File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/include/llvm/ADT/PointerIntPair.h |
Warning: | line 190, column 52 The result of the left shift is undefined due to shifting '0' by '2', which is unrepresentable in the unsigned version of the return type 'intptr_t' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==// | ||||
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 implementation for the loop memory dependence that was originally | ||||
10 | // developed for the loop vectorizer. | ||||
11 | // | ||||
12 | //===----------------------------------------------------------------------===// | ||||
13 | |||||
14 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||
15 | #include "llvm/ADT/APInt.h" | ||||
16 | #include "llvm/ADT/DenseMap.h" | ||||
17 | #include "llvm/ADT/DepthFirstIterator.h" | ||||
18 | #include "llvm/ADT/EquivalenceClasses.h" | ||||
19 | #include "llvm/ADT/PointerIntPair.h" | ||||
20 | #include "llvm/ADT/STLExtras.h" | ||||
21 | #include "llvm/ADT/SetVector.h" | ||||
22 | #include "llvm/ADT/SmallPtrSet.h" | ||||
23 | #include "llvm/ADT/SmallSet.h" | ||||
24 | #include "llvm/ADT/SmallVector.h" | ||||
25 | #include "llvm/ADT/iterator_range.h" | ||||
26 | #include "llvm/Analysis/AliasAnalysis.h" | ||||
27 | #include "llvm/Analysis/AliasSetTracker.h" | ||||
28 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||
29 | #include "llvm/Analysis/LoopInfo.h" | ||||
30 | #include "llvm/Analysis/MemoryLocation.h" | ||||
31 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||
32 | #include "llvm/Analysis/ScalarEvolution.h" | ||||
33 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||
34 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||
35 | #include "llvm/Analysis/ValueTracking.h" | ||||
36 | #include "llvm/Analysis/VectorUtils.h" | ||||
37 | #include "llvm/IR/BasicBlock.h" | ||||
38 | #include "llvm/IR/Constants.h" | ||||
39 | #include "llvm/IR/DataLayout.h" | ||||
40 | #include "llvm/IR/DebugLoc.h" | ||||
41 | #include "llvm/IR/DerivedTypes.h" | ||||
42 | #include "llvm/IR/DiagnosticInfo.h" | ||||
43 | #include "llvm/IR/Dominators.h" | ||||
44 | #include "llvm/IR/Function.h" | ||||
45 | #include "llvm/IR/InstrTypes.h" | ||||
46 | #include "llvm/IR/Instruction.h" | ||||
47 | #include "llvm/IR/Instructions.h" | ||||
48 | #include "llvm/IR/Operator.h" | ||||
49 | #include "llvm/IR/PassManager.h" | ||||
50 | #include "llvm/IR/Type.h" | ||||
51 | #include "llvm/IR/Value.h" | ||||
52 | #include "llvm/IR/ValueHandle.h" | ||||
53 | #include "llvm/InitializePasses.h" | ||||
54 | #include "llvm/Pass.h" | ||||
55 | #include "llvm/Support/Casting.h" | ||||
56 | #include "llvm/Support/CommandLine.h" | ||||
57 | #include "llvm/Support/Debug.h" | ||||
58 | #include "llvm/Support/ErrorHandling.h" | ||||
59 | #include "llvm/Support/raw_ostream.h" | ||||
60 | #include <algorithm> | ||||
61 | #include <cassert> | ||||
62 | #include <cstdint> | ||||
63 | #include <iterator> | ||||
64 | #include <utility> | ||||
65 | #include <vector> | ||||
66 | |||||
67 | using namespace llvm; | ||||
68 | |||||
69 | #define DEBUG_TYPE"loop-accesses" "loop-accesses" | ||||
70 | |||||
71 | static cl::opt<unsigned, true> | ||||
72 | VectorizationFactor("force-vector-width", cl::Hidden, | ||||
73 | cl::desc("Sets the SIMD width. Zero is autoselect."), | ||||
74 | cl::location(VectorizerParams::VectorizationFactor)); | ||||
75 | unsigned VectorizerParams::VectorizationFactor; | ||||
76 | |||||
77 | static cl::opt<unsigned, true> | ||||
78 | VectorizationInterleave("force-vector-interleave", cl::Hidden, | ||||
79 | cl::desc("Sets the vectorization interleave count. " | ||||
80 | "Zero is autoselect."), | ||||
81 | cl::location( | ||||
82 | VectorizerParams::VectorizationInterleave)); | ||||
83 | unsigned VectorizerParams::VectorizationInterleave; | ||||
84 | |||||
85 | static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold( | ||||
86 | "runtime-memory-check-threshold", cl::Hidden, | ||||
87 | cl::desc("When performing memory disambiguation checks at runtime do not " | ||||
88 | "generate more than this number of comparisons (default = 8)."), | ||||
89 | cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8)); | ||||
90 | unsigned VectorizerParams::RuntimeMemoryCheckThreshold; | ||||
91 | |||||
92 | /// The maximum iterations used to merge memory checks | ||||
93 | static cl::opt<unsigned> MemoryCheckMergeThreshold( | ||||
94 | "memory-check-merge-threshold", cl::Hidden, | ||||
95 | cl::desc("Maximum number of comparisons done when trying to merge " | ||||
96 | "runtime memory checks. (default = 100)"), | ||||
97 | cl::init(100)); | ||||
98 | |||||
99 | /// Maximum SIMD width. | ||||
100 | const unsigned VectorizerParams::MaxVectorWidth = 64; | ||||
101 | |||||
102 | /// We collect dependences up to this threshold. | ||||
103 | static cl::opt<unsigned> | ||||
104 | MaxDependences("max-dependences", cl::Hidden, | ||||
105 | cl::desc("Maximum number of dependences collected by " | ||||
106 | "loop-access analysis (default = 100)"), | ||||
107 | cl::init(100)); | ||||
108 | |||||
109 | /// This enables versioning on the strides of symbolically striding memory | ||||
110 | /// accesses in code like the following. | ||||
111 | /// for (i = 0; i < N; ++i) | ||||
112 | /// A[i * Stride1] += B[i * Stride2] ... | ||||
