xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/LoopPeel.cpp (revision 734e82fe33aa764367791a7d603b383996c6b40b)
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/SmallVector.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/Loads.h"
17 #include "llvm/Analysis/LoopInfo.h"
18 #include "llvm/Analysis/LoopIterator.h"
19 #include "llvm/Analysis/ScalarEvolution.h"
20 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/InstrTypes.h"
26 #include "llvm/IR/Instruction.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/MDBuilder.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/IR/ProfDataUtils.h"
32 #include "llvm/Support/Casting.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/LoopSimplify.h"
39 #include "llvm/Transforms/Utils/LoopUtils.h"
40 #include "llvm/Transforms/Utils/ValueMapper.h"
41 #include <algorithm>
42 #include <cassert>
43 #include <cstdint>
44 #include <optional>
45 
46 using namespace llvm;
47 using namespace llvm::PatternMatch;
48 
49 #define DEBUG_TYPE "loop-peel"
50 
51 STATISTIC(NumPeeled, "Number of loops peeled");
52 
53 static cl::opt<unsigned> UnrollPeelCount(
54     "unroll-peel-count", cl::Hidden,
55     cl::desc("Set the unroll peeling count, for testing purposes"));
56 
57 static cl::opt<bool>
58     UnrollAllowPeeling("unroll-allow-peeling", cl::init(true), cl::Hidden,
59                        cl::desc("Allows loops to be peeled when the dynamic "
60                                 "trip count is known to be low."));
61 
62 static cl::opt<bool>
63     UnrollAllowLoopNestsPeeling("unroll-allow-loop-nests-peeling",
64                                 cl::init(false), cl::Hidden,
65                                 cl::desc("Allows loop nests to be peeled."));
66 
67 static cl::opt<unsigned> UnrollPeelMaxCount(
68     "unroll-peel-max-count", cl::init(7), cl::Hidden,
69     cl::desc("Max average trip count which will cause loop peeling."));
70 
71 static cl::opt<unsigned> UnrollForcePeelCount(
72     "unroll-force-peel-count", cl::init(0), cl::Hidden,
73     cl::desc("Force a peel count regardless of profiling information."));
74 
75 static cl::opt<bool> DisableAdvancedPeeling(
76     "disable-advanced-peeling", cl::init(false), cl::Hidden,
77     cl::desc(
78         "Disable advance peeling. Issues for convergent targets (D134803)."));
79 
80 static const char *PeeledCountMetaData = "llvm.loop.peeled.count";
81 
82 // Check whether we are capable of peeling this loop.
83 bool llvm::canPeel(const Loop *L) {
84   // Make sure the loop is in simplified form
85   if (!L->isLoopSimplifyForm())
86     return false;
87   if (!DisableAdvancedPeeling)
88     return true;
89 
90   SmallVector<BasicBlock *, 4> Exits;
91   L->getUniqueNonLatchExitBlocks(Exits);
92   // The latch must either be the only exiting block or all non-latch exit
93   // blocks have either a deopt or unreachable terminator or compose a chain of
94   // blocks where the last one is either deopt or unreachable terminated. Both
95   // deopt and unreachable terminators are a strong indication they are not
96   // taken. Note that this is a profitability check, not a legality check. Also
97   // note that LoopPeeling currently can only update the branch weights of latch
98   // blocks and branch weights to blocks with deopt or unreachable do not need
99   // updating.
100   return llvm::all_of(Exits, IsBlockFollowedByDeoptOrUnreachable);
101 }
102 
103 namespace {
104 
105 // As a loop is peeled, it may be the case that Phi nodes become
106 // loop-invariant (ie, known because there is only one choice).
107 // For example, consider the following function:
108 //   void g(int);
109 //   void binary() {
110 //     int x = 0;
111 //     int y = 0;
112 //     int a = 0;
113 //     for(int i = 0; i <100000; ++i) {
114 //       g(x);
115 //       x = y;
116 //       g(a);
117 //       y = a + 1;
118 //       a = 5;
119 //     }
120 //   }
121 // Peeling 3 iterations is beneficial because the values for x, y and a
122 // become known.  The IR for this loop looks something like the following:
123 //
124 //   %i = phi i32 [ 0, %entry ], [ %inc, %if.end ]
125 //   %a = phi i32 [ 0, %entry ], [ 5, %if.end ]
126 //   %y = phi i32 [ 0, %entry ], [ %add, %if.end ]
127 //   %x = phi i32 [ 0, %entry ], [ %y, %if.end ]
128 //   ...
129 //   tail call void @_Z1gi(i32 signext %x)
130 //   tail call void @_Z1gi(i32 signext %a)
131 //   %add = add nuw nsw i32 %a, 1
132 //   %inc = add nuw nsw i32 %i, 1
133 //   %exitcond = icmp eq i32 %inc, 100000
134 //   br i1 %exitcond, label %for.cond.cleanup, label %for.body
135 //
136 // The arguments for the calls to g will become known after 3 iterations
137 // of the loop, because the phi nodes values become known after 3 iterations
138 // of the loop (ie, they are known on the 4th iteration, so peel 3 iterations).
139 // The first iteration has g(0), g(0); the second has g(0), g(5); the
140 // third has g(1), g(5) and the fourth (and all subsequent) have g(6), g(5).
141 // Now consider the phi nodes:
142 //   %a is a phi with constants so it is determined after iteration 1.
143 //   %y is a phi based on a constant and %a so it is determined on
144 //     the iteration after %a is determined, so iteration 2.
145 //   %x is a phi based on a constant and %y so it is determined on
146 //     the iteration after %y, so iteration 3.
147 //   %i is based on itself (and is an induction variable) so it is
148 //     never determined.
