1 | //===- LoopInfo.cpp - Natural Loop Calculator -----------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the LoopInfo class that is used to identify natural loops |
10 | // and determine the loop depth of various nodes of the CFG. Note that the |
11 | // loops identified may actually be several natural loops that share the same |
12 | // header node... not just a single natural loop. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "llvm/Analysis/LoopInfo.h" |
17 | #include "llvm/ADT/ScopeExit.h" |
18 | #include "llvm/ADT/SmallPtrSet.h" |
19 | #include "llvm/Analysis/IVDescriptors.h" |
20 | #include "llvm/Analysis/LoopIterator.h" |
21 | #include "llvm/Analysis/LoopNestAnalysis.h" |
22 | #include "llvm/Analysis/MemorySSA.h" |
23 | #include "llvm/Analysis/MemorySSAUpdater.h" |
24 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
25 | #include "llvm/Analysis/ValueTracking.h" |
26 | #include "llvm/Config/llvm-config.h" |
27 | #include "llvm/IR/CFG.h" |
28 | #include "llvm/IR/Constants.h" |
29 | #include "llvm/IR/DebugLoc.h" |
30 | #include "llvm/IR/Dominators.h" |
31 | #include "llvm/IR/Instructions.h" |
32 | #include "llvm/IR/LLVMContext.h" |
33 | #include "llvm/IR/Metadata.h" |
34 | #include "llvm/IR/Module.h" |
35 | #include "llvm/IR/PassManager.h" |
36 | #include "llvm/IR/PrintPasses.h" |
37 | #include "llvm/InitializePasses.h" |
38 | #include "llvm/Support/CommandLine.h" |
39 | #include "llvm/Support/Compiler.h" |
40 | #include "llvm/Support/GenericLoopInfoImpl.h" |
41 | #include "llvm/Support/raw_ostream.h" |
42 | using namespace llvm; |
43 | |
44 | // Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops. |
45 | template class LLVM_EXPORT_TEMPLATE llvm::LoopBase<BasicBlock, Loop>; |
46 | template class LLVM_EXPORT_TEMPLATE llvm::LoopInfoBase<BasicBlock, Loop>; |
47 | |
48 | // Always verify loopinfo if expensive checking is enabled. |
49 | #ifdef EXPENSIVE_CHECKS |
50 | bool llvm::VerifyLoopInfo = true; |
51 | #else |
52 | bool llvm::VerifyLoopInfo = false; |
53 | #endif |
54 | static cl::opt<bool, true> |
55 | VerifyLoopInfoX("verify-loop-info" , cl::location(L&: VerifyLoopInfo), |
56 | cl::Hidden, cl::desc("Verify loop info (time consuming)" )); |
57 | |
58 | //===----------------------------------------------------------------------===// |
59 | // Loop implementation |
60 | // |
61 | |
62 | bool Loop::isLoopInvariant(const Value *V) const { |
63 | if (const Instruction *I = dyn_cast<Instruction>(Val: V)) |
64 | return !contains(Inst: I); |
65 | return true; // All non-instructions are loop invariant |
66 | } |
67 | |
68 | bool Loop::hasLoopInvariantOperands(const Instruction *I) const { |
69 | return all_of(Range: I->operands(), P: [this](Value *V) { return isLoopInvariant(V); }); |
70 | } |
71 | |
72 | bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt, |
73 | MemorySSAUpdater *MSSAU, |
74 | ScalarEvolution *SE) const { |
75 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
76 | return makeLoopInvariant(I, Changed, InsertPt, MSSAU, SE); |
77 | return true; // All non-instructions are loop-invariant. |
78 | } |
79 | |
80 | bool Loop::makeLoopInvariant(Instruction *I, bool &Changed, |
81 | Instruction *InsertPt, MemorySSAUpdater *MSSAU, |
82 | ScalarEvolution *SE) const { |
83 | // Test if the value is already loop-invariant. |
84 | if (isLoopInvariant(V: I)) |
85 | return true; |
86 | if (!isSafeToSpeculativelyExecute(I)) |
87 | return false; |
88 | if (I->mayReadFromMemory()) |
89 | return false; |
90 | // EH block instructions are immobile. |
91 | if (I->isEHPad()) |
92 | return false; |
93 | // Determine the insertion point, unless one was given. |
94 | if (!InsertPt) { |
95 | BasicBlock * = getLoopPreheader(); |
96 | // Without a preheader, hoisting is not feasible. |
97 | if (!Preheader) |
98 | return false; |
99 | InsertPt = Preheader->getTerminator(); |
100 | } |
101 | // Don't hoist instructions with loop-variant operands. |
102 | for (Value *Operand : I->operands()) |
103 | if (!makeLoopInvariant(V: Operand, Changed, InsertPt, MSSAU, SE)) |
104 | return false; |
105 | |
106 | // Hoist. |
107 | I->moveBefore(InsertPos: InsertPt->getIterator()); |
108 | if (MSSAU) |
109 | if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I)) |
110 | MSSAU->moveToPlace(What: MUD, BB: InsertPt->getParent(), |
111 | Where: MemorySSA::BeforeTerminator); |
112 | |
113 | // There is possibility of hoisting this instruction above some arbitrary |
114 | // condition. Any metadata defined on it can be control dependent on this |
115 | // condition. Conservatively strip it here so that we don't give any wrong |
116 | // information to the optimizer. |
117 | I->dropUnknownNonDebugMetadata(); |
118 | |
119 | if (SE) |
120 | SE->forgetBlockAndLoopDispositions(V: I); |
121 | |
122 | Changed = true; |
123 | return true; |
124 | } |
125 | |
126 | bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming, |
127 | BasicBlock *&Backedge) const { |
128 | BasicBlock *H = getHeader(); |
129 | |
130 | Incoming = nullptr; |
131 | Backedge = nullptr; |
132 | pred_iterator PI = pred_begin(BB: H); |
133 | assert(PI != pred_end(H) && "Loop must have at least one backedge!" ); |
134 | Backedge = *PI++; |
135 | if (PI == pred_end(BB: H)) |
136 | return false; // dead loop |
137 | Incoming = *PI++; |
138 | if (PI != pred_end(BB: H)) |
139 | return false; // multiple backedges? |
140 | |
141 | if (contains(BB: Incoming)) { |
142 | if (contains(BB: Backedge)) |
143 | return false; |
144 | std::swap(a&: Incoming, b&: Backedge); |
145 | } else if (!contains(BB: Backedge)) |
146 | return false; |
147 | |
148 | assert(Incoming && Backedge && "expected non-null incoming and backedges" ); |
149 | return true; |
150 | } |
151 | |
152 | PHINode *Loop::getCanonicalInductionVariable() const { |
153 | BasicBlock *H = getHeader(); |
154 | |
155 | BasicBlock *Incoming = nullptr, *Backedge = nullptr; |
156 | if (!getIncomingAndBackEdge(Incoming, Backedge)) |
157 | return nullptr; |
158 | |
159 | // Loop over all of the PHI nodes, looking for a canonical indvar. |
160 | for (BasicBlock::iterator I = H->begin(); isa<PHINode>(Val: I); ++I) { |
