1 | //===- StructurizeCFG.cpp -------------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "llvm/Transforms/Scalar/StructurizeCFG.h" |
10 | #include "llvm/ADT/DenseMap.h" |
11 | #include "llvm/ADT/EquivalenceClasses.h" |
12 | #include "llvm/ADT/MapVector.h" |
13 | #include "llvm/ADT/SCCIterator.h" |
14 | #include "llvm/ADT/STLExtras.h" |
15 | #include "llvm/ADT/SmallPtrSet.h" |
16 | #include "llvm/ADT/SmallSet.h" |
17 | #include "llvm/ADT/SmallVector.h" |
18 | #include "llvm/Analysis/InstructionSimplify.h" |
19 | #include "llvm/Analysis/RegionInfo.h" |
20 | #include "llvm/Analysis/RegionIterator.h" |
21 | #include "llvm/Analysis/RegionPass.h" |
22 | #include "llvm/Analysis/UniformityAnalysis.h" |
23 | #include "llvm/IR/BasicBlock.h" |
24 | #include "llvm/IR/CFG.h" |
25 | #include "llvm/IR/Constants.h" |
26 | #include "llvm/IR/Dominators.h" |
27 | #include "llvm/IR/Function.h" |
28 | #include "llvm/IR/InstrTypes.h" |
29 | #include "llvm/IR/Instruction.h" |
30 | #include "llvm/IR/Instructions.h" |
31 | #include "llvm/IR/Metadata.h" |
32 | #include "llvm/IR/PassManager.h" |
33 | #include "llvm/IR/PatternMatch.h" |
34 | #include "llvm/IR/ProfDataUtils.h" |
35 | #include "llvm/IR/Type.h" |
36 | #include "llvm/IR/Use.h" |
37 | #include "llvm/IR/Value.h" |
38 | #include "llvm/IR/ValueHandle.h" |
39 | #include "llvm/InitializePasses.h" |
40 | #include "llvm/Pass.h" |
41 | #include "llvm/Support/Casting.h" |
42 | #include "llvm/Support/CommandLine.h" |
43 | #include "llvm/Support/Debug.h" |
44 | #include "llvm/Support/raw_ostream.h" |
45 | #include "llvm/Transforms/Scalar.h" |
46 | #include "llvm/Transforms/Utils.h" |
47 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
48 | #include "llvm/Transforms/Utils/Local.h" |
49 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
50 | #include <cassert> |
51 | #include <utility> |
52 | |
53 | using namespace llvm; |
54 | using namespace llvm::PatternMatch; |
55 | |
56 | #define DEBUG_TYPE "structurizecfg" |
57 | |
58 | // The name for newly created blocks. |
59 | const char FlowBlockName[] = "Flow" ; |
60 | |
61 | namespace { |
62 | |
63 | static cl::opt<bool> ForceSkipUniformRegions( |
64 | "structurizecfg-skip-uniform-regions" , |
65 | cl::Hidden, |
66 | cl::desc("Force whether the StructurizeCFG pass skips uniform regions" ), |
67 | cl::init(Val: false)); |
68 | |
69 | static cl::opt<bool> |
70 | RelaxedUniformRegions("structurizecfg-relaxed-uniform-regions" , cl::Hidden, |
71 | cl::desc("Allow relaxed uniform region checks" ), |
72 | cl::init(Val: true)); |
73 | |
74 | // Definition of the complex types used in this pass. |
75 | |
76 | using BBValuePair = std::pair<BasicBlock *, Value *>; |
77 | |
78 | using RNVector = SmallVector<RegionNode *, 8>; |
79 | using BBVector = SmallVector<BasicBlock *, 8>; |
80 | using BranchVector = SmallVector<BranchInst *, 8>; |
81 | using BBValueVector = SmallVector<BBValuePair, 2>; |
82 | |
83 | using BBSet = SmallPtrSet<BasicBlock *, 8>; |
84 | |
85 | using PhiMap = MapVector<PHINode *, BBValueVector>; |
86 | using BB2BBVecMap = MapVector<BasicBlock *, BBVector>; |
87 | |
88 | using BBPhiMap = DenseMap<BasicBlock *, PhiMap>; |
89 | |
90 | using MaybeCondBranchWeights = std::optional<class CondBranchWeights>; |
91 | |
92 | class CondBranchWeights { |
93 | uint32_t TrueWeight; |
94 | uint32_t FalseWeight; |
95 | |
96 | CondBranchWeights(uint32_t T, uint32_t F) : TrueWeight(T), FalseWeight(F) {} |
97 | |
98 | public: |
99 | static MaybeCondBranchWeights tryParse(const BranchInst &Br) { |
100 | assert(Br.isConditional()); |
101 | |
102 | uint64_t T, F; |
103 | if (!extractBranchWeights(I: Br, TrueVal&: T, FalseVal&: F)) |
104 | return std::nullopt; |
105 | |
106 | return CondBranchWeights(T, F); |
107 | } |
108 | |
109 | static void setMetadata(BranchInst &Br, |
110 | const MaybeCondBranchWeights &Weights) { |
111 | assert(Br.isConditional()); |
112 | if (!Weights) |
113 | return; |
114 | uint32_t Arr[] = {Weights->TrueWeight, Weights->FalseWeight}; |
115 | setBranchWeights(I&: Br, Weights: Arr, IsExpected: false); |
116 | } |
117 | |
118 | CondBranchWeights invert() const { |
119 | return CondBranchWeights{FalseWeight, TrueWeight}; |
120 | } |
121 | }; |
122 | |
123 | struct PredInfo { |
124 | Value *Pred; |
125 | MaybeCondBranchWeights Weights; |
126 | }; |
127 | |
128 | using BBPredicates = DenseMap<BasicBlock *, PredInfo>; |
129 | using PredMap = DenseMap<BasicBlock *, BBPredicates>; |
130 | using BB2BBMap = DenseMap<BasicBlock *, BasicBlock *>; |
131 | |
132 | // A traits type that is intended to be used in graph algorithms. The graph |
133 | // traits starts at an entry node, and traverses the RegionNodes that are in |
134 | // the Nodes set. |
135 | struct SubGraphTraits { |
136 | using NodeRef = std::pair<RegionNode *, SmallDenseSet<RegionNode *> *>; |
137 | using BaseSuccIterator = GraphTraits<RegionNode *>::ChildIteratorType; |
138 | |
139 | // This wraps a set of Nodes into the iterator, so we know which edges to |
140 | // filter out. |
141 | class WrappedSuccIterator |
142 | : public iterator_adaptor_base< |
143 | WrappedSuccIterator, BaseSuccIterator, |
144 | typename std::iterator_traits<BaseSuccIterator>::iterator_category, |
145 | NodeRef, std::ptrdiff_t, NodeRef *, NodeRef> { |
146 | SmallDenseSet<RegionNode *> *Nodes; |
147 | |
148 | public: |
149 | WrappedSuccIterator(BaseSuccIterator It, SmallDenseSet<RegionNode *> *Nodes) |
150 | : iterator_adaptor_base(It), Nodes(Nodes) {} |
151 | |
152 | NodeRef operator*() const { return {*I, Nodes}; } |
153 | }; |
154 | |
155 | static bool filterAll(const NodeRef &N) { return true; } |
156 | static bool filterSet(const NodeRef &N) { return N.second->count(V: N.first); } |
157 | |
158 | using ChildIteratorType = |
159 | filter_iterator<WrappedSuccIterator, bool (*)(const NodeRef &)>; |
160 | |
161 | static NodeRef getEntryNode(Region *R) { |
162 | return {GraphTraits<Region *>::getEntryNode(R), nullptr}; |
163 | } |
164 | |
165 | static NodeRef getEntryNode(NodeRef N) { return N; } |
166 | |
167 | static iterator_range<ChildIteratorType> children(const NodeRef &N) { |
168 | auto *filter = N.second ? &filterSet : &filterAll; |
169 | return make_filter_range( |
170 | Range: make_range<WrappedSuccIterator>( |
171 | x: {GraphTraits<RegionNode *>::child_begin(N: N.first), N.second}, |
172 | y: {GraphTraits<RegionNode *>::child_end(N: N.first), N.second}), |
173 | Pred: filter); |
174 | } |
175 | |
176 | static ChildIteratorType child_begin(const NodeRef &N) { |
