LLVM  9.0.0svn
X86CmovConversion.cpp
Go to the documentation of this file.
1 //====- X86CmovConversion.cpp - Convert Cmov to Branch --------------------===//
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 /// \file
10 /// This file implements a pass that converts X86 cmov instructions into
11 /// branches when profitable. This pass is conservative. It transforms if and
12 /// only if it can guarantee a gain with high confidence.
13 ///
14 /// Thus, the optimization applies under the following conditions:
15 /// 1. Consider as candidates only CMOVs in innermost loops (assume that
16 /// most hotspots are represented by these loops).
17 /// 2. Given a group of CMOV instructions that are using the same EFLAGS def
18 /// instruction:
19 /// a. Consider them as candidates only if all have the same code condition
20 /// or the opposite one to prevent generating more than one conditional
21 /// jump per EFLAGS def instruction.
22 /// b. Consider them as candidates only if all are profitable to be
23 /// converted (assume that one bad conversion may cause a degradation).
24 /// 3. Apply conversion only for loops that are found profitable and only for
25 /// CMOV candidates that were found profitable.
26 /// a. A loop is considered profitable only if conversion will reduce its
27 /// depth cost by some threshold.
28 /// b. CMOV is considered profitable if the cost of its condition is higher
29 /// than the average cost of its true-value and false-value by 25% of
30 /// branch-misprediction-penalty. This assures no degradation even with
31 /// 25% branch misprediction.
32 ///
33 /// Note: This pass is assumed to run on SSA machine code.
34 //
35 //===----------------------------------------------------------------------===//
36 //
37 // External interfaces:
38 // FunctionPass *llvm::createX86CmovConverterPass();
39 // bool X86CmovConverterPass::runOnMachineFunction(MachineFunction &MF);
40 //
41 //===----------------------------------------------------------------------===//
42 
43 #include "X86.h"
44 #include "X86InstrInfo.h"
45 #include "llvm/ADT/ArrayRef.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/STLExtras.h"
48 #include "llvm/ADT/SmallPtrSet.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
63 #include "llvm/IR/DebugLoc.h"
64 #include "llvm/MC/MCSchedule.h"
65 #include "llvm/Pass.h"
67 #include "llvm/Support/Debug.h"
69 #include <algorithm>
70 #include <cassert>
71 #include <iterator>
72 #include <utility>
73 
74 using namespace llvm;
75 
76 #define DEBUG_TYPE "x86-cmov-conversion"
77 
78 STATISTIC(NumOfSkippedCmovGroups, "Number of unsupported CMOV-groups");
79 STATISTIC(NumOfCmovGroupCandidate, "Number of CMOV-group candidates");
80 STATISTIC(NumOfLoopCandidate, "Number of CMOV-conversion profitable loops");
81 STATISTIC(NumOfOptimizedCmovGroups, "Number of optimized CMOV-groups");
82 
83 // This internal switch can be used to turn off the cmov/branch optimization.
84 static cl::opt<bool>
85  EnableCmovConverter("x86-cmov-converter",
86  cl::desc("Enable the X86 cmov-to-branch optimization."),
87  cl::init(true), cl::Hidden);
88 
89 static cl::opt<unsigned>
90  GainCycleThreshold("x86-cmov-converter-threshold",
91  cl::desc("Minimum gain per loop (in cycles) threshold."),
92  cl::init(4), cl::Hidden);
93 
95  "x86-cmov-converter-force-mem-operand",
96  cl::desc("Convert cmovs to branches whenever they have memory operands."),
97  cl::init(true), cl::Hidden);
98 
99 namespace {
100 
101 /// Converts X86 cmov instructions into branches when profitable.
102 class X86CmovConverterPass : public MachineFunctionPass {
103 public:
104  X86CmovConverterPass() : MachineFunctionPass(ID) {
106  }
107 
108  StringRef getPassName() const override { return "X86 cmov Conversion"; }
109  bool runOnMachineFunction(MachineFunction &MF) override;
110  void getAnalysisUsage(AnalysisUsage &AU) const override;
111 
112  /// Pass identification, replacement for typeid.
113  static char ID;
114 
115 private:
117  const TargetInstrInfo *TII;
118  const TargetRegisterInfo *TRI;
119  TargetSchedModel TSchedModel;
120 
121  /// List of consecutive CMOV instructions.
122  using CmovGroup = SmallVector<MachineInstr *, 2>;
123  using CmovGroups = SmallVector<CmovGroup, 2>;
124 
125  /// Collect all CMOV-group-candidates in \p CurrLoop and update \p
126  /// CmovInstGroups accordingly.
127  ///
128  /// \param Blocks List of blocks to process.
129  /// \param CmovInstGroups List of consecutive CMOV instructions in CurrLoop.
130  /// \returns true iff it found any CMOV-group-candidate.
131  bool collectCmovCandidates(ArrayRef<MachineBasicBlock *> Blocks,
132  CmovGroups &CmovInstGroups,
133  bool IncludeLoads = false);
134 
135  /// Check if it is profitable to transform each CMOV-group-candidates into
136  /// branch. Remove all groups that are not profitable from \p CmovInstGroups.
137  ///
138  /// \param Blocks List of blocks to process.
