LLVM  9.0.0svn
X86DomainReassignment.cpp
Go to the documentation of this file.
1 //===--- X86DomainReassignment.cpp - Selectively switch register classes---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass attempts to find instruction chains (closures) in one domain,
10 // and convert them to equivalent instructions in a different domain,
11 // if profitable.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "X86.h"
16 #include "X86InstrInfo.h"
17 #include "X86Subtarget.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DenseMapInfo.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/Printable.h"
29 #include <bitset>
30 
31 using namespace llvm;
32 
33 #define DEBUG_TYPE "x86-domain-reassignment"
34 
35 STATISTIC(NumClosuresConverted, "Number of closures converted by the pass");
36 
38  "disable-x86-domain-reassignment", cl::Hidden,
39  cl::desc("X86: Disable Virtual Register Reassignment."), cl::init(false));
40 
41 namespace {
42 enum RegDomain { NoDomain = -1, GPRDomain, MaskDomain, OtherDomain, NumDomains };
43 
44 static bool isGPR(const TargetRegisterClass *RC) {
45  return X86::GR64RegClass.hasSubClassEq(RC) ||
46  X86::GR32RegClass.hasSubClassEq(RC) ||
47  X86::GR16RegClass.hasSubClassEq(RC) ||
48  X86::GR8RegClass.hasSubClassEq(RC);
49 }
50 
51 static bool isMask(const TargetRegisterClass *RC,
52  const TargetRegisterInfo *TRI) {
53  return X86::VK16RegClass.hasSubClassEq(RC);
54 }
55 
56 static RegDomain getDomain(const TargetRegisterClass *RC,
57  const TargetRegisterInfo *TRI) {
58  if (isGPR(RC))
59  return GPRDomain;
60  if (isMask(RC, TRI))
61  return MaskDomain;
62  return OtherDomain;
63 }
64 
65 /// Return a register class equivalent to \p SrcRC, in \p Domain.
66 static const TargetRegisterClass *getDstRC(const TargetRegisterClass *SrcRC,
67  RegDomain Domain) {
68  assert(Domain == MaskDomain && "add domain");
69  if (X86::GR8RegClass.hasSubClassEq(SrcRC))
70  return &X86::VK8RegClass;
71  if (X86::GR16RegClass.hasSubClassEq(SrcRC))
72  return &X86::VK16RegClass;
73  if (X86::GR32RegClass.hasSubClassEq(SrcRC))
74  return &X86::VK32RegClass;
75  if (X86::GR64RegClass.hasSubClassEq(SrcRC))
76  return &X86::VK64RegClass;
77  llvm_unreachable("add register class");
78  return nullptr;
79 }
80 
81 /// Abstract Instruction Converter class.
82 class InstrConverterBase {
83 protected:
84  unsigned SrcOpcode;
85 
86 public:
87  InstrConverterBase(unsigned SrcOpcode) : SrcOpcode(SrcOpcode) {}
88 
89  virtual ~InstrConverterBase() {}
90 
91  /// \returns true if \p MI is legal to convert.
92  virtual bool isLegal(const MachineInstr *MI,
93  const TargetInstrInfo *TII) const {
94  assert(MI->getOpcode() == SrcOpcode &&
95  "Wrong instruction passed to converter");
96  return true;
97  }
98 
99  /// Applies conversion to \p MI.
100  ///
101  /// \returns true if \p MI is no longer need, and can be deleted.
102  virtual bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
103  MachineRegisterInfo *MRI) const = 0;
104 
105  /// \returns the cost increment incurred by converting \p MI.
106  virtual double getExtraCost(const MachineInstr *MI,
107  MachineRegisterInfo *MRI) const = 0;
108 };
109 
110 /// An Instruction Converter which ignores the given instruction.
111 /// For example, PHI instructions can be safely ignored since only the registers
112 /// need to change.
113 class InstrIgnore : public InstrConverterBase {
114 public:
115  InstrIgnore(unsigned SrcOpcode) : InstrConverterBase(SrcOpcode) {}
116 
117  bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
118  MachineRegisterInfo *MRI) const override {
119  assert(isLegal(MI, TII) && "Cannot convert instruction");
120  return false;
121  }
122 
123  double getExtraCost(const MachineInstr *MI,
124  MachineRegisterInfo *MRI) const override {
125  return 0;
126  }
127 };
128 
129 /// An Instruction Converter which replaces an instruction with another.
130 class InstrReplacer : public InstrConverterBase {
131 public:
132  /// Opcode of the destination instruction.