113 | /// | ||||
114 | /// Will be roughly translated to | ||||
115 | /// if (Stride1 == 1 && Stride2 == 1) { | ||||
116 | /// for (i = 0; i < N; i+=4) | ||||
117 | /// A[i:i+3] += ... | ||||
118 | /// } else | ||||
119 | /// ... | ||||
120 | static cl::opt<bool> EnableMemAccessVersioning( | ||||
121 | "enable-mem-access-versioning", cl::init(true), cl::Hidden, | ||||
122 | cl::desc("Enable symbolic stride memory access versioning")); | ||||
123 | |||||
124 | /// Enable store-to-load forwarding conflict detection. This option can | ||||
125 | /// be disabled for correctness testing. | ||||
126 | static cl::opt<bool> EnableForwardingConflictDetection( | ||||
127 | "store-to-load-forwarding-conflict-detection", cl::Hidden, | ||||
128 | cl::desc("Enable conflict detection in loop-access analysis"), | ||||
129 | cl::init(true)); | ||||
130 | |||||
131 | bool VectorizerParams::isInterleaveForced() { | ||||
132 | return ::VectorizationInterleave.getNumOccurrences() > 0; | ||||
133 | } | ||||
134 | |||||
135 | Value *llvm::stripIntegerCast(Value *V) { | ||||
136 | if (auto *CI = dyn_cast<CastInst>(V)) | ||||
137 | if (CI->getOperand(0)->getType()->isIntegerTy()) | ||||
138 | return CI->getOperand(0); | ||||
139 | return V; | ||||
140 | } | ||||
141 | |||||
142 | const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, | ||||
143 | const ValueToValueMap &PtrToStride, | ||||
144 | Value *Ptr) { | ||||
145 | const SCEV *OrigSCEV = PSE.getSCEV(Ptr); | ||||
146 | |||||
147 | // If there is an entry in the map return the SCEV of the pointer with the | ||||
148 | // symbolic stride replaced by one. | ||||
149 | ValueToValueMap::const_iterator SI = PtrToStride.find(Ptr); | ||||
150 | if (SI == PtrToStride.end()) | ||||
151 | // For a non-symbolic stride, just return the original expression. | ||||
152 | return OrigSCEV; | ||||
153 | |||||
154 | Value *StrideVal = stripIntegerCast(SI->second); | ||||
155 | |||||
156 | ScalarEvolution *SE = PSE.getSE(); | ||||
157 | const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal)); | ||||
158 | const auto *CT = | ||||
159 | static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType())); | ||||
160 | |||||
161 | PSE.addPredicate(*SE->getEqualPredicate(U, CT)); | ||||
162 | auto *Expr = PSE.getSCEV(Ptr); | ||||
163 | |||||
164 | LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEVdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr << "\n"; } } while (false) | ||||
165 | << " by: " << *Expr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr << "\n"; } } while (false); | ||||
166 | return Expr; | ||||
167 | } | ||||
168 | |||||
169 | RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup( | ||||
170 | unsigned Index, RuntimePointerChecking &RtCheck) | ||||
171 | : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start), | ||||
172 | AddressSpace(RtCheck.Pointers[Index] | ||||
173 | .PointerValue->getType() | ||||
174 | ->getPointerAddressSpace()) { | ||||
175 | Members.push_back(Index); | ||||
176 | } | ||||
177 | |||||
178 | /// Calculate Start and End points of memory access. | ||||
179 | /// Let's assume A is the first access and B is a memory access on N-th loop | ||||
180 | /// iteration. Then B is calculated as: | ||||
181 | /// B = A + Step*N . | ||||
182 | /// Step value may be positive or negative. | ||||
183 | /// N is a calculated back-edge taken count: | ||||
184 | /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0 | ||||
185 | /// Start and End points are calculated in the following way: | ||||
186 | /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt, | ||||
187 | /// where SizeOfElt is the size of single memory access in bytes. | ||||
188 | /// | ||||
189 | /// There is no conflict when the intervals are disjoint: | ||||
190 | /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End) | ||||
191 | void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, Type *AccessTy, | ||||
192 | bool WritePtr, unsigned DepSetId, | ||||
193 | unsigned ASId, | ||||
194 | const ValueToValueMap &Strides, | ||||
195 | PredicatedScalarEvolution &PSE) { | ||||
196 | // Get the stride replaced scev. | ||||
197 | const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); | ||||
198 | ScalarEvolution *SE = PSE.getSE(); | ||||
199 | |||||
200 | const SCEV *ScStart; | ||||
201 | const SCEV *ScEnd; | ||||
202 | |||||
203 | if (SE->isLoopInvariant(Sc, Lp)) { | ||||
204 | ScStart = ScEnd = Sc; | ||||
205 | } else { | ||||
206 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); | ||||
207 | assert(AR && "Invalid addrec expression")(static_cast <bool> (AR && "Invalid addrec expression" ) ? void (0) : __assert_fail ("AR && \"Invalid addrec expression\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 207, __extension__ __PRETTY_FUNCTION__)); | ||||
208 | const SCEV *Ex = PSE.getBackedgeTakenCount(); | ||||
209 | |||||
210 | ScStart = AR->getStart(); | ||||
211 | ScEnd = AR->evaluateAtIteration(Ex, *SE); | ||||
212 | const SCEV *Step = AR->getStepRecurrence(*SE); | ||||
213 | |||||
214 | // For expressions with negative step, the upper bound is ScStart and the | ||||
215 | // lower bound is ScEnd. | ||||
216 | if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) { | ||||
217 | if (CStep->getValue()->isNegative()) | ||||
218 | std::swap(ScStart, ScEnd); | ||||