161 | PHINode *PN = cast<PHINode>(Val&: I); |
162 | if (ConstantInt *CI = |
163 | dyn_cast<ConstantInt>(Val: PN->getIncomingValueForBlock(BB: Incoming))) |
164 | if (CI->isZero()) |
165 | if (Instruction *Inc = |
166 | dyn_cast<Instruction>(Val: PN->getIncomingValueForBlock(BB: Backedge))) |
167 | if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(i: 0) == PN) |
168 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Inc->getOperand(i: 1))) |
169 | if (CI->isOne()) |
170 | return PN; |
171 | } |
172 | return nullptr; |
173 | } |
174 | |
175 | /// Get the latch condition instruction. |
176 | ICmpInst *Loop::getLatchCmpInst() const { |
177 | if (BasicBlock *Latch = getLoopLatch()) |
178 | if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Val: Latch->getTerminator())) |
179 | if (BI->isConditional()) |
180 | return dyn_cast<ICmpInst>(Val: BI->getCondition()); |
181 | |
182 | return nullptr; |
183 | } |
184 | |
185 | /// Return the final value of the loop induction variable if found. |
186 | static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar, |
187 | const Instruction &StepInst) { |
188 | ICmpInst *LatchCmpInst = L.getLatchCmpInst(); |
189 | if (!LatchCmpInst) |
190 | return nullptr; |
191 | |
192 | Value *Op0 = LatchCmpInst->getOperand(i_nocapture: 0); |
193 | Value *Op1 = LatchCmpInst->getOperand(i_nocapture: 1); |
194 | if (Op0 == &IndVar || Op0 == &StepInst) |
195 | return Op1; |
196 | |
197 | if (Op1 == &IndVar || Op1 == &StepInst) |
198 | return Op0; |
199 | |
200 | return nullptr; |
201 | } |
202 | |
203 | std::optional<Loop::LoopBounds> |
204 | Loop::LoopBounds::getBounds(const Loop &L, PHINode &IndVar, |
205 | ScalarEvolution &SE) { |
206 | InductionDescriptor IndDesc; |
207 | if (!InductionDescriptor::isInductionPHI(Phi: &IndVar, L: &L, SE: &SE, D&: IndDesc)) |
208 | return std::nullopt; |
209 | |
210 | Value *InitialIVValue = IndDesc.getStartValue(); |
211 | Instruction *StepInst = IndDesc.getInductionBinOp(); |
212 | if (!InitialIVValue || !StepInst) |
213 | return std::nullopt; |
214 | |
215 | const SCEV *Step = IndDesc.getStep(); |
216 | Value *StepInstOp1 = StepInst->getOperand(i: 1); |
217 | Value *StepInstOp0 = StepInst->getOperand(i: 0); |
218 | Value *StepValue = nullptr; |
219 | if (SE.getSCEV(V: StepInstOp1) == Step) |
220 | StepValue = StepInstOp1; |
221 | else if (SE.getSCEV(V: StepInstOp0) == Step) |
222 | StepValue = StepInstOp0; |
223 | |
224 | Value *FinalIVValue = findFinalIVValue(L, IndVar, StepInst: *StepInst); |
225 | if (!FinalIVValue) |
226 | return std::nullopt; |
227 | |
228 | return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue, |
229 | SE); |
230 | } |
231 | |
232 | using Direction = Loop::LoopBounds::Direction; |
233 | |
234 | ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const { |
235 | BasicBlock *Latch = L.getLoopLatch(); |
236 | assert(Latch && "Expecting valid latch" ); |
237 | |
238 | BranchInst *BI = dyn_cast_or_null<BranchInst>(Val: Latch->getTerminator()); |
239 | assert(BI && BI->isConditional() && "Expecting conditional latch branch" ); |
240 | |
241 | ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(Val: BI->getCondition()); |
242 | assert(LatchCmpInst && |
243 | "Expecting the latch compare instruction to be a CmpInst" ); |
244 | |
245 | // Need to inverse the predicate when first successor is not the loop |
246 | // header |
247 | ICmpInst::Predicate Pred = (BI->getSuccessor(i: 0) == L.getHeader()) |
248 | ? LatchCmpInst->getPredicate() |
249 | : LatchCmpInst->getInversePredicate(); |
250 | |
251 | if (LatchCmpInst->getOperand(i_nocapture: 0) == &getFinalIVValue()) |
252 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
253 | |
254 | // Need to flip strictness of the predicate when the latch compare instruction |
255 | // is not using StepInst |
256 | if (LatchCmpInst->getOperand(i_nocapture: 0) == &getStepInst() || |
257 | LatchCmpInst->getOperand(i_nocapture: 1) == &getStepInst()) |
258 | return Pred; |
259 | |
260 | // Cannot flip strictness of NE and EQ |
261 | if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) |
262 | return ICmpInst::getFlippedStrictnessPredicate(pred: Pred); |
263 | |
264 | Direction D = getDirection(); |
265 | if (D == Direction::Increasing) |
266 | return ICmpInst::ICMP_SLT; |
267 | |
268 | if (D == Direction::Decreasing) |
269 | return ICmpInst::ICMP_SGT; |
270 | |
271 | // If cannot determine the direction, then unable to find the canonical |
272 | // predicate |
273 | return ICmpInst::BAD_ICMP_PREDICATE; |
274 | } |
275 | |
276 | Direction Loop::LoopBounds::getDirection() const { |
277 | if (const SCEVAddRecExpr *StepAddRecExpr = |
278 | dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: &getStepInst()))) |
279 | if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) { |
280 | if (SE.isKnownPositive(S: StepRecur)) |
281 | return Direction::Increasing; |
282 | if (SE.isKnownNegative(S: StepRecur)) |
283 | return Direction::Decreasing; |
284 | } |
285 | |
286 | return Direction::Unknown; |
287 | } |
288 | |
289 | std::optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const { |
290 | if (PHINode *IndVar = getInductionVariable(SE)) |
291 | return LoopBounds::getBounds(L: *this, IndVar&: *IndVar, SE); |
292 | |
293 | return std::nullopt; |
294 | } |
295 | |
296 | PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const { |
297 | if (!isLoopSimplifyForm()) |
298 | return nullptr; |
299 | |
300 | BasicBlock * = getHeader(); |
301 | assert(Header && "Expected a valid loop header" ); |
302 | ICmpInst *CmpInst = getLatchCmpInst(); |
303 | if (!CmpInst) |
304 | return nullptr; |
305 | |
306 | Value *LatchCmpOp0 = CmpInst->getOperand(i_nocapture: 0); |
307 | Value *LatchCmpOp1 = CmpInst->getOperand(i_nocapture: 1); |
308 | |
309 | for (PHINode &IndVar : Header->phis()) { |
310 | InductionDescriptor IndDesc; |
311 | if (!InductionDescriptor::isInductionPHI(Phi: &IndVar, L: this, SE: &SE, D&: IndDesc)) |
312 | continue; |
313 | |
314 | BasicBlock *Latch = getLoopLatch(); |
315 | Value *StepInst = IndVar.getIncomingValueForBlock(BB: Latch); |
316 | |
317 | // case 1: |
318 | // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}] |
