LLVM  9.0.0svn
X86FlagsCopyLowering.cpp
Go to the documentation of this file.
1 //====- X86FlagsCopyLowering.cpp - Lowers COPY nodes of EFLAGS ------------===//
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 /// \file
9 ///
10 /// Lowers COPY nodes of EFLAGS by directly extracting and preserving individual
11 /// flag bits.
12 ///
13 /// We have to do this by carefully analyzing and rewriting the usage of the
14 /// copied EFLAGS register because there is no general way to rematerialize the
15 /// entire EFLAGS register safely and efficiently. Using `popf` both forces
16 /// dynamic stack adjustment and can create correctness issues due to IF, TF,
17 /// and other non-status flags being overwritten. Using sequences involving
18 /// SAHF don't work on all x86 processors and are often quite slow compared to
19 /// directly testing a single status preserved in its own GPR.
20 ///
21 //===----------------------------------------------------------------------===//
22 
23 #include "X86.h"
24 #include "X86InstrBuilder.h"
25 #include "X86InstrInfo.h"
26 #include "X86Subtarget.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/ScopeExit.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallSet.h"
34 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/Statistic.h"
52 #include "llvm/IR/DebugLoc.h"
53 #include "llvm/MC/MCSchedule.h"
54 #include "llvm/Pass.h"
56 #include "llvm/Support/Debug.h"
58 #include <algorithm>
59 #include <cassert>
60 #include <iterator>
61 #include <utility>
62 
63 using namespace llvm;
64 
65 #define PASS_KEY "x86-flags-copy-lowering"
66 #define DEBUG_TYPE PASS_KEY
67 
68 STATISTIC(NumCopiesEliminated, "Number of copies of EFLAGS eliminated");
69 STATISTIC(NumSetCCsInserted, "Number of setCC instructions inserted");
70 STATISTIC(NumTestsInserted, "Number of test instructions inserted");
71 STATISTIC(NumAddsInserted, "Number of adds instructions inserted");
72 
73 namespace llvm {
74 
76 
77 } // end namespace llvm
78 
79 namespace {
80 
81 // Convenient array type for storing registers associated with each condition.
82 using CondRegArray = std::array<unsigned, X86::LAST_VALID_COND + 1>;
83 
84 class X86FlagsCopyLoweringPass : public MachineFunctionPass {
85 public:
86  X86FlagsCopyLoweringPass() : MachineFunctionPass(ID) {
88  }
89 
90  StringRef getPassName() const override { return "X86 EFLAGS copy lowering"; }
91  bool runOnMachineFunction(MachineFunction &MF) override;
92  void getAnalysisUsage(AnalysisUsage &AU) const override;
93 
94  /// Pass identification, replacement for typeid.
95  static char ID;
96 
97 private:
99  const X86Subtarget *Subtarget;
100  const X86InstrInfo *TII;
101  const TargetRegisterInfo *TRI;
102  const TargetRegisterClass *PromoteRC;
104 
105  CondRegArray collectCondsInRegs(MachineBasicBlock &MBB,
106  MachineBasicBlock::iterator CopyDefI);
107 
108  unsigned promoteCondToReg(MachineBasicBlock &MBB,
110  DebugLoc TestLoc, X86::CondCode Cond);
111  std::pair<unsigned, bool>
112  getCondOrInverseInReg(MachineBasicBlock &TestMBB,
113  MachineBasicBlock::iterator TestPos, DebugLoc TestLoc,
114  X86::CondCode Cond, CondRegArray &CondRegs);
115  void insertTest(MachineBasicBlock &MBB, MachineBasicBlock::iterator Pos,
116  DebugLoc Loc, unsigned Reg);
117 
118  void rewriteArithmetic(MachineBasicBlock &TestMBB,
119  MachineBasicBlock::iterator TestPos, DebugLoc TestLoc,
120  MachineInstr &MI, MachineOperand &FlagUse,
121  CondRegArray &CondRegs);
122  void rewriteCMov(MachineBasicBlock &TestMBB,
123  MachineBasicBlock::iterator TestPos, DebugLoc TestLoc,
124  MachineInstr &CMovI, MachineOperand &FlagUse,
125  CondRegArray &CondRegs);
126  void rewriteCondJmp(MachineBasicBlock &TestMBB,
127  MachineBasicBlock::iterator TestPos, DebugLoc TestLoc,
128  MachineInstr &JmpI, CondRegArray &CondRegs);
129  void rewriteCopy(MachineInstr &MI, MachineOperand &FlagUse,
130  MachineInstr &CopyDefI);
131  void rewriteSetCarryExtended(MachineBasicBlock &TestMBB,
133  DebugLoc TestLoc, MachineInstr &SetBI,
134  MachineOperand &FlagUse, CondRegArray &CondRegs);
135  void rewriteSetCC(MachineBasicBlock &TestMBB,
136  MachineBasicBlock::iterator TestPos, DebugLoc TestLoc,
137  MachineInstr &SetCCI, MachineOperand &FlagUse,
138  CondRegArray &CondRegs);
139 };
140 
141 } // end anonymous namespace
142 
143 INITIALIZE_PASS_BEGIN(X86FlagsCopyLoweringPass, DEBUG_TYPE,
144  "X86 EFLAGS copy lowering", false, false)
145 INITIALIZE_PASS_END(X86FlagsCopyLoweringPass, DEBUG_TYPE,
146  "X86 EFLAGS copy lowering", false, false)
147 
149  return new X86FlagsCopyLoweringPass();
150 }
151 
153 
154 void X86FlagsCopyLoweringPass::getAnalysisUsage(AnalysisUsage &AU) const {
157 }
158 
159 namespace {
160 /// An enumeration of the arithmetic instruction mnemonics which have
161 /// interesting flag semantics.
162 ///
163 /// We can map instruction opcodes into these mnemonics to make it easy to
164 /// dispatch with specific functionality.
