1 //===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements inline cost analysis.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/InlineCost.h"
15 #include "llvm/Support/CallSite.h"
16 #include "llvm/CallingConv.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/ADT/SmallPtrSet.h"
21 /// callIsSmall - If a call is likely to lower to a single target instruction,
22 /// or is otherwise deemed small return true.
23 /// TODO: Perhaps calls like memcpy, strcpy, etc?
24 bool llvm::callIsSmall(const Function *F) {
27 if (F->hasLocalLinkage()) return false;
29 if (!F->hasName()) return false;
31 StringRef Name = F->getName();
33 // These will all likely lower to a single selection DAG node.
34 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
35 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
36 Name == "sin" || Name == "sinf" || Name == "sinl" ||
37 Name == "cos" || Name == "cosf" || Name == "cosl" ||
38 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
41 // These are all likely to be optimized into something smaller.
42 if (Name == "pow" || Name == "powf" || Name == "powl" ||
43 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
44 Name == "floor" || Name == "floorf" || Name == "ceil" ||
45 Name == "round" || Name == "ffs" || Name == "ffsl" ||
46 Name == "abs" || Name == "labs" || Name == "llabs")
52 /// analyzeBasicBlock - Fill in the current structure with information gleaned
53 /// from the specified block.
54 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
56 unsigned NumInstsBeforeThisBB = NumInsts;
57 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
59 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
61 // Special handling for calls.
62 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
63 if (isa<DbgInfoIntrinsic>(II))
64 continue; // Debug intrinsics don't count as size.
66 ImmutableCallSite CS(cast<Instruction>(II));
68 // If this function contains a call to setjmp or _setjmp, never inline
69 // it. This is a hack because we depend on the user marking their local
70 // variables as volatile if they are live across a setjmp call, and they
71 // probably won't do this in callers.
72 if (const Function *F = CS.getCalledFunction()) {
73 if (F->isDeclaration() &&
74 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
77 // If this call is to function itself, then the function is recursive.
78 // Inlining it into other functions is a bad idea, because this is
79 // basically just a form of loop peeling, and our metrics aren't useful
81 if (F == BB->getParent())
85 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
86 // Each argument to a call takes on average one instruction to set up.
87 NumInsts += CS.arg_size();
89 // We don't want inline asm to count as a call - that would prevent loop
90 // unrolling. The argument setup cost is still real, though.
91 if (!isa<InlineAsm>(CS.getCalledValue()))
96 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
97 if (!AI->isStaticAlloca())
98 this->usesDynamicAlloca = true;
101 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
104 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
105 // Noop casts, including ptr <-> int, don't count.
106 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
107 isa<PtrToIntInst>(CI))
109 // Result of a cmp instruction is often extended (to be used by other
110 // cmp instructions, logical or return instructions). These are usually
111 // nop on most sane targets.
112 if (isa<CmpInst>(CI->getOperand(0)))
114 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
115 // If a GEP has all constant indices, it will probably be folded with
117 if (GEPI->hasAllConstantIndices())
124 if (isa<ReturnInst>(BB->getTerminator()))
127 // We never want to inline functions that contain an indirectbr. This is
128 // incorrect because all the blockaddress's (in static global initializers
129 // for example) would be referring to the original function, and this indirect
130 // jump would jump from the inlined copy of the function into the original
131 // function which is extremely undefined behavior.
132 if (isa<IndirectBrInst>(BB->getTerminator()))
133 containsIndirectBr = true;
135 // Remember NumInsts for this BB.
136 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
139 // CountCodeReductionForConstant - Figure out an approximation for how many
140 // instructions will be constant folded if the specified value is constant.
142 unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) {
143 unsigned Reduction = 0;
144 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
146 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
147 // We will be able to eliminate all but one of the successors.
148 const TerminatorInst &TI = cast<TerminatorInst>(*U);
149 const unsigned NumSucc = TI.getNumSuccessors();
151 for (unsigned I = 0; I != NumSucc; ++I)
152 Instrs += NumBBInsts[TI.getSuccessor(I)];
153 // We don't know which blocks will be eliminated, so use the average size.
154 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
155 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
156 // Turning an indirect call into a direct call is a BIG win
157 if (CI->getCalledValue() == V)
158 Reduction += InlineConstants::IndirectCallBonus;
159 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
160 // Turning an indirect call into a direct call is a BIG win
161 if (II->getCalledValue() == V)
162 Reduction += InlineConstants::IndirectCallBonus;
164 // Figure out if this instruction will be removed due to simple constant
166 Instruction &Inst = cast<Instruction>(*U);
168 // We can't constant propagate instructions which have effects or
171 // FIXME: It would be nice to capture the fact that a load from a
172 // pointer-to-constant-global is actually a *really* good thing to zap.