319 | // StepInst = IndVar + step |
320 | // cmp = StepInst < FinalValue |
321 | if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1) |
322 | return &IndVar; |
323 | |
324 | // case 2: |
325 | // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}] |
326 | // StepInst = IndVar + step |
327 | // cmp = IndVar < FinalValue |
328 | if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1) |
329 | return &IndVar; |
330 | } |
331 | |
332 | return nullptr; |
333 | } |
334 | |
335 | bool Loop::getInductionDescriptor(ScalarEvolution &SE, |
336 | InductionDescriptor &IndDesc) const { |
337 | if (PHINode *IndVar = getInductionVariable(SE)) |
338 | return InductionDescriptor::isInductionPHI(Phi: IndVar, L: this, SE: &SE, D&: IndDesc); |
339 | |
340 | return false; |
341 | } |
342 | |
343 | bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar, |
344 | ScalarEvolution &SE) const { |
345 | // Located in the loop header |
346 | BasicBlock * = getHeader(); |
347 | if (AuxIndVar.getParent() != Header) |
348 | return false; |
349 | |
350 | // No uses outside of the loop |
351 | for (User *U : AuxIndVar.users()) |
352 | if (const Instruction *I = dyn_cast<Instruction>(Val: U)) |
353 | if (!contains(Inst: I)) |
354 | return false; |
355 | |
356 | InductionDescriptor IndDesc; |
357 | if (!InductionDescriptor::isInductionPHI(Phi: &AuxIndVar, L: this, SE: &SE, D&: IndDesc)) |
358 | return false; |
359 | |
360 | // The step instruction opcode should be add or sub. |
361 | if (IndDesc.getInductionOpcode() != Instruction::Add && |
362 | IndDesc.getInductionOpcode() != Instruction::Sub) |
363 | return false; |
364 | |
365 | // Incremented by a loop invariant step for each loop iteration |
366 | return SE.isLoopInvariant(S: IndDesc.getStep(), L: this); |
367 | } |
368 | |
369 | BranchInst *Loop::getLoopGuardBranch() const { |
370 | if (!isLoopSimplifyForm()) |
371 | return nullptr; |
372 | |
373 | BasicBlock * = getLoopPreheader(); |
374 | assert(Preheader && getLoopLatch() && |
375 | "Expecting a loop with valid preheader and latch" ); |
376 | |
377 | // Loop should be in rotate form. |
378 | if (!isRotatedForm()) |
379 | return nullptr; |
380 | |
381 | // Disallow loops with more than one unique exit block, as we do not verify |
382 | // that GuardOtherSucc post dominates all exit blocks. |
383 | BasicBlock *ExitFromLatch = getUniqueExitBlock(); |
384 | if (!ExitFromLatch) |
385 | return nullptr; |
386 | |
387 | BasicBlock *GuardBB = Preheader->getUniquePredecessor(); |
388 | if (!GuardBB) |
389 | return nullptr; |
390 | |
391 | assert(GuardBB->getTerminator() && "Expecting valid guard terminator" ); |
392 | |
393 | BranchInst *GuardBI = dyn_cast<BranchInst>(Val: GuardBB->getTerminator()); |
394 | if (!GuardBI || GuardBI->isUnconditional()) |
395 | return nullptr; |
396 | |
397 | BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(i: 0) == Preheader) |
398 | ? GuardBI->getSuccessor(i: 1) |
399 | : GuardBI->getSuccessor(i: 0); |
400 | |
401 | // Check if ExitFromLatch (or any BasicBlock which is an empty unique |
402 | // successor of ExitFromLatch) is equal to GuardOtherSucc. If |
403 | // skipEmptyBlockUntil returns GuardOtherSucc, then the guard branch for the |
404 | // loop is GuardBI (return GuardBI), otherwise return nullptr. |
405 | if (&LoopNest::skipEmptyBlockUntil(From: ExitFromLatch, End: GuardOtherSucc, |
406 | /*CheckUniquePred=*/true) == |
407 | GuardOtherSucc) |
408 | return GuardBI; |
409 | else |
410 | return nullptr; |
411 | } |
412 | |
413 | bool Loop::isCanonical(ScalarEvolution &SE) const { |
414 | InductionDescriptor IndDesc; |
415 | if (!getInductionDescriptor(SE, IndDesc)) |
416 | return false; |
417 | |
418 | ConstantInt *Init = dyn_cast_or_null<ConstantInt>(Val: IndDesc.getStartValue()); |
419 | if (!Init || !Init->isZero()) |
420 | return false; |
421 | |
422 | if (IndDesc.getInductionOpcode() != Instruction::Add) |
423 | return false; |
424 | |
425 | ConstantInt *Step = IndDesc.getConstIntStepValue(); |
426 | if (!Step || !Step->isOne()) |
427 | return false; |
428 | |
429 | return true; |
430 | } |
431 | |
432 | // Check that 'BB' doesn't have any uses outside of the 'L' |
433 | static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB, |
434 | const DominatorTree &DT, bool IgnoreTokens) { |
435 | for (const Instruction &I : BB) { |
436 | // Tokens can't be used in PHI nodes and live-out tokens prevent loop |
437 | // optimizations, so for the purposes of considered LCSSA form, we |
438 | // can ignore them. |
439 | if (IgnoreTokens && I.getType()->isTokenTy()) |
440 | continue; |
441 | |
442 | for (const Use &U : I.uses()) { |
443 | const Instruction *UI = cast<Instruction>(Val: U.getUser()); |
444 | const BasicBlock *UserBB = UI->getParent(); |
445 | |
446 | // For practical purposes, we consider that the use in a PHI |
447 | // occurs in the respective predecessor block. For more info, |
448 | // see the `phi` doc in LangRef and the LCSSA doc. |
449 | if (const PHINode *P = dyn_cast<PHINode>(Val: UI)) |
450 | UserBB = P->getIncomingBlock(U); |
451 | |
452 | // Check the current block, as a fast-path, before checking whether |
453 | // the use is anywhere in the loop. Most values are used in the same |
454 | // block they are defined in. Also, blocks not reachable from the |
455 | // entry are special; uses in them don't need to go through PHIs. |
456 | if (UserBB != &BB && !L.contains(BB: UserBB) && |
457 | DT.isReachableFromEntry(A: UserBB)) |
458 | return false; |
459 | } |
460 | } |
461 | return true; |
462 | } |
463 | |
464 | bool Loop::isLCSSAForm(const DominatorTree &DT, bool IgnoreTokens) const { |
465 | // For each block we check that it doesn't have any uses outside of this loop. |
466 | return all_of(Range: this->blocks(), P: [&](const BasicBlock *BB) { |
467 | return isBlockInLCSSAForm(L: *this, BB: *BB, DT, IgnoreTokens); |
468 | }); |
469 | } |
470 | |
471 | bool Loop::isRecursivelyLCSSAForm(const DominatorTree &DT, const LoopInfo &LI, |
472 | bool IgnoreTokens) const { |
473 | // For each block we check that it doesn't have any uses outside of its |
474 | // innermost loop. This process will transitively guarantee that the current |
