llvm.org GIT mirror llvm / release_30 lib / Analysis / InlineCost.cpp
release_30

Tree @release_30 (Download .tar.gz)

InlineCost.cpp @release_30raw · history · blame

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
//===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements inline cost analysis.
//
//===----------------------------------------------------------------------===//

#include "llvm/Analysis/InlineCost.h"
#include "llvm/Support/CallSite.h"
#include "llvm/CallingConv.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/SmallPtrSet.h"

using namespace llvm;

/// callIsSmall - If a call is likely to lower to a single target instruction,
/// or is otherwise deemed small return true.
/// TODO: Perhaps calls like memcpy, strcpy, etc?
bool llvm::callIsSmall(const Function *F) {
  if (!F) return false;

  if (F->hasLocalLinkage()) return false;

  if (!F->hasName()) return false;

  StringRef Name = F->getName();

  // These will all likely lower to a single selection DAG node.
  if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
      Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
      Name == "sin" || Name == "sinf" || Name == "sinl" ||
      Name == "cos" || Name == "cosf" || Name == "cosl" ||
      Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
    return true;

  // These are all likely to be optimized into something smaller.
  if (Name == "pow" || Name == "powf" || Name == "powl" ||
      Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
      Name == "floor" || Name == "floorf" || Name == "ceil" ||
      Name == "round" || Name == "ffs" || Name == "ffsl" ||
      Name == "abs" || Name == "labs" || Name == "llabs")
    return true;

  return false;
}

/// analyzeBasicBlock - Fill in the current structure with information gleaned
/// from the specified block.
void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB,
                                    const TargetData *TD) {
  ++NumBlocks;
  unsigned NumInstsBeforeThisBB = NumInsts;
  for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
       II != E; ++II) {
    if (isa<PHINode>(II)) continue;           // PHI nodes don't count.

    // Special handling for calls.
    if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
      if (isa<DbgInfoIntrinsic>(II))
        continue;  // Debug intrinsics don't count as size.

      ImmutableCallSite CS(cast<Instruction>(II));

      if (const Function *F = CS.getCalledFunction()) {
        // If a function is both internal and has a single use, then it is
        // extremely likely to get inlined in the future (it was probably
        // exposed by an interleaved devirtualization pass).
        if (F->hasInternalLinkage() && F->hasOneUse())
          ++NumInlineCandidates;

        // If this call is to function itself, then the function is recursive.
        // Inlining it into other functions is a bad idea, because this is
        // basically just a form of loop peeling, and our metrics aren't useful
        // for that case.
        if (F == BB->getParent())
          isRecursive = true;
      }

      if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
        // Each argument to a call takes on average one instruction to set up.
        NumInsts += CS.arg_size();

        // We don't want inline asm to count as a call - that would prevent loop
        // unrolling. The argument setup cost is still real, though.
        if (!isa<InlineAsm>(CS.getCalledValue()))
          ++NumCalls;
      }
    }

    if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
      if (!AI->isStaticAlloca())
        this->usesDynamicAlloca = true;
    }

    if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
      ++NumVectorInsts;

    if (const CastInst *CI = dyn_cast<CastInst>(II)) {
      // Noop casts, including ptr <-> int,  don't count.
      if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
          isa<PtrToIntInst>(CI))
        continue;
      // trunc to a native type is free (assuming the target has compare and
      // shift-right of the same width).
      if (isa<TruncInst>(CI) && TD &&
          TD->isLegalInteger(TD->getTypeSizeInBits(CI->getType())))
        continue;
      // Result of a cmp instruction is often extended (to be used by other
      // cmp instructions, logical or return instructions). These are usually
      // nop on most sane targets.
      if (isa<CmpInst>(CI->getOperand(0)))
        continue;
    } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
      // If a GEP has all constant indices, it will probably be folded with
      // a load/store.
      if (GEPI->hasAllConstantIndices())
        continue;
    }

    ++NumInsts;
  }

  if (isa<ReturnInst>(BB->getTerminator()))
    ++NumRets;