133  unsigned DstOpcode;
134 
135  InstrReplacer(unsigned SrcOpcode, unsigned DstOpcode)
136  : InstrConverterBase(SrcOpcode), DstOpcode(DstOpcode) {}
137 
138  bool isLegal(const MachineInstr *MI,
139  const TargetInstrInfo *TII) const override {
140  if (!InstrConverterBase::isLegal(MI, TII))
141  return false;
142  // It's illegal to replace an instruction that implicitly defines a register
143  // with an instruction that doesn't, unless that register dead.
144  for (auto &MO : MI->implicit_operands())
145  if (MO.isReg() && MO.isDef() && !MO.isDead() &&
146  !TII->get(DstOpcode).hasImplicitDefOfPhysReg(MO.getReg()))
147  return false;
148  return true;
149  }
150 
151  bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
152  MachineRegisterInfo *MRI) const override {
153  assert(isLegal(MI, TII) && "Cannot convert instruction");
154  MachineInstrBuilder Bld =
155  BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII->get(DstOpcode));
156  // Transfer explicit operands from original instruction. Implicit operands
157  // are handled by BuildMI.
158  for (auto &Op : MI->explicit_operands())
159  Bld.add(Op);
160  return true;
161  }
162 
163  double getExtraCost(const MachineInstr *MI,
164  MachineRegisterInfo *MRI) const override {
165  // Assuming instructions have the same cost.
166  return 0;
167  }
168 };
169 
170 /// An Instruction Converter which replaces an instruction with another, and
171 /// adds a COPY from the new instruction's destination to the old one's.
172 class InstrReplacerDstCOPY : public InstrConverterBase {
173 public:
174  unsigned DstOpcode;
175 
176  InstrReplacerDstCOPY(unsigned SrcOpcode, unsigned DstOpcode)
177  : InstrConverterBase(SrcOpcode), DstOpcode(DstOpcode) {}
178 
179  bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
180  MachineRegisterInfo *MRI) const override {
181  assert(isLegal(MI, TII) && "Cannot convert instruction");
182  MachineBasicBlock *MBB = MI->getParent();
183  auto &DL = MI->getDebugLoc();
184 
185  unsigned Reg = MRI->createVirtualRegister(
186  TII->getRegClass(TII->get(DstOpcode), 0, MRI->getTargetRegisterInfo(),
187  *MBB->getParent()));
188  MachineInstrBuilder Bld = BuildMI(*MBB, MI, DL, TII->get(DstOpcode), Reg);
189  for (unsigned Idx = 1, End = MI->getNumOperands(); Idx < End; ++Idx)
190  Bld.add(MI->getOperand(Idx));
191 
192  BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::COPY))
193  .add(MI->getOperand(0))
194  .addReg(Reg);
195 
196  return true;
197  }
198 
199  double getExtraCost(const MachineInstr *MI,
200  MachineRegisterInfo *MRI) const override {
201  // Assuming instructions have the same cost, and that COPY is in the same
202  // domain so it will be eliminated.
203  return 0;
204  }
205 };
206 
207 /// An Instruction Converter for replacing COPY instructions.
208 class InstrCOPYReplacer : public InstrReplacer {
209 public:
210  RegDomain DstDomain;
211 
212  InstrCOPYReplacer(unsigned SrcOpcode, RegDomain DstDomain, unsigned DstOpcode)
213  : InstrReplacer(SrcOpcode, DstOpcode), DstDomain(DstDomain) {}
214 
215  bool isLegal(const MachineInstr *MI,
216  const TargetInstrInfo *TII) const override {
217  if (!InstrConverterBase::isLegal(MI, TII))
218  return false;
219 
220  // Don't allow copies to/flow GR8/GR16 physical registers.
221  // FIXME: Is there some better way to support this?
222  unsigned DstReg = MI->getOperand(0).getReg();
224  (X86::GR8RegClass.contains(DstReg) ||
225  X86::GR16RegClass.contains(DstReg)))
226  return false;
227  unsigned SrcReg = MI->getOperand(1).getReg();
229  (X86::GR8RegClass.contains(SrcReg) ||
230  X86::GR16RegClass.contains(SrcReg)))
231  return false;
232 
233  return true;
234  }
235 
236  double getExtraCost(const MachineInstr *MI,
237  MachineRegisterInfo *MRI) const override {
238  assert(MI->getOpcode() == TargetOpcode::COPY && "Expected a COPY");
239 
240  for (auto &MO : MI->operands()) {
241  // Physical registers will not be converted. Assume that converting the
242  // COPY to the destination domain will eventually result in a actual
243  // instruction.
