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 #define DEBUG_TYPE "inline-cost"
15 #include "llvm/Analysis/InlineCost.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/IR/CallingConv.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Operator.h"
29 #include "llvm/InstVisitor.h"
30 #include "llvm/Support/CallSite.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/raw_ostream.h"
37 STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed");
41 class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
42 typedef InstVisitor<CallAnalyzer, bool> Base;
43 friend class InstVisitor<CallAnalyzer, bool>;
45 // DataLayout if available, or null.
46 const DataLayout *const TD;
48 /// The TargetTransformInfo available for this compilation.
49 const TargetTransformInfo &TTI;
51 // The called function.
57 bool IsCallerRecursive;
59 bool ExposesReturnsTwice;
60 bool HasDynamicAlloca;
61 bool ContainsNoDuplicateCall;
63 /// Number of bytes allocated statically by the callee.
64 uint64_t AllocatedSize;
65 unsigned NumInstructions, NumVectorInstructions;
66 int FiftyPercentVectorBonus, TenPercentVectorBonus;
69 // While we walk the potentially-inlined instructions, we build up and
70 // maintain a mapping of simplified values specific to this callsite. The
71 // idea is to propagate any special information we have about arguments to
72 // this call through the inlinable section of the function, and account for
73 // likely simplifications post-inlining. The most important aspect we track
74 // is CFG altering simplifications -- when we prove a basic block dead, that
75 // can cause dramatic shifts in the cost of inlining a function.
76 DenseMap<Value *, Constant *> SimplifiedValues;
78 // Keep track of the values which map back (through function arguments) to
79 // allocas on the caller stack which could be simplified through SROA.
80 DenseMap<Value *, Value *> SROAArgValues;
82 // The mapping of caller Alloca values to their accumulated cost savings. If
83 // we have to disable SROA for one of the allocas, this tells us how much
84 // cost must be added.
85 DenseMap<Value *, int> SROAArgCosts;
87 // Keep track of values which map to a pointer base and constant offset.
88 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs;
90 // Custom simplification helper routines.
91 bool isAllocaDerivedArg(Value *V);
92 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
93 DenseMap<Value *, int>::iterator &CostIt);
94 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
95 void disableSROA(Value *V);
96 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
98 bool handleSROACandidate(bool IsSROAValid,
99 DenseMap<Value *, int>::iterator CostIt,
100 int InstructionCost);
101 bool isGEPOffsetConstant(GetElementPtrInst &GEP);
102 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
103 bool simplifyCallSite(Function *F, CallSite CS);
104 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
106 // Custom analysis routines.
107 bool analyzeBlock(BasicBlock *BB);
109 // Disable several entry points to the visitor so we don't accidentally use
110 // them by declaring but not defining them here.
111 void visit(Module *); void visit(Module &);
112 void visit(Function *); void visit(Function &);
113 void visit(BasicBlock *); void visit(BasicBlock &);
115 // Provide base case for our instruction visit.
116 bool visitInstruction(Instruction &I);
118 // Our visit overrides.
119 bool visitAlloca(AllocaInst &I);
120 bool visitPHI(PHINode &I);
121 bool visitGetElementPtr(GetElementPtrInst &I);
122 bool visitBitCast(BitCastInst &I);
123 bool visitPtrToInt(PtrToIntInst &I);
124 bool visitIntToPtr(IntToPtrInst &I);
125 bool visitCastInst(CastInst &I);
126 bool visitUnaryInstruction(UnaryInstruction &I);
127 bool visitICmp(ICmpInst &I);
128 bool visitSub(BinaryOperator &I);
129 bool visitBinaryOperator(BinaryOperator &I);
130 bool visitLoad(LoadInst &I);
131 bool visitStore(StoreInst &I);
132 bool visitExtractValue(ExtractValueInst &I);
133 bool visitInsertValue(InsertValueInst &I);
134 bool visitCallSite(CallSite CS);
137 CallAnalyzer(const DataLayout *TD, const TargetTransformInfo &TTI,
138 Function &Callee, int Threshold)
139 : TD(TD), TTI(TTI), F(Callee), Threshold(Threshold), Cost(0),
140 IsCallerRecursive(false), IsRecursiveCall(false),
141 ExposesReturnsTwice(false), HasDynamicAlloca(false),
142 ContainsNoDuplicateCall(false), AllocatedSize(0), NumInstructions(0),
143 NumVectorInstructions(0), FiftyPercentVectorBonus(0),
144 TenPercentVectorBonus(0), VectorBonus(0), NumConstantArgs(0),
145 NumConstantOffsetPtrArgs(0), NumAllocaArgs(0), NumConstantPtrCmps(0),
146 NumConstantPtrDiffs(0), NumInstructionsSimplified(0),
147 SROACostSavings(0), SROACostSavingsLost(0) {}
149 bool analyzeCall(CallSite CS);
151 int getThreshold() { return Threshold; }
152 int getCost() { return Cost; }
154 // Keep a bunch of stats about the cost savings found so we can print them
155 // out when debugging.