  // We never want to inline functions that contain an indirectbr.  This is
  // incorrect because all the blockaddress's (in static global initializers
  // for example) would be referring to the original function, and this indirect
  // jump would jump from the inlined copy of the function into the original
  // function which is extremely undefined behavior.
  if (isa<IndirectBrInst>(BB->getTerminator()))
    containsIndirectBr = true;

  // Remember NumInsts for this BB.
  NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
}

// CountCodeReductionForConstant - Figure out an approximation for how many
// instructions will be constant folded if the specified value is constant.
//
unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) {
  unsigned Reduction = 0;
  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    User *U = *UI;
    if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
      // We will be able to eliminate all but one of the successors.
      const TerminatorInst &TI = cast<TerminatorInst>(*U);
      const unsigned NumSucc = TI.getNumSuccessors();
      unsigned Instrs = 0;
      for (unsigned I = 0; I != NumSucc; ++I)
        Instrs += NumBBInsts[TI.getSuccessor(I)];
      // We don't know which blocks will be eliminated, so use the average size.
      Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
    } else {
      // Figure out if this instruction will be removed due to simple constant
      // propagation.
      Instruction &Inst = cast<Instruction>(*U);

      // We can't constant propagate instructions which have effects or
      // read memory.
      //
      // FIXME: It would be nice to capture the fact that a load from a
      // pointer-to-constant-global is actually a *really* good thing to zap.
      // Unfortunately, we don't know the pointer that may get propagated here,
      // so we can't make this decision.
      if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
          isa<AllocaInst>(Inst))
        continue;

      bool AllOperandsConstant = true;
      for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
        if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
          AllOperandsConstant = false;
          break;
        }

      if (AllOperandsConstant) {
        // We will get to remove this instruction...
        Reduction += InlineConstants::InstrCost;

        // And any other instructions that use it which become constants
        // themselves.
        Reduction += CountCodeReductionForConstant(&Inst);
      }
    }
  }
  return Reduction;
}

// CountCodeReductionForAlloca - Figure out an approximation of how much smaller
// the function will be if it is inlined into a context where an argument
// becomes an alloca.
//
unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) {
  if (!V->getType()->isPointerTy()) return 0;  // Not a pointer
  unsigned Reduction = 0;
  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    Instruction *I = cast<Instruction>(*UI);
    if (isa<LoadInst>(I) || isa<StoreInst>(I))
      Reduction += InlineConstants::InstrCost;
    else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
      // If the GEP has variable indices, we won't be able to do much with it.
      if (GEP->hasAllConstantIndices())
        Reduction += CountCodeReductionForAlloca(GEP);
    } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
      // Track pointer through bitcasts.
      Reduction += CountCodeReductionForAlloca(BCI);
    } else {
      // If there is some other strange instruction, we're not going to be able
      // to do much if we inline this.
      return 0;
    }
  }

  return Reduction;
}

/// analyzeFunction - Fill in the current structure with information gleaned
/// from the specified function.
void CodeMetrics::analyzeFunction(Function *F, const TargetData *TD) {
  // If this function contains a call to setjmp or _setjmp, never inline
  // it.  This is a hack because we depend on the user marking their local
  // variables as volatile if they are live across a setjmp call, and they
  // probably won't do this in callers.
  if (F->callsFunctionThatReturnsTwice())
    callsSetJmp = true;

  // Look at the size of the callee.
  for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
    analyzeBasicBlock(&*BB, TD);
}

/// analyzeFunction - Fill in the current structure with information gleaned
/// from the specified function.
void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F,
                                                       const TargetData *TD) {
  Metrics.analyzeFunction(F, TD);

  // A function with exactly one return has it removed during the inlining
  // process (see InlineFunction), so don't count it.
  // FIXME: This knowledge should really be encoded outside of FunctionInfo.
  if (Metrics.NumRets==1)
    --Metrics.NumInsts;

  // Check out all of the arguments to the function, figuring out how much
  // code can be eliminated if one of the arguments is a constant.
  ArgumentWeights.reserve(F->arg_size());
  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
    ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
                                      Metrics.CountCodeReductionForAlloca(I)));
}

