1 | //===-- LoopUtils.cpp - Loop Utility functions -------------------------===// |
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 common loop utility functions. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Transforms/Utils/LoopUtils.h" |
14 | #include "llvm/ADT/DenseSet.h" |
15 | #include "llvm/ADT/PriorityWorklist.h" |
16 | #include "llvm/ADT/ScopeExit.h" |
17 | #include "llvm/ADT/SetVector.h" |
18 | #include "llvm/ADT/SmallPtrSet.h" |
19 | #include "llvm/ADT/SmallVector.h" |
20 | #include "llvm/Analysis/AliasAnalysis.h" |
21 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
22 | #include "llvm/Analysis/DomTreeUpdater.h" |
23 | #include "llvm/Analysis/GlobalsModRef.h" |
24 | #include "llvm/Analysis/InstSimplifyFolder.h" |
25 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
26 | #include "llvm/Analysis/LoopInfo.h" |
27 | #include "llvm/Analysis/LoopPass.h" |
28 | #include "llvm/Analysis/MemorySSA.h" |
29 | #include "llvm/Analysis/MemorySSAUpdater.h" |
30 | #include "llvm/Analysis/ScalarEvolution.h" |
31 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" |
32 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
33 | #include "llvm/IR/DIBuilder.h" |
34 | #include "llvm/IR/Dominators.h" |
35 | #include "llvm/IR/Instructions.h" |
36 | #include "llvm/IR/IntrinsicInst.h" |
37 | #include "llvm/IR/MDBuilder.h" |
38 | #include "llvm/IR/Module.h" |
39 | #include "llvm/IR/PatternMatch.h" |
40 | #include "llvm/IR/ProfDataUtils.h" |
41 | #include "llvm/IR/ValueHandle.h" |
42 | #include "llvm/InitializePasses.h" |
43 | #include "llvm/Pass.h" |
44 | #include "llvm/Support/Compiler.h" |
45 | #include "llvm/Support/Debug.h" |
46 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
47 | #include "llvm/Transforms/Utils/Local.h" |
48 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
49 | |
50 | using namespace llvm; |
51 | using namespace llvm::PatternMatch; |
52 | |
53 | #define DEBUG_TYPE "loop-utils" |
54 | |
55 | static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced" ; |
56 | static const char *LLVMLoopDisableLICM = "llvm.licm.disable" ; |
57 | |
58 | bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, |
59 | MemorySSAUpdater *MSSAU, |
60 | bool PreserveLCSSA) { |
61 | bool Changed = false; |
62 | |
63 | // We re-use a vector for the in-loop predecesosrs. |
64 | SmallVector<BasicBlock *, 4> InLoopPredecessors; |
65 | |
66 | auto RewriteExit = [&](BasicBlock *BB) { |
67 | assert(InLoopPredecessors.empty() && |
68 | "Must start with an empty predecessors list!" ); |
69 | auto Cleanup = make_scope_exit(F: [&] { InLoopPredecessors.clear(); }); |
70 | |
71 | // See if there are any non-loop predecessors of this exit block and |
72 | // keep track of the in-loop predecessors. |
73 | bool IsDedicatedExit = true; |
74 | for (auto *PredBB : predecessors(BB)) |
75 | if (L->contains(BB: PredBB)) { |
76 | if (isa<IndirectBrInst>(Val: PredBB->getTerminator())) |
77 | // We cannot rewrite exiting edges from an indirectbr. |
78 | return false; |
79 | |
80 | InLoopPredecessors.push_back(Elt: PredBB); |
81 | } else { |
82 | IsDedicatedExit = false; |
83 | } |
84 | |
85 | assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!" ); |
86 | |
87 | // Nothing to do if this is already a dedicated exit. |
88 | if (IsDedicatedExit) |
89 | return false; |
90 | |
91 | auto *NewExitBB = SplitBlockPredecessors( |
92 | BB, Preds: InLoopPredecessors, Suffix: ".loopexit" , DT, LI, MSSAU, PreserveLCSSA); |
93 | |
94 | if (!NewExitBB) |
95 | LLVM_DEBUG( |
96 | dbgs() << "WARNING: Can't create a dedicated exit block for loop: " |
97 | << *L << "\n" ); |
98 | else |
99 | LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block " |
100 | << NewExitBB->getName() << "\n" ); |
101 | return true; |
102 | }; |
103 | |
104 | // Walk the exit blocks directly rather than building up a data structure for |
105 | // them, but only visit each one once. |
106 | SmallPtrSet<BasicBlock *, 4> Visited; |
107 | for (auto *BB : L->blocks()) |
108 | for (auto *SuccBB : successors(BB)) { |
109 | // We're looking for exit blocks so skip in-loop successors. |
110 | if (L->contains(BB: SuccBB)) |
111 | continue; |
112 | |
113 | // Visit each exit block exactly once. |
114 | if (!Visited.insert(Ptr: SuccBB).second) |
115 | continue; |
116 | |
117 | Changed |= RewriteExit(SuccBB); |
118 | } |
119 | |
120 | return Changed; |
121 | } |
122 | |
123 | /// Returns the instructions that use values defined in the loop. |
124 | SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) { |
125 | SmallVector<Instruction *, 8> UsedOutside; |
126 | |
127 | for (auto *Block : L->getBlocks()) |
128 | // FIXME: I believe that this could use copy_if if the Inst reference could |
129 | // be adapted into a pointer. |
130 | for (auto &Inst : *Block) { |
131 | auto Users = Inst.users(); |
132 | if (any_of(Range&: Users, P: [&](User *U) { |
133 | auto *Use = cast<Instruction>(Val: U); |
134 | return !L->contains(BB: Use->getParent()); |
135 | })) |
136 | UsedOutside.push_back(Elt: &Inst); |
137 | } |
138 | |
139 | return UsedOutside; |
140 | } |
141 | |
142 | void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) { |
143 | // By definition, all loop passes need the LoopInfo analysis and the |
144 | // Dominator tree it depends on. Because they all participate in the loop |
145 | // pass manager, they must also preserve these. |
146 | AU.addRequired<DominatorTreeWrapperPass>(); |
147 | AU.addPreserved<DominatorTreeWrapperPass>(); |
148 | AU.addRequired<LoopInfoWrapperPass>(); |
149 | AU.addPreserved<LoopInfoWrapperPass>(); |
150 | |
151 | // We must also preserve LoopSimplify and LCSSA. We locally access their IDs |
152 | // here because users shouldn't directly get them from this header. |
153 | extern char &LoopSimplifyID; |
154 | extern char &LCSSAID; |
155 | AU.addRequiredID(ID&: LoopSimplifyID); |
156 | AU.addPreservedID(ID&: LoopSimplifyID); |
157 | AU.addRequiredID(ID&: LCSSAID); |
158 | AU.addPreservedID(ID&: LCSSAID); |
159 | // This is used in the LPPassManager to perform LCSSA verification on passes |
160 | // which preserve lcssa form |
161 | AU.addRequired<LCSSAVerificationPass>(); |
162 | AU.addPreserved<LCSSAVerificationPass>(); |
163 | |
164 | // Loop passes are designed to run inside of a loop pass manager which means |
165 | // that any function analyses they require must be required by the first loop |
166 | // pass in the manager (so that it is computed before the loop pass manager |
167 | // runs) and preserved by all loop pasess in the manager. To make this |
168 | // reasonably robust, the set needed for most loop passes is maintained here. |
169 | // If your loop pass requires an analysis not listed here, you will need to |
170 | // carefully audit the loop pass manager nesting structure that results. |
171 | AU.addRequired<AAResultsWrapperPass>(); |
172 | AU.addPreserved<AAResultsWrapperPass>(); |
173 | AU.addPreserved<BasicAAWrapperPass>(); |
174 | AU.addPreserved<GlobalsAAWrapperPass>(); |
175 | AU.addPreserved<SCEVAAWrapperPass>(); |
176 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
177 | AU.addPreserved<ScalarEvolutionWrapperPass>(); |
178 | // FIXME: When all loop passes preserve MemorySSA, it can be required and |
179 | // preserved here instead of the individual handling in each pass. |
180 | } |
181 | |
182 | /// Manually defined generic "LoopPass" dependency initialization. This is used |
183 | /// to initialize the exact set of passes from above in \c |
184 | /// getLoopAnalysisUsage. It can be used within a loop pass's initialization |
185 | /// with: |
186 | /// |
187 | /// INITIALIZE_PASS_DEPENDENCY(LoopPass) |
188 | /// |
189 | /// As-if "LoopPass" were a pass. |
190 | void llvm::initializeLoopPassPass(PassRegistry &Registry) { |
191 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
192 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
193 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify) |
194 | INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass) |
195 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
196 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass) |
197 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) |
198 | INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass) |
199 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) |
200 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) |
201 | } |
202 | |
203 | /// Create MDNode for input string. |
204 | static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) { |
205 | LLVMContext &Context = TheLoop->getHeader()->getContext(); |
206 | Metadata *MDs[] = { |
207 | MDString::get(Context, Str: Name), |
208 | ConstantAsMetadata::get(C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Context), V))}; |
209 | return MDNode::get(Context, MDs); |
210 | } |
211 | |
212 | /// Set input string into loop metadata by keeping other values intact. |
213 | /// If the string is already in loop metadata update value if it is |
214 | /// different. |
215 | void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD, |
216 | unsigned V) { |
217 | SmallVector<Metadata *, 4> MDs(1); |
218 | // If the loop already has metadata, retain it. |
219 | MDNode *LoopID = TheLoop->getLoopID(); |
220 | if (LoopID) { |
221 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { |
222 | MDNode *Node = cast<MDNode>(Val: LoopID->getOperand(I: i)); |
223 | // If it is of form key = value, try to parse it. |
224 | if (Node->getNumOperands() == 2) { |
225 | MDString *S = dyn_cast<MDString>(Val: Node->getOperand(I: 0)); |
226 | if (S && S->getString() == StringMD) { |
227 | ConstantInt *IntMD = |
228 | mdconst::extract_or_null<ConstantInt>(MD: Node->getOperand(I: 1)); |
229 | if (IntMD && IntMD->getSExtValue() == V) |
230 | // It is already in place. Do nothing. |
231 | return; |
232 | // We need to update the value, so just skip it here and it will |
233 | // be added after copying other existed nodes. |
234 | continue; |
235 | } |
236 | } |
237 | MDs.push_back(Elt: Node); |
238 | } |
239 | } |
240 | // Add new metadata. |
241 | MDs.push_back(Elt: createStringMetadata(TheLoop, Name: StringMD, V)); |
242 | // Replace current metadata node with new one. |
243 | LLVMContext &Context = TheLoop->getHeader()->getContext(); |
244 | MDNode *NewLoopID = MDNode::get(Context, MDs); |
245 | // Set operand 0 to refer to the loop id itself. |
246 | NewLoopID->replaceOperandWith(I: 0, New: NewLoopID); |
247 | TheLoop->setLoopID(NewLoopID); |
248 | } |
249 | |
250 | std::optional<ElementCount> |
251 | llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) { |
252 | std::optional<int> Width = |
253 | getOptionalIntLoopAttribute(TheLoop, Name: "llvm.loop.vectorize.width" ); |
254 | |
255 | if (Width) { |
256 | std::optional<int> IsScalable = getOptionalIntLoopAttribute( |
257 | TheLoop, Name: "llvm.loop.vectorize.scalable.enable" ); |
258 | return ElementCount::get(MinVal: *Width, Scalable: IsScalable.value_or(u: false)); |
259 | } |
260 | |
261 | return std::nullopt; |
262 | } |
263 | |
264 | std::optional<MDNode *> llvm::makeFollowupLoopID( |
265 | MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions, |
266 | const char *InheritOptionsExceptPrefix, bool AlwaysNew) { |
267 | if (!OrigLoopID) { |
268 | if (AlwaysNew) |
269 | return nullptr; |
270 | return std::nullopt; |
271 | } |
272 | |
273 | assert(OrigLoopID->getOperand(0) == OrigLoopID); |
274 | |
275 | bool InheritAllAttrs = !InheritOptionsExceptPrefix; |
276 | bool InheritSomeAttrs = |
277 | InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0'; |
278 | SmallVector<Metadata *, 8> MDs; |
279 | MDs.push_back(Elt: nullptr); |
280 | |
281 | bool Changed = false; |
282 | if (InheritAllAttrs || InheritSomeAttrs) { |
283 | for (const MDOperand &Existing : drop_begin(RangeOrContainer: OrigLoopID->operands())) { |
284 | MDNode *Op = cast<MDNode>(Val: Existing.get()); |
285 | |
286 | auto InheritThisAttribute = [InheritSomeAttrs, |
287 | InheritOptionsExceptPrefix](MDNode *Op) { |
288 | if (!InheritSomeAttrs) |
289 | return false; |
290 | |
291 | // Skip malformatted attribute metadata nodes. |
292 | if (Op->getNumOperands() == 0) |
