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/Instructions.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define DEBUG_TYPE "x86-isel"
40 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
42 //===----------------------------------------------------------------------===//
43 // Pattern Matcher Implementation
44 //===----------------------------------------------------------------------===//
47 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
48 /// SDValue's instead of register numbers for the leaves of the matched
50 struct X86ISelAddressMode {
56 // This is really a union, discriminated by BaseType!
64 const GlobalValue *GV;
66 const BlockAddress *BlockAddr;
69 unsigned Align; // CP alignment.
70 unsigned char SymbolFlags; // X86II::MO_*
73 : BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
74 Segment(), GV(nullptr), CP(nullptr), BlockAddr(nullptr), ES(nullptr),
75 JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {
78 bool hasSymbolicDisplacement() const {
79 return GV != nullptr || CP != nullptr || ES != nullptr ||
80 JT != -1 || BlockAddr != nullptr;
83 bool hasBaseOrIndexReg() const {
84 return BaseType == FrameIndexBase ||
85 IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
88 /// isRIPRelative - Return true if this addressing mode is already RIP
90 bool isRIPRelative() const {
91 if (BaseType != RegBase) return false;
92 if (RegisterSDNode *RegNode =
93 dyn_cast_or_null<RegisterSDNode>(Base_Reg.getNode()))
94 return RegNode->getReg() == X86::RIP;
98 void setBaseReg(SDValue Reg) {
103 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
105 dbgs() << "X86ISelAddressMode " << this << '\n';
106 dbgs() << "Base_Reg ";
107 if (Base_Reg.getNode() != 0)
108 Base_Reg.getNode()->dump();
111 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
112 << " Scale" << Scale << '\n'
114 if (IndexReg.getNode() != 0)
115 IndexReg.getNode()->dump();
118 dbgs() << " Disp " << Disp << '\n'
135 dbgs() << " JT" << JT << " Align" << Align << '\n';
142 //===--------------------------------------------------------------------===//
143 /// ISel - X86 specific code to select X86 machine instructions for
144 /// SelectionDAG operations.
146 class X86DAGToDAGISel final : public SelectionDAGISel {
147 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
148 /// make the right decision when generating code for different targets.
149 const X86Subtarget *Subtarget;
151 /// OptForSize - If true, selector should try to optimize for code size
152 /// instead of performance.
156 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
157 : SelectionDAGISel(tm, OptLevel),
158 Subtarget(&tm.getSubtarget<X86Subtarget>()),
161 const char *getPassName() const override {
162 return "X86 DAG->DAG Instruction Selection";
165 void EmitFunctionEntryCode() override;
167 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
169 void PreprocessISelDAG() override;
171 inline bool immSext8(SDNode *N) const {
172 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
175 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
176 // sign extended field.
177 inline bool i64immSExt32(SDNode *N) const {
178 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
179 return (int64_t)v == (int32_t)v;
182 // Include the pieces autogenerated from the target description.
183 #include "X86GenDAGISel.inc"
186 SDNode *Select(SDNode *N) override;
187 SDNode *SelectGather(SDNode *N, unsigned Opc);
188 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
189 SDNode *SelectAtomicLoadArith(SDNode *Node, MVT NVT);
191 bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
192 bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
193 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
194 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
195 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
197 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
198 bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
199 SDValue &Scale, SDValue &Index, SDValue &Disp,
201 bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
202 bool SelectLEAAddr(SDValue N, SDValue &Base,
203 SDValue &Scale, SDValue &Index, SDValue &Disp,
205 bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
206 SDValue &Scale, SDValue &Index, SDValue &Disp,
208 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
209 SDValue &Scale, SDValue &Index, SDValue &Disp,
211 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
212 SDValue &Base, SDValue &Scale,
213 SDValue &Index, SDValue &Disp,
215 SDValue &NodeWithChain);
217 bool TryFoldLoad(SDNode *P, SDValue N,
218 SDValue &Base, SDValue &Scale,
219 SDValue &Index, SDValue &Disp,
222 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
223 /// inline asm expressions.
224 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
226 std::vector<SDValue> &OutOps) override;
228 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
230 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
231 SDValue &Scale, SDValue &Index,
232 SDValue &Disp, SDValue &Segment) {
233 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
234 CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
235 getTargetLowering()->getPointerTy()) :
237 Scale = getI8Imm(AM.Scale);
239 // These are 32-bit even in 64-bit mode since RIP relative offset
242 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
246 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
247 AM.Align, AM.Disp, AM.SymbolFlags);
249 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
250 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
251 } else if (AM.JT != -1) {
252 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
253 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
254 } else if (AM.BlockAddr)
255 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
258 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
260 if (AM.Segment.getNode())
261 Segment = AM.Segment;
263 Segment = CurDAG->getRegister(0, MVT::i32);
266 /// getI8Imm - Return a target constant with the specified value, of type
268 inline SDValue getI8Imm(unsigned Imm) {
269 return CurDAG->getTargetConstant(Imm, MVT::i8);
272 /// getI32Imm - Return a target constant with the specified value, of type
274 inline SDValue getI32Imm(unsigned Imm) {
275 return CurDAG->getTargetConstant(Imm, MVT::i32);
278 /// getGlobalBaseReg - Return an SDNode that returns the value of
279 /// the global base register. Output instructions required to
280 /// initialize the global base register, if necessary.
282 SDNode *getGlobalBaseReg();
284 /// getTargetMachine - Return a reference to the TargetMachine, casted
285 /// to the target-specific type.
286 const X86TargetMachine &getTargetMachine() const {
287 return static_cast<const X86TargetMachine &>(TM);
290 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
291 /// to the target-specific type.
292 const X86InstrInfo *getInstrInfo() const {
293 return getTargetMachine().getInstrInfo();
300 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
301 if (OptLevel == CodeGenOpt::None) return false;
306 if (N.getOpcode() != ISD::LOAD)
309 // If N is a load, do additional profitability checks.
311 switch (U->getOpcode()) {
324 SDValue Op1 = U->getOperand(1);
326 // If the other operand is a 8-bit immediate we should fold the immediate
327 // instead. This reduces code size.
329 // movl 4(%esp), %eax
333 // addl 4(%esp), %eax
334 // The former is 2 bytes shorter. In case where the increment is 1, then
335 // the saving can be 4 bytes (by using incl %eax).
336 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
337 if (Imm->getAPIntValue().isSignedIntN(8))
340 // If the other operand is a TLS address, we should fold it instead.
343 // leal i@NTPOFF(%eax), %eax
345 // movl $i@NTPOFF, %eax
347 // if the block also has an access to a second TLS address this will save
349 // FIXME: This is probably also true for non-TLS addresses.
350 if (Op1.getOpcode() == X86ISD::Wrapper) {
351 SDValue Val = Op1.getOperand(0);
352 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
362 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
363 /// load's chain operand and move load below the call's chain operand.
364 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
365 SDValue Call, SDValue OrigChain) {
366 SmallVector<SDValue, 8> Ops;
367 SDValue Chain = OrigChain.getOperand(0);
368 if (Chain.getNode() == Load.getNode())
369 Ops.push_back(Load.getOperand(0));
371 assert(Chain.getOpcode() == ISD::TokenFactor &&
372 "Unexpected chain operand");
373 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
374 if (Chain.getOperand(i).getNode() == Load.getNode())
375 Ops.push_back(Load.getOperand(0));
377 Ops.push_back(Chain.getOperand(i));
379 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load),
380 MVT::Other, &Ops[0], Ops.size());
382 Ops.push_back(NewChain);
384 for (unsigned i = 1, e = OrigChain.getNumOperands(); i != e; ++i)
385 Ops.push_back(OrigChain.getOperand(i));
386 CurDAG->UpdateNodeOperands(OrigChain.getNode(), &Ops[0], Ops.size());
387 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
388 Load.getOperand(1), Load.getOperand(2));
390 unsigned NumOps = Call.getNode()->getNumOperands();
392 Ops.push_back(SDValue(Load.getNode(), 1));
393 for (unsigned i = 1, e = NumOps; i != e; ++i)
394 Ops.push_back(Call.getOperand(i));
395 CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], NumOps);
398 /// isCalleeLoad - Return true if call address is a load and it can be
399 /// moved below CALLSEQ_START and the chains leading up to the call.
400 /// Return the CALLSEQ_START by reference as a second output.
401 /// In the case of a tail call, there isn't a callseq node between the call
402 /// chain and the load.
