LLVM  8.0.0svn
X86CmovConversion.cpp
Go to the documentation of this file.
1 //====- X86CmovConversion.cpp - Convert Cmov to Branch --------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// This file implements a pass that converts X86 cmov instructions into
12 /// branches when profitable. This pass is conservative. It transforms if and
13 /// only if it can guarantee a gain with high confidence.
14 ///
15 /// Thus, the optimization applies under the following conditions:
16 /// 1. Consider as candidates only CMOVs in innermost loops (assume that
17 /// most hotspots are represented by these loops).
18 /// 2. Given a group of CMOV instructions that are using the same EFLAGS def
19 /// instruction:
20 /// a. Consider them as candidates only if all have the same code condition
21 /// or the opposite one to prevent generating more than one conditional
22 /// jump per EFLAGS def instruction.
23 /// b. Consider them as candidates only if all are profitable to be
24 /// converted (assume that one bad conversion may cause a degradation).
25 /// 3. Apply conversion only for loops that are found profitable and only for
26 /// CMOV candidates that were found profitable.
27 /// a. A loop is considered profitable only if conversion will reduce its
28 /// depth cost by some threshold.
29 /// b. CMOV is considered profitable if the cost of its condition is higher
30 /// than the average cost of its true-value and false-value by 25% of
31 /// branch-misprediction-penalty. This assures no degradation even with
32 /// 25% branch misprediction.
33 ///
34 /// Note: This pass is assumed to run on SSA machine code.
35 //
36 //===----------------------------------------------------------------------===//
37 //
38 // External interfaces:
39 // FunctionPass *llvm::createX86CmovConverterPass();
40 // bool X86CmovConverterPass::runOnMachineFunction(MachineFunction &MF);
41 //
42 //===----------------------------------------------------------------------===//
43 
44 #include "X86.h"
45 #include "X86InstrInfo.h"
46 #include "llvm/ADT/ArrayRef.h"
47 #include "llvm/ADT/DenseMap.h"
48 #include "llvm/ADT/STLExtras.h"
49 #include "llvm/ADT/SmallPtrSet.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/Statistic.h"
64 #include "llvm/IR/DebugLoc.h"
65 #include "llvm/MC/MCSchedule.h"
66 #include "llvm/Pass.h"
68 #include "llvm/Support/Debug.h"
70 #include <algorithm>
71 #include <cassert>
72 #include <iterator>
73 #include <utility>
74 
75 using namespace llvm;
76 
77 #define DEBUG_TYPE "x86-cmov-conversion"
78 
79 STATISTIC(NumOfSkippedCmovGroups, "Number of unsupported CMOV-groups");
80 STATISTIC(NumOfCmovGroupCandidate, "Number of CMOV-group candidates");
81 STATISTIC(NumOfLoopCandidate, "Number of CMOV-conversion profitable loops");
82 STATISTIC(NumOfOptimizedCmovGroups, "Number of optimized CMOV-groups");
83 
84 // This internal switch can be used to turn off the cmov/branch optimization.
85 static cl::opt<bool>
86  EnableCmovConverter("x86-cmov-converter",
87  cl::desc("Enable the X86 cmov-to-branch optimization."),
88  cl::init(true), cl::Hidden);
89 
90 static cl::opt<unsigned>
91  GainCycleThreshold("x86-cmov-converter-threshold",
92  cl::desc("Minimum gain per loop (in cycles) threshold."),
93  cl::init(4), cl::Hidden);
94 
96  "x86-cmov-converter-force-mem-operand",
97  cl::desc("Convert cmovs to branches whenever they have memory operands."),
98  cl::init(true), cl::Hidden);
99 
100 namespace {
101 
102 /// Converts X86 cmov instructions into branches when profitable.
103 class X86CmovConverterPass : public MachineFunctionPass {
104 public:
105  X86CmovConverterPass() : MachineFunctionPass(ID) {
107  }
108 
109  StringRef getPassName() const override { return "X86 cmov Conversion"; }
110  bool runOnMachineFunction(MachineFunction &MF) override;
111  void getAnalysisUsage(AnalysisUsage &AU) const override;
112 
113  /// Pass identification, replacement for typeid.
114  static char ID;
115 
116 private:
118  const TargetInstrInfo *TII;
119  const TargetRegisterInfo *TRI;
120  TargetSchedModel TSchedModel;
121 
122  /// List of consecutive CMOV instructions.
123  using CmovGroup = SmallVector<MachineInstr *, 2>;
124  using CmovGroups = SmallVector<CmovGroup, 2>;
125 
126  /// Collect all CMOV-group-candidates in \p CurrLoop and update \p
127  /// CmovInstGroups accordingly.
128  ///
129  /// \param Blocks List of blocks to process.
