1//===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
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// This file transforms calls of the current function (self recursion) followed
10// by a return instruction with a branch to the entry of the function, creating
11// a loop. This pass also implements the following extensions to the basic
12// algorithm:
13//
14// 1. Trivial instructions between the call and return do not prevent the
15// transformation from taking place, though currently the analysis cannot
16// support moving any really useful instructions (only dead ones).
17// 2. This pass transforms functions that are prevented from being tail
18// recursive by an associative and commutative expression to use an
19// accumulator variable, thus compiling the typical naive factorial or
20// 'fib' implementation into efficient code.
21// 3. TRE is performed if the function returns void, if the return
22// returns the result returned by the call, or if the function returns a
23// run-time constant on all exits from the function. It is possible, though
24// unlikely, that the return returns something else (like constant 0), and
25// can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in
26// the function return the exact same value.
27// 4. If it can prove that callees do not access their caller stack frame,
28// they are marked as eligible for tail call elimination (by the code
29// generator).
30//
31// There are several improvements that could be made:
32//
33// 1. If the function has any alloca instructions, these instructions will be
34// moved out of the entry block of the function, causing them to be
35// evaluated each time through the tail recursion. Safely keeping allocas
36// in the entry block requires analysis to proves that the tail-called
37// function does not read or write the stack object.
38// 2. Tail recursion is only performed if the call immediately precedes the
39// return instruction. It's possible that there could be a jump between
40// the call and the return.
41// 3. There can be intervening operations between the call and the return that
42// prevent the TRE from occurring. For example, there could be GEP's and
43// stores to memory that will not be read or written by the call. This
44// requires some substantial analysis (such as with DSA) to prove safe to
45// move ahead of the call, but doing so could allow many more TREs to be
46// performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark.
47// 4. The algorithm we use to detect if callees access their caller stack
48// frames is very primitive.
49//
50//===----------------------------------------------------------------------===//
51
52#include "llvm/Transforms/Scalar/TailRecursionElimination.h"
53#include "llvm/ADT/STLExtras.h"
54#include "llvm/ADT/SmallPtrSet.h"
55#include "llvm/ADT/Statistic.h"
56#include "llvm/Analysis/BlockFrequencyInfo.h"
57#include "llvm/Analysis/DomTreeUpdater.h"
58#include "llvm/Analysis/GlobalsModRef.h"
59#include "llvm/Analysis/InstructionSimplify.h"
60#include "llvm/Analysis/Loads.h"
61#include "llvm/Analysis/OptimizationRemarkEmitter.h"
62#include "llvm/Analysis/PostDominators.h"
63#include "llvm/Analysis/ProfileSummaryInfo.h"
64#include "llvm/Analysis/TargetTransformInfo.h"
65#include "llvm/Analysis/ValueTracking.h"
66#include "llvm/IR/CFG.h"
67#include "llvm/IR/Constants.h"
68#include "llvm/IR/DataLayout.h"
69#include "llvm/IR/DerivedTypes.h"
70#include "llvm/IR/DiagnosticInfo.h"
71#include "llvm/IR/Dominators.h"
72#include "llvm/IR/Function.h"
73#include "llvm/IR/IRBuilder.h"
74#include "llvm/IR/InstIterator.h"
75#include "llvm/IR/Instructions.h"
76#include "llvm/IR/IntrinsicInst.h"
77#include "llvm/IR/Module.h"
78#include "llvm/InitializePasses.h"
79#include "llvm/Pass.h"
80#include "llvm/Support/CommandLine.h"
81#include "llvm/Support/Debug.h"
82#include "llvm/Support/raw_ostream.h"
83#include "llvm/Transforms/Scalar.h"
84#include "llvm/Transforms/Utils/BasicBlockUtils.h"
85#include <cmath>
86using namespace llvm;
87
88#define DEBUG_TYPE "tailcallelim"
89
90STATISTIC(NumEliminated, "Number of tail calls removed");
91STATISTIC(NumRetDuped, "Number of return duplicated");
92STATISTIC(NumAccumAdded, "Number of accumulators introduced");
93STATISTIC(NumTREPreventedCold,
94 "Number of tail calls/recursion eliminations prevented due to cold "
95 "calling convention or attribute");
96
97static cl::opt<bool> DisableEntryCountRecompute(
98 "tre-disable-entrycount-recompute", cl::init(Val: false), cl::Hidden,
99 cl::desc("Force disabling recomputing of function entry count, on "
100 "successful tail recursion elimination."));
101
102static cl::opt<bool> DisableTailCallElimForColdCalls(
103 "disable-tail-call-elim-for-cold-calls", cl::Hidden, cl::init(Val: false),
104 cl::desc("Disable tail call elimination and optimization for cold calls or "
105 "in cold functions"));
106
107static bool shouldDisableTailCallsForCold(const CallBase *CB,
108 const Function *Caller,
109 const ProfileSummaryInfo *PSI,
110 BlockFrequencyInfo *BFI) {
111 if (!DisableTailCallElimForColdCalls)
112 return false;
113
114 if (CB && CB->isMustTailCall())
115 return false;
116
117 if (Caller && (Caller->hasFnAttribute(Kind: Attribute::Cold) ||
118 Caller->getCallingConv() == CallingConv::Cold))
119 return true;
120
121 if (!PSI || !PSI->hasProfileSummary())
122 return false;
123
124 // We require both the function entry and the call site/block/callee to be
125 // cold.
126 // 1. Checking that the function entry is cold ensures we don't disable tail
127 // call elimination in hot functions (with calls on cold conditional
128 // paths), which would force stack frame setup and teardown on hot paths.
129 // 2. Checking that the call site/block/callee is also cold ensures that if a
130 // function has a cold entry count but contains a hot loop, we don't
131 // disable tail call elimination for calls within that hot loop.
132 if (Caller && PSI->isFunctionEntryCold(F: Caller) && CB) {
133 if (CB->hasFnAttr(Kind: Attribute::Cold) ||
134 CB->getCallingConv() == CallingConv::Cold)
135 return true;
136 if (BFI && (PSI->isColdCallSite(CB: *CB, BFI) ||
137 PSI->isColdBlock(BB: CB->getParent(), BFI)))
138 return true;
139 }
140
141 return false;
142}
143
144/// Scan the specified function for alloca instructions.
145/// If it contains any dynamic allocas, returns false.
146static bool canTRE(Function &F) {
147 // TODO: We don't do TRE if dynamic allocas are used.