403 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
404 // The transformation is somewhat dangerous if the call's chain was glued to
405 // the call. After MoveBelowOrigChain the load is moved between the call and
406 // the chain, this can create a cycle if the load is not folded. So it is
407 // *really* important that we are sure the load will be folded.
408 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
410 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
413 LD->getAddressingMode() != ISD::UNINDEXED ||
414 LD->getExtensionType() != ISD::NON_EXTLOAD)
417 // Now let's find the callseq_start.
418 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
419 if (!Chain.hasOneUse())
421 Chain = Chain.getOperand(0);
424 if (!Chain.getNumOperands())
426 // Since we are not checking for AA here, conservatively abort if the chain
427 // writes to memory. It's not safe to move the callee (a load) across a store.
428 if (isa<MemSDNode>(Chain.getNode()) &&
429 cast<MemSDNode>(Chain.getNode())->writeMem())
431 if (Chain.getOperand(0).getNode() == Callee.getNode())
433 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
434 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
435 Callee.getValue(1).hasOneUse())
440 void X86DAGToDAGISel::PreprocessISelDAG() {
441 // OptForSize is used in pattern predicates that isel is matching.
442 OptForSize = MF->getFunction()->getAttributes().
443 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
445 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
446 E = CurDAG->allnodes_end(); I != E; ) {
447 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
449 if (OptLevel != CodeGenOpt::None &&
450 // Only does this when target favors doesn't favor register indirect
452 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
453 (N->getOpcode() == X86ISD::TC_RETURN &&
454 // Only does this if load can be folded into TC_RETURN.
455 (Subtarget->is64Bit() ||
456 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
457 /// Also try moving call address load from outside callseq_start to just
458 /// before the call to allow it to be folded.
476 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
477 SDValue Chain = N->getOperand(0);
478 SDValue Load = N->getOperand(1);
479 if (!isCalleeLoad(Load, Chain, HasCallSeq))
481 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
486 // Lower fpround and fpextend nodes that target the FP stack to be store and
487 // load to the stack. This is a gross hack. We would like to simply mark
488 // these as being illegal, but when we do that, legalize produces these when
489 // it expands calls, then expands these in the same legalize pass. We would
490 // like dag combine to be able to hack on these between the call expansion
491 // and the node legalization. As such this pass basically does "really
492 // late" legalization of these inline with the X86 isel pass.
493 // FIXME: This should only happen when not compiled with -O0.
494 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
497 MVT SrcVT = N->getOperand(0).getSimpleValueType();
498 MVT DstVT = N->getSimpleValueType(0);
500 // If any of the sources are vectors, no fp stack involved.
501 if (SrcVT.isVector() || DstVT.isVector())
504 // If the source and destination are SSE registers, then this is a legal
505 // conversion that should not be lowered.
506 const X86TargetLowering *X86Lowering =
507 static_cast<const X86TargetLowering *>(getTargetLowering());
508 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
509 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
510 if (SrcIsSSE && DstIsSSE)
513 if (!SrcIsSSE && !DstIsSSE) {
514 // If this is an FPStack extension, it is a noop.
515 if (N->getOpcode() == ISD::FP_EXTEND)
517 // If this is a value-preserving FPStack truncation, it is a noop.
518 if (N->getConstantOperandVal(1))
522 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
523 // FPStack has extload and truncstore. SSE can fold direct loads into other
524 // operations. Based on this, decide what we want to do.
526 if (N->getOpcode() == ISD::FP_ROUND)
527 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
529 MemVT = SrcIsSSE ? SrcVT : DstVT;
531 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
534 // FIXME: optimize the case where the src/dest is a load or store?
535 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
537 MemTmp, MachinePointerInfo(), MemVT,
539 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
540 MachinePointerInfo(),
541 MemVT, false, false, 0);
543 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
544 // extload we created. This will cause general havok on the dag because
545 // anything below the conversion could be folded into other existing nodes.
546 // To avoid invalidating 'I', back it up to the convert node.
548 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
550 // Now that we did that, the node is dead. Increment the iterator to the
551 // next node to process, then delete N.
553 CurDAG->DeleteNode(N);
558 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
559 /// the main function.
560 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
561 MachineFrameInfo *MFI) {
562 const TargetInstrInfo *TII = TM.getInstrInfo();
563 if (Subtarget->isTargetCygMing()) {
565 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
566 BuildMI(BB, DebugLoc(),
567 TII->get(CallOp)).addExternalSymbol("__main");
571 void X86DAGToDAGISel::EmitFunctionEntryCode() {
572 // If this is main, emit special code for main.
573 if (const Function *Fn = MF->getFunction())
574 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
575 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
578 static bool isDispSafeForFrameIndex(int64_t Val) {
579 // On 64-bit platforms, we can run into an issue where a frame index
580 // includes a displacement that, when added to the explicit displacement,
581 // will overflow the displacement field. Assuming that the frame index
582 // displacement fits into a 31-bit integer (which is only slightly more
583 // aggressive than the current fundamental assumption that it fits into
584 // a 32-bit integer), a 31-bit disp should always be safe.
585 return isInt<31>(Val);
588 bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
589 X86ISelAddressMode &AM) {
590 int64_t Val = AM.Disp + Offset;
591 CodeModel::Model M = TM.getCodeModel();
592 if (Subtarget->is64Bit()) {
593 if (!X86::isOffsetSuitableForCodeModel(Val, M,
594 AM.hasSymbolicDisplacement()))
596 // In addition to the checks required for a register base, check that
597 // we do not try to use an unsafe Disp with a frame index.
598 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
599 !isDispSafeForFrameIndex(Val))
607 bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
608 SDValue Address = N->getOperand(1);
610 // load gs:0 -> GS segment register.
611 // load fs:0 -> FS segment register.
613 // This optimization is valid because the GNU TLS model defines that
614 // gs:0 (or fs:0 on X86-64) contains its own address.
615 // For more information see http://people.redhat.com/drepper/tls.pdf
616 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
617 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
618 Subtarget->isTargetLinux())
619 switch (N->getPointerInfo().getAddrSpace()) {
621 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
624 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
631 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
632 /// into an addressing mode. These wrap things that will resolve down into a
633 /// symbol reference. If no match is possible, this returns true, otherwise it
635 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
636 // If the addressing mode already has a symbol as the displacement, we can
637 // never match another symbol.
638 if (AM.hasSymbolicDisplacement())
641 SDValue N0 = N.getOperand(0);
642 CodeModel::Model M = TM.getCodeModel();
644 // Handle X86-64 rip-relative addresses. We check this before checking direct
645 // folding because RIP is preferable to non-RIP accesses.
646 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
647 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
648 // they cannot be folded into immediate fields.
649 // FIXME: This can be improved for kernel and other models?
650 (M == CodeModel::Small || M == CodeModel::Kernel)) {
651 // Base and index reg must be 0 in order to use %rip as base.
652 if (AM.hasBaseOrIndexReg())
654 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
655 X86ISelAddressMode Backup = AM;
656 AM.GV = G->getGlobal();
657 AM.SymbolFlags = G->getTargetFlags();
658 if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
662 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
663 X86ISelAddressMode Backup = AM;
664 AM.CP = CP->getConstVal();
665 AM.Align = CP->getAlignment();
666 AM.SymbolFlags = CP->getTargetFlags();
667 if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
671 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
672 AM.ES = S->getSymbol();
673 AM.SymbolFlags = S->getTargetFlags();
674 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
675 AM.JT = J->getIndex();
676 AM.SymbolFlags = J->getTargetFlags();
677 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
678 X86ISelAddressMode Backup = AM;
679 AM.BlockAddr = BA->getBlockAddress();
680 AM.SymbolFlags = BA->getTargetFlags();
681 if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
686 llvm_unreachable("Unhandled symbol reference node.");
688 if (N.getOpcode() == X86ISD::WrapperRIP)
689 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
693 // Handle the case when globals fit in our immediate field: This is true for
694 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
695 // mode, this only applies to a non-RIP-relative computation.