293 | return true; |
294 | Metadata *NameMD = Op->getOperand(I: 0).get(); |
295 | if (!isa<MDString>(Val: NameMD)) |
296 | return true; |
297 | StringRef AttrName = cast<MDString>(Val: NameMD)->getString(); |
298 | |
299 | // Do not inherit excluded attributes. |
300 | return !AttrName.starts_with(Prefix: InheritOptionsExceptPrefix); |
301 | }; |
302 | |
303 | if (InheritThisAttribute(Op)) |
304 | MDs.push_back(Elt: Op); |
305 | else |
306 | Changed = true; |
307 | } |
308 | } else { |
309 | // Modified if we dropped at least one attribute. |
310 | Changed = OrigLoopID->getNumOperands() > 1; |
311 | } |
312 | |
313 | bool HasAnyFollowup = false; |
314 | for (StringRef OptionName : FollowupOptions) { |
315 | MDNode *FollowupNode = findOptionMDForLoopID(LoopID: OrigLoopID, Name: OptionName); |
316 | if (!FollowupNode) |
317 | continue; |
318 | |
319 | HasAnyFollowup = true; |
320 | for (const MDOperand &Option : drop_begin(RangeOrContainer: FollowupNode->operands())) { |
321 | MDs.push_back(Elt: Option.get()); |
322 | Changed = true; |
323 | } |
324 | } |
325 | |
326 | // Attributes of the followup loop not specified explicity, so signal to the |
327 | // transformation pass to add suitable attributes. |
328 | if (!AlwaysNew && !HasAnyFollowup) |
329 | return std::nullopt; |
330 | |
331 | // If no attributes were added or remove, the previous loop Id can be reused. |
332 | if (!AlwaysNew && !Changed) |
333 | return OrigLoopID; |
334 | |
335 | // No attributes is equivalent to having no !llvm.loop metadata at all. |
336 | if (MDs.size() == 1) |
337 | return nullptr; |
338 | |
339 | // Build the new loop ID. |
340 | MDTuple *FollowupLoopID = MDNode::get(Context&: OrigLoopID->getContext(), MDs); |
341 | FollowupLoopID->replaceOperandWith(I: 0, New: FollowupLoopID); |
342 | return FollowupLoopID; |
343 | } |
344 | |
345 | bool llvm::hasDisableAllTransformsHint(const Loop *L) { |
346 | return getBooleanLoopAttribute(TheLoop: L, Name: LLVMLoopDisableNonforced); |
347 | } |
348 | |
349 | bool llvm::hasDisableLICMTransformsHint(const Loop *L) { |
350 | return getBooleanLoopAttribute(TheLoop: L, Name: LLVMLoopDisableLICM); |
351 | } |
352 | |
353 | TransformationMode llvm::hasUnrollTransformation(const Loop *L) { |
354 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll.disable" )) |
355 | return TM_SuppressedByUser; |
356 | |
357 | std::optional<int> Count = |
358 | getOptionalIntLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll.count" ); |
359 | if (Count) |
360 | return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser; |
361 | |
362 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll.enable" )) |
363 | return TM_ForcedByUser; |
364 | |
365 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll.full" )) |
366 | return TM_ForcedByUser; |
367 | |
368 | if (hasDisableAllTransformsHint(L)) |
369 | return TM_Disable; |
370 | |
371 | return TM_Unspecified; |
372 | } |
373 | |
374 | TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) { |
375 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll_and_jam.disable" )) |
376 | return TM_SuppressedByUser; |
377 | |
378 | std::optional<int> Count = |
379 | getOptionalIntLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll_and_jam.count" ); |
380 | if (Count) |
381 | return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser; |
382 | |
383 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.unroll_and_jam.enable" )) |
384 | return TM_ForcedByUser; |
385 | |
386 | if (hasDisableAllTransformsHint(L)) |
387 | return TM_Disable; |
388 | |
389 | return TM_Unspecified; |
390 | } |
391 | |
392 | TransformationMode llvm::hasVectorizeTransformation(const Loop *L) { |
393 | std::optional<bool> Enable = |
394 | getOptionalBoolLoopAttribute(TheLoop: L, Name: "llvm.loop.vectorize.enable" ); |
395 | |
396 | if (Enable == false) |
397 | return TM_SuppressedByUser; |
398 | |
399 | std::optional<ElementCount> VectorizeWidth = |
400 | getOptionalElementCountLoopAttribute(TheLoop: L); |
401 | std::optional<int> InterleaveCount = |
402 | getOptionalIntLoopAttribute(TheLoop: L, Name: "llvm.loop.interleave.count" ); |
403 | |
404 | // 'Forcing' vector width and interleave count to one effectively disables |
405 | // this tranformation. |
406 | if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() && |
407 | InterleaveCount == 1) |
408 | return TM_SuppressedByUser; |
409 | |
410 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.isvectorized" )) |
411 | return TM_Disable; |
412 | |
413 | if (Enable == true) |
414 | return TM_ForcedByUser; |
415 | |
416 | if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1) |
417 | return TM_Disable; |
418 | |
419 | if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1) |
420 | return TM_Enable; |
421 | |
422 | if (hasDisableAllTransformsHint(L)) |
423 | return TM_Disable; |
424 | |
425 | return TM_Unspecified; |
426 | } |
427 | |
428 | TransformationMode llvm::hasDistributeTransformation(const Loop *L) { |
429 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.distribute.enable" )) |
430 | return TM_ForcedByUser; |
431 | |
432 | if (hasDisableAllTransformsHint(L)) |
433 | return TM_Disable; |
434 | |
435 | return TM_Unspecified; |
436 | } |
437 | |
438 | TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) { |
439 | if (getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.licm_versioning.disable" )) |
440 | return TM_SuppressedByUser; |
441 | |
442 | if (hasDisableAllTransformsHint(L)) |
443 | return TM_Disable; |
444 | |
445 | return TM_Unspecified; |
446 | } |
447 | |
448 | /// Does a BFS from a given node to all of its children inside a given loop. |
449 | /// The returned vector of basic blocks includes the starting point. |
450 | SmallVector<BasicBlock *, 16> llvm::collectChildrenInLoop(DominatorTree *DT, |
451 | DomTreeNode *N, |
452 | const Loop *CurLoop) { |
453 | SmallVector<BasicBlock *, 16> Worklist; |
454 | auto AddRegionToWorklist = [&](DomTreeNode *DTN) { |
455 | // Only include subregions in the top level loop. |
456 | BasicBlock *BB = DTN->getBlock(); |
457 | if (CurLoop->contains(BB)) |
458 | Worklist.push_back(Elt: DTN->getBlock()); |
459 | }; |
460 | |
461 | AddRegionToWorklist(N); |
462 | |
463 | for (size_t I = 0; I < Worklist.size(); I++) { |
464 | for (DomTreeNode *Child : DT->getNode(BB: Worklist[I])->children()) |
465 | AddRegionToWorklist(Child); |
466 | } |
467 | |
468 | return Worklist; |
469 | } |
470 | |
471 | bool llvm::isAlmostDeadIV(PHINode *PN, BasicBlock *LatchBlock, Value *Cond) { |
472 | int LatchIdx = PN->getBasicBlockIndex(BB: LatchBlock); |
473 | assert(LatchIdx != -1 && "LatchBlock is not a case in this PHINode" ); |
474 | Value *IncV = PN->getIncomingValue(i: LatchIdx); |
475 | |
476 | for (User *U : PN->users()) |
477 | if (U != Cond && U != IncV) return false; |
478 | |
479 | for (User *U : IncV->users()) |
480 | if (U != Cond && U != PN) return false; |
481 | return true; |
482 | } |
483 | |
484 | |
485 | void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE, |
486 | LoopInfo *LI, MemorySSA *MSSA) { |
487 | assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!" ); |
488 | auto * = L->getLoopPreheader(); |
489 | assert(Preheader && "Preheader should exist!" ); |
490 | |
491 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
492 | if (MSSA) |
493 | MSSAU = std::make_unique<MemorySSAUpdater>(args&: MSSA); |
494 | |
495 | // Now that we know the removal is safe, remove the loop by changing the |
496 | // branch from the preheader to go to the single exit block. |
497 | // |
498 | // Because we're deleting a large chunk of code at once, the sequence in which |
499 | // we remove things is very important to avoid invalidation issues. |
500 | |
501 | // Tell ScalarEvolution that the loop is deleted. Do this before |
502 | // deleting the loop so that ScalarEvolution can look at the loop |
503 | // to determine what it needs to clean up. |
504 | if (SE) { |
505 | SE->forgetLoop(L); |
506 | SE->forgetBlockAndLoopDispositions(); |
507 | } |
508 | |
509 | Instruction *OldTerm = Preheader->getTerminator(); |
510 | assert(!OldTerm->mayHaveSideEffects() && |
511 | "Preheader must end with a side-effect-free terminator" ); |
512 | assert(OldTerm->getNumSuccessors() == 1 && |
513 | "Preheader must have a single successor" ); |
514 | // Connect the preheader to the exit block. Keep the old edge to the header |
515 | // around to perform the dominator tree update in two separate steps |
516 | // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge |
517 | // preheader -> header. |
518 | // |
519 | // |
520 | // 0. Preheader 1. Preheader 2. Preheader |
521 | // | | | | |
522 | // V | V | |
523 | // Header <--\ | Header <--\ | Header <--\ |
524 | // | | | | | | | | | | | |
525 | // | V | | | V | | | V | |
526 | // | Body --/ | | Body --/ | | Body --/ |
527 | // V V V V V |
528 | // Exit Exit Exit |
529 | // |
530 | // By doing this is two separate steps we can perform the dominator tree |
531 | // update without using the batch update API. |
532 | // |
533 | // Even when the loop is never executed, we cannot remove the edge from the |
534 | // source block to the exit block. Consider the case where the unexecuted loop |
535 | // branches back to an outer loop. If we deleted the loop and removed the edge |
536 | // coming to this inner loop, this will break the outer loop structure (by |
537 | // deleting the backedge of the outer loop). If the outer loop is indeed a |
538 | // non-loop, it will be deleted in a future iteration of loop deletion pass. |
539 | IRBuilder<> Builder(OldTerm); |
540 | |
541 | auto *ExitBlock = L->getUniqueExitBlock(); |
542 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
543 | if (ExitBlock) { |
544 | assert(ExitBlock && "Should have a unique exit block!" ); |
545 | assert(L->hasDedicatedExits() && "Loop should have dedicated exits!" ); |
546 | |
547 | Builder.CreateCondBr(Cond: Builder.getFalse(), True: L->getHeader(), False: ExitBlock); |
548 | // Remove the old branch. The conditional branch becomes a new terminator. |
549 | OldTerm->eraseFromParent(); |
550 | |
551 | // Rewrite phis in the exit block to get their inputs from the Preheader |
552 | // instead of the exiting block. |
553 | for (PHINode &P : ExitBlock->phis()) { |
554 | // Set the zero'th element of Phi to be from the preheader and remove all |
555 | // other incoming values. Given the loop has dedicated exits, all other |
556 | // incoming values must be from the exiting blocks. |
557 | int PredIndex = 0; |
558 | P.setIncomingBlock(i: PredIndex, BB: Preheader); |
559 | // Removes all incoming values from all other exiting blocks (including |
560 | // duplicate values from an exiting block). |
561 | // Nuke all entries except the zero'th entry which is the preheader entry. |
562 | P.removeIncomingValueIf(Predicate: [](unsigned Idx) { return Idx != 0; }, |
563 | /* DeletePHIIfEmpty */ false); |
564 | |
565 | assert((P.getNumIncomingValues() == 1 && |
566 | P.getIncomingBlock(PredIndex) == Preheader) && |
567 | "Should have exactly one value and that's from the preheader!" ); |
568 | } |
569 | |
570 | if (DT) { |
571 | DTU.applyUpdates(Updates: {{DominatorTree::Insert, Preheader, ExitBlock}}); |
572 | if (MSSA) { |
573 | MSSAU->applyUpdates(Updates: {{DominatorTree::Insert, Preheader, ExitBlock}}, |
574 | DT&: *DT); |
575 | if (VerifyMemorySSA) |
576 | MSSA->verifyMemorySSA(); |
577 | } |
578 | } |
579 | |
580 | // Disconnect the loop body by branching directly to its exit. |
581 | Builder.SetInsertPoint(Preheader->getTerminator()); |
582 | Builder.CreateBr(Dest: ExitBlock); |
583 | // Remove the old branch. |
584 | Preheader->getTerminator()->eraseFromParent(); |
585 | } else { |
586 | assert(L->hasNoExitBlocks() && |
587 | "Loop should have either zero or one exit blocks." ); |
588 | |
589 | Builder.SetInsertPoint(OldTerm); |
590 | Builder.CreateUnreachable(); |
591 | Preheader->getTerminator()->eraseFromParent(); |
592 | } |
593 | |
594 | if (DT) { |
595 | DTU.applyUpdates(Updates: {{DominatorTree::Delete, Preheader, L->getHeader()}}); |
596 | if (MSSA) { |
