
1936 case ISD::ATOMIC_LOAD_ADD:
1941 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val, Subtarget);
1942 bool isUnOp = !Val.getNode();
1943 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1946 switch (NVT.SimpleTy) {
1947 default: return nullptr;
1950 Opc = AtomicOpcTbl[Op][ConstantI8];
1952 Opc = AtomicOpcTbl[Op][I8];
1956 if (immSext8(Val.getNode()))
1957 Opc = AtomicOpcTbl[Op][SextConstantI16];
1959 Opc = AtomicOpcTbl[Op][ConstantI16];
1961 Opc = AtomicOpcTbl[Op][I16];
1965 if (immSext8(Val.getNode()))
1966 Opc = AtomicOpcTbl[Op][SextConstantI32];
1968 Opc = AtomicOpcTbl[Op][ConstantI32];
1970 Opc = AtomicOpcTbl[Op][I32];
1974 if (immSext8(Val.getNode()))
1975 Opc = AtomicOpcTbl[Op][SextConstantI64];
1976 else if (i64immSExt32(Val.getNode()))
1977 Opc = AtomicOpcTbl[Op][ConstantI64];
1979 llvm_unreachable("True 64 bits constant in SelectAtomicLoadArith");
1981 Opc = AtomicOpcTbl[Op][I64];
1985 assert(Opc != 0 && "Invalid arith lock transform!");
1987 // Building the new node.
1990 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Chain };
1991 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1993 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Val, Chain };
1994 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1997 // Copying the MachineMemOperand.
1998 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1999 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
2000 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
2002 // We need to have two outputs as that is what the original instruction had.
2003 // So we add a dummy, undefined output. This is safe as we checked first
2004 // that no-one uses our output anyway.
2005 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
2007 SDValue RetVals[] = { Undef, Ret };
2008 return CurDAG->getMergeValues(RetVals, dl).getNode();
2011 /// Test whether the given X86ISD::CMP node has any uses which require the SF
2012 /// or OF bits to be accurate.
2013 static bool hasNoSignedComparisonUses(SDNode *N) {
2014 // Examine each user of the node.
2015 for (SDNode::use_iterator UI = N->use_begin(),
2016 UE = N->use_end(); UI != UE; ++UI) {
2017 // Only examine CopyToReg uses.
2018 if (UI->getOpcode() != ISD::CopyToReg)
2020 // Only examine CopyToReg uses that copy to EFLAGS.
2021 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
2024 // Examine each user of the CopyToReg use.
2025 for (SDNode::use_iterator FlagUI = UI->use_begin(),
2026 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
2027 // Only examine the Flag result.
2028 if (FlagUI.getUse().getResNo() != 1) continue;
2029 // Anything unusual: assume conservatively.
2030 if (!FlagUI->isMachineOpcode()) return false;
2031 // Examine the opcode of the user.
2032 switch (FlagUI->getMachineOpcode()) {
2033 // These comparisons don't treat the most significant bit specially.
2034 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
2035 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
2036 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
2037 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
2038 case X86::JA_1: case X86::JAE_1: case X86::JB_1: case X86::JBE_1:
2039 case X86::JE_1: case X86::JNE_1: case X86::JP_1: case X86::JNP_1:
2040 case X86::CMOVA16rr: case X86::CMOVA16rm:
2041 case X86::CMOVA32rr: case X86::CMOVA32rm:
2042 case X86::CMOVA64rr: case X86::CMOVA64rm:
2043 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
2044 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
2045 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
2046 case X86::CMOVB16rr: case X86::CMOVB16rm:
2047 case X86::CMOVB32rr: case X86::CMOVB32rm:
2048 case X86::CMOVB64rr: case X86::CMOVB64rm:
2049 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
2050 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
2051 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
2052 case X86::CMOVE16rr: case X86::CMOVE16rm:
2053 case X86::CMOVE32rr: case X86::CMOVE32rm:
2054 case X86::CMOVE64rr: case X86::CMOVE64rm:
2055 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
2056 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
2057 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
2058 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
2059 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
2060 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
2061 case X86::CMOVP16rr: case X86::CMOVP16rm:
2062 case X86::CMOVP32rr: case X86::CMOVP32rm:
2063 case X86::CMOVP64rr: case X86::CMOVP64rm:
2065 // Anything else: assume conservatively.