696 if (!Subtarget->is64Bit() ||
697 M == CodeModel::Small || M == CodeModel::Kernel) {
698 assert(N.getOpcode() != X86ISD::WrapperRIP &&
699 "RIP-relative addressing already handled");
700 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
701 AM.GV = G->getGlobal();
702 AM.Disp += G->getOffset();
703 AM.SymbolFlags = G->getTargetFlags();
704 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
705 AM.CP = CP->getConstVal();
706 AM.Align = CP->getAlignment();
707 AM.Disp += CP->getOffset();
708 AM.SymbolFlags = CP->getTargetFlags();
709 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
710 AM.ES = S->getSymbol();
711 AM.SymbolFlags = S->getTargetFlags();
712 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
713 AM.JT = J->getIndex();
714 AM.SymbolFlags = J->getTargetFlags();
715 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
716 AM.BlockAddr = BA->getBlockAddress();
717 AM.Disp += BA->getOffset();
718 AM.SymbolFlags = BA->getTargetFlags();
720 llvm_unreachable("Unhandled symbol reference node.");
727 /// MatchAddress - Add the specified node to the specified addressing mode,
728 /// returning true if it cannot be done. This just pattern matches for the
730 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
731 if (MatchAddressRecursively(N, AM, 0))
734 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
735 // a smaller encoding and avoids a scaled-index.
737 AM.BaseType == X86ISelAddressMode::RegBase &&
738 AM.Base_Reg.getNode() == nullptr) {
739 AM.Base_Reg = AM.IndexReg;
743 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
744 // because it has a smaller encoding.
745 // TODO: Which other code models can use this?
746 if (TM.getCodeModel() == CodeModel::Small &&
747 Subtarget->is64Bit() &&
749 AM.BaseType == X86ISelAddressMode::RegBase &&
750 AM.Base_Reg.getNode() == nullptr &&
751 AM.IndexReg.getNode() == nullptr &&
752 AM.SymbolFlags == X86II::MO_NO_FLAG &&
753 AM.hasSymbolicDisplacement())
754 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
759 // Insert a node into the DAG at least before the Pos node's position. This
760 // will reposition the node as needed, and will assign it a node ID that is <=
761 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
762 // IDs! The selection DAG must no longer depend on their uniqueness when this
764 static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
765 if (N.getNode()->getNodeId() == -1 ||
766 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
767 DAG.RepositionNode(Pos.getNode(), N.getNode());
768 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
772 // Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if safe. This
773 // allows us to convert the shift and and into an h-register extract and
774 // a scaled index. Returns false if the simplification is performed.
775 static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
777 SDValue Shift, SDValue X,
778 X86ISelAddressMode &AM) {
779 if (Shift.getOpcode() != ISD::SRL ||
780 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
784 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
785 if (ScaleLog <= 0 || ScaleLog >= 4 ||
786 Mask != (0xffu << ScaleLog))
789 MVT VT = N.getSimpleValueType();
791 SDValue Eight = DAG.getConstant(8, MVT::i8);
792 SDValue NewMask = DAG.getConstant(0xff, VT);
793 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
794 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
795 SDValue ShlCount = DAG.getConstant(ScaleLog, MVT::i8);
796 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
798 // Insert the new nodes into the topological ordering. We must do this in
799 // a valid topological ordering as nothing is going to go back and re-sort
800 // these nodes. We continually insert before 'N' in sequence as this is
801 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
802 // hierarchy left to express.
803 InsertDAGNode(DAG, N, Eight);
804 InsertDAGNode(DAG, N, Srl);
805 InsertDAGNode(DAG, N, NewMask);
806 InsertDAGNode(DAG, N, And);
807 InsertDAGNode(DAG, N, ShlCount);
808 InsertDAGNode(DAG, N, Shl);
809 DAG.ReplaceAllUsesWith(N, Shl);
811 AM.Scale = (1 << ScaleLog);
815 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
816 // allows us to fold the shift into this addressing mode. Returns false if the
817 // transform succeeded.
818 static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
820 SDValue Shift, SDValue X,
821 X86ISelAddressMode &AM) {
822 if (Shift.getOpcode() != ISD::SHL ||
823 !isa<ConstantSDNode>(Shift.getOperand(1)))
826 // Not likely to be profitable if either the AND or SHIFT node has more
827 // than one use (unless all uses are for address computation). Besides,
828 // isel mechanism requires their node ids to be reused.
829 if (!N.hasOneUse() || !Shift.hasOneUse())
832 // Verify that the shift amount is something we can fold.
833 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
834 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
837 MVT VT = N.getSimpleValueType();
839 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
840 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
841 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
843 // Insert the new nodes into the topological ordering. We must do this in
844 // a valid topological ordering as nothing is going to go back and re-sort
845 // these nodes. We continually insert before 'N' in sequence as this is
846 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
847 // hierarchy left to express.
848 InsertDAGNode(DAG, N, NewMask);
849 InsertDAGNode(DAG, N, NewAnd);
850 InsertDAGNode(DAG, N, NewShift);
851 DAG.ReplaceAllUsesWith(N, NewShift);
853 AM.Scale = 1 << ShiftAmt;
854 AM.IndexReg = NewAnd;
858 // Implement some heroics to detect shifts of masked values where the mask can
859 // be replaced by extending the shift and undoing that in the addressing mode
860 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
861 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
862 // the addressing mode. This results in code such as:
864 // int f(short *y, int *lookup_table) {
866 // return *y + lookup_table[*y >> 11];
870 // movzwl (%rdi), %eax
873 // addl (%rsi,%rcx,4), %eax
876 // movzwl (%rdi), %eax
880 // addl (%rsi,%rcx), %eax
882 // Note that this function assumes the mask is provided as a mask *after* the
883 // value is shifted. The input chain may or may not match that, but computing
884 // such a mask is trivial.
885 static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
887 SDValue Shift, SDValue X,
888 X86ISelAddressMode &AM) {
889 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
890 !isa<ConstantSDNode>(Shift.getOperand(1)))
893 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
894 unsigned MaskLZ = countLeadingZeros(Mask);
895 unsigned MaskTZ = countTrailingZeros(Mask);
897 // The amount of shift we're trying to fit into the addressing mode is taken
898 // from the trailing zeros of the mask.
899 unsigned AMShiftAmt = MaskTZ;
901 // There is nothing we can do here unless the mask is removing some bits.
902 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
903 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
905 // We also need to ensure that mask is a continuous run of bits.
906 if (CountTrailingOnes_64(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
908 // Scale the leading zero count down based on the actual size of the value.
909 // Also scale it down based on the size of the shift.
910 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
912 // The final check is to ensure that any masked out high bits of X are
913 // already known to be zero. Otherwise, the mask has a semantic impact
914 // other than masking out a couple of low bits. Unfortunately, because of
915 // the mask, zero extensions will be removed from operands in some cases.
916 // This code works extra hard to look through extensions because we can
917 // replace them with zero extensions cheaply if necessary.
918 bool ReplacingAnyExtend = false;
919 if (X.getOpcode() == ISD::ANY_EXTEND) {
920 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
921 X.getOperand(0).getSimpleValueType().getSizeInBits();
922 // Assume that we'll replace the any-extend with a zero-extend, and
923 // narrow the search to the extended value.
925 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
926 ReplacingAnyExtend = true;
928 APInt MaskedHighBits =
929 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
930 APInt KnownZero, KnownOne;
931 DAG.ComputeMaskedBits(X, KnownZero, KnownOne);
932 if (MaskedHighBits != KnownZero) return true;
934 // We've identified a pattern that can be transformed into a single shift
935 // and an addressing mode. Make it so.
936 MVT VT = N.getSimpleValueType();
937 if (ReplacingAnyExtend) {
938 assert(X.getValueType() != VT);
939 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
940 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
941 InsertDAGNode(DAG, N, NewX);
945 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
946 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
947 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
948 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
950 // Insert the new nodes into the topological ordering. We must do this in
951 // a valid topological ordering as nothing is going to go back and re-sort
952 // these nodes. We continually insert before 'N' in sequence as this is
953 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
954 // hierarchy left to express.
955 InsertDAGNode(DAG, N, NewSRLAmt);
956 InsertDAGNode(DAG, N, NewSRL);
957 InsertDAGNode(DAG, N, NewSHLAmt);
958 InsertDAGNode(DAG, N, NewSHL);
959 DAG.ReplaceAllUsesWith(N, NewSHL);
961 AM.Scale = 1 << AMShiftAmt;
962 AM.IndexReg = NewSRL;
966 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
970 dbgs() << "MatchAddress: ";
975 return MatchAddressBase(N, AM);
977 // If this is already a %rip relative address, we can only merge immediates
978 // into it. Instead of handling this in every case, we handle it here.
979 // RIP relative addressing: %rip + 32-bit displacement!
