File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp |
Warning: | line 2883, column 21 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | ///===- SimpleLoopUnswitch.cpp - Hoist loop-invariant control flow ---------===// | ||||||||||
2 | // | ||||||||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||||||
6 | // | ||||||||||
7 | //===----------------------------------------------------------------------===// | ||||||||||
8 | |||||||||||
9 | #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" | ||||||||||
10 | #include "llvm/ADT/DenseMap.h" | ||||||||||
11 | #include "llvm/ADT/STLExtras.h" | ||||||||||
12 | #include "llvm/ADT/Sequence.h" | ||||||||||
13 | #include "llvm/ADT/SetVector.h" | ||||||||||
14 | #include "llvm/ADT/SmallPtrSet.h" | ||||||||||
15 | #include "llvm/ADT/SmallVector.h" | ||||||||||
16 | #include "llvm/ADT/Statistic.h" | ||||||||||
17 | #include "llvm/ADT/Twine.h" | ||||||||||
18 | #include "llvm/Analysis/AssumptionCache.h" | ||||||||||
19 | #include "llvm/Analysis/CFG.h" | ||||||||||
20 | #include "llvm/Analysis/CodeMetrics.h" | ||||||||||
21 | #include "llvm/Analysis/GuardUtils.h" | ||||||||||
22 | #include "llvm/Analysis/InstructionSimplify.h" | ||||||||||
23 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||||||||
24 | #include "llvm/Analysis/LoopInfo.h" | ||||||||||
25 | #include "llvm/Analysis/LoopIterator.h" | ||||||||||
26 | #include "llvm/Analysis/LoopPass.h" | ||||||||||
27 | #include "llvm/Analysis/MemorySSA.h" | ||||||||||
28 | #include "llvm/Analysis/MemorySSAUpdater.h" | ||||||||||
29 | #include "llvm/Analysis/MustExecute.h" | ||||||||||
30 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||||||
31 | #include "llvm/IR/BasicBlock.h" | ||||||||||
32 | #include "llvm/IR/Constant.h" | ||||||||||
33 | #include "llvm/IR/Constants.h" | ||||||||||
34 | #include "llvm/IR/Dominators.h" | ||||||||||
35 | #include "llvm/IR/Function.h" | ||||||||||
36 | #include "llvm/IR/IRBuilder.h" | ||||||||||
37 | #include "llvm/IR/InstrTypes.h" | ||||||||||
38 | #include "llvm/IR/Instruction.h" | ||||||||||
39 | #include "llvm/IR/Instructions.h" | ||||||||||
40 | #include "llvm/IR/IntrinsicInst.h" | ||||||||||
41 | #include "llvm/IR/PatternMatch.h" | ||||||||||
42 | #include "llvm/IR/Use.h" | ||||||||||
43 | #include "llvm/IR/Value.h" | ||||||||||
44 | #include "llvm/InitializePasses.h" | ||||||||||
45 | #include "llvm/Pass.h" | ||||||||||
46 | #include "llvm/Support/Casting.h" | ||||||||||
47 | #include "llvm/Support/CommandLine.h" | ||||||||||
48 | #include "llvm/Support/Debug.h" | ||||||||||
49 | #include "llvm/Support/ErrorHandling.h" | ||||||||||
50 | #include "llvm/Support/GenericDomTree.h" | ||||||||||
51 | #include "llvm/Support/raw_ostream.h" | ||||||||||
52 | #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" | ||||||||||
53 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||||||
54 | #include "llvm/Transforms/Utils/Cloning.h" | ||||||||||
55 | #include "llvm/Transforms/Utils/Local.h" | ||||||||||
56 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||||||
57 | #include "llvm/Transforms/Utils/ValueMapper.h" | ||||||||||
58 | #include <algorithm> | ||||||||||
59 | #include <cassert> | ||||||||||
60 | #include <iterator> | ||||||||||
61 | #include <numeric> | ||||||||||
62 | #include <utility> | ||||||||||
63 | |||||||||||
64 | #define DEBUG_TYPE"simple-loop-unswitch" "simple-loop-unswitch" | ||||||||||
65 | |||||||||||
66 | using namespace llvm; | ||||||||||
67 | using namespace llvm::PatternMatch; | ||||||||||
68 | |||||||||||
69 | STATISTIC(NumBranches, "Number of branches unswitched")static llvm::Statistic NumBranches = {"simple-loop-unswitch", "NumBranches", "Number of branches unswitched"}; | ||||||||||
70 | STATISTIC(NumSwitches, "Number of switches unswitched")static llvm::Statistic NumSwitches = {"simple-loop-unswitch", "NumSwitches", "Number of switches unswitched"}; | ||||||||||
71 | STATISTIC(NumGuards, "Number of guards turned into branches for unswitching")static llvm::Statistic NumGuards = {"simple-loop-unswitch", "NumGuards" , "Number of guards turned into branches for unswitching"}; | ||||||||||
72 | STATISTIC(NumTrivial, "Number of unswitches that are trivial")static llvm::Statistic NumTrivial = {"simple-loop-unswitch", "NumTrivial" , "Number of unswitches that are trivial"}; | ||||||||||
73 | STATISTIC(static llvm::Statistic NumCostMultiplierSkipped = {"simple-loop-unswitch" , "NumCostMultiplierSkipped", "Number of unswitch candidates that had their cost multiplier skipped" } | ||||||||||
74 | NumCostMultiplierSkipped,static llvm::Statistic NumCostMultiplierSkipped = {"simple-loop-unswitch" , "NumCostMultiplierSkipped", "Number of unswitch candidates that had their cost multiplier skipped" } | ||||||||||
75 | "Number of unswitch candidates that had their cost multiplier skipped")static llvm::Statistic NumCostMultiplierSkipped = {"simple-loop-unswitch" , "NumCostMultiplierSkipped", "Number of unswitch candidates that had their cost multiplier skipped" }; | ||||||||||
76 | |||||||||||
77 | static cl::opt<bool> EnableNonTrivialUnswitch( | ||||||||||
78 | "enable-nontrivial-unswitch", cl::init(false), cl::Hidden, | ||||||||||
79 | cl::desc("Forcibly enables non-trivial loop unswitching rather than " | ||||||||||
80 | "following the configuration passed into the pass.")); | ||||||||||
81 | |||||||||||
82 | static cl::opt<int> | ||||||||||
83 | UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden, | ||||||||||
84 | cl::desc("The cost threshold for unswitching a loop.")); | ||||||||||
85 | |||||||||||
86 | static cl::opt<bool> EnableUnswitchCostMultiplier( | ||||||||||
87 | "enable-unswitch-cost-multiplier", cl::init(true), cl::Hidden, | ||||||||||
88 | cl::desc("Enable unswitch cost multiplier that prohibits exponential " | ||||||||||
89 | "explosion in nontrivial unswitch.")); | ||||||||||
90 | static cl::opt<int> UnswitchSiblingsToplevelDiv( | ||||||||||
91 | "unswitch-siblings-toplevel-div", cl::init(2), cl::Hidden, | ||||||||||
92 | cl::desc("Toplevel siblings divisor for cost multiplier.")); | ||||||||||
93 | static cl::opt<int> UnswitchNumInitialUnscaledCandidates( | ||||||||||
94 | "unswitch-num-initial-unscaled-candidates", cl::init(8), cl::Hidden, | ||||||||||
95 | cl::desc("Number of unswitch candidates that are ignored when calculating " | ||||||||||
96 | "cost multiplier.")); | ||||||||||
97 | static cl::opt<bool> UnswitchGuards( | ||||||||||
98 | "simple-loop-unswitch-guards", cl::init(true), cl::Hidden, | ||||||||||
99 | cl::desc("If enabled, simple loop unswitching will also consider " | ||||||||||
100 | "llvm.experimental.guard intrinsics as unswitch candidates.")); | ||||||||||
101 | static cl::opt<bool> DropNonTrivialImplicitNullChecks( | ||||||||||
102 | "simple-loop-unswitch-drop-non-trivial-implicit-null-checks", | ||||||||||
103 | cl::init(false), cl::Hidden, | ||||||||||
104 | cl::desc("If enabled, drop make.implicit metadata in unswitched implicit " | ||||||||||
105 | "null checks to save time analyzing if we can keep it.")); | ||||||||||
106 | static cl::opt<unsigned> | ||||||||||
107 | MSSAThreshold("simple-loop-unswitch-memoryssa-threshold", | ||||||||||
108 | cl::desc("Max number of memory uses to explore during " | ||||||||||
109 | "partial unswitching analysis"), | ||||||||||
110 | cl::init(100), cl::Hidden); | ||||||||||
111 | |||||||||||
112 | /// Collect all of the loop invariant input values transitively used by the | ||||||||||
113 | /// homogeneous instruction graph from a given root. | ||||||||||
114 | /// | ||||||||||
115 | /// This essentially walks from a root recursively through loop variant operands | ||||||||||
116 | /// which have the exact same opcode and finds all inputs which are loop | ||||||||||
117 | /// invariant. For some operations these can be re-associated and unswitched out | ||||||||||
118 | /// of the loop entirely. | ||||||||||
119 | static TinyPtrVector<Value *> | ||||||||||
120 | collectHomogenousInstGraphLoopInvariants(Loop &L, Instruction &Root, | ||||||||||
121 | LoopInfo &LI) { | ||||||||||
122 | assert(!L.isLoopInvariant(&Root) &&((void)0) | ||||||||||
123 | "Only need to walk the graph if root itself is not invariant.")((void)0); | ||||||||||
124 | TinyPtrVector<Value *> Invariants; | ||||||||||
125 | |||||||||||
126 | bool IsRootAnd = match(&Root, m_LogicalAnd()); | ||||||||||
127 | bool IsRootOr = match(&Root, m_LogicalOr()); | ||||||||||
128 | |||||||||||
129 | // Build a worklist and recurse through operators collecting invariants. | ||||||||||
130 | SmallVector<Instruction *, 4> Worklist; | ||||||||||
131 | SmallPtrSet<Instruction *, 8> Visited; | ||||||||||
132 | Worklist.push_back(&Root); | ||||||||||
133 | Visited.insert(&Root); | ||||||||||
134 | do { | ||||||||||
135 | Instruction &I = *Worklist.pop_back_val(); | ||||||||||
136 | for (Value *OpV : I.operand_values()) { | ||||||||||
137 | // Skip constants as unswitching isn't interesting for them. | ||||||||||
138 | if (isa<Constant>(OpV)) | ||||||||||
139 | continue; | ||||||||||
140 | |||||||||||
141 | // Add it to our result if loop invariant. | ||||||||||
142 | if (L.isLoopInvariant(OpV)) { | ||||||||||
143 | Invariants.push_back(OpV); | ||||||||||
144 | continue; | ||||||||||
145 | } | ||||||||||
146 | |||||||||||
147 | // If not an instruction with the same opcode, nothing we can do. | ||||||||||
148 | Instruction *OpI = dyn_cast<Instruction>(OpV); | ||||||||||
149 | |||||||||||
150 | if (OpI && ((IsRootAnd && match(OpI, m_LogicalAnd())) || | ||||||||||
151 | (IsRootOr && match(OpI, m_LogicalOr())))) { | ||||||||||
152 | // Visit this operand. | ||||||||||
153 | if (Visited.insert(OpI).second) | ||||||||||
154 | Worklist.push_back(OpI); | ||||||||||
155 | } | ||||||||||
156 | } | ||||||||||
157 | } while (!Worklist.empty()); | ||||||||||
158 | |||||||||||
159 | return Invariants; | ||||||||||
160 | } | ||||||||||
161 | |||||||||||
162 | static void replaceLoopInvariantUses(Loop &L, Value *Invariant, | ||||||||||
163 | Constant &Replacement) { | ||||||||||
164 | assert(!isa<Constant>(Invariant) && "Why are we unswitching on a constant?")((void)0); | ||||||||||
165 | |||||||||||
166 | // Replace uses of LIC in the loop with the given constant. | ||||||||||
167 | // We use make_early_inc_range as set invalidates the iterator. | ||||||||||
168 | for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { | ||||||||||
169 | Instruction *UserI = dyn_cast<Instruction>(U.getUser()); | ||||||||||
170 | |||||||||||
171 | // Replace this use within the loop body. | ||||||||||
172 | if (UserI && L.contains(UserI)) | ||||||||||
173 | U.set(&Replacement); | ||||||||||
174 | } | ||||||||||
175 | } | ||||||||||
176 | |||||||||||
177 | /// Check that all the LCSSA PHI nodes in the loop exit block have trivial | ||||||||||
178 | /// incoming values along this edge. | ||||||||||
179 | static bool areLoopExitPHIsLoopInvariant(Loop &L, BasicBlock &ExitingBB, | ||||||||||
180 | BasicBlock &ExitBB) { | ||||||||||
181 | for (Instruction &I : ExitBB) { | ||||||||||
182 | auto *PN = dyn_cast<PHINode>(&I); | ||||||||||
183 | if (!PN) | ||||||||||
184 | // No more PHIs to check. | ||||||||||
185 | return true; | ||||||||||
186 | |||||||||||
187 | // If the incoming value for this edge isn't loop invariant the unswitch | ||||||||||
188 | // won't be trivial. | ||||||||||
189 | if (!L.isLoopInvariant(PN->getIncomingValueForBlock(&ExitingBB))) | ||||||||||
190 | return false; | ||||||||||
191 | } | ||||||||||
192 | llvm_unreachable("Basic blocks should never be empty!")__builtin_unreachable(); | ||||||||||
193 | } | ||||||||||
194 | |||||||||||
195 | /// Copy a set of loop invariant values \p ToDuplicate and insert them at the | ||||||||||
196 | /// end of \p BB and conditionally branch on the copied condition. We only | ||||||||||
197 | /// branch on a single value. | ||||||||||
198 | static void buildPartialUnswitchConditionalBranch(BasicBlock &BB, | ||||||||||
199 | ArrayRef<Value *> Invariants, | ||||||||||
200 | bool Direction, | ||||||||||
201 | BasicBlock &UnswitchedSucc, | ||||||||||
202 | BasicBlock &NormalSucc) { | ||||||||||
203 | IRBuilder<> IRB(&BB); | ||||||||||
204 | |||||||||||
205 | Value *Cond = Direction ? IRB.CreateOr(Invariants) : | ||||||||||
206 | IRB.CreateAnd(Invariants); | ||||||||||
207 | IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, | ||||||||||
208 | Direction ? &NormalSucc : &UnswitchedSucc); | ||||||||||
209 | } | ||||||||||
210 | |||||||||||
211 | /// Copy a set of loop invariant values, and conditionally branch on them. | ||||||||||
212 | static void buildPartialInvariantUnswitchConditionalBranch( | ||||||||||
213 | BasicBlock &BB, ArrayRef<Value *> ToDuplicate, bool Direction, | ||||||||||
214 | BasicBlock &UnswitchedSucc, BasicBlock &NormalSucc, Loop &L, | ||||||||||
215 | MemorySSAUpdater *MSSAU) { | ||||||||||
216 | ValueToValueMapTy VMap; | ||||||||||
217 | for (auto *Val : reverse(ToDuplicate)) { | ||||||||||
218 | Instruction *Inst = cast<Instruction>(Val); | ||||||||||
219 | Instruction *NewInst = Inst->clone(); | ||||||||||
220 | BB.getInstList().insert(BB.end(), NewInst); | ||||||||||
221 | RemapInstruction(NewInst, VMap, | ||||||||||
222 | RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); | ||||||||||
223 | VMap[Val] = NewInst; | ||||||||||
224 | |||||||||||
225 | if (!MSSAU) | ||||||||||
226 | continue; | ||||||||||
227 | |||||||||||
228 | MemorySSA *MSSA = MSSAU->getMemorySSA(); | ||||||||||
229 | if (auto *MemUse = | ||||||||||
230 | dyn_cast_or_null<MemoryUse>(MSSA->getMemoryAccess(Inst))) { | ||||||||||
231 | auto *DefiningAccess = MemUse->getDefiningAccess(); | ||||||||||
232 | // Get the first defining access before the loop. | ||||||||||
233 | while (L.contains(DefiningAccess->getBlock())) { | ||||||||||
234 | // If the defining access is a MemoryPhi, get the incoming | ||||||||||
235 | // value for the pre-header as defining access. | ||||||||||
236 | if (auto *MemPhi = dyn_cast<MemoryPhi>(DefiningAccess)) | ||||||||||
237 | DefiningAccess = | ||||||||||
238 | MemPhi->getIncomingValueForBlock(L.getLoopPreheader()); | ||||||||||
239 | else | ||||||||||
240 | DefiningAccess = cast<MemoryDef>(DefiningAccess)->getDefiningAccess(); | ||||||||||
241 | } | ||||||||||
242 | MSSAU->createMemoryAccessInBB(NewInst, DefiningAccess, | ||||||||||
243 | NewInst->getParent(), | ||||||||||
244 | MemorySSA::BeforeTerminator); | ||||||||||
245 | } | ||||||||||
246 | } | ||||||||||
247 | |||||||||||
248 | IRBuilder<> IRB(&BB); | ||||||||||
249 | Value *Cond = VMap[ToDuplicate[0]]; | ||||||||||
250 | IRB.CreateCondBr(Cond, Direction ? &UnswitchedSucc : &NormalSucc, | ||||||||||
251 | Direction ? &NormalSucc : &UnswitchedSucc); | ||||||||||
252 | } | ||||||||||
253 | |||||||||||
254 | /// Rewrite the PHI nodes in an unswitched loop exit basic block. | ||||||||||
255 | /// | ||||||||||
256 | /// Requires that the loop exit and unswitched basic block are the same, and | ||||||||||
257 | /// that the exiting block was a unique predecessor of that block. Rewrites the | ||||||||||
258 | /// PHI nodes in that block such that what were LCSSA PHI nodes become trivial | ||||||||||
259 | /// PHI nodes from the old preheader that now contains the unswitched | ||||||||||
260 | /// terminator. | ||||||||||
261 | static void rewritePHINodesForUnswitchedExitBlock(BasicBlock &UnswitchedBB, | ||||||||||
262 | BasicBlock &OldExitingBB, | ||||||||||
263 | BasicBlock &OldPH) { | ||||||||||
264 | for (PHINode &PN : UnswitchedBB.phis()) { | ||||||||||
265 | // When the loop exit is directly unswitched we just need to update the | ||||||||||
266 | // incoming basic block. We loop to handle weird cases with repeated | ||||||||||
267 | // incoming blocks, but expect to typically only have one operand here. | ||||||||||
268 | for (auto i : seq<int>(0, PN.getNumOperands())) { | ||||||||||
269 | assert(PN.getIncomingBlock(i) == &OldExitingBB &&((void)0) | ||||||||||
270 | "Found incoming block different from unique predecessor!")((void)0); | ||||||||||
271 | PN.setIncomingBlock(i, &OldPH); | ||||||||||
272 | } | ||||||||||
273 | } | ||||||||||
274 | } | ||||||||||
275 | |||||||||||
276 | /// Rewrite the PHI nodes in the loop exit basic block and the split off | ||||||||||
277 | /// unswitched block. | ||||||||||
278 | /// | ||||||||||
279 | /// Because the exit block remains an exit from the loop, this rewrites the | ||||||||||
280 | /// LCSSA PHI nodes in it to remove the unswitched edge and introduces PHI | ||||||||||
281 | /// nodes into the unswitched basic block to select between the value in the | ||||||||||
282 | /// old preheader and the loop exit. | ||||||||||
283 | static void rewritePHINodesForExitAndUnswitchedBlocks(BasicBlock &ExitBB, | ||||||||||
284 | BasicBlock &UnswitchedBB, | ||||||||||
285 | BasicBlock &OldExitingBB, | ||||||||||
286 | BasicBlock &OldPH, | ||||||||||
287 | bool FullUnswitch) { | ||||||||||
288 | assert(&ExitBB != &UnswitchedBB &&((void)0) | ||||||||||
289 | "Must have different loop exit and unswitched blocks!")((void)0); | ||||||||||
290 | Instruction *InsertPt = &*UnswitchedBB.begin(); | ||||||||||
291 | for (PHINode &PN : ExitBB.phis()) { | ||||||||||
292 | auto *NewPN = PHINode::Create(PN.getType(), /*NumReservedValues*/ 2, | ||||||||||
293 | PN.getName() + ".split", InsertPt); | ||||||||||
294 | |||||||||||
295 | // Walk backwards over the old PHI node's inputs to minimize the cost of | ||||||||||
296 | // removing each one. We have to do this weird loop manually so that we | ||||||||||
297 | // create the same number of new incoming edges in the new PHI as we expect | ||||||||||
298 | // each case-based edge to be included in the unswitched switch in some | ||||||||||
299 | // cases. | ||||||||||
300 | // FIXME: This is really, really gross. It would be much cleaner if LLVM | ||||||||||
301 | // allowed us to create a single entry for a predecessor block without | ||||||||||
302 | // having separate entries for each "edge" even though these edges are | ||||||||||
303 | // required to produce identical results. | ||||||||||
304 | for (int i = PN.getNumIncomingValues() - 1; i >= 0; --i) { | ||||||||||
305 | if (PN.getIncomingBlock(i) != &OldExitingBB) | ||||||||||
306 | continue; | ||||||||||
307 | |||||||||||
308 | Value *Incoming = PN.getIncomingValue(i); | ||||||||||
309 | if (FullUnswitch) | ||||||||||
310 | // No more edge from the old exiting block to the exit block. | ||||||||||
311 | PN.removeIncomingValue(i); | ||||||||||
312 | |||||||||||
313 | NewPN->addIncoming(Incoming, &OldPH); | ||||||||||
314 | } | ||||||||||
315 | |||||||||||
316 | // Now replace the old PHI with the new one and wire the old one in as an | ||||||||||
317 | // input to the new one. | ||||||||||
318 | PN.replaceAllUsesWith(NewPN); | ||||||||||
319 | NewPN->addIncoming(&PN, &ExitBB); | ||||||||||
320 | } | ||||||||||
321 | } | ||||||||||
322 | |||||||||||
323 | /// Hoist the current loop up to the innermost loop containing a remaining exit. | ||||||||||
324 | /// | ||||||||||
325 | /// Because we've removed an exit from the loop, we may have changed the set of | ||||||||||
326 | /// loops reachable and need to move the current loop up the loop nest or even | ||||||||||
327 | /// to an entirely separate nest. | ||||||||||
328 | static void hoistLoopToNewParent(Loop &L, BasicBlock &Preheader, | ||||||||||
329 | DominatorTree &DT, LoopInfo &LI, | ||||||||||
330 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE) { | ||||||||||
331 | // If the loop is already at the top level, we can't hoist it anywhere. | ||||||||||
332 | Loop *OldParentL = L.getParentLoop(); | ||||||||||
333 | if (!OldParentL) | ||||||||||
334 | return; | ||||||||||
335 | |||||||||||
336 | SmallVector<BasicBlock *, 4> Exits; | ||||||||||
337 | L.getExitBlocks(Exits); | ||||||||||
338 | Loop *NewParentL = nullptr; | ||||||||||
339 | for (auto *ExitBB : Exits) | ||||||||||
340 | if (Loop *ExitL = LI.getLoopFor(ExitBB)) | ||||||||||
341 | if (!NewParentL || NewParentL->contains(ExitL)) | ||||||||||
342 | NewParentL = ExitL; | ||||||||||
343 | |||||||||||
344 | if (NewParentL == OldParentL) | ||||||||||
345 | return; | ||||||||||
346 | |||||||||||
347 | // The new parent loop (if different) should always contain the old one. | ||||||||||
348 | if (NewParentL) | ||||||||||
349 | assert(NewParentL->contains(OldParentL) &&((void)0) | ||||||||||
350 | "Can only hoist this loop up the nest!")((void)0); | ||||||||||
351 | |||||||||||
352 | // The preheader will need to move with the body of this loop. However, | ||||||||||
353 | // because it isn't in this loop we also need to update the primary loop map. | ||||||||||
354 | assert(OldParentL == LI.getLoopFor(&Preheader) &&((void)0) | ||||||||||
355 | "Parent loop of this loop should contain this loop's preheader!")((void)0); | ||||||||||
356 | LI.changeLoopFor(&Preheader, NewParentL); | ||||||||||
357 | |||||||||||
358 | // Remove this loop from its old parent. | ||||||||||
359 | OldParentL->removeChildLoop(&L); | ||||||||||
360 | |||||||||||
361 | // Add the loop either to the new parent or as a top-level loop. | ||||||||||
362 | if (NewParentL) | ||||||||||
363 | NewParentL->addChildLoop(&L); | ||||||||||
364 | else | ||||||||||
365 | LI.addTopLevelLoop(&L); | ||||||||||
366 | |||||||||||
367 | // Remove this loops blocks from the old parent and every other loop up the | ||||||||||
368 | // nest until reaching the new parent. Also update all of these | ||||||||||
369 | // no-longer-containing loops to reflect the nesting change. | ||||||||||
370 | for (Loop *OldContainingL = OldParentL; OldContainingL != NewParentL; | ||||||||||
371 | OldContainingL = OldContainingL->getParentLoop()) { | ||||||||||
372 | llvm::erase_if(OldContainingL->getBlocksVector(), | ||||||||||
373 | [&](const BasicBlock *BB) { | ||||||||||
374 | return BB == &Preheader || L.contains(BB); | ||||||||||
375 | }); | ||||||||||
376 | |||||||||||
377 | OldContainingL->getBlocksSet().erase(&Preheader); | ||||||||||
378 | for (BasicBlock *BB : L.blocks()) | ||||||||||
379 | OldContainingL->getBlocksSet().erase(BB); | ||||||||||
380 | |||||||||||
381 | // Because we just hoisted a loop out of this one, we have essentially | ||||||||||
382 | // created new exit paths from it. That means we need to form LCSSA PHI | ||||||||||
383 | // nodes for values used in the no-longer-nested loop. | ||||||||||
384 | formLCSSA(*OldContainingL, DT, &LI, SE); | ||||||||||
385 | |||||||||||
386 | // We shouldn't need to form dedicated exits because the exit introduced | ||||||||||
387 | // here is the (just split by unswitching) preheader. However, after trivial | ||||||||||
388 | // unswitching it is possible to get new non-dedicated exits out of parent | ||||||||||
389 | // loop so let's conservatively form dedicated exit blocks and figure out | ||||||||||
390 | // if we can optimize later. | ||||||||||
391 | formDedicatedExitBlocks(OldContainingL, &DT, &LI, MSSAU, | ||||||||||
392 | /*PreserveLCSSA*/ true); | ||||||||||
393 | } | ||||||||||
394 | } | ||||||||||
395 | |||||||||||
396 | // Return the top-most loop containing ExitBB and having ExitBB as exiting block | ||||||||||
397 | // or the loop containing ExitBB, if there is no parent loop containing ExitBB | ||||||||||
398 | // as exiting block. | ||||||||||
399 | static Loop *getTopMostExitingLoop(BasicBlock *ExitBB, LoopInfo &LI) { | ||||||||||
400 | Loop *TopMost = LI.getLoopFor(ExitBB); | ||||||||||
401 | Loop *Current = TopMost; | ||||||||||
402 | while (Current) { | ||||||||||
403 | if (Current->isLoopExiting(ExitBB)) | ||||||||||
404 | TopMost = Current; | ||||||||||
405 | Current = Current->getParentLoop(); | ||||||||||
406 | } | ||||||||||
407 | return TopMost; | ||||||||||
408 | } | ||||||||||
409 | |||||||||||
410 | /// Unswitch a trivial branch if the condition is loop invariant. | ||||||||||
411 | /// | ||||||||||
412 | /// This routine should only be called when loop code leading to the branch has | ||||||||||
413 | /// been validated as trivial (no side effects). This routine checks if the | ||||||||||
414 | /// condition is invariant and one of the successors is a loop exit. This | ||||||||||
415 | /// allows us to unswitch without duplicating the loop, making it trivial. | ||||||||||
416 | /// | ||||||||||
417 | /// If this routine fails to unswitch the branch it returns false. | ||||||||||
418 | /// | ||||||||||
419 | /// If the branch can be unswitched, this routine splits the preheader and | ||||||||||
420 | /// hoists the branch above that split. Preserves loop simplified form | ||||||||||
421 | /// (splitting the exit block as necessary). It simplifies the branch within | ||||||||||
422 | /// the loop to an unconditional branch but doesn't remove it entirely. Further | ||||||||||
423 | /// cleanup can be done with some simplifycfg like pass. | ||||||||||
424 | /// | ||||||||||
425 | /// If `SE` is not null, it will be updated based on the potential loop SCEVs | ||||||||||
426 | /// invalidated by this. | ||||||||||
427 | static bool unswitchTrivialBranch(Loop &L, BranchInst &BI, DominatorTree &DT, | ||||||||||
428 | LoopInfo &LI, ScalarEvolution *SE, | ||||||||||
429 | MemorySSAUpdater *MSSAU) { | ||||||||||
430 | assert(BI.isConditional() && "Can only unswitch a conditional branch!")((void)0); | ||||||||||
431 | LLVM_DEBUG(dbgs() << " Trying to unswitch branch: " << BI << "\n")do { } while (false); | ||||||||||
432 | |||||||||||
433 | // The loop invariant values that we want to unswitch. | ||||||||||
434 | TinyPtrVector<Value *> Invariants; | ||||||||||
435 | |||||||||||
436 | // When true, we're fully unswitching the branch rather than just unswitching | ||||||||||
437 | // some input conditions to the branch. | ||||||||||
438 | bool FullUnswitch = false; | ||||||||||
439 | |||||||||||
440 | if (L.isLoopInvariant(BI.getCondition())) { | ||||||||||
441 | Invariants.push_back(BI.getCondition()); | ||||||||||
442 | FullUnswitch = true; | ||||||||||
443 | } else { | ||||||||||
444 | if (auto *CondInst = dyn_cast<Instruction>(BI.getCondition())) | ||||||||||
445 | Invariants = collectHomogenousInstGraphLoopInvariants(L, *CondInst, LI); | ||||||||||
446 | if (Invariants.empty()) { | ||||||||||
447 | LLVM_DEBUG(dbgs() << " Couldn't find invariant inputs!\n")do { } while (false); | ||||||||||
448 | return false; | ||||||||||
449 | } | ||||||||||
450 | } | ||||||||||
451 | |||||||||||
452 | // Check that one of the branch's successors exits, and which one. | ||||||||||
453 | bool ExitDirection = true; | ||||||||||
454 | int LoopExitSuccIdx = 0; | ||||||||||
455 | auto *LoopExitBB = BI.getSuccessor(0); | ||||||||||
456 | if (L.contains(LoopExitBB)) { | ||||||||||
457 | ExitDirection = false; | ||||||||||
458 | LoopExitSuccIdx = 1; | ||||||||||
459 | LoopExitBB = BI.getSuccessor(1); | ||||||||||
460 | if (L.contains(LoopExitBB)) { | ||||||||||
461 | LLVM_DEBUG(dbgs() << " Branch doesn't exit the loop!\n")do { } while (false); | ||||||||||
462 | return false; | ||||||||||
463 | } | ||||||||||
464 | } | ||||||||||
465 | auto *ContinueBB = BI.getSuccessor(1 - LoopExitSuccIdx); | ||||||||||
466 | auto *ParentBB = BI.getParent(); | ||||||||||
467 | if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, *LoopExitBB)) { | ||||||||||
468 | LLVM_DEBUG(dbgs() << " Loop exit PHI's aren't loop-invariant!\n")do { } while (false); | ||||||||||
469 | return false; | ||||||||||
470 | } | ||||||||||
471 | |||||||||||
472 | // When unswitching only part of the branch's condition, we need the exit | ||||||||||
473 | // block to be reached directly from the partially unswitched input. This can | ||||||||||
474 | // be done when the exit block is along the true edge and the branch condition | ||||||||||
475 | // is a graph of `or` operations, or the exit block is along the false edge | ||||||||||
476 | // and the condition is a graph of `and` operations. | ||||||||||
477 | if (!FullUnswitch) { | ||||||||||
478 | if (ExitDirection ? !match(BI.getCondition(), m_LogicalOr()) | ||||||||||
479 | : !match(BI.getCondition(), m_LogicalAnd())) { | ||||||||||
480 | LLVM_DEBUG(dbgs() << " Branch condition is in improper form for "do { } while (false) | ||||||||||
481 | "non-full unswitch!\n")do { } while (false); | ||||||||||
482 | return false; | ||||||||||
483 | } | ||||||||||
484 | } | ||||||||||
485 | |||||||||||
486 | LLVM_DEBUG({do { } while (false) | ||||||||||
487 | dbgs() << " unswitching trivial invariant conditions for: " << BIdo { } while (false) | ||||||||||
488 | << "\n";do { } while (false) | ||||||||||
489 | for (Value *Invariant : Invariants) {do { } while (false) | ||||||||||
490 | dbgs() << " " << *Invariant << " == true";do { } while (false) | ||||||||||
491 | if (Invariant != Invariants.back())do { } while (false) | ||||||||||
492 | dbgs() << " ||";do { } while (false) | ||||||||||
493 | dbgs() << "\n";do { } while (false) | ||||||||||
494 | }do { } while (false) | ||||||||||
495 | })do { } while (false); | ||||||||||
496 | |||||||||||
497 | // If we have scalar evolutions, we need to invalidate them including this | ||||||||||
498 | // loop, the loop containing the exit block and the topmost parent loop | ||||||||||
499 | // exiting via LoopExitBB. | ||||||||||
500 | if (SE) { | ||||||||||
501 | if (Loop *ExitL = getTopMostExitingLoop(LoopExitBB, LI)) | ||||||||||
502 | SE->forgetLoop(ExitL); | ||||||||||
503 | else | ||||||||||
504 | // Forget the entire nest as this exits the entire nest. | ||||||||||
505 | SE->forgetTopmostLoop(&L); | ||||||||||
506 | } | ||||||||||
507 | |||||||||||
508 | if (MSSAU && VerifyMemorySSA) | ||||||||||
509 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
510 | |||||||||||
511 | // Split the preheader, so that we know that there is a safe place to insert | ||||||||||
512 | // the conditional branch. We will change the preheader to have a conditional | ||||||||||
513 | // branch on LoopCond. | ||||||||||
514 | BasicBlock *OldPH = L.getLoopPreheader(); | ||||||||||
515 | BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); | ||||||||||
516 | |||||||||||
517 | // Now that we have a place to insert the conditional branch, create a place | ||||||||||
518 | // to branch to: this is the exit block out of the loop that we are | ||||||||||
519 | // unswitching. We need to split this if there are other loop predecessors. | ||||||||||
520 | // Because the loop is in simplified form, *any* other predecessor is enough. | ||||||||||
521 | BasicBlock *UnswitchedBB; | ||||||||||
522 | if (FullUnswitch && LoopExitBB->getUniquePredecessor()) { | ||||||||||
523 | assert(LoopExitBB->getUniquePredecessor() == BI.getParent() &&((void)0) | ||||||||||
524 | "A branch's parent isn't a predecessor!")((void)0); | ||||||||||
525 | UnswitchedBB = LoopExitBB; | ||||||||||
526 | } else { | ||||||||||
527 | UnswitchedBB = | ||||||||||
528 | SplitBlock(LoopExitBB, &LoopExitBB->front(), &DT, &LI, MSSAU); | ||||||||||
529 | } | ||||||||||
530 | |||||||||||
531 | if (MSSAU && VerifyMemorySSA) | ||||||||||
532 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
533 | |||||||||||
534 | // Actually move the invariant uses into the unswitched position. If possible, | ||||||||||
535 | // we do this by moving the instructions, but when doing partial unswitching | ||||||||||
536 | // we do it by building a new merge of the values in the unswitched position. | ||||||||||
537 | OldPH->getTerminator()->eraseFromParent(); | ||||||||||
538 | if (FullUnswitch) { | ||||||||||
539 | // If fully unswitching, we can use the existing branch instruction. | ||||||||||
540 | // Splice it into the old PH to gate reaching the new preheader and re-point | ||||||||||
541 | // its successors. | ||||||||||
542 | OldPH->getInstList().splice(OldPH->end(), BI.getParent()->getInstList(), | ||||||||||
543 | BI); | ||||||||||
544 | if (MSSAU) { | ||||||||||
545 | // Temporarily clone the terminator, to make MSSA update cheaper by | ||||||||||
546 | // separating "insert edge" updates from "remove edge" ones. | ||||||||||
547 | ParentBB->getInstList().push_back(BI.clone()); | ||||||||||
548 | } else { | ||||||||||
549 | // Create a new unconditional branch that will continue the loop as a new | ||||||||||
550 | // terminator. | ||||||||||
551 | BranchInst::Create(ContinueBB, ParentBB); | ||||||||||
552 | } | ||||||||||
553 | BI.setSuccessor(LoopExitSuccIdx, UnswitchedBB); | ||||||||||
554 | BI.setSuccessor(1 - LoopExitSuccIdx, NewPH); | ||||||||||
555 | } else { | ||||||||||
556 | // Only unswitching a subset of inputs to the condition, so we will need to | ||||||||||
557 | // build a new branch that merges the invariant inputs. | ||||||||||
558 | if (ExitDirection) | ||||||||||
559 | assert(match(BI.getCondition(), m_LogicalOr()) &&((void)0) | ||||||||||
560 | "Must have an `or` of `i1`s or `select i1 X, true, Y`s for the "((void)0) | ||||||||||
561 | "condition!")((void)0); | ||||||||||
562 | else | ||||||||||
563 | assert(match(BI.getCondition(), m_LogicalAnd()) &&((void)0) | ||||||||||
564 | "Must have an `and` of `i1`s or `select i1 X, Y, false`s for the"((void)0) | ||||||||||
565 | " condition!")((void)0); | ||||||||||
566 | buildPartialUnswitchConditionalBranch(*OldPH, Invariants, ExitDirection, | ||||||||||
567 | *UnswitchedBB, *NewPH); | ||||||||||
568 | } | ||||||||||
569 | |||||||||||
570 | // Update the dominator tree with the added edge. | ||||||||||
571 | DT.insertEdge(OldPH, UnswitchedBB); | ||||||||||
572 | |||||||||||
573 | // After the dominator tree was updated with the added edge, update MemorySSA | ||||||||||
574 | // if available. | ||||||||||
575 | if (MSSAU) { | ||||||||||
576 | SmallVector<CFGUpdate, 1> Updates; | ||||||||||
577 | Updates.push_back({cfg::UpdateKind::Insert, OldPH, UnswitchedBB}); | ||||||||||
578 | MSSAU->applyInsertUpdates(Updates, DT); | ||||||||||
579 | } | ||||||||||
580 | |||||||||||
581 | // Finish updating dominator tree and memory ssa for full unswitch. | ||||||||||
582 | if (FullUnswitch) { | ||||||||||
583 | if (MSSAU) { | ||||||||||
584 | // Remove the cloned branch instruction. | ||||||||||
585 | ParentBB->getTerminator()->eraseFromParent(); | ||||||||||
586 | // Create unconditional branch now. | ||||||||||
587 | BranchInst::Create(ContinueBB, ParentBB); | ||||||||||
588 | MSSAU->removeEdge(ParentBB, LoopExitBB); | ||||||||||
589 | } | ||||||||||
590 | DT.deleteEdge(ParentBB, LoopExitBB); | ||||||||||
591 | } | ||||||||||
592 | |||||||||||
593 | if (MSSAU && VerifyMemorySSA) | ||||||||||
594 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
595 | |||||||||||
596 | // Rewrite the relevant PHI nodes. | ||||||||||
597 | if (UnswitchedBB == LoopExitBB) | ||||||||||
598 | rewritePHINodesForUnswitchedExitBlock(*UnswitchedBB, *ParentBB, *OldPH); | ||||||||||
599 | else | ||||||||||
600 | rewritePHINodesForExitAndUnswitchedBlocks(*LoopExitBB, *UnswitchedBB, | ||||||||||
601 | *ParentBB, *OldPH, FullUnswitch); | ||||||||||
602 | |||||||||||
603 | // The constant we can replace all of our invariants with inside the loop | ||||||||||
604 | // body. If any of the invariants have a value other than this the loop won't | ||||||||||
605 | // be entered. | ||||||||||
606 | ConstantInt *Replacement = ExitDirection | ||||||||||
607 | ? ConstantInt::getFalse(BI.getContext()) | ||||||||||
608 | : ConstantInt::getTrue(BI.getContext()); | ||||||||||
609 | |||||||||||
610 | // Since this is an i1 condition we can also trivially replace uses of it | ||||||||||
611 | // within the loop with a constant. | ||||||||||
612 | for (Value *Invariant : Invariants) | ||||||||||
613 | replaceLoopInvariantUses(L, Invariant, *Replacement); | ||||||||||
614 | |||||||||||
615 | // If this was full unswitching, we may have changed the nesting relationship | ||||||||||
616 | // for this loop so hoist it to its correct parent if needed. | ||||||||||
617 | if (FullUnswitch) | ||||||||||
618 | hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); | ||||||||||
619 | |||||||||||
620 | if (MSSAU && VerifyMemorySSA) | ||||||||||
621 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
622 | |||||||||||
623 | LLVM_DEBUG(dbgs() << " done: unswitching trivial branch...\n")do { } while (false); | ||||||||||
624 | ++NumTrivial; | ||||||||||
625 | ++NumBranches; | ||||||||||
626 | return true; | ||||||||||
627 | } | ||||||||||
628 | |||||||||||
629 | /// Unswitch a trivial switch if the condition is loop invariant. | ||||||||||
630 | /// | ||||||||||
631 | /// This routine should only be called when loop code leading to the switch has | ||||||||||
632 | /// been validated as trivial (no side effects). This routine checks if the | ||||||||||
633 | /// condition is invariant and that at least one of the successors is a loop | ||||||||||
634 | /// exit. This allows us to unswitch without duplicating the loop, making it | ||||||||||
635 | /// trivial. | ||||||||||
636 | /// | ||||||||||
