File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp |
Warning: | line 605, column 5 Value stored to 'ToRemoveStart' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- DeadStoreElimination.cpp - MemorySSA Backed Dead Store Elimination -===// |
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 | // The code below implements dead store elimination using MemorySSA. It uses |
10 | // the following general approach: given a MemoryDef, walk upwards to find |
11 | // clobbering MemoryDefs that may be killed by the starting def. Then check |
12 | // that there are no uses that may read the location of the original MemoryDef |
13 | // in between both MemoryDefs. A bit more concretely: |
14 | // |
15 | // For all MemoryDefs StartDef: |
16 | // 1. Get the next dominating clobbering MemoryDef (EarlierAccess) by walking |
17 | // upwards. |
18 | // 2. Check that there are no reads between EarlierAccess and the StartDef by |
19 | // checking all uses starting at EarlierAccess and walking until we see |
20 | // StartDef. |
21 | // 3. For each found CurrentDef, check that: |
22 | // 1. There are no barrier instructions between CurrentDef and StartDef (like |
23 | // throws or stores with ordering constraints). |
24 | // 2. StartDef is executed whenever CurrentDef is executed. |
25 | // 3. StartDef completely overwrites CurrentDef. |
26 | // 4. Erase CurrentDef from the function and MemorySSA. |
27 | // |
28 | //===----------------------------------------------------------------------===// |
29 | |
30 | #include "llvm/Transforms/Scalar/DeadStoreElimination.h" |
31 | #include "llvm/ADT/APInt.h" |
32 | #include "llvm/ADT/DenseMap.h" |
33 | #include "llvm/ADT/MapVector.h" |
34 | #include "llvm/ADT/PostOrderIterator.h" |
35 | #include "llvm/ADT/SetVector.h" |
36 | #include "llvm/ADT/SmallPtrSet.h" |
37 | #include "llvm/ADT/SmallVector.h" |
38 | #include "llvm/ADT/Statistic.h" |
39 | #include "llvm/ADT/StringRef.h" |
40 | #include "llvm/Analysis/AliasAnalysis.h" |
41 | #include "llvm/Analysis/CaptureTracking.h" |
42 | #include "llvm/Analysis/GlobalsModRef.h" |
43 | #include "llvm/Analysis/LoopInfo.h" |
44 | #include "llvm/Analysis/MemoryBuiltins.h" |
45 | #include "llvm/Analysis/MemoryLocation.h" |
46 | #include "llvm/Analysis/MemorySSA.h" |
47 | #include "llvm/Analysis/MemorySSAUpdater.h" |
48 | #include "llvm/Analysis/MustExecute.h" |
49 | #include "llvm/Analysis/PostDominators.h" |
50 | #include "llvm/Analysis/TargetLibraryInfo.h" |
51 | #include "llvm/Analysis/ValueTracking.h" |
52 | #include "llvm/IR/Argument.h" |
53 | #include "llvm/IR/BasicBlock.h" |
54 | #include "llvm/IR/Constant.h" |
55 | #include "llvm/IR/Constants.h" |
56 | #include "llvm/IR/DataLayout.h" |
57 | #include "llvm/IR/Dominators.h" |
58 | #include "llvm/IR/Function.h" |
59 | #include "llvm/IR/InstIterator.h" |
60 | #include "llvm/IR/InstrTypes.h" |
61 | #include "llvm/IR/Instruction.h" |
62 | #include "llvm/IR/Instructions.h" |
63 | #include "llvm/IR/IntrinsicInst.h" |
64 | #include "llvm/IR/Intrinsics.h" |
65 | #include "llvm/IR/LLVMContext.h" |
66 | #include "llvm/IR/Module.h" |
67 | #include "llvm/IR/PassManager.h" |
68 | #include "llvm/IR/PatternMatch.h" |
69 | #include "llvm/IR/Value.h" |
70 | #include "llvm/InitializePasses.h" |
71 | #include "llvm/Pass.h" |
72 | #include "llvm/Support/Casting.h" |
73 | #include "llvm/Support/CommandLine.h" |
74 | #include "llvm/Support/Debug.h" |
75 | #include "llvm/Support/DebugCounter.h" |
76 | #include "llvm/Support/ErrorHandling.h" |
77 | #include "llvm/Support/MathExtras.h" |
78 | #include "llvm/Support/raw_ostream.h" |
79 | #include "llvm/Transforms/Scalar.h" |
80 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
81 | #include "llvm/Transforms/Utils/Local.h" |
82 | #include <algorithm> |
83 | #include <cassert> |
84 | #include <cstddef> |
85 | #include <cstdint> |
86 | #include <iterator> |
87 | #include <map> |
88 | #include <utility> |
89 | |
90 | using namespace llvm; |
91 | using namespace PatternMatch; |
92 | |
93 | #define DEBUG_TYPE"dse" "dse" |
94 | |
95 | STATISTIC(NumRemainingStores, "Number of stores remaining after DSE")static llvm::Statistic NumRemainingStores = {"dse", "NumRemainingStores" , "Number of stores remaining after DSE"}; |
96 | STATISTIC(NumRedundantStores, "Number of redundant stores deleted")static llvm::Statistic NumRedundantStores = {"dse", "NumRedundantStores" , "Number of redundant stores deleted"}; |
97 | STATISTIC(NumFastStores, "Number of stores deleted")static llvm::Statistic NumFastStores = {"dse", "NumFastStores" , "Number of stores deleted"}; |
98 | STATISTIC(NumFastOther, "Number of other instrs removed")static llvm::Statistic NumFastOther = {"dse", "NumFastOther", "Number of other instrs removed"}; |
99 | STATISTIC(NumCompletePartials, "Number of stores dead by later partials")static llvm::Statistic NumCompletePartials = {"dse", "NumCompletePartials" , "Number of stores dead by later partials"}; |
100 | STATISTIC(NumModifiedStores, "Number of stores modified")static llvm::Statistic NumModifiedStores = {"dse", "NumModifiedStores" , "Number of stores modified"}; |
101 | STATISTIC(NumCFGChecks, "Number of stores modified")static llvm::Statistic NumCFGChecks = {"dse", "NumCFGChecks", "Number of stores modified"}; |
102 | STATISTIC(NumCFGTries, "Number of stores modified")static llvm::Statistic NumCFGTries = {"dse", "NumCFGTries", "Number of stores modified" }; |
103 | STATISTIC(NumCFGSuccess, "Number of stores modified")static llvm::Statistic NumCFGSuccess = {"dse", "NumCFGSuccess" , "Number of stores modified"}; |
104 | STATISTIC(NumGetDomMemoryDefPassed,static llvm::Statistic NumGetDomMemoryDefPassed = {"dse", "NumGetDomMemoryDefPassed" , "Number of times a valid candidate is returned from getDomMemoryDef" } |
105 | "Number of times a valid candidate is returned from getDomMemoryDef")static llvm::Statistic NumGetDomMemoryDefPassed = {"dse", "NumGetDomMemoryDefPassed" , "Number of times a valid candidate is returned from getDomMemoryDef" }; |
106 | STATISTIC(NumDomMemDefChecks,static llvm::Statistic NumDomMemDefChecks = {"dse", "NumDomMemDefChecks" , "Number iterations check for reads in getDomMemoryDef"} |
107 | "Number iterations check for reads in getDomMemoryDef")static llvm::Statistic NumDomMemDefChecks = {"dse", "NumDomMemDefChecks" , "Number iterations check for reads in getDomMemoryDef"}; |
108 | |
109 | DEBUG_COUNTER(MemorySSACounter, "dse-memoryssa",static const unsigned MemorySSACounter = DebugCounter::registerCounter ("dse-memoryssa", "Controls which MemoryDefs are eliminated." ) |
110 | "Controls which MemoryDefs are eliminated.")static const unsigned MemorySSACounter = DebugCounter::registerCounter ("dse-memoryssa", "Controls which MemoryDefs are eliminated." ); |
111 | |
112 | static cl::opt<bool> |
113 | EnablePartialOverwriteTracking("enable-dse-partial-overwrite-tracking", |
114 | cl::init(true), cl::Hidden, |
115 | cl::desc("Enable partial-overwrite tracking in DSE")); |
116 | |
117 | static cl::opt<bool> |
118 | EnablePartialStoreMerging("enable-dse-partial-store-merging", |
119 | cl::init(true), cl::Hidden, |
120 | cl::desc("Enable partial store merging in DSE")); |
121 | |
122 | static cl::opt<unsigned> |
123 | MemorySSAScanLimit("dse-memoryssa-scanlimit", cl::init(150), cl::Hidden, |
124 | cl::desc("The number of memory instructions to scan for " |
125 | "dead store elimination (default = 100)")); |
126 | static cl::opt<unsigned> MemorySSAUpwardsStepLimit( |
127 | "dse-memoryssa-walklimit", cl::init(90), cl::Hidden, |
128 | cl::desc("The maximum number of steps while walking upwards to find " |
129 | "MemoryDefs that may be killed (default = 90)")); |
130 | |
131 | static cl::opt<unsigned> MemorySSAPartialStoreLimit( |
132 | "dse-memoryssa-partial-store-limit", cl::init(5), cl::Hidden, |
133 | cl::desc("The maximum number candidates that only partially overwrite the " |
134 | "killing MemoryDef to consider" |
135 | " (default = 5)")); |
136 | |
137 | static cl::opt<unsigned> MemorySSADefsPerBlockLimit( |
138 | "dse-memoryssa-defs-per-block-limit", cl::init(5000), cl::Hidden, |
139 | cl::desc("The number of MemoryDefs we consider as candidates to eliminated " |
140 | "other stores per basic block (default = 5000)")); |
141 | |
142 | static cl::opt<unsigned> MemorySSASameBBStepCost( |
143 | "dse-memoryssa-samebb-cost", cl::init(1), cl::Hidden, |
144 | cl::desc( |
145 | "The cost of a step in the same basic block as the killing MemoryDef" |
146 | "(default = 1)")); |
147 | |
148 | static cl::opt<unsigned> |
149 | MemorySSAOtherBBStepCost("dse-memoryssa-otherbb-cost", cl::init(5), |
150 | cl::Hidden, |
151 | cl::desc("The cost of a step in a different basic " |
152 | "block than the killing MemoryDef" |
153 | "(default = 5)")); |
154 | |
155 | static cl::opt<unsigned> MemorySSAPathCheckLimit( |
156 | "dse-memoryssa-path-check-limit", cl::init(50), cl::Hidden, |
157 | cl::desc("The maximum number of blocks to check when trying to prove that " |
158 | "all paths to an exit go through a killing block (default = 50)")); |
159 | |
160 | //===----------------------------------------------------------------------===// |
161 | // Helper functions |
162 | //===----------------------------------------------------------------------===// |
163 | using OverlapIntervalsTy = std::map<int64_t, int64_t>; |
164 | using InstOverlapIntervalsTy = DenseMap<Instruction *, OverlapIntervalsTy>; |
165 | |
166 | /// Does this instruction write some memory? This only returns true for things |
167 | /// that we can analyze with other helpers below. |
168 | static bool hasAnalyzableMemoryWrite(Instruction *I, |
169 | const TargetLibraryInfo &TLI) { |
170 | if (isa<StoreInst>(I)) |
171 | return true; |
172 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
173 | switch (II->getIntrinsicID()) { |
174 | default: |
175 | return false; |
176 | case Intrinsic::memset: |
177 | case Intrinsic::memmove: |
178 | case Intrinsic::memcpy: |
179 | case Intrinsic::memcpy_inline: |
180 | case Intrinsic::memcpy_element_unordered_atomic: |
181 | case Intrinsic::memmove_element_unordered_atomic: |
182 | case Intrinsic::memset_element_unordered_atomic: |
183 | case Intrinsic::init_trampoline: |
184 | case Intrinsic::lifetime_end: |
185 | case Intrinsic::masked_store: |
186 | return true; |
187 | } |
188 | } |
189 | if (auto *CB = dyn_cast<CallBase>(I)) { |
190 | LibFunc LF; |
191 | if (TLI.getLibFunc(*CB, LF) && TLI.has(LF)) { |
192 | switch (LF) { |
193 | case LibFunc_strcpy: |
194 | case LibFunc_strncpy: |
195 | case LibFunc_strcat: |
196 | case LibFunc_strncat: |
197 | return true; |
198 | default: |
199 | return false; |
200 | } |
201 | } |
202 | } |
203 | return false; |
204 | } |
205 | |
206 | /// Return a Location stored to by the specified instruction. If isRemovable |
207 | /// returns true, this function and getLocForRead completely describe the memory |
208 | /// operations for this instruction. |
209 | static MemoryLocation getLocForWrite(Instruction *Inst, |
210 | const TargetLibraryInfo &TLI) { |
211 | if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) |
212 | return MemoryLocation::get(SI); |
213 | |
214 | // memcpy/memmove/memset. |
215 | if (auto *MI = dyn_cast<AnyMemIntrinsic>(Inst)) |
216 | return MemoryLocation::getForDest(MI); |
217 | |
218 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { |
219 | switch (II->getIntrinsicID()) { |
220 | default: |
221 | return MemoryLocation(); // Unhandled intrinsic. |
222 | case Intrinsic::init_trampoline: |
223 | return MemoryLocation::getAfter(II->getArgOperand(0)); |
224 | case Intrinsic::masked_store: |
225 | return MemoryLocation::getForArgument(II, 1, TLI); |
226 | case Intrinsic::lifetime_end: { |
227 | uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); |
228 | return MemoryLocation(II->getArgOperand(1), Len); |
229 | } |
230 | } |
231 | } |
232 | if (auto *CB = dyn_cast<CallBase>(Inst)) |
233 | // All the supported TLI functions so far happen to have dest as their |
234 | // first argument. |
235 | return MemoryLocation::getAfter(CB->getArgOperand(0)); |
236 | return MemoryLocation(); |
237 | } |
238 | |
