1//===- LoopPeel.cpp -------------------------------------------------------===//
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// Loop Peeling Utilities.
10//===----------------------------------------------------------------------===//
11
12#include "llvm/Transforms/Utils/LoopPeel.h"
13#include "llvm/ADT/DenseMap.h"
14#include "llvm/ADT/MapVector.h"
15#include "llvm/ADT/SmallVector.h"
16#include "llvm/ADT/Statistic.h"
17#include "llvm/Analysis/Loads.h"
18#include "llvm/Analysis/LoopInfo.h"
19#include "llvm/Analysis/LoopIterator.h"
20#include "llvm/Analysis/ScalarEvolution.h"
21#include "llvm/Analysis/ScalarEvolutionExpressions.h"
22#include "llvm/Analysis/ScalarEvolutionPatternMatch.h"
23#include "llvm/Analysis/TargetTransformInfo.h"
24#include "llvm/IR/BasicBlock.h"
25#include "llvm/IR/Dominators.h"
26#include "llvm/IR/Function.h"
27#include "llvm/IR/InstrTypes.h"
28#include "llvm/IR/Instruction.h"
29#include "llvm/IR/Instructions.h"
30#include "llvm/IR/LLVMContext.h"
31#include "llvm/IR/MDBuilder.h"
32#include "llvm/IR/PatternMatch.h"
33#include "llvm/IR/ProfDataUtils.h"
34#include "llvm/Support/Casting.h"
35#include "llvm/Support/CheckedArithmetic.h"
36#include "llvm/Support/CommandLine.h"
37#include "llvm/Support/Debug.h"
38#include "llvm/Support/raw_ostream.h"
39#include "llvm/Transforms/Utils/BasicBlockUtils.h"
40#include "llvm/Transforms/Utils/Cloning.h"
41#include "llvm/Transforms/Utils/LoopSimplify.h"
42#include "llvm/Transforms/Utils/LoopUtils.h"
43#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
44#include "llvm/Transforms/Utils/ValueMapper.h"
45#include <algorithm>
46#include <cassert>
47#include <cstdint>
48#include <optional>
49
50using namespace llvm;
51using namespace llvm::PatternMatch;
52using namespace llvm::SCEVPatternMatch;
53
54#define DEBUG_TYPE "loop-peel"
55
56STATISTIC(NumPeeled, "Number of loops peeled");
57STATISTIC(NumPeeledEnd, "Number of loops peeled from end");
58
59namespace llvm {
60static cl::opt<unsigned> UnrollPeelCount(
61 "unroll-peel-count", cl::Hidden,
62 cl::desc("Set the unroll peeling count, for testing purposes"));
63
64static cl::opt<bool>
65 UnrollAllowPeeling("unroll-allow-peeling", cl::init(Val: true), cl::Hidden,
66 cl::desc("Allows loops to be peeled when the dynamic "
67 "trip count is known to be low."));
68
69static cl::opt<bool>
70 UnrollAllowLoopNestsPeeling("unroll-allow-loop-nests-peeling",
71 cl::init(Val: false), cl::Hidden,
72 cl::desc("Allows loop nests to be peeled."));
73
74static cl::opt<unsigned> UnrollPeelMaxCount(
75 "unroll-peel-max-count", cl::init(Val: 7), cl::Hidden,
76 cl::desc("Max average trip count which will cause loop peeling."));
77
78static cl::opt<unsigned> UnrollForcePeelCount(
79 "unroll-force-peel-count", cl::init(Val: 0), cl::Hidden,
80 cl::desc("Force a peel count regardless of profiling information."));
81
82static cl::opt<bool> DisableAdvancedPeeling(
83 "disable-advanced-peeling", cl::init(Val: false), cl::Hidden,
84 cl::desc(
85 "Disable advance peeling. Issues for convergent targets (D134803)."));
86
87static cl::opt<bool> EnablePeelingForIV(
88 "enable-peeling-for-iv", cl::init(Val: false), cl::Hidden,
89 cl::desc("Enable peeling to convert Phi nodes into IVs"));
90
91static const char *PeeledCountMetaData = "llvm.loop.peeled.count";
92
93} // namespace llvm
94
95// Check whether we are capable of peeling this loop.
96bool llvm::canPeel(const Loop *L) {
97 // Make sure the loop is in simplified form
98 if (!L->isLoopSimplifyForm())
99 return false;
100 if (!DisableAdvancedPeeling)
101 return true;
102
103 SmallVector<BasicBlock *, 4> Exits;
104 L->getUniqueNonLatchExitBlocks(ExitBlocks&: Exits);
105 // The latch must either be the only exiting block or all non-latch exit
106 // blocks have either a deopt or unreachable terminator or compose a chain of
107 // blocks where the last one is either deopt or unreachable terminated. Both
108 // deopt and unreachable terminators are a strong indication they are not
109 // taken. Note that this is a profitability check, not a legality check. Also
110 // note that LoopPeeling currently can only update the branch weights of latch
111 // blocks and branch weights to blocks with deopt or unreachable do not need
112 // updating.
113 return llvm::all_of(Range&: Exits, P: IsBlockFollowedByDeoptOrUnreachable);
114}
115
116namespace {
117
118// As a loop is peeled, it may be the case that Phi nodes become
119// loop-invariant (ie, known because there is only one choice).
120// For example, consider the following function:
121// void g(int);
122// void binary() {
123// int x = 0;
124// int y = 0;
125// int a = 0;
126// for(int i = 0; i <100000; ++i) {
127// g(x);
128// x = y;
129// g(a);
130// y = a + 1;
131// a = 5;
132// }
133// }
134// Peeling 3 iterations is beneficial because the values for x, y and a
135// become known. The IR for this loop looks something like the following:
136//
137// %i = phi i32 [ 0, %entry ], [ %inc, %if.end ]
138// %a = phi i32 [ 0, %entry ], [ 5, %if.end ]
139// %y = phi i32 [ 0, %entry ], [ %add, %if.end ]
140// %x = phi i32 [ 0, %entry ], [ %y, %if.end ]
141// ...
142// tail call void @_Z1gi(i32 signext %x)
143// tail call void @_Z1gi(i32 signext %a)
144// %add = add nuw nsw i32 %a, 1
145// %inc = add nuw nsw i32 %i, 1
146// %exitcond = icmp eq i32 %inc, 100000
147// br i1 %exitcond, label %for.cond.cleanup, label %for.body
148//
149// The arguments for the calls to g will become known after 3 iterations
150// of the loop, because the phi nodes values become known after 3 iterations
151// of the loop (ie, they are known on the 4th iteration, so peel 3 iterations).
152// The first iteration has g(0), g(0); the second has g(0), g(5); the
153// third has g(1), g(5) and the fourth (and all subsequent) have g(6), g(5).
154// Now consider the phi nodes:
155// %a is a phi with constants so it is determined after iteration 1.
156// %y is a phi based on a constant and %a so it is determined on
157// the iteration after %a is determined, so iteration 2.
158// %x is a phi based on a constant and %y so it is determined on
159// the iteration after %y, so iteration 3.
160// %i is based on itself (and is an induction variable) so it is
161// never determined.
162// This means that peeling off 3 iterations will result in being able to
163// remove the phi nodes for %a, %y, and %x. The arguments for the
164// corresponding calls to g are determined and the code for computing
165// x, y, and a can be removed.
166//
167// Similarly, there are cases where peeling makes Phi nodes loop-inductions
168// (i.e., the value is increased or decreased by a fixed amount on every
169// iteration). For example, consider the following function.
170//
171// #define N 100
172// void f(int a[], int b[]) {
173// int im = N - 1;
174// for (int i = 0; i < N; i++) {
175// a[i] = b[i] + b[im];
176// im = i;
177// }
178// }
179//
180// The IR of the loop will look something like the following.
181//
182// %i = phi i32 [ 0, %entry ], [ %i.next, %for.body ]
183// %im = phi i32 [ 99, %entry ], [ %i, %for.body ]
184// ...
185// %i.next = add nuw nsw i32 %i, 1
186// ...
187//
188// In this case, %im becomes a loop-induction variable by peeling 1 iteration,
189// because %i is a loop-induction one. The peeling count can be determined by
190// the same algorithm with loop-invariant case. Such peeling is profitable for
191// loop-vectorization.
192//
193// The PhiAnalyzer class calculates how many times a loop should be
194// peeled based on the above analysis of the phi nodes in the loop while
195// respecting the maximum specified.