173 // Unfortunately, we don't know the pointer that may get propagated here,
174 // so we can't make this decision.
175 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
176 isa<AllocaInst>(Inst))
179 bool AllOperandsConstant = true;
180 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
181 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
182 AllOperandsConstant = false;
186 if (AllOperandsConstant) {
187 // We will get to remove this instruction...
188 Reduction += InlineConstants::InstrCost;
190 // And any other instructions that use it which become constants
192 Reduction += CountCodeReductionForConstant(&Inst);
199 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
200 // the function will be if it is inlined into a context where an argument
201 // becomes an alloca.
203 unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) {
204 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
205 unsigned Reduction = 0;
206 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
207 Instruction *I = cast<Instruction>(*UI);
208 if (isa<LoadInst>(I) || isa<StoreInst>(I))
209 Reduction += InlineConstants::InstrCost;
210 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
211 // If the GEP has variable indices, we won't be able to do much with it.
212 if (GEP->hasAllConstantIndices())
213 Reduction += CountCodeReductionForAlloca(GEP);
214 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
215 // Track pointer through bitcasts.
216 Reduction += CountCodeReductionForAlloca(BCI);
218 // If there is some other strange instruction, we're not going to be able
219 // to do much if we inline this.
227 /// analyzeFunction - Fill in the current structure with information gleaned
228 /// from the specified function.
229 void CodeMetrics::analyzeFunction(Function *F) {
230 // Look at the size of the callee.
231 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
232 analyzeBasicBlock(&*BB);
235 /// analyzeFunction - Fill in the current structure with information gleaned
236 /// from the specified function.
237 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
238 Metrics.analyzeFunction(F);
240 // A function with exactly one return has it removed during the inlining
241 // process (see InlineFunction), so don't count it.
242 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
243 if (Metrics.NumRets==1)
246 // Don't bother calculating argument weights if we are never going to inline
247 // the function anyway.
251 // Check out all of the arguments to the function, figuring out how much
252 // code can be eliminated if one of the arguments is a constant.
253 ArgumentWeights.reserve(F->arg_size());
254 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
255 ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
256 Metrics.CountCodeReductionForAlloca(I)));
259 /// NeverInline - returns true if the function should never be inlined into
261 bool InlineCostAnalyzer::FunctionInfo::NeverInline()
263 return (Metrics.callsSetJmp || Metrics.isRecursive ||
264 Metrics.containsIndirectBr);
267 // getInlineCost - The heuristic used to determine if we should inline the
268 // function call or not.
270 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
271 SmallPtrSet<const Function*, 16> &NeverInline) {
272 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
275 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
277 SmallPtrSet<const Function*, 16> &NeverInline) {
278 Instruction *TheCall = CS.getInstruction();
279 Function *Caller = TheCall->getParent()->getParent();
280 bool isDirectCall = CS.getCalledFunction() == Callee;
282 // Don't inline functions which can be redefined at link-time to mean
283 // something else. Don't inline functions marked noinline or call sites
285 if (Callee->mayBeOverridden() ||
286 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
288 return llvm::InlineCost::getNever();
290 // InlineCost - This value measures how good of an inline candidate this call
291 // site is to inline. A lower inline cost make is more likely for the call to
292 // be inlined. This value may go negative.
296 // If there is only one call of the function, and it has internal linkage,
297 // make it almost guaranteed to be inlined.
299 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
300 InlineCost += InlineConstants::LastCallToStaticBonus;
302 // If this function uses the coldcc calling convention, prefer not to inline
304 if (Callee->getCallingConv() == CallingConv::Cold)
305 InlineCost += InlineConstants::ColdccPenalty;
307 // If the instruction after the call, or if the normal destination of the
308 // invoke is an unreachable instruction, the function is noreturn. As such,
309 // there is little point in inlining this.
310 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
311 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
312 InlineCost += InlineConstants::NoreturnPenalty;
313 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
314 InlineCost += InlineConstants::NoreturnPenalty;
316 // Get information about the callee.
317 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
319 // If we haven't calculated this information yet, do so now.
320 if (CalleeFI->Metrics.NumBlocks == 0)
321 CalleeFI->analyzeFunction(Callee);
323 // If we should never inline this, return a huge cost.