239 | /// If the value of this instruction and the memory it writes to is unused, may |
240 | /// we delete this instruction? |
241 | static bool isRemovable(Instruction *I) { |
242 | // Don't remove volatile/atomic stores. |
243 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) |
244 | return SI->isUnordered(); |
245 | |
246 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
247 | switch (II->getIntrinsicID()) { |
248 | default: llvm_unreachable("doesn't pass 'hasAnalyzableMemoryWrite' predicate")__builtin_unreachable(); |
249 | case Intrinsic::lifetime_end: |
250 | // Never remove dead lifetime_end's, e.g. because it is followed by a |
251 | // free. |
252 | return false; |
253 | case Intrinsic::init_trampoline: |
254 | // Always safe to remove init_trampoline. |
255 | return true; |
256 | case Intrinsic::memset: |
257 | case Intrinsic::memmove: |
258 | case Intrinsic::memcpy: |
259 | case Intrinsic::memcpy_inline: |
260 | // Don't remove volatile memory intrinsics. |
261 | return !cast<MemIntrinsic>(II)->isVolatile(); |
262 | case Intrinsic::memcpy_element_unordered_atomic: |
263 | case Intrinsic::memmove_element_unordered_atomic: |
264 | case Intrinsic::memset_element_unordered_atomic: |
265 | case Intrinsic::masked_store: |
266 | return true; |
267 | } |
268 | } |
269 | |
270 | // note: only get here for calls with analyzable writes - i.e. libcalls |
271 | if (auto *CB = dyn_cast<CallBase>(I)) |
272 | return CB->use_empty(); |
273 | |
274 | return false; |
275 | } |
276 | |
277 | /// Returns true if the end of this instruction can be safely shortened in |
278 | /// length. |
279 | static bool isShortenableAtTheEnd(Instruction *I) { |
280 | // Don't shorten stores for now |
281 | if (isa<StoreInst>(I)) |
282 | return false; |
283 | |
284 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
285 | switch (II->getIntrinsicID()) { |
286 | default: return false; |
287 | case Intrinsic::memset: |
288 | case Intrinsic::memcpy: |
289 | case Intrinsic::memcpy_element_unordered_atomic: |
290 | case Intrinsic::memset_element_unordered_atomic: |
291 | // Do shorten memory intrinsics. |
292 | // FIXME: Add memmove if it's also safe to transform. |
293 | return true; |
294 | } |
295 | } |
296 | |
297 | // Don't shorten libcalls calls for now. |
298 | |
299 | return false; |
300 | } |
301 | |
302 | /// Returns true if the beginning of this instruction can be safely shortened |
303 | /// in length. |
304 | static bool isShortenableAtTheBeginning(Instruction *I) { |
305 | // FIXME: Handle only memset for now. Supporting memcpy/memmove should be |
306 | // easily done by offsetting the source address. |
307 | return isa<AnyMemSetInst>(I); |
308 | } |
309 | |
310 | static uint64_t getPointerSize(const Value *V, const DataLayout &DL, |
311 | const TargetLibraryInfo &TLI, |
312 | const Function *F) { |
313 | uint64_t Size; |
314 | ObjectSizeOpts Opts; |
315 | Opts.NullIsUnknownSize = NullPointerIsDefined(F); |
316 | |
317 | if (getObjectSize(V, Size, DL, &TLI, Opts)) |
318 | return Size; |
319 | return MemoryLocation::UnknownSize; |
320 | } |
321 | |
322 | namespace { |
323 | |
324 | enum OverwriteResult { |
325 | OW_Begin, |
326 | OW_Complete, |
327 | OW_End, |
328 | OW_PartialEarlierWithFullLater, |
329 | OW_MaybePartial, |
330 | OW_Unknown |
331 | }; |
332 | |
333 | } // end anonymous namespace |
334 | |
335 | /// Check if two instruction are masked stores that completely |
336 | /// overwrite one another. More specifically, \p Later has to |
337 | /// overwrite \p Earlier. |
338 | static OverwriteResult isMaskedStoreOverwrite(const Instruction *Later, |
339 | const Instruction *Earlier, |
340 | BatchAAResults &AA) { |
341 | const auto *IIL = dyn_cast<IntrinsicInst>(Later); |
342 | const auto *IIE = dyn_cast<IntrinsicInst>(Earlier); |
343 | if (IIL == nullptr || IIE == nullptr) |
344 | return OW_Unknown; |
345 | if (IIL->getIntrinsicID() != Intrinsic::masked_store || |
346 | IIE->getIntrinsicID() != Intrinsic::masked_store) |
347 | return OW_Unknown; |
348 | // Pointers. |
349 | Value *LP = IIL->getArgOperand(1)->stripPointerCasts(); |
350 | Value *EP = IIE->getArgOperand(1)->stripPointerCasts(); |
351 | if (LP != EP && !AA.isMustAlias(LP, EP)) |
352 | return OW_Unknown; |
353 | // Masks. |
354 | // TODO: check that Later's mask is a superset of the Earlier's mask. |
355 | if (IIL->getArgOperand(3) != IIE->getArgOperand(3)) |
356 | return OW_Unknown; |
357 | return OW_Complete; |
358 | } |
359 | |
360 | /// Return 'OW_Complete' if a store to the 'Later' location completely |
361 | /// overwrites a store to the 'Earlier' location, 'OW_End' if the end of the |
362 | /// 'Earlier' location is completely overwritten by 'Later', 'OW_Begin' if the |
363 | /// beginning of the 'Earlier' location is overwritten by 'Later'. |
364 | /// 'OW_PartialEarlierWithFullLater' means that an earlier (big) store was |
365 | /// overwritten by a latter (smaller) store which doesn't write outside the big |
366 | /// store's memory locations. Returns 'OW_Unknown' if nothing can be determined. |
367 | /// NOTE: This function must only be called if both \p Later and \p Earlier |
368 | /// write to the same underlying object with valid \p EarlierOff and \p |
369 | /// LaterOff. |
370 | static OverwriteResult isPartialOverwrite(const MemoryLocation &Later, |
371 | const MemoryLocation &Earlier, |
372 | int64_t EarlierOff, int64_t LaterOff, |
373 | Instruction *DepWrite, |
374 | InstOverlapIntervalsTy &IOL) { |
375 | const uint64_t LaterSize = Later.Size.getValue(); |
376 | const uint64_t EarlierSize = Earlier.Size.getValue(); |
377 | // We may now overlap, although the overlap is not complete. There might also |
378 | // be other incomplete overlaps, and together, they might cover the complete |
379 | // earlier write. |
380 | // Note: The correctness of this logic depends on the fact that this function |
381 | // is not even called providing DepWrite when there are any intervening reads. |
382 | if (EnablePartialOverwriteTracking && |
383 | LaterOff < int64_t(EarlierOff + EarlierSize) && |
384 | int64_t(LaterOff + LaterSize) >= EarlierOff) { |
385 | |
386 | // Insert our part of the overlap into the map. |
387 | auto &IM = IOL[DepWrite]; |
388 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOffdo { } while (false) |
389 | << ", " << int64_t(EarlierOff + EarlierSize)do { } while (false) |
390 | << ") Later [" << LaterOff << ", "do { } while (false) |
391 | << int64_t(LaterOff + LaterSize) << ")\n")do { } while (false); |
392 | |
393 | // Make sure that we only insert non-overlapping intervals and combine |
394 | // adjacent intervals. The intervals are stored in the map with the ending |
395 | // offset as the key (in the half-open sense) and the starting offset as |
396 | // the value. |
397 | int64_t LaterIntStart = LaterOff, LaterIntEnd = LaterOff + LaterSize; |
398 | |
399 | // Find any intervals ending at, or after, LaterIntStart which start |
400 | // before LaterIntEnd. |
401 | auto ILI = IM.lower_bound(LaterIntStart); |
402 | if (ILI != IM.end() && ILI->second <= LaterIntEnd) { |
403 | // This existing interval is overlapped with the current store somewhere |
404 | // in [LaterIntStart, LaterIntEnd]. Merge them by erasing the existing |
405 | // intervals and adjusting our start and end. |
406 | LaterIntStart = std::min(LaterIntStart, ILI->second); |
407 | LaterIntEnd = std::max(LaterIntEnd, ILI->first); |
408 | ILI = IM.erase(ILI); |
409 | |
410 | // Continue erasing and adjusting our end in case other previous |
411 | // intervals are also overlapped with the current store. |
412 | // |
413 | // |--- ealier 1 ---| |--- ealier 2 ---| |
414 | // |------- later---------| |
415 | // |
416 | while (ILI != IM.end() && ILI->second <= LaterIntEnd) { |
417 | assert(ILI->second > LaterIntStart && "Unexpected interval")((void)0); |
418 | LaterIntEnd = std::max(LaterIntEnd, ILI->first); |
419 | ILI = IM.erase(ILI); |
420 | } |
421 | } |
422 | |
423 | IM[LaterIntEnd] = LaterIntStart; |
424 | |
425 | ILI = IM.begin(); |
426 | if (ILI->second <= EarlierOff && |
427 | ILI->first >= int64_t(EarlierOff + EarlierSize)) { |
428 | LLVM_DEBUG(dbgs() << "DSE: Full overwrite from partials: Earlier ["do { } while (false) |
429 | << EarlierOff << ", "do { } while (false) |
430 | << int64_t(EarlierOff + EarlierSize)do { } while (false) |
431 | << ") Composite Later [" << ILI->second << ", "do { } while (false) |
432 | << ILI->first << ")\n")do { } while (false); |
433 | ++NumCompletePartials; |
434 | return OW_Complete; |
435 | } |
436 | } |
437 | |
438 | // Check for an earlier store which writes to all the memory locations that |
439 | // the later store writes to. |
440 | if (EnablePartialStoreMerging && LaterOff >= EarlierOff && |
441 | int64_t(EarlierOff + EarlierSize) > LaterOff && |
442 | uint64_t(LaterOff - EarlierOff) + LaterSize <= EarlierSize) { |
443 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite an earlier load ["do { } while (false) |
444 | << EarlierOff << ", "do { } while (false) |
445 | << int64_t(EarlierOff + EarlierSize)do { } while (false) |
446 | << ") by a later store [" << LaterOff << ", "do { } while (false) |
447 | << int64_t(LaterOff + LaterSize) << ")\n")do { } while (false); |
448 | // TODO: Maybe come up with a better name? |
449 | return OW_PartialEarlierWithFullLater; |
450 | } |
451 | |
452 | // Another interesting case is if the later store overwrites the end of the |
453 | // earlier store. |
454 | // |
455 | // |--earlier--| |
456 | // |-- later --| |
457 | // |
458 | // In this case we may want to trim the size of earlier to avoid generating |
459 | // writes to addresses which will definitely be overwritten later |
460 | if (!EnablePartialOverwriteTracking && |
461 | (LaterOff > EarlierOff && LaterOff < int64_t(EarlierOff + EarlierSize) && |
462 | int64_t(LaterOff + LaterSize) >= int64_t(EarlierOff + EarlierSize))) |
463 | return OW_End; |
464 | |
465 | // Finally, we also need to check if the later store overwrites the beginning |
466 | // of the earlier store. |
467 | // |
468 | // |--earlier--| |
469 | // |-- later --| |
470 | // |
471 | // In this case we may want to move the destination address and trim the size |
472 | // of earlier to avoid generating writes to addresses which will definitely |
473 | // be overwritten later. |
474 | if (!EnablePartialOverwriteTracking && |
475 | (LaterOff <= EarlierOff && int64_t(LaterOff + LaterSize) > EarlierOff)) { |
476 | assert(int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize) &&((void)0) |
477 | "Expect to be handled as OW_Complete")((void)0); |
478 | return OW_Begin; |
479 | } |
480 | // Otherwise, they don't completely overlap. |
481 | return OW_Unknown; |
482 | } |
483 | |
484 | /// Returns true if the memory which is accessed by the second instruction is not |
485 | /// modified between the first and the second instruction. |
486 | /// Precondition: Second instruction must be dominated by the first |
487 | /// instruction. |
488 | static bool |
489 | memoryIsNotModifiedBetween(Instruction *FirstI, Instruction *SecondI, |
490 | BatchAAResults &AA, const DataLayout &DL, |
491 | DominatorTree *DT) { |
492 | // Do a backwards scan through the CFG from SecondI to FirstI. Look for |
493 | // instructions which can modify the memory location accessed by SecondI. |
494 | // |
495 | // While doing the walk keep track of the address to check. It might be |
496 | // different in different basic blocks due to PHI translation. |
497 | using BlockAddressPair = std::pair<BasicBlock *, PHITransAddr>; |
498 | SmallVector<BlockAddressPair, 16> WorkList; |
499 | // Keep track of the address we visited each block with. Bail out if we |
500 | // visit a block with different addresses. |
501 | DenseMap<BasicBlock *, Value *> Visited; |
502 | |
503 | BasicBlock::iterator FirstBBI(FirstI); |
504 | ++FirstBBI; |
505 | BasicBlock::iterator SecondBBI(SecondI); |
506 | BasicBlock *FirstBB = FirstI->getParent(); |
507 | BasicBlock *SecondBB = SecondI->getParent(); |
508 | MemoryLocation MemLoc = MemoryLocation::get(SecondI); |
509 | auto *MemLocPtr = const_cast<Value *>(MemLoc.Ptr); |
510 | |
511 | // Start checking the SecondBB. |