148 // Dynamic allocas allocate stack space which should be
149 // deallocated before new iteration started. That is
150 // currently not implemented.
151 return llvm::all_of(Range: instructions(F), P: [](Instruction &I) {
152 auto *AI = dyn_cast<AllocaInst>(Val: &I);
153 return !AI || AI->isStaticAlloca();
154 });
155}
156
157namespace {
158struct AllocaDerivedValueTracker {
159 // Start at a root value and walk its use-def chain to mark calls that use the
160 // value or a derived value in AllocaUsers, and places where it may escape in
161 // EscapePoints.
162 void walk(Value *Root) {
163 SmallVector<Use *, 32> Worklist;
164 SmallPtrSet<Use *, 32> Visited;
165
166 auto AddUsesToWorklist = [&](Value *V) {
167 for (auto &U : V->uses()) {
168 if (!Visited.insert(Ptr: &U).second)
169 continue;
170 Worklist.push_back(Elt: &U);
171 }
172 };
173
174 AddUsesToWorklist(Root);
175
176 while (!Worklist.empty()) {
177 Use *U = Worklist.pop_back_val();
178 Instruction *I = cast<Instruction>(Val: U->getUser());
179
180 switch (I->getOpcode()) {
181 case Instruction::Call:
182 case Instruction::Invoke: {
183 auto &CB = cast<CallBase>(Val&: *I);
184 // If the alloca-derived argument is passed byval it is not an escape
185 // point, or a use of an alloca. Calling with byval copies the contents
186 // of the alloca into argument registers or stack slots, which exist
187 // beyond the lifetime of the current frame.
188 if (CB.isArgOperand(U) && CB.isByValArgument(ArgNo: CB.getArgOperandNo(U)))
189 continue;
190 bool IsNocapture =
191 CB.isDataOperand(U) && CB.doesNotCapture(OpNo: CB.getDataOperandNo(U));
192 callUsesLocalStack(CB, IsNocapture);
193 if (IsNocapture) {
194 // If the alloca-derived argument is passed in as nocapture, then it
195 // can't propagate to the call's return. That would be capturing.
196 continue;
197 }
198 break;
199 }
200 case Instruction::Load: {
201 // The result of a load is not alloca-derived (unless an alloca has
202 // otherwise escaped, but this is a local analysis).
203 continue;
204 }
205 case Instruction::Store: {
206 if (U->getOperandNo() == 0)
207 EscapePoints.insert(Ptr: I);
208 continue; // Stores have no users to analyze.
209 }
210 case Instruction::BitCast:
211 case Instruction::GetElementPtr:
212 case Instruction::PHI:
213 case Instruction::Select:
214 case Instruction::AddrSpaceCast:
215 break;
216 default:
217 EscapePoints.insert(Ptr: I);
218 break;
219 }
220
221 AddUsesToWorklist(I);
222 }
223 }
224
225 void callUsesLocalStack(CallBase &CB, bool IsNocapture) {
226 // Add it to the list of alloca users.
227 AllocaUsers.insert(Ptr: &CB);
228
229 // If it's nocapture then it can't capture this alloca.
230 if (IsNocapture)
231 return;
232
233 // If it can write to memory, it can leak the alloca value.
234 if (!CB.onlyReadsMemory())
235 EscapePoints.insert(Ptr: &CB);
236 }
237
238 SmallPtrSet<Instruction *, 32> AllocaUsers;
239 SmallPtrSet<Instruction *, 32> EscapePoints;
240};
241} // namespace
242
243static bool markTails(Function &F, OptimizationRemarkEmitter *ORE,
244 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
245 if (F.callsFunctionThatReturnsTwice())
246 return false;
247
248 // The local stack holds all alloca instructions and all byval arguments.
249 AllocaDerivedValueTracker Tracker;
250 for (Argument &Arg : F.args()) {
251 if (Arg.hasByValAttr())
252 Tracker.walk(Root: &Arg);
253 }
254 for (auto &BB : F) {
255 for (auto &I : BB)
256 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: &I))
257 Tracker.walk(Root: AI);
258 }
259
260 bool Modified = false;
261
262 // Track whether a block is reachable after an alloca has escaped. Blocks that
263 // contain the escaping instruction will be marked as being visited without an
264 // escaped alloca, since that is how the block began.
265 enum VisitType {
266 UNVISITED,
267 UNESCAPED,
268 ESCAPED
269 };
270 DenseMap<BasicBlock *, VisitType> Visited;
271
272 // We propagate the fact that an alloca has escaped from block to successor.
273 // Visit the blocks that are propagating the escapedness first. To do this, we
274 // maintain two worklists.
275 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
276
277 // We may enter a block and visit it thinking that no alloca has escaped yet,
278 // then see an escape point and go back around a loop edge and come back to
279 // the same block twice. Because of this, we defer setting tail on calls when
280 // we first encounter them in a block. Every entry in this list does not
281 // statically use an alloca via use-def chain analysis, but may find an alloca
282 // through other means if the block turns out to be reachable after an escape
283 // point.
284 SmallVector<CallInst *, 32> DeferredTails;
285
286 BasicBlock *BB = &F.getEntryBlock();
287 VisitType Escaped = UNESCAPED;
288 do {
289 for (auto &I : *BB) {
290 if (Tracker.EscapePoints.count(Ptr: &I))
291 Escaped = ESCAPED;
292
293 CallInst *CI = dyn_cast<CallInst>(Val: &I);
294 // A PseudoProbeInst has the IntrInaccessibleMemOnly tag hence it is
295 // considered accessing memory and will be marked as a tail call if we
296 // don't bail out here.
297 if (!CI || CI->isTailCall() || isa<PseudoProbeInst>(Val: &I))
298 continue;
299
300 // Bail out for intrinsic stackrestore call because it can modify
301 // unescaped allocas.
302 if (auto *II = dyn_cast<IntrinsicInst>(Val: CI))
303 if (II->getIntrinsicID() == Intrinsic::stackrestore)
304 continue;
305
306 // Special-case operand bundles "clang.arc.attachedcall", "ptrauth", and
307 // "kcfi".