637 | /// If this routine fails to unswitch the switch it returns false. | ||||||||||
638 | /// | ||||||||||
639 | /// If the switch can be unswitched, this routine splits the preheader and | ||||||||||
640 | /// copies the switch above that split. If the default case is one of the | ||||||||||
641 | /// exiting cases, it copies the non-exiting cases and points them at the new | ||||||||||
642 | /// preheader. If the default case is not exiting, it copies the exiting cases | ||||||||||
643 | /// and points the default at the preheader. It preserves loop simplified form | ||||||||||
644 | /// (splitting the exit blocks as necessary). It simplifies the switch within | ||||||||||
645 | /// the loop by removing now-dead cases. If the default case is one of those | ||||||||||
646 | /// unswitched, it replaces its destination with a new basic block containing | ||||||||||
647 | /// only unreachable. Such basic blocks, while technically loop exits, are not | ||||||||||
648 | /// considered for unswitching so this is a stable transform and the same | ||||||||||
649 | /// switch will not be revisited. If after unswitching there is only a single | ||||||||||
650 | /// in-loop successor, the switch is further simplified to an unconditional | ||||||||||
651 | /// branch. Still more cleanup can be done with some simplifycfg like pass. | ||||||||||
652 | /// | ||||||||||
653 | /// If `SE` is not null, it will be updated based on the potential loop SCEVs | ||||||||||
654 | /// invalidated by this. | ||||||||||
655 | static bool unswitchTrivialSwitch(Loop &L, SwitchInst &SI, DominatorTree &DT, | ||||||||||
656 | LoopInfo &LI, ScalarEvolution *SE, | ||||||||||
657 | MemorySSAUpdater *MSSAU) { | ||||||||||
658 | LLVM_DEBUG(dbgs() << " Trying to unswitch switch: " << SI << "\n")do { } while (false); | ||||||||||
659 | Value *LoopCond = SI.getCondition(); | ||||||||||
660 | |||||||||||
661 | // If this isn't switching on an invariant condition, we can't unswitch it. | ||||||||||
662 | if (!L.isLoopInvariant(LoopCond)) | ||||||||||
663 | return false; | ||||||||||
664 | |||||||||||
665 | auto *ParentBB = SI.getParent(); | ||||||||||
666 | |||||||||||
667 | // The same check must be used both for the default and the exit cases. We | ||||||||||
668 | // should never leave edges from the switch instruction to a basic block that | ||||||||||
669 | // we are unswitching, hence the condition used to determine the default case | ||||||||||
670 | // needs to also be used to populate ExitCaseIndices, which is then used to | ||||||||||
671 | // remove cases from the switch. | ||||||||||
672 | auto IsTriviallyUnswitchableExitBlock = [&](BasicBlock &BBToCheck) { | ||||||||||
673 | // BBToCheck is not an exit block if it is inside loop L. | ||||||||||
674 | if (L.contains(&BBToCheck)) | ||||||||||
675 | return false; | ||||||||||
676 | // BBToCheck is not trivial to unswitch if its phis aren't loop invariant. | ||||||||||
677 | if (!areLoopExitPHIsLoopInvariant(L, *ParentBB, BBToCheck)) | ||||||||||
678 | return false; | ||||||||||
679 | // We do not unswitch a block that only has an unreachable statement, as | ||||||||||
680 | // it's possible this is a previously unswitched block. Only unswitch if | ||||||||||
681 | // either the terminator is not unreachable, or, if it is, it's not the only | ||||||||||
682 | // instruction in the block. | ||||||||||
683 | auto *TI = BBToCheck.getTerminator(); | ||||||||||
684 | bool isUnreachable = isa<UnreachableInst>(TI); | ||||||||||
685 | return !isUnreachable || | ||||||||||
686 | (isUnreachable && (BBToCheck.getFirstNonPHIOrDbg() != TI)); | ||||||||||
687 | }; | ||||||||||
688 | |||||||||||
689 | SmallVector<int, 4> ExitCaseIndices; | ||||||||||
690 | for (auto Case : SI.cases()) | ||||||||||
691 | if (IsTriviallyUnswitchableExitBlock(*Case.getCaseSuccessor())) | ||||||||||
692 | ExitCaseIndices.push_back(Case.getCaseIndex()); | ||||||||||
693 | BasicBlock *DefaultExitBB = nullptr; | ||||||||||
694 | SwitchInstProfUpdateWrapper::CaseWeightOpt DefaultCaseWeight = | ||||||||||
695 | SwitchInstProfUpdateWrapper::getSuccessorWeight(SI, 0); | ||||||||||
696 | if (IsTriviallyUnswitchableExitBlock(*SI.getDefaultDest())) { | ||||||||||
697 | DefaultExitBB = SI.getDefaultDest(); | ||||||||||
698 | } else if (ExitCaseIndices.empty()) | ||||||||||
699 | return false; | ||||||||||
700 | |||||||||||
701 | LLVM_DEBUG(dbgs() << " unswitching trivial switch...\n")do { } while (false); | ||||||||||
702 | |||||||||||
703 | if (MSSAU && VerifyMemorySSA) | ||||||||||
704 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
705 | |||||||||||
706 | // We may need to invalidate SCEVs for the outermost loop reached by any of | ||||||||||
707 | // the exits. | ||||||||||
708 | Loop *OuterL = &L; | ||||||||||
709 | |||||||||||
710 | if (DefaultExitBB) { | ||||||||||
711 | // Clear out the default destination temporarily to allow accurate | ||||||||||
712 | // predecessor lists to be examined below. | ||||||||||
713 | SI.setDefaultDest(nullptr); | ||||||||||
714 | // Check the loop containing this exit. | ||||||||||
715 | Loop *ExitL = LI.getLoopFor(DefaultExitBB); | ||||||||||
716 | if (!ExitL || ExitL->contains(OuterL)) | ||||||||||
717 | OuterL = ExitL; | ||||||||||
718 | } | ||||||||||
719 | |||||||||||
720 | // Store the exit cases into a separate data structure and remove them from | ||||||||||
721 | // the switch. | ||||||||||
722 | SmallVector<std::tuple<ConstantInt *, BasicBlock *, | ||||||||||
723 | SwitchInstProfUpdateWrapper::CaseWeightOpt>, | ||||||||||
724 | 4> ExitCases; | ||||||||||
725 | ExitCases.reserve(ExitCaseIndices.size()); | ||||||||||
726 | SwitchInstProfUpdateWrapper SIW(SI); | ||||||||||
727 | // We walk the case indices backwards so that we remove the last case first | ||||||||||
728 | // and don't disrupt the earlier indices. | ||||||||||
729 | for (unsigned Index : reverse(ExitCaseIndices)) { | ||||||||||
730 | auto CaseI = SI.case_begin() + Index; | ||||||||||
731 | // Compute the outer loop from this exit. | ||||||||||
732 | Loop *ExitL = LI.getLoopFor(CaseI->getCaseSuccessor()); | ||||||||||
733 | if (!ExitL || ExitL->contains(OuterL)) | ||||||||||
734 | OuterL = ExitL; | ||||||||||
735 | // Save the value of this case. | ||||||||||
736 | auto W = SIW.getSuccessorWeight(CaseI->getSuccessorIndex()); | ||||||||||
737 | ExitCases.emplace_back(CaseI->getCaseValue(), CaseI->getCaseSuccessor(), W); | ||||||||||
738 | // Delete the unswitched cases. | ||||||||||
739 | SIW.removeCase(CaseI); | ||||||||||
740 | } | ||||||||||
741 | |||||||||||
742 | if (SE) { | ||||||||||
743 | if (OuterL) | ||||||||||
744 | SE->forgetLoop(OuterL); | ||||||||||
745 | else | ||||||||||
746 | SE->forgetTopmostLoop(&L); | ||||||||||
747 | } | ||||||||||
748 | |||||||||||
749 | // Check if after this all of the remaining cases point at the same | ||||||||||
750 | // successor. | ||||||||||
751 | BasicBlock *CommonSuccBB = nullptr; | ||||||||||
752 | if (SI.getNumCases() > 0 && | ||||||||||
753 | all_of(drop_begin(SI.cases()), [&SI](const SwitchInst::CaseHandle &Case) { | ||||||||||
754 | return Case.getCaseSuccessor() == SI.case_begin()->getCaseSuccessor(); | ||||||||||
755 | })) | ||||||||||
756 | CommonSuccBB = SI.case_begin()->getCaseSuccessor(); | ||||||||||
757 | if (!DefaultExitBB) { | ||||||||||
758 | // If we're not unswitching the default, we need it to match any cases to | ||||||||||
759 | // have a common successor or if we have no cases it is the common | ||||||||||
760 | // successor. | ||||||||||
761 | if (SI.getNumCases() == 0) | ||||||||||
762 | CommonSuccBB = SI.getDefaultDest(); | ||||||||||
763 | else if (SI.getDefaultDest() != CommonSuccBB) | ||||||||||
764 | CommonSuccBB = nullptr; | ||||||||||
765 | } | ||||||||||
766 | |||||||||||
767 | // Split the preheader, so that we know that there is a safe place to insert | ||||||||||
768 | // the switch. | ||||||||||
769 | BasicBlock *OldPH = L.getLoopPreheader(); | ||||||||||
770 | BasicBlock *NewPH = SplitEdge(OldPH, L.getHeader(), &DT, &LI, MSSAU); | ||||||||||
771 | OldPH->getTerminator()->eraseFromParent(); | ||||||||||
772 | |||||||||||
773 | // Now add the unswitched switch. | ||||||||||
774 | auto *NewSI = SwitchInst::Create(LoopCond, NewPH, ExitCases.size(), OldPH); | ||||||||||
775 | SwitchInstProfUpdateWrapper NewSIW(*NewSI); | ||||||||||
776 | |||||||||||
777 | // Rewrite the IR for the unswitched basic blocks. This requires two steps. | ||||||||||
778 | // First, we split any exit blocks with remaining in-loop predecessors. Then | ||||||||||
779 | // we update the PHIs in one of two ways depending on if there was a split. | ||||||||||
780 | // We walk in reverse so that we split in the same order as the cases | ||||||||||
781 | // appeared. This is purely for convenience of reading the resulting IR, but | ||||||||||
782 | // it doesn't cost anything really. | ||||||||||
783 | SmallPtrSet<BasicBlock *, 2> UnswitchedExitBBs; | ||||||||||
784 | SmallDenseMap<BasicBlock *, BasicBlock *, 2> SplitExitBBMap; | ||||||||||
785 | // Handle the default exit if necessary. | ||||||||||
786 | // FIXME: It'd be great if we could merge this with the loop below but LLVM's | ||||||||||
787 | // ranges aren't quite powerful enough yet. | ||||||||||
788 | if (DefaultExitBB) { | ||||||||||
789 | if (pred_empty(DefaultExitBB)) { | ||||||||||
790 | UnswitchedExitBBs.insert(DefaultExitBB); | ||||||||||
791 | rewritePHINodesForUnswitchedExitBlock(*DefaultExitBB, *ParentBB, *OldPH); | ||||||||||
792 | } else { | ||||||||||
793 | auto *SplitBB = | ||||||||||
794 | SplitBlock(DefaultExitBB, &DefaultExitBB->front(), &DT, &LI, MSSAU); | ||||||||||
795 | rewritePHINodesForExitAndUnswitchedBlocks(*DefaultExitBB, *SplitBB, | ||||||||||
796 | *ParentBB, *OldPH, | ||||||||||
797 | /*FullUnswitch*/ true); | ||||||||||
798 | DefaultExitBB = SplitExitBBMap[DefaultExitBB] = SplitBB; | ||||||||||
799 | } | ||||||||||
800 | } | ||||||||||
801 | // Note that we must use a reference in the for loop so that we update the | ||||||||||
802 | // container. | ||||||||||
803 | for (auto &ExitCase : reverse(ExitCases)) { | ||||||||||
804 | // Grab a reference to the exit block in the pair so that we can update it. | ||||||||||
805 | BasicBlock *ExitBB = std::get<1>(ExitCase); | ||||||||||
806 | |||||||||||
807 | // If this case is the last edge into the exit block, we can simply reuse it | ||||||||||
808 | // as it will no longer be a loop exit. No mapping necessary. | ||||||||||
809 | if (pred_empty(ExitBB)) { | ||||||||||
810 | // Only rewrite once. | ||||||||||
811 | if (UnswitchedExitBBs.insert(ExitBB).second) | ||||||||||
812 | rewritePHINodesForUnswitchedExitBlock(*ExitBB, *ParentBB, *OldPH); | ||||||||||
813 | continue; | ||||||||||
814 | } | ||||||||||
815 | |||||||||||
816 | // Otherwise we need to split the exit block so that we retain an exit | ||||||||||
817 | // block from the loop and a target for the unswitched condition. | ||||||||||
818 | BasicBlock *&SplitExitBB = SplitExitBBMap[ExitBB]; | ||||||||||
819 | if (!SplitExitBB) { | ||||||||||
820 | // If this is the first time we see this, do the split and remember it. | ||||||||||
821 | SplitExitBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU); | ||||||||||
822 | rewritePHINodesForExitAndUnswitchedBlocks(*ExitBB, *SplitExitBB, | ||||||||||
823 | *ParentBB, *OldPH, | ||||||||||
824 | /*FullUnswitch*/ true); | ||||||||||
825 | } | ||||||||||
826 | // Update the case pair to point to the split block. | ||||||||||
827 | std::get<1>(ExitCase) = SplitExitBB; | ||||||||||
828 | } | ||||||||||
829 | |||||||||||
830 | // Now add the unswitched cases. We do this in reverse order as we built them | ||||||||||
831 | // in reverse order. | ||||||||||
832 | for (auto &ExitCase : reverse(ExitCases)) { | ||||||||||
833 | ConstantInt *CaseVal = std::get<0>(ExitCase); | ||||||||||
834 | BasicBlock *UnswitchedBB = std::get<1>(ExitCase); | ||||||||||
835 | |||||||||||
836 | NewSIW.addCase(CaseVal, UnswitchedBB, std::get<2>(ExitCase)); | ||||||||||
837 | } | ||||||||||
838 | |||||||||||
839 | // If the default was unswitched, re-point it and add explicit cases for | ||||||||||
840 | // entering the loop. | ||||||||||
841 | if (DefaultExitBB) { | ||||||||||
842 | NewSIW->setDefaultDest(DefaultExitBB); | ||||||||||
843 | NewSIW.setSuccessorWeight(0, DefaultCaseWeight); | ||||||||||
844 | |||||||||||
845 | // We removed all the exit cases, so we just copy the cases to the | ||||||||||
846 | // unswitched switch. | ||||||||||
847 | for (const auto &Case : SI.cases()) | ||||||||||
848 | NewSIW.addCase(Case.getCaseValue(), NewPH, | ||||||||||
849 | SIW.getSuccessorWeight(Case.getSuccessorIndex())); | ||||||||||
850 | } else if (DefaultCaseWeight) { | ||||||||||
851 | // We have to set branch weight of the default case. | ||||||||||
852 | uint64_t SW = *DefaultCaseWeight; | ||||||||||
853 | for (const auto &Case : SI.cases()) { | ||||||||||
854 | auto W = SIW.getSuccessorWeight(Case.getSuccessorIndex()); | ||||||||||
855 | assert(W &&((void)0) | ||||||||||
856 | "case weight must be defined as default case weight is defined")((void)0); | ||||||||||
857 | SW += *W; | ||||||||||
858 | } | ||||||||||
859 | NewSIW.setSuccessorWeight(0, SW); | ||||||||||
860 | } | ||||||||||
861 | |||||||||||
862 | // If we ended up with a common successor for every path through the switch | ||||||||||
863 | // after unswitching, rewrite it to an unconditional branch to make it easy | ||||||||||
864 | // to recognize. Otherwise we potentially have to recognize the default case | ||||||||||
865 | // pointing at unreachable and other complexity. | ||||||||||
866 | if (CommonSuccBB) { | ||||||||||
867 | BasicBlock *BB = SI.getParent(); | ||||||||||
868 | // We may have had multiple edges to this common successor block, so remove | ||||||||||
869 | // them as predecessors. We skip the first one, either the default or the | ||||||||||
870 | // actual first case. | ||||||||||
871 | bool SkippedFirst = DefaultExitBB == nullptr; | ||||||||||
872 | for (auto Case : SI.cases()) { | ||||||||||
873 | assert(Case.getCaseSuccessor() == CommonSuccBB &&((void)0) | ||||||||||
874 | "Non-common successor!")((void)0); | ||||||||||
875 | (void)Case; | ||||||||||
876 | if (!SkippedFirst) { | ||||||||||
877 | SkippedFirst = true; | ||||||||||
878 | continue; | ||||||||||
879 | } | ||||||||||
880 | CommonSuccBB->removePredecessor(BB, | ||||||||||
881 | /*KeepOneInputPHIs*/ true); | ||||||||||
882 | } | ||||||||||
883 | // Now nuke the switch and replace it with a direct branch. | ||||||||||
884 | SIW.eraseFromParent(); | ||||||||||
885 | BranchInst::Create(CommonSuccBB, BB); | ||||||||||
886 | } else if (DefaultExitBB) { | ||||||||||
887 | assert(SI.getNumCases() > 0 &&((void)0) | ||||||||||
888 | "If we had no cases we'd have a common successor!")((void)0); | ||||||||||
889 | // Move the last case to the default successor. This is valid as if the | ||||||||||
890 | // default got unswitched it cannot be reached. This has the advantage of | ||||||||||
891 | // being simple and keeping the number of edges from this switch to | ||||||||||
892 | // successors the same, and avoiding any PHI update complexity. | ||||||||||
893 | auto LastCaseI = std::prev(SI.case_end()); | ||||||||||
894 | |||||||||||
895 | SI.setDefaultDest(LastCaseI->getCaseSuccessor()); | ||||||||||
896 | SIW.setSuccessorWeight( | ||||||||||
897 | 0, SIW.getSuccessorWeight(LastCaseI->getSuccessorIndex())); | ||||||||||
898 | SIW.removeCase(LastCaseI); | ||||||||||
899 | } | ||||||||||
900 | |||||||||||
901 | // Walk the unswitched exit blocks and the unswitched split blocks and update | ||||||||||
902 | // the dominator tree based on the CFG edits. While we are walking unordered | ||||||||||
903 | // containers here, the API for applyUpdates takes an unordered list of | ||||||||||
904 | // updates and requires them to not contain duplicates. | ||||||||||
905 | SmallVector<DominatorTree::UpdateType, 4> DTUpdates; | ||||||||||
906 | for (auto *UnswitchedExitBB : UnswitchedExitBBs) { | ||||||||||
907 | DTUpdates.push_back({DT.Delete, ParentBB, UnswitchedExitBB}); | ||||||||||
908 | DTUpdates.push_back({DT.Insert, OldPH, UnswitchedExitBB}); | ||||||||||
909 | } | ||||||||||
910 | for (auto SplitUnswitchedPair : SplitExitBBMap) { | ||||||||||
911 | DTUpdates.push_back({DT.Delete, ParentBB, SplitUnswitchedPair.first}); | ||||||||||
912 | DTUpdates.push_back({DT.Insert, OldPH, SplitUnswitchedPair.second}); | ||||||||||
913 | } | ||||||||||
914 | |||||||||||
915 | if (MSSAU) { | ||||||||||
916 | MSSAU->applyUpdates(DTUpdates, DT, /*UpdateDT=*/true); | ||||||||||
917 | if (VerifyMemorySSA) | ||||||||||
918 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
919 | } else { | ||||||||||
920 | DT.applyUpdates(DTUpdates); | ||||||||||
921 | } | ||||||||||
922 | |||||||||||
923 | assert(DT.verify(DominatorTree::VerificationLevel::Fast))((void)0); | ||||||||||
924 | |||||||||||
925 | // We may have changed the nesting relationship for this loop so hoist it to | ||||||||||
926 | // its correct parent if needed. | ||||||||||
927 | hoistLoopToNewParent(L, *NewPH, DT, LI, MSSAU, SE); | ||||||||||
928 | |||||||||||
929 | if (MSSAU && VerifyMemorySSA) | ||||||||||
930 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
931 | |||||||||||
932 | ++NumTrivial; | ||||||||||
933 | ++NumSwitches; | ||||||||||
934 | LLVM_DEBUG(dbgs() << " done: unswitching trivial switch...\n")do { } while (false); | ||||||||||
935 | return true; | ||||||||||
936 | } | ||||||||||
937 | |||||||||||
938 | /// This routine scans the loop to find a branch or switch which occurs before | ||||||||||
939 | /// any side effects occur. These can potentially be unswitched without | ||||||||||
940 | /// duplicating the loop. If a branch or switch is successfully unswitched the | ||||||||||
941 | /// scanning continues to see if subsequent branches or switches have become | ||||||||||
942 | /// trivial. Once all trivial candidates have been unswitched, this routine | ||||||||||
943 | /// returns. | ||||||||||
944 | /// | ||||||||||
945 | /// The return value indicates whether anything was unswitched (and therefore | ||||||||||
946 | /// changed). | ||||||||||
947 | /// | ||||||||||
948 | /// If `SE` is not null, it will be updated based on the potential loop SCEVs | ||||||||||
949 | /// invalidated by this. | ||||||||||
950 | static bool unswitchAllTrivialConditions(Loop &L, DominatorTree &DT, | ||||||||||
951 | LoopInfo &LI, ScalarEvolution *SE, | ||||||||||
952 | MemorySSAUpdater *MSSAU) { | ||||||||||
953 | bool Changed = false; | ||||||||||
954 | |||||||||||
955 | // If loop header has only one reachable successor we should keep looking for | ||||||||||
956 | // trivial condition candidates in the successor as well. An alternative is | ||||||||||
957 | // to constant fold conditions and merge successors into loop header (then we | ||||||||||
958 | // only need to check header's terminator). The reason for not doing this in | ||||||||||
959 | // LoopUnswitch pass is that it could potentially break LoopPassManager's | ||||||||||
960 | // invariants. Folding dead branches could either eliminate the current loop | ||||||||||
961 | // or make other loops unreachable. LCSSA form might also not be preserved | ||||||||||
962 | // after deleting branches. The following code keeps traversing loop header's | ||||||||||
963 | // successors until it finds the trivial condition candidate (condition that | ||||||||||
964 | // is not a constant). Since unswitching generates branches with constant | ||||||||||
965 | // conditions, this scenario could be very common in practice. | ||||||||||
966 | BasicBlock *CurrentBB = L.getHeader(); | ||||||||||
967 | SmallPtrSet<BasicBlock *, 8> Visited; | ||||||||||
968 | Visited.insert(CurrentBB); | ||||||||||
969 | do { | ||||||||||
970 | // Check if there are any side-effecting instructions (e.g. stores, calls, | ||||||||||
971 | // volatile loads) in the part of the loop that the code *would* execute | ||||||||||
972 | // without unswitching. | ||||||||||
973 | if (MSSAU) // Possible early exit with MSSA | ||||||||||
974 | if (auto *Defs = MSSAU->getMemorySSA()->getBlockDefs(CurrentBB)) | ||||||||||
975 | if (!isa<MemoryPhi>(*Defs->begin()) || (++Defs->begin() != Defs->end())) | ||||||||||
976 | return Changed; | ||||||||||
977 | if (llvm::any_of(*CurrentBB, | ||||||||||
978 | [](Instruction &I) { return I.mayHaveSideEffects(); })) | ||||||||||
979 | return Changed; | ||||||||||
980 | |||||||||||
981 | Instruction *CurrentTerm = CurrentBB->getTerminator(); | ||||||||||
982 | |||||||||||
983 | if (auto *SI = dyn_cast<SwitchInst>(CurrentTerm)) { | ||||||||||
984 | // Don't bother trying to unswitch past a switch with a constant | ||||||||||
985 | // condition. This should be removed prior to running this pass by | ||||||||||
986 | // simplifycfg. | ||||||||||
987 | if (isa<Constant>(SI->getCondition())) | ||||||||||
988 | return Changed; | ||||||||||
989 | |||||||||||
990 | if (!unswitchTrivialSwitch(L, *SI, DT, LI, SE, MSSAU)) | ||||||||||
991 | // Couldn't unswitch this one so we're done. | ||||||||||
992 | return Changed; | ||||||||||
993 | |||||||||||
994 | // Mark that we managed to unswitch something. | ||||||||||
995 | Changed = true; | ||||||||||
996 | |||||||||||
997 | // If unswitching turned the terminator into an unconditional branch then | ||||||||||
998 | // we can continue. The unswitching logic specifically works to fold any | ||||||||||
999 | // cases it can into an unconditional branch to make it easier to | ||||||||||
1000 | // recognize here. | ||||||||||
1001 | auto *BI = dyn_cast<BranchInst>(CurrentBB->getTerminator()); | ||||||||||
1002 | if (!BI || BI->isConditional()) | ||||||||||
1003 | return Changed; | ||||||||||
1004 | |||||||||||
1005 | CurrentBB = BI->getSuccessor(0); | ||||||||||
1006 | continue; | ||||||||||
1007 | } | ||||||||||
1008 | |||||||||||
1009 | auto *BI = dyn_cast<BranchInst>(CurrentTerm); | ||||||||||
1010 | if (!BI) | ||||||||||
1011 | // We do not understand other terminator instructions. | ||||||||||
1012 | return Changed; | ||||||||||
1013 | |||||||||||
1014 | // Don't bother trying to unswitch past an unconditional branch or a branch | ||||||||||
1015 | // with a constant value. These should be removed by simplifycfg prior to | ||||||||||
1016 | // running this pass. | ||||||||||
1017 | if (!BI->isConditional() || isa<Constant>(BI->getCondition())) | ||||||||||
1018 | return Changed; | ||||||||||
1019 | |||||||||||
1020 | // Found a trivial condition candidate: non-foldable conditional branch. If | ||||||||||
1021 | // we fail to unswitch this, we can't do anything else that is trivial. | ||||||||||
1022 | if (!unswitchTrivialBranch(L, *BI, DT, LI, SE, MSSAU)) | ||||||||||
1023 | return Changed; | ||||||||||
1024 | |||||||||||
1025 | // Mark that we managed to unswitch something. | ||||||||||
1026 | Changed = true; | ||||||||||
1027 | |||||||||||
1028 | // If we only unswitched some of the conditions feeding the branch, we won't | ||||||||||
1029 | // have collapsed it to a single successor. | ||||||||||
1030 | BI = cast<BranchInst>(CurrentBB->getTerminator()); | ||||||||||
1031 | if (BI->isConditional()) | ||||||||||
1032 | return Changed; | ||||||||||
1033 | |||||||||||
1034 | // Follow the newly unconditional branch into its successor. | ||||||||||
1035 | CurrentBB = BI->getSuccessor(0); | ||||||||||
1036 | |||||||||||
1037 | // When continuing, if we exit the loop or reach a previous visited block, | ||||||||||
1038 | // then we can not reach any trivial condition candidates (unfoldable | ||||||||||
1039 | // branch instructions or switch instructions) and no unswitch can happen. | ||||||||||
1040 | } while (L.contains(CurrentBB) && Visited.insert(CurrentBB).second); | ||||||||||
1041 | |||||||||||
1042 | return Changed; | ||||||||||
1043 | } | ||||||||||
1044 | |||||||||||
1045 | /// Build the cloned blocks for an unswitched copy of the given loop. | ||||||||||
1046 | /// | ||||||||||
1047 | /// The cloned blocks are inserted before the loop preheader (`LoopPH`) and | ||||||||||
1048 | /// after the split block (`SplitBB`) that will be used to select between the | ||||||||||
1049 | /// cloned and original loop. | ||||||||||
1050 | /// | ||||||||||
1051 | /// This routine handles cloning all of the necessary loop blocks and exit | ||||||||||
1052 | /// blocks including rewriting their instructions and the relevant PHI nodes. | ||||||||||
1053 | /// Any loop blocks or exit blocks which are dominated by a different successor | ||||||||||
1054 | /// than the one for this clone of the loop blocks can be trivially skipped. We | ||||||||||
1055 | /// use the `DominatingSucc` map to determine whether a block satisfies that | ||||||||||
1056 | /// property with a simple map lookup. | ||||||||||
1057 | /// | ||||||||||
1058 | /// It also correctly creates the unconditional branch in the cloned | ||||||||||
1059 | /// unswitched parent block to only point at the unswitched successor. | ||||||||||
1060 | /// | ||||||||||
1061 | /// This does not handle most of the necessary updates to `LoopInfo`. Only exit | ||||||||||
1062 | /// block splitting is correctly reflected in `LoopInfo`, essentially all of | ||||||||||
1063 | /// the cloned blocks (and their loops) are left without full `LoopInfo` | ||||||||||
1064 | /// updates. This also doesn't fully update `DominatorTree`. It adds the cloned | ||||||||||
1065 | /// blocks to them but doesn't create the cloned `DominatorTree` structure and | ||||||||||
1066 | /// instead the caller must recompute an accurate DT. It *does* correctly | ||||||||||
1067 | /// update the `AssumptionCache` provided in `AC`. | ||||||||||
1068 | static BasicBlock *buildClonedLoopBlocks( | ||||||||||
1069 | Loop &L, BasicBlock *LoopPH, BasicBlock *SplitBB, | ||||||||||
1070 | ArrayRef<BasicBlock *> ExitBlocks, BasicBlock *ParentBB, | ||||||||||
1071 | BasicBlock *UnswitchedSuccBB, BasicBlock *ContinueSuccBB, | ||||||||||
1072 | const SmallDenseMap<BasicBlock *, BasicBlock *, 16> &DominatingSucc, | ||||||||||
1073 | ValueToValueMapTy &VMap, | ||||||||||
1074 | SmallVectorImpl<DominatorTree::UpdateType> &DTUpdates, AssumptionCache &AC, | ||||||||||
1075 | DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) { | ||||||||||
1076 | SmallVector<BasicBlock *, 4> NewBlocks; | ||||||||||
1077 | NewBlocks.reserve(L.getNumBlocks() + ExitBlocks.size()); | ||||||||||
1078 | |||||||||||
1079 | // We will need to clone a bunch of blocks, wrap up the clone operation in | ||||||||||
1080 | // a helper. | ||||||||||
1081 | auto CloneBlock = [&](BasicBlock *OldBB) { | ||||||||||
1082 | // Clone the basic block and insert it before the new preheader. | ||||||||||
1083 | BasicBlock *NewBB = CloneBasicBlock(OldBB, VMap, ".us", OldBB->getParent()); | ||||||||||
1084 | NewBB->moveBefore(LoopPH); | ||||||||||
1085 | |||||||||||
1086 | // Record this block and the mapping. | ||||||||||
1087 | NewBlocks.push_back(NewBB); | ||||||||||
1088 | VMap[OldBB] = NewBB; | ||||||||||
1089 | |||||||||||
1090 | return NewBB; | ||||||||||
1091 | }; | ||||||||||
1092 | |||||||||||
1093 | // We skip cloning blocks when they have a dominating succ that is not the | ||||||||||
1094 | // succ we are cloning for. | ||||||||||
1095 | auto SkipBlock = [&](BasicBlock *BB) { | ||||||||||
1096 | auto It = DominatingSucc.find(BB); | ||||||||||
1097 | return It != DominatingSucc.end() && It->second != UnswitchedSuccBB; | ||||||||||
1098 | }; | ||||||||||
1099 | |||||||||||
1100 | // First, clone the preheader. | ||||||||||
1101 | auto *ClonedPH = CloneBlock(LoopPH); | ||||||||||
1102 | |||||||||||
1103 | // Then clone all the loop blocks, skipping the ones that aren't necessary. | ||||||||||
1104 | for (auto *LoopBB : L.blocks()) | ||||||||||
1105 | if (!SkipBlock(LoopBB)) | ||||||||||
1106 | CloneBlock(LoopBB); | ||||||||||
1107 | |||||||||||
1108 | // Split all the loop exit edges so that when we clone the exit blocks, if | ||||||||||
1109 | // any of the exit blocks are *also* a preheader for some other loop, we | ||||||||||
1110 | // don't create multiple predecessors entering the loop header. | ||||||||||
1111 | for (auto *ExitBB : ExitBlocks) { | ||||||||||
1112 | if (SkipBlock(ExitBB)) | ||||||||||
1113 | continue; | ||||||||||
1114 | |||||||||||
1115 | // When we are going to clone an exit, we don't need to clone all the | ||||||||||
1116 | // instructions in the exit block and we want to ensure we have an easy | ||||||||||
1117 | // place to merge the CFG, so split the exit first. This is always safe to | ||||||||||
1118 | // do because there cannot be any non-loop predecessors of a loop exit in | ||||||||||
1119 | // loop simplified form. | ||||||||||
1120 | auto *MergeBB = SplitBlock(ExitBB, &ExitBB->front(), &DT, &LI, MSSAU); | ||||||||||
1121 | |||||||||||
1122 | // Rearrange the names to make it easier to write test cases by having the | ||||||||||
1123 | // exit block carry the suffix rather than the merge block carrying the | ||||||||||
1124 | // suffix. | ||||||||||
1125 | MergeBB->takeName(ExitBB); | ||||||||||
1126 | ExitBB->setName(Twine(MergeBB->getName()) + ".split"); | ||||||||||
1127 | |||||||||||
1128 | // Now clone the original exit block. | ||||||||||
1129 | auto *ClonedExitBB = CloneBlock(ExitBB); | ||||||||||
1130 | assert(ClonedExitBB->getTerminator()->getNumSuccessors() == 1 &&((void)0) | ||||||||||
1131 | "Exit block should have been split to have one successor!")((void)0); | ||||||||||
1132 | assert(ClonedExitBB->getTerminator()->getSuccessor(0) == MergeBB &&((void)0) | ||||||||||
1133 | "Cloned exit block has the wrong successor!")((void)0); | ||||||||||
1134 | |||||||||||
1135 | // Remap any cloned instructions and create a merge phi node for them. | ||||||||||
1136 | for (auto ZippedInsts : llvm::zip_first( | ||||||||||
1137 | llvm::make_range(ExitBB->begin(), std::prev(ExitBB->end())), | ||||||||||
1138 | llvm::make_range(ClonedExitBB->begin(), | ||||||||||
1139 | std::prev(ClonedExitBB->end())))) { | ||||||||||
1140 | Instruction &I = std::get<0>(ZippedInsts); | ||||||||||
1141 | Instruction &ClonedI = std::get<1>(ZippedInsts); | ||||||||||
1142 | |||||||||||
1143 | // The only instructions in the exit block should be PHI nodes and | ||||||||||
1144 | // potentially a landing pad. | ||||||||||
1145 | assert(((void)0) | ||||||||||
1146 | (isa<PHINode>(I) || isa<LandingPadInst>(I) || isa<CatchPadInst>(I)) &&((void)0) | ||||||||||
1147 | "Bad instruction in exit block!")((void)0); | ||||||||||
1148 | // We should have a value map between the instruction and its clone. | ||||||||||
1149 | assert(VMap.lookup(&I) == &ClonedI && "Mismatch in the value map!")((void)0); | ||||||||||
1150 | |||||||||||
1151 | auto *MergePN = | ||||||||||
1152 | PHINode::Create(I.getType(), /*NumReservedValues*/ 2, ".us-phi", | ||||||||||
1153 | &*MergeBB->getFirstInsertionPt()); | ||||||||||
1154 | I.replaceAllUsesWith(MergePN); | ||||||||||
1155 | MergePN->addIncoming(&I, ExitBB); | ||||||||||
1156 | MergePN->addIncoming(&ClonedI, ClonedExitBB); | ||||||||||
1157 | } | ||||||||||
1158 | } | ||||||||||
1159 | |||||||||||
1160 | // Rewrite the instructions in the cloned blocks to refer to the instructions | ||||||||||
1161 | // in the cloned blocks. We have to do this as a second pass so that we have | ||||||||||
1162 | // everything available. Also, we have inserted new instructions which may | ||||||||||
1163 | // include assume intrinsics, so we update the assumption cache while | ||||||||||
1164 | // processing this. | ||||||||||
1165 | for (auto *ClonedBB : NewBlocks) | ||||||||||
1166 | for (Instruction &I : *ClonedBB) { | ||||||||||
1167 | RemapInstruction(&I, VMap, | ||||||||||
1168 | RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); | ||||||||||
1169 | if (auto *II = dyn_cast<AssumeInst>(&I)) | ||||||||||
1170 | AC.registerAssumption(II); | ||||||||||
1171 | } | ||||||||||
1172 | |||||||||||
1173 | // Update any PHI nodes in the cloned successors of the skipped blocks to not | ||||||||||
1174 | // have spurious incoming values. | ||||||||||
1175 | for (auto *LoopBB : L.blocks()) | ||||||||||
1176 | if (SkipBlock(LoopBB)) | ||||||||||
1177 | for (auto *SuccBB : successors(LoopBB)) | ||||||||||
1178 | if (auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB))) | ||||||||||
1179 | for (PHINode &PN : ClonedSuccBB->phis()) | ||||||||||
1180 | PN.removeIncomingValue(LoopBB, /*DeletePHIIfEmpty*/ false); | ||||||||||
1181 | |||||||||||
1182 | // Remove the cloned parent as a predecessor of any successor we ended up | ||||||||||
1183 | // cloning other than the unswitched one. | ||||||||||
1184 | auto *ClonedParentBB = cast<BasicBlock>(VMap.lookup(ParentBB)); | ||||||||||
1185 | for (auto *SuccBB : successors(ParentBB)) { | ||||||||||
1186 | if (SuccBB == UnswitchedSuccBB) | ||||||||||
1187 | continue; | ||||||||||
1188 | |||||||||||
1189 | auto *ClonedSuccBB = cast_or_null<BasicBlock>(VMap.lookup(SuccBB)); | ||||||||||
1190 | if (!ClonedSuccBB) | ||||||||||
1191 | continue; | ||||||||||
1192 | |||||||||||
1193 | ClonedSuccBB->removePredecessor(ClonedParentBB, | ||||||||||
1194 | /*KeepOneInputPHIs*/ true); | ||||||||||
1195 | } | ||||||||||
1196 | |||||||||||
1197 | // Replace the cloned branch with an unconditional branch to the cloned | ||||||||||
1198 | // unswitched successor. | ||||||||||
1199 | auto *ClonedSuccBB = cast<BasicBlock>(VMap.lookup(UnswitchedSuccBB)); | ||||||||||
1200 | Instruction *ClonedTerminator = ClonedParentBB->getTerminator(); | ||||||||||
1201 | // Trivial Simplification. If Terminator is a conditional branch and | ||||||||||
1202 | // condition becomes dead - erase it. | ||||||||||
1203 | Value *ClonedConditionToErase = nullptr; | ||||||||||
1204 | if (auto *BI = dyn_cast<BranchInst>(ClonedTerminator)) | ||||||||||
1205 | ClonedConditionToErase = BI->getCondition(); | ||||||||||
1206 | else if (auto *SI = dyn_cast<SwitchInst>(ClonedTerminator)) | ||||||||||
1207 | ClonedConditionToErase = SI->getCondition(); | ||||||||||
1208 | |||||||||||
1209 | ClonedTerminator->eraseFromParent(); | ||||||||||
1210 | BranchInst::Create(ClonedSuccBB, ClonedParentBB); | ||||||||||
1211 | |||||||||||
1212 | if (ClonedConditionToErase) | ||||||||||
1213 | RecursivelyDeleteTriviallyDeadInstructions(ClonedConditionToErase, nullptr, | ||||||||||
1214 | MSSAU); | ||||||||||
1215 | |||||||||||
1216 | // If there are duplicate entries in the PHI nodes because of multiple edges | ||||||||||
1217 | // to the unswitched successor, we need to nuke all but one as we replaced it | ||||||||||
1218 | // with a direct branch. | ||||||||||
1219 | for (PHINode &PN : ClonedSuccBB->phis()) { | ||||||||||
1220 | bool Found = false; | ||||||||||
1221 | // Loop over the incoming operands backwards so we can easily delete as we | ||||||||||
1222 | // go without invalidating the index. | ||||||||||
1223 | for (int i = PN.getNumOperands() - 1; i >= 0; --i) { | ||||||||||
1224 | if (PN.getIncomingBlock(i) != ClonedParentBB) | ||||||||||
1225 | continue; | ||||||||||
1226 | if (!Found) { | ||||||||||
1227 | Found = true; | ||||||||||
1228 | continue; | ||||||||||
1229 | } | ||||||||||
1230 | PN.removeIncomingValue(i, /*DeletePHIIfEmpty*/ false); | ||||||||||
1231 | } | ||||||||||
1232 | } | ||||||||||
1233 | |||||||||||
1234 | // Record the domtree updates for the new blocks. | ||||||||||
1235 | SmallPtrSet<BasicBlock *, 4> SuccSet; | ||||||||||
1236 | for (auto *ClonedBB : NewBlocks) { | ||||||||||
1237 | for (auto *SuccBB : successors(ClonedBB)) | ||||||||||
1238 | if (SuccSet.insert(SuccBB).second) | ||||||||||
1239 | DTUpdates.push_back({DominatorTree::Insert, ClonedBB, SuccBB}); | ||||||||||
1240 | SuccSet.clear(); | ||||||||||
1241 | } | ||||||||||
1242 | |||||||||||
1243 | return ClonedPH; | ||||||||||
1244 | } | ||||||||||
1245 | |||||||||||
1246 | /// Recursively clone the specified loop and all of its children. | ||||||||||
1247 | /// | ||||||||||
1248 | /// The target parent loop for the clone should be provided, or can be null if | ||||||||||
1249 | /// the clone is a top-level loop. While cloning, all the blocks are mapped | ||||||||||
1250 | /// with the provided value map. The entire original loop must be present in | ||||||||||
1251 | /// the value map. The cloned loop is returned. | ||||||||||
1252 | static Loop *cloneLoopNest(Loop &OrigRootL, Loop *RootParentL, | ||||||||||
1253 | const ValueToValueMapTy &VMap, LoopInfo &LI) { | ||||||||||
1254 | auto AddClonedBlocksToLoop = [&](Loop &OrigL, Loop &ClonedL) { | ||||||||||
1255 | assert(ClonedL.getBlocks().empty() && "Must start with an empty loop!")((void)0); | ||||||||||
1256 | ClonedL.reserveBlocks(OrigL.getNumBlocks()); | ||||||||||
1257 | for (auto *BB : OrigL.blocks()) { | ||||||||||
1258 | auto *ClonedBB = cast<BasicBlock>(VMap.lookup(BB)); | ||||||||||
1259 | ClonedL.addBlockEntry(ClonedBB); | ||||||||||
