File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/GenericDomTree.h |
Warning: | line 494, column 12 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===// | ||||||
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 pass performs several transformations to transform natural loops into a | ||||||
10 | // simpler form, which makes subsequent analyses and transformations simpler and | ||||||
11 | // more effective. | ||||||
12 | // | ||||||
13 | // Loop pre-header insertion guarantees that there is a single, non-critical | ||||||
14 | // entry edge from outside of the loop to the loop header. This simplifies a | ||||||
15 | // number of analyses and transformations, such as LICM. | ||||||
16 | // | ||||||
17 | // Loop exit-block insertion guarantees that all exit blocks from the loop | ||||||
18 | // (blocks which are outside of the loop that have predecessors inside of the | ||||||
19 | // loop) only have predecessors from inside of the loop (and are thus dominated | ||||||
20 | // by the loop header). This simplifies transformations such as store-sinking | ||||||
21 | // that are built into LICM. | ||||||
22 | // | ||||||
23 | // This pass also guarantees that loops will have exactly one backedge. | ||||||
24 | // | ||||||
25 | // Indirectbr instructions introduce several complications. If the loop | ||||||
26 | // contains or is entered by an indirectbr instruction, it may not be possible | ||||||
27 | // to transform the loop and make these guarantees. Client code should check | ||||||
28 | // that these conditions are true before relying on them. | ||||||
29 | // | ||||||
30 | // Similar complications arise from callbr instructions, particularly in | ||||||
31 | // asm-goto where blockaddress expressions are used. | ||||||
32 | // | ||||||
33 | // Note that the simplifycfg pass will clean up blocks which are split out but | ||||||
34 | // end up being unnecessary, so usage of this pass should not pessimize | ||||||
35 | // generated code. | ||||||
36 | // | ||||||
37 | // This pass obviously modifies the CFG, but updates loop information and | ||||||
38 | // dominator information. | ||||||
39 | // | ||||||
40 | //===----------------------------------------------------------------------===// | ||||||
41 | |||||||
42 | #include "llvm/Transforms/Utils/LoopSimplify.h" | ||||||
43 | #include "llvm/ADT/DepthFirstIterator.h" | ||||||
44 | #include "llvm/ADT/SetOperations.h" | ||||||
45 | #include "llvm/ADT/SetVector.h" | ||||||
46 | #include "llvm/ADT/SmallVector.h" | ||||||
47 | #include "llvm/ADT/Statistic.h" | ||||||
48 | #include "llvm/Analysis/AliasAnalysis.h" | ||||||
49 | #include "llvm/Analysis/AssumptionCache.h" | ||||||
50 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||
51 | #include "llvm/Analysis/BranchProbabilityInfo.h" | ||||||
52 | #include "llvm/Analysis/DependenceAnalysis.h" | ||||||
53 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||
54 | #include "llvm/Analysis/InstructionSimplify.h" | ||||||
55 | #include "llvm/Analysis/LoopInfo.h" | ||||||
56 | #include "llvm/Analysis/MemorySSA.h" | ||||||
57 | #include "llvm/Analysis/MemorySSAUpdater.h" | ||||||
58 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||
59 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" | ||||||
60 | #include "llvm/IR/CFG.h" | ||||||
61 | #include "llvm/IR/Constants.h" | ||||||
62 | #include "llvm/IR/DataLayout.h" | ||||||
63 | #include "llvm/IR/Dominators.h" | ||||||
64 | #include "llvm/IR/Function.h" | ||||||
65 | #include "llvm/IR/Instructions.h" | ||||||
66 | #include "llvm/IR/IntrinsicInst.h" | ||||||
67 | #include "llvm/IR/LLVMContext.h" | ||||||
68 | #include "llvm/IR/Module.h" | ||||||
69 | #include "llvm/IR/Type.h" | ||||||
70 | #include "llvm/InitializePasses.h" | ||||||
71 | #include "llvm/Support/Debug.h" | ||||||
72 | #include "llvm/Support/raw_ostream.h" | ||||||
73 | #include "llvm/Transforms/Utils.h" | ||||||
74 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||
75 | #include "llvm/Transforms/Utils/Local.h" | ||||||
76 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||
77 | using namespace llvm; | ||||||
78 | |||||||
79 | #define DEBUG_TYPE"loop-simplify" "loop-simplify" | ||||||
80 | |||||||
81 | STATISTIC(NumNested , "Number of nested loops split out")static llvm::Statistic NumNested = {"loop-simplify", "NumNested" , "Number of nested loops split out"}; | ||||||
82 | |||||||
83 | // If the block isn't already, move the new block to right after some 'outside | ||||||
84 | // block' block. This prevents the preheader from being placed inside the loop | ||||||
85 | // body, e.g. when the loop hasn't been rotated. | ||||||
86 | static void placeSplitBlockCarefully(BasicBlock *NewBB, | ||||||
87 | SmallVectorImpl<BasicBlock *> &SplitPreds, | ||||||
88 | Loop *L) { | ||||||
89 | // Check to see if NewBB is already well placed. | ||||||
90 | Function::iterator BBI = --NewBB->getIterator(); | ||||||
91 | for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { | ||||||
92 | if (&*BBI == SplitPreds[i]) | ||||||
93 | return; | ||||||
94 | } | ||||||
95 | |||||||
96 | // If it isn't already after an outside block, move it after one. This is | ||||||
97 | // always good as it makes the uncond branch from the outside block into a | ||||||
98 | // fall-through. | ||||||
99 | |||||||
100 | // Figure out *which* outside block to put this after. Prefer an outside | ||||||
101 | // block that neighbors a BB actually in the loop. | ||||||
102 | BasicBlock *FoundBB = nullptr; | ||||||
103 | for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { | ||||||
104 | Function::iterator BBI = SplitPreds[i]->getIterator(); | ||||||
105 | if (++BBI != NewBB->getParent()->end() && L->contains(&*BBI)) { | ||||||
106 | FoundBB = SplitPreds[i]; | ||||||
107 | break; | ||||||
108 | } | ||||||
109 | } | ||||||
110 | |||||||
111 | // If our heuristic for a *good* bb to place this after doesn't find | ||||||
112 | // anything, just pick something. It's likely better than leaving it within | ||||||
113 | // the loop. | ||||||
114 | if (!FoundBB) | ||||||
115 | FoundBB = SplitPreds[0]; | ||||||
116 | NewBB->moveAfter(FoundBB); | ||||||
117 | } | ||||||
118 | |||||||
119 | /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a | ||||||
120 | /// preheader, this method is called to insert one. This method has two phases: | ||||||
121 | /// preheader insertion and analysis updating. | ||||||
122 | /// | ||||||
123 | BasicBlock *llvm::InsertPreheaderForLoop(Loop *L, DominatorTree *DT, | ||||||
124 | LoopInfo *LI, MemorySSAUpdater *MSSAU, | ||||||
125 | bool PreserveLCSSA) { | ||||||
126 | BasicBlock *Header = L->getHeader(); | ||||||
127 | |||||||
128 | // Compute the set of predecessors of the loop that are not in the loop. | ||||||
129 | SmallVector<BasicBlock*, 8> OutsideBlocks; | ||||||
130 | for (BasicBlock *P : predecessors(Header)) { | ||||||
131 | if (!L->contains(P)) { // Coming in from outside the loop? | ||||||
132 | // If the loop is branched to from an indirect terminator, we won't | ||||||
133 | // be able to fully transform the loop, because it prohibits | ||||||
134 | // edge splitting. | ||||||
135 | if (P->getTerminator()->isIndirectTerminator()) | ||||||
136 | return nullptr; | ||||||
137 | |||||||
138 | // Keep track of it. | ||||||
139 | OutsideBlocks.push_back(P); | ||||||
140 | } | ||||||
141 | } | ||||||
142 | |||||||
143 | // Split out the loop pre-header. | ||||||
144 | BasicBlock *PreheaderBB; | ||||||
145 | PreheaderBB = SplitBlockPredecessors(Header, OutsideBlocks, ".preheader", DT, | ||||||
146 | LI, MSSAU, PreserveLCSSA); | ||||||
147 | if (!PreheaderBB) | ||||||
148 | return nullptr; | ||||||
149 | |||||||
150 | LLVM_DEBUG(dbgs() << "LoopSimplify: Creating pre-header "do { } while (false) | ||||||
151 | << PreheaderBB->getName() << "\n")do { } while (false); | ||||||
152 | |||||||
153 | // Make sure that NewBB is put someplace intelligent, which doesn't mess up | ||||||
154 | // code layout too horribly. | ||||||
155 | placeSplitBlockCarefully(PreheaderBB, OutsideBlocks, L); | ||||||
156 | |||||||
157 | return PreheaderBB; | ||||||
158 | } | ||||||
159 | |||||||
160 | /// Add the specified block, and all of its predecessors, to the specified set, | ||||||
161 | /// if it's not already in there. Stop predecessor traversal when we reach | ||||||
162 | /// StopBlock. | ||||||
163 | static void addBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock, | ||||||
164 | SmallPtrSetImpl<BasicBlock *> &Blocks) { | ||||||
165 | SmallVector<BasicBlock *, 8> Worklist; | ||||||
166 | Worklist.push_back(InputBB); | ||||||
167 | do { | ||||||
168 | BasicBlock *BB = Worklist.pop_back_val(); | ||||||
169 | if (Blocks.insert(BB).second && BB != StopBlock) | ||||||
170 | // If BB is not already processed and it is not a stop block then | ||||||
171 | // insert its predecessor in the work list | ||||||
172 | append_range(Worklist, predecessors(BB)); | ||||||
173 | } while (!Worklist.empty()); | ||||||
174 | } | ||||||
175 | |||||||
176 | /// The first part of loop-nestification is to find a PHI node that tells | ||||||
177 | /// us how to partition the loops. | ||||||
178 | static PHINode *findPHIToPartitionLoops(Loop *L, DominatorTree *DT, | ||||||
179 | AssumptionCache *AC) { | ||||||
180 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); | ||||||
181 | for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { | ||||||
182 | PHINode *PN = cast<PHINode>(I); | ||||||
183 | ++I; | ||||||
184 | if (Value *V = SimplifyInstruction(PN, {DL, nullptr, DT, AC})) { | ||||||
185 | // This is a degenerate PHI already, don't modify it! | ||||||
186 | PN->replaceAllUsesWith(V); | ||||||
187 | PN->eraseFromParent(); | ||||||
188 | continue; | ||||||
189 | } | ||||||
190 | |||||||
191 | // Scan this PHI node looking for a use of the PHI node by itself. | ||||||
192 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | ||||||
193 | if (PN->getIncomingValue(i) == PN && | ||||||
194 | L->contains(PN->getIncomingBlock(i))) | ||||||
195 | // We found something tasty to remove. | ||||||
196 | return PN; | ||||||
197 | } | ||||||
198 | return nullptr; | ||||||
199 | } | ||||||
200 | |||||||
201 | /// If this loop has multiple backedges, try to pull one of them out into | ||||||
202 | /// a nested loop. | ||||||
203 | /// | ||||||
204 | /// This is important for code that looks like | ||||||
205 | /// this: | ||||||
206 | /// | ||||||
207 | /// Loop: | ||||||
208 | /// ... | ||||||
209 | /// br cond, Loop, Next | ||||||
210 | /// ... | ||||||
211 | /// br cond2, Loop, Out | ||||||
212 | /// | ||||||
213 | /// To identify this common case, we look at the PHI nodes in the header of the | ||||||
214 | /// loop. PHI nodes with unchanging values on one backedge correspond to values | ||||||
215 | /// that change in the "outer" loop, but not in the "inner" loop. | ||||||
216 | /// | ||||||
217 | /// If we are able to separate out a loop, return the new outer loop that was | ||||||
218 | /// created. | ||||||
219 | /// | ||||||
220 | static Loop *separateNestedLoop(Loop *L, BasicBlock *Preheader, | ||||||
221 | DominatorTree *DT, LoopInfo *LI, | ||||||
222 | ScalarEvolution *SE, bool PreserveLCSSA, | ||||||
223 | AssumptionCache *AC, MemorySSAUpdater *MSSAU) { | ||||||
224 | // Don't try to separate loops without a preheader. | ||||||
225 | if (!Preheader) | ||||||
226 | return nullptr; | ||||||
227 | |||||||
228 | // Treat the presence of convergent functions conservatively. The | ||||||
229 | // transformation is invalid if calls to certain convergent | ||||||
230 | // functions (like an AMDGPU barrier) get included in the resulting | ||||||
231 | // inner loop. But blocks meant for the inner loop will be | ||||||
232 | // identified later at a point where it's too late to abort the | ||||||
233 | // transformation. Also, the convergent attribute is not really | ||||||
234 | // sufficient to express the semantics of functions that are | ||||||
235 | // affected by this transformation. So we choose to back off if such | ||||||
236 | // a function call is present until a better alternative becomes | ||||||
237 | // available. This is similar to the conservative treatment of | ||||||
238 | // convergent function calls in GVNHoist and JumpThreading. | ||||||
239 | for (auto BB : L->blocks()) { | ||||||
240 | for (auto &II : *BB) { | ||||||
241 | if (auto CI = dyn_cast<CallBase>(&II)) { | ||||||
242 | if (CI->isConvergent()) { | ||||||
243 | return nullptr; | ||||||
244 | } | ||||||
245 | } | ||||||
246 | } | ||||||
247 | } | ||||||
248 | |||||||
249 | // The header is not a landing pad; preheader insertion should ensure this. | ||||||
250 | BasicBlock *Header = L->getHeader(); | ||||||
251 | assert(!Header->isEHPad() && "Can't insert backedge to EH pad")((void)0); | ||||||
252 | |||||||
253 | PHINode *PN = findPHIToPartitionLoops(L, DT, AC); | ||||||
254 | if (!PN) return nullptr; // No known way to partition. | ||||||
255 | |||||||
256 | // Pull out all predecessors that have varying values in the loop. This | ||||||
257 | // handles the case when a PHI node has multiple instances of itself as | ||||||
258 | // arguments. | ||||||
259 | SmallVector<BasicBlock*, 8> OuterLoopPreds; | ||||||
260 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | ||||||
261 | if (PN->getIncomingValue(i) != PN || | ||||||
262 | !L->contains(PN->getIncomingBlock(i))) { | ||||||
263 | // We can't split indirect control flow edges. | ||||||
264 | if (PN->getIncomingBlock(i)->getTerminator()->isIndirectTerminator()) | ||||||
265 | return nullptr; | ||||||
266 | OuterLoopPreds.push_back(PN->getIncomingBlock(i)); | ||||||
267 | } | ||||||
268 | } | ||||||
269 | LLVM_DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n")do { } while (false); | ||||||
270 | |||||||
271 | // If ScalarEvolution is around and knows anything about values in | ||||||
272 | // this loop, tell it to forget them, because we're about to | ||||||
273 | // substantially change it. | ||||||
274 | if (SE) | ||||||
275 | SE->forgetLoop(L); | ||||||
276 | |||||||
277 | BasicBlock *NewBB = SplitBlockPredecessors(Header, OuterLoopPreds, ".outer", | ||||||
278 | DT, LI, MSSAU, PreserveLCSSA); | ||||||
279 | |||||||
280 | // Make sure that NewBB is put someplace intelligent, which doesn't mess up | ||||||
281 | // code layout too horribly. | ||||||
282 | placeSplitBlockCarefully(NewBB, OuterLoopPreds, L); | ||||||
283 | |||||||
284 | // Create the new outer loop. | ||||||
285 | Loop *NewOuter = LI->AllocateLoop(); | ||||||
286 | |||||||
287 | // Change the parent loop to use the outer loop as its child now. | ||||||
288 | if (Loop *Parent = L->getParentLoop()) | ||||||
289 | Parent->replaceChildLoopWith(L, NewOuter); | ||||||
290 | else | ||||||
291 | LI->changeTopLevelLoop(L, NewOuter); | ||||||
292 | |||||||
293 | // L is now a subloop of our outer loop. | ||||||
294 | NewOuter->addChildLoop(L); | ||||||
295 | |||||||
296 | for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); | ||||||
297 | I != E; ++I) | ||||||
298 | NewOuter->addBlockEntry(*I); | ||||||
299 | |||||||
300 | // Now reset the header in L, which had been moved by | ||||||
301 | // SplitBlockPredecessors for the outer loop. | ||||||
302 | L->moveToHeader(Header); | ||||||
303 | |||||||
304 | // Determine which blocks should stay in L and which should be moved out to | ||||||
305 | // the Outer loop now. | ||||||
306 | SmallPtrSet<BasicBlock *, 4> BlocksInL; | ||||||
307 | for (BasicBlock *P : predecessors(Header)) { | ||||||
308 | if (DT->dominates(Header, P)) | ||||||
309 | addBlockAndPredsToSet(P, Header, BlocksInL); | ||||||
310 | } | ||||||
311 | |||||||
312 | // Scan all of the loop children of L, moving them to OuterLoop if they are | ||||||
313 | // not part of the inner loop. | ||||||
314 | const std::vector<Loop*> &SubLoops = L->getSubLoops(); | ||||||
315 | for (size_t I = 0; I != SubLoops.size(); ) | ||||||
316 | if (BlocksInL.count(SubLoops[I]->getHeader())) | ||||||
317 | ++I; // Loop remains in L | ||||||
318 | else | ||||||
319 | NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I)); | ||||||
320 | |||||||
321 | SmallVector<BasicBlock *, 8> OuterLoopBlocks; | ||||||
322 | OuterLoopBlocks.push_back(NewBB); | ||||||
323 | // Now that we know which blocks are in L and which need to be moved to | ||||||
324 | // OuterLoop, move any blocks that need it. | ||||||
325 | for (unsigned i = 0; i != L->getBlocks().size(); ++i) { | ||||||
326 | BasicBlock *BB = L->getBlocks()[i]; | ||||||
327 | if (!BlocksInL.count(BB)) { | ||||||
328 | // Move this block to the parent, updating the exit blocks sets | ||||||
329 | L->removeBlockFromLoop(BB); | ||||||
330 | if ((*LI)[BB] == L) { | ||||||
331 | LI->changeLoopFor(BB, NewOuter); | ||||||
332 | OuterLoopBlocks.push_back(BB); | ||||||
333 | } | ||||||
334 | --i; | ||||||
335 | } | ||||||
336 | } | ||||||
337 | |||||||
338 | // Split edges to exit blocks from the inner loop, if they emerged in the | ||||||
339 | // process of separating the outer one. | ||||||
340 | formDedicatedExitBlocks(L, DT, LI, MSSAU, PreserveLCSSA); | ||||||
341 | |||||||
342 | if (PreserveLCSSA) { | ||||||
343 | // Fix LCSSA form for L. Some values, which previously were only used inside | ||||||