475 | // loop and all of the nested loops are in LCSSA form. |
476 | return all_of(Range: this->blocks(), P: [&](const BasicBlock *BB) { |
477 | return isBlockInLCSSAForm(L: *LI.getLoopFor(BB), BB: *BB, DT, IgnoreTokens); |
478 | }); |
479 | } |
480 | |
481 | bool Loop::isLoopSimplifyForm() const { |
482 | // Normal-form loops have a preheader, a single backedge, and all of their |
483 | // exits have all their predecessors inside the loop. |
484 | return getLoopPreheader() && getLoopLatch() && hasDedicatedExits(); |
485 | } |
486 | |
487 | // Routines that reform the loop CFG and split edges often fail on indirectbr. |
488 | bool Loop::isSafeToClone() const { |
489 | // Return false if any loop blocks contain indirectbrs, or there are any calls |
490 | // to noduplicate functions. |
491 | for (BasicBlock *BB : this->blocks()) { |
492 | if (isa<IndirectBrInst>(Val: BB->getTerminator())) |
493 | return false; |
494 | |
495 | for (Instruction &I : *BB) |
496 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
497 | if (CB->cannotDuplicate()) |
498 | return false; |
499 | } |
500 | return true; |
501 | } |
502 | |
503 | MDNode *Loop::getLoopID() const { |
504 | MDNode *LoopID = nullptr; |
505 | |
506 | // Go through the latch blocks and check the terminator for the metadata. |
507 | SmallVector<BasicBlock *, 4> LatchesBlocks; |
508 | getLoopLatches(LoopLatches&: LatchesBlocks); |
509 | for (BasicBlock *BB : LatchesBlocks) { |
510 | Instruction *TI = BB->getTerminator(); |
511 | MDNode *MD = TI->getMetadata(KindID: LLVMContext::MD_loop); |
512 | |
513 | if (!MD) |
514 | return nullptr; |
515 | |
516 | if (!LoopID) |
517 | LoopID = MD; |
518 | else if (MD != LoopID) |
519 | return nullptr; |
520 | } |
521 | if (!LoopID || LoopID->getNumOperands() == 0 || |
522 | LoopID->getOperand(I: 0) != LoopID) |
523 | return nullptr; |
524 | return LoopID; |
525 | } |
526 | |
527 | void Loop::setLoopID(MDNode *LoopID) const { |
528 | assert((!LoopID || LoopID->getNumOperands() > 0) && |
529 | "Loop ID needs at least one operand" ); |
530 | assert((!LoopID || LoopID->getOperand(0) == LoopID) && |
531 | "Loop ID should refer to itself" ); |
532 | |
533 | SmallVector<BasicBlock *, 4> LoopLatches; |
534 | getLoopLatches(LoopLatches); |
535 | for (BasicBlock *BB : LoopLatches) |
536 | BB->getTerminator()->setMetadata(KindID: LLVMContext::MD_loop, Node: LoopID); |
537 | } |
538 | |
539 | void Loop::setLoopAlreadyUnrolled() { |
540 | LLVMContext &Context = getHeader()->getContext(); |
541 | |
542 | MDNode *DisableUnrollMD = |
543 | MDNode::get(Context, MDs: MDString::get(Context, Str: "llvm.loop.unroll.disable" )); |
544 | MDNode *LoopID = getLoopID(); |
545 | MDNode *NewLoopID = makePostTransformationMetadata( |
546 | Context, OrigLoopID: LoopID, RemovePrefixes: {"llvm.loop.unroll." }, AddAttrs: {DisableUnrollMD}); |
547 | setLoopID(NewLoopID); |
548 | } |
549 | |
550 | void Loop::setLoopMustProgress() { |
551 | LLVMContext &Context = getHeader()->getContext(); |
552 | |
553 | MDNode *MustProgress = findOptionMDForLoop(TheLoop: this, Name: "llvm.loop.mustprogress" ); |
554 | |
555 | if (MustProgress) |
556 | return; |
557 | |
558 | MDNode *MustProgressMD = |
559 | MDNode::get(Context, MDs: MDString::get(Context, Str: "llvm.loop.mustprogress" )); |
560 | MDNode *LoopID = getLoopID(); |
561 | MDNode *NewLoopID = |
562 | makePostTransformationMetadata(Context, OrigLoopID: LoopID, RemovePrefixes: {}, AddAttrs: {MustProgressMD}); |
563 | setLoopID(NewLoopID); |
564 | } |
565 | |
566 | bool Loop::isAnnotatedParallel() const { |
567 | MDNode *DesiredLoopIdMetadata = getLoopID(); |
568 | |
569 | if (!DesiredLoopIdMetadata) |
570 | return false; |
571 | |
572 | MDNode *ParallelAccesses = |
573 | findOptionMDForLoop(TheLoop: this, Name: "llvm.loop.parallel_accesses" ); |
574 | SmallPtrSet<MDNode *, 4> |
575 | ParallelAccessGroups; // For scalable 'contains' check. |
576 | if (ParallelAccesses) { |
577 | for (const MDOperand &MD : drop_begin(RangeOrContainer: ParallelAccesses->operands())) { |
578 | MDNode *AccGroup = cast<MDNode>(Val: MD.get()); |
579 | assert(isValidAsAccessGroup(AccGroup) && |
580 | "List item must be an access group" ); |
581 | ParallelAccessGroups.insert(Ptr: AccGroup); |
582 | } |
583 | } |
584 | |
585 | // The loop branch contains the parallel loop metadata. In order to ensure |
586 | // that any parallel-loop-unaware optimization pass hasn't added loop-carried |
587 | // dependencies (thus converted the loop back to a sequential loop), check |
588 | // that all the memory instructions in the loop belong to an access group that |
589 | // is parallel to this loop. |
590 | for (BasicBlock *BB : this->blocks()) { |
591 | for (Instruction &I : *BB) { |
592 | if (!I.mayReadOrWriteMemory()) |
593 | continue; |
594 | |
595 | if (MDNode *AccessGroup = I.getMetadata(KindID: LLVMContext::MD_access_group)) { |
596 | auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool { |
597 | if (AG->getNumOperands() == 0) { |
598 | assert(isValidAsAccessGroup(AG) && "Item must be an access group" ); |
599 | return ParallelAccessGroups.count(Ptr: AG); |
600 | } |
601 | |
602 | for (const MDOperand &AccessListItem : AG->operands()) { |
603 | MDNode *AccGroup = cast<MDNode>(Val: AccessListItem.get()); |
604 | assert(isValidAsAccessGroup(AccGroup) && |
605 | "List item must be an access group" ); |
606 | if (ParallelAccessGroups.count(Ptr: AccGroup)) |
607 | return true; |
608 | } |
609 | return false; |
610 | }; |
611 | |
612 | if (ContainsAccessGroup(AccessGroup)) |
613 | continue; |
614 | } |
615 | |
616 | // The memory instruction can refer to the loop identifier metadata |
617 | // directly or indirectly through another list metadata (in case of |
618 | // nested parallel loops). The loop identifier metadata refers to |
619 | // itself so we can check both cases with the same routine. |
620 | MDNode *LoopIdMD = |
621 | I.getMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access); |
622 | |
623 | if (!LoopIdMD) |
624 | return false; |
625 | |
626 | if (!llvm::is_contained(Range: LoopIdMD->operands(), Element: DesiredLoopIdMetadata)) |