512 | WorkList.push_back( |
513 | std::make_pair(SecondBB, PHITransAddr(MemLocPtr, DL, nullptr))); |
514 | bool isFirstBlock = true; |
515 | |
516 | // Check all blocks going backward until we reach the FirstBB. |
517 | while (!WorkList.empty()) { |
518 | BlockAddressPair Current = WorkList.pop_back_val(); |
519 | BasicBlock *B = Current.first; |
520 | PHITransAddr &Addr = Current.second; |
521 | Value *Ptr = Addr.getAddr(); |
522 | |
523 | // Ignore instructions before FirstI if this is the FirstBB. |
524 | BasicBlock::iterator BI = (B == FirstBB ? FirstBBI : B->begin()); |
525 | |
526 | BasicBlock::iterator EI; |
527 | if (isFirstBlock) { |
528 | // Ignore instructions after SecondI if this is the first visit of SecondBB. |
529 | assert(B == SecondBB && "first block is not the store block")((void)0); |
530 | EI = SecondBBI; |
531 | isFirstBlock = false; |
532 | } else { |
533 | // It's not SecondBB or (in case of a loop) the second visit of SecondBB. |
534 | // In this case we also have to look at instructions after SecondI. |
535 | EI = B->end(); |
536 | } |
537 | for (; BI != EI; ++BI) { |
538 | Instruction *I = &*BI; |
539 | if (I->mayWriteToMemory() && I != SecondI) |
540 | if (isModSet(AA.getModRefInfo(I, MemLoc.getWithNewPtr(Ptr)))) |
541 | return false; |
542 | } |
543 | if (B != FirstBB) { |
544 | assert(B != &FirstBB->getParent()->getEntryBlock() &&((void)0) |
545 | "Should not hit the entry block because SI must be dominated by LI")((void)0); |
546 | for (BasicBlock *Pred : predecessors(B)) { |
547 | PHITransAddr PredAddr = Addr; |
548 | if (PredAddr.NeedsPHITranslationFromBlock(B)) { |
549 | if (!PredAddr.IsPotentiallyPHITranslatable()) |
550 | return false; |
551 | if (PredAddr.PHITranslateValue(B, Pred, DT, false)) |
552 | return false; |
553 | } |
554 | Value *TranslatedPtr = PredAddr.getAddr(); |
555 | auto Inserted = Visited.insert(std::make_pair(Pred, TranslatedPtr)); |
556 | if (!Inserted.second) { |
557 | // We already visited this block before. If it was with a different |
558 | // address - bail out! |
559 | if (TranslatedPtr != Inserted.first->second) |
560 | return false; |
561 | // ... otherwise just skip it. |
562 | continue; |
563 | } |
564 | WorkList.push_back(std::make_pair(Pred, PredAddr)); |
565 | } |
566 | } |
567 | } |
568 | return true; |
569 | } |
570 | |
571 | static bool tryToShorten(Instruction *EarlierWrite, int64_t &EarlierStart, |
572 | uint64_t &EarlierSize, int64_t LaterStart, |
573 | uint64_t LaterSize, bool IsOverwriteEnd) { |
574 | auto *EarlierIntrinsic = cast<AnyMemIntrinsic>(EarlierWrite); |
575 | Align PrefAlign = EarlierIntrinsic->getDestAlign().valueOrOne(); |
576 | |
577 | // We assume that memet/memcpy operates in chunks of the "largest" native |
578 | // type size and aligned on the same value. That means optimal start and size |
579 | // of memset/memcpy should be modulo of preferred alignment of that type. That |
580 | // is it there is no any sense in trying to reduce store size any further |
581 | // since any "extra" stores comes for free anyway. |
582 | // On the other hand, maximum alignment we can achieve is limited by alignment |
583 | // of initial store. |
584 | |
585 | // TODO: Limit maximum alignment by preferred (or abi?) alignment of the |
586 | // "largest" native type. |
587 | // Note: What is the proper way to get that value? |
588 | // Should TargetTransformInfo::getRegisterBitWidth be used or anything else? |
589 | // PrefAlign = std::min(DL.getPrefTypeAlign(LargestType), PrefAlign); |
590 | |
591 | int64_t ToRemoveStart = 0; |
592 | uint64_t ToRemoveSize = 0; |
593 | // Compute start and size of the region to remove. Make sure 'PrefAlign' is |
594 | // maintained on the remaining store. |
595 | if (IsOverwriteEnd) { |
596 | // Calculate required adjustment for 'LaterStart'in order to keep remaining |
597 | // store size aligned on 'PerfAlign'. |
598 | uint64_t Off = |
599 | offsetToAlignment(uint64_t(LaterStart - EarlierStart), PrefAlign); |
600 | ToRemoveStart = LaterStart + Off; |
601 | if (EarlierSize <= uint64_t(ToRemoveStart - EarlierStart)) |
602 | return false; |
603 | ToRemoveSize = EarlierSize - uint64_t(ToRemoveStart - EarlierStart); |
604 | } else { |
605 | ToRemoveStart = EarlierStart; |
Value stored to 'ToRemoveStart' is never read | |
606 | assert(LaterSize >= uint64_t(EarlierStart - LaterStart) &&((void)0) |
607 | "Not overlapping accesses?")((void)0); |
608 | ToRemoveSize = LaterSize - uint64_t(EarlierStart - LaterStart); |
609 | // Calculate required adjustment for 'ToRemoveSize'in order to keep |
610 | // start of the remaining store aligned on 'PerfAlign'. |
611 | uint64_t Off = offsetToAlignment(ToRemoveSize, PrefAlign); |
612 | if (Off != 0) { |
613 | if (ToRemoveSize <= (PrefAlign.value() - Off)) |
614 | return false; |
615 | ToRemoveSize -= PrefAlign.value() - Off; |
616 | } |
617 | assert(isAligned(PrefAlign, ToRemoveSize) &&((void)0) |
618 | "Should preserve selected alignment")((void)0); |
619 | } |
620 | |
621 | assert(ToRemoveSize > 0 && "Shouldn't reach here if nothing to remove")((void)0); |
622 | assert(EarlierSize > ToRemoveSize && "Can't remove more than original size")((void)0); |
623 | |
624 | uint64_t NewSize = EarlierSize - ToRemoveSize; |
625 | if (auto *AMI = dyn_cast<AtomicMemIntrinsic>(EarlierWrite)) { |
626 | // When shortening an atomic memory intrinsic, the newly shortened |
627 | // length must remain an integer multiple of the element size. |
628 | const uint32_t ElementSize = AMI->getElementSizeInBytes(); |
629 | if (0 != NewSize % ElementSize) |
630 | return false; |
631 | } |
632 | |
633 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW "do { } while (false) |
634 | << (IsOverwriteEnd ? "END" : "BEGIN") << ": "do { } while (false) |
635 | << *EarlierWrite << "\n KILLER [" << ToRemoveStart << ", "do { } while (false) |
636 | << int64_t(ToRemoveStart + ToRemoveSize) << ")\n")do { } while (false); |
637 | |
638 | Value *EarlierWriteLength = EarlierIntrinsic->getLength(); |
639 | Value *TrimmedLength = |
640 | ConstantInt::get(EarlierWriteLength->getType(), NewSize); |
641 | EarlierIntrinsic->setLength(TrimmedLength); |
642 | EarlierIntrinsic->setDestAlignment(PrefAlign); |
643 | |
644 | if (!IsOverwriteEnd) { |
645 | Value *OrigDest = EarlierIntrinsic->getRawDest(); |
646 | Type *Int8PtrTy = |
647 | Type::getInt8PtrTy(EarlierIntrinsic->getContext(), |
648 | OrigDest->getType()->getPointerAddressSpace()); |
649 | Value *Dest = OrigDest; |
650 | if (OrigDest->getType() != Int8PtrTy) |
651 | Dest = CastInst::CreatePointerCast(OrigDest, Int8PtrTy, "", EarlierWrite); |
652 | Value *Indices[1] = { |
653 | ConstantInt::get(EarlierWriteLength->getType(), ToRemoveSize)}; |
654 | Instruction *NewDestGEP = GetElementPtrInst::CreateInBounds( |
655 | Type::getInt8Ty(EarlierIntrinsic->getContext()), |
656 | Dest, Indices, "", EarlierWrite); |
657 | NewDestGEP->setDebugLoc(EarlierIntrinsic->getDebugLoc()); |
658 | if (NewDestGEP->getType() != OrigDest->getType()) |
659 | NewDestGEP = CastInst::CreatePointerCast(NewDestGEP, OrigDest->getType(), |
660 | "", EarlierWrite); |
661 | EarlierIntrinsic->setDest(NewDestGEP); |
662 | } |
663 | |
664 | // Finally update start and size of earlier access. |
665 | if (!IsOverwriteEnd) |
666 | EarlierStart += ToRemoveSize; |
667 | EarlierSize = NewSize; |
668 | |
669 | return true; |
670 | } |
671 | |
672 | static bool tryToShortenEnd(Instruction *EarlierWrite, |
673 | OverlapIntervalsTy &IntervalMap, |
674 | int64_t &EarlierStart, uint64_t &EarlierSize) { |
675 | if (IntervalMap.empty() || !isShortenableAtTheEnd(EarlierWrite)) |
676 | return false; |
677 | |
678 | OverlapIntervalsTy::iterator OII = --IntervalMap.end(); |
679 | int64_t LaterStart = OII->second; |
680 | uint64_t LaterSize = OII->first - LaterStart; |
681 | |
682 | assert(OII->first - LaterStart >= 0 && "Size expected to be positive")((void)0); |
683 | |
684 | if (LaterStart > EarlierStart && |
685 | // Note: "LaterStart - EarlierStart" is known to be positive due to |
686 | // preceding check. |
687 | (uint64_t)(LaterStart - EarlierStart) < EarlierSize && |
688 | // Note: "EarlierSize - (uint64_t)(LaterStart - EarlierStart)" is known to |
689 | // be non negative due to preceding checks. |
690 | LaterSize >= EarlierSize - (uint64_t)(LaterStart - EarlierStart)) { |
691 | if (tryToShorten(EarlierWrite, EarlierStart, EarlierSize, LaterStart, |
692 | LaterSize, true)) { |
693 | IntervalMap.erase(OII); |
694 | return true; |
695 | } |
696 | } |
697 | return false; |
698 | } |
699 | |
700 | static bool tryToShortenBegin(Instruction *EarlierWrite, |
701 | OverlapIntervalsTy &IntervalMap, |
702 | int64_t &EarlierStart, uint64_t &EarlierSize) { |
703 | if (IntervalMap.empty() || !isShortenableAtTheBeginning(EarlierWrite)) |
704 | return false; |
705 | |
706 | OverlapIntervalsTy::iterator OII = IntervalMap.begin(); |
707 | int64_t LaterStart = OII->second; |
708 | uint64_t LaterSize = OII->first - LaterStart; |
709 | |
710 | assert(OII->first - LaterStart >= 0 && "Size expected to be positive")((void)0); |
711 | |
712 | if (LaterStart <= EarlierStart && |
713 | // Note: "EarlierStart - LaterStart" is known to be non negative due to |
714 | // preceding check. |
715 | LaterSize > (uint64_t)(EarlierStart - LaterStart)) { |
716 | // Note: "LaterSize - (uint64_t)(EarlierStart - LaterStart)" is known to be |
717 | // positive due to preceding checks. |
718 | assert(LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize &&((void)0) |
719 | "Should have been handled as OW_Complete")((void)0); |
720 | if (tryToShorten(EarlierWrite, EarlierStart, EarlierSize, LaterStart, |
721 | LaterSize, false)) { |
722 | IntervalMap.erase(OII); |
723 | return true; |
724 | } |
725 | } |
726 | return false; |
727 | } |
728 | |
729 | static bool removePartiallyOverlappedStores(const DataLayout &DL, |
730 | InstOverlapIntervalsTy &IOL, |
731 | const TargetLibraryInfo &TLI) { |
732 | bool Changed = false; |
733 | for (auto OI : IOL) { |
734 | Instruction *EarlierWrite = OI.first; |
735 | MemoryLocation Loc = getLocForWrite(EarlierWrite, TLI); |
736 | assert(isRemovable(EarlierWrite) && "Expect only removable instruction")((void)0); |
737 | |
738 | const Value *Ptr = Loc.Ptr->stripPointerCasts(); |
739 | int64_t EarlierStart = 0; |
740 | uint64_t EarlierSize = Loc.Size.getValue(); |
741 | GetPointerBaseWithConstantOffset(Ptr, EarlierStart, DL); |
742 | OverlapIntervalsTy &IntervalMap = OI.second; |
743 | Changed |= |
744 | tryToShortenEnd(EarlierWrite, IntervalMap, EarlierStart, EarlierSize); |
745 | if (IntervalMap.empty()) |
746 | continue; |
747 | Changed |= |
748 | tryToShortenBegin(EarlierWrite, IntervalMap, EarlierStart, EarlierSize); |
749 | } |
750 | return Changed; |
751 | } |
752 | |
753 | static Constant *tryToMergePartialOverlappingStores( |
754 | StoreInst *Earlier, StoreInst *Later, int64_t InstWriteOffset, |
755 | int64_t DepWriteOffset, const DataLayout &DL, BatchAAResults &AA, |
756 | DominatorTree *DT) { |
757 | |
758 | if (Earlier && isa<ConstantInt>(Earlier->getValueOperand()) && |
759 | DL.typeSizeEqualsStoreSize(Earlier->getValueOperand()->getType()) && |
760 | Later && isa<ConstantInt>(Later->getValueOperand()) && |
761 | DL.typeSizeEqualsStoreSize(Later->getValueOperand()->getType()) && |
762 | memoryIsNotModifiedBetween(Earlier, Later, AA, DL, DT)) { |
763 | // If the store we find is: |
764 | // a) partially overwritten by the store to 'Loc' |
765 | // b) the later store is fully contained in the earlier one and |
766 | // c) they both have a constant value |
767 | // d) none of the two stores need padding |
768 | // Merge the two stores, replacing the earlier store's value with a |
769 | // merge of both values. |
770 | // TODO: Deal with other constant types (vectors, etc), and probably |
771 | // some mem intrinsics (if needed) |
772 | |
773 | APInt EarlierValue = |
774 | cast<ConstantInt>(Earlier->getValueOperand())->getValue(); |
775 | APInt LaterValue = cast<ConstantInt>(Later->getValueOperand())->getValue(); |
776 | unsigned LaterBits = LaterValue.getBitWidth(); |
777 | assert(EarlierValue.getBitWidth() > LaterValue.getBitWidth())((void)0); |