130  /// \param CmovInstGroups List of consecutive CMOV instructions in CurrLoop.
131  /// \returns true iff it found any CMOV-group-candidate.
132  bool collectCmovCandidates(ArrayRef<MachineBasicBlock *> Blocks,
133  CmovGroups &CmovInstGroups,
134  bool IncludeLoads = false);
135 
136  /// Check if it is profitable to transform each CMOV-group-candidates into
137  /// branch. Remove all groups that are not profitable from \p CmovInstGroups.
138  ///
139  /// \param Blocks List of blocks to process.
140  /// \param CmovInstGroups List of consecutive CMOV instructions in CurrLoop.
141  /// \returns true iff any CMOV-group-candidate remain.
142  bool checkForProfitableCmovCandidates(ArrayRef<MachineBasicBlock *> Blocks,
143  CmovGroups &CmovInstGroups);
144 
145  /// Convert the given list of consecutive CMOV instructions into a branch.
146  ///
147  /// \param Group Consecutive CMOV instructions to be converted into branch.
148  void convertCmovInstsToBranches(SmallVectorImpl<MachineInstr *> &Group) const;
149 };
150 
151 } // end anonymous namespace
152 
153 char X86CmovConverterPass::ID = 0;
154 
155 void X86CmovConverterPass::getAnalysisUsage(AnalysisUsage &AU) const {
158 }
159 
160 bool X86CmovConverterPass::runOnMachineFunction(MachineFunction &MF) {
161  if (skipFunction(MF.getFunction()))
162  return false;
163  if (!EnableCmovConverter)
164  return false;
165 
166  LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
167  << "**********\n");
168 
169  bool Changed = false;
170  MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
171  const TargetSubtargetInfo &STI = MF.getSubtarget();
172  MRI = &MF.getRegInfo();
173  TII = STI.getInstrInfo();
174  TRI = STI.getRegisterInfo();
175  TSchedModel.init(&STI);
176 
177  // Before we handle the more subtle cases of register-register CMOVs inside
178  // of potentially hot loops, we want to quickly remove all CMOVs with
179  // a memory operand. The CMOV will risk a stall waiting for the load to
180  // complete that speculative execution behind a branch is better suited to
181  // handle on modern x86 chips.
182  if (ForceMemOperand) {
183  CmovGroups AllCmovGroups;
185  for (auto &MBB : MF)
186  Blocks.push_back(&MBB);
187  if (collectCmovCandidates(Blocks, AllCmovGroups, /*IncludeLoads*/ true)) {
188  for (auto &Group : AllCmovGroups) {
189  // Skip any group that doesn't do at least one memory operand cmov.
190  if (!llvm::any_of(Group, [&](MachineInstr *I) { return I->mayLoad(); }))
191  continue;
192 
193  // For CMOV groups which we can rewrite and which contain a memory load,
194  // always rewrite them. On x86, a CMOV will dramatically amplify any
195  // memory latency by blocking speculative execution.
196  Changed = true;
197  convertCmovInstsToBranches(Group);
198  }
199  }
200  }
201 
202  //===--------------------------------------------------------------------===//
203  // Register-operand Conversion Algorithm
204  // ---------
205  // For each inner most loop
206  // collectCmovCandidates() {
207  // Find all CMOV-group-candidates.
208  // }
209  //
210  // checkForProfitableCmovCandidates() {
211  // * Calculate both loop-depth and optimized-loop-depth.
212  // * Use these depth to check for loop transformation profitability.
213  // * Check for CMOV-group-candidate transformation profitability.
214  // }
215  //
216  // For each profitable CMOV-group-candidate
217  // convertCmovInstsToBranches() {
218  // * Create FalseBB, SinkBB, Conditional branch to SinkBB.
219  // * Replace each CMOV instruction with a PHI instruction in SinkBB.
220  // }
221  //
222  // Note: For more details, see each function description.
223  //===--------------------------------------------------------------------===//
224 
225  // Build up the loops in pre-order.
227  // Note that we need to check size on each iteration as we accumulate child
228  // loops.
229  for (int i = 0; i < (int)Loops.size(); ++i)
230  for (MachineLoop *Child : Loops[i]->getSubLoops())
231  Loops.push_back(Child);
232 
233  for (MachineLoop *CurrLoop : Loops) {
234  // Optimize only inner most loops.
235  if (!CurrLoop->getSubLoops().empty())
236  continue;
237 
238  // List of consecutive CMOV instructions to be processed.