/// NeverInline - returns true if the function should never be inlined into
/// any caller
bool InlineCostAnalyzer::FunctionInfo::NeverInline() {
  return (Metrics.callsSetJmp || Metrics.isRecursive ||
          Metrics.containsIndirectBr);
}
// getSpecializationBonus - The heuristic used to determine the per-call
// performance boost for using a specialization of Callee with argument
// specializedArgNo replaced by a constant.
int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
         SmallVectorImpl<unsigned> &SpecializedArgNos)
{
  if (Callee->mayBeOverridden())
    return 0;

  int Bonus = 0;
  // If this function uses the coldcc calling convention, prefer not to
  // specialize it.
  if (Callee->getCallingConv() == CallingConv::Cold)
    Bonus -= InlineConstants::ColdccPenalty;

  // Get information about the callee.
  FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI->Metrics.NumBlocks == 0)
    CalleeFI->analyzeFunction(Callee, TD);

  unsigned ArgNo = 0;
  unsigned i = 0;
  for (Function::arg_iterator I = Callee->arg_begin(), E = Callee->arg_end();
       I != E; ++I, ++ArgNo)
    if (ArgNo == SpecializedArgNos[i]) {
      ++i;
      Bonus += CountBonusForConstant(I);
    }

  // Calls usually take a long time, so they make the specialization gain
  // smaller.
  Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;

  return Bonus;
}

// ConstantFunctionBonus - Figure out how much of a bonus we can get for
// possibly devirtualizing a function. We'll subtract the size of the function
// we may wish to inline from the indirect call bonus providing a limit on
// growth. Leave an upper limit of 0 for the bonus - we don't want to penalize
// inlining because we decide we don't want to give a bonus for
// devirtualizing.
int InlineCostAnalyzer::ConstantFunctionBonus(CallSite CS, Constant *C) {

  // This could just be NULL.
  if (!C) return 0;

  Function *F = dyn_cast<Function>(C);
  if (!F) return 0;

  int Bonus = InlineConstants::IndirectCallBonus + getInlineSize(CS, F);
  return (Bonus > 0) ? 0 : Bonus;
}

// CountBonusForConstant - Figure out an approximation for how much per-call
// performance boost we can expect if the specified value is constant.
int InlineCostAnalyzer::CountBonusForConstant(Value *V, Constant *C) {
  unsigned Bonus = 0;
  for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    User *U = *UI;
    if (CallInst *CI = dyn_cast<CallInst>(U)) {
      // Turning an indirect call into a direct call is a BIG win
      if (CI->getCalledValue() == V)
        Bonus += ConstantFunctionBonus(CallSite(CI), C);
    } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
      // Turning an indirect call into a direct call is a BIG win
      if (II->getCalledValue() == V)
        Bonus += ConstantFunctionBonus(CallSite(II), C);
    }
    // FIXME: Eliminating conditional branches and switches should
    // also yield a per-call performance boost.
    else {
      // Figure out the bonuses that wll accrue due to simple constant
      // propagation.
      Instruction &Inst = cast<Instruction>(*U);

      // We can't constant propagate instructions which have effects or
      // read memory.
      //
      // FIXME: It would be nice to capture the fact that a load from a
      // pointer-to-constant-global is actually a *really* good thing to zap.
      // Unfortunately, we don't know the pointer that may get propagated here,
      // so we can't make this decision.
      if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
          isa<AllocaInst>(Inst))
        continue;

      bool AllOperandsConstant = true;
      for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
        if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
          AllOperandsConstant = false;
          break;
        }

      if (AllOperandsConstant)
        Bonus += CountBonusForConstant(&Inst);
    }
  }

  return Bonus;
}

int InlineCostAnalyzer::getInlineSize(CallSite CS, Function *Callee) {
  // Get information about the callee.
  FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI->Metrics.NumBlocks == 0)
    CalleeFI->analyzeFunction(Callee, TD);

  // InlineCost - This value measures how good of an inline candidate this call
  // site is to inline.  A lower inline cost make is more likely for the call to
  // be inlined.  This value may go negative.
  //
  int InlineCost = 0;

  // Compute any size reductions we can expect due to arguments being passed into
  // the function.
  //
  unsigned ArgNo = 0;
  CallSite::arg_iterator I = CS.arg_begin();
  for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
       FI != FE; ++I, ++FI, ++ArgNo) {