2066 default: return false;
2073 /// Check whether or not the chain ending in StoreNode is suitable for doing
2074 /// the {load; increment or decrement; store} to modify transformation.
2075 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
2076 SDValue StoredVal, SelectionDAG *CurDAG,
2077 LoadSDNode* &LoadNode, SDValue &InputChain) {
2079 // is the value stored the result of a DEC or INC?
2080 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
2082 // is the stored value result 0 of the load?
2083 if (StoredVal.getResNo() != 0) return false;
2085 // are there other uses of the loaded value than the inc or dec?
2086 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
2088 // is the store non-extending and non-indexed?
2089 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
2092 SDValue Load = StoredVal->getOperand(0);
2093 // Is the stored value a non-extending and non-indexed load?
2094 if (!ISD::isNormalLoad(Load.getNode())) return false;
2096 // Return LoadNode by reference.
2097 LoadNode = cast<LoadSDNode>(Load);
2098 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
2099 EVT LdVT = LoadNode->getMemoryVT();
2100 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
2104 // Is store the only read of the loaded value?
2105 if (!Load.hasOneUse())
2108 // Is the address of the store the same as the load?
2109 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
2110 LoadNode->getOffset() != StoreNode->getOffset())
2113 // Check if the chain is produced by the load or is a TokenFactor with
2114 // the load output chain as an operand. Return InputChain by reference.
2115 SDValue Chain = StoreNode->getChain();
2117 bool ChainCheck = false;
2118 if (Chain == Load.getValue(1)) {
2120 InputChain = LoadNode->getChain();
2121 } else if (Chain.getOpcode() == ISD::TokenFactor) {
2122 SmallVector<SDValue, 4> ChainOps;
2123 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
2124 SDValue Op = Chain.getOperand(i);
2125 if (Op == Load.getValue(1)) {
2130 // Make sure using Op as part of the chain would not cause a cycle here.
2131 // In theory, we could check whether the chain node is a predecessor of
2132 // the load. But that can be very expensive. Instead visit the uses and
2133 // make sure they all have smaller node id than the load.
2134 int LoadId = LoadNode->getNodeId();
2135 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
2136 UE = UI->use_end(); UI != UE; ++UI) {
2137 if (UI.getUse().getResNo() != 0)
2139 if (UI->getNodeId() > LoadId)
2143 ChainOps.push_back(Op);
2147 // Make a new TokenFactor with all the other input chains except
2149 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
2150 MVT::Other, ChainOps);
2158 /// Get the appropriate X86 opcode for an in-memory increment or decrement.
2159 /// Opc should be X86ISD::DEC or X86ISD::INC.
2160 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2161 if (Opc == X86ISD::DEC) {
2162 if (LdVT == MVT::i64) return X86::DEC64m;
2163 if (LdVT == MVT::i32) return X86::DEC32m;
2164 if (LdVT == MVT::i16) return X86::DEC16m;
2165 if (LdVT == MVT::i8) return X86::DEC8m;
2167 assert(Opc == X86ISD::INC && "unrecognized opcode");
2168 if (LdVT == MVT::i64) return X86::INC64m;
2169 if (LdVT == MVT::i32) return X86::INC32m;
2170 if (LdVT == MVT::i16) return X86::INC16m;
2171 if (LdVT == MVT::i8) return X86::INC8m;
2173 llvm_unreachable("unrecognized size for LdVT");
2176 /// Customized ISel for GATHER operations.
2177 SDNode *X86DAGToDAGISel::selectGather(SDNode *Node, unsigned Opc) {
2178 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2179 SDValue Chain = Node->getOperand(0);
2180 SDValue VSrc = Node->getOperand(2);
2181 SDValue Base = Node->getOperand(3);
2182 SDValue VIdx = Node->getOperand(4);
2183 SDValue VMask = Node->getOperand(5);
2184 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2188 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2193 // Memory Operands: Base, Scale, Index, Disp, Segment
2194 SDValue Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
2195 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2196 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue(), DL), VIdx,
2197 Disp, Segment, VMask, Chain};
2198 SDNode *ResNode = CurDAG->getMachineNode(Opc, DL, VTs, Ops);
2199 // Node has 2 outputs: VDst and MVT::Other.