177 | return children(N).begin(); |
178 | } |
179 | |
180 | static ChildIteratorType child_end(const NodeRef &N) { |
181 | return children(N).end(); |
182 | } |
183 | }; |
184 | |
185 | /// Finds the nearest common dominator of a set of BasicBlocks. |
186 | /// |
187 | /// For every BB you add to the set, you can specify whether we "remember" the |
188 | /// block. When you get the common dominator, you can also ask whether it's one |
189 | /// of the blocks we remembered. |
190 | class NearestCommonDominator { |
191 | DominatorTree *DT; |
192 | BasicBlock *Result = nullptr; |
193 | bool ResultIsRemembered = false; |
194 | |
195 | /// Add BB to the resulting dominator. |
196 | void addBlock(BasicBlock *BB, bool Remember) { |
197 | if (!Result) { |
198 | Result = BB; |
199 | ResultIsRemembered = Remember; |
200 | return; |
201 | } |
202 | |
203 | BasicBlock *NewResult = DT->findNearestCommonDominator(A: Result, B: BB); |
204 | if (NewResult != Result) |
205 | ResultIsRemembered = false; |
206 | if (NewResult == BB) |
207 | ResultIsRemembered |= Remember; |
208 | Result = NewResult; |
209 | } |
210 | |
211 | public: |
212 | explicit NearestCommonDominator(DominatorTree *DomTree) : DT(DomTree) {} |
213 | |
214 | void addBlock(BasicBlock *BB) { |
215 | addBlock(BB, /* Remember = */ false); |
216 | } |
217 | |
218 | void addAndRememberBlock(BasicBlock *BB) { |
219 | addBlock(BB, /* Remember = */ true); |
220 | } |
221 | |
222 | /// Get the nearest common dominator of all the BBs added via addBlock() and |
223 | /// addAndRememberBlock(). |
224 | BasicBlock *result() { return Result; } |
225 | |
226 | /// Is the BB returned by getResult() one of the blocks we added to the set |
227 | /// with addAndRememberBlock()? |
228 | bool resultIsRememberedBlock() { return ResultIsRemembered; } |
229 | }; |
230 | |
231 | /// Transforms the control flow graph on one single entry/exit region |
232 | /// at a time. |
233 | /// |
234 | /// After the transform all "If"/"Then"/"Else" style control flow looks like |
235 | /// this: |
236 | /// |
237 | /// \verbatim |
238 | /// 1 |
239 | /// || |
240 | /// | | |
241 | /// 2 | |
242 | /// | / |
243 | /// |/ |
244 | /// 3 |
245 | /// || Where: |
246 | /// | | 1 = "If" block, calculates the condition |
247 | /// 4 | 2 = "Then" subregion, runs if the condition is true |
248 | /// | / 3 = "Flow" blocks, newly inserted flow blocks, rejoins the flow |
249 | /// |/ 4 = "Else" optional subregion, runs if the condition is false |
250 | /// 5 5 = "End" block, also rejoins the control flow |
251 | /// \endverbatim |
252 | /// |
253 | /// Control flow is expressed as a branch where the true exit goes into the |
254 | /// "Then"/"Else" region, while the false exit skips the region |
255 | /// The condition for the optional "Else" region is expressed as a PHI node. |
256 | /// The incoming values of the PHI node are true for the "If" edge and false |
257 | /// for the "Then" edge. |
258 | /// |
259 | /// Additionally to that even complicated loops look like this: |
260 | /// |
261 | /// \verbatim |
262 | /// 1 |
263 | /// || |
264 | /// | | |
265 | /// 2 ^ Where: |
266 | /// | / 1 = "Entry" block |
267 | /// |/ 2 = "Loop" optional subregion, with all exits at "Flow" block |
268 | /// 3 3 = "Flow" block, with back edge to entry block |
269 | /// | |
270 | /// \endverbatim |
271 | /// |
272 | /// The back edge of the "Flow" block is always on the false side of the branch |
273 | /// while the true side continues the general flow. So the loop condition |
274 | /// consist of a network of PHI nodes where the true incoming values expresses |
275 | /// breaks and the false values expresses continue states. |
276 | |
277 | class StructurizeCFG { |
278 | Type *Boolean; |
279 | ConstantInt *BoolTrue; |
280 | ConstantInt *BoolFalse; |
281 | Value *BoolPoison; |
282 | |
283 | Function *Func; |
284 | Region *ParentRegion; |
285 | |
286 | UniformityInfo *UA = nullptr; |
287 | DominatorTree *DT; |
288 | |
289 | SmallVector<RegionNode *, 8> Order; |
290 | BBSet Visited; |
291 | BBSet FlowSet; |
292 | |
293 | SmallVector<WeakVH, 8> AffectedPhis; |
294 | BBPhiMap DeletedPhis; |
295 | BB2BBVecMap AddedPhis; |
296 | |
297 | PredMap Predicates; |
298 | BranchVector Conditions; |
299 | |
300 | BB2BBMap Loops; |
301 | PredMap LoopPreds; |
302 | BranchVector LoopConds; |
303 | |
304 | RegionNode *PrevNode; |
305 | |
306 | void orderNodes(); |
307 | |
308 | void analyzeLoops(RegionNode *N); |
309 | |
310 | PredInfo buildCondition(BranchInst *Term, unsigned Idx, bool Invert); |
311 | |
312 | void gatherPredicates(RegionNode *N); |
313 | |
314 | void collectInfos(); |
315 | |
316 | void insertConditions(bool Loops); |
317 | |
318 | void simplifyConditions(); |
319 | |
320 | void delPhiValues(BasicBlock *From, BasicBlock *To); |
321 | |
322 | void addPhiValues(BasicBlock *From, BasicBlock *To); |
323 | |
324 | void findUndefBlocks(BasicBlock *PHIBlock, |
325 | const SmallSet<BasicBlock *, 8> &Incomings, |
326 | SmallVector<BasicBlock *> &UndefBlks) const; |
327 | |
328 | void mergeIfCompatible(EquivalenceClasses<PHINode *> &PhiClasses, PHINode *A, |
329 | PHINode *B); |
330 | |
331 | void setPhiValues(); |
332 | |
333 | void simplifyAffectedPhis(); |
334 | |
335 | DebugLoc killTerminator(BasicBlock *BB); |
336 | |
337 | void changeExit(RegionNode *Node, BasicBlock *NewExit, |
338 | bool IncludeDominator); |
339 | |
340 | BasicBlock *getNextFlow(BasicBlock *Dominator); |
341 | |
342 | std::pair<BasicBlock *, DebugLoc> needPrefix(bool NeedEmpty); |
343 | |
344 | BasicBlock *needPostfix(BasicBlock *Flow, bool ExitUseAllowed); |
345 | |
346 | void setPrevNode(BasicBlock *BB); |
347 | |
348 | bool dominatesPredicates(BasicBlock *BB, RegionNode *Node); |
349 | |
350 | bool isPredictableTrue(RegionNode *Node); |
351 | |
352 | void wireFlow(bool ExitUseAllowed, BasicBlock *LoopEnd); |
353 | |
354 | void handleLoops(bool ExitUseAllowed, BasicBlock *LoopEnd); |
355 | |
356 | void createFlow(); |
357 | |
358 | void rebuildSSA(); |
359 | |
360 | public: |
361 | void init(Region *R); |
362 | bool run(Region *R, DominatorTree *DT); |
363 | bool makeUniformRegion(Region *R, UniformityInfo &UA); |
364 | }; |
365 | |
366 | class StructurizeCFGLegacyPass : public RegionPass { |
367 | bool SkipUniformRegions; |
368 | |
369 | public: |
370 | static char ID; |
371 | |
372 | explicit StructurizeCFGLegacyPass(bool SkipUniformRegions_ = false) |
373 | : RegionPass(ID), SkipUniformRegions(SkipUniformRegions_) { |