196class PhiAnalyzer {
197public:
198 PhiAnalyzer(const Loop &L, unsigned MaxIterations, bool PeelForIV);
199
200 // Calculate the sufficient minimum number of iterations of the loop to peel
201 // such that phi instructions become determined (subject to allowable limits)
202 std::optional<unsigned> calculateIterationsToPeel();
203
204protected:
205 enum class PeelCounterType {
206 Invariant,
207 Induction,
208 };
209
210 using PeelCounterValue = std::pair<unsigned, PeelCounterType>;
211 using PeelCounter = std::optional<PeelCounterValue>;
212 const PeelCounter Unknown = std::nullopt;
213
214 // Add 1 respecting Unknown and return Unknown if result over MaxIterations
215 PeelCounter addOne(PeelCounter PC) const {
216 if (PC == Unknown)
217 return Unknown;
218 auto [Val, Ty] = *PC;
219 return (Val + 1 <= MaxIterations) ? PeelCounter({Val + 1, Ty}) : Unknown;
220 }
221
222 // Return a value representing zero for the given counter type.
223 PeelCounter makeZero(PeelCounterType Ty) const {
224 return PeelCounter({0, Ty});
225 }
226
227 // Calculate the number of iterations after which the given value becomes an
228 // invariant or an induction.
229 PeelCounter calculate(const Value &);
230
231 // Auxiliary function to calculate the number of iterations for a comparison
232 // instruction or a binary operator.
233 PeelCounter mergeTwoCounters(const Instruction &CmpOrBinaryOp,
234 const PeelCounterValue &LHS,
235 const PeelCounterValue &RHS) const;
236
237 // Returns true if the \p Phi is an induction in the target loop. This is a
238 // lightweight check and possible to detect an IV in some cases.
239 bool isInductionPHI(const PHINode *Phi) const;
240
241 const Loop &L;
242 const unsigned MaxIterations;
243 const bool PeelForIV;
244
245 // Map of Values to number of iterations to invariance or induction
246 SmallDenseMap<const Value *, PeelCounter> IterationsToInvarianceOrInduction;
247};
248
249PhiAnalyzer::PhiAnalyzer(const Loop &L, unsigned MaxIterations, bool PeelForIV)
250 : L(L), MaxIterations(MaxIterations), PeelForIV(PeelForIV) {
251 assert(canPeel(&L) && "loop is not suitable for peeling");
252 assert(MaxIterations > 0 && "no peeling is allowed?");
253}
254
255/// Test whether \p Phi is an induction variable. Although this can be
256/// determined using SCEV analysis, it is expensive to compute here. Instead,
257/// we perform cheaper checks that may not detect complex cases but are
258/// sufficient for some situations.
259bool PhiAnalyzer::isInductionPHI(const PHINode *Phi) const {
260 // Currently we only support a loop that has single latch.
261 BasicBlock *Latch = L.getLoopLatch();
262 if (Latch == nullptr)
263 return false;
264
265 Value *Cur = Phi->getIncomingValueForBlock(BB: Latch);
266 SmallPtrSet<Value *, 4> Visited;
267 bool VisitBinOp = false;
268
269 // Starting from the incoming value of the Phi, we follow the use-def chain.
270 // We consider Phi to be an IV if we can reach it again by traversing only
271 // add, sub, or cast instructions.
272 while (true) {
273 if (Cur == Phi)
274 break;
275
276 // Avoid infinite loop.
277 if (!Visited.insert(Ptr: Cur).second)
278 return false;
279
280 auto *I = dyn_cast<Instruction>(Val: Cur);
281 if (!I || !L.contains(Inst: I))
282 return false;
283
284 if (auto *Cast = dyn_cast<CastInst>(Val: I)) {
285 Cur = Cast->getOperand(i_nocapture: 0);
286 } else if (auto *BinOp = dyn_cast<BinaryOperator>(Val: I)) {
287 if (BinOp->getOpcode() != Instruction::Add &&
288 BinOp->getOpcode() != Instruction::Sub)
289 return false;
290 if (!isa<ConstantInt>(Val: BinOp->getOperand(i_nocapture: 1)))
291 return false;
292
293 VisitBinOp = true;
294 Cur = BinOp->getOperand(i_nocapture: 0);
295 } else {
296 return false;
297 }
298 }
299
300 // Ignore cases where no binary operations are visited.
301 return VisitBinOp;
302}
303
304/// When either \p LHS or \p RHS is an IV, the result of \p CmpOrBinaryOp is
305/// considered an IV only if it is an addition or a subtraction. Otherwise the
306/// result can be a value that is neither a loop-invariant nor an IV.
307///
308/// If both \p LHS and \p RHS are loop-invariants, then the result of
309/// \CmpOrBinaryOp is also a loop-invariant.
310PhiAnalyzer::PeelCounter
311PhiAnalyzer::mergeTwoCounters(const Instruction &CmpOrBinaryOp,
312 const PeelCounterValue &LHS,
313 const PeelCounterValue &RHS) const {
314 auto &[LVal, LTy] = LHS;
315 auto &[RVal, RTy] = RHS;
316 unsigned NewVal = std::max(a: LVal, b: RVal);
317
318 if (LTy == PeelCounterType::Induction || RTy == PeelCounterType::Induction) {
319 if (const auto *BinOp = dyn_cast<BinaryOperator>(Val: &CmpOrBinaryOp)) {
320 if (BinOp->getOpcode() == Instruction::Add ||
321 BinOp->getOpcode() == Instruction::Sub)
322 return PeelCounter({NewVal, PeelCounterType::Induction});
323 }
324 return Unknown;
325 }
326 return PeelCounter({NewVal, PeelCounterType::Invariant});
327}
328
329// This function calculates the number of iterations after which the value
330// becomes an invariant. The pre-calculated values are memorized in a map.
331// N.B. This number will be Unknown or <= MaxIterations.
332// The function is calculated according to the following definition:
333// Given %x = phi <Inputs from above the loop>, ..., [%y, %back.edge].
334// F(%x) = G(%y) + 1 (N.B. [MaxIterations | Unknown] + 1 => Unknown)
335// G(%y) = 0 if %y is a loop invariant
336// G(%y) = G(%BackEdgeValue) if %y is a phi in the header block
337// G(%y) = TODO: if %y is an expression based on phis and loop invariants
338// The example looks like:
339// %x = phi(0, %a) <-- becomes invariant starting from 3rd iteration.
340// %y = phi(0, 5)
341// %a = %y + 1
342// G(%y) = Unknown otherwise (including phi not in header block)
343PhiAnalyzer::PeelCounter PhiAnalyzer::calculate(const Value &V) {
344 // If we already know the answer, take it from the map.
345 // Otherwise, place Unknown to map to avoid infinite recursion. Such
346 // cycles can never stop on an invariant.
347 auto [I, Inserted] =
348 IterationsToInvarianceOrInduction.try_emplace(Key: &V, Args: Unknown);
349 if (!Inserted)
350 return I->second;
351
352 if (L.isLoopInvariant(V: &V))
353 // Loop invariant so known at start.
354 return (IterationsToInvarianceOrInduction[&V] =
355 makeZero(Ty: PeelCounterType::Invariant));
356 if (const PHINode *Phi = dyn_cast<PHINode>(Val: &V)) {
357 if (Phi->getParent() != L.getHeader()) {
358 // Phi is not in header block so Unknown.
359 assert(IterationsToInvarianceOrInduction[&V] == Unknown &&
360 "unexpected value saved");
361 return Unknown;
362 }
363
364 // If Phi is an induction, register it as a starting point.
365 if (PeelForIV && isInductionPHI(Phi))
366 return (IterationsToInvarianceOrInduction[&V] =
367 makeZero(Ty: PeelCounterType::Induction));
368
369 // We need to analyze the input from the back edge and add 1.
370 Value *Input = Phi->getIncomingValueForBlock(BB: L.getLoopLatch());
371 PeelCounter Iterations = calculate(V: *Input);
372 assert(IterationsToInvarianceOrInduction[Input] == Iterations &&
373 "unexpected value saved");
374 return (IterationsToInvarianceOrInduction[Phi] = addOne(PC: Iterations));
375 }
376 if (const Instruction *I = dyn_cast<Instruction>(Val: &V)) {
377 if (isa<CmpInst>(Val: I) || I->isBinaryOp()) {
378 // Binary instructions get the max of the operands.
379 PeelCounter LHS = calculate(V: *I->getOperand(i: 0));
380 if (LHS == Unknown)
381 return Unknown;
382 PeelCounter RHS = calculate(V: *I->getOperand(i: 1));
383 if (RHS == Unknown)
384 return Unknown;
385 return (IterationsToInvarianceOrInduction[I] =
386 mergeTwoCounters(CmpOrBinaryOp: *I, LHS: *LHS, RHS: *RHS));
387 }
388 if (I->isCast())
389 // Cast instructions get the value of the operand.
390 return (IterationsToInvarianceOrInduction[I] =
391 calculate(V: *I->getOperand(i: 0)));
392 }
393 // TODO: handle more expressions
394
395 // Everything else is Unknown.