245  return 1;
246 
247  RegDomain OpDomain = getDomain(MRI->getRegClass(MO.getReg()),
248  MRI->getTargetRegisterInfo());
249  // Converting a cross domain COPY to a same domain COPY should eliminate
250  // an insturction
251  if (OpDomain == DstDomain)
252  return -1;
253  }
254  return 0;
255  }
256 };
257 
258 /// An Instruction Converter which replaces an instruction with a COPY.
259 class InstrReplaceWithCopy : public InstrConverterBase {
260 public:
261  // Source instruction operand Index, to be used as the COPY source.
262  unsigned SrcOpIdx;
263 
264  InstrReplaceWithCopy(unsigned SrcOpcode, unsigned SrcOpIdx)
265  : InstrConverterBase(SrcOpcode), SrcOpIdx(SrcOpIdx) {}
266 
267  bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
268  MachineRegisterInfo *MRI) const override {
269  assert(isLegal(MI, TII) && "Cannot convert instruction");
270  BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
271  TII->get(TargetOpcode::COPY))
272  .add({MI->getOperand(0), MI->getOperand(SrcOpIdx)});
273  return true;
274  }
275 
276  double getExtraCost(const MachineInstr *MI,
277  MachineRegisterInfo *MRI) const override {
278  return 0;
279  }
280 };
281 
282 // Key type to be used by the Instruction Converters map.
283 // A converter is identified by <destination domain, source opcode>
284 typedef std::pair<int, unsigned> InstrConverterBaseKeyTy;
285 
287  InstrConverterBaseMap;
288 
289 /// A closure is a set of virtual register representing all of the edges in
290 /// the closure, as well as all of the instructions connected by those edges.
291 ///
292 /// A closure may encompass virtual registers in the same register bank that
293 /// have different widths. For example, it may contain 32-bit GPRs as well as
294 /// 64-bit GPRs.
295 ///
296 /// A closure that computes an address (i.e. defines a virtual register that is
297 /// used in a memory operand) excludes the instructions that contain memory
298 /// operands using the address. Such an instruction will be included in a
299 /// different closure that manipulates the loaded or stored value.
300 class Closure {
301 private:
302  /// Virtual registers in the closure.
303  DenseSet<unsigned> Edges;
304 
305  /// Instructions in the closure.
307 
308  /// Domains which this closure can legally be reassigned to.
309  std::bitset<NumDomains> LegalDstDomains;
310 
311  /// An ID to uniquely identify this closure, even when it gets
312  /// moved around
313  unsigned ID;
314 
315 public:
316  Closure(unsigned ID, std::initializer_list<RegDomain> LegalDstDomainList) : ID(ID) {
317  for (RegDomain D : LegalDstDomainList)
318  LegalDstDomains.set(D);
319  }
320 
321  /// Mark this closure as illegal for reassignment to all domains.
322  void setAllIllegal() { LegalDstDomains.reset(); }
323 
324  /// \returns true if this closure has domains which are legal to reassign to.
325  bool hasLegalDstDomain() const { return LegalDstDomains.any(); }
326 
327  /// \returns true if is legal to reassign this closure to domain \p RD.
328  bool isLegal(RegDomain RD) const { return LegalDstDomains[RD]; }
329 
330  /// Mark this closure as illegal for reassignment to domain \p RD.
331  void setIllegal(RegDomain RD) { LegalDstDomains[RD] = false; }
332 
333  bool empty() const { return Edges.empty(); }
334 
335  bool insertEdge(unsigned Reg) {
336  return Edges.insert(Reg).second;
337  }
338 
339  using const_edge_iterator = DenseSet<unsigned>::const_iterator;
341  return iterator_range<const_edge_iterator>(Edges.begin(), Edges.end());
342  }
343 
344  void addInstruction(MachineInstr *I) {
345  Instrs.push_back(I);
346  }
347 
349  return Instrs;
350  }
351 
352  LLVM_DUMP_METHOD void dump(const MachineRegisterInfo *MRI) const {
353  dbgs() << "Registers: ";
354  bool First = true;
355  for (unsigned Reg : Edges) {
356  if (!First)
357  dbgs() << ", ";
358  First = false;
359  dbgs() << printReg(Reg, MRI->getTargetRegisterInfo(), 0, MRI);
360  }
361  dbgs() << "\n" << "Instructions:";
362  for (MachineInstr *MI : Instrs) {
363  dbgs() << "\n ";
364  MI->print(dbgs());
365  }
366  dbgs() << "\n";
367  }
368 
369  unsigned getID() const {
370  return ID;
371  }
372 
373 };
374 
375 class X86DomainReassignment : public MachineFunctionPass {
376  const X86Subtarget *STI;
378  const X86InstrInfo *TII;
379 
380  /// All edges that are included in some closure
381  DenseSet<unsigned> EnclosedEdges;
382 
383  /// All instructions that are included in some closure.