344 | // L, can now be used in NewOuter loop. We need to insert phi-nodes for them | ||||||
345 | // in corresponding exit blocks. | ||||||
346 | // We don't need to form LCSSA recursively, because there cannot be uses | ||||||
347 | // inside a newly created loop of defs from inner loops as those would | ||||||
348 | // already be a use of an LCSSA phi node. | ||||||
349 | formLCSSA(*L, *DT, LI, SE); | ||||||
350 | |||||||
351 | assert(NewOuter->isRecursivelyLCSSAForm(*DT, *LI) &&((void)0) | ||||||
352 | "LCSSA is broken after separating nested loops!")((void)0); | ||||||
353 | } | ||||||
354 | |||||||
355 | return NewOuter; | ||||||
356 | } | ||||||
357 | |||||||
358 | /// This method is called when the specified loop has more than one | ||||||
359 | /// backedge in it. | ||||||
360 | /// | ||||||
361 | /// If this occurs, revector all of these backedges to target a new basic block | ||||||
362 | /// and have that block branch to the loop header. This ensures that loops | ||||||
363 | /// have exactly one backedge. | ||||||
364 | static BasicBlock *insertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader, | ||||||
365 | DominatorTree *DT, LoopInfo *LI, | ||||||
366 | MemorySSAUpdater *MSSAU) { | ||||||
367 | assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!")((void)0); | ||||||
368 | |||||||
369 | // Get information about the loop | ||||||
370 | BasicBlock *Header = L->getHeader(); | ||||||
371 | Function *F = Header->getParent(); | ||||||
372 | |||||||
373 | // Unique backedge insertion currently depends on having a preheader. | ||||||
374 | if (!Preheader
| ||||||
375 | return nullptr; | ||||||
376 | |||||||
377 | // The header is not an EH pad; preheader insertion should ensure this. | ||||||
378 | assert(!Header->isEHPad() && "Can't insert backedge to EH pad")((void)0); | ||||||
379 | |||||||
380 | // Figure out which basic blocks contain back-edges to the loop header. | ||||||
381 | std::vector<BasicBlock*> BackedgeBlocks; | ||||||
382 | for (BasicBlock *P : predecessors(Header)) { | ||||||
383 | // Indirect edges cannot be split, so we must fail if we find one. | ||||||
384 | if (P->getTerminator()->isIndirectTerminator()) | ||||||
385 | return nullptr; | ||||||
386 | |||||||
387 | if (P != Preheader) BackedgeBlocks.push_back(P); | ||||||
388 | } | ||||||
389 | |||||||
390 | // Create and insert the new backedge block... | ||||||
391 | BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(), | ||||||
392 | Header->getName() + ".backedge", F); | ||||||
393 | BranchInst *BETerminator = BranchInst::Create(Header, BEBlock); | ||||||
394 | BETerminator->setDebugLoc(Header->getFirstNonPHI()->getDebugLoc()); | ||||||
395 | |||||||
396 | LLVM_DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "do { } while (false) | ||||||
397 | << BEBlock->getName() << "\n")do { } while (false); | ||||||
398 | |||||||
399 | // Move the new backedge block to right after the last backedge block. | ||||||
400 | Function::iterator InsertPos = ++BackedgeBlocks.back()->getIterator(); | ||||||
401 | F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); | ||||||
402 | |||||||
403 | // Now that the block has been inserted into the function, create PHI nodes in | ||||||
404 | // the backedge block which correspond to any PHI nodes in the header block. | ||||||
405 | for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { | ||||||
406 | PHINode *PN = cast<PHINode>(I); | ||||||
407 | PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(), | ||||||
408 | PN->getName()+".be", BETerminator); | ||||||
409 | |||||||
410 | // Loop over the PHI node, moving all entries except the one for the | ||||||
411 | // preheader over to the new PHI node. | ||||||
412 | unsigned PreheaderIdx = ~0U; | ||||||
413 | bool HasUniqueIncomingValue = true; | ||||||
414 | Value *UniqueValue = nullptr; | ||||||
415 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | ||||||
416 | BasicBlock *IBB = PN->getIncomingBlock(i); | ||||||
417 | Value *IV = PN->getIncomingValue(i); | ||||||
418 | if (IBB == Preheader) { | ||||||
419 | PreheaderIdx = i; | ||||||
420 | } else { | ||||||
421 | NewPN->addIncoming(IV, IBB); | ||||||
422 | if (HasUniqueIncomingValue) { | ||||||
423 | if (!UniqueValue) | ||||||
424 | UniqueValue = IV; | ||||||
425 | else if (UniqueValue != IV) | ||||||
426 | HasUniqueIncomingValue = false; | ||||||
427 | } | ||||||
428 | } | ||||||
429 | } | ||||||
430 | |||||||
431 | // Delete all of the incoming values from the old PN except the preheader's | ||||||
432 | assert(PreheaderIdx != ~0U && "PHI has no preheader entry??")((void)0); | ||||||
433 | if (PreheaderIdx != 0) { | ||||||
434 | PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); | ||||||
435 | PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); | ||||||
436 | } | ||||||
437 | // Nuke all entries except the zero'th. | ||||||
438 | for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) | ||||||
439 | PN->removeIncomingValue(e-i, false); | ||||||
440 | |||||||
441 | // Finally, add the newly constructed PHI node as the entry for the BEBlock. | ||||||
442 | PN->addIncoming(NewPN, BEBlock); | ||||||
443 | |||||||
444 | // As an optimization, if all incoming values in the new PhiNode (which is a | ||||||
445 | // subset of the incoming values of the old PHI node) have the same value, | ||||||
446 | // eliminate the PHI Node. | ||||||
447 | if (HasUniqueIncomingValue) { | ||||||
448 | NewPN->replaceAllUsesWith(UniqueValue); | ||||||
449 | BEBlock->getInstList().erase(NewPN); | ||||||
450 | } | ||||||
451 | } | ||||||
452 | |||||||
453 | // Now that all of the PHI nodes have been inserted and adjusted, modify the | ||||||
454 | // backedge blocks to jump to the BEBlock instead of the header. | ||||||
455 | // If one of the backedges has llvm.loop metadata attached, we remove | ||||||
456 | // it from the backedge and add it to BEBlock. | ||||||
457 | unsigned LoopMDKind = BEBlock->getContext().getMDKindID("llvm.loop"); | ||||||
458 | MDNode *LoopMD = nullptr; | ||||||
459 | for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { | ||||||
460 | Instruction *TI = BackedgeBlocks[i]->getTerminator(); | ||||||
461 | if (!LoopMD) | ||||||
462 | LoopMD = TI->getMetadata(LoopMDKind); | ||||||
463 | TI->setMetadata(LoopMDKind, nullptr); | ||||||
464 | TI->replaceSuccessorWith(Header, BEBlock); | ||||||
465 | } | ||||||
466 | BEBlock->getTerminator()->setMetadata(LoopMDKind, LoopMD); | ||||||
467 | |||||||
468 | //===--- Update all analyses which we must preserve now -----------------===// | ||||||
469 | |||||||
470 | // Update Loop Information - we know that this block is now in the current | ||||||
471 | // loop and all parent loops. | ||||||
472 | L->addBasicBlockToLoop(BEBlock, *LI); | ||||||
473 | |||||||
474 | // Update dominator information | ||||||
475 | DT->splitBlock(BEBlock); | ||||||
476 | |||||||
477 | if (MSSAU) | ||||||
478 | MSSAU->updatePhisWhenInsertingUniqueBackedgeBlock(Header, Preheader, | ||||||
479 | BEBlock); | ||||||
480 | |||||||
481 | return BEBlock; | ||||||
482 | } | ||||||
483 | |||||||
484 | /// Simplify one loop and queue further loops for simplification. | ||||||
485 | static bool simplifyOneLoop(Loop *L, SmallVectorImpl<Loop *> &Worklist, | ||||||
486 | DominatorTree *DT, LoopInfo *LI, | ||||||
487 | ScalarEvolution *SE, AssumptionCache *AC, | ||||||
488 | MemorySSAUpdater *MSSAU, bool PreserveLCSSA) { | ||||||
489 | bool Changed = false; | ||||||
490 | if (MSSAU && VerifyMemorySSA) | ||||||
| |||||||
491 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||
492 | |||||||
493 | ReprocessLoop: | ||||||
494 | |||||||
495 | // Check to see that no blocks (other than the header) in this loop have | ||||||
496 | // predecessors that are not in the loop. This is not valid for natural | ||||||
497 | // loops, but can occur if the blocks are unreachable. Since they are | ||||||
498 | // unreachable we can just shamelessly delete those CFG edges! | ||||||
499 | for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); | ||||||
500 | BB != E; ++BB) { | ||||||
501 | if (*BB == L->getHeader()) continue; | ||||||
502 | |||||||
503 | SmallPtrSet<BasicBlock*, 4> BadPreds; | ||||||
504 | for (BasicBlock *P : predecessors(*BB)) | ||||||
505 | if (!L->contains(P)) | ||||||
506 | BadPreds.insert(P); | ||||||
507 | |||||||
508 | // Delete each unique out-of-loop (and thus dead) predecessor. | ||||||
509 | for (BasicBlock *P : BadPreds) { | ||||||
510 | |||||||
511 | LLVM_DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "do { } while (false) | ||||||
512 | << P->getName() << "\n")do { } while (false); | ||||||
513 | |||||||
514 | // Zap the dead pred's terminator and replace it with unreachable. | ||||||
515 | Instruction *TI = P->getTerminator(); | ||||||
516 | changeToUnreachable(TI, PreserveLCSSA, | ||||||
517 | /*DTU=*/nullptr, MSSAU); | ||||||
518 | Changed = true; | ||||||
519 | } | ||||||
520 | } | ||||||
521 | |||||||
522 | if (MSSAU
| ||||||
523 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||
524 | |||||||
525 | // If there are exiting blocks with branches on undef, resolve the undef in | ||||||
526 | // the direction which will exit the loop. This will help simplify loop | ||||||
527 | // trip count computations. | ||||||
528 | SmallVector<BasicBlock*, 8> ExitingBlocks; | ||||||
529 | L->getExitingBlocks(ExitingBlocks); | ||||||
530 | for (BasicBlock *ExitingBlock : ExitingBlocks) | ||||||
531 | if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator())) | ||||||
532 | if (BI->isConditional()) { | ||||||
533 | if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) { | ||||||
534 | |||||||
535 | LLVM_DEBUG(dbgs()do { } while (false) | ||||||
536 | << "LoopSimplify: Resolving \"br i1 undef\" to exit in "do { } while (false) | ||||||
537 | << ExitingBlock->getName() << "\n")do { } while (false); | ||||||
538 | |||||||
539 | BI->setCondition(ConstantInt::get(Cond->getType(), | ||||||
540 | !L->contains(BI->getSuccessor(0)))); | ||||||
541 | |||||||
542 | Changed = true; | ||||||
543 | } | ||||||
544 | } | ||||||
545 | |||||||
546 | // Does the loop already have a preheader? If so, don't insert one. | ||||||
547 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||
548 | if (!Preheader) { | ||||||
549 | Preheader = InsertPreheaderForLoop(L, DT, LI, MSSAU, PreserveLCSSA); | ||||||
550 | if (Preheader) | ||||||
551 | Changed = true; | ||||||
552 | } | ||||||
553 | |||||||
554 | // Next, check to make sure that all exit nodes of the loop only have | ||||||
555 | // predecessors that are inside of the loop. This check guarantees that the | ||||||
556 | // loop preheader/header will dominate the exit blocks. If the exit block has | ||||||
557 | // predecessors from outside of the loop, split the edge now. | ||||||
558 | if (formDedicatedExitBlocks(L, DT, LI, MSSAU, PreserveLCSSA)) | ||||||
559 | Changed = true; | ||||||
560 | |||||||
561 | if (MSSAU
| ||||||
562 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||
563 | |||||||
564 | // If the header has more than two predecessors at this point (from the | ||||||
565 | // preheader and from multiple backedges), we must adjust the loop. | ||||||
566 | BasicBlock *LoopLatch = L->getLoopLatch(); | ||||||
567 | if (!LoopLatch) { | ||||||
568 | // If this is really a nested loop, rip it out into a child loop. Don't do | ||||||
569 | // this for loops with a giant number of backedges, just factor them into a | ||||||
570 | // common backedge instead. | ||||||
571 | if (L->getNumBackEdges() < 8) { | ||||||
572 | if (Loop *OuterL = separateNestedLoop(L, Preheader, DT, LI, SE, | ||||||
573 | PreserveLCSSA, AC, MSSAU)) { | ||||||
574 | ++NumNested; | ||||||
575 | // Enqueue the outer loop as it should be processed next in our | ||||||
576 | // depth-first nest walk. | ||||||
577 | Worklist.push_back(OuterL); | ||||||
578 | |||||||
579 | // This is a big restructuring change, reprocess the whole loop. | ||||||
580 | Changed = true; | ||||||
581 | // GCC doesn't tail recursion eliminate this. | ||||||
582 | // FIXME: It isn't clear we can't rely on LLVM to TRE this. | ||||||
583 | goto ReprocessLoop; | ||||||
584 | } | ||||||
585 | } | ||||||
586 | |||||||
587 | // If we either couldn't, or didn't want to, identify nesting of the loops, | ||||||
588 | // insert a new block that all backedges target, then make it jump to the | ||||||
589 | // loop header. | ||||||
590 | LoopLatch = insertUniqueBackedgeBlock(L, Preheader, DT, LI, MSSAU); | ||||||
591 | if (LoopLatch) | ||||||
592 | Changed = true; | ||||||
593 | } | ||||||
594 | |||||||
595 | if (MSSAU && VerifyMemorySSA) | ||||||
596 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||
597 | |||||||
598 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); | ||||||
599 | |||||||
600 | // Scan over the PHI nodes in the loop header. Since they now have only two | ||||||
601 | // incoming values (the loop is canonicalized), we may have simplified the PHI | ||||||
602 | // down to 'X = phi [X, Y]', which should be replaced with 'Y'. | ||||||
603 | PHINode *PN; | ||||||
604 | for (BasicBlock::iterator I = L->getHeader()->begin(); | ||||||
605 | (PN = dyn_cast<PHINode>(I++)); ) | ||||||
606 | if (Value *V = SimplifyInstruction(PN, {DL, nullptr, DT, AC})) { | ||||||
607 | if (SE) SE->forgetValue(PN); | ||||||
608 | if (!PreserveLCSSA || LI->replacementPreservesLCSSAForm(PN, V)) { | ||||||
609 | PN->replaceAllUsesWith(V); | ||||||
610 | PN->eraseFromParent(); | ||||||
611 | Changed = true; | ||||||
612 | } | ||||||
613 | } | ||||||
614 | |||||||
615 | // If this loop has multiple exits and the exits all go to the same | ||||||
616 | // block, attempt to merge the exits. This helps several passes, such | ||||||
617 | // as LoopRotation, which do not support loops with multiple exits. | ||||||
618 | // SimplifyCFG also does this (and this code uses the same utility | ||||||
619 | // function), however this code is loop-aware, where SimplifyCFG is | ||||||
620 | // not. That gives it the advantage of being able to hoist | ||||||
621 | // loop-invariant instructions out of the way to open up more | ||||||
622 | // opportunities, and the disadvantage of having the responsibility | ||||||
623 | // to preserve dominator information. | ||||||
624 | auto HasUniqueExitBlock = [&]() { | ||||||
625 | BasicBlock *UniqueExit = nullptr; | ||||||
626 | for (auto *ExitingBB : ExitingBlocks) | ||||||
627 | for (auto *SuccBB : successors(ExitingBB)) { | ||||||
628 | if (L->contains(SuccBB)) | ||||||
629 | continue; | ||||||
630 | |||||||
631 | if (!UniqueExit) | ||||||
632 | UniqueExit = SuccBB; | ||||||
633 | else if (UniqueExit != SuccBB) | ||||||
634 | return false; | ||||||
635 | } | ||||||
636 | |||||||
637 | return true; | ||||||
638 | }; | ||||||
639 | if (HasUniqueExitBlock()) { | ||||||
640 | for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { | ||||||
641 | BasicBlock *ExitingBlock = ExitingBlocks[i]; | ||||||
642 | if (!ExitingBlock->getSinglePredecessor()) continue; | ||||||
643 | BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); | ||||||
644 | if (!BI || !BI->isConditional()) continue; | ||||||
645 | CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition()); | ||||||
646 | if (!CI || CI->getParent() != ExitingBlock) continue; | ||||||
647 | |||||||
648 | // Attempt to hoist out all instructions except for the | ||||||
649 | // comparison and the branch. | ||||||
650 | bool AllInvariant = true; | ||||||
651 | bool AnyInvariant = false; | ||||||
652 | for (auto I = ExitingBlock->instructionsWithoutDebug().begin(); &*I != BI; ) { | ||||||
653 | Instruction *Inst = &*I++; | ||||||
654 | if (Inst == CI) | ||||||
655 | continue; | ||||||
656 | if (!L->makeLoopInvariant( | ||||||
657 | Inst, AnyInvariant, | ||||||
658 | Preheader ? Preheader->getTerminator() : nullptr, MSSAU)) { | ||||||
659 | AllInvariant = false; | ||||||
660 | break; | ||||||
661 | } | ||||||
662 | } | ||||||
663 | if (AnyInvariant) { | ||||||
664 | Changed = true; | ||||||
665 | // The loop disposition of all SCEV expressions that depend on any | ||||||
666 | // hoisted values have also changed. | ||||||
667 | if (SE) | ||||||
668 | SE->forgetLoopDispositions(L); | ||||||
669 | } | ||||||
670 | if (!AllInvariant) continue; | ||||||
671 | |||||||
672 | // The block has now been cleared of all instructions except for | ||||||
673 | // a comparison and a conditional branch. SimplifyCFG may be able | ||||||
674 | // to fold it now. | ||||||
675 | if (!FoldBranchToCommonDest(BI, /*DTU=*/nullptr, MSSAU)) | ||||||
676 | continue; | ||||||
677 | |||||||
678 | // Success. The block is now dead, so remove it from the loop, | ||||||
679 | // update the dominator tree and delete it. | ||||||
680 | LLVM_DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "do { } while (false) | ||||||
681 | << ExitingBlock->getName() << "\n")do { } while (false); | ||||||
682 | |||||||
683 | assert(pred_empty(ExitingBlock))((void)0); | ||||||
684 | Changed = true; | ||||||
685 | LI->removeBlock(ExitingBlock); | ||||||
686 | |||||||
687 | DomTreeNode *Node = DT->getNode(ExitingBlock); | ||||||
688 | while (!Node->isLeaf()) { | ||||||
689 | DomTreeNode *Child = Node->back(); | ||||||
