1//===- FunctionSpecialization.cpp - Function Specialization ---------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#include "llvm/Transforms/IPO/FunctionSpecialization.h"
10#include "llvm/ADT/Statistic.h"
11#include "llvm/Analysis/CodeMetrics.h"
12#include "llvm/Analysis/ConstantFolding.h"
13#include "llvm/Analysis/InlineCost.h"
14#include "llvm/Analysis/InstructionSimplify.h"
15#include "llvm/Analysis/TargetTransformInfo.h"
16#include "llvm/Analysis/ValueLattice.h"
17#include "llvm/Analysis/ValueLatticeUtils.h"
18#include "llvm/Analysis/ValueTracking.h"
19#include "llvm/IR/IntrinsicInst.h"
20#include "llvm/Transforms/Scalar/SCCP.h"
21#include "llvm/Transforms/Utils/Cloning.h"
22#include "llvm/Transforms/Utils/SCCPSolver.h"
23#include "llvm/Transforms/Utils/SizeOpts.h"
24#include <cmath>
25
26using namespace llvm;
27
28#define DEBUG_TYPE "function-specialization"
29
30STATISTIC(NumSpecsCreated, "Number of specializations created");
31
32static cl::opt<bool> ForceSpecialization(
33 "force-specialization", cl::init(Val: false), cl::Hidden, cl::desc(
34 "Force function specialization for every call site with a constant "
35 "argument"));
36
37static cl::opt<unsigned> MaxClones(
38 "funcspec-max-clones", cl::init(Val: 3), cl::Hidden, cl::desc(
39 "The maximum number of clones allowed for a single function "
40 "specialization"));
41
42static cl::opt<unsigned>
43 MaxDiscoveryIterations("funcspec-max-discovery-iterations", cl::init(Val: 100),
44 cl::Hidden,
45 cl::desc("The maximum number of iterations allowed "
46 "when searching for transitive "
47 "phis"));
48
49static cl::opt<unsigned> MaxIncomingPhiValues(
50 "funcspec-max-incoming-phi-values", cl::init(Val: 8), cl::Hidden,
51 cl::desc("The maximum number of incoming values a PHI node can have to be "
52 "considered during the specialization bonus estimation"));
53
54static cl::opt<unsigned> MaxBlockPredecessors(
55 "funcspec-max-block-predecessors", cl::init(Val: 2), cl::Hidden, cl::desc(
56 "The maximum number of predecessors a basic block can have to be "
57 "considered during the estimation of dead code"));
58
59static cl::opt<unsigned> MinFunctionSize(
60 "funcspec-min-function-size", cl::init(Val: 300), cl::Hidden, cl::desc(
61 "Don't specialize functions that have less than this number of "
62 "instructions"));
63
64static cl::opt<unsigned> MaxCodeSizeGrowth(
65 "funcspec-max-codesize-growth", cl::init(Val: 3), cl::Hidden, cl::desc(
66 "Maximum codesize growth allowed per function"));
67
68static cl::opt<unsigned> MinCodeSizeSavings(
69 "funcspec-min-codesize-savings", cl::init(Val: 20), cl::Hidden, cl::desc(
70 "Reject specializations whose codesize savings are less than this"
71 "much percent of the original function size"));
72
73static cl::opt<unsigned> MinLatencySavings(
74 "funcspec-min-latency-savings", cl::init(Val: 40), cl::Hidden,
75 cl::desc("Reject specializations whose latency savings are less than this"
76 "much percent of the original function size"));
77
78static cl::opt<unsigned> MinInliningBonus(
79 "funcspec-min-inlining-bonus", cl::init(Val: 300), cl::Hidden, cl::desc(
80 "Reject specializations whose inlining bonus is less than this"
81 "much percent of the original function size"));
82
83static cl::opt<bool> SpecializeOnAddress(
84 "funcspec-on-address", cl::init(Val: false), cl::Hidden, cl::desc(
85 "Enable function specialization on the address of global values"));
86
87// Disabled by default as it can significantly increase compilation times.
88//
89// https://llvm-compile-time-tracker.com
90// https://github.com/nikic/llvm-compile-time-tracker
91static cl::opt<bool> SpecializeLiteralConstant(
92 "funcspec-for-literal-constant", cl::init(Val: false), cl::Hidden, cl::desc(
93 "Enable specialization of functions that take a literal constant as an "
94 "argument"));
95
96bool InstCostVisitor::canEliminateSuccessor(BasicBlock *BB, BasicBlock *Succ,
97 DenseSet<BasicBlock *> &DeadBlocks) {
98 unsigned I = 0;
99 return all_of(Range: predecessors(BB: Succ),
100 P: [&I, BB, Succ, &DeadBlocks] (BasicBlock *Pred) {
101 return I++ < MaxBlockPredecessors &&
102 (Pred == BB || Pred == Succ || DeadBlocks.contains(V: Pred));
103 });
104}
105
106// Estimates the codesize savings due to dead code after constant propagation.
107// \p WorkList represents the basic blocks of a specialization which will
108// eventually become dead once we replace instructions that are known to be
109// constants. The successors of such blocks are added to the list as long as
110// the \p Solver found they were executable prior to specialization, and only
111// if all their predecessors are dead.
112Cost InstCostVisitor::estimateBasicBlocks(
113 SmallVectorImpl<BasicBlock *> &WorkList) {
114 Cost CodeSize = 0;
115 // Accumulate the instruction cost of each basic block weighted by frequency.
116 while (!WorkList.empty()) {
117 BasicBlock *BB = WorkList.pop_back_val();
118
119 // These blocks are considered dead as far as the InstCostVisitor
120 // is concerned. They haven't been proven dead yet by the Solver,
121 // but may become if we propagate the specialization arguments.
122 if (!DeadBlocks.insert(V: BB).second)
123 continue;
124
125 for (Instruction &I : *BB) {
126 // Disregard SSA copies.
127 if (auto *II = dyn_cast<IntrinsicInst>(Val: &I))
128 if (II->getIntrinsicID() == Intrinsic::ssa_copy)
129 continue;
130 // If it's a known constant we have already accounted for it.