396 assert(IterationsToInvarianceOrInduction[&V] == Unknown &&
397 "unexpected value saved");
398 return Unknown;
399}
400
401std::optional<unsigned> PhiAnalyzer::calculateIterationsToPeel() {
402 unsigned Iterations = 0;
403 for (auto &PHI : L.getHeader()->phis()) {
404 PeelCounter ToInvarianceOrInduction = calculate(V: PHI);
405 if (ToInvarianceOrInduction != Unknown) {
406 unsigned Val = ToInvarianceOrInduction->first;
407 assert(Val <= MaxIterations && "bad result in phi analysis");
408 Iterations = std::max(a: Iterations, b: Val);
409 if (Iterations == MaxIterations)
410 break;
411 }
412 }
413 assert((Iterations <= MaxIterations) && "bad result in phi analysis");
414 return Iterations ? std::optional<unsigned>(Iterations) : std::nullopt;
415}
416
417} // unnamed namespace
418
419// Try to find any invariant memory reads that will become dereferenceable in
420// the remainder loop after peeling. The load must also be used (transitively)
421// by an exit condition. Returns the number of iterations to peel off (at the
422// moment either 0 or 1).
423static unsigned peelToTurnInvariantLoadsDereferenceable(Loop &L,
424 DominatorTree &DT,
425 AssumptionCache *AC) {
426 // Skip loops with a single exiting block, because there should be no benefit
427 // for the heuristic below.
428 if (L.getExitingBlock())
429 return 0;
430
431 // All non-latch exit blocks must have an UnreachableInst terminator.
432 // Otherwise the heuristic below may not be profitable.
433 SmallVector<BasicBlock *, 4> Exits;
434 L.getUniqueNonLatchExitBlocks(ExitBlocks&: Exits);
435 if (any_of(Range&: Exits, P: [](const BasicBlock *BB) {
436 return !isa<UnreachableInst>(Val: BB->getTerminator());
437 }))
438 return 0;
439
440 // Now look for invariant loads that dominate the latch and are not known to
441 // be dereferenceable. If there are such loads and no writes, they will become
442 // dereferenceable in the loop if the first iteration is peeled off. Also
443 // collect the set of instructions controlled by such loads. Only peel if an
444 // exit condition uses (transitively) such a load.
445 BasicBlock *Header = L.getHeader();
446 BasicBlock *Latch = L.getLoopLatch();
447 SmallPtrSet<Value *, 8> LoadUsers;
448 const DataLayout &DL = L.getHeader()->getDataLayout();
449 for (BasicBlock *BB : L.blocks()) {
450 for (Instruction &I : *BB) {
451 // Calls that only access inaccessible memory can never alias with loads.
452 if (I.mayWriteToMemory() &&
453 !(isa<CallBase>(Val: I) &&
454 cast<CallBase>(Val&: I).onlyAccessesInaccessibleMemory()))
455 return 0;
456
457 if (LoadUsers.contains(Ptr: &I))
458 LoadUsers.insert_range(R: I.users());
459 // Do not look for reads in the header; they can already be hoisted
460 // without peeling.
461 if (BB == Header)
462 continue;
463 if (auto *LI = dyn_cast<LoadInst>(Val: &I)) {
464 Value *Ptr = LI->getPointerOperand();
465 if (DT.dominates(A: BB, B: Latch) && L.isLoopInvariant(V: Ptr) &&
466 !isDereferenceablePointer(V: Ptr, Ty: LI->getType(),
467 Q: SimplifyQuery(DL, &DT, AC, LI)))
468 LoadUsers.insert_range(R: I.users());
469 }
470 }
471 }
472 SmallVector<BasicBlock *> ExitingBlocks;
473 L.getExitingBlocks(ExitingBlocks);
474 if (any_of(Range&: ExitingBlocks, P: [&LoadUsers](BasicBlock *Exiting) {
475 return LoadUsers.contains(Ptr: Exiting->getTerminator());
476 }))
477 return 1;
478 return 0;
479}
480
481bool llvm::canPeelLastIteration(const Loop &L, ScalarEvolution &SE) {
482 const SCEV *BTC = SE.getBackedgeTakenCount(L: &L);
483 if (isa<SCEVCouldNotCompute>(Val: BTC))
484 return false;
485
486 // Check if the exit condition of the loop can be adjusted by the peeling
487 // codegen. For now, it must
488 // * exit via the latch,
489 // * the exit condition must be a NE/EQ compare of an induction with step
490 // of 1 and must only be used by the exiting branch.
491 BasicBlock *Latch = L.getLoopLatch();
492 Value *Inc;
493 Value *Bound;
494 CmpPredicate Pred;
495 BasicBlock *Succ1;
496 BasicBlock *Succ2;
497 return Latch && Latch == L.getExitingBlock() &&
498 match(V: Latch->getTerminator(),
499 P: m_Br(C: m_OneUse(SubPattern: m_ICmp(Pred, L: m_Value(V&: Inc), R: m_Value(V&: Bound))),
500 T: m_BasicBlock(V&: Succ1), F: m_BasicBlock(V&: Succ2))) &&
501 ((Pred == CmpInst::ICMP_EQ && Succ2 == L.getHeader()) ||
502 (Pred == CmpInst::ICMP_NE && Succ1 == L.getHeader())) &&
503 Bound->getType()->isIntegerTy() &&
504 SE.isLoopInvariant(S: SE.getSCEV(V: Bound), L: &L) &&
505 match(S: SE.getSCEV(V: Inc),
506 P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_scev_One(), L: m_SpecificLoop(L: &L)));
507}
508
509/// Returns true if the last iteration can be peeled off and the condition (Pred
510/// LeftAR, RightSCEV) is known at the last iteration and the inverse condition
511/// is known at the second-to-last.
512static bool shouldPeelLastIteration(Loop &L, CmpPredicate Pred,
513 const SCEVAddRecExpr *LeftAR,
514 const SCEV *RightSCEV, ScalarEvolution &SE,
515 const TargetTransformInfo &TTI) {
516 if (!canPeelLastIteration(L, SE))
517 return false;
518
519 const SCEV *BTC = SE.getBackedgeTakenCount(L: &L);
520 SCEVExpander Expander(SE, "loop-peel");
521 if (!SE.isKnownNonZero(S: BTC) &&
522 Expander.isHighCostExpansion(Exprs: BTC, L: &L, Budget: SCEVCheapExpansionBudget, TTI: &TTI,
523 At: L.getLoopPredecessor()->getTerminator()))
524 return false;
525
526 auto Guards = ScalarEvolution::LoopGuards::collect(L: &L, SE);
527 BTC = SE.applyLoopGuards(Expr: BTC, Guards);
528 RightSCEV = SE.applyLoopGuards(Expr: RightSCEV, Guards);
529 const SCEV *ValAtLastIter = LeftAR->evaluateAtIteration(It: BTC, SE);
530 const SCEV *ValAtSecondToLastIter = LeftAR->evaluateAtIteration(
531 It: SE.getMinusSCEV(LHS: BTC, RHS: SE.getOne(Ty: BTC->getType())), SE);
532
533 return SE.isKnownPredicate(Pred: ICmpInst::getInversePredicate(pred: Pred), LHS: ValAtLastIter,
534 RHS: RightSCEV) &&
535 SE.isKnownPredicate(Pred, LHS: ValAtSecondToLastIter, RHS: RightSCEV);
536}
537
538// Return the number of iterations to peel off from the beginning and end of the
539// loop respectively, that make conditions in the body true/false. For example,
540// if we peel 2 iterations off the loop below, the condition i < 2 can be
541// evaluated at compile time.
542//
543// for (i = 0; i < n; i++)
544// if (i < 2)
545// ..
546// else
547// ..
548// }
549static std::pair<unsigned, unsigned>
550countToEliminateCompares(Loop &L, unsigned MaxPeelCount, ScalarEvolution &SE,
551 const TargetTransformInfo &TTI) {
552 assert(L.isLoopSimplifyForm() && "Loop needs to be in loop simplify form");
553 unsigned DesiredPeelCount = 0;
554 unsigned DesiredPeelCountLast = 0;
555
556 // Do not peel the entire loop.
557 const SCEV *BE = SE.getConstantMaxBackedgeTakenCount(L: &L);
558 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Val: BE))
559 MaxPeelCount =
560 std::min(a: (unsigned)SC->getAPInt().getLimitedValue() - 1, b: MaxPeelCount);
561
562 // Increase PeelCount while (IterVal Pred BoundSCEV) condition is satisfied;
563 // return true if inversed condition become known before reaching the
564 // MaxPeelCount limit.