165 enum class FlagArithMnemonic {
166  ADC,
167  ADCX,
168  ADOX,
169  RCL,
170  RCR,
171  SBB,
172 };
173 } // namespace
174 
175 static FlagArithMnemonic getMnemonicFromOpcode(unsigned Opcode) {
176  switch (Opcode) {
177  default:
178  report_fatal_error("No support for lowering a copy into EFLAGS when used "
179  "by this instruction!");
180 
181 #define LLVM_EXPAND_INSTR_SIZES(MNEMONIC, SUFFIX) \
182  case X86::MNEMONIC##8##SUFFIX: \
183  case X86::MNEMONIC##16##SUFFIX: \
184  case X86::MNEMONIC##32##SUFFIX: \
185  case X86::MNEMONIC##64##SUFFIX:
186 
187 #define LLVM_EXPAND_ADC_SBB_INSTR(MNEMONIC) \
188  LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr) \
189  LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr_REV) \
190  LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rm) \
191  LLVM_EXPAND_INSTR_SIZES(MNEMONIC, mr) \
192  case X86::MNEMONIC##8ri: \
193  case X86::MNEMONIC##16ri8: \
194  case X86::MNEMONIC##32ri8: \
195  case X86::MNEMONIC##64ri8: \
196  case X86::MNEMONIC##16ri: \
197  case X86::MNEMONIC##32ri: \
198  case X86::MNEMONIC##64ri32: \
199  case X86::MNEMONIC##8mi: \
200  case X86::MNEMONIC##16mi8: \
201  case X86::MNEMONIC##32mi8: \
202  case X86::MNEMONIC##64mi8: \
203  case X86::MNEMONIC##16mi: \
204  case X86::MNEMONIC##32mi: \
205  case X86::MNEMONIC##64mi32: \
206  case X86::MNEMONIC##8i8: \
207  case X86::MNEMONIC##16i16: \
208  case X86::MNEMONIC##32i32: \
209  case X86::MNEMONIC##64i32:
210 
212  return FlagArithMnemonic::ADC;
213 
215  return FlagArithMnemonic::SBB;
216 
217 #undef LLVM_EXPAND_ADC_SBB_INSTR
218 
219  LLVM_EXPAND_INSTR_SIZES(RCL, rCL)
221  LLVM_EXPAND_INSTR_SIZES(RCL, ri)
222  return FlagArithMnemonic::RCL;
223 
224  LLVM_EXPAND_INSTR_SIZES(RCR, rCL)
226  LLVM_EXPAND_INSTR_SIZES(RCR, ri)
227  return FlagArithMnemonic::RCR;
228 
229 #undef LLVM_EXPAND_INSTR_SIZES
230 
231  case X86::ADCX32rr:
232  case X86::ADCX64rr:
233  case X86::ADCX32rm:
234  case X86::ADCX64rm:
235  return FlagArithMnemonic::ADCX;
236 
237  case X86::ADOX32rr:
238  case X86::ADOX64rr:
239  case X86::ADOX32rm:
240  case X86::ADOX64rm:
241  return FlagArithMnemonic::ADOX;
242  }
243 }
244 
246  MachineInstr &SplitI,
247  const X86InstrInfo &TII) {
248  MachineFunction &MF = *MBB.getParent();
249 
250  assert(SplitI.getParent() == &MBB &&
251  "Split instruction must be in the split block!");
252  assert(SplitI.isBranch() &&
253  "Only designed to split a tail of branch instructions!");
255  "Must split on an actual jCC instruction!");
256 
257  // Dig out the previous instruction to the split point.
258  MachineInstr &PrevI = *std::prev(SplitI.getIterator());
259  assert(PrevI.isBranch() && "Must split after a branch!");
261  "Must split after an actual jCC instruction!");
262  assert(!std::prev(PrevI.getIterator())->isTerminator() &&
263  "Must only have this one terminator prior to the split!");
264 
265  // Grab the one successor edge that will stay in `MBB`.
266  MachineBasicBlock &UnsplitSucc = *PrevI.getOperand(0).getMBB();
267 
268  // Analyze the original block to see if we are actually splitting an edge
269  // into two edges. This can happen when we have multiple conditional jumps to
270  // the same successor.
271  bool IsEdgeSplit =
272  std::any_of(SplitI.getIterator(), MBB.instr_end(),
273  [&](MachineInstr &MI) {
274  assert(MI.isTerminator() &&
275  "Should only have spliced terminators!");
276  return llvm::any_of(
277  MI.operands(), [&](MachineOperand &MOp) {
278  return MOp.isMBB() && MOp.getMBB() == &UnsplitSucc;
279  });
280  }) ||
281  MBB.getFallThrough() == &UnsplitSucc;
282 
284 
285  // Insert the new block immediately after the current one. Any existing
286  // fallthrough will be sunk into this new block anyways.
287  MF.insert(std::next(MachineFunction::iterator(&MBB)), &NewMBB);
288 
289  // Splice the tail of instructions into the new block.
290  NewMBB.splice(NewMBB.end(), &MBB, SplitI.getIterator(), MBB.end());
291 
292  // Copy the necessary succesors (and their probability info) into the new
293  // block.
294  for (auto SI = MBB.succ_begin(), SE = MBB.succ_end(); SI != SE; ++SI)
295  if (IsEdgeSplit || *SI != &UnsplitSucc)
296  NewMBB.copySuccessor(&MBB, SI);
297  // Normalize the probabilities if we didn't end up splitting the edge.
298  if (!IsEdgeSplit)
299  NewMBB.normalizeSuccProbs();
300 
301  // Now replace all of the moved successors in the original block with the new
302  // block. This will merge their probabilities.
303  for (MachineBasicBlock *Succ : NewMBB.successors())
304  if (Succ != &UnsplitSucc)
305  MBB.replaceSuccessor(Succ, &NewMBB);
306 
307  // We should always end up replacing at least one successor.
308  assert(MBB.isSuccessor(&NewMBB) &&
309  "Failed to make the new block a successor!");
310 
311  // Now update all the PHIs.
312  for (MachineBasicBlock *Succ : NewMBB.successors()) {
313  for (MachineInstr &MI : *Succ) {
314  if (!MI.isPHI())
315  break;
316 
317  for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
318  OpIdx += 2) {
319  MachineOperand &OpV = MI.getOperand(OpIdx);
320  MachineOperand &OpMBB = MI.getOperand(OpIdx + 1);
321  assert(OpMBB.isMBB() && "Block operand to a PHI is not a block!");
322  if (OpMBB.getMBB() != &MBB)
323  continue;
324 
325  // Replace the operand for unsplit successors
326  if (!IsEdgeSplit || Succ != &UnsplitSucc) {
327  OpMBB.setMBB(&NewMBB);
328 
329  // We have to continue scanning as there may be multiple entries in
330  // the PHI.
331  continue;
332  }
333 
334  // When we have split the edge append a new successor.