374 | if (ForceSkipUniformRegions.getNumOccurrences()) |
375 | SkipUniformRegions = ForceSkipUniformRegions.getValue(); |
376 | initializeStructurizeCFGLegacyPassPass(*PassRegistry::getPassRegistry()); |
377 | } |
378 | |
379 | bool runOnRegion(Region *R, RGPassManager &RGM) override { |
380 | StructurizeCFG SCFG; |
381 | SCFG.init(R); |
382 | if (SkipUniformRegions) { |
383 | UniformityInfo &UA = |
384 | getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo(); |
385 | if (SCFG.makeUniformRegion(R, UA)) |
386 | return false; |
387 | } |
388 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
389 | return SCFG.run(R, DT); |
390 | } |
391 | |
392 | StringRef getPassName() const override { return "Structurize control flow" ; } |
393 | |
394 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
395 | if (SkipUniformRegions) |
396 | AU.addRequired<UniformityInfoWrapperPass>(); |
397 | AU.addRequired<DominatorTreeWrapperPass>(); |
398 | |
399 | AU.addPreserved<DominatorTreeWrapperPass>(); |
400 | RegionPass::getAnalysisUsage(AU); |
401 | } |
402 | }; |
403 | |
404 | } // end anonymous namespace |
405 | |
406 | char StructurizeCFGLegacyPass::ID = 0; |
407 | |
408 | INITIALIZE_PASS_BEGIN(StructurizeCFGLegacyPass, "structurizecfg" , |
409 | "Structurize the CFG" , false, false) |
410 | INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass) |
411 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
412 | INITIALIZE_PASS_DEPENDENCY(RegionInfoPass) |
413 | INITIALIZE_PASS_END(StructurizeCFGLegacyPass, "structurizecfg" , |
414 | "Structurize the CFG" , false, false) |
415 | |
416 | /// Build up the general order of nodes, by performing a topological sort of the |
417 | /// parent region's nodes, while ensuring that there is no outer cycle node |
418 | /// between any two inner cycle nodes. |
419 | void StructurizeCFG::orderNodes() { |
420 | Order.resize(N: std::distance(first: GraphTraits<Region *>::nodes_begin(R: ParentRegion), |
421 | last: GraphTraits<Region *>::nodes_end(R: ParentRegion))); |
422 | if (Order.empty()) |
423 | return; |
424 | |
425 | SmallDenseSet<RegionNode *> Nodes; |
426 | auto EntryNode = SubGraphTraits::getEntryNode(R: ParentRegion); |
427 | |
428 | // A list of range indices of SCCs in Order, to be processed. |
429 | SmallVector<std::pair<unsigned, unsigned>, 8> WorkList; |
430 | unsigned I = 0, E = Order.size(); |
431 | while (true) { |
432 | // Run through all the SCCs in the subgraph starting with Entry. |
433 | for (auto SCCI = |
434 | scc_iterator<SubGraphTraits::NodeRef, SubGraphTraits>::begin( |
435 | G: EntryNode); |
436 | !SCCI.isAtEnd(); ++SCCI) { |
437 | auto &SCC = *SCCI; |
438 | |
439 | // An SCC up to the size of 2, can be reduced to an entry (the last node), |
440 | // and a possible additional node. Therefore, it is already in order, and |
441 | // there is no need to add it to the work-list. |
442 | unsigned Size = SCC.size(); |
443 | if (Size > 2) |
444 | WorkList.emplace_back(Args&: I, Args: I + Size); |
445 | |
446 | // Add the SCC nodes to the Order array. |
447 | for (const auto &N : SCC) { |
448 | assert(I < E && "SCC size mismatch!" ); |
449 | Order[I++] = N.first; |
450 | } |
451 | } |
452 | assert(I == E && "SCC size mismatch!" ); |
453 | |
454 | // If there are no more SCCs to order, then we are done. |
455 | if (WorkList.empty()) |
456 | break; |
457 | |
458 | std::tie(args&: I, args&: E) = WorkList.pop_back_val(); |
459 | |
460 | // Collect the set of nodes in the SCC's subgraph. These are only the |
461 | // possible child nodes; we do not add the entry (last node) otherwise we |
462 | // will have the same exact SCC all over again. |
463 | Nodes.clear(); |
464 | Nodes.insert(I: Order.begin() + I, E: Order.begin() + E - 1); |
465 | |
466 | // Update the entry node. |
467 | EntryNode.first = Order[E - 1]; |
468 | EntryNode.second = &Nodes; |
469 | } |
470 | } |
471 | |
472 | /// Determine the end of the loops |
473 | void StructurizeCFG::analyzeLoops(RegionNode *N) { |
474 | if (N->isSubRegion()) { |
475 | // Test for exit as back edge |
476 | BasicBlock *Exit = N->getNodeAs<Region>()->getExit(); |
477 | if (Visited.count(Ptr: Exit)) |
478 | Loops[Exit] = N->getEntry(); |
479 | |
480 | } else { |
481 | // Test for successors as back edge |
482 | BasicBlock *BB = N->getNodeAs<BasicBlock>(); |
483 | BranchInst *Term = cast<BranchInst>(Val: BB->getTerminator()); |
484 | |
485 | for (BasicBlock *Succ : Term->successors()) |
486 | if (Visited.count(Ptr: Succ)) |
487 | Loops[Succ] = BB; |
488 | } |
489 | } |
490 | |
491 | /// Build the condition for one edge |
492 | PredInfo StructurizeCFG::buildCondition(BranchInst *Term, unsigned Idx, |
493 | bool Invert) { |
494 | Value *Cond = Invert ? BoolFalse : BoolTrue; |
495 | MaybeCondBranchWeights Weights; |
496 | |
497 | if (Term->isConditional()) { |
498 | Cond = Term->getCondition(); |
499 | Weights = CondBranchWeights::tryParse(Br: *Term); |
500 | |
501 | if (Idx != (unsigned)Invert) { |
502 | Cond = invertCondition(Condition: Cond); |
503 | if (Weights) |
504 | Weights = Weights->invert(); |
505 | } |
506 | } |
507 | return {.Pred: Cond, .Weights: Weights}; |
508 | } |
509 | |
510 | /// Analyze the predecessors of each block and build up predicates |
511 | void StructurizeCFG::gatherPredicates(RegionNode *N) { |
512 | RegionInfo *RI = ParentRegion->getRegionInfo(); |
513 | BasicBlock *BB = N->getEntry(); |
514 | BBPredicates &Pred = Predicates[BB]; |
515 | BBPredicates &LPred = LoopPreds[BB]; |
516 | |
517 | for (BasicBlock *P : predecessors(BB)) { |
518 | // Ignore it if it's a branch from outside into our region entry |
519 | if (!ParentRegion->contains(BB: P)) |
520 | continue; |
521 | |
522 | Region *R = RI->getRegionFor(BB: P); |
523 | if (R == ParentRegion) { |
524 | // It's a top level block in our region |
525 | BranchInst *Term = cast<BranchInst>(Val: P->getTerminator()); |
526 | for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) { |
527 | BasicBlock *Succ = Term->getSuccessor(i); |
528 | if (Succ != BB) |
529 | continue; |
530 | |
531 | if (Visited.count(Ptr: P)) { |
532 | // Normal forward edge |
533 | if (Term->isConditional()) { |
534 | // Try to treat it like an ELSE block |
535 | BasicBlock *Other = Term->getSuccessor(i: !i); |
536 | if (Visited.count(Ptr: Other) && !Loops.count(Val: Other) && |
537 | !Pred.count(Val: Other) && !Pred.count(Val: P)) { |
538 | |
539 | Pred[Other] = {.Pred: BoolFalse, .Weights: std::nullopt}; |