131 if (KnownConstants.contains(Val: &I))
132 continue;
133
134 Cost C = TTI.getInstructionCost(U: &I, CostKind: TargetTransformInfo::TCK_CodeSize);
135
136 LLVM_DEBUG(dbgs() << "FnSpecialization: CodeSize " << C
137 << " for user " << I << "\n");
138 CodeSize += C;
139 }
140
141 // Keep adding dead successors to the list as long as they are
142 // executable and only reachable from dead blocks.
143 for (BasicBlock *SuccBB : successors(BB))
144 if (isBlockExecutable(BB: SuccBB) &&
145 canEliminateSuccessor(BB, Succ: SuccBB, DeadBlocks))
146 WorkList.push_back(Elt: SuccBB);
147 }
148 return CodeSize;
149}
150
151static Constant *findConstantFor(Value *V, ConstMap &KnownConstants) {
152 if (auto *C = dyn_cast<Constant>(Val: V))
153 return C;
154 return KnownConstants.lookup(Val: V);
155}
156
157Bonus InstCostVisitor::getBonusFromPendingPHIs() {
158 Bonus B;
159 while (!PendingPHIs.empty()) {
160 Instruction *Phi = PendingPHIs.pop_back_val();
161 // The pending PHIs could have been proven dead by now.
162 if (isBlockExecutable(BB: Phi->getParent()))
163 B += getUserBonus(User: Phi);
164 }
165 return B;
166}
167
168/// Compute a bonus for replacing argument \p A with constant \p C.
169Bonus InstCostVisitor::getSpecializationBonus(Argument *A, Constant *C) {
170 LLVM_DEBUG(dbgs() << "FnSpecialization: Analysing bonus for constant: "
171 << C->getNameOrAsOperand() << "\n");
172 Bonus B;
173 for (auto *U : A->users())
174 if (auto *UI = dyn_cast<Instruction>(Val: U))
175 if (isBlockExecutable(BB: UI->getParent()))
176 B += getUserBonus(User: UI, Use: A, C);
177
178 LLVM_DEBUG(dbgs() << "FnSpecialization: Accumulated bonus {CodeSize = "
179 << B.CodeSize << ", Latency = " << B.Latency
180 << "} for argument " << *A << "\n");
181 return B;
182}
183
184Bonus InstCostVisitor::getUserBonus(Instruction *User, Value *Use, Constant *C) {
185 // We have already propagated a constant for this user.
186 if (KnownConstants.contains(Val: User))
187 return {0, 0};
188
189 // Cache the iterator before visiting.
190 LastVisited = Use ? KnownConstants.insert(KV: {Use, C}).first
191 : KnownConstants.end();
192
193 Cost CodeSize = 0;
194 if (auto *I = dyn_cast<SwitchInst>(Val: User)) {
195 CodeSize = estimateSwitchInst(I&: *I);
196 } else if (auto *I = dyn_cast<BranchInst>(Val: User)) {
197 CodeSize = estimateBranchInst(I&: *I);
198 } else {
199 C = visit(I&: *User);
200 if (!C)
201 return {0, 0};
202 }
203
204 // Even though it doesn't make sense to bind switch and branch instructions
205 // with a constant, unlike any other instruction type, it prevents estimating
206 // their bonus multiple times.
207 KnownConstants.insert(KV: {User, C});
208
209 CodeSize += TTI.getInstructionCost(U: User, CostKind: TargetTransformInfo::TCK_CodeSize);
210
211 uint64_t Weight = BFI.getBlockFreq(BB: User->getParent()).getFrequency() /
212 BFI.getEntryFreq().getFrequency();
213
214 Cost Latency = Weight *
215 TTI.getInstructionCost(U: User, CostKind: TargetTransformInfo::TCK_Latency);
216
217 LLVM_DEBUG(dbgs() << "FnSpecialization: {CodeSize = " << CodeSize
218 << ", Latency = " << Latency << "} for user "
219 << *User << "\n");
220
221 Bonus B(CodeSize, Latency);
222 for (auto *U : User->users())
223 if (auto *UI = dyn_cast<Instruction>(Val: U))
224 if (UI != User && isBlockExecutable(BB: UI->getParent()))
225 B += getUserBonus(User: UI, Use: User, C);
226
227 return B;
228}
229
230Cost InstCostVisitor::estimateSwitchInst(SwitchInst &I) {
231 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
232
233 if (I.getCondition() != LastVisited->first)
234 return 0;
235
236 auto *C = dyn_cast<ConstantInt>(Val: LastVisited->second);
237 if (!C)
238 return 0;
239
240 BasicBlock *Succ = I.findCaseValue(C)->getCaseSuccessor();
241 // Initialize the worklist with the dead basic blocks. These are the
242 // destination labels which are different from the one corresponding
243 // to \p C. They should be executable and have a unique predecessor.
244 SmallVector<BasicBlock *> WorkList;
245 for (const auto &Case : I.cases()) {
246 BasicBlock *BB = Case.getCaseSuccessor();
247 if (BB != Succ && isBlockExecutable(BB) &&
248 canEliminateSuccessor(BB: I.getParent(), Succ: BB, DeadBlocks))
249 WorkList.push_back(Elt: BB);
250 }
251
252 return estimateBasicBlocks(WorkList);
253}
254
255Cost InstCostVisitor::estimateBranchInst(BranchInst &I) {
256 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
257
258 if (I.getCondition() != LastVisited->first)
259 return 0;
260
261 BasicBlock *Succ = I.getSuccessor(i: LastVisited->second->isOneValue());
262 // Initialize the worklist with the dead successor as long as
263 // it is executable and has a unique predecessor.