597 | MSSAU->applyUpdates(Updates: {{DominatorTree::Delete, Preheader, L->getHeader()}}, |
598 | DT&: *DT); |
599 | SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(), |
600 | L->block_end()); |
601 | MSSAU->removeBlocks(DeadBlocks: DeadBlockSet); |
602 | if (VerifyMemorySSA) |
603 | MSSA->verifyMemorySSA(); |
604 | } |
605 | } |
606 | |
607 | // Use a map to unique and a vector to guarantee deterministic ordering. |
608 | llvm::SmallDenseSet<DebugVariable, 4> DeadDebugSet; |
609 | llvm::SmallVector<DbgVariableRecord *, 4> DeadDbgVariableRecords; |
610 | |
611 | if (ExitBlock) { |
612 | // Given LCSSA form is satisfied, we should not have users of instructions |
613 | // within the dead loop outside of the loop. However, LCSSA doesn't take |
614 | // unreachable uses into account. We handle them here. |
615 | // We could do it after drop all references (in this case all users in the |
616 | // loop will be already eliminated and we have less work to do but according |
617 | // to API doc of User::dropAllReferences only valid operation after dropping |
618 | // references, is deletion. So let's substitute all usages of |
619 | // instruction from the loop with poison value of corresponding type first. |
620 | for (auto *Block : L->blocks()) |
621 | for (Instruction &I : *Block) { |
622 | auto *Poison = PoisonValue::get(T: I.getType()); |
623 | for (Use &U : llvm::make_early_inc_range(Range: I.uses())) { |
624 | if (auto *Usr = dyn_cast<Instruction>(Val: U.getUser())) |
625 | if (L->contains(BB: Usr->getParent())) |
626 | continue; |
627 | // If we have a DT then we can check that uses outside a loop only in |
628 | // unreachable block. |
629 | if (DT) |
630 | assert(!DT->isReachableFromEntry(U) && |
631 | "Unexpected user in reachable block" ); |
632 | U.set(Poison); |
633 | } |
634 | |
635 | // For one of each variable encountered, preserve a debug record (set |
636 | // to Poison) and transfer it to the loop exit. This terminates any |
637 | // variable locations that were set during the loop. |
638 | for (DbgVariableRecord &DVR : |
639 | llvm::make_early_inc_range(Range: filterDbgVars(R: I.getDbgRecordRange()))) { |
640 | DebugVariable Key(DVR.getVariable(), DVR.getExpression(), |
641 | DVR.getDebugLoc().get()); |
642 | if (!DeadDebugSet.insert(V: Key).second) |
643 | continue; |
644 | // Unlinks the DVR from it's container, for later insertion. |
645 | DVR.removeFromParent(); |
646 | DeadDbgVariableRecords.push_back(Elt: &DVR); |
647 | } |
648 | } |
649 | |
650 | // After the loop has been deleted all the values defined and modified |
651 | // inside the loop are going to be unavailable. Values computed in the |
652 | // loop will have been deleted, automatically causing their debug uses |
653 | // be be replaced with undef. Loop invariant values will still be available. |
654 | // Move dbg.values out the loop so that earlier location ranges are still |
655 | // terminated and loop invariant assignments are preserved. |
656 | DIBuilder DIB(*ExitBlock->getModule()); |
657 | BasicBlock::iterator InsertDbgValueBefore = |
658 | ExitBlock->getFirstInsertionPt(); |
659 | assert(InsertDbgValueBefore != ExitBlock->end() && |
660 | "There should be a non-PHI instruction in exit block, else these " |
661 | "instructions will have no parent." ); |
662 | |
663 | // Due to the "head" bit in BasicBlock::iterator, we're going to insert |
664 | // each DbgVariableRecord right at the start of the block, wheras dbg.values |
665 | // would be repeatedly inserted before the first instruction. To replicate |
666 | // this behaviour, do it backwards. |
667 | for (DbgVariableRecord *DVR : llvm::reverse(C&: DeadDbgVariableRecords)) |
668 | ExitBlock->insertDbgRecordBefore(DR: DVR, Here: InsertDbgValueBefore); |
669 | } |
670 | |
671 | // Remove the block from the reference counting scheme, so that we can |
672 | // delete it freely later. |
673 | for (auto *Block : L->blocks()) |
674 | Block->dropAllReferences(); |
675 | |
676 | if (MSSA && VerifyMemorySSA) |
677 | MSSA->verifyMemorySSA(); |
678 | |
679 | if (LI) { |
680 | // Erase the instructions and the blocks without having to worry |
681 | // about ordering because we already dropped the references. |
682 | // NOTE: This iteration is safe because erasing the block does not remove |
683 | // its entry from the loop's block list. We do that in the next section. |
684 | for (BasicBlock *BB : L->blocks()) |
685 | BB->eraseFromParent(); |
686 | |
687 | // Finally, the blocks from loopinfo. This has to happen late because |
688 | // otherwise our loop iterators won't work. |
689 | |
690 | SmallPtrSet<BasicBlock *, 8> blocks(llvm::from_range, L->blocks()); |
691 | for (BasicBlock *BB : blocks) |
692 | LI->removeBlock(BB); |
693 | |
694 | // The last step is to update LoopInfo now that we've eliminated this loop. |
695 | // Note: LoopInfo::erase remove the given loop and relink its subloops with |
696 | // its parent. While removeLoop/removeChildLoop remove the given loop but |
697 | // not relink its subloops, which is what we want. |
698 | if (Loop *ParentLoop = L->getParentLoop()) { |
699 | Loop::iterator I = find(Range&: *ParentLoop, Val: L); |
700 | assert(I != ParentLoop->end() && "Couldn't find loop" ); |
701 | ParentLoop->removeChildLoop(I); |
702 | } else { |
703 | Loop::iterator I = find(Range&: *LI, Val: L); |
704 | assert(I != LI->end() && "Couldn't find loop" ); |
705 | LI->removeLoop(I); |
706 | } |
707 | LI->destroy(L); |
708 | } |
709 | } |
710 | |
711 | void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE, |
712 | LoopInfo &LI, MemorySSA *MSSA) { |
713 | auto *Latch = L->getLoopLatch(); |
714 | assert(Latch && "multiple latches not yet supported" ); |
715 | auto * = L->getHeader(); |
716 | Loop *OutermostLoop = L->getOutermostLoop(); |
717 | |
718 | SE.forgetLoop(L); |
719 | SE.forgetBlockAndLoopDispositions(); |
720 | |
721 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
722 | if (MSSA) |
723 | MSSAU = std::make_unique<MemorySSAUpdater>(args&: MSSA); |
724 | |
725 | // Update the CFG and domtree. We chose to special case a couple of |
726 | // of common cases for code quality and test readability reasons. |
727 | [&]() -> void { |
728 | if (auto *BI = dyn_cast<BranchInst>(Val: Latch->getTerminator())) { |
729 | if (!BI->isConditional()) { |
730 | DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); |
731 | (void)changeToUnreachable(I: BI, /*PreserveLCSSA*/ true, DTU: &DTU, |
732 | MSSAU: MSSAU.get()); |
733 | return; |
734 | } |
735 | |
736 | // Conditional latch/exit - note that latch can be shared by inner |
737 | // and outer loop so the other target doesn't need to an exit |
738 | if (L->isLoopExiting(BB: Latch)) { |
739 | // TODO: Generalize ConstantFoldTerminator so that it can be used |
740 | // here without invalidating LCSSA or MemorySSA. (Tricky case for |
741 | // LCSSA: header is an exit block of a preceeding sibling loop w/o |
742 | // dedicated exits.) |
743 | const unsigned ExitIdx = L->contains(BB: BI->getSuccessor(i: 0)) ? 1 : 0; |
744 | BasicBlock *ExitBB = BI->getSuccessor(i: ExitIdx); |
745 | |
746 | DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); |
747 | Header->removePredecessor(Pred: Latch, KeepOneInputPHIs: true); |
748 | |
749 | IRBuilder<> Builder(BI); |
750 | auto *NewBI = Builder.CreateBr(Dest: ExitBB); |
751 | // Transfer the metadata to the new branch instruction (minus the |
752 | // loop info since this is no longer a loop) |
753 | NewBI->copyMetadata(SrcInst: *BI, WL: {LLVMContext::MD_dbg, |
754 | LLVMContext::MD_annotation}); |
755 | |
756 | BI->eraseFromParent(); |
757 | DTU.applyUpdates(Updates: {{DominatorTree::Delete, Latch, Header}}); |
758 | if (MSSA) |
759 | MSSAU->applyUpdates(Updates: {{DominatorTree::Delete, Latch, Header}}, DT); |
760 | return; |
761 | } |
762 | } |
763 | |
764 | // General case. By splitting the backedge, and then explicitly making it |
765 | // unreachable we gracefully handle corner cases such as switch and invoke |
766 | // termiantors. |
767 | auto *BackedgeBB = SplitEdge(From: Latch, To: Header, DT: &DT, LI: &LI, MSSAU: MSSAU.get()); |
768 | |
769 | DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager); |
770 | (void)changeToUnreachable(I: BackedgeBB->getTerminator(), |
771 | /*PreserveLCSSA*/ true, DTU: &DTU, MSSAU: MSSAU.get()); |
772 | }(); |
773 | |
774 | // Erase (and destroy) this loop instance. Handles relinking sub-loops |
775 | // and blocks within the loop as needed. |
776 | LI.erase(L); |
777 | |
778 | // If the loop we broke had a parent, then changeToUnreachable might have |
779 | // caused a block to be removed from the parent loop (see loop_nest_lcssa |
780 | // test case in zero-btc.ll for an example), thus changing the parent's |
781 | // exit blocks. If that happened, we need to rebuild LCSSA on the outermost |
782 | // loop which might have a had a block removed. |
783 | if (OutermostLoop != L) |
784 | formLCSSARecursively(L&: *OutermostLoop, DT, LI: &LI, SE: &SE); |
785 | } |
786 | |
787 | |
788 | /// Checks if \p L has an exiting latch branch. There may also be other |
789 | /// exiting blocks. Returns branch instruction terminating the loop |
790 | /// latch if above check is successful, nullptr otherwise. |
791 | static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) { |
792 | BasicBlock *Latch = L->getLoopLatch(); |
793 | if (!Latch) |
794 | return nullptr; |
795 | |
796 | BranchInst *LatchBR = dyn_cast<BranchInst>(Val: Latch->getTerminator()); |
797 | if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(BB: Latch)) |
798 | return nullptr; |
799 | |
800 | assert((LatchBR->getSuccessor(0) == L->getHeader() || |
801 | LatchBR->getSuccessor(1) == L->getHeader()) && |
802 | "At least one edge out of the latch must go to the header" ); |
803 | |
804 | return LatchBR; |
805 | } |
806 | |
807 | /// Return the estimated trip count for any exiting branch which dominates |
808 | /// the loop latch. |
809 | static std::optional<unsigned> getEstimatedTripCount(BranchInst *ExitingBranch, |
810 | Loop *L, |
811 | uint64_t &OrigExitWeight) { |
812 | // To estimate the number of times the loop body was executed, we want to |
813 | // know the number of times the backedge was taken, vs. the number of times |
814 | // we exited the loop. |
815 | uint64_t LoopWeight, ExitWeight; |
816 | if (!extractBranchWeights(I: *ExitingBranch, TrueVal&: LoopWeight, FalseVal&: ExitWeight)) |
817 | return std::nullopt; |
818 | |
819 | if (L->contains(BB: ExitingBranch->getSuccessor(i: 1))) |
820 | std::swap(a&: LoopWeight, b&: ExitWeight); |
821 | |
822 | if (!ExitWeight) |
823 | // Don't have a way to return predicated infinite |
824 | return std::nullopt; |
825 | |
826 | OrigExitWeight = ExitWeight; |
827 | |
828 | // Estimated exit count is a ratio of the loop weight by the weight of the |
829 | // edge exiting the loop, rounded to nearest. |
830 | uint64_t ExitCount = llvm::divideNearest(Numerator: LoopWeight, Denominator: ExitWeight); |
831 | |
832 | // When ExitCount + 1 would wrap in unsigned, saturate at UINT_MAX. |
833 | if (ExitCount >= std::numeric_limits<unsigned>::max()) |
834 | return std::numeric_limits<unsigned>::max(); |
835 | |
836 | // Estimated trip count is one plus estimated exit count. |
837 | return ExitCount + 1; |
838 | } |
839 | |
840 | std::optional<unsigned> |
841 | llvm::getLoopEstimatedTripCount(Loop *L, |
842 | unsigned *EstimatedLoopInvocationWeight) { |
843 | // Currently we take the estimate exit count only from the loop latch, |
844 | // ignoring other exiting blocks. This can overestimate the trip count |
845 | // if we exit through another exit, but can never underestimate it. |
846 | // TODO: incorporate information from other exits |
847 | if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) { |
848 | uint64_t ExitWeight; |
849 | if (std::optional<uint64_t> EstTripCount = |
850 | getEstimatedTripCount(ExitingBranch: LatchBranch, L, OrigExitWeight&: ExitWeight)) { |
851 | if (EstimatedLoopInvocationWeight) |
852 | *EstimatedLoopInvocationWeight = ExitWeight; |
853 | return *EstTripCount; |
854 | } |
855 | } |
856 | return std::nullopt; |
857 | } |
858 | |