627 | return false; |
628 | } |
629 | } |
630 | return true; |
631 | } |
632 | |
633 | DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); } |
634 | |
635 | Loop::LocRange Loop::getLocRange() const { |
636 | // If we have a debug location in the loop ID, then use it. |
637 | if (MDNode *LoopID = getLoopID()) { |
638 | DebugLoc Start; |
639 | // We use the first DebugLoc in the header as the start location of the loop |
640 | // and if there is a second DebugLoc in the header we use it as end location |
641 | // of the loop. |
642 | for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) { |
643 | if (DILocation *L = dyn_cast<DILocation>(Val: MDO)) { |
644 | if (!Start) |
645 | Start = DebugLoc(L); |
646 | else |
647 | return LocRange(Start, DebugLoc(L)); |
648 | } |
649 | } |
650 | |
651 | if (Start) |
652 | return LocRange(Start); |
653 | } |
654 | |
655 | // Try the pre-header first. |
656 | if (BasicBlock *PHeadBB = getLoopPreheader()) |
657 | if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc()) |
658 | return LocRange(DL); |
659 | |
660 | // If we have no pre-header or there are no instructions with debug |
661 | // info in it, try the header. |
662 | if (BasicBlock *HeadBB = getHeader()) |
663 | return LocRange(HeadBB->getTerminator()->getDebugLoc()); |
664 | |
665 | return LocRange(); |
666 | } |
667 | |
668 | std::string Loop::getLocStr() const { |
669 | std::string Result; |
670 | raw_string_ostream OS(Result); |
671 | if (const DebugLoc LoopDbgLoc = getStartLoc()) |
672 | LoopDbgLoc.print(OS); |
673 | else |
674 | // Just print the module name. |
675 | OS << getHeader()->getParent()->getParent()->getModuleIdentifier(); |
676 | return Result; |
677 | } |
678 | |
679 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
680 | LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); } |
681 | |
682 | LLVM_DUMP_METHOD void Loop::dumpVerbose() const { |
683 | print(dbgs(), /*Verbose=*/true); |
684 | } |
685 | #endif |
686 | |
687 | //===----------------------------------------------------------------------===// |
688 | // UnloopUpdater implementation |
689 | // |
690 | |
691 | namespace { |
692 | /// Find the new parent loop for all blocks within the "unloop" whose last |
693 | /// backedges has just been removed. |
694 | class UnloopUpdater { |
695 | Loop &Unloop; |
696 | LoopInfo *LI; |
697 | |
698 | LoopBlocksDFS DFS; |
699 | |
700 | // Map unloop's immediate subloops to their nearest reachable parents. Nested |
701 | // loops within these subloops will not change parents. However, an immediate |
702 | // subloop's new parent will be the nearest loop reachable from either its own |
703 | // exits *or* any of its nested loop's exits. |
704 | DenseMap<Loop *, Loop *> SubloopParents; |
705 | |
706 | // Flag the presence of an irreducible backedge whose destination is a block |
707 | // directly contained by the original unloop. |
708 | bool FoundIB = false; |
709 | |
710 | public: |
711 | UnloopUpdater(Loop *UL, LoopInfo *LInfo) : Unloop(*UL), LI(LInfo), DFS(UL) {} |
712 | |
713 | void updateBlockParents(); |
714 | |
715 | void removeBlocksFromAncestors(); |
716 | |
717 | void updateSubloopParents(); |
718 | |
719 | protected: |
720 | Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop); |
721 | }; |
722 | } // end anonymous namespace |
723 | |
724 | /// Update the parent loop for all blocks that are directly contained within the |
725 | /// original "unloop". |
726 | void UnloopUpdater::updateBlockParents() { |
727 | if (Unloop.getNumBlocks()) { |
728 | // Perform a post order CFG traversal of all blocks within this loop, |
729 | // propagating the nearest loop from successors to predecessors. |
730 | LoopBlocksTraversal Traversal(DFS, LI); |
731 | for (BasicBlock *POI : Traversal) { |
732 | |
733 | Loop *L = LI->getLoopFor(BB: POI); |
734 | Loop *NL = getNearestLoop(BB: POI, BBLoop: L); |
735 | |
736 | if (NL != L) { |
737 | // For reducible loops, NL is now an ancestor of Unloop. |
738 | assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) && |
739 | "uninitialized successor" ); |
740 | LI->changeLoopFor(BB: POI, L: NL); |
741 | } else { |
742 | // Or the current block is part of a subloop, in which case its parent |
743 | // is unchanged. |
744 | assert((FoundIB || Unloop.contains(L)) && "uninitialized successor" ); |
745 | } |
746 | } |
747 | } |
748 | // Each irreducible loop within the unloop induces a round of iteration using |
749 | // the DFS result cached by Traversal. |
750 | bool Changed = FoundIB; |
751 | for (unsigned NIters = 0; Changed; ++NIters) { |
752 | assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm" ); |
753 | (void)NIters; |
754 | |
755 | // Iterate over the postorder list of blocks, propagating the nearest loop |
756 | // from successors to predecessors as before. |
757 | Changed = false; |
758 | for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(), |
759 | POE = DFS.endPostorder(); |
760 | POI != POE; ++POI) { |
761 | |
762 | Loop *L = LI->getLoopFor(BB: *POI); |
763 | Loop *NL = getNearestLoop(BB: *POI, BBLoop: L); |
764 | if (NL != L) { |
765 | assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) && |
766 | "uninitialized successor" ); |
767 | LI->changeLoopFor(BB: *POI, L: NL); |
768 | Changed = true; |
769 | } |
770 | } |
771 | } |
772 | } |
773 | |
774 | /// Remove unloop's blocks from all ancestors below their new parents. |
775 | void UnloopUpdater::removeBlocksFromAncestors() { |
776 | // Remove all unloop's blocks (including those in nested subloops) from |
777 | // ancestors below the new parent loop. |
778 | for (BasicBlock *BB : Unloop.blocks()) { |
779 | Loop *OuterParent = LI->getLoopFor(BB); |
780 | if (Unloop.contains(L: OuterParent)) { |
781 | while (OuterParent->getParentLoop() != &Unloop) |
782 | OuterParent = OuterParent->getParentLoop(); |
783 | OuterParent = SubloopParents[OuterParent]; |
784 | } |
785 | // Remove blocks from former Ancestors except Unloop itself which will be |
786 | // deleted. |
787 | for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent; |
788 | OldParent = OldParent->getParentLoop()) { |