156 unsigned NumConstantArgs;
157 unsigned NumConstantOffsetPtrArgs;
158 unsigned NumAllocaArgs;
159 unsigned NumConstantPtrCmps;
160 unsigned NumConstantPtrDiffs;
161 unsigned NumInstructionsSimplified;
162 unsigned SROACostSavings;
163 unsigned SROACostSavingsLost;
170 /// \brief Test whether the given value is an Alloca-derived function argument.
171 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
172 return SROAArgValues.count(V);
175 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
176 /// Returns false if V does not map to a SROA-candidate.
177 bool CallAnalyzer::lookupSROAArgAndCost(
178 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
179 if (SROAArgValues.empty() || SROAArgCosts.empty())
182 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
183 if (ArgIt == SROAArgValues.end())
187 CostIt = SROAArgCosts.find(Arg);
188 return CostIt != SROAArgCosts.end();
191 /// \brief Disable SROA for the candidate marked by this cost iterator.
193 /// This marks the candidate as no longer viable for SROA, and adds the cost
194 /// savings associated with it back into the inline cost measurement.
195 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
196 // If we're no longer able to perform SROA we need to undo its cost savings
197 // and prevent subsequent analysis.
198 Cost += CostIt->second;
199 SROACostSavings -= CostIt->second;
200 SROACostSavingsLost += CostIt->second;
201 SROAArgCosts.erase(CostIt);
204 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
205 void CallAnalyzer::disableSROA(Value *V) {
207 DenseMap<Value *, int>::iterator CostIt;
208 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
212 /// \brief Accumulate the given cost for a particular SROA candidate.
213 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
214 int InstructionCost) {
215 CostIt->second += InstructionCost;
216 SROACostSavings += InstructionCost;
219 /// \brief Helper for the common pattern of handling a SROA candidate.
220 /// Either accumulates the cost savings if the SROA remains valid, or disables
221 /// SROA for the candidate.
222 bool CallAnalyzer::handleSROACandidate(bool IsSROAValid,
223 DenseMap<Value *, int>::iterator CostIt,
224 int InstructionCost) {
226 accumulateSROACost(CostIt, InstructionCost);
234 /// \brief Check whether a GEP's indices are all constant.
236 /// Respects any simplified values known during the analysis of this callsite.
237 bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) {
238 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
239 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
245 /// \brief Accumulate a constant GEP offset into an APInt if possible.
247 /// Returns false if unable to compute the offset for any reason. Respects any
248 /// simplified values known during the analysis of this callsite.
249 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
253 unsigned IntPtrWidth = TD->getPointerSizeInBits();
254 assert(IntPtrWidth == Offset.getBitWidth());
256 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
258 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
260 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
261 OpC = dyn_cast<ConstantInt>(SimpleOp);
264 if (OpC->isZero()) continue;
266 // Handle a struct index, which adds its field offset to the pointer.
267 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
268 unsigned ElementIdx = OpC->getZExtValue();
269 const StructLayout *SL = TD->getStructLayout(STy);
270 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
274 APInt TypeSize(IntPtrWidth, TD->getTypeAllocSize(GTI.getIndexedType()));
275 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
280 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
281 // FIXME: Check whether inlining will turn a dynamic alloca into a static
282 // alloca, and handle that case.
284 // Accumulate the allocated size.
285 if (I.isStaticAlloca()) {
286 Type *Ty = I.getAllocatedType();
287 AllocatedSize += (TD ? TD->getTypeAllocSize(Ty) :
288 Ty->getPrimitiveSizeInBits());
291 // We will happily inline static alloca instructions.
292 if (I.isStaticAlloca())
293 return Base::visitAlloca(I);
295 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
296 // a variety of reasons, and so we would like to not inline them into
297 // functions which don't currently have a dynamic alloca. This simply
298 // disables inlining altogether in the presence of a dynamic alloca.