264 SmallVector<BasicBlock *> WorkList;
265 if (isBlockExecutable(BB: Succ) &&
266 canEliminateSuccessor(BB: I.getParent(), Succ, DeadBlocks))
267 WorkList.push_back(Elt: Succ);
268
269 return estimateBasicBlocks(WorkList);
270}
271
272bool InstCostVisitor::discoverTransitivelyIncomingValues(
273 Constant *Const, PHINode *Root, DenseSet<PHINode *> &TransitivePHIs) {
274
275 SmallVector<PHINode *, 64> WorkList;
276 WorkList.push_back(Elt: Root);
277 unsigned Iter = 0;
278
279 while (!WorkList.empty()) {
280 PHINode *PN = WorkList.pop_back_val();
281
282 if (++Iter > MaxDiscoveryIterations ||
283 PN->getNumIncomingValues() > MaxIncomingPhiValues)
284 return false;
285
286 if (!TransitivePHIs.insert(V: PN).second)
287 continue;
288
289 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
290 Value *V = PN->getIncomingValue(i: I);
291
292 // Disregard self-references and dead incoming values.
293 if (auto *Inst = dyn_cast<Instruction>(Val: V))
294 if (Inst == PN || DeadBlocks.contains(V: PN->getIncomingBlock(i: I)))
295 continue;
296
297 if (Constant *C = findConstantFor(V, KnownConstants)) {
298 // Not all incoming values are the same constant. Bail immediately.
299 if (C != Const)
300 return false;
301 continue;
302 }
303
304 if (auto *Phi = dyn_cast<PHINode>(Val: V)) {
305 WorkList.push_back(Elt: Phi);
306 continue;
307 }
308
309 // We can't reason about anything else.
310 return false;
311 }
312 }
313 return true;
314}
315
316Constant *InstCostVisitor::visitPHINode(PHINode &I) {
317 if (I.getNumIncomingValues() > MaxIncomingPhiValues)
318 return nullptr;
319
320 bool Inserted = VisitedPHIs.insert(V: &I).second;
321 Constant *Const = nullptr;
322 bool HaveSeenIncomingPHI = false;
323
324 for (unsigned Idx = 0, E = I.getNumIncomingValues(); Idx != E; ++Idx) {
325 Value *V = I.getIncomingValue(i: Idx);
326
327 // Disregard self-references and dead incoming values.
328 if (auto *Inst = dyn_cast<Instruction>(Val: V))
329 if (Inst == &I || DeadBlocks.contains(V: I.getIncomingBlock(i: Idx)))
330 continue;
331
332 if (Constant *C = findConstantFor(V, KnownConstants)) {
333 if (!Const)
334 Const = C;
335 // Not all incoming values are the same constant. Bail immediately.
336 if (C != Const)
337 return nullptr;
338 continue;
339 }
340
341 if (Inserted) {
342 // First time we are seeing this phi. We will retry later, after
343 // all the constant arguments have been propagated. Bail for now.
344 PendingPHIs.push_back(Elt: &I);
345 return nullptr;
346 }
347
348 if (isa<PHINode>(Val: V)) {
349 // Perhaps it is a Transitive Phi. We will confirm later.
350 HaveSeenIncomingPHI = true;
351 continue;
352 }
353
354 // We can't reason about anything else.
355 return nullptr;
356 }
357
358 if (!Const)
359 return nullptr;
360
361 if (!HaveSeenIncomingPHI)
362 return Const;
363
364 DenseSet<PHINode *> TransitivePHIs;
365 if (!discoverTransitivelyIncomingValues(Const, Root: &I, TransitivePHIs))
366 return nullptr;
367
368 return Const;
369}
370
371Constant *InstCostVisitor::visitFreezeInst(FreezeInst &I) {
372 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
373
374 if (isGuaranteedNotToBeUndefOrPoison(V: LastVisited->second))
375 return LastVisited->second;
376 return nullptr;
377}
378
379Constant *InstCostVisitor::visitCallBase(CallBase &I) {
380 Function *F = I.getCalledFunction();
381 if (!F || !canConstantFoldCallTo(Call: &I, F))
382 return nullptr;
383
384 SmallVector<Constant *, 8> Operands;
385 Operands.reserve(N: I.getNumOperands());
386
387 for (unsigned Idx = 0, E = I.getNumOperands() - 1; Idx != E; ++Idx) {
388 Value *V = I.getOperand(i_nocapture: Idx);
389 Constant *C = findConstantFor(V, KnownConstants);
390 if (!C)
391 return nullptr;
392 Operands.push_back(Elt: C);
393 }
394
395 auto Ops = ArrayRef(Operands.begin(), Operands.end());
396 return ConstantFoldCall(Call: &I, F, Operands: Ops);
397}
398
399Constant *InstCostVisitor::visitLoadInst(LoadInst &I) {
400 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
401
402 if (isa<ConstantPointerNull>(Val: LastVisited->second))
403 return nullptr;
404 return ConstantFoldLoadFromConstPtr(C: LastVisited->second, Ty: I.getType(), DL);
405}
406
407Constant *InstCostVisitor::visitGetElementPtrInst(GetElementPtrInst &I) {
408 SmallVector<Constant *, 8> Operands;
409 Operands.reserve(N: I.getNumOperands());
410
411 for (unsigned Idx = 0, E = I.getNumOperands(); Idx != E; ++Idx) {
412 Value *V = I.getOperand(i_nocapture: Idx);
413 Constant *C = findConstantFor(V, KnownConstants);
414 if (!C)
415 return nullptr;
416 Operands.push_back(Elt: C);
417 }
418
419 auto Ops = ArrayRef(Operands.begin(), Operands.end());
420 return ConstantFoldInstOperands(I: &I, Ops, DL);
421}
422
423Constant *InstCostVisitor::visitSelectInst(SelectInst &I) {
424 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
425
426 if (I.getCondition() != LastVisited->first)
427 return nullptr;
428
429 Value *V = LastVisited->second->isZeroValue() ? I.getFalseValue()
430 : I.getTrueValue();
431 Constant *C = findConstantFor(V, KnownConstants);
432 return C;
433}
434
435Constant *InstCostVisitor::visitCastInst(CastInst &I) {
436 return ConstantFoldCastOperand(Opcode: I.getOpcode(), C: LastVisited->second,
437 DestTy: I.getType(), DL);
438}
439
440Constant *InstCostVisitor::visitCmpInst(CmpInst &I) {
441 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
442
443 bool Swap = I.getOperand(i_nocapture: 1) == LastVisited->first;
444 Value *V = Swap ? I.getOperand(i_nocapture: 0) : I.getOperand(i_nocapture: 1);
445 Constant *Other = findConstantFor(V, KnownConstants);
446 if (!Other)
447 return nullptr;
448
449 Constant *Const = LastVisited->second;
450 return Swap ?