308 bool DisableForCold = shouldDisableTailCallsForCold(CB: CI, Caller: &F, PSI, BFI);
309 bool IsNoTail = CI->isNoTailCall() || DisableForCold ||
310 CI->hasOperandBundlesOtherThan(
311 IDs: {LLVMContext::OB_clang_arc_attachedcall,
312 LLVMContext::OB_ptrauth, LLVMContext::OB_kcfi});
313 if (!CI->isNoTailCall() && DisableForCold)
314 ++NumTREPreventedCold;
315
316 if (!IsNoTail && CI->doesNotAccessMemory()) {
317 // A call to a readnone function whose arguments are all things computed
318 // outside this function can be marked tail. Even if you stored the
319 // alloca address into a global, a readnone function can't load the
320 // global anyhow.
321 //
322 // Note that this runs whether we know an alloca has escaped or not. If
323 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
324 bool SafeToTail = true;
325 for (auto &Arg : CI->args()) {
326 if (isa<Constant>(Val: Arg.getUser()))
327 continue;
328 if (Argument *A = dyn_cast<Argument>(Val: Arg.getUser()))
329 if (!A->hasByValAttr())
330 continue;
331 SafeToTail = false;
332 break;
333 }
334 if (SafeToTail) {
335 using namespace ore;
336 ORE->emit(RemarkBuilder: [&]() {
337 return OptimizationRemark(DEBUG_TYPE, "tailcall-readnone", CI)
338 << "marked as tail call candidate (readnone)";
339 });
340 CI->setTailCall();
341 Modified = true;
342 continue;
343 }
344 }
345
346 if (!IsNoTail && Escaped == UNESCAPED && !Tracker.AllocaUsers.count(Ptr: CI))
347 DeferredTails.push_back(Elt: CI);
348 }
349
350 for (auto *SuccBB : successors(BB)) {
351 auto &State = Visited[SuccBB];
352 if (State < Escaped) {
353 State = Escaped;
354 if (State == ESCAPED)
355 WorklistEscaped.push_back(Elt: SuccBB);
356 else
357 WorklistUnescaped.push_back(Elt: SuccBB);
358 }
359 }
360
361 if (!WorklistEscaped.empty()) {
362 BB = WorklistEscaped.pop_back_val();
363 Escaped = ESCAPED;
364 } else {
365 BB = nullptr;
366 while (!WorklistUnescaped.empty()) {
367 auto *NextBB = WorklistUnescaped.pop_back_val();
368 if (Visited[NextBB] == UNESCAPED) {
369 BB = NextBB;
370 Escaped = UNESCAPED;
371 break;
372 }
373 }
374 }
375 } while (BB);
376
377 for (CallInst *CI : DeferredTails) {
378 if (Visited[CI->getParent()] != ESCAPED) {
379 // If the escape point was part way through the block, calls after the
380 // escape point wouldn't have been put into DeferredTails.
381 LLVM_DEBUG(dbgs() << "Marked as tail call candidate: " << *CI << "\n");
382 CI->setTailCall();
383 Modified = true;
384 }
385 }
386
387 return Modified;
388}
389
390/// Return true if it is safe to move the specified
391/// instruction from after the call to before the call, assuming that all
392/// instructions between the call and this instruction are movable.
393///
394static bool canMoveAboveCall(Instruction *I, CallInst *CI, AliasAnalysis *AA) {
395 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I))
396 if (II->getIntrinsicID() == Intrinsic::lifetime_end)
397 return true;
398
399 // FIXME: We can move load/store/call/free instructions above the call if the
400 // call does not mod/ref the memory location being processed.
401 if (I->mayHaveSideEffects()) // This also handles volatile loads.
402 return false;
403
404 if (LoadInst *L = dyn_cast<LoadInst>(Val: I)) {
405 // Loads may always be moved above calls without side effects.
406 if (CI->mayHaveSideEffects()) {
407 // Non-volatile loads may be moved above a call with side effects if it
408 // does not write to memory and the load provably won't trap.
409 // Writes to memory only matter if they may alias the pointer
410 // being loaded from.
411 const DataLayout &DL = L->getDataLayout();
412 if (isModSet(MRI: AA->getModRefInfo(I: CI, OptLoc: MemoryLocation::get(LI: L))) ||
413 !isSafeToLoadUnconditionally(V: L->getPointerOperand(), Ty: L->getType(),
414 Alignment: L->getAlign(), SQ: SimplifyQuery(DL, L)))
415 return false;
416 }
417 }
418
419 // Otherwise, if this is a side-effect free instruction, check to make sure
420 // that it does not use the return value of the call. If it doesn't use the
421 // return value of the call, it must only use things that are defined before
422 // the call, or movable instructions between the call and the instruction
423 // itself.
424 return !is_contained(Range: I->operands(), Element: CI);
425}
426
427// Return true if I is a unary accumulator recurrence: a chain of
428// applications of a unary function `g` composed with itself,
429// `g(g(...g(Base)...))`, which is equivalent to a single application of the
430// N-times-composed function when `g` is pure. Neither associative nor
431// commutative, this differs from the ordinary accumulator recurrence handled
432// below, which requires I to be associative and commutative.
433//
434// TODO: Generalize this beyond shifts by a constant amount to arbitrary pure
435// unary functions (e.g., `f(x) = x == 0 ? Base : g(f(x - 1))` for any pure
436// unary `g`).
437static bool isUnaryAccumulatorRecurrence(Instruction *I) {
438 if (!I->isShift())
439 return false;
440
441 // A chain of shifts by a constant amount C is equivalent to a single shift
442 // by the sum of the amounts:
443 // ... (Base << C) << C) ... << C == Base << (C * Iterations)
444 // This relation applies to left shifts as well as arithmetic/logical right
445 // shifts when the shift amount is a constant.
446 return isa<ConstantInt>(Val: I->getOperand(i: 1));
447}
448
449// Find the base-case return value for function F, given the accumulator
450// recursion instruction AccRecInstr that is about to be eliminated. Every
451// return other than the one fed by AccRecInstr survives the transformation and
452// will be rewritten to return the accumulator, so all of them have to yield the
453// same base-case constant. Return that constant, or nullptr on failure.
454//
455// FIXME: There is a room for improvement here in the future, e.g., consider
456// non-constant values and multiple base cases -- e.g., we want to be able to
457// handle code like:
458// ```
459// int f(int x) {
460// if (x == 1) return 1;
461// if (x == 10) return 10;
462// return f(x-1) << 1;
463// }
464// ```
465static Constant *findBaseCaseRetConstant(Function &F,
466 Instruction *AccRecInstr) {
467 Constant *BaseCaseVal = nullptr;
468
469 for (BasicBlock &BB : F) {
470 auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator());
471 if (!RI || !RI->getReturnValue())
472 continue;
473
474 Value *RV = RI->getReturnValue();
475
476 // This is the recursive case being turned into a loop: the return goes
477 // away along with AccRecInstr.