219 | } else { | ||||
220 | // Fallback case: the step is not constant, but we can still | ||||
221 | // get the upper and lower bounds of the interval by using min/max | ||||
222 | // expressions. | ||||
223 | ScStart = SE->getUMinExpr(ScStart, ScEnd); | ||||
224 | ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd); | ||||
225 | } | ||||
226 | } | ||||
227 | // Add the size of the pointed element to ScEnd. | ||||
228 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); | ||||
229 | Type *IdxTy = DL.getIndexType(Ptr->getType()); | ||||
230 | const SCEV *EltSizeSCEV = SE->getStoreSizeOfExpr(IdxTy, AccessTy); | ||||
231 | ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV); | ||||
232 | |||||
233 | Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc); | ||||
234 | } | ||||
235 | |||||
236 | SmallVector<RuntimePointerCheck, 4> | ||||
237 | RuntimePointerChecking::generateChecks() const { | ||||
238 | SmallVector<RuntimePointerCheck, 4> Checks; | ||||
239 | |||||
240 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { | ||||
241 | for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) { | ||||
242 | const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I]; | ||||
243 | const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J]; | ||||
244 | |||||
245 | if (needsChecking(CGI, CGJ)) | ||||
246 | Checks.push_back(std::make_pair(&CGI, &CGJ)); | ||||
247 | } | ||||
248 | } | ||||
249 | return Checks; | ||||
250 | } | ||||
251 | |||||
252 | void RuntimePointerChecking::generateChecks( | ||||
253 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { | ||||
254 | assert(Checks.empty() && "Checks is not empty")(static_cast <bool> (Checks.empty() && "Checks is not empty" ) ? void (0) : __assert_fail ("Checks.empty() && \"Checks is not empty\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 254, __extension__ __PRETTY_FUNCTION__)); | ||||
255 | groupChecks(DepCands, UseDependencies); | ||||
256 | Checks = generateChecks(); | ||||
257 | } | ||||
258 | |||||
259 | bool RuntimePointerChecking::needsChecking( | ||||
260 | const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const { | ||||
261 | for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I) | ||||
262 | for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J) | ||||
263 | if (needsChecking(M.Members[I], N.Members[J])) | ||||
264 | return true; | ||||
265 | return false; | ||||
266 | } | ||||
267 | |||||
268 | /// Compare \p I and \p J and return the minimum. | ||||
269 | /// Return nullptr in case we couldn't find an answer. | ||||
270 | static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J, | ||||
271 | ScalarEvolution *SE) { | ||||
272 | const SCEV *Diff = SE->getMinusSCEV(J, I); | ||||
273 | const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff); | ||||
274 | |||||
275 | if (!C) | ||||
276 | return nullptr; | ||||
277 | if (C->getValue()->isNegative()) | ||||
278 | return J; | ||||
279 | return I; | ||||
280 | } | ||||
281 | |||||
282 | bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, | ||||
283 | RuntimePointerChecking &RtCheck) { | ||||
284 | return addPointer( | ||||
285 | Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End, | ||||
286 | RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(), | ||||
287 | *RtCheck.SE); | ||||
288 | } | ||||
289 | |||||
290 | bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start, | ||||
291 | const SCEV *End, unsigned AS, | ||||
292 | ScalarEvolution &SE) { | ||||
293 | assert(AddressSpace == AS &&(static_cast <bool> (AddressSpace == AS && "all pointers in a checking group must be in the same address space" ) ? void (0) : __assert_fail ("AddressSpace == AS && \"all pointers in a checking group must be in the same address space\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 294, __extension__ __PRETTY_FUNCTION__)) | ||||
294 | "all pointers in a checking group must be in the same address space")(static_cast <bool> (AddressSpace == AS && "all pointers in a checking group must be in the same address space" ) ? void (0) : __assert_fail ("AddressSpace == AS && \"all pointers in a checking group must be in the same address space\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 294, __extension__ __PRETTY_FUNCTION__)); | ||||
295 | |||||
296 | // Compare the starts and ends with the known minimum and maximum | ||||
297 | // of this set. We need to know how we compare against the min/max | ||||
298 | // of the set in order to be able to emit memchecks. | ||||
299 | const SCEV *Min0 = getMinFromExprs(Start, Low, &SE); | ||||
300 | if (!Min0) | ||||
301 | return false; | ||||
302 | |||||
303 | const SCEV *Min1 = getMinFromExprs(End, High, &SE); | ||||
304 | if (!Min1) | ||||
305 | return false; | ||||
306 | |||||
307 | // Update the low bound expression if we've found a new min value. | ||||
308 | if (Min0 == Start) | ||||
309 | Low = Start; | ||||
310 | |||||
311 | // Update the high bound expression if we've found a new max value. | ||||
312 | if (Min1 != End) | ||||
313 | High = End; | ||||
314 | |||||
315 | Members.push_back(Index); | ||||
316 | return true; | ||||
317 | } | ||||
318 | |||||
319 | void RuntimePointerChecking::groupChecks( | ||||
320 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { | ||||
321 | // We build the groups from dependency candidates equivalence classes | ||||