540 | Pred[P] = {.Pred: BoolTrue, .Weights: std::nullopt}; |
541 | continue; |
542 | } |
543 | } |
544 | Pred[P] = buildCondition(Term, Idx: i, Invert: false); |
545 | } else { |
546 | // Back edge |
547 | LPred[P] = buildCondition(Term, Idx: i, Invert: true); |
548 | } |
549 | } |
550 | } else { |
551 | // It's an exit from a sub region |
552 | while (R->getParent() != ParentRegion) |
553 | R = R->getParent(); |
554 | |
555 | // Edge from inside a subregion to its entry, ignore it |
556 | if (*R == *N) |
557 | continue; |
558 | |
559 | BasicBlock *Entry = R->getEntry(); |
560 | if (Visited.count(Ptr: Entry)) |
561 | Pred[Entry] = {.Pred: BoolTrue, .Weights: std::nullopt}; |
562 | else |
563 | LPred[Entry] = {.Pred: BoolFalse, .Weights: std::nullopt}; |
564 | } |
565 | } |
566 | } |
567 | |
568 | /// Collect various loop and predicate infos |
569 | void StructurizeCFG::collectInfos() { |
570 | // Reset predicate |
571 | Predicates.clear(); |
572 | |
573 | // and loop infos |
574 | Loops.clear(); |
575 | LoopPreds.clear(); |
576 | |
577 | // Reset the visited nodes |
578 | Visited.clear(); |
579 | |
580 | for (RegionNode *RN : reverse(C&: Order)) { |
581 | LLVM_DEBUG(dbgs() << "Visiting: " |
582 | << (RN->isSubRegion() ? "SubRegion with entry: " : "" ) |
583 | << RN->getEntry()->getName() << "\n" ); |
584 | |
585 | // Analyze all the conditions leading to a node |
586 | gatherPredicates(N: RN); |
587 | |
588 | // Remember that we've seen this node |
589 | Visited.insert(Ptr: RN->getEntry()); |
590 | |
591 | // Find the last back edges |
592 | analyzeLoops(N: RN); |
593 | } |
594 | } |
595 | |
596 | /// Insert the missing branch conditions |
597 | void StructurizeCFG::insertConditions(bool Loops) { |
598 | BranchVector &Conds = Loops ? LoopConds : Conditions; |
599 | Value *Default = Loops ? BoolTrue : BoolFalse; |
600 | SSAUpdater PhiInserter; |
601 | |
602 | for (BranchInst *Term : Conds) { |
603 | assert(Term->isConditional()); |
604 | |
605 | BasicBlock *Parent = Term->getParent(); |
606 | BasicBlock *SuccTrue = Term->getSuccessor(i: 0); |
607 | BasicBlock *SuccFalse = Term->getSuccessor(i: 1); |
608 | |
609 | PhiInserter.Initialize(Ty: Boolean, Name: "" ); |
610 | PhiInserter.AddAvailableValue(BB: Loops ? SuccFalse : Parent, V: Default); |
611 | |
612 | BBPredicates &Preds = Loops ? LoopPreds[SuccFalse] : Predicates[SuccTrue]; |
613 | |
614 | NearestCommonDominator Dominator(DT); |
615 | Dominator.addBlock(BB: Parent); |
616 | |
617 | PredInfo ParentInfo{.Pred: nullptr, .Weights: std::nullopt}; |
618 | for (auto [BB, PI] : Preds) { |
619 | if (BB == Parent) { |
620 | ParentInfo = PI; |
621 | break; |
622 | } |
623 | PhiInserter.AddAvailableValue(BB, V: PI.Pred); |
624 | Dominator.addAndRememberBlock(BB); |
625 | } |
626 | |
627 | if (ParentInfo.Pred) { |
628 | Term->setCondition(ParentInfo.Pred); |
629 | CondBranchWeights::setMetadata(Br&: *Term, Weights: ParentInfo.Weights); |
630 | } else { |
631 | if (!Dominator.resultIsRememberedBlock()) |
632 | PhiInserter.AddAvailableValue(BB: Dominator.result(), V: Default); |
633 | |
634 | Term->setCondition(PhiInserter.GetValueInMiddleOfBlock(BB: Parent)); |
635 | } |
636 | } |
637 | } |
638 | |
639 | /// Simplify any inverted conditions that were built by buildConditions. |
640 | void StructurizeCFG::simplifyConditions() { |
641 | SmallVector<Instruction *> InstToErase; |
642 | for (auto &I : concat<PredMap::value_type>(Ranges&: Predicates, Ranges&: LoopPreds)) { |
643 | auto &Preds = I.second; |
644 | for (auto [BB, PI] : Preds) { |
645 | Instruction *Inverted; |
646 | if (match(V: PI.Pred, P: m_Not(V: m_OneUse(SubPattern: m_Instruction(I&: Inverted)))) && |
647 | !PI.Pred->use_empty()) { |
648 | if (auto *InvertedCmp = dyn_cast<CmpInst>(Val: Inverted)) { |
649 | InvertedCmp->setPredicate(InvertedCmp->getInversePredicate()); |
650 | PI.Pred->replaceAllUsesWith(V: InvertedCmp); |
651 | InstToErase.push_back(Elt: cast<Instruction>(Val: PI.Pred)); |
652 | } |
653 | } |
654 | } |
655 | } |
656 | for (auto *I : InstToErase) |
657 | I->eraseFromParent(); |
658 | } |
659 | |
660 | /// Remove all PHI values coming from "From" into "To" and remember |
661 | /// them in DeletedPhis |
662 | void StructurizeCFG::delPhiValues(BasicBlock *From, BasicBlock *To) { |
663 | PhiMap &Map = DeletedPhis[To]; |
664 | for (PHINode &Phi : To->phis()) { |
665 | bool Recorded = false; |
666 | while (Phi.getBasicBlockIndex(BB: From) != -1) { |
667 | Value *Deleted = Phi.removeIncomingValue(BB: From, DeletePHIIfEmpty: false); |
668 | Map[&Phi].push_back(Elt: std::make_pair(x&: From, y&: Deleted)); |
669 | if (!Recorded) { |
670 | AffectedPhis.push_back(Elt: &Phi); |
671 | Recorded = true; |
672 | } |
673 | } |
674 | } |
675 | } |
676 | |
677 | /// Add a dummy PHI value as soon as we knew the new predecessor |
678 | void StructurizeCFG::addPhiValues(BasicBlock *From, BasicBlock *To) { |
679 | for (PHINode &Phi : To->phis()) { |
680 | Value *Poison = PoisonValue::get(T: Phi.getType()); |
681 | Phi.addIncoming(V: Poison, BB: From); |
682 | } |
683 | AddedPhis[To].push_back(Elt: From); |
684 | } |
685 | |
686 | /// When we are reconstructing a PHI inside \p PHIBlock with incoming values |
687 | /// from predecessors \p Incomings, we have a chance to mark the available value |
688 | /// from some blocks as undefined. The function will find out all such blocks |
689 | /// and return in \p UndefBlks. |
690 | void StructurizeCFG::findUndefBlocks( |
691 | BasicBlock *PHIBlock, const SmallSet<BasicBlock *, 8> &Incomings, |
692 | SmallVector<BasicBlock *> &UndefBlks) const { |
693 | // We may get a post-structured CFG like below: |
694 | // |
695 | // | P1 |
696 | // |/ |
697 | // F1 |
698 | // |\ |
699 | // | N |
700 | // |/ |
701 | // F2 |
702 | // |\ |
703 | // | P2 |
704 | // |/ |
705 | // F3 |
706 | // |\ |
707 | // B |
708 | // |
709 | // B is the block that has a PHI being reconstructed. P1/P2 are predecessors |
710 | // of B before structurization. F1/F2/F3 are flow blocks inserted during |
711 | // structurization process. Block N is not a predecessor of B before |
712 | // structurization, but are placed between the predecessors(P1/P2) of B after |
713 | // structurization. This usually means that threads went to N never take the |
714 | // path N->F2->F3->B. For example, the threads take the branch F1->N may |
715 | // always take the branch F2->P2. So, when we are reconstructing a PHI |
716 | // originally in B, we can safely say the incoming value from N is undefined. |