478 if (RV == AccRecInstr)
479 continue;
480
481 // Anything else has to be the base case. In particular a return still
482 // computing from a recursive call (e.g. a second recursion site that is
483 // not eliminated) must be rejected: returning the accumulator in its place
484 // would drop that computation.
485 auto *C = dyn_cast<Constant>(Val: RV);
486 if (!C)
487 return nullptr;
488
489 if (!BaseCaseVal)
490 BaseCaseVal = C;
491 else if (BaseCaseVal != C)
492 return nullptr;
493 }
494
495 return BaseCaseVal;
496}
497
498// This function checks whether the instruction I can be used
499// to perform accumulator recursion elimination for the
500// call instruction CI.
501static Constant *canTransformAccumulatorRecursion(Instruction *I,
502 CallInst *CI) {
503 bool IsUnaryAccumulatorRecurrence = isUnaryAccumulatorRecurrence(I);
504 if ((!I->isAssociative() || !I->isCommutative()) &&
505 !IsUnaryAccumulatorRecurrence)
506 return nullptr;
507
508 assert(I->getNumOperands() >= 2 &&
509 "Associative/commutative operations should have at least 2 args!");
510
511 Constant *AccInitVal = nullptr;
512 if (IsUnaryAccumulatorRecurrence) {
513 // For unary accumulator recurrences, we require that the recursive call
514 // is always on the first operand.
515 if (I->getOperand(i: 0) != CI)
516 return nullptr;
517
518 // findTRECandidate guarantees CI is a recursive call to its own
519 // function, so scan the enclosing function for the base-case return.
520 AccInitVal = findBaseCaseRetConstant(F&: *CI->getFunction(), /*AccRecInstr=*/I);
521 if (!AccInitVal)
522 return nullptr;
523 } else {
524 AccInitVal = ConstantExpr::getIdentity(I, Ty: I->getType());
525 if (!AccInitVal)
526 return nullptr;
527
528 // Exactly one operand should be the result of the call instruction.
529 if ((I->getOperand(i: 0) == CI && I->getOperand(i: 1) == CI) ||
530 (I->getOperand(i: 0) != CI && I->getOperand(i: 1) != CI))
531 return nullptr;
532 }
533
534 // The only user of this instruction we allow is a single return instruction.
535 if (!I->hasOneUse() || !isa<ReturnInst>(Val: I->user_back()))
536 return nullptr;
537
538 return AccInitVal;
539}
540
541namespace {
542class TailRecursionEliminator {
543 Function &F;
544 const TargetTransformInfo *TTI;
545 AliasAnalysis *AA;
546 OptimizationRemarkEmitter *ORE;
547 DomTreeUpdater &DTU;
548 BlockFrequencyInfo *const BFI;
549 ProfileSummaryInfo *const PSI;
550 const bool UpdateFunctionEntryCount;
551 const uint64_t OrigEntryBBFreq;
552 const uint64_t OrigEntryCount;
553
554 // The below are shared state we want to have available when eliminating any
555 // calls in the function. There values should be populated by
556 // createTailRecurseLoopHeader the first time we find a call we can eliminate.
557 BasicBlock *HeaderBB = nullptr;
558 SmallVector<PHINode *, 8> ArgumentPHIs;
559
560 // PHI node to store our return value.
561 PHINode *RetPN = nullptr;
562
563 // i1 PHI node to track if we have a valid return value stored in RetPN.
564 PHINode *RetKnownPN = nullptr;
565
566 // Vector of select instructions we insereted. These selects use RetKnownPN
567 // to either propagate RetPN or select a new return value.
568 SmallVector<SelectInst *, 8> RetSelects;
569
570 // The below are shared state needed when performing accumulator recursion.
571 // There values should be populated by insertAccumulator the first time we
572 // find an elimination that requires an accumulator.
573
574 // PHI node to store our current accumulated value.
575 PHINode *AccPN = nullptr;
576
577 // The instruction doing the accumulating.
578 Instruction *AccumulatorRecursionInstr = nullptr;
579
580 Constant *AccumulatorInitialValue = nullptr;
581
582 TailRecursionEliminator(Function &F, const TargetTransformInfo *TTI,
583 AliasAnalysis *AA, OptimizationRemarkEmitter *ORE,
584 DomTreeUpdater &DTU, BlockFrequencyInfo *BFI,
585 ProfileSummaryInfo *PSI,
586 bool UpdateFunctionEntryCount)
587 : F(F), TTI(TTI), AA(AA), ORE(ORE), DTU(DTU), BFI(BFI), PSI(PSI),
588 UpdateFunctionEntryCount(UpdateFunctionEntryCount),
589 OrigEntryBBFreq(
590 BFI ? BFI->getBlockFreq(BB: &F.getEntryBlock()).getFrequency() : 0U),
591 OrigEntryCount(F.getEntryCount() ? *F.getEntryCount() : 0) {
592 if (BFI) {
593 // The assert is meant as API documentation for the caller.
594 assert(OrigEntryBBFreq != 0 &&
595 "If a BFI was provided, the function should have an entry "
596 "basic block with a non-zero frequency.");
597 }
598 }
599
600 CallInst *findTRECandidate(BasicBlock *BB);
601
602 void createTailRecurseLoopHeader(CallInst *CI);
603
604 void insertAccumulator(Instruction *AccRecInstr);
605
606 bool eliminateCall(CallInst *CI);
607
608 void cleanupAndFinalize();
609
610 bool processBlock(BasicBlock &BB);
611
612 void copyByValueOperandIntoLocalTemp(CallInst *CI, int OpndIdx);
613
614 void copyLocalTempOfByValueOperandIntoArguments(CallInst *CI, int OpndIdx);
615
616public:
617 static bool eliminate(Function &F, const TargetTransformInfo *TTI,
618 AliasAnalysis *AA, OptimizationRemarkEmitter *ORE,
619 DomTreeUpdater &DTU, BlockFrequencyInfo *BFI,
620 ProfileSummaryInfo *PSI, bool UpdateFunctionEntryCount);
621};
622} // namespace
623
624CallInst *TailRecursionEliminator::findTRECandidate(BasicBlock *BB) {
625 Instruction *TI = BB->getTerminator();
626
627 if (&BB->front() == TI) // Make sure there is something before the terminator.
