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 // CountCodeReductionForConstant - Figure out an approximation for how many
22 // instructions will be constant folded if the specified value is constant.
24 unsigned InlineCostAnalyzer::FunctionInfo::
25 CountCodeReductionForConstant(Value *V) {
26 unsigned Reduction = 0;
27 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
29 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
30 // We will be able to eliminate all but one of the successors.
31 const TerminatorInst &TI = cast<TerminatorInst>(*U);
32 const unsigned NumSucc = TI.getNumSuccessors();
34 for (unsigned I = 0; I != NumSucc; ++I)
35 Instrs += Metrics.NumBBInsts[TI.getSuccessor(I)];
36 // We don't know which blocks will be eliminated, so use the average size.
37 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
38 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
39 // Turning an indirect call into a direct call is a BIG win
40 if (CI->getCalledValue() == V)
41 Reduction += InlineConstants::IndirectCallBonus;
42 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
43 // Turning an indirect call into a direct call is a BIG win
44 if (II->getCalledValue() == V)
45 Reduction += InlineConstants::IndirectCallBonus;
47 // Figure out if this instruction will be removed due to simple constant
49 Instruction &Inst = cast<Instruction>(*U);
51 // We can't constant propagate instructions which have effects or
54 // FIXME: It would be nice to capture the fact that a load from a
55 // pointer-to-constant-global is actually a *really* good thing to zap.
56 // Unfortunately, we don't know the pointer that may get propagated here,
57 // so we can't make this decision.
58 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
59 isa<AllocaInst>(Inst))
62 bool AllOperandsConstant = true;
63 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
64 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
65 AllOperandsConstant = false;
69 if (AllOperandsConstant) {
70 // We will get to remove this instruction...
71 Reduction += InlineConstants::InstrCost;
73 // And any other instructions that use it which become constants
75 Reduction += CountCodeReductionForConstant(&Inst);
82 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
83 // the function will be if it is inlined into a context where an argument
86 unsigned InlineCostAnalyzer::FunctionInfo::
87 CountCodeReductionForAlloca(Value *V) {
88 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
89 unsigned Reduction = 0;
90 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
91 Instruction *I = cast<Instruction>(*UI);
92 if (isa<LoadInst>(I) || isa<StoreInst>(I))
93 Reduction += InlineConstants::InstrCost;
94 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
95 // If the GEP has variable indices, we won't be able to do much with it.
96 if (GEP->hasAllConstantIndices())
97 Reduction += CountCodeReductionForAlloca(GEP);
98 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
99 // Track pointer through bitcasts.
100 Reduction += CountCodeReductionForAlloca(BCI);
102 // If there is some other strange instruction, we're not going to be able
103 // to do much if we inline this.
111 /// callIsSmall - If a call is likely to lower to a single target instruction,
112 /// or is otherwise deemed small return true.
113 /// TODO: Perhaps calls like memcpy, strcpy, etc?
114 bool llvm::callIsSmall(const Function *F) {
115 if (!F) return false;
117 if (F->hasLocalLinkage()) return false;
119 if (!F->hasName()) return false;
121 StringRef Name = F->getName();
123 // These will all likely lower to a single selection DAG node.
124 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
125 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
126 Name == "sin" || Name == "sinf" || Name == "sinl" ||
127 Name == "cos" || Name == "cosf" || Name == "cosl" ||
128 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
131 // These are all likely to be optimized into something smaller.
132 if (Name == "pow" || Name == "powf" || Name == "powl" ||
133 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
134 Name == "floor" || Name == "floorf" || Name == "ceil" ||
135 Name == "round" || Name == "ffs" || Name == "ffsl" ||
136 Name == "abs" || Name == "labs" || Name == "llabs")
142 /// analyzeBasicBlock - Fill in the current structure with information gleaned
143 /// from the specified block.
144 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
146 unsigned NumInstsBeforeThisBB = NumInsts;
147 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
149 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
151 // Special handling for calls.
152 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
153 if (isa<DbgInfoIntrinsic>(II))
154 continue; // Debug intrinsics don't count as size.
156 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
158 // If this function contains a call to setjmp or _setjmp, never inline
159 // it. This is a hack because we depend on the user marking their local
160 // variables as volatile if they are live across a setjmp call, and they
161 // probably won't do this in callers.
162 if (Function *F = CS.getCalledFunction()) {
163 if (F->isDeclaration() &&
164 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
167 // If this call is to function itself, then the function is recursive.