239  CmovGroups CmovInstGroups;
240 
241  if (!collectCmovCandidates(CurrLoop->getBlocks(), CmovInstGroups))
242  continue;
243 
244  if (!checkForProfitableCmovCandidates(CurrLoop->getBlocks(),
245  CmovInstGroups))
246  continue;
247 
248  Changed = true;
249  for (auto &Group : CmovInstGroups)
250  convertCmovInstsToBranches(Group);
251  }
252 
253  return Changed;
254 }
255 
256 bool X86CmovConverterPass::collectCmovCandidates(
257  ArrayRef<MachineBasicBlock *> Blocks, CmovGroups &CmovInstGroups,
258  bool IncludeLoads) {
259  //===--------------------------------------------------------------------===//
260  // Collect all CMOV-group-candidates and add them into CmovInstGroups.
261  //
262  // CMOV-group:
263  // CMOV instructions, in same MBB, that uses same EFLAGS def instruction.
264  //
265  // CMOV-group-candidate:
266  // CMOV-group where all the CMOV instructions are
267  // 1. consecutive.
268  // 2. have same condition code or opposite one.
269  // 3. have only operand registers (X86::CMOVrr).
270  //===--------------------------------------------------------------------===//
271  // List of possible improvement (TODO's):
272  // --------------------------------------
273  // TODO: Add support for X86::CMOVrm instructions.
274  // TODO: Add support for X86::SETcc instructions.
275  // TODO: Add support for CMOV-groups with non consecutive CMOV instructions.
276  //===--------------------------------------------------------------------===//
277 
278  // Current processed CMOV-Group.
279  CmovGroup Group;
280  for (auto *MBB : Blocks) {
281  Group.clear();
282  // Condition code of first CMOV instruction current processed range and its
283  // opposite condition code.
284  X86::CondCode FirstCC, FirstOppCC, MemOpCC;
285  // Indicator of a non CMOVrr instruction in the current processed range.
286  bool FoundNonCMOVInst = false;
287  // Indicator for current processed CMOV-group if it should be skipped.
288  bool SkipGroup = false;
289 
290  for (auto &I : *MBB) {
291  // Skip debug instructions.
292  if (I.isDebugInstr())
293  continue;
294  X86::CondCode CC = X86::getCondFromCMovOpc(I.getOpcode());
295  // Check if we found a X86::CMOVrr instruction.
296  if (CC != X86::COND_INVALID && (IncludeLoads || !I.mayLoad())) {
297  if (Group.empty()) {
298  // We found first CMOV in the range, reset flags.
299  FirstCC = CC;
300  FirstOppCC = X86::GetOppositeBranchCondition(CC);
301  // Clear out the prior group's memory operand CC.
302  MemOpCC = X86::COND_INVALID;
303  FoundNonCMOVInst = false;
304  SkipGroup = false;
305  }
306  Group.push_back(&I);
307  // Check if it is a non-consecutive CMOV instruction or it has different
308  // condition code than FirstCC or FirstOppCC.
309  if (FoundNonCMOVInst || (CC != FirstCC && CC != FirstOppCC))
310  // Mark the SKipGroup indicator to skip current processed CMOV-Group.
311  SkipGroup = true;
312  if (I.mayLoad()) {
313  if (MemOpCC == X86::COND_INVALID)
314  // The first memory operand CMOV.
315  MemOpCC = CC;
316  else if (CC != MemOpCC)
317  // Can't handle mixed conditions with memory operands.
318  SkipGroup = true;
319  }
320  // Check if we were relying on zero-extending behavior of the CMOV.
321  if (!SkipGroup &&
322  llvm::any_of(
323  MRI->use_nodbg_instructions(I.defs().begin()->getReg()),
324  [&](MachineInstr &UseI) {
325  return UseI.getOpcode() == X86::SUBREG_TO_REG;
326  }))
327  // FIXME: We should model the cost of using an explicit MOV to handle
328  // the zero-extension rather than just refusing to handle this.
329  SkipGroup = true;
330  continue;
331  }
332  // If Group is empty, keep looking for first CMOV in the range.
333  if (Group.empty())
334  continue;
335 
336  // We found a non X86::CMOVrr instruction.
337  FoundNonCMOVInst = true;
338  // Check if this instruction define EFLAGS, to determine end of processed
339  // range, as there would be no more instructions using current EFLAGS def.
340  if (I.definesRegister(X86::EFLAGS)) {
341  // Check if current processed CMOV-group should not be skipped and add
342  // it as a CMOV-group-candidate.
343  if (!SkipGroup)
344  CmovInstGroups.push_back(Group);
345  else
346  ++NumOfSkippedCmovGroups;
347  Group.clear();
348  }
349  }
350  // End of basic block is considered end of range, check if current processed
351  // CMOV-group should not be skipped and add it as a CMOV-group-candidate.