451 ConstantFoldCompareInstOperands(Predicate: I.getPredicate(), LHS: Other, RHS: Const, DL)
452 : ConstantFoldCompareInstOperands(Predicate: I.getPredicate(), LHS: Const, RHS: Other, DL);
453}
454
455Constant *InstCostVisitor::visitUnaryOperator(UnaryOperator &I) {
456 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
457
458 return ConstantFoldUnaryOpOperand(Opcode: I.getOpcode(), Op: LastVisited->second, DL);
459}
460
461Constant *InstCostVisitor::visitBinaryOperator(BinaryOperator &I) {
462 assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
463
464 bool Swap = I.getOperand(i_nocapture: 1) == LastVisited->first;
465 Value *V = Swap ? I.getOperand(i_nocapture: 0) : I.getOperand(i_nocapture: 1);
466 Constant *Other = findConstantFor(V, KnownConstants);
467 if (!Other)
468 return nullptr;
469
470 Constant *Const = LastVisited->second;
471 return dyn_cast_or_null<Constant>(Val: Swap ?
472 simplifyBinOp(Opcode: I.getOpcode(), LHS: Other, RHS: Const, Q: SimplifyQuery(DL))
473 : simplifyBinOp(Opcode: I.getOpcode(), LHS: Const, RHS: Other, Q: SimplifyQuery(DL)));
474}
475
476Constant *FunctionSpecializer::getPromotableAlloca(AllocaInst *Alloca,
477 CallInst *Call) {
478 Value *StoreValue = nullptr;
479 for (auto *User : Alloca->users()) {
480 // We can't use llvm::isAllocaPromotable() as that would fail because of
481 // the usage in the CallInst, which is what we check here.
482 if (User == Call)
483 continue;
484 if (auto *Bitcast = dyn_cast<BitCastInst>(Val: User)) {
485 if (!Bitcast->hasOneUse() || *Bitcast->user_begin() != Call)
486 return nullptr;
487 continue;
488 }
489
490 if (auto *Store = dyn_cast<StoreInst>(Val: User)) {
491 // This is a duplicate store, bail out.
492 if (StoreValue || Store->isVolatile())
493 return nullptr;
494 StoreValue = Store->getValueOperand();
495 continue;
496 }
497 // Bail if there is any other unknown usage.
498 return nullptr;
499 }
500
501 if (!StoreValue)
502 return nullptr;
503
504 return getCandidateConstant(V: StoreValue);
505}
506
507// A constant stack value is an AllocaInst that has a single constant
508// value stored to it. Return this constant if such an alloca stack value
509// is a function argument.
510Constant *FunctionSpecializer::getConstantStackValue(CallInst *Call,
511 Value *Val) {
512 if (!Val)
513 return nullptr;
514 Val = Val->stripPointerCasts();
515 if (auto *ConstVal = dyn_cast<ConstantInt>(Val))
516 return ConstVal;
517 auto *Alloca = dyn_cast<AllocaInst>(Val);
518 if (!Alloca || !Alloca->getAllocatedType()->isIntegerTy())
519 return nullptr;
520 return getPromotableAlloca(Alloca, Call);
521}
522
523// To support specializing recursive functions, it is important to propagate
524// constant arguments because after a first iteration of specialisation, a
525// reduced example may look like this:
526//
527// define internal void @RecursiveFn(i32* arg1) {
528// %temp = alloca i32, align 4
529// store i32 2 i32* %temp, align 4
530// call void @RecursiveFn.1(i32* nonnull %temp)
531// ret void
532// }
533//
534// Before a next iteration, we need to propagate the constant like so
535// which allows further specialization in next iterations.
536//
537// @funcspec.arg = internal constant i32 2
538//
539// define internal void @someFunc(i32* arg1) {
540// call void @otherFunc(i32* nonnull @funcspec.arg)
541// ret void
542// }
543//
544// See if there are any new constant values for the callers of \p F via
545// stack variables and promote them to global variables.
546void FunctionSpecializer::promoteConstantStackValues(Function *F) {
547 for (User *U : F->users()) {
548
549 auto *Call = dyn_cast<CallInst>(Val: U);
550 if (!Call)
551 continue;
552
553 if (!Solver.isBlockExecutable(BB: Call->getParent()))
554 continue;
555
556 for (const Use &U : Call->args()) {
557 unsigned Idx = Call->getArgOperandNo(U: &U);
558 Value *ArgOp = Call->getArgOperand(i: Idx);
559 Type *ArgOpType = ArgOp->getType();
560
561 if (!Call->onlyReadsMemory(OpNo: Idx) || !ArgOpType->isPointerTy())
562 continue;
563
564 auto *ConstVal = getConstantStackValue(Call, Val: ArgOp);
565 if (!ConstVal)
566 continue;
567
568 Value *GV = new GlobalVariable(M, ConstVal->getType(), true,
569 GlobalValue::InternalLinkage, ConstVal,
570 "specialized.arg." + Twine(++NGlobals));
571 Call->setArgOperand(i: Idx, v: GV);
572 }
573 }
574}
575
576// ssa_copy intrinsics are introduced by the SCCP solver. These intrinsics
577// interfere with the promoteConstantStackValues() optimization.
578static void removeSSACopy(Function &F) {
579 for (BasicBlock &BB : F) {
580 for (Instruction &Inst : llvm::make_early_inc_range(Range&: BB)) {
581 auto *II = dyn_cast<IntrinsicInst>(Val: &Inst);
582 if (!II)
583 continue;
584 if (II->getIntrinsicID() != Intrinsic::ssa_copy)
585 continue;
586 Inst.replaceAllUsesWith(V: II->getOperand(i_nocapture: 0));
587 Inst.eraseFromParent();
588 }
589 }
590}
591
592/// Remove any ssa_copy intrinsics that may have been introduced.