690 | DT->changeImmediateDominator(Child, Node->getIDom()); | ||||||
691 | } | ||||||
692 | DT->eraseNode(ExitingBlock); | ||||||
693 | if (MSSAU) { | ||||||
694 | SmallSetVector<BasicBlock *, 8> ExitBlockSet; | ||||||
695 | ExitBlockSet.insert(ExitingBlock); | ||||||
696 | MSSAU->removeBlocks(ExitBlockSet); | ||||||
697 | } | ||||||
698 | |||||||
699 | BI->getSuccessor(0)->removePredecessor( | ||||||
700 | ExitingBlock, /* KeepOneInputPHIs */ PreserveLCSSA); | ||||||
701 | BI->getSuccessor(1)->removePredecessor( | ||||||
702 | ExitingBlock, /* KeepOneInputPHIs */ PreserveLCSSA); | ||||||
703 | ExitingBlock->eraseFromParent(); | ||||||
704 | } | ||||||
705 | } | ||||||
706 | |||||||
707 | // Changing exit conditions for blocks may affect exit counts of this loop and | ||||||
708 | // any of its paretns, so we must invalidate the entire subtree if we've made | ||||||
709 | // any changes. | ||||||
710 | if (Changed && SE) | ||||||
711 | SE->forgetTopmostLoop(L); | ||||||
712 | |||||||
713 | if (MSSAU && VerifyMemorySSA) | ||||||
714 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||
715 | |||||||
716 | return Changed; | ||||||
717 | } | ||||||
718 | |||||||
719 | bool llvm::simplifyLoop(Loop *L, DominatorTree *DT, LoopInfo *LI, | ||||||
720 | ScalarEvolution *SE, AssumptionCache *AC, | ||||||
721 | MemorySSAUpdater *MSSAU, bool PreserveLCSSA) { | ||||||
722 | bool Changed = false; | ||||||
723 | |||||||
724 | #ifndef NDEBUG1 | ||||||
725 | // If we're asked to preserve LCSSA, the loop nest needs to start in LCSSA | ||||||
726 | // form. | ||||||
727 | if (PreserveLCSSA) { | ||||||
728 | assert(DT && "DT not available.")((void)0); | ||||||
729 | assert(LI && "LI not available.")((void)0); | ||||||
730 | assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&((void)0) | ||||||
731 | "Requested to preserve LCSSA, but it's already broken.")((void)0); | ||||||
732 | } | ||||||
733 | #endif | ||||||
734 | |||||||
735 | // Worklist maintains our depth-first queue of loops in this nest to process. | ||||||
736 | SmallVector<Loop *, 4> Worklist; | ||||||
737 | Worklist.push_back(L); | ||||||
738 | |||||||
739 | // Walk the worklist from front to back, pushing newly found sub loops onto | ||||||
740 | // the back. This will let us process loops from back to front in depth-first | ||||||
741 | // order. We can use this simple process because loops form a tree. | ||||||
742 | for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { | ||||||
743 | Loop *L2 = Worklist[Idx]; | ||||||
744 | Worklist.append(L2->begin(), L2->end()); | ||||||
745 | } | ||||||
746 | |||||||
747 | while (!Worklist.empty()) | ||||||
748 | Changed |= simplifyOneLoop(Worklist.pop_back_val(), Worklist, DT, LI, SE, | ||||||
749 | AC, MSSAU, PreserveLCSSA); | ||||||
750 | |||||||
751 | return Changed; | ||||||
752 | } | ||||||
753 | |||||||
754 | namespace { | ||||||
755 | struct LoopSimplify : public FunctionPass { | ||||||
756 | static char ID; // Pass identification, replacement for typeid | ||||||
757 | LoopSimplify() : FunctionPass(ID) { | ||||||
758 | initializeLoopSimplifyPass(*PassRegistry::getPassRegistry()); | ||||||
759 | } | ||||||
760 | |||||||
761 | bool runOnFunction(Function &F) override; | ||||||
762 | |||||||
763 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||
764 | AU.addRequired<AssumptionCacheTracker>(); | ||||||
765 | |||||||
766 | // We need loop information to identify the loops... | ||||||
767 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||
768 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||
769 | |||||||
770 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||
771 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||
772 | |||||||
773 | AU.addPreserved<BasicAAWrapperPass>(); | ||||||
774 | AU.addPreserved<AAResultsWrapperPass>(); | ||||||
775 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||||
776 | AU.addPreserved<ScalarEvolutionWrapperPass>(); | ||||||
777 | AU.addPreserved<SCEVAAWrapperPass>(); | ||||||
778 | AU.addPreservedID(LCSSAID); | ||||||
779 | AU.addPreserved<DependenceAnalysisWrapperPass>(); | ||||||
780 | AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. | ||||||
781 | AU.addPreserved<BranchProbabilityInfoWrapperPass>(); | ||||||
782 | if (EnableMSSALoopDependency) | ||||||
783 | AU.addPreserved<MemorySSAWrapperPass>(); | ||||||
784 | } | ||||||
785 | |||||||
786 | /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees. | ||||||
787 | void verifyAnalysis() const override; | ||||||
788 | }; | ||||||
789 | } | ||||||
790 | |||||||
791 | char LoopSimplify::ID = 0; | ||||||
792 | INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify",static void *initializeLoopSimplifyPassOnce(PassRegistry & Registry) { | ||||||
793 | "Canonicalize natural loops", false, false)static void *initializeLoopSimplifyPassOnce(PassRegistry & Registry) { | ||||||
794 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||
795 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||
796 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||
797 | INITIALIZE_PASS_END(LoopSimplify, "loop-simplify",PassInfo *PI = new PassInfo( "Canonicalize natural loops", "loop-simplify" , &LoopSimplify::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopSimplify>), false, false); Registry.registerPass(* PI, true); return PI; } static llvm::once_flag InitializeLoopSimplifyPassFlag ; void llvm::initializeLoopSimplifyPass(PassRegistry &Registry ) { llvm::call_once(InitializeLoopSimplifyPassFlag, initializeLoopSimplifyPassOnce , std::ref(Registry)); } | ||||||
798 | "Canonicalize natural loops", false, false)PassInfo *PI = new PassInfo( "Canonicalize natural loops", "loop-simplify" , &LoopSimplify::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopSimplify>), false, false); Registry.registerPass(* PI, true); return PI; } static llvm::once_flag InitializeLoopSimplifyPassFlag ; void llvm::initializeLoopSimplifyPass(PassRegistry &Registry ) { llvm::call_once(InitializeLoopSimplifyPassFlag, initializeLoopSimplifyPassOnce , std::ref(Registry)); } | ||||||
799 | |||||||
800 | // Publicly exposed interface to pass... | ||||||
801 | char &llvm::LoopSimplifyID = LoopSimplify::ID; | ||||||
802 | Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } | ||||||
803 | |||||||
804 | /// runOnFunction - Run down all loops in the CFG (recursively, but we could do | ||||||
805 | /// it in any convenient order) inserting preheaders... | ||||||
806 | /// | ||||||
807 | bool LoopSimplify::runOnFunction(Function &F) { | ||||||
808 | bool Changed = false; | ||||||
809 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||
810 | DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||
811 | auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | ||||||
812 | ScalarEvolution *SE = SEWP ? &SEWP->getSE() : nullptr; | ||||||
813 | AssumptionCache *AC = | ||||||
814 | &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||
815 | MemorySSA *MSSA = nullptr; | ||||||
816 | std::unique_ptr<MemorySSAUpdater> MSSAU; | ||||||
817 | if (EnableMSSALoopDependency) { | ||||||
818 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); | ||||||
819 | if (MSSAAnalysis) { | ||||||
820 | MSSA = &MSSAAnalysis->getMSSA(); | ||||||
821 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); | ||||||
822 | } | ||||||
823 | } | ||||||
824 | |||||||
825 | bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID); | ||||||
826 | |||||||
827 | // Simplify each loop nest in the function. | ||||||
828 | for (auto *L : *LI) | ||||||
829 | Changed |= simplifyLoop(L, DT, LI, SE, AC, MSSAU.get(), PreserveLCSSA); | ||||||
830 | |||||||
831 | #ifndef NDEBUG1 | ||||||
832 | if (PreserveLCSSA) { | ||||||
833 | bool InLCSSA = all_of( | ||||||
834 | *LI, [&](Loop *L) { return L->isRecursivelyLCSSAForm(*DT, *LI); }); | ||||||
835 | assert(InLCSSA && "LCSSA is broken after loop-simplify.")((void)0); | ||||||
836 | } | ||||||
837 | #endif | ||||||
838 | return Changed; | ||||||
839 | } | ||||||
840 | |||||||
841 | PreservedAnalyses LoopSimplifyPass::run(Function &F, | ||||||
842 | FunctionAnalysisManager &AM) { | ||||||
843 | bool Changed = false; | ||||||
844 | LoopInfo *LI = &AM.getResult<LoopAnalysis>(F); | ||||||
845 | DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F); | ||||||
846 | ScalarEvolution *SE = AM.getCachedResult<ScalarEvolutionAnalysis>(F); | ||||||
847 | AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F); | ||||||
848 | auto *MSSAAnalysis = AM.getCachedResult<MemorySSAAnalysis>(F); | ||||||
849 | std::unique_ptr<MemorySSAUpdater> MSSAU; | ||||||
850 | if (MSSAAnalysis) { | ||||||
851 | auto *MSSA = &MSSAAnalysis->getMSSA(); | ||||||
852 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); | ||||||
853 | } | ||||||
854 | |||||||
855 | |||||||
856 | // Note that we don't preserve LCSSA in the new PM, if you need it run LCSSA | ||||||
857 | // after simplifying the loops. MemorySSA is preserved if it exists. | ||||||
858 | for (auto *L : *LI) | ||||||
859 | Changed |= | ||||||
860 | simplifyLoop(L, DT, LI, SE, AC, MSSAU.get(), /*PreserveLCSSA*/ false); | ||||||
861 | |||||||
862 | if (!Changed) | ||||||
863 | return PreservedAnalyses::all(); | ||||||
864 | |||||||
865 | PreservedAnalyses PA; | ||||||
866 | PA.preserve<DominatorTreeAnalysis>(); | ||||||
867 | PA.preserve<LoopAnalysis>(); | ||||||
868 | PA.preserve<ScalarEvolutionAnalysis>(); | ||||||
869 | PA.preserve<DependenceAnalysis>(); | ||||||
870 | if (MSSAAnalysis) | ||||||
871 | PA.preserve<MemorySSAAnalysis>(); | ||||||
872 | // BPI maps conditional terminators to probabilities, LoopSimplify can insert | ||||||
873 | // blocks, but it does so only by splitting existing blocks and edges. This | ||||||
874 | // results in the interesting property that all new terminators inserted are | ||||||
875 | // unconditional branches which do not appear in BPI. All deletions are | ||||||
876 | // handled via ValueHandle callbacks w/in BPI. | ||||||
877 | PA.preserve<BranchProbabilityAnalysis>(); | ||||||
878 | return PA; | ||||||
879 | } | ||||||
880 | |||||||
881 | // FIXME: Restore this code when we re-enable verification in verifyAnalysis | ||||||
882 | // below. | ||||||
883 | #if 0 | ||||||
884 | static void verifyLoop(Loop *L) { | ||||||
885 | // Verify subloops. | ||||||
886 | for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) | ||||||
887 | verifyLoop(*I); | ||||||
888 | |||||||
889 | // It used to be possible to just assert L->isLoopSimplifyForm(), however | ||||||
890 | // with the introduction of indirectbr, there are now cases where it's | ||||||
891 | // not possible to transform a loop as necessary. We can at least check | ||||||
892 | // that there is an indirectbr near any time there's trouble. | ||||||
893 | |||||||
894 | // Indirectbr can interfere with preheader and unique backedge insertion. | ||||||
895 | if (!L->getLoopPreheader() || !L->getLoopLatch()) { | ||||||
896 | bool HasIndBrPred = false; | ||||||
897 | for (BasicBlock *Pred : predecessors(L->getHeader())) | ||||||
898 | if (isa<IndirectBrInst>(Pred->getTerminator())) { | ||||||
899 | HasIndBrPred = true; | ||||||
900 | break; | ||||||
901 | } | ||||||
902 | assert(HasIndBrPred &&((void)0) | ||||||
903 | "LoopSimplify has no excuse for missing loop header info!")((void)0); | ||||||
904 | (void)HasIndBrPred; | ||||||
905 | } | ||||||
906 | |||||||
907 | // Indirectbr can interfere with exit block canonicalization. | ||||||
908 | if (!L->hasDedicatedExits()) { | ||||||
909 | bool HasIndBrExiting = false; | ||||||
910 | SmallVector<BasicBlock*, 8> ExitingBlocks; | ||||||
911 | L->getExitingBlocks(ExitingBlocks); | ||||||
912 | for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { | ||||||
913 | if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) { | ||||||
914 | HasIndBrExiting = true; | ||||||
915 | break; | ||||||
916 | } | ||||||
917 | } | ||||||
918 | |||||||
919 | assert(HasIndBrExiting &&((void)0) | ||||||
920 | "LoopSimplify has no excuse for missing exit block info!")((void)0); | ||||||
921 | (void)HasIndBrExiting; | ||||||
922 | } | ||||||
923 | } | ||||||
924 | #endif | ||||||
925 | |||||||
926 | void LoopSimplify::verifyAnalysis() const { | ||||||
927 | // FIXME: This routine is being called mid-way through the loop pass manager | ||||||
928 | // as loop passes destroy this analysis. That's actually fine, but we have no | ||||||
929 | // way of expressing that here. Once all of the passes that destroy this are | ||||||
930 | // hoisted out of the loop pass manager we can add back verification here. | ||||||
931 | #if 0 | ||||||
932 | for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) | ||||||
933 | verifyLoop(*I); | ||||||
934 | #endif | ||||||
935 | } |
1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- C++ -*-===// | ||||||
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 | /// \file | ||||||
9 | /// | ||||||
10 | /// This file defines a set of templates that efficiently compute a dominator | ||||||
11 | /// tree over a generic graph. This is used typically in LLVM for fast | ||||||
12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying | ||||||
13 | /// graph types. | ||||||
14 | /// | ||||||
15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements | ||||||
16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, | ||||||
17 | /// NodeRef->getParent() must return the parent node that is also a pointer. | ||||||
18 | /// | ||||||
19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. | ||||||
20 | /// | ||||||
21 | //===----------------------------------------------------------------------===// | ||||||
22 | |||||||
23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H | ||||||
24 | #define LLVM_SUPPORT_GENERICDOMTREE_H | ||||||
25 | |||||||
26 | #include "llvm/ADT/DenseMap.h" | ||||||
27 | #include "llvm/ADT/GraphTraits.h" | ||||||
28 | #include "llvm/ADT/STLExtras.h" | ||||||
29 | #include "llvm/ADT/SmallPtrSet.h" | ||||||
30 | #include "llvm/ADT/SmallVector.h" | ||||||
31 | #include "llvm/Support/CFGDiff.h" | ||||||
32 | #include "llvm/Support/CFGUpdate.h" | ||||||
33 | #include "llvm/Support/raw_ostream.h" | ||||||
34 | #include <algorithm> | ||||||
35 | #include <cassert> | ||||||
36 | #include <cstddef> | ||||||
37 | #include <iterator> | ||||||
38 | #include <memory> | ||||||
39 | #include <type_traits> | ||||||
40 | #include <utility> | ||||||
41 | |||||||
42 | namespace llvm { | ||||||
43 | |||||||
44 | template <typename NodeT, bool IsPostDom> | ||||||
45 | class DominatorTreeBase; | ||||||
46 | |||||||
47 | namespace DomTreeBuilder { | ||||||
48 | template <typename DomTreeT> | ||||||
49 | struct SemiNCAInfo; | ||||||
50 | } // namespace DomTreeBuilder | ||||||
51 | |||||||
52 | /// Base class for the actual dominator tree node. | ||||||
53 | template <class NodeT> class DomTreeNodeBase { | ||||||
54 | friend class PostDominatorTree; | ||||||
55 | friend class DominatorTreeBase<NodeT, false>; | ||||||
56 | friend class DominatorTreeBase<NodeT, true>; | ||||||
57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; | ||||||
58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; | ||||||
59 | |||||||
60 | NodeT *TheBB; | ||||||
61 | DomTreeNodeBase *IDom; | ||||||
62 | unsigned Level; | ||||||
63 | SmallVector<DomTreeNodeBase *, 4> Children; | ||||||
64 | mutable unsigned DFSNumIn = ~0; | ||||||
65 | mutable unsigned DFSNumOut = ~0; | ||||||
66 | |||||||
67 | public: | ||||||
68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) | ||||||
69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} | ||||||
70 | |||||||
71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; | ||||||
72 | using const_iterator = | ||||||
73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; | ||||||
74 | |||||||
75 | iterator begin() { return Children.begin(); } | ||||||
76 | iterator end() { return Children.end(); } | ||||||
77 | const_iterator begin() const { return Children.begin(); } | ||||||
78 | const_iterator end() const { return Children.end(); } | ||||||
79 | |||||||
80 | DomTreeNodeBase *const &back() const { return Children.back(); } | ||||||
81 | DomTreeNodeBase *&back() { return Children.back(); } | ||||||
82 | |||||||
83 | iterator_range<iterator> children() { return make_range(begin(), end()); } | ||||||
84 | iterator_range<const_iterator> children() const { | ||||||
85 | return make_range(begin(), end()); | ||||||
86 | } | ||||||
87 | |||||||
88 | NodeT *getBlock() const { return TheBB; } | ||||||
89 | DomTreeNodeBase *getIDom() const { return IDom; } | ||||||
90 | unsigned getLevel() const { return Level; } | ||||||
91 | |||||||
92 | std::unique_ptr<DomTreeNodeBase> addChild( | ||||||
93 | std::unique_ptr<DomTreeNodeBase> C) { | ||||||
94 | Children.push_back(C.get()); | ||||||
95 | return C; | ||||||
96 | } | ||||||
97 | |||||||
98 | bool isLeaf() const { return Children.empty(); } | ||||||
99 | size_t getNumChildren() const { return Children.size(); } | ||||||
100 | |||||||
101 | void clearAllChildren() { Children.clear(); } | ||||||
102 | |||||||
103 | bool compare(const DomTreeNodeBase *Other) const { | ||||||
104 | if (getNumChildren() != Other->getNumChildren()) | ||||||
105 | return true; | ||||||
106 | |||||||
107 | if (Level != Other->Level) return true; | ||||||