717 | SmallSet<BasicBlock *, 8> VisitedBlock; |
718 | SmallVector<BasicBlock *, 8> Stack; |
719 | if (PHIBlock == ParentRegion->getExit()) { |
720 | for (auto P : predecessors(BB: PHIBlock)) { |
721 | if (ParentRegion->contains(BB: P)) |
722 | Stack.push_back(Elt: P); |
723 | } |
724 | } else { |
725 | append_range(C&: Stack, R: predecessors(BB: PHIBlock)); |
726 | } |
727 | |
728 | // Do a backward traversal over the CFG, and stop further searching if |
729 | // the block is not a Flow. If a block is neither flow block nor the |
730 | // incoming predecessor, then the incoming value from the block is |
731 | // undefined value for the PHI being reconstructed. |
732 | while (!Stack.empty()) { |
733 | BasicBlock *Current = Stack.pop_back_val(); |
734 | if (!VisitedBlock.insert(Ptr: Current).second) |
735 | continue; |
736 | |
737 | if (FlowSet.contains(Ptr: Current)) |
738 | llvm::append_range(C&: Stack, R: predecessors(BB: Current)); |
739 | else if (!Incomings.contains(Ptr: Current)) |
740 | UndefBlks.push_back(Elt: Current); |
741 | } |
742 | } |
743 | |
744 | // If two phi nodes have compatible incoming values (for each |
745 | // incoming block, either they have the same incoming value or only one phi |
746 | // node has an incoming value), let them share the merged incoming values. The |
747 | // merge process is guided by the equivalence information from \p PhiClasses. |
748 | // The function will possibly update the incoming values of leader phi in |
749 | // DeletedPhis. |
750 | void StructurizeCFG::mergeIfCompatible( |
751 | EquivalenceClasses<PHINode *> &PhiClasses, PHINode *A, PHINode *B) { |
752 | auto ItA = PhiClasses.findLeader(ECV: PhiClasses.insert(Data: A)); |
753 | auto ItB = PhiClasses.findLeader(ECV: PhiClasses.insert(Data: B)); |
754 | // They are already in the same class, no work needed. |
755 | if (ItA == ItB) |
756 | return; |
757 | |
758 | PHINode *LeaderA = *ItA; |
759 | PHINode *LeaderB = *ItB; |
760 | BBValueVector &IncomingA = DeletedPhis[LeaderA->getParent()][LeaderA]; |
761 | BBValueVector &IncomingB = DeletedPhis[LeaderB->getParent()][LeaderB]; |
762 | |
763 | DenseMap<BasicBlock *, Value *> Mergeable(IncomingA.begin(), IncomingA.end()); |
764 | for (auto [BB, V] : IncomingB) { |
765 | auto BBIt = Mergeable.find(Val: BB); |
766 | if (BBIt != Mergeable.end() && BBIt->second != V) |
767 | return; |
768 | // Either IncomingA does not have this value or IncomingA has the same |
769 | // value. |
770 | Mergeable.insert(KV: {BB, V}); |
771 | } |
772 | |
773 | // Update the incoming value of leaderA. |
774 | IncomingA.assign(in_start: Mergeable.begin(), in_end: Mergeable.end()); |
775 | PhiClasses.unionSets(L1: ItA, L2: ItB); |
776 | } |
777 | |
778 | /// Add the real PHI value as soon as everything is set up |
779 | void StructurizeCFG::setPhiValues() { |
780 | SmallVector<PHINode *, 8> InsertedPhis; |
781 | SSAUpdater Updater(&InsertedPhis); |
782 | DenseMap<BasicBlock *, SmallVector<BasicBlock *>> UndefBlksMap; |
783 | |
784 | // Find phi nodes that have compatible incoming values (either they have |
785 | // the same value for the same block or only one phi node has an incoming |
786 | // value, see example below). We only search again the phi's that are |
787 | // referenced by another phi, which is the case we care about. |
788 | // |
789 | // For example (-- means no incoming value): |
790 | // phi1 : BB1:phi2 BB2:v BB3:-- |
791 | // phi2: BB1:-- BB2:v BB3:w |
792 | // |
793 | // Then we can merge these incoming values and let phi1, phi2 use the |
794 | // same set of incoming values: |
795 | // |
796 | // phi1&phi2: BB1:phi2 BB2:v BB3:w |
797 | // |
798 | // By doing this, phi1 and phi2 would share more intermediate phi nodes. |
799 | // This would help reduce the number of phi nodes during SSA reconstruction |
800 | // and ultimately result in fewer COPY instructions. |
801 | // |
802 | // This should be correct, because if a phi node does not have incoming |
803 | // value from certain block, this means the block is not the predecessor |
804 | // of the parent block, so we actually don't care about its incoming value. |
805 | EquivalenceClasses<PHINode *> PhiClasses; |
806 | for (const auto &[To, From] : AddedPhis) { |
807 | auto OldPhiIt = DeletedPhis.find(Val: To); |
808 | if (OldPhiIt == DeletedPhis.end()) |
809 | continue; |
810 | |
811 | PhiMap &BlkPhis = OldPhiIt->second; |
812 | SmallVector<BasicBlock *> &UndefBlks = UndefBlksMap[To]; |
813 | SmallSet<BasicBlock *, 8> Incomings; |
814 | |
815 | // Get the undefined blocks shared by all the phi nodes. |
816 | if (!BlkPhis.empty()) { |
817 | Incomings.insert_range(R: llvm::make_first_range(c&: BlkPhis.front().second)); |
818 | findUndefBlocks(PHIBlock: To, Incomings, UndefBlks); |
819 | } |
820 | |
821 | for (const auto &[Phi, Incomings] : OldPhiIt->second) { |
822 | SmallVector<PHINode *> IncomingPHIs; |
823 | for (const auto &[BB, V] : Incomings) { |
824 | // First, for each phi, check whether it has incoming value which is |
825 | // another phi. |
826 | if (PHINode *P = dyn_cast<PHINode>(Val: V)) |
827 | IncomingPHIs.push_back(Elt: P); |
828 | } |
829 | |
830 | for (auto *OtherPhi : IncomingPHIs) { |
831 | // Skip phis that are unrelated to the phi reconstruction for now. |
832 | if (!DeletedPhis.contains(Val: OtherPhi->getParent())) |
833 | continue; |
834 | mergeIfCompatible(PhiClasses, A: Phi, B: OtherPhi); |
835 | } |
836 | } |
837 | } |
838 | |
839 | for (const auto &AddedPhi : AddedPhis) { |
840 | BasicBlock *To = AddedPhi.first; |
841 | const BBVector &From = AddedPhi.second; |
842 | |
843 | auto It = DeletedPhis.find(Val: To); |
844 | if (It == DeletedPhis.end()) |
845 | continue; |
846 | |
847 | PhiMap &Map = It->second; |
848 | SmallVector<BasicBlock *> &UndefBlks = UndefBlksMap[To]; |
849 | for (const auto &[Phi, Incoming] : Map) { |
850 | Value *Poison = PoisonValue::get(T: Phi->getType()); |
851 | Updater.Initialize(Ty: Phi->getType(), Name: "" ); |
852 | Updater.AddAvailableValue(BB: &Func->getEntryBlock(), V: Poison); |
853 | Updater.AddAvailableValue(BB: To, V: Poison); |
854 | |
855 | // Use leader phi's incoming if there is. |
856 | auto LeaderIt = PhiClasses.findLeader(V: Phi); |
857 | bool UseIncomingOfLeader = |
858 | LeaderIt != PhiClasses.member_end() && *LeaderIt != Phi; |
859 | const auto &IncomingMap = |
860 | UseIncomingOfLeader ? DeletedPhis[(*LeaderIt)->getParent()][*LeaderIt] |
861 | : Incoming; |
862 | |
863 | SmallVector<BasicBlock *> ConstantPreds; |