628 return nullptr;
629
630 // Scan backwards from the return, checking to see if there is a tail call in
631 // this block. If so, set CI to it.
632 CallInst *CI = nullptr;
633 BasicBlock::iterator BBI(TI);
634 while (true) {
635 CI = dyn_cast<CallInst>(Val&: BBI);
636 if (CI && CI->getCalledFunction() == &F)
637 break;
638
639 if (BBI == BB->begin())
640 return nullptr; // Didn't find a potential tail call.
641 --BBI;
642 }
643
644 assert((!CI->isTailCall() || !CI->isNoTailCall()) &&
645 "Incompatible call site attributes(Tail,NoTail)");
646 if (!CI->isTailCall() || shouldDisableTailCallsForCold(CB: CI, Caller: &F, PSI, BFI))
647 return nullptr;
648
649 // As a special case, detect code like this:
650 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
651 // and disable this xform in this case, because the code generator will
652 // lower the call to fabs into inline code.
653 if (BB == &F.getEntryBlock() && &BB->front() == CI &&
654 &*std::next(x: BB->begin()) == TI && CI->getCalledFunction() &&
655 !TTI->isLoweredToCall(F: CI->getCalledFunction())) {
656 // A single-block function with just a call and a return. Check that
657 // the arguments match.
658 auto I = CI->arg_begin(), E = CI->arg_end();
659 Function::arg_iterator FI = F.arg_begin(), FE = F.arg_end();
660 for (; I != E && FI != FE; ++I, ++FI)
661 if (*I != &*FI) break;
662 if (I == E && FI == FE)
663 return nullptr;
664 }
665
666 return CI;
667}
668
669void TailRecursionEliminator::createTailRecurseLoopHeader(CallInst *CI) {
670 HeaderBB = &F.getEntryBlock();
671 BasicBlock *NewEntry = BasicBlock::Create(Context&: F.getContext(), Name: "", Parent: &F, InsertBefore: HeaderBB);
672 NewEntry->takeName(V: HeaderBB);
673 HeaderBB->setName("tailrecurse");
674 auto *BI = UncondBrInst::Create(Target: HeaderBB, InsertBefore: NewEntry);
675 BI->setDebugLoc(DebugLoc::getCompilerGenerated());
676 // If the new branch preserves the debug location of CI, it could result in
677 // misleading stepping, if CI is located in a conditional branch.
678 // So, here we don't give any debug location to the new branch.
679
680 // Move all fixed sized allocas from HeaderBB to NewEntry.
681 for (BasicBlock::iterator OEBI = HeaderBB->begin(), E = HeaderBB->end(),
682 NEBI = NewEntry->begin();
683 OEBI != E;)
684 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: OEBI++))
685 if (isa<ConstantInt>(Val: AI->getArraySize()))
686 AI->moveBefore(InsertPos: NEBI);
687
688 // Now that we have created a new block, which jumps to the entry
689 // block, insert a PHI node for each argument of the function.
690 // For now, we initialize each PHI to only have the real arguments
691 // which are passed in.
692 BasicBlock::iterator InsertPos = HeaderBB->begin();
693 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
694 PHINode *PN = PHINode::Create(Ty: I->getType(), NumReservedValues: 2, NameStr: I->getName() + ".tr");
695 PN->insertBefore(InsertPos);
696 I->replaceAllUsesWith(V: PN); // Everyone use the PHI node now!
697 PN->addIncoming(V: &*I, BB: NewEntry);
698 ArgumentPHIs.push_back(Elt: PN);
699 }
700
701 // If the function doen't return void, create the RetPN and RetKnownPN PHI
702 // nodes to track our return value. We initialize RetPN with poison and
703 // RetKnownPN with false since we can't know our return value at function
704 // entry.
705 Type *RetType = F.getReturnType();
706 if (!RetType->isVoidTy()) {
707 Type *BoolType = Type::getInt1Ty(C&: F.getContext());
708 RetPN = PHINode::Create(Ty: RetType, NumReservedValues: 2, NameStr: "ret.tr");
709 RetPN->insertBefore(InsertPos);
710 RetKnownPN = PHINode::Create(Ty: BoolType, NumReservedValues: 2, NameStr: "ret.known.tr");
711 RetKnownPN->insertBefore(InsertPos);
712
713 RetPN->addIncoming(V: PoisonValue::get(T: RetType), BB: NewEntry);
714 RetKnownPN->addIncoming(V: ConstantInt::getFalse(Ty: BoolType), BB: NewEntry);
715 }
716
717 // The entry block was changed from HeaderBB to NewEntry.
718 // The forward DominatorTree needs to be recalculated when the EntryBB is
719 // changed. In this corner-case we recalculate the entire tree.
720 DTU.recalculate(F&: *NewEntry->getParent());
721}
722
723void TailRecursionEliminator::insertAccumulator(Instruction *AccRecInstr) {
724 assert(!AccPN && "Trying to insert multiple accumulators");
725
726 AccumulatorRecursionInstr = AccRecInstr;
727
728 // Start by inserting a new PHI node for the accumulator.
729 pred_iterator PB = pred_begin(BB: HeaderBB), PE = pred_end(BB: HeaderBB);
730 AccPN = PHINode::Create(Ty: F.getReturnType(), NumReservedValues: std::distance(first: PB, last: PE) + 1,
731 NameStr: "accumulator.tr");
732 AccPN->insertBefore(InsertPos: HeaderBB->begin());
733
734 // Loop over all of the predecessors of the tail recursion block. For the
735 // real entry into the function we seed the PHI with the identity constant for
736 // the accumulation operation. For any other existing branches to this block
737 // (due to other tail recursions eliminated) the accumulator is not modified.
738 // Because we haven't added the branch in the current block to HeaderBB yet,
739 // it will not show up as a predecessor.
740 for (pred_iterator PI = PB; PI != PE; ++PI) {
741 BasicBlock *P = *PI;
742 if (P == &F.getEntryBlock()) {
743 AccPN->addIncoming(V: AccumulatorInitialValue, BB: P);
744 } else {
745 AccPN->addIncoming(V: AccPN, BB: P);
746 }
747 }
748
749 ++NumAccumAdded;
750}
751
752// Creates a copy of contents of ByValue operand of the specified
753// call instruction into the newly created temporarily variable.