1260 | if (LI.getLoopFor(BB) == &OrigL) | ||||||||||
1261 | LI.changeLoopFor(ClonedBB, &ClonedL); | ||||||||||
1262 | } | ||||||||||
1263 | }; | ||||||||||
1264 | |||||||||||
1265 | // We specially handle the first loop because it may get cloned into | ||||||||||
1266 | // a different parent and because we most commonly are cloning leaf loops. | ||||||||||
1267 | Loop *ClonedRootL = LI.AllocateLoop(); | ||||||||||
1268 | if (RootParentL) | ||||||||||
1269 | RootParentL->addChildLoop(ClonedRootL); | ||||||||||
1270 | else | ||||||||||
1271 | LI.addTopLevelLoop(ClonedRootL); | ||||||||||
1272 | AddClonedBlocksToLoop(OrigRootL, *ClonedRootL); | ||||||||||
1273 | |||||||||||
1274 | if (OrigRootL.isInnermost()) | ||||||||||
1275 | return ClonedRootL; | ||||||||||
1276 | |||||||||||
1277 | // If we have a nest, we can quickly clone the entire loop nest using an | ||||||||||
1278 | // iterative approach because it is a tree. We keep the cloned parent in the | ||||||||||
1279 | // data structure to avoid repeatedly querying through a map to find it. | ||||||||||
1280 | SmallVector<std::pair<Loop *, Loop *>, 16> LoopsToClone; | ||||||||||
1281 | // Build up the loops to clone in reverse order as we'll clone them from the | ||||||||||
1282 | // back. | ||||||||||
1283 | for (Loop *ChildL : llvm::reverse(OrigRootL)) | ||||||||||
1284 | LoopsToClone.push_back({ClonedRootL, ChildL}); | ||||||||||
1285 | do { | ||||||||||
1286 | Loop *ClonedParentL, *L; | ||||||||||
1287 | std::tie(ClonedParentL, L) = LoopsToClone.pop_back_val(); | ||||||||||
1288 | Loop *ClonedL = LI.AllocateLoop(); | ||||||||||
1289 | ClonedParentL->addChildLoop(ClonedL); | ||||||||||
1290 | AddClonedBlocksToLoop(*L, *ClonedL); | ||||||||||
1291 | for (Loop *ChildL : llvm::reverse(*L)) | ||||||||||
1292 | LoopsToClone.push_back({ClonedL, ChildL}); | ||||||||||
1293 | } while (!LoopsToClone.empty()); | ||||||||||
1294 | |||||||||||
1295 | return ClonedRootL; | ||||||||||
1296 | } | ||||||||||
1297 | |||||||||||
1298 | /// Build the cloned loops of an original loop from unswitching. | ||||||||||
1299 | /// | ||||||||||
1300 | /// Because unswitching simplifies the CFG of the loop, this isn't a trivial | ||||||||||
1301 | /// operation. We need to re-verify that there even is a loop (as the backedge | ||||||||||
1302 | /// may not have been cloned), and even if there are remaining backedges the | ||||||||||
1303 | /// backedge set may be different. However, we know that each child loop is | ||||||||||
1304 | /// undisturbed, we only need to find where to place each child loop within | ||||||||||
1305 | /// either any parent loop or within a cloned version of the original loop. | ||||||||||
1306 | /// | ||||||||||
1307 | /// Because child loops may end up cloned outside of any cloned version of the | ||||||||||
1308 | /// original loop, multiple cloned sibling loops may be created. All of them | ||||||||||
1309 | /// are returned so that the newly introduced loop nest roots can be | ||||||||||
1310 | /// identified. | ||||||||||
1311 | static void buildClonedLoops(Loop &OrigL, ArrayRef<BasicBlock *> ExitBlocks, | ||||||||||
1312 | const ValueToValueMapTy &VMap, LoopInfo &LI, | ||||||||||
1313 | SmallVectorImpl<Loop *> &NonChildClonedLoops) { | ||||||||||
1314 | Loop *ClonedL = nullptr; | ||||||||||
1315 | |||||||||||
1316 | auto *OrigPH = OrigL.getLoopPreheader(); | ||||||||||
1317 | auto *OrigHeader = OrigL.getHeader(); | ||||||||||
1318 | |||||||||||
1319 | auto *ClonedPH = cast<BasicBlock>(VMap.lookup(OrigPH)); | ||||||||||
1320 | auto *ClonedHeader = cast<BasicBlock>(VMap.lookup(OrigHeader)); | ||||||||||
1321 | |||||||||||
1322 | // We need to know the loops of the cloned exit blocks to even compute the | ||||||||||
1323 | // accurate parent loop. If we only clone exits to some parent of the | ||||||||||
1324 | // original parent, we want to clone into that outer loop. We also keep track | ||||||||||
1325 | // of the loops that our cloned exit blocks participate in. | ||||||||||
1326 | Loop *ParentL = nullptr; | ||||||||||
1327 | SmallVector<BasicBlock *, 4> ClonedExitsInLoops; | ||||||||||
1328 | SmallDenseMap<BasicBlock *, Loop *, 16> ExitLoopMap; | ||||||||||
1329 | ClonedExitsInLoops.reserve(ExitBlocks.size()); | ||||||||||
1330 | for (auto *ExitBB : ExitBlocks) | ||||||||||
1331 | if (auto *ClonedExitBB = cast_or_null<BasicBlock>(VMap.lookup(ExitBB))) | ||||||||||
1332 | if (Loop *ExitL = LI.getLoopFor(ExitBB)) { | ||||||||||
1333 | ExitLoopMap[ClonedExitBB] = ExitL; | ||||||||||
1334 | ClonedExitsInLoops.push_back(ClonedExitBB); | ||||||||||
1335 | if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) | ||||||||||
1336 | ParentL = ExitL; | ||||||||||
1337 | } | ||||||||||
1338 | assert((!ParentL || ParentL == OrigL.getParentLoop() ||((void)0) | ||||||||||
1339 | ParentL->contains(OrigL.getParentLoop())) &&((void)0) | ||||||||||
1340 | "The computed parent loop should always contain (or be) the parent of "((void)0) | ||||||||||
1341 | "the original loop.")((void)0); | ||||||||||
1342 | |||||||||||
1343 | // We build the set of blocks dominated by the cloned header from the set of | ||||||||||
1344 | // cloned blocks out of the original loop. While not all of these will | ||||||||||
1345 | // necessarily be in the cloned loop, it is enough to establish that they | ||||||||||
1346 | // aren't in unreachable cycles, etc. | ||||||||||
1347 | SmallSetVector<BasicBlock *, 16> ClonedLoopBlocks; | ||||||||||
1348 | for (auto *BB : OrigL.blocks()) | ||||||||||
1349 | if (auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB))) | ||||||||||
1350 | ClonedLoopBlocks.insert(ClonedBB); | ||||||||||
1351 | |||||||||||
1352 | // Rebuild the set of blocks that will end up in the cloned loop. We may have | ||||||||||
1353 | // skipped cloning some region of this loop which can in turn skip some of | ||||||||||
1354 | // the backedges so we have to rebuild the blocks in the loop based on the | ||||||||||
1355 | // backedges that remain after cloning. | ||||||||||
1356 | SmallVector<BasicBlock *, 16> Worklist; | ||||||||||
1357 | SmallPtrSet<BasicBlock *, 16> BlocksInClonedLoop; | ||||||||||
1358 | for (auto *Pred : predecessors(ClonedHeader)) { | ||||||||||
1359 | // The only possible non-loop header predecessor is the preheader because | ||||||||||
1360 | // we know we cloned the loop in simplified form. | ||||||||||
1361 | if (Pred == ClonedPH) | ||||||||||
1362 | continue; | ||||||||||
1363 | |||||||||||
1364 | // Because the loop was in simplified form, the only non-loop predecessor | ||||||||||
1365 | // should be the preheader. | ||||||||||
1366 | assert(ClonedLoopBlocks.count(Pred) && "Found a predecessor of the loop "((void)0) | ||||||||||
1367 | "header other than the preheader "((void)0) | ||||||||||
1368 | "that is not part of the loop!")((void)0); | ||||||||||
1369 | |||||||||||
1370 | // Insert this block into the loop set and on the first visit (and if it | ||||||||||
1371 | // isn't the header we're currently walking) put it into the worklist to | ||||||||||
1372 | // recurse through. | ||||||||||
1373 | if (BlocksInClonedLoop.insert(Pred).second && Pred != ClonedHeader) | ||||||||||
1374 | Worklist.push_back(Pred); | ||||||||||
1375 | } | ||||||||||
1376 | |||||||||||
1377 | // If we had any backedges then there *is* a cloned loop. Put the header into | ||||||||||
1378 | // the loop set and then walk the worklist backwards to find all the blocks | ||||||||||
1379 | // that remain within the loop after cloning. | ||||||||||
1380 | if (!BlocksInClonedLoop.empty()) { | ||||||||||
1381 | BlocksInClonedLoop.insert(ClonedHeader); | ||||||||||
1382 | |||||||||||
1383 | while (!Worklist.empty()) { | ||||||||||
1384 | BasicBlock *BB = Worklist.pop_back_val(); | ||||||||||
1385 | assert(BlocksInClonedLoop.count(BB) &&((void)0) | ||||||||||
1386 | "Didn't put block into the loop set!")((void)0); | ||||||||||
1387 | |||||||||||
1388 | // Insert any predecessors that are in the possible set into the cloned | ||||||||||
1389 | // set, and if the insert is successful, add them to the worklist. Note | ||||||||||
1390 | // that we filter on the blocks that are definitely reachable via the | ||||||||||
1391 | // backedge to the loop header so we may prune out dead code within the | ||||||||||
1392 | // cloned loop. | ||||||||||
1393 | for (auto *Pred : predecessors(BB)) | ||||||||||
1394 | if (ClonedLoopBlocks.count(Pred) && | ||||||||||
1395 | BlocksInClonedLoop.insert(Pred).second) | ||||||||||
1396 | Worklist.push_back(Pred); | ||||||||||
1397 | } | ||||||||||
1398 | |||||||||||
1399 | ClonedL = LI.AllocateLoop(); | ||||||||||
1400 | if (ParentL) { | ||||||||||
1401 | ParentL->addBasicBlockToLoop(ClonedPH, LI); | ||||||||||
1402 | ParentL->addChildLoop(ClonedL); | ||||||||||
1403 | } else { | ||||||||||
1404 | LI.addTopLevelLoop(ClonedL); | ||||||||||
1405 | } | ||||||||||
1406 | NonChildClonedLoops.push_back(ClonedL); | ||||||||||
1407 | |||||||||||
1408 | ClonedL->reserveBlocks(BlocksInClonedLoop.size()); | ||||||||||
1409 | // We don't want to just add the cloned loop blocks based on how we | ||||||||||
1410 | // discovered them. The original order of blocks was carefully built in | ||||||||||
1411 | // a way that doesn't rely on predecessor ordering. Rather than re-invent | ||||||||||
1412 | // that logic, we just re-walk the original blocks (and those of the child | ||||||||||
1413 | // loops) and filter them as we add them into the cloned loop. | ||||||||||
1414 | for (auto *BB : OrigL.blocks()) { | ||||||||||
1415 | auto *ClonedBB = cast_or_null<BasicBlock>(VMap.lookup(BB)); | ||||||||||
1416 | if (!ClonedBB || !BlocksInClonedLoop.count(ClonedBB)) | ||||||||||
1417 | continue; | ||||||||||
1418 | |||||||||||
1419 | // Directly add the blocks that are only in this loop. | ||||||||||
1420 | if (LI.getLoopFor(BB) == &OrigL) { | ||||||||||
1421 | ClonedL->addBasicBlockToLoop(ClonedBB, LI); | ||||||||||
1422 | continue; | ||||||||||
1423 | } | ||||||||||
1424 | |||||||||||
1425 | // We want to manually add it to this loop and parents. | ||||||||||
1426 | // Registering it with LoopInfo will happen when we clone the top | ||||||||||
1427 | // loop for this block. | ||||||||||
1428 | for (Loop *PL = ClonedL; PL; PL = PL->getParentLoop()) | ||||||||||
1429 | PL->addBlockEntry(ClonedBB); | ||||||||||
1430 | } | ||||||||||
1431 | |||||||||||
1432 | // Now add each child loop whose header remains within the cloned loop. All | ||||||||||
1433 | // of the blocks within the loop must satisfy the same constraints as the | ||||||||||
1434 | // header so once we pass the header checks we can just clone the entire | ||||||||||
1435 | // child loop nest. | ||||||||||
1436 | for (Loop *ChildL : OrigL) { | ||||||||||
1437 | auto *ClonedChildHeader = | ||||||||||
1438 | cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); | ||||||||||
1439 | if (!ClonedChildHeader || !BlocksInClonedLoop.count(ClonedChildHeader)) | ||||||||||
1440 | continue; | ||||||||||
1441 | |||||||||||
1442 | #ifndef NDEBUG1 | ||||||||||
1443 | // We should never have a cloned child loop header but fail to have | ||||||||||
1444 | // all of the blocks for that child loop. | ||||||||||
1445 | for (auto *ChildLoopBB : ChildL->blocks()) | ||||||||||
1446 | assert(BlocksInClonedLoop.count(((void)0) | ||||||||||
1447 | cast<BasicBlock>(VMap.lookup(ChildLoopBB))) &&((void)0) | ||||||||||
1448 | "Child cloned loop has a header within the cloned outer "((void)0) | ||||||||||
1449 | "loop but not all of its blocks!")((void)0); | ||||||||||
1450 | #endif | ||||||||||
1451 | |||||||||||
1452 | cloneLoopNest(*ChildL, ClonedL, VMap, LI); | ||||||||||
1453 | } | ||||||||||
1454 | } | ||||||||||
1455 | |||||||||||
1456 | // Now that we've handled all the components of the original loop that were | ||||||||||
1457 | // cloned into a new loop, we still need to handle anything from the original | ||||||||||
1458 | // loop that wasn't in a cloned loop. | ||||||||||
1459 | |||||||||||
1460 | // Figure out what blocks are left to place within any loop nest containing | ||||||||||
1461 | // the unswitched loop. If we never formed a loop, the cloned PH is one of | ||||||||||
1462 | // them. | ||||||||||
1463 | SmallPtrSet<BasicBlock *, 16> UnloopedBlockSet; | ||||||||||
1464 | if (BlocksInClonedLoop.empty()) | ||||||||||
1465 | UnloopedBlockSet.insert(ClonedPH); | ||||||||||
1466 | for (auto *ClonedBB : ClonedLoopBlocks) | ||||||||||
1467 | if (!BlocksInClonedLoop.count(ClonedBB)) | ||||||||||
1468 | UnloopedBlockSet.insert(ClonedBB); | ||||||||||
1469 | |||||||||||
1470 | // Copy the cloned exits and sort them in ascending loop depth, we'll work | ||||||||||
1471 | // backwards across these to process them inside out. The order shouldn't | ||||||||||
1472 | // matter as we're just trying to build up the map from inside-out; we use | ||||||||||
1473 | // the map in a more stably ordered way below. | ||||||||||
1474 | auto OrderedClonedExitsInLoops = ClonedExitsInLoops; | ||||||||||
1475 | llvm::sort(OrderedClonedExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { | ||||||||||
1476 | return ExitLoopMap.lookup(LHS)->getLoopDepth() < | ||||||||||
1477 | ExitLoopMap.lookup(RHS)->getLoopDepth(); | ||||||||||
1478 | }); | ||||||||||
1479 | |||||||||||
1480 | // Populate the existing ExitLoopMap with everything reachable from each | ||||||||||
1481 | // exit, starting from the inner most exit. | ||||||||||
1482 | while (!UnloopedBlockSet.empty() && !OrderedClonedExitsInLoops.empty()) { | ||||||||||
1483 | assert(Worklist.empty() && "Didn't clear worklist!")((void)0); | ||||||||||
1484 | |||||||||||
1485 | BasicBlock *ExitBB = OrderedClonedExitsInLoops.pop_back_val(); | ||||||||||
1486 | Loop *ExitL = ExitLoopMap.lookup(ExitBB); | ||||||||||
1487 | |||||||||||
1488 | // Walk the CFG back until we hit the cloned PH adding everything reachable | ||||||||||
1489 | // and in the unlooped set to this exit block's loop. | ||||||||||
1490 | Worklist.push_back(ExitBB); | ||||||||||
1491 | do { | ||||||||||
1492 | BasicBlock *BB = Worklist.pop_back_val(); | ||||||||||
1493 | // We can stop recursing at the cloned preheader (if we get there). | ||||||||||
1494 | if (BB == ClonedPH) | ||||||||||
1495 | continue; | ||||||||||
1496 | |||||||||||
1497 | for (BasicBlock *PredBB : predecessors(BB)) { | ||||||||||
1498 | // If this pred has already been moved to our set or is part of some | ||||||||||
1499 | // (inner) loop, no update needed. | ||||||||||
1500 | if (!UnloopedBlockSet.erase(PredBB)) { | ||||||||||
1501 | assert(((void)0) | ||||||||||
1502 | (BlocksInClonedLoop.count(PredBB) || ExitLoopMap.count(PredBB)) &&((void)0) | ||||||||||
1503 | "Predecessor not mapped to a loop!")((void)0); | ||||||||||
1504 | continue; | ||||||||||
1505 | } | ||||||||||
1506 | |||||||||||
1507 | // We just insert into the loop set here. We'll add these blocks to the | ||||||||||
1508 | // exit loop after we build up the set in an order that doesn't rely on | ||||||||||
1509 | // predecessor order (which in turn relies on use list order). | ||||||||||
1510 | bool Inserted = ExitLoopMap.insert({PredBB, ExitL}).second; | ||||||||||
1511 | (void)Inserted; | ||||||||||
1512 | assert(Inserted && "Should only visit an unlooped block once!")((void)0); | ||||||||||
1513 | |||||||||||
1514 | // And recurse through to its predecessors. | ||||||||||
1515 | Worklist.push_back(PredBB); | ||||||||||
1516 | } | ||||||||||
1517 | } while (!Worklist.empty()); | ||||||||||
1518 | } | ||||||||||
1519 | |||||||||||
1520 | // Now that the ExitLoopMap gives as mapping for all the non-looping cloned | ||||||||||
1521 | // blocks to their outer loops, walk the cloned blocks and the cloned exits | ||||||||||
1522 | // in their original order adding them to the correct loop. | ||||||||||
1523 | |||||||||||
1524 | // We need a stable insertion order. We use the order of the original loop | ||||||||||
1525 | // order and map into the correct parent loop. | ||||||||||
1526 | for (auto *BB : llvm::concat<BasicBlock *const>( | ||||||||||
1527 | makeArrayRef(ClonedPH), ClonedLoopBlocks, ClonedExitsInLoops)) | ||||||||||
1528 | if (Loop *OuterL = ExitLoopMap.lookup(BB)) | ||||||||||
1529 | OuterL->addBasicBlockToLoop(BB, LI); | ||||||||||
1530 | |||||||||||
1531 | #ifndef NDEBUG1 | ||||||||||
1532 | for (auto &BBAndL : ExitLoopMap) { | ||||||||||
1533 | auto *BB = BBAndL.first; | ||||||||||
1534 | auto *OuterL = BBAndL.second; | ||||||||||
1535 | assert(LI.getLoopFor(BB) == OuterL &&((void)0) | ||||||||||
1536 | "Failed to put all blocks into outer loops!")((void)0); | ||||||||||
1537 | } | ||||||||||
1538 | #endif | ||||||||||
1539 | |||||||||||
1540 | // Now that all the blocks are placed into the correct containing loop in the | ||||||||||
1541 | // absence of child loops, find all the potentially cloned child loops and | ||||||||||
1542 | // clone them into whatever outer loop we placed their header into. | ||||||||||
1543 | for (Loop *ChildL : OrigL) { | ||||||||||
1544 | auto *ClonedChildHeader = | ||||||||||
1545 | cast_or_null<BasicBlock>(VMap.lookup(ChildL->getHeader())); | ||||||||||
1546 | if (!ClonedChildHeader || BlocksInClonedLoop.count(ClonedChildHeader)) | ||||||||||
1547 | continue; | ||||||||||
1548 | |||||||||||
1549 | #ifndef NDEBUG1 | ||||||||||
1550 | for (auto *ChildLoopBB : ChildL->blocks()) | ||||||||||
1551 | assert(VMap.count(ChildLoopBB) &&((void)0) | ||||||||||
1552 | "Cloned a child loop header but not all of that loops blocks!")((void)0); | ||||||||||
1553 | #endif | ||||||||||
1554 | |||||||||||
1555 | NonChildClonedLoops.push_back(cloneLoopNest( | ||||||||||
1556 | *ChildL, ExitLoopMap.lookup(ClonedChildHeader), VMap, LI)); | ||||||||||
1557 | } | ||||||||||
1558 | } | ||||||||||
1559 | |||||||||||
1560 | static void | ||||||||||
1561 | deleteDeadClonedBlocks(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, | ||||||||||
1562 | ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, | ||||||||||
1563 | DominatorTree &DT, MemorySSAUpdater *MSSAU) { | ||||||||||
1564 | // Find all the dead clones, and remove them from their successors. | ||||||||||
1565 | SmallVector<BasicBlock *, 16> DeadBlocks; | ||||||||||
1566 | for (BasicBlock *BB : llvm::concat<BasicBlock *const>(L.blocks(), ExitBlocks)) | ||||||||||
1567 | for (auto &VMap : VMaps) | ||||||||||
1568 | if (BasicBlock *ClonedBB = cast_or_null<BasicBlock>(VMap->lookup(BB))) | ||||||||||
1569 | if (!DT.isReachableFromEntry(ClonedBB)) { | ||||||||||
1570 | for (BasicBlock *SuccBB : successors(ClonedBB)) | ||||||||||
1571 | SuccBB->removePredecessor(ClonedBB); | ||||||||||
1572 | DeadBlocks.push_back(ClonedBB); | ||||||||||
1573 | } | ||||||||||
1574 | |||||||||||
1575 | // Remove all MemorySSA in the dead blocks | ||||||||||
1576 | if (MSSAU) { | ||||||||||
1577 | SmallSetVector<BasicBlock *, 8> DeadBlockSet(DeadBlocks.begin(), | ||||||||||
1578 | DeadBlocks.end()); | ||||||||||
1579 | MSSAU->removeBlocks(DeadBlockSet); | ||||||||||
1580 | } | ||||||||||
1581 | |||||||||||
1582 | // Drop any remaining references to break cycles. | ||||||||||
1583 | for (BasicBlock *BB : DeadBlocks) | ||||||||||
1584 | BB->dropAllReferences(); | ||||||||||
1585 | // Erase them from the IR. | ||||||||||
1586 | for (BasicBlock *BB : DeadBlocks) | ||||||||||
1587 | BB->eraseFromParent(); | ||||||||||
1588 | } | ||||||||||
1589 | |||||||||||
1590 | static void | ||||||||||
1591 | deleteDeadBlocksFromLoop(Loop &L, | ||||||||||
1592 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | ||||||||||
1593 | DominatorTree &DT, LoopInfo &LI, | ||||||||||
1594 | MemorySSAUpdater *MSSAU, | ||||||||||
1595 | function_ref<void(Loop &, StringRef)> DestroyLoopCB) { | ||||||||||
1596 | // Find all the dead blocks tied to this loop, and remove them from their | ||||||||||
1597 | // successors. | ||||||||||
1598 | SmallSetVector<BasicBlock *, 8> DeadBlockSet; | ||||||||||
1599 | |||||||||||
1600 | // Start with loop/exit blocks and get a transitive closure of reachable dead | ||||||||||
1601 | // blocks. | ||||||||||
1602 | SmallVector<BasicBlock *, 16> DeathCandidates(ExitBlocks.begin(), | ||||||||||
1603 | ExitBlocks.end()); | ||||||||||
1604 | DeathCandidates.append(L.blocks().begin(), L.blocks().end()); | ||||||||||
1605 | while (!DeathCandidates.empty()) { | ||||||||||
1606 | auto *BB = DeathCandidates.pop_back_val(); | ||||||||||
1607 | if (!DeadBlockSet.count(BB) && !DT.isReachableFromEntry(BB)) { | ||||||||||
1608 | for (BasicBlock *SuccBB : successors(BB)) { | ||||||||||
1609 | SuccBB->removePredecessor(BB); | ||||||||||
1610 | DeathCandidates.push_back(SuccBB); | ||||||||||
1611 | } | ||||||||||
1612 | DeadBlockSet.insert(BB); | ||||||||||
1613 | } | ||||||||||
1614 | } | ||||||||||
1615 | |||||||||||
1616 | // Remove all MemorySSA in the dead blocks | ||||||||||
1617 | if (MSSAU) | ||||||||||
1618 | MSSAU->removeBlocks(DeadBlockSet); | ||||||||||
1619 | |||||||||||
1620 | // Filter out the dead blocks from the exit blocks list so that it can be | ||||||||||
1621 | // used in the caller. | ||||||||||
1622 | llvm::erase_if(ExitBlocks, | ||||||||||
1623 | [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); | ||||||||||
1624 | |||||||||||
1625 | // Walk from this loop up through its parents removing all of the dead blocks. | ||||||||||
1626 | for (Loop *ParentL = &L; ParentL; ParentL = ParentL->getParentLoop()) { | ||||||||||
1627 | for (auto *BB : DeadBlockSet) | ||||||||||
1628 | ParentL->getBlocksSet().erase(BB); | ||||||||||
1629 | llvm::erase_if(ParentL->getBlocksVector(), | ||||||||||
1630 | [&](BasicBlock *BB) { return DeadBlockSet.count(BB); }); | ||||||||||
1631 | } | ||||||||||
1632 | |||||||||||
1633 | // Now delete the dead child loops. This raw delete will clear them | ||||||||||
1634 | // recursively. | ||||||||||
1635 | llvm::erase_if(L.getSubLoopsVector(), [&](Loop *ChildL) { | ||||||||||
1636 | if (!DeadBlockSet.count(ChildL->getHeader())) | ||||||||||
1637 | return false; | ||||||||||
1638 | |||||||||||
1639 | assert(llvm::all_of(ChildL->blocks(),((void)0) | ||||||||||
1640 | [&](BasicBlock *ChildBB) {((void)0) | ||||||||||
1641 | return DeadBlockSet.count(ChildBB);((void)0) | ||||||||||
1642 | }) &&((void)0) | ||||||||||
1643 | "If the child loop header is dead all blocks in the child loop must "((void)0) | ||||||||||
1644 | "be dead as well!")((void)0); | ||||||||||
1645 | DestroyLoopCB(*ChildL, ChildL->getName()); | ||||||||||
1646 | LI.destroy(ChildL); | ||||||||||
1647 | return true; | ||||||||||
1648 | }); | ||||||||||
1649 | |||||||||||
1650 | // Remove the loop mappings for the dead blocks and drop all the references | ||||||||||
1651 | // from these blocks to others to handle cyclic references as we start | ||||||||||
1652 | // deleting the blocks themselves. | ||||||||||
1653 | for (auto *BB : DeadBlockSet) { | ||||||||||
1654 | // Check that the dominator tree has already been updated. | ||||||||||
1655 | assert(!DT.getNode(BB) && "Should already have cleared domtree!")((void)0); | ||||||||||
1656 | LI.changeLoopFor(BB, nullptr); | ||||||||||
1657 | // Drop all uses of the instructions to make sure we won't have dangling | ||||||||||
1658 | // uses in other blocks. | ||||||||||
1659 | for (auto &I : *BB) | ||||||||||
1660 | if (!I.use_empty()) | ||||||||||
1661 | I.replaceAllUsesWith(UndefValue::get(I.getType())); | ||||||||||
1662 | BB->dropAllReferences(); | ||||||||||
1663 | } | ||||||||||
1664 | |||||||||||
1665 | // Actually delete the blocks now that they've been fully unhooked from the | ||||||||||
1666 | // IR. | ||||||||||
1667 | for (auto *BB : DeadBlockSet) | ||||||||||
1668 | BB->eraseFromParent(); | ||||||||||
1669 | } | ||||||||||
1670 | |||||||||||
1671 | /// Recompute the set of blocks in a loop after unswitching. | ||||||||||
1672 | /// | ||||||||||
1673 | /// This walks from the original headers predecessors to rebuild the loop. We | ||||||||||
1674 | /// take advantage of the fact that new blocks can't have been added, and so we | ||||||||||
1675 | /// filter by the original loop's blocks. This also handles potentially | ||||||||||
1676 | /// unreachable code that we don't want to explore but might be found examining | ||||||||||
1677 | /// the predecessors of the header. | ||||||||||
1678 | /// | ||||||||||
1679 | /// If the original loop is no longer a loop, this will return an empty set. If | ||||||||||
1680 | /// it remains a loop, all the blocks within it will be added to the set | ||||||||||
1681 | /// (including those blocks in inner loops). | ||||||||||
1682 | static SmallPtrSet<const BasicBlock *, 16> recomputeLoopBlockSet(Loop &L, | ||||||||||
1683 | LoopInfo &LI) { | ||||||||||
1684 | SmallPtrSet<const BasicBlock *, 16> LoopBlockSet; | ||||||||||
1685 | |||||||||||
1686 | auto *PH = L.getLoopPreheader(); | ||||||||||
1687 | auto *Header = L.getHeader(); | ||||||||||
1688 | |||||||||||
1689 | // A worklist to use while walking backwards from the header. | ||||||||||
1690 | SmallVector<BasicBlock *, 16> Worklist; | ||||||||||
1691 | |||||||||||
1692 | // First walk the predecessors of the header to find the backedges. This will | ||||||||||
1693 | // form the basis of our walk. | ||||||||||
1694 | for (auto *Pred : predecessors(Header)) { | ||||||||||
1695 | // Skip the preheader. | ||||||||||
1696 | if (Pred == PH) | ||||||||||
1697 | continue; | ||||||||||
1698 | |||||||||||
1699 | // Because the loop was in simplified form, the only non-loop predecessor | ||||||||||
1700 | // is the preheader. | ||||||||||
1701 | assert(L.contains(Pred) && "Found a predecessor of the loop header other "((void)0) | ||||||||||
1702 | "than the preheader that is not part of the "((void)0) | ||||||||||
1703 | "loop!")((void)0); | ||||||||||
1704 | |||||||||||
1705 | // Insert this block into the loop set and on the first visit and, if it | ||||||||||
1706 | // isn't the header we're currently walking, put it into the worklist to | ||||||||||
1707 | // recurse through. | ||||||||||
1708 | if (LoopBlockSet.insert(Pred).second && Pred != Header) | ||||||||||
1709 | Worklist.push_back(Pred); | ||||||||||
1710 | } | ||||||||||
1711 | |||||||||||
1712 | // If no backedges were found, we're done. | ||||||||||
1713 | if (LoopBlockSet.empty()) | ||||||||||
1714 | return LoopBlockSet; | ||||||||||
1715 | |||||||||||
1716 | // We found backedges, recurse through them to identify the loop blocks. | ||||||||||
1717 | while (!Worklist.empty()) { | ||||||||||
1718 | BasicBlock *BB = Worklist.pop_back_val(); | ||||||||||
1719 | assert(LoopBlockSet.count(BB) && "Didn't put block into the loop set!")((void)0); | ||||||||||
1720 | |||||||||||
1721 | // No need to walk past the header. | ||||||||||
1722 | if (BB == Header) | ||||||||||
1723 | continue; | ||||||||||
1724 | |||||||||||
1725 | // Because we know the inner loop structure remains valid we can use the | ||||||||||
1726 | // loop structure to jump immediately across the entire nested loop. | ||||||||||
1727 | // Further, because it is in loop simplified form, we can directly jump | ||||||||||
1728 | // to its preheader afterward. | ||||||||||
1729 | if (Loop *InnerL = LI.getLoopFor(BB)) | ||||||||||
1730 | if (InnerL != &L) { | ||||||||||
1731 | assert(L.contains(InnerL) &&((void)0) | ||||||||||
1732 | "Should not reach a loop *outside* this loop!")((void)0); | ||||||||||
1733 | // The preheader is the only possible predecessor of the loop so | ||||||||||
1734 | // insert it into the set and check whether it was already handled. | ||||||||||
1735 | auto *InnerPH = InnerL->getLoopPreheader(); | ||||||||||
1736 | assert(L.contains(InnerPH) && "Cannot contain an inner loop block "((void)0) | ||||||||||
1737 | "but not contain the inner loop "((void)0) | ||||||||||
1738 | "preheader!")((void)0); | ||||||||||
1739 | if (!LoopBlockSet.insert(InnerPH).second) | ||||||||||
1740 | // The only way to reach the preheader is through the loop body | ||||||||||
1741 | // itself so if it has been visited the loop is already handled. | ||||||||||
1742 | continue; | ||||||||||
1743 | |||||||||||
1744 | // Insert all of the blocks (other than those already present) into | ||||||||||
1745 | // the loop set. We expect at least the block that led us to find the | ||||||||||
1746 | // inner loop to be in the block set, but we may also have other loop | ||||||||||
1747 | // blocks if they were already enqueued as predecessors of some other | ||||||||||
1748 | // outer loop block. | ||||||||||
1749 | for (auto *InnerBB : InnerL->blocks()) { | ||||||||||
1750 | if (InnerBB == BB) { | ||||||||||
1751 | assert(LoopBlockSet.count(InnerBB) &&((void)0) | ||||||||||
1752 | "Block should already be in the set!")((void)0); | ||||||||||
1753 | continue; | ||||||||||
1754 | } | ||||||||||
1755 | |||||||||||
1756 | LoopBlockSet.insert(InnerBB); | ||||||||||
1757 | } | ||||||||||
1758 | |||||||||||
1759 | // Add the preheader to the worklist so we will continue past the | ||||||||||
1760 | // loop body. | ||||||||||
1761 | Worklist.push_back(InnerPH); | ||||||||||
1762 | continue; | ||||||||||
1763 | } | ||||||||||
1764 | |||||||||||
1765 | // Insert any predecessors that were in the original loop into the new | ||||||||||
1766 | // set, and if the insert is successful, add them to the worklist. | ||||||||||
1767 | for (auto *Pred : predecessors(BB)) | ||||||||||
1768 | if (L.contains(Pred) && LoopBlockSet.insert(Pred).second) | ||||||||||
1769 | Worklist.push_back(Pred); | ||||||||||
1770 | } | ||||||||||
1771 | |||||||||||
1772 | assert(LoopBlockSet.count(Header) && "Cannot fail to add the header!")((void)0); | ||||||||||
1773 | |||||||||||
1774 | // We've found all the blocks participating in the loop, return our completed | ||||||||||
1775 | // set. | ||||||||||
1776 | return LoopBlockSet; | ||||||||||
1777 | } | ||||||||||
1778 | |||||||||||
1779 | /// Rebuild a loop after unswitching removes some subset of blocks and edges. | ||||||||||
1780 | /// | ||||||||||
1781 | /// The removal may have removed some child loops entirely but cannot have | ||||||||||
1782 | /// disturbed any remaining child loops. However, they may need to be hoisted | ||||||||||
1783 | /// to the parent loop (or to be top-level loops). The original loop may be | ||||||||||
1784 | /// completely removed. | ||||||||||
1785 | /// | ||||||||||
1786 | /// The sibling loops resulting from this update are returned. If the original | ||||||||||
1787 | /// loop remains a valid loop, it will be the first entry in this list with all | ||||||||||
1788 | /// of the newly sibling loops following it. | ||||||||||
1789 | /// | ||||||||||
1790 | /// Returns true if the loop remains a loop after unswitching, and false if it | ||||||||||
1791 | /// is no longer a loop after unswitching (and should not continue to be | ||||||||||
1792 | /// referenced). | ||||||||||
1793 | static bool rebuildLoopAfterUnswitch(Loop &L, ArrayRef<BasicBlock *> ExitBlocks, | ||||||||||
1794 | LoopInfo &LI, | ||||||||||
1795 | SmallVectorImpl<Loop *> &HoistedLoops) { | ||||||||||
1796 | auto *PH = L.getLoopPreheader(); | ||||||||||
1797 | |||||||||||
1798 | // Compute the actual parent loop from the exit blocks. Because we may have | ||||||||||
1799 | // pruned some exits the loop may be different from the original parent. | ||||||||||
1800 | Loop *ParentL = nullptr; | ||||||||||
1801 | SmallVector<Loop *, 4> ExitLoops; | ||||||||||
1802 | SmallVector<BasicBlock *, 4> ExitsInLoops; | ||||||||||
1803 | ExitsInLoops.reserve(ExitBlocks.size()); | ||||||||||
1804 | for (auto *ExitBB : ExitBlocks) | ||||||||||
1805 | if (Loop *ExitL = LI.getLoopFor(ExitBB)) { | ||||||||||
1806 | ExitLoops.push_back(ExitL); | ||||||||||
1807 | ExitsInLoops.push_back(ExitBB); | ||||||||||
1808 | if (!ParentL || (ParentL != ExitL && ParentL->contains(ExitL))) | ||||||||||
1809 | ParentL = ExitL; | ||||||||||
1810 | } | ||||||||||
1811 | |||||||||||
1812 | // Recompute the blocks participating in this loop. This may be empty if it | ||||||||||
1813 | // is no longer a loop. | ||||||||||
1814 | auto LoopBlockSet = recomputeLoopBlockSet(L, LI); | ||||||||||
1815 | |||||||||||
1816 | // If we still have a loop, we need to re-set the loop's parent as the exit | ||||||||||
1817 | // block set changing may have moved it within the loop nest. Note that this | ||||||||||
1818 | // can only happen when this loop has a parent as it can only hoist the loop | ||||||||||
1819 | // *up* the nest. | ||||||||||
1820 | if (!LoopBlockSet.empty() && L.getParentLoop() != ParentL) { | ||||||||||
1821 | // Remove this loop's (original) blocks from all of the intervening loops. | ||||||||||
1822 | for (Loop *IL = L.getParentLoop(); IL != ParentL; | ||||||||||
1823 | IL = IL->getParentLoop()) { | ||||||||||
1824 | IL->getBlocksSet().erase(PH); | ||||||||||
1825 | for (auto *BB : L.blocks()) | ||||||||||
1826 | IL->getBlocksSet().erase(BB); | ||||||||||
1827 | llvm::erase_if(IL->getBlocksVector(), [&](BasicBlock *BB) { | ||||||||||
1828 | return BB == PH || L.contains(BB); | ||||||||||
1829 | }); | ||||||||||
1830 | } | ||||||||||
1831 | |||||||||||
1832 | LI.changeLoopFor(PH, ParentL); | ||||||||||
1833 | L.getParentLoop()->removeChildLoop(&L); | ||||||||||
1834 | if (ParentL) | ||||||||||
1835 | ParentL->addChildLoop(&L); | ||||||||||
1836 | else | ||||||||||
1837 | LI.addTopLevelLoop(&L); | ||||||||||
1838 | } | ||||||||||
1839 | |||||||||||
1840 | // Now we update all the blocks which are no longer within the loop. | ||||||||||
1841 | auto &Blocks = L.getBlocksVector(); | ||||||||||
1842 | auto BlocksSplitI = | ||||||||||
1843 | LoopBlockSet.empty() | ||||||||||
1844 | ? Blocks.begin() | ||||||||||
1845 | : std::stable_partition( | ||||||||||
1846 | Blocks.begin(), Blocks.end(), | ||||||||||
1847 | [&](BasicBlock *BB) { return LoopBlockSet.count(BB); }); | ||||||||||
1848 | |||||||||||
1849 | // Before we erase the list of unlooped blocks, build a set of them. | ||||||||||
1850 | SmallPtrSet<BasicBlock *, 16> UnloopedBlocks(BlocksSplitI, Blocks.end()); | ||||||||||
1851 | if (LoopBlockSet.empty()) | ||||||||||
1852 | UnloopedBlocks.insert(PH); | ||||||||||
1853 | |||||||||||
1854 | // Now erase these blocks from the loop. | ||||||||||
1855 | for (auto *BB : make_range(BlocksSplitI, Blocks.end())) | ||||||||||
1856 | L.getBlocksSet().erase(BB); | ||||||||||
1857 | Blocks.erase(BlocksSplitI, Blocks.end()); | ||||||||||
1858 | |||||||||||
1859 | // Sort the exits in ascending loop depth, we'll work backwards across these | ||||||||||
1860 | // to process them inside out. | ||||||||||
1861 | llvm::stable_sort(ExitsInLoops, [&](BasicBlock *LHS, BasicBlock *RHS) { | ||||||||||
1862 | return LI.getLoopDepth(LHS) < LI.getLoopDepth(RHS); | ||||||||||
1863 | }); | ||||||||||
1864 | |||||||||||
1865 | // We'll build up a set for each exit loop. | ||||||||||
1866 | SmallPtrSet<BasicBlock *, 16> NewExitLoopBlocks; | ||||||||||
1867 | Loop *PrevExitL = L.getParentLoop(); // The deepest possible exit loop. | ||||||||||
1868 | |||||||||||
1869 | auto RemoveUnloopedBlocksFromLoop = | ||||||||||
1870 | [](Loop &L, SmallPtrSetImpl<BasicBlock *> &UnloopedBlocks) { | ||||||||||
1871 | for (auto *BB : UnloopedBlocks) | ||||||||||
1872 | L.getBlocksSet().erase(BB); | ||||||||||
1873 | llvm::erase_if(L.getBlocksVector(), [&](BasicBlock *BB) { | ||||||||||
1874 | return UnloopedBlocks.count(BB); | ||||||||||
1875 | }); | ||||||||||
1876 | }; | ||||||||||
1877 | |||||||||||
1878 | SmallVector<BasicBlock *, 16> Worklist; | ||||||||||
1879 | while (!UnloopedBlocks.empty() && !ExitsInLoops.empty()) { | ||||||||||
1880 | assert(Worklist.empty() && "Didn't clear worklist!")((void)0); | ||||||||||
1881 | assert(NewExitLoopBlocks.empty() && "Didn't clear loop set!")((void)0); | ||||||||||
1882 | |||||||||||
1883 | // Grab the next exit block, in decreasing loop depth order. | ||||||||||
1884 | BasicBlock *ExitBB = ExitsInLoops.pop_back_val(); | ||||||||||
1885 | Loop &ExitL = *LI.getLoopFor(ExitBB); | ||||||||||
1886 | assert(ExitL.contains(&L) && "Exit loop must contain the inner loop!")((void)0); | ||||||||||
1887 | |||||||||||
1888 | // Erase all of the unlooped blocks from the loops between the previous | ||||||||||
1889 | // exit loop and this exit loop. This works because the ExitInLoops list is | ||||||||||
1890 | // sorted in increasing order of loop depth and thus we visit loops in | ||||||||||
1891 | // decreasing order of loop depth. | ||||||||||
1892 | for (; PrevExitL != &ExitL; PrevExitL = PrevExitL->getParentLoop()) | ||||||||||
1893 | RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); | ||||||||||
1894 | |||||||||||
1895 | // Walk the CFG back until we hit the cloned PH adding everything reachable | ||||||||||
1896 | // and in the unlooped set to this exit block's loop. | ||||||||||
1897 | Worklist.push_back(ExitBB); | ||||||||||
1898 | do { | ||||||||||
1899 | BasicBlock *BB = Worklist.pop_back_val(); | ||||||||||