168 // Inlining it into other functions is a bad idea, because this is
169 // basically just a form of loop peeling, and our metrics aren't useful
171 if (F == BB->getParent())
175 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
176 // Each argument to a call takes on average one instruction to set up.
177 NumInsts += CS.arg_size();
179 // We don't want inline asm to count as a call - that would prevent loop
180 // unrolling. The argument setup cost is still real, though.
181 if (!isa<InlineAsm>(CS.getCalledValue()))
186 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
187 if (!AI->isStaticAlloca())
188 this->usesDynamicAlloca = true;
191 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
194 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
195 // Noop casts, including ptr <-> int, don't count.
196 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
197 isa<PtrToIntInst>(CI))
199 // Result of a cmp instruction is often extended (to be used by other
200 // cmp instructions, logical or return instructions). These are usually
201 // nop on most sane targets.
202 if (isa<CmpInst>(CI->getOperand(0)))
204 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
205 // If a GEP has all constant indices, it will probably be folded with
207 if (GEPI->hasAllConstantIndices())
214 if (isa<ReturnInst>(BB->getTerminator()))
217 // We never want to inline functions that contain an indirectbr. This is
218 // incorrect because all the blockaddress's (in static global initializers
219 // for example) would be referring to the original function, and this indirect
220 // jump would jump from the inlined copy of the function into the original
221 // function which is extremely undefined behavior.
222 if (isa<IndirectBrInst>(BB->getTerminator()))
223 containsIndirectBr = true;
225 // Remember NumInsts for this BB.
226 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
229 /// analyzeFunction - Fill in the current structure with information gleaned
230 /// from the specified function.
231 void CodeMetrics::analyzeFunction(Function *F) {
232 // Look at the size of the callee.
233 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
234 analyzeBasicBlock(&*BB);
237 /// analyzeFunction - Fill in the current structure with information gleaned
238 /// from the specified function.
239 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
240 Metrics.analyzeFunction(F);
242 // A function with exactly one return has it removed during the inlining
243 // process (see InlineFunction), so don't count it.
244 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
245 if (Metrics.NumRets==1)
248 // Don't bother calculating argument weights if we are never going to inline
249 // the function anyway.
253 // Check out all of the arguments to the function, figuring out how much
254 // code can be eliminated if one of the arguments is a constant.
255 ArgumentWeights.reserve(F->arg_size());
256 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
257 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
258 CountCodeReductionForAlloca(I)));
261 /// NeverInline - returns true if the function should never be inlined into
263 bool InlineCostAnalyzer::FunctionInfo::NeverInline()
265 return (Metrics.callsSetJmp || Metrics.isRecursive ||
266 Metrics.containsIndirectBr);
269 // getInlineCost - The heuristic used to determine if we should inline the
270 // function call or not.
272 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
273 SmallPtrSet<const Function*, 16> &NeverInline) {
274 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
277 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
279 SmallPtrSet<const Function*, 16> &NeverInline) {
280 Instruction *TheCall = CS.getInstruction();
281 Function *Caller = TheCall->getParent()->getParent();
282 bool isDirectCall = CS.getCalledFunction() == Callee;
284 // Don't inline functions which can be redefined at link-time to mean
285 // something else. Don't inline functions marked noinline or call sites
287 if (Callee->mayBeOverridden() ||
288 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
290 return llvm::InlineCost::getNever();
292 // InlineCost - This value measures how good of an inline candidate this call
293 // site is to inline. A lower inline cost make is more likely for the call to
294 // be inlined. This value may go negative.
298 // If there is only one call of the function, and it has internal linkage,
299 // make it almost guaranteed to be inlined.
301 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
302 InlineCost += InlineConstants::LastCallToStaticBonus;
304 // If this function uses the coldcc calling convention, prefer not to inline
306 if (Callee->getCallingConv() == CallingConv::Cold)
307 InlineCost += InlineConstants::ColdccPenalty;
309 // If the instruction after the call, or if the normal destination of the
310 // invoke is an unreachable instruction, the function is noreturn. As such,
311 // there is little point in inlining this.
312 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
313 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
314 InlineCost += InlineConstants::NoreturnPenalty;
315 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
316 InlineCost += InlineConstants::NoreturnPenalty;
318 // Get information about the callee.
319 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
321 // If we haven't calculated this information yet, do so now.
322 if (CalleeFI->Metrics.NumBlocks == 0)
323 CalleeFI->analyzeFunction(Callee);
325 // If we should never inline this, return a huge cost.