565 auto PeelWhilePredicateIsKnown =
566 [&](unsigned &PeelCount, const SCEV *&IterVal, const SCEV *BoundSCEV,
567 const SCEV *Step, ICmpInst::Predicate Pred) {
568 while (PeelCount < MaxPeelCount &&
569 SE.isKnownPredicate(Pred, LHS: IterVal, RHS: BoundSCEV)) {
570 IterVal = SE.getAddExpr(LHS: IterVal, RHS: Step);
571 ++PeelCount;
572 }
573 return SE.isKnownPredicate(Pred: ICmpInst::getInversePredicate(pred: Pred), LHS: IterVal,
574 RHS: BoundSCEV);
575 };
576
577 const unsigned MaxDepth = 4;
578 std::function<void(Value *, unsigned)> ComputePeelCount =
579 [&](Value *Condition, unsigned Depth) -> void {
580 if (!Condition->getType()->isIntegerTy() || Depth >= MaxDepth)
581 return;
582
583 Value *LeftVal, *RightVal;
584 if (match(V: Condition, P: m_And(L: m_Value(V&: LeftVal), R: m_Value(V&: RightVal))) ||
585 match(V: Condition, P: m_Or(L: m_Value(V&: LeftVal), R: m_Value(V&: RightVal)))) {
586 ComputePeelCount(LeftVal, Depth + 1);
587 ComputePeelCount(RightVal, Depth + 1);
588 return;
589 }
590
591 CmpPredicate Pred;
592 if (!match(V: Condition, P: m_ICmp(Pred, L: m_Value(V&: LeftVal), R: m_Value(V&: RightVal))))
593 return;
594
595 const SCEV *LeftSCEV = SE.getSCEV(V: LeftVal);
596 const SCEV *RightSCEV = SE.getSCEV(V: RightVal);
597
598 // Do not consider predicates that are known to be true or false
599 // independently of the loop iteration.
600 if (SE.evaluatePredicate(Pred, LHS: LeftSCEV, RHS: RightSCEV))
601 return;
602
603 // Check if we have a condition with one AddRec and one non AddRec
604 // expression. Normalize LeftSCEV to be the AddRec.
605 if (!isa<SCEVAddRecExpr>(Val: LeftSCEV)) {
606 if (isa<SCEVAddRecExpr>(Val: RightSCEV)) {
607 std::swap(a&: LeftSCEV, b&: RightSCEV);
608 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
609 } else
610 return;
611 }
612
613 const SCEVAddRecExpr *LeftAR = cast<SCEVAddRecExpr>(Val: LeftSCEV);
614
615 // Avoid huge SCEV computations in the loop below, make sure we only
616 // consider AddRecs of the loop we are trying to peel.
617 if (!LeftAR->isAffine() || LeftAR->getLoop() != &L)
618 return;
619 if (!(ICmpInst::isEquality(P: Pred) && LeftAR->hasNoSelfWrap()) &&
620 !SE.getMonotonicPredicateType(LHS: LeftAR, Pred))
621 return;
622
623 // Check if extending the current DesiredPeelCount lets us evaluate Pred
624 // or !Pred in the loop body statically.
625 unsigned NewPeelCount = DesiredPeelCount;
626
627 const SCEV *IterVal = LeftAR->evaluateAtIteration(
628 It: SE.getConstant(Ty: LeftSCEV->getType(), V: NewPeelCount), SE);
629
630 // If the original condition is not known, get the negated predicate
631 // (which holds on the else branch) and check if it is known. This allows
632 // us to peel of iterations that make the original condition false.
633 if (!SE.isKnownPredicate(Pred, LHS: IterVal, RHS: RightSCEV))
634 Pred = ICmpInst::getInversePredicate(pred: Pred);
635
636 const SCEV *Step = LeftAR->getStepRecurrence(SE);
637 if (!PeelWhilePredicateIsKnown(NewPeelCount, IterVal, RightSCEV, Step,
638 Pred)) {
639 if (shouldPeelLastIteration(L, Pred, LeftAR, RightSCEV, SE, TTI))
640 DesiredPeelCountLast = 1;
641 return;
642 }
643
644 // However, for equality comparisons, that isn't always sufficient to
645 // eliminate the comparsion in loop body, we may need to peel one more
646 // iteration. See if that makes !Pred become unknown again.
647 const SCEV *NextIterVal = SE.getAddExpr(LHS: IterVal, RHS: Step);
648 if (ICmpInst::isEquality(P: Pred) &&
649 !SE.isKnownPredicate(Pred: ICmpInst::getInversePredicate(pred: Pred), LHS: NextIterVal,
650 RHS: RightSCEV) &&
651 !SE.isKnownPredicate(Pred, LHS: IterVal, RHS: RightSCEV) &&
652 SE.isKnownPredicate(Pred, LHS: NextIterVal, RHS: RightSCEV)) {
653 if (NewPeelCount >= MaxPeelCount)
654 return; // Need to peel one more iteration, but can't. Give up.
655 ++NewPeelCount; // Great!
656 }
657
658 DesiredPeelCount = std::max(a: DesiredPeelCount, b: NewPeelCount);
659 DesiredPeelCountLast = std::max(a: DesiredPeelCountLast, b: NewPeelCount);
660 };
661
662 auto ComputePeelCountMinMax = [&](MinMaxIntrinsic *MinMax) {
663 if (!MinMax->getType()->isIntegerTy())
664 return;
665 Value *LHS = MinMax->getLHS(), *RHS = MinMax->getRHS();
666 const SCEV *BoundSCEV, *IterSCEV;
667 if (L.isLoopInvariant(V: LHS)) {
668 BoundSCEV = SE.getSCEV(V: LHS);
669 IterSCEV = SE.getSCEV(V: RHS);
670 } else if (L.isLoopInvariant(V: RHS)) {
671 BoundSCEV = SE.getSCEV(V: RHS);
672 IterSCEV = SE.getSCEV(V: LHS);
673 } else
674 return;
675 const auto *AddRec = dyn_cast<SCEVAddRecExpr>(Val: IterSCEV);
676 // For simplicity, we support only affine recurrences.
677 if (!AddRec || !AddRec->isAffine() || AddRec->getLoop() != &L)
678 return;
679 const SCEV *Step = AddRec->getStepRecurrence(SE);
680 bool IsSigned = MinMax->isSigned();
681 // To minimize number of peeled iterations, we use strict relational
682 // predicates here.
683 ICmpInst::Predicate Pred;
684 if (SE.isKnownPositive(S: Step))
685 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
686 else if (SE.isKnownNegative(S: Step))
687 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
688 else
689 return;
690 // Check that AddRec is not wrapping.
691 if (!(IsSigned ? AddRec->hasNoSignedWrap() : AddRec->hasNoUnsignedWrap()))
692 return;
693 unsigned NewPeelCount = DesiredPeelCount;
694 const SCEV *IterVal = AddRec->evaluateAtIteration(
695 It: SE.getConstant(Ty: AddRec->getType(), V: NewPeelCount), SE);
696 if (!PeelWhilePredicateIsKnown(NewPeelCount, IterVal, BoundSCEV, Step,
697 Pred)) {
698 if (shouldPeelLastIteration(L, Pred, LeftAR: AddRec, RightSCEV: BoundSCEV, SE, TTI))
699 DesiredPeelCountLast = 1;
700 return;
701 }
702 DesiredPeelCount = NewPeelCount;
703 };
704
705 for (BasicBlock *BB : L.blocks()) {
706 for (Instruction &I : *BB) {
707 if (SelectInst *SI = dyn_cast<SelectInst>(Val: &I))
708 ComputePeelCount(SI->getCondition(), 0);
709 if (MinMaxIntrinsic *MinMax = dyn_cast<MinMaxIntrinsic>(Val: &I))
710 ComputePeelCountMinMax(MinMax);
711 }
712
713 auto *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator());
714 if (!BI)
715 continue;
716
717 // Ignore loop exit condition.
718 if (L.getLoopLatch() == BB)
719 continue;
720
721 ComputePeelCount(BI->getCondition(), 0);
722 }
723
724 return {DesiredPeelCount, DesiredPeelCountLast};
725}
726
727/// This "heuristic" exactly matches implicit behavior which used to exist
728/// inside getLoopEstimatedTripCount. It was added here to keep an
729/// improvement inside that API from causing peeling to become more aggressive.
730/// This should probably be removed.