1900 | // We can stop recursing at the cloned preheader (if we get there). | ||||||||||
1901 | if (BB == PH) | ||||||||||
1902 | continue; | ||||||||||
1903 | |||||||||||
1904 | for (BasicBlock *PredBB : predecessors(BB)) { | ||||||||||
1905 | // If this pred has already been moved to our set or is part of some | ||||||||||
1906 | // (inner) loop, no update needed. | ||||||||||
1907 | if (!UnloopedBlocks.erase(PredBB)) { | ||||||||||
1908 | assert((NewExitLoopBlocks.count(PredBB) ||((void)0) | ||||||||||
1909 | ExitL.contains(LI.getLoopFor(PredBB))) &&((void)0) | ||||||||||
1910 | "Predecessor not in a nested loop (or already visited)!")((void)0); | ||||||||||
1911 | continue; | ||||||||||
1912 | } | ||||||||||
1913 | |||||||||||
1914 | // We just insert into the loop set here. We'll add these blocks to the | ||||||||||
1915 | // exit loop after we build up the set in a deterministic order rather | ||||||||||
1916 | // than the predecessor-influenced visit order. | ||||||||||
1917 | bool Inserted = NewExitLoopBlocks.insert(PredBB).second; | ||||||||||
1918 | (void)Inserted; | ||||||||||
1919 | assert(Inserted && "Should only visit an unlooped block once!")((void)0); | ||||||||||
1920 | |||||||||||
1921 | // And recurse through to its predecessors. | ||||||||||
1922 | Worklist.push_back(PredBB); | ||||||||||
1923 | } | ||||||||||
1924 | } while (!Worklist.empty()); | ||||||||||
1925 | |||||||||||
1926 | // If blocks in this exit loop were directly part of the original loop (as | ||||||||||
1927 | // opposed to a child loop) update the map to point to this exit loop. This | ||||||||||
1928 | // just updates a map and so the fact that the order is unstable is fine. | ||||||||||
1929 | for (auto *BB : NewExitLoopBlocks) | ||||||||||
1930 | if (Loop *BBL = LI.getLoopFor(BB)) | ||||||||||
1931 | if (BBL == &L || !L.contains(BBL)) | ||||||||||
1932 | LI.changeLoopFor(BB, &ExitL); | ||||||||||
1933 | |||||||||||
1934 | // We will remove the remaining unlooped blocks from this loop in the next | ||||||||||
1935 | // iteration or below. | ||||||||||
1936 | NewExitLoopBlocks.clear(); | ||||||||||
1937 | } | ||||||||||
1938 | |||||||||||
1939 | // Any remaining unlooped blocks are no longer part of any loop unless they | ||||||||||
1940 | // are part of some child loop. | ||||||||||
1941 | for (; PrevExitL; PrevExitL = PrevExitL->getParentLoop()) | ||||||||||
1942 | RemoveUnloopedBlocksFromLoop(*PrevExitL, UnloopedBlocks); | ||||||||||
1943 | for (auto *BB : UnloopedBlocks) | ||||||||||
1944 | if (Loop *BBL = LI.getLoopFor(BB)) | ||||||||||
1945 | if (BBL == &L || !L.contains(BBL)) | ||||||||||
1946 | LI.changeLoopFor(BB, nullptr); | ||||||||||
1947 | |||||||||||
1948 | // Sink all the child loops whose headers are no longer in the loop set to | ||||||||||
1949 | // the parent (or to be top level loops). We reach into the loop and directly | ||||||||||
1950 | // update its subloop vector to make this batch update efficient. | ||||||||||
1951 | auto &SubLoops = L.getSubLoopsVector(); | ||||||||||
1952 | auto SubLoopsSplitI = | ||||||||||
1953 | LoopBlockSet.empty() | ||||||||||
1954 | ? SubLoops.begin() | ||||||||||
1955 | : std::stable_partition( | ||||||||||
1956 | SubLoops.begin(), SubLoops.end(), [&](Loop *SubL) { | ||||||||||
1957 | return LoopBlockSet.count(SubL->getHeader()); | ||||||||||
1958 | }); | ||||||||||
1959 | for (auto *HoistedL : make_range(SubLoopsSplitI, SubLoops.end())) { | ||||||||||
1960 | HoistedLoops.push_back(HoistedL); | ||||||||||
1961 | HoistedL->setParentLoop(nullptr); | ||||||||||
1962 | |||||||||||
1963 | // To compute the new parent of this hoisted loop we look at where we | ||||||||||
1964 | // placed the preheader above. We can't lookup the header itself because we | ||||||||||
1965 | // retained the mapping from the header to the hoisted loop. But the | ||||||||||
1966 | // preheader and header should have the exact same new parent computed | ||||||||||
1967 | // based on the set of exit blocks from the original loop as the preheader | ||||||||||
1968 | // is a predecessor of the header and so reached in the reverse walk. And | ||||||||||
1969 | // because the loops were all in simplified form the preheader of the | ||||||||||
1970 | // hoisted loop can't be part of some *other* loop. | ||||||||||
1971 | if (auto *NewParentL = LI.getLoopFor(HoistedL->getLoopPreheader())) | ||||||||||
1972 | NewParentL->addChildLoop(HoistedL); | ||||||||||
1973 | else | ||||||||||
1974 | LI.addTopLevelLoop(HoistedL); | ||||||||||
1975 | } | ||||||||||
1976 | SubLoops.erase(SubLoopsSplitI, SubLoops.end()); | ||||||||||
1977 | |||||||||||
1978 | // Actually delete the loop if nothing remained within it. | ||||||||||
1979 | if (Blocks.empty()) { | ||||||||||
1980 | assert(SubLoops.empty() &&((void)0) | ||||||||||
1981 | "Failed to remove all subloops from the original loop!")((void)0); | ||||||||||
1982 | if (Loop *ParentL = L.getParentLoop()) | ||||||||||
1983 | ParentL->removeChildLoop(llvm::find(*ParentL, &L)); | ||||||||||
1984 | else | ||||||||||
1985 | LI.removeLoop(llvm::find(LI, &L)); | ||||||||||
1986 | // markLoopAsDeleted for L should be triggered by the caller (it is typically | ||||||||||
1987 | // done by using the UnswitchCB callback). | ||||||||||
1988 | LI.destroy(&L); | ||||||||||
1989 | return false; | ||||||||||
1990 | } | ||||||||||
1991 | |||||||||||
1992 | return true; | ||||||||||
1993 | } | ||||||||||
1994 | |||||||||||
1995 | /// Helper to visit a dominator subtree, invoking a callable on each node. | ||||||||||
1996 | /// | ||||||||||
1997 | /// Returning false at any point will stop walking past that node of the tree. | ||||||||||
1998 | template <typename CallableT> | ||||||||||
1999 | void visitDomSubTree(DominatorTree &DT, BasicBlock *BB, CallableT Callable) { | ||||||||||
2000 | SmallVector<DomTreeNode *, 4> DomWorklist; | ||||||||||
2001 | DomWorklist.push_back(DT[BB]); | ||||||||||
2002 | #ifndef NDEBUG1 | ||||||||||
2003 | SmallPtrSet<DomTreeNode *, 4> Visited; | ||||||||||
2004 | Visited.insert(DT[BB]); | ||||||||||
2005 | #endif | ||||||||||
2006 | do { | ||||||||||
2007 | DomTreeNode *N = DomWorklist.pop_back_val(); | ||||||||||
2008 | |||||||||||
2009 | // Visit this node. | ||||||||||
2010 | if (!Callable(N->getBlock())) | ||||||||||
2011 | continue; | ||||||||||
2012 | |||||||||||
2013 | // Accumulate the child nodes. | ||||||||||
2014 | for (DomTreeNode *ChildN : *N) { | ||||||||||
2015 | assert(Visited.insert(ChildN).second &&((void)0) | ||||||||||
2016 | "Cannot visit a node twice when walking a tree!")((void)0); | ||||||||||
2017 | DomWorklist.push_back(ChildN); | ||||||||||
2018 | } | ||||||||||
2019 | } while (!DomWorklist.empty()); | ||||||||||
2020 | } | ||||||||||
2021 | |||||||||||
2022 | static void unswitchNontrivialInvariants( | ||||||||||
2023 | Loop &L, Instruction &TI, ArrayRef<Value *> Invariants, | ||||||||||
2024 | SmallVectorImpl<BasicBlock *> &ExitBlocks, IVConditionInfo &PartialIVInfo, | ||||||||||
2025 | DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, | ||||||||||
2026 | function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB, | ||||||||||
2027 | ScalarEvolution *SE, MemorySSAUpdater *MSSAU, | ||||||||||
2028 | function_ref<void(Loop &, StringRef)> DestroyLoopCB) { | ||||||||||
2029 | auto *ParentBB = TI.getParent(); | ||||||||||
2030 | BranchInst *BI = dyn_cast<BranchInst>(&TI); | ||||||||||
2031 | SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI); | ||||||||||
2032 | |||||||||||
2033 | // We can only unswitch switches, conditional branches with an invariant | ||||||||||
2034 | // condition, or combining invariant conditions with an instruction or | ||||||||||
2035 | // partially invariant instructions. | ||||||||||
2036 | assert((SI || (BI && BI->isConditional())) &&((void)0) | ||||||||||
2037 | "Can only unswitch switches and conditional branch!")((void)0); | ||||||||||
2038 | bool PartiallyInvariant = !PartialIVInfo.InstToDuplicate.empty(); | ||||||||||
2039 | bool FullUnswitch = | ||||||||||
2040 | SI || (BI->getCondition() == Invariants[0] && !PartiallyInvariant); | ||||||||||
2041 | if (FullUnswitch) | ||||||||||
2042 | assert(Invariants.size() == 1 &&((void)0) | ||||||||||
2043 | "Cannot have other invariants with full unswitching!")((void)0); | ||||||||||
2044 | else | ||||||||||
2045 | assert(isa<Instruction>(BI->getCondition()) &&((void)0) | ||||||||||
2046 | "Partial unswitching requires an instruction as the condition!")((void)0); | ||||||||||
2047 | |||||||||||
2048 | if (MSSAU && VerifyMemorySSA) | ||||||||||
2049 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2050 | |||||||||||
2051 | // Constant and BBs tracking the cloned and continuing successor. When we are | ||||||||||
2052 | // unswitching the entire condition, this can just be trivially chosen to | ||||||||||
2053 | // unswitch towards `true`. However, when we are unswitching a set of | ||||||||||
2054 | // invariants combined with `and` or `or` or partially invariant instructions, | ||||||||||
2055 | // the combining operation determines the best direction to unswitch: we want | ||||||||||
2056 | // to unswitch the direction that will collapse the branch. | ||||||||||
2057 | bool Direction = true; | ||||||||||
2058 | int ClonedSucc = 0; | ||||||||||
2059 | if (!FullUnswitch) { | ||||||||||
2060 | Value *Cond = BI->getCondition(); | ||||||||||
2061 | (void)Cond; | ||||||||||
2062 | assert(((match(Cond, m_LogicalAnd()) ^ match(Cond, m_LogicalOr())) ||((void)0) | ||||||||||
2063 | PartiallyInvariant) &&((void)0) | ||||||||||
2064 | "Only `or`, `and`, an `select`, partially invariant instructions "((void)0) | ||||||||||
2065 | "can combine invariants being unswitched.")((void)0); | ||||||||||
2066 | if (!match(BI->getCondition(), m_LogicalOr())) { | ||||||||||
2067 | if (match(BI->getCondition(), m_LogicalAnd()) || | ||||||||||
2068 | (PartiallyInvariant && !PartialIVInfo.KnownValue->isOneValue())) { | ||||||||||
2069 | Direction = false; | ||||||||||
2070 | ClonedSucc = 1; | ||||||||||
2071 | } | ||||||||||
2072 | } | ||||||||||
2073 | } | ||||||||||
2074 | |||||||||||
2075 | BasicBlock *RetainedSuccBB = | ||||||||||
2076 | BI ? BI->getSuccessor(1 - ClonedSucc) : SI->getDefaultDest(); | ||||||||||
2077 | SmallSetVector<BasicBlock *, 4> UnswitchedSuccBBs; | ||||||||||
2078 | if (BI) | ||||||||||
2079 | UnswitchedSuccBBs.insert(BI->getSuccessor(ClonedSucc)); | ||||||||||
2080 | else | ||||||||||
2081 | for (auto Case : SI->cases()) | ||||||||||
2082 | if (Case.getCaseSuccessor() != RetainedSuccBB) | ||||||||||
2083 | UnswitchedSuccBBs.insert(Case.getCaseSuccessor()); | ||||||||||
2084 | |||||||||||
2085 | assert(!UnswitchedSuccBBs.count(RetainedSuccBB) &&((void)0) | ||||||||||
2086 | "Should not unswitch the same successor we are retaining!")((void)0); | ||||||||||
2087 | |||||||||||
2088 | // The branch should be in this exact loop. Any inner loop's invariant branch | ||||||||||
2089 | // should be handled by unswitching that inner loop. The caller of this | ||||||||||
2090 | // routine should filter out any candidates that remain (but were skipped for | ||||||||||
2091 | // whatever reason). | ||||||||||
2092 | assert(LI.getLoopFor(ParentBB) == &L && "Branch in an inner loop!")((void)0); | ||||||||||
2093 | |||||||||||
2094 | // Compute the parent loop now before we start hacking on things. | ||||||||||
2095 | Loop *ParentL = L.getParentLoop(); | ||||||||||
2096 | // Get blocks in RPO order for MSSA update, before changing the CFG. | ||||||||||
2097 | LoopBlocksRPO LBRPO(&L); | ||||||||||
2098 | if (MSSAU) | ||||||||||
2099 | LBRPO.perform(&LI); | ||||||||||
2100 | |||||||||||
2101 | // Compute the outer-most loop containing one of our exit blocks. This is the | ||||||||||
2102 | // furthest up our loopnest which can be mutated, which we will use below to | ||||||||||
2103 | // update things. | ||||||||||
2104 | Loop *OuterExitL = &L; | ||||||||||
2105 | for (auto *ExitBB : ExitBlocks) { | ||||||||||
2106 | Loop *NewOuterExitL = LI.getLoopFor(ExitBB); | ||||||||||
2107 | if (!NewOuterExitL) { | ||||||||||
2108 | // We exited the entire nest with this block, so we're done. | ||||||||||
2109 | OuterExitL = nullptr; | ||||||||||
2110 | break; | ||||||||||
2111 | } | ||||||||||
2112 | if (NewOuterExitL != OuterExitL && NewOuterExitL->contains(OuterExitL)) | ||||||||||
2113 | OuterExitL = NewOuterExitL; | ||||||||||
2114 | } | ||||||||||
2115 | |||||||||||
2116 | // At this point, we're definitely going to unswitch something so invalidate | ||||||||||
2117 | // any cached information in ScalarEvolution for the outer most loop | ||||||||||
2118 | // containing an exit block and all nested loops. | ||||||||||
2119 | if (SE) { | ||||||||||
2120 | if (OuterExitL) | ||||||||||
2121 | SE->forgetLoop(OuterExitL); | ||||||||||
2122 | else | ||||||||||
2123 | SE->forgetTopmostLoop(&L); | ||||||||||
2124 | } | ||||||||||
2125 | |||||||||||
2126 | // If the edge from this terminator to a successor dominates that successor, | ||||||||||
2127 | // store a map from each block in its dominator subtree to it. This lets us | ||||||||||
2128 | // tell when cloning for a particular successor if a block is dominated by | ||||||||||
2129 | // some *other* successor with a single data structure. We use this to | ||||||||||
2130 | // significantly reduce cloning. | ||||||||||
2131 | SmallDenseMap<BasicBlock *, BasicBlock *, 16> DominatingSucc; | ||||||||||
2132 | for (auto *SuccBB : llvm::concat<BasicBlock *const>( | ||||||||||
2133 | makeArrayRef(RetainedSuccBB), UnswitchedSuccBBs)) | ||||||||||
2134 | if (SuccBB->getUniquePredecessor() || | ||||||||||
2135 | llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { | ||||||||||
2136 | return PredBB == ParentBB || DT.dominates(SuccBB, PredBB); | ||||||||||
2137 | })) | ||||||||||
2138 | visitDomSubTree(DT, SuccBB, [&](BasicBlock *BB) { | ||||||||||
2139 | DominatingSucc[BB] = SuccBB; | ||||||||||
2140 | return true; | ||||||||||
2141 | }); | ||||||||||
2142 | |||||||||||
2143 | // Split the preheader, so that we know that there is a safe place to insert | ||||||||||
2144 | // the conditional branch. We will change the preheader to have a conditional | ||||||||||
2145 | // branch on LoopCond. The original preheader will become the split point | ||||||||||
2146 | // between the unswitched versions, and we will have a new preheader for the | ||||||||||
2147 | // original loop. | ||||||||||
2148 | BasicBlock *SplitBB = L.getLoopPreheader(); | ||||||||||
2149 | BasicBlock *LoopPH = SplitEdge(SplitBB, L.getHeader(), &DT, &LI, MSSAU); | ||||||||||
2150 | |||||||||||
2151 | // Keep track of the dominator tree updates needed. | ||||||||||
2152 | SmallVector<DominatorTree::UpdateType, 4> DTUpdates; | ||||||||||
2153 | |||||||||||
2154 | // Clone the loop for each unswitched successor. | ||||||||||
2155 | SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; | ||||||||||
2156 | VMaps.reserve(UnswitchedSuccBBs.size()); | ||||||||||
2157 | SmallDenseMap<BasicBlock *, BasicBlock *, 4> ClonedPHs; | ||||||||||
2158 | for (auto *SuccBB : UnswitchedSuccBBs) { | ||||||||||
2159 | VMaps.emplace_back(new ValueToValueMapTy()); | ||||||||||
2160 | ClonedPHs[SuccBB] = buildClonedLoopBlocks( | ||||||||||
2161 | L, LoopPH, SplitBB, ExitBlocks, ParentBB, SuccBB, RetainedSuccBB, | ||||||||||
2162 | DominatingSucc, *VMaps.back(), DTUpdates, AC, DT, LI, MSSAU); | ||||||||||
2163 | } | ||||||||||
2164 | |||||||||||
2165 | // Drop metadata if we may break its semantics by moving this instr into the | ||||||||||
2166 | // split block. | ||||||||||
2167 | if (TI.getMetadata(LLVMContext::MD_make_implicit)) { | ||||||||||
2168 | if (DropNonTrivialImplicitNullChecks) | ||||||||||
2169 | // Do not spend time trying to understand if we can keep it, just drop it | ||||||||||
2170 | // to save compile time. | ||||||||||
2171 | TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); | ||||||||||
2172 | else { | ||||||||||
2173 | // It is only legal to preserve make.implicit metadata if we are | ||||||||||
2174 | // guaranteed no reach implicit null check after following this branch. | ||||||||||
2175 | ICFLoopSafetyInfo SafetyInfo; | ||||||||||
2176 | SafetyInfo.computeLoopSafetyInfo(&L); | ||||||||||
2177 | if (!SafetyInfo.isGuaranteedToExecute(TI, &DT, &L)) | ||||||||||
2178 | TI.setMetadata(LLVMContext::MD_make_implicit, nullptr); | ||||||||||
2179 | } | ||||||||||
2180 | } | ||||||||||
2181 | |||||||||||
2182 | // The stitching of the branched code back together depends on whether we're | ||||||||||
2183 | // doing full unswitching or not with the exception that we always want to | ||||||||||
2184 | // nuke the initial terminator placed in the split block. | ||||||||||
2185 | SplitBB->getTerminator()->eraseFromParent(); | ||||||||||
2186 | if (FullUnswitch) { | ||||||||||
2187 | // Splice the terminator from the original loop and rewrite its | ||||||||||
2188 | // successors. | ||||||||||
2189 | SplitBB->getInstList().splice(SplitBB->end(), ParentBB->getInstList(), TI); | ||||||||||
2190 | |||||||||||
2191 | // Keep a clone of the terminator for MSSA updates. | ||||||||||
2192 | Instruction *NewTI = TI.clone(); | ||||||||||
2193 | ParentBB->getInstList().push_back(NewTI); | ||||||||||
2194 | |||||||||||
2195 | // First wire up the moved terminator to the preheaders. | ||||||||||
2196 | if (BI) { | ||||||||||
2197 | BasicBlock *ClonedPH = ClonedPHs.begin()->second; | ||||||||||
2198 | BI->setSuccessor(ClonedSucc, ClonedPH); | ||||||||||
2199 | BI->setSuccessor(1 - ClonedSucc, LoopPH); | ||||||||||
2200 | DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); | ||||||||||
2201 | } else { | ||||||||||
2202 | assert(SI && "Must either be a branch or switch!")((void)0); | ||||||||||
2203 | |||||||||||
2204 | // Walk the cases and directly update their successors. | ||||||||||
2205 | assert(SI->getDefaultDest() == RetainedSuccBB &&((void)0) | ||||||||||
2206 | "Not retaining default successor!")((void)0); | ||||||||||
2207 | SI->setDefaultDest(LoopPH); | ||||||||||
2208 | for (auto &Case : SI->cases()) | ||||||||||
2209 | if (Case.getCaseSuccessor() == RetainedSuccBB) | ||||||||||
2210 | Case.setSuccessor(LoopPH); | ||||||||||
2211 | else | ||||||||||
2212 | Case.setSuccessor(ClonedPHs.find(Case.getCaseSuccessor())->second); | ||||||||||
2213 | |||||||||||
2214 | // We need to use the set to populate domtree updates as even when there | ||||||||||
2215 | // are multiple cases pointing at the same successor we only want to | ||||||||||
2216 | // remove and insert one edge in the domtree. | ||||||||||
2217 | for (BasicBlock *SuccBB : UnswitchedSuccBBs) | ||||||||||
2218 | DTUpdates.push_back( | ||||||||||
2219 | {DominatorTree::Insert, SplitBB, ClonedPHs.find(SuccBB)->second}); | ||||||||||
2220 | } | ||||||||||
2221 | |||||||||||
2222 | if (MSSAU) { | ||||||||||
2223 | DT.applyUpdates(DTUpdates); | ||||||||||
2224 | DTUpdates.clear(); | ||||||||||
2225 | |||||||||||
2226 | // Remove all but one edge to the retained block and all unswitched | ||||||||||
2227 | // blocks. This is to avoid having duplicate entries in the cloned Phis, | ||||||||||
2228 | // when we know we only keep a single edge for each case. | ||||||||||
2229 | MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, RetainedSuccBB); | ||||||||||
2230 | for (BasicBlock *SuccBB : UnswitchedSuccBBs) | ||||||||||
2231 | MSSAU->removeDuplicatePhiEdgesBetween(ParentBB, SuccBB); | ||||||||||
2232 | |||||||||||
2233 | for (auto &VMap : VMaps) | ||||||||||
2234 | MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, | ||||||||||
2235 | /*IgnoreIncomingWithNoClones=*/true); | ||||||||||
2236 | MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); | ||||||||||
2237 | |||||||||||
2238 | // Remove all edges to unswitched blocks. | ||||||||||
2239 | for (BasicBlock *SuccBB : UnswitchedSuccBBs) | ||||||||||
2240 | MSSAU->removeEdge(ParentBB, SuccBB); | ||||||||||
2241 | } | ||||||||||
2242 | |||||||||||
2243 | // Now unhook the successor relationship as we'll be replacing | ||||||||||
2244 | // the terminator with a direct branch. This is much simpler for branches | ||||||||||
2245 | // than switches so we handle those first. | ||||||||||
2246 | if (BI) { | ||||||||||
2247 | // Remove the parent as a predecessor of the unswitched successor. | ||||||||||
2248 | assert(UnswitchedSuccBBs.size() == 1 &&((void)0) | ||||||||||
2249 | "Only one possible unswitched block for a branch!")((void)0); | ||||||||||
2250 | BasicBlock *UnswitchedSuccBB = *UnswitchedSuccBBs.begin(); | ||||||||||
2251 | UnswitchedSuccBB->removePredecessor(ParentBB, | ||||||||||
2252 | /*KeepOneInputPHIs*/ true); | ||||||||||
2253 | DTUpdates.push_back({DominatorTree::Delete, ParentBB, UnswitchedSuccBB}); | ||||||||||
2254 | } else { | ||||||||||
2255 | // Note that we actually want to remove the parent block as a predecessor | ||||||||||
2256 | // of *every* case successor. The case successor is either unswitched, | ||||||||||
2257 | // completely eliminating an edge from the parent to that successor, or it | ||||||||||
2258 | // is a duplicate edge to the retained successor as the retained successor | ||||||||||
2259 | // is always the default successor and as we'll replace this with a direct | ||||||||||
2260 | // branch we no longer need the duplicate entries in the PHI nodes. | ||||||||||
2261 | SwitchInst *NewSI = cast<SwitchInst>(NewTI); | ||||||||||
2262 | assert(NewSI->getDefaultDest() == RetainedSuccBB &&((void)0) | ||||||||||
2263 | "Not retaining default successor!")((void)0); | ||||||||||
2264 | for (auto &Case : NewSI->cases()) | ||||||||||
2265 | Case.getCaseSuccessor()->removePredecessor( | ||||||||||
2266 | ParentBB, | ||||||||||
2267 | /*KeepOneInputPHIs*/ true); | ||||||||||
2268 | |||||||||||
2269 | // We need to use the set to populate domtree updates as even when there | ||||||||||
2270 | // are multiple cases pointing at the same successor we only want to | ||||||||||
2271 | // remove and insert one edge in the domtree. | ||||||||||
2272 | for (BasicBlock *SuccBB : UnswitchedSuccBBs) | ||||||||||
2273 | DTUpdates.push_back({DominatorTree::Delete, ParentBB, SuccBB}); | ||||||||||
2274 | } | ||||||||||
2275 | |||||||||||
2276 | // After MSSAU update, remove the cloned terminator instruction NewTI. | ||||||||||
2277 | ParentBB->getTerminator()->eraseFromParent(); | ||||||||||
2278 | |||||||||||
2279 | // Create a new unconditional branch to the continuing block (as opposed to | ||||||||||
2280 | // the one cloned). | ||||||||||
2281 | BranchInst::Create(RetainedSuccBB, ParentBB); | ||||||||||
2282 | } else { | ||||||||||
2283 | assert(BI && "Only branches have partial unswitching.")((void)0); | ||||||||||
2284 | assert(UnswitchedSuccBBs.size() == 1 &&((void)0) | ||||||||||
2285 | "Only one possible unswitched block for a branch!")((void)0); | ||||||||||
2286 | BasicBlock *ClonedPH = ClonedPHs.begin()->second; | ||||||||||
2287 | // When doing a partial unswitch, we have to do a bit more work to build up | ||||||||||
2288 | // the branch in the split block. | ||||||||||
2289 | if (PartiallyInvariant) | ||||||||||
2290 | buildPartialInvariantUnswitchConditionalBranch( | ||||||||||
2291 | *SplitBB, Invariants, Direction, *ClonedPH, *LoopPH, L, MSSAU); | ||||||||||
2292 | else | ||||||||||
2293 | buildPartialUnswitchConditionalBranch(*SplitBB, Invariants, Direction, | ||||||||||
2294 | *ClonedPH, *LoopPH); | ||||||||||
2295 | DTUpdates.push_back({DominatorTree::Insert, SplitBB, ClonedPH}); | ||||||||||
2296 | |||||||||||
2297 | if (MSSAU) { | ||||||||||
2298 | DT.applyUpdates(DTUpdates); | ||||||||||
2299 | DTUpdates.clear(); | ||||||||||
2300 | |||||||||||
2301 | // Perform MSSA cloning updates. | ||||||||||
2302 | for (auto &VMap : VMaps) | ||||||||||
2303 | MSSAU->updateForClonedLoop(LBRPO, ExitBlocks, *VMap, | ||||||||||
2304 | /*IgnoreIncomingWithNoClones=*/true); | ||||||||||
2305 | MSSAU->updateExitBlocksForClonedLoop(ExitBlocks, VMaps, DT); | ||||||||||
2306 | } | ||||||||||
2307 | } | ||||||||||
2308 | |||||||||||
2309 | // Apply the updates accumulated above to get an up-to-date dominator tree. | ||||||||||
2310 | DT.applyUpdates(DTUpdates); | ||||||||||
2311 | |||||||||||
2312 | // Now that we have an accurate dominator tree, first delete the dead cloned | ||||||||||
2313 | // blocks so that we can accurately build any cloned loops. It is important to | ||||||||||
2314 | // not delete the blocks from the original loop yet because we still want to | ||||||||||
2315 | // reference the original loop to understand the cloned loop's structure. | ||||||||||
2316 | deleteDeadClonedBlocks(L, ExitBlocks, VMaps, DT, MSSAU); | ||||||||||
2317 | |||||||||||
2318 | // Build the cloned loop structure itself. This may be substantially | ||||||||||
2319 | // different from the original structure due to the simplified CFG. This also | ||||||||||
2320 | // handles inserting all the cloned blocks into the correct loops. | ||||||||||
2321 | SmallVector<Loop *, 4> NonChildClonedLoops; | ||||||||||
2322 | for (std::unique_ptr<ValueToValueMapTy> &VMap : VMaps) | ||||||||||
2323 | buildClonedLoops(L, ExitBlocks, *VMap, LI, NonChildClonedLoops); | ||||||||||
2324 | |||||||||||
2325 | // Now that our cloned loops have been built, we can update the original loop. | ||||||||||
2326 | // First we delete the dead blocks from it and then we rebuild the loop | ||||||||||
2327 | // structure taking these deletions into account. | ||||||||||
2328 | deleteDeadBlocksFromLoop(L, ExitBlocks, DT, LI, MSSAU, DestroyLoopCB); | ||||||||||
2329 | |||||||||||
2330 | if (MSSAU && VerifyMemorySSA) | ||||||||||
2331 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2332 | |||||||||||
2333 | SmallVector<Loop *, 4> HoistedLoops; | ||||||||||
2334 | bool IsStillLoop = rebuildLoopAfterUnswitch(L, ExitBlocks, LI, HoistedLoops); | ||||||||||
2335 | |||||||||||
2336 | if (MSSAU && VerifyMemorySSA) | ||||||||||
2337 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2338 | |||||||||||
2339 | // This transformation has a high risk of corrupting the dominator tree, and | ||||||||||
2340 | // the below steps to rebuild loop structures will result in hard to debug | ||||||||||
2341 | // errors in that case so verify that the dominator tree is sane first. | ||||||||||
2342 | // FIXME: Remove this when the bugs stop showing up and rely on existing | ||||||||||
2343 | // verification steps. | ||||||||||
2344 | assert(DT.verify(DominatorTree::VerificationLevel::Fast))((void)0); | ||||||||||
2345 | |||||||||||
2346 | if (BI && !PartiallyInvariant) { | ||||||||||
2347 | // If we unswitched a branch which collapses the condition to a known | ||||||||||
2348 | // constant we want to replace all the uses of the invariants within both | ||||||||||
2349 | // the original and cloned blocks. We do this here so that we can use the | ||||||||||
2350 | // now updated dominator tree to identify which side the users are on. | ||||||||||
2351 | assert(UnswitchedSuccBBs.size() == 1 &&((void)0) | ||||||||||
2352 | "Only one possible unswitched block for a branch!")((void)0); | ||||||||||
2353 | BasicBlock *ClonedPH = ClonedPHs.begin()->second; | ||||||||||
2354 | |||||||||||
2355 | // When considering multiple partially-unswitched invariants | ||||||||||
2356 | // we cant just go replace them with constants in both branches. | ||||||||||
2357 | // | ||||||||||
2358 | // For 'AND' we infer that true branch ("continue") means true | ||||||||||
2359 | // for each invariant operand. | ||||||||||
2360 | // For 'OR' we can infer that false branch ("continue") means false | ||||||||||
2361 | // for each invariant operand. | ||||||||||
2362 | // So it happens that for multiple-partial case we dont replace | ||||||||||
2363 | // in the unswitched branch. | ||||||||||
2364 | bool ReplaceUnswitched = | ||||||||||
2365 | FullUnswitch || (Invariants.size() == 1) || PartiallyInvariant; | ||||||||||
2366 | |||||||||||
2367 | ConstantInt *UnswitchedReplacement = | ||||||||||
2368 | Direction ? ConstantInt::getTrue(BI->getContext()) | ||||||||||
2369 | : ConstantInt::getFalse(BI->getContext()); | ||||||||||
2370 | ConstantInt *ContinueReplacement = | ||||||||||
2371 | Direction ? ConstantInt::getFalse(BI->getContext()) | ||||||||||
2372 | : ConstantInt::getTrue(BI->getContext()); | ||||||||||
2373 | for (Value *Invariant : Invariants) | ||||||||||
2374 | // Use make_early_inc_range here as set invalidates the iterator. | ||||||||||
2375 | for (Use &U : llvm::make_early_inc_range(Invariant->uses())) { | ||||||||||
2376 | Instruction *UserI = dyn_cast<Instruction>(U.getUser()); | ||||||||||
2377 | if (!UserI) | ||||||||||
2378 | continue; | ||||||||||
2379 | |||||||||||
2380 | // Replace it with the 'continue' side if in the main loop body, and the | ||||||||||
2381 | // unswitched if in the cloned blocks. | ||||||||||
2382 | if (DT.dominates(LoopPH, UserI->getParent())) | ||||||||||
2383 | U.set(ContinueReplacement); | ||||||||||
2384 | else if (ReplaceUnswitched && | ||||||||||
2385 | DT.dominates(ClonedPH, UserI->getParent())) | ||||||||||
2386 | U.set(UnswitchedReplacement); | ||||||||||
2387 | } | ||||||||||
2388 | } | ||||||||||
2389 | |||||||||||
2390 | // We can change which blocks are exit blocks of all the cloned sibling | ||||||||||
2391 | // loops, the current loop, and any parent loops which shared exit blocks | ||||||||||
2392 | // with the current loop. As a consequence, we need to re-form LCSSA for | ||||||||||
2393 | // them. But we shouldn't need to re-form LCSSA for any child loops. | ||||||||||
2394 | // FIXME: This could be made more efficient by tracking which exit blocks are | ||||||||||
2395 | // new, and focusing on them, but that isn't likely to be necessary. | ||||||||||
2396 | // | ||||||||||
2397 | // In order to reasonably rebuild LCSSA we need to walk inside-out across the | ||||||||||
2398 | // loop nest and update every loop that could have had its exits changed. We | ||||||||||
2399 | // also need to cover any intervening loops. We add all of these loops to | ||||||||||
2400 | // a list and sort them by loop depth to achieve this without updating | ||||||||||
2401 | // unnecessary loops. | ||||||||||
2402 | auto UpdateLoop = [&](Loop &UpdateL) { | ||||||||||
2403 | #ifndef NDEBUG1 | ||||||||||
2404 | UpdateL.verifyLoop(); | ||||||||||
2405 | for (Loop *ChildL : UpdateL) { | ||||||||||
2406 | ChildL->verifyLoop(); | ||||||||||
2407 | assert(ChildL->isRecursivelyLCSSAForm(DT, LI) &&((void)0) | ||||||||||
2408 | "Perturbed a child loop's LCSSA form!")((void)0); | ||||||||||
2409 | } | ||||||||||
2410 | #endif | ||||||||||
2411 | // First build LCSSA for this loop so that we can preserve it when | ||||||||||
2412 | // forming dedicated exits. We don't want to perturb some other loop's | ||||||||||
2413 | // LCSSA while doing that CFG edit. | ||||||||||
2414 | formLCSSA(UpdateL, DT, &LI, SE); | ||||||||||
2415 | |||||||||||
2416 | // For loops reached by this loop's original exit blocks we may | ||||||||||
2417 | // introduced new, non-dedicated exits. At least try to re-form dedicated | ||||||||||
2418 | // exits for these loops. This may fail if they couldn't have dedicated | ||||||||||
2419 | // exits to start with. | ||||||||||
2420 | formDedicatedExitBlocks(&UpdateL, &DT, &LI, MSSAU, /*PreserveLCSSA*/ true); | ||||||||||
2421 | }; | ||||||||||
2422 | |||||||||||
2423 | // For non-child cloned loops and hoisted loops, we just need to update LCSSA | ||||||||||
2424 | // and we can do it in any order as they don't nest relative to each other. | ||||||||||
2425 | // | ||||||||||
2426 | // Also check if any of the loops we have updated have become top-level loops | ||||||||||
2427 | // as that will necessitate widening the outer loop scope. | ||||||||||
2428 | for (Loop *UpdatedL : | ||||||||||
2429 | llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) { | ||||||||||
2430 | UpdateLoop(*UpdatedL); | ||||||||||
2431 | if (UpdatedL->isOutermost()) | ||||||||||
2432 | OuterExitL = nullptr; | ||||||||||
2433 | } | ||||||||||
2434 | if (IsStillLoop) { | ||||||||||
2435 | UpdateLoop(L); | ||||||||||
2436 | if (L.isOutermost()) | ||||||||||
2437 | OuterExitL = nullptr; | ||||||||||
2438 | } | ||||||||||
2439 | |||||||||||
2440 | // If the original loop had exit blocks, walk up through the outer most loop | ||||||||||
2441 | // of those exit blocks to update LCSSA and form updated dedicated exits. | ||||||||||
2442 | if (OuterExitL != &L) | ||||||||||
2443 | for (Loop *OuterL = ParentL; OuterL != OuterExitL; | ||||||||||
2444 | OuterL = OuterL->getParentLoop()) | ||||||||||
2445 | UpdateLoop(*OuterL); | ||||||||||
2446 | |||||||||||
2447 | #ifndef NDEBUG1 | ||||||||||
2448 | // Verify the entire loop structure to catch any incorrect updates before we | ||||||||||
2449 | // progress in the pass pipeline. | ||||||||||
2450 | LI.verify(DT); | ||||||||||
2451 | #endif | ||||||||||
2452 | |||||||||||
2453 | // Now that we've unswitched something, make callbacks to report the changes. | ||||||||||
2454 | // For that we need to merge together the updated loops and the cloned loops | ||||||||||
2455 | // and check whether the original loop survived. | ||||||||||
2456 | SmallVector<Loop *, 4> SibLoops; | ||||||||||
2457 | for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops)) | ||||||||||
2458 | if (UpdatedL->getParentLoop() == ParentL) | ||||||||||
2459 | SibLoops.push_back(UpdatedL); | ||||||||||
2460 | UnswitchCB(IsStillLoop, PartiallyInvariant, SibLoops); | ||||||||||
2461 | |||||||||||
2462 | if (MSSAU && VerifyMemorySSA) | ||||||||||
2463 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2464 | |||||||||||
2465 | if (BI) | ||||||||||
2466 | ++NumBranches; | ||||||||||
2467 | else | ||||||||||
2468 | ++NumSwitches; | ||||||||||
2469 | } | ||||||||||
2470 | |||||||||||
2471 | /// Recursively compute the cost of a dominator subtree based on the per-block | ||||||||||
2472 | /// cost map provided. | ||||||||||
2473 | /// | ||||||||||
2474 | /// The recursive computation is memozied into the provided DT-indexed cost map | ||||||||||
2475 | /// to allow querying it for most nodes in the domtree without it becoming | ||||||||||
2476 | /// quadratic. | ||||||||||
2477 | static InstructionCost computeDomSubtreeCost( | ||||||||||
2478 | DomTreeNode &N, | ||||||||||
2479 | const SmallDenseMap<BasicBlock *, InstructionCost, 4> &BBCostMap, | ||||||||||
2480 | SmallDenseMap<DomTreeNode *, InstructionCost, 4> &DTCostMap) { | ||||||||||
2481 | // Don't accumulate cost (or recurse through) blocks not in our block cost | ||||||||||
2482 | // map and thus not part of the duplication cost being considered. | ||||||||||
2483 | auto BBCostIt = BBCostMap.find(N.getBlock()); | ||||||||||
2484 | if (BBCostIt == BBCostMap.end()) | ||||||||||
2485 | return 0; | ||||||||||
2486 | |||||||||||
2487 | // Lookup this node to see if we already computed its cost. | ||||||||||
2488 | auto DTCostIt = DTCostMap.find(&N); | ||||||||||
2489 | if (DTCostIt != DTCostMap.end()) | ||||||||||
2490 | return DTCostIt->second; | ||||||||||
2491 | |||||||||||
2492 | // If not, we have to compute it. We can't use insert above and update | ||||||||||
2493 | // because computing the cost may insert more things into the map. | ||||||||||
2494 | InstructionCost Cost = std::accumulate( | ||||||||||
2495 | N.begin(), N.end(), BBCostIt->second, | ||||||||||
2496 | [&](InstructionCost Sum, DomTreeNode *ChildN) -> InstructionCost { | ||||||||||
2497 | return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap); | ||||||||||
2498 | }); | ||||||||||
2499 | bool Inserted = DTCostMap.insert({&N, Cost}).second; | ||||||||||
2500 | (void)Inserted; | ||||||||||
2501 | assert(Inserted && "Should not insert a node while visiting children!")((void)0); | ||||||||||
2502 | return Cost; | ||||||||||
2503 | } | ||||||||||
2504 | |||||||||||
2505 | /// Turns a llvm.experimental.guard intrinsic into implicit control flow branch, | ||||||||||
2506 | /// making the following replacement: | ||||||||||
2507 | /// | ||||||||||
2508 | /// --code before guard-- | ||||||||||
2509 | /// call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ] | ||||||||||
2510 | /// --code after guard-- | ||||||||||
2511 | /// | ||||||||||
2512 | /// into | ||||||||||
2513 | /// | ||||||||||
2514 | /// --code before guard-- | ||||||||||
2515 | /// br i1 %cond, label %guarded, label %deopt | ||||||||||
2516 | /// | ||||||||||
2517 | /// guarded: | ||||||||||
2518 | /// --code after guard-- | ||||||||||
2519 | /// | ||||||||||
2520 | /// deopt: | ||||||||||
2521 | /// call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ] | ||||||||||
2522 | /// unreachable | ||||||||||
2523 | /// | ||||||||||
2524 | /// It also makes all relevant DT and LI updates, so that all structures are in | ||||||||||
2525 | /// valid state after this transform. | ||||||||||
2526 | static BranchInst * | ||||||||||