384  DenseMap<MachineInstr *, unsigned> EnclosedInstrs;
385 
386 public:
387  static char ID;
388 
389  X86DomainReassignment() : MachineFunctionPass(ID) {
391  }
392 
393  bool runOnMachineFunction(MachineFunction &MF) override;
394 
395  void getAnalysisUsage(AnalysisUsage &AU) const override {
396  AU.setPreservesCFG();
398  }
399 
400  StringRef getPassName() const override {
401  return "X86 Domain Reassignment Pass";
402  }
403 
404 private:
405  /// A map of available Instruction Converters.
406  InstrConverterBaseMap Converters;
407 
408  /// Initialize Converters map.
409  void initConverters();
410 
411  /// Starting from \Reg, expand the closure as much as possible.
412  void buildClosure(Closure &, unsigned Reg);
413 
414  /// Enqueue \p Reg to be considered for addition to the closure.
415  void visitRegister(Closure &, unsigned Reg, RegDomain &Domain,
416  SmallVectorImpl<unsigned> &Worklist);
417 
418  /// Reassign the closure to \p Domain.
419  void reassign(const Closure &C, RegDomain Domain) const;
420 
421  /// Add \p MI to the closure.
422  void encloseInstr(Closure &C, MachineInstr *MI);
423 
424  /// /returns true if it is profitable to reassign the closure to \p Domain.
425  bool isReassignmentProfitable(const Closure &C, RegDomain Domain) const;
426 
427  /// Calculate the total cost of reassigning the closure to \p Domain.
428  double calculateCost(const Closure &C, RegDomain Domain) const;
429 };
430 
432 
433 } // End anonymous namespace.
434 
435 void X86DomainReassignment::visitRegister(Closure &C, unsigned Reg,
436  RegDomain &Domain,
437  SmallVectorImpl<unsigned> &Worklist) {
438  if (EnclosedEdges.count(Reg))
439  return;
440 
442  return;
443 
444  if (!MRI->hasOneDef(Reg))
445  return;
446 
447  RegDomain RD = getDomain(MRI->getRegClass(Reg), MRI->getTargetRegisterInfo());
448  // First edge in closure sets the domain.
449  if (Domain == NoDomain)
450  Domain = RD;
451 
452  if (Domain != RD)
453  return;
454 
455  Worklist.push_back(Reg);
456 }
457 
458 void X86DomainReassignment::encloseInstr(Closure &C, MachineInstr *MI) {
459  auto I = EnclosedInstrs.find(MI);
460  if (I != EnclosedInstrs.end()) {
461  if (I->second != C.getID())
462  // Instruction already belongs to another closure, avoid conflicts between
463  // closure and mark this closure as illegal.
464  C.setAllIllegal();
465  return;
466  }
467 
468  EnclosedInstrs[MI] = C.getID();
469  C.addInstruction(MI);
470 
471  // Mark closure as illegal for reassignment to domains, if there is no
472  // converter for the instruction or if the converter cannot convert the
473  // instruction.
474  for (int i = 0; i != NumDomains; ++i) {
475  if (C.isLegal((RegDomain)i)) {
476  InstrConverterBase *IC = Converters.lookup({i, MI->getOpcode()});
477  if (!IC || !IC->isLegal(MI, TII))
478  C.setIllegal((RegDomain)i);
479  }
480  }
481 }
482 
483 double X86DomainReassignment::calculateCost(const Closure &C,
484  RegDomain DstDomain) const {
485  assert(C.isLegal(DstDomain) && "Cannot calculate cost for illegal closure");
486 
487  double Cost = 0.0;
488  for (auto *MI : C.instructions())
489  Cost +=
490  Converters.lookup({DstDomain, MI->getOpcode()})->getExtraCost(MI, MRI);
491  return Cost;
492 }
493 
494 bool X86DomainReassignment::isReassignmentProfitable(const Closure &C,
495  RegDomain Domain) const {
496  return calculateCost(C, Domain) < 0.0;
497 }
498 
499 void X86DomainReassignment::reassign(const Closure &C, RegDomain Domain) const {
500  assert(C.isLegal(Domain) && "Cannot convert illegal closure");
501 
502  // Iterate all instructions in the closure, convert each one using the
503  // appropriate converter.
505  for (auto *MI : C.instructions())
506  if (Converters.lookup({Domain, MI->getOpcode()})
507  ->convertInstr(MI, TII, MRI))
508  ToErase.push_back(MI);
509 
510  // Iterate all registers in the closure, replace them with registers in the
511  // destination domain.