352  if (Group.empty())
353  continue;
354  if (!SkipGroup)
355  CmovInstGroups.push_back(Group);
356  else
357  ++NumOfSkippedCmovGroups;
358  }
359 
360  NumOfCmovGroupCandidate += CmovInstGroups.size();
361  return !CmovInstGroups.empty();
362 }
363 
364 /// \returns Depth of CMOV instruction as if it was converted into branch.
365 /// \param TrueOpDepth depth cost of CMOV true value operand.
366 /// \param FalseOpDepth depth cost of CMOV false value operand.
367 static unsigned getDepthOfOptCmov(unsigned TrueOpDepth, unsigned FalseOpDepth) {
368  //===--------------------------------------------------------------------===//
369  // With no info about branch weight, we assume 50% for each value operand.
370  // Thus, depth of optimized CMOV instruction is the rounded up average of
371  // its True-Operand-Value-Depth and False-Operand-Value-Depth.
372  //===--------------------------------------------------------------------===//
373  return (TrueOpDepth + FalseOpDepth + 1) / 2;
374 }
375 
376 bool X86CmovConverterPass::checkForProfitableCmovCandidates(
377  ArrayRef<MachineBasicBlock *> Blocks, CmovGroups &CmovInstGroups) {
378  struct DepthInfo {
379  /// Depth of original loop.
380  unsigned Depth;
381  /// Depth of optimized loop.
382  unsigned OptDepth;
383  };
384  /// Number of loop iterations to calculate depth for ?!
385  static const unsigned LoopIterations = 2;
387  DepthInfo LoopDepth[LoopIterations] = {{0, 0}, {0, 0}};
388  enum { PhyRegType = 0, VirRegType = 1, RegTypeNum = 2 };
389  /// For each register type maps the register to its last def instruction.
390  DenseMap<unsigned, MachineInstr *> RegDefMaps[RegTypeNum];
391  /// Maps register operand to its def instruction, which can be nullptr if it
392  /// is unknown (e.g., operand is defined outside the loop).
394 
395  // Set depth of unknown instruction (i.e., nullptr) to zero.
396  DepthMap[nullptr] = {0, 0};
397 
398  SmallPtrSet<MachineInstr *, 4> CmovInstructions;
399  for (auto &Group : CmovInstGroups)
400  CmovInstructions.insert(Group.begin(), Group.end());
401 
402  //===--------------------------------------------------------------------===//
403  // Step 1: Calculate instruction depth and loop depth.
404  // Optimized-Loop:
405  // loop with CMOV-group-candidates converted into branches.
406  //
407  // Instruction-Depth:
408  // instruction latency + max operand depth.
409  // * For CMOV instruction in optimized loop the depth is calculated as:
410  // CMOV latency + getDepthOfOptCmov(True-Op-Depth, False-Op-depth)
411  // TODO: Find a better way to estimate the latency of the branch instruction
412  // rather than using the CMOV latency.
413  //
414  // Loop-Depth:
415  // max instruction depth of all instructions in the loop.
416  // Note: instruction with max depth represents the critical-path in the loop.
417  //
418  // Loop-Depth[i]:
419  // Loop-Depth calculated for first `i` iterations.
420  // Note: it is enough to calculate depth for up to two iterations.
421  //
422  // Depth-Diff[i]:
423  // Number of cycles saved in first 'i` iterations by optimizing the loop.
424  //===--------------------------------------------------------------------===//
425  for (unsigned I = 0; I < LoopIterations; ++I) {
426  DepthInfo &MaxDepth = LoopDepth[I];
427  for (auto *MBB : Blocks) {
428  // Clear physical registers Def map.
429  RegDefMaps[PhyRegType].clear();
430  for (MachineInstr &MI : *MBB) {
431  // Skip debug instructions.
432  if (MI.isDebugInstr())
433  continue;
434  unsigned MIDepth = 0;
435  unsigned MIDepthOpt = 0;
436  bool IsCMOV = CmovInstructions.count(&MI);
437  for (auto &MO : MI.uses()) {
438  // Checks for "isUse()" as "uses()" returns also implicit definitions.
439  if (!MO.isReg() || !MO.isUse())
440  continue;
441  unsigned Reg = MO.getReg();
442  auto &RDM = RegDefMaps[TargetRegisterInfo::isVirtualRegister(Reg)];
443  if (MachineInstr *DefMI = RDM.lookup(Reg)) {
444  OperandToDefMap[&MO] = DefMI;
445  DepthInfo Info = DepthMap.lookup(DefMI);
446  MIDepth = std::max(MIDepth, Info.Depth);
447  if (!IsCMOV)
448  MIDepthOpt = std::max(MIDepthOpt, Info.OptDepth);
449  }
450  }
451 
452  if (IsCMOV)
453  MIDepthOpt = getDepthOfOptCmov(
454  DepthMap[OperandToDefMap.lookup(&MI.getOperand(1))].OptDepth,
455  DepthMap[OperandToDefMap.lookup(&MI.getOperand(2))].OptDepth);
456 
457  // Iterates over all operands to handle implicit definitions as well.