108 | |||||||
109 | SmallPtrSet<const NodeT *, 4> OtherChildren; | ||||||
110 | for (const DomTreeNodeBase *I : *Other) { | ||||||
111 | const NodeT *Nd = I->getBlock(); | ||||||
112 | OtherChildren.insert(Nd); | ||||||
113 | } | ||||||
114 | |||||||
115 | for (const DomTreeNodeBase *I : *this) { | ||||||
116 | const NodeT *N = I->getBlock(); | ||||||
117 | if (OtherChildren.count(N) == 0) | ||||||
118 | return true; | ||||||
119 | } | ||||||
120 | return false; | ||||||
121 | } | ||||||
122 | |||||||
123 | void setIDom(DomTreeNodeBase *NewIDom) { | ||||||
124 | assert(IDom && "No immediate dominator?")((void)0); | ||||||
125 | if (IDom == NewIDom) return; | ||||||
126 | |||||||
127 | auto I = find(IDom->Children, this); | ||||||
128 | assert(I != IDom->Children.end() &&((void)0) | ||||||
129 | "Not in immediate dominator children set!")((void)0); | ||||||
130 | // I am no longer your child... | ||||||
131 | IDom->Children.erase(I); | ||||||
132 | |||||||
133 | // Switch to new dominator | ||||||
134 | IDom = NewIDom; | ||||||
135 | IDom->Children.push_back(this); | ||||||
136 | |||||||
137 | UpdateLevel(); | ||||||
138 | } | ||||||
139 | |||||||
140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes | ||||||
141 | /// in the dominator tree. They are only guaranteed valid if | ||||||
142 | /// updateDFSNumbers() has been called. | ||||||
143 | unsigned getDFSNumIn() const { return DFSNumIn; } | ||||||
144 | unsigned getDFSNumOut() const { return DFSNumOut; } | ||||||
145 | |||||||
146 | private: | ||||||
147 | // Return true if this node is dominated by other. Use this only if DFS info | ||||||
148 | // is valid. | ||||||
149 | bool DominatedBy(const DomTreeNodeBase *other) const { | ||||||
150 | return this->DFSNumIn >= other->DFSNumIn && | ||||||
151 | this->DFSNumOut <= other->DFSNumOut; | ||||||
152 | } | ||||||
153 | |||||||
154 | void UpdateLevel() { | ||||||
155 | assert(IDom)((void)0); | ||||||
156 | if (Level == IDom->Level + 1) return; | ||||||
157 | |||||||
158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; | ||||||
159 | |||||||
160 | while (!WorkStack.empty()) { | ||||||
161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); | ||||||
162 | Current->Level = Current->IDom->Level + 1; | ||||||
163 | |||||||
164 | for (DomTreeNodeBase *C : *Current) { | ||||||
165 | assert(C->IDom)((void)0); | ||||||
166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); | ||||||
167 | } | ||||||
168 | } | ||||||
169 | } | ||||||
170 | }; | ||||||
171 | |||||||
172 | template <class NodeT> | ||||||
173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { | ||||||
174 | if (Node->getBlock()) | ||||||
175 | Node->getBlock()->printAsOperand(O, false); | ||||||
176 | else | ||||||
177 | O << " <<exit node>>"; | ||||||
178 | |||||||
179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" | ||||||
180 | << Node->getLevel() << "]\n"; | ||||||
181 | |||||||
182 | return O; | ||||||
183 | } | ||||||
184 | |||||||
185 | template <class NodeT> | ||||||
186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, | ||||||
187 | unsigned Lev) { | ||||||
188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; | ||||||
189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), | ||||||
190 | E = N->end(); | ||||||
191 | I != E; ++I) | ||||||
192 | PrintDomTree<NodeT>(*I, O, Lev + 1); | ||||||
193 | } | ||||||
194 | |||||||
195 | namespace DomTreeBuilder { | ||||||
196 | // The routines below are provided in a separate header but referenced here. | ||||||
197 | template <typename DomTreeT> | ||||||
198 | void Calculate(DomTreeT &DT); | ||||||
199 | |||||||
200 | template <typename DomTreeT> | ||||||
201 | void CalculateWithUpdates(DomTreeT &DT, | ||||||
202 | ArrayRef<typename DomTreeT::UpdateType> Updates); | ||||||
203 | |||||||
204 | template <typename DomTreeT> | ||||||
205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | ||||||
206 | typename DomTreeT::NodePtr To); | ||||||
207 | |||||||
208 | template <typename DomTreeT> | ||||||
209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | ||||||
210 | typename DomTreeT::NodePtr To); | ||||||
211 | |||||||
212 | template <typename DomTreeT> | ||||||
213 | void ApplyUpdates(DomTreeT &DT, | ||||||
214 | GraphDiff<typename DomTreeT::NodePtr, | ||||||
215 | DomTreeT::IsPostDominator> &PreViewCFG, | ||||||
216 | GraphDiff<typename DomTreeT::NodePtr, | ||||||
217 | DomTreeT::IsPostDominator> *PostViewCFG); | ||||||
218 | |||||||
219 | template <typename DomTreeT> | ||||||
220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); | ||||||
221 | } // namespace DomTreeBuilder | ||||||
222 | |||||||
223 | /// Core dominator tree base class. | ||||||
224 | /// | ||||||
225 | /// This class is a generic template over graph nodes. It is instantiated for | ||||||
226 | /// various graphs in the LLVM IR or in the code generator. | ||||||
227 | template <typename NodeT, bool IsPostDom> | ||||||
228 | class DominatorTreeBase { | ||||||
229 | public: | ||||||
230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, | ||||||
231 | "Currently DominatorTreeBase supports only pointer nodes"); | ||||||
232 | using NodeType = NodeT; | ||||||
233 | using NodePtr = NodeT *; | ||||||
234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); | ||||||
235 | static_assert(std::is_pointer<ParentPtr>::value, | ||||||
236 | "Currently NodeT's parent must be a pointer type"); | ||||||
237 | using ParentType = std::remove_pointer_t<ParentPtr>; | ||||||
238 | static constexpr bool IsPostDominator = IsPostDom; | ||||||
239 | |||||||
240 | using UpdateType = cfg::Update<NodePtr>; | ||||||
241 | using UpdateKind = cfg::UpdateKind; | ||||||
242 | static constexpr UpdateKind Insert = UpdateKind::Insert; | ||||||
243 | static constexpr UpdateKind Delete = UpdateKind::Delete; | ||||||
244 | |||||||
245 | enum class VerificationLevel { Fast, Basic, Full }; | ||||||
246 | |||||||
247 | protected: | ||||||
248 | // Dominators always have a single root, postdominators can have more. | ||||||
249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; | ||||||
250 | |||||||
251 | using DomTreeNodeMapType = | ||||||
252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; | ||||||
253 | DomTreeNodeMapType DomTreeNodes; | ||||||
254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; | ||||||
255 | ParentPtr Parent = nullptr; | ||||||
256 | |||||||
257 | mutable bool DFSInfoValid = false; | ||||||
258 | mutable unsigned int SlowQueries = 0; | ||||||
259 | |||||||
260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; | ||||||
261 | |||||||
262 | public: | ||||||
263 | DominatorTreeBase() {} | ||||||
264 | |||||||
265 | DominatorTreeBase(DominatorTreeBase &&Arg) | ||||||
266 | : Roots(std::move(Arg.Roots)), | ||||||
267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), | ||||||
268 | RootNode(Arg.RootNode), | ||||||
269 | Parent(Arg.Parent), | ||||||
270 | DFSInfoValid(Arg.DFSInfoValid), | ||||||
271 | SlowQueries(Arg.SlowQueries) { | ||||||
272 | Arg.wipe(); | ||||||
273 | } | ||||||
274 | |||||||
275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { | ||||||
276 | Roots = std::move(RHS.Roots); | ||||||
277 | DomTreeNodes = std::move(RHS.DomTreeNodes); | ||||||
278 | RootNode = RHS.RootNode; | ||||||
279 | Parent = RHS.Parent; | ||||||
280 | DFSInfoValid = RHS.DFSInfoValid; | ||||||
281 | SlowQueries = RHS.SlowQueries; | ||||||
282 | RHS.wipe(); | ||||||
283 | return *this; | ||||||
284 | } | ||||||
285 | |||||||
286 | DominatorTreeBase(const DominatorTreeBase &) = delete; | ||||||
287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; | ||||||
288 | |||||||
289 | /// Iteration over roots. | ||||||
290 | /// | ||||||
291 | /// This may include multiple blocks if we are computing post dominators. | ||||||
292 | /// For forward dominators, this will always be a single block (the entry | ||||||
293 | /// block). | ||||||
294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; | ||||||
295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; | ||||||
296 | |||||||
297 | root_iterator root_begin() { return Roots.begin(); } | ||||||
298 | const_root_iterator root_begin() const { return Roots.begin(); } | ||||||
299 | root_iterator root_end() { return Roots.end(); } | ||||||
300 | const_root_iterator root_end() const { return Roots.end(); } | ||||||
301 | |||||||
302 | size_t root_size() const { return Roots.size(); } | ||||||
303 | |||||||
304 | iterator_range<root_iterator> roots() { | ||||||
305 | return make_range(root_begin(), root_end()); | ||||||
306 | } | ||||||
307 | iterator_range<const_root_iterator> roots() const { | ||||||
308 | return make_range(root_begin(), root_end()); | ||||||
309 | } | ||||||
310 | |||||||
311 | /// isPostDominator - Returns true if analysis based of postdoms | ||||||
312 | /// | ||||||
313 | bool isPostDominator() const { return IsPostDominator; } | ||||||
314 | |||||||
315 | /// compare - Return false if the other dominator tree base matches this | ||||||
316 | /// dominator tree base. Otherwise return true. | ||||||
317 | bool compare(const DominatorTreeBase &Other) const { | ||||||
318 | if (Parent != Other.Parent) return true; | ||||||
319 | |||||||
320 | if (Roots.size() != Other.Roots.size()) | ||||||
321 | return true; | ||||||
322 | |||||||
323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) | ||||||
324 | return true; | ||||||
325 | |||||||
326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; | ||||||
327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) | ||||||
328 | return true; | ||||||
329 | |||||||
330 | for (const auto &DomTreeNode : DomTreeNodes) { | ||||||
331 | NodeT *BB = DomTreeNode.first; | ||||||
332 | typename DomTreeNodeMapType::const_iterator OI = | ||||||
333 | OtherDomTreeNodes.find(BB); | ||||||
334 | if (OI == OtherDomTreeNodes.end()) | ||||||
335 | return true; | ||||||
336 | |||||||
337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; | ||||||
338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; | ||||||
339 | |||||||
340 | if (MyNd.compare(&OtherNd)) | ||||||
341 | return true; | ||||||
342 | } | ||||||
343 | |||||||
344 | return false; | ||||||
345 | } | ||||||
346 | |||||||
347 | /// getNode - return the (Post)DominatorTree node for the specified basic | ||||||
348 | /// block. This is the same as using operator[] on this class. The result | ||||||
349 | /// may (but is not required to) be null for a forward (backwards) | ||||||
350 | /// statically unreachable block. | ||||||
351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { | ||||||
352 | auto I = DomTreeNodes.find(BB); | ||||||
353 | if (I != DomTreeNodes.end()) | ||||||
354 | return I->second.get(); | ||||||
355 | return nullptr; | ||||||
356 | } | ||||||
357 | |||||||
358 | /// See getNode. | ||||||
359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { | ||||||
360 | return getNode(BB); | ||||||
361 | } | ||||||
362 | |||||||
363 | /// getRootNode - This returns the entry node for the CFG of the function. If | ||||||
364 | /// this tree represents the post-dominance relations for a function, however, | ||||||
365 | /// this root may be a node with the block == NULL. This is the case when | ||||||
366 | /// there are multiple exit nodes from a particular function. Consumers of | ||||||
367 | /// post-dominance information must be capable of dealing with this | ||||||
368 | /// possibility. | ||||||
369 | /// | ||||||
370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } | ||||||
371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } | ||||||
372 | |||||||
373 | /// Get all nodes dominated by R, including R itself. | ||||||
374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { | ||||||
375 | Result.clear(); | ||||||
376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); | ||||||
377 | if (!RN) | ||||||
378 | return; // If R is unreachable, it will not be present in the DOM tree. | ||||||
379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; | ||||||
380 | WL.push_back(RN); | ||||||
381 | |||||||
382 | while (!WL.empty()) { | ||||||
383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); | ||||||
384 | Result.push_back(N->getBlock()); | ||||||
385 | WL.append(N->begin(), N->end()); | ||||||
386 | } | ||||||
387 | } | ||||||
388 | |||||||
389 | /// properlyDominates - Returns true iff A dominates B and A != B. | ||||||
390 | /// Note that this is not a constant time operation! | ||||||
391 | /// | ||||||
392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, | ||||||
393 | const DomTreeNodeBase<NodeT> *B) const { | ||||||
394 | if (!A || !B) | ||||||
395 | return false; | ||||||
396 | if (A == B) | ||||||
397 | return false; | ||||||
398 | return dominates(A, B); | ||||||
399 | } | ||||||
400 | |||||||
401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; | ||||||
402 | |||||||
403 | /// isReachableFromEntry - Return true if A is dominated by the entry | ||||||
404 | /// block of the function containing it. | ||||||
405 | bool isReachableFromEntry(const NodeT *A) const { | ||||||
406 | assert(!this->isPostDominator() &&((void)0) | ||||||
407 | "This is not implemented for post dominators")((void)0); | ||||||
408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); | ||||||
409 | } | ||||||
410 | |||||||
411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } | ||||||
412 | |||||||
413 | /// dominates - Returns true iff A dominates B. Note that this is not a | ||||||
414 | /// constant time operation! | ||||||
415 | /// | ||||||
416 | bool dominates(const DomTreeNodeBase<NodeT> *A, | ||||||
417 | const DomTreeNodeBase<NodeT> *B) const { | ||||||
418 | // A node trivially dominates itself. | ||||||
419 | if (B == A) | ||||||
420 | return true; | ||||||
421 | |||||||
422 | // An unreachable node is dominated by anything. | ||||||
423 | if (!isReachableFromEntry(B)) | ||||||
424 | return true; | ||||||
425 | |||||||
426 | // And dominates nothing. | ||||||
427 | if (!isReachableFromEntry(A)) | ||||||
428 | return false; | ||||||
429 | |||||||
430 | if (B->getIDom() == A) return true; | ||||||
431 | |||||||
432 | if (A->getIDom() == B) return false; | ||||||
433 | |||||||
434 | // A can only dominate B if it is higher in the tree. | ||||||
435 | if (A->getLevel() >= B->getLevel()) return false; | ||||||
436 | |||||||
437 | // Compare the result of the tree walk and the dfs numbers, if expensive | ||||||
438 | // checks are enabled. | ||||||
439 | #ifdef EXPENSIVE_CHECKS | ||||||
440 | assert((!DFSInfoValid ||((void)0) | ||||||
441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&((void)0) | ||||||
442 | "Tree walk disagrees with dfs numbers!")((void)0); | ||||||
443 | #endif | ||||||
444 | |||||||
445 | if (DFSInfoValid) | ||||||
446 | return B->DominatedBy(A); | ||||||
447 | |||||||
448 | // If we end up with too many slow queries, just update the | ||||||
449 | // DFS numbers on the theory that we are going to keep querying. | ||||||
450 | SlowQueries++; | ||||||
451 | if (SlowQueries > 32) { | ||||||
452 | updateDFSNumbers(); | ||||||
453 | return B->DominatedBy(A); | ||||||
454 | } | ||||||
455 | |||||||
456 | return dominatedBySlowTreeWalk(A, B); | ||||||
457 | } | ||||||
458 | |||||||
459 | bool dominates(const NodeT *A, const NodeT *B) const; | ||||||
460 | |||||||
461 | NodeT *getRoot() const { | ||||||
462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((void)0); | ||||||
463 | return this->Roots[0]; | ||||||
464 | } | ||||||
465 | |||||||
466 | /// Find nearest common dominator basic block for basic block A and B. A and B | ||||||
467 | /// must have tree nodes. | ||||||
468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { | ||||||
469 | assert(A && B && "Pointers are not valid")((void)0); | ||||||
470 | assert(A->getParent() == B->getParent() &&((void)0) | ||||||
471 | "Two blocks are not in same function")((void)0); | ||||||
472 | |||||||
473 | // If either A or B is a entry block then it is nearest common dominator | ||||||
474 | // (for forward-dominators). | ||||||
475 | if (!isPostDominator()) { | ||||||
476 | NodeT &Entry = A->getParent()->front(); | ||||||
477 | if (A == &Entry || B == &Entry) | ||||||
478 | return &Entry; | ||||||
479 | } | ||||||
480 | |||||||
481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); | ||||||
482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); | ||||||
483 | assert(NodeA && "A must be in the tree")((void)0); | ||||||
484 | assert(NodeB && "B must be in the tree")((void)0); | ||||||
485 | |||||||
486 | // Use level information to go up the tree until the levels match. Then | ||||||
487 | // continue going up til we arrive at the same node. | ||||||
488 | while (NodeA != NodeB) { | ||||||
489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); | ||||||
490 | |||||||
491 | NodeA = NodeA->IDom; | ||||||
492 | } | ||||||
493 | |||||||
494 | return NodeA->getBlock(); | ||||||
| |||||||
495 | } | ||||||
496 | |||||||
497 | const NodeT *findNearestCommonDominator(const NodeT *A, | ||||||
498 | const NodeT *B) const { | ||||||
499 | // Cast away the const qualifiers here. This is ok since | ||||||
500 | // const is re-introduced on the return type. | ||||||
501 | return findNearestCommonDominator(const_cast<NodeT *>(A), | ||||||
502 | const_cast<NodeT *>(B)); | ||||||