322 | // because: | ||||
323 | // - We know that pointers in the same equivalence class share | ||||
324 | // the same underlying object and therefore there is a chance | ||||
325 | // that we can compare pointers | ||||
326 | // - We wouldn't be able to merge two pointers for which we need | ||||
327 | // to emit a memcheck. The classes in DepCands are already | ||||
328 | // conveniently built such that no two pointers in the same | ||||
329 | // class need checking against each other. | ||||
330 | |||||
331 | // We use the following (greedy) algorithm to construct the groups | ||||
332 | // For every pointer in the equivalence class: | ||||
333 | // For each existing group: | ||||
334 | // - if the difference between this pointer and the min/max bounds | ||||
335 | // of the group is a constant, then make the pointer part of the | ||||
336 | // group and update the min/max bounds of that group as required. | ||||
337 | |||||
338 | CheckingGroups.clear(); | ||||
339 | |||||
340 | // If we need to check two pointers to the same underlying object | ||||
341 | // with a non-constant difference, we shouldn't perform any pointer | ||||
342 | // grouping with those pointers. This is because we can easily get | ||||
343 | // into cases where the resulting check would return false, even when | ||||
344 | // the accesses are safe. | ||||
345 | // | ||||
346 | // The following example shows this: | ||||
347 | // for (i = 0; i < 1000; ++i) | ||||
348 | // a[5000 + i * m] = a[i] + a[i + 9000] | ||||
349 | // | ||||
350 | // Here grouping gives a check of (5000, 5000 + 1000 * m) against | ||||
351 | // (0, 10000) which is always false. However, if m is 1, there is no | ||||
352 | // dependence. Not grouping the checks for a[i] and a[i + 9000] allows | ||||
353 | // us to perform an accurate check in this case. | ||||
354 | // | ||||
355 | // The above case requires that we have an UnknownDependence between | ||||
356 | // accesses to the same underlying object. This cannot happen unless | ||||
357 | // FoundNonConstantDistanceDependence is set, and therefore UseDependencies | ||||
358 | // is also false. In this case we will use the fallback path and create | ||||
359 | // separate checking groups for all pointers. | ||||
360 | |||||
361 | // If we don't have the dependency partitions, construct a new | ||||
362 | // checking pointer group for each pointer. This is also required | ||||
363 | // for correctness, because in this case we can have checking between | ||||
364 | // pointers to the same underlying object. | ||||
365 | if (!UseDependencies) { | ||||
366 | for (unsigned I = 0; I < Pointers.size(); ++I) | ||||
367 | CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this)); | ||||
368 | return; | ||||
369 | } | ||||
370 | |||||
371 | unsigned TotalComparisons = 0; | ||||
372 | |||||
373 | DenseMap<Value *, unsigned> PositionMap; | ||||
374 | for (unsigned Index = 0; Index < Pointers.size(); ++Index) | ||||
375 | PositionMap[Pointers[Index].PointerValue] = Index; | ||||
376 | |||||
377 | // We need to keep track of what pointers we've already seen so we | ||||
378 | // don't process them twice. | ||||
379 | SmallSet<unsigned, 2> Seen; | ||||
380 | |||||
381 | // Go through all equivalence classes, get the "pointer check groups" | ||||
382 | // and add them to the overall solution. We use the order in which accesses | ||||
383 | // appear in 'Pointers' to enforce determinism. | ||||
384 | for (unsigned I = 0; I < Pointers.size(); ++I) { | ||||
385 | // We've seen this pointer before, and therefore already processed | ||||
386 | // its equivalence class. | ||||
387 | if (Seen.count(I)) | ||||
388 | continue; | ||||
389 | |||||
390 | MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue, | ||||
391 | Pointers[I].IsWritePtr); | ||||
392 | |||||
393 | SmallVector<RuntimeCheckingPtrGroup, 2> Groups; | ||||
394 | auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access)); | ||||
395 | |||||
396 | // Because DepCands is constructed by visiting accesses in the order in | ||||
397 | // which they appear in alias sets (which is deterministic) and the | ||||
398 | // iteration order within an equivalence class member is only dependent on | ||||
399 | // the order in which unions and insertions are performed on the | ||||
400 | // equivalence class, the iteration order is deterministic. | ||||
401 | for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end(); | ||||
402 | MI != ME; ++MI) { | ||||
403 | auto PointerI = PositionMap.find(MI->getPointer()); | ||||
404 | assert(PointerI != PositionMap.end() &&(static_cast <bool> (PointerI != PositionMap.end() && "pointer in equivalence class not found in PositionMap") ? void (0) : __assert_fail ("PointerI != PositionMap.end() && \"pointer in equivalence class not found in PositionMap\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 405, __extension__ __PRETTY_FUNCTION__)) | ||||
405 | "pointer in equivalence class not found in PositionMap")(static_cast <bool> (PointerI != PositionMap.end() && "pointer in equivalence class not found in PositionMap") ? void (0) : __assert_fail ("PointerI != PositionMap.end() && \"pointer in equivalence class not found in PositionMap\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 405, __extension__ __PRETTY_FUNCTION__)); | ||||