731static bool violatesLegacyMultiExitLoopCheck(Loop *L) {
732 BasicBlock *Latch = L->getLoopLatch();
733 if (!Latch)
734 return true;
735
736 CondBrInst *LatchBR = dyn_cast<CondBrInst>(Val: Latch->getTerminator());
737 if (!LatchBR || !L->isLoopExiting(BB: Latch))
738 return true;
739
740 assert((LatchBR->getSuccessor(0) == L->getHeader() ||
741 LatchBR->getSuccessor(1) == L->getHeader()) &&
742 "At least one edge out of the latch must go to the header");
743
744 SmallVector<BasicBlock *, 4> ExitBlocks;
745 L->getUniqueNonLatchExitBlocks(ExitBlocks);
746 return any_of(Range&: ExitBlocks, P: [](const BasicBlock *EB) {
747 return !EB->getTerminatingDeoptimizeCall();
748 });
749}
750
751
752// Return the number of iterations we want to peel off.
753void llvm::computePeelCount(Loop *L, unsigned LoopSize,
754 TargetTransformInfo::PeelingPreferences &PP,
755 unsigned TripCount, DominatorTree &DT,
756 ScalarEvolution &SE, const TargetTransformInfo &TTI,
757 AssumptionCache *AC, unsigned Threshold) {
758 assert(LoopSize > 0 && "Zero loop size is not allowed!");
759 // Save the PP.PeelCount value set by the target in
760 // TTI.getPeelingPreferences or by the flag -unroll-peel-count.
761 unsigned TargetPeelCount = PP.PeelCount;
762 PP.PeelCount = 0;
763 PP.PeelLast = false;
764 if (!canPeel(L))
765 return;
766
767 // Only try to peel innermost loops by default.
768 // The constraint can be relaxed by the target in TTI.getPeelingPreferences
769 // or by the flag -unroll-allow-loop-nests-peeling.
770 if (!PP.AllowLoopNestsPeeling && !L->isInnermost())
771 return;
772
773 // If the user provided a peel count, use that.
774 bool UserPeelCount = UnrollForcePeelCount.getNumOccurrences() > 0;
775 if (UserPeelCount) {
776 LLVM_DEBUG(dbgs() << "Force-peeling first " << UnrollForcePeelCount
777 << " iterations.\n");
778 PP.PeelCount = UnrollForcePeelCount;
779 PP.PeelProfiledIterations = true;
780 return;
781 }
782
783 // Skip peeling if it's disabled.
784 if (!PP.AllowPeeling)
785 return;
786
787 // Check that we can peel at least one iteration.
788 if (2 * LoopSize > Threshold)
789 return;
790
791 unsigned AlreadyPeeled = 0;
792 if (auto Peeled = getOptionalIntLoopAttribute(TheLoop: L, Name: PeeledCountMetaData))
793 AlreadyPeeled = *Peeled;
794 // Stop if we already peeled off the maximum number of iterations.
795 if (AlreadyPeeled >= UnrollPeelMaxCount)
796 return;
797
798 // Pay respect to limitations implied by loop size and the max peel count.
799 unsigned MaxPeelCount = UnrollPeelMaxCount;
800 MaxPeelCount = std::min(a: MaxPeelCount, b: Threshold / LoopSize - 1);
801
802 // Start the max computation with the PP.PeelCount value set by the target
803 // in TTI.getPeelingPreferences or by the flag -unroll-peel-count.
804 unsigned DesiredPeelCount = TargetPeelCount;
805
806 // Here we try to get rid of Phis which become invariants or inductions after
807 // 1, 2, ..., N iterations of the loop. For this we compute the number for
808 // iterations after which every Phi is guaranteed to become an invariant or an
809 // induction, and try to peel the maximum number of iterations among these
810 // values, thus turning all those Phis into invariants or inductions.
811 if (MaxPeelCount > DesiredPeelCount) {
812 // Check how many iterations are useful for resolving Phis
813 auto NumPeels = PhiAnalyzer(*L, MaxPeelCount, EnablePeelingForIV)
814 .calculateIterationsToPeel();
815 if (NumPeels)
816 DesiredPeelCount = std::max(a: DesiredPeelCount, b: *NumPeels);
817 }
818
819 const auto &[CountToEliminateCmps, CountToEliminateCmpsLast] =
820 countToEliminateCompares(L&: *L, MaxPeelCount, SE, TTI);
821 DesiredPeelCount = std::max(a: DesiredPeelCount, b: CountToEliminateCmps);
822
823 if (DesiredPeelCount == 0)
824 DesiredPeelCount = peelToTurnInvariantLoadsDereferenceable(L&: *L, DT, AC);
825
826 if (DesiredPeelCount > 0) {
827 DesiredPeelCount = std::min(a: DesiredPeelCount, b: MaxPeelCount);
828 // Consider max peel count limitation.
829 assert(DesiredPeelCount > 0 && "Wrong loop size estimation?");
830 if (DesiredPeelCount + AlreadyPeeled <= UnrollPeelMaxCount) {
831 LLVM_DEBUG(dbgs() << "Peel " << DesiredPeelCount
832 << " iteration(s) to turn"
833 << " some Phis into invariants or inductions.\n");
834 PP.PeelCount = DesiredPeelCount;
835 PP.PeelProfiledIterations = false;
836 PP.PeelLast = false;
837 return;
838 }
839 }
840
841 if (CountToEliminateCmpsLast > 0) {
842 unsigned DesiredPeelCountLast =
843 std::min(a: CountToEliminateCmpsLast, b: MaxPeelCount);
844 // Consider max peel count limitation.
845 assert(DesiredPeelCountLast > 0 && "Wrong loop size estimation?");
846 if (DesiredPeelCountLast + AlreadyPeeled <= UnrollPeelMaxCount) {
847 LLVM_DEBUG(dbgs() << "Peel " << DesiredPeelCount
848 << " iteration(s) to turn"
849 << " some Phis into invariants.\n");
850 PP.PeelCount = DesiredPeelCountLast;
851 PP.PeelProfiledIterations = false;
852 PP.PeelLast = true;
853 return;
854 }
855 }
856
857 // Bail if we know the statically calculated trip count.
858 // In this case we rather prefer partial unrolling.
859 if (TripCount)
860 return;
861
862 // Do not apply profile base peeling if it is disabled.
863 if (!PP.PeelProfiledIterations)
864 return;
865 // If we don't know the trip count, but have reason to believe the average
866 // trip count is low, peeling should be beneficial, since we will usually
867 // hit the peeled section.
868 // We only do this in the presence of profile information, since otherwise
869 // our estimates of the trip count are not reliable enough.
870 if (L->getHeader()->getParent()->hasProfileData()) {
871 if (violatesLegacyMultiExitLoopCheck(L))
872 return;
873 std::optional<unsigned> EstimatedTripCount = getLoopEstimatedTripCount(L);
874 if (!EstimatedTripCount)
875 return;
876
877 LLVM_DEBUG(dbgs() << "Profile-based estimated trip count is "
878 << *EstimatedTripCount << "\n");
879
880 std::optional<unsigned> TotalPeeled =
881 llvm::checkedAddUnsigned(LHS: *EstimatedTripCount, RHS: AlreadyPeeled);
882 if (TotalPeeled && *TotalPeeled <= MaxPeelCount) {
883 unsigned PeelCount = *EstimatedTripCount;
884 LLVM_DEBUG(dbgs() << "Peeling first " << PeelCount << " iterations.\n");
885 PP.PeelCount = PeelCount;
886 return;
887 }
888 LLVM_DEBUG(dbgs() << "Already peel count: " << AlreadyPeeled << "\n");
889 LLVM_DEBUG(dbgs() << "Max peel count: " << UnrollPeelMaxCount << "\n");
890 LLVM_DEBUG(dbgs() << "Loop cost: " << LoopSize << "\n");
891 LLVM_DEBUG(dbgs() << "Max peel cost: " << Threshold << "\n");
892 LLVM_DEBUG(dbgs() << "Max peel count by cost: "
893 << (Threshold / LoopSize - 1) << "\n");
894 }
895}
896
897/// Clones the body of the loop L, putting it between \p InsertTop and \p
898/// InsertBot.
899/// \param IterNumber The serial number of the iteration currently being
900/// peeled off.
901/// \param PeelLast Peel off the last iterations from \p L.
902/// \param ExitEdges The exit edges of the original loop.
903/// \param[out] NewBlocks A list of the blocks in the newly created clone
904/// \param[out] VMap The value map between the loop and the new clone.
905/// \param LoopBlocks A helper for DFS-traversal of the loop.
906/// \param LVMap A value-map that maps instructions from the original loop to
907/// instructions in the last peeled-off iteration.