503 | } | ||||||
504 | |||||||
505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { | ||||||
506 | return isPostDominator() && !A->getBlock(); | ||||||
507 | } | ||||||
508 | |||||||
509 | //===--------------------------------------------------------------------===// | ||||||
510 | // API to update (Post)DominatorTree information based on modifications to | ||||||
511 | // the CFG... | ||||||
512 | |||||||
513 | /// Inform the dominator tree about a sequence of CFG edge insertions and | ||||||
514 | /// deletions and perform a batch update on the tree. | ||||||
515 | /// | ||||||
516 | /// This function should be used when there were multiple CFG updates after | ||||||
517 | /// the last dominator tree update. It takes care of performing the updates | ||||||
518 | /// in sync with the CFG and optimizes away the redundant operations that | ||||||
519 | /// cancel each other. | ||||||
520 | /// The functions expects the sequence of updates to be balanced. Eg.: | ||||||
521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because | ||||||
522 | /// logically it results in a single insertions. | ||||||
523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make | ||||||
524 | /// sense to insert the same edge twice. | ||||||
525 | /// | ||||||
526 | /// What's more, the functions assumes that it's safe to ask every node in the | ||||||
527 | /// CFG about its children and inverse children. This implies that deletions | ||||||
528 | /// of CFG edges must not delete the CFG nodes before calling this function. | ||||||
529 | /// | ||||||
530 | /// The applyUpdates function can reorder the updates and remove redundant | ||||||
531 | /// ones internally. The batch updater is also able to detect sequences of | ||||||
532 | /// zero and exactly one update -- it's optimized to do less work in these | ||||||
533 | /// cases. | ||||||
534 | /// | ||||||
535 | /// Note that for postdominators it automatically takes care of applying | ||||||
536 | /// updates on reverse edges internally (so there's no need to swap the | ||||||
537 | /// From and To pointers when constructing DominatorTree::UpdateType). | ||||||
538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> | ||||||
539 | /// with the same template parameter T. | ||||||
540 | /// | ||||||
541 | /// \param Updates An unordered sequence of updates to perform. The current | ||||||
542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | ||||||
543 | /// | ||||||
544 | void applyUpdates(ArrayRef<UpdateType> Updates) { | ||||||
545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( | ||||||
546 | Updates, /*ReverseApplyUpdates=*/true); | ||||||
547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); | ||||||
548 | } | ||||||
549 | |||||||
550 | /// \param Updates An unordered sequence of updates to perform. The current | ||||||
551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | ||||||
552 | /// \param PostViewUpdates An unordered sequence of update to perform in order | ||||||
553 | /// to obtain a post-view of the CFG. The DT will be updated assuming the | ||||||
554 | /// obtained PostViewCFG is the desired end state. | ||||||
555 | void applyUpdates(ArrayRef<UpdateType> Updates, | ||||||
556 | ArrayRef<UpdateType> PostViewUpdates) { | ||||||
557 | if (Updates.empty()) { | ||||||
558 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | ||||||
559 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); | ||||||
560 | } else { | ||||||
561 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in | ||||||
562 | // Updates need to be reversed, and match the direction in PostViewCFG. | ||||||
563 | // The PostViewCFG is created with updates reversed (equivalent to changes | ||||||
564 | // made to the CFG), so the PreViewCFG needs all the updates reverse | ||||||
565 | // applied. | ||||||
566 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); | ||||||
567 | append_range(AllUpdates, PostViewUpdates); | ||||||
568 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, | ||||||
569 | /*ReverseApplyUpdates=*/true); | ||||||
570 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | ||||||
571 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); | ||||||
572 | } | ||||||
573 | } | ||||||
574 | |||||||
575 | /// Inform the dominator tree about a CFG edge insertion and update the tree. | ||||||
576 | /// | ||||||
577 | /// This function has to be called just before or just after making the update | ||||||
578 | /// on the actual CFG. There cannot be any other updates that the dominator | ||||||
579 | /// tree doesn't know about. | ||||||
580 | /// | ||||||
581 | /// Note that for postdominators it automatically takes care of inserting | ||||||
582 | /// a reverse edge internally (so there's no need to swap the parameters). | ||||||
583 | /// | ||||||
584 | void insertEdge(NodeT *From, NodeT *To) { | ||||||
585 | assert(From)((void)0); | ||||||
586 | assert(To)((void)0); | ||||||
587 | assert(From->getParent() == Parent)((void)0); | ||||||
588 | assert(To->getParent() == Parent)((void)0); | ||||||
589 | DomTreeBuilder::InsertEdge(*this, From, To); | ||||||
590 | } | ||||||
591 | |||||||
592 | /// Inform the dominator tree about a CFG edge deletion and update the tree. | ||||||
593 | /// | ||||||
594 | /// This function has to be called just after making the update on the actual | ||||||
595 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in | ||||||
596 | /// DEBUG mode. There cannot be any other updates that the | ||||||
597 | /// dominator tree doesn't know about. | ||||||
598 | /// | ||||||
599 | /// Note that for postdominators it automatically takes care of deleting | ||||||
600 | /// a reverse edge internally (so there's no need to swap the parameters). | ||||||
601 | /// | ||||||
602 | void deleteEdge(NodeT *From, NodeT *To) { | ||||||
603 | assert(From)((void)0); | ||||||
604 | assert(To)((void)0); | ||||||
605 | assert(From->getParent() == Parent)((void)0); | ||||||
606 | assert(To->getParent() == Parent)((void)0); | ||||||
607 | DomTreeBuilder::DeleteEdge(*this, From, To); | ||||||
608 | } | ||||||
609 | |||||||
610 | /// Add a new node to the dominator tree information. | ||||||
611 | /// | ||||||
612 | /// This creates a new node as a child of DomBB dominator node, linking it | ||||||
613 | /// into the children list of the immediate dominator. | ||||||
614 | /// | ||||||
615 | /// \param BB New node in CFG. | ||||||
616 | /// \param DomBB CFG node that is dominator for BB. | ||||||
617 | /// \returns New dominator tree node that represents new CFG node. | ||||||
618 | /// | ||||||
619 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { | ||||||
620 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | ||||||
621 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); | ||||||
622 | assert(IDomNode && "Not immediate dominator specified for block!")((void)0); | ||||||
623 | DFSInfoValid = false; | ||||||
624 | return createChild(BB, IDomNode); | ||||||
625 | } | ||||||
626 | |||||||
627 | /// Add a new node to the forward dominator tree and make it a new root. | ||||||
628 | /// | ||||||
629 | /// \param BB New node in CFG. | ||||||
630 | /// \returns New dominator tree node that represents new CFG node. | ||||||
631 | /// | ||||||
632 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { | ||||||
633 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | ||||||
634 | assert(!this->isPostDominator() &&((void)0) | ||||||
635 | "Cannot change root of post-dominator tree")((void)0); | ||||||
636 | DFSInfoValid = false; | ||||||
637 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); | ||||||
638 | if (Roots.empty()) { | ||||||
639 | addRoot(BB); | ||||||
640 | } else { | ||||||
641 | assert(Roots.size() == 1)((void)0); | ||||||
642 | NodeT *OldRoot = Roots.front(); | ||||||
643 | auto &OldNode = DomTreeNodes[OldRoot]; | ||||||
644 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); | ||||||
645 | OldNode->IDom = NewNode; | ||||||
646 | OldNode->UpdateLevel(); | ||||||
647 | Roots[0] = BB; | ||||||
648 | } | ||||||
649 | return RootNode = NewNode; | ||||||
650 | } | ||||||
651 | |||||||
652 | /// changeImmediateDominator - This method is used to update the dominator | ||||||
653 | /// tree information when a node's immediate dominator changes. | ||||||
654 | /// | ||||||
655 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, | ||||||
656 | DomTreeNodeBase<NodeT> *NewIDom) { | ||||||
657 | assert(N && NewIDom && "Cannot change null node pointers!")((void)0); | ||||||
658 | DFSInfoValid = false; | ||||||
659 | N->setIDom(NewIDom); | ||||||
660 | } | ||||||
661 | |||||||
662 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { | ||||||
663 | changeImmediateDominator(getNode(BB), getNode(NewBB)); | ||||||
664 | } | ||||||
665 | |||||||
666 | /// eraseNode - Removes a node from the dominator tree. Block must not | ||||||
667 | /// dominate any other blocks. Removes node from its immediate dominator's | ||||||
668 | /// children list. Deletes dominator node associated with basic block BB. | ||||||
669 | void eraseNode(NodeT *BB) { | ||||||
670 | DomTreeNodeBase<NodeT> *Node = getNode(BB); | ||||||
671 | assert(Node && "Removing node that isn't in dominator tree.")((void)0); | ||||||
672 | assert(Node->isLeaf() && "Node is not a leaf node.")((void)0); | ||||||
673 | |||||||
674 | DFSInfoValid = false; | ||||||
675 | |||||||
676 | // Remove node from immediate dominator's children list. | ||||||
677 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); | ||||||
678 | if (IDom) { | ||||||
679 | const auto I = find(IDom->Children, Node); | ||||||
680 | assert(I != IDom->Children.end() &&((void)0) | ||||||
681 | "Not in immediate dominator children set!")((void)0); | ||||||
682 | // I am no longer your child... | ||||||
683 | IDom->Children.erase(I); | ||||||
684 | } | ||||||
685 | |||||||
686 | DomTreeNodes.erase(BB); | ||||||
687 | |||||||
688 | if (!IsPostDom) return; | ||||||
689 | |||||||
690 | // Remember to update PostDominatorTree roots. | ||||||
691 | auto RIt = llvm::find(Roots, BB); | ||||||
692 | if (RIt != Roots.end()) { | ||||||
693 | std::swap(*RIt, Roots.back()); | ||||||
694 | Roots.pop_back(); | ||||||
695 | } | ||||||
696 | } | ||||||
697 | |||||||
698 | /// splitBlock - BB is split and now it has one successor. Update dominator | ||||||
699 | /// tree to reflect this change. | ||||||
700 | void splitBlock(NodeT *NewBB) { | ||||||
701 | if (IsPostDominator
| ||||||
702 | Split<Inverse<NodeT *>>(NewBB); | ||||||
703 | else | ||||||
704 | Split<NodeT *>(NewBB); | ||||||
705 | } | ||||||
706 | |||||||
707 | /// print - Convert to human readable form | ||||||
708 | /// | ||||||
709 | void print(raw_ostream &O) const { | ||||||
710 | O << "=============================--------------------------------\n"; | ||||||
711 | if (IsPostDominator) | ||||||
712 | O << "Inorder PostDominator Tree: "; | ||||||
713 | else | ||||||
714 | O << "Inorder Dominator Tree: "; | ||||||
715 | if (!DFSInfoValid) | ||||||
716 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; | ||||||
717 | O << "\n"; | ||||||
718 | |||||||
719 | // The postdom tree can have a null root if there are no returns. | ||||||
720 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); | ||||||
721 | O << "Roots: "; | ||||||
722 | for (const NodePtr Block : Roots) { | ||||||
723 | Block->printAsOperand(O, false); | ||||||
724 | O << " "; | ||||||
725 | } | ||||||
726 | O << "\n"; | ||||||
727 | } | ||||||
728 | |||||||
729 | public: | ||||||
730 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking | ||||||
731 | /// dominator tree in dfs order. | ||||||
732 | void updateDFSNumbers() const { | ||||||
733 | if (DFSInfoValid) { | ||||||
734 | SlowQueries = 0; | ||||||
735 | return; | ||||||
736 | } | ||||||
737 | |||||||
738 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, | ||||||
739 | typename DomTreeNodeBase<NodeT>::const_iterator>, | ||||||
740 | 32> WorkStack; | ||||||
741 | |||||||
742 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); | ||||||
743 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")((void)0); | ||||||
744 | if (!ThisRoot) | ||||||
745 | return; | ||||||
746 | |||||||
747 | // Both dominators and postdominators have a single root node. In the case | ||||||
748 | // case of PostDominatorTree, this node is a virtual root. | ||||||
749 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); | ||||||
750 | |||||||
751 | unsigned DFSNum = 0; | ||||||
752 | ThisRoot->DFSNumIn = DFSNum++; | ||||||
753 | |||||||
754 | while (!WorkStack.empty()) { | ||||||
755 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; | ||||||
756 | const auto ChildIt = WorkStack.back().second; | ||||||
757 | |||||||
758 | // If we visited all of the children of this node, "recurse" back up the | ||||||
759 | // stack setting the DFOutNum. | ||||||
760 | if (ChildIt == Node->end()) { | ||||||
761 | Node->DFSNumOut = DFSNum++; | ||||||
762 | WorkStack.pop_back(); | ||||||
763 | } else { | ||||||
764 | // Otherwise, recursively visit this child. | ||||||
765 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; | ||||||
766 | ++WorkStack.back().second; | ||||||
767 | |||||||
768 | WorkStack.push_back({Child, Child->begin()}); | ||||||
769 | Child->DFSNumIn = DFSNum++; | ||||||
770 | } | ||||||
771 | } | ||||||
772 | |||||||
773 | SlowQueries = 0; | ||||||
774 | DFSInfoValid = true; | ||||||
775 | } | ||||||
776 | |||||||
777 | /// recalculate - compute a dominator tree for the given function | ||||||
778 | void recalculate(ParentType &Func) { | ||||||
779 | Parent = &Func; | ||||||
780 | DomTreeBuilder::Calculate(*this); | ||||||
781 | } | ||||||
782 | |||||||
783 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { | ||||||
784 | Parent = &Func; | ||||||
785 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); | ||||||
786 | } | ||||||
787 | |||||||
788 | /// verify - checks if the tree is correct. There are 3 level of verification: | ||||||
789 | /// - Full -- verifies if the tree is correct by making sure all the | ||||||
790 | /// properties (including the parent and the sibling property) | ||||||
791 | /// hold. | ||||||
792 | /// Takes O(N^3) time. | ||||||
793 | /// | ||||||
794 | /// - Basic -- checks if the tree is correct, but compares it to a freshly | ||||||
795 | /// constructed tree instead of checking the sibling property. | ||||||
796 | /// Takes O(N^2) time. | ||||||
797 | /// | ||||||
798 | /// - Fast -- checks basic tree structure and compares it with a freshly | ||||||
799 | /// constructed tree. | ||||||
800 | /// Takes O(N^2) time worst case, but is faster in practise (same | ||||||
801 | /// as tree construction). | ||||||
802 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { | ||||||
803 | return DomTreeBuilder::Verify(*this, VL); | ||||||
804 | } | ||||||
805 | |||||||
806 | void reset() { | ||||||
807 | DomTreeNodes.clear(); | ||||||
808 | Roots.clear(); | ||||||
809 | RootNode = nullptr; | ||||||
810 | Parent = nullptr; | ||||||
811 | DFSInfoValid = false; | ||||||
812 | SlowQueries = 0; | ||||||
813 | } | ||||||
814 | |||||||
815 | protected: | ||||||
816 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } | ||||||
817 | |||||||
818 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { | ||||||
819 | return (DomTreeNodes[BB] = IDom->addChild( | ||||||
820 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) | ||||||
821 | .get(); | ||||||
822 | } | ||||||
823 | |||||||
824 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { | ||||||
825 | return (DomTreeNodes[BB] = | ||||||
826 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) | ||||||
827 | .get(); | ||||||
828 | } | ||||||
829 | |||||||
830 | // NewBB is split and now it has one successor. Update dominator tree to | ||||||
831 | // reflect this change. | ||||||
832 | template <class N> | ||||||
833 | void Split(typename GraphTraits<N>::NodeRef NewBB) { | ||||||
834 | using GraphT = GraphTraits<N>; | ||||||
835 | using NodeRef = typename GraphT::NodeRef; | ||||||
836 | assert(std::distance(GraphT::child_begin(NewBB),((void)0) | ||||||
837 | GraphT::child_end(NewBB)) == 1 &&((void)0) | ||||||
838 | "NewBB should have a single successor!")((void)0); | ||||||
839 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); | ||||||
840 | |||||||
841 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); | ||||||
842 | |||||||
843 | assert(!PredBlocks.empty() && "No predblocks?")((void)0); | ||||||
844 | |||||||
845 | bool NewBBDominatesNewBBSucc = true; | ||||||
846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { | ||||||
847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && | ||||||
848 | isReachableFromEntry(Pred)) { | ||||||
849 | NewBBDominatesNewBBSucc = false; | ||||||
850 | break; | ||||||
851 | } | ||||||
852 | } | ||||||
853 | |||||||
854 | // Find NewBB's immediate dominator and create new dominator tree node for | ||||||
855 | // NewBB. | ||||||
856 | NodeT *NewBBIDom = nullptr; | ||||||
857 | unsigned i = 0; | ||||||
858 | for (i = 0; i < PredBlocks.size(); ++i) | ||||||
859 | if (isReachableFromEntry(PredBlocks[i])) { | ||||||
860 | NewBBIDom = PredBlocks[i]; | ||||||
861 | break; | ||||||
862 | } | ||||||
863 | |||||||