406 | unsigned Pointer = PointerI->second; | ||||
407 | bool Merged = false; | ||||
408 | // Mark this pointer as seen. | ||||
409 | Seen.insert(Pointer); | ||||
410 | |||||
411 | // Go through all the existing sets and see if we can find one | ||||
412 | // which can include this pointer. | ||||
413 | for (RuntimeCheckingPtrGroup &Group : Groups) { | ||||
414 | // Don't perform more than a certain amount of comparisons. | ||||
415 | // This should limit the cost of grouping the pointers to something | ||||
416 | // reasonable. If we do end up hitting this threshold, the algorithm | ||||
417 | // will create separate groups for all remaining pointers. | ||||
418 | if (TotalComparisons > MemoryCheckMergeThreshold) | ||||
419 | break; | ||||
420 | |||||
421 | TotalComparisons++; | ||||
422 | |||||
423 | if (Group.addPointer(Pointer, *this)) { | ||||
424 | Merged = true; | ||||
425 | break; | ||||
426 | } | ||||
427 | } | ||||
428 | |||||
429 | if (!Merged) | ||||
430 | // We couldn't add this pointer to any existing set or the threshold | ||||
431 | // for the number of comparisons has been reached. Create a new group | ||||
432 | // to hold the current pointer. | ||||
433 | Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this)); | ||||
434 | } | ||||
435 | |||||
436 | // We've computed the grouped checks for this partition. | ||||
437 | // Save the results and continue with the next one. | ||||
438 | llvm::copy(Groups, std::back_inserter(CheckingGroups)); | ||||
439 | } | ||||
440 | } | ||||
441 | |||||
442 | bool RuntimePointerChecking::arePointersInSamePartition( | ||||
443 | const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1, | ||||
444 | unsigned PtrIdx2) { | ||||
445 | return (PtrToPartition[PtrIdx1] != -1 && | ||||
446 | PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]); | ||||
447 | } | ||||
448 | |||||
449 | bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const { | ||||
450 | const PointerInfo &PointerI = Pointers[I]; | ||||
451 | const PointerInfo &PointerJ = Pointers[J]; | ||||
452 | |||||
453 | // No need to check if two readonly pointers intersect. | ||||
454 | if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr) | ||||
455 | return false; | ||||
456 | |||||
457 | // Only need to check pointers between two different dependency sets. | ||||
458 | if (PointerI.DependencySetId == PointerJ.DependencySetId) | ||||
459 | return false; | ||||
460 | |||||
461 | // Only need to check pointers in the same alias set. | ||||
462 | if (PointerI.AliasSetId != PointerJ.AliasSetId) | ||||
463 | return false; | ||||
464 | |||||
465 | return true; | ||||
466 | } | ||||
467 | |||||
468 | void RuntimePointerChecking::printChecks( | ||||
469 | raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks, | ||||
470 | unsigned Depth) const { | ||||
471 | unsigned N = 0; | ||||
472 | for (const auto &Check : Checks) { | ||||
473 | const auto &First = Check.first->Members, &Second = Check.second->Members; | ||||
474 | |||||
475 | OS.indent(Depth) << "Check " << N++ << ":\n"; | ||||
476 | |||||
477 | OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n"; | ||||
478 | for (unsigned K = 0; K < First.size(); ++K) | ||||
479 | OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n"; | ||||
480 | |||||
481 | OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n"; | ||||
482 | for (unsigned K = 0; K < Second.size(); ++K) | ||||
483 | OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n"; | ||||
484 | } | ||||
485 | } | ||||
486 | |||||
487 | void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const { | ||||
488 | |||||
489 | OS.indent(Depth) << "Run-time memory checks:\n"; | ||||
490 | printChecks(OS, Checks, Depth); | ||||
491 | |||||
492 | OS.indent(Depth) << "Grouped accesses:\n"; | ||||
493 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { | ||||
494 | const auto &CG = CheckingGroups[I]; | ||||
495 | |||||
496 | OS.indent(Depth + 2) << "Group " << &CG << ":\n"; | ||||
497 | OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High | ||||
498 | << ")\n"; | ||||
499 | for (unsigned J = 0; J < CG.Members.size(); ++J) { | ||||
500 | OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr | ||||
501 | << "\n"; | ||||
502 | } | ||||
503 | } | ||||
504 | } | ||||
505 | |||||
506 | namespace { | ||||
507 | |||||
508 | /// Analyses memory accesses in a loop. | ||||
509 | /// | ||||
510 | /// Checks whether run time pointer checks are needed and builds sets for data | ||||
511 | /// dependence checking. | ||||
512 | class AccessAnalysis { | ||||
513 | public: | ||||
514 | /// Read or write access location. | ||||
515 | typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; | ||||
516 | typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; | ||||
517 | |||||
518 | AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI, | ||||
519 | MemoryDepChecker::DepCandidates &DA, | ||||
520 | PredicatedScalarEvolution &PSE) | ||||
521 | : TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA), PSE(PSE) {} | ||||
522 | |||||
523 | /// Register a load and whether it is only read from. | ||||
524 | void addLoad(MemoryLocation &Loc, Type *AccessTy, bool IsReadOnly) { | ||||
525 | Value *Ptr = const_cast<Value*>(Loc.Ptr); | ||||
526 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); | ||||