908static void cloneLoopBlocks(
909 Loop *L, unsigned IterNumber, bool PeelLast, BasicBlock *InsertTop,
910 BasicBlock *InsertBot, BasicBlock *OrigPreHeader,
911 SmallVectorImpl<std::pair<BasicBlock *, BasicBlock *>> &ExitEdges,
912 SmallVectorImpl<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks,
913 ValueToValueMapTy &VMap, ValueToValueMapTy &LVMap, DominatorTree *DT,
914 LoopInfo *LI, ArrayRef<MDNode *> LoopLocalNoAliasDeclScopes,
915 ScalarEvolution &SE) {
916 BasicBlock *Header = L->getHeader();
917 BasicBlock *Latch = L->getLoopLatch();
918 BasicBlock *PreHeader = L->getLoopPreheader();
919
920 Function *F = Header->getParent();
921 LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
922 LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
923 Loop *ParentLoop = L->getParentLoop();
924
925 // For each block in the original loop, create a new copy,
926 // and update the value map with the newly created values.
927 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
928 BasicBlock *NewBB = CloneBasicBlock(BB: *BB, VMap, NameSuffix: ".peel", F);
929 NewBlocks.push_back(Elt: NewBB);
930
931 // If an original block is an immediate child of the loop L, its copy
932 // is a child of a ParentLoop after peeling. If a block is a child of
933 // a nested loop, it is handled in the cloneLoop() call below.
934 if (ParentLoop && LI->getLoopFor(BB: *BB) == L)
935 ParentLoop->addBasicBlockToLoop(NewBB, LI&: *LI);
936
937 VMap[*BB] = NewBB;
938
939 // If dominator tree is available, insert nodes to represent cloned blocks.
940 if (DT) {
941 if (Header == *BB)
942 DT->addNewBlock(BB: NewBB, DomBB: InsertTop);
943 else {
944 DomTreeNode *IDom = DT->getNode(BB: *BB)->getIDom();
945 // VMap must contain entry for IDom, as the iteration order is RPO.
946 DT->addNewBlock(BB: NewBB, DomBB: cast<BasicBlock>(Val&: VMap[IDom->getBlock()]));
947 }
948 }
949 }
950
951 {
952 // Identify what other metadata depends on the cloned version. After
953 // cloning, replace the metadata with the corrected version for both
954 // memory instructions and noalias intrinsics.
955 std::string Ext = (Twine("Peel") + Twine(IterNumber)).str();
956 cloneAndAdaptNoAliasScopes(NoAliasDeclScopes: LoopLocalNoAliasDeclScopes, NewBlocks,
957 Context&: Header->getContext(), Ext);
958 }
959
960 // Recursively create the new Loop objects for nested loops, if any,
961 // to preserve LoopInfo.
962 for (Loop *ChildLoop : *L) {
963 cloneLoop(L: ChildLoop, PL: ParentLoop, VM&: VMap, LI, LPM: nullptr);
964 }
965
966 // Hook-up the control flow for the newly inserted blocks.
967 // The new header is hooked up directly to the "top", which is either
968 // the original loop preheader (for the first iteration) or the previous
969 // iteration's exiting block (for every other iteration)
970 InsertTop->getTerminator()->setSuccessor(Idx: 0, BB: cast<BasicBlock>(Val&: VMap[Header]));
971
972 // Similarly, for the latch:
973 // The original exiting edge is still hooked up to the loop exit.
974 BasicBlock *NewLatch = cast<BasicBlock>(Val&: VMap[Latch]);
975 if (PeelLast) {
976 // This is the last iteration and we definitely will go to the exit. Just
977 // set both successors to InsertBot and let the branch be simplified later.
978 assert(IterNumber == 0 && "Only peeling a single iteration implemented.");
979 auto *LatchTerm = cast<CondBrInst>(Val: NewLatch->getTerminator());
980 LatchTerm->setSuccessor(idx: 0, NewSucc: InsertBot);
981 LatchTerm->setSuccessor(idx: 1, NewSucc: InsertBot);
982 } else {
983 auto *LatchTerm = cast<Instruction>(Val: NewLatch->getTerminator());
984 // The backedge now goes to the "bottom", which is either the loop's real
985 // header (for the last peeled iteration) or the copied header of the next
986 // iteration (for every other iteration)
987 for (unsigned idx = 0, e = LatchTerm->getNumSuccessors(); idx < e; ++idx) {
988 if (LatchTerm->getSuccessor(Idx: idx) == Header) {
989 LatchTerm->setSuccessor(Idx: idx, BB: InsertBot);
990 break;
991 }
992 }
993 }
994 if (DT)
995 DT->changeImmediateDominator(BB: InsertBot, NewBB: NewLatch);
996
997 // The new copy of the loop body starts with a bunch of PHI nodes
998 // that pick an incoming value from either the preheader, or the previous
999 // loop iteration. Since this copy is no longer part of the loop, we
1000 // resolve this statically:
1001 if (PeelLast) {
1002 // For the last iteration, we introduce new phis for each header phi in
1003 // InsertTop, using the incoming value from the preheader for the original
1004 // preheader (when skipping the main loop) and the incoming value from the
1005 // latch for the latch (when continuing from the main loop).
1006 IRBuilder<> B(InsertTop, InsertTop->getFirstNonPHIIt());
1007 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(Val: I); ++I) {
1008 PHINode *NewPHI = cast<PHINode>(Val&: VMap[&*I]);
1009 PHINode *PN = B.CreatePHI(Ty: NewPHI->getType(), NumReservedValues: 2);
1010 NewPHI->eraseFromParent();
1011 if (OrigPreHeader)
1012 PN->addIncoming(V: cast<PHINode>(Val: &*I)->getIncomingValueForBlock(BB: PreHeader),
1013 BB: OrigPreHeader);
1014
1015 PN->addIncoming(V: cast<PHINode>(Val: &*I)->getIncomingValueForBlock(BB: Latch),
1016 BB: Latch);
1017 VMap[&*I] = PN;
1018 }
1019 } else {
1020 // For the first iteration, we use the value from the preheader directly.
1021 // For any other iteration, we replace the phi with the value generated by
1022 // the immediately preceding clone of the loop body (which represents
1023 // the previous iteration).
1024 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(Val: I); ++I) {
1025 PHINode *NewPHI = cast<PHINode>(Val&: VMap[&*I]);
1026 if (IterNumber == 0) {
1027 VMap[&*I] = NewPHI->getIncomingValueForBlock(BB: PreHeader);
1028 } else {
1029 Value *LatchVal = NewPHI->getIncomingValueForBlock(BB: Latch);
1030 Instruction *LatchInst = dyn_cast<Instruction>(Val: LatchVal);
1031 if (LatchInst && L->contains(Inst: LatchInst))
1032 VMap[&*I] = LVMap[LatchInst];
1033 else
1034 VMap[&*I] = LatchVal;
1035 }
1036 NewPHI->eraseFromParent();
1037 }
1038 }
1039
1040 // Fix up the outgoing values - we need to add a value for the iteration
1041 // we've just created. Note that this must happen *after* the incoming
1042 // values are adjusted, since the value going out of the latch may also be
1043 // a value coming into the header.
1044 for (auto Edge : ExitEdges)
1045 for (PHINode &PHI : Edge.second->phis()) {
1046 Value *LatchVal = PHI.getIncomingValueForBlock(BB: Edge.first);
1047 Instruction *LatchInst = dyn_cast<Instruction>(Val: LatchVal);
1048 if (LatchInst && L->contains(Inst: LatchInst))
1049 LatchVal = VMap[LatchVal];
1050 PHI.addIncoming(V: LatchVal, BB: cast<BasicBlock>(Val&: VMap[Edge.first]));
1051 SE.forgetLcssaPhiWithNewPredecessor(L, V: &PHI);
1052 }
1053
1054 // LastValueMap is updated with the values for the current loop
1055 // which are used the next time this function is called.
1056 for (auto KV : VMap)
1057 LVMap[KV.first] = KV.second;
1058}
1059
1060TargetTransformInfo::PeelingPreferences
1061llvm::gatherPeelingPreferences(Loop *L, ScalarEvolution &SE,
1062 const TargetTransformInfo &TTI,
1063 std::optional<bool> UserAllowPeeling,
1064 std::optional<bool> UserAllowProfileBasedPeeling,
1065 bool UnrollingSpecficValues) {
1066 TargetTransformInfo::PeelingPreferences PP;
1067
1068 // Set the default values.
1069 PP.PeelCount = 0;
1070 PP.AllowPeeling = true;
1071 PP.AllowLoopNestsPeeling = false;
1072 PP.PeelLast = false;
1073 PP.PeelProfiledIterations = true;
1074
1075 // Get the target specifc values.
1076 TTI.getPeelingPreferences(L, SE, PP);
1077
1078 // User specified values using cl::opt.
1079 if (UnrollingSpecficValues) {
1080 if (UnrollPeelCount.getNumOccurrences() > 0)
1081 PP.PeelCount = UnrollPeelCount;
1082 if (UnrollAllowPeeling.getNumOccurrences() > 0)
1083 PP.AllowPeeling = UnrollAllowPeeling;
1084 if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0)
1085 PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling;
1086 }
1087
1088 // User specifed values provided by argument.