859 | bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount, |
860 | unsigned EstimatedloopInvocationWeight) { |
861 | // At the moment, we currently support changing the estimate trip count of |
862 | // the latch branch only. We could extend this API to manipulate estimated |
863 | // trip counts for any exit. |
864 | BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L); |
865 | if (!LatchBranch) |
866 | return false; |
867 | |
868 | // Calculate taken and exit weights. |
869 | unsigned LatchExitWeight = 0; |
870 | unsigned BackedgeTakenWeight = 0; |
871 | |
872 | if (EstimatedTripCount > 0) { |
873 | LatchExitWeight = EstimatedloopInvocationWeight; |
874 | BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight; |
875 | } |
876 | |
877 | // Make a swap if back edge is taken when condition is "false". |
878 | if (LatchBranch->getSuccessor(i: 0) != L->getHeader()) |
879 | std::swap(a&: BackedgeTakenWeight, b&: LatchExitWeight); |
880 | |
881 | MDBuilder MDB(LatchBranch->getContext()); |
882 | |
883 | // Set/Update profile metadata. |
884 | LatchBranch->setMetadata( |
885 | KindID: LLVMContext::MD_prof, |
886 | Node: MDB.createBranchWeights(TrueWeight: BackedgeTakenWeight, FalseWeight: LatchExitWeight)); |
887 | |
888 | return true; |
889 | } |
890 | |
891 | bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop, |
892 | ScalarEvolution &SE) { |
893 | Loop *OuterL = InnerLoop->getParentLoop(); |
894 | if (!OuterL) |
895 | return true; |
896 | |
897 | // Get the backedge taken count for the inner loop |
898 | BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); |
899 | const SCEV *InnerLoopBECountSC = SE.getExitCount(L: InnerLoop, ExitingBlock: InnerLoopLatch); |
900 | if (isa<SCEVCouldNotCompute>(Val: InnerLoopBECountSC) || |
901 | !InnerLoopBECountSC->getType()->isIntegerTy()) |
902 | return false; |
903 | |
904 | // Get whether count is invariant to the outer loop |
905 | ScalarEvolution::LoopDisposition LD = |
906 | SE.getLoopDisposition(S: InnerLoopBECountSC, L: OuterL); |
907 | if (LD != ScalarEvolution::LoopInvariant) |
908 | return false; |
909 | |
910 | return true; |
911 | } |
912 | |
913 | constexpr Intrinsic::ID llvm::getReductionIntrinsicID(RecurKind RK) { |
914 | switch (RK) { |
915 | default: |
916 | llvm_unreachable("Unexpected recurrence kind" ); |
917 | case RecurKind::Add: |
918 | return Intrinsic::vector_reduce_add; |
919 | case RecurKind::Mul: |
920 | return Intrinsic::vector_reduce_mul; |
921 | case RecurKind::And: |
922 | return Intrinsic::vector_reduce_and; |
923 | case RecurKind::Or: |
924 | return Intrinsic::vector_reduce_or; |
925 | case RecurKind::Xor: |
926 | return Intrinsic::vector_reduce_xor; |
927 | case RecurKind::FMulAdd: |
928 | case RecurKind::FAdd: |
929 | return Intrinsic::vector_reduce_fadd; |
930 | case RecurKind::FMul: |
931 | return Intrinsic::vector_reduce_fmul; |
932 | case RecurKind::SMax: |
933 | return Intrinsic::vector_reduce_smax; |
934 | case RecurKind::SMin: |
935 | return Intrinsic::vector_reduce_smin; |
936 | case RecurKind::UMax: |
937 | return Intrinsic::vector_reduce_umax; |
938 | case RecurKind::UMin: |
939 | return Intrinsic::vector_reduce_umin; |
940 | case RecurKind::FMax: |
941 | return Intrinsic::vector_reduce_fmax; |
942 | case RecurKind::FMin: |
943 | return Intrinsic::vector_reduce_fmin; |
944 | case RecurKind::FMaximum: |
945 | return Intrinsic::vector_reduce_fmaximum; |
946 | case RecurKind::FMinimum: |
947 | return Intrinsic::vector_reduce_fminimum; |
948 | case RecurKind::FMaximumNum: |
949 | return Intrinsic::vector_reduce_fmax; |
950 | case RecurKind::FMinimumNum: |
951 | return Intrinsic::vector_reduce_fmin; |
952 | } |
953 | } |
954 | |
955 | // This is the inverse to getReductionForBinop |
956 | unsigned llvm::getArithmeticReductionInstruction(Intrinsic::ID RdxID) { |
957 | switch (RdxID) { |
958 | case Intrinsic::vector_reduce_fadd: |
959 | return Instruction::FAdd; |
960 | case Intrinsic::vector_reduce_fmul: |
961 | return Instruction::FMul; |
962 | case Intrinsic::vector_reduce_add: |
963 | return Instruction::Add; |
964 | case Intrinsic::vector_reduce_mul: |
965 | return Instruction::Mul; |
966 | case Intrinsic::vector_reduce_and: |
967 | return Instruction::And; |
968 | case Intrinsic::vector_reduce_or: |
969 | return Instruction::Or; |
970 | case Intrinsic::vector_reduce_xor: |
971 | return Instruction::Xor; |
972 | case Intrinsic::vector_reduce_smax: |
973 | case Intrinsic::vector_reduce_smin: |
974 | case Intrinsic::vector_reduce_umax: |
975 | case Intrinsic::vector_reduce_umin: |
976 | return Instruction::ICmp; |
977 | case Intrinsic::vector_reduce_fmax: |
978 | case Intrinsic::vector_reduce_fmin: |
979 | return Instruction::FCmp; |
980 | default: |
981 | llvm_unreachable("Unexpected ID" ); |
982 | } |
983 | } |
984 | |
985 | // This is the inverse to getArithmeticReductionInstruction |
986 | Intrinsic::ID llvm::getReductionForBinop(Instruction::BinaryOps Opc) { |
987 | switch (Opc) { |
988 | default: |
989 | break; |
990 | case Instruction::Add: |
991 | return Intrinsic::vector_reduce_add; |
992 | case Instruction::Mul: |
993 | return Intrinsic::vector_reduce_mul; |
994 | case Instruction::And: |
995 | return Intrinsic::vector_reduce_and; |
996 | case Instruction::Or: |
997 | return Intrinsic::vector_reduce_or; |
998 | case Instruction::Xor: |
999 | return Intrinsic::vector_reduce_xor; |
1000 | } |
1001 | return Intrinsic::not_intrinsic; |
1002 | } |
1003 | |
1004 | Intrinsic::ID llvm::getMinMaxReductionIntrinsicOp(Intrinsic::ID RdxID) { |
1005 | switch (RdxID) { |
1006 | default: |
1007 | llvm_unreachable("Unknown min/max recurrence kind" ); |
1008 | case Intrinsic::vector_reduce_umin: |
1009 | return Intrinsic::umin; |
1010 | case Intrinsic::vector_reduce_umax: |
1011 | return Intrinsic::umax; |
1012 | case Intrinsic::vector_reduce_smin: |
1013 | return Intrinsic::smin; |
1014 | case Intrinsic::vector_reduce_smax: |
1015 | return Intrinsic::smax; |
1016 | case Intrinsic::vector_reduce_fmin: |
1017 | return Intrinsic::minnum; |
1018 | case Intrinsic::vector_reduce_fmax: |
1019 | return Intrinsic::maxnum; |
1020 | case Intrinsic::vector_reduce_fminimum: |
1021 | return Intrinsic::minimum; |
1022 | case Intrinsic::vector_reduce_fmaximum: |
1023 | return Intrinsic::maximum; |
1024 | } |
1025 | } |
1026 | |
1027 | Intrinsic::ID llvm::getMinMaxReductionIntrinsicOp(RecurKind RK) { |
1028 | switch (RK) { |
1029 | default: |
1030 | llvm_unreachable("Unknown min/max recurrence kind" ); |
1031 | case RecurKind::UMin: |
1032 | return Intrinsic::umin; |
1033 | case RecurKind::UMax: |
1034 | return Intrinsic::umax; |
1035 | case RecurKind::SMin: |
1036 | return Intrinsic::smin; |
1037 | case RecurKind::SMax: |
1038 | return Intrinsic::smax; |
1039 | case RecurKind::FMin: |
1040 | return Intrinsic::minnum; |
1041 | case RecurKind::FMax: |
1042 | return Intrinsic::maxnum; |
1043 | case RecurKind::FMinimum: |
1044 | return Intrinsic::minimum; |
1045 | case RecurKind::FMaximum: |
1046 | return Intrinsic::maximum; |
1047 | case RecurKind::FMinimumNum: |
1048 | return Intrinsic::minimumnum; |
1049 | case RecurKind::FMaximumNum: |
1050 | return Intrinsic::maximumnum; |
1051 | } |
1052 | } |
1053 | |
1054 | RecurKind llvm::getMinMaxReductionRecurKind(Intrinsic::ID RdxID) { |
1055 | switch (RdxID) { |
1056 | case Intrinsic::vector_reduce_smax: |
1057 | return RecurKind::SMax; |
1058 | case Intrinsic::vector_reduce_smin: |
1059 | return RecurKind::SMin; |
1060 | case Intrinsic::vector_reduce_umax: |
1061 | return RecurKind::UMax; |
1062 | case Intrinsic::vector_reduce_umin: |
1063 | return RecurKind::UMin; |
1064 | case Intrinsic::vector_reduce_fmax: |
1065 | return RecurKind::FMax; |
1066 | case Intrinsic::vector_reduce_fmin: |
1067 | return RecurKind::FMin; |
1068 | default: |
1069 | return RecurKind::None; |
1070 | } |
1071 | } |
1072 | |
1073 | CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) { |
1074 | switch (RK) { |
1075 | default: |
1076 | llvm_unreachable("Unknown min/max recurrence kind" ); |
1077 | case RecurKind::UMin: |
1078 | return CmpInst::ICMP_ULT; |
1079 | case RecurKind::UMax: |
1080 | return CmpInst::ICMP_UGT; |
1081 | case RecurKind::SMin: |
1082 | return CmpInst::ICMP_SLT; |
1083 | case RecurKind::SMax: |
1084 | return CmpInst::ICMP_SGT; |
1085 | case RecurKind::FMin: |
1086 | return CmpInst::FCMP_OLT; |
1087 | case RecurKind::FMax: |
1088 | return CmpInst::FCMP_OGT; |
1089 | // We do not add FMinimum/FMaximum recurrence kind here since there is no |
1090 | // equivalent predicate which compares signed zeroes according to the |
1091 | // semantics of the intrinsics (llvm.minimum/maximum). |
1092 | } |
1093 | } |
1094 | |
1095 | Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left, |
1096 | Value *Right) { |
1097 | Type *Ty = Left->getType(); |
1098 | if (Ty->isIntOrIntVectorTy() || |
1099 | (RK == RecurKind::FMinimum || RK == RecurKind::FMaximum || |
1100 | RK == RecurKind::FMinimumNum || RK == RecurKind::FMaximumNum)) { |
1101 | // TODO: Add float minnum/maxnum support when FMF nnan is set. |
1102 | Intrinsic::ID Id = getMinMaxReductionIntrinsicOp(RK); |
1103 | return Builder.CreateIntrinsic(RetTy: Ty, ID: Id, Args: {Left, Right}, FMFSource: nullptr, |
1104 | Name: "rdx.minmax" ); |
1105 | } |
1106 | CmpInst::Predicate Pred = getMinMaxReductionPredicate(RK); |
1107 | Value *Cmp = Builder.CreateCmp(Pred, LHS: Left, RHS: Right, Name: "rdx.minmax.cmp" ); |
1108 | Value *Select = Builder.CreateSelect(C: Cmp, True: Left, False: Right, Name: "rdx.minmax.select" ); |
1109 | return Select; |
1110 | } |
1111 | |
1112 | // Helper to generate an ordered reduction. |
1113 | Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src, |
1114 | unsigned Op, RecurKind RdxKind) { |
1115 | unsigned VF = cast<FixedVectorType>(Val: Src->getType())->getNumElements(); |
1116 | |
1117 | // Extract and apply reduction ops in ascending order: |
1118 | // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1] |
1119 | Value *Result = Acc; |
1120 | for (unsigned = 0; ExtractIdx != VF; ++ExtractIdx) { |
1121 | Value *Ext = |
1122 | Builder.CreateExtractElement(Vec: Src, Idx: Builder.getInt32(C: ExtractIdx)); |
1123 | |
1124 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { |
1125 | Result = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)Op, LHS: Result, RHS: Ext, |
1126 | Name: "bin.rdx" ); |
1127 | } else { |
1128 | assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) && |
1129 | "Invalid min/max" ); |
1130 | Result = createMinMaxOp(Builder, RK: RdxKind, Left: Result, Right: Ext); |
1131 | } |
1132 | } |
1133 | |
1134 | return Result; |
1135 | } |
1136 | |
1137 | // Helper to generate a log2 shuffle reduction. |
1138 | Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src, |
1139 | unsigned Op, |
1140 | TargetTransformInfo::ReductionShuffle RS, |
1141 | RecurKind RdxKind) { |
1142 | unsigned VF = cast<FixedVectorType>(Val: Src->getType())->getNumElements(); |
1143 | // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles |
1144 | // and vector ops, reducing the set of values being computed by half each |
1145 | // round. |
1146 | assert(isPowerOf2_32(VF) && |
1147 | "Reduction emission only supported for pow2 vectors!" ); |
1148 | // Note: fast-math-flags flags are controlled by the builder configuration |
1149 | // and are assumed to apply to all generated arithmetic instructions. Other |
1150 | // poison generating flags (nsw/nuw/inbounds/inrange/exact) are not part |
1151 | // of the builder configuration, and since they're not passed explicitly, |
1152 | // will never be relevant here. Note that it would be generally unsound to |
1153 | // propagate these from an intrinsic call to the expansion anyways as we/ |
1154 | // change the order of operations. |
1155 | auto BuildShuffledOp = [&Builder, &Op, |
1156 | &RdxKind](SmallVectorImpl<int> &ShuffleMask, |
1157 | Value *&TmpVec) -> void { |
1158 | Value *Shuf = Builder.CreateShuffleVector(V: TmpVec, Mask: ShuffleMask, Name: "rdx.shuf" ); |
1159 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { |
1160 | TmpVec = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)Op, LHS: TmpVec, RHS: Shuf, |