864 | for (const auto &[BB, V] : IncomingMap) { |
865 | Updater.AddAvailableValue(BB, V); |
866 | if (isa<Constant>(Val: V)) |
867 | ConstantPreds.push_back(Elt: BB); |
868 | } |
869 | |
870 | for (auto UB : UndefBlks) { |
871 | // If this undef block is dominated by any predecessor(before |
872 | // structurization) of reconstructed PHI with constant incoming value, |
873 | // don't mark the available value as undefined. Setting undef to such |
874 | // block will stop us from getting optimal phi insertion. |
875 | if (any_of(Range&: ConstantPreds, |
876 | P: [&](BasicBlock *CP) { return DT->dominates(A: CP, B: UB); })) |
877 | continue; |
878 | // Maybe already get a value through sharing with other phi nodes. |
879 | if (Updater.HasValueForBlock(BB: UB)) |
880 | continue; |
881 | |
882 | Updater.AddAvailableValue(BB: UB, V: Poison); |
883 | } |
884 | |
885 | for (BasicBlock *FI : From) |
886 | Phi->setIncomingValueForBlock(BB: FI, V: Updater.GetValueAtEndOfBlock(BB: FI)); |
887 | AffectedPhis.push_back(Elt: Phi); |
888 | } |
889 | } |
890 | |
891 | AffectedPhis.append(in_start: InsertedPhis.begin(), in_end: InsertedPhis.end()); |
892 | } |
893 | |
894 | void StructurizeCFG::simplifyAffectedPhis() { |
895 | bool Changed; |
896 | do { |
897 | Changed = false; |
898 | SimplifyQuery Q(Func->getDataLayout()); |
899 | Q.DT = DT; |
900 | // Setting CanUseUndef to true might extend value liveness, set it to false |
901 | // to achieve better register pressure. |
902 | Q.CanUseUndef = false; |
903 | for (WeakVH VH : AffectedPhis) { |
904 | if (auto Phi = dyn_cast_or_null<PHINode>(Val&: VH)) { |
905 | if (auto NewValue = simplifyInstruction(I: Phi, Q)) { |
906 | Phi->replaceAllUsesWith(V: NewValue); |
907 | Phi->eraseFromParent(); |
908 | Changed = true; |
909 | } |
910 | } |
911 | } |
912 | } while (Changed); |
913 | } |
914 | |
915 | /// Remove phi values from all successors and then remove the terminator. |
916 | DebugLoc StructurizeCFG::killTerminator(BasicBlock *BB) { |
917 | Instruction *Term = BB->getTerminator(); |
918 | if (!Term) |
919 | return DebugLoc(); |
920 | |
921 | for (BasicBlock *Succ : successors(BB)) |
922 | delPhiValues(From: BB, To: Succ); |
923 | |
924 | DebugLoc DL = Term->getDebugLoc(); |
925 | Term->eraseFromParent(); |
926 | return DL; |
927 | } |
928 | |
929 | /// Let node exit(s) point to NewExit |
930 | void StructurizeCFG::changeExit(RegionNode *Node, BasicBlock *NewExit, |
931 | bool IncludeDominator) { |
932 | if (Node->isSubRegion()) { |
933 | Region *SubRegion = Node->getNodeAs<Region>(); |
934 | BasicBlock *OldExit = SubRegion->getExit(); |
935 | BasicBlock *Dominator = nullptr; |
936 | |
937 | // Find all the edges from the sub region to the exit. |
938 | // We use make_early_inc_range here because we modify BB's terminator. |
939 | for (BasicBlock *BB : llvm::make_early_inc_range(Range: predecessors(BB: OldExit))) { |
940 | if (!SubRegion->contains(BB)) |
941 | continue; |
942 | |
943 | // Modify the edges to point to the new exit |
944 | delPhiValues(From: BB, To: OldExit); |
945 | BB->getTerminator()->replaceUsesOfWith(From: OldExit, To: NewExit); |
946 | addPhiValues(From: BB, To: NewExit); |
947 | |
948 | // Find the new dominator (if requested) |
949 | if (IncludeDominator) { |
950 | if (!Dominator) |
951 | Dominator = BB; |
952 | else |
953 | Dominator = DT->findNearestCommonDominator(A: Dominator, B: BB); |
954 | } |
955 | } |
956 | |
957 | // Change the dominator (if requested) |
958 | if (Dominator) |
959 | DT->changeImmediateDominator(BB: NewExit, NewBB: Dominator); |
960 | |
961 | // Update the region info |
962 | SubRegion->replaceExit(BB: NewExit); |
963 | } else { |
964 | BasicBlock *BB = Node->getNodeAs<BasicBlock>(); |
965 | DebugLoc DL = killTerminator(BB); |
966 | BranchInst *Br = BranchInst::Create(IfTrue: NewExit, InsertBefore: BB); |
967 | Br->setDebugLoc(DL); |
968 | addPhiValues(From: BB, To: NewExit); |
969 | if (IncludeDominator) |
970 | DT->changeImmediateDominator(BB: NewExit, NewBB: BB); |
971 | } |
972 | } |
973 | |
974 | /// Create a new flow node and update dominator tree and region info |
975 | BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) { |
976 | LLVMContext &Context = Func->getContext(); |
977 | BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() : |
978 | Order.back()->getEntry(); |
979 | BasicBlock *Flow = BasicBlock::Create(Context, Name: FlowBlockName, |
980 | Parent: Func, InsertBefore: Insert); |
981 | FlowSet.insert(Ptr: Flow); |
982 | DT->addNewBlock(BB: Flow, DomBB: Dominator); |
983 | ParentRegion->getRegionInfo()->setRegionFor(BB: Flow, R: ParentRegion); |
984 | return Flow; |
985 | } |
986 | |
987 | /// Create a new or reuse the previous node as flow node. Returns a block and a |
988 | /// debug location to be used for new instructions in that block. |
989 | std::pair<BasicBlock *, DebugLoc> StructurizeCFG::needPrefix(bool NeedEmpty) { |
990 | BasicBlock *Entry = PrevNode->getEntry(); |
991 | |
992 | if (!PrevNode->isSubRegion()) { |
993 | DebugLoc DL = killTerminator(BB: Entry); |
994 | if (!NeedEmpty || Entry->getFirstInsertionPt() == Entry->end()) |
995 | return {Entry, DL}; |
996 | } |
997 | |
998 | // create a new flow node |
999 | BasicBlock *Flow = getNextFlow(Dominator: Entry); |
1000 | |
1001 | // and wire it up |
1002 | changeExit(Node: PrevNode, NewExit: Flow, IncludeDominator: true); |
1003 | PrevNode = ParentRegion->getBBNode(BB: Flow); |
1004 | return {Flow, DebugLoc()}; |
1005 | } |
1006 | |
1007 | /// Returns the region exit if possible, otherwise just a new flow node |
1008 | BasicBlock *StructurizeCFG::needPostfix(BasicBlock *Flow, |
1009 | bool ExitUseAllowed) { |
1010 | if (!Order.empty() || !ExitUseAllowed) |
1011 | return getNextFlow(Dominator: Flow); |
1012 | |
1013 | BasicBlock *Exit = ParentRegion->getExit(); |
1014 | DT->changeImmediateDominator(BB: Exit, NewBB: Flow); |
1015 | addPhiValues(From: Flow, To: Exit); |
1016 | return Exit; |
1017 | } |
1018 | |
1019 | /// Set the previous node |
1020 | void StructurizeCFG::setPrevNode(BasicBlock *BB) { |
1021 | PrevNode = ParentRegion->contains(BB) ? ParentRegion->getBBNode(BB) |
1022 | : nullptr; |
1023 | } |
1024 | |
1025 | /// Does BB dominate all the predicates of Node? |
1026 | bool StructurizeCFG::dominatesPredicates(BasicBlock *BB, RegionNode *Node) { |
1027 | BBPredicates &Preds = Predicates[Node->getEntry()]; |
1028 | return llvm::all_of(Range&: Preds, P: [&](std::pair<BasicBlock *, PredInfo> Pred) { |