778 | LaterValue = LaterValue.zext(EarlierValue.getBitWidth()); |
779 | |
780 | // Offset of the smaller store inside the larger store |
781 | unsigned BitOffsetDiff = (InstWriteOffset - DepWriteOffset) * 8; |
782 | unsigned LShiftAmount = DL.isBigEndian() ? EarlierValue.getBitWidth() - |
783 | BitOffsetDiff - LaterBits |
784 | : BitOffsetDiff; |
785 | APInt Mask = APInt::getBitsSet(EarlierValue.getBitWidth(), LShiftAmount, |
786 | LShiftAmount + LaterBits); |
787 | // Clear the bits we'll be replacing, then OR with the smaller |
788 | // store, shifted appropriately. |
789 | APInt Merged = (EarlierValue & ~Mask) | (LaterValue << LShiftAmount); |
790 | LLVM_DEBUG(dbgs() << "DSE: Merge Stores:\n Earlier: " << *Earlierdo { } while (false) |
791 | << "\n Later: " << *Laterdo { } while (false) |
792 | << "\n Merged Value: " << Merged << '\n')do { } while (false); |
793 | return ConstantInt::get(Earlier->getValueOperand()->getType(), Merged); |
794 | } |
795 | return nullptr; |
796 | } |
797 | |
798 | namespace { |
799 | // Returns true if \p I is an intrisnic that does not read or write memory. |
800 | bool isNoopIntrinsic(Instruction *I) { |
801 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
802 | switch (II->getIntrinsicID()) { |
803 | case Intrinsic::lifetime_start: |
804 | case Intrinsic::lifetime_end: |
805 | case Intrinsic::invariant_end: |
806 | case Intrinsic::launder_invariant_group: |
807 | case Intrinsic::assume: |
808 | return true; |
809 | case Intrinsic::dbg_addr: |
810 | case Intrinsic::dbg_declare: |
811 | case Intrinsic::dbg_label: |
812 | case Intrinsic::dbg_value: |
813 | llvm_unreachable("Intrinsic should not be modeled in MemorySSA")__builtin_unreachable(); |
814 | default: |
815 | return false; |
816 | } |
817 | } |
818 | return false; |
819 | } |
820 | |
821 | // Check if we can ignore \p D for DSE. |
822 | bool canSkipDef(MemoryDef *D, bool DefVisibleToCaller) { |
823 | Instruction *DI = D->getMemoryInst(); |
824 | // Calls that only access inaccessible memory cannot read or write any memory |
825 | // locations we consider for elimination. |
826 | if (auto *CB = dyn_cast<CallBase>(DI)) |
827 | if (CB->onlyAccessesInaccessibleMemory()) |
828 | return true; |
829 | |
830 | // We can eliminate stores to locations not visible to the caller across |
831 | // throwing instructions. |
832 | if (DI->mayThrow() && !DefVisibleToCaller) |
833 | return true; |
834 | |
835 | // We can remove the dead stores, irrespective of the fence and its ordering |
836 | // (release/acquire/seq_cst). Fences only constraints the ordering of |
837 | // already visible stores, it does not make a store visible to other |
838 | // threads. So, skipping over a fence does not change a store from being |
839 | // dead. |
840 | if (isa<FenceInst>(DI)) |
841 | return true; |
842 | |
843 | // Skip intrinsics that do not really read or modify memory. |
844 | if (isNoopIntrinsic(D->getMemoryInst())) |
845 | return true; |
846 | |
847 | return false; |
848 | } |
849 | |
850 | struct DSEState { |
851 | Function &F; |
852 | AliasAnalysis &AA; |
853 | |
854 | /// The single BatchAA instance that is used to cache AA queries. It will |
855 | /// not be invalidated over the whole run. This is safe, because: |
856 | /// 1. Only memory writes are removed, so the alias cache for memory |
857 | /// locations remains valid. |
858 | /// 2. No new instructions are added (only instructions removed), so cached |
859 | /// information for a deleted value cannot be accessed by a re-used new |
860 | /// value pointer. |
861 | BatchAAResults BatchAA; |
862 | |
863 | MemorySSA &MSSA; |
864 | DominatorTree &DT; |
865 | PostDominatorTree &PDT; |
866 | const TargetLibraryInfo &TLI; |
867 | const DataLayout &DL; |
868 | const LoopInfo &LI; |
869 | |
870 | // Whether the function contains any irreducible control flow, useful for |
871 | // being accurately able to detect loops. |
872 | bool ContainsIrreducibleLoops; |
873 | |
874 | // All MemoryDefs that potentially could kill other MemDefs. |
875 | SmallVector<MemoryDef *, 64> MemDefs; |
876 | // Any that should be skipped as they are already deleted |
877 | SmallPtrSet<MemoryAccess *, 4> SkipStores; |
878 | // Keep track of all of the objects that are invisible to the caller before |
879 | // the function returns. |
880 | // SmallPtrSet<const Value *, 16> InvisibleToCallerBeforeRet; |
881 | DenseMap<const Value *, bool> InvisibleToCallerBeforeRet; |
882 | // Keep track of all of the objects that are invisible to the caller after |
883 | // the function returns. |
884 | DenseMap<const Value *, bool> InvisibleToCallerAfterRet; |
885 | // Keep track of blocks with throwing instructions not modeled in MemorySSA. |
886 | SmallPtrSet<BasicBlock *, 16> ThrowingBlocks; |
887 | // Post-order numbers for each basic block. Used to figure out if memory |
888 | // accesses are executed before another access. |
889 | DenseMap<BasicBlock *, unsigned> PostOrderNumbers; |
890 | |
891 | /// Keep track of instructions (partly) overlapping with killing MemoryDefs per |
892 | /// basic block. |
893 | DenseMap<BasicBlock *, InstOverlapIntervalsTy> IOLs; |
894 | |
895 | DSEState(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, DominatorTree &DT, |
896 | PostDominatorTree &PDT, const TargetLibraryInfo &TLI, |
897 | const LoopInfo &LI) |
898 | : F(F), AA(AA), BatchAA(AA), MSSA(MSSA), DT(DT), PDT(PDT), TLI(TLI), |
899 | DL(F.getParent()->getDataLayout()), LI(LI) {} |
900 | |
901 | static DSEState get(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, |
902 | DominatorTree &DT, PostDominatorTree &PDT, |
903 | const TargetLibraryInfo &TLI, const LoopInfo &LI) { |
904 | DSEState State(F, AA, MSSA, DT, PDT, TLI, LI); |
905 | // Collect blocks with throwing instructions not modeled in MemorySSA and |
906 | // alloc-like objects. |
907 | unsigned PO = 0; |
908 | for (BasicBlock *BB : post_order(&F)) { |
909 | State.PostOrderNumbers[BB] = PO++; |
910 | for (Instruction &I : *BB) { |
911 | MemoryAccess *MA = MSSA.getMemoryAccess(&I); |
912 | if (I.mayThrow() && !MA) |
913 | State.ThrowingBlocks.insert(I.getParent()); |
914 | |
915 | auto *MD = dyn_cast_or_null<MemoryDef>(MA); |
916 | if (MD && State.MemDefs.size() < MemorySSADefsPerBlockLimit && |
917 | (State.getLocForWriteEx(&I) || State.isMemTerminatorInst(&I))) |
918 | State.MemDefs.push_back(MD); |
919 | } |
920 | } |
921 | |
922 | // Treat byval or inalloca arguments the same as Allocas, stores to them are |
923 | // dead at the end of the function. |
924 | for (Argument &AI : F.args()) |
925 | if (AI.hasPassPointeeByValueCopyAttr()) { |
926 | // For byval, the caller doesn't know the address of the allocation. |
927 | if (AI.hasByValAttr()) |
928 | State.InvisibleToCallerBeforeRet.insert({&AI, true}); |
929 | State.InvisibleToCallerAfterRet.insert({&AI, true}); |
930 | } |
931 | |
932 | // Collect whether there is any irreducible control flow in the function. |
933 | State.ContainsIrreducibleLoops = mayContainIrreducibleControl(F, &LI); |
934 | |
935 | return State; |
936 | } |
937 | |
938 | /// Return 'OW_Complete' if a store to the 'Later' location (by \p LaterI |
939 | /// instruction) completely overwrites a store to the 'Earlier' location. |
940 | /// (by \p EarlierI instruction). |
941 | /// Return OW_MaybePartial if \p Later does not completely overwrite |
942 | /// \p Earlier, but they both write to the same underlying object. In that |
943 | /// case, use isPartialOverwrite to check if \p Later partially overwrites |
944 | /// \p Earlier. Returns 'OW_Unknown' if nothing can be determined. |
945 | OverwriteResult |
946 | isOverwrite(const Instruction *LaterI, const Instruction *EarlierI, |
947 | const MemoryLocation &Later, const MemoryLocation &Earlier, |
948 | int64_t &EarlierOff, int64_t &LaterOff) { |
949 | // AliasAnalysis does not always account for loops. Limit overwrite checks |
950 | // to dependencies for which we can guarantee they are independant of any |
951 | // loops they are in. |
952 | if (!isGuaranteedLoopIndependent(EarlierI, LaterI, Earlier)) |
953 | return OW_Unknown; |
954 | |
955 | // FIXME: Vet that this works for size upper-bounds. Seems unlikely that we'll |
956 | // get imprecise values here, though (except for unknown sizes). |
957 | if (!Later.Size.isPrecise() || !Earlier.Size.isPrecise()) { |
958 | // In case no constant size is known, try to an IR values for the number |
959 | // of bytes written and check if they match. |
960 | const auto *LaterMemI = dyn_cast<MemIntrinsic>(LaterI); |
961 | const auto *EarlierMemI = dyn_cast<MemIntrinsic>(EarlierI); |
962 | if (LaterMemI && EarlierMemI) { |
963 | const Value *LaterV = LaterMemI->getLength(); |
964 | const Value *EarlierV = EarlierMemI->getLength(); |
965 | if (LaterV == EarlierV && BatchAA.isMustAlias(Earlier, Later)) |
966 | return OW_Complete; |
967 | } |
968 | |
969 | // Masked stores have imprecise locations, but we can reason about them |
970 | // to some extent. |
971 | return isMaskedStoreOverwrite(LaterI, EarlierI, BatchAA); |
972 | } |
973 | |
974 | const uint64_t LaterSize = Later.Size.getValue(); |
975 | const uint64_t EarlierSize = Earlier.Size.getValue(); |
976 | |
977 | // Query the alias information |
978 | AliasResult AAR = BatchAA.alias(Later, Earlier); |
979 | |
980 | // If the start pointers are the same, we just have to compare sizes to see if |
981 | // the later store was larger than the earlier store. |
982 | if (AAR == AliasResult::MustAlias) { |
983 | // Make sure that the Later size is >= the Earlier size. |
984 | if (LaterSize >= EarlierSize) |
985 | return OW_Complete; |
986 | } |
987 | |
988 | // If we hit a partial alias we may have a full overwrite |
989 | if (AAR == AliasResult::PartialAlias && AAR.hasOffset()) { |
990 | int32_t Off = AAR.getOffset(); |
991 | if (Off >= 0 && (uint64_t)Off + EarlierSize <= LaterSize) |
992 | return OW_Complete; |
993 | } |
994 | |
995 | // Check to see if the later store is to the entire object (either a global, |
996 | // an alloca, or a byval/inalloca argument). If so, then it clearly |
997 | // overwrites any other store to the same object. |
998 | const Value *P1 = Earlier.Ptr->stripPointerCasts(); |
999 | const Value *P2 = Later.Ptr->stripPointerCasts(); |
1000 | const Value *UO1 = getUnderlyingObject(P1), *UO2 = getUnderlyingObject(P2); |
1001 | |
1002 | // If we can't resolve the same pointers to the same object, then we can't |
1003 | // analyze them at all. |
1004 | if (UO1 != UO2) |
1005 | return OW_Unknown; |
1006 | |
1007 | // If the "Later" store is to a recognizable object, get its size. |
1008 | uint64_t ObjectSize = getPointerSize(UO2, DL, TLI, &F); |
1009 | if (ObjectSize != MemoryLocation::UnknownSize) |
1010 | if (ObjectSize == LaterSize && ObjectSize >= EarlierSize) |
1011 | return OW_Complete; |
1012 | |
1013 | // Okay, we have stores to two completely different pointers. Try to |
1014 | // decompose the pointer into a "base + constant_offset" form. If the base |
1015 | // pointers are equal, then we can reason about the two stores. |
1016 | EarlierOff = 0; |
1017 | LaterOff = 0; |
1018 | const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, DL); |
1019 | const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, DL); |
1020 | |
1021 | // If the base pointers still differ, we have two completely different stores. |
1022 | if (BP1 != BP2) |
1023 | return OW_Unknown; |
1024 | |
1025 | // The later access completely overlaps the earlier store if and only if |
1026 | // both start and end of the earlier one is "inside" the later one: |
1027 | // |<->|--earlier--|<->| |
1028 | // |-------later-------| |
1029 | // Accesses may overlap if and only if start of one of them is "inside" |
1030 | // another one: |
1031 | // |<->|--earlier--|<----->| |
1032 | // |-------later-------| |
1033 | // OR |
1034 | // |----- earlier -----| |
1035 | // |<->|---later---|<----->| |
1036 | // |
1037 | // We have to be careful here as *Off is signed while *.Size is unsigned. |
1038 | |
1039 | // Check if the earlier access starts "not before" the later one. |
1040 | if (EarlierOff >= LaterOff) { |