2200 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2201 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2203 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2204 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2208 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2209 MVT NVT = Node->getSimpleValueType(0);
2211 unsigned Opcode = Node->getOpcode();
2214 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2216 if (Node->isMachineOpcode()) {
2217 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2218 Node->setNodeId(-1);
2219 return nullptr; // Already selected.
2225 if (Subtarget->isTargetNaCl())
2226 // NaCl has its own pass where jmp %r32 are converted to jmp %r64. We
2227 // leave the instruction alone.
2229 if (Subtarget->isTarget64BitILP32()) {
2230 // Converts a 32-bit register to a 64-bit, zero-extended version of
2231 // it. This is needed because x86-64 can do many things, but jmp %r32
2232 // ain't one of them.
2233 const SDValue &Target = Node->getOperand(1);
2234 assert(Target.getSimpleValueType() == llvm::MVT::i32);
2235 SDValue ZextTarget = CurDAG->getZExtOrTrunc(Target, dl, EVT(MVT::i64));
2236 SDValue Brind = CurDAG->getNode(ISD::BRIND, dl, MVT::Other,
2237 Node->getOperand(0), ZextTarget);
2238 ReplaceUses(SDValue(Node, 0), Brind);
2239 SelectCode(ZextTarget.getNode());
2240 SelectCode(Brind.getNode());
2245 case ISD::INTRINSIC_W_CHAIN: {
2246 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2249 case Intrinsic::x86_avx2_gather_d_pd:
2250 case Intrinsic::x86_avx2_gather_d_pd_256:
2251 case Intrinsic::x86_avx2_gather_q_pd:
2252 case Intrinsic::x86_avx2_gather_q_pd_256:
2253 case Intrinsic::x86_avx2_gather_d_ps:
2254 case Intrinsic::x86_avx2_gather_d_ps_256:
2255 case Intrinsic::x86_avx2_gather_q_ps:
2256 case Intrinsic::x86_avx2_gather_q_ps_256:
2257 case Intrinsic::x86_avx2_gather_d_q:
2258 case Intrinsic::x86_avx2_gather_d_q_256:
2259 case Intrinsic::x86_avx2_gather_q_q:
2260 case Intrinsic::x86_avx2_gather_q_q_256:
2261 case Intrinsic::x86_avx2_gather_d_d:
2262 case Intrinsic::x86_avx2_gather_d_d_256:
2263 case Intrinsic::x86_avx2_gather_q_d:
2264 case Intrinsic::x86_avx2_gather_q_d_256: {
2265 if (!Subtarget->hasAVX2())
2269 default: llvm_unreachable("Impossible intrinsic");
2270 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2271 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2272 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2273 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2274 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2275 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2276 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2277 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2278 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2279 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2280 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2281 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2282 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2283 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2284 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2285 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2287 SDNode *RetVal = selectGather(Node, Opc);
2289 // We already called ReplaceUses inside SelectGather.
2296 case X86ISD::GlobalBaseReg:
2297 return getGlobalBaseReg();
2299 case X86ISD::SHRUNKBLEND: {
2300 // SHRUNKBLEND selects like a regular VSELECT.
2301 SDValue VSelect = CurDAG->getNode(
2302 ISD::VSELECT, SDLoc(Node), Node->getValueType(0), Node->getOperand(0),
2303 Node->getOperand(1), Node->getOperand(2));
2304 ReplaceUses(SDValue(Node, 0), VSelect);
2305 SelectCode(VSelect.getNode());
2306 // We already called ReplaceUses.
2310 case ISD::ATOMIC_LOAD_XOR:
2311 case ISD::ATOMIC_LOAD_AND:
2312 case ISD::ATOMIC_LOAD_OR:
2313 case ISD::ATOMIC_LOAD_ADD: {
2314 SDNode *RetVal = selectAtomicLoadArith(Node, NVT);
2322 // For operations of the form (x << C1) op C2, check if we can use a smaller
2323 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2324 SDValue N0 = Node->getOperand(0);
2325 SDValue N1 = Node->getOperand(1);
2327 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2330 // i8 is unshrinkable, i16 should be promoted to i32.