335  MI.addOperand(MF, OpV);
336  MI.addOperand(MF, MachineOperand::CreateMBB(&NewMBB));
337  break;
338  }
339  }
340  }
341 
342  return NewMBB;
343 }
344 
345 bool X86FlagsCopyLoweringPass::runOnMachineFunction(MachineFunction &MF) {
346  LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
347  << " **********\n");
348 
349  Subtarget = &MF.getSubtarget<X86Subtarget>();
350  MRI = &MF.getRegInfo();
351  TII = Subtarget->getInstrInfo();
352  TRI = Subtarget->getRegisterInfo();
353  MDT = &getAnalysis<MachineDominatorTree>();
354  PromoteRC = &X86::GR8RegClass;
355 
356  if (MF.begin() == MF.end())
357  // Nothing to do for a degenerate empty function...
358  return false;
359 
360  // Collect the copies in RPO so that when there are chains where a copy is in
361  // turn copied again we visit the first one first. This ensures we can find
362  // viable locations for testing the original EFLAGS that dominate all the
363  // uses across complex CFGs.
366  for (MachineBasicBlock *MBB : RPOT)
367  for (MachineInstr &MI : *MBB)
368  if (MI.getOpcode() == TargetOpcode::COPY &&
369  MI.getOperand(0).getReg() == X86::EFLAGS)
370  Copies.push_back(&MI);
371 
372  for (MachineInstr *CopyI : Copies) {
373  MachineBasicBlock &MBB = *CopyI->getParent();
374 
375  MachineOperand &VOp = CopyI->getOperand(1);
376  assert(VOp.isReg() &&
377  "The input to the copy for EFLAGS should always be a register!");
378  MachineInstr &CopyDefI = *MRI->getVRegDef(VOp.getReg());
379  if (CopyDefI.getOpcode() != TargetOpcode::COPY) {
380  // FIXME: The big likely candidate here are PHI nodes. We could in theory
381  // handle PHI nodes, but it gets really, really hard. Insanely hard. Hard
382  // enough that it is probably better to change every other part of LLVM
383  // to avoid creating them. The issue is that once we have PHIs we won't
384  // know which original EFLAGS value we need to capture with our setCCs
385  // below. The end result will be computing a complete set of setCCs that
386  // we *might* want, computing them in every place where we copy *out* of
387  // EFLAGS and then doing SSA formation on all of them to insert necessary
388  // PHI nodes and consume those here. Then hoping that somehow we DCE the
389  // unnecessary ones. This DCE seems very unlikely to be successful and so
390  // we will almost certainly end up with a glut of dead setCC
391  // instructions. Until we have a motivating test case and fail to avoid
392  // it by changing other parts of LLVM's lowering, we refuse to handle
393  // this complex case here.
394  LLVM_DEBUG(
395  dbgs() << "ERROR: Encountered unexpected def of an eflags copy: ";
396  CopyDefI.dump());
398  "Cannot lower EFLAGS copy unless it is defined in turn by a copy!");
399  }
400 
401  auto Cleanup = make_scope_exit([&] {
402  // All uses of the EFLAGS copy are now rewritten, kill the copy into
403  // eflags and if dead the copy from.
404  CopyI->eraseFromParent();
405  if (MRI->use_empty(CopyDefI.getOperand(0).getReg()))
406  CopyDefI.eraseFromParent();
407  ++NumCopiesEliminated;
408  });
409 
410  MachineOperand &DOp = CopyI->getOperand(0);
411  assert(DOp.isDef() && "Expected register def!");
412  assert(DOp.getReg() == X86::EFLAGS && "Unexpected copy def register!");
413  if (DOp.isDead())
414  continue;
415 
416  MachineBasicBlock *TestMBB = CopyDefI.getParent();
417  auto TestPos = CopyDefI.getIterator();
418  DebugLoc TestLoc = CopyDefI.getDebugLoc();
419 
420  LLVM_DEBUG(dbgs() << "Rewriting copy: "; CopyI->dump());
421 
422  // Walk up across live-in EFLAGS to find where they were actually def'ed.
423  //
424  // This copy's def may just be part of a region of blocks covered by
425  // a single def of EFLAGS and we want to find the top of that region where
426  // possible.
427  //
428  // This is essentially a search for a *candidate* reaching definition
429  // location. We don't need to ever find the actual reaching definition here,
430  // but we want to walk up the dominator tree to find the highest point which
431  // would be viable for such a definition.
432  auto HasEFLAGSClobber = [&](MachineBasicBlock::iterator Begin,
434  // Scan backwards as we expect these to be relatively short and often find
435  // a clobber near the end.
436  return llvm::any_of(
437  llvm::reverse(llvm::make_range(Begin, End)), [&](MachineInstr &MI) {
438  // Flag any instruction (other than the copy we are
439  // currently rewriting) that defs EFLAGS.
440  return &MI != CopyI && MI.findRegisterDefOperand(X86::EFLAGS);
441  });
442  };
443  auto HasEFLAGSClobberPath = [&](MachineBasicBlock *BeginMBB,
444  MachineBasicBlock *EndMBB) {
445  assert(MDT->dominates(BeginMBB, EndMBB) &&
446  "Only support paths down the dominator tree!");
449  // We terminate at the beginning. No need to scan it.
450  Visited.insert(BeginMBB);
451  Worklist.push_back(EndMBB);
452  do {
453  auto *MBB = Worklist.pop_back_val();
454  for (auto *PredMBB : MBB->predecessors()) {
455  if (!Visited.insert(PredMBB).second)
456  continue;
457  if (HasEFLAGSClobber(PredMBB->begin(), PredMBB->end()))
458  return true;
459  // Enqueue this block to walk its predecessors.
460  Worklist.push_back(PredMBB);
461  }
462  } while (!Worklist.empty());
463  // No clobber found along a path from the begin to end.
464  return false;
465  };
466  while (TestMBB->isLiveIn(X86::EFLAGS) && !TestMBB->pred_empty() &&
467  !HasEFLAGSClobber(TestMBB->begin(), TestPos)) {
468  // Find the nearest common dominator of the predecessors, as
469  // that will be the best candidate to hoist into.
470  MachineBasicBlock *HoistMBB =
471  std::accumulate(std::next(TestMBB->pred_begin()), TestMBB->pred_end(),
472  *TestMBB->pred_begin(),
473  [&](MachineBasicBlock *LHS, MachineBasicBlock *RHS) {
474  return MDT->findNearestCommonDominator(LHS, RHS);
475  });
476 
477  // Now we need to scan all predecessors that may be reached along paths to
478  // the hoist block. A clobber anywhere in any of these blocks the hoist.