1029 | return DT->dominates(A: BB, B: Pred.first); |
1030 | }); |
1031 | } |
1032 | |
1033 | /// Can we predict that this node will always be called? |
1034 | bool StructurizeCFG::isPredictableTrue(RegionNode *Node) { |
1035 | BBPredicates &Preds = Predicates[Node->getEntry()]; |
1036 | bool Dominated = false; |
1037 | |
1038 | // Regionentry is always true |
1039 | if (!PrevNode) |
1040 | return true; |
1041 | |
1042 | for (auto [BB, PI] : Preds) { |
1043 | if (PI.Pred != BoolTrue) |
1044 | return false; |
1045 | |
1046 | if (!Dominated && DT->dominates(A: BB, B: PrevNode->getEntry())) |
1047 | Dominated = true; |
1048 | } |
1049 | |
1050 | // TODO: The dominator check is too strict |
1051 | return Dominated; |
1052 | } |
1053 | |
1054 | /// Take one node from the order vector and wire it up |
1055 | void StructurizeCFG::wireFlow(bool ExitUseAllowed, |
1056 | BasicBlock *LoopEnd) { |
1057 | RegionNode *Node = Order.pop_back_val(); |
1058 | Visited.insert(Ptr: Node->getEntry()); |
1059 | |
1060 | if (isPredictableTrue(Node)) { |
1061 | // Just a linear flow |
1062 | if (PrevNode) { |
1063 | changeExit(Node: PrevNode, NewExit: Node->getEntry(), IncludeDominator: true); |
1064 | } |
1065 | PrevNode = Node; |
1066 | } else { |
1067 | // Insert extra prefix node (or reuse last one) |
1068 | auto [Flow, DL] = needPrefix(NeedEmpty: false); |
1069 | |
1070 | // Insert extra postfix node (or use exit instead) |
1071 | BasicBlock *Entry = Node->getEntry(); |
1072 | BasicBlock *Next = needPostfix(Flow, ExitUseAllowed); |
1073 | |
1074 | // let it point to entry and next block |
1075 | BranchInst *Br = BranchInst::Create(IfTrue: Entry, IfFalse: Next, Cond: BoolPoison, InsertBefore: Flow); |
1076 | Br->setDebugLoc(DL); |
1077 | Conditions.push_back(Elt: Br); |
1078 | addPhiValues(From: Flow, To: Entry); |
1079 | DT->changeImmediateDominator(BB: Entry, NewBB: Flow); |
1080 | |
1081 | PrevNode = Node; |
1082 | while (!Order.empty() && !Visited.count(Ptr: LoopEnd) && |
1083 | dominatesPredicates(BB: Entry, Node: Order.back())) { |
1084 | handleLoops(ExitUseAllowed: false, LoopEnd); |
1085 | } |
1086 | |
1087 | changeExit(Node: PrevNode, NewExit: Next, IncludeDominator: false); |
1088 | setPrevNode(Next); |
1089 | } |
1090 | } |
1091 | |
1092 | void StructurizeCFG::handleLoops(bool ExitUseAllowed, |
1093 | BasicBlock *LoopEnd) { |
1094 | RegionNode *Node = Order.back(); |
1095 | BasicBlock *LoopStart = Node->getEntry(); |
1096 | |
1097 | if (!Loops.count(Val: LoopStart)) { |
1098 | wireFlow(ExitUseAllowed, LoopEnd); |
1099 | return; |
1100 | } |
1101 | |
1102 | if (!isPredictableTrue(Node)) |
1103 | LoopStart = needPrefix(NeedEmpty: true).first; |
1104 | |
1105 | LoopEnd = Loops[Node->getEntry()]; |
1106 | wireFlow(ExitUseAllowed: false, LoopEnd); |
1107 | while (!Visited.count(Ptr: LoopEnd)) { |
1108 | handleLoops(ExitUseAllowed: false, LoopEnd); |
1109 | } |
1110 | |
1111 | assert(LoopStart != &LoopStart->getParent()->getEntryBlock()); |
1112 | |
1113 | // Create an extra loop end node |
1114 | DebugLoc DL; |
1115 | std::tie(args&: LoopEnd, args&: DL) = needPrefix(NeedEmpty: false); |
1116 | BasicBlock *Next = needPostfix(Flow: LoopEnd, ExitUseAllowed); |
1117 | BranchInst *Br = BranchInst::Create(IfTrue: Next, IfFalse: LoopStart, Cond: BoolPoison, InsertBefore: LoopEnd); |
1118 | Br->setDebugLoc(DL); |
1119 | LoopConds.push_back(Elt: Br); |
1120 | addPhiValues(From: LoopEnd, To: LoopStart); |
1121 | setPrevNode(Next); |
1122 | } |
1123 | |
1124 | /// After this function control flow looks like it should be, but |
1125 | /// branches and PHI nodes only have undefined conditions. |
1126 | void StructurizeCFG::createFlow() { |
1127 | BasicBlock *Exit = ParentRegion->getExit(); |
1128 | bool EntryDominatesExit = DT->dominates(A: ParentRegion->getEntry(), B: Exit); |
1129 | |
1130 | AffectedPhis.clear(); |
1131 | DeletedPhis.clear(); |
1132 | AddedPhis.clear(); |
1133 | Conditions.clear(); |
1134 | LoopConds.clear(); |
1135 | |
1136 | PrevNode = nullptr; |
1137 | Visited.clear(); |
1138 | |
1139 | while (!Order.empty()) { |
1140 | handleLoops(ExitUseAllowed: EntryDominatesExit, LoopEnd: nullptr); |
1141 | } |
1142 | |
1143 | if (PrevNode) |
1144 | changeExit(Node: PrevNode, NewExit: Exit, IncludeDominator: EntryDominatesExit); |
1145 | else |
1146 | assert(EntryDominatesExit); |
1147 | } |
1148 | |
1149 | /// Handle a rare case where the disintegrated nodes instructions |
1150 | /// no longer dominate all their uses. Not sure if this is really necessary |
1151 | void StructurizeCFG::rebuildSSA() { |
1152 | SSAUpdater Updater; |
1153 | for (BasicBlock *BB : ParentRegion->blocks()) |
1154 | for (Instruction &I : *BB) { |
1155 | bool Initialized = false; |
1156 | // We may modify the use list as we iterate over it, so we use |
1157 | // make_early_inc_range. |
1158 | for (Use &U : llvm::make_early_inc_range(Range: I.uses())) { |
1159 | Instruction *User = cast<Instruction>(Val: U.getUser()); |
1160 | if (User->getParent() == BB) { |
1161 | continue; |
1162 | } else if (PHINode *UserPN = dyn_cast<PHINode>(Val: User)) { |
1163 | if (UserPN->getIncomingBlock(U) == BB) |
1164 | continue; |
1165 | } |
1166 | |
1167 | if (DT->dominates(Def: &I, User)) |
1168 | continue; |
1169 | |
1170 | if (!Initialized) { |
1171 | Value *Poison = PoisonValue::get(T: I.getType()); |
1172 | Updater.Initialize(Ty: I.getType(), Name: "" ); |
1173 | Updater.AddAvailableValue(BB: &Func->getEntryBlock(), V: Poison); |
1174 | Updater.AddAvailableValue(BB, V: &I); |
1175 | Initialized = true; |
1176 | } |
1177 | Updater.RewriteUseAfterInsertions(U); |
1178 | } |
1179 | } |
1180 | } |
1181 | |
1182 | static bool hasOnlyUniformBranches(Region *R, unsigned UniformMDKindID, |
1183 | const UniformityInfo &UA) { |
1184 | // Bool for if all sub-regions are uniform. |
1185 | bool SubRegionsAreUniform = true; |
1186 | // Count of how many direct children are conditional. |
1187 | unsigned ConditionalDirectChildren = 0; |
1188 | |
1189 | for (auto *E : R->elements()) { |
1190 | if (!E->isSubRegion()) { |
1191 | auto Br = dyn_cast<BranchInst>(Val: E->getEntry()->getTerminator()); |
1192 | if (!Br || !Br->isConditional()) |
1193 | continue; |
1194 | |
1195 | if (!UA.isUniform(I: Br)) |
1196 | return false; |
1197 | |
1198 | // One of our direct children is conditional. |
1199 | ConditionalDirectChildren++; |
1200 | |
1201 | LLVM_DEBUG(dbgs() << "BB: " << Br->getParent()->getName() |
1202 | << " has uniform terminator\n" ); |