2331 if (NVT != MVT::i32 && NVT != MVT::i64)
2334 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2335 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2336 if (!Cst || !ShlCst)
2339 int64_t Val = Cst->getSExtValue();
2340 uint64_t ShlVal = ShlCst->getZExtValue();
2342 // Make sure that we don't change the operation by removing bits.
2343 // This only matters for OR and XOR, AND is unaffected.
2344 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2345 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2348 unsigned ShlOp, AddOp, Op;
2351 // Check the minimum bitwidth for the new constant.
2352 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2353 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2354 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2355 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2357 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2360 // Bail if there is no smaller encoding.
2364 switch (NVT.SimpleTy) {
2365 default: llvm_unreachable("Unsupported VT!");
2367 assert(CstVT == MVT::i8);
2368 ShlOp = X86::SHL32ri;
2369 AddOp = X86::ADD32rr;
2372 default: llvm_unreachable("Impossible opcode");
2373 case ISD::AND: Op = X86::AND32ri8; break;
2374 case ISD::OR: Op = X86::OR32ri8; break;
2375 case ISD::XOR: Op = X86::XOR32ri8; break;
2379 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2380 ShlOp = X86::SHL64ri;
2381 AddOp = X86::ADD64rr;
2384 default: llvm_unreachable("Impossible opcode");
2385 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2386 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2387 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2392 // Emit the smaller op and the shift.
2393 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, dl, CstVT);
2394 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2396 return CurDAG->SelectNodeTo(Node, AddOp, NVT, SDValue(New, 0),
2398 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2399 getI8Imm(ShlVal, dl));
2402 case X86ISD::SMUL8: {
2403 SDValue N0 = Node->getOperand(0);
2404 SDValue N1 = Node->getOperand(1);
2406 Opc = (Opcode == X86ISD::SMUL8 ? X86::IMUL8r : X86::MUL8r);
2408 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::AL,
2409 N0, SDValue()).getValue(1);
2411 SDVTList VTs = CurDAG->getVTList(NVT, MVT::i32);
2412 SDValue Ops[] = {N1, InFlag};
2413 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2415 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2416 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2420 case X86ISD::UMUL: {
2421 SDValue N0 = Node->getOperand(0);
2422 SDValue N1 = Node->getOperand(1);
2425 switch (NVT.SimpleTy) {
2426 default: llvm_unreachable("Unsupported VT!");
2427 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2428 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2429 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2430 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2433 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2434 N0, SDValue()).getValue(1);
2436 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2437 SDValue Ops[] = {N1, InFlag};
2438 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2440 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2441 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2442 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2446 case ISD::SMUL_LOHI:
2447 case ISD::UMUL_LOHI: {
2448 SDValue N0 = Node->getOperand(0);
2449 SDValue N1 = Node->getOperand(1);
2451 bool isSigned = Opcode == ISD::SMUL_LOHI;
2452 bool hasBMI2 = Subtarget->hasBMI2();
2454 switch (NVT.SimpleTy) {
2455 default: llvm_unreachable("Unsupported VT!");
2456 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2457 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2458 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2459 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2460 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2461 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2464 switch (NVT.SimpleTy) {
2465 default: llvm_unreachable("Unsupported VT!");
2466 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2467 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2468 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2469 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2473 unsigned SrcReg, LoReg, HiReg;
2475 default: llvm_unreachable("Unknown MUL opcode!");
2478 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2482 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2486 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2490 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2493 SrcReg = X86::EDX; LoReg = HiReg = 0;
2496 SrcReg = X86::RDX; LoReg = HiReg = 0;
2500 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2501 bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2502 // Multiply is commmutative.
2504 foldedLoad = tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2509 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2510 N0, SDValue()).getValue(1);
2511 SDValue ResHi, ResLo;
2515 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2517 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2518 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2519 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2520 ResHi = SDValue(CNode, 0);
2521 ResLo = SDValue(CNode, 1);
2522 Chain = SDValue(CNode, 2);
2523 InFlag = SDValue(CNode, 3);
2525 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2526 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2527 Chain = SDValue(CNode, 0);
2528 InFlag = SDValue(CNode, 1);
2531 // Update the chain.