754void TailRecursionEliminator::copyByValueOperandIntoLocalTemp(CallInst *CI,
755 int OpndIdx) {
756 Type *AggTy = CI->getParamByValType(ArgNo: OpndIdx);
757 assert(AggTy);
758 const DataLayout &DL = F.getDataLayout();
759
760 // Get alignment of byVal operand.
761 Align Alignment(CI->getParamAlign(ArgNo: OpndIdx).valueOrOne());
762
763 // Create alloca for temporarily byval operands.
764 // Put alloca into the entry block.
765 Value *NewAlloca = new AllocaInst(
766 AggTy, DL.getAllocaAddrSpace(), nullptr, Alignment,
767 CI->getArgOperand(i: OpndIdx)->getName(), F.getEntryBlock().begin());
768
769 IRBuilder<> Builder(CI);
770 Value *Size = Builder.getInt64(C: DL.getTypeAllocSize(Ty: AggTy));
771
772 // Copy data from byvalue operand into the temporarily variable.
773 Builder.CreateMemCpy(Dst: NewAlloca, /*DstAlign*/ Alignment,
774 Src: CI->getArgOperand(i: OpndIdx),
775 /*SrcAlign*/ Alignment, Size);
776 CI->setArgOperand(i: OpndIdx, v: NewAlloca);
777}
778
779// Creates a copy from temporarily variable(keeping value of ByVal argument)
780// into the corresponding function argument location.
781void TailRecursionEliminator::copyLocalTempOfByValueOperandIntoArguments(
782 CallInst *CI, int OpndIdx) {
783 Type *AggTy = CI->getParamByValType(ArgNo: OpndIdx);
784 assert(AggTy);
785 const DataLayout &DL = F.getDataLayout();
786
787 // Get alignment of byVal operand.
788 Align Alignment(CI->getParamAlign(ArgNo: OpndIdx).valueOrOne());
789
790 IRBuilder<> Builder(CI);
791 Value *Size = Builder.getInt64(C: DL.getTypeAllocSize(Ty: AggTy));
792
793 // Copy data from the temporarily variable into corresponding
794 // function argument location.
795 Builder.CreateMemCpy(Dst: F.getArg(i: OpndIdx), /*DstAlign*/ Alignment,
796 Src: CI->getArgOperand(i: OpndIdx),
797 /*SrcAlign*/ Alignment, Size);
798}
799
800bool TailRecursionEliminator::eliminateCall(CallInst *CI) {
801 ReturnInst *Ret = cast<ReturnInst>(Val: CI->getParent()->getTerminator());
802
803 // Ok, we found a potential tail call. We can currently only transform the
804 // tail call if all of the instructions between the call and the return are
805 // movable to above the call itself, leaving the call next to the return.
806 // Check that this is the case now.
807 Instruction *AccRecInstr = nullptr;
808 BasicBlock::iterator BBI(CI);
809 for (++BBI; &*BBI != Ret; ++BBI) {
810 if (canMoveAboveCall(I: &*BBI, CI, AA))
811 continue;
812
813 // If we can't move the instruction above the call, it might be because it
814 // is an (associative and commutative) or unary accumulator recurrence
815 // arithmetic operation that could be transformed using accumulator
816 // recursion elimination. Check to see if this is the case, and if so,
817 // remember which instruction accumulates for later.
818 Constant *AccInitVal = canTransformAccumulatorRecursion(I: &*BBI, CI);
819
820 if (AccPN || !AccInitVal)
821 return false; // We cannot eliminate the tail recursion!
822
823 // Yes, this is accumulator recursion. Remember which instruction
824 // accumulates.
825 AccRecInstr = &*BBI;
826
827 // Keep track of the base case (i.e., initial value) of the accumulator
828 // return value if any.
829 AccumulatorInitialValue = AccInitVal;
830 }
831
832 BasicBlock *BB = Ret->getParent();
833
834 using namespace ore;
835 ORE->emit(RemarkBuilder: [&]() {
836 return OptimizationRemark(DEBUG_TYPE, "tailcall-recursion", CI)
837 << "transforming tail recursion into loop";
838 });
839
840 // OK! We can transform this tail call. If this is the first one found,
841 // create the new entry block, allowing us to branch back to the old entry.
842 if (!HeaderBB)
843 createTailRecurseLoopHeader(CI);
844
845 // Copy values of ByVal operands into local temporarily variables.
846 for (unsigned I = 0, E = CI->arg_size(); I != E; ++I) {
847 if (CI->isByValArgument(ArgNo: I))
848 copyByValueOperandIntoLocalTemp(CI, OpndIdx: I);
849 }
850
851 // Ok, now that we know we have a pseudo-entry block WITH all of the
852 // required PHI nodes, add entries into the PHI node for the actual
853 // parameters passed into the tail-recursive call.
854 for (unsigned I = 0, E = CI->arg_size(); I != E; ++I) {
855 if (CI->isByValArgument(ArgNo: I)) {
856 copyLocalTempOfByValueOperandIntoArguments(CI, OpndIdx: I);
857 // When eliminating a tail call, we modify the values of the arguments.
858 // Therefore, if the byval parameter has a readonly attribute, we have to
859 // remove it. It is safe because, from the perspective of a caller, the
860 // byval parameter is always treated as "readonly," even if the readonly
861 // attribute is removed.
862 F.removeParamAttr(ArgNo: I, Kind: Attribute::ReadOnly);
863 ArgumentPHIs[I]->addIncoming(V: F.getArg(i: I), BB);
864 } else
865 ArgumentPHIs[I]->addIncoming(V: CI->getArgOperand(i: I), BB);
866 }
867
868 if (AccRecInstr) {
869 insertAccumulator(AccRecInstr);
870
871 // Rewrite the accumulator recursion instruction so that it does not use
872 // the result of the call anymore, instead, use the PHI node we just
873 // inserted.
874 AccRecInstr->setOperand(i: AccRecInstr->getOperand(i: 0) != CI, Val: AccPN);
875
876 // Reassociating into the loop reorders the operands, so flags from the
877 // original order (nsw/nuw/exact/...) may no longer hold.
878 AccRecInstr->dropPoisonGeneratingFlags();
879 }
880
881 // Update our return value tracking
882 if (RetPN) {
883 if (Ret->getReturnValue() == CI || AccRecInstr) {
884 // Defer selecting a return value
885 RetPN->addIncoming(V: RetPN, BB);
886 RetKnownPN->addIncoming(V: RetKnownPN, BB);
887 } else {
888 // We found a return value we want to use, insert a select instruction to
889 // select it if we don't already know what our return value will be and
890 // store the result in our return value PHI node.