299 HasDynamicAlloca = true;
303 bool CallAnalyzer::visitPHI(PHINode &I) {
304 // FIXME: We should potentially be tracking values through phi nodes,
305 // especially when they collapse to a single value due to deleted CFG edges
308 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
309 // though we don't want to propagate it's bonuses. The idea is to disable
310 // SROA if it *might* be used in an inappropriate manner.
312 // Phi nodes are always zero-cost.
316 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
318 DenseMap<Value *, int>::iterator CostIt;
319 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(),
322 // Try to fold GEPs of constant-offset call site argument pointers. This
323 // requires target data and inbounds GEPs.
324 if (TD && I.isInBounds()) {
325 // Check if we have a base + offset for the pointer.
326 Value *Ptr = I.getPointerOperand();
327 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
328 if (BaseAndOffset.first) {
329 // Check if the offset of this GEP is constant, and if so accumulate it
331 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
332 // Non-constant GEPs aren't folded, and disable SROA.
338 // Add the result as a new mapping to Base + Offset.
339 ConstantOffsetPtrs[&I] = BaseAndOffset;
341 // Also handle SROA candidates here, we already know that the GEP is
342 // all-constant indexed.
344 SROAArgValues[&I] = SROAArg;
350 if (isGEPOffsetConstant(I)) {
352 SROAArgValues[&I] = SROAArg;
354 // Constant GEPs are modeled as free.
358 // Variable GEPs will require math and will disable SROA.
364 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
365 // Propagate constants through bitcasts.
366 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
368 COp = SimplifiedValues.lookup(I.getOperand(0));
370 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) {
371 SimplifiedValues[&I] = C;
375 // Track base/offsets through casts
376 std::pair<Value *, APInt> BaseAndOffset
377 = ConstantOffsetPtrs.lookup(I.getOperand(0));
378 // Casts don't change the offset, just wrap it up.
379 if (BaseAndOffset.first)
380 ConstantOffsetPtrs[&I] = BaseAndOffset;
382 // Also look for SROA candidates here.
384 DenseMap<Value *, int>::iterator CostIt;
385 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
386 SROAArgValues[&I] = SROAArg;
388 // Bitcasts are always zero cost.
392 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
393 // Propagate constants through ptrtoint.
394 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
396 COp = SimplifiedValues.lookup(I.getOperand(0));
398 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) {
399 SimplifiedValues[&I] = C;
403 // Track base/offset pairs when converted to a plain integer provided the
404 // integer is large enough to represent the pointer.
405 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
406 if (TD && IntegerSize >= TD->getPointerSizeInBits()) {
407 std::pair<Value *, APInt> BaseAndOffset
408 = ConstantOffsetPtrs.lookup(I.getOperand(0));
409 if (BaseAndOffset.first)
410 ConstantOffsetPtrs[&I] = BaseAndOffset;
413 // This is really weird. Technically, ptrtoint will disable SROA. However,
414 // unless that ptrtoint is *used* somewhere in the live basic blocks after
415 // inlining, it will be nuked, and SROA should proceed. All of the uses which
416 // would block SROA would also block SROA if applied directly to a pointer,
417 // and so we can just add the integer in here. The only places where SROA is
418 // preserved either cannot fire on an integer, or won't in-and-of themselves
419 // disable SROA (ext) w/o some later use that we would see and disable.
421 DenseMap<Value *, int>::iterator CostIt;
422 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
423 SROAArgValues[&I] = SROAArg;
425 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
428 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
429 // Propagate constants through ptrtoint.
430 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
432 COp = SimplifiedValues.lookup(I.getOperand(0));
434 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) {
435 SimplifiedValues[&I] = C;
439 // Track base/offset pairs when round-tripped through a pointer without
440 // modifications provided the integer is not too large.
441 Value *Op = I.getOperand(0);
442 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
443 if (TD && IntegerSize <= TD->getPointerSizeInBits()) {
444 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
445 if (BaseAndOffset.first)
446 ConstantOffsetPtrs[&I] = BaseAndOffset;
449 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
451 DenseMap<Value *, int>::iterator CostIt;
452 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
453 SROAArgValues[&I] = SROAArg;
455 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
458 bool CallAnalyzer::visitCastInst(CastInst &I) {
459 // Propagate constants through ptrtoint.