789 | assert(OldParent && "new loop is not an ancestor of the original" ); |
790 | OldParent->removeBlockFromLoop(BB); |
791 | } |
792 | } |
793 | } |
794 | |
795 | /// Update the parent loop for all subloops directly nested within unloop. |
796 | void UnloopUpdater::updateSubloopParents() { |
797 | while (!Unloop.isInnermost()) { |
798 | Loop *Subloop = *std::prev(x: Unloop.end()); |
799 | Unloop.removeChildLoop(I: std::prev(x: Unloop.end())); |
800 | |
801 | assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop" ); |
802 | if (Loop *Parent = SubloopParents[Subloop]) |
803 | Parent->addChildLoop(NewChild: Subloop); |
804 | else |
805 | LI->addTopLevelLoop(New: Subloop); |
806 | } |
807 | } |
808 | |
809 | /// Return the nearest parent loop among this block's successors. If a successor |
810 | /// is a subloop header, consider its parent to be the nearest parent of the |
811 | /// subloop's exits. |
812 | /// |
813 | /// For subloop blocks, simply update SubloopParents and return NULL. |
814 | Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) { |
815 | |
816 | // Initially for blocks directly contained by Unloop, NearLoop == Unloop and |
817 | // is considered uninitialized. |
818 | Loop *NearLoop = BBLoop; |
819 | |
820 | Loop *Subloop = nullptr; |
821 | if (NearLoop != &Unloop && Unloop.contains(L: NearLoop)) { |
822 | Subloop = NearLoop; |
823 | // Find the subloop ancestor that is directly contained within Unloop. |
824 | while (Subloop->getParentLoop() != &Unloop) { |
825 | Subloop = Subloop->getParentLoop(); |
826 | assert(Subloop && "subloop is not an ancestor of the original loop" ); |
827 | } |
828 | // Get the current nearest parent of the Subloop exits, initially Unloop. |
829 | NearLoop = SubloopParents.insert(KV: {Subloop, &Unloop}).first->second; |
830 | } |
831 | |
832 | if (succ_empty(BB)) { |
833 | assert(!Subloop && "subloop blocks must have a successor" ); |
834 | NearLoop = nullptr; // unloop blocks may now exit the function. |
835 | } |
836 | for (BasicBlock *Succ : successors(BB)) { |
837 | if (Succ == BB) |
838 | continue; // self loops are uninteresting |
839 | |
840 | Loop *L = LI->getLoopFor(BB: Succ); |
841 | if (L == &Unloop) { |
842 | // This successor has not been processed. This path must lead to an |
843 | // irreducible backedge. |
844 | assert((FoundIB || !DFS.hasPostorder(Succ)) && "should have seen IB" ); |
845 | FoundIB = true; |
846 | } |
847 | if (L != &Unloop && Unloop.contains(L)) { |
848 | // Successor is in a subloop. |
849 | if (Subloop) |
850 | continue; // Branching within subloops. Ignore it. |
851 | |
852 | // BB branches from the original into a subloop header. |
853 | assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops" ); |
854 | |
855 | // Get the current nearest parent of the Subloop's exits. |
856 | L = SubloopParents[L]; |
857 | // L could be Unloop if the only exit was an irreducible backedge. |
858 | } |
859 | if (L == &Unloop) { |
860 | continue; |
861 | } |
862 | // Handle critical edges from Unloop into a sibling loop. |
863 | if (L && !L->contains(L: &Unloop)) { |
864 | L = L->getParentLoop(); |
865 | } |
866 | // Remember the nearest parent loop among successors or subloop exits. |
867 | if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L)) |
868 | NearLoop = L; |
869 | } |
870 | if (Subloop) { |
871 | SubloopParents[Subloop] = NearLoop; |
872 | return BBLoop; |
873 | } |
874 | return NearLoop; |
875 | } |
876 | |
877 | LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); } |
878 | |
879 | bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA, |
880 | FunctionAnalysisManager::Invalidator &) { |
881 | // Check whether the analysis, all analyses on functions, or the function's |
882 | // CFG have been preserved. |
883 | auto PAC = PA.getChecker<LoopAnalysis>(); |
884 | return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() || |
885 | PAC.preservedSet<CFGAnalyses>()); |
886 | } |
887 | |
888 | void LoopInfo::erase(Loop *Unloop) { |
889 | assert(!Unloop->isInvalid() && "Loop has already been erased!" ); |
890 | |
891 | auto InvalidateOnExit = make_scope_exit(F: [&]() { destroy(L: Unloop); }); |
892 | |
893 | // First handle the special case of no parent loop to simplify the algorithm. |
894 | if (Unloop->isOutermost()) { |
895 | // Since BBLoop had no parent, Unloop blocks are no longer in a loop. |
896 | for (BasicBlock *BB : Unloop->blocks()) { |
897 | // Don't reparent blocks in subloops. |
898 | if (getLoopFor(BB) != Unloop) |
899 | continue; |
900 | |
901 | // Blocks no longer have a parent but are still referenced by Unloop until |
902 | // the Unloop object is deleted. |
903 | changeLoopFor(BB, L: nullptr); |
904 | } |
905 | |
906 | // Remove the loop from the top-level LoopInfo object. |
907 | for (iterator I = begin();; ++I) { |
908 | assert(I != end() && "Couldn't find loop" ); |
909 | if (*I == Unloop) { |
910 | removeLoop(I); |
911 | break; |
912 | } |
913 | } |
914 | |
915 | // Move all of the subloops to the top-level. |
916 | while (!Unloop->isInnermost()) |
917 | addTopLevelLoop(New: Unloop->removeChildLoop(I: std::prev(x: Unloop->end()))); |
918 | |
919 | return; |
920 | } |
921 | |
922 | // Update the parent loop for all blocks within the loop. Blocks within |
923 | // subloops will not change parents. |
924 | UnloopUpdater Updater(Unloop, this); |
925 | Updater.updateBlockParents(); |
926 | |
927 | // Remove blocks from former ancestor loops. |
928 | Updater.removeBlocksFromAncestors(); |
929 | |
930 | // Add direct subloops as children in their new parent loop. |
931 | Updater.updateSubloopParents(); |
932 | |
933 | // Remove unloop from its parent loop. |
934 | Loop *ParentLoop = Unloop->getParentLoop(); |
935 | for (Loop::iterator I = ParentLoop->begin();; ++I) { |
936 | assert(I != ParentLoop->end() && "Couldn't find loop" ); |
937 | if (*I == Unloop) { |
938 | ParentLoop->removeChildLoop(I); |
939 | break; |
940 | } |
941 | } |
942 | } |
943 | |
944 | bool LoopInfo::wouldBeOutOfLoopUseRequiringLCSSA( |
945 | const Value *V, const BasicBlock *ExitBB) const { |
946 | if (V->getType()->isTokenTy()) |
947 | // We can't form PHIs of token type, so the definition of LCSSA excludes |