458  for (auto &MO : MI.operands()) {
459  if (!MO.isReg() || !MO.isDef())
460  continue;
461  unsigned Reg = MO.getReg();
462  RegDefMaps[TargetRegisterInfo::isVirtualRegister(Reg)][Reg] = &MI;
463  }
464 
465  unsigned Latency = TSchedModel.computeInstrLatency(&MI);
466  DepthMap[&MI] = {MIDepth += Latency, MIDepthOpt += Latency};
467  MaxDepth.Depth = std::max(MaxDepth.Depth, MIDepth);
468  MaxDepth.OptDepth = std::max(MaxDepth.OptDepth, MIDepthOpt);
469  }
470  }
471  }
472 
473  unsigned Diff[LoopIterations] = {LoopDepth[0].Depth - LoopDepth[0].OptDepth,
474  LoopDepth[1].Depth - LoopDepth[1].OptDepth};
475 
476  //===--------------------------------------------------------------------===//
477  // Step 2: Check if Loop worth to be optimized.
478  // Worth-Optimize-Loop:
479  // case 1: Diff[1] == Diff[0]
480  // Critical-path is iteration independent - there is no dependency
481  // of critical-path instructions on critical-path instructions of
482  // previous iteration.
483  // Thus, it is enough to check gain percent of 1st iteration -
484  // To be conservative, the optimized loop need to have a depth of
485  // 12.5% cycles less than original loop, per iteration.
486  //
487  // case 2: Diff[1] > Diff[0]
488  // Critical-path is iteration dependent - there is dependency of
489  // critical-path instructions on critical-path instructions of
490  // previous iteration.
491  // Thus, check the gain percent of the 2nd iteration (similar to the
492  // previous case), but it is also required to check the gradient of
493  // the gain - the change in Depth-Diff compared to the change in
494  // Loop-Depth between 1st and 2nd iterations.
495  // To be conservative, the gradient need to be at least 50%.
496  //
497  // In addition, In order not to optimize loops with very small gain, the
498  // gain (in cycles) after 2nd iteration should not be less than a given
499  // threshold. Thus, the check (Diff[1] >= GainCycleThreshold) must apply.
500  //
501  // If loop is not worth optimizing, remove all CMOV-group-candidates.
502  //===--------------------------------------------------------------------===//
503  if (Diff[1] < GainCycleThreshold)
504  return false;
505 
506  bool WorthOptLoop = false;
507  if (Diff[1] == Diff[0])
508  WorthOptLoop = Diff[0] * 8 >= LoopDepth[0].Depth;
509  else if (Diff[1] > Diff[0])
510  WorthOptLoop =
511  (Diff[1] - Diff[0]) * 2 >= (LoopDepth[1].Depth - LoopDepth[0].Depth) &&
512  (Diff[1] * 8 >= LoopDepth[1].Depth);
513 
514  if (!WorthOptLoop)
515  return false;
516 
517  ++NumOfLoopCandidate;
518 
519  //===--------------------------------------------------------------------===//
520  // Step 3: Check for each CMOV-group-candidate if it worth to be optimized.
521  // Worth-Optimize-Group:
522  // Iff it worths to optimize all CMOV instructions in the group.
523  //
524  // Worth-Optimize-CMOV:
525  // Predicted branch is faster than CMOV by the difference between depth of
526  // condition operand and depth of taken (predicted) value operand.
527  // To be conservative, the gain of such CMOV transformation should cover at
528  // at least 25% of branch-misprediction-penalty.
529  //===--------------------------------------------------------------------===//
530  unsigned MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty;
531  CmovGroups TempGroups;
532  std::swap(TempGroups, CmovInstGroups);
533  for (auto &Group : TempGroups) {
534  bool WorthOpGroup = true;
535  for (auto *MI : Group) {
536  // Avoid CMOV instruction which value is used as a pointer to load from.
537  // This is another conservative check to avoid converting CMOV instruction
538  // used with tree-search like algorithm, where the branch is unpredicted.