460 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
462 COp = SimplifiedValues.lookup(I.getOperand(0));
464 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
465 SimplifiedValues[&I] = C;
469 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
470 disableSROA(I.getOperand(0));
472 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
475 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
476 Value *Operand = I.getOperand(0);
477 Constant *COp = dyn_cast<Constant>(Operand);
479 COp = SimplifiedValues.lookup(Operand);
481 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(),
483 SimplifiedValues[&I] = C;
487 // Disable any SROA on the argument to arbitrary unary operators.
488 disableSROA(Operand);
493 bool CallAnalyzer::visitICmp(ICmpInst &I) {
494 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
495 // First try to handle simplified comparisons.
496 if (!isa<Constant>(LHS))
497 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
499 if (!isa<Constant>(RHS))
500 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
502 if (Constant *CLHS = dyn_cast<Constant>(LHS))
503 if (Constant *CRHS = dyn_cast<Constant>(RHS))
504 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
505 SimplifiedValues[&I] = C;
509 // Otherwise look for a comparison between constant offset pointers with
511 Value *LHSBase, *RHSBase;
512 APInt LHSOffset, RHSOffset;
513 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
515 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
516 if (RHSBase && LHSBase == RHSBase) {
517 // We have common bases, fold the icmp to a constant based on the
519 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
520 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
521 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
522 SimplifiedValues[&I] = C;
523 ++NumConstantPtrCmps;
529 // If the comparison is an equality comparison with null, we can simplify it
530 // for any alloca-derived argument.
531 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)))
532 if (isAllocaDerivedArg(I.getOperand(0))) {
533 // We can actually predict the result of comparisons between an
534 // alloca-derived value and null. Note that this fires regardless of
536 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
537 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
538 : ConstantInt::getFalse(I.getType());
542 // Finally check for SROA candidates in comparisons.
544 DenseMap<Value *, int>::iterator CostIt;
545 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
546 if (isa<ConstantPointerNull>(I.getOperand(1))) {
547 accumulateSROACost(CostIt, InlineConstants::InstrCost);
557 bool CallAnalyzer::visitSub(BinaryOperator &I) {
558 // Try to handle a special case: we can fold computing the difference of two
559 // constant-related pointers.
560 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
561 Value *LHSBase, *RHSBase;
562 APInt LHSOffset, RHSOffset;
563 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
565 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
566 if (RHSBase && LHSBase == RHSBase) {
567 // We have common bases, fold the subtract to a constant based on the
569 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
570 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
571 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
572 SimplifiedValues[&I] = C;
573 ++NumConstantPtrDiffs;
579 // Otherwise, fall back to the generic logic for simplifying and handling
581 return Base::visitSub(I);
584 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
585 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
586 if (!isa<Constant>(LHS))
587 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
589 if (!isa<Constant>(RHS))
590 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
592 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, TD);
593 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
594 SimplifiedValues[&I] = C;
598 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
605 bool CallAnalyzer::visitLoad(LoadInst &I) {
607 DenseMap<Value *, int>::iterator CostIt;
608 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
610 accumulateSROACost(CostIt, InlineConstants::InstrCost);
620 bool CallAnalyzer::visitStore(StoreInst &I) {
622 DenseMap<Value *, int>::iterator CostIt;
623 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
625 accumulateSROACost(CostIt, InlineConstants::InstrCost);
635 bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) {
636 // Constant folding for extract value is trivial.
637 Constant *C = dyn_cast<Constant>(I.getAggregateOperand());
639 C = SimplifiedValues.lookup(I.getAggregateOperand());
641 SimplifiedValues[&I] = ConstantExpr::getExtractValue(C, I.getIndices());
645 // SROA can look through these but give them a cost.
649 bool CallAnalyzer::visitInsertValue(InsertValueInst &I) {
650 // Constant folding for insert value is trivial.
651 Constant *AggC = dyn_cast<Constant>(I.getAggregateOperand());
653 AggC = SimplifiedValues.lookup(I.getAggregateOperand());
654 Constant *InsertedC = dyn_cast<Constant>(I.getInsertedValueOperand());
656 InsertedC = SimplifiedValues.lookup(I.getInsertedValueOperand());
657 if (AggC && InsertedC) {
658 SimplifiedValues[&I] = ConstantExpr::getInsertValue(AggC, InsertedC,
663 // SROA can look through these but give them a cost.
667 /// \brief Try to simplify a call site.
669 /// Takes a concrete function and callsite and tries to actually simplify it by
670 /// analyzing the arguments and call itself with instsimplify. Returns true if
671 /// it has simplified the callsite to some other entity (a constant), making it
673 bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) {
674 // FIXME: Using the instsimplify logic directly for this is inefficient
675 // because we have to continually rebuild the argument list even when no
676 // simplifications can be performed. Until that is fixed with remapping
677 // inside of instsimplify, directly constant fold calls here.