139  /// \param CmovInstGroups List of consecutive CMOV instructions in CurrLoop.
140  /// \returns true iff any CMOV-group-candidate remain.
141  bool checkForProfitableCmovCandidates(ArrayRef<MachineBasicBlock *> Blocks,
142  CmovGroups &CmovInstGroups);
143 
144  /// Convert the given list of consecutive CMOV instructions into a branch.
145  ///
146  /// \param Group Consecutive CMOV instructions to be converted into branch.
147  void convertCmovInstsToBranches(SmallVectorImpl<MachineInstr *> &Group) const;
148 };
149 
150 } // end anonymous namespace
151 
152 char X86CmovConverterPass::ID = 0;
153 
154 void X86CmovConverterPass::getAnalysisUsage(AnalysisUsage &AU) const {
157 }
158 
159 bool X86CmovConverterPass::runOnMachineFunction(MachineFunction &MF) {
160  if (skipFunction(MF.getFunction()))
161  return false;
162  if (!EnableCmovConverter)
163  return false;
164 
165  LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
166  << "**********\n");
167 
168  bool Changed = false;
169  MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
170  const TargetSubtargetInfo &STI = MF.getSubtarget();
171  MRI = &MF.getRegInfo();
172  TII = STI.getInstrInfo();
173  TRI = STI.getRegisterInfo();
174  TSchedModel.init(&STI);
175 
176  // Before we handle the more subtle cases of register-register CMOVs inside
177  // of potentially hot loops, we want to quickly remove all CMOVs with
178  // a memory operand. The CMOV will risk a stall waiting for the load to
179  // complete that speculative execution behind a branch is better suited to
180  // handle on modern x86 chips.
181  if (ForceMemOperand) {
182  CmovGroups AllCmovGroups;
184  for (auto &MBB : MF)
185  Blocks.push_back(&MBB);
186  if (collectCmovCandidates(Blocks, AllCmovGroups, /*IncludeLoads*/ true)) {
187  for (auto &Group : AllCmovGroups) {
188  // Skip any group that doesn't do at least one memory operand cmov.
189  if (!llvm::any_of(Group, [&](MachineInstr *I) { return I->mayLoad(); }))
190  continue;
191 
192  // For CMOV groups which we can rewrite and which contain a memory load,
193  // always rewrite them. On x86, a CMOV will dramatically amplify any
194  // memory latency by blocking speculative execution.
195  Changed = true;
196  convertCmovInstsToBranches(Group);
197  }
198  }
199  }
200 
201  //===--------------------------------------------------------------------===//
202  // Register-operand Conversion Algorithm
203  // ---------
204  // For each inner most loop
205  // collectCmovCandidates() {
206  // Find all CMOV-group-candidates.
207  // }
208  //
209  // checkForProfitableCmovCandidates() {
210  // * Calculate both loop-depth and optimized-loop-depth.
211  // * Use these depth to check for loop transformation profitability.
212  // * Check for CMOV-group-candidate transformation profitability.
213  // }
214  //
215  // For each profitable CMOV-group-candidate
216  // convertCmovInstsToBranches() {
217  // * Create FalseBB, SinkBB, Conditional branch to SinkBB.
218  // * Replace each CMOV instruction with a PHI instruction in SinkBB.
219  // }
220  //
221  // Note: For more details, see each function description.
222  //===--------------------------------------------------------------------===//
223 
224  // Build up the loops in pre-order.
226  // Note that we need to check size on each iteration as we accumulate child
227  // loops.
228  for (int i = 0; i < (int)Loops.size(); ++i)
229  for (MachineLoop *Child : Loops[i]->getSubLoops())
230  Loops.push_back(Child);
231 
232  for (MachineLoop *CurrLoop : Loops) {
233  // Optimize only inner most loops.
234  if (!CurrLoop->getSubLoops().empty())
235  continue;
236 
237  // List of consecutive CMOV instructions to be processed.
238  CmovGroups CmovInstGroups;
239 
240  if (!collectCmovCandidates(CurrLoop->getBlocks(), CmovInstGroups))
241  continue;
242 
243  if (!checkForProfitableCmovCandidates(CurrLoop->getBlocks(),
244  CmovInstGroups))
245  continue;
246 
247  Changed = true;
248  for (auto &Group : CmovInstGroups)
249  convertCmovInstsToBranches(Group);
250  }
251 
252  return Changed;
253 }
254 
255 bool X86CmovConverterPass::collectCmovCandidates(
256  ArrayRef<MachineBasicBlock *> Blocks, CmovGroups &CmovInstGroups,
257  bool IncludeLoads) {
258  //===--------------------------------------------------------------------===//
259  // Collect all CMOV-group-candidates and add them into CmovInstGroups.
260  //
261  // CMOV-group:
262  // CMOV instructions, in same MBB, that uses same EFLAGS def instruction.
263  //
264  // CMOV-group-candidate:
265  // CMOV-group where all the CMOV instructions are
266  // 1. consecutive.
267  // 2. have same condition code or opposite one.
268  // 3. have only operand registers (X86::CMOVrr).
269  //===--------------------------------------------------------------------===//
270  // List of possible improvement (TODO's):
271  // --------------------------------------
272  // TODO: Add support for X86::CMOVrm instructions.