326 if (CalleeFI->NeverInline())
327 return InlineCost::getNever();
329 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
330 // could move this up and avoid computing the FunctionInfo for
331 // things we are going to just return always inline for. This
332 // requires handling setjmp somewhere else, however.
333 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
334 return InlineCost::getAlways();
336 if (CalleeFI->Metrics.usesDynamicAlloca) {
337 // Get infomation about the caller.
338 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
340 // If we haven't calculated this information yet, do so now.
341 if (CallerFI.Metrics.NumBlocks == 0) {
342 CallerFI.analyzeFunction(Caller);
344 // Recompute the CalleeFI pointer, getting Caller could have invalidated
346 CalleeFI = &CachedFunctionInfo[Callee];
349 // Don't inline a callee with dynamic alloca into a caller without them.
350 // Functions containing dynamic alloca's are inefficient in various ways;
351 // don't create more inefficiency.
352 if (!CallerFI.Metrics.usesDynamicAlloca)
353 return InlineCost::getNever();
356 // Add to the inline quality for properties that make the call valuable to
357 // inline. This includes factors that indicate that the result of inlining
358 // the function will be optimizable. Currently this just looks at arguments
359 // passed into the function.
362 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
363 I != E; ++I, ++ArgNo) {
364 // Each argument passed in has a cost at both the caller and the callee
365 // sides. Measurements show that each argument costs about the same as an
367 InlineCost -= InlineConstants::InstrCost;
369 // If an alloca is passed in, inlining this function is likely to allow
370 // significant future optimization possibilities (like scalar promotion, and
371 // scalarization), so encourage the inlining of the function.
373 if (isa<AllocaInst>(I)) {
374 if (ArgNo < CalleeFI->ArgumentWeights.size())
375 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
377 // If this is a constant being passed into the function, use the argument
378 // weights calculated for the callee to determine how much will be folded
379 // away with this information.
380 } else if (isa<Constant>(I)) {
381 if (ArgNo < CalleeFI->ArgumentWeights.size())
382 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
386 // Now that we have considered all of the factors that make the call site more
387 // likely to be inlined, look at factors that make us not want to inline it.
389 // Calls usually take a long time, so they make the inlining gain smaller.
390 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
392 // Look at the size of the callee. Each instruction counts as 5.
393 InlineCost += CalleeFI->Metrics.NumInsts*InlineConstants::InstrCost;
395 return llvm::InlineCost::get(InlineCost);
398 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
399 // higher threshold to determine if the function call should be inlined.
400 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
401 Function *Callee = CS.getCalledFunction();
403 // Get information about the callee.
404 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
406 // If we haven't calculated this information yet, do so now.
407 if (CalleeFI.Metrics.NumBlocks == 0)
408 CalleeFI.analyzeFunction(Callee);
411 // Single BB functions are often written to be inlined.
412 if (CalleeFI.Metrics.NumBlocks == 1)
415 // Be more aggressive if the function contains a good chunk (if it mades up
416 // at least 10% of the instructions) of vector instructions.
417 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
419 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
424 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
427 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
428 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
430 // For small functions we prefer to recalculate the cost for better accuracy.
431 if (CallerMetrics.NumBlocks < 10 || CallerMetrics.NumInsts < 1000) {
432 resetCachedCostInfo(Caller);
436 // For large functions, we can save a lot of computation time by skipping
438 if (CallerMetrics.NumCalls > 0)
439 --CallerMetrics.NumCalls;
441 if (Callee == 0) return;
443 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
445 // If we don't have metrics for the callee, don't recalculate them just to
446 // update an approximation in the caller. Instead, just recalculate the
447 // caller info from scratch.
448 if (CalleeMetrics.NumBlocks == 0) {
449 resetCachedCostInfo(Caller);
453 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
454 // reference isn't invalidated: both were in the DenseMap.
455 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
457 // FIXME: If any of these three are true for the callee, the callee was
458 // not inlined into the caller, so I think they're redundant here.
459 CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp;
460 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
461 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
463 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
464 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
465 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
466 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
467 CallerMetrics.NumRets += CalleeMetrics.NumRets;
469 // analyzeBasicBlock counts each function argument as an inst.
470 if (CallerMetrics.NumInsts >= Callee->arg_size())
471 CallerMetrics.NumInsts -= Callee->arg_size();
473 CallerMetrics.NumInsts = 0;
475 // We are not updating the argument weights. We have already determined that
476 // Caller is a fairly large function, so we accept the loss of precision.
479 /// clear - empty the cache of inline costs
480 void InlineCostAnalyzer::clear() {
481 CachedFunctionInfo.clear();