593void FunctionSpecializer::cleanUpSSA() {
594 for (Function *F : Specializations)
595 removeSSACopy(F&: *F);
596}
597
598
599template <> struct llvm::DenseMapInfo<SpecSig> {
600 static inline SpecSig getEmptyKey() { return {.Key: ~0U, .Args: {}}; }
601
602 static inline SpecSig getTombstoneKey() { return {.Key: ~1U, .Args: {}}; }
603
604 static unsigned getHashValue(const SpecSig &S) {
605 return static_cast<unsigned>(hash_value(S));
606 }
607
608 static bool isEqual(const SpecSig &LHS, const SpecSig &RHS) {
609 return LHS == RHS;
610 }
611};
612
613FunctionSpecializer::~FunctionSpecializer() {
614 LLVM_DEBUG(
615 if (NumSpecsCreated > 0)
616 dbgs() << "FnSpecialization: Created " << NumSpecsCreated
617 << " specializations in module " << M.getName() << "\n");
618 // Eliminate dead code.
619 removeDeadFunctions();
620 cleanUpSSA();
621}
622
623/// Attempt to specialize functions in the module to enable constant
624/// propagation across function boundaries.
625///
626/// \returns true if at least one function is specialized.
627bool FunctionSpecializer::run() {
628 // Find possible specializations for each function.
629 SpecMap SM;
630 SmallVector<Spec, 32> AllSpecs;
631 unsigned NumCandidates = 0;
632 for (Function &F : M) {
633 if (!isCandidateFunction(F: &F))
634 continue;
635
636 auto [It, Inserted] = FunctionMetrics.try_emplace(Key: &F);
637 CodeMetrics &Metrics = It->second;
638 //Analyze the function.
639 if (Inserted) {
640 SmallPtrSet<const Value *, 32> EphValues;
641 CodeMetrics::collectEphemeralValues(L: &F, AC: &GetAC(F), EphValues);
642 for (BasicBlock &BB : F)
643 Metrics.analyzeBasicBlock(BB: &BB, TTI: GetTTI(F), EphValues);
644 }
645
646 // If the code metrics reveal that we shouldn't duplicate the function,
647 // or if the code size implies that this function is easy to get inlined,
648 // then we shouldn't specialize it.
649 if (Metrics.notDuplicatable || !Metrics.NumInsts.isValid() ||
650 (!ForceSpecialization && !F.hasFnAttribute(Kind: Attribute::NoInline) &&
651 Metrics.NumInsts < MinFunctionSize))
652 continue;
653
654 // TODO: For now only consider recursive functions when running multiple
655 // times. This should change if specialization on literal constants gets
656 // enabled.
657 if (!Inserted && !Metrics.isRecursive && !SpecializeLiteralConstant)
658 continue;
659
660 int64_t Sz = *Metrics.NumInsts.getValue();
661 assert(Sz > 0 && "CodeSize should be positive");
662 // It is safe to down cast from int64_t, NumInsts is always positive.
663 unsigned FuncSize = static_cast<unsigned>(Sz);
664
665 LLVM_DEBUG(dbgs() << "FnSpecialization: Specialization cost for "
666 << F.getName() << " is " << FuncSize << "\n");
667
668 if (Inserted && Metrics.isRecursive)
669 promoteConstantStackValues(F: &F);
670
671 if (!findSpecializations(F: &F, FuncSize, AllSpecs, SM)) {
672 LLVM_DEBUG(
673 dbgs() << "FnSpecialization: No possible specializations found for "
674 << F.getName() << "\n");
675 continue;
676 }
677
678 ++NumCandidates;
679 }
680
681 if (!NumCandidates) {
682 LLVM_DEBUG(
683 dbgs()
684 << "FnSpecialization: No possible specializations found in module\n");
685 return false;
686 }
687
688 // Choose the most profitable specialisations, which fit in the module
689 // specialization budget, which is derived from maximum number of
690 // specializations per specialization candidate function.
691 auto CompareScore = [&AllSpecs](unsigned I, unsigned J) {
692 if (AllSpecs[I].Score != AllSpecs[J].Score)
693 return AllSpecs[I].Score > AllSpecs[J].Score;
694 return I > J;
695 };
696 const unsigned NSpecs =
697 std::min(a: NumCandidates * MaxClones, b: unsigned(AllSpecs.size()));
698 SmallVector<unsigned> BestSpecs(NSpecs + 1);
699 std::iota(first: BestSpecs.begin(), last: BestSpecs.begin() + NSpecs, value: 0);
700 if (AllSpecs.size() > NSpecs) {
701 LLVM_DEBUG(dbgs() << "FnSpecialization: Number of candidates exceed "
702 << "the maximum number of clones threshold.\n"
703 << "FnSpecialization: Specializing the "
704 << NSpecs
705 << " most profitable candidates.\n");
706 std::make_heap(first: BestSpecs.begin(), last: BestSpecs.begin() + NSpecs, comp: CompareScore);
707 for (unsigned I = NSpecs, N = AllSpecs.size(); I < N; ++I) {
708 BestSpecs[NSpecs] = I;
709 std::push_heap(first: BestSpecs.begin(), last: BestSpecs.end(), comp: CompareScore);
710 std::pop_heap(first: BestSpecs.begin(), last: BestSpecs.end(), comp: CompareScore);
711 }
712 }
713
714 LLVM_DEBUG(dbgs() << "FnSpecialization: List of specializations \n";
715 for (unsigned I = 0; I < NSpecs; ++I) {
716 const Spec &S = AllSpecs[BestSpecs[I]];
717 dbgs() << "FnSpecialization: Function " << S.F->getName()
718 << " , score " << S.Score << "\n";
719 for (const ArgInfo &Arg : S.Sig.Args)
720 dbgs() << "FnSpecialization: FormalArg = "
721 << Arg.Formal->getNameOrAsOperand()
722 << ", ActualArg = " << Arg.Actual->getNameOrAsOperand()
723 << "\n";
724 });
725
726 // Create the chosen specializations.