273  // TODO: Add support for X86::SETcc instructions.
274  // TODO: Add support for CMOV-groups with non consecutive CMOV instructions.
275  //===--------------------------------------------------------------------===//
276 
277  // Current processed CMOV-Group.
278  CmovGroup Group;
279  for (auto *MBB : Blocks) {
280  Group.clear();
281  // Condition code of first CMOV instruction current processed range and its
282  // opposite condition code.
283  X86::CondCode FirstCC, FirstOppCC, MemOpCC;
284  // Indicator of a non CMOVrr instruction in the current processed range.
285  bool FoundNonCMOVInst = false;
286  // Indicator for current processed CMOV-group if it should be skipped.
287  bool SkipGroup = false;
288 
289  for (auto &I : *MBB) {
290  // Skip debug instructions.
291  if (I.isDebugInstr())
292  continue;
293  X86::CondCode CC = X86::getCondFromCMovOpc(I.getOpcode());
294  // Check if we found a X86::CMOVrr instruction.
295  if (CC != X86::COND_INVALID && (IncludeLoads || !I.mayLoad())) {
296  if (Group.empty()) {
297  // We found first CMOV in the range, reset flags.
298  FirstCC = CC;
299  FirstOppCC = X86::GetOppositeBranchCondition(CC);
300  // Clear out the prior group's memory operand CC.
301  MemOpCC = X86::COND_INVALID;
302  FoundNonCMOVInst = false;
303  SkipGroup = false;
304  }
305  Group.push_back(&I);
306  // Check if it is a non-consecutive CMOV instruction or it has different
307  // condition code than FirstCC or FirstOppCC.
308  if (FoundNonCMOVInst || (CC != FirstCC && CC != FirstOppCC))
309  // Mark the SKipGroup indicator to skip current processed CMOV-Group.
310  SkipGroup = true;
311  if (I.mayLoad()) {
312  if (MemOpCC == X86::COND_INVALID)
313  // The first memory operand CMOV.
314  MemOpCC = CC;
315  else if (CC != MemOpCC)
316  // Can't handle mixed conditions with memory operands.
317  SkipGroup = true;
318  }
319  // Check if we were relying on zero-extending behavior of the CMOV.
320  if (!SkipGroup &&
321  llvm::any_of(
322  MRI->use_nodbg_instructions(I.defs().begin()->getReg()),
323  [&](MachineInstr &UseI) {
324  return UseI.getOpcode() == X86::SUBREG_TO_REG;
325  }))
326  // FIXME: We should model the cost of using an explicit MOV to handle
327  // the zero-extension rather than just refusing to handle this.
328  SkipGroup = true;
329  continue;
330  }
331  // If Group is empty, keep looking for first CMOV in the range.
332  if (Group.empty())
333  continue;
334 
335  // We found a non X86::CMOVrr instruction.
336  FoundNonCMOVInst = true;
337  // Check if this instruction define EFLAGS, to determine end of processed
338  // range, as there would be no more instructions using current EFLAGS def.
339  if (I.definesRegister(X86::EFLAGS)) {
340  // Check if current processed CMOV-group should not be skipped and add
341  // it as a CMOV-group-candidate.
342  if (!SkipGroup)
343  CmovInstGroups.push_back(Group);
344  else
345  ++NumOfSkippedCmovGroups;
346  Group.clear();
347  }
348  }
349  // End of basic block is considered end of range, check if current processed
350  // CMOV-group should not be skipped and add it as a CMOV-group-candidate.
351  if (Group.empty())
352  continue;
353  if (!SkipGroup)
354  CmovInstGroups.push_back(Group);
355  else
356  ++NumOfSkippedCmovGroups;
357  }
358 
359  NumOfCmovGroupCandidate += CmovInstGroups.size();
360  return !CmovInstGroups.empty();
361 }
362 
363 /// \returns Depth of CMOV instruction as if it was converted into branch.
364 /// \param TrueOpDepth depth cost of CMOV true value operand.
365 /// \param FalseOpDepth depth cost of CMOV false value operand.
366 static unsigned getDepthOfOptCmov(unsigned TrueOpDepth, unsigned FalseOpDepth) {
367  //===--------------------------------------------------------------------===//
368  // With no info about branch weight, we assume 50% for each value operand.
369  // Thus, depth of optimized CMOV instruction is the rounded up average of
370  // its True-Operand-Value-Depth and False-Operand-Value-Depth.
371  //===--------------------------------------------------------------------===//
372  return (TrueOpDepth + FalseOpDepth + 1) / 2;
373 }
374 
375 bool X86CmovConverterPass::checkForProfitableCmovCandidates(
376  ArrayRef<MachineBasicBlock *> Blocks, CmovGroups &CmovInstGroups) {
377  struct DepthInfo {
378  /// Depth of original loop.
379  unsigned Depth;
380  /// Depth of optimized loop.
381  unsigned OptDepth;
382  };
383  /// Number of loop iterations to calculate depth for ?!
384  static const unsigned LoopIterations = 2;
386  DepthInfo LoopDepth[LoopIterations] = {{0, 0}, {0, 0}};
387  enum { PhyRegType = 0, VirRegType = 1, RegTypeNum = 2 };
388  /// For each register type maps the register to its last def instruction.