    // If an alloca is passed in, inlining this function is likely to allow
    // significant future optimization possibilities (like scalar promotion, and
    // scalarization), so encourage the inlining of the function.
    //
    if (isa<AllocaInst>(I))
      InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;

    // If this is a constant being passed into the function, use the argument
    // weights calculated for the callee to determine how much will be folded
    // away with this information.
    else if (isa<Constant>(I))
      InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
  }

  // Each argument passed in has a cost at both the caller and the callee
  // sides.  Measurements show that each argument costs about the same as an
  // instruction.
  InlineCost -= (CS.arg_size() * InlineConstants::InstrCost);

  // Now that we have considered all of the factors that make the call site more
  // likely to be inlined, look at factors that make us not want to inline it.

  // Calls usually take a long time, so they make the inlining gain smaller.
  InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;

  // Look at the size of the callee. Each instruction counts as 5.
  InlineCost += CalleeFI->Metrics.NumInsts*InlineConstants::InstrCost;

  return InlineCost;
}

int InlineCostAnalyzer::getInlineBonuses(CallSite CS, Function *Callee) {
  // Get information about the callee.
  FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI->Metrics.NumBlocks == 0)
    CalleeFI->analyzeFunction(Callee, TD);

  bool isDirectCall = CS.getCalledFunction() == Callee;
  Instruction *TheCall = CS.getInstruction();
  int Bonus = 0;

  // If there is only one call of the function, and it has internal linkage,
  // make it almost guaranteed to be inlined.
  //
  if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
    Bonus += InlineConstants::LastCallToStaticBonus;

  // If the instruction after the call, or if the normal destination of the
  // invoke is an unreachable instruction, the function is noreturn.  As such,
  // there is little point in inlining this.
  if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
    if (isa<UnreachableInst>(II->getNormalDest()->begin()))
      Bonus += InlineConstants::NoreturnPenalty;
  } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
    Bonus += InlineConstants::NoreturnPenalty;

  // If this function uses the coldcc calling convention, prefer not to inline
  // it.
  if (Callee->getCallingConv() == CallingConv::Cold)
    Bonus += InlineConstants::ColdccPenalty;

  // Add to the inline quality for properties that make the call valuable to
  // inline.  This includes factors that indicate that the result of inlining
  // the function will be optimizable.  Currently this just looks at arguments
  // passed into the function.
  //
  CallSite::arg_iterator I = CS.arg_begin();
  for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
       FI != FE; ++I, ++FI)
    // Compute any constant bonus due to inlining we want to give here.
    if (isa<Constant>(I))
      Bonus += CountBonusForConstant(FI, cast<Constant>(I));

  return Bonus;
}

// getInlineCost - The heuristic used to determine if we should inline the
// function call or not.
//
InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
                               SmallPtrSet<const Function*, 16> &NeverInline) {
  return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
}

InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
                               Function *Callee,
                               SmallPtrSet<const Function*, 16> &NeverInline) {
  Instruction *TheCall = CS.getInstruction();
  Function *Caller = TheCall->getParent()->getParent();

  // Don't inline functions which can be redefined at link-time to mean
  // something else.  Don't inline functions marked noinline or call sites
  // marked noinline.
  if (Callee->mayBeOverridden() ||
      Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
      CS.isNoInline())
    return llvm::InlineCost::getNever();

  // Get information about the callee.
  FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI->Metrics.NumBlocks == 0)
    CalleeFI->analyzeFunction(Callee, TD);

  // If we should never inline this, return a huge cost.
  if (CalleeFI->NeverInline())
    return InlineCost::getNever();

  // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
  // could move this up and avoid computing the FunctionInfo for
  // things we are going to just return always inline for. This
  // requires handling setjmp somewhere else, however.
  if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
    return InlineCost::getAlways();

  if (CalleeFI->Metrics.usesDynamicAlloca) {
    // Get information about the caller.
    FunctionInfo &CallerFI = CachedFunctionInfo[Caller];

    // If we haven't calculated this information yet, do so now.
    if (CallerFI.Metrics.NumBlocks == 0) {
      CallerFI.analyzeFunction(Caller, TD);