864 | // It's possible that none of the predecessors of NewBB are reachable; | ||||||
865 | // in that case, NewBB itself is unreachable, so nothing needs to be | ||||||
866 | // changed. | ||||||
867 | if (!NewBBIDom) return; | ||||||
868 | |||||||
869 | for (i = i + 1; i < PredBlocks.size(); ++i) { | ||||||
870 | if (isReachableFromEntry(PredBlocks[i])) | ||||||
871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); | ||||||
872 | } | ||||||
873 | |||||||
874 | // Create the new dominator tree node... and set the idom of NewBB. | ||||||
875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); | ||||||
876 | |||||||
877 | // If NewBB strictly dominates other blocks, then it is now the immediate | ||||||
878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. | ||||||
879 | if (NewBBDominatesNewBBSucc) { | ||||||
880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); | ||||||
881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); | ||||||
882 | } | ||||||
883 | } | ||||||
884 | |||||||
885 | private: | ||||||
886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, | ||||||
887 | const DomTreeNodeBase<NodeT> *B) const { | ||||||
888 | assert(A != B)((void)0); | ||||||
889 | assert(isReachableFromEntry(B))((void)0); | ||||||
890 | assert(isReachableFromEntry(A))((void)0); | ||||||
891 | |||||||
892 | const unsigned ALevel = A->getLevel(); | ||||||
893 | const DomTreeNodeBase<NodeT> *IDom; | ||||||
894 | |||||||
895 | // Don't walk nodes above A's subtree. When we reach A's level, we must | ||||||
896 | // either find A or be in some other subtree not dominated by A. | ||||||
897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) | ||||||
898 | B = IDom; // Walk up the tree | ||||||
899 | |||||||
900 | return B == A; | ||||||
901 | } | ||||||
902 | |||||||
903 | /// Wipe this tree's state without releasing any resources. | ||||||
904 | /// | ||||||
905 | /// This is essentially a post-move helper only. It leaves the object in an | ||||||
906 | /// assignable and destroyable state, but otherwise invalid. | ||||||
907 | void wipe() { | ||||||
908 | DomTreeNodes.clear(); | ||||||
909 | RootNode = nullptr; | ||||||
910 | Parent = nullptr; | ||||||
911 | } | ||||||
912 | }; | ||||||
913 | |||||||
914 | template <typename T> | ||||||
915 | using DomTreeBase = DominatorTreeBase<T, false>; | ||||||
916 | |||||||
917 | template <typename T> | ||||||
918 | using PostDomTreeBase = DominatorTreeBase<T, true>; | ||||||
919 | |||||||
920 | // These two functions are declared out of line as a workaround for building | ||||||
921 | // with old (< r147295) versions of clang because of pr11642. | ||||||
922 | template <typename NodeT, bool IsPostDom> | ||||||
923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, | ||||||
924 | const NodeT *B) const { | ||||||
925 | if (A == B) | ||||||
926 | return true; | ||||||
927 | |||||||
928 | // Cast away the const qualifiers here. This is ok since | ||||||
929 | // this function doesn't actually return the values returned | ||||||
930 | // from getNode. | ||||||
931 | return dominates(getNode(const_cast<NodeT *>(A)), | ||||||
932 | getNode(const_cast<NodeT *>(B))); | ||||||
933 | } | ||||||
934 | template <typename NodeT, bool IsPostDom> | ||||||
935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( | ||||||
936 | const NodeT *A, const NodeT *B) const { | ||||||
937 | if (A == B) | ||||||
938 | return false; | ||||||
939 | |||||||
940 | // Cast away the const qualifiers here. This is ok since | ||||||
941 | // this function doesn't actually return the values returned | ||||||
942 | // from getNode. | ||||||
943 | return dominates(getNode(const_cast<NodeT *>(A)), | ||||||
944 | getNode(const_cast<NodeT *>(B))); | ||||||
945 | } | ||||||
946 | |||||||
947 | } // end namespace llvm | ||||||
948 | |||||||
949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |
1 | //===- llvm/ADT/DenseMap.h - Dense probed hash table ------------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the DenseMap class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_ADT_DENSEMAP_H |
14 | #define LLVM_ADT_DENSEMAP_H |
15 | |
16 | #include "llvm/ADT/DenseMapInfo.h" |
17 | #include "llvm/ADT/EpochTracker.h" |
18 | #include "llvm/Support/AlignOf.h" |
19 | #include "llvm/Support/Compiler.h" |
20 | #include "llvm/Support/MathExtras.h" |
21 | #include "llvm/Support/MemAlloc.h" |
22 | #include "llvm/Support/ReverseIteration.h" |
23 | #include "llvm/Support/type_traits.h" |
24 | #include <algorithm> |
25 | #include <cassert> |
26 | #include <cstddef> |
27 | #include <cstring> |
28 | #include <initializer_list> |
29 | #include <iterator> |
30 | #include <new> |
31 | #include <type_traits> |
32 | #include <utility> |
33 | |
34 | namespace llvm { |
35 | |
36 | namespace detail { |
37 | |
38 | // We extend a pair to allow users to override the bucket type with their own |
39 | // implementation without requiring two members. |
40 | template <typename KeyT, typename ValueT> |
41 | struct DenseMapPair : public std::pair<KeyT, ValueT> { |
42 | using std::pair<KeyT, ValueT>::pair; |
43 | |
44 | KeyT &getFirst() { return std::pair<KeyT, ValueT>::first; } |
45 | const KeyT &getFirst() const { return std::pair<KeyT, ValueT>::first; } |
46 | ValueT &getSecond() { return std::pair<KeyT, ValueT>::second; } |
47 | const ValueT &getSecond() const { return std::pair<KeyT, ValueT>::second; } |
48 | }; |
49 | |
50 | } // end namespace detail |
51 | |
52 | template <typename KeyT, typename ValueT, |
53 | typename KeyInfoT = DenseMapInfo<KeyT>, |
54 | typename Bucket = llvm::detail::DenseMapPair<KeyT, ValueT>, |
55 | bool IsConst = false> |
56 | class DenseMapIterator; |
57 | |
58 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
59 | typename BucketT> |
60 | class DenseMapBase : public DebugEpochBase { |
61 | template <typename T> |
62 | using const_arg_type_t = typename const_pointer_or_const_ref<T>::type; |
63 | |
64 | public: |
65 | using size_type = unsigned; |
66 | using key_type = KeyT; |
67 | using mapped_type = ValueT; |
68 | using value_type = BucketT; |
69 | |
70 | using iterator = DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT>; |
71 | using const_iterator = |
72 | DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT, true>; |
73 | |
74 | inline iterator begin() { |
75 | // When the map is empty, avoid the overhead of advancing/retreating past |
76 | // empty buckets. |
77 | if (empty()) |
78 | return end(); |
79 | if (shouldReverseIterate<KeyT>()) |
80 | return makeIterator(getBucketsEnd() - 1, getBuckets(), *this); |
81 | return makeIterator(getBuckets(), getBucketsEnd(), *this); |
82 | } |
83 | inline iterator end() { |
84 | return makeIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
85 | } |
86 | inline const_iterator begin() const { |
87 | if (empty()) |
88 | return end(); |
89 | if (shouldReverseIterate<KeyT>()) |
90 | return makeConstIterator(getBucketsEnd() - 1, getBuckets(), *this); |
91 | return makeConstIterator(getBuckets(), getBucketsEnd(), *this); |
92 | } |
93 | inline const_iterator end() const { |
94 | return makeConstIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
95 | } |
96 | |
97 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { |
98 | return getNumEntries() == 0; |
99 | } |
100 | unsigned size() const { return getNumEntries(); } |
101 | |
102 | /// Grow the densemap so that it can contain at least \p NumEntries items |
103 | /// before resizing again. |
104 | void reserve(size_type NumEntries) { |
105 | auto NumBuckets = getMinBucketToReserveForEntries(NumEntries); |
106 | incrementEpoch(); |
107 | if (NumBuckets > getNumBuckets()) |
108 | grow(NumBuckets); |
109 | } |
110 | |
111 | void clear() { |
112 | incrementEpoch(); |
113 | if (getNumEntries() == 0 && getNumTombstones() == 0) return; |
114 | |
115 | // If the capacity of the array is huge, and the # elements used is small, |
116 | // shrink the array. |
117 | if (getNumEntries() * 4 < getNumBuckets() && getNumBuckets() > 64) { |
118 | shrink_and_clear(); |
119 | return; |
120 | } |
121 | |
122 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
123 | if (std::is_trivially_destructible<ValueT>::value) { |
124 | // Use a simpler loop when values don't need destruction. |
125 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) |
126 | P->getFirst() = EmptyKey; |
127 | } else { |
128 | unsigned NumEntries = getNumEntries(); |
129 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
130 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey)) { |
131 | if (!KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
132 | P->getSecond().~ValueT(); |
133 | --NumEntries; |
134 | } |
135 | P->getFirst() = EmptyKey; |
136 | } |
137 | } |
138 | assert(NumEntries == 0 && "Node count imbalance!")((void)0); |
139 | } |
140 | setNumEntries(0); |
141 | setNumTombstones(0); |
142 | } |
143 | |
144 | /// Return 1 if the specified key is in the map, 0 otherwise. |
145 | size_type count(const_arg_type_t<KeyT> Val) const { |
146 | const BucketT *TheBucket; |
147 | return LookupBucketFor(Val, TheBucket) ? 1 : 0; |
148 | } |
149 | |
150 | iterator find(const_arg_type_t<KeyT> Val) { |
151 | BucketT *TheBucket; |
152 | if (LookupBucketFor(Val, TheBucket)) |
153 | return makeIterator(TheBucket, |
154 | shouldReverseIterate<KeyT>() ? getBuckets() |
155 | : getBucketsEnd(), |
156 | *this, true); |
157 | return end(); |
158 | } |
159 | const_iterator find(const_arg_type_t<KeyT> Val) const { |
160 | const BucketT *TheBucket; |
161 | if (LookupBucketFor(Val, TheBucket)) |
162 | return makeConstIterator(TheBucket, |
163 | shouldReverseIterate<KeyT>() ? getBuckets() |
164 | : getBucketsEnd(), |
165 | *this, true); |
166 | return end(); |
167 | } |
168 | |
169 | /// Alternate version of find() which allows a different, and possibly |
170 | /// less expensive, key type. |
171 | /// The DenseMapInfo is responsible for supplying methods |
172 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
173 | /// type used. |
174 | template<class LookupKeyT> |
175 | iterator find_as(const LookupKeyT &Val) { |
176 | BucketT *TheBucket; |
177 | if (LookupBucketFor(Val, TheBucket)) |
178 | return makeIterator(TheBucket, |
179 | shouldReverseIterate<KeyT>() ? getBuckets() |
180 | : getBucketsEnd(), |
181 | *this, true); |
182 | return end(); |
183 | } |
184 | template<class LookupKeyT> |
185 | const_iterator find_as(const LookupKeyT &Val) const { |
186 | const BucketT *TheBucket; |
187 | if (LookupBucketFor(Val, TheBucket)) |
188 | return makeConstIterator(TheBucket, |
189 | shouldReverseIterate<KeyT>() ? getBuckets() |
190 | : getBucketsEnd(), |
191 | *this, true); |
192 | return end(); |
193 | } |
194 | |
195 | /// lookup - Return the entry for the specified key, or a default |
196 | /// constructed value if no such entry exists. |
197 | ValueT lookup(const_arg_type_t<KeyT> Val) const { |
198 | const BucketT *TheBucket; |
199 | if (LookupBucketFor(Val, TheBucket)) |
200 | return TheBucket->getSecond(); |
201 | return ValueT(); |
202 | } |
203 | |
204 | // Inserts key,value pair into the map if the key isn't already in the map. |
205 | // If the key is already in the map, it returns false and doesn't update the |
206 | // value. |
207 | std::pair<iterator, bool> insert(const std::pair<KeyT, ValueT> &KV) { |
208 | return try_emplace(KV.first, KV.second); |
209 | } |
210 | |
211 | // Inserts key,value pair into the map if the key isn't already in the map. |
212 | // If the key is already in the map, it returns false and doesn't update the |
213 | // value. |
214 | std::pair<iterator, bool> insert(std::pair<KeyT, ValueT> &&KV) { |
215 | return try_emplace(std::move(KV.first), std::move(KV.second)); |
216 | } |
217 | |
218 | // Inserts key,value pair into the map if the key isn't already in the map. |
219 | // The value is constructed in-place if the key is not in the map, otherwise |
220 | // it is not moved. |
221 | template <typename... Ts> |
222 | std::pair<iterator, bool> try_emplace(KeyT &&Key, Ts &&... Args) { |
223 | BucketT *TheBucket; |
224 | if (LookupBucketFor(Key, TheBucket)) |
225 | return std::make_pair(makeIterator(TheBucket, |
226 | shouldReverseIterate<KeyT>() |
227 | ? getBuckets() |
228 | : getBucketsEnd(), |
229 | *this, true), |
230 | false); // Already in map. |
231 | |
232 | // Otherwise, insert the new element. |
233 | TheBucket = |
234 | InsertIntoBucket(TheBucket, std::move(Key), std::forward<Ts>(Args)...); |
235 | return std::make_pair(makeIterator(TheBucket, |
236 | shouldReverseIterate<KeyT>() |
237 | ? getBuckets() |
238 | : getBucketsEnd(), |
239 | *this, true), |
240 | true); |
241 | } |
242 | |
243 | // Inserts key,value pair into the map if the key isn't already in the map. |
244 | // The value is constructed in-place if the key is not in the map, otherwise |
245 | // it is not moved. |
246 | template <typename... Ts> |
247 | std::pair<iterator, bool> try_emplace(const KeyT &Key, Ts &&... Args) { |
248 | BucketT *TheBucket; |
249 | if (LookupBucketFor(Key, TheBucket)) |
250 | return std::make_pair(makeIterator(TheBucket, |
251 | shouldReverseIterate<KeyT>() |
252 | ? getBuckets() |
253 | : getBucketsEnd(), |
254 | *this, true), |
255 | false); // Already in map. |
256 | |
257 | // Otherwise, insert the new element. |
258 | TheBucket = InsertIntoBucket(TheBucket, Key, std::forward<Ts>(Args)...); |
259 | return std::make_pair(makeIterator(TheBucket, |
260 | shouldReverseIterate<KeyT>() |
261 | ? getBuckets() |
262 | : getBucketsEnd(), |
263 | *this, true), |
264 | true); |
265 | } |
266 | |
267 | /// Alternate version of insert() which allows a different, and possibly |
268 | /// less expensive, key type. |
269 | /// The DenseMapInfo is responsible for supplying methods |
270 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
271 | /// type used. |
272 | template <typename LookupKeyT> |
273 | std::pair<iterator, bool> insert_as(std::pair<KeyT, ValueT> &&KV, |
274 | const LookupKeyT &Val) { |
275 | BucketT *TheBucket; |
276 | if (LookupBucketFor(Val, TheBucket)) |
277 | return std::make_pair(makeIterator(TheBucket, |
278 | shouldReverseIterate<KeyT>() |
279 | ? getBuckets() |
280 | : getBucketsEnd(), |
281 | *this, true), |
282 | false); // Already in map. |
283 | |
284 | // Otherwise, insert the new element. |
285 | TheBucket = InsertIntoBucketWithLookup(TheBucket, std::move(KV.first), |
286 | std::move(KV.second), Val); |
287 | return std::make_pair(makeIterator(TheBucket, |
288 | shouldReverseIterate<KeyT>() |
289 | ? getBuckets() |
290 | : getBucketsEnd(), |
291 | *this, true), |
292 | true); |
293 | } |
294 | |
295 | /// insert - Range insertion of pairs. |
296 | template<typename InputIt> |
297 | void insert(InputIt I, InputIt E) { |
298 | for (; I != E; ++I) |
299 | insert(*I); |
300 | } |
301 | |
302 | bool erase(const KeyT &Val) { |
303 | BucketT *TheBucket; |
304 | if (!LookupBucketFor(Val, TheBucket)) |
305 | return false; // not in map. |
306 | |
307 | TheBucket->getSecond().~ValueT(); |
308 | TheBucket->getFirst() = getTombstoneKey(); |
309 | decrementNumEntries(); |
310 | incrementNumTombstones(); |
311 | return true; |
312 | } |
313 | void erase(iterator I) { |
314 | BucketT *TheBucket = &*I; |
315 | TheBucket->getSecond().~ValueT(); |
316 | TheBucket->getFirst() = getTombstoneKey(); |
317 | decrementNumEntries(); |
318 | incrementNumTombstones(); |
319 | } |
320 | |
321 | value_type& FindAndConstruct(const KeyT &Key) { |
322 | BucketT *TheBucket; |
323 | if (LookupBucketFor(Key, TheBucket)) |
324 | return *TheBucket; |
325 | |
326 | return *InsertIntoBucket(TheBucket, Key); |
327 | } |
328 | |
329 | ValueT &operator[](const KeyT &Key) { |
330 | return FindAndConstruct(Key).second; |
331 | } |
332 | |
333 | value_type& FindAndConstruct(KeyT &&Key) { |
334 | BucketT *TheBucket; |
335 | if (LookupBucketFor(Key, TheBucket)) |
336 | return *TheBucket; |
337 | |
338 | return *InsertIntoBucket(TheBucket, std::move(Key)); |
339 | } |
340 | |
341 | ValueT &operator[](KeyT &&Key) { |
342 | return FindAndConstruct(std::move(Key)).second; |
343 | } |
344 | |
345 | /// isPointerIntoBucketsArray - Return true if the specified pointer points |
346 | /// somewhere into the DenseMap's array of buckets (i.e. either to a key or |
347 | /// value in the DenseMap). |
348 | bool isPointerIntoBucketsArray(const void *Ptr) const { |
349 | return Ptr >= getBuckets() && Ptr < getBucketsEnd(); |
350 | } |
351 | |
352 | /// getPointerIntoBucketsArray() - Return an opaque pointer into the buckets |
353 | /// array. In conjunction with the previous method, this can be used to |
354 | /// determine whether an insertion caused the DenseMap to reallocate. |
355 | const void *getPointerIntoBucketsArray() const { return getBuckets(); } |
356 | |
357 | protected: |
358 | DenseMapBase() = default; |
359 | |
360 | void destroyAll() { |
361 | if (getNumBuckets() == 0) // Nothing to do. |
362 | return; |
363 | |
364 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
365 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
366 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
367 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) |
368 | P->getSecond().~ValueT(); |
369 | P->getFirst().~KeyT(); |
370 | } |
371 | } |
372 | |