678 if (!canConstantFoldCallTo(F))
681 // Try to re-map the arguments to constants.
682 SmallVector<Constant *, 4> ConstantArgs;
683 ConstantArgs.reserve(CS.arg_size());
684 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
686 Constant *C = dyn_cast<Constant>(*I);
688 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I));
690 return false; // This argument doesn't map to a constant.
692 ConstantArgs.push_back(C);
694 if (Constant *C = ConstantFoldCall(F, ConstantArgs)) {
695 SimplifiedValues[CS.getInstruction()] = C;
702 bool CallAnalyzer::visitCallSite(CallSite CS) {
703 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->canReturnTwice() &&
704 !F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
705 Attribute::ReturnsTwice)) {
706 // This aborts the entire analysis.
707 ExposesReturnsTwice = true;
711 cast<CallInst>(CS.getInstruction())->hasFnAttr(Attribute::NoDuplicate))
712 ContainsNoDuplicateCall = true;
714 if (Function *F = CS.getCalledFunction()) {
715 // When we have a concrete function, first try to simplify it directly.
716 if (simplifyCallSite(F, CS))
719 // Next check if it is an intrinsic we know about.
720 // FIXME: Lift this into part of the InstVisitor.
721 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
722 switch (II->getIntrinsicID()) {
724 return Base::visitCallSite(CS);
726 case Intrinsic::memset:
727 case Intrinsic::memcpy:
728 case Intrinsic::memmove:
729 // SROA can usually chew through these intrinsics, but they aren't free.
734 if (F == CS.getInstruction()->getParent()->getParent()) {
735 // This flag will fully abort the analysis, so don't bother with anything
737 IsRecursiveCall = true;
741 if (TTI.isLoweredToCall(F)) {
742 // We account for the average 1 instruction per call argument setup
744 Cost += CS.arg_size() * InlineConstants::InstrCost;
746 // Everything other than inline ASM will also have a significant cost
747 // merely from making the call.
748 if (!isa<InlineAsm>(CS.getCalledValue()))
749 Cost += InlineConstants::CallPenalty;
752 return Base::visitCallSite(CS);
755 // Otherwise we're in a very special case -- an indirect function call. See
756 // if we can be particularly clever about this.
757 Value *Callee = CS.getCalledValue();
759 // First, pay the price of the argument setup. We account for the average
760 // 1 instruction per call argument setup here.
761 Cost += CS.arg_size() * InlineConstants::InstrCost;
763 // Next, check if this happens to be an indirect function call to a known
764 // function in this inline context. If not, we've done all we can.
765 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
767 return Base::visitCallSite(CS);
769 // If we have a constant that we are calling as a function, we can peer
770 // through it and see the function target. This happens not infrequently
771 // during devirtualization and so we want to give it a hefty bonus for
772 // inlining, but cap that bonus in the event that inlining wouldn't pan
773 // out. Pretend to inline the function, with a custom threshold.
774 CallAnalyzer CA(TD, TTI, *F, InlineConstants::IndirectCallThreshold);
775 if (CA.analyzeCall(CS)) {
776 // We were able to inline the indirect call! Subtract the cost from the
777 // bonus we want to apply, but don't go below zero.
778 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost());
781 return Base::visitCallSite(CS);
784 bool CallAnalyzer::visitInstruction(Instruction &I) {
785 // Some instructions are free. All of the free intrinsics can also be
786 // handled by SROA, etc.
787 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I))
790 // We found something we don't understand or can't handle. Mark any SROA-able
791 // values in the operand list as no longer viable.
792 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
799 /// \brief Analyze a basic block for its contribution to the inline cost.
801 /// This method walks the analyzer over every instruction in the given basic
802 /// block and accounts for their cost during inlining at this callsite. It
803 /// aborts early if the threshold has been exceeded or an impossible to inline
804 /// construct has been detected. It returns false if inlining is no longer
805 /// viable, and true if inlining remains viable.
806 bool CallAnalyzer::analyzeBlock(BasicBlock *BB) {
807 for (BasicBlock::iterator I = BB->begin(), E = llvm::prior(BB->end());
810 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
811 ++NumVectorInstructions;
813 // If the instruction simplified to a constant, there is no cost to this
814 // instruction. Visit the instructions using our InstVisitor to account for
815 // all of the per-instruction logic. The visit tree returns true if we
816 // consumed the instruction in any way, and false if the instruction's base
817 // cost should count against inlining.