149 // This means that peeling off 3 iterations will result in being able to
150 // remove the phi nodes for %a, %y, and %x.  The arguments for the
151 // corresponding calls to g are determined and the code for computing
152 // x, y, and a can be removed.
153 //
154 // The PhiAnalyzer class calculates how many times a loop should be
155 // peeled based on the above analysis of the phi nodes in the loop while
156 // respecting the maximum specified.
157 class PhiAnalyzer {
158 public:
159   PhiAnalyzer(const Loop &L, unsigned MaxIterations);
160 
161   // Calculate the sufficient minimum number of iterations of the loop to peel
162   // such that phi instructions become determined (subject to allowable limits)
163   std::optional<unsigned> calculateIterationsToPeel();
164 
165 protected:
166   using PeelCounter = std::optional<unsigned>;
167   const PeelCounter Unknown = std::nullopt;
168 
169   // Add 1 respecting Unknown and return Unknown if result over MaxIterations
170   PeelCounter addOne(PeelCounter PC) const {
171     if (PC == Unknown)
172       return Unknown;
173     return (*PC + 1 <= MaxIterations) ? PeelCounter{*PC + 1} : Unknown;
174   }
175 
176   // Calculate the number of iterations after which the given value
177   // becomes an invariant.
178   PeelCounter calculate(const Value &);
179 
180   const Loop &L;
181   const unsigned MaxIterations;
182 
183   // Map of Values to number of iterations to invariance
184   SmallDenseMap<const Value *, PeelCounter> IterationsToInvariance;
185 };
186 
187 PhiAnalyzer::PhiAnalyzer(const Loop &L, unsigned MaxIterations)
188     : L(L), MaxIterations(MaxIterations) {
189   assert(canPeel(&L) && "loop is not suitable for peeling");
190   assert(MaxIterations > 0 && "no peeling is allowed?");
191 }
192 
193 // This function calculates the number of iterations after which the value
194 // becomes an invariant. The pre-calculated values are memorized in a map.
195 // N.B. This number will be Unknown or <= MaxIterations.
196 // The function is calculated according to the following definition:
197 // Given %x = phi <Inputs from above the loop>, ..., [%y, %back.edge].
198 //   F(%x) = G(%y) + 1 (N.B. [MaxIterations | Unknown] + 1 => Unknown)
199 //   G(%y) = 0 if %y is a loop invariant
200 //   G(%y) = G(%BackEdgeValue) if %y is a phi in the header block
201 //   G(%y) = TODO: if %y is an expression based on phis and loop invariants
202 //           The example looks like:
203 //           %x = phi(0, %a) <-- becomes invariant starting from 3rd iteration.
204 //           %y = phi(0, 5)
205 //           %a = %y + 1
206 //   G(%y) = Unknown otherwise (including phi not in header block)
207 PhiAnalyzer::PeelCounter PhiAnalyzer::calculate(const Value &V) {
208   // If we already know the answer, take it from the map.
209   auto I = IterationsToInvariance.find(&V);
210   if (I != IterationsToInvariance.end())
211     return I->second;
212 
213   // Place Unknown to map to avoid infinite recursion. Such
214   // cycles can never stop on an invariant.
215   IterationsToInvariance[&V] = Unknown;
216 
217   if (L.isLoopInvariant(&V))
218     // Loop invariant so known at start.
219     return (IterationsToInvariance[&V] = 0);
220   if (const PHINode *Phi = dyn_cast<PHINode>(&V)) {
221     if (Phi->getParent() != L.getHeader()) {
222       // Phi is not in header block so Unknown.
223       assert(IterationsToInvariance[&V] == Unknown && "unexpected value saved");
224       return Unknown;
225     }
226     // We need to analyze the input from the back edge and add 1.
227     Value *Input = Phi->getIncomingValueForBlock(L.getLoopLatch());
228     PeelCounter Iterations = calculate(*Input);
229     assert(IterationsToInvariance[Input] == Iterations &&
230            "unexpected value saved");
231     return (IterationsToInvariance[Phi] = addOne(Iterations));
232   }
233   if (const Instruction *I = dyn_cast<Instruction>(&V)) {
234     if (isa<CmpInst>(I) || I->isBinaryOp()) {
235       // Binary instructions get the max of the operands.
236       PeelCounter LHS = calculate(*I->getOperand(0));
237       if (LHS == Unknown)
238         return Unknown;
239       PeelCounter RHS = calculate(*I->getOperand(1));
240       if (RHS == Unknown)
241         return Unknown;
242       return (IterationsToInvariance[I] = {std::max(*LHS, *RHS)});
243     }
244     if (I->isCast())
245       // Cast instructions get the value of the operand.
246       return (IterationsToInvariance[I] = calculate(*I->getOperand(0)));
247   }
248   // TODO: handle more expressions
249 
250   // Everything else is Unknown.
251   assert(IterationsToInvariance[&V] == Unknown && "unexpected value saved");
252   return Unknown;
253 }
254 
255 std::optional<unsigned> PhiAnalyzer::calculateIterationsToPeel() {
256   unsigned Iterations = 0;
257   for (auto &PHI : L.getHeader()->phis()) {
258     PeelCounter ToInvariance = calculate(PHI);
259     if (ToInvariance != Unknown) {
260       assert(*ToInvariance <= MaxIterations && "bad result in phi analysis");
261       Iterations = std::max(Iterations, *ToInvariance);
262       if (Iterations == MaxIterations)
263         break;
264     }
265   }
266   assert((Iterations <= MaxIterations) && "bad result in phi analysis");
267   return Iterations ? std::optional<unsigned>(Iterations) : std::nullopt;
268 }
269 
270 } // unnamed namespace
271 
272 // Try to find any invariant memory reads that will become dereferenceable in
273 // the remainder loop after peeling. The load must also be used (transitively)
274 // by an exit condition. Returns the number of iterations to peel off (at the
275 // moment either 0 or 1).