527 | Accesses[MemAccessInfo(Ptr, false)].insert(AccessTy); | ||||
528 | if (IsReadOnly) | ||||
529 | ReadOnlyPtr.insert(Ptr); | ||||
530 | } | ||||
531 | |||||
532 | /// Register a store. | ||||
533 | void addStore(MemoryLocation &Loc, Type *AccessTy) { | ||||
534 | Value *Ptr = const_cast<Value*>(Loc.Ptr); | ||||
535 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); | ||||
536 | Accesses[MemAccessInfo(Ptr, true)].insert(AccessTy); | ||||
537 | } | ||||
538 | |||||
539 | /// Check if we can emit a run-time no-alias check for \p Access. | ||||
540 | /// | ||||
541 | /// Returns true if we can emit a run-time no alias check for \p Access. | ||||
542 | /// If we can check this access, this also adds it to a dependence set and | ||||
543 | /// adds a run-time to check for it to \p RtCheck. If \p Assume is true, | ||||
544 | /// we will attempt to use additional run-time checks in order to get | ||||
545 | /// the bounds of the pointer. | ||||
546 | bool createCheckForAccess(RuntimePointerChecking &RtCheck, | ||||
547 | MemAccessInfo Access, Type *AccessTy, | ||||
548 | const ValueToValueMap &Strides, | ||||
549 | DenseMap<Value *, unsigned> &DepSetId, | ||||
550 | Loop *TheLoop, unsigned &RunningDepId, | ||||
551 | unsigned ASId, bool ShouldCheckStride, bool Assume); | ||||
552 | |||||
553 | /// Check whether we can check the pointers at runtime for | ||||
554 | /// non-intersection. | ||||
555 | /// | ||||
556 | /// Returns true if we need no check or if we do and we can generate them | ||||
557 | /// (i.e. the pointers have computable bounds). | ||||
558 | bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE, | ||||
559 | Loop *TheLoop, const ValueToValueMap &Strides, | ||||
560 | Value *&UncomputablePtr, bool ShouldCheckWrap = false); | ||||
561 | |||||
562 | /// Goes over all memory accesses, checks whether a RT check is needed | ||||
563 | /// and builds sets of dependent accesses. | ||||
564 | void buildDependenceSets() { | ||||
565 | processMemAccesses(); | ||||
566 | } | ||||
567 | |||||
568 | /// Initial processing of memory accesses determined that we need to | ||||
569 | /// perform dependency checking. | ||||
570 | /// | ||||
571 | /// Note that this can later be cleared if we retry memcheck analysis without | ||||
572 | /// dependency checking (i.e. FoundNonConstantDistanceDependence). | ||||
573 | bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } | ||||
574 | |||||
575 | /// We decided that no dependence analysis would be used. Reset the state. | ||||
576 | void resetDepChecks(MemoryDepChecker &DepChecker) { | ||||
577 | CheckDeps.clear(); | ||||
578 | DepChecker.clearDependences(); | ||||
579 | } | ||||
580 | |||||
581 | MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; } | ||||
582 | |||||
583 | private: | ||||
584 | typedef MapVector<MemAccessInfo, SmallSetVector<Type *, 1>> PtrAccessMap; | ||||
585 | |||||
586 | /// Go over all memory access and check whether runtime pointer checks | ||||
587 | /// are needed and build sets of dependency check candidates. | ||||
588 | void processMemAccesses(); | ||||
589 | |||||
590 | /// Map of all accesses. Values are the types used to access memory pointed to | ||||
591 | /// by the pointer. | ||||
592 | PtrAccessMap Accesses; | ||||
593 | |||||
594 | /// The loop being checked. | ||||
595 | const Loop *TheLoop; | ||||
596 | |||||
597 | /// List of accesses that need a further dependence check. | ||||
598 | MemAccessInfoList CheckDeps; | ||||
599 | |||||
600 | /// Set of pointers that are read only. | ||||
601 | SmallPtrSet<Value*, 16> ReadOnlyPtr; | ||||
602 | |||||
603 | /// An alias set tracker to partition the access set by underlying object and | ||||
604 | //intrinsic property (such as TBAA metadata). | ||||
605 | AliasSetTracker AST; | ||||
606 | |||||
607 | LoopInfo *LI; | ||||
608 | |||||
609 | /// Sets of potentially dependent accesses - members of one set share an | ||||
610 | /// underlying pointer. The set "CheckDeps" identfies which sets really need a | ||||
611 | /// dependence check. | ||||
612 | MemoryDepChecker::DepCandidates &DepCands; | ||||
613 | |||||
614 | /// Initial processing of memory accesses determined that we may need | ||||
615 | /// to add memchecks. Perform the analysis to determine the necessary checks. | ||||
616 | /// | ||||
617 | /// Note that, this is different from isDependencyCheckNeeded. When we retry | ||||
618 | /// memcheck analysis without dependency checking | ||||
619 | /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is | ||||
620 | /// cleared while this remains set if we have potentially dependent accesses. | ||||
621 | bool IsRTCheckAnalysisNeeded = false; | ||||
622 | |||||
623 | /// The SCEV predicate containing all the SCEV-related assumptions. | ||||
624 | PredicatedScalarEvolution &PSE; | ||||
625 | }; | ||||
626 | |||||
627 | } // end anonymous namespace | ||||
628 | |||||
629 | /// Check whether a pointer can participate in a runtime bounds check. | ||||
630 | /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr | ||||
631 | /// by adding run-time checks (overflow checks) if necessary. | ||||
632 | static bool hasComputableBounds(PredicatedScalarEvolution &PSE, | ||||
633 | const ValueToValueMap &Strides, Value *Ptr, | ||||
634 | Loop *L, bool Assume) { | ||||