1203 | } else { |
1204 | // Explicitly refuse to treat regions as uniform if they have non-uniform |
1205 | // subregions. We cannot rely on UniformityAnalysis for branches in |
1206 | // subregions because those branches may have been removed and re-created, |
1207 | // so we look for our metadata instead. |
1208 | // |
1209 | // Warning: It would be nice to treat regions as uniform based only on |
1210 | // their direct child basic blocks' terminators, regardless of whether |
1211 | // subregions are uniform or not. However, this requires a very careful |
1212 | // look at SIAnnotateControlFlow to make sure nothing breaks there. |
1213 | for (auto *BB : E->getNodeAs<Region>()->blocks()) { |
1214 | auto Br = dyn_cast<BranchInst>(Val: BB->getTerminator()); |
1215 | if (!Br || !Br->isConditional()) |
1216 | continue; |
1217 | |
1218 | if (!Br->getMetadata(KindID: UniformMDKindID)) { |
1219 | // Early exit if we cannot have relaxed uniform regions. |
1220 | if (!RelaxedUniformRegions) |
1221 | return false; |
1222 | |
1223 | SubRegionsAreUniform = false; |
1224 | break; |
1225 | } |
1226 | } |
1227 | } |
1228 | } |
1229 | |
1230 | // Our region is uniform if: |
1231 | // 1. All conditional branches that are direct children are uniform (checked |
1232 | // above). |
1233 | // 2. And either: |
1234 | // a. All sub-regions are uniform. |
1235 | // b. There is one or less conditional branches among the direct children. |
1236 | return SubRegionsAreUniform || (ConditionalDirectChildren <= 1); |
1237 | } |
1238 | |
1239 | void StructurizeCFG::init(Region *R) { |
1240 | LLVMContext &Context = R->getEntry()->getContext(); |
1241 | |
1242 | Boolean = Type::getInt1Ty(C&: Context); |
1243 | BoolTrue = ConstantInt::getTrue(Context); |
1244 | BoolFalse = ConstantInt::getFalse(Context); |
1245 | BoolPoison = PoisonValue::get(T: Boolean); |
1246 | |
1247 | this->UA = nullptr; |
1248 | } |
1249 | |
1250 | bool StructurizeCFG::makeUniformRegion(Region *R, UniformityInfo &UA) { |
1251 | if (R->isTopLevelRegion()) |
1252 | return false; |
1253 | |
1254 | this->UA = &UA; |
1255 | |
1256 | // TODO: We could probably be smarter here with how we handle sub-regions. |
1257 | // We currently rely on the fact that metadata is set by earlier invocations |
1258 | // of the pass on sub-regions, and that this metadata doesn't get lost -- |
1259 | // but we shouldn't rely on metadata for correctness! |
1260 | unsigned UniformMDKindID = |
1261 | R->getEntry()->getContext().getMDKindID(Name: "structurizecfg.uniform" ); |
1262 | |
1263 | if (hasOnlyUniformBranches(R, UniformMDKindID, UA)) { |
1264 | LLVM_DEBUG(dbgs() << "Skipping region with uniform control flow: " << *R |
1265 | << '\n'); |
1266 | |
1267 | // Mark all direct child block terminators as having been treated as |
1268 | // uniform. To account for a possible future in which non-uniform |
1269 | // sub-regions are treated more cleverly, indirect children are not |
1270 | // marked as uniform. |
1271 | MDNode *MD = MDNode::get(Context&: R->getEntry()->getParent()->getContext(), MDs: {}); |
1272 | for (RegionNode *E : R->elements()) { |
1273 | if (E->isSubRegion()) |
1274 | continue; |
1275 | |
1276 | if (Instruction *Term = E->getEntry()->getTerminator()) |
1277 | Term->setMetadata(KindID: UniformMDKindID, Node: MD); |
1278 | } |
1279 | |
1280 | return true; |
1281 | } |
1282 | return false; |
1283 | } |
1284 | |
1285 | /// Run the transformation for each region found |
1286 | bool StructurizeCFG::run(Region *R, DominatorTree *DT) { |
1287 | if (R->isTopLevelRegion()) |
1288 | return false; |
1289 | |
1290 | this->DT = DT; |
1291 | |
1292 | Func = R->getEntry()->getParent(); |
1293 | assert(hasOnlySimpleTerminator(*Func) && "Unsupported block terminator." ); |
1294 | |
1295 | ParentRegion = R; |
1296 | |
1297 | orderNodes(); |
1298 | collectInfos(); |
1299 | createFlow(); |
1300 | insertConditions(Loops: false); |
1301 | insertConditions(Loops: true); |
1302 | setPhiValues(); |
1303 | simplifyConditions(); |
1304 | simplifyAffectedPhis(); |
1305 | rebuildSSA(); |
1306 | |
1307 | // Cleanup |
1308 | Order.clear(); |
1309 | Visited.clear(); |
1310 | DeletedPhis.clear(); |
1311 | AddedPhis.clear(); |
1312 | Predicates.clear(); |
1313 | Conditions.clear(); |
1314 | Loops.clear(); |
1315 | LoopPreds.clear(); |
1316 | LoopConds.clear(); |
1317 | FlowSet.clear(); |
1318 | |
1319 | return true; |
1320 | } |
1321 | |
1322 | Pass *llvm::createStructurizeCFGPass(bool SkipUniformRegions) { |
1323 | return new StructurizeCFGLegacyPass(SkipUniformRegions); |
1324 | } |
1325 | |
1326 | static void addRegionIntoQueue(Region &R, std::vector<Region *> &Regions) { |
1327 | Regions.push_back(x: &R); |
1328 | for (const auto &E : R) |
1329 | addRegionIntoQueue(R&: *E, Regions); |
1330 | } |
1331 | |
1332 | StructurizeCFGPass::StructurizeCFGPass(bool SkipUniformRegions_) |
1333 | : SkipUniformRegions(SkipUniformRegions_) { |
1334 | if (ForceSkipUniformRegions.getNumOccurrences()) |
1335 | SkipUniformRegions = ForceSkipUniformRegions.getValue(); |
1336 | } |
1337 | |
1338 | void StructurizeCFGPass::printPipeline( |
1339 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
1340 | static_cast<PassInfoMixin<StructurizeCFGPass> *>(this)->printPipeline( |
1341 | OS, MapClassName2PassName); |
1342 | if (SkipUniformRegions) |
1343 | OS << "<skip-uniform-regions>" ; |
1344 | } |
1345 | |
1346 | PreservedAnalyses StructurizeCFGPass::run(Function &F, |
1347 | FunctionAnalysisManager &AM) { |
1348 | |
1349 | bool Changed = false; |
1350 | DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F); |
1351 | auto &RI = AM.getResult<RegionInfoAnalysis>(IR&: F); |
1352 | |
1353 | UniformityInfo *UI = nullptr; |
1354 | if (SkipUniformRegions) |
1355 | UI = &AM.getResult<UniformityInfoAnalysis>(IR&: F); |
1356 | |
1357 | std::vector<Region *> Regions; |
1358 | addRegionIntoQueue(R&: *RI.getTopLevelRegion(), Regions); |
1359 | while (!Regions.empty()) { |
1360 | Region *R = Regions.back(); |
1361 | Regions.pop_back(); |
1362 | |
1363 | StructurizeCFG SCFG; |
1364 | SCFG.init(R); |
1365 | |
1366 | if (SkipUniformRegions && SCFG.makeUniformRegion(R, UA&: *UI)) { |
1367 | Changed = true; // May have added metadata. |
1368 | continue; |
1369 | } |
1370 | |
1371 | Changed |= SCFG.run(R, DT); |
1372 | } |
1373 | if (!Changed) |
1374 | return PreservedAnalyses::all(); |
1375 | PreservedAnalyses PA; |
1376 | PA.preserve<DominatorTreeAnalysis>(); |
1377 | return PA; |
1378 | } |
1379 | |