980 if (AM.isRIPRelative()) {
981 // FIXME: JumpTable and ExternalSymbol address currently don't like
982 // displacements. It isn't very important, but this should be fixed for
984 if (!AM.ES && AM.JT != -1) return true;
986 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
987 if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
992 switch (N.getOpcode()) {
994 case ISD::Constant: {
995 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
996 if (!FoldOffsetIntoAddress(Val, AM))
1001 case X86ISD::Wrapper:
1002 case X86ISD::WrapperRIP:
1003 if (!MatchWrapper(N, AM))
1008 if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
1012 case ISD::FrameIndex:
1013 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1014 AM.Base_Reg.getNode() == nullptr &&
1015 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1016 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1017 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1023 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1027 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1028 unsigned Val = CN->getZExtValue();
1029 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1030 // that the base operand remains free for further matching. If
1031 // the base doesn't end up getting used, a post-processing step
1032 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1033 if (Val == 1 || Val == 2 || Val == 3) {
1034 AM.Scale = 1 << Val;
1035 SDValue ShVal = N.getNode()->getOperand(0);
1037 // Okay, we know that we have a scale by now. However, if the scaled
1038 // value is an add of something and a constant, we can fold the
1039 // constant into the disp field here.
1040 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1041 AM.IndexReg = ShVal.getNode()->getOperand(0);
1042 ConstantSDNode *AddVal =
1043 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1044 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1045 if (!FoldOffsetIntoAddress(Disp, AM))
1049 AM.IndexReg = ShVal;
1056 // Scale must not be used already.
1057 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1059 SDValue And = N.getOperand(0);
1060 if (And.getOpcode() != ISD::AND) break;
1061 SDValue X = And.getOperand(0);
1063 // We only handle up to 64-bit values here as those are what matter for
1064 // addressing mode optimizations.
1065 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1067 // The mask used for the transform is expected to be post-shift, but we
1068 // found the shift first so just apply the shift to the mask before passing
1070 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1071 !isa<ConstantSDNode>(And.getOperand(1)))
1073 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1075 // Try to fold the mask and shift into the scale, and return false if we
1077 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1082 case ISD::SMUL_LOHI:
1083 case ISD::UMUL_LOHI:
1084 // A mul_lohi where we need the low part can be folded as a plain multiply.
1085 if (N.getResNo() != 0) break;
1088 case X86ISD::MUL_IMM:
1089 // X*[3,5,9] -> X+X*[2,4,8]
1090 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1091 AM.Base_Reg.getNode() == nullptr &&
1092 AM.IndexReg.getNode() == nullptr) {
1094 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1095 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1096 CN->getZExtValue() == 9) {
1097 AM.Scale = unsigned(CN->getZExtValue())-1;
1099 SDValue MulVal = N.getNode()->getOperand(0);
1102 // Okay, we know that we have a scale by now. However, if the scaled
1103 // value is an add of something and a constant, we can fold the
1104 // constant into the disp field here.
1105 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1106 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1107 Reg = MulVal.getNode()->getOperand(0);
1108 ConstantSDNode *AddVal =
1109 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1110 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1111 if (FoldOffsetIntoAddress(Disp, AM))
1112 Reg = N.getNode()->getOperand(0);
1114 Reg = N.getNode()->getOperand(0);
1117 AM.IndexReg = AM.Base_Reg = Reg;
1124 // Given A-B, if A can be completely folded into the address and
1125 // the index field with the index field unused, use -B as the index.
1126 // This is a win if a has multiple parts that can be folded into
1127 // the address. Also, this saves a mov if the base register has
1128 // other uses, since it avoids a two-address sub instruction, however
1129 // it costs an additional mov if the index register has other uses.
1131 // Add an artificial use to this node so that we can keep track of
1132 // it if it gets CSE'd with a different node.
1133 HandleSDNode Handle(N);
1135 // Test if the LHS of the sub can be folded.
1136 X86ISelAddressMode Backup = AM;
1137 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1141 // Test if the index field is free for use.
1142 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1148 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1149 // If the RHS involves a register with multiple uses, this
1150 // transformation incurs an extra mov, due to the neg instruction
1151 // clobbering its operand.
1152 if (!RHS.getNode()->hasOneUse() ||
1153 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1154 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1155 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1156 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1157 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1159 // If the base is a register with multiple uses, this
1160 // transformation may save a mov.
1161 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1162 AM.Base_Reg.getNode() &&
1163 !AM.Base_Reg.getNode()->hasOneUse()) ||
1164 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1166 // If the folded LHS was interesting, this transformation saves
1167 // address arithmetic.
1168 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1169 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1170 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1172 // If it doesn't look like it may be an overall win, don't do it.
1178 // Ok, the transformation is legal and appears profitable. Go for it.
1179 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
1180 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1184 // Insert the new nodes into the topological ordering.
1185 InsertDAGNode(*CurDAG, N, Zero);
1186 InsertDAGNode(*CurDAG, N, Neg);
1191 // Add an artificial use to this node so that we can keep track of
1192 // it if it gets CSE'd with a different node.
1193 HandleSDNode Handle(N);
1195 X86ISelAddressMode Backup = AM;
1196 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1197 !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1201 // Try again after commuting the operands.
1202 if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1203 !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1207 // If we couldn't fold both operands into the address at the same time,
1208 // see if we can just put each operand into a register and fold at least
1210 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1211 !AM.Base_Reg.getNode() &&
1212 !AM.IndexReg.getNode()) {
1213 N = Handle.getValue();
1214 AM.Base_Reg = N.getOperand(0);
1215 AM.IndexReg = N.getOperand(1);
1219 N = Handle.getValue();
1224 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1225 if (CurDAG->isBaseWithConstantOffset(N)) {
1226 X86ISelAddressMode Backup = AM;
1227 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1229 // Start with the LHS as an addr mode.
1230 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1231 !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
1238 // Perform some heroic transforms on an and of a constant-count shift
1239 // with a constant to enable use of the scaled offset field.
1241 // Scale must not be used already.
1242 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1244 SDValue Shift = N.getOperand(0);
1245 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1246 SDValue X = Shift.getOperand(0);
1248 // We only handle up to 64-bit values here as those are what matter for
1249 // addressing mode optimizations.
1250 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1252 if (!isa<ConstantSDNode>(N.getOperand(1)))
1254 uint64_t Mask = N.getConstantOperandVal(1);
1256 // Try to fold the mask and shift into an extract and scale.
1257 if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1260 // Try to fold the mask and shift directly into the scale.
1261 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1264 // Try to swap the mask and shift to place shifts which can be done as
1265 // a scale on the outside of the mask.
1266 if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1272 return MatchAddressBase(N, AM);
1275 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1276 /// specified addressing mode without any further recursion.
1277 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1278 // Is the base register already occupied?
1279 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1280 // If so, check to see if the scale index register is set.
1281 if (!AM.IndexReg.getNode()) {
1287 // Otherwise, we cannot select it.
1291 // Default, generate it as a register.
1292 AM.BaseType = X86ISelAddressMode::RegBase;
1297 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1298 /// It returns the operands which make up the maximal addressing mode it can
1299 /// match by reference.
1301 /// Parent is the parent node of the addr operand that is being matched. It
1302 /// is always a load, store, atomic node, or null. It is only null when
1303 /// checking memory operands for inline asm nodes.
1304 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1305 SDValue &Scale, SDValue &Index,
1306 SDValue &Disp, SDValue &Segment) {
1307 X86ISelAddressMode AM;
1310 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1311 // that are not a MemSDNode, and thus don't have proper addrspace info.
1312 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1313 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1314 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1315 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1316 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1317 unsigned AddrSpace =
1318 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1319 // AddrSpace 256 -> GS, 257 -> FS.
1320 if (AddrSpace == 256)
1321 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1322 if (AddrSpace == 257)
1323 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1326 if (MatchAddress(N, AM))
1329 MVT VT = N.getSimpleValueType();
1330 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1331 if (!AM.Base_Reg.getNode())
1332 AM.Base_Reg = CurDAG->getRegister(0, VT);
1335 if (!AM.IndexReg.getNode())
1336 AM.IndexReg = CurDAG->getRegister(0, VT);
1338 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1342 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1343 /// match a load whose top elements are either undef or zeros. The load flavor
1344 /// is derived from the type of N, which is either v4f32 or v2f64.
1347 /// PatternChainNode: this is the matched node that has a chain input and
1349 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1350 SDValue N, SDValue &Base,
1351 SDValue &Scale, SDValue &Index,
1352 SDValue &Disp, SDValue &Segment,
1353 SDValue &PatternNodeWithChain) {
1354 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1355 PatternNodeWithChain = N.getOperand(0);
1356 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1357 PatternNodeWithChain.hasOneUse() &&
1358 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1359 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1360 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1361 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1367 // Also handle the case where we explicitly require zeros in the top
1368 // elements. This is a vector shuffle from the zero vector.