891 SelectInst *SI =
892 SelectInst::Create(C: RetKnownPN, S1: RetPN, S2: Ret->getReturnValue(),
893 NameStr: "current.ret.tr", InsertBefore: Ret->getIterator());
894 SI->setDebugLoc(Ret->getDebugLoc());
895 RetSelects.push_back(Elt: SI);
896
897 RetPN->addIncoming(V: SI, BB);
898 RetKnownPN->addIncoming(V: ConstantInt::getTrue(Ty: RetKnownPN->getType()), BB);
899 }
900
901 if (AccPN)
902 AccPN->addIncoming(V: AccRecInstr ? AccRecInstr : AccPN, BB);
903 }
904
905 // Now that all of the PHI nodes are in place, remove the call and
906 // ret instructions, replacing them with an unconditional branch.
907 UncondBrInst *NewBI = UncondBrInst::Create(Target: HeaderBB, InsertBefore: Ret->getIterator());
908 NewBI->setDebugLoc(CI->getDebugLoc());
909
910 Ret->eraseFromParent(); // Remove return.
911 CI->eraseFromParent(); // Remove call.
912 DTU.applyUpdates(Updates: {{DominatorTree::Insert, BB, HeaderBB}});
913 ++NumEliminated;
914 if (!DisableEntryCountRecompute && UpdateFunctionEntryCount &&
915 OrigEntryBBFreq) {
916 assert(F.getEntryCount().has_value());
917 // This pass is not expected to remove BBs, only add an entry BB. For that
918 // reason, and because the BB here isn't the new entry BB, the BFI lookup is
919 // expected to succeed.
920 assert(&F.getEntryBlock() != BB);
921 auto RelativeBBFreq =
922 static_cast<double>(BFI->getBlockFreq(BB).getFrequency()) /
923 static_cast<double>(OrigEntryBBFreq);
924 auto ToSubtract =
925 static_cast<uint64_t>(std::round(x: RelativeBBFreq * OrigEntryCount));
926 auto OldEntryCount = *F.getEntryCount();
927 if (OldEntryCount <= ToSubtract) {
928 LLVM_DEBUG(
929 errs() << "[TRE] The entrycount attributable to the recursive call, "
930 << ToSubtract
931 << ", should be strictly lower than the function entry count, "
932 << OldEntryCount << "\n");
933 } else {
934 F.setEntryCount(Count: OldEntryCount - ToSubtract);
935 }
936 }
937 return true;
938}
939
940void TailRecursionEliminator::cleanupAndFinalize() {
941 // If we eliminated any tail recursions, it's possible that we inserted some
942 // silly PHI nodes which just merge an initial value (the incoming operand)
943 // with themselves. Check to see if we did and clean up our mess if so. This
944 // occurs when a function passes an argument straight through to its tail
945 // call.
946 for (PHINode *PN : ArgumentPHIs) {
947 // If the PHI Node is a dynamic constant, replace it with the value it is.
948 if (Value *PNV = simplifyInstruction(I: PN, Q: F.getDataLayout())) {
949 PN->replaceAllUsesWith(V: PNV);
950 PN->eraseFromParent();
951 }
952 }
953
954 if (RetPN) {
955 Instruction *AccRecInstr = AccumulatorRecursionInstr;
956 auto MaterializeAccumulator = [&](Value *OtherVal,
957 BasicBlock::iterator InsertPt) {
958 Instruction *New = AccRecInstr->clone();
959 New->setName("accumulator.ret.tr");
960 New->setOperand(i: AccRecInstr->getOperand(i: 0) == AccPN, Val: OtherVal);
961 New->insertBefore(InsertPos: InsertPt);
962 New->dropLocation();
963 return New;
964 };
965
966 if (RetSelects.empty()) {
967 // If we didn't insert any select instructions, then we know we didn't
968 // store a return value and we can remove the PHI nodes we inserted.
969 RetPN->dropAllReferences();
970 RetPN->eraseFromParent();
971
972 RetKnownPN->dropAllReferences();
973 RetKnownPN->eraseFromParent();
974
975 if (AccPN) {
976 // We need to insert a copy of our accumulator instruction before any
977 // return in the function, and return its result instead.
978 for (BasicBlock &BB : F) {
979 ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator());
980 if (!RI)
981 continue;
982
983 if (isUnaryAccumulatorRecurrence(I: AccRecInstr)) {
984 // Base-case initialization: the accumulator PHI already holds the
985 // final result, so return it directly.
986 RI->setOperand(i_nocapture: 0, Val_nocapture: AccPN);
987 } else {
988 // Since the accumulator starts with the identity value, before the
989 // return we need to apply the accumulation instruction one more
990 // time to combine the last value with the result of the recursive
991 // call.
992 RI->setOperand(i_nocapture: 0, Val_nocapture: MaterializeAccumulator(RI->getOperand(i_nocapture: 0),
993 RI->getIterator()));
994 }
995 }
996 }
997 } else {
998 // We need to insert a select instruction before any return left in the
999 // function to select our stored return value if we have one.
1000 for (BasicBlock &BB : F) {
1001 ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator());
1002 if (!RI)
1003 continue;
1004
1005 SelectInst *SI =
1006 SelectInst::Create(C: RetKnownPN, S1: RetPN, S2: RI->getOperand(i_nocapture: 0),
1007 NameStr: "current.ret.tr", InsertBefore: RI->getIterator());
1008 SI->setDebugLoc(DebugLoc::getCompilerGenerated());
1009 RetSelects.push_back(Elt: SI);
1010 RI->setOperand(i_nocapture: 0, Val_nocapture: SI);
1011 }
1012
1013 if (AccPN) {
1014 // We need to insert a copy of our accumulator instruction before any
1015 // of the selects we inserted, and select its result instead.