276 static unsigned peelToTurnInvariantLoadsDerefencebale(Loop &L,
277                                                       DominatorTree &DT,
278                                                       AssumptionCache *AC) {
279   // Skip loops with a single exiting block, because there should be no benefit
280   // for the heuristic below.
281   if (L.getExitingBlock())
282     return 0;
283 
284   // All non-latch exit blocks must have an UnreachableInst terminator.
285   // Otherwise the heuristic below may not be profitable.
286   SmallVector<BasicBlock *, 4> Exits;
287   L.getUniqueNonLatchExitBlocks(Exits);
288   if (any_of(Exits, [](const BasicBlock *BB) {
289         return !isa<UnreachableInst>(BB->getTerminator());
290       }))
291     return 0;
292 
293   // Now look for invariant loads that dominate the latch and are not known to
294   // be dereferenceable. If there are such loads and no writes, they will become
295   // dereferenceable in the loop if the first iteration is peeled off. Also
296   // collect the set of instructions controlled by such loads. Only peel if an
297   // exit condition uses (transitively) such a load.
298   BasicBlock *Header = L.getHeader();
299   BasicBlock *Latch = L.getLoopLatch();
300   SmallPtrSet<Value *, 8> LoadUsers;
301   const DataLayout &DL = L.getHeader()->getModule()->getDataLayout();
302   for (BasicBlock *BB : L.blocks()) {
303     for (Instruction &I : *BB) {
304       if (I.mayWriteToMemory())
305         return 0;
306 
307       auto Iter = LoadUsers.find(&I);
308       if (Iter != LoadUsers.end()) {
309         for (Value *U : I.users())
310           LoadUsers.insert(U);
311       }
312       // Do not look for reads in the header; they can already be hoisted
313       // without peeling.
314       if (BB == Header)
315         continue;
316       if (auto *LI = dyn_cast<LoadInst>(&I)) {
317         Value *Ptr = LI->getPointerOperand();
318         if (DT.dominates(BB, Latch) && L.isLoopInvariant(Ptr) &&
319             !isDereferenceablePointer(Ptr, LI->getType(), DL, LI, AC, &DT))
320           for (Value *U : I.users())
321             LoadUsers.insert(U);
322       }
323     }
324   }
325   SmallVector<BasicBlock *> ExitingBlocks;
326   L.getExitingBlocks(ExitingBlocks);
327   if (any_of(ExitingBlocks, [&LoadUsers](BasicBlock *Exiting) {
328         return LoadUsers.contains(Exiting->getTerminator());
329       }))
330     return 1;
331   return 0;
332 }
333 
334 // Return the number of iterations to peel off that make conditions in the
335 // body true/false. For example, if we peel 2 iterations off the loop below,
336 // the condition i < 2 can be evaluated at compile time.
337 //  for (i = 0; i < n; i++)
338 //    if (i < 2)
339 //      ..
340 //    else
341 //      ..
342 //   }
343 static unsigned countToEliminateCompares(Loop &L, unsigned MaxPeelCount,
344                                          ScalarEvolution &SE) {
345   assert(L.isLoopSimplifyForm() && "Loop needs to be in loop simplify form");
346   unsigned DesiredPeelCount = 0;
347 
348   for (auto *BB : L.blocks()) {
349     auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
350     if (!BI || BI->isUnconditional())
351       continue;
352 
353     // Ignore loop exit condition.
354     if (L.getLoopLatch() == BB)
355       continue;
356 
357     Value *Condition = BI->getCondition();
358     Value *LeftVal, *RightVal;
359     CmpInst::Predicate Pred;
360     if (!match(Condition, m_ICmp(Pred, m_Value(LeftVal), m_Value(RightVal))))
361       continue;
362 
363     const SCEV *LeftSCEV = SE.getSCEV(LeftVal);
364     const SCEV *RightSCEV = SE.getSCEV(RightVal);
365 
366     // Do not consider predicates that are known to be true or false
367     // independently of the loop iteration.
368     if (SE.evaluatePredicate(Pred, LeftSCEV, RightSCEV))
369       continue;
370 
371     // Check if we have a condition with one AddRec and one non AddRec
372     // expression. Normalize LeftSCEV to be the AddRec.
373     if (!isa<SCEVAddRecExpr>(LeftSCEV)) {
374       if (isa<SCEVAddRecExpr>(RightSCEV)) {
375         std::swap(LeftSCEV, RightSCEV);
376         Pred = ICmpInst::getSwappedPredicate(Pred);
377       } else
378         continue;
379     }
380 
381     const SCEVAddRecExpr *LeftAR = cast<SCEVAddRecExpr>(LeftSCEV);
382 
383     // Avoid huge SCEV computations in the loop below, make sure we only
384     // consider AddRecs of the loop we are trying to peel.
385     if (!LeftAR->isAffine() || LeftAR->getLoop() != &L)
386       continue;
387     if (!(ICmpInst::isEquality(Pred) && LeftAR->hasNoSelfWrap()) &&
388         !SE.getMonotonicPredicateType(LeftAR, Pred))
389       continue;
390 
391     // Check if extending the current DesiredPeelCount lets us evaluate Pred
392     // or !Pred in the loop body statically.
393     unsigned NewPeelCount = DesiredPeelCount;
394 
395     const SCEV *IterVal = LeftAR->evaluateAtIteration(
396         SE.getConstant(LeftSCEV->getType(), NewPeelCount), SE);
397 
398     // If the original condition is not known, get the negated predicate
399     // (which holds on the else branch) and check if it is known. This allows
400     // us to peel of iterations that make the original condition false.