389  DenseMap<unsigned, MachineInstr *> RegDefMaps[RegTypeNum];
390  /// Maps register operand to its def instruction, which can be nullptr if it
391  /// is unknown (e.g., operand is defined outside the loop).
393 
394  // Set depth of unknown instruction (i.e., nullptr) to zero.
395  DepthMap[nullptr] = {0, 0};
396 
397  SmallPtrSet<MachineInstr *, 4> CmovInstructions;
398  for (auto &Group : CmovInstGroups)
399  CmovInstructions.insert(Group.begin(), Group.end());
400 
401  //===--------------------------------------------------------------------===//
402  // Step 1: Calculate instruction depth and loop depth.
403  // Optimized-Loop:
404  // loop with CMOV-group-candidates converted into branches.
405  //
406  // Instruction-Depth:
407  // instruction latency + max operand depth.
408  // * For CMOV instruction in optimized loop the depth is calculated as:
409  // CMOV latency + getDepthOfOptCmov(True-Op-Depth, False-Op-depth)
410  // TODO: Find a better way to estimate the latency of the branch instruction
411  // rather than using the CMOV latency.
412  //
413  // Loop-Depth:
414  // max instruction depth of all instructions in the loop.
415  // Note: instruction with max depth represents the critical-path in the loop.
416  //
417  // Loop-Depth[i]:
418  // Loop-Depth calculated for first `i` iterations.
419  // Note: it is enough to calculate depth for up to two iterations.
420  //
421  // Depth-Diff[i]:
422  // Number of cycles saved in first 'i` iterations by optimizing the loop.
423  //===--------------------------------------------------------------------===//
424  for (unsigned I = 0; I < LoopIterations; ++I) {
425  DepthInfo &MaxDepth = LoopDepth[I];
426  for (auto *MBB : Blocks) {
427  // Clear physical registers Def map.
428  RegDefMaps[PhyRegType].clear();
429  for (MachineInstr &MI : *MBB) {
430  // Skip debug instructions.
431  if (MI.isDebugInstr())
432  continue;
433  unsigned MIDepth = 0;
434  unsigned MIDepthOpt = 0;
435  bool IsCMOV = CmovInstructions.count(&MI);
436  for (auto &MO : MI.uses()) {
437  // Checks for "isUse()" as "uses()" returns also implicit definitions.
438  if (!MO.isReg() || !MO.isUse())
439  continue;
440  unsigned Reg = MO.getReg();
441  auto &RDM = RegDefMaps[TargetRegisterInfo::isVirtualRegister(Reg)];
442  if (MachineInstr *DefMI = RDM.lookup(Reg)) {
443  OperandToDefMap[&MO] = DefMI;
444  DepthInfo Info = DepthMap.lookup(DefMI);
445  MIDepth = std::max(MIDepth, Info.Depth);
446  if (!IsCMOV)
447  MIDepthOpt = std::max(MIDepthOpt, Info.OptDepth);
448  }
449  }
450 
451  if (IsCMOV)
452  MIDepthOpt = getDepthOfOptCmov(
453  DepthMap[OperandToDefMap.lookup(&MI.getOperand(1))].OptDepth,
454  DepthMap[OperandToDefMap.lookup(&MI.getOperand(2))].OptDepth);
455 
456  // Iterates over all operands to handle implicit definitions as well.
457  for (auto &MO : MI.operands()) {
458  if (!MO.isReg() || !MO.isDef())
459  continue;
460  unsigned Reg = MO.getReg();
461  RegDefMaps[TargetRegisterInfo::isVirtualRegister(Reg)][Reg] = &MI;
462  }
463 
464  unsigned Latency = TSchedModel.computeInstrLatency(&MI);
465  DepthMap[&MI] = {MIDepth += Latency, MIDepthOpt += Latency};
466  MaxDepth.Depth = std::max(MaxDepth.Depth, MIDepth);
467  MaxDepth.OptDepth = std::max(MaxDepth.OptDepth, MIDepthOpt);
468  }
469  }
470  }
471 
472  unsigned Diff[LoopIterations] = {LoopDepth[0].Depth - LoopDepth[0].OptDepth,
473  LoopDepth[1].Depth - LoopDepth[1].OptDepth};
474 
475  //===--------------------------------------------------------------------===//
476  // Step 2: Check if Loop worth to be optimized.
477  // Worth-Optimize-Loop:
478  // case 1: Diff[1] == Diff[0]
479  // Critical-path is iteration independent - there is no dependency
480  // of critical-path instructions on critical-path instructions of
481  // previous iteration.
482  // Thus, it is enough to check gain percent of 1st iteration -
483  // To be conservative, the optimized loop need to have a depth of
484  // 12.5% cycles less than original loop, per iteration.
485  //
486  // case 2: Diff[1] > Diff[0]
487  // Critical-path is iteration dependent - there is dependency of
488  // critical-path instructions on critical-path instructions of
489  // previous iteration.
490  // Thus, check the gain percent of the 2nd iteration (similar to the
491  // previous case), but it is also required to check the gradient of
492  // the gain - the change in Depth-Diff compared to the change in
493  // Loop-Depth between 1st and 2nd iterations.