1161 | Name: "bin.rdx" ); |
1162 | } else { |
1163 | assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) && |
1164 | "Invalid min/max" ); |
1165 | TmpVec = createMinMaxOp(Builder, RK: RdxKind, Left: TmpVec, Right: Shuf); |
1166 | } |
1167 | }; |
1168 | |
1169 | Value *TmpVec = Src; |
1170 | if (TargetTransformInfo::ReductionShuffle::Pairwise == RS) { |
1171 | SmallVector<int, 32> ShuffleMask(VF); |
1172 | for (unsigned stride = 1; stride < VF; stride <<= 1) { |
1173 | // Initialise the mask with undef. |
1174 | llvm::fill(Range&: ShuffleMask, Value: -1); |
1175 | for (unsigned j = 0; j < VF; j += stride << 1) { |
1176 | ShuffleMask[j] = j + stride; |
1177 | } |
1178 | BuildShuffledOp(ShuffleMask, TmpVec); |
1179 | } |
1180 | } else { |
1181 | SmallVector<int, 32> ShuffleMask(VF); |
1182 | for (unsigned i = VF; i != 1; i >>= 1) { |
1183 | // Move the upper half of the vector to the lower half. |
1184 | for (unsigned j = 0; j != i / 2; ++j) |
1185 | ShuffleMask[j] = i / 2 + j; |
1186 | |
1187 | // Fill the rest of the mask with undef. |
1188 | std::fill(first: &ShuffleMask[i / 2], last: ShuffleMask.end(), value: -1); |
1189 | BuildShuffledOp(ShuffleMask, TmpVec); |
1190 | } |
1191 | } |
1192 | // The result is in the first element of the vector. |
1193 | return Builder.CreateExtractElement(Vec: TmpVec, Idx: Builder.getInt32(C: 0)); |
1194 | } |
1195 | |
1196 | Value *llvm::createAnyOfReduction(IRBuilderBase &Builder, Value *Src, |
1197 | Value *InitVal, PHINode *OrigPhi) { |
1198 | Value *NewVal = nullptr; |
1199 | |
1200 | // First use the original phi to determine the new value we're trying to |
1201 | // select from in the loop. |
1202 | SelectInst *SI = nullptr; |
1203 | for (auto *U : OrigPhi->users()) { |
1204 | if ((SI = dyn_cast<SelectInst>(Val: U))) |
1205 | break; |
1206 | } |
1207 | assert(SI && "One user of the original phi should be a select" ); |
1208 | |
1209 | if (SI->getTrueValue() == OrigPhi) |
1210 | NewVal = SI->getFalseValue(); |
1211 | else { |
1212 | assert(SI->getFalseValue() == OrigPhi && |
1213 | "At least one input to the select should be the original Phi" ); |
1214 | NewVal = SI->getTrueValue(); |
1215 | } |
1216 | |
1217 | // If any predicate is true it means that we want to select the new value. |
1218 | Value *AnyOf = |
1219 | Src->getType()->isVectorTy() ? Builder.CreateOrReduce(Src) : Src; |
1220 | // The compares in the loop may yield poison, which propagates through the |
1221 | // bitwise ORs. Freeze it here before the condition is used. |
1222 | AnyOf = Builder.CreateFreeze(V: AnyOf); |
1223 | return Builder.CreateSelect(C: AnyOf, True: NewVal, False: InitVal, Name: "rdx.select" ); |
1224 | } |
1225 | |
1226 | Value *llvm::createFindLastIVReduction(IRBuilderBase &Builder, Value *Src, |
1227 | RecurKind RdxKind, Value *Start, |
1228 | Value *Sentinel) { |
1229 | bool IsSigned = RecurrenceDescriptor::isSignedRecurrenceKind(Kind: RdxKind); |
1230 | bool IsMaxRdx = RecurrenceDescriptor::isFindLastIVRecurrenceKind(Kind: RdxKind); |
1231 | Value *MaxRdx = Src->getType()->isVectorTy() |
1232 | ? (IsMaxRdx ? Builder.CreateIntMaxReduce(Src, IsSigned) |
1233 | : Builder.CreateIntMinReduce(Src, IsSigned)) |
1234 | : Src; |
1235 | // Correct the final reduction result back to the start value if the maximum |
1236 | // reduction is sentinel value. |
1237 | Value *Cmp = |
1238 | Builder.CreateCmp(Pred: CmpInst::ICMP_NE, LHS: MaxRdx, RHS: Sentinel, Name: "rdx.select.cmp" ); |
1239 | return Builder.CreateSelect(C: Cmp, True: MaxRdx, False: Start, Name: "rdx.select" ); |
1240 | } |
1241 | |
1242 | Value *llvm::getReductionIdentity(Intrinsic::ID RdxID, Type *Ty, |
1243 | FastMathFlags Flags) { |
1244 | bool Negative = false; |
1245 | switch (RdxID) { |
1246 | default: |
1247 | llvm_unreachable("Expecting a reduction intrinsic" ); |
1248 | case Intrinsic::vector_reduce_add: |
1249 | case Intrinsic::vector_reduce_mul: |
1250 | case Intrinsic::vector_reduce_or: |
1251 | case Intrinsic::vector_reduce_xor: |
1252 | case Intrinsic::vector_reduce_and: |
1253 | case Intrinsic::vector_reduce_fadd: |
1254 | case Intrinsic::vector_reduce_fmul: { |
1255 | unsigned Opc = getArithmeticReductionInstruction(RdxID); |
1256 | return ConstantExpr::getBinOpIdentity(Opcode: Opc, Ty, AllowRHSConstant: false, |
1257 | NSZ: Flags.noSignedZeros()); |
1258 | } |
1259 | case Intrinsic::vector_reduce_umax: |
1260 | case Intrinsic::vector_reduce_umin: |
1261 | case Intrinsic::vector_reduce_smin: |
1262 | case Intrinsic::vector_reduce_smax: { |
1263 | Intrinsic::ID ScalarID = getMinMaxReductionIntrinsicOp(RdxID); |
1264 | return ConstantExpr::getIntrinsicIdentity(ScalarID, Ty); |
1265 | } |
1266 | case Intrinsic::vector_reduce_fmax: |
1267 | case Intrinsic::vector_reduce_fmaximum: |
1268 | Negative = true; |
1269 | [[fallthrough]]; |
1270 | case Intrinsic::vector_reduce_fmin: |
1271 | case Intrinsic::vector_reduce_fminimum: { |
1272 | bool PropagatesNaN = RdxID == Intrinsic::vector_reduce_fminimum || |
1273 | RdxID == Intrinsic::vector_reduce_fmaximum; |
1274 | const fltSemantics &Semantics = Ty->getFltSemantics(); |
1275 | return (!Flags.noNaNs() && !PropagatesNaN) |
1276 | ? ConstantFP::getQNaN(Ty, Negative) |
1277 | : !Flags.noInfs() |
1278 | ? ConstantFP::getInfinity(Ty, Negative) |
1279 | : ConstantFP::get(Ty, V: APFloat::getLargest(Sem: Semantics, Negative)); |
1280 | } |
1281 | } |
1282 | } |
1283 | |
1284 | Value *llvm::getRecurrenceIdentity(RecurKind K, Type *Tp, FastMathFlags FMF) { |
1285 | assert((!(K == RecurKind::FMin || K == RecurKind::FMax) || |
1286 | (FMF.noNaNs() && FMF.noSignedZeros())) && |
1287 | "nnan, nsz is expected to be set for FP min/max reduction." ); |
1288 | Intrinsic::ID RdxID = getReductionIntrinsicID(RK: K); |
1289 | return getReductionIdentity(RdxID, Ty: Tp, Flags: FMF); |
1290 | } |
1291 | |
1292 | Value *llvm::createSimpleReduction(IRBuilderBase &Builder, Value *Src, |
1293 | RecurKind RdxKind) { |
1294 | auto *SrcVecEltTy = cast<VectorType>(Val: Src->getType())->getElementType(); |
1295 | auto getIdentity = [&]() { |
1296 | return getRecurrenceIdentity(K: RdxKind, Tp: SrcVecEltTy, |
1297 | FMF: Builder.getFastMathFlags()); |
1298 | }; |
1299 | switch (RdxKind) { |
1300 | case RecurKind::Add: |
1301 | case RecurKind::Mul: |
1302 | case RecurKind::And: |
1303 | case RecurKind::Or: |
1304 | case RecurKind::Xor: |
1305 | case RecurKind::SMax: |
1306 | case RecurKind::SMin: |
1307 | case RecurKind::UMax: |
1308 | case RecurKind::UMin: |
1309 | case RecurKind::FMax: |
1310 | case RecurKind::FMin: |
1311 | case RecurKind::FMinimum: |
1312 | case RecurKind::FMaximum: |
1313 | case RecurKind::FMinimumNum: |
1314 | case RecurKind::FMaximumNum: |
1315 | return Builder.CreateUnaryIntrinsic(ID: getReductionIntrinsicID(RK: RdxKind), V: Src); |
1316 | case RecurKind::FMulAdd: |
1317 | case RecurKind::FAdd: |
1318 | return Builder.CreateFAddReduce(Acc: getIdentity(), Src); |
1319 | case RecurKind::FMul: |
1320 | return Builder.CreateFMulReduce(Acc: getIdentity(), Src); |
1321 | default: |
1322 | llvm_unreachable("Unhandled opcode" ); |
1323 | } |
1324 | } |
1325 | |
1326 | Value *llvm::createSimpleReduction(IRBuilderBase &Builder, Value *Src, |
1327 | RecurKind Kind, Value *Mask, Value *EVL) { |
1328 | assert(!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) && |
1329 | !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) && |
1330 | "AnyOf and FindIV reductions are not supported." ); |
1331 | Intrinsic::ID Id = getReductionIntrinsicID(RK: Kind); |
1332 | auto VPID = VPIntrinsic::getForIntrinsic(Id); |
1333 | assert(VPReductionIntrinsic::isVPReduction(VPID) && |
1334 | "No VPIntrinsic for this reduction" ); |
1335 | auto *EltTy = cast<VectorType>(Val: Src->getType())->getElementType(); |
1336 | Value *Iden = getRecurrenceIdentity(K: Kind, Tp: EltTy, FMF: Builder.getFastMathFlags()); |
1337 | Value *Ops[] = {Iden, Src, Mask, EVL}; |
1338 | return Builder.CreateIntrinsic(RetTy: EltTy, ID: VPID, Args: Ops); |
1339 | } |
1340 | |
1341 | Value *llvm::createOrderedReduction(IRBuilderBase &B, RecurKind Kind, |
1342 | Value *Src, Value *Start) { |
1343 | assert((Kind == RecurKind::FAdd || Kind == RecurKind::FMulAdd) && |
1344 | "Unexpected reduction kind" ); |
1345 | assert(Src->getType()->isVectorTy() && "Expected a vector type" ); |
1346 | assert(!Start->getType()->isVectorTy() && "Expected a scalar type" ); |
1347 | |
1348 | return B.CreateFAddReduce(Acc: Start, Src); |
1349 | } |
1350 | |
1351 | Value *llvm::createOrderedReduction(IRBuilderBase &Builder, RecurKind Kind, |
1352 | Value *Src, Value *Start, Value *Mask, |
1353 | Value *EVL) { |
1354 | assert((Kind == RecurKind::FAdd || Kind == RecurKind::FMulAdd) && |
1355 | "Unexpected reduction kind" ); |
1356 | assert(Src->getType()->isVectorTy() && "Expected a vector type" ); |
1357 | assert(!Start->getType()->isVectorTy() && "Expected a scalar type" ); |
1358 | |
1359 | Intrinsic::ID Id = getReductionIntrinsicID(RK: RecurKind::FAdd); |
1360 | auto VPID = VPIntrinsic::getForIntrinsic(Id); |
1361 | assert(VPReductionIntrinsic::isVPReduction(VPID) && |
1362 | "No VPIntrinsic for this reduction" ); |
1363 | auto *EltTy = cast<VectorType>(Val: Src->getType())->getElementType(); |
1364 | Value *Ops[] = {Start, Src, Mask, EVL}; |
1365 | return Builder.CreateIntrinsic(RetTy: EltTy, ID: VPID, Args: Ops); |
1366 | } |
1367 | |
1368 | void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue, |
1369 | bool IncludeWrapFlags) { |
1370 | auto *VecOp = dyn_cast<Instruction>(Val: I); |
1371 | if (!VecOp) |
1372 | return; |
1373 | auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(Val: VL[0]) |
1374 | : dyn_cast<Instruction>(Val: OpValue); |
1375 | if (!Intersection) |
1376 | return; |
1377 | const unsigned Opcode = Intersection->getOpcode(); |
1378 | VecOp->copyIRFlags(V: Intersection, IncludeWrapFlags); |
1379 | for (auto *V : VL) { |
1380 | auto *Instr = dyn_cast<Instruction>(Val: V); |
1381 | if (!Instr) |
1382 | continue; |
1383 | if (OpValue == nullptr || Opcode == Instr->getOpcode()) |
1384 | VecOp->andIRFlags(V); |
1385 | } |
1386 | } |
1387 | |
1388 | bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L, |
1389 | ScalarEvolution &SE) { |
1390 | const SCEV *Zero = SE.getZero(Ty: S->getType()); |
1391 | return SE.isAvailableAtLoopEntry(S, L) && |
1392 | SE.isLoopEntryGuardedByCond(L, Pred: ICmpInst::ICMP_SLT, LHS: S, RHS: Zero); |
1393 | } |
1394 | |
1395 | bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L, |
1396 | ScalarEvolution &SE) { |
1397 | const SCEV *Zero = SE.getZero(Ty: S->getType()); |
1398 | return SE.isAvailableAtLoopEntry(S, L) && |
1399 | SE.isLoopEntryGuardedByCond(L, Pred: ICmpInst::ICMP_SGE, LHS: S, RHS: Zero); |
1400 | } |
1401 | |
1402 | bool llvm::isKnownPositiveInLoop(const SCEV *S, const Loop *L, |
1403 | ScalarEvolution &SE) { |
1404 | const SCEV *Zero = SE.getZero(Ty: S->getType()); |
1405 | return SE.isAvailableAtLoopEntry(S, L) && |
1406 | SE.isLoopEntryGuardedByCond(L, Pred: ICmpInst::ICMP_SGT, LHS: S, RHS: Zero); |
1407 | } |
1408 | |
1409 | bool llvm::isKnownNonPositiveInLoop(const SCEV *S, const Loop *L, |
1410 | ScalarEvolution &SE) { |
1411 | const SCEV *Zero = SE.getZero(Ty: S->getType()); |
1412 | return SE.isAvailableAtLoopEntry(S, L) && |
1413 | SE.isLoopEntryGuardedByCond(L, Pred: ICmpInst::ICMP_SLE, LHS: S, RHS: Zero); |
1414 | } |
1415 | |
1416 | bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, |
1417 | bool Signed) { |
1418 | unsigned BitWidth = cast<IntegerType>(Val: S->getType())->getBitWidth(); |
1419 | APInt Min = Signed ? APInt::getSignedMinValue(numBits: BitWidth) : |
1420 | APInt::getMinValue(numBits: BitWidth); |
1421 | auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; |
1422 | return SE.isAvailableAtLoopEntry(S, L) && |
1423 | SE.isLoopEntryGuardedByCond(L, Pred: Predicate, LHS: S, |
1424 | RHS: SE.getConstant(Val: Min)); |
1425 | } |
1426 | |
1427 | bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, |
1428 | bool Signed) { |
1429 | unsigned BitWidth = cast<IntegerType>(Val: S->getType())->getBitWidth(); |
1430 | APInt Max = Signed ? APInt::getSignedMaxValue(numBits: BitWidth) : |