539  auto UIs = MRI->use_instructions(MI->defs().begin()->getReg());
540  if (UIs.begin() != UIs.end() && ++UIs.begin() == UIs.end()) {
541  unsigned Op = UIs.begin()->getOpcode();
542  if (Op == X86::MOV64rm || Op == X86::MOV32rm) {
543  WorthOpGroup = false;
544  break;
545  }
546  }
547 
548  unsigned CondCost =
549  DepthMap[OperandToDefMap.lookup(&MI->getOperand(3))].Depth;
550  unsigned ValCost = getDepthOfOptCmov(
551  DepthMap[OperandToDefMap.lookup(&MI->getOperand(1))].Depth,
552  DepthMap[OperandToDefMap.lookup(&MI->getOperand(2))].Depth);
553  if (ValCost > CondCost || (CondCost - ValCost) * 4 < MispredictPenalty) {
554  WorthOpGroup = false;
555  break;
556  }
557  }
558 
559  if (WorthOpGroup)
560  CmovInstGroups.push_back(Group);
561  }
562 
563  return !CmovInstGroups.empty();
564 }
565 
567  if (MI->killsRegister(X86::EFLAGS))
568  return false;
569 
570  // The EFLAGS operand of MI might be missing a kill marker.
571  // Figure out whether EFLAGS operand should LIVE after MI instruction.
572  MachineBasicBlock *BB = MI->getParent();
574 
575  // Scan forward through BB for a use/def of EFLAGS.
576  for (auto I = std::next(ItrMI), E = BB->end(); I != E; ++I) {
577  if (I->readsRegister(X86::EFLAGS))
578  return true;
579  if (I->definesRegister(X86::EFLAGS))
580  return false;
581  }
582 
583  // We hit the end of the block, check whether EFLAGS is live into a successor.
584  for (auto I = BB->succ_begin(), E = BB->succ_end(); I != E; ++I) {
585  if ((*I)->isLiveIn(X86::EFLAGS))
586  return true;
587  }
588 
589  return false;
590 }
591 
592 /// Given /p First CMOV instruction and /p Last CMOV instruction representing a
593 /// group of CMOV instructions, which may contain debug instructions in between,
594 /// move all debug instructions to after the last CMOV instruction, making the
595 /// CMOV group consecutive.
596 static void packCmovGroup(MachineInstr *First, MachineInstr *Last) {
598  "Last instruction in a CMOV group must be a CMOV instruction");
599 
600  SmallVector<MachineInstr *, 2> DBGInstructions;
601  for (auto I = First->getIterator(), E = Last->getIterator(); I != E; I++) {
602  if (I->isDebugInstr())
603  DBGInstructions.push_back(&*I);
604  }
605 
606  // Splice the debug instruction after the cmov group.
607  MachineBasicBlock *MBB = First->getParent();
608  for (auto *MI : DBGInstructions)
609  MBB->insertAfter(Last, MI->removeFromParent());
610 }
611 
612 void X86CmovConverterPass::convertCmovInstsToBranches(
613  SmallVectorImpl<MachineInstr *> &Group) const {
614  assert(!Group.empty() && "No CMOV instructions to convert");
615  ++NumOfOptimizedCmovGroups;
616 
617  // If the CMOV group is not packed, e.g., there are debug instructions between
618  // first CMOV and last CMOV, then pack the group and make the CMOV instruction
619  // consecutive by moving the debug instructions to after the last CMOV.
620  packCmovGroup(Group.front(), Group.back());
621 
622  // To convert a CMOVcc instruction, we actually have to insert the diamond
623  // control-flow pattern. The incoming instruction knows the destination vreg
624  // to set, the condition code register to branch on, the true/false values to
625  // select between, and a branch opcode to use.
626 
627  // Before
628  // -----
629  // MBB:
630  // cond = cmp ...
631  // v1 = CMOVge t1, f1, cond
632  // v2 = CMOVlt t2, f2, cond
633  // v3 = CMOVge v1, f3, cond
634  //
635  // After
636  // -----
637  // MBB:
638  // cond = cmp ...
639  // jge %SinkMBB
640  //
641  // FalseMBB:
642  // jmp %SinkMBB
643  //
644  // SinkMBB:
645  // %v1 = phi[%f1, %FalseMBB], [%t1, %MBB]
646  // %v2 = phi[%t2, %FalseMBB], [%f2, %MBB] ; For CMOV with OppCC switch
647  // ; true-value with false-value
648  // %v3 = phi[%f3, %FalseMBB], [%t1, %MBB] ; Phi instruction cannot use
649  // ; previous Phi instruction result
650 
651  MachineInstr &MI = *Group.front();
652  MachineInstr *LastCMOV = Group.back();
653  DebugLoc DL = MI.getDebugLoc();
654 
657  // Potentially swap the condition codes so that any memory operand to a CMOV
658  // is in the *false* position instead of the *true* position. We can invert
659  // any non-memory operand CMOV instructions to cope with this and we ensure
660  // memory operand CMOVs are only included with a single condition code.