2527 | turnGuardIntoBranch(IntrinsicInst *GI, Loop &L, | ||||||||||
2528 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | ||||||||||
2529 | DominatorTree &DT, LoopInfo &LI, MemorySSAUpdater *MSSAU) { | ||||||||||
2530 | SmallVector<DominatorTree::UpdateType, 4> DTUpdates; | ||||||||||
2531 | LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n")do { } while (false); | ||||||||||
2532 | BasicBlock *CheckBB = GI->getParent(); | ||||||||||
2533 | |||||||||||
2534 | if (MSSAU && VerifyMemorySSA) | ||||||||||
2535 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2536 | |||||||||||
2537 | // Remove all CheckBB's successors from DomTree. A block can be seen among | ||||||||||
2538 | // successors more than once, but for DomTree it should be added only once. | ||||||||||
2539 | SmallPtrSet<BasicBlock *, 4> Successors; | ||||||||||
2540 | for (auto *Succ : successors(CheckBB)) | ||||||||||
2541 | if (Successors.insert(Succ).second) | ||||||||||
2542 | DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ}); | ||||||||||
2543 | |||||||||||
2544 | Instruction *DeoptBlockTerm = | ||||||||||
2545 | SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true); | ||||||||||
2546 | BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator()); | ||||||||||
2547 | // SplitBlockAndInsertIfThen inserts control flow that branches to | ||||||||||
2548 | // DeoptBlockTerm if the condition is true. We want the opposite. | ||||||||||
2549 | CheckBI->swapSuccessors(); | ||||||||||
2550 | |||||||||||
2551 | BasicBlock *GuardedBlock = CheckBI->getSuccessor(0); | ||||||||||
2552 | GuardedBlock->setName("guarded"); | ||||||||||
2553 | CheckBI->getSuccessor(1)->setName("deopt"); | ||||||||||
2554 | BasicBlock *DeoptBlock = CheckBI->getSuccessor(1); | ||||||||||
2555 | |||||||||||
2556 | // We now have a new exit block. | ||||||||||
2557 | ExitBlocks.push_back(CheckBI->getSuccessor(1)); | ||||||||||
2558 | |||||||||||
2559 | if (MSSAU) | ||||||||||
2560 | MSSAU->moveAllAfterSpliceBlocks(CheckBB, GuardedBlock, GI); | ||||||||||
2561 | |||||||||||
2562 | GI->moveBefore(DeoptBlockTerm); | ||||||||||
2563 | GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext())); | ||||||||||
2564 | |||||||||||
2565 | // Add new successors of CheckBB into DomTree. | ||||||||||
2566 | for (auto *Succ : successors(CheckBB)) | ||||||||||
2567 | DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ}); | ||||||||||
2568 | |||||||||||
2569 | // Now the blocks that used to be CheckBB's successors are GuardedBlock's | ||||||||||
2570 | // successors. | ||||||||||
2571 | for (auto *Succ : Successors) | ||||||||||
2572 | DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ}); | ||||||||||
2573 | |||||||||||
2574 | // Make proper changes to DT. | ||||||||||
2575 | DT.applyUpdates(DTUpdates); | ||||||||||
2576 | // Inform LI of a new loop block. | ||||||||||
2577 | L.addBasicBlockToLoop(GuardedBlock, LI); | ||||||||||
2578 | |||||||||||
2579 | if (MSSAU) { | ||||||||||
2580 | MemoryDef *MD = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(GI)); | ||||||||||
2581 | MSSAU->moveToPlace(MD, DeoptBlock, MemorySSA::BeforeTerminator); | ||||||||||
2582 | if (VerifyMemorySSA) | ||||||||||
2583 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||||
2584 | } | ||||||||||
2585 | |||||||||||
2586 | ++NumGuards; | ||||||||||
2587 | return CheckBI; | ||||||||||
2588 | } | ||||||||||
2589 | |||||||||||
2590 | /// Cost multiplier is a way to limit potentially exponential behavior | ||||||||||
2591 | /// of loop-unswitch. Cost is multipied in proportion of 2^number of unswitch | ||||||||||
2592 | /// candidates available. Also accounting for the number of "sibling" loops with | ||||||||||
2593 | /// the idea to account for previous unswitches that already happened on this | ||||||||||
2594 | /// cluster of loops. There was an attempt to keep this formula simple, | ||||||||||
2595 | /// just enough to limit the worst case behavior. Even if it is not that simple | ||||||||||
2596 | /// now it is still not an attempt to provide a detailed heuristic size | ||||||||||
2597 | /// prediction. | ||||||||||
2598 | /// | ||||||||||
2599 | /// TODO: Make a proper accounting of "explosion" effect for all kinds of | ||||||||||
2600 | /// unswitch candidates, making adequate predictions instead of wild guesses. | ||||||||||
2601 | /// That requires knowing not just the number of "remaining" candidates but | ||||||||||
2602 | /// also costs of unswitching for each of these candidates. | ||||||||||
2603 | static int CalculateUnswitchCostMultiplier( | ||||||||||
2604 | Instruction &TI, Loop &L, LoopInfo &LI, DominatorTree &DT, | ||||||||||
2605 | ArrayRef<std::pair<Instruction *, TinyPtrVector<Value *>>> | ||||||||||
2606 | UnswitchCandidates) { | ||||||||||
2607 | |||||||||||
2608 | // Guards and other exiting conditions do not contribute to exponential | ||||||||||
2609 | // explosion as soon as they dominate the latch (otherwise there might be | ||||||||||
2610 | // another path to the latch remaining that does not allow to eliminate the | ||||||||||
2611 | // loop copy on unswitch). | ||||||||||
2612 | BasicBlock *Latch = L.getLoopLatch(); | ||||||||||
2613 | BasicBlock *CondBlock = TI.getParent(); | ||||||||||
2614 | if (DT.dominates(CondBlock, Latch) && | ||||||||||
2615 | (isGuard(&TI) || | ||||||||||
2616 | llvm::count_if(successors(&TI), [&L](BasicBlock *SuccBB) { | ||||||||||
2617 | return L.contains(SuccBB); | ||||||||||
2618 | }) <= 1)) { | ||||||||||
2619 | NumCostMultiplierSkipped++; | ||||||||||
2620 | return 1; | ||||||||||
2621 | } | ||||||||||
2622 | |||||||||||
2623 | auto *ParentL = L.getParentLoop(); | ||||||||||
2624 | int SiblingsCount = (ParentL ? ParentL->getSubLoopsVector().size() | ||||||||||
2625 | : std::distance(LI.begin(), LI.end())); | ||||||||||
2626 | // Count amount of clones that all the candidates might cause during | ||||||||||
2627 | // unswitching. Branch/guard counts as 1, switch counts as log2 of its cases. | ||||||||||
2628 | int UnswitchedClones = 0; | ||||||||||
2629 | for (auto Candidate : UnswitchCandidates) { | ||||||||||
2630 | Instruction *CI = Candidate.first; | ||||||||||
2631 | BasicBlock *CondBlock = CI->getParent(); | ||||||||||
2632 | bool SkipExitingSuccessors = DT.dominates(CondBlock, Latch); | ||||||||||
2633 | if (isGuard(CI)) { | ||||||||||
2634 | if (!SkipExitingSuccessors) | ||||||||||
2635 | UnswitchedClones++; | ||||||||||
2636 | continue; | ||||||||||
2637 | } | ||||||||||
2638 | int NonExitingSuccessors = llvm::count_if( | ||||||||||
2639 | successors(CondBlock), [SkipExitingSuccessors, &L](BasicBlock *SuccBB) { | ||||||||||
2640 | return !SkipExitingSuccessors || L.contains(SuccBB); | ||||||||||
2641 | }); | ||||||||||
2642 | UnswitchedClones += Log2_32(NonExitingSuccessors); | ||||||||||
2643 | } | ||||||||||
2644 | |||||||||||
2645 | // Ignore up to the "unscaled candidates" number of unswitch candidates | ||||||||||
2646 | // when calculating the power-of-two scaling of the cost. The main idea | ||||||||||
2647 | // with this control is to allow a small number of unswitches to happen | ||||||||||
2648 | // and rely more on siblings multiplier (see below) when the number | ||||||||||
2649 | // of candidates is small. | ||||||||||
2650 | unsigned ClonesPower = | ||||||||||
2651 | std::max(UnswitchedClones - (int)UnswitchNumInitialUnscaledCandidates, 0); | ||||||||||
2652 | |||||||||||
2653 | // Allowing top-level loops to spread a bit more than nested ones. | ||||||||||
2654 | int SiblingsMultiplier = | ||||||||||
2655 | std::max((ParentL ? SiblingsCount | ||||||||||
2656 | : SiblingsCount / (int)UnswitchSiblingsToplevelDiv), | ||||||||||
2657 | 1); | ||||||||||
2658 | // Compute the cost multiplier in a way that won't overflow by saturating | ||||||||||
2659 | // at an upper bound. | ||||||||||
2660 | int CostMultiplier; | ||||||||||
2661 | if (ClonesPower > Log2_32(UnswitchThreshold) || | ||||||||||
2662 | SiblingsMultiplier > UnswitchThreshold) | ||||||||||
2663 | CostMultiplier = UnswitchThreshold; | ||||||||||
2664 | else | ||||||||||
2665 | CostMultiplier = std::min(SiblingsMultiplier * (1 << ClonesPower), | ||||||||||
2666 | (int)UnswitchThreshold); | ||||||||||
2667 | |||||||||||
2668 | LLVM_DEBUG(dbgs() << " Computed multiplier " << CostMultiplierdo { } while (false) | ||||||||||
2669 | << " (siblings " << SiblingsMultiplier << " * clones "do { } while (false) | ||||||||||
2670 | << (1 << ClonesPower) << ")"do { } while (false) | ||||||||||
2671 | << " for unswitch candidate: " << TI << "\n")do { } while (false); | ||||||||||
2672 | return CostMultiplier; | ||||||||||
2673 | } | ||||||||||
2674 | |||||||||||
2675 | static bool unswitchBestCondition( | ||||||||||
2676 | Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, | ||||||||||
2677 | AAResults &AA, TargetTransformInfo &TTI, | ||||||||||
2678 | function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB, | ||||||||||
2679 | ScalarEvolution *SE, MemorySSAUpdater *MSSAU, | ||||||||||
2680 | function_ref<void(Loop &, StringRef)> DestroyLoopCB) { | ||||||||||
2681 | // Collect all invariant conditions within this loop (as opposed to an inner | ||||||||||
2682 | // loop which would be handled when visiting that inner loop). | ||||||||||
2683 | SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4> | ||||||||||
2684 | UnswitchCandidates; | ||||||||||
2685 | |||||||||||
2686 | // Whether or not we should also collect guards in the loop. | ||||||||||
2687 | bool CollectGuards = false; | ||||||||||
2688 | if (UnswitchGuards) { | ||||||||||
| |||||||||||
2689 | auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction( | ||||||||||
2690 | Intrinsic::getName(Intrinsic::experimental_guard)); | ||||||||||
2691 | if (GuardDecl && !GuardDecl->use_empty()) | ||||||||||
2692 | CollectGuards = true; | ||||||||||
2693 | } | ||||||||||
2694 | |||||||||||
2695 | IVConditionInfo PartialIVInfo; | ||||||||||
2696 | for (auto *BB : L.blocks()) { | ||||||||||
2697 | if (LI.getLoopFor(BB) != &L) | ||||||||||
2698 | continue; | ||||||||||
2699 | |||||||||||
2700 | if (CollectGuards) | ||||||||||
2701 | for (auto &I : *BB) | ||||||||||
2702 | if (isGuard(&I)) { | ||||||||||
2703 | auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0); | ||||||||||
2704 | // TODO: Support AND, OR conditions and partial unswitching. | ||||||||||
2705 | if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond)) | ||||||||||
2706 | UnswitchCandidates.push_back({&I, {Cond}}); | ||||||||||
2707 | } | ||||||||||
2708 | |||||||||||
2709 | if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { | ||||||||||
2710 | // We can only consider fully loop-invariant switch conditions as we need | ||||||||||
2711 | // to completely eliminate the switch after unswitching. | ||||||||||
2712 | if (!isa<Constant>(SI->getCondition()) && | ||||||||||
2713 | L.isLoopInvariant(SI->getCondition()) && !BB->getUniqueSuccessor()) | ||||||||||
2714 | UnswitchCandidates.push_back({SI, {SI->getCondition()}}); | ||||||||||
2715 | continue; | ||||||||||
2716 | } | ||||||||||
2717 | |||||||||||
2718 | auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); | ||||||||||
2719 | if (!BI || !BI->isConditional() || isa<Constant>(BI->getCondition()) || | ||||||||||
2720 | BI->getSuccessor(0) == BI->getSuccessor(1)) | ||||||||||
2721 | continue; | ||||||||||
2722 | |||||||||||
2723 | // If BI's condition is 'select _, true, false', simplify it to confuse | ||||||||||
2724 | // matchers | ||||||||||
2725 | Value *Cond = BI->getCondition(), *CondNext; | ||||||||||
2726 | while (match(Cond, m_Select(m_Value(CondNext), m_One(), m_Zero()))) | ||||||||||
2727 | Cond = CondNext; | ||||||||||
2728 | BI->setCondition(Cond); | ||||||||||
2729 | |||||||||||
2730 | if (L.isLoopInvariant(BI->getCondition())) { | ||||||||||
2731 | UnswitchCandidates.push_back({BI, {BI->getCondition()}}); | ||||||||||
2732 | continue; | ||||||||||
2733 | } | ||||||||||
2734 | |||||||||||
2735 | Instruction &CondI = *cast<Instruction>(BI->getCondition()); | ||||||||||
2736 | if (match(&CondI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()))) { | ||||||||||
2737 | TinyPtrVector<Value *> Invariants = | ||||||||||
2738 | collectHomogenousInstGraphLoopInvariants(L, CondI, LI); | ||||||||||
2739 | if (Invariants.empty()) | ||||||||||
2740 | continue; | ||||||||||
2741 | |||||||||||
2742 | UnswitchCandidates.push_back({BI, std::move(Invariants)}); | ||||||||||
2743 | continue; | ||||||||||
2744 | } | ||||||||||
2745 | } | ||||||||||
2746 | |||||||||||
2747 | Instruction *PartialIVCondBranch = nullptr; | ||||||||||
2748 | if (MSSAU && !findOptionMDForLoop(&L, "llvm.loop.unswitch.partial.disable") && | ||||||||||
2749 | !any_of(UnswitchCandidates, [&L](auto &TerminatorAndInvariants) { | ||||||||||
2750 | return TerminatorAndInvariants.first == L.getHeader()->getTerminator(); | ||||||||||
2751 | })) { | ||||||||||
2752 | MemorySSA *MSSA = MSSAU->getMemorySSA(); | ||||||||||
2753 | if (auto Info = hasPartialIVCondition(L, MSSAThreshold, *MSSA, AA)) { | ||||||||||
2754 | LLVM_DEBUG(do { } while (false) | ||||||||||
2755 | dbgs() << "simple-loop-unswitch: Found partially invariant condition "do { } while (false) | ||||||||||
2756 | << *Info->InstToDuplicate[0] << "\n")do { } while (false); | ||||||||||
2757 | PartialIVInfo = *Info; | ||||||||||
2758 | PartialIVCondBranch = L.getHeader()->getTerminator(); | ||||||||||
2759 | TinyPtrVector<Value *> ValsToDuplicate; | ||||||||||
2760 | for (auto *Inst : Info->InstToDuplicate) | ||||||||||
2761 | ValsToDuplicate.push_back(Inst); | ||||||||||
2762 | UnswitchCandidates.push_back( | ||||||||||
2763 | {L.getHeader()->getTerminator(), std::move(ValsToDuplicate)}); | ||||||||||
2764 | } | ||||||||||
2765 | } | ||||||||||
2766 | |||||||||||
2767 | // If we didn't find any candidates, we're done. | ||||||||||
2768 | if (UnswitchCandidates.empty()) | ||||||||||
2769 | return false; | ||||||||||
2770 | |||||||||||
2771 | // Check if there are irreducible CFG cycles in this loop. If so, we cannot | ||||||||||
2772 | // easily unswitch non-trivial edges out of the loop. Doing so might turn the | ||||||||||
2773 | // irreducible control flow into reducible control flow and introduce new | ||||||||||
2774 | // loops "out of thin air". If we ever discover important use cases for doing | ||||||||||
2775 | // this, we can add support to loop unswitch, but it is a lot of complexity | ||||||||||
2776 | // for what seems little or no real world benefit. | ||||||||||
2777 | LoopBlocksRPO RPOT(&L); | ||||||||||
2778 | RPOT.perform(&LI); | ||||||||||
2779 | if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) | ||||||||||
2780 | return false; | ||||||||||
2781 | |||||||||||
2782 | SmallVector<BasicBlock *, 4> ExitBlocks; | ||||||||||
2783 | L.getUniqueExitBlocks(ExitBlocks); | ||||||||||
2784 | |||||||||||
2785 | // We cannot unswitch if exit blocks contain a cleanuppad/catchswitch | ||||||||||
2786 | // instruction as we don't know how to split those exit blocks. | ||||||||||
2787 | // FIXME: We should teach SplitBlock to handle this and remove this | ||||||||||
2788 | // restriction. | ||||||||||
2789 | for (auto *ExitBB : ExitBlocks) { | ||||||||||
2790 | auto *I = ExitBB->getFirstNonPHI(); | ||||||||||
2791 | if (isa<CleanupPadInst>(I) || isa<CatchSwitchInst>(I)) { | ||||||||||
2792 | LLVM_DEBUG(dbgs() << "Cannot unswitch because of cleanuppad/catchswitch "do { } while (false) | ||||||||||
2793 | "in exit block\n")do { } while (false); | ||||||||||
2794 | return false; | ||||||||||
2795 | } | ||||||||||
2796 | } | ||||||||||
2797 | |||||||||||
2798 | LLVM_DEBUG(do { } while (false) | ||||||||||
2799 | dbgs() << "Considering " << UnswitchCandidates.size()do { } while (false) | ||||||||||
2800 | << " non-trivial loop invariant conditions for unswitching.\n")do { } while (false); | ||||||||||
2801 | |||||||||||
2802 | // Given that unswitching these terminators will require duplicating parts of | ||||||||||
2803 | // the loop, so we need to be able to model that cost. Compute the ephemeral | ||||||||||
2804 | // values and set up a data structure to hold per-BB costs. We cache each | ||||||||||
2805 | // block's cost so that we don't recompute this when considering different | ||||||||||
2806 | // subsets of the loop for duplication during unswitching. | ||||||||||
2807 | SmallPtrSet<const Value *, 4> EphValues; | ||||||||||
2808 | CodeMetrics::collectEphemeralValues(&L, &AC, EphValues); | ||||||||||
2809 | SmallDenseMap<BasicBlock *, InstructionCost, 4> BBCostMap; | ||||||||||
2810 | |||||||||||
2811 | // Compute the cost of each block, as well as the total loop cost. Also, bail | ||||||||||
2812 | // out if we see instructions which are incompatible with loop unswitching | ||||||||||
2813 | // (convergent, noduplicate, or cross-basic-block tokens). | ||||||||||
2814 | // FIXME: We might be able to safely handle some of these in non-duplicated | ||||||||||
2815 | // regions. | ||||||||||
2816 | TargetTransformInfo::TargetCostKind CostKind = | ||||||||||
2817 | L.getHeader()->getParent()->hasMinSize() | ||||||||||
2818 | ? TargetTransformInfo::TCK_CodeSize | ||||||||||
2819 | : TargetTransformInfo::TCK_SizeAndLatency; | ||||||||||
2820 | InstructionCost LoopCost = 0; | ||||||||||
2821 | for (auto *BB : L.blocks()) { | ||||||||||
2822 | InstructionCost Cost = 0; | ||||||||||
2823 | for (auto &I : *BB) { | ||||||||||
2824 | if (EphValues.count(&I)) | ||||||||||
2825 | continue; | ||||||||||
2826 | |||||||||||
2827 | if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB)) | ||||||||||
2828 | return false; | ||||||||||
2829 | if (auto *CB = dyn_cast<CallBase>(&I)) | ||||||||||
2830 | if (CB->isConvergent() || CB->cannotDuplicate()) | ||||||||||
2831 | return false; | ||||||||||
2832 | |||||||||||
2833 | Cost += TTI.getUserCost(&I, CostKind); | ||||||||||
2834 | } | ||||||||||
2835 | assert(Cost >= 0 && "Must not have negative costs!")((void)0); | ||||||||||
2836 | LoopCost += Cost; | ||||||||||
2837 | assert(LoopCost >= 0 && "Must not have negative loop costs!")((void)0); | ||||||||||
2838 | BBCostMap[BB] = Cost; | ||||||||||
2839 | } | ||||||||||
2840 | LLVM_DEBUG(dbgs() << " Total loop cost: " << LoopCost << "\n")do { } while (false); | ||||||||||
2841 | |||||||||||
2842 | // Now we find the best candidate by searching for the one with the following | ||||||||||
2843 | // properties in order: | ||||||||||
2844 | // | ||||||||||
2845 | // 1) An unswitching cost below the threshold | ||||||||||
2846 | // 2) The smallest number of duplicated unswitch candidates (to avoid | ||||||||||
2847 | // creating redundant subsequent unswitching) | ||||||||||
2848 | // 3) The smallest cost after unswitching. | ||||||||||
2849 | // | ||||||||||
2850 | // We prioritize reducing fanout of unswitch candidates provided the cost | ||||||||||
2851 | // remains below the threshold because this has a multiplicative effect. | ||||||||||
2852 | // | ||||||||||
2853 | // This requires memoizing each dominator subtree to avoid redundant work. | ||||||||||
2854 | // | ||||||||||
2855 | // FIXME: Need to actually do the number of candidates part above. | ||||||||||
2856 | SmallDenseMap<DomTreeNode *, InstructionCost, 4> DTCostMap; | ||||||||||
2857 | // Given a terminator which might be unswitched, computes the non-duplicated | ||||||||||
2858 | // cost for that terminator. | ||||||||||
2859 | auto ComputeUnswitchedCost = [&](Instruction &TI, | ||||||||||
2860 | bool FullUnswitch) -> InstructionCost { | ||||||||||
2861 | BasicBlock &BB = *TI.getParent(); | ||||||||||
2862 | SmallPtrSet<BasicBlock *, 4> Visited; | ||||||||||
2863 | |||||||||||
2864 | InstructionCost Cost = 0; | ||||||||||
2865 | for (BasicBlock *SuccBB : successors(&BB)) { | ||||||||||
2866 | // Don't count successors more than once. | ||||||||||
2867 | if (!Visited.insert(SuccBB).second) | ||||||||||
2868 | continue; | ||||||||||
2869 | |||||||||||
2870 | // If this is a partial unswitch candidate, then it must be a conditional | ||||||||||
2871 | // branch with a condition of either `or`, `and`, their corresponding | ||||||||||
2872 | // select forms or partially invariant instructions. In that case, one of | ||||||||||
2873 | // the successors is necessarily duplicated, so don't even try to remove | ||||||||||
2874 | // its cost. | ||||||||||
2875 | if (!FullUnswitch
| ||||||||||
2876 | auto &BI = cast<BranchInst>(TI); | ||||||||||
2877 | if (match(BI.getCondition(), m_LogicalAnd())) { | ||||||||||
2878 | if (SuccBB == BI.getSuccessor(1)) | ||||||||||
2879 | continue; | ||||||||||
2880 | } else if (match(BI.getCondition(), m_LogicalOr())) { | ||||||||||
2881 | if (SuccBB == BI.getSuccessor(0)) | ||||||||||
2882 | continue; | ||||||||||
2883 | } else if ((PartialIVInfo.KnownValue->isOneValue() && | ||||||||||
| |||||||||||
2884 | SuccBB == BI.getSuccessor(0)) || | ||||||||||
2885 | (!PartialIVInfo.KnownValue->isOneValue() && | ||||||||||
2886 | SuccBB == BI.getSuccessor(1))) | ||||||||||
2887 | continue; | ||||||||||
2888 | } | ||||||||||
2889 | |||||||||||
2890 | // This successor's domtree will not need to be duplicated after | ||||||||||
2891 | // unswitching if the edge to the successor dominates it (and thus the | ||||||||||
2892 | // entire tree). This essentially means there is no other path into this | ||||||||||
2893 | // subtree and so it will end up live in only one clone of the loop. | ||||||||||
2894 | if (SuccBB->getUniquePredecessor() || | ||||||||||
2895 | llvm::all_of(predecessors(SuccBB), [&](BasicBlock *PredBB) { | ||||||||||
2896 | return PredBB == &BB || DT.dominates(SuccBB, PredBB); | ||||||||||
2897 | })) { | ||||||||||
2898 | Cost += computeDomSubtreeCost(*DT[SuccBB], BBCostMap, DTCostMap); | ||||||||||
2899 | assert(Cost <= LoopCost &&((void)0) | ||||||||||
2900 | "Non-duplicated cost should never exceed total loop cost!")((void)0); | ||||||||||
2901 | } | ||||||||||
2902 | } | ||||||||||
2903 | |||||||||||
2904 | // Now scale the cost by the number of unique successors minus one. We | ||||||||||
2905 | // subtract one because there is already at least one copy of the entire | ||||||||||
2906 | // loop. This is computing the new cost of unswitching a condition. | ||||||||||
2907 | // Note that guards always have 2 unique successors that are implicit and | ||||||||||
2908 | // will be materialized if we decide to unswitch it. | ||||||||||
2909 | int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size(); | ||||||||||
2910 | assert(SuccessorsCount > 1 &&((void)0) | ||||||||||
2911 | "Cannot unswitch a condition without multiple distinct successors!")((void)0); | ||||||||||
2912 | return (LoopCost - Cost) * (SuccessorsCount - 1); | ||||||||||
2913 | }; | ||||||||||
2914 | Instruction *BestUnswitchTI = nullptr; | ||||||||||
2915 | InstructionCost BestUnswitchCost = 0; | ||||||||||
2916 | ArrayRef<Value *> BestUnswitchInvariants; | ||||||||||
2917 | for (auto &TerminatorAndInvariants : UnswitchCandidates) { | ||||||||||
2918 | Instruction &TI = *TerminatorAndInvariants.first; | ||||||||||
2919 | ArrayRef<Value *> Invariants = TerminatorAndInvariants.second; | ||||||||||
2920 | BranchInst *BI = dyn_cast<BranchInst>(&TI); | ||||||||||
2921 | InstructionCost CandidateCost = ComputeUnswitchedCost( | ||||||||||
2922 | TI, /*FullUnswitch*/ !BI
| ||||||||||
2923 | Invariants[0] == BI->getCondition())); | ||||||||||
2924 | // Calculate cost multiplier which is a tool to limit potentially | ||||||||||
2925 | // exponential behavior of loop-unswitch. | ||||||||||
2926 | if (EnableUnswitchCostMultiplier) { | ||||||||||
2927 | int CostMultiplier = | ||||||||||
2928 | CalculateUnswitchCostMultiplier(TI, L, LI, DT, UnswitchCandidates); | ||||||||||
2929 | assert(((void)0) | ||||||||||
2930 | (CostMultiplier > 0 && CostMultiplier <= UnswitchThreshold) &&((void)0) | ||||||||||
2931 | "cost multiplier needs to be in the range of 1..UnswitchThreshold")((void)0); | ||||||||||
2932 | CandidateCost *= CostMultiplier; | ||||||||||
2933 | LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCostdo { } while (false) | ||||||||||
2934 | << " (multiplier: " << CostMultiplier << ")"do { } while (false) | ||||||||||
2935 | << " for unswitch candidate: " << TI << "\n")do { } while (false); | ||||||||||
2936 | } else { | ||||||||||
2937 | LLVM_DEBUG(dbgs() << " Computed cost of " << CandidateCostdo { } while (false) | ||||||||||
2938 | << " for unswitch candidate: " << TI << "\n")do { } while (false); | ||||||||||
2939 | } | ||||||||||
2940 | |||||||||||
2941 | if (!BestUnswitchTI || CandidateCost < BestUnswitchCost) { | ||||||||||
2942 | BestUnswitchTI = &TI; | ||||||||||
2943 | BestUnswitchCost = CandidateCost; | ||||||||||
2944 | BestUnswitchInvariants = Invariants; | ||||||||||
2945 | } | ||||||||||
2946 | } | ||||||||||
2947 | assert(BestUnswitchTI && "Failed to find loop unswitch candidate")((void)0); | ||||||||||
2948 | |||||||||||
2949 | if (BestUnswitchCost >= UnswitchThreshold) { | ||||||||||
2950 | LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "do { } while (false) | ||||||||||
2951 | << BestUnswitchCost << "\n")do { } while (false); | ||||||||||
2952 | return false; | ||||||||||
2953 | } | ||||||||||
2954 | |||||||||||
2955 | if (BestUnswitchTI != PartialIVCondBranch) | ||||||||||
2956 | PartialIVInfo.InstToDuplicate.clear(); | ||||||||||
2957 | |||||||||||
2958 | // If the best candidate is a guard, turn it into a branch. | ||||||||||
2959 | if (isGuard(BestUnswitchTI)) | ||||||||||
2960 | BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L, | ||||||||||
2961 | ExitBlocks, DT, LI, MSSAU); | ||||||||||
2962 | |||||||||||
2963 | LLVM_DEBUG(dbgs() << " Unswitching non-trivial (cost = "do { } while (false) | ||||||||||
2964 | << BestUnswitchCost << ") terminator: " << *BestUnswitchTIdo { } while (false) | ||||||||||
2965 | << "\n")do { } while (false); | ||||||||||
2966 | unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants, | ||||||||||
2967 | ExitBlocks, PartialIVInfo, DT, LI, AC, | ||||||||||
2968 | UnswitchCB, SE, MSSAU, DestroyLoopCB); | ||||||||||
2969 | return true; | ||||||||||
2970 | } | ||||||||||
2971 | |||||||||||
2972 | /// Unswitch control flow predicated on loop invariant conditions. | ||||||||||
2973 | /// | ||||||||||
2974 | /// This first hoists all branches or switches which are trivial (IE, do not | ||||||||||
2975 | /// require duplicating any part of the loop) out of the loop body. It then | ||||||||||
2976 | /// looks at other loop invariant control flows and tries to unswitch those as | ||||||||||
2977 | /// well by cloning the loop if the result is small enough. | ||||||||||
2978 | /// | ||||||||||
2979 | /// The `DT`, `LI`, `AC`, `AA`, `TTI` parameters are required analyses that are | ||||||||||
2980 | /// also updated based on the unswitch. The `MSSA` analysis is also updated if | ||||||||||
2981 | /// valid (i.e. its use is enabled). | ||||||||||
2982 | /// | ||||||||||
2983 | /// If either `NonTrivial` is true or the flag `EnableNonTrivialUnswitch` is | ||||||||||
2984 | /// true, we will attempt to do non-trivial unswitching as well as trivial | ||||||||||
2985 | /// unswitching. | ||||||||||
2986 | /// | ||||||||||
2987 | /// The `UnswitchCB` callback provided will be run after unswitching is | ||||||||||
2988 | /// complete, with the first parameter set to `true` if the provided loop | ||||||||||
2989 | /// remains a loop, and a list of new sibling loops created. | ||||||||||
2990 | /// | ||||||||||
2991 | /// If `SE` is non-null, we will update that analysis based on the unswitching | ||||||||||
2992 | /// done. | ||||||||||
2993 | static bool | ||||||||||
2994 | unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC, | ||||||||||
2995 | AAResults &AA, TargetTransformInfo &TTI, bool Trivial, | ||||||||||
2996 | bool NonTrivial, | ||||||||||
2997 | function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB, | ||||||||||
2998 | ScalarEvolution *SE, MemorySSAUpdater *MSSAU, | ||||||||||
2999 | function_ref<void(Loop &, StringRef)> DestroyLoopCB) { | ||||||||||
3000 | assert(L.isRecursivelyLCSSAForm(DT, LI) &&((void)0) | ||||||||||
3001 | "Loops must be in LCSSA form before unswitching.")((void)0); | ||||||||||
3002 | |||||||||||
3003 | // Must be in loop simplified form: we need a preheader and dedicated exits. | ||||||||||
3004 | if (!L.isLoopSimplifyForm()) | ||||||||||
3005 | return false; | ||||||||||
3006 | |||||||||||
3007 | // Try trivial unswitch first before loop over other basic blocks in the loop. | ||||||||||
3008 | if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) { | ||||||||||
3009 | // If we unswitched successfully we will want to clean up the loop before | ||||||||||
3010 | // processing it further so just mark it as unswitched and return. | ||||||||||
3011 | UnswitchCB(/*CurrentLoopValid*/ true, false, {}); | ||||||||||
3012 | return true; | ||||||||||
3013 | } | ||||||||||
3014 | |||||||||||
3015 | // Check whether we should continue with non-trivial conditions. | ||||||||||
3016 | // EnableNonTrivialUnswitch: Global variable that forces non-trivial | ||||||||||
3017 | // unswitching for testing and debugging. | ||||||||||
3018 | // NonTrivial: Parameter that enables non-trivial unswitching for this | ||||||||||
3019 | // invocation of the transform. But this should be allowed only | ||||||||||
3020 | // for targets without branch divergence. | ||||||||||
3021 | // | ||||||||||
3022 | // FIXME: If divergence analysis becomes available to a loop | ||||||||||
3023 | // transform, we should allow unswitching for non-trivial uniform | ||||||||||
3024 | // branches even on targets that have divergence. | ||||||||||
3025 | // https://bugs.llvm.org/show_bug.cgi?id=48819 | ||||||||||
3026 | bool ContinueWithNonTrivial = | ||||||||||
3027 | EnableNonTrivialUnswitch || (NonTrivial && !TTI.hasBranchDivergence()); | ||||||||||
3028 | if (!ContinueWithNonTrivial) | ||||||||||
3029 | return false; | ||||||||||
3030 | |||||||||||
3031 | // Skip non-trivial unswitching for optsize functions. | ||||||||||
3032 | if (L.getHeader()->getParent()->hasOptSize()) | ||||||||||
3033 | return false; | ||||||||||
3034 | |||||||||||
3035 | // Skip non-trivial unswitching for loops that cannot be cloned. | ||||||||||
3036 | if (!L.isSafeToClone()) | ||||||||||
3037 | return false; | ||||||||||
3038 | |||||||||||
3039 | // For non-trivial unswitching, because it often creates new loops, we rely on | ||||||||||
3040 | // the pass manager to iterate on the loops rather than trying to immediately | ||||||||||
3041 | // reach a fixed point. There is no substantial advantage to iterating | ||||||||||
3042 | // internally, and if any of the new loops are simplified enough to contain | ||||||||||
3043 | // trivial unswitching we want to prefer those. | ||||||||||
3044 | |||||||||||
3045 | // Try to unswitch the best invariant condition. We prefer this full unswitch to | ||||||||||
3046 | // a partial unswitch when possible below the threshold. | ||||||||||
3047 | if (unswitchBestCondition(L, DT, LI, AC, AA, TTI, UnswitchCB, SE, MSSAU, | ||||||||||
3048 | DestroyLoopCB)) | ||||||||||
3049 | return true; | ||||||||||
3050 | |||||||||||
3051 | // No other opportunities to unswitch. | ||||||||||
3052 | return false; | ||||||||||
3053 | } | ||||||||||
3054 | |||||||||||
3055 | PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM, | ||||||||||
3056 | LoopStandardAnalysisResults &AR, | ||||||||||
3057 | LPMUpdater &U) { | ||||||||||
3058 | Function &F = *L.getHeader()->getParent(); | ||||||||||
3059 | (void)F; | ||||||||||
3060 | |||||||||||
3061 | LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << Ldo { } while (false) | ||||||||||
3062 | << "\n")do { } while (false); | ||||||||||
3063 | |||||||||||
3064 | // Save the current loop name in a variable so that we can report it even | ||||||||||
3065 | // after it has been deleted. | ||||||||||
3066 | std::string LoopName = std::string(L.getName()); | ||||||||||
3067 | |||||||||||
3068 | auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid, | ||||||||||
3069 | bool PartiallyInvariant, | ||||||||||
3070 | ArrayRef<Loop *> NewLoops) { | ||||||||||
3071 | // If we did a non-trivial unswitch, we have added new (cloned) loops. | ||||||||||
3072 | if (!NewLoops.empty()) | ||||||||||
3073 | U.addSiblingLoops(NewLoops); | ||||||||||
3074 | |||||||||||
3075 | // If the current loop remains valid, we should revisit it to catch any | ||||||||||
3076 | // other unswitch opportunities. Otherwise, we need to mark it as deleted. | ||||||||||
3077 | if (CurrentLoopValid) { | ||||||||||
3078 | if (PartiallyInvariant) { | ||||||||||
3079 | // Mark the new loop as partially unswitched, to avoid unswitching on | ||||||||||
3080 | // the same condition again. | ||||||||||
3081 | auto &Context = L.getHeader()->getContext(); | ||||||||||
3082 | MDNode *DisableUnswitchMD = MDNode::get( | ||||||||||
3083 | Context, | ||||||||||
3084 | MDString::get(Context, "llvm.loop.unswitch.partial.disable")); | ||||||||||
3085 | MDNode *NewLoopID = makePostTransformationMetadata( | ||||||||||
3086 | Context, L.getLoopID(), {"llvm.loop.unswitch.partial"}, | ||||||||||
3087 | {DisableUnswitchMD}); | ||||||||||
3088 | L.setLoopID(NewLoopID); | ||||||||||
3089 | } else | ||||||||||
3090 | U.revisitCurrentLoop(); | ||||||||||
3091 | } else | ||||||||||
3092 | U.markLoopAsDeleted(L, LoopName); | ||||||||||
3093 | }; | ||||||||||
3094 | |||||||||||
3095 | auto DestroyLoopCB = [&U](Loop &L, StringRef Name) { | ||||||||||
3096 | U.markLoopAsDeleted(L, Name); | ||||||||||
3097 | }; | ||||||||||
3098 | |||||||||||
3099 | Optional<MemorySSAUpdater> MSSAU; | ||||||||||
3100 | if (AR.MSSA) { | ||||||||||
3101 | MSSAU = MemorySSAUpdater(AR.MSSA); | ||||||||||
3102 | if (VerifyMemorySSA) | ||||||||||
3103 | AR.MSSA->verifyMemorySSA(); | ||||||||||
3104 | } | ||||||||||
3105 | if (!unswitchLoop(L, AR.DT, AR.LI, AR.AC, AR.AA, AR.TTI, Trivial, NonTrivial, | ||||||||||
3106 | UnswitchCB, &AR.SE, | ||||||||||
3107 | MSSAU.hasValue() ? MSSAU.getPointer() : nullptr, | ||||||||||
3108 | DestroyLoopCB)) | ||||||||||
3109 | return PreservedAnalyses::all(); | ||||||||||
3110 | |||||||||||
3111 | if (AR.MSSA && VerifyMemorySSA) | ||||||||||
3112 | AR.MSSA->verifyMemorySSA(); | ||||||||||
3113 | |||||||||||
3114 | // Historically this pass has had issues with the dominator tree so verify it | ||||||||||
3115 | // in asserts builds. | ||||||||||
3116 | assert(AR.DT.verify(DominatorTree::VerificationLevel::Fast))((void)0); | ||||||||||
3117 | |||||||||||
3118 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||||
3119 | if (AR.MSSA) | ||||||||||
3120 | PA.preserve<MemorySSAAnalysis>(); | ||||||||||
3121 | return PA; | ||||||||||
3122 | } | ||||||||||
3123 | |||||||||||
3124 | namespace { | ||||||||||
3125 | |||||||||||
3126 | class SimpleLoopUnswitchLegacyPass : public LoopPass { | ||||||||||
3127 | bool NonTrivial; | ||||||||||
3128 | |||||||||||
3129 | public: | ||||||||||
3130 | static char ID; // Pass ID, replacement for typeid | ||||||||||
3131 | |||||||||||
3132 | explicit SimpleLoopUnswitchLegacyPass(bool NonTrivial = false) | ||||||||||
3133 | : LoopPass(ID), NonTrivial(NonTrivial) { | ||||||||||
3134 | initializeSimpleLoopUnswitchLegacyPassPass( | ||||||||||
3135 | *PassRegistry::getPassRegistry()); | ||||||||||
3136 | } | ||||||||||
3137 | |||||||||||
3138 | bool runOnLoop(Loop *L, LPPassManager &LPM) override; | ||||||||||
3139 | |||||||||||
3140 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||||
3141 | AU.addRequired<AssumptionCacheTracker>(); | ||||||||||
3142 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||||
3143 | if (EnableMSSALoopDependency) { | ||||||||||
3144 | AU.addRequired<MemorySSAWrapperPass>(); | ||||||||||
3145 | AU.addPreserved<MemorySSAWrapperPass>(); | ||||||||||
3146 | } | ||||||||||
3147 | getLoopAnalysisUsage(AU); | ||||||||||
3148 | } | ||||||||||
3149 | }; | ||||||||||
3150 | |||||||||||
3151 | } // end anonymous namespace | ||||||||||
3152 | |||||||||||
3153 | bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { | ||||||||||
3154 | if (skipLoop(L)) | ||||||||||
3155 | return false; | ||||||||||
3156 | |||||||||||
3157 | Function &F = *L->getHeader()->getParent(); | ||||||||||
3158 | |||||||||||
3159 | LLVM_DEBUG(dbgs() << "Unswitching loop in " << F.getName() << ": " << *Ldo { } while (false) | ||||||||||
3160 | << "\n")do { } while (false); | ||||||||||
3161 | |||||||||||
3162 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||||||
3163 | auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||||||
3164 | auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||||||
3165 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||||||||
3166 | auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||||||||
3167 | MemorySSA *MSSA = nullptr; | ||||||||||
3168 | Optional<MemorySSAUpdater> MSSAU; | ||||||||||