1369 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1370 // Check to see if the top elements are all zeros (or bitcast of zeros).
1371 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1372 N.getOperand(0).getNode()->hasOneUse() &&
1373 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1374 N.getOperand(0).getOperand(0).hasOneUse() &&
1375 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1376 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1377 // Okay, this is a zero extending load. Fold it.
1378 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1379 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1381 PatternNodeWithChain = SDValue(LD, 0);
1388 bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
1389 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1390 uint64_t ImmVal = CN->getZExtValue();
1391 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1394 Imm = CurDAG->getTargetConstant(ImmVal, MVT::i64);
1398 // In static codegen with small code model, we can get the address of a label
1399 // into a register with 'movl'. TableGen has already made sure we're looking
1400 // at a label of some kind.
1401 assert(N->getOpcode() == X86ISD::Wrapper &&
1402 "Unexpected node type for MOV32ri64");
1403 N = N.getOperand(0);
1405 if (N->getOpcode() != ISD::TargetConstantPool &&
1406 N->getOpcode() != ISD::TargetJumpTable &&
1407 N->getOpcode() != ISD::TargetGlobalAddress &&
1408 N->getOpcode() != ISD::TargetExternalSymbol &&
1409 N->getOpcode() != ISD::TargetBlockAddress)
1413 return TM.getCodeModel() == CodeModel::Small;
1416 bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
1417 SDValue &Scale, SDValue &Index,
1418 SDValue &Disp, SDValue &Segment) {
1419 if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1423 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1424 if (RN && RN->getReg() == 0)
1425 Base = CurDAG->getRegister(0, MVT::i64);
1426 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(N)) {
1427 // Base could already be %rip, particularly in the x32 ABI.
1428 Base = SDValue(CurDAG->getMachineNode(
1429 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1430 CurDAG->getTargetConstant(0, MVT::i64),
1432 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
1436 RN = dyn_cast<RegisterSDNode>(Index);
1437 if (RN && RN->getReg() == 0)
1438 Index = CurDAG->getRegister(0, MVT::i64);
1440 assert(Index.getValueType() == MVT::i32 &&
1441 "Expect to be extending 32-bit registers for use in LEA");
1442 Index = SDValue(CurDAG->getMachineNode(
1443 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1444 CurDAG->getTargetConstant(0, MVT::i64),
1446 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
1453 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1454 /// mode it matches can be cost effectively emitted as an LEA instruction.
1455 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1456 SDValue &Base, SDValue &Scale,
1457 SDValue &Index, SDValue &Disp,
1459 X86ISelAddressMode AM;
1461 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1463 SDValue Copy = AM.Segment;
1464 SDValue T = CurDAG->getRegister(0, MVT::i32);
1466 if (MatchAddress(N, AM))
1468 assert (T == AM.Segment);
1471 MVT VT = N.getSimpleValueType();
1472 unsigned Complexity = 0;
1473 if (AM.BaseType == X86ISelAddressMode::RegBase)
1474 if (AM.Base_Reg.getNode())
1477 AM.Base_Reg = CurDAG->getRegister(0, VT);
1478 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1481 if (AM.IndexReg.getNode())
1484 AM.IndexReg = CurDAG->getRegister(0, VT);
1486 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1491 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1492 // to a LEA. This is determined with some expermentation but is by no means
1493 // optimal (especially for code size consideration). LEA is nice because of
1494 // its three-address nature. Tweak the cost function again when we can run
1495 // convertToThreeAddress() at register allocation time.
1496 if (AM.hasSymbolicDisplacement()) {
1497 // For X86-64, we should always use lea to materialize RIP relative
1499 if (Subtarget->is64Bit())
1505 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1508 // If it isn't worth using an LEA, reject it.
1509 if (Complexity <= 2)
1512 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1516 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1517 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1518 SDValue &Scale, SDValue &Index,
1519 SDValue &Disp, SDValue &Segment) {
1520 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1521 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1523 X86ISelAddressMode AM;
1524 AM.GV = GA->getGlobal();
1525 AM.Disp += GA->getOffset();
1526 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1527 AM.SymbolFlags = GA->getTargetFlags();
1529 if (N.getValueType() == MVT::i32) {
1531 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1533 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1536 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1541 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1542 SDValue &Base, SDValue &Scale,
1543 SDValue &Index, SDValue &Disp,
1545 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1546 !IsProfitableToFold(N, P, P) ||
1547 !IsLegalToFold(N, P, P, OptLevel))
1550 return SelectAddr(N.getNode(),
1551 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1554 /// getGlobalBaseReg - Return an SDNode that returns the value of
1555 /// the global base register. Output instructions required to
1556 /// initialize the global base register, if necessary.
1558 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1559 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1560 return CurDAG->getRegister(GlobalBaseReg,
1561 getTargetLowering()->getPointerTy()).getNode();
1564 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1565 SDValue Chain = Node->getOperand(0);
1566 SDValue In1 = Node->getOperand(1);
1567 SDValue In2L = Node->getOperand(2);
1568 SDValue In2H = Node->getOperand(3);
1570 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1571 if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1573 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1574 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1575 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1576 SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node),
1577 MVT::i32, MVT::i32, MVT::Other, Ops);
1578 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1582 /// Atomic opcode table
1610 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1621 X86::LOCK_ADD64mi32,
1634 X86::LOCK_SUB64mi32,
1686 X86::LOCK_AND64mi32,
1699 X86::LOCK_XOR64mi32,
1704 // Return the target constant operand for atomic-load-op and do simple
1705 // translations, such as from atomic-load-add to lock-sub. The return value is
1706 // one of the following 3 cases:
1707 // + target-constant, the operand could be supported as a target constant.
1708 // + empty, the operand is not needed any more with the new op selected.
1709 // + non-empty, otherwise.
1710 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1712 enum AtomicOpc &Op, MVT NVT,
1714 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1715 int64_t CNVal = CN->getSExtValue();
1716 // Quit if not 32-bit imm.
1717 if ((int32_t)CNVal != CNVal)
1719 // For atomic-load-add, we could do some optimizations.
1721 // Translate to INC/DEC if ADD by 1 or -1.
1722 if ((CNVal == 1) || (CNVal == -1)) {
1723 Op = (CNVal == 1) ? INC : DEC;
1724 // No more constant operand after being translated into INC/DEC.
1727 // Translate to SUB if ADD by negative value.
1733 return CurDAG->getTargetConstant(CNVal, NVT);
1736 // If the value operand is single-used, try to optimize it.
1737 if (Op == ADD && Val.hasOneUse()) {
1738 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1739 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1741 return Val.getOperand(1);
1743 // A special case for i16, which needs truncating as, in most cases, it's
1744 // promoted to i32. We will translate
1745 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1746 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1747 Val.getOperand(0).getOpcode() == ISD::SUB &&
1748 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1750 Val = Val.getOperand(0);
1751 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1759 SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
1760 if (Node->hasAnyUseOfValue(0))
1765 // Optimize common patterns for __sync_or_and_fetch and similar arith
1766 // operations where the result is not used. This allows us to use the "lock"
1767 // version of the arithmetic instruction.
1768 SDValue Chain = Node->getOperand(0);
1769 SDValue Ptr = Node->getOperand(1);
1770 SDValue Val = Node->getOperand(2);
1771 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1772 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1775 // Which index into the table.
1777 switch (Node->getOpcode()) {
1780 case ISD::ATOMIC_LOAD_OR:
1783 case ISD::ATOMIC_LOAD_AND:
1786 case ISD::ATOMIC_LOAD_XOR:
1789 case ISD::ATOMIC_LOAD_ADD:
1794 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val);
1795 bool isUnOp = !Val.getNode();
1796 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1799 switch (NVT.SimpleTy) {
1800 default: return nullptr;
1803 Opc = AtomicOpcTbl[Op][ConstantI8];
1805 Opc = AtomicOpcTbl[Op][I8];
1809 if (immSext8(Val.getNode()))
1810 Opc = AtomicOpcTbl[Op][SextConstantI16];
1812 Opc = AtomicOpcTbl[Op][ConstantI16];
1814 Opc = AtomicOpcTbl[Op][I16];
1818 if (immSext8(Val.getNode()))
1819 Opc = AtomicOpcTbl[Op][SextConstantI32];
1821 Opc = AtomicOpcTbl[Op][ConstantI32];
1823 Opc = AtomicOpcTbl[Op][I32];
1826 Opc = AtomicOpcTbl[Op][I64];
1828 if (immSext8(Val.getNode()))
1829 Opc = AtomicOpcTbl[Op][SextConstantI64];
1830 else if (i64immSExt32(Val.getNode()))
1831 Opc = AtomicOpcTbl[Op][ConstantI64];
1836 assert(Opc != 0 && "Invalid arith lock transform!");
1839 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1841 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1842 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1844 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1845 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1847 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1848 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1850 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1851 SDValue RetVals[] = { Undef, Ret };
1852 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1855 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1856 /// any uses which require the SF or OF bits to be accurate.