661  if (llvm::any_of(Group, [&](MachineInstr *I) {
662  return I->mayLoad() && X86::getCondFromCMovOpc(I->getOpcode()) == CC;
663  }))
664  std::swap(CC, OppCC);
665 
666  MachineBasicBlock *MBB = MI.getParent();
668  MachineFunction *F = MBB->getParent();
669  const BasicBlock *BB = MBB->getBasicBlock();
670 
671  MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB);
672  MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(BB);
673  F->insert(It, FalseMBB);
674  F->insert(It, SinkMBB);
675 
676  // If the EFLAGS register isn't dead in the terminator, then claim that it's
677  // live into the sink and copy blocks.
678  if (checkEFLAGSLive(LastCMOV)) {
679  FalseMBB->addLiveIn(X86::EFLAGS);
680  SinkMBB->addLiveIn(X86::EFLAGS);
681  }
682 
683  // Transfer the remainder of BB and its successor edges to SinkMBB.
684  SinkMBB->splice(SinkMBB->begin(), MBB,
685  std::next(MachineBasicBlock::iterator(LastCMOV)), MBB->end());
686  SinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
687 
688  // Add the false and sink blocks as its successors.
689  MBB->addSuccessor(FalseMBB);
690  MBB->addSuccessor(SinkMBB);
691 
692  // Create the conditional branch instruction.
693  BuildMI(MBB, DL, TII->get(X86::GetCondBranchFromCond(CC))).addMBB(SinkMBB);
694 
695  // Add the sink block to the false block successors.
696  FalseMBB->addSuccessor(SinkMBB);
697 
701  std::next(MachineBasicBlock::iterator(LastCMOV));
702  MachineBasicBlock::iterator FalseInsertionPoint = FalseMBB->begin();
703  MachineBasicBlock::iterator SinkInsertionPoint = SinkMBB->begin();
704 
705  // First we need to insert an explicit load on the false path for any memory
706  // operand. We also need to potentially do register rewriting here, but it is
707  // simpler as the memory operands are always on the false path so we can
708  // simply take that input, whatever it is.
709  DenseMap<unsigned, unsigned> FalseBBRegRewriteTable;
710  for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd;) {
711  auto &MI = *MIIt++;
712  // Skip any CMOVs in this group which don't load from memory.
713  if (!MI.mayLoad()) {
714  // Remember the false-side register input.
715  unsigned FalseReg =
716  MI.getOperand(X86::getCondFromCMovOpc(MI.getOpcode()) == CC ? 1 : 2)
717  .getReg();
718  // Walk back through any intermediate cmovs referenced.
719  while (true) {
720  auto FRIt = FalseBBRegRewriteTable.find(FalseReg);
721  if (FRIt == FalseBBRegRewriteTable.end())
722  break;
723  FalseReg = FRIt->second;
724  }
725  FalseBBRegRewriteTable[MI.getOperand(0).getReg()] = FalseReg;
726  continue;
727  }
728 
729  // The condition must be the *opposite* of the one we've decided to branch
730  // on as the branch will go *around* the load and the load should happen
731  // when the CMOV condition is false.
732  assert(X86::getCondFromCMovOpc(MI.getOpcode()) == OppCC &&
733  "Can only handle memory-operand cmov instructions with a condition "
734  "opposite to the selected branch direction.");
735 
736  // The goal is to rewrite the cmov from:
737  //
738  // MBB:
739  // %A = CMOVcc %B (tied), (mem)
740  //
741  // to
742  //
743  // MBB:
744  // %A = CMOVcc %B (tied), %C
745  // FalseMBB:
746  // %C = MOV (mem)
747  //
748  // Which will allow the next loop to rewrite the CMOV in terms of a PHI:
749  //
750  // MBB:
751  // JMP!cc SinkMBB
752  // FalseMBB:
753  // %C = MOV (mem)
754  // SinkMBB:
755  // %A = PHI [ %C, FalseMBB ], [ %B, MBB]
756 
757  // Get a fresh register to use as the destination of the MOV.
758  const TargetRegisterClass *RC = MRI->getRegClass(MI.getOperand(0).getReg());
759  unsigned TmpReg = MRI->createVirtualRegister(RC);
760 
762  bool Unfolded = TII->unfoldMemoryOperand(*MBB->getParent(), MI, TmpReg,
763  /*UnfoldLoad*/ true,
764  /*UnfoldStore*/ false, NewMIs);
765  (void)Unfolded;
766  assert(Unfolded && "Should never fail to unfold a loading cmov!");
767 
768  // Move the new CMOV to just before the old one and reset any impacted
769  // iterator.
770  auto *NewCMOV = NewMIs.pop_back_val();
771  assert(X86::getCondFromCMovOpc(NewCMOV->getOpcode()) == OppCC &&
772  "Last new instruction isn't the expected CMOV!");
773  LLVM_DEBUG(dbgs() << "\tRewritten cmov: "; NewCMOV->dump());
774  MBB->insert(MachineBasicBlock::iterator(MI), NewCMOV);
775  if (&*MIItBegin == &MI)
776  MIItBegin = MachineBasicBlock::iterator(NewCMOV);
777 
778  // Sink whatever instructions were needed to produce the unfolded operand
779  // into the false block.