1089 if (UserAllowPeeling)
1090 PP.AllowPeeling = *UserAllowPeeling;
1091 if (UserAllowProfileBasedPeeling)
1092 PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling;
1093
1094 return PP;
1095}
1096
1097/// Peel off the first \p PeelCount iterations of loop \p L.
1098///
1099/// Note that this does not peel them off as a single straight-line block.
1100/// Rather, each iteration is peeled off separately, and needs to check the
1101/// exit condition.
1102/// For loops that dynamically execute \p PeelCount iterations or less
1103/// this provides a benefit, since the peeled off iterations, which account
1104/// for the bulk of dynamic execution, can be further simplified by scalar
1105/// optimizations.
1106void llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
1107 ScalarEvolution *SE, DominatorTree &DT, AssumptionCache *AC,
1108 bool PreserveLCSSA, ValueToValueMapTy &LVMap) {
1109 assert(PeelCount > 0 && "Attempt to peel out zero iterations?");
1110 assert(canPeel(L) && "Attempt to peel a loop which is not peelable?");
1111 assert((!PeelLast || (canPeelLastIteration(*L, *SE) && PeelCount == 1)) &&
1112 "when peeling the last iteration, the loop must be supported and can "
1113 "only peel a single iteration");
1114
1115 LoopBlocksDFS LoopBlocks(L);
1116 LoopBlocks.perform(LI);
1117
1118 BasicBlock *Header = L->getHeader();
1119 BasicBlock *PreHeader = L->getLoopPreheader();
1120 BasicBlock *Latch = L->getLoopLatch();
1121 SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitEdges;
1122 LI->getExitEdges(L: *L, ExitEdges);
1123
1124 // Remember dominators of blocks we might reach through exits to change them
1125 // later. Immediate dominator of such block might change, because we add more
1126 // routes which can lead to the exit: we can reach it from the peeled
1127 // iterations too.
1128 MapVector<BasicBlock *, BasicBlock *> NonLoopBlocksIDom;
1129 for (auto *BB : L->blocks()) {
1130 auto *BBDomNode = DT.getNode(BB);
1131 SmallVector<BasicBlock *, 16> ChildrenToUpdate;
1132 for (auto *ChildDomNode : BBDomNode->children()) {
1133 auto *ChildBB = ChildDomNode->getBlock();
1134 if (!L->contains(BB: ChildBB))
1135 ChildrenToUpdate.push_back(Elt: ChildBB);
1136 }
1137 // The new idom of the block will be the nearest common dominator
1138 // of all copies of the previous idom. This is equivalent to the
1139 // nearest common dominator of the previous idom and the first latch,
1140 // which dominates all copies of the previous idom.
1141 BasicBlock *NewIDom = DT.findNearestCommonDominator(A: BB, B: Latch);
1142 for (auto *ChildBB : ChildrenToUpdate)
1143 NonLoopBlocksIDom[ChildBB] = NewIDom;
1144 }
1145
1146 Function *F = Header->getParent();
1147
1148 // Set up all the necessary basic blocks.
1149 BasicBlock *InsertTop;
1150 BasicBlock *InsertBot;
1151 BasicBlock *NewPreHeader = nullptr;
1152 DenseMap<Instruction *, Value *> ExitValues;
1153 if (PeelLast) {
1154 // It is convenient to split the single exit block from the latch the
1155 // into 3 parts - two blocks to anchor the peeled copy of the loop body,
1156 // and a new final exit block.
1157
1158 // Peeling the last iteration transforms.
1159 //
1160 // PreHeader:
1161 // ...
1162 // Header:
1163 // LoopBody
1164 // If (cond) goto Header
1165 // Exit:
1166 //
1167 // into
1168 //
1169 // Header:
1170 // LoopBody
1171 // If (cond) goto Header
1172 // InsertTop:
1173 // LoopBody
1174 // If (!cond) goto InsertBot
1175 // InsertBot:
1176 // Exit:
1177 // ...
1178 BasicBlock *Exit = L->getExitBlock();
1179 for (PHINode &P : Exit->phis())
1180 ExitValues[&P] = P.getIncomingValueForBlock(BB: Latch);
1181
1182 const SCEV *BTC = SE->getBackedgeTakenCount(L);
1183
1184 InsertTop = SplitEdge(From: Latch, To: Exit, DT: &DT, LI);
1185 InsertBot = SplitBlock(Old: InsertTop, SplitPt: InsertTop->getTerminator(), DT: &DT, LI);
1186
1187 InsertTop->setName(Exit->getName() + ".peel.begin");
1188 InsertBot->setName(Exit->getName() + ".peel.next");
1189 NewPreHeader = nullptr;
1190
1191 // If the original loop may only execute a single iteration we need to
1192 // insert a trip count check and skip the original loop with the last
1193 // iteration peeled off if necessary. Either way, we must update branch
1194 // weights to maintain the loop body frequency.
1195 if (SE->isKnownNonZero(S: BTC)) {
1196 // We have just proven that, when reached, the original loop always
1197 // executes at least two iterations. Thus, we unconditionally execute
1198 // both the remaining loop's initial iteration and the peeled iteration.
1199 // But that increases the latter's frequency above its frequency in the
1200 // original loop. To maintain the total frequency, we compensate by
1201 // decreasing the remaining loop body's frequency to indicate one less
1202 // iteration.
1203 //
1204 // We use this formula to convert probability to/from frequency:
1205 // Sum(i=0..inf)(P^i) = 1/(1-P) = Freq.
1206 if (BranchProbability P = getLoopProbability(L); !P.isUnknown()) {
1207 // Trying to subtract one from an infinite loop is pointless, and our
1208 // formulas then produce division by zero, so skip that case.
1209 if (BranchProbability ExitP = P.getCompl(); !ExitP.isZero()) {
1210 double Freq = 1 / ExitP.toDouble();
1211 // No branch weights can produce a frequency of less than one given
1212 // the initial iteration, and our formulas produce a negative
1213 // probability if we try.
1214 assert(Freq >= 1.0 && "expected freq >= 1 due to initial iteration");
1215 double NewFreq = std::max(a: Freq - 1, b: 1.0);
1216 setLoopProbability(
1217 L, P: BranchProbability::getBranchProbability(Prob: 1 - 1 / NewFreq));
1218 }
1219 }
1220 } else {
1221 NewPreHeader = SplitEdge(From: PreHeader, To: Header, DT: &DT, LI);
1222 SCEVExpander Expander(*SE, "loop-peel");
1223
1224 Instruction *PreHeaderBR = PreHeader->getTerminator();
1225 Value *BTCValue =
1226 Expander.expandCodeFor(SH: BTC, Ty: BTC->getType(), I: PreHeaderBR);
1227 IRBuilder<> B(PreHeaderBR);
1228 Value *Cond =
1229 B.CreateICmpNE(LHS: BTCValue, RHS: ConstantInt::get(Ty: BTCValue->getType(), V: 0));
1230 auto *BI = B.CreateCondBr(Cond, True: NewPreHeader, False: InsertTop);
1231 SmallVector<uint32_t> Weights;
1232 auto *OrigLatchBr = Latch->getTerminator();
1233 auto HasBranchWeights = extractBranchWeights(I: *OrigLatchBr, Weights);
1234 if (HasBranchWeights) {
1235 // The probability that the new guard skips the loop to execute just one
1236 // iteration is the original loop's probability of exiting at the latch
1237 // after any iteration. That should maintain the original loop body
1238 // frequency. Upon arriving at the loop, due to the guard, the
1239 // probability of reaching iteration i of the new loop is the
1240 // probability of reaching iteration i+1 of the original loop. The
1241 // probability of reaching the peeled iteration is 1, which is the
1242 // probability of reaching iteration 0 of the original loop.
1243 if (L->getExitBlock() == OrigLatchBr->getSuccessor(Idx: 0))
1244 std::swap(a&: Weights[0], b&: Weights[1]);
1245 setBranchWeights(I&: *BI, Weights, /*IsExpected=*/false);
1246 }
1247 PreHeaderBR->eraseFromParent();
1248
1249 // PreHeader now dominates InsertTop.
1250 DT.changeImmediateDominator(BB: InsertTop, NewBB: PreHeader);
1251 }
1252 } else {
1253 // It is convenient to split the preheader into 3 parts - two blocks to
1254 // anchor the peeled copy of the loop body, and a new preheader for the
1255 // "real" loop.
1256
1257 // Peeling the first iteration transforms.
1258 //
1259 // PreHeader:
1260 // ...
1261 // Header:
1262 // LoopBody
1263 // If (cond) goto Header
1264 // Exit:
1265 //
1266 // into
1267 //
1268 // InsertTop:
1269 // LoopBody
1270 // If (!cond) goto Exit
1271 // InsertBot:
1272 // NewPreHeader:
1273 // ...