512  for (unsigned Reg : C.edges()) {
513  MRI->setRegClass(Reg, getDstRC(MRI->getRegClass(Reg), Domain));
514  for (auto &MO : MRI->use_operands(Reg)) {
515  if (MO.isReg())
516  // Remove all subregister references as they are not valid in the
517  // destination domain.
518  MO.setSubReg(0);
519  }
520  }
521 
522  for (auto MI : ToErase)
523  MI->eraseFromParent();
524 }
525 
526 /// \returns true when \p Reg is used as part of an address calculation in \p
527 /// MI.
528 static bool usedAsAddr(const MachineInstr &MI, unsigned Reg,
529  const TargetInstrInfo *TII) {
530  if (!MI.mayLoadOrStore())
531  return false;
532 
533  const MCInstrDesc &Desc = TII->get(MI.getOpcode());
534  int MemOpStart = X86II::getMemoryOperandNo(Desc.TSFlags);
535  if (MemOpStart == -1)
536  return false;
537 
538  MemOpStart += X86II::getOperandBias(Desc);
539  for (unsigned MemOpIdx = MemOpStart,
540  MemOpEnd = MemOpStart + X86::AddrNumOperands;
541  MemOpIdx < MemOpEnd; ++MemOpIdx) {
542  auto &Op = MI.getOperand(MemOpIdx);
543  if (Op.isReg() && Op.getReg() == Reg)
544  return true;
545  }
546  return false;
547 }
548 
549 void X86DomainReassignment::buildClosure(Closure &C, unsigned Reg) {
550  SmallVector<unsigned, 4> Worklist;
551  RegDomain Domain = NoDomain;
552  visitRegister(C, Reg, Domain, Worklist);
553  while (!Worklist.empty()) {
554  unsigned CurReg = Worklist.pop_back_val();
555 
556  // Register already in this closure.
557  if (!C.insertEdge(CurReg))
558  continue;
559 
560  MachineInstr *DefMI = MRI->getVRegDef(CurReg);
561  encloseInstr(C, DefMI);
562 
563  // Add register used by the defining MI to the worklist.
564  // Do not add registers which are used in address calculation, they will be
565  // added to a different closure.
566  int OpEnd = DefMI->getNumOperands();
567  const MCInstrDesc &Desc = DefMI->getDesc();
568  int MemOp = X86II::getMemoryOperandNo(Desc.TSFlags);
569  if (MemOp != -1)
570  MemOp += X86II::getOperandBias(Desc);
571  for (int OpIdx = 0; OpIdx < OpEnd; ++OpIdx) {
572  if (OpIdx == MemOp) {
573  // skip address calculation.
574  OpIdx += (X86::AddrNumOperands - 1);
575  continue;
576  }
577  auto &Op = DefMI->getOperand(OpIdx);
578  if (!Op.isReg() || !Op.isUse())
579  continue;
580  visitRegister(C, Op.getReg(), Domain, Worklist);
581  }
582 
583  // Expand closure through register uses.
584  for (auto &UseMI : MRI->use_nodbg_instructions(CurReg)) {
585  // We would like to avoid converting closures which calculare addresses,
586  // as this should remain in GPRs.
587  if (usedAsAddr(UseMI, CurReg, TII)) {
588  C.setAllIllegal();
589  continue;
590  }
591  encloseInstr(C, &UseMI);
592 
593  for (auto &DefOp : UseMI.defs()) {
594  if (!DefOp.isReg())
595  continue;
596 
597  unsigned DefReg = DefOp.getReg();
599  C.setAllIllegal();
600  continue;
601  }
602  visitRegister(C, DefReg, Domain, Worklist);
603  }
604  }
605  }
606 }
607 
608 void X86DomainReassignment::initConverters() {
609  Converters[{MaskDomain, TargetOpcode::PHI}] =
610  new InstrIgnore(TargetOpcode::PHI);
611 
612  Converters[{MaskDomain, TargetOpcode::IMPLICIT_DEF}] =
613  new InstrIgnore(TargetOpcode::IMPLICIT_DEF);
614 
615  Converters[{MaskDomain, TargetOpcode::INSERT_SUBREG}] =
616  new InstrReplaceWithCopy(TargetOpcode::INSERT_SUBREG, 2);
617 
618  Converters[{MaskDomain, TargetOpcode::COPY}] =
619  new InstrCOPYReplacer(TargetOpcode::COPY, MaskDomain, TargetOpcode::COPY);
620 