1431 | APInt::getMaxValue(numBits: BitWidth); |
1432 | auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; |
1433 | return SE.isAvailableAtLoopEntry(S, L) && |
1434 | SE.isLoopEntryGuardedByCond(L, Pred: Predicate, LHS: S, |
1435 | RHS: SE.getConstant(Val: Max)); |
1436 | } |
1437 | |
1438 | //===----------------------------------------------------------------------===// |
1439 | // rewriteLoopExitValues - Optimize IV users outside the loop. |
1440 | // As a side effect, reduces the amount of IV processing within the loop. |
1441 | //===----------------------------------------------------------------------===// |
1442 | |
1443 | static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) { |
1444 | SmallPtrSet<const Instruction *, 8> Visited; |
1445 | SmallVector<const Instruction *, 8> WorkList; |
1446 | Visited.insert(Ptr: I); |
1447 | WorkList.push_back(Elt: I); |
1448 | while (!WorkList.empty()) { |
1449 | const Instruction *Curr = WorkList.pop_back_val(); |
1450 | // This use is outside the loop, nothing to do. |
1451 | if (!L->contains(Inst: Curr)) |
1452 | continue; |
1453 | // Do we assume it is a "hard" use which will not be eliminated easily? |
1454 | if (Curr->mayHaveSideEffects()) |
1455 | return true; |
1456 | // Otherwise, add all its users to worklist. |
1457 | for (const auto *U : Curr->users()) { |
1458 | auto *UI = cast<Instruction>(Val: U); |
1459 | if (Visited.insert(Ptr: UI).second) |
1460 | WorkList.push_back(Elt: UI); |
1461 | } |
1462 | } |
1463 | return false; |
1464 | } |
1465 | |
1466 | // Collect information about PHI nodes which can be transformed in |
1467 | // rewriteLoopExitValues. |
1468 | struct RewritePhi { |
1469 | PHINode *PN; // For which PHI node is this replacement? |
1470 | unsigned Ith; // For which incoming value? |
1471 | const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting. |
1472 | Instruction *ExpansionPoint; // Where we'd like to expand that SCEV? |
1473 | bool HighCost; // Is this expansion a high-cost? |
1474 | |
1475 | RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt, |
1476 | bool H) |
1477 | : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt), |
1478 | HighCost(H) {} |
1479 | }; |
1480 | |
1481 | // Check whether it is possible to delete the loop after rewriting exit |
1482 | // value. If it is possible, ignore ReplaceExitValue and do rewriting |
1483 | // aggressively. |
1484 | static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) { |
1485 | BasicBlock * = L->getLoopPreheader(); |
1486 | // If there is no preheader, the loop will not be deleted. |
1487 | if (!Preheader) |
1488 | return false; |
1489 | |
1490 | // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1. |
1491 | // We obviate multiple ExitingBlocks case for simplicity. |
1492 | // TODO: If we see testcase with multiple ExitingBlocks can be deleted |
1493 | // after exit value rewriting, we can enhance the logic here. |
1494 | SmallVector<BasicBlock *, 4> ExitingBlocks; |
1495 | L->getExitingBlocks(ExitingBlocks); |
1496 | SmallVector<BasicBlock *, 8> ExitBlocks; |
1497 | L->getUniqueExitBlocks(ExitBlocks); |
1498 | if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1) |
1499 | return false; |
1500 | |
1501 | BasicBlock *ExitBlock = ExitBlocks[0]; |
1502 | BasicBlock::iterator BI = ExitBlock->begin(); |
1503 | while (PHINode *P = dyn_cast<PHINode>(Val&: BI)) { |
1504 | Value *Incoming = P->getIncomingValueForBlock(BB: ExitingBlocks[0]); |
1505 | |
1506 | // If the Incoming value of P is found in RewritePhiSet, we know it |
1507 | // could be rewritten to use a loop invariant value in transformation |
1508 | // phase later. Skip it in the loop invariant check below. |
1509 | bool found = false; |
1510 | for (const RewritePhi &Phi : RewritePhiSet) { |
1511 | unsigned i = Phi.Ith; |
1512 | if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) { |
1513 | found = true; |
1514 | break; |
1515 | } |
1516 | } |
1517 | |
1518 | Instruction *I; |
1519 | if (!found && (I = dyn_cast<Instruction>(Val: Incoming))) |
1520 | if (!L->hasLoopInvariantOperands(I)) |
1521 | return false; |
1522 | |
1523 | ++BI; |
1524 | } |
1525 | |
1526 | for (auto *BB : L->blocks()) |
1527 | if (llvm::any_of(Range&: *BB, P: [](Instruction &I) { |
1528 | return I.mayHaveSideEffects(); |
1529 | })) |
1530 | return false; |
1531 | |
1532 | return true; |
1533 | } |
1534 | |
1535 | /// Checks if it is safe to call InductionDescriptor::isInductionPHI for \p Phi, |
1536 | /// and returns true if this Phi is an induction phi in the loop. When |
1537 | /// isInductionPHI returns true, \p ID will be also be set by isInductionPHI. |
1538 | static bool checkIsIndPhi(PHINode *Phi, Loop *L, ScalarEvolution *SE, |
1539 | InductionDescriptor &ID) { |
1540 | if (!Phi) |
1541 | return false; |
1542 | if (!L->getLoopPreheader()) |
1543 | return false; |
1544 | if (Phi->getParent() != L->getHeader()) |
1545 | return false; |
1546 | return InductionDescriptor::isInductionPHI(Phi, L, SE, D&: ID); |
1547 | } |
1548 | |
1549 | int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI, |
1550 | ScalarEvolution *SE, |
1551 | const TargetTransformInfo *TTI, |
1552 | SCEVExpander &Rewriter, DominatorTree *DT, |
1553 | ReplaceExitVal ReplaceExitValue, |
1554 | SmallVector<WeakTrackingVH, 16> &DeadInsts) { |
1555 | // Check a pre-condition. |
1556 | assert(L->isRecursivelyLCSSAForm(*DT, *LI) && |
1557 | "Indvars did not preserve LCSSA!" ); |
1558 | |
1559 | SmallVector<BasicBlock*, 8> ExitBlocks; |
1560 | L->getUniqueExitBlocks(ExitBlocks); |
1561 | |
1562 | SmallVector<RewritePhi, 8> RewritePhiSet; |
1563 | // Find all values that are computed inside the loop, but used outside of it. |
1564 | // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan |
1565 | // the exit blocks of the loop to find them. |
1566 | for (BasicBlock *ExitBB : ExitBlocks) { |
1567 | // If there are no PHI nodes in this exit block, then no values defined |
1568 | // inside the loop are used on this path, skip it. |
1569 | PHINode *PN = dyn_cast<PHINode>(Val: ExitBB->begin()); |
1570 | if (!PN) continue; |
1571 | |
1572 | unsigned NumPreds = PN->getNumIncomingValues(); |
1573 | |
1574 | // Iterate over all of the PHI nodes. |
1575 | BasicBlock::iterator BBI = ExitBB->begin(); |
1576 | while ((PN = dyn_cast<PHINode>(Val: BBI++))) { |
1577 | if (PN->use_empty()) |
1578 | continue; // dead use, don't replace it |
1579 | |
1580 | if (!SE->isSCEVable(Ty: PN->getType())) |
1581 | continue; |
1582 | |
1583 | // Iterate over all of the values in all the PHI nodes. |
1584 | for (unsigned i = 0; i != NumPreds; ++i) { |
1585 | // If the value being merged in is not integer or is not defined |
1586 | // in the loop, skip it. |
1587 | Value *InVal = PN->getIncomingValue(i); |
1588 | if (!isa<Instruction>(Val: InVal)) |
1589 | continue; |
1590 | |
1591 | // If this pred is for a subloop, not L itself, skip it. |
1592 | if (LI->getLoopFor(BB: PN->getIncomingBlock(i)) != L) |
1593 | continue; // The Block is in a subloop, skip it. |
1594 | |
1595 | // Check that InVal is defined in the loop. |
1596 | Instruction *Inst = cast<Instruction>(Val: InVal); |
1597 | if (!L->contains(Inst)) |
1598 | continue; |
1599 | |
1600 | // Find exit values which are induction variables in the loop, and are |
1601 | // unused in the loop, with the only use being the exit block PhiNode, |
1602 | // and the induction variable update binary operator. |
1603 | // The exit value can be replaced with the final value when it is cheap |
1604 | // to do so. |
1605 | if (ReplaceExitValue == UnusedIndVarInLoop) { |
1606 | InductionDescriptor ID; |
1607 | PHINode *IndPhi = dyn_cast<PHINode>(Val: Inst); |
1608 | if (IndPhi) { |
1609 | if (!checkIsIndPhi(Phi: IndPhi, L, SE, ID)) |
1610 | continue; |
1611 | // This is an induction PHI. Check that the only users are PHI |
1612 | // nodes, and induction variable update binary operators. |
1613 | if (llvm::any_of(Range: Inst->users(), P: [&](User *U) { |
1614 | if (!isa<PHINode>(Val: U) && !isa<BinaryOperator>(Val: U)) |
1615 | return true; |
1616 | BinaryOperator *B = dyn_cast<BinaryOperator>(Val: U); |
1617 | if (B && B != ID.getInductionBinOp()) |
1618 | return true; |
1619 | return false; |
1620 | })) |
1621 | continue; |
1622 | } else { |
1623 | // If it is not an induction phi, it must be an induction update |
1624 | // binary operator with an induction phi user. |
1625 | BinaryOperator *B = dyn_cast<BinaryOperator>(Val: Inst); |
1626 | if (!B) |
1627 | continue; |
1628 | if (llvm::any_of(Range: Inst->users(), P: [&](User *U) { |
1629 | PHINode *Phi = dyn_cast<PHINode>(Val: U); |
1630 | if (Phi != PN && !checkIsIndPhi(Phi, L, SE, ID)) |
1631 | return true; |
1632 | return false; |
1633 | })) |
1634 | continue; |
1635 | if (B != ID.getInductionBinOp()) |
1636 | continue; |
1637 | } |
1638 | } |
1639 | |
1640 | // Okay, this instruction has a user outside of the current loop |
1641 | // and varies predictably *inside* the loop. Evaluate the value it |
1642 | // contains when the loop exits, if possible. We prefer to start with |
1643 | // expressions which are true for all exits (so as to maximize |
1644 | // expression reuse by the SCEVExpander), but resort to per-exit |
1645 | // evaluation if that fails. |
1646 | const SCEV *ExitValue = SE->getSCEVAtScope(V: Inst, L: L->getParentLoop()); |
1647 | if (isa<SCEVCouldNotCompute>(Val: ExitValue) || |
1648 | !SE->isLoopInvariant(S: ExitValue, L) || |
1649 | !Rewriter.isSafeToExpand(S: ExitValue)) { |
1650 | // TODO: This should probably be sunk into SCEV in some way; maybe a |
1651 | // getSCEVForExit(SCEV*, L, ExitingBB)? It can be generalized for |
1652 | // most SCEV expressions and other recurrence types (e.g. shift |
1653 | // recurrences). Is there existing code we can reuse? |
1654 | const SCEV *ExitCount = SE->getExitCount(L, ExitingBlock: PN->getIncomingBlock(i)); |
1655 | if (isa<SCEVCouldNotCompute>(Val: ExitCount)) |
1656 | continue; |
1657 | if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: Inst))) |
1658 | if (AddRec->getLoop() == L) |
1659 | ExitValue = AddRec->evaluateAtIteration(It: ExitCount, SE&: *SE); |
1660 | if (isa<SCEVCouldNotCompute>(Val: ExitValue) || |
1661 | !SE->isLoopInvariant(S: ExitValue, L) || |
1662 | !Rewriter.isSafeToExpand(S: ExitValue)) |
1663 | continue; |
1664 | } |
1665 | |
1666 | // Computing the value outside of the loop brings no benefit if it is |
1667 | // definitely used inside the loop in a way which can not be optimized |
1668 | // away. Avoid doing so unless we know we have a value which computes |
1669 | // the ExitValue already. TODO: This should be merged into SCEV |
1670 | // expander to leverage its knowledge of existing expressions. |
1671 | if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(Val: ExitValue) && |
1672 | !isa<SCEVUnknown>(Val: ExitValue) && hasHardUserWithinLoop(L, I: Inst)) |
1673 | continue; |
1674 | |
1675 | // Check if expansions of this SCEV would count as being high cost. |
1676 | bool HighCost = Rewriter.isHighCostExpansion( |
1677 | Exprs: ExitValue, L, Budget: SCEVCheapExpansionBudget, TTI, At: Inst); |
1678 | |
1679 | // Note that we must not perform expansions until after |
1680 | // we query *all* the costs, because if we perform temporary expansion |
1681 | // inbetween, one that we might not intend to keep, said expansion |
1682 | // *may* affect cost calculation of the next SCEV's we'll query, |
1683 | // and next SCEV may errneously get smaller cost. |
1684 | |
1685 | // Collect all the candidate PHINodes to be rewritten. |
1686 | Instruction *InsertPt = |
1687 | (isa<PHINode>(Val: Inst) || isa<LandingPadInst>(Val: Inst)) ? |
1688 | &*Inst->getParent()->getFirstInsertionPt() : Inst; |
1689 | RewritePhiSet.emplace_back(Args&: PN, Args&: i, Args&: ExitValue, Args&: InsertPt, Args&: HighCost); |
1690 | } |
1691 | } |
1692 | } |
1693 | |
1694 | // TODO: evaluate whether it is beneficial to change how we calculate |
1695 | // high-cost: if we have SCEV 'A' which we know we will expand, should we |
1696 | // calculate the cost of other SCEV's after expanding SCEV 'A', thus |
1697 | // potentially giving cost bonus to those other SCEV's? |