1857 static bool HasNoSignedComparisonUses(SDNode *N) {
1858 // Examine each user of the node.
1859 for (SDNode::use_iterator UI = N->use_begin(),
1860 UE = N->use_end(); UI != UE; ++UI) {
1861 // Only examine CopyToReg uses.
1862 if (UI->getOpcode() != ISD::CopyToReg)
1864 // Only examine CopyToReg uses that copy to EFLAGS.
1865 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1868 // Examine each user of the CopyToReg use.
1869 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1870 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1871 // Only examine the Flag result.
1872 if (FlagUI.getUse().getResNo() != 1) continue;
1873 // Anything unusual: assume conservatively.
1874 if (!FlagUI->isMachineOpcode()) return false;
1875 // Examine the opcode of the user.
1876 switch (FlagUI->getMachineOpcode()) {
1877 // These comparisons don't treat the most significant bit specially.
1878 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1879 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1880 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1881 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1882 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1883 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1884 case X86::CMOVA16rr: case X86::CMOVA16rm:
1885 case X86::CMOVA32rr: case X86::CMOVA32rm:
1886 case X86::CMOVA64rr: case X86::CMOVA64rm:
1887 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1888 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1889 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1890 case X86::CMOVB16rr: case X86::CMOVB16rm:
1891 case X86::CMOVB32rr: case X86::CMOVB32rm:
1892 case X86::CMOVB64rr: case X86::CMOVB64rm:
1893 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1894 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1895 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1896 case X86::CMOVE16rr: case X86::CMOVE16rm:
1897 case X86::CMOVE32rr: case X86::CMOVE32rm:
1898 case X86::CMOVE64rr: case X86::CMOVE64rm:
1899 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1900 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1901 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1902 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1903 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1904 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1905 case X86::CMOVP16rr: case X86::CMOVP16rm:
1906 case X86::CMOVP32rr: case X86::CMOVP32rm:
1907 case X86::CMOVP64rr: case X86::CMOVP64rm:
1909 // Anything else: assume conservatively.
1910 default: return false;
1917 /// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
1918 /// is suitable for doing the {load; increment or decrement; store} to modify
1920 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
1921 SDValue StoredVal, SelectionDAG *CurDAG,
1922 LoadSDNode* &LoadNode, SDValue &InputChain) {
1924 // is the value stored the result of a DEC or INC?
1925 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
1927 // is the stored value result 0 of the load?
1928 if (StoredVal.getResNo() != 0) return false;
1930 // are there other uses of the loaded value than the inc or dec?
1931 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
1933 // is the store non-extending and non-indexed?
1934 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
1937 SDValue Load = StoredVal->getOperand(0);
1938 // Is the stored value a non-extending and non-indexed load?
1939 if (!ISD::isNormalLoad(Load.getNode())) return false;
1941 // Return LoadNode by reference.
1942 LoadNode = cast<LoadSDNode>(Load);
1943 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
1944 EVT LdVT = LoadNode->getMemoryVT();
1945 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
1949 // Is store the only read of the loaded value?
1950 if (!Load.hasOneUse())
1953 // Is the address of the store the same as the load?
1954 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
1955 LoadNode->getOffset() != StoreNode->getOffset())
1958 // Check if the chain is produced by the load or is a TokenFactor with
1959 // the load output chain as an operand. Return InputChain by reference.
1960 SDValue Chain = StoreNode->getChain();
1962 bool ChainCheck = false;
1963 if (Chain == Load.getValue(1)) {
1965 InputChain = LoadNode->getChain();
1966 } else if (Chain.getOpcode() == ISD::TokenFactor) {
1967 SmallVector<SDValue, 4> ChainOps;
1968 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
1969 SDValue Op = Chain.getOperand(i);
1970 if (Op == Load.getValue(1)) {
1975 // Make sure using Op as part of the chain would not cause a cycle here.
1976 // In theory, we could check whether the chain node is a predecessor of
1977 // the load. But that can be very expensive. Instead visit the uses and
1978 // make sure they all have smaller node id than the load.
1979 int LoadId = LoadNode->getNodeId();
1980 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
1981 UE = UI->use_end(); UI != UE; ++UI) {
1982 if (UI.getUse().getResNo() != 0)
1984 if (UI->getNodeId() > LoadId)
1988 ChainOps.push_back(Op);
1992 // Make a new TokenFactor with all the other input chains except
1994 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
1995 MVT::Other, &ChainOps[0], ChainOps.size());
2003 /// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
2004 /// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
2005 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2006 if (Opc == X86ISD::DEC) {
2007 if (LdVT == MVT::i64) return X86::DEC64m;
2008 if (LdVT == MVT::i32) return X86::DEC32m;
2009 if (LdVT == MVT::i16) return X86::DEC16m;
2010 if (LdVT == MVT::i8) return X86::DEC8m;
2012 assert(Opc == X86ISD::INC && "unrecognized opcode");
2013 if (LdVT == MVT::i64) return X86::INC64m;
2014 if (LdVT == MVT::i32) return X86::INC32m;
2015 if (LdVT == MVT::i16) return X86::INC16m;
2016 if (LdVT == MVT::i8) return X86::INC8m;
2018 llvm_unreachable("unrecognized size for LdVT");
2021 /// SelectGather - Customized ISel for GATHER operations.
2023 SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
2024 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2025 SDValue Chain = Node->getOperand(0);
2026 SDValue VSrc = Node->getOperand(2);
2027 SDValue Base = Node->getOperand(3);
2028 SDValue VIdx = Node->getOperand(4);
2029 SDValue VMask = Node->getOperand(5);
2030 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2034 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2037 // Memory Operands: Base, Scale, Index, Disp, Segment
2038 SDValue Disp = CurDAG->getTargetConstant(0, MVT::i32);
2039 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2040 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue()), VIdx,
2041 Disp, Segment, VMask, Chain};
2042 SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node), VTs, Ops);
2043 // Node has 2 outputs: VDst and MVT::Other.
2044 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2045 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2047 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2048 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2052 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2053 MVT NVT = Node->getSimpleValueType(0);
2055 unsigned Opcode = Node->getOpcode();
2058 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2060 if (Node->isMachineOpcode()) {
2061 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2062 Node->setNodeId(-1);
2063 return nullptr; // Already selected.
2068 case ISD::INTRINSIC_W_CHAIN: {
2069 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2072 case Intrinsic::x86_avx2_gather_d_pd:
2073 case Intrinsic::x86_avx2_gather_d_pd_256:
2074 case Intrinsic::x86_avx2_gather_q_pd:
2075 case Intrinsic::x86_avx2_gather_q_pd_256:
2076 case Intrinsic::x86_avx2_gather_d_ps:
2077 case Intrinsic::x86_avx2_gather_d_ps_256:
2078 case Intrinsic::x86_avx2_gather_q_ps:
2079 case Intrinsic::x86_avx2_gather_q_ps_256:
2080 case Intrinsic::x86_avx2_gather_d_q:
2081 case Intrinsic::x86_avx2_gather_d_q_256:
2082 case Intrinsic::x86_avx2_gather_q_q:
2083 case Intrinsic::x86_avx2_gather_q_q_256:
2084 case Intrinsic::x86_avx2_gather_d_d:
2085 case Intrinsic::x86_avx2_gather_d_d_256:
2086 case Intrinsic::x86_avx2_gather_q_d:
2087 case Intrinsic::x86_avx2_gather_q_d_256: {
2088 if (!Subtarget->hasAVX2())
2092 default: llvm_unreachable("Impossible intrinsic");
2093 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2094 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2095 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2096 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2097 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2098 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2099 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2100 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2101 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2102 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2103 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2104 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2105 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2106 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2107 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2108 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2110 SDNode *RetVal = SelectGather(Node, Opc);
2112 // We already called ReplaceUses inside SelectGather.