401     if (!SE.isKnownPredicate(Pred, IterVal, RightSCEV))
402       Pred = ICmpInst::getInversePredicate(Pred);
403 
404     const SCEV *Step = LeftAR->getStepRecurrence(SE);
405     const SCEV *NextIterVal = SE.getAddExpr(IterVal, Step);
406     auto PeelOneMoreIteration = [&IterVal, &NextIterVal, &SE, Step,
407                                  &NewPeelCount]() {
408       IterVal = NextIterVal;
409       NextIterVal = SE.getAddExpr(IterVal, Step);
410       NewPeelCount++;
411     };
412 
413     auto CanPeelOneMoreIteration = [&NewPeelCount, &MaxPeelCount]() {
414       return NewPeelCount < MaxPeelCount;
415     };
416 
417     while (CanPeelOneMoreIteration() &&
418            SE.isKnownPredicate(Pred, IterVal, RightSCEV))
419       PeelOneMoreIteration();
420 
421     // With *that* peel count, does the predicate !Pred become known in the
422     // first iteration of the loop body after peeling?
423     if (!SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), IterVal,
424                              RightSCEV))
425       continue; // If not, give up.
426 
427     // However, for equality comparisons, that isn't always sufficient to
428     // eliminate the comparsion in loop body, we may need to peel one more
429     // iteration. See if that makes !Pred become unknown again.
430     if (ICmpInst::isEquality(Pred) &&
431         !SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), NextIterVal,
432                              RightSCEV) &&
433         !SE.isKnownPredicate(Pred, IterVal, RightSCEV) &&
434         SE.isKnownPredicate(Pred, NextIterVal, RightSCEV)) {
435       if (!CanPeelOneMoreIteration())
436         continue; // Need to peel one more iteration, but can't. Give up.
437       PeelOneMoreIteration(); // Great!
438     }
439 
440     DesiredPeelCount = std::max(DesiredPeelCount, NewPeelCount);
441   }
442 
443   return DesiredPeelCount;
444 }
445 
446 /// This "heuristic" exactly matches implicit behavior which used to exist
447 /// inside getLoopEstimatedTripCount.  It was added here to keep an
448 /// improvement inside that API from causing peeling to become more aggressive.
449 /// This should probably be removed.
450 static bool violatesLegacyMultiExitLoopCheck(Loop *L) {
451   BasicBlock *Latch = L->getLoopLatch();
452   if (!Latch)
453     return true;
454 
455   BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator());
456   if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch))
457     return true;
458 
459   assert((LatchBR->getSuccessor(0) == L->getHeader() ||
460           LatchBR->getSuccessor(1) == L->getHeader()) &&
461          "At least one edge out of the latch must go to the header");
462 
463   SmallVector<BasicBlock *, 4> ExitBlocks;
464   L->getUniqueNonLatchExitBlocks(ExitBlocks);
465   return any_of(ExitBlocks, [](const BasicBlock *EB) {
466       return !EB->getTerminatingDeoptimizeCall();
467     });
468 }
469 
470 
471 // Return the number of iterations we want to peel off.
472 void llvm::computePeelCount(Loop *L, unsigned LoopSize,
473                             TargetTransformInfo::PeelingPreferences &PP,
474                             unsigned TripCount, DominatorTree &DT,
475                             ScalarEvolution &SE, AssumptionCache *AC,
476                             unsigned Threshold) {
477   assert(LoopSize > 0 && "Zero loop size is not allowed!");
478   // Save the PP.PeelCount value set by the target in
479   // TTI.getPeelingPreferences or by the flag -unroll-peel-count.
480   unsigned TargetPeelCount = PP.PeelCount;
481   PP.PeelCount = 0;
482   if (!canPeel(L))
483     return;
484 
485   // Only try to peel innermost loops by default.
486   // The constraint can be relaxed by the target in TTI.getPeelingPreferences
487   // or by the flag -unroll-allow-loop-nests-peeling.
488   if (!PP.AllowLoopNestsPeeling && !L->isInnermost())
489     return;
490 
491   // If the user provided a peel count, use that.
492   bool UserPeelCount = UnrollForcePeelCount.getNumOccurrences() > 0;
493   if (UserPeelCount) {
494     LLVM_DEBUG(dbgs() << "Force-peeling first " << UnrollForcePeelCount
495                       << " iterations.\n");
496     PP.PeelCount = UnrollForcePeelCount;
497     PP.PeelProfiledIterations = true;
498     return;
499   }
500 
501   // Skip peeling if it's disabled.
502   if (!PP.AllowPeeling)
503     return;
504 
505   // Check that we can peel at least one iteration.
506   if (2 * LoopSize > Threshold)
507     return;
508 
509   unsigned AlreadyPeeled = 0;
510   if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData))
511     AlreadyPeeled = *Peeled;
512   // Stop if we already peeled off the maximum number of iterations.
513   if (AlreadyPeeled >= UnrollPeelMaxCount)
514     return;
515 
516   // Pay respect to limitations implied by loop size and the max peel count.
517   unsigned MaxPeelCount = UnrollPeelMaxCount;
518   MaxPeelCount = std::min(MaxPeelCount, Threshold / LoopSize - 1);
519 
520   // Start the max computation with the PP.PeelCount value set by the target
521   // in TTI.getPeelingPreferences or by the flag -unroll-peel-count.
522   unsigned DesiredPeelCount = TargetPeelCount;
523 
524   // Here we try to get rid of Phis which become invariants after 1, 2, ..., N
525   // iterations of the loop. For this we compute the number for iterations after
526   // which every Phi is guaranteed to become an invariant, and try to peel the
527   // maximum number of iterations among these values, thus turning all those
528   // Phis into invariants.