1698 | |
1699 | bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet); |
1700 | int NumReplaced = 0; |
1701 | |
1702 | // Transformation. |
1703 | for (const RewritePhi &Phi : RewritePhiSet) { |
1704 | PHINode *PN = Phi.PN; |
1705 | |
1706 | // Only do the rewrite when the ExitValue can be expanded cheaply. |
1707 | // If LoopCanBeDel is true, rewrite exit value aggressively. |
1708 | if ((ReplaceExitValue == OnlyCheapRepl || |
1709 | ReplaceExitValue == UnusedIndVarInLoop) && |
1710 | !LoopCanBeDel && Phi.HighCost) |
1711 | continue; |
1712 | |
1713 | Value *ExitVal = Rewriter.expandCodeFor( |
1714 | SH: Phi.ExpansionSCEV, Ty: Phi.PN->getType(), I: Phi.ExpansionPoint); |
1715 | |
1716 | LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = " << *ExitVal |
1717 | << '\n' |
1718 | << " LoopVal = " << *(Phi.ExpansionPoint) << "\n" ); |
1719 | |
1720 | #ifndef NDEBUG |
1721 | // If we reuse an instruction from a loop which is neither L nor one of |
1722 | // its containing loops, we end up breaking LCSSA form for this loop by |
1723 | // creating a new use of its instruction. |
1724 | if (auto *ExitInsn = dyn_cast<Instruction>(ExitVal)) |
1725 | if (auto *EVL = LI->getLoopFor(ExitInsn->getParent())) |
1726 | if (EVL != L) |
1727 | assert(EVL->contains(L) && "LCSSA breach detected!" ); |
1728 | #endif |
1729 | |
1730 | NumReplaced++; |
1731 | Instruction *Inst = cast<Instruction>(Val: PN->getIncomingValue(i: Phi.Ith)); |
1732 | PN->setIncomingValue(i: Phi.Ith, V: ExitVal); |
1733 | // It's necessary to tell ScalarEvolution about this explicitly so that |
1734 | // it can walk the def-use list and forget all SCEVs, as it may not be |
1735 | // watching the PHI itself. Once the new exit value is in place, there |
1736 | // may not be a def-use connection between the loop and every instruction |
1737 | // which got a SCEVAddRecExpr for that loop. |
1738 | SE->forgetValue(V: PN); |
1739 | |
1740 | // If this instruction is dead now, delete it. Don't do it now to avoid |
1741 | // invalidating iterators. |
1742 | if (isInstructionTriviallyDead(I: Inst, TLI)) |
1743 | DeadInsts.push_back(Elt: Inst); |
1744 | |
1745 | // Replace PN with ExitVal if that is legal and does not break LCSSA. |
1746 | if (PN->getNumIncomingValues() == 1 && |
1747 | LI->replacementPreservesLCSSAForm(From: PN, To: ExitVal)) { |
1748 | PN->replaceAllUsesWith(V: ExitVal); |
1749 | PN->eraseFromParent(); |
1750 | } |
1751 | } |
1752 | |
1753 | // The insertion point instruction may have been deleted; clear it out |
1754 | // so that the rewriter doesn't trip over it later. |
1755 | Rewriter.clearInsertPoint(); |
1756 | return NumReplaced; |
1757 | } |
1758 | |
1759 | /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for |
1760 | /// \p OrigLoop. |
1761 | void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop, |
1762 | Loop *RemainderLoop, uint64_t UF) { |
1763 | assert(UF > 0 && "Zero unrolled factor is not supported" ); |
1764 | assert(UnrolledLoop != RemainderLoop && |
1765 | "Unrolled and Remainder loops are expected to distinct" ); |
1766 | |
1767 | // Get number of iterations in the original scalar loop. |
1768 | unsigned OrigLoopInvocationWeight = 0; |
1769 | std::optional<unsigned> OrigAverageTripCount = |
1770 | getLoopEstimatedTripCount(L: OrigLoop, EstimatedLoopInvocationWeight: &OrigLoopInvocationWeight); |
1771 | if (!OrigAverageTripCount) |
1772 | return; |
1773 | |
1774 | // Calculate number of iterations in unrolled loop. |
1775 | unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF; |
1776 | // Calculate number of iterations for remainder loop. |
1777 | unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF; |
1778 | |
1779 | setLoopEstimatedTripCount(L: UnrolledLoop, EstimatedTripCount: UnrolledAverageTripCount, |
1780 | EstimatedloopInvocationWeight: OrigLoopInvocationWeight); |
1781 | setLoopEstimatedTripCount(L: RemainderLoop, EstimatedTripCount: RemainderAverageTripCount, |
1782 | EstimatedloopInvocationWeight: OrigLoopInvocationWeight); |
1783 | } |
1784 | |
1785 | /// Utility that implements appending of loops onto a worklist. |
1786 | /// Loops are added in preorder (analogous for reverse postorder for trees), |
1787 | /// and the worklist is processed LIFO. |
1788 | template <typename RangeT> |
1789 | void llvm::appendReversedLoopsToWorklist( |
1790 | RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) { |
1791 | // We use an internal worklist to build up the preorder traversal without |
1792 | // recursion. |
1793 | SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist; |
1794 | |
1795 | // We walk the initial sequence of loops in reverse because we generally want |
1796 | // to visit defs before uses and the worklist is LIFO. |
1797 | for (Loop *RootL : Loops) { |
1798 | assert(PreOrderLoops.empty() && "Must start with an empty preorder walk." ); |
1799 | assert(PreOrderWorklist.empty() && |
1800 | "Must start with an empty preorder walk worklist." ); |
1801 | PreOrderWorklist.push_back(Elt: RootL); |
1802 | do { |
1803 | Loop *L = PreOrderWorklist.pop_back_val(); |
1804 | PreOrderWorklist.append(in_start: L->begin(), in_end: L->end()); |
1805 | PreOrderLoops.push_back(Elt: L); |
1806 | } while (!PreOrderWorklist.empty()); |
1807 | |
1808 | Worklist.insert(Input: std::move(PreOrderLoops)); |
1809 | PreOrderLoops.clear(); |
1810 | } |
1811 | } |
1812 | |
1813 | template <typename RangeT> |
1814 | void llvm::appendLoopsToWorklist(RangeT &&Loops, |
1815 | SmallPriorityWorklist<Loop *, 4> &Worklist) { |
1816 | appendReversedLoopsToWorklist(reverse(Loops), Worklist); |
1817 | } |
1818 | |
1819 | template LLVM_EXPORT_TEMPLATE void |
1820 | llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>( |
1821 | ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist); |
1822 | |
1823 | template LLVM_EXPORT_TEMPLATE void |
1824 | llvm::appendLoopsToWorklist<Loop &>(Loop &L, |
1825 | SmallPriorityWorklist<Loop *, 4> &Worklist); |
1826 | |
1827 | void llvm::appendLoopsToWorklist(LoopInfo &LI, |
1828 | SmallPriorityWorklist<Loop *, 4> &Worklist) { |
1829 | appendReversedLoopsToWorklist(Loops&: LI, Worklist); |
1830 | } |
1831 | |
1832 | Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, |
1833 | LoopInfo *LI, LPPassManager *LPM) { |
1834 | Loop &New = *LI->AllocateLoop(); |
1835 | if (PL) |
1836 | PL->addChildLoop(NewChild: &New); |
1837 | else |
1838 | LI->addTopLevelLoop(New: &New); |
1839 | |
1840 | if (LPM) |
1841 | LPM->addLoop(L&: New); |
1842 | |
1843 | // Add all of the blocks in L to the new loop. |
1844 | for (BasicBlock *BB : L->blocks()) |
1845 | if (LI->getLoopFor(BB) == L) |
1846 | New.addBasicBlockToLoop(NewBB: cast<BasicBlock>(Val&: VM[BB]), LI&: *LI); |
1847 | |
1848 | // Add all of the subloops to the new loop. |
1849 | for (Loop *I : *L) |
1850 | cloneLoop(L: I, PL: &New, VM, LI, LPM); |
1851 | |
1852 | return &New; |
1853 | } |
1854 | |
1855 | /// IR Values for the lower and upper bounds of a pointer evolution. We |
1856 | /// need to use value-handles because SCEV expansion can invalidate previously |
1857 | /// expanded values. Thus expansion of a pointer can invalidate the bounds for |
1858 | /// a previous one. |
1859 | struct PointerBounds { |
1860 | TrackingVH<Value> Start; |
1861 | TrackingVH<Value> End; |
1862 | Value *StrideToCheck; |
1863 | }; |
1864 | |
1865 | /// Expand code for the lower and upper bound of the pointer group \p CG |
1866 | /// in \p TheLoop. \return the values for the bounds. |
1867 | static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG, |
1868 | Loop *TheLoop, Instruction *Loc, |
1869 | SCEVExpander &Exp, bool HoistRuntimeChecks) { |
1870 | LLVMContext &Ctx = Loc->getContext(); |
1871 | Type *PtrArithTy = PointerType::get(C&: Ctx, AddressSpace: CG->AddressSpace); |
1872 | |
1873 | Value *Start = nullptr, *End = nullptr; |
1874 | LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n" ); |
1875 | const SCEV *Low = CG->Low, *High = CG->High, *Stride = nullptr; |
1876 | |
1877 | // If the Low and High values are themselves loop-variant, then we may want |
1878 | // to expand the range to include those covered by the outer loop as well. |
1879 | // There is a trade-off here with the advantage being that creating checks |
1880 | // using the expanded range permits the runtime memory checks to be hoisted |
1881 | // out of the outer loop. This reduces the cost of entering the inner loop, |
1882 | // which can be significant for low trip counts. The disadvantage is that |
1883 | // there is a chance we may now never enter the vectorized inner loop, |
1884 | // whereas using a restricted range check could have allowed us to enter at |
1885 | // least once. This is why the behaviour is not currently the default and is |
1886 | // controlled by the parameter 'HoistRuntimeChecks'. |
1887 | if (HoistRuntimeChecks && TheLoop->getParentLoop() && |
1888 | isa<SCEVAddRecExpr>(Val: High) && isa<SCEVAddRecExpr>(Val: Low)) { |
1889 | auto *HighAR = cast<SCEVAddRecExpr>(Val: High); |
1890 | auto *LowAR = cast<SCEVAddRecExpr>(Val: Low); |
1891 | const Loop *OuterLoop = TheLoop->getParentLoop(); |
1892 | ScalarEvolution &SE = *Exp.getSE(); |
1893 | const SCEV *Recur = LowAR->getStepRecurrence(SE); |
1894 | if (Recur == HighAR->getStepRecurrence(SE) && |
1895 | HighAR->getLoop() == OuterLoop && LowAR->getLoop() == OuterLoop) { |
1896 | BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch(); |
1897 | const SCEV *OuterExitCount = SE.getExitCount(L: OuterLoop, ExitingBlock: OuterLoopLatch); |
1898 | if (!isa<SCEVCouldNotCompute>(Val: OuterExitCount) && |
1899 | OuterExitCount->getType()->isIntegerTy()) { |
1900 | const SCEV *NewHigh = |
1901 | cast<SCEVAddRecExpr>(Val: High)->evaluateAtIteration(It: OuterExitCount, SE); |
1902 | if (!isa<SCEVCouldNotCompute>(Val: NewHigh)) { |
1903 | LLVM_DEBUG(dbgs() << "LAA: Expanded RT check for range to include " |
1904 | "outer loop in order to permit hoisting\n" ); |
1905 | High = NewHigh; |
1906 | Low = cast<SCEVAddRecExpr>(Val: Low)->getStart(); |
1907 | // If there is a possibility that the stride is negative then we have |
1908 | // to generate extra checks to ensure the stride is positive. |
1909 | if (!SE.isKnownNonNegative( |
1910 | S: SE.applyLoopGuards(Expr: Recur, L: HighAR->getLoop()))) { |
1911 | Stride = Recur; |
1912 | LLVM_DEBUG(dbgs() << "LAA: ... but need to check stride is " |
1913 | "positive: " |
1914 | << *Stride << '\n'); |
1915 | } |
1916 | } |
1917 | } |
1918 | } |
1919 | } |
1920 | |
1921 | Start = Exp.expandCodeFor(SH: Low, Ty: PtrArithTy, I: Loc); |
1922 | End = Exp.expandCodeFor(SH: High, Ty: PtrArithTy, I: Loc); |
1923 | if (CG->NeedsFreeze) { |
1924 | IRBuilder<> Builder(Loc); |
1925 | Start = Builder.CreateFreeze(V: Start, Name: Start->getName() + ".fr" ); |
1926 | End = Builder.CreateFreeze(V: End, Name: End->getName() + ".fr" ); |
1927 | } |
1928 | Value *StrideVal = |
1929 | Stride ? Exp.expandCodeFor(SH: Stride, Ty: Stride->getType(), I: Loc) : nullptr; |
1930 | LLVM_DEBUG(dbgs() << "Start: " << *Low << " End: " << *High << "\n" ); |
1931 | return {.Start: Start, .End: End, .StrideToCheck: StrideVal}; |
1932 | } |
1933 | |
1934 | /// Turns a collection of checks into a collection of expanded upper and |
1935 | /// lower bounds for both pointers in the check. |
1936 | static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> |
1937 | expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L, |
1938 | Instruction *Loc, SCEVExpander &Exp, bool HoistRuntimeChecks) { |
1939 | SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds; |
1940 | |
1941 | // Here we're relying on the SCEV Expander's cache to only emit code for the |
1942 | // same bounds once. |
1943 | transform(Range: PointerChecks, d_first: std::back_inserter(x&: ChecksWithBounds), |
1944 | F: [&](const RuntimePointerCheck &Check) { |
1945 | PointerBounds First = expandBounds(CG: Check.first, TheLoop: L, Loc, Exp, |
1946 | HoistRuntimeChecks), |
1947 | Second = expandBounds(CG: Check.second, TheLoop: L, Loc, Exp, |
1948 | HoistRuntimeChecks); |
1949 | return std::make_pair(x&: First, y&: Second); |
1950 | }); |
1951 | |