621  auto createReplacerDstCOPY = [&](unsigned From, unsigned To) {
622  Converters[{MaskDomain, From}] = new InstrReplacerDstCOPY(From, To);
623  };
624 
625  createReplacerDstCOPY(X86::MOVZX32rm16, X86::KMOVWkm);
626  createReplacerDstCOPY(X86::MOVZX64rm16, X86::KMOVWkm);
627 
628  createReplacerDstCOPY(X86::MOVZX32rr16, X86::KMOVWkk);
629  createReplacerDstCOPY(X86::MOVZX64rr16, X86::KMOVWkk);
630 
631  if (STI->hasDQI()) {
632  createReplacerDstCOPY(X86::MOVZX16rm8, X86::KMOVBkm);
633  createReplacerDstCOPY(X86::MOVZX32rm8, X86::KMOVBkm);
634  createReplacerDstCOPY(X86::MOVZX64rm8, X86::KMOVBkm);
635 
636  createReplacerDstCOPY(X86::MOVZX16rr8, X86::KMOVBkk);
637  createReplacerDstCOPY(X86::MOVZX32rr8, X86::KMOVBkk);
638  createReplacerDstCOPY(X86::MOVZX64rr8, X86::KMOVBkk);
639  }
640 
641  auto createReplacer = [&](unsigned From, unsigned To) {
642  Converters[{MaskDomain, From}] = new InstrReplacer(From, To);
643  };
644 
645  createReplacer(X86::MOV16rm, X86::KMOVWkm);
646  createReplacer(X86::MOV16mr, X86::KMOVWmk);
647  createReplacer(X86::MOV16rr, X86::KMOVWkk);
648  createReplacer(X86::SHR16ri, X86::KSHIFTRWri);
649  createReplacer(X86::SHL16ri, X86::KSHIFTLWri);
650  createReplacer(X86::NOT16r, X86::KNOTWrr);
651  createReplacer(X86::OR16rr, X86::KORWrr);
652  createReplacer(X86::AND16rr, X86::KANDWrr);
653  createReplacer(X86::XOR16rr, X86::KXORWrr);
654 
655  if (STI->hasBWI()) {
656  createReplacer(X86::MOV32rm, X86::KMOVDkm);
657  createReplacer(X86::MOV64rm, X86::KMOVQkm);
658 
659  createReplacer(X86::MOV32mr, X86::KMOVDmk);
660  createReplacer(X86::MOV64mr, X86::KMOVQmk);
661 
662  createReplacer(X86::MOV32rr, X86::KMOVDkk);
663  createReplacer(X86::MOV64rr, X86::KMOVQkk);
664 
665  createReplacer(X86::SHR32ri, X86::KSHIFTRDri);
666  createReplacer(X86::SHR64ri, X86::KSHIFTRQri);
667 
668  createReplacer(X86::SHL32ri, X86::KSHIFTLDri);
669  createReplacer(X86::SHL64ri, X86::KSHIFTLQri);
670 
671  createReplacer(X86::ADD32rr, X86::KADDDrr);
672  createReplacer(X86::ADD64rr, X86::KADDQrr);
673 
674  createReplacer(X86::NOT32r, X86::KNOTDrr);
675  createReplacer(X86::NOT64r, X86::KNOTQrr);
676 
677  createReplacer(X86::OR32rr, X86::KORDrr);
678  createReplacer(X86::OR64rr, X86::KORQrr);
679 
680  createReplacer(X86::AND32rr, X86::KANDDrr);
681  createReplacer(X86::AND64rr, X86::KANDQrr);
682 
683  createReplacer(X86::ANDN32rr, X86::KANDNDrr);
684  createReplacer(X86::ANDN64rr, X86::KANDNQrr);
685 
686  createReplacer(X86::XOR32rr, X86::KXORDrr);
687  createReplacer(X86::XOR64rr, X86::KXORQrr);
688 
689  // TODO: KTEST is not a replacement for TEST due to flag differences. Need
690  // to prove only Z flag is used.
691  //createReplacer(X86::TEST32rr, X86::KTESTDrr);
692  //createReplacer(X86::TEST64rr, X86::KTESTQrr);
693  }
694 
695  if (STI->hasDQI()) {
696  createReplacer(X86::ADD8rr, X86::KADDBrr);
697  createReplacer(X86::ADD16rr, X86::KADDWrr);
698 
699  createReplacer(X86::AND8rr, X86::KANDBrr);
700 
701  createReplacer(X86::MOV8rm, X86::KMOVBkm);
702  createReplacer(X86::MOV8mr, X86::KMOVBmk);
703  createReplacer(X86::MOV8rr, X86::KMOVBkk);
704 
705  createReplacer(X86::NOT8r, X86::KNOTBrr);
706 
707  createReplacer(X86::OR8rr, X86::KORBrr);
708 
709  createReplacer(X86::SHR8ri, X86::KSHIFTRBri);
710  createReplacer(X86::SHL8ri, X86::KSHIFTLBri);
711 
712  // TODO: KTEST is not a replacement for TEST due to flag differences. Need
713  // to prove only Z flag is used.