727 SmallPtrSet<Function *, 8> OriginalFuncs;
728 SmallVector<Function *> Clones;
729 for (unsigned I = 0; I < NSpecs; ++I) {
730 Spec &S = AllSpecs[BestSpecs[I]];
731 S.Clone = createSpecialization(F: S.F, S: S.Sig);
732
733 // Update the known call sites to call the clone.
734 for (CallBase *Call : S.CallSites) {
735 LLVM_DEBUG(dbgs() << "FnSpecialization: Redirecting " << *Call
736 << " to call " << S.Clone->getName() << "\n");
737 Call->setCalledFunction(S.Clone);
738 }
739
740 Clones.push_back(Elt: S.Clone);
741 OriginalFuncs.insert(Ptr: S.F);
742 }
743
744 Solver.solveWhileResolvedUndefsIn(WorkList&: Clones);
745
746 // Update the rest of the call sites - these are the recursive calls, calls
747 // to discarded specialisations and calls that may match a specialisation
748 // after the solver runs.
749 for (Function *F : OriginalFuncs) {
750 auto [Begin, End] = SM[F];
751 updateCallSites(F, Begin: AllSpecs.begin() + Begin, End: AllSpecs.begin() + End);
752 }
753
754 for (Function *F : Clones) {
755 if (F->getReturnType()->isVoidTy())
756 continue;
757 if (F->getReturnType()->isStructTy()) {
758 auto *STy = cast<StructType>(Val: F->getReturnType());
759 if (!Solver.isStructLatticeConstant(F, STy))
760 continue;
761 } else {
762 auto It = Solver.getTrackedRetVals().find(Key: F);
763 assert(It != Solver.getTrackedRetVals().end() &&
764 "Return value ought to be tracked");
765 if (SCCPSolver::isOverdefined(LV: It->second))
766 continue;
767 }
768 for (User *U : F->users()) {
769 if (auto *CS = dyn_cast<CallBase>(Val: U)) {
770 //The user instruction does not call our function.
771 if (CS->getCalledFunction() != F)
772 continue;
773 Solver.resetLatticeValueFor(Call: CS);
774 }
775 }
776 }
777
778 // Rerun the solver to notify the users of the modified callsites.
779 Solver.solveWhileResolvedUndefs();
780
781 for (Function *F : OriginalFuncs)
782 if (FunctionMetrics[F].isRecursive)
783 promoteConstantStackValues(F);
784
785 return true;
786}
787
788void FunctionSpecializer::removeDeadFunctions() {
789 for (Function *F : FullySpecialized) {
790 LLVM_DEBUG(dbgs() << "FnSpecialization: Removing dead function "
791 << F->getName() << "\n");
792 if (FAM)
793 FAM->clear(IR&: *F, Name: F->getName());
794 F->eraseFromParent();
795 }
796 FullySpecialized.clear();
797}
798
799/// Clone the function \p F and remove the ssa_copy intrinsics added by
800/// the SCCPSolver in the cloned version.
801static Function *cloneCandidateFunction(Function *F, unsigned NSpecs) {
802 ValueToValueMapTy Mappings;
803 Function *Clone = CloneFunction(F, VMap&: Mappings);
804 Clone->setName(F->getName() + ".specialized." + Twine(NSpecs));
805 removeSSACopy(F&: *Clone);
806 return Clone;
807}
808
809bool FunctionSpecializer::findSpecializations(Function *F, unsigned FuncSize,
810 SmallVectorImpl<Spec> &AllSpecs,
811 SpecMap &SM) {
812 // A mapping from a specialisation signature to the index of the respective
813 // entry in the all specialisation array. Used to ensure uniqueness of
814 // specialisations.
815 DenseMap<SpecSig, unsigned> UniqueSpecs;
816
817 // Get a list of interesting arguments.
818 SmallVector<Argument *> Args;
819 for (Argument &Arg : F->args())
820 if (isArgumentInteresting(A: &Arg))
821 Args.push_back(Elt: &Arg);
822
823 if (Args.empty())
824 return false;
825
826 for (User *U : F->users()) {
827 if (!isa<CallInst>(Val: U) && !isa<InvokeInst>(Val: U))
828 continue;
829 auto &CS = *cast<CallBase>(Val: U);
830
831 // The user instruction does not call our function.
832 if (CS.getCalledFunction() != F)
833 continue;
834
835 // If the call site has attribute minsize set, that callsite won't be
836 // specialized.
837 if (CS.hasFnAttr(Kind: Attribute::MinSize))
838 continue;
839
840 // If the parent of the call site will never be executed, we don't need
841 // to worry about the passed value.
842 if (!Solver.isBlockExecutable(BB: CS.getParent()))
843 continue;
844
845 // Examine arguments and create a specialisation candidate from the
846 // constant operands of this call site.
847 SpecSig S;
848 for (Argument *A : Args) {
849 Constant *C = getCandidateConstant(V: CS.getArgOperand(i: A->getArgNo()));
850 if (!C)
851 continue;
852 LLVM_DEBUG(dbgs() << "FnSpecialization: Found interesting argument "
853 << A->getName() << " : " << C->getNameOrAsOperand()
854 << "\n");
855 S.Args.push_back(Elt: {A, C});
856 }
857
858 if (S.Args.empty())
859 continue;
860
861 // Check if we have encountered the same specialisation already.
862 if (auto It = UniqueSpecs.find(Val: S); It != UniqueSpecs.end()) {
863 // Existing specialisation. Add the call to the list to rewrite, unless
864 // it's a recursive call. A specialisation, generated because of a
865 // recursive call may end up as not the best specialisation for all
866 // the cloned instances of this call, which result from specialising
867 // functions. Hence we don't rewrite the call directly, but match it with
868 // the best specialisation once all specialisations are known.