479  // Note that this even handles loops because we require *no* clobbers.
480  if (HasEFLAGSClobberPath(HoistMBB, TestMBB))
481  break;
482 
483  // We also need the terminators to not sneakily clobber flags.
484  if (HasEFLAGSClobber(HoistMBB->getFirstTerminator()->getIterator(),
485  HoistMBB->instr_end()))
486  break;
487 
488  // We found a viable location, hoist our test position to it.
489  TestMBB = HoistMBB;
490  TestPos = TestMBB->getFirstTerminator()->getIterator();
491  // Clear the debug location as it would just be confusing after hoisting.
492  TestLoc = DebugLoc();
493  }
494  LLVM_DEBUG({
495  auto DefIt = llvm::find_if(
496  llvm::reverse(llvm::make_range(TestMBB->instr_begin(), TestPos)),
497  [&](MachineInstr &MI) {
498  return MI.findRegisterDefOperand(X86::EFLAGS);
499  });
500  if (DefIt.base() != TestMBB->instr_begin()) {
501  dbgs() << " Using EFLAGS defined by: ";
502  DefIt->dump();
503  } else {
504  dbgs() << " Using live-in flags for BB:\n";
505  TestMBB->dump();
506  }
507  });
508 
509  // While rewriting uses, we buffer jumps and rewrite them in a second pass
510  // because doing so will perturb the CFG that we are walking to find the
511  // uses in the first place.
513 
514  // Gather the condition flags that have already been preserved in
515  // registers. We do this from scratch each time as we expect there to be
516  // very few of them and we expect to not revisit the same copy definition
517  // many times. If either of those change sufficiently we could build a map
518  // of these up front instead.
519  CondRegArray CondRegs = collectCondsInRegs(*TestMBB, TestPos);
520 
521  // Collect the basic blocks we need to scan. Typically this will just be
522  // a single basic block but we may have to scan multiple blocks if the
523  // EFLAGS copy lives into successors.
526  Blocks.push_back(&MBB);
527 
528  do {
529  MachineBasicBlock &UseMBB = *Blocks.pop_back_val();
530 
531  // Track when if/when we find a kill of the flags in this block.
532  bool FlagsKilled = false;
533 
534  // In most cases, we walk from the beginning to the end of the block. But
535  // when the block is the same block as the copy is from, we will visit it
536  // twice. The first time we start from the copy and go to the end. The
537  // second time we start from the beginning and go to the copy. This lets
538  // us handle copies inside of cycles.
539  // FIXME: This loop is *super* confusing. This is at least in part
540  // a symptom of all of this routine needing to be refactored into
541  // documentable components. Once done, there may be a better way to write
542  // this loop.
543  for (auto MII = (&UseMBB == &MBB && !VisitedBlocks.count(&UseMBB))
544  ? std::next(CopyI->getIterator())
545  : UseMBB.instr_begin(),
546  MIE = UseMBB.instr_end();
547  MII != MIE;) {
548  MachineInstr &MI = *MII++;
549  // If we are in the original copy block and encounter either the copy
550  // def or the copy itself, break so that we don't re-process any part of
551  // the block or process the instructions in the range that was copied
552  // over.
553  if (&MI == CopyI || &MI == &CopyDefI) {
554  assert(&UseMBB == &MBB && VisitedBlocks.count(&MBB) &&
555  "Should only encounter these on the second pass over the "
556  "original block.");
557  break;
558  }
559 
560  MachineOperand *FlagUse = MI.findRegisterUseOperand(X86::EFLAGS);
561  if (!FlagUse) {
562  if (MI.findRegisterDefOperand(X86::EFLAGS)) {
563  // If EFLAGS are defined, it's as-if they were killed. We can stop
564  // scanning here.
565  //
566  // NB!!! Many instructions only modify some flags. LLVM currently
567  // models this as clobbering all flags, but if that ever changes
568  // this will need to be carefully updated to handle that more
569  // complex logic.
570  FlagsKilled = true;
571  break;
572  }
573  continue;
574  }
575 
576  LLVM_DEBUG(dbgs() << " Rewriting use: "; MI.dump());
577 
578  // Check the kill flag before we rewrite as that may change it.
579  if (FlagUse->isKill())
580  FlagsKilled = true;
581 
582  // Once we encounter a branch, the rest of the instructions must also be
583  // branches. We can't rewrite in place here, so we handle them below.
584  //
585  // Note that we don't have to handle tail calls here, even conditional
586  // tail calls, as those are not introduced into the X86 MI until post-RA
587  // branch folding or black placement. As a consequence, we get to deal
588  // with the simpler formulation of conditional branches followed by tail
589  // calls.
591  auto JmpIt = MI.getIterator();
592  do {
593  JmpIs.push_back(&*JmpIt);
594  ++JmpIt;
595  } while (JmpIt != UseMBB.instr_end() &&
596  X86::getCondFromBranch(*JmpIt) !=
598  break;
599  }
600 
601  // Otherwise we can just rewrite in-place.
603  rewriteCMov(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs);
604  } else if (X86::getCondFromSETCC(MI) != X86::COND_INVALID) {
605  rewriteSetCC(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs);
606  } else if (MI.getOpcode() == TargetOpcode::COPY) {
607  rewriteCopy(MI, *FlagUse, CopyDefI);
608  } else {
609  // We assume all other instructions that use flags also def them.
610  assert(MI.findRegisterDefOperand(X86::EFLAGS) &&
611  "Expected a def of EFLAGS for this instruction!");
612 
613  // NB!!! Several arithmetic instructions only *partially* update
614  // flags. Theoretically, we could generate MI code sequences that
615  // would rely on this fact and observe different flags independently.
616  // But currently LLVM models all of these instructions as clobbering
617  // all the flags in an undef way. We rely on that to simplify the
618  // logic.
619  FlagsKilled = true;
620 
621  switch (MI.getOpcode()) {
622  case X86::SETB_C8r:
623  case X86::SETB_C16r:
624  case X86::SETB_C32r:
625  case X86::SETB_C64r:
626  // Use custom lowering for arithmetic that is merely extending the
627  // carry flag. We model this as the SETB_C* pseudo instructions.
628  rewriteSetCarryExtended(*TestMBB, TestPos, TestLoc, MI, *FlagUse,
629  CondRegs);
630  break;
631 
632  default:
633  // Generically handle remaining uses as arithmetic instructions.