1041 | // If the earlier access ends "not after" the later access then the earlier |
1042 | // one is completely overwritten by the later one. |
1043 | if (uint64_t(EarlierOff - LaterOff) + EarlierSize <= LaterSize) |
1044 | return OW_Complete; |
1045 | // If start of the earlier access is "before" end of the later access then |
1046 | // accesses overlap. |
1047 | else if ((uint64_t)(EarlierOff - LaterOff) < LaterSize) |
1048 | return OW_MaybePartial; |
1049 | } |
1050 | // If start of the later access is "before" end of the earlier access then |
1051 | // accesses overlap. |
1052 | else if ((uint64_t)(LaterOff - EarlierOff) < EarlierSize) { |
1053 | return OW_MaybePartial; |
1054 | } |
1055 | |
1056 | // Can reach here only if accesses are known not to overlap. There is no |
1057 | // dedicated code to indicate no overlap so signal "unknown". |
1058 | return OW_Unknown; |
1059 | } |
1060 | |
1061 | bool isInvisibleToCallerAfterRet(const Value *V) { |
1062 | if (isa<AllocaInst>(V)) |
1063 | return true; |
1064 | auto I = InvisibleToCallerAfterRet.insert({V, false}); |
1065 | if (I.second) { |
1066 | if (!isInvisibleToCallerBeforeRet(V)) { |
1067 | I.first->second = false; |
1068 | } else { |
1069 | auto *Inst = dyn_cast<Instruction>(V); |
1070 | if (Inst && isAllocLikeFn(Inst, &TLI)) |
1071 | I.first->second = !PointerMayBeCaptured(V, true, false); |
1072 | } |
1073 | } |
1074 | return I.first->second; |
1075 | } |
1076 | |
1077 | bool isInvisibleToCallerBeforeRet(const Value *V) { |
1078 | if (isa<AllocaInst>(V)) |
1079 | return true; |
1080 | auto I = InvisibleToCallerBeforeRet.insert({V, false}); |
1081 | if (I.second) { |
1082 | auto *Inst = dyn_cast<Instruction>(V); |
1083 | if (Inst && isAllocLikeFn(Inst, &TLI)) |
1084 | // NOTE: This could be made more precise by PointerMayBeCapturedBefore |
1085 | // with the killing MemoryDef. But we refrain from doing so for now to |
1086 | // limit compile-time and this does not cause any changes to the number |
1087 | // of stores removed on a large test set in practice. |
1088 | I.first->second = !PointerMayBeCaptured(V, false, true); |
1089 | } |
1090 | return I.first->second; |
1091 | } |
1092 | |
1093 | Optional<MemoryLocation> getLocForWriteEx(Instruction *I) const { |
1094 | if (!I->mayWriteToMemory()) |
1095 | return None; |
1096 | |
1097 | if (auto *MTI = dyn_cast<AnyMemIntrinsic>(I)) |
1098 | return {MemoryLocation::getForDest(MTI)}; |
1099 | |
1100 | if (auto *CB = dyn_cast<CallBase>(I)) { |
1101 | // If the functions may write to memory we do not know about, bail out. |
1102 | if (!CB->onlyAccessesArgMemory() && |
1103 | !CB->onlyAccessesInaccessibleMemOrArgMem()) |
1104 | return None; |
1105 | |
1106 | LibFunc LF; |
1107 | if (TLI.getLibFunc(*CB, LF) && TLI.has(LF)) { |
1108 | switch (LF) { |
1109 | case LibFunc_strcpy: |
1110 | case LibFunc_strncpy: |
1111 | case LibFunc_strcat: |
1112 | case LibFunc_strncat: |
1113 | return {MemoryLocation::getAfter(CB->getArgOperand(0))}; |
1114 | default: |
1115 | break; |
1116 | } |
1117 | } |
1118 | switch (CB->getIntrinsicID()) { |
1119 | case Intrinsic::init_trampoline: |
1120 | return {MemoryLocation::getAfter(CB->getArgOperand(0))}; |
1121 | case Intrinsic::masked_store: |
1122 | return {MemoryLocation::getForArgument(CB, 1, TLI)}; |
1123 | default: |
1124 | break; |
1125 | } |
1126 | return None; |
1127 | } |
1128 | |
1129 | return MemoryLocation::getOrNone(I); |
1130 | } |
1131 | |
1132 | /// Returns true if \p UseInst completely overwrites \p DefLoc |
1133 | /// (stored by \p DefInst). |
1134 | bool isCompleteOverwrite(const MemoryLocation &DefLoc, Instruction *DefInst, |
1135 | Instruction *UseInst) { |
1136 | // UseInst has a MemoryDef associated in MemorySSA. It's possible for a |
1137 | // MemoryDef to not write to memory, e.g. a volatile load is modeled as a |
1138 | // MemoryDef. |
1139 | if (!UseInst->mayWriteToMemory()) |
1140 | return false; |
1141 | |
1142 | if (auto *CB = dyn_cast<CallBase>(UseInst)) |
1143 | if (CB->onlyAccessesInaccessibleMemory()) |
1144 | return false; |
1145 | |
1146 | int64_t InstWriteOffset, DepWriteOffset; |
1147 | if (auto CC = getLocForWriteEx(UseInst)) |
1148 | return isOverwrite(UseInst, DefInst, *CC, DefLoc, DepWriteOffset, |
1149 | InstWriteOffset) == OW_Complete; |
1150 | return false; |
1151 | } |
1152 | |
1153 | /// Returns true if \p Def is not read before returning from the function. |
1154 | bool isWriteAtEndOfFunction(MemoryDef *Def) { |
1155 | LLVM_DEBUG(dbgs() << " Check if def " << *Def << " ("do { } while (false) |
1156 | << *Def->getMemoryInst()do { } while (false) |
1157 | << ") is at the end the function \n")do { } while (false); |
1158 | |
1159 | auto MaybeLoc = getLocForWriteEx(Def->getMemoryInst()); |
1160 | if (!MaybeLoc) { |
1161 | LLVM_DEBUG(dbgs() << " ... could not get location for write.\n")do { } while (false); |
1162 | return false; |
1163 | } |
1164 | |
1165 | SmallVector<MemoryAccess *, 4> WorkList; |
1166 | SmallPtrSet<MemoryAccess *, 8> Visited; |
1167 | auto PushMemUses = [&WorkList, &Visited](MemoryAccess *Acc) { |
1168 | if (!Visited.insert(Acc).second) |
1169 | return; |
1170 | for (Use &U : Acc->uses()) |
1171 | WorkList.push_back(cast<MemoryAccess>(U.getUser())); |
1172 | }; |
1173 | PushMemUses(Def); |
1174 | for (unsigned I = 0; I < WorkList.size(); I++) { |
1175 | if (WorkList.size() >= MemorySSAScanLimit) { |
1176 | LLVM_DEBUG(dbgs() << " ... hit exploration limit.\n")do { } while (false); |
1177 | return false; |
1178 | } |
1179 | |
1180 | MemoryAccess *UseAccess = WorkList[I]; |
1181 | // Simply adding the users of MemoryPhi to the worklist is not enough, |
1182 | // because we might miss read clobbers in different iterations of a loop, |
1183 | // for example. |
1184 | // TODO: Add support for phi translation to handle the loop case. |
1185 | if (isa<MemoryPhi>(UseAccess)) |
1186 | return false; |
1187 | |
1188 | // TODO: Checking for aliasing is expensive. Consider reducing the amount |
1189 | // of times this is called and/or caching it. |
1190 | Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); |
1191 | if (isReadClobber(*MaybeLoc, UseInst)) { |
1192 | LLVM_DEBUG(dbgs() << " ... hit read clobber " << *UseInst << ".\n")do { } while (false); |
1193 | return false; |
1194 | } |
1195 | |
1196 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) |
1197 | PushMemUses(UseDef); |
1198 | } |
1199 | return true; |
1200 | } |
1201 | |
1202 | /// If \p I is a memory terminator like llvm.lifetime.end or free, return a |
1203 | /// pair with the MemoryLocation terminated by \p I and a boolean flag |
1204 | /// indicating whether \p I is a free-like call. |
1205 | Optional<std::pair<MemoryLocation, bool>> |
1206 | getLocForTerminator(Instruction *I) const { |
1207 | uint64_t Len; |
1208 | Value *Ptr; |
1209 | if (match(I, m_Intrinsic<Intrinsic::lifetime_end>(m_ConstantInt(Len), |
1210 | m_Value(Ptr)))) |
1211 | return {std::make_pair(MemoryLocation(Ptr, Len), false)}; |
1212 | |
1213 | if (auto *CB = dyn_cast<CallBase>(I)) { |
1214 | if (isFreeCall(I, &TLI)) |
1215 | return {std::make_pair(MemoryLocation::getAfter(CB->getArgOperand(0)), |
1216 | true)}; |
1217 | } |
1218 | |
1219 | return None; |
1220 | } |
1221 | |
1222 | /// Returns true if \p I is a memory terminator instruction like |
1223 | /// llvm.lifetime.end or free. |
1224 | bool isMemTerminatorInst(Instruction *I) const { |
1225 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(I); |
1226 | return (II && II->getIntrinsicID() == Intrinsic::lifetime_end) || |
1227 | isFreeCall(I, &TLI); |
1228 | } |
1229 | |
1230 | /// Returns true if \p MaybeTerm is a memory terminator for \p Loc from |
1231 | /// instruction \p AccessI. |
1232 | bool isMemTerminator(const MemoryLocation &Loc, Instruction *AccessI, |
1233 | Instruction *MaybeTerm) { |
1234 | Optional<std::pair<MemoryLocation, bool>> MaybeTermLoc = |
1235 | getLocForTerminator(MaybeTerm); |
1236 | |
1237 | if (!MaybeTermLoc) |
1238 | return false; |
1239 | |
1240 | // If the terminator is a free-like call, all accesses to the underlying |
1241 | // object can be considered terminated. |
1242 | if (getUnderlyingObject(Loc.Ptr) != |
1243 | getUnderlyingObject(MaybeTermLoc->first.Ptr)) |
1244 | return false; |
1245 | |
1246 | auto TermLoc = MaybeTermLoc->first; |
1247 | if (MaybeTermLoc->second) { |
1248 | const Value *LocUO = getUnderlyingObject(Loc.Ptr); |
1249 | return BatchAA.isMustAlias(TermLoc.Ptr, LocUO); |
1250 | } |
1251 | int64_t InstWriteOffset, DepWriteOffset; |
1252 | return isOverwrite(MaybeTerm, AccessI, TermLoc, Loc, DepWriteOffset, |
1253 | InstWriteOffset) == OW_Complete; |
1254 | } |
1255 | |
1256 | // Returns true if \p Use may read from \p DefLoc. |
1257 | bool isReadClobber(const MemoryLocation &DefLoc, Instruction *UseInst) { |
1258 | if (isNoopIntrinsic(UseInst)) |
1259 | return false; |
1260 | |
1261 | // Monotonic or weaker atomic stores can be re-ordered and do not need to be |
1262 | // treated as read clobber. |
1263 | if (auto SI = dyn_cast<StoreInst>(UseInst)) |
1264 | return isStrongerThan(SI->getOrdering(), AtomicOrdering::Monotonic); |
1265 | |
1266 | if (!UseInst->mayReadFromMemory()) |
1267 | return false; |
1268 | |
1269 | if (auto *CB = dyn_cast<CallBase>(UseInst)) |
1270 | if (CB->onlyAccessesInaccessibleMemory()) |
1271 | return false; |
1272 | |
1273 | // NOTE: For calls, the number of stores removed could be slightly improved |
1274 | // by using AA.callCapturesBefore(UseInst, DefLoc, &DT), but that showed to |
1275 | // be expensive compared to the benefits in practice. For now, avoid more |
1276 | // expensive analysis to limit compile-time. |
1277 | return isRefSet(BatchAA.getModRefInfo(UseInst, DefLoc)); |
1278 | } |
1279 | |
1280 | /// Returns true if a dependency between \p Current and \p KillingDef is |
1281 | /// guaranteed to be loop invariant for the loops that they are in. Either |
1282 | /// because they are known to be in the same block, in the same loop level or |
1283 | /// by guaranteeing that \p CurrentLoc only references a single MemoryLocation |
1284 | /// during execution of the containing function. |
1285 | bool isGuaranteedLoopIndependent(const Instruction *Current, |
1286 | const Instruction *KillingDef, |
1287 | const MemoryLocation &CurrentLoc) { |
1288 | // If the dependency is within the same block or loop level (being careful |
1289 | // of irreducible loops), we know that AA will return a valid result for the |
1290 | // memory dependency. (Both at the function level, outside of any loop, |
1291 | // would also be valid but we currently disable that to limit compile time). |
1292 | if (Current->getParent() == KillingDef->getParent()) |
1293 | return true; |
1294 | const Loop *CurrentLI = LI.getLoopFor(Current->getParent()); |
1295 | if (!ContainsIrreducibleLoops && CurrentLI && |
1296 | CurrentLI == LI.getLoopFor(KillingDef->getParent())) |
1297 | return true; |
1298 | // Otherwise check the memory location is invariant to any loops. |
1299 | return isGuaranteedLoopInvariant(CurrentLoc.Ptr); |
1300 | } |
1301 | |
1302 | /// Returns true if \p Ptr is guaranteed to be loop invariant for any possible |
1303 | /// loop. In particular, this guarantees that it only references a single |
1304 | /// MemoryLocation during execution of the containing function. |
1305 | bool isGuaranteedLoopInvariant(const Value *Ptr) { |
1306 | auto IsGuaranteedLoopInvariantBase = [this](const Value *Ptr) { |
1307 | Ptr = Ptr->stripPointerCasts(); |
1308 | if (auto *I = dyn_cast<Instruction>(Ptr)) { |
1309 | if (isa<AllocaInst>(Ptr)) |
1310 | return true; |
1311 | |
1312 | if (isAllocLikeFn(I, &TLI)) |
1313 | return true; |
1314 | |
1315 | return false; |
1316 | } |
1317 | return true; |
1318 | }; |
1319 | |
1320 | Ptr = Ptr->stripPointerCasts(); |
1321 | if (auto *I = dyn_cast<Instruction>(Ptr)) { |
1322 | if (I->getParent()->isEntryBlock()) |
1323 | return true; |
1324 | } |
1325 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { |
1326 | return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && |
1327 | GEP->hasAllConstantIndices(); |
1328 | } |
1329 | return IsGuaranteedLoopInvariantBase(Ptr); |