3169 | if (EnableMSSALoopDependency) { | ||||||||||
3170 | MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA(); | ||||||||||
3171 | MSSAU = MemorySSAUpdater(MSSA); | ||||||||||
3172 | } | ||||||||||
3173 | |||||||||||
3174 | auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | ||||||||||
3175 | auto *SE = SEWP ? &SEWP->getSE() : nullptr; | ||||||||||
3176 | |||||||||||
3177 | auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid, bool PartiallyInvariant, | ||||||||||
3178 | ArrayRef<Loop *> NewLoops) { | ||||||||||
3179 | // If we did a non-trivial unswitch, we have added new (cloned) loops. | ||||||||||
3180 | for (auto *NewL : NewLoops) | ||||||||||
3181 | LPM.addLoop(*NewL); | ||||||||||
3182 | |||||||||||
3183 | // If the current loop remains valid, re-add it to the queue. This is | ||||||||||
3184 | // a little wasteful as we'll finish processing the current loop as well, | ||||||||||
3185 | // but it is the best we can do in the old PM. | ||||||||||
3186 | if (CurrentLoopValid) { | ||||||||||
3187 | // If the current loop has been unswitched using a partially invariant | ||||||||||
3188 | // condition, we should not re-add the current loop to avoid unswitching | ||||||||||
3189 | // on the same condition again. | ||||||||||
3190 | if (!PartiallyInvariant) | ||||||||||
3191 | LPM.addLoop(*L); | ||||||||||
3192 | } else | ||||||||||
3193 | LPM.markLoopAsDeleted(*L); | ||||||||||
3194 | }; | ||||||||||
3195 | |||||||||||
3196 | auto DestroyLoopCB = [&LPM](Loop &L, StringRef /* Name */) { | ||||||||||
3197 | LPM.markLoopAsDeleted(L); | ||||||||||
3198 | }; | ||||||||||
3199 | |||||||||||
3200 | if (MSSA && VerifyMemorySSA) | ||||||||||
3201 | MSSA->verifyMemorySSA(); | ||||||||||
3202 | |||||||||||
3203 | bool Changed = | ||||||||||
3204 | unswitchLoop(*L, DT, LI, AC, AA, TTI, true, NonTrivial, UnswitchCB, SE, | ||||||||||
3205 | MSSAU.hasValue() ? MSSAU.getPointer() : nullptr, | ||||||||||
3206 | DestroyLoopCB); | ||||||||||
3207 | |||||||||||
3208 | if (MSSA && VerifyMemorySSA) | ||||||||||
3209 | MSSA->verifyMemorySSA(); | ||||||||||
3210 | |||||||||||
3211 | // Historically this pass has had issues with the dominator tree so verify it | ||||||||||
3212 | // in asserts builds. | ||||||||||
3213 | assert(DT.verify(DominatorTree::VerificationLevel::Fast))((void)0); | ||||||||||
3214 | |||||||||||
3215 | return Changed; | ||||||||||
3216 | } | ||||||||||
3217 | |||||||||||
3218 | char SimpleLoopUnswitchLegacyPass::ID = 0; | ||||||||||
3219 | INITIALIZE_PASS_BEGIN(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",static void *initializeSimpleLoopUnswitchLegacyPassPassOnce(PassRegistry &Registry) { | ||||||||||
3220 | "Simple unswitch loops", false, false)static void *initializeSimpleLoopUnswitchLegacyPassPassOnce(PassRegistry &Registry) { | ||||||||||
3221 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||||||
3222 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||||||
3223 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||||||
3224 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | ||||||||||
3225 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | ||||||||||
3226 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||||
3227 | INITIALIZE_PASS_END(SimpleLoopUnswitchLegacyPass, "simple-loop-unswitch",PassInfo *PI = new PassInfo( "Simple unswitch loops", "simple-loop-unswitch" , &SimpleLoopUnswitchLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<SimpleLoopUnswitchLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSimpleLoopUnswitchLegacyPassPassFlag ; void llvm::initializeSimpleLoopUnswitchLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeSimpleLoopUnswitchLegacyPassPassFlag , initializeSimpleLoopUnswitchLegacyPassPassOnce, std::ref(Registry )); } | ||||||||||
3228 | "Simple unswitch loops", false, false)PassInfo *PI = new PassInfo( "Simple unswitch loops", "simple-loop-unswitch" , &SimpleLoopUnswitchLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<SimpleLoopUnswitchLegacyPass>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSimpleLoopUnswitchLegacyPassPassFlag ; void llvm::initializeSimpleLoopUnswitchLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeSimpleLoopUnswitchLegacyPassPassFlag , initializeSimpleLoopUnswitchLegacyPassPassOnce, std::ref(Registry )); } | ||||||||||
3229 | |||||||||||
3230 | Pass *llvm::createSimpleLoopUnswitchLegacyPass(bool NonTrivial) { | ||||||||||
3231 | return new SimpleLoopUnswitchLegacyPass(NonTrivial); | ||||||||||
3232 | } |
1 | //===- llvm/Transforms/Utils/LoopUtils.h - Loop utilities -------*- 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 some loop transformation utilities. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_TRANSFORMS_UTILS_LOOPUTILS_H |
14 | #define LLVM_TRANSFORMS_UTILS_LOOPUTILS_H |
15 | |
16 | #include "llvm/ADT/StringRef.h" |
17 | #include "llvm/Analysis/IVDescriptors.h" |
18 | #include "llvm/Analysis/TargetTransformInfo.h" |
19 | #include "llvm/Transforms/Utils/ValueMapper.h" |
20 | |
21 | namespace llvm { |
22 | |
23 | template <typename T> class DomTreeNodeBase; |
24 | using DomTreeNode = DomTreeNodeBase<BasicBlock>; |
25 | class AAResults; |
26 | class AliasSet; |
27 | class AliasSetTracker; |
28 | class BasicBlock; |
29 | class BlockFrequencyInfo; |
30 | class ICFLoopSafetyInfo; |
31 | class IRBuilderBase; |
32 | class Loop; |
33 | class LoopInfo; |
34 | class MemoryAccess; |
35 | class MemorySSA; |
36 | class MemorySSAUpdater; |
37 | class OptimizationRemarkEmitter; |
38 | class PredIteratorCache; |
39 | class ScalarEvolution; |
40 | class ScalarEvolutionExpander; |
41 | class SCEV; |
42 | class SCEVExpander; |
43 | class TargetLibraryInfo; |
44 | class LPPassManager; |
45 | class Instruction; |
46 | struct RuntimeCheckingPtrGroup; |
47 | typedef std::pair<const RuntimeCheckingPtrGroup *, |
48 | const RuntimeCheckingPtrGroup *> |
49 | RuntimePointerCheck; |
50 | |
51 | template <typename T> class Optional; |
52 | template <typename T, unsigned N> class SmallSetVector; |
53 | template <typename T, unsigned N> class SmallVector; |
54 | template <typename T> class SmallVectorImpl; |
55 | template <typename T, unsigned N> class SmallPriorityWorklist; |
56 | |
57 | BasicBlock *InsertPreheaderForLoop(Loop *L, DominatorTree *DT, LoopInfo *LI, |
58 | MemorySSAUpdater *MSSAU, bool PreserveLCSSA); |
59 | |
60 | /// Ensure that all exit blocks of the loop are dedicated exits. |
61 | /// |
62 | /// For any loop exit block with non-loop predecessors, we split the loop |
63 | /// predecessors to use a dedicated loop exit block. We update the dominator |
64 | /// tree and loop info if provided, and will preserve LCSSA if requested. |
65 | bool formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, |
66 | MemorySSAUpdater *MSSAU, bool PreserveLCSSA); |
67 | |
68 | /// Ensures LCSSA form for every instruction from the Worklist in the scope of |
69 | /// innermost containing loop. |
70 | /// |
71 | /// For the given instruction which have uses outside of the loop, an LCSSA PHI |
72 | /// node is inserted and the uses outside the loop are rewritten to use this |
73 | /// node. |
74 | /// |
75 | /// LoopInfo and DominatorTree are required and, since the routine makes no |
76 | /// changes to CFG, preserved. |
77 | /// |
78 | /// Returns true if any modifications are made. |
79 | /// |
80 | /// This function may introduce unused PHI nodes. If \p PHIsToRemove is not |
81 | /// nullptr, those are added to it (before removing, the caller has to check if |
82 | /// they still do not have any uses). Otherwise the PHIs are directly removed. |
83 | bool formLCSSAForInstructions( |
84 | SmallVectorImpl<Instruction *> &Worklist, const DominatorTree &DT, |
85 | const LoopInfo &LI, ScalarEvolution *SE, IRBuilderBase &Builder, |
86 | SmallVectorImpl<PHINode *> *PHIsToRemove = nullptr); |
87 | |
88 | /// Put loop into LCSSA form. |
89 | /// |
90 | /// Looks at all instructions in the loop which have uses outside of the |
91 | /// current loop. For each, an LCSSA PHI node is inserted and the uses outside |
92 | /// the loop are rewritten to use this node. Sub-loops must be in LCSSA form |
93 | /// already. |
94 | /// |
95 | /// LoopInfo and DominatorTree are required and preserved. |
96 | /// |
97 | /// If ScalarEvolution is passed in, it will be preserved. |
98 | /// |
99 | /// Returns true if any modifications are made to the loop. |
100 | bool formLCSSA(Loop &L, const DominatorTree &DT, const LoopInfo *LI, |
101 | ScalarEvolution *SE); |
102 | |
103 | /// Put a loop nest into LCSSA form. |
104 | /// |
105 | /// This recursively forms LCSSA for a loop nest. |
106 | /// |
107 | /// LoopInfo and DominatorTree are required and preserved. |
108 | /// |
109 | /// If ScalarEvolution is passed in, it will be preserved. |
110 | /// |
111 | /// Returns true if any modifications are made to the loop. |
112 | bool formLCSSARecursively(Loop &L, const DominatorTree &DT, const LoopInfo *LI, |
113 | ScalarEvolution *SE); |
114 | |
115 | /// Flags controlling how much is checked when sinking or hoisting |
116 | /// instructions. The number of memory access in the loop (and whether there |
117 | /// are too many) is determined in the constructors when using MemorySSA. |
118 | class SinkAndHoistLICMFlags { |
119 | public: |
120 | // Explicitly set limits. |
121 | SinkAndHoistLICMFlags(unsigned LicmMssaOptCap, |
122 | unsigned LicmMssaNoAccForPromotionCap, bool IsSink, |
123 | Loop *L = nullptr, MemorySSA *MSSA = nullptr); |
124 | // Use default limits. |
125 | SinkAndHoistLICMFlags(bool IsSink, Loop *L = nullptr, |
126 | MemorySSA *MSSA = nullptr); |
127 | |
128 | void setIsSink(bool B) { IsSink = B; } |
129 | bool getIsSink() { return IsSink; } |
130 | bool tooManyMemoryAccesses() { return NoOfMemAccTooLarge; } |
131 | bool tooManyClobberingCalls() { return LicmMssaOptCounter >= LicmMssaOptCap; } |
132 | void incrementClobberingCalls() { ++LicmMssaOptCounter; } |
133 | |
134 | protected: |
135 | bool NoOfMemAccTooLarge = false; |
136 | unsigned LicmMssaOptCounter = 0; |
137 | unsigned LicmMssaOptCap; |
138 | unsigned LicmMssaNoAccForPromotionCap; |
139 | bool IsSink; |
140 | }; |
141 | |
142 | /// Walk the specified region of the CFG (defined by all blocks |
143 | /// dominated by the specified block, and that are in the current loop) in |
144 | /// reverse depth first order w.r.t the DominatorTree. This allows us to visit |
145 | /// uses before definitions, allowing us to sink a loop body in one pass without |
146 | /// iteration. Takes DomTreeNode, AAResults, LoopInfo, DominatorTree, |
147 | /// BlockFrequencyInfo, TargetLibraryInfo, Loop, AliasSet information for all |
148 | /// instructions of the loop and loop safety information as |
149 | /// arguments. Diagnostics is emitted via \p ORE. It returns changed status. |
150 | bool sinkRegion(DomTreeNode *, AAResults *, LoopInfo *, DominatorTree *, |
151 | BlockFrequencyInfo *, TargetLibraryInfo *, |
152 | TargetTransformInfo *, Loop *, AliasSetTracker *, |
153 | MemorySSAUpdater *, ICFLoopSafetyInfo *, |
154 | SinkAndHoistLICMFlags &, OptimizationRemarkEmitter *); |
155 | |
156 | /// Walk the specified region of the CFG (defined by all blocks |
157 | /// dominated by the specified block, and that are in the current loop) in depth |
158 | /// first order w.r.t the DominatorTree. This allows us to visit definitions |
159 | /// before uses, allowing us to hoist a loop body in one pass without iteration. |
160 | /// Takes DomTreeNode, AAResults, LoopInfo, DominatorTree, |
161 | /// BlockFrequencyInfo, TargetLibraryInfo, Loop, AliasSet information for all |
162 | /// instructions of the loop and loop safety information as arguments. |
163 | /// Diagnostics is emitted via \p ORE. It returns changed status. |
164 | bool hoistRegion(DomTreeNode *, AAResults *, LoopInfo *, DominatorTree *, |
165 | BlockFrequencyInfo *, TargetLibraryInfo *, Loop *, |
166 | AliasSetTracker *, MemorySSAUpdater *, ScalarEvolution *, |
167 | ICFLoopSafetyInfo *, SinkAndHoistLICMFlags &, |
168 | OptimizationRemarkEmitter *, bool); |
169 | |
170 | /// This function deletes dead loops. The caller of this function needs to |
171 | /// guarantee that the loop is infact dead. |
172 | /// The function requires a bunch or prerequisites to be present: |
173 | /// - The loop needs to be in LCSSA form |
174 | /// - The loop needs to have a Preheader |
175 | /// - A unique dedicated exit block must exist |
176 | /// |
177 | /// This also updates the relevant analysis information in \p DT, \p SE, \p LI |
178 | /// and \p MSSA if pointers to those are provided. |
179 | /// It also updates the loop PM if an updater struct is provided. |
180 | |
181 | void deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE, |
182 | LoopInfo *LI, MemorySSA *MSSA = nullptr); |
183 | |
184 | /// Remove the backedge of the specified loop. Handles loop nests and general |
185 | /// loop structures subject to the precondition that the loop has no parent |
186 | /// loop and has a single latch block. Preserves all listed analyses. |
187 | void breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE, |
188 | LoopInfo &LI, MemorySSA *MSSA); |
189 | |
190 | /// Try to promote memory values to scalars by sinking stores out of |
191 | /// the loop and moving loads to before the loop. We do this by looping over |
192 | /// the stores in the loop, looking for stores to Must pointers which are |
193 | /// loop invariant. It takes a set of must-alias values, Loop exit blocks |
194 | /// vector, loop exit blocks insertion point vector, PredIteratorCache, |
195 | /// LoopInfo, DominatorTree, Loop, AliasSet information for all instructions |
196 | /// of the loop and loop safety information as arguments. |
197 | /// Diagnostics is emitted via \p ORE. It returns changed status. |
198 | bool promoteLoopAccessesToScalars( |
199 | const SmallSetVector<Value *, 8> &, SmallVectorImpl<BasicBlock *> &, |
200 | SmallVectorImpl<Instruction *> &, SmallVectorImpl<MemoryAccess *> &, |
201 | PredIteratorCache &, LoopInfo *, DominatorTree *, const TargetLibraryInfo *, |
202 | Loop *, AliasSetTracker *, MemorySSAUpdater *, ICFLoopSafetyInfo *, |
203 | OptimizationRemarkEmitter *); |
204 | |
205 | /// Does a BFS from a given node to all of its children inside a given loop. |
206 | /// The returned vector of nodes includes the starting point. |
207 | SmallVector<DomTreeNode *, 16> collectChildrenInLoop(DomTreeNode *N, |
208 | const Loop *CurLoop); |
209 | |
210 | /// Returns the instructions that use values defined in the loop. |
211 | SmallVector<Instruction *, 8> findDefsUsedOutsideOfLoop(Loop *L); |
212 | |
213 | /// Find a combination of metadata ("llvm.loop.vectorize.width" and |
214 | /// "llvm.loop.vectorize.scalable.enable") for a loop and use it to construct a |
215 | /// ElementCount. If the metadata "llvm.loop.vectorize.width" cannot be found |
216 | /// then None is returned. |
217 | Optional<ElementCount> |
218 | getOptionalElementCountLoopAttribute(const Loop *TheLoop); |
219 | |
220 | /// Create a new loop identifier for a loop created from a loop transformation. |
221 | /// |
222 | /// @param OrigLoopID The loop ID of the loop before the transformation. |
223 | /// @param FollowupAttrs List of attribute names that contain attributes to be |
224 | /// added to the new loop ID. |
225 | /// @param InheritOptionsAttrsPrefix Selects which attributes should be inherited |
226 | /// from the original loop. The following values |
227 | /// are considered: |
228 | /// nullptr : Inherit all attributes from @p OrigLoopID. |
229 | /// "" : Do not inherit any attribute from @p OrigLoopID; only use |
230 | /// those specified by a followup attribute. |
231 | /// "<prefix>": Inherit all attributes except those which start with |
232 | /// <prefix>; commonly used to remove metadata for the |
233 | /// applied transformation. |
234 | /// @param AlwaysNew If true, do not try to reuse OrigLoopID and never return |
235 | /// None. |
236 | /// |
237 | /// @return The loop ID for the after-transformation loop. The following values |
238 | /// can be returned: |
239 | /// None : No followup attribute was found; it is up to the |
240 | /// transformation to choose attributes that make sense. |
241 | /// @p OrigLoopID: The original identifier can be reused. |
242 | /// nullptr : The new loop has no attributes. |
243 | /// MDNode* : A new unique loop identifier. |
244 | Optional<MDNode *> |
245 | makeFollowupLoopID(MDNode *OrigLoopID, ArrayRef<StringRef> FollowupAttrs, |
246 | const char *InheritOptionsAttrsPrefix = "", |
247 | bool AlwaysNew = false); |
248 | |
249 | /// Look for the loop attribute that disables all transformation heuristic. |
250 | bool hasDisableAllTransformsHint(const Loop *L); |
251 | |
252 | /// Look for the loop attribute that disables the LICM transformation heuristics. |
253 | bool hasDisableLICMTransformsHint(const Loop *L); |
254 | |
255 | /// The mode sets how eager a transformation should be applied. |
256 | enum TransformationMode { |
257 | /// The pass can use heuristics to determine whether a transformation should |
258 | /// be applied. |
259 | TM_Unspecified, |
260 | |
261 | /// The transformation should be applied without considering a cost model. |
262 | TM_Enable, |
263 | |
264 | /// The transformation should not be applied. |
265 | TM_Disable, |
266 | |
267 | /// Force is a flag and should not be used alone. |
268 | TM_Force = 0x04, |
269 | |
270 | /// The transformation was directed by the user, e.g. by a #pragma in |
271 | /// the source code. If the transformation could not be applied, a |
272 | /// warning should be emitted. |
273 | TM_ForcedByUser = TM_Enable | TM_Force, |
274 | |
275 | /// The transformation must not be applied. For instance, `#pragma clang loop |
276 | /// unroll(disable)` explicitly forbids any unrolling to take place. Unlike |
277 | /// general loop metadata, it must not be dropped. Most passes should not |
278 | /// behave differently under TM_Disable and TM_SuppressedByUser. |
279 | TM_SuppressedByUser = TM_Disable | TM_Force |
280 | }; |
281 | |
282 | /// @{ |
283 | /// Get the mode for LLVM's supported loop transformations. |
284 | TransformationMode hasUnrollTransformation(const Loop *L); |
285 | TransformationMode hasUnrollAndJamTransformation(const Loop *L); |
286 | TransformationMode hasVectorizeTransformation(const Loop *L); |
287 | TransformationMode hasDistributeTransformation(const Loop *L); |
288 | TransformationMode hasLICMVersioningTransformation(const Loop *L); |
289 | /// @} |
290 | |
291 | /// Set input string into loop metadata by keeping other values intact. |
292 | /// If the string is already in loop metadata update value if it is |
293 | /// different. |
294 | void addStringMetadataToLoop(Loop *TheLoop, const char *MDString, |
295 | unsigned V = 0); |
296 | |
297 | /// Returns a loop's estimated trip count based on branch weight metadata. |
298 | /// In addition if \p EstimatedLoopInvocationWeight is not null it is |
299 | /// initialized with weight of loop's latch leading to the exit. |
300 | /// Returns 0 when the count is estimated to be 0, or None when a meaningful |
301 | /// estimate can not be made. |
302 | Optional<unsigned> |
303 | getLoopEstimatedTripCount(Loop *L, |
304 | unsigned *EstimatedLoopInvocationWeight = nullptr); |
305 | |
306 | /// Set a loop's branch weight metadata to reflect that loop has \p |
307 | /// EstimatedTripCount iterations and \p EstimatedLoopInvocationWeight exits |
308 | /// through latch. Returns true if metadata is successfully updated, false |
309 | /// otherwise. Note that loop must have a latch block which controls loop exit |
310 | /// in order to succeed. |
311 | bool setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount, |
312 | unsigned EstimatedLoopInvocationWeight); |
313 | |
314 | /// Check inner loop (L) backedge count is known to be invariant on all |
315 | /// iterations of its outer loop. If the loop has no parent, this is trivially |
316 | /// true. |
317 | bool hasIterationCountInvariantInParent(Loop *L, ScalarEvolution &SE); |
318 | |
319 | /// Helper to consistently add the set of standard passes to a loop pass's \c |
320 | /// AnalysisUsage. |
321 | /// |
322 | /// All loop passes should call this as part of implementing their \c |
323 | /// getAnalysisUsage. |
324 | void getLoopAnalysisUsage(AnalysisUsage &AU); |
325 | |
326 | /// Returns true if is legal to hoist or sink this instruction disregarding the |
327 | /// possible introduction of faults. Reasoning about potential faulting |
328 | /// instructions is the responsibility of the caller since it is challenging to |
329 | /// do efficiently from within this routine. |
330 | /// \p TargetExecutesOncePerLoop is true only when it is guaranteed that the |
331 | /// target executes at most once per execution of the loop body. This is used |
332 | /// to assess the legality of duplicating atomic loads. Generally, this is |
333 | /// true when moving out of loop and not true when moving into loops. |
334 | /// If \p ORE is set use it to emit optimization remarks. |
335 | bool canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, |
336 | Loop *CurLoop, AliasSetTracker *CurAST, |
337 | MemorySSAUpdater *MSSAU, bool TargetExecutesOncePerLoop, |
338 | SinkAndHoistLICMFlags *LICMFlags = nullptr, |
339 | OptimizationRemarkEmitter *ORE = nullptr); |
340 | |
341 | /// Returns a Min/Max operation corresponding to MinMaxRecurrenceKind. |
342 | /// The Builder's fast-math-flags must be set to propagate the expected values. |
343 | Value *createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left, |
344 | Value *Right); |
345 | |
346 | /// Generates an ordered vector reduction using extracts to reduce the value. |
347 | Value *getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src, |
348 | unsigned Op, RecurKind MinMaxKind = RecurKind::None, |
349 | ArrayRef<Value *> RedOps = None); |
350 | |
351 | /// Generates a vector reduction using shufflevectors to reduce the value. |
352 | /// Fast-math-flags are propagated using the IRBuilder's setting. |
353 | Value *getShuffleReduction(IRBuilderBase &Builder, Value *Src, unsigned Op, |
354 | RecurKind MinMaxKind = RecurKind::None, |
355 | ArrayRef<Value *> RedOps = None); |
356 | |
357 | /// Create a target reduction of the given vector. The reduction operation |
358 | /// is described by the \p Opcode parameter. min/max reductions require |
359 | /// additional information supplied in \p RdxKind. |
360 | /// The target is queried to determine if intrinsics or shuffle sequences are |
361 | /// required to implement the reduction. |
362 | /// Fast-math-flags are propagated using the IRBuilder's setting. |
363 | Value *createSimpleTargetReduction(IRBuilderBase &B, |
364 | const TargetTransformInfo *TTI, Value *Src, |
365 | RecurKind RdxKind, |
366 | ArrayRef<Value *> RedOps = None); |
367 | |
368 | /// Create a generic target reduction using a recurrence descriptor \p Desc |
369 | /// The target is queried to determine if intrinsics or shuffle sequences are |
370 | /// required to implement the reduction. |
371 | /// Fast-math-flags are propagated using the RecurrenceDescriptor. |
372 | Value *createTargetReduction(IRBuilderBase &B, const TargetTransformInfo *TTI, |
373 | const RecurrenceDescriptor &Desc, Value *Src); |
374 | |
375 | /// Create an ordered reduction intrinsic using the given recurrence |
376 | /// descriptor \p Desc. |
377 | Value *createOrderedReduction(IRBuilderBase &B, |
378 | const RecurrenceDescriptor &Desc, Value *Src, |
379 | Value *Start); |
380 | |
381 | /// Get the intersection (logical and) of all of the potential IR flags |
382 | /// of each scalar operation (VL) that will be converted into a vector (I). |
383 | /// If OpValue is non-null, we only consider operations similar to OpValue |
384 | /// when intersecting. |
385 | /// Flag set: NSW, NUW, exact, and all of fast-math. |
386 | void propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue = nullptr); |
387 | |
388 | /// Returns true if we can prove that \p S is defined and always negative in |
389 | /// loop \p L. |
390 | bool isKnownNegativeInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE); |
391 | |
392 | /// Returns true if we can prove that \p S is defined and always non-negative in |
393 | /// loop \p L. |
394 | bool isKnownNonNegativeInLoop(const SCEV *S, const Loop *L, |
395 | ScalarEvolution &SE); |
396 | |
397 | /// Returns true if \p S is defined and never is equal to signed/unsigned max. |
398 | bool cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, |
399 | bool Signed); |
400 | |
401 | /// Returns true if \p S is defined and never is equal to signed/unsigned min. |
402 | bool cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE, |
403 | bool Signed); |
404 | |
405 | enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, NoHardUse, AlwaysRepl }; |
406 | |
407 | /// If the final value of any expressions that are recurrent in the loop can |
408 | /// be computed, substitute the exit values from the loop into any instructions |
409 | /// outside of the loop that use the final values of the current expressions. |
410 | /// Return the number of loop exit values that have been replaced, and the |
411 | /// corresponding phi node will be added to DeadInsts. |
412 | int rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI, |
413 | ScalarEvolution *SE, const TargetTransformInfo *TTI, |
414 | SCEVExpander &Rewriter, DominatorTree *DT, |
415 | ReplaceExitVal ReplaceExitValue, |
416 | SmallVector<WeakTrackingVH, 16> &DeadInsts); |
417 | |
418 | /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for |
419 | /// \p OrigLoop and the following distribution of \p OrigLoop iteration among \p |
420 | /// UnrolledLoop and \p RemainderLoop. \p UnrolledLoop receives weights that |
421 | /// reflect TC/UF iterations, and \p RemainderLoop receives weights that reflect |
422 | /// the remaining TC%UF iterations. |
423 | /// |
424 | /// Note that \p OrigLoop may be equal to either \p UnrolledLoop or \p |
425 | /// RemainderLoop in which case weights for \p OrigLoop are updated accordingly. |
426 | /// Note also behavior is undefined if \p UnrolledLoop and \p RemainderLoop are |
427 | /// equal. \p UF must be greater than zero. |
428 | /// If \p OrigLoop has no profile info associated nothing happens. |
429 | /// |
430 | /// This utility may be useful for such optimizations as unroller and |
431 | /// vectorizer as it's typical transformation for them. |
432 | void setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop, |
433 | Loop *RemainderLoop, uint64_t UF); |
434 | |
435 | /// Utility that implements appending of loops onto a worklist given a range. |
436 | /// We want to process loops in postorder, but the worklist is a LIFO data |
437 | /// structure, so we append to it in *reverse* postorder. |
438 | /// For trees, a preorder traversal is a viable reverse postorder, so we |
439 | /// actually append using a preorder walk algorithm. |
440 | template <typename RangeT> |
441 | void appendLoopsToWorklist(RangeT &&, SmallPriorityWorklist<Loop *, 4> &); |
442 | /// Utility that implements appending of loops onto a worklist given a range. |
443 | /// It has the same behavior as appendLoopsToWorklist, but assumes the range of |
444 | /// loops has already been reversed, so it processes loops in the given order. |
445 | template <typename RangeT> |
446 | void appendReversedLoopsToWorklist(RangeT &&, |
447 | SmallPriorityWorklist<Loop *, 4> &); |
448 | |
449 | /// Utility that implements appending of loops onto a worklist given LoopInfo. |
450 | /// Calls the templated utility taking a Range of loops, handing it the Loops |
451 | /// in LoopInfo, iterated in reverse. This is because the loops are stored in |
452 | /// RPO w.r.t. the control flow graph in LoopInfo. For the purpose of unrolling, |
453 | /// loop deletion, and LICM, we largely want to work forward across the CFG so |
454 | /// that we visit defs before uses and can propagate simplifications from one |
455 | /// loop nest into the next. Calls appendReversedLoopsToWorklist with the |
456 | /// already reversed loops in LI. |
457 | /// FIXME: Consider changing the order in LoopInfo. |
458 | void appendLoopsToWorklist(LoopInfo &, SmallPriorityWorklist<Loop *, 4> &); |
459 | |
460 | /// Recursively clone the specified loop and all of its children, |
461 | /// mapping the blocks with the specified map. |
462 | Loop *cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, |
463 | LoopInfo *LI, LPPassManager *LPM); |
464 | |
465 | /// Add code that checks at runtime if the accessed arrays in \p PointerChecks |
466 | /// overlap. |
467 | /// |
468 | /// Returns a pair of instructions where the first element is the first |
469 | /// instruction generated in possibly a sequence of instructions and the |
470 | /// second value is the final comparator value or NULL if no check is needed. |
471 | std::pair<Instruction *, Instruction *> |
472 | addRuntimeChecks(Instruction *Loc, Loop *TheLoop, |
473 | const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, |
474 | SCEVExpander &Expander); |
475 | |
476 | /// Struct to hold information about a partially invariant condition. |
477 | struct IVConditionInfo { |
478 | /// Instructions that need to be duplicated and checked for the unswitching |
479 | /// condition. |
480 | SmallVector<Instruction *> InstToDuplicate; |
481 | |
482 | /// Constant to indicate for which value the condition is invariant. |
483 | Constant *KnownValue = nullptr; |
484 | |
485 | /// True if the partially invariant path is no-op (=does not have any |
486 | /// side-effects and no loop value is used outside the loop). |
487 | bool PathIsNoop = true; |
488 | |
489 | /// If the partially invariant path reaches a single exit block, ExitForPath |
490 | /// is set to that block. Otherwise it is nullptr. |
491 | BasicBlock *ExitForPath = nullptr; |
492 | }; |
493 | |
494 | /// Check if the loop header has a conditional branch that is not |
495 | /// loop-invariant, because it involves load instructions. If all paths from |
496 | /// either the true or false successor to the header or loop exists do not |
497 | /// modify the memory feeding the condition, perform 'partial unswitching'. That |
498 | /// is, duplicate the instructions feeding the condition in the pre-header. Then |
499 | /// unswitch on the duplicated condition. The condition is now known in the |
500 | /// unswitched version for the 'invariant' path through the original loop. |
501 | /// |
502 | /// If the branch condition of the header is partially invariant, return a pair |
503 | /// containing the instructions to duplicate and a boolean Constant to update |
504 | /// the condition in the loops created for the true or false successors. |
505 | Optional<IVConditionInfo> hasPartialIVCondition(Loop &L, unsigned MSSAThreshold, |
506 | MemorySSA &MSSA, AAResults &AA); |
507 | |
508 | } // end namespace llvm |
509 | |
510 | #endif // LLVM_TRANSFORMS_UTILS_LOOPUTILS_H |
1 | //===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- 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 SmallVector class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_ADT_SMALLVECTOR_H |
14 | #define LLVM_ADT_SMALLVECTOR_H |
15 | |
16 | #include "llvm/ADT/iterator_range.h" |
17 | #include "llvm/Support/Compiler.h" |
18 | #include "llvm/Support/ErrorHandling.h" |
19 | #include "llvm/Support/MemAlloc.h" |
20 | #include "llvm/Support/type_traits.h" |
21 | #include <algorithm> |
22 | #include <cassert> |
23 | #include <cstddef> |
24 | #include <cstdlib> |
25 | #include <cstring> |
26 | #include <functional> |
27 | #include <initializer_list> |
28 | #include <iterator> |
29 | #include <limits> |
30 | #include <memory> |
31 | #include <new> |
32 | #include <type_traits> |
33 | #include <utility> |
34 | |
35 | namespace llvm { |
36 | |
37 | /// This is all the stuff common to all SmallVectors. |
38 | /// |
39 | /// The template parameter specifies the type which should be used to hold the |
40 | /// Size and Capacity of the SmallVector, so it can be adjusted. |
41 | /// Using 32 bit size is desirable to shrink the size of the SmallVector. |
42 | /// Using 64 bit size is desirable for cases like SmallVector<char>, where a |
43 | /// 32 bit size would limit the vector to ~4GB. SmallVectors are used for |
44 | /// buffering bitcode output - which can exceed 4GB. |
45 | template <class Size_T> class SmallVectorBase { |
46 | protected: |
47 | void *BeginX; |
48 | Size_T Size = 0, Capacity; |
49 | |
50 | /// The maximum value of the Size_T used. |
51 | static constexpr size_t SizeTypeMax() { |
52 | return std::numeric_limits<Size_T>::max(); |
53 | } |
54 | |
55 | SmallVectorBase() = delete; |
56 | SmallVectorBase(void *FirstEl, size_t TotalCapacity) |
57 | : BeginX(FirstEl), Capacity(TotalCapacity) {} |
58 | |
59 | /// This is a helper for \a grow() that's out of line to reduce code |
60 | /// duplication. This function will report a fatal error if it can't grow at |
61 | /// least to \p MinSize. |
62 | void *mallocForGrow(size_t MinSize, size_t TSize, size_t &NewCapacity); |
63 | |
64 | /// This is an implementation of the grow() method which only works |
65 | /// on POD-like data types and is out of line to reduce code duplication. |
66 | /// This function will report a fatal error if it cannot increase capacity. |
67 | void grow_pod(void *FirstEl, size_t MinSize, size_t TSize); |
68 | |
69 | public: |
70 | size_t size() const { return Size; } |
71 | size_t capacity() const { return Capacity; } |
72 | |
73 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { return !Size; } |
74 | |
75 | /// Set the array size to \p N, which the current array must have enough |
76 | /// capacity for. |
77 | /// |
78 | /// This does not construct or destroy any elements in the vector. |
79 | /// |
80 | /// Clients can use this in conjunction with capacity() to write past the end |
81 | /// of the buffer when they know that more elements are available, and only |
82 | /// update the size later. This avoids the cost of value initializing elements |
83 | /// which will only be overwritten. |
84 | void set_size(size_t N) { |
85 | assert(N <= capacity())((void)0); |
86 | Size = N; |
87 | } |
88 | }; |
89 | |
90 | template <class T> |
91 | using SmallVectorSizeType = |
92 | typename std::conditional<sizeof(T) < 4 && sizeof(void *) >= 8, uint64_t, |
93 | uint32_t>::type; |
94 | |
95 | /// Figure out the offset of the first element. |
96 | template <class T, typename = void> struct SmallVectorAlignmentAndSize { |
97 | alignas(SmallVectorBase<SmallVectorSizeType<T>>) char Base[sizeof( |
98 | SmallVectorBase<SmallVectorSizeType<T>>)]; |
99 | alignas(T) char FirstEl[sizeof(T)]; |
100 | }; |
101 | |
102 | /// This is the part of SmallVectorTemplateBase which does not depend on whether |
103 | /// the type T is a POD. The extra dummy template argument is used by ArrayRef |
104 | /// to avoid unnecessarily requiring T to be complete. |
105 | template <typename T, typename = void> |
106 | class SmallVectorTemplateCommon |
107 | : public SmallVectorBase<SmallVectorSizeType<T>> { |
108 | using Base = SmallVectorBase<SmallVectorSizeType<T>>; |
109 | |
110 | /// Find the address of the first element. For this pointer math to be valid |
111 | /// with small-size of 0 for T with lots of alignment, it's important that |
112 | /// SmallVectorStorage is properly-aligned even for small-size of 0. |
113 | void *getFirstEl() const { |
114 | return const_cast<void *>(reinterpret_cast<const void *>( |
115 | reinterpret_cast<const char *>(this) + |
116 | offsetof(SmallVectorAlignmentAndSize<T>, FirstEl)__builtin_offsetof(SmallVectorAlignmentAndSize<T>, FirstEl ))); |
117 | } |
118 | // Space after 'FirstEl' is clobbered, do not add any instance vars after it. |
119 | |
120 | protected: |
121 | SmallVectorTemplateCommon(size_t Size) : Base(getFirstEl(), Size) {} |
122 | |
123 | void grow_pod(size_t MinSize, size_t TSize) { |
124 | Base::grow_pod(getFirstEl(), MinSize, TSize); |
125 | } |
126 | |
127 | /// Return true if this is a smallvector which has not had dynamic |
128 | /// memory allocated for it. |
129 | bool isSmall() const { return this->BeginX == getFirstEl(); } |
130 | |
131 | /// Put this vector in a state of being small. |
132 | void resetToSmall() { |
133 | this->BeginX = getFirstEl(); |
134 | this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect. |
135 | } |
136 | |
137 | /// Return true if V is an internal reference to the given range. |
138 | bool isReferenceToRange(const void *V, const void *First, const void *Last) const { |
139 | // Use std::less to avoid UB. |
140 | std::less<> LessThan; |
141 | return !LessThan(V, First) && LessThan(V, Last); |
142 | } |
143 | |
144 | /// Return true if V is an internal reference to this vector. |