2532 ReplaceUses(N1.getValue(1), Chain);
2534 SDValue Ops[] = { N1, InFlag };
2535 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2536 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2537 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2538 ResHi = SDValue(CNode, 0);
2539 ResLo = SDValue(CNode, 1);
2540 InFlag = SDValue(CNode, 2);
2542 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2543 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2544 InFlag = SDValue(CNode, 0);
2548 // Prevent use of AH in a REX instruction by referencing AX instead.
2549 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2550 !SDValue(Node, 1).use_empty()) {
2551 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2552 X86::AX, MVT::i16, InFlag);
2553 InFlag = Result.getValue(2);
2554 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2556 if (!SDValue(Node, 0).use_empty())
2557 ReplaceUses(SDValue(Node, 1),
2558 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2560 // Shift AX down 8 bits.
2561 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2563 CurDAG->getTargetConstant(8, dl, MVT::i8)),
2565 // Then truncate it down to i8.
2566 ReplaceUses(SDValue(Node, 1),
2567 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2569 // Copy the low half of the result, if it is needed.
2570 if (!SDValue(Node, 0).use_empty()) {
2571 if (!ResLo.getNode()) {
2572 assert(LoReg && "Register for low half is not defined!");
2573 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2575 InFlag = ResLo.getValue(2);
2577 ReplaceUses(SDValue(Node, 0), ResLo);
2578 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2580 // Copy the high half of the result, if it is needed.
2581 if (!SDValue(Node, 1).use_empty()) {
2582 if (!ResHi.getNode()) {
2583 assert(HiReg && "Register for high half is not defined!");
2584 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2586 InFlag = ResHi.getValue(2);
2588 ReplaceUses(SDValue(Node, 1), ResHi);
2589 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2597 case X86ISD::SDIVREM8_SEXT_HREG:
2598 case X86ISD::UDIVREM8_ZEXT_HREG: {
2599 SDValue N0 = Node->getOperand(0);
2600 SDValue N1 = Node->getOperand(1);
2602 bool isSigned = (Opcode == ISD::SDIVREM ||
2603 Opcode == X86ISD::SDIVREM8_SEXT_HREG);
2605 switch (NVT.SimpleTy) {
2606 default: llvm_unreachable("Unsupported VT!");
2607 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2608 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2609 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2610 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2613 switch (NVT.SimpleTy) {
2614 default: llvm_unreachable("Unsupported VT!");
2615 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2616 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2617 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2618 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2622 unsigned LoReg, HiReg, ClrReg;
2623 unsigned SExtOpcode;
2624 switch (NVT.SimpleTy) {
2625 default: llvm_unreachable("Unsupported VT!");
2627 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2628 SExtOpcode = X86::CBW;
2631 LoReg = X86::AX; HiReg = X86::DX;
2633 SExtOpcode = X86::CWD;
2636 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2637 SExtOpcode = X86::CDQ;
2640 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2641 SExtOpcode = X86::CQO;
2645 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2646 bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2647 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2650 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2651 // Special case for div8, just use a move with zero extension to AX to
2652 // clear the upper 8 bits (AH).
2653 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2654 if (tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2655 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2657 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2658 MVT::Other, Ops), 0);
2659 Chain = Move.getValue(1);
2660 ReplaceUses(N0.getValue(1), Chain);
2663 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2664 Chain = CurDAG->getEntryNode();
2666 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2667 InFlag = Chain.getValue(1);
2670 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2671 LoReg, N0, SDValue()).getValue(1);
2672 if (isSigned && !signBitIsZero) {
2673 // Sign extend the low part into the high part.
2675 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2677 // Zero out the high part, effectively zero extending the input.
2678 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2679 switch (NVT.SimpleTy) {
2682 SDValue(CurDAG->getMachineNode(
2683 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2684 CurDAG->getTargetConstant(X86::sub_16bit, dl,
2692 SDValue(CurDAG->getMachineNode(
2693 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2694 CurDAG->getTargetConstant(0, dl, MVT::i64), ClrNode,
2695 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2700 llvm_unreachable("Unexpected division source");
2703 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2704 ClrNode, InFlag).getValue(1);
2709 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2712 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2713 InFlag = SDValue(CNode, 1);
2714 // Update the chain.