494  // To be conservative, the gradient need to be at least 50%.
495  //
496  // In addition, In order not to optimize loops with very small gain, the
497  // gain (in cycles) after 2nd iteration should not be less than a given
498  // threshold. Thus, the check (Diff[1] >= GainCycleThreshold) must apply.
499  //
500  // If loop is not worth optimizing, remove all CMOV-group-candidates.
501  //===--------------------------------------------------------------------===//
502  if (Diff[1] < GainCycleThreshold)
503  return false;
504 
505  bool WorthOptLoop = false;
506  if (Diff[1] == Diff[0])
507  WorthOptLoop = Diff[0] * 8 >= LoopDepth[0].Depth;
508  else if (Diff[1] > Diff[0])
509  WorthOptLoop =
510  (Diff[1] - Diff[0]) * 2 >= (LoopDepth[1].Depth - LoopDepth[0].Depth) &&
511  (Diff[1] * 8 >= LoopDepth[1].Depth);
512 
513  if (!WorthOptLoop)
514  return false;
515 
516  ++NumOfLoopCandidate;
517 
518  //===--------------------------------------------------------------------===//
519  // Step 3: Check for each CMOV-group-candidate if it worth to be optimized.
520  // Worth-Optimize-Group:
521  // Iff it worths to optimize all CMOV instructions in the group.
522  //
523  // Worth-Optimize-CMOV:
524  // Predicted branch is faster than CMOV by the difference between depth of
525  // condition operand and depth of taken (predicted) value operand.
526  // To be conservative, the gain of such CMOV transformation should cover at
527  // at least 25% of branch-misprediction-penalty.
528  //===--------------------------------------------------------------------===//
529  unsigned MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty;
530  CmovGroups TempGroups;
531  std::swap(TempGroups, CmovInstGroups);
532  for (auto &Group : TempGroups) {
533  bool WorthOpGroup = true;
534  for (auto *MI : Group) {
535  // Avoid CMOV instruction which value is used as a pointer to load from.
536  // This is another conservative check to avoid converting CMOV instruction
537  // used with tree-search like algorithm, where the branch is unpredicted.
538  auto UIs = MRI->use_instructions(MI->defs().begin()->getReg());
539  if (UIs.begin() != UIs.end() && ++UIs.begin() == UIs.end()) {
540  unsigned Op = UIs.begin()->getOpcode();
541  if (Op == X86::MOV64rm || Op == X86::MOV32rm) {
542  WorthOpGroup = false;
543  break;
544  }
545  }
546 
547  unsigned CondCost =
548  DepthMap[OperandToDefMap.lookup(&MI->getOperand(3))].Depth;
549  unsigned ValCost = getDepthOfOptCmov(
550  DepthMap[OperandToDefMap.lookup(&MI->getOperand(1))].Depth,
551  DepthMap[OperandToDefMap.lookup(&MI->getOperand(2))].Depth);
552  if (ValCost > CondCost || (CondCost - ValCost) * 4 < MispredictPenalty) {
553  WorthOpGroup = false;
554  break;
555  }
556  }
557 
558  if (WorthOpGroup)
559  CmovInstGroups.push_back(Group);
560  }
561 
562  return !CmovInstGroups.empty();
563 }
564 
566  if (MI->killsRegister(X86::EFLAGS))
567  return false;
568 
569  // The EFLAGS operand of MI might be missing a kill marker.
570  // Figure out whether EFLAGS operand should LIVE after MI instruction.
571  MachineBasicBlock *BB = MI->getParent();
573 
574  // Scan forward through BB for a use/def of EFLAGS.
575  for (auto I = std::next(ItrMI), E = BB->end(); I != E; ++I) {
576  if (I->readsRegister(X86::EFLAGS))
577  return true;
578  if (I->definesRegister(X86::EFLAGS))
579  return false;
580  }
581 
582  // We hit the end of the block, check whether EFLAGS is live into a successor.
583  for (auto I = BB->succ_begin(), E = BB->succ_end(); I != E; ++I) {
584  if ((*I)->isLiveIn(X86::EFLAGS))
585  return true;
586  }
587 
588  return false;
589 }
590 
591 /// Given /p First CMOV instruction and /p Last CMOV instruction representing a
592 /// group of CMOV instructions, which may contain debug instructions in between,
593 /// move all debug instructions to after the last CMOV instruction, making the
594 /// CMOV group consecutive.
595 static void packCmovGroup(MachineInstr *First, MachineInstr *Last) {
597  "Last instruction in a CMOV group must be a CMOV instruction");
598 
599  SmallVector<MachineInstr *, 2> DBGInstructions;
600  for (auto I = First->getIterator(), E = Last->getIterator(); I != E; I++) {
601  if (I->isDebugInstr())
602  DBGInstructions.push_back(&*I);
603  }
604 
605  // Splice the debug instruction after the cmov group.
606  MachineBasicBlock *MBB = First->getParent();
607  for (auto *MI : DBGInstructions)
608  MBB->insertAfter(Last, MI->removeFromParent());
609 }
610 
611 void X86CmovConverterPass::convertCmovInstsToBranches(
612  SmallVectorImpl<MachineInstr *> &Group) const {
613  assert(!Group.empty() && "No CMOV instructions to convert");
614  ++NumOfOptimizedCmovGroups;
615 
616  // If the CMOV group is not packed, e.g., there are debug instructions between
617  // first CMOV and last CMOV, then pack the group and make the CMOV instruction
618  // consecutive by moving the debug instructions to after the last CMOV.