1016 for (SelectInst *SI : RetSelects) {
1017 if (isUnaryAccumulatorRecurrence(I: AccRecInstr)) {
1018 SI->setFalseValue(AccPN);
1019 } else {
1020 SI->setFalseValue(
1021 MaterializeAccumulator(SI->getFalseValue(), SI->getIterator()));
1022 }
1023 }
1024 }
1025 }
1026 }
1027}
1028
1029bool TailRecursionEliminator::processBlock(BasicBlock &BB) {
1030 Instruction *TI = BB.getTerminator();
1031
1032 if (UncondBrInst *BI = dyn_cast<UncondBrInst>(Val: TI)) {
1033 BasicBlock *Succ = BI->getSuccessor();
1034 ReturnInst *Ret = dyn_cast<ReturnInst>(Val: Succ->getFirstNonPHIOrDbg(SkipPseudoOp: true));
1035
1036 if (!Ret)
1037 return false;
1038
1039 CallInst *CI = findTRECandidate(BB: &BB);
1040
1041 if (!CI)
1042 return false;
1043
1044 LLVM_DEBUG(dbgs() << "FOLDING: " << *Succ
1045 << "INTO UNCOND BRANCH PRED: " << BB);
1046 FoldReturnIntoUncondBranch(RI: Ret, BB: Succ, Pred: &BB, DTU: &DTU);
1047 ++NumRetDuped;
1048
1049 // If all predecessors of Succ have been eliminated by
1050 // FoldReturnIntoUncondBranch, delete it. It is important to empty it,
1051 // because the ret instruction in there is still using a value which
1052 // eliminateCall will attempt to remove. This block can only contain
1053 // instructions that can't have uses, therefore it is safe to remove.
1054 if (pred_empty(BB: Succ))
1055 DTU.deleteBB(DelBB: Succ);
1056
1057 eliminateCall(CI);
1058 return true;
1059 }
1060
1061 if (isa<ReturnInst>(Val: TI)) {
1062 CallInst *CI = findTRECandidate(BB: &BB);
1063
1064 if (CI)
1065 return eliminateCall(CI);
1066 }
1067
1068 return false;
1069}
1070
1071bool TailRecursionEliminator::eliminate(
1072 Function &F, const TargetTransformInfo *TTI, AliasAnalysis *AA,
1073 OptimizationRemarkEmitter *ORE, DomTreeUpdater &DTU,
1074 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
1075 bool UpdateFunctionEntryCount) {
1076 if (F.getFnAttribute(Kind: "disable-tail-calls").getValueAsBool())
1077 return false;
1078
1079 bool MadeChange = false;
1080 MadeChange |= markTails(F, ORE, PSI, BFI);
1081
1082 // If this function is a varargs function, we won't be able to PHI the args
1083 // right, so don't even try to convert it...
1084 if (F.getFunctionType()->isVarArg())
1085 return MadeChange;
1086
1087 if (!canTRE(F))
1088 return MadeChange;
1089
1090 // Change any tail recursive calls to loops.
1091 TailRecursionEliminator TRE(F, TTI, AA, ORE, DTU, BFI, PSI,
1092 UpdateFunctionEntryCount);
1093
1094 for (BasicBlock &BB : F)
1095 MadeChange |= TRE.processBlock(BB);
1096
1097 TRE.cleanupAndFinalize();
1098
1099 return MadeChange;
1100}
1101
1102namespace {
1103struct TailCallElim : public FunctionPass {
1104 static char ID; // Pass identification, replacement for typeid
1105 TailCallElim() : FunctionPass(ID) {
1106 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
1107 }
1108
1109 void getAnalysisUsage(AnalysisUsage &AU) const override {
1110 AU.addRequired<TargetTransformInfoWrapperPass>();
1111 AU.addRequired<AAResultsWrapperPass>();
1112 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1113 AU.addPreserved<GlobalsAAWrapperPass>();
1114 AU.addPreserved<DominatorTreeWrapperPass>();
1115 AU.addPreserved<PostDominatorTreeWrapperPass>();
1116 }
1117
1118 bool runOnFunction(Function &F) override {
1119 if (skipFunction(F))
1120 return false;
1121
1122 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1123 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
1124 auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>();
1125 auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr;
1126 // There is no noticable performance difference here between Lazy and Eager
1127 // UpdateStrategy based on some test results. It is feasible to switch the
1128 // UpdateStrategy to Lazy if we find it profitable later.
1129 DomTreeUpdater DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Eager);
1130
1131 return TailRecursionEliminator::eliminate(
1132 F, TTI: &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F),
1133 AA: &getAnalysis<AAResultsWrapperPass>().getAAResults(),
1134 ORE: &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(), DTU,
1135 /*BFI=*/nullptr, /*PSI=*/nullptr, /*UpdateFunctionEntryCount=*/false);
1136 }
1137};
1138} // namespace
1139
1140char TailCallElim::ID = 0;
1141INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim", "Tail Call Elimination",
1142 false, false)
1143INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1144INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1145INITIALIZE_PASS_END(TailCallElim, "tailcallelim", "Tail Call Elimination",
1146 false, false)
1147
1148// Public interface to the TailCallElimination pass
1149FunctionPass *llvm::createTailCallEliminationPass() {
1150 return new TailCallElim();
1151}
1152
1153PreservedAnalyses TailCallElimPass::run(Function &F,
1154 FunctionAnalysisManager &AM) {
1155
1156 TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(IR&: F);
1157 AliasAnalysis &AA = AM.getResult<AAManager>(IR&: F);
1158 // This must come first. It needs the 2 analyses, meaning, if it came after
1159 // the lines asking for the cached result, should they be nullptr (which, in
1160 // the case of the PDT, is likely), updates to the trees would be missed.
1161 auto *BFI = F.getEntryCount().has_value()
1162 ? &AM.getResult<BlockFrequencyAnalysis>(IR&: F)
1163 : nullptr;
1164 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
1165 auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
1166 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F);
1167 auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(IR&: F);
1168 auto *PDT = AM.getCachedResult<PostDominatorTreeAnalysis>(IR&: F);
1169 // There is no noticable performance difference here between Lazy and Eager
1170 // UpdateStrategy based on some test results. It is feasible to switch the
1171 // UpdateStrategy to Lazy if we find it profitable later.
1172 DomTreeUpdater DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Eager);
1173 bool Changed = TailRecursionEliminator::eliminate(
1174 F, TTI: &TTI, AA: &AA, ORE: &ORE, DTU, BFI, PSI, UpdateFunctionEntryCount);
1175
1176 if (!Changed)
1177 return PreservedAnalyses::all();
1178 PreservedAnalyses PA;
1179 PA.preserve<DominatorTreeAnalysis>();
1180 PA.preserve<PostDominatorTreeAnalysis>();
1181 return PA;
1182}
1183