529   if (MaxPeelCount > DesiredPeelCount) {
530     // Check how many iterations are useful for resolving Phis
531     auto NumPeels = PhiAnalyzer(*L, MaxPeelCount).calculateIterationsToPeel();
532     if (NumPeels)
533       DesiredPeelCount = std::max(DesiredPeelCount, *NumPeels);
534   }
535 
536   DesiredPeelCount = std::max(DesiredPeelCount,
537                               countToEliminateCompares(*L, MaxPeelCount, SE));
538 
539   if (DesiredPeelCount == 0)
540     DesiredPeelCount = peelToTurnInvariantLoadsDerefencebale(*L, DT, AC);
541 
542   if (DesiredPeelCount > 0) {
543     DesiredPeelCount = std::min(DesiredPeelCount, MaxPeelCount);
544     // Consider max peel count limitation.
545     assert(DesiredPeelCount > 0 && "Wrong loop size estimation?");
546     if (DesiredPeelCount + AlreadyPeeled <= UnrollPeelMaxCount) {
547       LLVM_DEBUG(dbgs() << "Peel " << DesiredPeelCount
548                         << " iteration(s) to turn"
549                         << " some Phis into invariants.\n");
550       PP.PeelCount = DesiredPeelCount;
551       PP.PeelProfiledIterations = false;
552       return;
553     }
554   }
555 
556   // Bail if we know the statically calculated trip count.
557   // In this case we rather prefer partial unrolling.
558   if (TripCount)
559     return;
560 
561   // Do not apply profile base peeling if it is disabled.
562   if (!PP.PeelProfiledIterations)
563     return;
564   // If we don't know the trip count, but have reason to believe the average
565   // trip count is low, peeling should be beneficial, since we will usually
566   // hit the peeled section.
567   // We only do this in the presence of profile information, since otherwise
568   // our estimates of the trip count are not reliable enough.
569   if (L->getHeader()->getParent()->hasProfileData()) {
570     if (violatesLegacyMultiExitLoopCheck(L))
571       return;
572     std::optional<unsigned> EstimatedTripCount = getLoopEstimatedTripCount(L);
573     if (!EstimatedTripCount)
574       return;
575 
576     LLVM_DEBUG(dbgs() << "Profile-based estimated trip count is "
577                       << *EstimatedTripCount << "\n");
578 
579     if (*EstimatedTripCount) {
580       if (*EstimatedTripCount + AlreadyPeeled <= MaxPeelCount) {
581         unsigned PeelCount = *EstimatedTripCount;
582         LLVM_DEBUG(dbgs() << "Peeling first " << PeelCount << " iterations.\n");
583         PP.PeelCount = PeelCount;
584         return;
585       }
586       LLVM_DEBUG(dbgs() << "Already peel count: " << AlreadyPeeled << "\n");
587       LLVM_DEBUG(dbgs() << "Max peel count: " << UnrollPeelMaxCount << "\n");
588       LLVM_DEBUG(dbgs() << "Loop cost: " << LoopSize << "\n");
589       LLVM_DEBUG(dbgs() << "Max peel cost: " << Threshold << "\n");
590       LLVM_DEBUG(dbgs() << "Max peel count by cost: "
591                         << (Threshold / LoopSize - 1) << "\n");
592     }
593   }
594 }
595 
596 struct WeightInfo {
597   // Weights for current iteration.
598   SmallVector<uint32_t> Weights;
599   // Weights to subtract after each iteration.
600   const SmallVector<uint32_t> SubWeights;
601 };
602 
603 /// Update the branch weights of an exiting block of a peeled-off loop
604 /// iteration.
605 /// Let F is a weight of the edge to continue (fallthrough) into the loop.
606 /// Let E is a weight of the edge to an exit.
607 /// F/(F+E) is a probability to go to loop and E/(F+E) is a probability to
608 /// go to exit.
609 /// Then, Estimated ExitCount = F / E.
610 /// For I-th (counting from 0) peeled off iteration we set the the weights for
611 /// the peeled exit as (EC - I, 1). It gives us reasonable distribution,
612 /// The probability to go to exit 1/(EC-I) increases. At the same time
613 /// the estimated exit count in the remainder loop reduces by I.
614 /// To avoid dealing with division rounding we can just multiple both part
615 /// of weights to E and use weight as (F - I * E, E).
616 static void updateBranchWeights(Instruction *Term, WeightInfo &Info) {
617   MDBuilder MDB(Term->getContext());
618   Term->setMetadata(LLVMContext::MD_prof,
619                     MDB.createBranchWeights(Info.Weights));
620   for (auto [Idx, SubWeight] : enumerate(Info.SubWeights))
621     if (SubWeight != 0)
622       Info.Weights[Idx] = Info.Weights[Idx] > SubWeight
623                               ? Info.Weights[Idx] - SubWeight
624                               : 1;
625 }
626 
627 /// Initialize the weights for all exiting blocks.
628 static void initBranchWeights(DenseMap<Instruction *, WeightInfo> &WeightInfos,
629                               Loop *L) {
630   SmallVector<BasicBlock *> ExitingBlocks;
631   L->getExitingBlocks(ExitingBlocks);
632   for (BasicBlock *ExitingBlock : ExitingBlocks) {
633     Instruction *Term = ExitingBlock->getTerminator();
634     SmallVector<uint32_t> Weights;
635     if (!extractBranchWeights(*Term, Weights))
636       continue;
637 
638     // See the comment on updateBranchWeights() for an explanation of what we
639     // do here.
640     uint32_t FallThroughWeights = 0;
641     uint32_t ExitWeights = 0;
642     for (auto [Succ, Weight] : zip(successors(Term), Weights)) {
643       if (L->contains(Succ))
644         FallThroughWeights += Weight;
645       else
646         ExitWeights += Weight;
647     }
648 
649     // Don't try to update weights for degenerate case.