619  packCmovGroup(Group.front(), Group.back());
620 
621  // To convert a CMOVcc instruction, we actually have to insert the diamond
622  // control-flow pattern. The incoming instruction knows the destination vreg
623  // to set, the condition code register to branch on, the true/false values to
624  // select between, and a branch opcode to use.
625 
626  // Before
627  // -----
628  // MBB:
629  // cond = cmp ...
630  // v1 = CMOVge t1, f1, cond
631  // v2 = CMOVlt t2, f2, cond
632  // v3 = CMOVge v1, f3, cond
633  //
634  // After
635  // -----
636  // MBB:
637  // cond = cmp ...
638  // jge %SinkMBB
639  //
640  // FalseMBB:
641  // jmp %SinkMBB
642  //
643  // SinkMBB:
644  // %v1 = phi[%f1, %FalseMBB], [%t1, %MBB]
645  // %v2 = phi[%t2, %FalseMBB], [%f2, %MBB] ; For CMOV with OppCC switch
646  // ; true-value with false-value
647  // %v3 = phi[%f3, %FalseMBB], [%t1, %MBB] ; Phi instruction cannot use
648  // ; previous Phi instruction result
649 
650  MachineInstr &MI = *Group.front();
651  MachineInstr *LastCMOV = Group.back();
652  DebugLoc DL = MI.getDebugLoc();
653 
656  // Potentially swap the condition codes so that any memory operand to a CMOV
657  // is in the *false* position instead of the *true* position. We can invert
658  // any non-memory operand CMOV instructions to cope with this and we ensure
659  // memory operand CMOVs are only included with a single condition code.
660  if (llvm::any_of(Group, [&](MachineInstr *I) {
661  return I->mayLoad() && X86::getCondFromCMovOpc(I->getOpcode()) == CC;
662  }))
663  std::swap(CC, OppCC);
664 
665  MachineBasicBlock *MBB = MI.getParent();
667  MachineFunction *F = MBB->getParent();
668  const BasicBlock *BB = MBB->getBasicBlock();
669 
670  MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB);
671  MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(BB);
672  F->insert(It, FalseMBB);
673  F->insert(It, SinkMBB);
674 
675  // If the EFLAGS register isn't dead in the terminator, then claim that it's
676  // live into the sink and copy blocks.
677  if (checkEFLAGSLive(LastCMOV)) {
678  FalseMBB->addLiveIn(X86::EFLAGS);
679  SinkMBB->addLiveIn(X86::EFLAGS);
680  }
681 
682  // Transfer the remainder of BB and its successor edges to SinkMBB.
683  SinkMBB->splice(SinkMBB->begin(), MBB,
684  std::next(MachineBasicBlock::iterator(LastCMOV)), MBB->end());
685  SinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
686 
687  // Add the false and sink blocks as its successors.
688  MBB->addSuccessor(FalseMBB);
689  MBB->addSuccessor(SinkMBB);
690 
691  // Create the conditional branch instruction.
692  BuildMI(MBB, DL, TII->get(X86::GetCondBranchFromCond(CC))).addMBB(SinkMBB);
693 
694  // Add the sink block to the false block successors.
695  FalseMBB->addSuccessor(SinkMBB);
696 
700  std::next(MachineBasicBlock::iterator(LastCMOV));
701  MachineBasicBlock::iterator FalseInsertionPoint = FalseMBB->begin();
702  MachineBasicBlock::iterator SinkInsertionPoint = SinkMBB->begin();
703 
704  // First we need to insert an explicit load on the false path for any memory
705  // operand. We also need to potentially do register rewriting here, but it is
706  // simpler as the memory operands are always on the false path so we can
707  // simply take that input, whatever it is.
708  DenseMap<unsigned, unsigned> FalseBBRegRewriteTable;
709  for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd;) {
710  auto &MI = *MIIt++;
711  // Skip any CMOVs in this group which don't load from memory.
712  if (!MI.mayLoad()) {
713  // Remember the false-side register input.
714  unsigned FalseReg =
715  MI.getOperand(X86::getCondFromCMovOpc(MI.getOpcode()) == CC ? 1 : 2)
716  .getReg();
717  // Walk back through any intermediate cmovs referenced.
718  while (true) {
719  auto FRIt = FalseBBRegRewriteTable.find(FalseReg);
720  if (FRIt == FalseBBRegRewriteTable.end())
721  break;
722  FalseReg = FRIt->second;
723  }
724  FalseBBRegRewriteTable[MI.getOperand(0).getReg()] = FalseReg;
725  continue;
726  }
727 
728  // The condition must be the *opposite* of the one we've decided to branch
729  // on as the branch will go *around* the load and the load should happen
730  // when the CMOV condition is false.