869 if (CS.getFunction() == F)
870 continue;
871 const unsigned Index = It->second;
872 AllSpecs[Index].CallSites.push_back(Elt: &CS);
873 } else {
874 // Calculate the specialisation gain.
875 Bonus B;
876 unsigned Score = 0;
877 InstCostVisitor Visitor = getInstCostVisitorFor(F);
878 for (ArgInfo &A : S.Args) {
879 B += Visitor.getSpecializationBonus(A: A.Formal, C: A.Actual);
880 Score += getInliningBonus(A: A.Formal, C: A.Actual);
881 }
882 B += Visitor.getBonusFromPendingPHIs();
883
884
885 LLVM_DEBUG(dbgs() << "FnSpecialization: Specialization bonus {CodeSize = "
886 << B.CodeSize << ", Latency = " << B.Latency
887 << ", Inlining = " << Score << "}\n");
888
889 FunctionGrowth[F] += FuncSize - B.CodeSize;
890
891 auto IsProfitable = [](Bonus &B, unsigned Score, unsigned FuncSize,
892 unsigned FuncGrowth) -> bool {
893 // No check required.
894 if (ForceSpecialization)
895 return true;
896 // Minimum inlining bonus.
897 if (Score > MinInliningBonus * FuncSize / 100)
898 return true;
899 // Minimum codesize savings.
900 if (B.CodeSize < MinCodeSizeSavings * FuncSize / 100)
901 return false;
902 // Minimum latency savings.
903 if (B.Latency < MinLatencySavings * FuncSize / 100)
904 return false;
905 // Maximum codesize growth.
906 if (FuncGrowth / FuncSize > MaxCodeSizeGrowth)
907 return false;
908 return true;
909 };
910
911 // Discard unprofitable specialisations.
912 if (!IsProfitable(B, Score, FuncSize, FunctionGrowth[F]))
913 continue;
914
915 // Create a new specialisation entry.
916 Score += std::max(a: B.CodeSize, b: B.Latency);
917 auto &Spec = AllSpecs.emplace_back(Args&: F, Args&: S, Args&: Score);
918 if (CS.getFunction() != F)
919 Spec.CallSites.push_back(Elt: &CS);
920 const unsigned Index = AllSpecs.size() - 1;
921 UniqueSpecs[S] = Index;
922 if (auto [It, Inserted] = SM.try_emplace(Key: F, Args: Index, Args: Index + 1); !Inserted)
923 It->second.second = Index + 1;
924 }
925 }
926
927 return !UniqueSpecs.empty();
928}
929
930bool FunctionSpecializer::isCandidateFunction(Function *F) {
931 if (F->isDeclaration() || F->arg_empty())
932 return false;
933
934 if (F->hasFnAttribute(Kind: Attribute::NoDuplicate))
935 return false;
936
937 // Do not specialize the cloned function again.
938 if (Specializations.contains(Ptr: F))
939 return false;
940
941 // If we're optimizing the function for size, we shouldn't specialize it.
942 if (F->hasOptSize() ||
943 shouldOptimizeForSize(F, PSI: nullptr, BFI: nullptr, QueryType: PGSOQueryType::IRPass))
944 return false;
945
946 // Exit if the function is not executable. There's no point in specializing
947 // a dead function.
948 if (!Solver.isBlockExecutable(BB: &F->getEntryBlock()))
949 return false;
950
951 // It wastes time to specialize a function which would get inlined finally.
952 if (F->hasFnAttribute(Kind: Attribute::AlwaysInline))
953 return false;
954
955 LLVM_DEBUG(dbgs() << "FnSpecialization: Try function: " << F->getName()
956 << "\n");
957 return true;
958}
959
960Function *FunctionSpecializer::createSpecialization(Function *F,
961 const SpecSig &S) {
962 Function *Clone = cloneCandidateFunction(F, NSpecs: Specializations.size() + 1);
963
964 // The original function does not neccessarily have internal linkage, but the
965 // clone must.
966 Clone->setLinkage(GlobalValue::InternalLinkage);
967
968 // Initialize the lattice state of the arguments of the function clone,
969 // marking the argument on which we specialized the function constant
970 // with the given value.
971 Solver.setLatticeValueForSpecializationArguments(F: Clone, Args: S.Args);
972 Solver.markBlockExecutable(BB: &Clone->front());
973 Solver.addArgumentTrackedFunction(F: Clone);
974 Solver.addTrackedFunction(F: Clone);
975
976 // Mark all the specialized functions
977 Specializations.insert(Ptr: Clone);
978 ++NumSpecsCreated;
979
980 return Clone;
981}
982
983/// Compute the inlining bonus for replacing argument \p A with constant \p C.
984/// The below heuristic is only concerned with exposing inlining
985/// opportunities via indirect call promotion. If the argument is not a
986/// (potentially casted) function pointer, give up.
987unsigned FunctionSpecializer::getInliningBonus(Argument *A, Constant *C) {
988 Function *CalledFunction = dyn_cast<Function>(Val: C->stripPointerCasts());
989 if (!CalledFunction)
990 return 0;
991
992 // Get TTI for the called function (used for the inline cost).
993 auto &CalleeTTI = (GetTTI)(*CalledFunction);
994
995 // Look at all the call sites whose called value is the argument.
996 // Specializing the function on the argument would allow these indirect
997 // calls to be promoted to direct calls. If the indirect call promotion
998 // would likely enable the called function to be inlined, specializing is a
999 // good idea.
1000 int InliningBonus = 0;
1001 for (User *U : A->users()) {
1002 if (!isa<CallInst>(Val: U) && !isa<InvokeInst>(Val: U))
1003 continue;
1004 auto *CS = cast<CallBase>(Val: U);
1005 if (CS->getCalledOperand() != A)
1006 continue;
1007 if (CS->getFunctionType() != CalledFunction->getFunctionType())
1008 continue;
1009
1010 // Get the cost of inlining the called function at this call site. Note
1011 // that this is only an estimate. The called function may eventually
1012 // change in a way that leads to it not being inlined here, even though
1013 // inlining looks profitable now. For example, one of its called
1014 // functions may be inlined into it, making the called function too large
1015 // to be inlined into this call site.