819 ++NumInstructionsSimplified;
821 Cost += InlineConstants::InstrCost;
823 // If the visit this instruction detected an uninlinable pattern, abort.
824 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca)
827 // If the caller is a recursive function then we don't want to inline
828 // functions which allocate a lot of stack space because it would increase
829 // the caller stack usage dramatically.
830 if (IsCallerRecursive &&
831 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
834 if (NumVectorInstructions > NumInstructions/2)
835 VectorBonus = FiftyPercentVectorBonus;
836 else if (NumVectorInstructions > NumInstructions/10)
837 VectorBonus = TenPercentVectorBonus;
841 // Check if we've past the threshold so we don't spin in huge basic
842 // blocks that will never inline.
843 if (Cost > (Threshold + VectorBonus))
850 /// \brief Compute the base pointer and cumulative constant offsets for V.
852 /// This strips all constant offsets off of V, leaving it the base pointer, and
853 /// accumulates the total constant offset applied in the returned constant. It
854 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
855 /// no constant offsets applied.
856 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
857 if (!TD || !V->getType()->isPointerTy())
860 unsigned IntPtrWidth = TD->getPointerSizeInBits();
861 APInt Offset = APInt::getNullValue(IntPtrWidth);
863 // Even though we don't look through PHI nodes, we could be called on an
864 // instruction in an unreachable block, which may be on a cycle.
865 SmallPtrSet<Value *, 4> Visited;
868 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
869 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
871 V = GEP->getPointerOperand();
872 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
873 V = cast<Operator>(V)->getOperand(0);
874 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
875 if (GA->mayBeOverridden())
877 V = GA->getAliasee();
881 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
882 } while (Visited.insert(V));
884 Type *IntPtrTy = TD->getIntPtrType(V->getContext());
885 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
888 /// \brief Analyze a call site for potential inlining.
890 /// Returns true if inlining this call is viable, and false if it is not
891 /// viable. It computes the cost and adjusts the threshold based on numerous
892 /// factors and heuristics. If this method returns false but the computed cost
893 /// is below the computed threshold, then inlining was forcibly disabled by
894 /// some artifact of the routine.
895 bool CallAnalyzer::analyzeCall(CallSite CS) {
898 // Track whether the post-inlining function would have more than one basic
899 // block. A single basic block is often intended for inlining. Balloon the
900 // threshold by 50% until we pass the single-BB phase.
901 bool SingleBB = true;
902 int SingleBBBonus = Threshold / 2;
903 Threshold += SingleBBBonus;
905 // Perform some tweaks to the cost and threshold based on the direct
906 // callsite information.
908 // We want to more aggressively inline vector-dense kernels, so up the
909 // threshold, and we'll lower it if the % of vector instructions gets too
911 assert(NumInstructions == 0);
912 assert(NumVectorInstructions == 0);
913 FiftyPercentVectorBonus = Threshold;
914 TenPercentVectorBonus = Threshold / 2;
916 // Give out bonuses per argument, as the instructions setting them up will
917 // be gone after inlining.
918 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) {
919 if (TD && CS.isByValArgument(I)) {
920 // We approximate the number of loads and stores needed by dividing the
921 // size of the byval type by the target's pointer size.
922 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType());
923 unsigned TypeSize = TD->getTypeSizeInBits(PTy->getElementType());
924 unsigned PointerSize = TD->getPointerSizeInBits();
926 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize;
928 // If it generates more than 8 stores it is likely to be expanded as an
929 // inline memcpy so we take that as an upper bound. Otherwise we assume
930 // one load and one store per word copied.
931 // FIXME: The maxStoresPerMemcpy setting from the target should be used
932 // here instead of a magic number of 8, but it's not available via
934 NumStores = std::min(NumStores, 8U);
936 Cost -= 2 * NumStores * InlineConstants::InstrCost;
938 // For non-byval arguments subtract off one instruction per call
940 Cost -= InlineConstants::InstrCost;
944 // If there is only one call of the function, and it has internal linkage,
945 // the cost of inlining it drops dramatically.