2119 case X86ISD::GlobalBaseReg:
2120 return getGlobalBaseReg();
2123 case X86ISD::ATOMOR64_DAG:
2124 case X86ISD::ATOMXOR64_DAG:
2125 case X86ISD::ATOMADD64_DAG:
2126 case X86ISD::ATOMSUB64_DAG:
2127 case X86ISD::ATOMNAND64_DAG:
2128 case X86ISD::ATOMAND64_DAG:
2129 case X86ISD::ATOMMAX64_DAG:
2130 case X86ISD::ATOMMIN64_DAG:
2131 case X86ISD::ATOMUMAX64_DAG:
2132 case X86ISD::ATOMUMIN64_DAG:
2133 case X86ISD::ATOMSWAP64_DAG: {
2136 default: llvm_unreachable("Impossible opcode");
2137 case X86ISD::ATOMOR64_DAG: Opc = X86::ATOMOR6432; break;
2138 case X86ISD::ATOMXOR64_DAG: Opc = X86::ATOMXOR6432; break;
2139 case X86ISD::ATOMADD64_DAG: Opc = X86::ATOMADD6432; break;
2140 case X86ISD::ATOMSUB64_DAG: Opc = X86::ATOMSUB6432; break;
2141 case X86ISD::ATOMNAND64_DAG: Opc = X86::ATOMNAND6432; break;
2142 case X86ISD::ATOMAND64_DAG: Opc = X86::ATOMAND6432; break;
2143 case X86ISD::ATOMMAX64_DAG: Opc = X86::ATOMMAX6432; break;
2144 case X86ISD::ATOMMIN64_DAG: Opc = X86::ATOMMIN6432; break;
2145 case X86ISD::ATOMUMAX64_DAG: Opc = X86::ATOMUMAX6432; break;
2146 case X86ISD::ATOMUMIN64_DAG: Opc = X86::ATOMUMIN6432; break;
2147 case X86ISD::ATOMSWAP64_DAG: Opc = X86::ATOMSWAP6432; break;
2149 SDNode *RetVal = SelectAtomic64(Node, Opc);
2155 case ISD::ATOMIC_LOAD_XOR:
2156 case ISD::ATOMIC_LOAD_AND:
2157 case ISD::ATOMIC_LOAD_OR:
2158 case ISD::ATOMIC_LOAD_ADD: {
2159 SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
2167 // For operations of the form (x << C1) op C2, check if we can use a smaller
2168 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2169 SDValue N0 = Node->getOperand(0);
2170 SDValue N1 = Node->getOperand(1);
2172 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2175 // i8 is unshrinkable, i16 should be promoted to i32.
2176 if (NVT != MVT::i32 && NVT != MVT::i64)
2179 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2180 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2181 if (!Cst || !ShlCst)
2184 int64_t Val = Cst->getSExtValue();
2185 uint64_t ShlVal = ShlCst->getZExtValue();
2187 // Make sure that we don't change the operation by removing bits.
2188 // This only matters for OR and XOR, AND is unaffected.
2189 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2190 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2196 // Check the minimum bitwidth for the new constant.
2197 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2198 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2199 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2200 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2202 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2205 // Bail if there is no smaller encoding.
2209 switch (NVT.SimpleTy) {
2210 default: llvm_unreachable("Unsupported VT!");
2212 assert(CstVT == MVT::i8);
2213 ShlOp = X86::SHL32ri;
2216 default: llvm_unreachable("Impossible opcode");
2217 case ISD::AND: Op = X86::AND32ri8; break;
2218 case ISD::OR: Op = X86::OR32ri8; break;
2219 case ISD::XOR: Op = X86::XOR32ri8; break;
2223 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2224 ShlOp = X86::SHL64ri;
2227 default: llvm_unreachable("Impossible opcode");
2228 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2229 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2230 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2235 // Emit the smaller op and the shift.
2236 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, CstVT);
2237 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2238 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2241 case X86ISD::UMUL: {
2242 SDValue N0 = Node->getOperand(0);
2243 SDValue N1 = Node->getOperand(1);
2246 switch (NVT.SimpleTy) {
2247 default: llvm_unreachable("Unsupported VT!");
2248 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2249 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2250 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2251 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2254 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2255 N0, SDValue()).getValue(1);
2257 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2258 SDValue Ops[] = {N1, InFlag};
2259 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2261 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2262 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2263 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2267 case ISD::SMUL_LOHI:
2268 case ISD::UMUL_LOHI: {
2269 SDValue N0 = Node->getOperand(0);
2270 SDValue N1 = Node->getOperand(1);
2272 bool isSigned = Opcode == ISD::SMUL_LOHI;
2273 bool hasBMI2 = Subtarget->hasBMI2();
2275 switch (NVT.SimpleTy) {
2276 default: llvm_unreachable("Unsupported VT!");
2277 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2278 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2279 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2280 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2281 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2282 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2285 switch (NVT.SimpleTy) {
2286 default: llvm_unreachable("Unsupported VT!");
2287 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2288 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2289 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2290 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2294 unsigned SrcReg, LoReg, HiReg;
2296 default: llvm_unreachable("Unknown MUL opcode!");
2299 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2303 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2307 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2311 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2314 SrcReg = X86::EDX; LoReg = HiReg = 0;
2317 SrcReg = X86::RDX; LoReg = HiReg = 0;
2321 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2322 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2323 // Multiply is commmutative.
2325 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2330 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2331 N0, SDValue()).getValue(1);
2332 SDValue ResHi, ResLo;
2336 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2338 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2339 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2340 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2341 ResHi = SDValue(CNode, 0);
2342 ResLo = SDValue(CNode, 1);
2343 Chain = SDValue(CNode, 2);
2344 InFlag = SDValue(CNode, 3);
2346 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2347 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2348 Chain = SDValue(CNode, 0);
2349 InFlag = SDValue(CNode, 1);
2352 // Update the chain.
2353 ReplaceUses(N1.getValue(1), Chain);
2355 SDValue Ops[] = { N1, InFlag };
2356 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2357 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2358 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2359 ResHi = SDValue(CNode, 0);
2360 ResLo = SDValue(CNode, 1);
2361 InFlag = SDValue(CNode, 2);
2363 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2364 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2365 InFlag = SDValue(CNode, 0);
2369 // Prevent use of AH in a REX instruction by referencing AX instead.
2370 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2371 !SDValue(Node, 1).use_empty()) {
2372 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2373 X86::AX, MVT::i16, InFlag);
2374 InFlag = Result.getValue(2);
2375 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2377 if (!SDValue(Node, 0).use_empty())
2378 ReplaceUses(SDValue(Node, 1),
2379 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2381 // Shift AX down 8 bits.
2382 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2384 CurDAG->getTargetConstant(8, MVT::i8)), 0);
2385 // Then truncate it down to i8.
2386 ReplaceUses(SDValue(Node, 1),
2387 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2389 // Copy the low half of the result, if it is needed.
2390 if (!SDValue(Node, 0).use_empty()) {
2391 if (!ResLo.getNode()) {
2392 assert(LoReg && "Register for low half is not defined!");
2393 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2395 InFlag = ResLo.getValue(2);
2397 ReplaceUses(SDValue(Node, 0), ResLo);
2398 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2400 // Copy the high half of the result, if it is needed.
2401 if (!SDValue(Node, 1).use_empty()) {
2402 if (!ResHi.getNode()) {
2403 assert(HiReg && "Register for high half is not defined!");
2404 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2406 InFlag = ResHi.getValue(2);
2408 ReplaceUses(SDValue(Node, 1), ResHi);
2409 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2416 case ISD::UDIVREM: {
2417 SDValue N0 = Node->getOperand(0);
2418 SDValue N1 = Node->getOperand(1);
2420 bool isSigned = Opcode == ISD::SDIVREM;
2422 switch (NVT.SimpleTy) {
2423 default: llvm_unreachable("Unsupported VT!");
2424 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2425 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2426 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2427 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2430 switch (NVT.SimpleTy) {
2431 default: llvm_unreachable("Unsupported VT!");
2432 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2433 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2434 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2435 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2439 unsigned LoReg, HiReg, ClrReg;
2440 unsigned SExtOpcode;
2441 switch (NVT.SimpleTy) {
2442 default: llvm_unreachable("Unsupported VT!");
2444 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2445 SExtOpcode = X86::CBW;
2448 LoReg = X86::AX; HiReg = X86::DX;
2450 SExtOpcode = X86::CWD;
2453 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2454 SExtOpcode = X86::CDQ;
2457 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2458 SExtOpcode = X86::CQO;
2462 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2463 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2464 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2467 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2468 // Special case for div8, just use a move with zero extension to AX to
2469 // clear the upper 8 bits (AH).