145 | bool isReferenceToStorage(const void *V) const { |
146 | return isReferenceToRange(V, this->begin(), this->end()); |
147 | } |
148 | |
149 | /// Return true if First and Last form a valid (possibly empty) range in this |
150 | /// vector's storage. |
151 | bool isRangeInStorage(const void *First, const void *Last) const { |
152 | // Use std::less to avoid UB. |
153 | std::less<> LessThan; |
154 | return !LessThan(First, this->begin()) && !LessThan(Last, First) && |
155 | !LessThan(this->end(), Last); |
156 | } |
157 | |
158 | /// Return true unless Elt will be invalidated by resizing the vector to |
159 | /// NewSize. |
160 | bool isSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
161 | // Past the end. |
162 | if (LLVM_LIKELY(!isReferenceToStorage(Elt))__builtin_expect((bool)(!isReferenceToStorage(Elt)), true)) |
163 | return true; |
164 | |
165 | // Return false if Elt will be destroyed by shrinking. |
166 | if (NewSize <= this->size()) |
167 | return Elt < this->begin() + NewSize; |
168 | |
169 | // Return false if we need to grow. |
170 | return NewSize <= this->capacity(); |
171 | } |
172 | |
173 | /// Check whether Elt will be invalidated by resizing the vector to NewSize. |
174 | void assertSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
175 | assert(isSafeToReferenceAfterResize(Elt, NewSize) &&((void)0) |
176 | "Attempting to reference an element of the vector in an operation "((void)0) |
177 | "that invalidates it")((void)0); |
178 | } |
179 | |
180 | /// Check whether Elt will be invalidated by increasing the size of the |
181 | /// vector by N. |
182 | void assertSafeToAdd(const void *Elt, size_t N = 1) { |
183 | this->assertSafeToReferenceAfterResize(Elt, this->size() + N); |
184 | } |
185 | |
186 | /// Check whether any part of the range will be invalidated by clearing. |
187 | void assertSafeToReferenceAfterClear(const T *From, const T *To) { |
188 | if (From == To) |
189 | return; |
190 | this->assertSafeToReferenceAfterResize(From, 0); |
191 | this->assertSafeToReferenceAfterResize(To - 1, 0); |
192 | } |
193 | template < |
194 | class ItTy, |
195 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
196 | bool> = false> |
197 | void assertSafeToReferenceAfterClear(ItTy, ItTy) {} |
198 | |
199 | /// Check whether any part of the range will be invalidated by growing. |
200 | void assertSafeToAddRange(const T *From, const T *To) { |
201 | if (From == To) |
202 | return; |
203 | this->assertSafeToAdd(From, To - From); |
204 | this->assertSafeToAdd(To - 1, To - From); |
205 | } |
206 | template < |
207 | class ItTy, |
208 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
209 | bool> = false> |
210 | void assertSafeToAddRange(ItTy, ItTy) {} |
211 | |
212 | /// Reserve enough space to add one element, and return the updated element |
213 | /// pointer in case it was a reference to the storage. |
214 | template <class U> |
215 | static const T *reserveForParamAndGetAddressImpl(U *This, const T &Elt, |
216 | size_t N) { |
217 | size_t NewSize = This->size() + N; |
218 | if (LLVM_LIKELY(NewSize <= This->capacity())__builtin_expect((bool)(NewSize <= This->capacity()), true )) |
219 | return &Elt; |
220 | |
221 | bool ReferencesStorage = false; |
222 | int64_t Index = -1; |
223 | if (!U::TakesParamByValue) { |
224 | if (LLVM_UNLIKELY(This->isReferenceToStorage(&Elt))__builtin_expect((bool)(This->isReferenceToStorage(&Elt )), false)) { |
225 | ReferencesStorage = true; |
226 | Index = &Elt - This->begin(); |
227 | } |
228 | } |
229 | This->grow(NewSize); |
230 | return ReferencesStorage ? This->begin() + Index : &Elt; |
231 | } |
232 | |
233 | public: |
234 | using size_type = size_t; |
235 | using difference_type = ptrdiff_t; |
236 | using value_type = T; |
237 | using iterator = T *; |
238 | using const_iterator = const T *; |
239 | |
240 | using const_reverse_iterator = std::reverse_iterator<const_iterator>; |
241 | using reverse_iterator = std::reverse_iterator<iterator>; |
242 | |
243 | using reference = T &; |
244 | using const_reference = const T &; |
245 | using pointer = T *; |
246 | using const_pointer = const T *; |
247 | |
248 | using Base::capacity; |
249 | using Base::empty; |
250 | using Base::size; |
251 | |
252 | // forward iterator creation methods. |
253 | iterator begin() { return (iterator)this->BeginX; } |
254 | const_iterator begin() const { return (const_iterator)this->BeginX; } |
255 | iterator end() { return begin() + size(); } |
256 | const_iterator end() const { return begin() + size(); } |
257 | |
258 | // reverse iterator creation methods. |
259 | reverse_iterator rbegin() { return reverse_iterator(end()); } |
260 | const_reverse_iterator rbegin() const{ return const_reverse_iterator(end()); } |
261 | reverse_iterator rend() { return reverse_iterator(begin()); } |
262 | const_reverse_iterator rend() const { return const_reverse_iterator(begin());} |
263 | |
264 | size_type size_in_bytes() const { return size() * sizeof(T); } |
265 | size_type max_size() const { |
266 | return std::min(this->SizeTypeMax(), size_type(-1) / sizeof(T)); |
267 | } |
268 | |
269 | size_t capacity_in_bytes() const { return capacity() * sizeof(T); } |
270 | |
271 | /// Return a pointer to the vector's buffer, even if empty(). |
272 | pointer data() { return pointer(begin()); } |
273 | /// Return a pointer to the vector's buffer, even if empty(). |
274 | const_pointer data() const { return const_pointer(begin()); } |
275 | |
276 | reference operator[](size_type idx) { |
277 | assert(idx < size())((void)0); |
278 | return begin()[idx]; |
279 | } |
280 | const_reference operator[](size_type idx) const { |
281 | assert(idx < size())((void)0); |
282 | return begin()[idx]; |
283 | } |
284 | |
285 | reference front() { |
286 | assert(!empty())((void)0); |
287 | return begin()[0]; |
288 | } |
289 | const_reference front() const { |
290 | assert(!empty())((void)0); |
291 | return begin()[0]; |
292 | } |
293 | |
294 | reference back() { |
295 | assert(!empty())((void)0); |
296 | return end()[-1]; |
297 | } |
298 | const_reference back() const { |
299 | assert(!empty())((void)0); |
300 | return end()[-1]; |
301 | } |
302 | }; |
303 | |
304 | /// SmallVectorTemplateBase<TriviallyCopyable = false> - This is where we put |
305 | /// method implementations that are designed to work with non-trivial T's. |
306 | /// |
307 | /// We approximate is_trivially_copyable with trivial move/copy construction and |
308 | /// trivial destruction. While the standard doesn't specify that you're allowed |
309 | /// copy these types with memcpy, there is no way for the type to observe this. |
310 | /// This catches the important case of std::pair<POD, POD>, which is not |
311 | /// trivially assignable. |
312 | template <typename T, bool = (is_trivially_copy_constructible<T>::value) && |
313 | (is_trivially_move_constructible<T>::value) && |
314 | std::is_trivially_destructible<T>::value> |
315 | class SmallVectorTemplateBase : public SmallVectorTemplateCommon<T> { |
316 | friend class SmallVectorTemplateCommon<T>; |
317 | |
318 | protected: |
319 | static constexpr bool TakesParamByValue = false; |
320 | using ValueParamT = const T &; |
321 | |
322 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
323 | |
324 | static void destroy_range(T *S, T *E) { |
325 | while (S != E) { |
326 | --E; |
327 | E->~T(); |
328 | } |
329 | } |
330 | |
331 | /// Move the range [I, E) into the uninitialized memory starting with "Dest", |
332 | /// constructing elements as needed. |
333 | template<typename It1, typename It2> |
334 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
335 | std::uninitialized_copy(std::make_move_iterator(I), |
336 | std::make_move_iterator(E), Dest); |
337 | } |
338 | |
339 | /// Copy the range [I, E) onto the uninitialized memory starting with "Dest", |
340 | /// constructing elements as needed. |
341 | template<typename It1, typename It2> |
342 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
343 | std::uninitialized_copy(I, E, Dest); |
344 | } |
345 | |
346 | /// Grow the allocated memory (without initializing new elements), doubling |
347 | /// the size of the allocated memory. Guarantees space for at least one more |
348 | /// element, or MinSize more elements if specified. |
349 | void grow(size_t MinSize = 0); |
350 | |
351 | /// Create a new allocation big enough for \p MinSize and pass back its size |
352 | /// in \p NewCapacity. This is the first section of \a grow(). |
353 | T *mallocForGrow(size_t MinSize, size_t &NewCapacity) { |
354 | return static_cast<T *>( |
355 | SmallVectorBase<SmallVectorSizeType<T>>::mallocForGrow( |
356 | MinSize, sizeof(T), NewCapacity)); |
357 | } |
358 | |
359 | /// Move existing elements over to the new allocation \p NewElts, the middle |
360 | /// section of \a grow(). |
361 | void moveElementsForGrow(T *NewElts); |
362 | |
363 | /// Transfer ownership of the allocation, finishing up \a grow(). |
364 | void takeAllocationForGrow(T *NewElts, size_t NewCapacity); |
365 | |
366 | /// Reserve enough space to add one element, and return the updated element |
367 | /// pointer in case it was a reference to the storage. |
368 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
369 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
370 | } |
371 | |
372 | /// Reserve enough space to add one element, and return the updated element |
373 | /// pointer in case it was a reference to the storage. |
374 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
375 | return const_cast<T *>( |
376 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
377 | } |
378 | |
379 | static T &&forward_value_param(T &&V) { return std::move(V); } |
380 | static const T &forward_value_param(const T &V) { return V; } |
381 | |
382 | void growAndAssign(size_t NumElts, const T &Elt) { |
383 | // Grow manually in case Elt is an internal reference. |
384 | size_t NewCapacity; |
385 | T *NewElts = mallocForGrow(NumElts, NewCapacity); |
386 | std::uninitialized_fill_n(NewElts, NumElts, Elt); |
387 | this->destroy_range(this->begin(), this->end()); |
388 | takeAllocationForGrow(NewElts, NewCapacity); |
389 | this->set_size(NumElts); |
390 | } |
391 | |
392 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
393 | // Grow manually in case one of Args is an internal reference. |
394 | size_t NewCapacity; |
395 | T *NewElts = mallocForGrow(0, NewCapacity); |
396 | ::new ((void *)(NewElts + this->size())) T(std::forward<ArgTypes>(Args)...); |
397 | moveElementsForGrow(NewElts); |
398 | takeAllocationForGrow(NewElts, NewCapacity); |
399 | this->set_size(this->size() + 1); |
400 | return this->back(); |
401 | } |
402 | |
403 | public: |
404 | void push_back(const T &Elt) { |
405 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
406 | ::new ((void *)this->end()) T(*EltPtr); |
407 | this->set_size(this->size() + 1); |
408 | } |
409 | |
410 | void push_back(T &&Elt) { |
411 | T *EltPtr = reserveForParamAndGetAddress(Elt); |
412 | ::new ((void *)this->end()) T(::std::move(*EltPtr)); |
413 | this->set_size(this->size() + 1); |
414 | } |
415 | |
416 | void pop_back() { |
417 | this->set_size(this->size() - 1); |
418 | this->end()->~T(); |
419 | } |
420 | }; |
421 | |
422 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
423 | template <typename T, bool TriviallyCopyable> |
424 | void SmallVectorTemplateBase<T, TriviallyCopyable>::grow(size_t MinSize) { |
425 | size_t NewCapacity; |
426 | T *NewElts = mallocForGrow(MinSize, NewCapacity); |
427 | moveElementsForGrow(NewElts); |
428 | takeAllocationForGrow(NewElts, NewCapacity); |
429 | } |
430 | |
431 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
432 | template <typename T, bool TriviallyCopyable> |
433 | void SmallVectorTemplateBase<T, TriviallyCopyable>::moveElementsForGrow( |
434 | T *NewElts) { |
435 | // Move the elements over. |
436 | this->uninitialized_move(this->begin(), this->end(), NewElts); |
437 | |
438 | // Destroy the original elements. |
439 | destroy_range(this->begin(), this->end()); |
440 | } |
441 | |
442 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
443 | template <typename T, bool TriviallyCopyable> |
444 | void SmallVectorTemplateBase<T, TriviallyCopyable>::takeAllocationForGrow( |
445 | T *NewElts, size_t NewCapacity) { |
446 | // If this wasn't grown from the inline copy, deallocate the old space. |
447 | if (!this->isSmall()) |
448 | free(this->begin()); |
449 | |
450 | this->BeginX = NewElts; |
451 | this->Capacity = NewCapacity; |
452 | } |
453 | |
454 | /// SmallVectorTemplateBase<TriviallyCopyable = true> - This is where we put |
455 | /// method implementations that are designed to work with trivially copyable |
456 | /// T's. This allows using memcpy in place of copy/move construction and |
457 | /// skipping destruction. |
458 | template <typename T> |
459 | class SmallVectorTemplateBase<T, true> : public SmallVectorTemplateCommon<T> { |
460 | friend class SmallVectorTemplateCommon<T>; |
461 | |
462 | protected: |
463 | /// True if it's cheap enough to take parameters by value. Doing so avoids |
464 | /// overhead related to mitigations for reference invalidation. |
465 | static constexpr bool TakesParamByValue = sizeof(T) <= 2 * sizeof(void *); |
466 | |
467 | /// Either const T& or T, depending on whether it's cheap enough to take |
468 | /// parameters by value. |
469 | using ValueParamT = |
470 | typename std::conditional<TakesParamByValue, T, const T &>::type; |
471 | |
472 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
473 | |
474 | // No need to do a destroy loop for POD's. |
475 | static void destroy_range(T *, T *) {} |
476 | |
477 | /// Move the range [I, E) onto the uninitialized memory |
478 | /// starting with "Dest", constructing elements into it as needed. |
479 | template<typename It1, typename It2> |
480 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
481 | // Just do a copy. |
482 | uninitialized_copy(I, E, Dest); |
483 | } |
484 | |
485 | /// Copy the range [I, E) onto the uninitialized memory |
486 | /// starting with "Dest", constructing elements into it as needed. |
487 | template<typename It1, typename It2> |
488 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
489 | // Arbitrary iterator types; just use the basic implementation. |
490 | std::uninitialized_copy(I, E, Dest); |
491 | } |
492 | |
493 | /// Copy the range [I, E) onto the uninitialized memory |
494 | /// starting with "Dest", constructing elements into it as needed. |
495 | template <typename T1, typename T2> |
496 | static void uninitialized_copy( |
497 | T1 *I, T1 *E, T2 *Dest, |
498 | std::enable_if_t<std::is_same<typename std::remove_const<T1>::type, |
499 | T2>::value> * = nullptr) { |
500 | // Use memcpy for PODs iterated by pointers (which includes SmallVector |
501 | // iterators): std::uninitialized_copy optimizes to memmove, but we can |
502 | // use memcpy here. Note that I and E are iterators and thus might be |
503 | // invalid for memcpy if they are equal. |
504 | if (I != E) |
505 | memcpy(reinterpret_cast<void *>(Dest), I, (E - I) * sizeof(T)); |
506 | } |
507 | |
508 | /// Double the size of the allocated memory, guaranteeing space for at |
509 | /// least one more element or MinSize if specified. |
510 | void grow(size_t MinSize = 0) { this->grow_pod(MinSize, sizeof(T)); } |
511 | |
512 | /// Reserve enough space to add one element, and return the updated element |
513 | /// pointer in case it was a reference to the storage. |
514 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
515 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
516 | } |
517 | |
518 | /// Reserve enough space to add one element, and return the updated element |
519 | /// pointer in case it was a reference to the storage. |
520 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
521 | return const_cast<T *>( |
522 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
523 | } |
524 | |
525 | /// Copy \p V or return a reference, depending on \a ValueParamT. |
526 | static ValueParamT forward_value_param(ValueParamT V) { return V; } |
527 | |
528 | void growAndAssign(size_t NumElts, T Elt) { |
529 | // Elt has been copied in case it's an internal reference, side-stepping |
530 | // reference invalidation problems without losing the realloc optimization. |
531 | this->set_size(0); |
532 | this->grow(NumElts); |
533 | std::uninitialized_fill_n(this->begin(), NumElts, Elt); |
534 | this->set_size(NumElts); |
535 | } |
536 | |
537 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
538 | // Use push_back with a copy in case Args has an internal reference, |
539 | // side-stepping reference invalidation problems without losing the realloc |
540 | // optimization. |
541 | push_back(T(std::forward<ArgTypes>(Args)...)); |
542 | return this->back(); |
543 | } |
544 | |
545 | public: |
546 | void push_back(ValueParamT Elt) { |
547 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
548 | memcpy(reinterpret_cast<void *>(this->end()), EltPtr, sizeof(T)); |
549 | this->set_size(this->size() + 1); |
550 | } |
551 | |
552 | void pop_back() { this->set_size(this->size() - 1); } |
553 | }; |
554 | |
555 | /// This class consists of common code factored out of the SmallVector class to |
556 | /// reduce code duplication based on the SmallVector 'N' template parameter. |
557 | template <typename T> |
558 | class SmallVectorImpl : public SmallVectorTemplateBase<T> { |
559 | using SuperClass = SmallVectorTemplateBase<T>; |
560 | |
561 | public: |
562 | using iterator = typename SuperClass::iterator; |
563 | using const_iterator = typename SuperClass::const_iterator; |
564 | using reference = typename SuperClass::reference; |
565 | using size_type = typename SuperClass::size_type; |
566 | |
567 | protected: |
568 | using SmallVectorTemplateBase<T>::TakesParamByValue; |
569 | using ValueParamT = typename SuperClass::ValueParamT; |
570 | |
571 | // Default ctor - Initialize to empty. |
572 | explicit SmallVectorImpl(unsigned N) |
573 | : SmallVectorTemplateBase<T>(N) {} |
574 | |
575 | public: |
576 | SmallVectorImpl(const SmallVectorImpl &) = delete; |
577 | |
578 | ~SmallVectorImpl() { |
579 | // Subclass has already destructed this vector's elements. |
580 | // If this wasn't grown from the inline copy, deallocate the old space. |
581 | if (!this->isSmall()) |
582 | free(this->begin()); |
583 | } |
584 | |
585 | void clear() { |
586 | this->destroy_range(this->begin(), this->end()); |
587 | this->Size = 0; |
588 | } |
589 | |
590 | private: |
591 | template <bool ForOverwrite> void resizeImpl(size_type N) { |
592 | if (N < this->size()) { |
593 | this->pop_back_n(this->size() - N); |
594 | } else if (N > this->size()) { |
595 | this->reserve(N); |
596 | for (auto I = this->end(), E = this->begin() + N; I != E; ++I) |
597 | if (ForOverwrite) |
598 | new (&*I) T; |
599 | else |
600 | new (&*I) T(); |
601 | this->set_size(N); |
602 | } |
603 | } |
604 | |
605 | public: |
606 | void resize(size_type N) { resizeImpl<false>(N); } |
607 | |
608 | /// Like resize, but \ref T is POD, the new values won't be initialized. |
609 | void resize_for_overwrite(size_type N) { resizeImpl<true>(N); } |
610 | |
611 | void resize(size_type N, ValueParamT NV) { |
612 | if (N == this->size()) |
613 | return; |
614 | |
615 | if (N < this->size()) { |
616 | this->pop_back_n(this->size() - N); |
617 | return; |
618 | } |
619 | |
620 | // N > this->size(). Defer to append. |
621 | this->append(N - this->size(), NV); |
622 | } |
623 | |
624 | void reserve(size_type N) { |
625 | if (this->capacity() < N) |
626 | this->grow(N); |
627 | } |
628 | |
629 | void pop_back_n(size_type NumItems) { |
630 | assert(this->size() >= NumItems)((void)0); |
631 | this->destroy_range(this->end() - NumItems, this->end()); |
632 | this->set_size(this->size() - NumItems); |
633 | } |
634 | |
635 | LLVM_NODISCARD[[clang::warn_unused_result]] T pop_back_val() { |
636 | T Result = ::std::move(this->back()); |
637 | this->pop_back(); |
638 | return Result; |
639 | } |
640 | |
641 | void swap(SmallVectorImpl &RHS); |
642 | |
643 | /// Add the specified range to the end of the SmallVector. |
644 | template <typename in_iter, |
645 | typename = std::enable_if_t<std::is_convertible< |
646 | typename std::iterator_traits<in_iter>::iterator_category, |
647 | std::input_iterator_tag>::value>> |
648 | void append(in_iter in_start, in_iter in_end) { |
649 | this->assertSafeToAddRange(in_start, in_end); |
650 | size_type NumInputs = std::distance(in_start, in_end); |
651 | this->reserve(this->size() + NumInputs); |
652 | this->uninitialized_copy(in_start, in_end, this->end()); |
653 | this->set_size(this->size() + NumInputs); |
654 | } |
655 | |
656 | /// Append \p NumInputs copies of \p Elt to the end. |
657 | void append(size_type NumInputs, ValueParamT Elt) { |
658 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumInputs); |
659 | std::uninitialized_fill_n(this->end(), NumInputs, *EltPtr); |
660 | this->set_size(this->size() + NumInputs); |
661 | } |
662 | |
663 | void append(std::initializer_list<T> IL) { |
664 | append(IL.begin(), IL.end()); |
665 | } |
666 | |
667 | void append(const SmallVectorImpl &RHS) { append(RHS.begin(), RHS.end()); } |
668 | |
669 | void assign(size_type NumElts, ValueParamT Elt) { |
670 | // Note that Elt could be an internal reference. |
671 | if (NumElts > this->capacity()) { |
672 | this->growAndAssign(NumElts, Elt); |
673 | return; |
674 | } |
675 | |
676 | // Assign over existing elements. |
677 | std::fill_n(this->begin(), std::min(NumElts, this->size()), Elt); |
678 | if (NumElts > this->size()) |
679 | std::uninitialized_fill_n(this->end(), NumElts - this->size(), Elt); |
680 | else if (NumElts < this->size()) |
681 | this->destroy_range(this->begin() + NumElts, this->end()); |
682 | this->set_size(NumElts); |
683 | } |
684 | |
685 | // FIXME: Consider assigning over existing elements, rather than clearing & |
686 | // re-initializing them - for all assign(...) variants. |
687 | |
688 | template <typename in_iter, |
689 | typename = std::enable_if_t<std::is_convertible< |
690 | typename std::iterator_traits<in_iter>::iterator_category, |
691 | std::input_iterator_tag>::value>> |
692 | void assign(in_iter in_start, in_iter in_end) { |
693 | this->assertSafeToReferenceAfterClear(in_start, in_end); |
694 | clear(); |
695 | append(in_start, in_end); |
696 | } |
697 | |
698 | void assign(std::initializer_list<T> IL) { |
699 | clear(); |
700 | append(IL); |
701 | } |
702 | |
703 | void assign(const SmallVectorImpl &RHS) { assign(RHS.begin(), RHS.end()); } |
704 | |
705 | iterator erase(const_iterator CI) { |
706 | // Just cast away constness because this is a non-const member function. |
707 | iterator I = const_cast<iterator>(CI); |
708 | |
709 | assert(this->isReferenceToStorage(CI) && "Iterator to erase is out of bounds.")((void)0); |
710 | |
711 | iterator N = I; |
712 | // Shift all elts down one. |
713 | std::move(I+1, this->end(), I); |
714 | // Drop the last elt. |
715 | this->pop_back(); |
716 | return(N); |
717 | } |
718 | |
719 | iterator erase(const_iterator CS, const_iterator CE) { |
720 | // Just cast away constness because this is a non-const member function. |
721 | iterator S = const_cast<iterator>(CS); |
722 | iterator E = const_cast<iterator>(CE); |
723 | |
724 | assert(this->isRangeInStorage(S, E) && "Range to erase is out of bounds.")((void)0); |
725 | |
726 | iterator N = S; |
727 | // Shift all elts down. |
728 | iterator I = std::move(E, this->end(), S); |
729 | // Drop the last elts. |
730 | this->destroy_range(I, this->end()); |
731 | this->set_size(I - this->begin()); |
732 | return(N); |
733 | } |
734 | |
735 | private: |
736 | template <class ArgType> iterator insert_one_impl(iterator I, ArgType &&Elt) { |
737 | // Callers ensure that ArgType is derived from T. |
738 | static_assert( |
739 | std::is_same<std::remove_const_t<std::remove_reference_t<ArgType>>, |
740 | T>::value, |
741 | "ArgType must be derived from T!"); |
742 | |
743 | if (I == this->end()) { // Important special case for empty vector. |
744 | this->push_back(::std::forward<ArgType>(Elt)); |
745 | return this->end()-1; |
746 | } |
747 | |
748 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
749 | |
750 | // Grow if necessary. |
751 | size_t Index = I - this->begin(); |
752 | std::remove_reference_t<ArgType> *EltPtr = |
753 | this->reserveForParamAndGetAddress(Elt); |
754 | I = this->begin() + Index; |
755 | |
756 | ::new ((void*) this->end()) T(::std::move(this->back())); |
757 | // Push everything else over. |
758 | std::move_backward(I, this->end()-1, this->end()); |
759 | this->set_size(this->size() + 1); |
760 | |
761 | // If we just moved the element we're inserting, be sure to update |
762 | // the reference (never happens if TakesParamByValue). |
763 | static_assert(!TakesParamByValue || std::is_same<ArgType, T>::value, |
764 | "ArgType must be 'T' when taking by value!"); |
765 | if (!TakesParamByValue && this->isReferenceToRange(EltPtr, I, this->end())) |
766 | ++EltPtr; |
767 | |
768 | *I = ::std::forward<ArgType>(*EltPtr); |
769 | return I; |
770 | } |
771 | |
772 | public: |
773 | iterator insert(iterator I, T &&Elt) { |
774 | return insert_one_impl(I, this->forward_value_param(std::move(Elt))); |
775 | } |
776 | |
777 | iterator insert(iterator I, const T &Elt) { |
778 | return insert_one_impl(I, this->forward_value_param(Elt)); |
779 | } |
780 | |
781 | iterator insert(iterator I, size_type NumToInsert, ValueParamT Elt) { |
782 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
783 | size_t InsertElt = I - this->begin(); |
784 | |
785 | if (I == this->end()) { // Important special case for empty vector. |
786 | append(NumToInsert, Elt); |
787 | return this->begin()+InsertElt; |
788 | } |
789 | |
790 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
791 | |
792 | // Ensure there is enough space, and get the (maybe updated) address of |
793 | // Elt. |
794 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumToInsert); |
795 | |
796 | // Uninvalidate the iterator. |
797 | I = this->begin()+InsertElt; |
798 | |
799 | // If there are more elements between the insertion point and the end of the |
800 | // range than there are being inserted, we can use a simple approach to |
801 | // insertion. Since we already reserved space, we know that this won't |
802 | // reallocate the vector. |
803 | if (size_t(this->end()-I) >= NumToInsert) { |
804 | T *OldEnd = this->end(); |
805 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
806 | std::move_iterator<iterator>(this->end())); |
807 | |
808 | // Copy the existing elements that get replaced. |
809 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
810 | |
811 | // If we just moved the element we're inserting, be sure to update |
812 | // the reference (never happens if TakesParamByValue). |
813 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
814 | EltPtr += NumToInsert; |
815 | |
816 | std::fill_n(I, NumToInsert, *EltPtr); |
817 | return I; |
818 | } |
819 | |
820 | // Otherwise, we're inserting more elements than exist already, and we're |
821 | // not inserting at the end. |
822 | |
823 | // Move over the elements that we're about to overwrite. |
824 | T *OldEnd = this->end(); |
825 | this->set_size(this->size() + NumToInsert); |
826 | size_t NumOverwritten = OldEnd-I; |
827 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
828 | |
829 | // If we just moved the element we're inserting, be sure to update |
830 | // the reference (never happens if TakesParamByValue). |
831 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
832 | EltPtr += NumToInsert; |
833 | |
834 | // Replace the overwritten part. |
835 | std::fill_n(I, NumOverwritten, *EltPtr); |
836 | |
837 | // Insert the non-overwritten middle part. |
838 | std::uninitialized_fill_n(OldEnd, NumToInsert - NumOverwritten, *EltPtr); |
839 | return I; |
840 | } |
841 | |
842 | template <typename ItTy, |
843 | typename = std::enable_if_t<std::is_convertible< |
844 | typename std::iterator_traits<ItTy>::iterator_category, |
845 | std::input_iterator_tag>::value>> |
846 | iterator insert(iterator I, ItTy From, ItTy To) { |
847 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
848 | size_t InsertElt = I - this->begin(); |
849 | |
850 | if (I == this->end()) { // Important special case for empty vector. |
851 | append(From, To); |
852 | return this->begin()+InsertElt; |
853 | } |
854 | |
855 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
856 | |
857 | // Check that the reserve that follows doesn't invalidate the iterators. |
858 | this->assertSafeToAddRange(From, To); |
859 | |
860 | size_t NumToInsert = std::distance(From, To); |
861 | |
862 | // Ensure there is enough space. |
863 | reserve(this->size() + NumToInsert); |
864 | |
865 | // Uninvalidate the iterator. |
866 | I = this->begin()+InsertElt; |
867 | |
868 | // If there are more elements between the insertion point and the end of the |
869 | // range than there are being inserted, we can use a simple approach to |
870 | // insertion. Since we already reserved space, we know that this won't |
871 | // reallocate the vector. |
872 | if (size_t(this->end()-I) >= NumToInsert) { |
873 | T *OldEnd = this->end(); |
874 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
875 | std::move_iterator<iterator>(this->end())); |
876 | |
877 | // Copy the existing elements that get replaced. |
878 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
879 | |
880 | std::copy(From, To, I); |
881 | return I; |
882 | } |
883 | |
884 | // Otherwise, we're inserting more elements than exist already, and we're |
885 | // not inserting at the end. |
886 | |
887 | // Move over the elements that we're about to overwrite. |
888 | T *OldEnd = this->end(); |
889 | this->set_size(this->size() + NumToInsert); |
890 | size_t NumOverwritten = OldEnd-I; |
891 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
892 | |
893 | // Replace the overwritten part. |
894 | for (T *J = I; NumOverwritten > 0; --NumOverwritten) { |
895 | *J = *From; |
896 | ++J; ++From; |
897 | } |
898 | |
899 | // Insert the non-overwritten middle part. |
900 | this->uninitialized_copy(From, To, OldEnd); |
901 | return I; |
902 | } |
903 | |
904 | void insert(iterator I, std::initializer_list<T> IL) { |
905 | insert(I, IL.begin(), IL.end()); |
906 | } |
907 | |
908 | template <typename... ArgTypes> reference emplace_back(ArgTypes &&... Args) { |
909 | if (LLVM_UNLIKELY(this->size() >= this->capacity())__builtin_expect((bool)(this->size() >= this->capacity ()), false)) |
910 | return this->growAndEmplaceBack(std::forward<ArgTypes>(Args)...); |
911 | |
912 | ::new ((void *)this->end()) T(std::forward<ArgTypes>(Args)...); |
913 | this->set_size(this->size() + 1); |
914 | return this->back(); |
915 | } |
916 | |
917 | SmallVectorImpl &operator=(const SmallVectorImpl &RHS); |
918 | |
919 | SmallVectorImpl &operator=(SmallVectorImpl &&RHS); |
920 | |
921 | bool operator==(const SmallVectorImpl &RHS) const { |
922 | if (this->size() != RHS.size()) return false; |
923 | return std::equal(this->begin(), this->end(), RHS.begin()); |
924 | } |
925 | bool operator!=(const SmallVectorImpl &RHS) const { |
926 | return !(*this == RHS); |
927 | } |
928 | |
929 | bool operator<(const SmallVectorImpl &RHS) const { |
930 | return std::lexicographical_compare(this->begin(), this->end(), |
931 | RHS.begin(), RHS.end()); |
932 | } |
933 | }; |
934 | |
935 | template <typename T> |
936 | void SmallVectorImpl<T>::swap(SmallVectorImpl<T> &RHS) { |
937 | if (this == &RHS) return; |
938 | |
939 | // We can only avoid copying elements if neither vector is small. |
940 | if (!this->isSmall() && !RHS.isSmall()) { |
941 | std::swap(this->BeginX, RHS.BeginX); |
942 | std::swap(this->Size, RHS.Size); |
943 | std::swap(this->Capacity, RHS.Capacity); |
944 | return; |
945 | } |
946 | this->reserve(RHS.size()); |
947 | RHS.reserve(this->size()); |
948 | |
949 | // Swap the shared elements. |
950 | size_t NumShared = this->size(); |
951 | if (NumShared > RHS.size()) NumShared = RHS.size(); |
952 | for (size_type i = 0; i != NumShared; ++i) |
953 | std::swap((*this)[i], RHS[i]); |
954 | |
955 | // Copy over the extra elts. |
956 | if (this->size() > RHS.size()) { |
957 | size_t EltDiff = this->size() - RHS.size(); |
958 | this->uninitialized_copy(this->begin()+NumShared, this->end(), RHS.end()); |
959 | RHS.set_size(RHS.size() + EltDiff); |
960 | this->destroy_range(this->begin()+NumShared, this->end()); |
961 | this->set_size(NumShared); |
962 | } else if (RHS.size() > this->size()) { |
963 | size_t EltDiff = RHS.size() - this->size(); |
964 | this->uninitialized_copy(RHS.begin()+NumShared, RHS.end(), this->end()); |
965 | this->set_size(this->size() + EltDiff); |
966 | this->destroy_range(RHS.begin()+NumShared, RHS.end()); |
967 | RHS.set_size(NumShared); |
968 | } |
969 | } |
970 | |
971 | template <typename T> |
972 | SmallVectorImpl<T> &SmallVectorImpl<T>:: |
973 | operator=(const SmallVectorImpl<T> &RHS) { |
974 | // Avoid self-assignment. |
975 | if (this == &RHS) return *this; |
976 | |
977 | // If we already have sufficient space, assign the common elements, then |
978 | // destroy any excess. |
979 | size_t RHSSize = RHS.size(); |
980 | size_t CurSize = this->size(); |
981 | if (CurSize >= RHSSize) { |
982 | // Assign common elements. |
983 | iterator NewEnd; |
984 | if (RHSSize) |
985 | NewEnd = std::copy(RHS.begin(), RHS.begin()+RHSSize, this->begin()); |
986 | else |
987 | NewEnd = this->begin(); |
988 | |
989 | // Destroy excess elements. |
990 | this->destroy_range(NewEnd, this->end()); |
991 | |
992 | // Trim. |
993 | this->set_size(RHSSize); |
994 | return *this; |
995 | } |
996 | |
997 | // If we have to grow to have enough elements, destroy the current elements. |
998 | // This allows us to avoid copying them during the grow. |
999 | // FIXME: don't do this if they're efficiently moveable. |
1000 | if (this->capacity() < RHSSize) { |
1001 | // Destroy current elements. |
1002 | this->clear(); |
1003 | CurSize = 0; |
1004 | this->grow(RHSSize); |
1005 | } else if (CurSize) { |
1006 | // Otherwise, use assignment for the already-constructed elements. |
1007 | std::copy(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
1008 | } |
1009 | |
1010 | // Copy construct the new elements in place. |
1011 | this->uninitialized_copy(RHS.begin()+CurSize, RHS.end(), |
1012 | this->begin()+CurSize); |
1013 | |
1014 | // Set end. |
1015 | this->set_size(RHSSize); |
1016 | return *this; |
1017 | } |
1018 | |
1019 | template <typename T> |
1020 | SmallVectorImpl<T> &SmallVectorImpl<T>::operator=(SmallVectorImpl<T> &&RHS) { |
1021 | // Avoid self-assignment. |
1022 | if (this == &RHS) return *this; |
1023 | |
1024 | // If the RHS isn't small, clear this vector and then steal its buffer. |
1025 | if (!RHS.isSmall()) { |
1026 | this->destroy_range(this->begin(), this->end()); |
1027 | if (!this->isSmall()) free(this->begin()); |
1028 | this->BeginX = RHS.BeginX; |
1029 | this->Size = RHS.Size; |
1030 | this->Capacity = RHS.Capacity; |
1031 | RHS.resetToSmall(); |
1032 | return *this; |
1033 | } |
1034 | |
1035 | // If we already have sufficient space, assign the common elements, then |
1036 | // destroy any excess. |
1037 | size_t RHSSize = RHS.size(); |
1038 | size_t CurSize = this->size(); |
1039 | if (CurSize >= RHSSize) { |
1040 | // Assign common elements. |
1041 | iterator NewEnd = this->begin(); |
1042 | if (RHSSize) |
1043 | NewEnd = std::move(RHS.begin(), RHS.end(), NewEnd); |
1044 | |
1045 | // Destroy excess elements and trim the bounds. |
1046 | this->destroy_range(NewEnd, this->end()); |
1047 | this->set_size(RHSSize); |
1048 | |
1049 | // Clear the RHS. |
1050 | RHS.clear(); |
1051 | |
1052 | return *this; |
1053 | } |
1054 | |
1055 | // If we have to grow to have enough elements, destroy the current elements. |
1056 | // This allows us to avoid copying them during the grow. |
1057 | // FIXME: this may not actually make any sense if we can efficiently move |
1058 | // elements. |
1059 | if (this->capacity() < RHSSize) { |
1060 | // Destroy current elements. |
1061 | this->clear(); |
1062 | CurSize = 0; |
1063 | this->grow(RHSSize); |
1064 | } else if (CurSize) { |
1065 | // Otherwise, use assignment for the already-constructed elements. |
1066 | std::move(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
1067 | } |
1068 | |
1069 | // Move-construct the new elements in place. |
1070 | this->uninitialized_move(RHS.begin()+CurSize, RHS.end(), |
1071 | this->begin()+CurSize); |
1072 | |
1073 | // Set end. |
1074 | this->set_size(RHSSize); |
1075 | |
1076 | RHS.clear(); |
1077 | return *this; |
1078 | } |
1079 | |
1080 | /// Storage for the SmallVector elements. This is specialized for the N=0 case |
1081 | /// to avoid allocating unnecessary storage. |
1082 | template <typename T, unsigned N> |
1083 | struct SmallVectorStorage { |