946 bool OnlyOneCallAndLocalLinkage = F.hasLocalLinkage() && F.hasOneUse() &&
947 &F == CS.getCalledFunction();
948 if (OnlyOneCallAndLocalLinkage)
949 Cost += InlineConstants::LastCallToStaticBonus;
951 // If the instruction after the call, or if the normal destination of the
952 // invoke is an unreachable instruction, the function is noreturn. As such,
953 // there is little point in inlining this unless there is literally zero
955 Instruction *Instr = CS.getInstruction();
956 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) {
957 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
959 } else if (isa<UnreachableInst>(++BasicBlock::iterator(Instr)))
962 // If this function uses the coldcc calling convention, prefer not to inline
964 if (F.getCallingConv() == CallingConv::Cold)
965 Cost += InlineConstants::ColdccPenalty;
967 // Check if we're done. This can happen due to bonuses and penalties.
968 if (Cost > Threshold)
974 Function *Caller = CS.getInstruction()->getParent()->getParent();
975 // Check if the caller function is recursive itself.
976 for (Value::use_iterator U = Caller->use_begin(), E = Caller->use_end();
978 CallSite Site(cast<Value>(*U));
981 Instruction *I = Site.getInstruction();
982 if (I->getParent()->getParent() == Caller) {
983 IsCallerRecursive = true;
988 // Track whether we've seen a return instruction. The first return
989 // instruction is free, as at least one will usually disappear in inlining.
990 bool HasReturn = false;
992 // Populate our simplified values by mapping from function arguments to call
993 // arguments with known important simplifications.
994 CallSite::arg_iterator CAI = CS.arg_begin();
995 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
996 FAI != FAE; ++FAI, ++CAI) {
997 assert(CAI != CS.arg_end());
998 if (Constant *C = dyn_cast<Constant>(CAI))
999 SimplifiedValues[FAI] = C;
1001 Value *PtrArg = *CAI;
1002 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
1003 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue());
1005 // We can SROA any pointer arguments derived from alloca instructions.
1006 if (isa<AllocaInst>(PtrArg)) {
1007 SROAArgValues[FAI] = PtrArg;
1008 SROAArgCosts[PtrArg] = 0;
1012 NumConstantArgs = SimplifiedValues.size();
1013 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
1014 NumAllocaArgs = SROAArgValues.size();
1016 // The worklist of live basic blocks in the callee *after* inlining. We avoid
1017 // adding basic blocks of the callee which can be proven to be dead for this
1018 // particular call site in order to get more accurate cost estimates. This
1019 // requires a somewhat heavyweight iteration pattern: we need to walk the
1020 // basic blocks in a breadth-first order as we insert live successors. To
1021 // accomplish this, prioritizing for small iterations because we exit after
1022 // crossing our threshold, we use a small-size optimized SetVector.
1023 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
1024 SmallPtrSet<BasicBlock *, 16> > BBSetVector;
1025 BBSetVector BBWorklist;
1026 BBWorklist.insert(&F.getEntryBlock());
1027 // Note that we *must not* cache the size, this loop grows the worklist.
1028 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
1029 // Bail out the moment we cross the threshold. This means we'll under-count
1030 // the cost, but only when undercounting doesn't matter.
1031 if (Cost > (Threshold + VectorBonus))
1034 BasicBlock *BB = BBWorklist[Idx];
1038 // Handle the terminator cost here where we can track returns and other
1039 // function-wide constructs.
1040 TerminatorInst *TI = BB->getTerminator();
1042 // We never want to inline functions that contain an indirectbr. This is
1043 // incorrect because all the blockaddress's (in static global initializers
1044 // for example) would be referring to the original function, and this
1045 // indirect jump would jump from the inlined copy of the function into the
1046 // original function which is extremely undefined behavior.
1047 // FIXME: This logic isn't really right; we can safely inline functions
1048 // with indirectbr's as long as no other function or global references the
1049 // blockaddress of a block within the current function. And as a QOI issue,
1050 // if someone is using a blockaddress without an indirectbr, and that
1051 // reference somehow ends up in another function or global, we probably
1052 // don't want to inline this function.
1053 if (isa<IndirectBrInst>(TI))
1056 if (!HasReturn && isa<ReturnInst>(TI))
1059 Cost += InlineConstants::InstrCost;
1061 // Analyze the cost of this block. If we blow through the threshold, this
1062 // returns false, and we can bail on out.
1063 if (!analyzeBlock(BB)) {
1064 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca)
1067 // If the caller is a recursive function then we don't want to inline
1068 // functions which allocate a lot of stack space because it would increase
1069 // the caller stack usage dramatically.
1070 if (IsCallerRecursive &&
1071 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
1077 // Add in the live successors by first checking whether we have terminator
1078 // that may be simplified based on the values simplified by this call.