635 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); | ||||
636 | |||||
637 | // The bounds for loop-invariant pointer is trivial. | ||||
638 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) | ||||
639 | return true; | ||||
640 | |||||
641 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); | ||||
642 | |||||
643 | if (!AR && Assume) | ||||
644 | AR = PSE.getAsAddRec(Ptr); | ||||
645 | |||||
646 | if (!AR) | ||||
647 | return false; | ||||
648 | |||||
649 | return AR->isAffine(); | ||||
650 | } | ||||
651 | |||||
652 | /// Check whether a pointer address cannot wrap. | ||||
653 | static bool isNoWrap(PredicatedScalarEvolution &PSE, | ||||
654 | const ValueToValueMap &Strides, Value *Ptr, Type *AccessTy, | ||||
655 | Loop *L) { | ||||
656 | const SCEV *PtrScev = PSE.getSCEV(Ptr); | ||||
657 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) | ||||
658 | return true; | ||||
659 | |||||
660 | int64_t Stride = getPtrStride(PSE, AccessTy, Ptr, L, Strides); | ||||
661 | if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) | ||||
662 | return true; | ||||
663 | |||||
664 | return false; | ||||
665 | } | ||||
666 | |||||
667 | static void visitPointers(Value *StartPtr, const Loop &InnermostLoop, | ||||
668 | function_ref<void(Value *)> AddPointer) { | ||||
669 | SmallPtrSet<Value *, 8> Visited; | ||||
670 | SmallVector<Value *> WorkList; | ||||
671 | WorkList.push_back(StartPtr); | ||||
672 | |||||
673 | while (!WorkList.empty()) { | ||||
674 | Value *Ptr = WorkList.pop_back_val(); | ||||
675 | if (!Visited.insert(Ptr).second) | ||||
676 | continue; | ||||
677 | auto *PN = dyn_cast<PHINode>(Ptr); | ||||
678 | // SCEV does not look through non-header PHIs inside the loop. Such phis | ||||
679 | // can be analyzed by adding separate accesses for each incoming pointer | ||||
680 | // value. | ||||
681 | if (PN && InnermostLoop.contains(PN->getParent()) && | ||||
682 | PN->getParent() != InnermostLoop.getHeader()) { | ||||
683 | for (const Use &Inc : PN->incoming_values()) | ||||
684 | WorkList.push_back(Inc); | ||||
685 | } else | ||||
686 | AddPointer(Ptr); | ||||
687 | } | ||||
688 | } | ||||
689 | |||||
690 | bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck, | ||||
691 | MemAccessInfo Access, Type *AccessTy, | ||||
692 | const ValueToValueMap &StridesMap, | ||||
693 | DenseMap<Value *, unsigned> &DepSetId, | ||||
694 | Loop *TheLoop, unsigned &RunningDepId, | ||||
695 | unsigned ASId, bool ShouldCheckWrap, | ||||
696 | bool Assume) { | ||||
697 | Value *Ptr = Access.getPointer(); | ||||
698 | |||||
699 | if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume)) | ||||
700 | return false; | ||||
701 | |||||
702 | // When we run after a failing dependency check we have to make sure | ||||
703 | // we don't have wrapping pointers. | ||||
704 | if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, AccessTy, TheLoop)) { | ||||
705 | auto *Expr = PSE.getSCEV(Ptr); | ||||
706 | if (!Assume || !isa<SCEVAddRecExpr>(Expr)) | ||||
707 | return false; | ||||
708 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
709 | } | ||||
710 | |||||
711 | // The id of the dependence set. | ||||
712 | unsigned DepId; | ||||
713 | |||||
714 | if (isDependencyCheckNeeded()) { | ||||
715 | Value *Leader = DepCands.getLeaderValue(Access).getPointer(); | ||||
716 | unsigned &LeaderId = DepSetId[Leader]; | ||||
717 | if (!LeaderId) | ||||
718 | LeaderId = RunningDepId++; | ||||
719 | DepId = LeaderId; | ||||
720 | } else | ||||
721 | // Each access has its own dependence set. | ||||
722 | DepId = RunningDepId++; | ||||
723 | |||||
724 | bool IsWrite = Access.getInt(); | ||||
725 | RtCheck.insert(TheLoop, Ptr, AccessTy, IsWrite, DepId, ASId, StridesMap, PSE); | ||||
726 | LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n'; } } while (false); | ||||
727 | |||||
728 | return true; | ||||
729 | } | ||||
730 | |||||
731 | bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck, | ||||
732 | ScalarEvolution *SE, Loop *TheLoop, | ||||
733 | const ValueToValueMap &StridesMap, | ||||
734 | Value *&UncomputablePtr, bool ShouldCheckWrap) { | ||||
735 | // Find pointers with computable bounds. We are going to use this information | ||||
736 | // to place a runtime bound check. | ||||
737 | bool CanDoRT = true; | ||||
738 | |||||
739 | bool MayNeedRTCheck = false; | ||||
740 | if (!IsRTCheckAnalysisNeeded) return true; | ||||
741 | |||||
742 | bool IsDepCheckNeeded = isDependencyCheckNeeded(); | ||||
743 | |||||
744 | // We assign a consecutive id to access from different alias sets. | ||||
745 | // Accesses between different groups doesn't need to be checked. | ||||
746 | unsigned ASId = 0; | ||||
747 | for (auto &AS : AST) { | ||||
748 | int NumReadPtrChecks = 0; | ||||
749 | int NumWritePtrChecks = 0; | ||||
750 | bool CanDoAliasSetRT = true; | ||||
751 | ++ASId; | ||||
752 | |||||
753 | // We assign consecutive id to access from different dependence sets. | ||||
754 | // Accesses within the same set don't need a runtime check. | ||||
755 | unsigned RunningDepId = 1; | ||||
756 | DenseMap<Value *, unsigned> DepSetId; | ||||
757 | |||||
758 | SmallVector<MemAccessInfo, 4> Retries; | ||||
759 | |||||
760 | // First, count how many write and read accesses are in the alias set. Also | ||||
761 | // collect MemAccessInfos for later. | ||||