650     if (FallThroughWeights == 0)
651       continue;
652 
653     SmallVector<uint32_t> SubWeights;
654     for (auto [Succ, Weight] : zip(successors(Term), Weights)) {
655       if (!L->contains(Succ)) {
656         // Exit weights stay the same.
657         SubWeights.push_back(0);
658         continue;
659       }
660 
661       // Subtract exit weights on each iteration, distributed across all
662       // fallthrough edges.
663       double W = (double)Weight / (double)FallThroughWeights;
664       SubWeights.push_back((uint32_t)(ExitWeights * W));
665     }
666 
667     WeightInfos.insert({Term, {std::move(Weights), std::move(SubWeights)}});
668   }
669 }
670 
671 /// Update the weights of original exiting block after peeling off all
672 /// iterations.
673 static void fixupBranchWeights(Instruction *Term, const WeightInfo &Info) {
674   MDBuilder MDB(Term->getContext());
675   Term->setMetadata(LLVMContext::MD_prof,
676                     MDB.createBranchWeights(Info.Weights));
677 }
678 
679 /// Clones the body of the loop L, putting it between \p InsertTop and \p
680 /// InsertBot.
681 /// \param IterNumber The serial number of the iteration currently being
682 /// peeled off.
683 /// \param ExitEdges The exit edges of the original loop.
684 /// \param[out] NewBlocks A list of the blocks in the newly created clone
685 /// \param[out] VMap The value map between the loop and the new clone.
686 /// \param LoopBlocks A helper for DFS-traversal of the loop.
687 /// \param LVMap A value-map that maps instructions from the original loop to
688 /// instructions in the last peeled-off iteration.
689 static void cloneLoopBlocks(
690     Loop *L, unsigned IterNumber, BasicBlock *InsertTop, BasicBlock *InsertBot,
691     SmallVectorImpl<std::pair<BasicBlock *, BasicBlock *>> &ExitEdges,
692     SmallVectorImpl<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks,
693     ValueToValueMapTy &VMap, ValueToValueMapTy &LVMap, DominatorTree *DT,
694     LoopInfo *LI, ArrayRef<MDNode *> LoopLocalNoAliasDeclScopes,
695     ScalarEvolution &SE) {
696   BasicBlock *Header = L->getHeader();
697   BasicBlock *Latch = L->getLoopLatch();
698   BasicBlock *PreHeader = L->getLoopPreheader();
699 
700   Function *F = Header->getParent();
701   LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
702   LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
703   Loop *ParentLoop = L->getParentLoop();
704 
705   // For each block in the original loop, create a new copy,
706   // and update the value map with the newly created values.
707   for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
708     BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, ".peel", F);
709     NewBlocks.push_back(NewBB);
710 
711     // If an original block is an immediate child of the loop L, its copy
712     // is a child of a ParentLoop after peeling. If a block is a child of
713     // a nested loop, it is handled in the cloneLoop() call below.
714     if (ParentLoop && LI->getLoopFor(*BB) == L)
715       ParentLoop->addBasicBlockToLoop(NewBB, *LI);
716 
717     VMap[*BB] = NewBB;
718 
719     // If dominator tree is available, insert nodes to represent cloned blocks.
720     if (DT) {
721       if (Header == *BB)
722         DT->addNewBlock(NewBB, InsertTop);
723       else {
724         DomTreeNode *IDom = DT->getNode(*BB)->getIDom();
725         // VMap must contain entry for IDom, as the iteration order is RPO.
726         DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDom->getBlock()]));
727       }
728     }
729   }
730 
731   {
732     // Identify what other metadata depends on the cloned version. After
733     // cloning, replace the metadata with the corrected version for both
734     // memory instructions and noalias intrinsics.
735     std::string Ext = (Twine("Peel") + Twine(IterNumber)).str();
736     cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks,
737                                Header->getContext(), Ext);
738   }
739 
740   // Recursively create the new Loop objects for nested loops, if any,
741   // to preserve LoopInfo.
742   for (Loop *ChildLoop : *L) {
743     cloneLoop(ChildLoop, ParentLoop, VMap, LI, nullptr);
744   }
745 
746   // Hook-up the control flow for the newly inserted blocks.
747   // The new header is hooked up directly to the "top", which is either
748   // the original loop preheader (for the first iteration) or the previous
749   // iteration's exiting block (for every other iteration)
750   InsertTop->getTerminator()->setSuccessor(0, cast<BasicBlock>(VMap[Header]));
751 
752   // Similarly, for the latch:
753   // The original exiting edge is still hooked up to the loop exit.
754   // The backedge now goes to the "bottom", which is either the loop's real
755   // header (for the last peeled iteration) or the copied header of the next
756   // iteration (for every other iteration)
757   BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]);
758   auto *LatchTerm = cast<Instruction>(NewLatch->getTerminator());
759   for (unsigned idx = 0, e = LatchTerm->getNumSuccessors(); idx < e; ++idx)
760     if (LatchTerm->getSuccessor(idx) == Header) {
761       LatchTerm->setSuccessor(idx, InsertBot);
762       break;
763     }
764   if (DT)
765     DT->changeImmediateDominator(InsertBot, NewLatch);
766 
767   // The new copy of the loop body starts with a bunch of PHI nodes
768   // that pick an incoming value from either the preheader, or the previous
769   // loop iteration. Since this copy is no longer part of the loop, we
770   // resolve this statically:
771   // For the first iteration, we use the value from the preheader directly.
772   // For any other iteration, we replace the phi with the value generated by
773   // the immediately preceding clone of the loop body (which represents
774   // the previous iteration).