634  rewriteArithmetic(*TestMBB, TestPos, TestLoc, MI, *FlagUse,
635  CondRegs);
636  break;
637  }
638  break;
639  }
640 
641  // If this was the last use of the flags, we're done.
642  if (FlagsKilled)
643  break;
644  }
645 
646  // If the flags were killed, we're done with this block.
647  if (FlagsKilled)
648  continue;
649 
650  // Otherwise we need to scan successors for ones where the flags live-in
651  // and queue those up for processing.
652  for (MachineBasicBlock *SuccMBB : UseMBB.successors())
653  if (SuccMBB->isLiveIn(X86::EFLAGS) &&
654  VisitedBlocks.insert(SuccMBB).second) {
655  // We currently don't do any PHI insertion and so we require that the
656  // test basic block dominates all of the use basic blocks. Further, we
657  // can't have a cycle from the test block back to itself as that would
658  // create a cycle requiring a PHI to break it.
659  //
660  // We could in theory do PHI insertion here if it becomes useful by
661  // just taking undef values in along every edge that we don't trace
662  // this EFLAGS copy along. This isn't as bad as fully general PHI
663  // insertion, but still seems like a great deal of complexity.
664  //
665  // Because it is theoretically possible that some earlier MI pass or
666  // other lowering transformation could induce this to happen, we do
667  // a hard check even in non-debug builds here.
668  if (SuccMBB == TestMBB || !MDT->dominates(TestMBB, SuccMBB)) {
669  LLVM_DEBUG({
670  dbgs()
671  << "ERROR: Encountered use that is not dominated by our test "
672  "basic block! Rewriting this would require inserting PHI "
673  "nodes to track the flag state across the CFG.\n\nTest "
674  "block:\n";
675  TestMBB->dump();
676  dbgs() << "Use block:\n";
677  SuccMBB->dump();
678  });
680  "Cannot lower EFLAGS copy when original copy def "
681  "does not dominate all uses.");
682  }
683 
684  Blocks.push_back(SuccMBB);
685  }
686  } while (!Blocks.empty());
687 
688  // Now rewrite the jumps that use the flags. These we handle specially
689  // because if there are multiple jumps in a single basic block we'll have
690  // to do surgery on the CFG.
691  MachineBasicBlock *LastJmpMBB = nullptr;
692  for (MachineInstr *JmpI : JmpIs) {
693  // Past the first jump within a basic block we need to split the blocks
694  // apart.
695  if (JmpI->getParent() == LastJmpMBB)
696  splitBlock(*JmpI->getParent(), *JmpI, *TII);
697  else
698  LastJmpMBB = JmpI->getParent();
699 
700  rewriteCondJmp(*TestMBB, TestPos, TestLoc, *JmpI, CondRegs);
701  }
702 
703  // FIXME: Mark the last use of EFLAGS before the copy's def as a kill if
704  // the copy's def operand is itself a kill.
705  }
706 
707 #ifndef NDEBUG
708  for (MachineBasicBlock &MBB : MF)
709  for (MachineInstr &MI : MBB)
710  if (MI.getOpcode() == TargetOpcode::COPY &&
711  (MI.getOperand(0).getReg() == X86::EFLAGS ||
712  MI.getOperand(1).getReg() == X86::EFLAGS)) {
713  LLVM_DEBUG(dbgs() << "ERROR: Found a COPY involving EFLAGS: ";
714  MI.dump());
715  llvm_unreachable("Unlowered EFLAGS copy!");
716  }
717 #endif
718 
719  return true;
720 }
721 
722 /// Collect any conditions that have already been set in registers so that we
723 /// can re-use them rather than adding duplicates.
724 CondRegArray X86FlagsCopyLoweringPass::collectCondsInRegs(
726  CondRegArray CondRegs = {};
727 
728  // Scan backwards across the range of instructions with live EFLAGS.
729  for (MachineInstr &MI :
730  llvm::reverse(llvm::make_range(MBB.begin(), TestPos))) {
732  if (Cond != X86::COND_INVALID && !MI.mayStore() && MI.getOperand(0).isReg() &&
733  TRI->isVirtualRegister(MI.getOperand(0).getReg())) {
734  assert(MI.getOperand(0).isDef() &&
735  "A non-storing SETcc should always define a register!");
736  CondRegs[Cond] = MI.getOperand(0).getReg();
737  }
738 
739  // Stop scanning when we see the first definition of the EFLAGS as prior to
740  // this we would potentially capture the wrong flag state.
741  if (MI.findRegisterDefOperand(X86::EFLAGS))
742  break;
743  }
744  return CondRegs;
745 }
746 
747 unsigned X86FlagsCopyLoweringPass::promoteCondToReg(
749  DebugLoc TestLoc, X86::CondCode Cond) {
750  unsigned Reg = MRI->createVirtualRegister(PromoteRC);
751  auto SetI = BuildMI(TestMBB, TestPos, TestLoc,
752  TII->get(X86::SETCCr), Reg).addImm(Cond);
753  (void)SetI;
754  LLVM_DEBUG(dbgs() << " save cond: "; SetI->dump());
755  ++NumSetCCsInserted;
756  return Reg;
757 }
758 
759 std::pair<unsigned, bool> X86FlagsCopyLoweringPass::getCondOrInverseInReg(
761  DebugLoc TestLoc, X86::CondCode Cond, CondRegArray &CondRegs) {
762  unsigned &CondReg = CondRegs[Cond];
763  unsigned &InvCondReg = CondRegs[X86::GetOppositeBranchCondition(Cond)];
764  if (!CondReg && !InvCondReg)
765  CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
766 
767  if (CondReg)
768  return {CondReg, false};
769  else
770  return {InvCondReg, true};
771 }
772 
773 void X86FlagsCopyLoweringPass::insertTest(MachineBasicBlock &MBB,
775  DebugLoc Loc, unsigned Reg) {
776  auto TestI =
777  BuildMI(MBB, Pos, Loc, TII->get(X86::TEST8rr)).addReg(Reg).addReg(Reg);
778  (void)TestI;
779  LLVM_DEBUG(dbgs() << " test cond: "; TestI->dump());
780  ++NumTestsInserted;
781 }
782 
783 void X86FlagsCopyLoweringPass::rewriteArithmetic(
785  DebugLoc TestLoc, MachineInstr &MI, MachineOperand &FlagUse,
786  CondRegArray &CondRegs) {
787  // Arithmetic is either reading CF or OF. Figure out which condition we need
788  // to preserve in a register.