714  //createReplacer(X86::TEST8rr, X86::KTESTBrr);
715  //createReplacer(X86::TEST16rr, X86::KTESTWrr);
716 
717  createReplacer(X86::XOR8rr, X86::KXORBrr);
718  }
719 }
720 
721 bool X86DomainReassignment::runOnMachineFunction(MachineFunction &MF) {
722  if (skipFunction(MF.getFunction()))
723  return false;
725  return false;
726 
727  LLVM_DEBUG(
728  dbgs() << "***** Machine Function before Domain Reassignment *****\n");
729  LLVM_DEBUG(MF.print(dbgs()));
730 
731  STI = &MF.getSubtarget<X86Subtarget>();
732  // GPR->K is the only transformation currently supported, bail out early if no
733  // AVX512.
734  // TODO: We're also bailing of AVX512BW isn't supported since we use VK32 and
735  // VK64 for GR32/GR64, but those aren't legal classes on KNL. If the register
736  // coalescer doesn't clean it up and we generate a spill we will crash.
737  if (!STI->hasAVX512() || !STI->hasBWI())
738  return false;
739 
740  MRI = &MF.getRegInfo();
741  assert(MRI->isSSA() && "Expected MIR to be in SSA form");
742 
743  TII = STI->getInstrInfo();
744  initConverters();
745  bool Changed = false;
746 
747  EnclosedEdges.clear();
748  EnclosedInstrs.clear();
749 
750  std::vector<Closure> Closures;
751 
752  // Go over all virtual registers and calculate a closure.
753  unsigned ClosureID = 0;
754  for (unsigned Idx = 0; Idx < MRI->getNumVirtRegs(); ++Idx) {
755  unsigned Reg = TargetRegisterInfo::index2VirtReg(Idx);
756 
757  // GPR only current source domain supported.
758  if (!isGPR(MRI->getRegClass(Reg)))
759  continue;
760 
761  // Register already in closure.
762  if (EnclosedEdges.count(Reg))
763  continue;
764 
765  // Calculate closure starting with Reg.
766  Closure C(ClosureID++, {MaskDomain});
767  buildClosure(C, Reg);
768 
769  // Collect all closures that can potentially be converted.
770  if (!C.empty() && C.isLegal(MaskDomain))
771  Closures.push_back(std::move(C));
772  }
773 
774  for (Closure &C : Closures) {
775  LLVM_DEBUG(C.dump(MRI));
776  if (isReassignmentProfitable(C, MaskDomain)) {
777  reassign(C, MaskDomain);
778  ++NumClosuresConverted;
779  Changed = true;
780  }
781  }
782 
783  DeleteContainerSeconds(Converters);
784 
785  LLVM_DEBUG(
786  dbgs() << "***** Machine Function after Domain Reassignment *****\n");
787  LLVM_DEBUG(MF.print(dbgs()));
788 
789  return Changed;
790 }
791 
792 INITIALIZE_PASS(X86DomainReassignment, "x86-domain-reassignment",
793  "X86 Domain Reassignment Pass", false, false)
794 
795 /// Returns an instance of the Domain Reassignment pass.
797  return new X86DomainReassignment();
798 }
void DeleteContainerSeconds(Container &C)
In a container of pairs (usually a map) whose second element is a pointer, deletes the second element...
Definition: STLExtras.h:1157
uint64_t CallInst * C
const MachineInstrBuilder & add(const MachineOperand &MO) const
static bool usedAsAddr(const MachineInstr &MI, unsigned Reg, const TargetInstrInfo *TII)
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
const TargetRegisterClass * getRegClass(unsigned Reg) const
Return the register class of the specified virtual register.
This class represents lattice values for constants.
Definition: AllocatorList.h:23
AddrNumOperands - Total number of operands in a memory reference.
Definition: X86BaseInfo.h:41
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:464
iterator_range< mop_iterator > explicit_operands()
Definition: MachineInstr.h:464
static unsigned index2VirtReg(unsigned Index)
Convert a 0-based index to a virtual register number.
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition: MachineInstr.h:382
Describe properties that are true of each instruction in the target description file.
Definition: MCInstrDesc.h:163
unsigned getReg() const
getReg - Returns the register number.
static bool isVirtualRegister(unsigned Reg)
Return true if the specified register number is in the virtual register namespace.
unsigned Reg
bool mayLoadOrStore(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly read or modify memory.
