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