789  X86::CondCode Cond;
790 
791  // The addend to use to reset CF or OF when added to the flag value.
792  int Addend;
793 
794  switch (getMnemonicFromOpcode(MI.getOpcode())) {
796  case FlagArithMnemonic::ADCX:
797  case FlagArithMnemonic::RCL:
798  case FlagArithMnemonic::RCR:
800  Cond = X86::COND_B; // CF == 1
801  // Set up an addend that when one is added will need a carry due to not
802  // having a higher bit available.
803  Addend = 255;
804  break;
805 
806  case FlagArithMnemonic::ADOX:
807  Cond = X86::COND_O; // OF == 1
808  // Set up an addend that when one is added will turn from positive to
809  // negative and thus overflow in the signed domain.
810  Addend = 127;
811  break;
812  }
813 
814  // Now get a register that contains the value of the flag input to the
815  // arithmetic. We require exactly this flag to simplify the arithmetic
816  // required to materialize it back into the flag.
817  unsigned &CondReg = CondRegs[Cond];
818  if (!CondReg)
819  CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
820 
821  MachineBasicBlock &MBB = *MI.getParent();
822 
823  // Insert an instruction that will set the flag back to the desired value.
824  unsigned TmpReg = MRI->createVirtualRegister(PromoteRC);
825  auto AddI =
826  BuildMI(MBB, MI.getIterator(), MI.getDebugLoc(), TII->get(X86::ADD8ri))
827  .addDef(TmpReg, RegState::Dead)
828  .addReg(CondReg)
829  .addImm(Addend);
830  (void)AddI;
831  LLVM_DEBUG(dbgs() << " add cond: "; AddI->dump());
832  ++NumAddsInserted;
833  FlagUse.setIsKill(true);
834 }
835 
836 void X86FlagsCopyLoweringPass::rewriteCMov(MachineBasicBlock &TestMBB,
838  DebugLoc TestLoc,
839  MachineInstr &CMovI,
840  MachineOperand &FlagUse,
841  CondRegArray &CondRegs) {
842  // First get the register containing this specific condition.
843  X86::CondCode Cond = X86::getCondFromCMov(CMovI);
844  unsigned CondReg;
845  bool Inverted;
846  std::tie(CondReg, Inverted) =
847  getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs);
848 
849  MachineBasicBlock &MBB = *CMovI.getParent();
850 
851  // Insert a direct test of the saved register.
852  insertTest(MBB, CMovI.getIterator(), CMovI.getDebugLoc(), CondReg);
853 
854  // Rewrite the CMov to use the !ZF flag from the test, and then kill its use
855  // of the flags afterward.
856  CMovI.getOperand(CMovI.getDesc().getNumOperands() - 1)
857  .setImm(Inverted ? X86::COND_E : X86::COND_NE);
858  FlagUse.setIsKill(true);
859  LLVM_DEBUG(dbgs() << " fixed cmov: "; CMovI.dump());
860 }
861 
862 void X86FlagsCopyLoweringPass::rewriteCondJmp(
864  DebugLoc TestLoc, MachineInstr &JmpI, CondRegArray &CondRegs) {
865  // First get the register containing this specific condition.
867  unsigned CondReg;
868  bool Inverted;
869  std::tie(CondReg, Inverted) =
870  getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs);
871 
872  MachineBasicBlock &JmpMBB = *JmpI.getParent();
873 
874  // Insert a direct test of the saved register.
875  insertTest(JmpMBB, JmpI.getIterator(), JmpI.getDebugLoc(), CondReg);
876 
877  // Rewrite the jump to use the !ZF flag from the test, and kill its use of
878  // flags afterward.
879  JmpI.getOperand(1).setImm(Inverted ? X86::COND_E : X86::COND_NE);
880  JmpI.findRegisterUseOperand(X86::EFLAGS)->setIsKill(true);
881  LLVM_DEBUG(dbgs() << " fixed jCC: "; JmpI.dump());
882 }
883 
884 void X86FlagsCopyLoweringPass::rewriteCopy(MachineInstr &MI,
885  MachineOperand &FlagUse,
886  MachineInstr &CopyDefI) {
887  // Just replace this copy with the original copy def.
888  MRI->replaceRegWith(MI.getOperand(0).getReg(),
889  CopyDefI.getOperand(0).getReg());
890  MI.eraseFromParent();
891 }
892 
893 void X86FlagsCopyLoweringPass::rewriteSetCarryExtended(
895  DebugLoc TestLoc, MachineInstr &SetBI, MachineOperand &FlagUse,
896  CondRegArray &CondRegs) {
897  // This routine is only used to handle pseudos for setting a register to zero
898  // or all ones based on CF. This is essentially the sign extended from 1-bit
899  // form of SETB and modeled with the SETB_C* pseudos. They require special
900  // handling as they aren't normal SETcc instructions and are lowered to an
901  // EFLAGS clobbering operation (SBB typically). One simplifying aspect is that
902  // they are only provided in reg-defining forms. A complicating factor is that
903  // they can define many different register widths.
904  assert(SetBI.getOperand(0).isReg() &&
905  "Cannot have a non-register defined operand to this variant of SETB!");
906 
907  // Little helper to do the common final step of replacing the register def'ed
908  // by this SETB instruction with a new register and removing the SETB
909  // instruction.
910  auto RewriteToReg = [&](unsigned Reg) {
911  MRI->replaceRegWith(SetBI.getOperand(0).getReg(), Reg);
912  SetBI.eraseFromParent();
913  };
914 
915  // Grab the register class used for this particular instruction.
916  auto &SetBRC = *MRI->getRegClass(SetBI.getOperand(0).getReg());
917 
918  MachineBasicBlock &MBB = *SetBI.getParent();
919  auto SetPos = SetBI.getIterator();
920  auto SetLoc = SetBI.getDebugLoc();
921 
922  auto AdjustReg = [&](unsigned Reg) {
923  auto &OrigRC = *MRI->getRegClass(Reg);
924  if (&OrigRC == &SetBRC)
925  return Reg;
926 
927  unsigned NewReg;
928 
929  int OrigRegSize = TRI->getRegSizeInBits(OrigRC) / 8;
930  int TargetRegSize = TRI->getRegSizeInBits(SetBRC) / 8;
931  assert(OrigRegSize <= 8 && "No GPRs larger than 64-bits!");
932  assert(TargetRegSize <= 8 && "No GPRs larger than 64-bits!");
933  int SubRegIdx[] = {X86::NoSubRegister, X86::sub_8bit, X86::sub_16bit,
934  X86::NoSubRegister, X86::sub_32bit};
935 
936  // If the original size is smaller than the target *and* is smaller than 4
937  // bytes, we need to explicitly zero extend it. We always extend to 4-bytes
938  // to maximize the chance of being able to CSE that operation and to avoid
939  // partial dependency stalls extending to 2-bytes.