1084 | alignas(T) char InlineElts[N * sizeof(T)]; |
1085 | }; |
1086 | |
1087 | /// We need the storage to be properly aligned even for small-size of 0 so that |
1088 | /// the pointer math in \a SmallVectorTemplateCommon::getFirstEl() is |
1089 | /// well-defined. |
1090 | template <typename T> struct alignas(T) SmallVectorStorage<T, 0> {}; |
1091 | |
1092 | /// Forward declaration of SmallVector so that |
1093 | /// calculateSmallVectorDefaultInlinedElements can reference |
1094 | /// `sizeof(SmallVector<T, 0>)`. |
1095 | template <typename T, unsigned N> class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector; |
1096 | |
1097 | /// Helper class for calculating the default number of inline elements for |
1098 | /// `SmallVector<T>`. |
1099 | /// |
1100 | /// This should be migrated to a constexpr function when our minimum |
1101 | /// compiler support is enough for multi-statement constexpr functions. |
1102 | template <typename T> struct CalculateSmallVectorDefaultInlinedElements { |
1103 | // Parameter controlling the default number of inlined elements |
1104 | // for `SmallVector<T>`. |
1105 | // |
1106 | // The default number of inlined elements ensures that |
1107 | // 1. There is at least one inlined element. |
1108 | // 2. `sizeof(SmallVector<T>) <= kPreferredSmallVectorSizeof` unless |
1109 | // it contradicts 1. |
1110 | static constexpr size_t kPreferredSmallVectorSizeof = 64; |
1111 | |
1112 | // static_assert that sizeof(T) is not "too big". |
1113 | // |
1114 | // Because our policy guarantees at least one inlined element, it is possible |
1115 | // for an arbitrarily large inlined element to allocate an arbitrarily large |
1116 | // amount of inline storage. We generally consider it an antipattern for a |
1117 | // SmallVector to allocate an excessive amount of inline storage, so we want |
1118 | // to call attention to these cases and make sure that users are making an |
1119 | // intentional decision if they request a lot of inline storage. |
1120 | // |
1121 | // We want this assertion to trigger in pathological cases, but otherwise |
1122 | // not be too easy to hit. To accomplish that, the cutoff is actually somewhat |
1123 | // larger than kPreferredSmallVectorSizeof (otherwise, |
1124 | // `SmallVector<SmallVector<T>>` would be one easy way to trip it, and that |
1125 | // pattern seems useful in practice). |
1126 | // |
1127 | // One wrinkle is that this assertion is in theory non-portable, since |
1128 | // sizeof(T) is in general platform-dependent. However, we don't expect this |
1129 | // to be much of an issue, because most LLVM development happens on 64-bit |
1130 | // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for |
1131 | // 32-bit hosts, dodging the issue. The reverse situation, where development |
1132 | // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a |
1133 | // 64-bit host, is expected to be very rare. |
1134 | static_assert( |
1135 | sizeof(T) <= 256, |
1136 | "You are trying to use a default number of inlined elements for " |
1137 | "`SmallVector<T>` but `sizeof(T)` is really big! Please use an " |
1138 | "explicit number of inlined elements with `SmallVector<T, N>` to make " |
1139 | "sure you really want that much inline storage."); |
1140 | |
1141 | // Discount the size of the header itself when calculating the maximum inline |
1142 | // bytes. |
1143 | static constexpr size_t PreferredInlineBytes = |
1144 | kPreferredSmallVectorSizeof - sizeof(SmallVector<T, 0>); |
1145 | static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T); |
1146 | static constexpr size_t value = |
1147 | NumElementsThatFit == 0 ? 1 : NumElementsThatFit; |
1148 | }; |
1149 | |
1150 | /// This is a 'vector' (really, a variable-sized array), optimized |
1151 | /// for the case when the array is small. It contains some number of elements |
1152 | /// in-place, which allows it to avoid heap allocation when the actual number of |
1153 | /// elements is below that threshold. This allows normal "small" cases to be |
1154 | /// fast without losing generality for large inputs. |
1155 | /// |
1156 | /// \note |
1157 | /// In the absence of a well-motivated choice for the number of inlined |
1158 | /// elements \p N, it is recommended to use \c SmallVector<T> (that is, |
1159 | /// omitting the \p N). This will choose a default number of inlined elements |
1160 | /// reasonable for allocation on the stack (for example, trying to keep \c |
1161 | /// sizeof(SmallVector<T>) around 64 bytes). |
1162 | /// |
1163 | /// \warning This does not attempt to be exception safe. |
1164 | /// |
1165 | /// \see https://llvm.org/docs/ProgrammersManual.html#llvm-adt-smallvector-h |
1166 | template <typename T, |
1167 | unsigned N = CalculateSmallVectorDefaultInlinedElements<T>::value> |
1168 | class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector : public SmallVectorImpl<T>, |
1169 | SmallVectorStorage<T, N> { |
1170 | public: |
1171 | SmallVector() : SmallVectorImpl<T>(N) {} |
1172 | |
1173 | ~SmallVector() { |
1174 | // Destroy the constructed elements in the vector. |
1175 | this->destroy_range(this->begin(), this->end()); |
1176 | } |
1177 | |
1178 | explicit SmallVector(size_t Size, const T &Value = T()) |
1179 | : SmallVectorImpl<T>(N) { |
1180 | this->assign(Size, Value); |
1181 | } |
1182 | |
1183 | template <typename ItTy, |
1184 | typename = std::enable_if_t<std::is_convertible< |
1185 | typename std::iterator_traits<ItTy>::iterator_category, |
1186 | std::input_iterator_tag>::value>> |
1187 | SmallVector(ItTy S, ItTy E) : SmallVectorImpl<T>(N) { |
1188 | this->append(S, E); |
1189 | } |
1190 | |
1191 | template <typename RangeTy> |
1192 | explicit SmallVector(const iterator_range<RangeTy> &R) |
1193 | : SmallVectorImpl<T>(N) { |
1194 | this->append(R.begin(), R.end()); |
1195 | } |
1196 | |
1197 | SmallVector(std::initializer_list<T> IL) : SmallVectorImpl<T>(N) { |
1198 | this->assign(IL); |
1199 | } |
1200 | |
1201 | SmallVector(const SmallVector &RHS) : SmallVectorImpl<T>(N) { |
1202 | if (!RHS.empty()) |
1203 | SmallVectorImpl<T>::operator=(RHS); |
1204 | } |
1205 | |
1206 | SmallVector &operator=(const SmallVector &RHS) { |
1207 | SmallVectorImpl<T>::operator=(RHS); |
1208 | return *this; |
1209 | } |
1210 | |
1211 | SmallVector(SmallVector &&RHS) : SmallVectorImpl<T>(N) { |
1212 | if (!RHS.empty()) |
1213 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1214 | } |
1215 | |
1216 | SmallVector(SmallVectorImpl<T> &&RHS) : SmallVectorImpl<T>(N) { |
1217 | if (!RHS.empty()) |
1218 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1219 | } |
1220 | |
1221 | SmallVector &operator=(SmallVector &&RHS) { |
1222 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1223 | return *this; |
1224 | } |
1225 | |
1226 | SmallVector &operator=(SmallVectorImpl<T> &&RHS) { |
1227 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1228 | return *this; |
1229 | } |
1230 | |
1231 | SmallVector &operator=(std::initializer_list<T> IL) { |
1232 | this->assign(IL); |
1233 | return *this; |
1234 | } |
1235 | }; |
1236 | |
1237 | template <typename T, unsigned N> |
1238 | inline size_t capacity_in_bytes(const SmallVector<T, N> &X) { |
1239 | return X.capacity_in_bytes(); |
1240 | } |
1241 | |
1242 | /// Given a range of type R, iterate the entire range and return a |
1243 | /// SmallVector with elements of the vector. This is useful, for example, |
1244 | /// when you want to iterate a range and then sort the results. |
1245 | template <unsigned Size, typename R> |
1246 | SmallVector<typename std::remove_const<typename std::remove_reference< |
1247 | decltype(*std::begin(std::declval<R &>()))>::type>::type, |
1248 | Size> |
1249 | to_vector(R &&Range) { |
1250 | return {std::begin(Range), std::end(Range)}; |
1251 | } |
1252 | |
1253 | } // end namespace llvm |
1254 | |
1255 | namespace std { |
1256 | |
1257 | /// Implement std::swap in terms of SmallVector swap. |
1258 | template<typename T> |
1259 | inline void |
1260 | swap(llvm::SmallVectorImpl<T> &LHS, llvm::SmallVectorImpl<T> &RHS) { |
1261 | LHS.swap(RHS); |
1262 | } |
1263 | |
1264 | /// Implement std::swap in terms of SmallVector swap. |
1265 | template<typename T, unsigned N> |
1266 | inline void |
1267 | swap(llvm::SmallVector<T, N> &LHS, llvm::SmallVector<T, N> &RHS) { |
1268 | LHS.swap(RHS); |
1269 | } |
1270 | |
1271 | } // end namespace std |
1272 | |
1273 | #endif // LLVM_ADT_SMALLVECTOR_H |
1 | //===-- Analysis/CFG.h - BasicBlock Analyses --------------------*- 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 family of functions performs analyses on basic blocks, and instructions |
10 | // contained within basic blocks. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef LLVM_ANALYSIS_CFG_H |
15 | #define LLVM_ANALYSIS_CFG_H |
16 | |
17 | #include "llvm/ADT/GraphTraits.h" |
18 | #include "llvm/ADT/SmallPtrSet.h" |
19 | #include <utility> |
20 | |
21 | namespace llvm { |
22 | |
23 | class BasicBlock; |
24 | class DominatorTree; |
25 | class Function; |
26 | class Instruction; |
27 | class LoopInfo; |
28 | template <typename T> class SmallVectorImpl; |
29 | |
30 | /// Analyze the specified function to find all of the loop backedges in the |
31 | /// function and return them. This is a relatively cheap (compared to |
32 | /// computing dominators and loop info) analysis. |
33 | /// |
34 | /// The output is added to Result, as pairs of <from,to> edge info. |
35 | void FindFunctionBackedges( |
36 | const Function &F, |
37 | SmallVectorImpl<std::pair<const BasicBlock *, const BasicBlock *> > & |
38 | Result); |
39 | |
40 | /// Search for the specified successor of basic block BB and return its position |
41 | /// in the terminator instruction's list of successors. It is an error to call |
42 | /// this with a block that is not a successor. |
43 | unsigned GetSuccessorNumber(const BasicBlock *BB, const BasicBlock *Succ); |
44 | |
45 | /// Return true if the specified edge is a critical edge. Critical edges are |
46 | /// edges from a block with multiple successors to a block with multiple |
47 | /// predecessors. |
48 | /// |
49 | bool isCriticalEdge(const Instruction *TI, unsigned SuccNum, |
50 | bool AllowIdenticalEdges = false); |
51 | bool isCriticalEdge(const Instruction *TI, const BasicBlock *Succ, |
52 | bool AllowIdenticalEdges = false); |
53 | |
54 | /// Determine whether instruction 'To' is reachable from 'From', without passing |
55 | /// through any blocks in ExclusionSet, returning true if uncertain. |
56 | /// |
57 | /// Determine whether there is a path from From to To within a single function. |
58 | /// Returns false only if we can prove that once 'From' has been executed then |
59 | /// 'To' can not be executed. Conservatively returns true. |
60 | /// |
61 | /// This function is linear with respect to the number of blocks in the CFG, |
62 | /// walking down successors from From to reach To, with a fixed threshold. |
63 | /// Using DT or LI allows us to answer more quickly. LI reduces the cost of |
64 | /// an entire loop of any number of blocks to be the same as the cost of a |
65 | /// single block. DT reduces the cost by allowing the search to terminate when |
66 | /// we find a block that dominates the block containing 'To'. DT is most useful |
67 | /// on branchy code but not loops, and LI is most useful on code with loops but |
68 | /// does not help on branchy code outside loops. |
69 | bool isPotentiallyReachable( |
70 | const Instruction *From, const Instruction *To, |
71 | const SmallPtrSetImpl<BasicBlock *> *ExclusionSet = nullptr, |
72 | const DominatorTree *DT = nullptr, const LoopInfo *LI = nullptr); |
73 | |
74 | /// Determine whether block 'To' is reachable from 'From', returning |
75 | /// true if uncertain. |
76 | /// |
77 | /// Determine whether there is a path from From to To within a single function. |
78 | /// Returns false only if we can prove that once 'From' has been reached then |
79 | /// 'To' can not be executed. Conservatively returns true. |
80 | bool isPotentiallyReachable( |
81 | const BasicBlock *From, const BasicBlock *To, |
82 | const SmallPtrSetImpl<BasicBlock *> *ExclusionSet = nullptr, |
83 | const DominatorTree *DT = nullptr, const LoopInfo *LI = nullptr); |
84 | |
85 | /// Determine whether there is at least one path from a block in |
86 | /// 'Worklist' to 'StopBB' without passing through any blocks in |
87 | /// 'ExclusionSet', returning true if uncertain. |
88 | /// |
89 | /// Determine whether there is a path from at least one block in Worklist to |
90 | /// StopBB within a single function without passing through any of the blocks |
91 | /// in 'ExclusionSet'. Returns false only if we can prove that once any block |
92 | /// in 'Worklist' has been reached then 'StopBB' can not be executed. |
93 | /// Conservatively returns true. |
94 | bool isPotentiallyReachableFromMany( |
95 | SmallVectorImpl<BasicBlock *> &Worklist, BasicBlock *StopBB, |
96 | const SmallPtrSetImpl<BasicBlock *> *ExclusionSet, |
97 | const DominatorTree *DT = nullptr, const LoopInfo *LI = nullptr); |
98 | |
99 | /// Return true if the control flow in \p RPOTraversal is irreducible. |
100 | /// |
101 | /// This is a generic implementation to detect CFG irreducibility based on loop |
102 | /// info analysis. It can be used for any kind of CFG (Loop, MachineLoop, |
103 | /// Function, MachineFunction, etc.) by providing an RPO traversal (\p |
104 | /// RPOTraversal) and the loop info analysis (\p LI) of the CFG. This utility |
105 | /// function is only recommended when loop info analysis is available. If loop |
106 | /// info analysis isn't available, please, don't compute it explicitly for this |
107 | /// purpose. There are more efficient ways to detect CFG irreducibility that |
108 | /// don't require recomputing loop info analysis (e.g., T1/T2 or Tarjan's |
109 | /// algorithm). |
110 | /// |
111 | /// Requirements: |
112 | /// 1) GraphTraits must be implemented for NodeT type. It is used to access |
113 | /// NodeT successors. |
114 | // 2) \p RPOTraversal must be a valid reverse post-order traversal of the |
115 | /// target CFG with begin()/end() iterator interfaces. |
116 | /// 3) \p LI must be a valid LoopInfoBase that contains up-to-date loop |
117 | /// analysis information of the CFG. |
118 | /// |
119 | /// This algorithm uses the information about reducible loop back-edges already |
120 | /// computed in \p LI. When a back-edge is found during the RPO traversal, the |
121 | /// algorithm checks whether the back-edge is one of the reducible back-edges in |
122 | /// loop info. If it isn't, the CFG is irreducible. For example, for the CFG |
123 | /// below (canonical irreducible graph) loop info won't contain any loop, so the |
124 | /// algorithm will return that the CFG is irreducible when checking the B <- |
125 | /// -> C back-edge. |
126 | /// |
127 | /// (A->B, A->C, B->C, C->B, C->D) |
128 | /// A |
129 | /// / \ |
130 | /// B<- ->C |
131 | /// | |
132 | /// D |
133 | /// |
134 | template <class NodeT, class RPOTraversalT, class LoopInfoT, |
135 | class GT = GraphTraits<NodeT>> |
136 | bool containsIrreducibleCFG(RPOTraversalT &RPOTraversal, const LoopInfoT &LI) { |
137 | /// Check whether the edge (\p Src, \p Dst) is a reducible loop backedge |
138 | /// according to LI. I.e., check if there exists a loop that contains Src and |
139 | /// where Dst is the loop header. |
140 | auto isProperBackedge = [&](NodeT Src, NodeT Dst) { |
141 | for (const auto *Lp = LI.getLoopFor(Src); Lp; Lp = Lp->getParentLoop()) { |
142 | if (Lp->getHeader() == Dst) |
143 | return true; |
144 | } |
145 | return false; |
146 | }; |
147 | |
148 | SmallPtrSet<NodeT, 32> Visited; |
149 | for (NodeT Node : RPOTraversal) { |
150 | Visited.insert(Node); |
151 | for (NodeT Succ : make_range(GT::child_begin(Node), GT::child_end(Node))) { |
152 | // Succ hasn't been visited yet |
153 | if (!Visited.count(Succ)) |
154 | continue; |
155 | // We already visited Succ, thus Node->Succ must be a backedge. Check that |
156 | // the head matches what we have in the loop information. Otherwise, we |
157 | // have an irreducible graph. |
158 | if (!isProperBackedge(Node, Succ)) |
159 | return true; |
160 | } |
161 | } |
162 | |
163 | return false; |
164 | } |
165 | } // End llvm namespace |
166 | |
167 | #endif |
1 | //===- PatternMatch.h - Match on the LLVM IR --------------------*- 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 provides a simple and efficient mechanism for performing general |
10 | // tree-based pattern matches on the LLVM IR. The power of these routines is |
11 | // that it allows you to write concise patterns that are expressive and easy to |
12 | // understand. The other major advantage of this is that it allows you to |
13 | // trivially capture/bind elements in the pattern to variables. For example, |
14 | // you can do something like this: |
15 | // |
16 | // Value *Exp = ... |
17 | // Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2) |
18 | // if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)), |
19 | // m_And(m_Value(Y), m_ConstantInt(C2))))) { |
20 | // ... Pattern is matched and variables are bound ... |
21 | // } |
22 | // |
23 | // This is primarily useful to things like the instruction combiner, but can |
24 | // also be useful for static analysis tools or code generators. |
25 | // |
26 | //===----------------------------------------------------------------------===// |
27 | |
28 | #ifndef LLVM_IR_PATTERNMATCH_H |
29 | #define LLVM_IR_PATTERNMATCH_H |
30 | |
31 | #include "llvm/ADT/APFloat.h" |
32 | #include "llvm/ADT/APInt.h" |
33 | #include "llvm/IR/Constant.h" |
34 | #include "llvm/IR/Constants.h" |
35 | #include "llvm/IR/DataLayout.h" |
36 | #include "llvm/IR/InstrTypes.h" |
37 | #include "llvm/IR/Instruction.h" |
38 | #include "llvm/IR/Instructions.h" |
39 | #include "llvm/IR/IntrinsicInst.h" |
40 | #include "llvm/IR/Intrinsics.h" |
41 | #include "llvm/IR/Operator.h" |
42 | #include "llvm/IR/Value.h" |
43 | #include "llvm/Support/Casting.h" |
44 | #include <cstdint> |
45 | |
46 | namespace llvm { |
47 | namespace PatternMatch { |
48 | |
49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { |
50 | return const_cast<Pattern &>(P).match(V); |
51 | } |
52 | |
53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { |
54 | return const_cast<Pattern &>(P).match(Mask); |
55 | } |
56 | |
57 | template <typename SubPattern_t> struct OneUse_match { |
58 | SubPattern_t SubPattern; |
59 | |
60 | OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {} |
61 | |
62 | template <typename OpTy> bool match(OpTy *V) { |
63 | return V->hasOneUse() && SubPattern.match(V); |
64 | } |
65 | }; |
66 | |
67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { |
68 | return SubPattern; |
69 | } |
70 | |
71 | template <typename Class> struct class_match { |
72 | template <typename ITy> bool match(ITy *V) { return isa<Class>(V); } |
73 | }; |
74 | |
75 | /// Match an arbitrary value and ignore it. |
76 | inline class_match<Value> m_Value() { return class_match<Value>(); } |
77 | |
78 | /// Match an arbitrary unary operation and ignore it. |
79 | inline class_match<UnaryOperator> m_UnOp() { |
80 | return class_match<UnaryOperator>(); |
81 | } |
82 | |
83 | /// Match an arbitrary binary operation and ignore it. |
84 | inline class_match<BinaryOperator> m_BinOp() { |
85 | return class_match<BinaryOperator>(); |
86 | } |
87 | |
88 | /// Matches any compare instruction and ignore it. |
89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } |
90 | |
91 | struct undef_match { |
92 | static bool check(const Value *V) { |
93 | if (isa<UndefValue>(V)) |
94 | return true; |
95 | |
96 | const auto *CA = dyn_cast<ConstantAggregate>(V); |
97 | if (!CA) |
98 | return false; |
99 | |
100 | SmallPtrSet<const ConstantAggregate *, 8> Seen; |
101 | SmallVector<const ConstantAggregate *, 8> Worklist; |
102 | |
103 | // Either UndefValue, PoisonValue, or an aggregate that only contains |
104 | // these is accepted by matcher. |
105 | // CheckValue returns false if CA cannot satisfy this constraint. |
106 | auto CheckValue = [&](const ConstantAggregate *CA) { |
107 | for (const Value *Op : CA->operand_values()) { |
108 | if (isa<UndefValue>(Op)) |
109 | continue; |
110 | |
111 | const auto *CA = dyn_cast<ConstantAggregate>(Op); |
112 | if (!CA) |
113 | return false; |
114 | if (Seen.insert(CA).second) |
115 | Worklist.emplace_back(CA); |
116 | } |
117 | |
118 | return true; |
119 | }; |
120 | |
121 | if (!CheckValue(CA)) |
122 | return false; |
123 | |
124 | while (!Worklist.empty()) { |
125 | if (!CheckValue(Worklist.pop_back_val())) |
126 | return false; |
127 | } |
128 | return true; |
129 | } |
130 | template <typename ITy> bool match(ITy *V) { return check(V); } |
131 | }; |
132 | |
133 | /// Match an arbitrary undef constant. This matches poison as well. |
134 | /// If this is an aggregate and contains a non-aggregate element that is |
135 | /// neither undef nor poison, the aggregate is not matched. |
136 | inline auto m_Undef() { return undef_match(); } |
137 | |
138 | /// Match an arbitrary poison constant. |
139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } |
140 | |
141 | /// Match an arbitrary Constant and ignore it. |
142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } |
143 | |
144 | /// Match an arbitrary ConstantInt and ignore it. |
145 | inline class_match<ConstantInt> m_ConstantInt() { |
146 | return class_match<ConstantInt>(); |
147 | } |
148 | |
149 | /// Match an arbitrary ConstantFP and ignore it. |
150 | inline class_match<ConstantFP> m_ConstantFP() { |
151 | return class_match<ConstantFP>(); |
152 | } |
153 | |
154 | /// Match an arbitrary ConstantExpr and ignore it. |
155 | inline class_match<ConstantExpr> m_ConstantExpr() { |
156 | return class_match<ConstantExpr>(); |
157 | } |
158 | |
159 | /// Match an arbitrary basic block value and ignore it. |
160 | inline class_match<BasicBlock> m_BasicBlock() { |
161 | return class_match<BasicBlock>(); |
162 | } |
163 | |
164 | /// Inverting matcher |
165 | template <typename Ty> struct match_unless { |
166 | Ty M; |
167 | |
168 | match_unless(const Ty &Matcher) : M(Matcher) {} |
169 | |
170 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } |
171 | }; |
172 | |
173 | /// Match if the inner matcher does *NOT* match. |
174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { |
175 | return match_unless<Ty>(M); |
176 | } |
177 | |
178 | /// Matching combinators |
179 | template <typename LTy, typename RTy> struct match_combine_or { |
180 | LTy L; |
181 | RTy R; |
182 | |
183 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
184 | |
185 | template <typename ITy> bool match(ITy *V) { |
186 | if (L.match(V)) |
187 | return true; |
188 | if (R.match(V)) |
189 | return true; |
190 | return false; |
191 | } |
192 | }; |
193 | |
194 | template <typename LTy, typename RTy> struct match_combine_and { |
195 | LTy L; |
196 | RTy R; |
197 | |
198 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
199 | |
200 | template <typename ITy> bool match(ITy *V) { |
201 | if (L.match(V)) |
202 | if (R.match(V)) |
203 | return true; |
204 | return false; |
205 | } |
206 | }; |
207 | |
208 | /// Combine two pattern matchers matching L || R |
209 | template <typename LTy, typename RTy> |
210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { |
211 | return match_combine_or<LTy, RTy>(L, R); |
212 | } |
213 | |
214 | /// Combine two pattern matchers matching L && R |
215 | template <typename LTy, typename RTy> |
216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { |
217 | return match_combine_and<LTy, RTy>(L, R); |
218 | } |
219 | |
220 | struct apint_match { |
221 | const APInt *&Res; |
222 | bool AllowUndef; |
223 | |
224 | apint_match(const APInt *&Res, bool AllowUndef) |
225 | : Res(Res), AllowUndef(AllowUndef) {} |
226 | |
227 | template <typename ITy> bool match(ITy *V) { |
228 | if (auto *CI = dyn_cast<ConstantInt>(V)) { |
229 | Res = &CI->getValue(); |
230 | return true; |
231 | } |
232 | if (V->getType()->isVectorTy()) |
233 | if (const auto *C = dyn_cast<Constant>(V)) |
234 | if (auto *CI = dyn_cast_or_null<ConstantInt>( |
235 | C->getSplatValue(AllowUndef))) { |
236 | Res = &CI->getValue(); |
237 | return true; |
238 | } |
239 | return false; |
240 | } |
241 | }; |
242 | // Either constexpr if or renaming ConstantFP::getValueAPF to |
243 | // ConstantFP::getValue is needed to do it via single template |
244 | // function for both apint/apfloat. |
245 | struct apfloat_match { |
246 | const APFloat *&Res; |
247 | bool AllowUndef; |
248 | |
249 | apfloat_match(const APFloat *&Res, bool AllowUndef) |
250 | : Res(Res), AllowUndef(AllowUndef) {} |
251 | |
252 | template <typename ITy> bool match(ITy *V) { |
253 | if (auto *CI = dyn_cast<ConstantFP>(V)) { |
254 | Res = &CI->getValueAPF(); |
255 | return true; |
256 | } |
257 | if (V->getType()->isVectorTy()) |
258 | if (const auto *C = dyn_cast<Constant>(V)) |
259 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
260 | C->getSplatValue(AllowUndef))) { |
261 | Res = &CI->getValueAPF(); |
262 | return true; |
263 | } |
264 | return false; |
265 | } |
266 | }; |
267 | |
268 | /// Match a ConstantInt or splatted ConstantVector, binding the |
269 | /// specified pointer to the contained APInt. |
270 | inline apint_match m_APInt(const APInt *&Res) { |
271 | // Forbid undefs by default to maintain previous behavior. |
272 | return apint_match(Res, /* AllowUndef */ false); |
273 | } |
274 | |
275 | /// Match APInt while allowing undefs in splat vector constants. |
276 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { |
277 | return apint_match(Res, /* AllowUndef */ true); |
278 | } |
279 | |
280 | /// Match APInt while forbidding undefs in splat vector constants. |
281 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { |
282 | return apint_match(Res, /* AllowUndef */ false); |
283 | } |
284 | |
285 | /// Match a ConstantFP or splatted ConstantVector, binding the |
286 | /// specified pointer to the contained APFloat. |
287 | inline apfloat_match m_APFloat(const APFloat *&Res) { |
288 | // Forbid undefs by default to maintain previous behavior. |
289 | return apfloat_match(Res, /* AllowUndef */ false); |
290 | } |
291 | |
292 | /// Match APFloat while allowing undefs in splat vector constants. |
293 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { |
294 | return apfloat_match(Res, /* AllowUndef */ true); |
295 | } |
296 | |
297 | /// Match APFloat while forbidding undefs in splat vector constants. |
298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { |
299 | return apfloat_match(Res, /* AllowUndef */ false); |
300 | } |
301 | |
302 | template <int64_t Val> struct constantint_match { |
303 | template <typename ITy> bool match(ITy *V) { |
304 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { |
305 | const APInt &CIV = CI->getValue(); |
306 | if (Val >= 0) |
307 | return CIV == static_cast<uint64_t>(Val); |
308 | // If Val is negative, and CI is shorter than it, truncate to the right |
309 | // number of bits. If it is larger, then we have to sign extend. Just |
310 | // compare their negated values. |
311 | return -CIV == -Val; |
312 | } |
313 | return false; |
314 | } |
315 | }; |
316 | |
317 | /// Match a ConstantInt with a specific value. |
318 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { |
319 | return constantint_match<Val>(); |
320 | } |
321 | |
322 | /// This helper class is used to match constant scalars, vector splats, |
323 | /// and fixed width vectors that satisfy a specified predicate. |
324 | /// For fixed width vector constants, undefined elements are ignored. |
325 | template <typename Predicate, typename ConstantVal> |
326 | struct cstval_pred_ty : public Predicate { |
327 | template <typename ITy> bool match(ITy *V) { |
328 | if (const auto *CV = dyn_cast<ConstantVal>(V)) |
329 | return this->isValue(CV->getValue()); |
330 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { |
331 | if (const auto *C = dyn_cast<Constant>(V)) { |
332 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) |
333 | return this->isValue(CV->getValue()); |
334 | |
335 | // Number of elements of a scalable vector unknown at compile time |
336 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); |
337 | if (!FVTy) |
338 | return false; |
339 | |
340 | // Non-splat vector constant: check each element for a match. |
341 | unsigned NumElts = FVTy->getNumElements(); |
342 | assert(NumElts != 0 && "Constant vector with no elements?")((void)0); |
343 | bool HasNonUndefElements = false; |
344 | for (unsigned i = 0; i != NumElts; ++i) { |
345 | Constant *Elt = C->getAggregateElement(i); |
346 | if (!Elt) |
347 | return false; |
348 | if (isa<UndefValue>(Elt)) |
349 | continue; |
350 | auto *CV = dyn_cast<ConstantVal>(Elt); |
351 | if (!CV || !this->isValue(CV->getValue())) |
352 | return false; |
353 | HasNonUndefElements = true; |
354 | } |
355 | return HasNonUndefElements; |
356 | } |
357 | } |
358 | return false; |
359 | } |
360 | }; |
361 | |
362 | /// specialization of cstval_pred_ty for ConstantInt |
363 | template <typename Predicate> |
364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; |
365 | |
366 | /// specialization of cstval_pred_ty for ConstantFP |
367 | template <typename Predicate> |
368 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; |
369 | |
370 | /// This helper class is used to match scalar and vector constants that |
371 | /// satisfy a specified predicate, and bind them to an APInt. |
372 | template <typename Predicate> struct api_pred_ty : public Predicate { |
373 | const APInt *&Res; |
374 | |
375 | api_pred_ty(const APInt *&R) : Res(R) {} |
376 | |
377 | template <typename ITy> bool match(ITy *V) { |
378 | if (const auto *CI = dyn_cast<ConstantInt>(V)) |
379 | if (this->isValue(CI->getValue())) { |
380 | Res = &CI->getValue(); |
381 | return true; |
382 | } |
383 | if (V->getType()->isVectorTy()) |
384 | if (const auto *C = dyn_cast<Constant>(V)) |
385 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) |
386 | if (this->isValue(CI->getValue())) { |
387 | Res = &CI->getValue(); |
388 | return true; |
389 | } |
390 | |
391 | return false; |
392 | } |
393 | }; |
394 | |
395 | /// This helper class is used to match scalar and vector constants that |
396 | /// satisfy a specified predicate, and bind them to an APFloat. |
397 | /// Undefs are allowed in splat vector constants. |
398 | template <typename Predicate> struct apf_pred_ty : public Predicate { |
399 | const APFloat *&Res; |
400 | |
401 | apf_pred_ty(const APFloat *&R) : Res(R) {} |
402 | |
403 | template <typename ITy> bool match(ITy *V) { |
404 | if (const auto *CI = dyn_cast<ConstantFP>(V)) |
405 | if (this->isValue(CI->getValue())) { |
406 | Res = &CI->getValue(); |
407 | return true; |
408 | } |
409 | if (V->getType()->isVectorTy()) |
410 | if (const auto *C = dyn_cast<Constant>(V)) |
411 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
412 | C->getSplatValue(/* AllowUndef */ true))) |
413 | if (this->isValue(CI->getValue())) { |
414 | Res = &CI->getValue(); |
415 | return true; |
416 | } |
417 | |
418 | return false; |
419 | } |
420 | }; |
421 | |
422 | /////////////////////////////////////////////////////////////////////////////// |
423 | // |
424 | // Encapsulate constant value queries for use in templated predicate matchers. |
425 | // This allows checking if constants match using compound predicates and works |
426 | // with vector constants, possibly with relaxed constraints. For example, ignore |
427 | // undef values. |
428 | // |
429 | /////////////////////////////////////////////////////////////////////////////// |
430 | |
431 | struct is_any_apint { |
432 | bool isValue(const APInt &C) { return true; } |
433 | }; |
434 | /// Match an integer or vector with any integral constant. |
435 | /// For vectors, this includes constants with undefined elements. |
436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { |
437 | return cst_pred_ty<is_any_apint>(); |
438 | } |
439 | |
440 | struct is_all_ones { |
441 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } |
442 | }; |
443 | /// Match an integer or vector with all bits set. |
444 | /// For vectors, this includes constants with undefined elements. |
445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { |
446 | return cst_pred_ty<is_all_ones>(); |
447 | } |
448 | |
449 | struct is_maxsignedvalue { |
450 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } |
451 | }; |
452 | /// Match an integer or vector with values having all bits except for the high |
453 | /// bit set (0x7f...). |
454 | /// For vectors, this includes constants with undefined elements. |
455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { |
456 | return cst_pred_ty<is_maxsignedvalue>(); |
457 | } |
458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { |
459 | return V; |
460 | } |
461 | |
462 | struct is_negative { |
463 | bool isValue(const APInt &C) { return C.isNegative(); } |
464 | }; |
465 | /// Match an integer or vector of negative values. |
466 | /// For vectors, this includes constants with undefined elements. |
467 | inline cst_pred_ty<is_negative> m_Negative() { |
468 | return cst_pred_ty<is_negative>(); |
469 | } |
470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { |
471 | return V; |
472 | } |
473 | |
474 | struct is_nonnegative { |
475 | bool isValue(const APInt &C) { return C.isNonNegative(); } |
476 | }; |
477 | /// Match an integer or vector of non-negative values. |
478 | /// For vectors, this includes constants with undefined elements. |
479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { |
480 | return cst_pred_ty<is_nonnegative>(); |
481 | } |
482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { |
483 | return V; |
484 | } |
485 | |
486 | struct is_strictlypositive { |
487 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } |
488 | }; |
489 | /// Match an integer or vector of strictly positive values. |
490 | /// For vectors, this includes constants with undefined elements. |
491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { |
492 | return cst_pred_ty<is_strictlypositive>(); |
493 | } |
494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { |
495 | return V; |
496 | } |
497 | |
498 | struct is_nonpositive { |
499 | bool isValue(const APInt &C) { return C.isNonPositive(); } |
500 | }; |
501 | /// Match an integer or vector of non-positive values. |
502 | /// For vectors, this includes constants with undefined elements. |
503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { |
504 | return cst_pred_ty<is_nonpositive>(); |
505 | } |
506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } |
507 | |
508 | struct is_one { |
509 | bool isValue(const APInt &C) { return C.isOneValue(); } |
510 | }; |
511 | /// Match an integer 1 or a vector with all elements equal to 1. |
512 | /// For vectors, this includes constants with undefined elements. |
513 | inline cst_pred_ty<is_one> m_One() { |
514 | return cst_pred_ty<is_one>(); |
515 | } |
516 | |
517 | struct is_zero_int { |
518 | bool isValue(const APInt &C) { return C.isNullValue(); } |
519 | }; |
520 | /// Match an integer 0 or a vector with all elements equal to 0. |
521 | /// For vectors, this includes constants with undefined elements. |
522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { |
523 | return cst_pred_ty<is_zero_int>(); |
524 | } |
525 | |
526 | struct is_zero { |
527 | template <typename ITy> bool match(ITy *V) { |
528 | auto *C = dyn_cast<Constant>(V); |
529 | // FIXME: this should be able to do something for scalable vectors |
530 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); |
531 | } |
532 | }; |
533 | /// Match any null constant or a vector with all elements equal to 0. |
534 | /// For vectors, this includes constants with undefined elements. |
535 | inline is_zero m_Zero() { |
536 | return is_zero(); |
537 | } |
538 | |
539 | struct is_power2 { |
540 | bool isValue(const APInt &C) { return C.isPowerOf2(); } |
541 | }; |
542 | /// Match an integer or vector power-of-2. |
543 | /// For vectors, this includes constants with undefined elements. |
544 | inline cst_pred_ty<is_power2> m_Power2() { |
545 | return cst_pred_ty<is_power2>(); |
546 | } |
547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { |
548 | return V; |
549 | } |
550 | |
551 | struct is_negated_power2 { |
552 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } |
553 | }; |
554 | /// Match a integer or vector negated power-of-2. |
555 | /// For vectors, this includes constants with undefined elements. |
556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { |
557 | return cst_pred_ty<is_negated_power2>(); |
558 | } |
559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { |
560 | return V; |
561 | } |
562 | |
563 | struct is_power2_or_zero { |
564 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } |