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
161 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
162 : SelectionDAGISel(tm, OptLevel), OptForSize(false) {}
164 const char *getPassName() const override {
165 return "X86 DAG->DAG Instruction Selection";
168 bool runOnMachineFunction(MachineFunction &MF) override {
169 // Reset the subtarget each time through.
170 Subtarget = &MF.getSubtarget<X86Subtarget>();
171 SelectionDAGISel::runOnMachineFunction(MF);
175 void EmitFunctionEntryCode() override;
177 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
179 void PreprocessISelDAG() override;
181 inline bool immSext8(SDNode *N) const {
182 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
185 // True if the 64-bit immediate fits in a 32-bit sign-extended field.
186 inline bool i64immSExt32(SDNode *N) const {
187 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
188 return (int64_t)v == (int32_t)v;
191 // Include the pieces autogenerated from the target description.
192 #include "X86GenDAGISel.inc"
195 SDNode *Select(SDNode *N) override;
196 SDNode *selectGather(SDNode *N, unsigned Opc);
197 SDNode *selectAtomicLoadArith(SDNode *Node, MVT NVT);
199 bool foldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
200 bool matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
201 bool matchWrapper(SDValue N, X86ISelAddressMode &AM);
202 bool matchAddress(SDValue N, X86ISelAddressMode &AM);
203 bool matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
205 bool matchAddressBase(SDValue N, X86ISelAddressMode &AM);
206 bool selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
207 SDValue &Scale, SDValue &Index, SDValue &Disp,
209 bool selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
210 SDValue &Scale, SDValue &Index, SDValue &Disp,
212 bool selectMOV64Imm32(SDValue N, SDValue &Imm);
213 bool selectLEAAddr(SDValue N, SDValue &Base,
214 SDValue &Scale, SDValue &Index, SDValue &Disp,
216 bool selectLEA64_32Addr(SDValue N, SDValue &Base,
217 SDValue &Scale, SDValue &Index, SDValue &Disp,
219 bool selectTLSADDRAddr(SDValue N, SDValue &Base,
220 SDValue &Scale, SDValue &Index, SDValue &Disp,
222 bool selectScalarSSELoad(SDNode *Root, SDValue N,
223 SDValue &Base, SDValue &Scale,
224 SDValue &Index, SDValue &Disp,
226 SDValue &NodeWithChain);
228 bool tryFoldLoad(SDNode *P, SDValue N,
229 SDValue &Base, SDValue &Scale,
230 SDValue &Index, SDValue &Disp,
233 /// Implement addressing mode selection for inline asm expressions.
234 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
235 unsigned ConstraintID,
236 std::vector<SDValue> &OutOps) override;
238 void emitSpecialCodeForMain();
240 inline void getAddressOperands(X86ISelAddressMode &AM, SDLoc DL,
241 SDValue &Base, SDValue &Scale,
242 SDValue &Index, SDValue &Disp,
244 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
245 ? CurDAG->getTargetFrameIndex(
247 TLI->getPointerTy(CurDAG->getDataLayout()))
249 Scale = getI8Imm(AM.Scale, DL);
251 // These are 32-bit even in 64-bit mode since RIP-relative offset
254 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
258 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
259 AM.Align, AM.Disp, AM.SymbolFlags);
261 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
262 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
263 } else if (AM.MCSym) {
264 assert(!AM.Disp && "Non-zero displacement is ignored with MCSym.");
265 assert(AM.SymbolFlags == 0 && "oo");
266 Disp = CurDAG->getMCSymbol(AM.MCSym, MVT::i32);
267 } else if (AM.JT != -1) {
268 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
269 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
270 } else if (AM.BlockAddr)
271 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
274 Disp = CurDAG->getTargetConstant(AM.Disp, DL, MVT::i32);
276 if (AM.Segment.getNode())
277 Segment = AM.Segment;
279 Segment = CurDAG->getRegister(0, MVT::i32);
282 // Utility function to determine whether we should avoid selecting
283 // immediate forms of instructions for better code size or not.
284 // At a high level, we'd like to avoid such instructions when
285 // we have similar constants used within the same basic block
286 // that can be kept in a register.
288 bool shouldAvoidImmediateInstFormsForSize(SDNode *N) const {
289 uint32_t UseCount = 0;
291 // Do not want to hoist if we're not optimizing for size.
292 // TODO: We'd like to remove this restriction.
293 // See the comment in X86InstrInfo.td for more info.
297 // Walk all the users of the immediate.
298 for (SDNode::use_iterator UI = N->use_begin(),
299 UE = N->use_end(); (UI != UE) && (UseCount < 2); ++UI) {
303 // This user is already selected. Count it as a legitimate use and
305 if (User->isMachineOpcode()) {
310 // We want to count stores of immediates as real uses.
311 if (User->getOpcode() == ISD::STORE &&
312 User->getOperand(1).getNode() == N) {
317 // We don't currently match users that have > 2 operands (except
318 // for stores, which are handled above)
319 // Those instruction won't match in ISEL, for now, and would
320 // be counted incorrectly.
321 // This may change in the future as we add additional instruction
323 if (User->getNumOperands() != 2)
326 // Immediates that are used for offsets as part of stack
327 // manipulation should be left alone. These are typically
328 // used to indicate SP offsets for argument passing and
329 // will get pulled into stores/pushes (implicitly).
330 if (User->getOpcode() == X86ISD::ADD ||
331 User->getOpcode() == ISD::ADD ||
332 User->getOpcode() == X86ISD::SUB ||
333 User->getOpcode() == ISD::SUB) {
335 // Find the other operand of the add/sub.
336 SDValue OtherOp = User->getOperand(0);
337 if (OtherOp.getNode() == N)
338 OtherOp = User->getOperand(1);
340 // Don't count if the other operand is SP.
341 RegisterSDNode *RegNode;
342 if (OtherOp->getOpcode() == ISD::CopyFromReg &&
343 (RegNode = dyn_cast_or_null<RegisterSDNode>(
344 OtherOp->getOperand(1).getNode())))
345 if ((RegNode->getReg() == X86::ESP) ||
346 (RegNode->getReg() == X86::RSP))
350 // ... otherwise, count this and move on.
354 // If we have more than 1 use, then recommend for hoisting.
355 return (UseCount > 1);
358 /// Return a target constant with the specified value of type i8.
359 inline SDValue getI8Imm(unsigned Imm, SDLoc DL) {
360 return CurDAG->getTargetConstant(Imm, DL, MVT::i8);
363 /// Return a target constant with the specified value, of type i32.
364 inline SDValue getI32Imm(unsigned Imm, SDLoc DL) {
365 return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
368 /// Return an SDNode that returns the value of the global base register.
369 /// Output instructions required to initialize the global base register,
371 SDNode *getGlobalBaseReg();
373 /// Return a reference to the TargetMachine, casted to the target-specific
375 const X86TargetMachine &getTargetMachine() const {
376 return static_cast<const X86TargetMachine &>(TM);
379 /// Return a reference to the TargetInstrInfo, casted to the target-specific
381 const X86InstrInfo *getInstrInfo() const {
382 return Subtarget->getInstrInfo();
385 /// \brief Address-mode matching performs shift-of-and to and-of-shift
386 /// reassociation in order to expose more scaled addressing
388 bool ComplexPatternFuncMutatesDAG() const override {
396 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
397 if (OptLevel == CodeGenOpt::None) return false;
402 if (N.getOpcode() != ISD::LOAD)
405 // If N is a load, do additional profitability checks.
407 switch (U->getOpcode()) {
420 SDValue Op1 = U->getOperand(1);
422 // If the other operand is a 8-bit immediate we should fold the immediate
423 // instead. This reduces code size.
425 // movl 4(%esp), %eax
429 // addl 4(%esp), %eax
430 // The former is 2 bytes shorter. In case where the increment is 1, then
431 // the saving can be 4 bytes (by using incl %eax).
432 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
433 if (Imm->getAPIntValue().isSignedIntN(8))
436 // If the other operand is a TLS address, we should fold it instead.
439 // leal i@NTPOFF(%eax), %eax
441 // movl $i@NTPOFF, %eax
443 // if the block also has an access to a second TLS address this will save
445 // FIXME: This is probably also true for non-TLS addresses.
446 if (Op1.getOpcode() == X86ISD::Wrapper) {
447 SDValue Val = Op1.getOperand(0);
448 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
458 /// Replace the original chain operand of the call with
459 /// load's chain operand and move load below the call's chain operand.
460 static void moveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
461 SDValue Call, SDValue OrigChain) {
462 SmallVector<SDValue, 8> Ops;
463 SDValue Chain = OrigChain.getOperand(0);
464 if (Chain.getNode() == Load.getNode())
465 Ops.push_back(Load.getOperand(0));
467 assert(Chain.getOpcode() == ISD::TokenFactor &&
468 "Unexpected chain operand");
469 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
470 if (Chain.getOperand(i).getNode() == Load.getNode())
471 Ops.push_back(Load.getOperand(0));
473 Ops.push_back(Chain.getOperand(i));
475 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
477 Ops.push_back(NewChain);
479 Ops.append(OrigChain->op_begin() + 1, OrigChain->op_end());
480 CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
481 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
482 Load.getOperand(1), Load.getOperand(2));
485 Ops.push_back(SDValue(Load.getNode(), 1));
486 Ops.append(Call->op_begin() + 1, Call->op_end());
487 CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
490 /// Return true if call address is a load and it can be
491 /// moved below CALLSEQ_START and the chains leading up to the call.
492 /// Return the CALLSEQ_START by reference as a second output.
493 /// In the case of a tail call, there isn't a callseq node between the call
494 /// chain and the load.
495 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
496 // The transformation is somewhat dangerous if the call's chain was glued to
497 // the call. After MoveBelowOrigChain the load is moved between the call and
498 // the chain, this can create a cycle if the load is not folded. So it is
499 // *really* important that we are sure the load will be folded.
500 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
502 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
505 LD->getAddressingMode() != ISD::UNINDEXED ||
506 LD->getExtensionType() != ISD::NON_EXTLOAD)
509 // Now let's find the callseq_start.
510 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
511 if (!Chain.hasOneUse())
513 Chain = Chain.getOperand(0);
516 if (!Chain.getNumOperands())
518 // Since we are not checking for AA here, conservatively abort if the chain
519 // writes to memory. It's not safe to move the callee (a load) across a store.
520 if (isa<MemSDNode>(Chain.getNode()) &&
521 cast<MemSDNode>(Chain.getNode())->writeMem())
523 if (Chain.getOperand(0).getNode() == Callee.getNode())
525 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
526 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
527 Callee.getValue(1).hasOneUse())
532 void X86DAGToDAGISel::PreprocessISelDAG() {
533 // OptForSize is used in pattern predicates that isel is matching.
534 OptForSize = MF->getFunction()->optForSize();
536 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
537 E = CurDAG->allnodes_end(); I != E; ) {
538 SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
540 if (OptLevel != CodeGenOpt::None &&
541 // Only does this when target favors doesn't favor register indirect
543 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
544 (N->getOpcode() == X86ISD::TC_RETURN &&
545 // Only does this if load can be folded into TC_RETURN.
546 (Subtarget->is64Bit() ||
547 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
548 /// Also try moving call address load from outside callseq_start to just
549 /// before the call to allow it to be folded.
567 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
568 SDValue Chain = N->getOperand(0);
569 SDValue Load = N->getOperand(1);
570 if (!isCalleeLoad(Load, Chain, HasCallSeq))
572 moveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
577 // Lower fpround and fpextend nodes that target the FP stack to be store and
578 // load to the stack. This is a gross hack. We would like to simply mark
579 // these as being illegal, but when we do that, legalize produces these when
580 // it expands calls, then expands these in the same legalize pass. We would
581 // like dag combine to be able to hack on these between the call expansion
582 // and the node legalization. As such this pass basically does "really
583 // late" legalization of these inline with the X86 isel pass.
584 // FIXME: This should only happen when not compiled with -O0.
585 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
588 MVT SrcVT = N->getOperand(0).getSimpleValueType();
589 MVT DstVT = N->getSimpleValueType(0);
591 // If any of the sources are vectors, no fp stack involved.
592 if (SrcVT.isVector() || DstVT.isVector())
595 // If the source and destination are SSE registers, then this is a legal
596 // conversion that should not be lowered.
597 const X86TargetLowering *X86Lowering =
598 static_cast<const X86TargetLowering *>(TLI);
599 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
600 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
601 if (SrcIsSSE && DstIsSSE)
604 if (!SrcIsSSE && !DstIsSSE) {
605 // If this is an FPStack extension, it is a noop.
606 if (N->getOpcode() == ISD::FP_EXTEND)
608 // If this is a value-preserving FPStack truncation, it is a noop.
609 if (N->getConstantOperandVal(1))
613 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
614 // FPStack has extload and truncstore. SSE can fold direct loads into other
615 // operations. Based on this, decide what we want to do.
617 if (N->getOpcode() == ISD::FP_ROUND)
618 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
620 MemVT = SrcIsSSE ? SrcVT : DstVT;
622 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
625 // FIXME: optimize the case where the src/dest is a load or store?
626 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
628 MemTmp, MachinePointerInfo(), MemVT,
630 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
631 MachinePointerInfo(),
632 MemVT, false, false, false, 0);
634 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
635 // extload we created. This will cause general havok on the dag because
636 // anything below the conversion could be folded into other existing nodes.
637 // To avoid invalidating 'I', back it up to the convert node.
639 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
641 // Now that we did that, the node is dead. Increment the iterator to the
642 // next node to process, then delete N.
644 CurDAG->DeleteNode(N);
649 /// Emit any code that needs to be executed only in the main function.
650 void X86DAGToDAGISel::emitSpecialCodeForMain() {
651 if (Subtarget->isTargetCygMing()) {
652 TargetLowering::ArgListTy Args;
653 auto &DL = CurDAG->getDataLayout();
655 TargetLowering::CallLoweringInfo CLI(*CurDAG);
656 CLI.setChain(CurDAG->getRoot())
657 .setCallee(CallingConv::C, Type::getVoidTy(*CurDAG->getContext()),
658 CurDAG->getExternalSymbol("__main", TLI->getPointerTy(DL)),
660 const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
661 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
662 CurDAG->setRoot(Result.second);
666 void X86DAGToDAGISel::EmitFunctionEntryCode() {
667 // If this is main, emit special code for main.
668 if (const Function *Fn = MF->getFunction())
669 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
670 emitSpecialCodeForMain();
673 static bool isDispSafeForFrameIndex(int64_t Val) {
674 // On 64-bit platforms, we can run into an issue where a frame index
675 // includes a displacement that, when added to the explicit displacement,
676 // will overflow the displacement field. Assuming that the frame index
677 // displacement fits into a 31-bit integer (which is only slightly more
678 // aggressive than the current fundamental assumption that it fits into
679 // a 32-bit integer), a 31-bit disp should always be safe.
680 return isInt<31>(Val);
683 bool X86DAGToDAGISel::foldOffsetIntoAddress(uint64_t Offset,
684 X86ISelAddressMode &AM) {
685 // Cannot combine ExternalSymbol displacements with integer offsets.
686 if (Offset != 0 && (AM.ES || AM.MCSym))
688 int64_t Val = AM.Disp + Offset;
689 CodeModel::Model M = TM.getCodeModel();
690 if (Subtarget->is64Bit()) {
691 if (!X86::isOffsetSuitableForCodeModel(Val, M,
692 AM.hasSymbolicDisplacement()))
694 // In addition to the checks required for a register base, check that
695 // we do not try to use an unsafe Disp with a frame index.
696 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
697 !isDispSafeForFrameIndex(Val))
705 bool X86DAGToDAGISel::matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
706 SDValue Address = N->getOperand(1);
708 // load gs:0 -> GS segment register.
709 // load fs:0 -> FS segment register.
711 // This optimization is valid because the GNU TLS model defines that
712 // gs:0 (or fs:0 on X86-64) contains its own address.
713 // For more information see http://people.redhat.com/drepper/tls.pdf
714 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
715 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
716 Subtarget->isTargetLinux())
717 switch (N->getPointerInfo().getAddrSpace()) {
719 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
722 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
729 /// Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes into an addressing
730 /// mode. These wrap things that will resolve down into a symbol reference.
731 /// If no match is possible, this returns true, otherwise it returns false.
732 bool X86DAGToDAGISel::matchWrapper(SDValue N, X86ISelAddressMode &AM) {
733 // If the addressing mode already has a symbol as the displacement, we can
734 // never match another symbol.
735 if (AM.hasSymbolicDisplacement())
738 SDValue N0 = N.getOperand(0);
739 CodeModel::Model M = TM.getCodeModel();
741 // Handle X86-64 rip-relative addresses. We check this before checking direct
742 // folding because RIP is preferable to non-RIP accesses.
743 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
744 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
745 // they cannot be folded into immediate fields.
746 // FIXME: This can be improved for kernel and other models?
747 (M == CodeModel::Small || M == CodeModel::Kernel)) {
748 // Base and index reg must be 0 in order to use %rip as base.
749 if (AM.hasBaseOrIndexReg())
751 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
752 X86ISelAddressMode Backup = AM;
753 AM.GV = G->getGlobal();
754 AM.SymbolFlags = G->getTargetFlags();
755 if (foldOffsetIntoAddress(G->getOffset(), AM)) {
759 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
760 X86ISelAddressMode Backup = AM;
761 AM.CP = CP->getConstVal();
762 AM.Align = CP->getAlignment();
763 AM.SymbolFlags = CP->getTargetFlags();
764 if (foldOffsetIntoAddress(CP->getOffset(), AM)) {
768 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
769 AM.ES = S->getSymbol();
770 AM.SymbolFlags = S->getTargetFlags();
771 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
772 AM.MCSym = S->getMCSymbol();
773 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
774 AM.JT = J->getIndex();
775 AM.SymbolFlags = J->getTargetFlags();
776 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
777 X86ISelAddressMode Backup = AM;
778 AM.BlockAddr = BA->getBlockAddress();
779 AM.SymbolFlags = BA->getTargetFlags();
780 if (foldOffsetIntoAddress(BA->getOffset(), AM)) {
785 llvm_unreachable("Unhandled symbol reference node.");
787 if (N.getOpcode() == X86ISD::WrapperRIP)
788 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
792 // Handle the case when globals fit in our immediate field: This is true for
793 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
794 // mode, this only applies to a non-RIP-relative computation.
795 if (!Subtarget->is64Bit() ||
796 M == CodeModel::Small || M == CodeModel::Kernel) {
797 assert(N.getOpcode() != X86ISD::WrapperRIP &&
798 "RIP-relative addressing already handled");
799 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
800 AM.GV = G->getGlobal();
801 AM.Disp += G->getOffset();
802 AM.SymbolFlags = G->getTargetFlags();
803 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
804 AM.CP = CP->getConstVal();
805 AM.Align = CP->getAlignment();
806 AM.Disp += CP->getOffset();
807 AM.SymbolFlags = CP->getTargetFlags();
808 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
809 AM.ES = S->getSymbol();
810 AM.SymbolFlags = S->getTargetFlags();
811 } else if (auto *S = dyn_cast<MCSymbolSDNode>(N0)) {
812 AM.MCSym = S->getMCSymbol();
813 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
814 AM.JT = J->getIndex();
815 AM.SymbolFlags = J->getTargetFlags();
816 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
817 AM.BlockAddr = BA->getBlockAddress();
818 AM.Disp += BA->getOffset();
819 AM.SymbolFlags = BA->getTargetFlags();
821 llvm_unreachable("Unhandled symbol reference node.");
828 /// Add the specified node to the specified addressing mode, returning true if
829 /// it cannot be done. This just pattern matches for the addressing mode.
830 bool X86DAGToDAGISel::matchAddress(SDValue N, X86ISelAddressMode &AM) {
831 if (matchAddressRecursively(N, AM, 0))
834 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
835 // a smaller encoding and avoids a scaled-index.
837 AM.BaseType == X86ISelAddressMode::RegBase &&
838 AM.Base_Reg.getNode() == nullptr) {
839 AM.Base_Reg = AM.IndexReg;
843 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
844 // because it has a smaller encoding.
845 // TODO: Which other code models can use this?
846 if (TM.getCodeModel() == CodeModel::Small &&
847 Subtarget->is64Bit() &&
849 AM.BaseType == X86ISelAddressMode::RegBase &&
850 AM.Base_Reg.getNode() == nullptr &&
851 AM.IndexReg.getNode() == nullptr &&
852 AM.SymbolFlags == X86II::MO_NO_FLAG &&
853 AM.hasSymbolicDisplacement())
854 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
859 // Insert a node into the DAG at least before the Pos node's position. This
860 // will reposition the node as needed, and will assign it a node ID that is <=
861 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
862 // IDs! The selection DAG must no longer depend on their uniqueness when this
864 static void insertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
865 if (N.getNode()->getNodeId() == -1 ||
866 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
867 DAG.RepositionNode(Pos.getNode()->getIterator(), N.getNode());
868 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
872 // Transform "(X >> (8-C1)) & (0xff << C1)" to "((X >> 8) & 0xff) << C1" if
873 // safe. This allows us to convert the shift and and into an h-register
874 // extract and a scaled index. Returns false if the simplification is
876 static bool foldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
878 SDValue Shift, SDValue X,
879 X86ISelAddressMode &AM) {
880 if (Shift.getOpcode() != ISD::SRL ||
881 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
885 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
886 if (ScaleLog <= 0 || ScaleLog >= 4 ||
887 Mask != (0xffu << ScaleLog))
890 MVT VT = N.getSimpleValueType();
892 SDValue Eight = DAG.getConstant(8, DL, MVT::i8);
893 SDValue NewMask = DAG.getConstant(0xff, DL, VT);
894 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
895 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
896 SDValue ShlCount = DAG.getConstant(ScaleLog, DL, MVT::i8);
897 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
899 // Insert the new nodes into the topological ordering. We must do this in
900 // a valid topological ordering as nothing is going to go back and re-sort
901 // these nodes. We continually insert before 'N' in sequence as this is
902 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
903 // hierarchy left to express.
904 insertDAGNode(DAG, N, Eight);
905 insertDAGNode(DAG, N, Srl);
906 insertDAGNode(DAG, N, NewMask);
907 insertDAGNode(DAG, N, And);
908 insertDAGNode(DAG, N, ShlCount);
909 insertDAGNode(DAG, N, Shl);
910 DAG.ReplaceAllUsesWith(N, Shl);
912 AM.Scale = (1 << ScaleLog);
916 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
917 // allows us to fold the shift into this addressing mode. Returns false if the
918 // transform succeeded.
919 static bool foldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
921 SDValue Shift, SDValue X,
922 X86ISelAddressMode &AM) {
923 if (Shift.getOpcode() != ISD::SHL ||
924 !isa<ConstantSDNode>(Shift.getOperand(1)))
927 // Not likely to be profitable if either the AND or SHIFT node has more
928 // than one use (unless all uses are for address computation). Besides,
929 // isel mechanism requires their node ids to be reused.
930 if (!N.hasOneUse() || !Shift.hasOneUse())
933 // Verify that the shift amount is something we can fold.
934 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
935 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
938 MVT VT = N.getSimpleValueType();
940 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, DL, VT);
941 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
942 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
944 // Insert the new nodes into the topological ordering. We must do this in
945 // a valid topological ordering as nothing is going to go back and re-sort
946 // these nodes. We continually insert before 'N' in sequence as this is
947 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
948 // hierarchy left to express.
949 insertDAGNode(DAG, N, NewMask);
950 insertDAGNode(DAG, N, NewAnd);
951 insertDAGNode(DAG, N, NewShift);
952 DAG.ReplaceAllUsesWith(N, NewShift);
954 AM.Scale = 1 << ShiftAmt;
955 AM.IndexReg = NewAnd;
959 // Implement some heroics to detect shifts of masked values where the mask can
960 // be replaced by extending the shift and undoing that in the addressing mode
961 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
962 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
963 // the addressing mode. This results in code such as:
965 // int f(short *y, int *lookup_table) {
967 // return *y + lookup_table[*y >> 11];
971 // movzwl (%rdi), %eax
974 // addl (%rsi,%rcx,4), %eax
977 // movzwl (%rdi), %eax
981 // addl (%rsi,%rcx), %eax
983 // Note that this function assumes the mask is provided as a mask *after* the
984 // value is shifted. The input chain may or may not match that, but computing
985 // such a mask is trivial.
986 static bool foldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
988 SDValue Shift, SDValue X,
989 X86ISelAddressMode &AM) {
990 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
991 !isa<ConstantSDNode>(Shift.getOperand(1)))
994 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
995 unsigned MaskLZ = countLeadingZeros(Mask);
996 unsigned MaskTZ = countTrailingZeros(Mask);
998 // The amount of shift we're trying to fit into the addressing mode is taken
999 // from the trailing zeros of the mask.
1000 unsigned AMShiftAmt = MaskTZ;
1002 // There is nothing we can do here unless the mask is removing some bits.
1003 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
1004 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
1006 // We also need to ensure that mask is a continuous run of bits.
1007 if (countTrailingOnes(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
1009 // Scale the leading zero count down based on the actual size of the value.
1010 // Also scale it down based on the size of the shift.
1011 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
1013 // The final check is to ensure that any masked out high bits of X are
1014 // already known to be zero. Otherwise, the mask has a semantic impact
1015 // other than masking out a couple of low bits. Unfortunately, because of
1016 // the mask, zero extensions will be removed from operands in some cases.
1017 // This code works extra hard to look through extensions because we can
1018 // replace them with zero extensions cheaply if necessary.
1019 bool ReplacingAnyExtend = false;
1020 if (X.getOpcode() == ISD::ANY_EXTEND) {
1021 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
1022 X.getOperand(0).getSimpleValueType().getSizeInBits();
1023 // Assume that we'll replace the any-extend with a zero-extend, and
1024 // narrow the search to the extended value.
1025 X = X.getOperand(0);
1026 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
1027 ReplacingAnyExtend = true;
1029 APInt MaskedHighBits =
1030 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
1031 APInt KnownZero, KnownOne;
1032 DAG.computeKnownBits(X, KnownZero, KnownOne);
1033 if (MaskedHighBits != KnownZero) return true;
1035 // We've identified a pattern that can be transformed into a single shift
1036 // and an addressing mode. Make it so.
1037 MVT VT = N.getSimpleValueType();
1038 if (ReplacingAnyExtend) {
1039 assert(X.getValueType() != VT);
1040 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
1041 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
1042 insertDAGNode(DAG, N, NewX);
1046 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, DL, MVT::i8);
1047 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
1048 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, DL, MVT::i8);
1049 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
1051 // Insert the new nodes into the topological ordering. We must do this in
1052 // a valid topological ordering as nothing is going to go back and re-sort
1053 // these nodes. We continually insert before 'N' in sequence as this is
1054 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
1055 // hierarchy left to express.
1056 insertDAGNode(DAG, N, NewSRLAmt);
1057 insertDAGNode(DAG, N, NewSRL);
1058 insertDAGNode(DAG, N, NewSHLAmt);
1059 insertDAGNode(DAG, N, NewSHL);
1060 DAG.ReplaceAllUsesWith(N, NewSHL);
1062 AM.Scale = 1 << AMShiftAmt;
1063 AM.IndexReg = NewSRL;
1067 bool X86DAGToDAGISel::matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
1071 dbgs() << "MatchAddress: ";
1076 return matchAddressBase(N, AM);
1078 // If this is already a %rip relative address, we can only merge immediates
1079 // into it. Instead of handling this in every case, we handle it here.
1080 // RIP relative addressing: %rip + 32-bit displacement!
1081 if (AM.isRIPRelative()) {
1082 // FIXME: JumpTable and ExternalSymbol address currently don't like
1083 // displacements. It isn't very important, but this should be fixed for
1085 if (!(AM.ES || AM.MCSym) && AM.JT != -1)
1088 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
1089 if (!foldOffsetIntoAddress(Cst->getSExtValue(), AM))
1094 switch (N.getOpcode()) {
1096 case ISD::LOCAL_RECOVER: {
1097 if (!AM.hasSymbolicDisplacement() && AM.Disp == 0)
1098 if (const auto *ESNode = dyn_cast<MCSymbolSDNode>(N.getOperand(0))) {
1099 // Use the symbol and don't prefix it.
1100 AM.MCSym = ESNode->getMCSymbol();
1105 case ISD::Constant: {
1106 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1107 if (!foldOffsetIntoAddress(Val, AM))
1112 case X86ISD::Wrapper:
1113 case X86ISD::WrapperRIP:
1114 if (!matchWrapper(N, AM))
1119 if (!matchLoadInAddress(cast<LoadSDNode>(N), AM))
1123 case ISD::FrameIndex:
1124 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1125 AM.Base_Reg.getNode() == nullptr &&
1126 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1127 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1128 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1134 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1138 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1139 unsigned Val = CN->getZExtValue();
1140 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1141 // that the base operand remains free for further matching. If
1142 // the base doesn't end up getting used, a post-processing step
1143 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1144 if (Val == 1 || Val == 2 || Val == 3) {
1145 AM.Scale = 1 << Val;
1146 SDValue ShVal = N.getNode()->getOperand(0);
1148 // Okay, we know that we have a scale by now. However, if the scaled
1149 // value is an add of something and a constant, we can fold the
1150 // constant into the disp field here.
1151 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1152 AM.IndexReg = ShVal.getNode()->getOperand(0);
1153 ConstantSDNode *AddVal =
1154 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1155 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1156 if (!foldOffsetIntoAddress(Disp, AM))
1160 AM.IndexReg = ShVal;
1167 // Scale must not be used already.
1168 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1170 SDValue And = N.getOperand(0);
1171 if (And.getOpcode() != ISD::AND) break;
1172 SDValue X = And.getOperand(0);
1174 // We only handle up to 64-bit values here as those are what matter for
1175 // addressing mode optimizations.
1176 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1178 // The mask used for the transform is expected to be post-shift, but we
1179 // found the shift first so just apply the shift to the mask before passing
1181 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1182 !isa<ConstantSDNode>(And.getOperand(1)))
1184 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1186 // Try to fold the mask and shift into the scale, and return false if we
1188 if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1193 case ISD::SMUL_LOHI:
1194 case ISD::UMUL_LOHI:
1195 // A mul_lohi where we need the low part can be folded as a plain multiply.
1196 if (N.getResNo() != 0) break;
1199 case X86ISD::MUL_IMM:
1200 // X*[3,5,9] -> X+X*[2,4,8]
1201 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1202 AM.Base_Reg.getNode() == nullptr &&
1203 AM.IndexReg.getNode() == nullptr) {
1205 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1206 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1207 CN->getZExtValue() == 9) {
1208 AM.Scale = unsigned(CN->getZExtValue())-1;
1210 SDValue MulVal = N.getNode()->getOperand(0);
1213 // Okay, we know that we have a scale by now. However, if the scaled
1214 // value is an add of something and a constant, we can fold the
1215 // constant into the disp field here.
1216 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1217 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1218 Reg = MulVal.getNode()->getOperand(0);
1219 ConstantSDNode *AddVal =
1220 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1221 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1222 if (foldOffsetIntoAddress(Disp, AM))
1223 Reg = N.getNode()->getOperand(0);
1225 Reg = N.getNode()->getOperand(0);
1228 AM.IndexReg = AM.Base_Reg = Reg;
1235 // Given A-B, if A can be completely folded into the address and
1236 // the index field with the index field unused, use -B as the index.
1237 // This is a win if a has multiple parts that can be folded into
1238 // the address. Also, this saves a mov if the base register has
1239 // other uses, since it avoids a two-address sub instruction, however
1240 // it costs an additional mov if the index register has other uses.
1242 // Add an artificial use to this node so that we can keep track of
1243 // it if it gets CSE'd with a different node.
1244 HandleSDNode Handle(N);
1246 // Test if the LHS of the sub can be folded.
1247 X86ISelAddressMode Backup = AM;
1248 if (matchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1252 // Test if the index field is free for use.
1253 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1259 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1260 // If the RHS involves a register with multiple uses, this
1261 // transformation incurs an extra mov, due to the neg instruction
1262 // clobbering its operand.
1263 if (!RHS.getNode()->hasOneUse() ||
1264 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1265 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1266 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1267 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1268 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1270 // If the base is a register with multiple uses, this
1271 // transformation may save a mov.
1272 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1273 AM.Base_Reg.getNode() &&
1274 !AM.Base_Reg.getNode()->hasOneUse()) ||
1275 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1277 // If the folded LHS was interesting, this transformation saves
1278 // address arithmetic.
1279 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1280 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1281 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1283 // If it doesn't look like it may be an overall win, don't do it.
1289 // Ok, the transformation is legal and appears profitable. Go for it.
1290 SDValue Zero = CurDAG->getConstant(0, dl, N.getValueType());
1291 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1295 // Insert the new nodes into the topological ordering.
1296 insertDAGNode(*CurDAG, N, Zero);
1297 insertDAGNode(*CurDAG, N, Neg);
1302 // Add an artificial use to this node so that we can keep track of
1303 // it if it gets CSE'd with a different node.
1304 HandleSDNode Handle(N);
1306 X86ISelAddressMode Backup = AM;
1307 if (!matchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1308 !matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1312 // Try again after commuting the operands.
1313 if (!matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1314 !matchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1318 // If we couldn't fold both operands into the address at the same time,
1319 // see if we can just put each operand into a register and fold at least
1321 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1322 !AM.Base_Reg.getNode() &&
1323 !AM.IndexReg.getNode()) {
1324 N = Handle.getValue();
1325 AM.Base_Reg = N.getOperand(0);
1326 AM.IndexReg = N.getOperand(1);
1330 N = Handle.getValue();
1335 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1336 if (CurDAG->isBaseWithConstantOffset(N)) {
1337 X86ISelAddressMode Backup = AM;
1338 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1340 // Start with the LHS as an addr mode.
1341 if (!matchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1342 !foldOffsetIntoAddress(CN->getSExtValue(), AM))
1349 // Perform some heroic transforms on an and of a constant-count shift
1350 // with a constant to enable use of the scaled offset field.
1352 // Scale must not be used already.
1353 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1355 SDValue Shift = N.getOperand(0);
1356 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1357 SDValue X = Shift.getOperand(0);
1359 // We only handle up to 64-bit values here as those are what matter for
1360 // addressing mode optimizations.
1361 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1363 if (!isa<ConstantSDNode>(N.getOperand(1)))
1365 uint64_t Mask = N.getConstantOperandVal(1);
1367 // Try to fold the mask and shift into an extract and scale.
1368 if (!foldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1371 // Try to fold the mask and shift directly into the scale.
1372 if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1375 // Try to swap the mask and shift to place shifts which can be done as
1376 // a scale on the outside of the mask.
1377 if (!foldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1383 return matchAddressBase(N, AM);
1386 /// Helper for MatchAddress. Add the specified node to the
1387 /// specified addressing mode without any further recursion.
1388 bool X86DAGToDAGISel::matchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1389 // Is the base register already occupied?
1390 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1391 // If so, check to see if the scale index register is set.
1392 if (!AM.IndexReg.getNode()) {
1398 // Otherwise, we cannot select it.
1402 // Default, generate it as a register.
1403 AM.BaseType = X86ISelAddressMode::RegBase;
1408 bool X86DAGToDAGISel::selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
1409 SDValue &Scale, SDValue &Index,
1410 SDValue &Disp, SDValue &Segment) {
1412 MaskedGatherScatterSDNode *Mgs = dyn_cast<MaskedGatherScatterSDNode>(Parent);
1415 X86ISelAddressMode AM;
1416 unsigned AddrSpace = Mgs->getPointerInfo().getAddrSpace();
1417 // AddrSpace 256 -> GS, 257 -> FS.
1418 if (AddrSpace == 256)
1419 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1420 if (AddrSpace == 257)
1421 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1424 Base = Mgs->getBasePtr();
1425 Index = Mgs->getIndex();
1426 unsigned ScalarSize = Mgs->getValue().getValueType().getScalarSizeInBits();
1427 Scale = getI8Imm(ScalarSize/8, DL);
1429 // If Base is 0, the whole address is in index and the Scale is 1
1430 if (isa<ConstantSDNode>(Base)) {
1431 assert(cast<ConstantSDNode>(Base)->isNullValue() &&
1432 "Unexpected base in gather/scatter");
1433 Scale = getI8Imm(1, DL);
1434 Base = CurDAG->getRegister(0, MVT::i32);
1436 if (AM.Segment.getNode())
1437 Segment = AM.Segment;
1439 Segment = CurDAG->getRegister(0, MVT::i32);
1440 Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
1444 /// Returns true if it is able to pattern match an addressing mode.
1445 /// It returns the operands which make up the maximal addressing mode it can
1446 /// match by reference.
1448 /// Parent is the parent node of the addr operand that is being matched. It
1449 /// is always a load, store, atomic node, or null. It is only null when
1450 /// checking memory operands for inline asm nodes.
1451 bool X86DAGToDAGISel::selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1452 SDValue &Scale, SDValue &Index,
1453 SDValue &Disp, SDValue &Segment) {
1454 X86ISelAddressMode AM;
1457 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1458 // that are not a MemSDNode, and thus don't have proper addrspace info.
1459 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1460 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1461 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1462 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1463 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1464 unsigned AddrSpace =
1465 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1466 // AddrSpace 256 -> GS, 257 -> FS.
1467 if (AddrSpace == 256)
1468 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1469 if (AddrSpace == 257)
1470 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1473 if (matchAddress(N, AM))
1476 MVT VT = N.getSimpleValueType();
1477 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1478 if (!AM.Base_Reg.getNode())
1479 AM.Base_Reg = CurDAG->getRegister(0, VT);
1482 if (!AM.IndexReg.getNode())
1483 AM.IndexReg = CurDAG->getRegister(0, VT);
1485 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1489 /// Match a scalar SSE load. In particular, we want to match a load whose top
1490 /// elements are either undef or zeros. The load flavor is derived from the
1491 /// type of N, which is either v4f32 or v2f64.
1494 /// PatternChainNode: this is the matched node that has a chain input and
1496 bool X86DAGToDAGISel::selectScalarSSELoad(SDNode *Root,
1497 SDValue N, SDValue &Base,
1498 SDValue &Scale, SDValue &Index,
1499 SDValue &Disp, SDValue &Segment,
1500 SDValue &PatternNodeWithChain) {
1501 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1502 PatternNodeWithChain = N.getOperand(0);
1503 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1504 PatternNodeWithChain.hasOneUse() &&
1505 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1506 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1507 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1508 if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1514 // Also handle the case where we explicitly require zeros in the top
1515 // elements. This is a vector shuffle from the zero vector.
1516 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1517 // Check to see if the top elements are all zeros (or bitcast of zeros).
1518 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1519 N.getOperand(0).getNode()->hasOneUse() &&
1520 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1521 N.getOperand(0).getOperand(0).hasOneUse() &&
1522 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1523 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1524 // Okay, this is a zero extending load. Fold it.
1525 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1526 if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1528 PatternNodeWithChain = SDValue(LD, 0);
1535 bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
1536 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1537 uint64_t ImmVal = CN->getZExtValue();
1538 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1541 Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i64);
1545 // In static codegen with small code model, we can get the address of a label
1546 // into a register with 'movl'. TableGen has already made sure we're looking
1547 // at a label of some kind.
1548 assert(N->getOpcode() == X86ISD::Wrapper &&
1549 "Unexpected node type for MOV32ri64");
1550 N = N.getOperand(0);
1552 if (N->getOpcode() != ISD::TargetConstantPool &&
1553 N->getOpcode() != ISD::TargetJumpTable &&
1554 N->getOpcode() != ISD::TargetGlobalAddress &&
1555 N->getOpcode() != ISD::TargetExternalSymbol &&
1556 N->getOpcode() != ISD::MCSymbol &&
1557 N->getOpcode() != ISD::TargetBlockAddress)
1561 return TM.getCodeModel() == CodeModel::Small;
1564 bool X86DAGToDAGISel::selectLEA64_32Addr(SDValue N, SDValue &Base,
1565 SDValue &Scale, SDValue &Index,
1566 SDValue &Disp, SDValue &Segment) {
1567 if (!selectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1571 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1572 if (RN && RN->getReg() == 0)
1573 Base = CurDAG->getRegister(0, MVT::i64);
1574 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
1575 // Base could already be %rip, particularly in the x32 ABI.
1576 Base = SDValue(CurDAG->getMachineNode(
1577 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1578 CurDAG->getTargetConstant(0, DL, MVT::i64),
1580 CurDAG->getTargetConstant(X86::sub_32bit, DL, MVT::i32)),
1584 RN = dyn_cast<RegisterSDNode>(Index);
1585 if (RN && RN->getReg() == 0)
1586 Index = CurDAG->getRegister(0, MVT::i64);
1588 assert(Index.getValueType() == MVT::i32 &&
1589 "Expect to be extending 32-bit registers for use in LEA");
1590 Index = SDValue(CurDAG->getMachineNode(
1591 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1592 CurDAG->getTargetConstant(0, DL, MVT::i64),
1594 CurDAG->getTargetConstant(X86::sub_32bit, DL,
1602 /// Calls SelectAddr and determines if the maximal addressing
1603 /// mode it matches can be cost effectively emitted as an LEA instruction.
1604 bool X86DAGToDAGISel::selectLEAAddr(SDValue N,
1605 SDValue &Base, SDValue &Scale,
1606 SDValue &Index, SDValue &Disp,
1608 X86ISelAddressMode AM;
1610 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1612 SDValue Copy = AM.Segment;
1613 SDValue T = CurDAG->getRegister(0, MVT::i32);
1615 if (matchAddress(N, AM))
1617 assert (T == AM.Segment);
1620 MVT VT = N.getSimpleValueType();
1621 unsigned Complexity = 0;
1622 if (AM.BaseType == X86ISelAddressMode::RegBase)
1623 if (AM.Base_Reg.getNode())
1626 AM.Base_Reg = CurDAG->getRegister(0, VT);
1627 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1630 if (AM.IndexReg.getNode())
1633 AM.IndexReg = CurDAG->getRegister(0, VT);
1635 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1640 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1641 // to a LEA. This is determined with some experimentation but is by no means
1642 // optimal (especially for code size consideration). LEA is nice because of
1643 // its three-address nature. Tweak the cost function again when we can run
1644 // convertToThreeAddress() at register allocation time.
1645 if (AM.hasSymbolicDisplacement()) {
1646 // For X86-64, always use LEA to materialize RIP-relative addresses.
1647 if (Subtarget->is64Bit())
1653 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1656 // If it isn't worth using an LEA, reject it.
1657 if (Complexity <= 2)
1660 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1664 /// This is only run on TargetGlobalTLSAddress nodes.
1665 bool X86DAGToDAGISel::selectTLSADDRAddr(SDValue N, SDValue &Base,
1666 SDValue &Scale, SDValue &Index,
1667 SDValue &Disp, SDValue &Segment) {
1668 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1669 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1671 X86ISelAddressMode AM;
1672 AM.GV = GA->getGlobal();
1673 AM.Disp += GA->getOffset();
1674 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1675 AM.SymbolFlags = GA->getTargetFlags();
1677 if (N.getValueType() == MVT::i32) {
1679 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1681 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1684 getAddressOperands(AM, SDLoc(N), Base, Scale, Index, Disp, Segment);
1689 bool X86DAGToDAGISel::tryFoldLoad(SDNode *P, SDValue N,
1690 SDValue &Base, SDValue &Scale,
1691 SDValue &Index, SDValue &Disp,
1693 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1694 !IsProfitableToFold(N, P, P) ||
1695 !IsLegalToFold(N, P, P, OptLevel))
1698 return selectAddr(N.getNode(),
1699 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1702 /// Return an SDNode that returns the value of the global base register.
1703 /// Output instructions required to initialize the global base register,
1705 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1706 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1707 auto &DL = MF->getDataLayout();
1708 return CurDAG->getRegister(GlobalBaseReg, TLI->getPointerTy(DL)).getNode();
1711 /// Atomic opcode table
1739 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1750 X86::LOCK_ADD64mi32,
1763 X86::LOCK_SUB64mi32,
1815 X86::LOCK_AND64mi32,
1828 X86::LOCK_XOR64mi32,
1833 // Return the target constant operand for atomic-load-op and do simple
1834 // translations, such as from atomic-load-add to lock-sub. The return value is
1835 // one of the following 3 cases:
1836 // + target-constant, the operand could be supported as a target constant.
1837 // + empty, the operand is not needed any more with the new op selected.
1838 // + non-empty, otherwise.
1839 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1841 enum AtomicOpc &Op, MVT NVT,
1843 const X86Subtarget *Subtarget) {
1844 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1845 int64_t CNVal = CN->getSExtValue();
1846 // Quit if not 32-bit imm.
1847 if ((int32_t)CNVal != CNVal)
1849 // Quit if INT32_MIN: it would be negated as it is negative and overflow,
1850 // producing an immediate that does not fit in the 32 bits available for
1851 // an immediate operand to sub. However, it still fits in 32 bits for the
1852 // add (since it is not negated) so we can return target-constant.
1853 if (CNVal == INT32_MIN)
1854 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1855 // For atomic-load-add, we could do some optimizations.
1857 // Translate to INC/DEC if ADD by 1 or -1.
1858 if (((CNVal == 1) || (CNVal == -1)) && !Subtarget->slowIncDec()) {
1859 Op = (CNVal == 1) ? INC : DEC;
1860 // No more constant operand after being translated into INC/DEC.
1863 // Translate to SUB if ADD by negative value.
1869 return CurDAG->getTargetConstant(CNVal, dl, NVT);
1872 // If the value operand is single-used, try to optimize it.
1873 if (Op == ADD && Val.hasOneUse()) {
1874 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1875 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1877 return Val.getOperand(1);
1879 // A special case for i16, which needs truncating as, in most cases, it's
1880 // promoted to i32. We will translate
1881 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1882 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1883 Val.getOperand(0).getOpcode() == ISD::SUB &&
1884 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1886 Val = Val.getOperand(0);
1887 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1895 SDNode *X86DAGToDAGISel::selectAtomicLoadArith(SDNode *Node, MVT NVT) {
1896 if (Node->hasAnyUseOfValue(0))
1901 // Optimize common patterns for __sync_or_and_fetch and similar arith
1902 // operations where the result is not used. This allows us to use the "lock"
1903 // version of the arithmetic instruction.
1904 SDValue Chain = Node->getOperand(0);
1905 SDValue Ptr = Node->getOperand(1);
1906 SDValue Val = Node->getOperand(2);
1907 SDValue Base, Scale, Index, Disp, Segment;
1908 if (!selectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
1911 // Which index into the table.
1913 switch (Node->getOpcode()) {
1916 case ISD::ATOMIC_LOAD_OR:
1919 case ISD::ATOMIC_LOAD_AND:
1922 case ISD::ATOMIC_LOAD_XOR:
1925 case ISD::ATOMIC_LOAD_ADD:
1930 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val, Subtarget);
1931 bool isUnOp = !Val.getNode();
1932 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1935 switch (NVT.SimpleTy) {
1936 default: return nullptr;
1939 Opc = AtomicOpcTbl[Op][ConstantI8];
1941 Opc = AtomicOpcTbl[Op][I8];
1945 if (immSext8(Val.getNode()))
1946 Opc = AtomicOpcTbl[Op][SextConstantI16];
1948 Opc = AtomicOpcTbl[Op][ConstantI16];
1950 Opc = AtomicOpcTbl[Op][I16];
1954 if (immSext8(Val.getNode()))
1955 Opc = AtomicOpcTbl[Op][SextConstantI32];
1957 Opc = AtomicOpcTbl[Op][ConstantI32];
1959 Opc = AtomicOpcTbl[Op][I32];
1963 if (immSext8(Val.getNode()))
1964 Opc = AtomicOpcTbl[Op][SextConstantI64];
1965 else if (i64immSExt32(Val.getNode()))
1966 Opc = AtomicOpcTbl[Op][ConstantI64];
1968 llvm_unreachable("True 64 bits constant in SelectAtomicLoadArith");
1970 Opc = AtomicOpcTbl[Op][I64];
1974 assert(Opc != 0 && "Invalid arith lock transform!");
1976 // Building the new node.
1979 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Chain };
1980 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1982 SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Val, Chain };
1983 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1986 // Copying the MachineMemOperand.
1987 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1988 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1989 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1991 // We need to have two outputs as that is what the original instruction had.
1992 // So we add a dummy, undefined output. This is safe as we checked first
1993 // that no-one uses our output anyway.
1994 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1996 SDValue RetVals[] = { Undef, Ret };
1997 return CurDAG->getMergeValues(RetVals, dl).getNode();
2000 /// Test whether the given X86ISD::CMP node has any uses which require the SF
2001 /// or OF bits to be accurate.
2002 static bool hasNoSignedComparisonUses(SDNode *N) {
2003 // Examine each user of the node.
2004 for (SDNode::use_iterator UI = N->use_begin(),
2005 UE = N->use_end(); UI != UE; ++UI) {
2006 // Only examine CopyToReg uses.
2007 if (UI->getOpcode() != ISD::CopyToReg)
2009 // Only examine CopyToReg uses that copy to EFLAGS.
2010 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
2013 // Examine each user of the CopyToReg use.
2014 for (SDNode::use_iterator FlagUI = UI->use_begin(),
2015 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
2016 // Only examine the Flag result.
2017 if (FlagUI.getUse().getResNo() != 1) continue;
2018 // Anything unusual: assume conservatively.
2019 if (!FlagUI->isMachineOpcode()) return false;
2020 // Examine the opcode of the user.
2021 switch (FlagUI->getMachineOpcode()) {
2022 // These comparisons don't treat the most significant bit specially.
2023 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
2024 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
2025 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
2026 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
2027 case X86::JA_1: case X86::JAE_1: case X86::JB_1: case X86::JBE_1:
2028 case X86::JE_1: case X86::JNE_1: case X86::JP_1: case X86::JNP_1:
2029 case X86::CMOVA16rr: case X86::CMOVA16rm:
2030 case X86::CMOVA32rr: case X86::CMOVA32rm:
2031 case X86::CMOVA64rr: case X86::CMOVA64rm:
2032 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
2033 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
2034 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
2035 case X86::CMOVB16rr: case X86::CMOVB16rm:
2036 case X86::CMOVB32rr: case X86::CMOVB32rm:
2037 case X86::CMOVB64rr: case X86::CMOVB64rm:
2038 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
2039 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
2040 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
2041 case X86::CMOVE16rr: case X86::CMOVE16rm:
2042 case X86::CMOVE32rr: case X86::CMOVE32rm:
2043 case X86::CMOVE64rr: case X86::CMOVE64rm:
2044 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
2045 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
2046 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
2047 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
2048 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
2049 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
2050 case X86::CMOVP16rr: case X86::CMOVP16rm:
2051 case X86::CMOVP32rr: case X86::CMOVP32rm:
2052 case X86::CMOVP64rr: case X86::CMOVP64rm:
2054 // Anything else: assume conservatively.
2055 default: return false;
2062 /// Check whether or not the chain ending in StoreNode is suitable for doing
2063 /// the {load; increment or decrement; store} to modify transformation.
2064 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
2065 SDValue StoredVal, SelectionDAG *CurDAG,
2066 LoadSDNode* &LoadNode, SDValue &InputChain) {
2068 // is the value stored the result of a DEC or INC?
2069 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
2071 // is the stored value result 0 of the load?
2072 if (StoredVal.getResNo() != 0) return false;
2074 // are there other uses of the loaded value than the inc or dec?
2075 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
2077 // is the store non-extending and non-indexed?
2078 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
2081 SDValue Load = StoredVal->getOperand(0);
2082 // Is the stored value a non-extending and non-indexed load?
2083 if (!ISD::isNormalLoad(Load.getNode())) return false;
2085 // Return LoadNode by reference.
2086 LoadNode = cast<LoadSDNode>(Load);
2087 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
2088 EVT LdVT = LoadNode->getMemoryVT();
2089 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
2093 // Is store the only read of the loaded value?
2094 if (!Load.hasOneUse())
2097 // Is the address of the store the same as the load?
2098 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
2099 LoadNode->getOffset() != StoreNode->getOffset())
2102 // Check if the chain is produced by the load or is a TokenFactor with
2103 // the load output chain as an operand. Return InputChain by reference.
2104 SDValue Chain = StoreNode->getChain();
2106 bool ChainCheck = false;
2107 if (Chain == Load.getValue(1)) {
2109 InputChain = LoadNode->getChain();
2110 } else if (Chain.getOpcode() == ISD::TokenFactor) {
2111 SmallVector<SDValue, 4> ChainOps;
2112 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
2113 SDValue Op = Chain.getOperand(i);
2114 if (Op == Load.getValue(1)) {
2119 // Make sure using Op as part of the chain would not cause a cycle here.
2120 // In theory, we could check whether the chain node is a predecessor of
2121 // the load. But that can be very expensive. Instead visit the uses and
2122 // make sure they all have smaller node id than the load.
2123 int LoadId = LoadNode->getNodeId();
2124 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
2125 UE = UI->use_end(); UI != UE; ++UI) {
2126 if (UI.getUse().getResNo() != 0)
2128 if (UI->getNodeId() > LoadId)
2132 ChainOps.push_back(Op);
2136 // Make a new TokenFactor with all the other input chains except
2138 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
2139 MVT::Other, ChainOps);
2147 /// Get the appropriate X86 opcode for an in-memory increment or decrement.
2148 /// Opc should be X86ISD::DEC or X86ISD::INC.
2149 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2150 if (Opc == X86ISD::DEC) {
2151 if (LdVT == MVT::i64) return X86::DEC64m;
2152 if (LdVT == MVT::i32) return X86::DEC32m;
2153 if (LdVT == MVT::i16) return X86::DEC16m;
2154 if (LdVT == MVT::i8) return X86::DEC8m;
2156 assert(Opc == X86ISD::INC && "unrecognized opcode");
2157 if (LdVT == MVT::i64) return X86::INC64m;
2158 if (LdVT == MVT::i32) return X86::INC32m;
2159 if (LdVT == MVT::i16) return X86::INC16m;
2160 if (LdVT == MVT::i8) return X86::INC8m;
2162 llvm_unreachable("unrecognized size for LdVT");
2165 /// Customized ISel for GATHER operations.
2166 SDNode *X86DAGToDAGISel::selectGather(SDNode *Node, unsigned Opc) {
2167 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2168 SDValue Chain = Node->getOperand(0);
2169 SDValue VSrc = Node->getOperand(2);
2170 SDValue Base = Node->getOperand(3);
2171 SDValue VIdx = Node->getOperand(4);
2172 SDValue VMask = Node->getOperand(5);
2173 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2177 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2182 // Memory Operands: Base, Scale, Index, Disp, Segment
2183 SDValue Disp = CurDAG->getTargetConstant(0, DL, MVT::i32);
2184 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2185 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue(), DL), VIdx,
2186 Disp, Segment, VMask, Chain};
2187 SDNode *ResNode = CurDAG->getMachineNode(Opc, DL, VTs, Ops);
2188 // Node has 2 outputs: VDst and MVT::Other.
2189 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2190 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2192 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2193 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2197 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2198 MVT NVT = Node->getSimpleValueType(0);
2200 unsigned Opcode = Node->getOpcode();
2203 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2205 if (Node->isMachineOpcode()) {
2206 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2207 Node->setNodeId(-1);
2208 return nullptr; // Already selected.
2214 if (Subtarget->isTargetNaCl())
2215 // NaCl has its own pass where jmp %r32 are converted to jmp %r64. We
2216 // leave the instruction alone.
2218 if (Subtarget->isTarget64BitILP32()) {
2219 // Converts a 32-bit register to a 64-bit, zero-extended version of
2220 // it. This is needed because x86-64 can do many things, but jmp %r32
2221 // ain't one of them.
2222 const SDValue &Target = Node->getOperand(1);
2223 assert(Target.getSimpleValueType() == llvm::MVT::i32);
2224 SDValue ZextTarget = CurDAG->getZExtOrTrunc(Target, dl, EVT(MVT::i64));
2225 SDValue Brind = CurDAG->getNode(ISD::BRIND, dl, MVT::Other,
2226 Node->getOperand(0), ZextTarget);
2227 ReplaceUses(SDValue(Node, 0), Brind);
2228 SelectCode(ZextTarget.getNode());
2229 SelectCode(Brind.getNode());
2234 case ISD::INTRINSIC_W_CHAIN: {
2235 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2238 case Intrinsic::x86_avx2_gather_d_pd:
2239 case Intrinsic::x86_avx2_gather_d_pd_256:
2240 case Intrinsic::x86_avx2_gather_q_pd:
2241 case Intrinsic::x86_avx2_gather_q_pd_256:
2242 case Intrinsic::x86_avx2_gather_d_ps:
2243 case Intrinsic::x86_avx2_gather_d_ps_256:
2244 case Intrinsic::x86_avx2_gather_q_ps:
2245 case Intrinsic::x86_avx2_gather_q_ps_256:
2246 case Intrinsic::x86_avx2_gather_d_q:
2247 case Intrinsic::x86_avx2_gather_d_q_256:
2248 case Intrinsic::x86_avx2_gather_q_q:
2249 case Intrinsic::x86_avx2_gather_q_q_256:
2250 case Intrinsic::x86_avx2_gather_d_d:
2251 case Intrinsic::x86_avx2_gather_d_d_256:
2252 case Intrinsic::x86_avx2_gather_q_d:
2253 case Intrinsic::x86_avx2_gather_q_d_256: {
2254 if (!Subtarget->hasAVX2())
2258 default: llvm_unreachable("Impossible intrinsic");
2259 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2260 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2261 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2262 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2263 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2264 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2265 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2266 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2267 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2268 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2269 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2270 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2271 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2272 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2273 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2274 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2276 SDNode *RetVal = selectGather(Node, Opc);
2278 // We already called ReplaceUses inside SelectGather.
2285 case X86ISD::GlobalBaseReg:
2286 return getGlobalBaseReg();
2288 case X86ISD::SHRUNKBLEND: {
2289 // SHRUNKBLEND selects like a regular VSELECT.
2290 SDValue VSelect = CurDAG->getNode(
2291 ISD::VSELECT, SDLoc(Node), Node->getValueType(0), Node->getOperand(0),
2292 Node->getOperand(1), Node->getOperand(2));
2293 ReplaceUses(SDValue(Node, 0), VSelect);
2294 SelectCode(VSelect.getNode());
2295 // We already called ReplaceUses.
2299 case ISD::ATOMIC_LOAD_XOR:
2300 case ISD::ATOMIC_LOAD_AND:
2301 case ISD::ATOMIC_LOAD_OR:
2302 case ISD::ATOMIC_LOAD_ADD: {
2303 SDNode *RetVal = selectAtomicLoadArith(Node, NVT);
2311 // For operations of the form (x << C1) op C2, check if we can use a smaller
2312 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2313 SDValue N0 = Node->getOperand(0);
2314 SDValue N1 = Node->getOperand(1);
2316 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2319 // i8 is unshrinkable, i16 should be promoted to i32.
2320 if (NVT != MVT::i32 && NVT != MVT::i64)
2323 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2324 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2325 if (!Cst || !ShlCst)
2328 int64_t Val = Cst->getSExtValue();
2329 uint64_t ShlVal = ShlCst->getZExtValue();
2331 // Make sure that we don't change the operation by removing bits.
2332 // This only matters for OR and XOR, AND is unaffected.
2333 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2334 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2337 unsigned ShlOp, AddOp, Op;
2340 // Check the minimum bitwidth for the new constant.
2341 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2342 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2343 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2344 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2346 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2349 // Bail if there is no smaller encoding.
2353 switch (NVT.SimpleTy) {
2354 default: llvm_unreachable("Unsupported VT!");
2356 assert(CstVT == MVT::i8);
2357 ShlOp = X86::SHL32ri;
2358 AddOp = X86::ADD32rr;
2361 default: llvm_unreachable("Impossible opcode");
2362 case ISD::AND: Op = X86::AND32ri8; break;
2363 case ISD::OR: Op = X86::OR32ri8; break;
2364 case ISD::XOR: Op = X86::XOR32ri8; break;
2368 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2369 ShlOp = X86::SHL64ri;
2370 AddOp = X86::ADD64rr;
2373 default: llvm_unreachable("Impossible opcode");
2374 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2375 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2376 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2381 // Emit the smaller op and the shift.
2382 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, dl, CstVT);
2383 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2385 return CurDAG->SelectNodeTo(Node, AddOp, NVT, SDValue(New, 0),
2387 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2388 getI8Imm(ShlVal, dl));
2391 case X86ISD::SMUL8: {
2392 SDValue N0 = Node->getOperand(0);
2393 SDValue N1 = Node->getOperand(1);
2395 Opc = (Opcode == X86ISD::SMUL8 ? X86::IMUL8r : X86::MUL8r);
2397 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::AL,
2398 N0, SDValue()).getValue(1);
2400 SDVTList VTs = CurDAG->getVTList(NVT, MVT::i32);
2401 SDValue Ops[] = {N1, InFlag};
2402 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2404 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2405 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2409 case X86ISD::UMUL: {
2410 SDValue N0 = Node->getOperand(0);
2411 SDValue N1 = Node->getOperand(1);
2414 switch (NVT.SimpleTy) {
2415 default: llvm_unreachable("Unsupported VT!");
2416 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2417 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2418 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2419 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2422 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2423 N0, SDValue()).getValue(1);
2425 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2426 SDValue Ops[] = {N1, InFlag};
2427 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2429 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2430 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2431 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2435 case ISD::SMUL_LOHI:
2436 case ISD::UMUL_LOHI: {
2437 SDValue N0 = Node->getOperand(0);
2438 SDValue N1 = Node->getOperand(1);
2440 bool isSigned = Opcode == ISD::SMUL_LOHI;
2441 bool hasBMI2 = Subtarget->hasBMI2();
2443 switch (NVT.SimpleTy) {
2444 default: llvm_unreachable("Unsupported VT!");
2445 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2446 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2447 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2448 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2449 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2450 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2453 switch (NVT.SimpleTy) {
2454 default: llvm_unreachable("Unsupported VT!");
2455 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2456 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2457 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2458 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2462 unsigned SrcReg, LoReg, HiReg;
2464 default: llvm_unreachable("Unknown MUL opcode!");
2467 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2471 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2475 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2479 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2482 SrcReg = X86::EDX; LoReg = HiReg = 0;
2485 SrcReg = X86::RDX; LoReg = HiReg = 0;
2489 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2490 bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2491 // Multiply is commmutative.
2493 foldedLoad = tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2498 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2499 N0, SDValue()).getValue(1);
2500 SDValue ResHi, ResLo;
2504 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2506 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2507 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2508 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2509 ResHi = SDValue(CNode, 0);
2510 ResLo = SDValue(CNode, 1);
2511 Chain = SDValue(CNode, 2);
2512 InFlag = SDValue(CNode, 3);
2514 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2515 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2516 Chain = SDValue(CNode, 0);
2517 InFlag = SDValue(CNode, 1);
2520 // Update the chain.
2521 ReplaceUses(N1.getValue(1), Chain);
2523 SDValue Ops[] = { N1, InFlag };
2524 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2525 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2526 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2527 ResHi = SDValue(CNode, 0);
2528 ResLo = SDValue(CNode, 1);
2529 InFlag = SDValue(CNode, 2);
2531 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2532 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2533 InFlag = SDValue(CNode, 0);
2537 // Prevent use of AH in a REX instruction by referencing AX instead.
2538 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2539 !SDValue(Node, 1).use_empty()) {
2540 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2541 X86::AX, MVT::i16, InFlag);
2542 InFlag = Result.getValue(2);
2543 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2545 if (!SDValue(Node, 0).use_empty())
2546 ReplaceUses(SDValue(Node, 1),
2547 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2549 // Shift AX down 8 bits.
2550 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2552 CurDAG->getTargetConstant(8, dl, MVT::i8)),
2554 // Then truncate it down to i8.
2555 ReplaceUses(SDValue(Node, 1),
2556 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2558 // Copy the low half of the result, if it is needed.
2559 if (!SDValue(Node, 0).use_empty()) {
2560 if (!ResLo.getNode()) {
2561 assert(LoReg && "Register for low half is not defined!");
2562 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2564 InFlag = ResLo.getValue(2);
2566 ReplaceUses(SDValue(Node, 0), ResLo);
2567 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2569 // Copy the high half of the result, if it is needed.
2570 if (!SDValue(Node, 1).use_empty()) {
2571 if (!ResHi.getNode()) {
2572 assert(HiReg && "Register for high half is not defined!");
2573 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2575 InFlag = ResHi.getValue(2);
2577 ReplaceUses(SDValue(Node, 1), ResHi);
2578 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2586 case X86ISD::SDIVREM8_SEXT_HREG:
2587 case X86ISD::UDIVREM8_ZEXT_HREG: {
2588 SDValue N0 = Node->getOperand(0);
2589 SDValue N1 = Node->getOperand(1);
2591 bool isSigned = (Opcode == ISD::SDIVREM ||
2592 Opcode == X86ISD::SDIVREM8_SEXT_HREG);
2594 switch (NVT.SimpleTy) {
2595 default: llvm_unreachable("Unsupported VT!");
2596 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2597 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2598 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2599 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2602 switch (NVT.SimpleTy) {
2603 default: llvm_unreachable("Unsupported VT!");
2604 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2605 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2606 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2607 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2611 unsigned LoReg, HiReg, ClrReg;
2612 unsigned SExtOpcode;
2613 switch (NVT.SimpleTy) {
2614 default: llvm_unreachable("Unsupported VT!");
2616 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2617 SExtOpcode = X86::CBW;
2620 LoReg = X86::AX; HiReg = X86::DX;
2622 SExtOpcode = X86::CWD;
2625 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2626 SExtOpcode = X86::CDQ;
2629 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2630 SExtOpcode = X86::CQO;
2634 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2635 bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2636 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2639 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2640 // Special case for div8, just use a move with zero extension to AX to
2641 // clear the upper 8 bits (AH).
2642 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2643 if (tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2644 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2646 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2647 MVT::Other, Ops), 0);
2648 Chain = Move.getValue(1);
2649 ReplaceUses(N0.getValue(1), Chain);
2652 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2653 Chain = CurDAG->getEntryNode();
2655 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2656 InFlag = Chain.getValue(1);
2659 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2660 LoReg, N0, SDValue()).getValue(1);
2661 if (isSigned && !signBitIsZero) {
2662 // Sign extend the low part into the high part.
2664 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2666 // Zero out the high part, effectively zero extending the input.
2667 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2668 switch (NVT.SimpleTy) {
2671 SDValue(CurDAG->getMachineNode(
2672 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2673 CurDAG->getTargetConstant(X86::sub_16bit, dl,
2681 SDValue(CurDAG->getMachineNode(
2682 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2683 CurDAG->getTargetConstant(0, dl, MVT::i64), ClrNode,
2684 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2689 llvm_unreachable("Unexpected division source");
2692 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2693 ClrNode, InFlag).getValue(1);
2698 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2701 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2702 InFlag = SDValue(CNode, 1);
2703 // Update the chain.
2704 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2707 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2710 // Prevent use of AH in a REX instruction by explicitly copying it to
2711 // an ABCD_L register.
2713 // The current assumption of the register allocator is that isel
2714 // won't generate explicit references to the GR8_ABCD_H registers. If
2715 // the allocator and/or the backend get enhanced to be more robust in
2716 // that regard, this can be, and should be, removed.
2717 if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
2718 SDValue AHCopy = CurDAG->getRegister(X86::AH, MVT::i8);
2719 unsigned AHExtOpcode =
2720 isSigned ? X86::MOVSX32_NOREXrr8 : X86::MOVZX32_NOREXrr8;
2722 SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
2723 MVT::Glue, AHCopy, InFlag);
2724 SDValue Result(RNode, 0);
2725 InFlag = SDValue(RNode, 1);
2727 if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
2728 Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
2729 if (Node->getValueType(1) == MVT::i64) {
2730 // It's not possible to directly movsx AH to a 64bit register, because
2731 // the latter needs the REX prefix, but the former can't have it.
2732 assert(Opcode != X86ISD::SDIVREM8_SEXT_HREG &&
2733 "Unexpected i64 sext of h-register");
2735 SDValue(CurDAG->getMachineNode(
2736 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2737 CurDAG->getTargetConstant(0, dl, MVT::i64), Result,
2738 CurDAG->getTargetConstant(X86::sub_32bit, dl,
2744 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
2746 ReplaceUses(SDValue(Node, 1), Result);
2747 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2749 // Copy the division (low) result, if it is needed.
2750 if (!SDValue(Node, 0).use_empty()) {
2751 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2752 LoReg, NVT, InFlag);
2753 InFlag = Result.getValue(2);
2754 ReplaceUses(SDValue(Node, 0), Result);
2755 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2757 // Copy the remainder (high) result, if it is needed.
2758 if (!SDValue(Node, 1).use_empty()) {
2759 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2760 HiReg, NVT, InFlag);
2761 InFlag = Result.getValue(2);
2762 ReplaceUses(SDValue(Node, 1), Result);
2763 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2770 // Sometimes a SUB is used to perform comparison.
2771 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2772 // This node is not a CMP.
2774 SDValue N0 = Node->getOperand(0);
2775 SDValue N1 = Node->getOperand(1);
2777 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2778 hasNoSignedComparisonUses(Node))
2779 N0 = N0.getOperand(0);
2781 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2782 // use a smaller encoding.
2783 // Look past the truncate if CMP is the only use of it.
2784 if ((N0.getNode()->getOpcode() == ISD::AND ||
2785 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2786 N0.getNode()->hasOneUse() &&
2787 N0.getValueType() != MVT::i8 &&
2788 X86::isZeroNode(N1)) {
2789 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2792 // For example, convert "testl %eax, $8" to "testb %al, $8"
2793 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2794 (!(C->getZExtValue() & 0x80) ||
2795 hasNoSignedComparisonUses(Node))) {
2796 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl, MVT::i8);
2797 SDValue Reg = N0.getNode()->getOperand(0);
2799 // On x86-32, only the ABCD registers have 8-bit subregisters.
2800 if (!Subtarget->is64Bit()) {
2801 const TargetRegisterClass *TRC;
2802 switch (N0.getSimpleValueType().SimpleTy) {
2803 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2804 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2805 default: llvm_unreachable("Unsupported TEST operand type!");
2807 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2808 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2809 Reg.getValueType(), Reg, RC), 0);
2812 // Extract the l-register.
2813 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2817 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2819 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2820 // one, do not call ReplaceAllUsesWith.
2821 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2822 SDValue(NewNode, 0));
2826 // For example, "testl %eax, $2048" to "testb %ah, $8".
2827 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2828 (!(C->getZExtValue() & 0x8000) ||
2829 hasNoSignedComparisonUses(Node))) {
2830 // Shift the immediate right by 8 bits.
2831 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2833 SDValue Reg = N0.getNode()->getOperand(0);
2835 // Put the value in an ABCD register.
2836 const TargetRegisterClass *TRC;
2837 switch (N0.getSimpleValueType().SimpleTy) {
2838 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2839 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2840 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2841 default: llvm_unreachable("Unsupported TEST operand type!");
2843 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
2844 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2845 Reg.getValueType(), Reg, RC), 0);
2847 // Extract the h-register.
2848 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2851 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2852 // target GR8_NOREX registers, so make sure the register class is
2854 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2855 MVT::i32, Subreg, ShiftedImm);
2856 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2857 // one, do not call ReplaceAllUsesWith.
2858 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2859 SDValue(NewNode, 0));
2863 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2864 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2865 N0.getValueType() != MVT::i16 &&
2866 (!(C->getZExtValue() & 0x8000) ||
2867 hasNoSignedComparisonUses(Node))) {
2868 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2870 SDValue Reg = N0.getNode()->getOperand(0);
2872 // Extract the 16-bit subregister.
2873 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2877 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2879 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2880 // one, do not call ReplaceAllUsesWith.
2881 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2882 SDValue(NewNode, 0));
2886 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2887 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2888 N0.getValueType() == MVT::i64 &&
2889 (!(C->getZExtValue() & 0x80000000) ||
2890 hasNoSignedComparisonUses(Node))) {
2891 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
2893 SDValue Reg = N0.getNode()->getOperand(0);
2895 // Extract the 32-bit subregister.
2896 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2900 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2902 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2903 // one, do not call ReplaceAllUsesWith.
2904 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2905 SDValue(NewNode, 0));
2912 // Change a chain of {load; incr or dec; store} of the same value into
2913 // a simple increment or decrement through memory of that value, if the
2914 // uses of the modified value and its address are suitable.
2915 // The DEC64m tablegen pattern is currently not able to match the case where
2916 // the EFLAGS on the original DEC are used. (This also applies to
2917 // {INC,DEC}X{64,32,16,8}.)
2918 // We'll need to improve tablegen to allow flags to be transferred from a
2919 // node in the pattern to the result node. probably with a new keyword
2920 // for example, we have this
2921 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2922 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2923 // (implicit EFLAGS)]>;
2924 // but maybe need something like this
2925 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2926 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2927 // (transferrable EFLAGS)]>;
2929 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2930 SDValue StoredVal = StoreNode->getOperand(1);
2931 unsigned Opc = StoredVal->getOpcode();
2933 LoadSDNode *LoadNode = nullptr;
2935 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2936 LoadNode, InputChain))
2939 SDValue Base, Scale, Index, Disp, Segment;
2940 if (!selectAddr(LoadNode, LoadNode->getBasePtr(),
2941 Base, Scale, Index, Disp, Segment))
2944 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2945 MemOp[0] = StoreNode->getMemOperand();
2946 MemOp[1] = LoadNode->getMemOperand();
2947 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2948 EVT LdVT = LoadNode->getMemoryVT();
2949 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2950 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2952 MVT::i32, MVT::Other, Ops);
2953 Result->setMemRefs(MemOp, MemOp + 2);
2955 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2956 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2962 SDNode *ResNode = SelectCode(Node);
2964 DEBUG(dbgs() << "=> ";
2965 if (ResNode == nullptr || ResNode == Node)
2968 ResNode->dump(CurDAG);
2974 bool X86DAGToDAGISel::
2975 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
2976 std::vector<SDValue> &OutOps) {
2977 SDValue Op0, Op1, Op2, Op3, Op4;
2978 switch (ConstraintID) {
2980 llvm_unreachable("Unexpected asm memory constraint");
2981 case InlineAsm::Constraint_i:
2982 // FIXME: It seems strange that 'i' is needed here since it's supposed to
2983 // be an immediate and not a memory constraint.
2985 case InlineAsm::Constraint_o: // offsetable ??
2986 case InlineAsm::Constraint_v: // not offsetable ??
2987 case InlineAsm::Constraint_m: // memory
2988 case InlineAsm::Constraint_X:
2989 if (!selectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
2994 OutOps.push_back(Op0);
2995 OutOps.push_back(Op1);
2996 OutOps.push_back(Op2);
2997 OutOps.push_back(Op3);
2998 OutOps.push_back(Op4);
3002 /// This pass converts a legalized DAG into a X86-specific DAG,
3003 /// ready for instruction scheduling.
3004 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
3005 CodeGenOpt::Level OptLevel) {
3006 return new X86DAGToDAGISel(TM, OptLevel);