373 | void initEmpty() { |
374 | setNumEntries(0); |
375 | setNumTombstones(0); |
376 | |
377 | assert((getNumBuckets() & (getNumBuckets()-1)) == 0 &&((void)0) |
378 | "# initial buckets must be a power of two!")((void)0); |
379 | const KeyT EmptyKey = getEmptyKey(); |
380 | for (BucketT *B = getBuckets(), *E = getBucketsEnd(); B != E; ++B) |
381 | ::new (&B->getFirst()) KeyT(EmptyKey); |
382 | } |
383 | |
384 | /// Returns the number of buckets to allocate to ensure that the DenseMap can |
385 | /// accommodate \p NumEntries without need to grow(). |
386 | unsigned getMinBucketToReserveForEntries(unsigned NumEntries) { |
387 | // Ensure that "NumEntries * 4 < NumBuckets * 3" |
388 | if (NumEntries == 0) |
389 | return 0; |
390 | // +1 is required because of the strict equality. |
391 | // For example if NumEntries is 48, we need to return 401. |
392 | return NextPowerOf2(NumEntries * 4 / 3 + 1); |
393 | } |
394 | |
395 | void moveFromOldBuckets(BucketT *OldBucketsBegin, BucketT *OldBucketsEnd) { |
396 | initEmpty(); |
397 | |
398 | // Insert all the old elements. |
399 | const KeyT EmptyKey = getEmptyKey(); |
400 | const KeyT TombstoneKey = getTombstoneKey(); |
401 | for (BucketT *B = OldBucketsBegin, *E = OldBucketsEnd; B != E; ++B) { |
402 | if (!KeyInfoT::isEqual(B->getFirst(), EmptyKey) && |
403 | !KeyInfoT::isEqual(B->getFirst(), TombstoneKey)) { |
404 | // Insert the key/value into the new table. |
405 | BucketT *DestBucket; |
406 | bool FoundVal = LookupBucketFor(B->getFirst(), DestBucket); |
407 | (void)FoundVal; // silence warning. |
408 | assert(!FoundVal && "Key already in new map?")((void)0); |
409 | DestBucket->getFirst() = std::move(B->getFirst()); |
410 | ::new (&DestBucket->getSecond()) ValueT(std::move(B->getSecond())); |
411 | incrementNumEntries(); |
412 | |
413 | // Free the value. |
414 | B->getSecond().~ValueT(); |
415 | } |
416 | B->getFirst().~KeyT(); |
417 | } |
418 | } |
419 | |
420 | template <typename OtherBaseT> |
421 | void copyFrom( |
422 | const DenseMapBase<OtherBaseT, KeyT, ValueT, KeyInfoT, BucketT> &other) { |
423 | assert(&other != this)((void)0); |
424 | assert(getNumBuckets() == other.getNumBuckets())((void)0); |
425 | |
426 | setNumEntries(other.getNumEntries()); |
427 | setNumTombstones(other.getNumTombstones()); |
428 | |
429 | if (std::is_trivially_copyable<KeyT>::value && |
430 | std::is_trivially_copyable<ValueT>::value) |
431 | memcpy(reinterpret_cast<void *>(getBuckets()), other.getBuckets(), |
432 | getNumBuckets() * sizeof(BucketT)); |
433 | else |
434 | for (size_t i = 0; i < getNumBuckets(); ++i) { |
435 | ::new (&getBuckets()[i].getFirst()) |
436 | KeyT(other.getBuckets()[i].getFirst()); |
437 | if (!KeyInfoT::isEqual(getBuckets()[i].getFirst(), getEmptyKey()) && |
438 | !KeyInfoT::isEqual(getBuckets()[i].getFirst(), getTombstoneKey())) |
439 | ::new (&getBuckets()[i].getSecond()) |
440 | ValueT(other.getBuckets()[i].getSecond()); |
441 | } |
442 | } |
443 | |
444 | static unsigned getHashValue(const KeyT &Val) { |
445 | return KeyInfoT::getHashValue(Val); |
446 | } |
447 | |
448 | template<typename LookupKeyT> |
449 | static unsigned getHashValue(const LookupKeyT &Val) { |
450 | return KeyInfoT::getHashValue(Val); |
451 | } |
452 | |
453 | static const KeyT getEmptyKey() { |
454 | static_assert(std::is_base_of<DenseMapBase, DerivedT>::value, |
455 | "Must pass the derived type to this template!"); |
456 | return KeyInfoT::getEmptyKey(); |
457 | } |
458 | |
459 | static const KeyT getTombstoneKey() { |
460 | return KeyInfoT::getTombstoneKey(); |
461 | } |
462 | |
463 | private: |
464 | iterator makeIterator(BucketT *P, BucketT *E, |
465 | DebugEpochBase &Epoch, |
466 | bool NoAdvance=false) { |
467 | if (shouldReverseIterate<KeyT>()) { |
468 | BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
469 | return iterator(B, E, Epoch, NoAdvance); |
470 | } |
471 | return iterator(P, E, Epoch, NoAdvance); |
472 | } |
473 | |
474 | const_iterator makeConstIterator(const BucketT *P, const BucketT *E, |
475 | const DebugEpochBase &Epoch, |
476 | const bool NoAdvance=false) const { |
477 | if (shouldReverseIterate<KeyT>()) { |
478 | const BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
479 | return const_iterator(B, E, Epoch, NoAdvance); |
480 | } |
481 | return const_iterator(P, E, Epoch, NoAdvance); |
482 | } |
483 | |
484 | unsigned getNumEntries() const { |
485 | return static_cast<const DerivedT *>(this)->getNumEntries(); |
486 | } |
487 | |
488 | void setNumEntries(unsigned Num) { |
489 | static_cast<DerivedT *>(this)->setNumEntries(Num); |
490 | } |
491 | |
492 | void incrementNumEntries() { |
493 | setNumEntries(getNumEntries() + 1); |
494 | } |
495 | |
496 | void decrementNumEntries() { |
497 | setNumEntries(getNumEntries() - 1); |
498 | } |
499 | |
500 | unsigned getNumTombstones() const { |
501 | return static_cast<const DerivedT *>(this)->getNumTombstones(); |
502 | } |
503 | |
504 | void setNumTombstones(unsigned Num) { |
505 | static_cast<DerivedT *>(this)->setNumTombstones(Num); |
506 | } |
507 | |
508 | void incrementNumTombstones() { |
509 | setNumTombstones(getNumTombstones() + 1); |
510 | } |
511 | |
512 | void decrementNumTombstones() { |
513 | setNumTombstones(getNumTombstones() - 1); |
514 | } |
515 | |
516 | const BucketT *getBuckets() const { |
517 | return static_cast<const DerivedT *>(this)->getBuckets(); |
518 | } |
519 | |
520 | BucketT *getBuckets() { |
521 | return static_cast<DerivedT *>(this)->getBuckets(); |
522 | } |
523 | |
524 | unsigned getNumBuckets() const { |
525 | return static_cast<const DerivedT *>(this)->getNumBuckets(); |
526 | } |
527 | |
528 | BucketT *getBucketsEnd() { |
529 | return getBuckets() + getNumBuckets(); |
530 | } |
531 | |
532 | const BucketT *getBucketsEnd() const { |
533 | return getBuckets() + getNumBuckets(); |
534 | } |
535 | |
536 | void grow(unsigned AtLeast) { |
537 | static_cast<DerivedT *>(this)->grow(AtLeast); |
538 | } |
539 | |
540 | void shrink_and_clear() { |
541 | static_cast<DerivedT *>(this)->shrink_and_clear(); |
542 | } |
543 | |
544 | template <typename KeyArg, typename... ValueArgs> |
545 | BucketT *InsertIntoBucket(BucketT *TheBucket, KeyArg &&Key, |
546 | ValueArgs &&... Values) { |
547 | TheBucket = InsertIntoBucketImpl(Key, Key, TheBucket); |
548 | |
549 | TheBucket->getFirst() = std::forward<KeyArg>(Key); |
550 | ::new (&TheBucket->getSecond()) ValueT(std::forward<ValueArgs>(Values)...); |
551 | return TheBucket; |
552 | } |
553 | |
554 | template <typename LookupKeyT> |
555 | BucketT *InsertIntoBucketWithLookup(BucketT *TheBucket, KeyT &&Key, |
556 | ValueT &&Value, LookupKeyT &Lookup) { |
557 | TheBucket = InsertIntoBucketImpl(Key, Lookup, TheBucket); |
558 | |
559 | TheBucket->getFirst() = std::move(Key); |
560 | ::new (&TheBucket->getSecond()) ValueT(std::move(Value)); |
561 | return TheBucket; |
562 | } |
563 | |
564 | template <typename LookupKeyT> |
565 | BucketT *InsertIntoBucketImpl(const KeyT &Key, const LookupKeyT &Lookup, |
566 | BucketT *TheBucket) { |
567 | incrementEpoch(); |
568 | |
569 | // If the load of the hash table is more than 3/4, or if fewer than 1/8 of |
570 | // the buckets are empty (meaning that many are filled with tombstones), |
571 | // grow the table. |
572 | // |
573 | // The later case is tricky. For example, if we had one empty bucket with |
574 | // tons of tombstones, failing lookups (e.g. for insertion) would have to |
575 | // probe almost the entire table until it found the empty bucket. If the |
576 | // table completely filled with tombstones, no lookup would ever succeed, |
577 | // causing infinite loops in lookup. |
578 | unsigned NewNumEntries = getNumEntries() + 1; |
579 | unsigned NumBuckets = getNumBuckets(); |
580 | if (LLVM_UNLIKELY(NewNumEntries * 4 >= NumBuckets * 3)__builtin_expect((bool)(NewNumEntries * 4 >= NumBuckets * 3 ), false)) { |
581 | this->grow(NumBuckets * 2); |
582 | LookupBucketFor(Lookup, TheBucket); |
583 | NumBuckets = getNumBuckets(); |
584 | } else if (LLVM_UNLIKELY(NumBuckets-(NewNumEntries+getNumTombstones()) <=__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false) |
585 | NumBuckets/8)__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false)) { |
586 | this->grow(NumBuckets); |
587 | LookupBucketFor(Lookup, TheBucket); |
588 | } |
589 | assert(TheBucket)((void)0); |
590 | |
591 | // Only update the state after we've grown our bucket space appropriately |
592 | // so that when growing buckets we have self-consistent entry count. |
593 | incrementNumEntries(); |
594 | |
595 | // If we are writing over a tombstone, remember this. |
596 | const KeyT EmptyKey = getEmptyKey(); |
597 | if (!KeyInfoT::isEqual(TheBucket->getFirst(), EmptyKey)) |
598 | decrementNumTombstones(); |
599 | |
600 | return TheBucket; |
601 | } |
602 | |
603 | /// LookupBucketFor - Lookup the appropriate bucket for Val, returning it in |
604 | /// FoundBucket. If the bucket contains the key and a value, this returns |
605 | /// true, otherwise it returns a bucket with an empty marker or tombstone and |
606 | /// returns false. |
607 | template<typename LookupKeyT> |
608 | bool LookupBucketFor(const LookupKeyT &Val, |
609 | const BucketT *&FoundBucket) const { |
610 | const BucketT *BucketsPtr = getBuckets(); |
611 | const unsigned NumBuckets = getNumBuckets(); |
612 | |
613 | if (NumBuckets == 0) { |
614 | FoundBucket = nullptr; |
615 | return false; |
616 | } |
617 | |
618 | // FoundTombstone - Keep track of whether we find a tombstone while probing. |
619 | const BucketT *FoundTombstone = nullptr; |
620 | const KeyT EmptyKey = getEmptyKey(); |
621 | const KeyT TombstoneKey = getTombstoneKey(); |
622 | assert(!KeyInfoT::isEqual(Val, EmptyKey) &&((void)0) |
623 | !KeyInfoT::isEqual(Val, TombstoneKey) &&((void)0) |
624 | "Empty/Tombstone value shouldn't be inserted into map!")((void)0); |
625 | |
626 | unsigned BucketNo = getHashValue(Val) & (NumBuckets-1); |
627 | unsigned ProbeAmt = 1; |
628 | while (true) { |
629 | const BucketT *ThisBucket = BucketsPtr + BucketNo; |
630 | // Found Val's bucket? If so, return it. |
631 | if (LLVM_LIKELY(KeyInfoT::isEqual(Val, ThisBucket->getFirst()))__builtin_expect((bool)(KeyInfoT::isEqual(Val, ThisBucket-> getFirst())), true)) { |
632 | FoundBucket = ThisBucket; |
633 | return true; |
634 | } |
635 | |
636 | // If we found an empty bucket, the key doesn't exist in the set. |
637 | // Insert it and return the default value. |
638 | if (LLVM_LIKELY(KeyInfoT::isEqual(ThisBucket->getFirst(), EmptyKey))__builtin_expect((bool)(KeyInfoT::isEqual(ThisBucket->getFirst (), EmptyKey)), true)) { |
639 | // If we've already seen a tombstone while probing, fill it in instead |
640 | // of the empty bucket we eventually probed to. |
641 | FoundBucket = FoundTombstone ? FoundTombstone : ThisBucket; |
642 | return false; |
643 | } |
644 | |
645 | // If this is a tombstone, remember it. If Val ends up not in the map, we |
646 | // prefer to return it than something that would require more probing. |
647 | if (KeyInfoT::isEqual(ThisBucket->getFirst(), TombstoneKey) && |
648 | !FoundTombstone) |
649 | FoundTombstone = ThisBucket; // Remember the first tombstone found. |
650 | |
651 | // Otherwise, it's a hash collision or a tombstone, continue quadratic |
652 | // probing. |
653 | BucketNo += ProbeAmt++; |
654 | BucketNo &= (NumBuckets-1); |
655 | } |
656 | } |
657 | |
658 | template <typename LookupKeyT> |
659 | bool LookupBucketFor(const LookupKeyT &Val, BucketT *&FoundBucket) { |
660 | const BucketT *ConstFoundBucket; |
661 | bool Result = const_cast<const DenseMapBase *>(this) |
662 | ->LookupBucketFor(Val, ConstFoundBucket); |
663 | FoundBucket = const_cast<BucketT *>(ConstFoundBucket); |
664 | return Result; |
665 | } |
666 | |
667 | public: |
668 | /// Return the approximate size (in bytes) of the actual map. |
669 | /// This is just the raw memory used by DenseMap. |
670 | /// If entries are pointers to objects, the size of the referenced objects |
671 | /// are not included. |
672 | size_t getMemorySize() const { |
673 | return getNumBuckets() * sizeof(BucketT); |
674 | } |
675 | }; |
676 | |
677 | /// Equality comparison for DenseMap. |
678 | /// |
679 | /// Iterates over elements of LHS confirming that each (key, value) pair in LHS |
680 | /// is also in RHS, and that no additional pairs are in RHS. |
681 | /// Equivalent to N calls to RHS.find and N value comparisons. Amortized |
682 | /// complexity is linear, worst case is O(N^2) (if every hash collides). |
683 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
684 | typename BucketT> |
685 | bool operator==( |
686 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
687 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
688 | if (LHS.size() != RHS.size()) |
689 | return false; |
690 | |
691 | for (auto &KV : LHS) { |
692 | auto I = RHS.find(KV.first); |
693 | if (I == RHS.end() || I->second != KV.second) |
694 | return false; |
695 | } |
696 | |
697 | return true; |
698 | } |
699 | |
700 | /// Inequality comparison for DenseMap. |
701 | /// |
702 | /// Equivalent to !(LHS == RHS). See operator== for performance notes. |
703 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
704 | typename BucketT> |
705 | bool operator!=( |
706 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
707 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
708 | return !(LHS == RHS); |
709 | } |
710 | |
711 | template <typename KeyT, typename ValueT, |
712 | typename KeyInfoT = DenseMapInfo<KeyT>, |
713 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
714 | class DenseMap : public DenseMapBase<DenseMap<KeyT, ValueT, KeyInfoT, BucketT>, |
715 | KeyT, ValueT, KeyInfoT, BucketT> { |
716 | friend class DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
717 | |
718 | // Lift some types from the dependent base class into this class for |
719 | // simplicity of referring to them. |
720 | using BaseT = DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
721 | |
722 | BucketT *Buckets; |
723 | unsigned NumEntries; |
724 | unsigned NumTombstones; |
725 | unsigned NumBuckets; |
726 | |
727 | public: |
728 | /// Create a DenseMap with an optional \p InitialReserve that guarantee that |
729 | /// this number of elements can be inserted in the map without grow() |
730 | explicit DenseMap(unsigned InitialReserve = 0) { init(InitialReserve); } |
731 | |
732 | DenseMap(const DenseMap &other) : BaseT() { |
733 | init(0); |
734 | copyFrom(other); |
735 | } |
736 | |
737 | DenseMap(DenseMap &&other) : BaseT() { |
738 | init(0); |
739 | swap(other); |
740 | } |
741 | |
742 | template<typename InputIt> |
743 | DenseMap(const InputIt &I, const InputIt &E) { |
744 | init(std::distance(I, E)); |
745 | this->insert(I, E); |
746 | } |
747 | |
748 | DenseMap(std::initializer_list<typename BaseT::value_type> Vals) { |
749 | init(Vals.size()); |
750 | this->insert(Vals.begin(), Vals.end()); |
751 | } |
752 | |
753 | ~DenseMap() { |
754 | this->destroyAll(); |
755 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
756 | } |
757 | |
758 | void swap(DenseMap& RHS) { |
759 | this->incrementEpoch(); |
760 | RHS.incrementEpoch(); |
761 | std::swap(Buckets, RHS.Buckets); |
762 | std::swap(NumEntries, RHS.NumEntries); |
763 | std::swap(NumTombstones, RHS.NumTombstones); |
764 | std::swap(NumBuckets, RHS.NumBuckets); |
765 | } |
766 | |
767 | DenseMap& operator=(const DenseMap& other) { |
768 | if (&other != this) |
769 | copyFrom(other); |
770 | return *this; |
771 | } |
772 | |
773 | DenseMap& operator=(DenseMap &&other) { |
774 | this->destroyAll(); |
775 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
776 | init(0); |
777 | swap(other); |
778 | return *this; |
779 | } |
780 | |
781 | void copyFrom(const DenseMap& other) { |
782 | this->destroyAll(); |
783 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
784 | if (allocateBuckets(other.NumBuckets)) { |
785 | this->BaseT::copyFrom(other); |
786 | } else { |
787 | NumEntries = 0; |
788 | NumTombstones = 0; |
789 | } |
790 | } |
791 | |
792 | void init(unsigned InitNumEntries) { |
793 | auto InitBuckets = BaseT::getMinBucketToReserveForEntries(InitNumEntries); |
794 | if (allocateBuckets(InitBuckets)) { |
795 | this->BaseT::initEmpty(); |
796 | } else { |
797 | NumEntries = 0; |
798 | NumTombstones = 0; |
799 | } |
800 | } |
801 | |
802 | void grow(unsigned AtLeast) { |
803 | unsigned OldNumBuckets = NumBuckets; |
804 | BucketT *OldBuckets = Buckets; |
805 | |
806 | allocateBuckets(std::max<unsigned>(64, static_cast<unsigned>(NextPowerOf2(AtLeast-1)))); |
807 | assert(Buckets)((void)0); |
808 | if (!OldBuckets) { |
809 | this->BaseT::initEmpty(); |
810 | return; |
811 | } |
812 | |
813 | this->moveFromOldBuckets(OldBuckets, OldBuckets+OldNumBuckets); |
814 | |
815 | // Free the old table. |
816 | deallocate_buffer(OldBuckets, sizeof(BucketT) * OldNumBuckets, |
817 | alignof(BucketT)); |
818 | } |
819 | |
820 | void shrink_and_clear() { |
821 | unsigned OldNumBuckets = NumBuckets; |
822 | unsigned OldNumEntries = NumEntries; |
823 | this->destroyAll(); |
824 | |
825 | // Reduce the number of buckets. |
826 | unsigned NewNumBuckets = 0; |
827 | if (OldNumEntries) |
828 | NewNumBuckets = std::max(64, 1 << (Log2_32_Ceil(OldNumEntries) + 1)); |