1016 //
1017 // We apply a boost for performing indirect call promotion by increasing
1018 // the default threshold by the threshold for indirect calls.
1019 auto Params = getInlineParams();
1020 Params.DefaultThreshold += InlineConstants::IndirectCallThreshold;
1021 InlineCost IC =
1022 getInlineCost(Call&: *CS, Callee: CalledFunction, Params, CalleeTTI, GetAssumptionCache: GetAC, GetTLI);
1023
1024 // We clamp the bonus for this call to be between zero and the default
1025 // threshold.
1026 if (IC.isAlways())
1027 InliningBonus += Params.DefaultThreshold;
1028 else if (IC.isVariable() && IC.getCostDelta() > 0)
1029 InliningBonus += IC.getCostDelta();
1030
1031 LLVM_DEBUG(dbgs() << "FnSpecialization: Inlining bonus " << InliningBonus
1032 << " for user " << *U << "\n");
1033 }
1034
1035 return InliningBonus > 0 ? static_cast<unsigned>(InliningBonus) : 0;
1036}
1037
1038/// Determine if it is possible to specialise the function for constant values
1039/// of the formal parameter \p A.
1040bool FunctionSpecializer::isArgumentInteresting(Argument *A) {
1041 // No point in specialization if the argument is unused.
1042 if (A->user_empty())
1043 return false;
1044
1045 Type *Ty = A->getType();
1046 if (!Ty->isPointerTy() && (!SpecializeLiteralConstant ||
1047 (!Ty->isIntegerTy() && !Ty->isFloatingPointTy() && !Ty->isStructTy())))
1048 return false;
1049
1050 // SCCP solver does not record an argument that will be constructed on
1051 // stack.
1052 if (A->hasByValAttr() && !A->getParent()->onlyReadsMemory())
1053 return false;
1054
1055 // For non-argument-tracked functions every argument is overdefined.
1056 if (!Solver.isArgumentTrackedFunction(F: A->getParent()))
1057 return true;
1058
1059 // Check the lattice value and decide if we should attemt to specialize,
1060 // based on this argument. No point in specialization, if the lattice value
1061 // is already a constant.
1062 bool IsOverdefined = Ty->isStructTy()
1063 ? any_of(Range: Solver.getStructLatticeValueFor(V: A), P: SCCPSolver::isOverdefined)
1064 : SCCPSolver::isOverdefined(LV: Solver.getLatticeValueFor(V: A));
1065
1066 LLVM_DEBUG(
1067 if (IsOverdefined)
1068 dbgs() << "FnSpecialization: Found interesting parameter "
1069 << A->getNameOrAsOperand() << "\n";
1070 else
1071 dbgs() << "FnSpecialization: Nothing to do, parameter "
1072 << A->getNameOrAsOperand() << " is already constant\n";
1073 );
1074 return IsOverdefined;
1075}
1076
1077/// Check if the value \p V (an actual argument) is a constant or can only
1078/// have a constant value. Return that constant.
1079Constant *FunctionSpecializer::getCandidateConstant(Value *V) {
1080 if (isa<PoisonValue>(Val: V))
1081 return nullptr;
1082
1083 // Select for possible specialisation values that are constants or
1084 // are deduced to be constants or constant ranges with a single element.
1085 Constant *C = dyn_cast<Constant>(Val: V);
1086 if (!C)
1087 C = Solver.getConstantOrNull(V);
1088
1089 // Don't specialize on (anything derived from) the address of a non-constant
1090 // global variable, unless explicitly enabled.
1091 if (C && C->getType()->isPointerTy() && !C->isNullValue())
1092 if (auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: C));
1093 GV && !(GV->isConstant() || SpecializeOnAddress))
1094 return nullptr;
1095
1096 return C;
1097}
1098
1099void FunctionSpecializer::updateCallSites(Function *F, const Spec *Begin,
1100 const Spec *End) {
1101 // Collect the call sites that need updating.
1102 SmallVector<CallBase *> ToUpdate;
1103 for (User *U : F->users())
1104 if (auto *CS = dyn_cast<CallBase>(Val: U);
1105 CS && CS->getCalledFunction() == F &&
1106 Solver.isBlockExecutable(BB: CS->getParent()))
1107 ToUpdate.push_back(Elt: CS);
1108
1109 unsigned NCallsLeft = ToUpdate.size();
1110 for (CallBase *CS : ToUpdate) {
1111 bool ShouldDecrementCount = CS->getFunction() == F;
1112
1113 // Find the best matching specialisation.
1114 const Spec *BestSpec = nullptr;
1115 for (const Spec &S : make_range(x: Begin, y: End)) {
1116 if (!S.Clone || (BestSpec && S.Score <= BestSpec->Score))
1117 continue;
1118
1119 if (any_of(Range: S.Sig.Args, P: [CS, this](const ArgInfo &Arg) {
1120 unsigned ArgNo = Arg.Formal->getArgNo();
1121 return getCandidateConstant(V: CS->getArgOperand(i: ArgNo)) != Arg.Actual;
1122 }))
1123 continue;
1124
1125 BestSpec = &S;
1126 }
1127
1128 if (BestSpec) {
1129 LLVM_DEBUG(dbgs() << "FnSpecialization: Redirecting " << *CS
1130 << " to call " << BestSpec->Clone->getName() << "\n");
1131 CS->setCalledFunction(BestSpec->Clone);
1132 ShouldDecrementCount = true;
1133 }
1134
1135 if (ShouldDecrementCount)
1136 --NCallsLeft;
1137 }
1138
1139 // If the function has been completely specialized, the original function
1140 // is no longer needed. Mark it unreachable.
1141 if (NCallsLeft == 0 && Solver.isArgumentTrackedFunction(F)) {
1142 Solver.markFunctionUnreachable(F);
1143 FullySpecialized.insert(Ptr: F);
1144 }
1145}
1146