1330 | } |
1331 | |
1332 | // Find a MemoryDef writing to \p DefLoc and dominating \p StartAccess, with |
1333 | // no read access between them or on any other path to a function exit block |
1334 | // if \p DefLoc is not accessible after the function returns. If there is no |
1335 | // such MemoryDef, return None. The returned value may not (completely) |
1336 | // overwrite \p DefLoc. Currently we bail out when we encounter an aliasing |
1337 | // MemoryUse (read). |
1338 | Optional<MemoryAccess *> |
1339 | getDomMemoryDef(MemoryDef *KillingDef, MemoryAccess *StartAccess, |
1340 | const MemoryLocation &DefLoc, const Value *DefUO, |
1341 | unsigned &ScanLimit, unsigned &WalkerStepLimit, |
1342 | bool IsMemTerm, unsigned &PartialLimit) { |
1343 | if (ScanLimit == 0 || WalkerStepLimit == 0) { |
1344 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n")do { } while (false); |
1345 | return None; |
1346 | } |
1347 | |
1348 | MemoryAccess *Current = StartAccess; |
1349 | Instruction *KillingI = KillingDef->getMemoryInst(); |
1350 | LLVM_DEBUG(dbgs() << " trying to get dominating access\n")do { } while (false); |
1351 | |
1352 | // Find the next clobbering Mod access for DefLoc, starting at StartAccess. |
1353 | Optional<MemoryLocation> CurrentLoc; |
1354 | for (;; Current = cast<MemoryDef>(Current)->getDefiningAccess()) { |
1355 | LLVM_DEBUG({do { } while (false) |
1356 | dbgs() << " visiting " << *Current;do { } while (false) |
1357 | if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef>(Current))do { } while (false) |
1358 | dbgs() << " (" << *cast<MemoryUseOrDef>(Current)->getMemoryInst()do { } while (false) |
1359 | << ")";do { } while (false) |
1360 | dbgs() << "\n";do { } while (false) |
1361 | })do { } while (false); |
1362 | |
1363 | // Reached TOP. |
1364 | if (MSSA.isLiveOnEntryDef(Current)) { |
1365 | LLVM_DEBUG(dbgs() << " ... found LiveOnEntryDef\n")do { } while (false); |
1366 | return None; |
1367 | } |
1368 | |
1369 | // Cost of a step. Accesses in the same block are more likely to be valid |
1370 | // candidates for elimination, hence consider them cheaper. |
1371 | unsigned StepCost = KillingDef->getBlock() == Current->getBlock() |
1372 | ? MemorySSASameBBStepCost |
1373 | : MemorySSAOtherBBStepCost; |
1374 | if (WalkerStepLimit <= StepCost) { |
1375 | LLVM_DEBUG(dbgs() << " ... hit walker step limit\n")do { } while (false); |
1376 | return None; |
1377 | } |
1378 | WalkerStepLimit -= StepCost; |
1379 | |
1380 | // Return for MemoryPhis. They cannot be eliminated directly and the |
1381 | // caller is responsible for traversing them. |
1382 | if (isa<MemoryPhi>(Current)) { |
1383 | LLVM_DEBUG(dbgs() << " ... found MemoryPhi\n")do { } while (false); |
1384 | return Current; |
1385 | } |
1386 | |
1387 | // Below, check if CurrentDef is a valid candidate to be eliminated by |
1388 | // KillingDef. If it is not, check the next candidate. |
1389 | MemoryDef *CurrentDef = cast<MemoryDef>(Current); |
1390 | Instruction *CurrentI = CurrentDef->getMemoryInst(); |
1391 | |
1392 | if (canSkipDef(CurrentDef, !isInvisibleToCallerBeforeRet(DefUO))) |
1393 | continue; |
1394 | |
1395 | // Before we try to remove anything, check for any extra throwing |
1396 | // instructions that block us from DSEing |
1397 | if (mayThrowBetween(KillingI, CurrentI, DefUO)) { |
1398 | LLVM_DEBUG(dbgs() << " ... skip, may throw!\n")do { } while (false); |
1399 | return None; |
1400 | } |
1401 | |
1402 | // Check for anything that looks like it will be a barrier to further |
1403 | // removal |
1404 | if (isDSEBarrier(DefUO, CurrentI)) { |
1405 | LLVM_DEBUG(dbgs() << " ... skip, barrier\n")do { } while (false); |
1406 | return None; |
1407 | } |
1408 | |
1409 | // If Current is known to be on path that reads DefLoc or is a read |
1410 | // clobber, bail out, as the path is not profitable. We skip this check |
1411 | // for intrinsic calls, because the code knows how to handle memcpy |
1412 | // intrinsics. |
1413 | if (!isa<IntrinsicInst>(CurrentI) && isReadClobber(DefLoc, CurrentI)) |
1414 | return None; |
1415 | |
1416 | // Quick check if there are direct uses that are read-clobbers. |
1417 | if (any_of(Current->uses(), [this, &DefLoc, StartAccess](Use &U) { |
1418 | if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(U.getUser())) |
1419 | return !MSSA.dominates(StartAccess, UseOrDef) && |
1420 | isReadClobber(DefLoc, UseOrDef->getMemoryInst()); |
1421 | return false; |
1422 | })) { |
1423 | LLVM_DEBUG(dbgs() << " ... found a read clobber\n")do { } while (false); |
1424 | return None; |
1425 | } |
1426 | |
1427 | // If Current cannot be analyzed or is not removable, check the next |
1428 | // candidate. |
1429 | if (!hasAnalyzableMemoryWrite(CurrentI, TLI) || !isRemovable(CurrentI)) |
1430 | continue; |
1431 | |
1432 | // If Current does not have an analyzable write location, skip it |
1433 | CurrentLoc = getLocForWriteEx(CurrentI); |
1434 | if (!CurrentLoc) |
1435 | continue; |
1436 | |
1437 | // AliasAnalysis does not account for loops. Limit elimination to |
1438 | // candidates for which we can guarantee they always store to the same |
1439 | // memory location and not located in different loops. |
1440 | if (!isGuaranteedLoopIndependent(CurrentI, KillingI, *CurrentLoc)) { |
1441 | LLVM_DEBUG(dbgs() << " ... not guaranteed loop independent\n")do { } while (false); |
1442 | WalkerStepLimit -= 1; |
1443 | continue; |
1444 | } |
1445 | |
1446 | if (IsMemTerm) { |
1447 | // If the killing def is a memory terminator (e.g. lifetime.end), check |
1448 | // the next candidate if the current Current does not write the same |
1449 | // underlying object as the terminator. |
1450 | if (!isMemTerminator(*CurrentLoc, CurrentI, KillingI)) |
1451 | continue; |
1452 | } else { |
1453 | int64_t InstWriteOffset, DepWriteOffset; |
1454 | auto OR = isOverwrite(KillingI, CurrentI, DefLoc, *CurrentLoc, |
1455 | DepWriteOffset, InstWriteOffset); |
1456 | // If Current does not write to the same object as KillingDef, check |
1457 | // the next candidate. |
1458 | if (OR == OW_Unknown) |
1459 | continue; |
1460 | else if (OR == OW_MaybePartial) { |
1461 | // If KillingDef only partially overwrites Current, check the next |
1462 | // candidate if the partial step limit is exceeded. This aggressively |
1463 | // limits the number of candidates for partial store elimination, |
1464 | // which are less likely to be removable in the end. |
1465 | if (PartialLimit <= 1) { |
1466 | WalkerStepLimit -= 1; |
1467 | continue; |
1468 | } |
1469 | PartialLimit -= 1; |
1470 | } |
1471 | } |
1472 | break; |
1473 | }; |
1474 | |
1475 | // Accesses to objects accessible after the function returns can only be |
1476 | // eliminated if the access is killed along all paths to the exit. Collect |
1477 | // the blocks with killing (=completely overwriting MemoryDefs) and check if |
1478 | // they cover all paths from EarlierAccess to any function exit. |
1479 | SmallPtrSet<Instruction *, 16> KillingDefs; |
1480 | KillingDefs.insert(KillingDef->getMemoryInst()); |
1481 | MemoryAccess *EarlierAccess = Current; |
1482 | Instruction *EarlierMemInst = |
1483 | cast<MemoryDef>(EarlierAccess)->getMemoryInst(); |
1484 | LLVM_DEBUG(dbgs() << " Checking for reads of " << *EarlierAccess << " ("do { } while (false) |
1485 | << *EarlierMemInst << ")\n")do { } while (false); |
1486 | |
1487 | SmallSetVector<MemoryAccess *, 32> WorkList; |
1488 | auto PushMemUses = [&WorkList](MemoryAccess *Acc) { |
1489 | for (Use &U : Acc->uses()) |
1490 | WorkList.insert(cast<MemoryAccess>(U.getUser())); |
1491 | }; |
1492 | PushMemUses(EarlierAccess); |
1493 | |
1494 | // Optimistically collect all accesses for reads. If we do not find any |
1495 | // read clobbers, add them to the cache. |
1496 | SmallPtrSet<MemoryAccess *, 16> KnownNoReads; |
1497 | if (!EarlierMemInst->mayReadFromMemory()) |
1498 | KnownNoReads.insert(EarlierAccess); |
1499 | // Check if EarlierDef may be read. |
1500 | for (unsigned I = 0; I < WorkList.size(); I++) { |
1501 | MemoryAccess *UseAccess = WorkList[I]; |
1502 | |
1503 | LLVM_DEBUG(dbgs() << " " << *UseAccess)do { } while (false); |
1504 | // Bail out if the number of accesses to check exceeds the scan limit. |
1505 | if (ScanLimit < (WorkList.size() - I)) { |
1506 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n")do { } while (false); |
1507 | return None; |
1508 | } |
1509 | --ScanLimit; |
1510 | NumDomMemDefChecks++; |
1511 | KnownNoReads.insert(UseAccess); |
1512 | |
1513 | if (isa<MemoryPhi>(UseAccess)) { |
1514 | if (any_of(KillingDefs, [this, UseAccess](Instruction *KI) { |
1515 | return DT.properlyDominates(KI->getParent(), |
1516 | UseAccess->getBlock()); |
1517 | })) { |
1518 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing block\n")do { } while (false); |
1519 | continue; |
1520 | } |
1521 | LLVM_DEBUG(dbgs() << "\n ... adding PHI uses\n")do { } while (false); |
1522 | PushMemUses(UseAccess); |
1523 | continue; |
1524 | } |
1525 | |
1526 | Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); |
1527 | LLVM_DEBUG(dbgs() << " (" << *UseInst << ")\n")do { } while (false); |
1528 | |
1529 | if (any_of(KillingDefs, [this, UseInst](Instruction *KI) { |
1530 | return DT.dominates(KI, UseInst); |
1531 | })) { |
1532 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing def\n")do { } while (false); |
1533 | continue; |
1534 | } |
1535 | |
1536 | // A memory terminator kills all preceeding MemoryDefs and all succeeding |
1537 | // MemoryAccesses. We do not have to check it's users. |
1538 | if (isMemTerminator(*CurrentLoc, EarlierMemInst, UseInst)) { |
1539 | LLVM_DEBUG(do { } while (false) |
1540 | dbgs()do { } while (false) |
1541 | << " ... skipping, memterminator invalidates following accesses\n")do { } while (false); |
1542 | continue; |
1543 | } |
1544 | |
1545 | if (isNoopIntrinsic(cast<MemoryUseOrDef>(UseAccess)->getMemoryInst())) { |
1546 | LLVM_DEBUG(dbgs() << " ... adding uses of intrinsic\n")do { } while (false); |
1547 | PushMemUses(UseAccess); |
1548 | continue; |
1549 | } |
1550 | |
1551 | if (UseInst->mayThrow() && !isInvisibleToCallerBeforeRet(DefUO)) { |
1552 | LLVM_DEBUG(dbgs() << " ... found throwing instruction\n")do { } while (false); |
1553 | return None; |
1554 | } |
1555 | |
1556 | // Uses which may read the original MemoryDef mean we cannot eliminate the |
1557 | // original MD. Stop walk. |
1558 | if (isReadClobber(*CurrentLoc, UseInst)) { |
1559 | LLVM_DEBUG(dbgs() << " ... found read clobber\n")do { } while (false); |
1560 | return None; |
1561 | } |
1562 | |
1563 | // If this worklist walks back to the original memory access (and the |
1564 | // pointer is not guarenteed loop invariant) then we cannot assume that a |
1565 | // store kills itself. |
1566 | if (EarlierAccess == UseAccess && |
1567 | !isGuaranteedLoopInvariant(CurrentLoc->Ptr)) { |
1568 | LLVM_DEBUG(dbgs() << " ... found not loop invariant self access\n")do { } while (false); |
1569 | return None; |
1570 | } |
1571 | // Otherwise, for the KillingDef and EarlierAccess we only have to check |
1572 | // if it reads the memory location. |
1573 | // TODO: It would probably be better to check for self-reads before |
1574 | // calling the function. |
1575 | if (KillingDef == UseAccess || EarlierAccess == UseAccess) { |
1576 | LLVM_DEBUG(dbgs() << " ... skipping killing def/dom access\n")do { } while (false); |
1577 | continue; |
1578 | } |
1579 | |
1580 | // Check all uses for MemoryDefs, except for defs completely overwriting |
1581 | // the original location. Otherwise we have to check uses of *all* |
1582 | // MemoryDefs we discover, including non-aliasing ones. Otherwise we might |
1583 | // miss cases like the following |
1584 | // 1 = Def(LoE) ; <----- EarlierDef stores [0,1] |
1585 | // 2 = Def(1) ; (2, 1) = NoAlias, stores [2,3] |
1586 | // Use(2) ; MayAlias 2 *and* 1, loads [0, 3]. |
1587 | // (The Use points to the *first* Def it may alias) |
1588 | // 3 = Def(1) ; <---- Current (3, 2) = NoAlias, (3,1) = MayAlias, |
1589 | // stores [0,1] |
1590 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) { |
1591 | if (isCompleteOverwrite(*CurrentLoc, EarlierMemInst, UseInst)) { |
1592 | BasicBlock *MaybeKillingBlock = UseInst->getParent(); |
1593 | if (PostOrderNumbers.find(MaybeKillingBlock)->second < |
1594 | PostOrderNumbers.find(EarlierAccess->getBlock())->second) { |
1595 | if (!isInvisibleToCallerAfterRet(DefUO)) { |
1596 | LLVM_DEBUG(dbgs()do { } while (false) |
1597 | << " ... found killing def " << *UseInst << "\n")do { } while (false); |
1598 | KillingDefs.insert(UseInst); |
1599 | } |
1600 | } else { |
1601 | LLVM_DEBUG(dbgs()do { } while (false) |
1602 | << " ... found preceeding def " << *UseInst << "\n")do { } while (false); |
1603 | return None; |
1604 | } |
1605 | } else |
1606 | PushMemUses(UseDef); |
1607 | } |
1608 | } |
1609 | |
1610 | // For accesses to locations visible after the function returns, make sure |
1611 | // that the location is killed (=overwritten) along all paths from |
1612 | // EarlierAccess to the exit. |
1613 | if (!isInvisibleToCallerAfterRet(DefUO)) { |
1614 | SmallPtrSet<BasicBlock *, 16> KillingBlocks; |
1615 | for (Instruction *KD : KillingDefs) |
1616 | KillingBlocks.insert(KD->getParent()); |
1617 | assert(!KillingBlocks.empty() &&((void)0) |
1618 | "Expected at least a single killing block")((void)0); |
1619 | |
1620 | // Find the common post-dominator of all killing blocks. |
1621 | BasicBlock *CommonPred = *KillingBlocks.begin(); |
1622 | for (auto I = std::next(KillingBlocks.begin()), E = KillingBlocks.end(); |
1623 | I != E; I++) { |
1624 | if (!CommonPred) |
1625 | break; |
1626 | CommonPred = PDT.findNearestCommonDominator(CommonPred, *I); |
1627 | } |
1628 | |
1629 | // If CommonPred is in the set of killing blocks, just check if it |
1630 | // post-dominates EarlierAccess. |
1631 | if (KillingBlocks.count(CommonPred)) { |
1632 | if (PDT.dominates(CommonPred, EarlierAccess->getBlock())) |
1633 | return {EarlierAccess}; |
1634 | return None; |
1635 | } |
1636 | |
1637 | // If the common post-dominator does not post-dominate EarlierAccess, |
1638 | // there is a path from EarlierAccess to an exit not going through a |
1639 | // killing block. |
1640 | if (PDT.dominates(CommonPred, EarlierAccess->getBlock())) { |
1641 | SetVector<BasicBlock *> WorkList; |
1642 | |
1643 | // If CommonPred is null, there are multiple exits from the function. |
1644 | // They all have to be added to the worklist. |
1645 | if (CommonPred) |
1646 | WorkList.insert(CommonPred); |
1647 | else |
1648 | for (BasicBlock *R : PDT.roots()) |
1649 | WorkList.insert(R); |
1650 | |
1651 | NumCFGTries++; |
1652 | // Check if all paths starting from an exit node go through one of the |
1653 | // killing blocks before reaching EarlierAccess. |
1654 | for (unsigned I = 0; I < WorkList.size(); I++) { |
1655 | NumCFGChecks++; |
1656 | BasicBlock *Current = WorkList[I]; |
1657 | if (KillingBlocks.count(Current)) |
1658 | continue; |
1659 | if (Current == EarlierAccess->getBlock()) |
1660 | return None; |
1661 | |
1662 | // EarlierAccess is reachable from the entry, so we don't have to |
1663 | // explore unreachable blocks further. |
1664 | if (!DT.isReachableFromEntry(Current)) |
1665 | continue; |
1666 | |
1667 | for (BasicBlock *Pred : predecessors(Current)) |
1668 | WorkList.insert(Pred); |
1669 | |
1670 | if (WorkList.size() >= MemorySSAPathCheckLimit) |
1671 | return None; |
1672 | } |
1673 | NumCFGSuccess++; |
1674 | return {EarlierAccess}; |
1675 | } |
1676 | return None; |
1677 | } |
1678 | |
1679 | // No aliasing MemoryUses of EarlierAccess found, EarlierAccess is |
1680 | // potentially dead. |
1681 | return {EarlierAccess}; |
1682 | } |
1683 | |
1684 | // Delete dead memory defs |
1685 | void deleteDeadInstruction(Instruction *SI) { |
1686 | MemorySSAUpdater Updater(&MSSA); |
1687 | SmallVector<Instruction *, 32> NowDeadInsts; |
1688 | NowDeadInsts.push_back(SI); |
1689 | --NumFastOther; |
1690 | |
1691 | while (!NowDeadInsts.empty()) { |
1692 | Instruction *DeadInst = NowDeadInsts.pop_back_val(); |
1693 | ++NumFastOther; |
1694 | |
1695 | // Try to preserve debug information attached to the dead instruction. |
1696 | salvageDebugInfo(*DeadInst); |
1697 | salvageKnowledge(DeadInst); |
1698 | |
1699 | // Remove the Instruction from MSSA. |
1700 | if (MemoryAccess *MA = MSSA.getMemoryAccess(DeadInst)) { |
1701 | if (MemoryDef *MD = dyn_cast<MemoryDef>(MA)) { |
1702 | SkipStores.insert(MD); |
1703 | } |
1704 | Updater.removeMemoryAccess(MA); |
1705 | } |
1706 | |
1707 | auto I = IOLs.find(DeadInst->getParent()); |
1708 | if (I != IOLs.end()) |
1709 | I->second.erase(DeadInst); |
1710 | // Remove its operands |
1711 | for (Use &O : DeadInst->operands()) |
1712 | if (Instruction *OpI = dyn_cast<Instruction>(O)) { |
1713 | O = nullptr; |
1714 | if (isInstructionTriviallyDead(OpI, &TLI)) |
1715 | NowDeadInsts.push_back(OpI); |
1716 | } |
1717 | |
1718 | DeadInst->eraseFromParent(); |
1719 | } |
1720 | } |
1721 | |
1722 | // Check for any extra throws between SI and NI that block DSE. This only |
1723 | // checks extra maythrows (those that aren't MemoryDef's). MemoryDef that may |
1724 | // throw are handled during the walk from one def to the next. |
1725 | bool mayThrowBetween(Instruction *SI, Instruction *NI, |
1726 | const Value *SILocUnd) { |
1727 | // First see if we can ignore it by using the fact that SI is an |
1728 | // alloca/alloca like object that is not visible to the caller during |
1729 | // execution of the function. |
1730 | if (SILocUnd && isInvisibleToCallerBeforeRet(SILocUnd)) |
1731 | return false; |
1732 | |
1733 | if (SI->getParent() == NI->getParent()) |
1734 | return ThrowingBlocks.count(SI->getParent()); |
1735 | return !ThrowingBlocks.empty(); |
1736 | } |
1737 | |
1738 | // Check if \p NI acts as a DSE barrier for \p SI. The following instructions |
1739 | // act as barriers: |
1740 | // * A memory instruction that may throw and \p SI accesses a non-stack |
1741 | // object. |
1742 | // * Atomic stores stronger that monotonic. |
1743 | bool isDSEBarrier(const Value *SILocUnd, Instruction *NI) { |
1744 | // If NI may throw it acts as a barrier, unless we are to an alloca/alloca |
1745 | // like object that does not escape. |
1746 | if (NI->mayThrow() && !isInvisibleToCallerBeforeRet(SILocUnd)) |
1747 | return true; |
1748 | |
1749 | // If NI is an atomic load/store stronger than monotonic, do not try to |
1750 | // eliminate/reorder it. |
1751 | if (NI->isAtomic()) { |
1752 | if (auto *LI = dyn_cast<LoadInst>(NI)) |
1753 | return isStrongerThanMonotonic(LI->getOrdering()); |
1754 | if (auto *SI = dyn_cast<StoreInst>(NI)) |
1755 | return isStrongerThanMonotonic(SI->getOrdering()); |
1756 | if (auto *ARMW = dyn_cast<AtomicRMWInst>(NI)) |
1757 | return isStrongerThanMonotonic(ARMW->getOrdering()); |
1758 | if (auto *CmpXchg = dyn_cast<AtomicCmpXchgInst>(NI)) |
1759 | return isStrongerThanMonotonic(CmpXchg->getSuccessOrdering()) || |
1760 | isStrongerThanMonotonic(CmpXchg->getFailureOrdering()); |
1761 | llvm_unreachable("other instructions should be skipped in MemorySSA")__builtin_unreachable(); |
1762 | } |
1763 | return false; |
1764 | } |
1765 | |
1766 | /// Eliminate writes to objects that are not visible in the caller and are not |
1767 | /// accessed before returning from the function. |
1768 | bool eliminateDeadWritesAtEndOfFunction() { |
1769 | bool MadeChange = false; |
1770 | LLVM_DEBUG(do { } while (false) |
1771 | dbgs()do { } while (false) |
1772 | << "Trying to eliminate MemoryDefs at the end of the function\n")do { } while (false); |
1773 | for (int I = MemDefs.size() - 1; I >= 0; I--) { |
1774 | MemoryDef *Def = MemDefs[I]; |
1775 | if (SkipStores.contains(Def) || !isRemovable(Def->getMemoryInst())) |
1776 | continue; |
1777 | |
1778 | Instruction *DefI = Def->getMemoryInst(); |
1779 | SmallVector<const Value *, 4> Pointers; |
1780 | auto DefLoc = getLocForWriteEx(DefI); |
1781 | if (!DefLoc) |
1782 | continue; |
1783 | |
1784 | // NOTE: Currently eliminating writes at the end of a function is limited |
1785 | // to MemoryDefs with a single underlying object, to save compile-time. In |
1786 | // practice it appears the case with multiple underlying objects is very |
1787 | // uncommon. If it turns out to be important, we can use |
1788 | // getUnderlyingObjects here instead. |
1789 | const Value *UO = getUnderlyingObject(DefLoc->Ptr); |
1790 | if (!UO || !isInvisibleToCallerAfterRet(UO)) |
1791 | continue; |
1792 | |
1793 | if (isWriteAtEndOfFunction(Def)) { |
1794 | // See through pointer-to-pointer bitcasts |
1795 | LLVM_DEBUG(dbgs() << " ... MemoryDef is not accessed until the end "do { } while (false) |
1796 | "of the function\n")do { } while (false); |
1797 | deleteDeadInstruction(DefI); |
1798 | ++NumFastStores; |
1799 | MadeChange = true; |
1800 | } |
1801 | } |
1802 | return MadeChange; |
1803 | } |
1804 | |
1805 | /// \returns true if \p Def is a no-op store, either because it |
1806 | /// directly stores back a loaded value or stores zero to a calloced object. |
1807 | bool storeIsNoop(MemoryDef *Def, const MemoryLocation &DefLoc, |
1808 | const Value *DefUO) { |
1809 | StoreInst *Store = dyn_cast<StoreInst>(Def->getMemoryInst()); |
1810 | MemSetInst *MemSet = dyn_cast<MemSetInst>(Def->getMemoryInst()); |
1811 | Constant *StoredConstant = nullptr; |
1812 | if (Store) |
1813 | StoredConstant = dyn_cast<Constant>(Store->getOperand(0)); |
1814 | if (MemSet) |
1815 | StoredConstant = dyn_cast<Constant>(MemSet->getValue()); |
1816 | |
1817 | if (StoredConstant && StoredConstant->isNullValue()) { |
1818 | auto *DefUOInst = dyn_cast<Instruction>(DefUO); |
1819 | if (DefUOInst && isCallocLikeFn(DefUOInst, &TLI)) { |
1820 | auto *UnderlyingDef = cast<MemoryDef>(MSSA.getMemoryAccess(DefUOInst)); |
1821 | // If UnderlyingDef is the clobbering access of Def, no instructions |
1822 | // between them can modify the memory location. |
1823 | auto *ClobberDef = |
1824 | MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(Def); |
1825 | return UnderlyingDef == ClobberDef; |
1826 | } |
1827 | } |
1828 | |
1829 | if (!Store) |
1830 | return false; |
1831 | |
1832 | if (auto *LoadI = dyn_cast<LoadInst>(Store->getOperand(0))) { |
1833 | if (LoadI->getPointerOperand() == Store->getOperand(1)) { |
1834 | // Get the defining access for the load. |
1835 | auto *LoadAccess = MSSA.getMemoryAccess(LoadI)->getDefiningAccess(); |
1836 | // Fast path: the defining accesses are the same. |
1837 | if (LoadAccess == Def->getDefiningAccess()) |
1838 | return true; |
1839 | |
1840 | // Look through phi accesses. Recursively scan all phi accesses by |
1841 | // adding them to a worklist. Bail when we run into a memory def that |
1842 | // does not match LoadAccess. |
1843 | SetVector<MemoryAccess *> ToCheck; |
1844 | MemoryAccess *Current = |
1845 | MSSA.getWalker()->getClobberingMemoryAccess(Def); |
1846 | // We don't want to bail when we run into the store memory def. But, |
1847 | // the phi access may point to it. So, pretend like we've already |
1848 | // checked it. |
1849 | ToCheck.insert(Def); |
1850 | ToCheck.insert(Current); |
1851 | // Start at current (1) to simulate already having checked Def. |
1852 | for (unsigned I = 1; I < ToCheck.size(); ++I) { |
1853 | Current = ToCheck[I]; |
1854 | if (auto PhiAccess = dyn_cast<MemoryPhi>(Current)) { |
1855 | // Check all the operands. |
1856 | for (auto &Use : PhiAccess->incoming_values()) |
1857 | ToCheck.insert(cast<MemoryAccess>(&Use)); |
1858 | continue; |
1859 | } |
1860 | |
1861 | // If we found a memory def, bail. This happens when we have an |
1862 | // unrelated write in between an otherwise noop store. |
1863 | assert(isa<MemoryDef>(Current) &&((void)0) |
1864 | "Only MemoryDefs should reach here.")((void)0); |
1865 | // TODO: Skip no alias MemoryDefs that have no aliasing reads. |
1866 | // We are searching for the definition of the store's destination. |
1867 | // So, if that is the same definition as the load, then this is a |
1868 | // noop. Otherwise, fail. |
1869 | if (LoadAccess != Current) |
1870 | return false; |
1871 | } |
1872 | return true; |
1873 | } |
1874 | } |
1875 | |
1876 | return false; |
1877 | } |
1878 | }; |
1879 | |
1880 | static bool eliminateDeadStores(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, |
1881 | DominatorTree &DT, PostDominatorTree &PDT, |
1882 | const TargetLibraryInfo &TLI, |
1883 | const LoopInfo &LI) { |
1884 | bool MadeChange = false; |
1885 | |
1886 | DSEState State = DSEState::get(F, AA, MSSA, DT, PDT, TLI, LI); |
1887 | // For each store: |
1888 | for (unsigned I = 0; I < State.MemDefs.size(); I++) { |
1889 | MemoryDef *KillingDef = State.MemDefs[I]; |
1890 | if (State.SkipStores.count(KillingDef)) |
1891 | continue; |
1892 | Instruction *SI = KillingDef->getMemoryInst(); |
1893 | |
1894 | Optional<MemoryLocation> MaybeSILoc; |
1895 | if (State.isMemTerminatorInst(SI)) |
1896 | MaybeSILoc = State.getLocForTerminator(SI).map( |
1897 | [](const std::pair<MemoryLocation, bool> &P) { return P.first; }); |
1898 | else |
1899 | MaybeSILoc = State.getLocForWriteEx(SI); |
1900 | |
1901 | if (!MaybeSILoc) { |
1902 | LLVM_DEBUG(dbgs() << "Failed to find analyzable write location for "do { } while (false) |
1903 | << *SI << "\n")do { } while (false); |
1904 | continue; |
1905 | } |
1906 | MemoryLocation SILoc = *MaybeSILoc; |
1907 | assert(SILoc.Ptr && "SILoc should not be null")((void)0); |
1908 | const Value *SILocUnd = getUnderlyingObject(SILoc.Ptr); |
1909 | |
1910 | MemoryAccess *Current = KillingDef; |
1911 | LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs killed by "do { } while (false) |
1912 | << *Current << " (" << *SI << ")\n")do { } while (false); |
1913 | |
1914 | unsigned ScanLimit = MemorySSAScanLimit; |
1915 | unsigned WalkerStepLimit = MemorySSAUpwardsStepLimit; |
1916 | unsigned PartialLimit = MemorySSAPartialStoreLimit; |
1917 | // Worklist of MemoryAccesses that may be killed by KillingDef. |
1918 | SetVector<MemoryAccess *> ToCheck; |
1919 | |
1920 | if (SILocUnd) |
1921 | ToCheck.insert(KillingDef->getDefiningAccess()); |
1922 | |
1923 | bool Shortend = false; |
1924 | bool IsMemTerm = State.isMemTerminatorInst(SI); |
1925 | // Check if MemoryAccesses in the worklist are killed by KillingDef. |
1926 | for (unsigned I = 0; I < ToCheck.size(); I++) { |
1927 | Current = ToCheck[I]; |
1928 | if (State.SkipStores.count(Current)) |
1929 | continue; |
1930 | |
1931 | Optional<MemoryAccess *> Next = State.getDomMemoryDef( |
1932 | KillingDef, Current, SILoc, SILocUnd, ScanLimit, WalkerStepLimit, |
1933 | IsMemTerm, PartialLimit); |
1934 | |
1935 | if (!Next) { |
1936 | LLVM_DEBUG(dbgs() << " finished walk\n")do { } while (false); |
1937 | continue; |
1938 | } |
1939 | |
1940 | MemoryAccess *EarlierAccess = *Next; |
1941 | LLVM_DEBUG(dbgs() << " Checking if we can kill " << *EarlierAccess)do { } while (false); |
1942 | if (isa<MemoryPhi>(EarlierAccess)) { |
1943 | LLVM_DEBUG(dbgs() << "\n ... adding incoming values to worklist\n")do { } while (false); |
1944 | for (Value *V : cast<MemoryPhi>(EarlierAccess)->incoming_values()) { |
1945 | MemoryAccess *IncomingAccess = cast<MemoryAccess>(V); |
1946 | BasicBlock *IncomingBlock = IncomingAccess->getBlock(); |
1947 | BasicBlock *PhiBlock = EarlierAccess->getBlock(); |
1948 | |
1949 | // We only consider incoming MemoryAccesses that come before the |
1950 | // MemoryPhi. Otherwise we could discover candidates that do not |
1951 | // strictly dominate our starting def. |
1952 | if (State.PostOrderNumbers[IncomingBlock] > |
1953 | State.PostOrderNumbers[PhiBlock]) |
1954 | ToCheck.insert(IncomingAccess); |
1955 | } |
1956 | continue; |
1957 | } |
1958 | auto *NextDef = cast<MemoryDef>(EarlierAccess); |
1959 | Instruction *NI = NextDef->getMemoryInst(); |
1960 | LLVM_DEBUG(dbgs() << " (" << *NI << ")\n")do { } while (false); |
1961 | ToCheck.insert(NextDef->getDefiningAccess()); |
1962 | NumGetDomMemoryDefPassed++; |
1963 | |
1964 | if (!DebugCounter::shouldExecute(MemorySSACounter)) |
1965 | continue; |
1966 | |
1967 | MemoryLocation NILoc = *State.getLocForWriteEx(NI); |
1968 | |
1969 | if (IsMemTerm) { |
1970 | const Value *NIUnd = getUnderlyingObject(NILoc.Ptr); |
1971 | if (SILocUnd != NIUnd) |
1972 | continue; |
1973 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NIdo { } while (false) |
1974 | << "\n KILLER: " << *SI << '\n')do { } while (false); |
1975 | State.deleteDeadInstruction(NI); |
1976 | ++NumFastStores; |
1977 | MadeChange = true; |
1978 | } else { |
1979 | // Check if NI overwrites SI. |
1980 | int64_t InstWriteOffset, DepWriteOffset; |
1981 | OverwriteResult OR = State.isOverwrite(SI, NI, SILoc, NILoc, |
1982 | DepWriteOffset, InstWriteOffset); |
1983 | if (OR == OW_MaybePartial) { |
1984 | auto Iter = State.IOLs.insert( |
1985 | std::make_pair<BasicBlock *, InstOverlapIntervalsTy>( |
1986 | NI->getParent(), InstOverlapIntervalsTy())); |
1987 | auto &IOL = Iter.first->second; |
1988 | OR = isPartialOverwrite(SILoc, NILoc, DepWriteOffset, InstWriteOffset, |
1989 | NI, IOL); |
1990 | } |
1991 | |
1992 | if (EnablePartialStoreMerging && OR == OW_PartialEarlierWithFullLater) { |
1993 | auto *Earlier = dyn_cast<StoreInst>(NI); |
1994 | auto *Later = dyn_cast<StoreInst>(SI); |
1995 | // We are re-using tryToMergePartialOverlappingStores, which requires |
1996 | // Earlier to domiante Later. |
1997 | // TODO: implement tryToMergeParialOverlappingStores using MemorySSA. |
1998 | if (Earlier && Later && DT.dominates(Earlier, Later)) { |
1999 | if (Constant *Merged = tryToMergePartialOverlappingStores( |
2000 | Earlier, Later, InstWriteOffset, DepWriteOffset, State.DL, |
2001 | State.BatchAA, &DT)) { |
2002 | |
2003 | // Update stored value of earlier store to merged constant. |
2004 | Earlier->setOperand(0, Merged); |
2005 | ++NumModifiedStores; |
2006 | MadeChange = true; |
2007 | |
2008 | Shortend = true; |
2009 | // Remove later store and remove any outstanding overlap intervals |
2010 | // for the updated store. |
2011 | State.deleteDeadInstruction(Later); |
2012 | auto I = State.IOLs.find(Earlier->getParent()); |
2013 | if (I != State.IOLs.end()) |
2014 | I->second.erase(Earlier); |
2015 | break; |
2016 | } |
2017 | } |
2018 | } |
2019 | |
2020 | if (OR == OW_Complete) { |
2021 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NIdo { } while (false) |
2022 | << "\n KILLER: " << *SI << '\n')do { } while (false); |
2023 | State.deleteDeadInstruction(NI); |
2024 | ++NumFastStores; |
2025 | MadeChange = true; |
2026 | } |
2027 | } |
2028 | } |
2029 | |
2030 | // Check if the store is a no-op. |
2031 | if (!Shortend && isRemovable(SI) && |
2032 | State.storeIsNoop(KillingDef, SILoc, SILocUnd)) { |
2033 | LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: " << *SI << '\n')do { } while (false); |
2034 | State.deleteDeadInstruction(SI); |
2035 | NumRedundantStores++; |
2036 | MadeChange = true; |
2037 | continue; |
2038 | } |
2039 | } |
2040 | |
2041 | if (EnablePartialOverwriteTracking) |
2042 | for (auto &KV : State.IOLs) |
2043 | MadeChange |= removePartiallyOverlappedStores(State.DL, KV.second, TLI); |
2044 | |
2045 | MadeChange |= State.eliminateDeadWritesAtEndOfFunction(); |
2046 | return MadeChange; |
2047 | } |
2048 | } // end anonymous namespace |
2049 | |
2050 | //===----------------------------------------------------------------------===// |
2051 | // DSE Pass |
2052 | //===----------------------------------------------------------------------===// |
2053 | PreservedAnalyses DSEPass::run(Function &F, FunctionAnalysisManager &AM) { |
2054 | AliasAnalysis &AA = AM.getResult<AAManager>(F); |
2055 | const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(F); |
2056 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F); |
2057 | MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); |
2058 | PostDominatorTree &PDT = AM.getResult<PostDominatorTreeAnalysis>(F); |
2059 | LoopInfo &LI = AM.getResult<LoopAnalysis>(F); |
2060 | |
2061 | bool Changed = eliminateDeadStores(F, AA, MSSA, DT, PDT, TLI, LI); |
2062 | |
2063 | #ifdef LLVM_ENABLE_STATS0 |
2064 | if (AreStatisticsEnabled()) |
2065 | for (auto &I : instructions(F)) |
2066 | NumRemainingStores += isa<StoreInst>(&I); |
2067 | #endif |
2068 | |
2069 | if (!Changed) |
2070 | return PreservedAnalyses::all(); |
2071 | |
2072 | PreservedAnalyses PA; |
2073 | PA.preserveSet<CFGAnalyses>(); |
2074 | PA.preserve<MemorySSAAnalysis>(); |
2075 | PA.preserve<LoopAnalysis>(); |
2076 | return PA; |
2077 | } |
2078 | |
2079 | namespace { |
2080 | |
2081 | /// A legacy pass for the legacy pass manager that wraps \c DSEPass. |
2082 | class DSELegacyPass : public FunctionPass { |
2083 | public: |
2084 | static char ID; // Pass identification, replacement for typeid |
2085 | |
2086 | DSELegacyPass() : FunctionPass(ID) { |
2087 | initializeDSELegacyPassPass(*PassRegistry::getPassRegistry()); |
2088 | } |
2089 | |
2090 | bool runOnFunction(Function &F) override { |
2091 | if (skipFunction(F)) |
2092 | return false; |
2093 | |
2094 | AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
2095 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
2096 | const TargetLibraryInfo &TLI = |
2097 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
2098 | MemorySSA &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); |
2099 | PostDominatorTree &PDT = |
2100 | getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree(); |
2101 | LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
2102 | |
2103 | bool Changed = eliminateDeadStores(F, AA, MSSA, DT, PDT, TLI, LI); |
2104 | |
2105 | #ifdef LLVM_ENABLE_STATS0 |
2106 | if (AreStatisticsEnabled()) |
2107 | for (auto &I : instructions(F)) |
2108 | NumRemainingStores += isa<StoreInst>(&I); |
2109 | #endif |
2110 | |
2111 | return Changed; |
2112 | } |
2113 | |
2114 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
2115 | AU.setPreservesCFG(); |
2116 | AU.addRequired<AAResultsWrapperPass>(); |
2117 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
2118 | AU.addPreserved<GlobalsAAWrapperPass>(); |
2119 | AU.addRequired<DominatorTreeWrapperPass>(); |
2120 | AU.addPreserved<DominatorTreeWrapperPass>(); |
2121 | AU.addRequired<PostDominatorTreeWrapperPass>(); |
2122 | AU.addRequired<MemorySSAWrapperPass>(); |
2123 | AU.addPreserved<PostDominatorTreeWrapperPass>(); |
2124 | AU.addPreserved<MemorySSAWrapperPass>(); |
2125 | AU.addRequired<LoopInfoWrapperPass>(); |
2126 | AU.addPreserved<LoopInfoWrapperPass>(); |
2127 | } |
2128 | }; |
2129 | |
2130 | } // end anonymous namespace |
2131 | |
2132 | char DSELegacyPass::ID = 0; |
2133 | |
2134 | INITIALIZE_PASS_BEGIN(DSELegacyPass, "dse", "Dead Store Elimination", false,static void *initializeDSELegacyPassPassOnce(PassRegistry & Registry) { |
2135 | false)static void *initializeDSELegacyPassPassOnce(PassRegistry & Registry) { |
2136 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
2137 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)initializePostDominatorTreeWrapperPassPass(Registry); |
2138 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
2139 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); |
2140 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); |
2141 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)initializeMemoryDependenceWrapperPassPass(Registry); |
2142 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
2143 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); |
2144 | INITIALIZE_PASS_END(DSELegacyPass, "dse", "Dead Store Elimination", false,PassInfo *PI = new PassInfo( "Dead Store Elimination", "dse", &DSELegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <DSELegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeDSELegacyPassPassFlag ; void llvm::initializeDSELegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeDSELegacyPassPassFlag, initializeDSELegacyPassPassOnce , std::ref(Registry)); } |
2145 | false)PassInfo *PI = new PassInfo( "Dead Store Elimination", "dse", &DSELegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <DSELegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeDSELegacyPassPassFlag ; void llvm::initializeDSELegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeDSELegacyPassPassFlag, initializeDSELegacyPassPassOnce , std::ref(Registry)); } |
2146 | |
2147 | FunctionPass *llvm::createDeadStoreEliminationPass() { |
2148 | return new DSELegacyPass(); |
2149 | } |