Definition: MachineInstr.h:829
STATISTIC(NumFunctions, "Total number of functions")
unsigned const TargetRegisterInfo * TRI
iterator_range< mop_iterator > operands()
Definition: MachineInstr.h:458
FunctionPass * createX86DomainReassignmentPass()
Return a Machine IR pass that reassigns instruction chains from one domain to another, when profitable.
return AArch64::GPR64RegClass contains(Reg)
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
const HexagonInstrInfo * TII
unsigned getNumOperands() const
Retuns the total number of operands.
Definition: MachineInstr.h:411
const TargetRegisterClass * getRegClass(const MCInstrDesc &MCID, unsigned OpNum, const TargetRegisterInfo *TRI, const MachineFunction &MF) const
Given a machine instruction descriptor, returns the register class constraint for OpNum...
Printable printReg(unsigned Reg, const TargetRegisterInfo *TRI=nullptr, unsigned SubIdx=0, const MachineRegisterInfo *MRI=nullptr)
Prints virtual and physical registers with or without a TRI instance.
void eraseFromParent()
Unlink &#39;this&#39; from the containing basic block and delete it.
INITIALIZE_PASS(X86DomainReassignment, "x86-domain-reassignment", "X86 Domain Reassignment Pass", false, false) FunctionPass *llvm
Returns an instance of the Domain Reassignment pass.
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:408
const MCInstrDesc & getDesc() const
Returns the target instruction descriptor of this MachineInstr.
Definition: MachineInstr.h:405
static int getID(struct InternalInstruction *insn, const void *miiArg)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
TargetInstrInfo - Interface to description of machine instruction set.
MachineInstrBuilder BuildMI(MachineFunction &MF, const DebugLoc &DL, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:422
const TargetRegisterInfo * getTargetRegisterInfo() const
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
unsigned const MachineRegisterInfo * MRI
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - Subclasses that override getAnalysisUsage must call this.
MachineInstrBuilder & UseMI
std::pair< iterator, bool > insert(const ValueT &V)
Definition: DenseSet.h:187
Represent the analysis usage information of a pass.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
unsigned getOperandBias(const MCInstrDesc &Desc)
getOperandBias - compute whether all of the def operands are repeated in the uses and therefore shoul...
Definition: X86BaseInfo.h:655
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
void print(raw_ostream &OS, const SlotIndexes *=nullptr) const
print - Print out the MachineFunction in a format suitable for debugging to the specified stream...
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void initializeX86DomainReassignmentPass(PassRegistry &)
constexpr bool empty(const T &RangeOrContainer)
Test whether RangeOrContainer is empty. Similar to C++17 std::empty.
Definition: STLExtras.h:209
BlockVerifier::State From
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:839
MachineInstrBuilder MachineInstrBuilder & DefMI
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:373
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:285
const Function & getFunction() const
Return the LLVM function that this machine code represents.
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
A range adaptor for a pair of iterators.
iterator_range< mop_iterator > implicit_operands()
Definition: MachineInstr.h:472
const MachineBasicBlock * getParent() const
Definition: MachineInstr.h:253
MachineRegisterInfo - Keep track of information for virtual and physical registers, including vreg register classes, use/def chains for registers, etc.
Representation of each machine instruction.
Definition: MachineInstr.h:63
static bool isPhysicalRegister(unsigned Reg)
Return true if the specified register number is in the physical register namespace.
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
static cl::opt< bool > DisableX86DomainReassignment("disable-x86-domain-reassignment", cl::Hidden, cl::desc("X86: Disable Virtual Register Reassignment."), cl::init(false))
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
const MCInstrDesc & get(unsigned Opcode) const
Return the machine instruction descriptor that corresponds to the specified instruction opcode...
Definition: MCInstrInfo.h:44
#define I(x, y, z)
Definition: MD5.cpp:58
virtual void print(raw_ostream &OS, const Module *M) const
print - Print out the internal state of the pass.
Definition: Pass.cpp:123
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
IRTranslator LLVM IR MI
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
inst_range instructions(Function *F)
Definition: InstIterator.h:133
bool hasImplicitDefOfPhysReg(unsigned Reg, const MCRegisterInfo *MRI=nullptr) const
Return true if this instruction implicitly defines the specified physical register.
Definition: MCInstrDesc.cpp:44
#define LLVM_DEBUG(X)
Definition: Debug.h:122
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:413
unsigned createVirtualRegister(const TargetRegisterClass *RegClass, StringRef Name="")
createVirtualRegister - Create and return a new virtual register in the function with the specified r...
int getMemoryOperandNo(uint64_t TSFlags)
getMemoryOperandNo - The function returns the MCInst operand # for the first field of the memory oper...
Definition: X86BaseInfo.h:696