1952 | return ChecksWithBounds; |
1953 | } |
1954 | |
1955 | Value *llvm::addRuntimeChecks( |
1956 | Instruction *Loc, Loop *TheLoop, |
1957 | const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, |
1958 | SCEVExpander &Exp, bool HoistRuntimeChecks) { |
1959 | // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible. |
1960 | // TODO: Pass RtPtrChecking instead of PointerChecks and SE separately, if possible |
1961 | auto ExpandedChecks = |
1962 | expandBounds(PointerChecks, L: TheLoop, Loc, Exp, HoistRuntimeChecks); |
1963 | |
1964 | LLVMContext &Ctx = Loc->getContext(); |
1965 | IRBuilder ChkBuilder(Ctx, InstSimplifyFolder(Loc->getDataLayout())); |
1966 | ChkBuilder.SetInsertPoint(Loc); |
1967 | // Our instructions might fold to a constant. |
1968 | Value *MemoryRuntimeCheck = nullptr; |
1969 | |
1970 | for (const auto &[A, B] : ExpandedChecks) { |
1971 | // Check if two pointers (A and B) conflict where conflict is computed as: |
1972 | // start(A) <= end(B) && start(B) <= end(A) |
1973 | |
1974 | assert((A.Start->getType()->getPointerAddressSpace() == |
1975 | B.End->getType()->getPointerAddressSpace()) && |
1976 | (B.Start->getType()->getPointerAddressSpace() == |
1977 | A.End->getType()->getPointerAddressSpace()) && |
1978 | "Trying to bounds check pointers with different address spaces" ); |
1979 | |
1980 | // [A|B].Start points to the first accessed byte under base [A|B]. |
1981 | // [A|B].End points to the last accessed byte, plus one. |
1982 | // There is no conflict when the intervals are disjoint: |
1983 | // NoConflict = (B.Start >= A.End) || (A.Start >= B.End) |
1984 | // |
1985 | // bound0 = (B.Start < A.End) |
1986 | // bound1 = (A.Start < B.End) |
1987 | // IsConflict = bound0 & bound1 |
1988 | Value *Cmp0 = ChkBuilder.CreateICmpULT(LHS: A.Start, RHS: B.End, Name: "bound0" ); |
1989 | Value *Cmp1 = ChkBuilder.CreateICmpULT(LHS: B.Start, RHS: A.End, Name: "bound1" ); |
1990 | Value *IsConflict = ChkBuilder.CreateAnd(LHS: Cmp0, RHS: Cmp1, Name: "found.conflict" ); |
1991 | if (A.StrideToCheck) { |
1992 | Value *IsNegativeStride = ChkBuilder.CreateICmpSLT( |
1993 | LHS: A.StrideToCheck, RHS: ConstantInt::get(Ty: A.StrideToCheck->getType(), V: 0), |
1994 | Name: "stride.check" ); |
1995 | IsConflict = ChkBuilder.CreateOr(LHS: IsConflict, RHS: IsNegativeStride); |
1996 | } |
1997 | if (B.StrideToCheck) { |
1998 | Value *IsNegativeStride = ChkBuilder.CreateICmpSLT( |
1999 | LHS: B.StrideToCheck, RHS: ConstantInt::get(Ty: B.StrideToCheck->getType(), V: 0), |
2000 | Name: "stride.check" ); |
2001 | IsConflict = ChkBuilder.CreateOr(LHS: IsConflict, RHS: IsNegativeStride); |
2002 | } |
2003 | if (MemoryRuntimeCheck) { |
2004 | IsConflict = |
2005 | ChkBuilder.CreateOr(LHS: MemoryRuntimeCheck, RHS: IsConflict, Name: "conflict.rdx" ); |
2006 | } |
2007 | MemoryRuntimeCheck = IsConflict; |
2008 | } |
2009 | |
2010 | return MemoryRuntimeCheck; |
2011 | } |
2012 | |
2013 | Value *llvm::addDiffRuntimeChecks( |
2014 | Instruction *Loc, ArrayRef<PointerDiffInfo> Checks, SCEVExpander &Expander, |
2015 | function_ref<Value *(IRBuilderBase &, unsigned)> GetVF, unsigned IC) { |
2016 | |
2017 | LLVMContext &Ctx = Loc->getContext(); |
2018 | IRBuilder ChkBuilder(Ctx, InstSimplifyFolder(Loc->getDataLayout())); |
2019 | ChkBuilder.SetInsertPoint(Loc); |
2020 | // Our instructions might fold to a constant. |
2021 | Value *MemoryRuntimeCheck = nullptr; |
2022 | |
2023 | auto &SE = *Expander.getSE(); |
2024 | // Map to keep track of created compares, The key is the pair of operands for |
2025 | // the compare, to allow detecting and re-using redundant compares. |
2026 | DenseMap<std::pair<Value *, Value *>, Value *> SeenCompares; |
2027 | for (const auto &[SrcStart, SinkStart, AccessSize, NeedsFreeze] : Checks) { |
2028 | Type *Ty = SinkStart->getType(); |
2029 | // Compute VF * IC * AccessSize. |
2030 | auto *VFTimesICTimesSize = |
2031 | ChkBuilder.CreateMul(LHS: GetVF(ChkBuilder, Ty->getScalarSizeInBits()), |
2032 | RHS: ConstantInt::get(Ty, V: IC * AccessSize)); |
2033 | Value *Diff = |
2034 | Expander.expandCodeFor(SH: SE.getMinusSCEV(LHS: SinkStart, RHS: SrcStart), Ty, I: Loc); |
2035 | |
2036 | // Check if the same compare has already been created earlier. In that case, |
2037 | // there is no need to check it again. |
2038 | Value *IsConflict = SeenCompares.lookup(Val: {Diff, VFTimesICTimesSize}); |
2039 | if (IsConflict) |
2040 | continue; |
2041 | |
2042 | IsConflict = |
2043 | ChkBuilder.CreateICmpULT(LHS: Diff, RHS: VFTimesICTimesSize, Name: "diff.check" ); |
2044 | SeenCompares.insert(KV: {{Diff, VFTimesICTimesSize}, IsConflict}); |
2045 | if (NeedsFreeze) |
2046 | IsConflict = |
2047 | ChkBuilder.CreateFreeze(V: IsConflict, Name: IsConflict->getName() + ".fr" ); |
2048 | if (MemoryRuntimeCheck) { |
2049 | IsConflict = |
2050 | ChkBuilder.CreateOr(LHS: MemoryRuntimeCheck, RHS: IsConflict, Name: "conflict.rdx" ); |
2051 | } |
2052 | MemoryRuntimeCheck = IsConflict; |
2053 | } |
2054 | |
2055 | return MemoryRuntimeCheck; |
2056 | } |
2057 | |
2058 | std::optional<IVConditionInfo> |
2059 | llvm::hasPartialIVCondition(const Loop &L, unsigned MSSAThreshold, |
2060 | const MemorySSA &MSSA, AAResults &AA) { |
2061 | auto *TI = dyn_cast<BranchInst>(Val: L.getHeader()->getTerminator()); |
2062 | if (!TI || !TI->isConditional()) |
2063 | return {}; |
2064 | |
2065 | auto *CondI = dyn_cast<Instruction>(Val: TI->getCondition()); |
2066 | // The case with the condition outside the loop should already be handled |
2067 | // earlier. |
2068 | // Allow CmpInst and TruncInsts as they may be users of load instructions |
2069 | // and have potential for partial unswitching |
2070 | if (!CondI || !isa<CmpInst, TruncInst>(Val: CondI) || !L.contains(Inst: CondI)) |
2071 | return {}; |
2072 | |
2073 | SmallVector<Instruction *> InstToDuplicate; |
2074 | InstToDuplicate.push_back(Elt: CondI); |
2075 | |
2076 | SmallVector<Value *, 4> WorkList; |
2077 | WorkList.append(in_start: CondI->op_begin(), in_end: CondI->op_end()); |
2078 | |
2079 | SmallVector<MemoryAccess *, 4> AccessesToCheck; |
2080 | SmallVector<MemoryLocation, 4> AccessedLocs; |
2081 | while (!WorkList.empty()) { |
2082 | Instruction *I = dyn_cast<Instruction>(Val: WorkList.pop_back_val()); |
2083 | if (!I || !L.contains(Inst: I)) |
2084 | continue; |
2085 | |
2086 | // TODO: support additional instructions. |
2087 | if (!isa<LoadInst>(Val: I) && !isa<GetElementPtrInst>(Val: I)) |
2088 | return {}; |
2089 | |
2090 | // Do not duplicate volatile and atomic loads. |
2091 | if (auto *LI = dyn_cast<LoadInst>(Val: I)) |
2092 | if (LI->isVolatile() || LI->isAtomic()) |
2093 | return {}; |
2094 | |
2095 | InstToDuplicate.push_back(Elt: I); |
2096 | if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) { |
2097 | if (auto *MemUse = dyn_cast_or_null<MemoryUse>(Val: MA)) { |
2098 | // Queue the defining access to check for alias checks. |
2099 | AccessesToCheck.push_back(Elt: MemUse->getDefiningAccess()); |
2100 | AccessedLocs.push_back(Elt: MemoryLocation::get(Inst: I)); |
2101 | } else { |
2102 | // MemoryDefs may clobber the location or may be atomic memory |
2103 | // operations. Bail out. |
2104 | return {}; |
2105 | } |
2106 | } |
2107 | WorkList.append(in_start: I->op_begin(), in_end: I->op_end()); |
2108 | } |
2109 | |
2110 | if (InstToDuplicate.empty()) |
2111 | return {}; |
2112 | |
2113 | SmallVector<BasicBlock *, 4> ExitingBlocks; |
2114 | L.getExitingBlocks(ExitingBlocks); |
2115 | auto HasNoClobbersOnPath = |
2116 | [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate, |
2117 | MSSAThreshold](BasicBlock *Succ, BasicBlock *, |
2118 | SmallVector<MemoryAccess *, 4> AccessesToCheck) |
2119 | -> std::optional<IVConditionInfo> { |
2120 | IVConditionInfo Info; |
2121 | // First, collect all blocks in the loop that are on a patch from Succ |
2122 | // to the header. |
2123 | SmallVector<BasicBlock *, 4> WorkList; |
2124 | WorkList.push_back(Elt: Succ); |
2125 | WorkList.push_back(Elt: Header); |
2126 | SmallPtrSet<BasicBlock *, 4> Seen; |
2127 | Seen.insert(Ptr: Header); |
2128 | Info.PathIsNoop &= |
2129 | all_of(Range&: *Header, P: [](Instruction &I) { return !I.mayHaveSideEffects(); }); |
2130 | |
2131 | while (!WorkList.empty()) { |
2132 | BasicBlock *Current = WorkList.pop_back_val(); |
2133 | if (!L.contains(BB: Current)) |
2134 | continue; |
2135 | const auto &SeenIns = Seen.insert(Ptr: Current); |
2136 | if (!SeenIns.second) |
2137 | continue; |
2138 | |
2139 | Info.PathIsNoop &= all_of( |
2140 | Range&: *Current, P: [](Instruction &I) { return !I.mayHaveSideEffects(); }); |
2141 | WorkList.append(in_start: succ_begin(BB: Current), in_end: succ_end(BB: Current)); |
2142 | } |
2143 | |
2144 | // Require at least 2 blocks on a path through the loop. This skips |
2145 | // paths that directly exit the loop. |
2146 | if (Seen.size() < 2) |
2147 | return {}; |
2148 | |
2149 | // Next, check if there are any MemoryDefs that are on the path through |
2150 | // the loop (in the Seen set) and they may-alias any of the locations in |
2151 | // AccessedLocs. If that is the case, they may modify the condition and |
2152 | // partial unswitching is not possible. |
2153 | SmallPtrSet<MemoryAccess *, 4> SeenAccesses; |
2154 | while (!AccessesToCheck.empty()) { |
2155 | MemoryAccess *Current = AccessesToCheck.pop_back_val(); |
2156 | auto SeenI = SeenAccesses.insert(Ptr: Current); |
2157 | if (!SeenI.second || !Seen.contains(Ptr: Current->getBlock())) |
2158 | continue; |
2159 | |
2160 | // Bail out if exceeded the threshold. |
2161 | if (SeenAccesses.size() >= MSSAThreshold) |
2162 | return {}; |
2163 | |
2164 | // MemoryUse are read-only accesses. |
2165 | if (isa<MemoryUse>(Val: Current)) |
2166 | continue; |
2167 | |
2168 | // For a MemoryDef, check if is aliases any of the location feeding |
2169 | // the original condition. |
2170 | if (auto *CurrentDef = dyn_cast<MemoryDef>(Val: Current)) { |
2171 | if (any_of(Range&: AccessedLocs, P: [&AA, CurrentDef](MemoryLocation &Loc) { |
2172 | return isModSet( |
2173 | MRI: AA.getModRefInfo(I: CurrentDef->getMemoryInst(), OptLoc: Loc)); |
2174 | })) |
2175 | return {}; |
2176 | } |
2177 | |
2178 | for (Use &U : Current->uses()) |
2179 | AccessesToCheck.push_back(Elt: cast<MemoryAccess>(Val: U.getUser())); |
2180 | } |
2181 | |
2182 | // We could also allow loops with known trip counts without mustprogress, |
2183 | // but ScalarEvolution may not be available. |
2184 | Info.PathIsNoop &= isMustProgress(L: &L); |
2185 | |
2186 | // If the path is considered a no-op so far, check if it reaches a |
2187 | // single exit block without any phis. This ensures no values from the |
2188 | // loop are used outside of the loop. |
2189 | if (Info.PathIsNoop) { |
2190 | for (auto *Exiting : ExitingBlocks) { |
2191 | if (!Seen.contains(Ptr: Exiting)) |
2192 | continue; |
2193 | for (auto *Succ : successors(BB: Exiting)) { |
2194 | if (L.contains(BB: Succ)) |
2195 | continue; |
2196 | |
2197 | Info.PathIsNoop &= Succ->phis().empty() && |
2198 | (!Info.ExitForPath || Info.ExitForPath == Succ); |
2199 | if (!Info.PathIsNoop) |
2200 | break; |
2201 | assert((!Info.ExitForPath || Info.ExitForPath == Succ) && |
2202 | "cannot have multiple exit blocks" ); |
2203 | Info.ExitForPath = Succ; |
2204 | } |
2205 | } |
2206 | } |
2207 | if (!Info.ExitForPath) |
2208 | Info.PathIsNoop = false; |
2209 | |
2210 | Info.InstToDuplicate = InstToDuplicate; |
2211 | return Info; |
2212 | }; |
2213 | |
2214 | // If we branch to the same successor, partial unswitching will not be |
2215 | // beneficial. |
2216 | if (TI->getSuccessor(i: 0) == TI->getSuccessor(i: 1)) |
2217 | return {}; |
2218 | |
2219 | if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(i: 0), L.getHeader(), |
2220 | AccessesToCheck)) { |
2221 | Info->KnownValue = ConstantInt::getTrue(Context&: TI->getContext()); |
2222 | return Info; |
2223 | } |
2224 | if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(i: 1), L.getHeader(), |
2225 | AccessesToCheck)) { |
2226 | Info->KnownValue = ConstantInt::getFalse(Context&: TI->getContext()); |
2227 | return Info; |
2228 | } |
2229 | |
2230 | return {}; |
2231 | } |
2232 | |