731  assert(X86::getCondFromCMovOpc(MI.getOpcode()) == OppCC &&
732  "Can only handle memory-operand cmov instructions with a condition "
733  "opposite to the selected branch direction.");
734 
735  // The goal is to rewrite the cmov from:
736  //
737  // MBB:
738  // %A = CMOVcc %B (tied), (mem)
739  //
740  // to
741  //
742  // MBB:
743  // %A = CMOVcc %B (tied), %C
744  // FalseMBB:
745  // %C = MOV (mem)
746  //
747  // Which will allow the next loop to rewrite the CMOV in terms of a PHI:
748  //
749  // MBB:
750  // JMP!cc SinkMBB
751  // FalseMBB:
752  // %C = MOV (mem)
753  // SinkMBB:
754  // %A = PHI [ %C, FalseMBB ], [ %B, MBB]
755 
756  // Get a fresh register to use as the destination of the MOV.
757  const TargetRegisterClass *RC = MRI->getRegClass(MI.getOperand(0).getReg());
758  unsigned TmpReg = MRI->createVirtualRegister(RC);
759 
761  bool Unfolded = TII->unfoldMemoryOperand(*MBB->getParent(), MI, TmpReg,
762  /*UnfoldLoad*/ true,
763  /*UnfoldStore*/ false, NewMIs);
764  (void)Unfolded;
765  assert(Unfolded && "Should never fail to unfold a loading cmov!");
766 
767  // Move the new CMOV to just before the old one and reset any impacted
768  // iterator.
769  auto *NewCMOV = NewMIs.pop_back_val();
770  assert(X86::getCondFromCMovOpc(NewCMOV->getOpcode()) == OppCC &&
771  "Last new instruction isn't the expected CMOV!");
772  LLVM_DEBUG(dbgs() << "\tRewritten cmov: "; NewCMOV->dump());
773  MBB->insert(MachineBasicBlock::iterator(MI), NewCMOV);
774  if (&*MIItBegin == &MI)
775  MIItBegin = MachineBasicBlock::iterator(NewCMOV);
776 
777  // Sink whatever instructions were needed to produce the unfolded operand
778  // into the false block.
779  for (auto *NewMI : NewMIs) {
780  LLVM_DEBUG(dbgs() << "\tRewritten load instr: "; NewMI->dump());
781  FalseMBB->insert(FalseInsertionPoint, NewMI);
782  // Re-map any operands that are from other cmovs to the inputs for this block.
783  for (auto &MOp : NewMI->uses()) {
784  if (!MOp.isReg())
785  continue;
786  auto It = FalseBBRegRewriteTable.find(MOp.getReg());
787  if (It == FalseBBRegRewriteTable.end())
788  continue;
789 
790  MOp.setReg(It->second);
791  // This might have been a kill when it referenced the cmov result, but
792  // it won't necessarily be once rewritten.
793  // FIXME: We could potentially improve this by tracking whether the
794  // operand to the cmov was also a kill, and then skipping the PHI node
795  // construction below.
796  MOp.setIsKill(false);
797  }
798  }
800  std::next(MachineBasicBlock::iterator(MI)));
801 
802  // Add this PHI to the rewrite table.
803  FalseBBRegRewriteTable[NewCMOV->getOperand(0).getReg()] = TmpReg;
804  }
805 
806  // As we are creating the PHIs, we have to be careful if there is more than
807  // one. Later CMOVs may reference the results of earlier CMOVs, but later
808  // PHIs have to reference the individual true/false inputs from earlier PHIs.
809  // That also means that PHI construction must work forward from earlier to
810  // later, and that the code must maintain a mapping from earlier PHI's
811  // destination registers, and the registers that went into the PHI.
813 
814  for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
815  unsigned DestReg = MIIt->getOperand(0).getReg();
816  unsigned Op1Reg = MIIt->getOperand(1).getReg();
817  unsigned Op2Reg = MIIt->getOperand(2).getReg();
818 
819  // If this CMOV we are processing is the opposite condition from the jump we
820  // generated, then we have to swap the operands for the PHI that is going to
821  // be generated.
822  if (X86::getCondFromCMovOpc(MIIt->getOpcode()) == OppCC)
823  std::swap(Op1Reg, Op2Reg);
824 
825  auto Op1Itr = RegRewriteTable.find(Op1Reg);
826  if (Op1Itr != RegRewriteTable.end())
827  Op1Reg = Op1Itr->second.first;
828 
829  auto Op2Itr = RegRewriteTable.find(Op2Reg);
830  if (Op2Itr != RegRewriteTable.end())
831  Op2Reg = Op2Itr->second.second;
832 
833  // SinkMBB:
834  // %Result = phi [ %FalseValue, FalseMBB ], [ %TrueValue, MBB ]
835  // ...
836  MIB = BuildMI(*SinkMBB, SinkInsertionPoint, DL, TII->get(X86::PHI), DestReg)
837  .addReg(Op1Reg)
838  .addMBB(FalseMBB)
839  .addReg(Op2Reg)
840  .addMBB(MBB);
841  (void)MIB;
842  LLVM_DEBUG(dbgs() << "\tFrom: "; MIIt->dump());
843  LLVM_DEBUG(dbgs() << "\tTo: "; MIB->dump());
844 
845  // Add this PHI to the rewrite table.
846  RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
847  }
848 
849  // Now remove the CMOV(s).