780  for (auto *NewMI : NewMIs) {
781  LLVM_DEBUG(dbgs() << "\tRewritten load instr: "; NewMI->dump());
782  FalseMBB->insert(FalseInsertionPoint, NewMI);
783  // Re-map any operands that are from other cmovs to the inputs for this block.
784  for (auto &MOp : NewMI->uses()) {
785  if (!MOp.isReg())
786  continue;
787  auto It = FalseBBRegRewriteTable.find(MOp.getReg());
788  if (It == FalseBBRegRewriteTable.end())
789  continue;
790 
791  MOp.setReg(It->second);
792  // This might have been a kill when it referenced the cmov result, but
793  // it won't necessarily be once rewritten.
794  // FIXME: We could potentially improve this by tracking whether the
795  // operand to the cmov was also a kill, and then skipping the PHI node
796  // construction below.
797  MOp.setIsKill(false);
798  }
799  }
801  std::next(MachineBasicBlock::iterator(MI)));
802 
803  // Add this PHI to the rewrite table.
804  FalseBBRegRewriteTable[NewCMOV->getOperand(0).getReg()] = TmpReg;
805  }
806 
807  // As we are creating the PHIs, we have to be careful if there is more than
808  // one. Later CMOVs may reference the results of earlier CMOVs, but later
809  // PHIs have to reference the individual true/false inputs from earlier PHIs.
810  // That also means that PHI construction must work forward from earlier to
811  // later, and that the code must maintain a mapping from earlier PHI's
812  // destination registers, and the registers that went into the PHI.
814 
815  for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
816  unsigned DestReg = MIIt->getOperand(0).getReg();
817  unsigned Op1Reg = MIIt->getOperand(1).getReg();
818  unsigned Op2Reg = MIIt->getOperand(2).getReg();
819 
820  // If this CMOV we are processing is the opposite condition from the jump we
821  // generated, then we have to swap the operands for the PHI that is going to
822  // be generated.
823  if (X86::getCondFromCMovOpc(MIIt->getOpcode()) == OppCC)
824  std::swap(Op1Reg, Op2Reg);
825 
826  auto Op1Itr = RegRewriteTable.find(Op1Reg);
827  if (Op1Itr != RegRewriteTable.end())
828  Op1Reg = Op1Itr->second.first;
829 
830  auto Op2Itr = RegRewriteTable.find(Op2Reg);
831  if (Op2Itr != RegRewriteTable.end())
832  Op2Reg = Op2Itr->second.second;
833 
834  // SinkMBB:
835  // %Result = phi [ %FalseValue, FalseMBB ], [ %TrueValue, MBB ]
836  // ...
837  MIB = BuildMI(*SinkMBB, SinkInsertionPoint, DL, TII->get(X86::PHI), DestReg)
838  .addReg(Op1Reg)
839  .addMBB(FalseMBB)
840  .addReg(Op2Reg)
841  .addMBB(MBB);
842  (void)MIB;
843  LLVM_DEBUG(dbgs() << "\tFrom: "; MIIt->dump());
844  LLVM_DEBUG(dbgs() << "\tTo: "; MIB->dump());
845 
846  // Add this PHI to the rewrite table.
847  RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
848  }
849 
850  // Now remove the CMOV(s).
851  MBB->erase(MIItBegin, MIItEnd);
852 }
853 
854 INITIALIZE_PASS_BEGIN(X86CmovConverterPass, DEBUG_TYPE, "X86 cmov Conversion",
855  false, false)
857 INITIALIZE_PASS_END(X86CmovConverterPass, DEBUG_TYPE, "X86 cmov Conversion",
858  false, false)
859 
861  return new X86CmovConverterPass();
862 }
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:24
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:34
F(f)
void initializeX86CmovConverterPassPass(PassRegistry &)
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
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:409
static bool checkEFLAGSLive(MachineInstr *MI)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
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.
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:419
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:58
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:371
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:1187
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
self_iterator getIterator()
Definition: ilist_node.h:82
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:418
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:847
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:381
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:133
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
iterator end() const
X86 cmov Conversion
const MachineBasicBlock * getParent() const
Definition: MachineInstr.h:254
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:64
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:56
#define I(x, y, z)
Definition: MD5.cpp:58
iterator end()
Definition: DenseMap.h:109
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:807
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:49
const MachineInstrBuilder & addMBB(MachineBasicBlock *MBB, unsigned char TargetFlags=0) const
#define LLVM_DEBUG(X)
Definition: Debug.h:123