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