2715 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2718 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2721 // Prevent use of AH in a REX instruction by explicitly copying it to
2722 // an ABCD_L register.
2724 // The current assumption of the register allocator is that isel
2725 // won't generate explicit references to the GR8_ABCD_H registers. If
2726 // the allocator and/or the backend get enhanced to be more robust in
2727 // that regard, this can be, and should be, removed.
2728 if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
2729 SDValue AHCopy = CurDAG->getRegister(X86::AH, MVT::i8);
2730 unsigned AHExtOpcode =
2731 isSigned ? X86::MOVSX32_NOREXrr8 : X86::MOVZX32_NOREXrr8;
2733 SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
2734 MVT::Glue, AHCopy, InFlag);
2735 SDValue Result(RNode, 0);
2736 InFlag = SDValue(RNode, 1);
2738 if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
2739 Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
2740 if (Node->getValueType(1) == MVT::i64) {
2741 // It's not possible to directly movsx AH to a 64bit register, because
2742 // the latter needs the REX prefix, but the former can't have it.
2743 assert(Opcode != X86ISD::SDIVREM8_SEXT_HREG &&
2744 "Unexpected i64 sext of h-register");
2746 SDValue(CurDAG->getMachineNode(
2747 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2748 CurDAG->getTargetConstant(0, dl, MVT::i64), Result,
2749 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2755 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
2757 ReplaceUses(SDValue(Node, 1), Result);
2758 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2760 // Copy the division (low) result, if it is needed.
2761 if (!SDValue(Node, 0).use_empty()) {
2762 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2763 LoReg, NVT, InFlag);
2764 InFlag = Result.getValue(2);
2765 ReplaceUses(SDValue(Node, 0), Result);
2766 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2768 // Copy the remainder (high) result, if it is needed.
2769 if (!SDValue(Node, 1).use_empty()) {
2770 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2771 HiReg, NVT, InFlag);
2772 InFlag = Result.getValue(2);
2773 ReplaceUses(SDValue(Node, 1), Result);
2774 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2781 // Sometimes a SUB is used to perform comparison.
2782 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2783 // This node is not a CMP.
2785 SDValue N0 = Node->getOperand(0);
2786 SDValue N1 = Node->getOperand(1);
2788 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2789 hasNoSignedComparisonUses(Node))
2790 N0 = N0.getOperand(0);
2792 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2793 // use a smaller encoding.
2794 // Look past the truncate if CMP is the only use of it.
2795 if ((N0.getNode()->getOpcode() == ISD::AND ||
2796 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2797 N0.getNode()->hasOneUse() &&
2798 N0.getValueType() != MVT::i8 &&
2799 X86::isZeroNode(N1)) {
2800 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2803 // For example, convert "testl %eax, $8" to "testb %al, $8"
2804 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2805 (!(C->getZExtValue() & 0x80) ||
2806 hasNoSignedComparisonUses(Node))) {
2807 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl, MVT::i8);
2808 SDValue Reg = N0.getNode()->getOperand(0);
2810 // On x86-32, only the ABCD registers have 8-bit subregisters.
2811 if (!Subtarget->is64Bit()) {
2812 const TargetRegisterClass *TRC;
2813 switch (N0.getSimpleValueType().SimpleTy) {
2814 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2815 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2816 default: llvm_unreachable("Unsupported TEST operand type!");
2818 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2819 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2820 Reg.getValueType(), Reg, RC), 0);
2823 // Extract the l-register.
2824 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2828 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2830 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2831 // one, do not call ReplaceAllUsesWith.
2832 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2833 SDValue(NewNode, 0));
2837 // For example, "testl %eax, $2048" to "testb %ah, $8".
2838 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2839 (!(C->getZExtValue() & 0x8000) ||
2840 hasNoSignedComparisonUses(Node))) {
2841 // Shift the immediate right by 8 bits.
2842 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2844 SDValue Reg = N0.getNode()->getOperand(0);
2846 // Put the value in an ABCD register.
2847 const TargetRegisterClass *TRC;
2848 switch (N0.getSimpleValueType().SimpleTy) {
2849 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2850 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2851 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2852 default: llvm_unreachable("Unsupported TEST operand type!");
2854 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2855 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2856 Reg.getValueType(), Reg, RC), 0);
2858 // Extract the h-register.