775   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
776     PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
777     if (IterNumber == 0) {
778       VMap[&*I] = NewPHI->getIncomingValueForBlock(PreHeader);
779     } else {
780       Value *LatchVal = NewPHI->getIncomingValueForBlock(Latch);
781       Instruction *LatchInst = dyn_cast<Instruction>(LatchVal);
782       if (LatchInst && L->contains(LatchInst))
783         VMap[&*I] = LVMap[LatchInst];
784       else
785         VMap[&*I] = LatchVal;
786     }
787     NewPHI->eraseFromParent();
788   }
789 
790   // Fix up the outgoing values - we need to add a value for the iteration
791   // we've just created. Note that this must happen *after* the incoming
792   // values are adjusted, since the value going out of the latch may also be
793   // a value coming into the header.
794   for (auto Edge : ExitEdges)
795     for (PHINode &PHI : Edge.second->phis()) {
796       Value *LatchVal = PHI.getIncomingValueForBlock(Edge.first);
797       Instruction *LatchInst = dyn_cast<Instruction>(LatchVal);
798       if (LatchInst && L->contains(LatchInst))
799         LatchVal = VMap[LatchVal];
800       PHI.addIncoming(LatchVal, cast<BasicBlock>(VMap[Edge.first]));
801       SE.forgetValue(&PHI);
802     }
803 
804   // LastValueMap is updated with the values for the current loop
805   // which are used the next time this function is called.
806   for (auto KV : VMap)
807     LVMap[KV.first] = KV.second;
808 }
809 
810 TargetTransformInfo::PeelingPreferences
811 llvm::gatherPeelingPreferences(Loop *L, ScalarEvolution &SE,
812                                const TargetTransformInfo &TTI,
813                                std::optional<bool> UserAllowPeeling,
814                                std::optional<bool> UserAllowProfileBasedPeeling,
815                                bool UnrollingSpecficValues) {
816   TargetTransformInfo::PeelingPreferences PP;
817 
818   // Set the default values.
819   PP.PeelCount = 0;
820   PP.AllowPeeling = true;
821   PP.AllowLoopNestsPeeling = false;
822   PP.PeelProfiledIterations = true;
823 
824   // Get the target specifc values.
825   TTI.getPeelingPreferences(L, SE, PP);
826 
827   // User specified values using cl::opt.
828   if (UnrollingSpecficValues) {
829     if (UnrollPeelCount.getNumOccurrences() > 0)
830       PP.PeelCount = UnrollPeelCount;
831     if (UnrollAllowPeeling.getNumOccurrences() > 0)
832       PP.AllowPeeling = UnrollAllowPeeling;
833     if (UnrollAllowLoopNestsPeeling.getNumOccurrences() > 0)
834       PP.AllowLoopNestsPeeling = UnrollAllowLoopNestsPeeling;
835   }
836 
837   // User specifed values provided by argument.
838   if (UserAllowPeeling)
839     PP.AllowPeeling = *UserAllowPeeling;
840   if (UserAllowProfileBasedPeeling)
841     PP.PeelProfiledIterations = *UserAllowProfileBasedPeeling;
842 
843   return PP;
844 }
845 
846 /// Peel off the first \p PeelCount iterations of loop \p L.
847 ///
848 /// Note that this does not peel them off as a single straight-line block.
849 /// Rather, each iteration is peeled off separately, and needs to check the
850 /// exit condition.
851 /// For loops that dynamically execute \p PeelCount iterations or less
852 /// this provides a benefit, since the peeled off iterations, which account
853 /// for the bulk of dynamic execution, can be further simplified by scalar
854 /// optimizations.
855 bool llvm::peelLoop(Loop *L, unsigned PeelCount, LoopInfo *LI,
856                     ScalarEvolution *SE, DominatorTree &DT, AssumptionCache *AC,
857                     bool PreserveLCSSA, ValueToValueMapTy &LVMap) {
858   assert(PeelCount > 0 && "Attempt to peel out zero iterations?");
859   assert(canPeel(L) && "Attempt to peel a loop which is not peelable?");
860 
861   LoopBlocksDFS LoopBlocks(L);
862   LoopBlocks.perform(LI);
863 
864   BasicBlock *Header = L->getHeader();
865   BasicBlock *PreHeader = L->getLoopPreheader();
866   BasicBlock *Latch = L->getLoopLatch();
867   SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> ExitEdges;
868   L->getExitEdges(ExitEdges);
869 
870   // Remember dominators of blocks we might reach through exits to change them
871   // later. Immediate dominator of such block might change, because we add more
872   // routes which can lead to the exit: we can reach it from the peeled
873   // iterations too.
874   DenseMap<BasicBlock *, BasicBlock *> NonLoopBlocksIDom;
875   for (auto *BB : L->blocks()) {
876     auto *BBDomNode = DT.getNode(BB);
877     SmallVector<BasicBlock *, 16> ChildrenToUpdate;
878     for (auto *ChildDomNode : BBDomNode->children()) {
879       auto *ChildBB = ChildDomNode->getBlock();
880       if (!L->contains(ChildBB))
881         ChildrenToUpdate.push_back(ChildBB);
882     }
883     // The new idom of the block will be the nearest common dominator
884     // of all copies of the previous idom. This is equivalent to the
885     // nearest common dominator of the previous idom and the first latch,
886     // which dominates all copies of the previous idom.
887     BasicBlock *NewIDom = DT.findNearestCommonDominator(BB, Latch);
888     for (auto *ChildBB : ChildrenToUpdate)
889       NonLoopBlocksIDom[ChildBB] = NewIDom;
890   }
891 
892   Function *F = Header->getParent();
893 
894   // Set up all the necessary basic blocks. It is convenient to split the
895   // preheader into 3 parts - two blocks to anchor the peeled copy of the loop
896   // body, and a new preheader for the "real" loop.