829 | if (NewNumBuckets == NumBuckets) { |
830 | this->BaseT::initEmpty(); |
831 | return; |
832 | } |
833 | |
834 | deallocate_buffer(Buckets, sizeof(BucketT) * OldNumBuckets, |
835 | alignof(BucketT)); |
836 | init(NewNumBuckets); |
837 | } |
838 | |
839 | private: |
840 | unsigned getNumEntries() const { |
841 | return NumEntries; |
842 | } |
843 | |
844 | void setNumEntries(unsigned Num) { |
845 | NumEntries = Num; |
846 | } |
847 | |
848 | unsigned getNumTombstones() const { |
849 | return NumTombstones; |
850 | } |
851 | |
852 | void setNumTombstones(unsigned Num) { |
853 | NumTombstones = Num; |
854 | } |
855 | |
856 | BucketT *getBuckets() const { |
857 | return Buckets; |
858 | } |
859 | |
860 | unsigned getNumBuckets() const { |
861 | return NumBuckets; |
862 | } |
863 | |
864 | bool allocateBuckets(unsigned Num) { |
865 | NumBuckets = Num; |
866 | if (NumBuckets == 0) { |
867 | Buckets = nullptr; |
868 | return false; |
869 | } |
870 | |
871 | Buckets = static_cast<BucketT *>( |
872 | allocate_buffer(sizeof(BucketT) * NumBuckets, alignof(BucketT))); |
873 | return true; |
874 | } |
875 | }; |
876 | |
877 | template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4, |
878 | typename KeyInfoT = DenseMapInfo<KeyT>, |
879 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
880 | class SmallDenseMap |
881 | : public DenseMapBase< |
882 | SmallDenseMap<KeyT, ValueT, InlineBuckets, KeyInfoT, BucketT>, KeyT, |
883 | ValueT, KeyInfoT, BucketT> { |
884 | friend class DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
885 | |
886 | // Lift some types from the dependent base class into this class for |
887 | // simplicity of referring to them. |
888 | using BaseT = DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
889 | |
890 | static_assert(isPowerOf2_64(InlineBuckets), |
891 | "InlineBuckets must be a power of 2."); |
892 | |
893 | unsigned Small : 1; |
894 | unsigned NumEntries : 31; |
895 | unsigned NumTombstones; |
896 | |
897 | struct LargeRep { |
898 | BucketT *Buckets; |
899 | unsigned NumBuckets; |
900 | }; |
901 | |
902 | /// A "union" of an inline bucket array and the struct representing |
903 | /// a large bucket. This union will be discriminated by the 'Small' bit. |
904 | AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage; |
905 | |
906 | public: |
907 | explicit SmallDenseMap(unsigned NumInitBuckets = 0) { |
908 | init(NumInitBuckets); |
909 | } |
910 | |
911 | SmallDenseMap(const SmallDenseMap &other) : BaseT() { |
912 | init(0); |
913 | copyFrom(other); |
914 | } |
915 | |
916 | SmallDenseMap(SmallDenseMap &&other) : BaseT() { |
917 | init(0); |
918 | swap(other); |
919 | } |
920 | |
921 | template<typename InputIt> |
922 | SmallDenseMap(const InputIt &I, const InputIt &E) { |
923 | init(NextPowerOf2(std::distance(I, E))); |
924 | this->insert(I, E); |
925 | } |
926 | |
927 | SmallDenseMap(std::initializer_list<typename BaseT::value_type> Vals) |
928 | : SmallDenseMap(Vals.begin(), Vals.end()) {} |
929 | |
930 | ~SmallDenseMap() { |
931 | this->destroyAll(); |
932 | deallocateBuckets(); |
933 | } |
934 | |
935 | void swap(SmallDenseMap& RHS) { |
936 | unsigned TmpNumEntries = RHS.NumEntries; |
937 | RHS.NumEntries = NumEntries; |
938 | NumEntries = TmpNumEntries; |
939 | std::swap(NumTombstones, RHS.NumTombstones); |
940 | |
941 | const KeyT EmptyKey = this->getEmptyKey(); |
942 | const KeyT TombstoneKey = this->getTombstoneKey(); |
943 | if (Small && RHS.Small) { |
944 | // If we're swapping inline bucket arrays, we have to cope with some of |
945 | // the tricky bits of DenseMap's storage system: the buckets are not |
946 | // fully initialized. Thus we swap every key, but we may have |
947 | // a one-directional move of the value. |
948 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
949 | BucketT *LHSB = &getInlineBuckets()[i], |
950 | *RHSB = &RHS.getInlineBuckets()[i]; |
951 | bool hasLHSValue = (!KeyInfoT::isEqual(LHSB->getFirst(), EmptyKey) && |
952 | !KeyInfoT::isEqual(LHSB->getFirst(), TombstoneKey)); |
953 | bool hasRHSValue = (!KeyInfoT::isEqual(RHSB->getFirst(), EmptyKey) && |
954 | !KeyInfoT::isEqual(RHSB->getFirst(), TombstoneKey)); |
955 | if (hasLHSValue && hasRHSValue) { |
956 | // Swap together if we can... |
957 | std::swap(*LHSB, *RHSB); |
958 | continue; |
959 | } |
960 | // Swap separately and handle any asymmetry. |
961 | std::swap(LHSB->getFirst(), RHSB->getFirst()); |
962 | if (hasLHSValue) { |
963 | ::new (&RHSB->getSecond()) ValueT(std::move(LHSB->getSecond())); |
964 | LHSB->getSecond().~ValueT(); |
965 | } else if (hasRHSValue) { |
966 | ::new (&LHSB->getSecond()) ValueT(std::move(RHSB->getSecond())); |
967 | RHSB->getSecond().~ValueT(); |
968 | } |
969 | } |
970 | return; |
971 | } |
972 | if (!Small && !RHS.Small) { |
973 | std::swap(getLargeRep()->Buckets, RHS.getLargeRep()->Buckets); |
974 | std::swap(getLargeRep()->NumBuckets, RHS.getLargeRep()->NumBuckets); |
975 | return; |
976 | } |
977 | |
978 | SmallDenseMap &SmallSide = Small ? *this : RHS; |
979 | SmallDenseMap &LargeSide = Small ? RHS : *this; |
980 | |
981 | // First stash the large side's rep and move the small side across. |
982 | LargeRep TmpRep = std::move(*LargeSide.getLargeRep()); |
983 | LargeSide.getLargeRep()->~LargeRep(); |
984 | LargeSide.Small = true; |
985 | // This is similar to the standard move-from-old-buckets, but the bucket |
986 | // count hasn't actually rotated in this case. So we have to carefully |
987 | // move construct the keys and values into their new locations, but there |
988 | // is no need to re-hash things. |
989 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
990 | BucketT *NewB = &LargeSide.getInlineBuckets()[i], |
991 | *OldB = &SmallSide.getInlineBuckets()[i]; |
992 | ::new (&NewB->getFirst()) KeyT(std::move(OldB->getFirst())); |
993 | OldB->getFirst().~KeyT(); |
994 | if (!KeyInfoT::isEqual(NewB->getFirst(), EmptyKey) && |
995 | !KeyInfoT::isEqual(NewB->getFirst(), TombstoneKey)) { |
996 | ::new (&NewB->getSecond()) ValueT(std::move(OldB->getSecond())); |
997 | OldB->getSecond().~ValueT(); |
998 | } |
999 | } |
1000 | |
1001 | // The hard part of moving the small buckets across is done, just move |
1002 | // the TmpRep into its new home. |
1003 | SmallSide.Small = false; |
1004 | new (SmallSide.getLargeRep()) LargeRep(std::move(TmpRep)); |
1005 | } |
1006 | |
1007 | SmallDenseMap& operator=(const SmallDenseMap& other) { |
1008 | if (&other != this) |
1009 | copyFrom(other); |
1010 | return *this; |
1011 | } |
1012 | |
1013 | SmallDenseMap& operator=(SmallDenseMap &&other) { |
1014 | this->destroyAll(); |
1015 | deallocateBuckets(); |
1016 | init(0); |
1017 | swap(other); |
1018 | return *this; |
1019 | } |
1020 | |
1021 | void copyFrom(const SmallDenseMap& other) { |
1022 | this->destroyAll(); |
1023 | deallocateBuckets(); |
1024 | Small = true; |
1025 | if (other.getNumBuckets() > InlineBuckets) { |
1026 | Small = false; |
1027 | new (getLargeRep()) LargeRep(allocateBuckets(other.getNumBuckets())); |
1028 | } |
1029 | this->BaseT::copyFrom(other); |
1030 | } |
1031 | |
1032 | void init(unsigned InitBuckets) { |
1033 | Small = true; |
1034 | if (InitBuckets > InlineBuckets) { |
1035 | Small = false; |
1036 | new (getLargeRep()) LargeRep(allocateBuckets(InitBuckets)); |
1037 | } |
1038 | this->BaseT::initEmpty(); |
1039 | } |
1040 | |
1041 | void grow(unsigned AtLeast) { |
1042 | if (AtLeast > InlineBuckets) |
1043 | AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast-1)); |
1044 | |
1045 | if (Small) { |
1046 | // First move the inline buckets into a temporary storage. |
1047 | AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage; |
1048 | BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage); |
1049 | BucketT *TmpEnd = TmpBegin; |
1050 | |
1051 | // Loop over the buckets, moving non-empty, non-tombstones into the |
1052 | // temporary storage. Have the loop move the TmpEnd forward as it goes. |
1053 | const KeyT EmptyKey = this->getEmptyKey(); |
1054 | const KeyT TombstoneKey = this->getTombstoneKey(); |
1055 | for (BucketT *P = getBuckets(), *E = P + InlineBuckets; P != E; ++P) { |
1056 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
1057 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
1058 | assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&((void)0) |
1059 | "Too many inline buckets!")((void)0); |
1060 | ::new (&TmpEnd->getFirst()) KeyT(std::move(P->getFirst())); |
1061 | ::new (&TmpEnd->getSecond()) ValueT(std::move(P->getSecond())); |
1062 | ++TmpEnd; |
1063 | P->getSecond().~ValueT(); |
1064 | } |
1065 | P->getFirst().~KeyT(); |
1066 | } |
1067 | |
1068 | // AtLeast == InlineBuckets can happen if there are many tombstones, |
1069 | // and grow() is used to remove them. Usually we always switch to the |
1070 | // large rep here. |
1071 | if (AtLeast > InlineBuckets) { |
1072 | Small = false; |
1073 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1074 | } |
1075 | this->moveFromOldBuckets(TmpBegin, TmpEnd); |
1076 | return; |
1077 | } |
1078 | |
1079 | LargeRep OldRep = std::move(*getLargeRep()); |
1080 | getLargeRep()->~LargeRep(); |
1081 | if (AtLeast <= InlineBuckets) { |
1082 | Small = true; |
1083 | } else { |
1084 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1085 | } |
1086 | |
1087 | this->moveFromOldBuckets(OldRep.Buckets, OldRep.Buckets+OldRep.NumBuckets); |
1088 | |
1089 | // Free the old table. |
1090 | deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets, |
1091 | alignof(BucketT)); |
1092 | } |
1093 | |
1094 | void shrink_and_clear() { |
1095 | unsigned OldSize = this->size(); |
1096 | this->destroyAll(); |
1097 | |
1098 | // Reduce the number of buckets. |
1099 | unsigned NewNumBuckets = 0; |
1100 | if (OldSize) { |
1101 | NewNumBuckets = 1 << (Log2_32_Ceil(OldSize) + 1); |
1102 | if (NewNumBuckets > InlineBuckets && NewNumBuckets < 64u) |
1103 | NewNumBuckets = 64; |
1104 | } |
1105 | if ((Small && NewNumBuckets <= InlineBuckets) || |
1106 | (!Small && NewNumBuckets == getLargeRep()->NumBuckets)) { |
1107 | this->BaseT::initEmpty(); |
1108 | return; |
1109 | } |
1110 | |
1111 | deallocateBuckets(); |
1112 | init(NewNumBuckets); |
1113 | } |
1114 | |
1115 | private: |
1116 | unsigned getNumEntries() const { |
1117 | return NumEntries; |
1118 | } |
1119 | |
1120 | void setNumEntries(unsigned Num) { |
1121 | // NumEntries is hardcoded to be 31 bits wide. |
1122 | assert(Num < (1U << 31) && "Cannot support more than 1<<31 entries")((void)0); |
1123 | NumEntries = Num; |
1124 | } |
1125 | |
1126 | unsigned getNumTombstones() const { |
1127 | return NumTombstones; |
1128 | } |
1129 | |
1130 | void setNumTombstones(unsigned Num) { |
1131 | NumTombstones = Num; |
1132 | } |
1133 | |
1134 | const BucketT *getInlineBuckets() const { |
1135 | assert(Small)((void)0); |
1136 | // Note that this cast does not violate aliasing rules as we assert that |
1137 | // the memory's dynamic type is the small, inline bucket buffer, and the |
1138 | // 'storage' is a POD containing a char buffer. |
1139 | return reinterpret_cast<const BucketT *>(&storage); |
1140 | } |
1141 | |
1142 | BucketT *getInlineBuckets() { |
1143 | return const_cast<BucketT *>( |
1144 | const_cast<const SmallDenseMap *>(this)->getInlineBuckets()); |
1145 | } |
1146 | |
1147 | const LargeRep *getLargeRep() const { |
1148 | assert(!Small)((void)0); |
1149 | // Note, same rule about aliasing as with getInlineBuckets. |
1150 | return reinterpret_cast<const LargeRep *>(&storage); |
1151 | } |
1152 | |
1153 | LargeRep *getLargeRep() { |
1154 | return const_cast<LargeRep *>( |
1155 | const_cast<const SmallDenseMap *>(this)->getLargeRep()); |
1156 | } |
1157 | |
1158 | const BucketT *getBuckets() const { |
1159 | return Small ? getInlineBuckets() : getLargeRep()->Buckets; |
1160 | } |
1161 | |
1162 | BucketT *getBuckets() { |
1163 | return const_cast<BucketT *>( |
1164 | const_cast<const SmallDenseMap *>(this)->getBuckets()); |
1165 | } |
1166 | |
1167 | unsigned getNumBuckets() const { |
1168 | return Small ? InlineBuckets : getLargeRep()->NumBuckets; |
1169 | } |
1170 | |
1171 | void deallocateBuckets() { |
1172 | if (Small) |
1173 | return; |
1174 | |
1175 | deallocate_buffer(getLargeRep()->Buckets, |
1176 | sizeof(BucketT) * getLargeRep()->NumBuckets, |
1177 | alignof(BucketT)); |
1178 | getLargeRep()->~LargeRep(); |
1179 | } |
1180 | |
1181 | LargeRep allocateBuckets(unsigned Num) { |
1182 | assert(Num > InlineBuckets && "Must allocate more buckets than are inline")((void)0); |
1183 | LargeRep Rep = {static_cast<BucketT *>(allocate_buffer( |
1184 | sizeof(BucketT) * Num, alignof(BucketT))), |
1185 | Num}; |
1186 | return Rep; |
1187 | } |
1188 | }; |
1189 | |
1190 | template <typename KeyT, typename ValueT, typename KeyInfoT, typename Bucket, |
1191 | bool IsConst> |
1192 | class DenseMapIterator : DebugEpochBase::HandleBase { |
1193 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, true>; |
1194 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, false>; |
1195 | |
1196 | public: |
1197 | using difference_type = ptrdiff_t; |
1198 | using value_type = |
1199 | typename std::conditional<IsConst, const Bucket, Bucket>::type; |
1200 | using pointer = value_type *; |
1201 | using reference = value_type &; |
1202 | using iterator_category = std::forward_iterator_tag; |
1203 | |
1204 | private: |
1205 | pointer Ptr = nullptr; |
1206 | pointer End = nullptr; |
1207 | |
1208 | public: |
1209 | DenseMapIterator() = default; |
1210 | |
1211 | DenseMapIterator(pointer Pos, pointer E, const DebugEpochBase &Epoch, |
1212 | bool NoAdvance = false) |
1213 | : DebugEpochBase::HandleBase(&Epoch), Ptr(Pos), End(E) { |
1214 | assert(isHandleInSync() && "invalid construction!")((void)0); |
1215 | |
1216 | if (NoAdvance) return; |
1217 | if (shouldReverseIterate<KeyT>()) { |
1218 | RetreatPastEmptyBuckets(); |
1219 | return; |
1220 | } |
1221 | AdvancePastEmptyBuckets(); |
1222 | } |
1223 | |
1224 | // Converting ctor from non-const iterators to const iterators. SFINAE'd out |
1225 | // for const iterator destinations so it doesn't end up as a user defined copy |
1226 | // constructor. |
1227 | template <bool IsConstSrc, |
1228 | typename = std::enable_if_t<!IsConstSrc && IsConst>> |
1229 | DenseMapIterator( |
1230 | const DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, IsConstSrc> &I) |
1231 | : DebugEpochBase::HandleBase(I), Ptr(I.Ptr), End(I.End) {} |
1232 | |
1233 | reference operator*() const { |
1234 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1235 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1236 | if (shouldReverseIterate<KeyT>()) |
1237 | return Ptr[-1]; |
1238 | return *Ptr; |
1239 | } |
1240 | pointer operator->() const { |
1241 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1242 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1243 | if (shouldReverseIterate<KeyT>()) |
1244 | return &(Ptr[-1]); |
1245 | return Ptr; |
1246 | } |
1247 | |
1248 | friend bool operator==(const DenseMapIterator &LHS, |
1249 | const DenseMapIterator &RHS) { |
1250 | assert((!LHS.Ptr || LHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1251 | assert((!RHS.Ptr || RHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1252 | assert(LHS.getEpochAddress() == RHS.getEpochAddress() &&((void)0) |
1253 | "comparing incomparable iterators!")((void)0); |
1254 | return LHS.Ptr == RHS.Ptr; |
1255 | } |
1256 | |
1257 | friend bool operator!=(const DenseMapIterator &LHS, |
1258 | const DenseMapIterator &RHS) { |
1259 | return !(LHS == RHS); |
1260 | } |
1261 | |
1262 | inline DenseMapIterator& operator++() { // Preincrement |
1263 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1264 | assert(Ptr != End && "incrementing end() iterator")((void)0); |
1265 | if (shouldReverseIterate<KeyT>()) { |
1266 | --Ptr; |
1267 | RetreatPastEmptyBuckets(); |
1268 | return *this; |
1269 | } |
1270 | ++Ptr; |
1271 | AdvancePastEmptyBuckets(); |
1272 | return *this; |
1273 | } |
1274 | DenseMapIterator operator++(int) { // Postincrement |
1275 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1276 | DenseMapIterator tmp = *this; ++*this; return tmp; |
1277 | } |
1278 | |
1279 | private: |
1280 | void AdvancePastEmptyBuckets() { |
1281 | assert(Ptr <= End)((void)0); |
1282 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1283 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1284 | |
1285 | while (Ptr != End && (KeyInfoT::isEqual(Ptr->getFirst(), Empty) || |
1286 | KeyInfoT::isEqual(Ptr->getFirst(), Tombstone))) |
1287 | ++Ptr; |
1288 | } |
1289 | |
1290 | void RetreatPastEmptyBuckets() { |
1291 | assert(Ptr >= End)((void)0); |
1292 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1293 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1294 | |
1295 | while (Ptr != End && (KeyInfoT::isEqual(Ptr[-1].getFirst(), Empty) || |
1296 | KeyInfoT::isEqual(Ptr[-1].getFirst(), Tombstone))) |
1297 | --Ptr; |
1298 | } |
1299 | }; |
1300 | |
1301 | template <typename KeyT, typename ValueT, typename KeyInfoT> |
1302 | inline size_t capacity_in_bytes(const DenseMap<KeyT, ValueT, KeyInfoT> &X) { |
1303 | return X.getMemorySize(); |
1304 | } |
1305 | |
1306 | } // end namespace llvm |
1307 | |
1308 | #endif // LLVM_ADT_DENSEMAP_H |