1079 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1080 if (BI->isConditional()) {
1081 Value *Cond = BI->getCondition();
1082 if (ConstantInt *SimpleCond
1083 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1084 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
1088 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1089 Value *Cond = SI->getCondition();
1090 if (ConstantInt *SimpleCond
1091 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1092 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor());
1097 // If we're unable to select a particular successor, just count all of
1099 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize;
1101 BBWorklist.insert(TI->getSuccessor(TIdx));
1103 // If we had any successors at this point, than post-inlining is likely to
1104 // have them as well. Note that we assume any basic blocks which existed
1105 // due to branches or switches which folded above will also fold after
1107 if (SingleBB && TI->getNumSuccessors() > 1) {
1108 // Take off the bonus we applied to the threshold.
1109 Threshold -= SingleBBBonus;
1114 // If this is a noduplicate call, we can still inline as long as
1115 // inlining this would cause the removal of the caller (so the instruction
1116 // is not actually duplicated, just moved).
1117 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall)
1120 Threshold += VectorBonus;
1122 return Cost < Threshold;
1125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1126 /// \brief Dump stats about this call's analysis.
1127 void CallAnalyzer::dump() {
1128 #define DEBUG_PRINT_STAT(x) llvm::dbgs() << " " #x ": " << x << "\n"
1129 DEBUG_PRINT_STAT(NumConstantArgs);
1130 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
1131 DEBUG_PRINT_STAT(NumAllocaArgs);
1132 DEBUG_PRINT_STAT(NumConstantPtrCmps);
1133 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
1134 DEBUG_PRINT_STAT(NumInstructionsSimplified);
1135 DEBUG_PRINT_STAT(SROACostSavings);
1136 DEBUG_PRINT_STAT(SROACostSavingsLost);
1137 DEBUG_PRINT_STAT(ContainsNoDuplicateCall);
1138 #undef DEBUG_PRINT_STAT
1142 INITIALIZE_PASS_BEGIN(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1144 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
1145 INITIALIZE_PASS_END(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1148 char InlineCostAnalysis::ID = 0;
1150 InlineCostAnalysis::InlineCostAnalysis() : CallGraphSCCPass(ID), TD(0) {}
1152 InlineCostAnalysis::~InlineCostAnalysis() {}
1154 void InlineCostAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1155 AU.setPreservesAll();
1156 AU.addRequired<TargetTransformInfo>();
1157 CallGraphSCCPass::getAnalysisUsage(AU);
1160 bool InlineCostAnalysis::runOnSCC(CallGraphSCC &SCC) {
1161 TD = getAnalysisIfAvailable<DataLayout>();
1162 TTI = &getAnalysis<TargetTransformInfo>();
1166 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, int Threshold) {
1167 return getInlineCost(CS, CS.getCalledFunction(), Threshold);
1170 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, Function *Callee,
1172 // Cannot inline indirect calls.
1174 return llvm::InlineCost::getNever();
1176 // Calls to functions with always-inline attributes should be inlined
1177 // whenever possible.
1178 if (Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1179 Attribute::AlwaysInline)) {
1180 if (isInlineViable(*Callee))
1181 return llvm::InlineCost::getAlways();
1182 return llvm::InlineCost::getNever();
1185 // Don't inline functions which can be redefined at link-time to mean
1186 // something else. Don't inline functions marked noinline or call sites
1188 if (Callee->mayBeOverridden() ||
1189 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1190 Attribute::NoInline) ||
1192 return llvm::InlineCost::getNever();
1194 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName()
1197 CallAnalyzer CA(TD, *TTI, *Callee, Threshold);
1198 bool ShouldInline = CA.analyzeCall(CS);
1202 // Check if there was a reason to force inlining or no inlining.
1203 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1204 return InlineCost::getNever();
1205 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1206 return InlineCost::getAlways();
1208 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());
1211 bool InlineCostAnalysis::isInlineViable(Function &F) {
1213 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1214 Attribute::ReturnsTwice);
1215 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1216 // Disallow inlining of functions which contain an indirect branch.
1217 if (isa<IndirectBrInst>(BI->getTerminator()))
1220 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
1226 // Disallow recursive calls.
1227 if (&F == CS.getCalledFunction())
1230 // Disallow calls which expose returns-twice to a function not previously
1231 // attributed as such.
1232 if (!ReturnsTwice && CS.isCall() &&
1233 cast<CallInst>(CS.getInstruction())->canReturnTwice())