897 
898   // Peeling the first iteration transforms.
899   //
900   // PreHeader:
901   // ...
902   // Header:
903   //   LoopBody
904   //   If (cond) goto Header
905   // Exit:
906   //
907   // into
908   //
909   // InsertTop:
910   //   LoopBody
911   //   If (!cond) goto Exit
912   // InsertBot:
913   // NewPreHeader:
914   // ...
915   // Header:
916   //  LoopBody
917   //  If (cond) goto Header
918   // Exit:
919   //
920   // Each following iteration will split the current bottom anchor in two,
921   // and put the new copy of the loop body between these two blocks. That is,
922   // after peeling another iteration from the example above, we'll split
923   // InsertBot, and get:
924   //
925   // InsertTop:
926   //   LoopBody
927   //   If (!cond) goto Exit
928   // InsertBot:
929   //   LoopBody
930   //   If (!cond) goto Exit
931   // InsertBot.next:
932   // NewPreHeader:
933   // ...
934   // Header:
935   //  LoopBody
936   //  If (cond) goto Header
937   // Exit:
938 
939   BasicBlock *InsertTop = SplitEdge(PreHeader, Header, &DT, LI);
940   BasicBlock *InsertBot =
941       SplitBlock(InsertTop, InsertTop->getTerminator(), &DT, LI);
942   BasicBlock *NewPreHeader =
943       SplitBlock(InsertBot, InsertBot->getTerminator(), &DT, LI);
944 
945   InsertTop->setName(Header->getName() + ".peel.begin");
946   InsertBot->setName(Header->getName() + ".peel.next");
947   NewPreHeader->setName(PreHeader->getName() + ".peel.newph");
948 
949   Instruction *LatchTerm =
950       cast<Instruction>(cast<BasicBlock>(Latch)->getTerminator());
951 
952   // If we have branch weight information, we'll want to update it for the
953   // newly created branches.
954   DenseMap<Instruction *, WeightInfo> Weights;
955   initBranchWeights(Weights, L);
956 
957   // Identify what noalias metadata is inside the loop: if it is inside the
958   // loop, the associated metadata must be cloned for each iteration.
959   SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes;
960   identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes);
961 
962   // For each peeled-off iteration, make a copy of the loop.
963   for (unsigned Iter = 0; Iter < PeelCount; ++Iter) {
964     SmallVector<BasicBlock *, 8> NewBlocks;
965     ValueToValueMapTy VMap;
966 
967     cloneLoopBlocks(L, Iter, InsertTop, InsertBot, ExitEdges, NewBlocks,
968                     LoopBlocks, VMap, LVMap, &DT, LI,
969                     LoopLocalNoAliasDeclScopes, *SE);
970 
971     // Remap to use values from the current iteration instead of the
972     // previous one.
973     remapInstructionsInBlocks(NewBlocks, VMap);
974 
975     // Update IDoms of the blocks reachable through exits.
976     if (Iter == 0)
977       for (auto BBIDom : NonLoopBlocksIDom)
978         DT.changeImmediateDominator(BBIDom.first,
979                                      cast<BasicBlock>(LVMap[BBIDom.second]));
980 #ifdef EXPENSIVE_CHECKS
981     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
982 #endif
983 
984     for (auto &[Term, Info] : Weights) {
985       auto *TermCopy = cast<Instruction>(VMap[Term]);
986       updateBranchWeights(TermCopy, Info);
987     }
988 
989     // Remove Loop metadata from the latch branch instruction
990     // because it is not the Loop's latch branch anymore.
991     auto *LatchTermCopy = cast<Instruction>(VMap[LatchTerm]);
992     LatchTermCopy->setMetadata(LLVMContext::MD_loop, nullptr);
993 
994     InsertTop = InsertBot;
995     InsertBot = SplitBlock(InsertBot, InsertBot->getTerminator(), &DT, LI);
996     InsertBot->setName(Header->getName() + ".peel.next");
997 
998     F->splice(InsertTop->getIterator(), F, NewBlocks[0]->getIterator(),
999               F->end());
1000   }
1001 
1002   // Now adjust the phi nodes in the loop header to get their initial values
1003   // from the last peeled-off iteration instead of the preheader.
1004   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
1005     PHINode *PHI = cast<PHINode>(I);
1006     Value *NewVal = PHI->getIncomingValueForBlock(Latch);
1007     Instruction *LatchInst = dyn_cast<Instruction>(NewVal);
1008     if (LatchInst && L->contains(LatchInst))
1009       NewVal = LVMap[LatchInst];
1010 
1011     PHI->setIncomingValueForBlock(NewPreHeader, NewVal);
1012   }
1013 
1014   for (const auto &[Term, Info] : Weights)
1015     fixupBranchWeights(Term, Info);
1016 
1017   // Update Metadata for count of peeled off iterations.
1018   unsigned AlreadyPeeled = 0;
1019   if (auto Peeled = getOptionalIntLoopAttribute(L, PeeledCountMetaData))
1020     AlreadyPeeled = *Peeled;
1021   addStringMetadataToLoop(L, PeeledCountMetaData, AlreadyPeeled + PeelCount);
1022 
1023   if (Loop *ParentLoop = L->getParentLoop())
1024     L = ParentLoop;
1025 
1026   // We modified the loop, update SE.
1027   SE->forgetTopmostLoop(L);
1028 
1029 #ifdef EXPENSIVE_CHECKS
1030   // Finally DomtTree must be correct.
1031   assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1032 #endif
1033 
1034   // FIXME: Incrementally update loop-simplify
1035   simplifyLoop(L, &DT, LI, SE, AC, nullptr, PreserveLCSSA);
1036 
1037   NumPeeled++;
1038 
1039   return true;
1040 }
1041