2859 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2862 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2863 // target GR8_NOREX registers, so make sure the register class is
2865 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2866 MVT::i32, Subreg, ShiftedImm);
2867 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2868 // one, do not call ReplaceAllUsesWith.
2869 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2870 SDValue(NewNode, 0));
2874 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2875 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2876 N0.getValueType() != MVT::i16 &&
2877 (!(C->getZExtValue() & 0x8000) ||
2878 hasNoSignedComparisonUses(Node))) {
2879 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2881 SDValue Reg = N0.getNode()->getOperand(0);
2883 // Extract the 16-bit subregister.
2884 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2888 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2890 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2891 // one, do not call ReplaceAllUsesWith.
2892 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2893 SDValue(NewNode, 0));
2897 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2898 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2899 N0.getValueType() == MVT::i64 &&
2900 (!(C->getZExtValue() & 0x80000000) ||
2901 hasNoSignedComparisonUses(Node))) {
2902 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2904 SDValue Reg = N0.getNode()->getOperand(0);
2906 // Extract the 32-bit subregister.
2907 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2911 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2913 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2914 // one, do not call ReplaceAllUsesWith.
2915 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2916 SDValue(NewNode, 0));
2923 // Change a chain of {load; incr or dec; store} of the same value into
2924 // a simple increment or decrement through memory of that value, if the
2925 // uses of the modified value and its address are suitable.
2926 // The DEC64m tablegen pattern is currently not able to match the case where
2927 // the EFLAGS on the original DEC are used. (This also applies to
2928 // {INC,DEC}X{64,32,16,8}.)
2929 // We'll need to improve tablegen to allow flags to be transferred from a
2930 // node in the pattern to the result node. probably with a new keyword
2931 // for example, we have this
2932 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2933 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2934 // (implicit EFLAGS)]>;
2935 // but maybe need something like this
2936 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2937 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2938 // (transferrable EFLAGS)]>;
2940 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2941 SDValue StoredVal = StoreNode->getOperand(1);
2942 unsigned Opc = StoredVal->getOpcode();
2944 LoadSDNode *LoadNode = nullptr;
2946 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2947 LoadNode, InputChain))
2950 SDValue Base, Scale, Index, Disp, Segment;
2951 if (!selectAddr(LoadNode, LoadNode->getBasePtr(),
2952 Base, Scale, Index, Disp, Segment))
2955 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2956 MemOp[0] = StoreNode->getMemOperand();
2957 MemOp[1] = LoadNode->getMemOperand();
2958 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2959 EVT LdVT = LoadNode->getMemoryVT();
2960 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2961 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2963 MVT::i32, MVT::Other, Ops);
2964 Result->setMemRefs(MemOp, MemOp + 2);
2966 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2967 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2973 SDNode *ResNode = SelectCode(Node);
2975 DEBUG(dbgs() << "=> ";
2976 if (ResNode == nullptr || ResNode == Node)
2979 ResNode->dump(CurDAG);
2985 bool X86DAGToDAGISel::
2986 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
2987 std::vector<SDValue> &OutOps) {
2988 SDValue Op0, Op1, Op2, Op3, Op4;
2989 switch (ConstraintID) {
2991 llvm_unreachable("Unexpected asm memory constraint");
2992 case InlineAsm::Constraint_i:
2993 // FIXME: It seems strange that 'i' is needed here since it's supposed to
2994 // be an immediate and not a memory constraint.
2996 case InlineAsm::Constraint_o: // offsetable ??
2997 case InlineAsm::Constraint_v: // not offsetable ??
2998 case InlineAsm::Constraint_m: // memory
2999 case InlineAsm::Constraint_X:
3000 if (!selectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
3005 OutOps.push_back(Op0);
3006 OutOps.push_back(Op1);
3007 OutOps.push_back(Op2);
3008 OutOps.push_back(Op3);
3009 OutOps.push_back(Op4);
3013 /// This pass converts a legalized DAG into a X86-specific DAG,
3014 /// ready for instruction scheduling.
3015 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
3016 CodeGenOpt::Level OptLevel) {
3017 return new X86DAGToDAGISel(TM, OptLevel);