948 | // values of that type. |
949 | return false; |
950 | |
951 | const Instruction *I = dyn_cast<Instruction>(Val: V); |
952 | if (!I) |
953 | return false; |
954 | const Loop *L = getLoopFor(BB: I->getParent()); |
955 | if (!L) |
956 | return false; |
957 | if (L->contains(BB: ExitBB)) |
958 | // Could be an exit bb of a subloop and contained in defining loop |
959 | return false; |
960 | |
961 | // We found a (new) out-of-loop use location, for a value defined in-loop. |
962 | // (Note that because of LCSSA, we don't have to account for values defined |
963 | // in sibling loops. Such values will have LCSSA phis of their own in the |
964 | // common parent loop.) |
965 | return true; |
966 | } |
967 | |
968 | AnalysisKey LoopAnalysis::Key; |
969 | |
970 | LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) { |
971 | // FIXME: Currently we create a LoopInfo from scratch for every function. |
972 | // This may prove to be too wasteful due to deallocating and re-allocating |
973 | // memory each time for the underlying map and vector datastructures. At some |
974 | // point it may prove worthwhile to use a freelist and recycle LoopInfo |
975 | // objects. I don't want to add that kind of complexity until the scope of |
976 | // the problem is better understood. |
977 | LoopInfo LI; |
978 | LI.analyze(DomTree: AM.getResult<DominatorTreeAnalysis>(IR&: F)); |
979 | return LI; |
980 | } |
981 | |
982 | PreservedAnalyses LoopPrinterPass::run(Function &F, |
983 | FunctionAnalysisManager &AM) { |
984 | auto &LI = AM.getResult<LoopAnalysis>(IR&: F); |
985 | OS << "Loop info for function '" << F.getName() << "':\n" ; |
986 | LI.print(OS); |
987 | return PreservedAnalyses::all(); |
988 | } |
989 | |
990 | void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) { |
991 | |
992 | if (forcePrintModuleIR()) { |
993 | // handling -print-module-scope |
994 | OS << Banner << " (loop: " ; |
995 | L.getHeader()->printAsOperand(O&: OS, PrintType: false); |
996 | OS << ")\n" ; |
997 | |
998 | // printing whole module |
999 | OS << *L.getHeader()->getModule(); |
1000 | return; |
1001 | } |
1002 | |
1003 | if (forcePrintFuncIR()) { |
1004 | // handling -print-loop-func-scope. |
1005 | // -print-module-scope overrides this. |
1006 | OS << Banner << " (loop: " ; |
1007 | L.getHeader()->printAsOperand(O&: OS, PrintType: false); |
1008 | OS << ")\n" ; |
1009 | |
1010 | // printing whole function. |
1011 | OS << *L.getHeader()->getParent(); |
1012 | return; |
1013 | } |
1014 | |
1015 | OS << Banner; |
1016 | |
1017 | auto * = L.getLoopPreheader(); |
1018 | if (PreHeader) { |
1019 | OS << "\n; Preheader:" ; |
1020 | PreHeader->print(OS); |
1021 | OS << "\n; Loop:" ; |
1022 | } |
1023 | |
1024 | for (auto *Block : L.blocks()) |
1025 | if (Block) |
1026 | Block->print(OS); |
1027 | else |
1028 | OS << "Printing <null> block" ; |
1029 | |
1030 | SmallVector<BasicBlock *, 8> ExitBlocks; |
1031 | L.getExitBlocks(ExitBlocks); |
1032 | if (!ExitBlocks.empty()) { |
1033 | OS << "\n; Exit blocks" ; |
1034 | for (auto *Block : ExitBlocks) |
1035 | if (Block) |
1036 | Block->print(OS); |
1037 | else |
1038 | OS << "Printing <null> block" ; |
1039 | } |
1040 | } |
1041 | |
1042 | MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) { |
1043 | // No loop metadata node, no loop properties. |
1044 | if (!LoopID) |
1045 | return nullptr; |
1046 | |
1047 | // First operand should refer to the metadata node itself, for legacy reasons. |
1048 | assert(LoopID->getNumOperands() > 0 && "requires at least one operand" ); |
1049 | assert(LoopID->getOperand(0) == LoopID && "invalid loop id" ); |
1050 | |
1051 | // Iterate over the metdata node operands and look for MDString metadata. |
1052 | for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) { |
1053 | MDNode *MD = dyn_cast<MDNode>(Val: MDO); |
1054 | if (!MD || MD->getNumOperands() < 1) |
1055 | continue; |
1056 | MDString *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
1057 | if (!S) |
1058 | continue; |
1059 | // Return the operand node if MDString holds expected metadata. |
1060 | if (Name == S->getString()) |
1061 | return MD; |
1062 | } |
1063 | |
1064 | // Loop property not found. |
1065 | return nullptr; |
1066 | } |
1067 | |
1068 | MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) { |
1069 | return findOptionMDForLoopID(LoopID: TheLoop->getLoopID(), Name); |
1070 | } |
1071 | |
1072 | /// Find string metadata for loop |
1073 | /// |
1074 | /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an |
1075 | /// operand or null otherwise. If the string metadata is not found return |
1076 | /// Optional's not-a-value. |
1077 | std::optional<const MDOperand *> |
1078 | llvm::findStringMetadataForLoop(const Loop *TheLoop, StringRef Name) { |
1079 | MDNode *MD = findOptionMDForLoop(TheLoop, Name); |
1080 | if (!MD) |
1081 | return std::nullopt; |
1082 | switch (MD->getNumOperands()) { |
1083 | case 1: |
1084 | return nullptr; |
1085 | case 2: |
1086 | return &MD->getOperand(I: 1); |
1087 | default: |
1088 | llvm_unreachable("loop metadata has 0 or 1 operand" ); |
1089 | } |
1090 | } |
1091 | |
1092 | std::optional<bool> llvm::getOptionalBoolLoopAttribute(const Loop *TheLoop, |
1093 | StringRef Name) { |
1094 | MDNode *MD = findOptionMDForLoop(TheLoop, Name); |
1095 | if (!MD) |
1096 | return std::nullopt; |
1097 | switch (MD->getNumOperands()) { |
1098 | case 1: |
1099 | // When the value is absent it is interpreted as 'attribute set'. |
1100 | return true; |
1101 | case 2: |
1102 | if (ConstantInt *IntMD = |
1103 | mdconst::extract_or_null<ConstantInt>(MD: MD->getOperand(I: 1).get())) |
1104 | return IntMD->getZExtValue(); |
1105 | return true; |
1106 | } |
1107 | llvm_unreachable("unexpected number of options" ); |
1108 | } |
1109 | |
1110 | bool llvm::getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) { |
1111 | return getOptionalBoolLoopAttribute(TheLoop, Name).value_or(u: false); |
1112 | } |
1113 | |
1114 | std::optional<int> llvm::getOptionalIntLoopAttribute(const Loop *TheLoop, |
1115 | StringRef Name) { |
1116 | const MDOperand *AttrMD = |
1117 | findStringMetadataForLoop(TheLoop, Name).value_or(u: nullptr); |