940  if (OrigRegSize < TargetRegSize && OrigRegSize < 4) {
941  NewReg = MRI->createVirtualRegister(&X86::GR32RegClass);
942  BuildMI(MBB, SetPos, SetLoc, TII->get(X86::MOVZX32rr8), NewReg)
943  .addReg(Reg);
944  if (&SetBRC == &X86::GR32RegClass)
945  return NewReg;
946  Reg = NewReg;
947  OrigRegSize = 4;
948  }
949 
950  NewReg = MRI->createVirtualRegister(&SetBRC);
951  if (OrigRegSize < TargetRegSize) {
952  BuildMI(MBB, SetPos, SetLoc, TII->get(TargetOpcode::SUBREG_TO_REG),
953  NewReg)
954  .addImm(0)
955  .addReg(Reg)
956  .addImm(SubRegIdx[OrigRegSize]);
957  } else if (OrigRegSize > TargetRegSize) {
958  if (TargetRegSize == 1 && !Subtarget->is64Bit()) {
959  // Need to constrain the register class.
960  MRI->constrainRegClass(Reg, &X86::GR32_ABCDRegClass);
961  }
962 
963  BuildMI(MBB, SetPos, SetLoc, TII->get(TargetOpcode::COPY),
964  NewReg)
965  .addReg(Reg, 0, SubRegIdx[TargetRegSize]);
966  } else {
967  BuildMI(MBB, SetPos, SetLoc, TII->get(TargetOpcode::COPY), NewReg)
968  .addReg(Reg);
969  }
970  return NewReg;
971  };
972 
973  unsigned &CondReg = CondRegs[X86::COND_B];
974  if (!CondReg)
975  CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, X86::COND_B);
976 
977  // Adjust the condition to have the desired register width by zero-extending
978  // as needed.
979  // FIXME: We should use a better API to avoid the local reference and using a
980  // different variable here.
981  unsigned ExtCondReg = AdjustReg(CondReg);
982 
983  // Now we need to turn this into a bitmask. We do this by subtracting it from
984  // zero.
985  unsigned ZeroReg = MRI->createVirtualRegister(&X86::GR32RegClass);
986  BuildMI(MBB, SetPos, SetLoc, TII->get(X86::MOV32r0), ZeroReg);
987  ZeroReg = AdjustReg(ZeroReg);
988 
989  unsigned Sub;
990  switch (SetBI.getOpcode()) {
991  case X86::SETB_C8r:
992  Sub = X86::SUB8rr;
993  break;
994 
995  case X86::SETB_C16r:
996  Sub = X86::SUB16rr;
997  break;
998 
999  case X86::SETB_C32r:
1000  Sub = X86::SUB32rr;
1001  break;
1002 
1003  case X86::SETB_C64r:
1004  Sub = X86::SUB64rr;
1005  break;
1006 
1007  default:
1008  llvm_unreachable("Invalid SETB_C* opcode!");
1009  }
1010  unsigned ResultReg = MRI->createVirtualRegister(&SetBRC);
1011  BuildMI(MBB, SetPos, SetLoc, TII->get(Sub), ResultReg)
1012  .addReg(ZeroReg)
1013  .addReg(ExtCondReg);
1014  return RewriteToReg(ResultReg);
1015 }
1016 
1017 void X86FlagsCopyLoweringPass::rewriteSetCC(MachineBasicBlock &TestMBB,
1019  DebugLoc TestLoc,
1020  MachineInstr &SetCCI,
1021  MachineOperand &FlagUse,
1022  CondRegArray &CondRegs) {
1023  X86::CondCode Cond = X86::getCondFromSETCC(SetCCI);
1024  // Note that we can't usefully rewrite this to the inverse without complex
1025  // analysis of the users of the setCC. Largely we rely on duplicates which
1026  // could have been avoided already being avoided here.
1027  unsigned &CondReg = CondRegs[Cond];
1028  if (!CondReg)
1029  CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
1030 
1031  // Rewriting a register def is trivial: we just replace the register and
1032  // remove the setcc.
1033  if (!SetCCI.mayStore()) {
1034  assert(SetCCI.getOperand(0).isReg() &&
1035  "Cannot have a non-register defined operand to SETcc!");
1036  MRI->replaceRegWith(SetCCI.getOperand(0).getReg(), CondReg);
1037  SetCCI.eraseFromParent();
1038  return;
1039  }
1040 
1041  // Otherwise, we need to emit a store.
1042  auto MIB = BuildMI(*SetCCI.getParent(), SetCCI.getIterator(),
1043  SetCCI.getDebugLoc(), TII->get(X86::MOV8mr));
1044  // Copy the address operands.
1045  for (int i = 0; i < X86::AddrNumOperands; ++i)
1046  MIB.add(SetCCI.getOperand(i));
1047 
1048  MIB.addReg(CondReg);
1049 
1050  MIB.setMemRefs(SetCCI.memoperands());
1051 
1052  SetCCI.eraseFromParent();
1053  return;
1054 }
const MachineInstrBuilder & setMemRefs(ArrayRef< MachineMemOperand *> MMOs) const
const MachineInstrBuilder & add(const MachineOperand &MO) const
MachineOperand * findRegisterDefOperand(unsigned Reg, bool isDead=false, bool Overlap=false, const TargetRegisterInfo *TRI=nullptr)
Wrapper for findRegisterDefOperandIdx, it returns a pointer to the MachineOperand rather than an inde...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
instr_iterator instr_begin()
instr_iterator instr_end()
MachineBasicBlock * getMBB() const
LLVM_ATTRIBUTE_NORETURN void report_fatal_error(Error Err, bool gen_crash_diag=true)
Report a serious error, calling any installed error handler.
Definition: Error.cpp:139
This class represents lattice values for constants.