      // Recompute the CalleeFI pointer, getting Caller could have invalidated
      // it.
      CalleeFI = &CachedFunctionInfo[Callee];
    }

    // Don't inline a callee with dynamic alloca into a caller without them.
    // Functions containing dynamic alloca's are inefficient in various ways;
    // don't create more inefficiency.
    if (!CallerFI.Metrics.usesDynamicAlloca)
      return InlineCost::getNever();
  }

  // InlineCost - This value measures how good of an inline candidate this call
  // site is to inline.  A lower inline cost make is more likely for the call to
  // be inlined.  This value may go negative due to the fact that bonuses
  // are negative numbers.
  //
  int InlineCost = getInlineSize(CS, Callee) + getInlineBonuses(CS, Callee);
  return llvm::InlineCost::get(InlineCost);
}

// getSpecializationCost - The heuristic used to determine the code-size
// impact of creating a specialized version of Callee with argument
// SpecializedArgNo replaced by a constant.
InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
                               SmallVectorImpl<unsigned> &SpecializedArgNos)
{
  // Don't specialize functions which can be redefined at link-time to mean
  // something else.
  if (Callee->mayBeOverridden())
    return llvm::InlineCost::getNever();

  // Get information about the callee.
  FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI->Metrics.NumBlocks == 0)
    CalleeFI->analyzeFunction(Callee, TD);

  int Cost = 0;

  // Look at the original size of the callee.  Each instruction counts as 5.
  Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;

  // Offset that with the amount of code that can be constant-folded
  // away with the given arguments replaced by constants.
  for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
       ae = SpecializedArgNos.end(); an != ae; ++an)
    Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;

  return llvm::InlineCost::get(Cost);
}

// getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
// higher threshold to determine if the function call should be inlined.
float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
  Function *Callee = CS.getCalledFunction();

  // Get information about the callee.
  FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];

  // If we haven't calculated this information yet, do so now.
  if (CalleeFI.Metrics.NumBlocks == 0)
    CalleeFI.analyzeFunction(Callee, TD);

  float Factor = 1.0f;
  // Single BB functions are often written to be inlined.
  if (CalleeFI.Metrics.NumBlocks == 1)
    Factor += 0.5f;

  // Be more aggressive if the function contains a good chunk (if it mades up
  // at least 10% of the instructions) of vector instructions.
  if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
    Factor += 2.0f;
  else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
    Factor += 1.5f;
  return Factor;
}

/// growCachedCostInfo - update the cached cost info for Caller after Callee has
/// been inlined.
void
InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
  CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;

  // For small functions we prefer to recalculate the cost for better accuracy.
  if (CallerMetrics.NumBlocks < 10 && CallerMetrics.NumInsts < 1000) {
    resetCachedCostInfo(Caller);
    return;
  }

  // For large functions, we can save a lot of computation time by skipping
  // recalculations.
  if (CallerMetrics.NumCalls > 0)
    --CallerMetrics.NumCalls;

  if (Callee == 0) return;

  CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;

  // If we don't have metrics for the callee, don't recalculate them just to
  // update an approximation in the caller.  Instead, just recalculate the
  // caller info from scratch.
  if (CalleeMetrics.NumBlocks == 0) {
    resetCachedCostInfo(Caller);
    return;
  }

  // Since CalleeMetrics were already calculated, we know that the CallerMetrics
  // reference isn't invalidated: both were in the DenseMap.
  CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;

  // FIXME: If any of these three are true for the callee, the callee was
  // not inlined into the caller, so I think they're redundant here.
  CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp;
  CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
  CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;

  CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
  CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
  CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
  CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
  CallerMetrics.NumRets += CalleeMetrics.NumRets;

  // analyzeBasicBlock counts each function argument as an inst.
  if (CallerMetrics.NumInsts >= Callee->arg_size())
    CallerMetrics.NumInsts -= Callee->arg_size();
  else
    CallerMetrics.NumInsts = 0;

  // We are not updating the argument weights. We have already determined that
  // Caller is a fairly large function, so we accept the loss of precision.
}

/// clear - empty the cache of inline costs
void InlineCostAnalyzer::clear() {
  CachedFunctionInfo.clear();
}