1118 | if (!AttrMD) |
1119 | return std::nullopt; |
1120 | |
1121 | ConstantInt *IntMD = mdconst::extract_or_null<ConstantInt>(MD: AttrMD->get()); |
1122 | if (!IntMD) |
1123 | return std::nullopt; |
1124 | |
1125 | return IntMD->getSExtValue(); |
1126 | } |
1127 | |
1128 | int llvm::getIntLoopAttribute(const Loop *TheLoop, StringRef Name, |
1129 | int Default) { |
1130 | return getOptionalIntLoopAttribute(TheLoop, Name).value_or(u&: Default); |
1131 | } |
1132 | |
1133 | CallBase *llvm::getLoopConvergenceHeart(const Loop *TheLoop) { |
1134 | BasicBlock *H = TheLoop->getHeader(); |
1135 | for (Instruction &II : *H) { |
1136 | if (auto *CB = dyn_cast<CallBase>(Val: &II)) { |
1137 | if (!CB->isConvergent()) |
1138 | continue; |
1139 | // This is the heart if it uses a token defined outside the loop. The |
1140 | // verifier has already checked that only the loop intrinsic can use such |
1141 | // a token. |
1142 | if (auto *Token = CB->getConvergenceControlToken()) { |
1143 | auto *TokenDef = cast<Instruction>(Val: Token); |
1144 | if (!TheLoop->contains(BB: TokenDef->getParent())) |
1145 | return CB; |
1146 | } |
1147 | return nullptr; |
1148 | } |
1149 | } |
1150 | return nullptr; |
1151 | } |
1152 | |
1153 | bool llvm::isFinite(const Loop *L) { |
1154 | return L->getHeader()->getParent()->willReturn(); |
1155 | } |
1156 | |
1157 | static const char *LLVMLoopMustProgress = "llvm.loop.mustprogress" ; |
1158 | |
1159 | bool llvm::hasMustProgress(const Loop *L) { |
1160 | return getBooleanLoopAttribute(TheLoop: L, Name: LLVMLoopMustProgress); |
1161 | } |
1162 | |
1163 | bool llvm::isMustProgress(const Loop *L) { |
1164 | return L->getHeader()->getParent()->mustProgress() || hasMustProgress(L); |
1165 | } |
1166 | |
1167 | bool llvm::isValidAsAccessGroup(MDNode *Node) { |
1168 | return Node->getNumOperands() == 0 && Node->isDistinct(); |
1169 | } |
1170 | |
1171 | MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context, |
1172 | MDNode *OrigLoopID, |
1173 | ArrayRef<StringRef> RemovePrefixes, |
1174 | ArrayRef<MDNode *> AddAttrs) { |
1175 | // First remove any existing loop metadata related to this transformation. |
1176 | SmallVector<Metadata *, 4> MDs; |
1177 | |
1178 | // Reserve first location for self reference to the LoopID metadata node. |
1179 | MDs.push_back(Elt: nullptr); |
1180 | |
1181 | // Remove metadata for the transformation that has been applied or that became |
1182 | // outdated. |
1183 | if (OrigLoopID) { |
1184 | for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: OrigLoopID->operands())) { |
1185 | bool IsVectorMetadata = false; |
1186 | Metadata *Op = MDO; |
1187 | if (MDNode *MD = dyn_cast<MDNode>(Val: Op)) { |
1188 | const MDString *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
1189 | if (S) |
1190 | IsVectorMetadata = |
1191 | llvm::any_of(Range&: RemovePrefixes, P: [S](StringRef Prefix) -> bool { |
1192 | return S->getString().starts_with(Prefix); |
1193 | }); |
1194 | } |
1195 | if (!IsVectorMetadata) |
1196 | MDs.push_back(Elt: Op); |
1197 | } |
1198 | } |
1199 | |
1200 | // Add metadata to avoid reapplying a transformation, such as |
1201 | // llvm.loop.unroll.disable and llvm.loop.isvectorized. |
1202 | MDs.append(in_start: AddAttrs.begin(), in_end: AddAttrs.end()); |
1203 | |
1204 | MDNode *NewLoopID = MDNode::getDistinct(Context, MDs); |
1205 | // Replace the temporary node with a self-reference. |
1206 | NewLoopID->replaceOperandWith(I: 0, New: NewLoopID); |
1207 | return NewLoopID; |
1208 | } |
1209 | |
1210 | //===----------------------------------------------------------------------===// |
1211 | // LoopInfo implementation |
1212 | // |
1213 | |
1214 | LoopInfoWrapperPass::LoopInfoWrapperPass() : FunctionPass(ID) {} |
1215 | |
1216 | char LoopInfoWrapperPass::ID = 0; |
1217 | INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops" , "Natural Loop Information" , |
1218 | true, true) |
1219 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
1220 | INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops" , "Natural Loop Information" , |
1221 | true, true) |
1222 | |
1223 | bool LoopInfoWrapperPass::runOnFunction(Function &) { |
1224 | releaseMemory(); |
1225 | LI.analyze(DomTree: getAnalysis<DominatorTreeWrapperPass>().getDomTree()); |
1226 | return false; |
1227 | } |
1228 | |
1229 | void LoopInfoWrapperPass::verifyAnalysis() const { |
1230 | // LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the |
1231 | // function each time verifyAnalysis is called is very expensive. The |
1232 | // -verify-loop-info option can enable this. In order to perform some |
1233 | // checking by default, LoopPass has been taught to call verifyLoop manually |
1234 | // during loop pass sequences. |
1235 | if (VerifyLoopInfo) { |
1236 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
1237 | LI.verify(DomTree: DT); |
1238 | } |
1239 | } |
1240 | |
1241 | void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { |
1242 | AU.setPreservesAll(); |
1243 | AU.addRequiredTransitive<DominatorTreeWrapperPass>(); |
1244 | } |
1245 | |
1246 | void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const { |
1247 | LI.print(OS); |
1248 | } |
1249 | |
1250 | PreservedAnalyses LoopVerifierPass::run(Function &F, |
1251 | FunctionAnalysisManager &AM) { |
1252 | LoopInfo &LI = AM.getResult<LoopAnalysis>(IR&: F); |
1253 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
1254 | LI.verify(DomTree: DT); |
1255 | return PreservedAnalyses::all(); |
1256 | } |
1257 | |
1258 | //===----------------------------------------------------------------------===// |
1259 | // LoopBlocksDFS implementation |
1260 | // |
1261 | |
1262 | /// Traverse the loop blocks and store the DFS result. |
1263 | /// Useful for clients that just want the final DFS result and don't need to |
1264 | /// visit blocks during the initial traversal. |
1265 | void LoopBlocksDFS::perform(const LoopInfo *LI) { |
1266 | LoopBlocksTraversal Traversal(*this, LI); |
1267 | for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(), |
1268 | POE = Traversal.end(); |
1269 | POI != POE; ++POI) |
1270 | ; |
1271 | } |
1272 | |