1274 // Header:
1275 // LoopBody
1276 // If (cond) goto Header
1277 // Exit:
1278 //
1279 // Each following iteration will split the current bottom anchor in two,
1280 // and put the new copy of the loop body between these two blocks. That
1281 // is, after peeling another iteration from the example above, we'll
1282 // split InsertBot, and get:
1283 //
1284 // InsertTop:
1285 // LoopBody
1286 // If (!cond) goto Exit
1287 // InsertBot:
1288 // LoopBody
1289 // If (!cond) goto Exit
1290 // InsertBot.next:
1291 // NewPreHeader:
1292 // ...
1293 // Header:
1294 // LoopBody
1295 // If (cond) goto Header
1296 // Exit:
1297 //
1298 InsertTop = SplitEdge(From: PreHeader, To: Header, DT: &DT, LI);
1299 InsertBot = SplitBlock(Old: InsertTop, SplitPt: InsertTop->getTerminator(), DT: &DT, LI);
1300 NewPreHeader = SplitBlock(Old: InsertBot, SplitPt: InsertBot->getTerminator(), DT: &DT, LI);
1301
1302 InsertTop->setName(Header->getName() + ".peel.begin");
1303 InsertBot->setName(Header->getName() + ".peel.next");
1304 NewPreHeader->setName(PreHeader->getName() + ".peel.newph");
1305 }
1306
1307 Instruction *LatchTerm =
1308 cast<Instruction>(Val: cast<BasicBlock>(Val: Latch)->getTerminator());
1309
1310 // Identify what noalias metadata is inside the loop: if it is inside the
1311 // loop, the associated metadata must be cloned for each iteration.
1312 SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes;
1313 identifyNoAliasScopesToClone(BBs: L->getBlocks(), NoAliasDeclScopes&: LoopLocalNoAliasDeclScopes);
1314
1315 // For each peeled-off iteration, make a copy of the loop.
1316 ValueToValueMapTy VMap;
1317 for (unsigned Iter = 0; Iter < PeelCount; ++Iter) {
1318 SmallVector<BasicBlock *, 8> NewBlocks;
1319
1320 cloneLoopBlocks(L, IterNumber: Iter, PeelLast, InsertTop, InsertBot,
1321 OrigPreHeader: NewPreHeader ? PreHeader : nullptr, ExitEdges, NewBlocks,
1322 LoopBlocks, VMap, LVMap, DT: &DT, LI,
1323 LoopLocalNoAliasDeclScopes, SE&: *SE);
1324
1325 // Remap to use values from the current iteration instead of the
1326 // previous one.
1327 remapInstructionsInBlocks(Blocks: NewBlocks, VMap);
1328
1329 if (Iter == 0) {
1330 if (PeelLast) {
1331 // Adjust the exit condition so the loop exits one iteration early.
1332 // For now we simply subtract one form the second operand of the
1333 // exit condition. This relies on the peel count computation to
1334 // check that this is actually legal. In particular, it ensures that
1335 // the first operand of the compare is an AddRec with step 1 and we
1336 // execute more than one iteration.
1337 auto *Cmp =
1338 cast<ICmpInst>(Val: L->getLoopLatch()->getTerminator()->getOperand(i: 0));
1339 IRBuilder B(Cmp);
1340 Cmp->setOperand(
1341 i_nocapture: 1, Val_nocapture: B.CreateSub(LHS: Cmp->getOperand(i_nocapture: 1),
1342 RHS: ConstantInt::get(Ty: Cmp->getOperand(i_nocapture: 1)->getType(), V: 1)));
1343 } else {
1344 // Update IDoms of the blocks reachable through exits.
1345 for (auto BBIDom : NonLoopBlocksIDom)
1346 DT.changeImmediateDominator(BB: BBIDom.first,
1347 NewBB: cast<BasicBlock>(Val&: LVMap[BBIDom.second]));
1348 }
1349 }
1350
1351#ifdef EXPENSIVE_CHECKS
1352 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1353#endif
1354
1355 // Remove Loop metadata from the latch branch instruction
1356 // because it is not the Loop's latch branch anymore.
1357 auto *LatchTermCopy = cast<Instruction>(Val&: VMap[LatchTerm]);
1358 LatchTermCopy->setMetadata(KindID: LLVMContext::MD_loop, Node: nullptr);
1359
1360 InsertTop = InsertBot;
1361 InsertBot = SplitBlock(Old: InsertBot, SplitPt: InsertBot->getTerminator(), DT: &DT, LI);
1362 InsertBot->setName(Header->getName() + ".peel.next");
1363
1364 F->splice(ToIt: InsertTop->getIterator(), FromF: F, FromBeginIt: NewBlocks[0]->getIterator(),
1365 FromEndIt: F->end());
1366 }
1367
1368 if (PeelLast) {
1369 // Now adjust users of the original exit values by replacing them with the
1370 // exit value from the peeled iteration and remove them.
1371 for (const auto &[P, E] : ExitValues) {
1372 Instruction *ExitInst = dyn_cast<Instruction>(Val: E);
1373 if (ExitInst && L->contains(Inst: ExitInst))
1374 P->replaceAllUsesWith(V: &*VMap[ExitInst]);
1375 else
1376 P->replaceAllUsesWith(V: E);
1377 P->eraseFromParent();
1378 }
1379 formLCSSA(L&: *L, DT, LI, SE);
1380 } else {
1381 // Now adjust the phi nodes in the loop header to get their initial values
1382 // from the last peeled-off iteration instead of the preheader.
1383 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(Val: I); ++I) {
1384 PHINode *PHI = cast<PHINode>(Val&: I);
1385 Value *NewVal = PHI->getIncomingValueForBlock(BB: Latch);
1386 Instruction *LatchInst = dyn_cast<Instruction>(Val: NewVal);
1387 if (LatchInst && L->contains(Inst: LatchInst))
1388 NewVal = LVMap[LatchInst];
1389
1390 PHI->setIncomingValueForBlock(BB: NewPreHeader, V: NewVal);
1391 }
1392 }
1393
1394 // Update Metadata for count of peeled off iterations.
1395 unsigned AlreadyPeeled = 0;
1396 if (auto Peeled = getOptionalIntLoopAttribute(TheLoop: L, Name: PeeledCountMetaData))
1397 AlreadyPeeled = *Peeled;
1398 unsigned TotalPeeled = AlreadyPeeled + PeelCount;
1399 addStringMetadataToLoop(TheLoop: L, MDString: PeeledCountMetaData, V: TotalPeeled);
1400
1401 // Update metadata for the estimated trip count. The original branch weight
1402 // metadata is already correct for both the remaining loop and the peeled loop
1403 // iterations, so do not adjust it.
1404 //
1405 // For example, consider what happens when peeling 2 iterations from a loop
1406 // with an estimated trip count of 10 and inserting them before the remaining
1407 // loop. Each of the peeled iterations and each iteration in the remaining
1408 // loop still has the same probability of exiting the *entire original* loop
1409 // as it did when in the original loop, and thus it should still have the same
1410 // branch weights. The peeled iterations' non-zero probabilities of exiting
1411 // already appropriately reduce the probability of reaching the remaining
1412 // iterations just as they did in the original loop. Trying to also adjust
1413 // the remaining loop's branch weights to reflect its new trip count of 8 will
1414 // erroneously further reduce its block frequencies. However, in case an
1415 // analysis later needs to determine the trip count of the remaining loop
1416 // while examining it in isolation without considering the probability of
1417 // actually reaching it, we store the new trip count as separate metadata.
1418 if (auto EstimatedTripCount = getLoopEstimatedTripCount(L)) {
1419 unsigned EstimatedTripCountNew = *EstimatedTripCount;
1420 if (EstimatedTripCountNew < TotalPeeled)
1421 EstimatedTripCountNew = 0;
1422 else
1423 EstimatedTripCountNew -= TotalPeeled;
1424 setLoopEstimatedTripCount(L, EstimatedTripCount: EstimatedTripCountNew);
1425 }
1426
1427 if (Loop *ParentLoop = L->getParentLoop())
1428 L = ParentLoop;
1429
1430 // We modified the loop, update SE.
1431 SE->forgetTopmostLoop(L);
1432 SE->forgetBlockAndLoopDispositions();
1433
1434#ifdef EXPENSIVE_CHECKS
1435 // Finally DomtTree must be correct.
1436 assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1437#endif
1438
1439 // FIXME: Incrementally update loop-simplify
1440 simplifyLoop(L, DT: &DT, LI, SE, AC, MSSAU: nullptr, PreserveLCSSA);
1441
1442 NumPeeled++;
1443 NumPeeledEnd += PeelLast;
1444}
1445