850  MBB->erase(MIItBegin, MIItEnd);
851 }
852 
853 INITIALIZE_PASS_BEGIN(X86CmovConverterPass, DEBUG_TYPE, "X86 cmov Conversion",
854  false, false)
856 INITIALIZE_PASS_END(X86CmovConverterPass, DEBUG_TYPE, "X86 cmov Conversion",
857  false, false)
858 
860  return new X86CmovConverterPass();
861 }
unsigned GetCondBranchFromCond(CondCode CC)
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
This class represents lattice values for constants.
Definition: AllocatorList.h:23
static void packCmovGroup(MachineInstr *First, MachineInstr *Last)
Given /p First CMOV instruction and /p Last CMOV instruction representing a group of CMOV instruction...
CondCode getCondFromCMovOpc(unsigned Opc)
Return condition code of a CMov opcode.
static bool isVirtualRegister(unsigned Reg)
Return true if the specified register number is in the virtual register namespace.
void transferSuccessorsAndUpdatePHIs(MachineBasicBlock *FromMBB)
Transfers all the successors, as in transferSuccessors, and update PHI operands in the successor bloc...
unsigned Reg
iterator insertAfter(iterator I, MachineInstr *MI)
Insert MI into the instruction list after I.
STATISTIC(NumFunctions, "Total number of functions")
unsigned const TargetRegisterInfo * TRI
A debug info location.
Definition: DebugLoc.h:33
F(f)
void initializeX86CmovConverterPassPass(PassRegistry &)
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
Hexagon Hardware Loops
instr_iterator erase(instr_iterator I)
Remove an instruction from the instruction list and delete it.
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
Provide an instruction scheduling machine model to CodeGen passes.
const HexagonInstrInfo * TII
static unsigned getDepthOfOptCmov(unsigned TrueOpDepth, unsigned FalseOpDepth)
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:408
static bool checkEFLAGSLive(MachineInstr *MI)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
instr_iterator insert(instr_iterator I, MachineInstr *M)
Insert MI into the instruction list before I, possibly inside a bundle.
#define DEBUG_TYPE
MachineBasicBlock * CreateMachineBasicBlock(const BasicBlock *bb=nullptr)
CreateMachineBasicBlock - Allocate a new MachineBasicBlock.
Analysis containing CSE Info
Definition: CSEInfo.cpp:20
StringRef getName() const
getName - Return the name of the corresponding LLVM function.
TargetInstrInfo - Interface to description of machine instruction set.
iterator begin() const
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:176
static cl::opt< bool > EnableCmovConverter("x86-cmov-converter", cl::desc("Enable the X86 cmov-to-branch optimization."), cl::init(true), cl::Hidden)
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
MachineInstrBundleIterator< MachineInstr > iterator
void addLiveIn(MCPhysReg PhysReg, LaneBitmask LaneMask=LaneBitmask::getAll())
Adds the specified register as a live in.
unsigned const MachineRegisterInfo * MRI
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
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.
static cl::opt< bool > ForceMemOperand("x86-cmov-converter-force-mem-operand", cl::desc("Convert cmovs to branches whenever they have memory operands."), cl::init(true), cl::Hidden)
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
static cl::opt< unsigned > GainCycleThreshold("x86-cmov-converter-threshold", cl::desc("Minimum gain per loop (in cycles) threshold."), cl::init(4), cl::Hidden)
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
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
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
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
const unsigned MaxDepth
FunctionPass * createX86CmovConverterPass()
This pass converts X86 cmov instructions into branch when profitable.
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
void addSuccessor(MachineBasicBlock *Succ, BranchProbability Prob=BranchProbability::getUnknown())
Add Succ as a successor of this MachineBasicBlock.
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:839
CondCode GetOppositeBranchCondition(CondCode CC)
GetOppositeBranchCondition - Return the inverse of the specified cond, e.g.
MachineInstrBuilder MachineInstrBuilder & DefMI
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:373
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
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
iterator end() const
X86 cmov Conversion
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.
TargetSubtargetInfo - Generic base class for all target subtargets.
Representation of each machine instruction.
Definition: MachineInstr.h:63
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
bool killsRegister(unsigned Reg, const TargetRegisterInfo *TRI=nullptr) const
Return true if the MachineInstr kills the specified register.
INITIALIZE_PASS_BEGIN(X86CmovConverterPass, DEBUG_TYPE, "X86 cmov Conversion", false, false) INITIALIZE_PASS_END(X86CmovConverterPass
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.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
#define I(x, y, z)
Definition: MD5.cpp:58
iterator end()
Definition: DenseMap.h:108
const BasicBlock * getBasicBlock() const
Return the LLVM basic block that this instance corresponded to originally.
const MachineInstrBuilder & addReg(unsigned RegNo, unsigned flags=0, unsigned SubReg=0) const
Add a new virtual register operand.
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:211
bool mayLoad(QueryType Type=AnyInBundle) const
Return true if this instruction could possibly read memory.
Definition: MachineInstr.h:806
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
void insert(iterator MBBI, MachineBasicBlock *MBB)
IRTranslator LLVM IR MI
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
const MachineInstrBuilder & addMBB(MachineBasicBlock *MBB, unsigned char TargetFlags=0) const
#define LLVM_DEBUG(X)
Definition: Debug.h:122