2470 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2471 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2472 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2474 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2475 MVT::Other, Ops), 0);
2476 Chain = Move.getValue(1);
2477 ReplaceUses(N0.getValue(1), Chain);
2480 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2481 Chain = CurDAG->getEntryNode();
2483 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2484 InFlag = Chain.getValue(1);
2487 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2488 LoReg, N0, SDValue()).getValue(1);
2489 if (isSigned && !signBitIsZero) {
2490 // Sign extend the low part into the high part.
2492 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2494 // Zero out the high part, effectively zero extending the input.
2495 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2496 switch (NVT.SimpleTy) {
2499 SDValue(CurDAG->getMachineNode(
2500 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2501 CurDAG->getTargetConstant(X86::sub_16bit, MVT::i32)),
2508 SDValue(CurDAG->getMachineNode(
2509 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2510 CurDAG->getTargetConstant(0, MVT::i64), ClrNode,
2511 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
2515 llvm_unreachable("Unexpected division source");
2518 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2519 ClrNode, InFlag).getValue(1);
2524 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2527 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2528 InFlag = SDValue(CNode, 1);
2529 // Update the chain.
2530 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2533 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2536 // Prevent use of AH in a REX instruction by referencing AX instead.
2537 // Shift it down 8 bits.
2539 // The current assumption of the register allocator is that isel
2540 // won't generate explicit references to the GPR8_NOREX registers. If
2541 // the allocator and/or the backend get enhanced to be more robust in
2542 // that regard, this can be, and should be, removed.
2543 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2544 !SDValue(Node, 1).use_empty()) {
2545 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2546 X86::AX, MVT::i16, InFlag);
2547 InFlag = Result.getValue(2);
2549 // If we also need AL (the quotient), get it by extracting a subreg from
2550 // Result. The fast register allocator does not like multiple CopyFromReg
2551 // nodes using aliasing registers.
2552 if (!SDValue(Node, 0).use_empty())
2553 ReplaceUses(SDValue(Node, 0),
2554 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2556 // Shift AX right by 8 bits instead of using AH.
2557 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2559 CurDAG->getTargetConstant(8, MVT::i8)),
2561 ReplaceUses(SDValue(Node, 1),
2562 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2564 // Copy the division (low) result, if it is needed.
2565 if (!SDValue(Node, 0).use_empty()) {
2566 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2567 LoReg, NVT, InFlag);
2568 InFlag = Result.getValue(2);
2569 ReplaceUses(SDValue(Node, 0), Result);
2570 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2572 // Copy the remainder (high) result, if it is needed.
2573 if (!SDValue(Node, 1).use_empty()) {
2574 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2575 HiReg, NVT, InFlag);
2576 InFlag = Result.getValue(2);
2577 ReplaceUses(SDValue(Node, 1), Result);
2578 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2585 // Sometimes a SUB is used to perform comparison.
2586 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2587 // This node is not a CMP.
2589 SDValue N0 = Node->getOperand(0);
2590 SDValue N1 = Node->getOperand(1);
2592 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2593 // use a smaller encoding.
2594 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2595 HasNoSignedComparisonUses(Node))
2596 // Look past the truncate if CMP is the only use of it.
2597 N0 = N0.getOperand(0);
2598 if ((N0.getNode()->getOpcode() == ISD::AND ||
2599 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2600 N0.getNode()->hasOneUse() &&
2601 N0.getValueType() != MVT::i8 &&
2602 X86::isZeroNode(N1)) {
2603 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2606 // For example, convert "testl %eax, $8" to "testb %al, $8"
2607 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2608 (!(C->getZExtValue() & 0x80) ||
2609 HasNoSignedComparisonUses(Node))) {
2610 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
2611 SDValue Reg = N0.getNode()->getOperand(0);
2613 // On x86-32, only the ABCD registers have 8-bit subregisters.
2614 if (!Subtarget->is64Bit()) {
2615 const TargetRegisterClass *TRC;
2616 switch (N0.getSimpleValueType().SimpleTy) {
2617 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2618 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2619 default: llvm_unreachable("Unsupported TEST operand type!");
2621 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2622 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2623 Reg.getValueType(), Reg, RC), 0);
2626 // Extract the l-register.
2627 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2631 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2633 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2634 // one, do not call ReplaceAllUsesWith.
2635 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2636 SDValue(NewNode, 0));
2640 // For example, "testl %eax, $2048" to "testb %ah, $8".
2641 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2642 (!(C->getZExtValue() & 0x8000) ||
2643 HasNoSignedComparisonUses(Node))) {
2644 // Shift the immediate right by 8 bits.
2645 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2647 SDValue Reg = N0.getNode()->getOperand(0);
2649 // Put the value in an ABCD register.
2650 const TargetRegisterClass *TRC;
2651 switch (N0.getSimpleValueType().SimpleTy) {
2652 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2653 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2654 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2655 default: llvm_unreachable("Unsupported TEST operand type!");
2657 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2658 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2659 Reg.getValueType(), Reg, RC), 0);
2661 // Extract the h-register.
2662 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2665 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2666 // target GR8_NOREX registers, so make sure the register class is
2668 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2669 MVT::i32, Subreg, ShiftedImm);
2670 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2671 // one, do not call ReplaceAllUsesWith.
2672 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2673 SDValue(NewNode, 0));
2677 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2678 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2679 N0.getValueType() != MVT::i16 &&
2680 (!(C->getZExtValue() & 0x8000) ||
2681 HasNoSignedComparisonUses(Node))) {
2682 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
2683 SDValue Reg = N0.getNode()->getOperand(0);
2685 // Extract the 16-bit subregister.
2686 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2690 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2692 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2693 // one, do not call ReplaceAllUsesWith.
2694 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2695 SDValue(NewNode, 0));
2699 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2700 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2701 N0.getValueType() == MVT::i64 &&
2702 (!(C->getZExtValue() & 0x80000000) ||
2703 HasNoSignedComparisonUses(Node))) {
2704 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
2705 SDValue Reg = N0.getNode()->getOperand(0);
2707 // Extract the 32-bit subregister.
2708 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2712 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2714 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2715 // one, do not call ReplaceAllUsesWith.
2716 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2717 SDValue(NewNode, 0));
2724 // Change a chain of {load; incr or dec; store} of the same value into
2725 // a simple increment or decrement through memory of that value, if the
2726 // uses of the modified value and its address are suitable.
2727 // The DEC64m tablegen pattern is currently not able to match the case where
2728 // the EFLAGS on the original DEC are used. (This also applies to
2729 // {INC,DEC}X{64,32,16,8}.)
2730 // We'll need to improve tablegen to allow flags to be transferred from a
2731 // node in the pattern to the result node. probably with a new keyword
2732 // for example, we have this
2733 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2734 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2735 // (implicit EFLAGS)]>;
2736 // but maybe need something like this
2737 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2738 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2739 // (transferrable EFLAGS)]>;
2741 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2742 SDValue StoredVal = StoreNode->getOperand(1);
2743 unsigned Opc = StoredVal->getOpcode();
2745 LoadSDNode *LoadNode = nullptr;
2747 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2748 LoadNode, InputChain))
2751 SDValue Base, Scale, Index, Disp, Segment;
2752 if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
2753 Base, Scale, Index, Disp, Segment))
2756 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2757 MemOp[0] = StoreNode->getMemOperand();
2758 MemOp[1] = LoadNode->getMemOperand();
2759 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2760 EVT LdVT = LoadNode->getMemoryVT();
2761 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2762 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2764 MVT::i32, MVT::Other, Ops);
2765 Result->setMemRefs(MemOp, MemOp + 2);
2767 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2768 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2774 SDNode *ResNode = SelectCode(Node);
2776 DEBUG(dbgs() << "=> ";
2777 if (ResNode == NULL || ResNode == Node)
2780 ResNode->dump(CurDAG);
2786 bool X86DAGToDAGISel::
2787 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2788 std::vector<SDValue> &OutOps) {
2789 SDValue Op0, Op1, Op2, Op3, Op4;
2790 switch (ConstraintCode) {
2791 case 'o': // offsetable ??
2792 case 'v': // not offsetable ??
2793 default: return true;
2795 if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
2800 OutOps.push_back(Op0);
2801 OutOps.push_back(Op1);
2802 OutOps.push_back(Op2);
2803 OutOps.push_back(Op3);
2804 OutOps.push_back(Op4);
2808 /// createX86ISelDag - This pass converts a legalized DAG into a
2809 /// X86-specific DAG, ready for instruction scheduling.
2811 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2812 CodeGenOpt::Level OptLevel) {
2813 return new X86DAGToDAGISel(TM, OptLevel);