324 if (CalleeFI->NeverInline())
325 return InlineCost::getNever();
327 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
328 // could move this up and avoid computing the FunctionInfo for
329 // things we are going to just return always inline for. This
330 // requires handling setjmp somewhere else, however.
331 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
332 return InlineCost::getAlways();
334 if (CalleeFI->Metrics.usesDynamicAlloca) {
335 // Get infomation about the caller.
336 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
338 // If we haven't calculated this information yet, do so now.
339 if (CallerFI.Metrics.NumBlocks == 0) {
340 CallerFI.analyzeFunction(Caller);
342 // Recompute the CalleeFI pointer, getting Caller could have invalidated
344 CalleeFI = &CachedFunctionInfo[Callee];
347 // Don't inline a callee with dynamic alloca into a caller without them.
348 // Functions containing dynamic alloca's are inefficient in various ways;
349 // don't create more inefficiency.
350 if (!CallerFI.Metrics.usesDynamicAlloca)
351 return InlineCost::getNever();
354 // Add to the inline quality for properties that make the call valuable to
355 // inline. This includes factors that indicate that the result of inlining
356 // the function will be optimizable. Currently this just looks at arguments
357 // passed into the function.
360 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
361 I != E; ++I, ++ArgNo) {
362 // Each argument passed in has a cost at both the caller and the callee
363 // sides. Measurements show that each argument costs about the same as an
365 InlineCost -= InlineConstants::InstrCost;
367 // If an alloca is passed in, inlining this function is likely to allow
368 // significant future optimization possibilities (like scalar promotion, and
369 // scalarization), so encourage the inlining of the function.
371 if (isa<AllocaInst>(I)) {
372 if (ArgNo < CalleeFI->ArgumentWeights.size())
373 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
375 // If this is a constant being passed into the function, use the argument
376 // weights calculated for the callee to determine how much will be folded
377 // away with this information.
378 } else if (isa<Constant>(I)) {
379 if (ArgNo < CalleeFI->ArgumentWeights.size())
380 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
384 // Now that we have considered all of the factors that make the call site more
385 // likely to be inlined, look at factors that make us not want to inline it.
387 // Calls usually take a long time, so they make the inlining gain smaller.
388 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
390 // Look at the size of the callee. Each instruction counts as 5.
391 InlineCost += CalleeFI->Metrics.NumInsts*InlineConstants::InstrCost;
393 return llvm::InlineCost::get(InlineCost);
396 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
397 // higher threshold to determine if the function call should be inlined.
398 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
399 Function *Callee = CS.getCalledFunction();
401 // Get information about the callee.
402 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
404 // If we haven't calculated this information yet, do so now.
405 if (CalleeFI.Metrics.NumBlocks == 0)
406 CalleeFI.analyzeFunction(Callee);
409 // Single BB functions are often written to be inlined.
410 if (CalleeFI.Metrics.NumBlocks == 1)
413 // Be more aggressive if the function contains a good chunk (if it mades up
414 // at least 10% of the instructions) of vector instructions.
415 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
417 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
422 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
425 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
426 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
428 // For small functions we prefer to recalculate the cost for better accuracy.
429 if (CallerMetrics.NumBlocks < 10 || CallerMetrics.NumInsts < 1000) {
430 resetCachedCostInfo(Caller);
434 // For large functions, we can save a lot of computation time by skipping
436 if (CallerMetrics.NumCalls > 0)
437 --CallerMetrics.NumCalls;
439 if (Callee == 0) return;
441 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
443 // If we don't have metrics for the callee, don't recalculate them just to
444 // update an approximation in the caller. Instead, just recalculate the
445 // caller info from scratch.
446 if (CalleeMetrics.NumBlocks == 0) {
447 resetCachedCostInfo(Caller);
451 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
452 // reference isn't invalidated: both were in the DenseMap.
453 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
455 // FIXME: If any of these three are true for the callee, the callee was
456 // not inlined into the caller, so I think they're redundant here.
457 CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp;
458 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
459 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
461 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
462 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
463 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
464 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
465 CallerMetrics.NumRets += CalleeMetrics.NumRets;
467 // analyzeBasicBlock counts each function argument as an inst.
468 if (CallerMetrics.NumInsts >= Callee->arg_size())
469 CallerMetrics.NumInsts -= Callee->arg_size();
471 CallerMetrics.NumInsts = 0;
473 // We are not updating the argument weights. We have already determined that
474 // Caller is a fairly large function, so we accept the loss of precision.
477 /// clear - empty the cache of inline costs
478 void InlineCostAnalyzer::clear() {
479 CachedFunctionInfo.clear();