Definition: AllocatorList.h:23
X86 EFLAGS copy lowering
void push_back(const T &Elt)
Definition: SmallVector.h:211
CondCode getCondFromCMov(const MachineInstr &MI)
Return condition code of a CMov opcode.
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition: MachineInstr.h:382
iterator getFirstTerminator()
Returns an iterator to the first terminator instruction of this basic block.
unsigned getReg() const
getReg - Returns the register number.
MachineOperand * findRegisterUseOperand(unsigned Reg, bool isKill=false, const TargetRegisterInfo *TRI=nullptr)
Wrapper for findRegisterUseOperandIdx, it returns a pointer to the MachineOperand rather than an inde...
unsigned Reg
LLVM_NODISCARD detail::scope_exit< typename std::decay< Callable >::type > make_scope_exit(Callable &&F)
Definition: ScopeExit.h:58
STATISTIC(NumFunctions, "Total number of functions")
unsigned const TargetRegisterInfo * TRI
A debug info location.
Definition: DebugLoc.h:33
iterator_range< succ_iterator > successors()
AnalysisUsage & addRequired()
This file declares the MachineConstantPool class which is an abstract constant pool to keep track of ...
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
unsigned getNumOperands() const
Return the number of declared MachineOperands for this MachineInstruction.
Definition: MCInstrDesc.h:210
const HexagonInstrInfo * TII
MachineBasicBlock * getFallThrough()
Return the fallthrough block if the block can implicitly transfer control to the block after it by fa...
void eraseFromParent()
Unlink &#39;this&#39; from the containing basic block and delete it.
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:408
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:266
const MCInstrDesc & getDesc() const
Returns the target instruction descriptor of this MachineInstr.
Definition: MachineInstr.h:405
void dump() const
Definition: Pass.cpp:134
MachineBasicBlock * CreateMachineBasicBlock(const BasicBlock *bb=nullptr)
CreateMachineBasicBlock - Allocate a new MachineBasicBlock.
bool isBranch(QueryType Type=AnyInBundle) const
Returns true if this is a conditional, unconditional, or indirect branch.
Definition: MachineInstr.h:656
void normalizeSuccProbs()
Normalize probabilities of all successors so that the sum of them becomes one.
StringRef getName() const
getName - Return the name of the corresponding LLVM function.
FlagArithMnemonic
An enumeration of the arithmetic instruction mnemonics which have interesting flag semantics...
MachineInstrBuilder BuildMI(MachineFunction &MF, const DebugLoc &DL, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
bool mayStore(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly modify memory.
Definition: MachineInstr.h:819
unsigned const MachineRegisterInfo * MRI
ArrayRef< MachineMemOperand * > memoperands() const
Access to memory operands of the instruction.
Definition: MachineInstr.h:515
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.
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:370
void setMBB(MachineBasicBlock *MBB)
#define LLVM_EXPAND_ADC_SBB_INSTR(MNEMONIC)
Represent the analysis usage information of a pass.
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1192
CondCode getCondFromSETCC(const MachineInstr &MI)
Return condition code of a SETCC opcode.
void setImm(int64_t immVal)
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:381
self_iterator getIterator()
Definition: ilist_node.h:81
iterator_range< pred_iterator > predecessors()
auto find_if(R &&Range, UnaryPredicate P) -> decltype(adl_begin(Range))
Provide wrappers to std::find_if which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1213
static void r1(uint32_t &A, uint32_t &B, uint32_t &C, uint32_t &D, uint32_t &E, int I, uint32_t *Buf)
Definition: SHA1.cpp:48
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void setIsKill(bool Val=true)
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
CondCode getCondFromBranch(const MachineInstr &MI)
Iterator for intrusive lists based on ilist_node.
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
MachineOperand class - Representation of each machine instruction operand.
CondCode GetOppositeBranchCondition(CondCode CC)
GetOppositeBranchCondition - Return the inverse of the specified cond, e.g.
SI Lower i1 Copies
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
FunctionPass * createX86FlagsCopyLoweringPass()
Return a pass that lowers EFLAGS copy pseudo instructions.
INITIALIZE_PASS_BEGIN(X86FlagsCopyLoweringPass, DEBUG_TYPE, "X86 EFLAGS copy lowering", false, false) INITIALIZE_PASS_END(X86FlagsCopyLoweringPass
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
bool isLiveIn(MCPhysReg Reg, LaneBitmask LaneMask=LaneBitmask::getAll()) const
Return true if the specified register is in the live in set.
void replaceSuccessor(MachineBasicBlock *Old, MachineBasicBlock *New)
Replace successor OLD with NEW and update probability info.
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:471
void copySuccessor(MachineBasicBlock *Orig, succ_iterator I)
Copy a successor (and any probability info) from original block to this block&#39;s.
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 MachineBasicBlock & splitBlock(MachineBasicBlock &MBB, MachineInstr &SplitI, const X86InstrInfo &TII)
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
static MachineOperand CreateMBB(MachineBasicBlock *MBB, unsigned char TargetFlags=0)
void splice(iterator Where, MachineBasicBlock *Other, iterator From)
Take an instruction from MBB &#39;Other&#39; at the position From, and insert it into this MBB right before &#39;...
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
#define LLVM_EXPAND_INSTR_SIZES(MNEMONIC, SUFFIX)
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
void initializeX86FlagsCopyLoweringPassPass(PassRegistry &)
static FlagArithMnemonic getMnemonicFromOpcode(unsigned Opcode)
const MachineInstrBuilder & addReg(unsigned RegNo, unsigned flags=0, unsigned SubReg=0) const
Add a new virtual register operand.
bool isSuccessor(const MachineBasicBlock *MBB) const
Return true if the specified MBB is a successor of this block.
bool isReg() const
isReg - Tests if this is a MO_Register operand.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
void insert(iterator MBBI, MachineBasicBlock *MBB)
IRTranslator LLVM IR MI
#define DEBUG_TYPE
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
PassRegistry - This class manages the registration and intitialization of the pass subsystem as appli...
Definition: PassRegistry.h:38
bool isMBB() const
isMBB - Tests if this is a MO_MachineBasicBlock operand.
#define LLVM_DEBUG(X)
Definition: Debug.h:122
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:413
OutputIt copy(R &&Range, OutputIt Out)
Definition: STLExtras.h:1237
DominatorTree Class - Concrete subclass of DominatorTreeBase that is used to compute a normal dominat...
AddrNumOperands - Total number of operands in a memory reference.
Definition: X86BaseInfo.h:41