762 | SmallVector<MemAccessInfo, 4> AccessInfos; | ||||
763 | for (const auto &A : AS) { | ||||
764 | Value *Ptr = A.getValue(); | ||||
765 | bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true)); | ||||
766 | |||||
767 | if (IsWrite) | ||||
768 | ++NumWritePtrChecks; | ||||
769 | else | ||||
770 | ++NumReadPtrChecks; | ||||
771 | AccessInfos.emplace_back(Ptr, IsWrite); | ||||
772 | } | ||||
773 | |||||
774 | // We do not need runtime checks for this alias set, if there are no writes | ||||
775 | // or a single write and no reads. | ||||
776 | if (NumWritePtrChecks == 0 || | ||||
777 | (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) { | ||||
778 | assert((AS.size() <= 1 ||(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
779 | all_of(AS,(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
780 | [this](auto AC) {(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
781 | MemAccessInfo AccessWrite(AC.getValue(), true);(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
782 | return DepCands.findValue(AccessWrite) == DepCands.end();(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
783 | })) &&(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
784 | "Can only skip updating CanDoRT below, if all entries in AS "(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)) | ||||
785 | "are reads or there is at most 1 entry")(static_cast <bool> ((AS.size() <= 1 || all_of(AS, [ this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true ); return DepCands.findValue(AccessWrite) == DepCands.end(); } )) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? void (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 785, __extension__ __PRETTY_FUNCTION__)); | ||||
786 | continue; | ||||
787 | } | ||||
788 | |||||
789 | for (auto &Access : AccessInfos) { | ||||
790 | for (auto &AccessTy : Accesses[Access]) { | ||||
791 | if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap, | ||||
792 | DepSetId, TheLoop, RunningDepId, ASId, | ||||
793 | ShouldCheckWrap, false)) { | ||||
794 | LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't find bounds for ptr:" << *Access.getPointer() << '\n'; } } while (false ) | ||||
795 | << *Access.getPointer() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't find bounds for ptr:" << *Access.getPointer() << '\n'; } } while (false ); | ||||
796 | Retries.push_back(Access); | ||||
797 | CanDoAliasSetRT = false; | ||||
798 | } | ||||
799 | } | ||||
800 | } | ||||
801 | |||||
802 | // Note that this function computes CanDoRT and MayNeedRTCheck | ||||
803 | // independently. For example CanDoRT=false, MayNeedRTCheck=false means that | ||||
804 | // we have a pointer for which we couldn't find the bounds but we don't | ||||
805 | // actually need to emit any checks so it does not matter. | ||||
806 | // | ||||
807 | // We need runtime checks for this alias set, if there are at least 2 | ||||
808 | // dependence sets (in which case RunningDepId > 2) or if we need to re-try | ||||
809 | // any bound checks (because in that case the number of dependence sets is | ||||
810 | // incomplete). | ||||
811 | bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty(); | ||||
812 | |||||
813 | // We need to perform run-time alias checks, but some pointers had bounds | ||||
814 | // that couldn't be checked. | ||||
815 | if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) { | ||||
816 | // Reset the CanDoSetRt flag and retry all accesses that have failed. | ||||
817 | // We know that we need these checks, so we can now be more aggressive | ||||
818 | // and add further checks if required (overflow checks). | ||||
819 | CanDoAliasSetRT = true; | ||||
820 | for (auto Access : Retries) { | ||||
821 | for (auto &AccessTy : Accesses[Access]) { | ||||
822 | if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap, | ||||
823 | DepSetId, TheLoop, RunningDepId, ASId, | ||||
824 | ShouldCheckWrap, /*Assume=*/true)) { | ||||
825 | CanDoAliasSetRT = false; | ||||
826 | UncomputablePtr = Access.getPointer(); | ||||
827 | break; | ||||
828 | } | ||||
829 | } | ||||
830 | } | ||||
831 | } | ||||
832 | |||||
833 | CanDoRT &= CanDoAliasSetRT; | ||||
834 | MayNeedRTCheck |= NeedsAliasSetRTCheck; | ||||
835 | ++ASId; | ||||
836 | } | ||||
837 | |||||
838 | // If the pointers that we would use for the bounds comparison have different | ||||
839 | // address spaces, assume the values aren't directly comparable, so we can't | ||||
840 | // use them for the runtime check. We also have to assume they could | ||||
841 | // overlap. In the future there should be metadata for whether address spaces | ||||
842 | // are disjoint. | ||||
843 | unsigned NumPointers = RtCheck.Pointers.size(); | ||||
844 | for (unsigned i = 0; i < NumPointers; ++i) { | ||||
845 | for (unsigned j = i + 1; j < NumPointers; ++j) { | ||||
846 | // Only need to check pointers between two different dependency sets. | ||||
847 | if (RtCheck.Pointers[i].DependencySetId == | ||||
848 | RtCheck.Pointers[j].DependencySetId) | ||||
849 | continue; | ||||
850 | // Only need to check pointers in the same alias set. | ||||
851 | if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId) | ||||
852 | continue; | ||||
853 | |||||
854 | Value *PtrI = RtCheck.Pointers[i].PointerValue; | ||||
855 | Value *PtrJ = RtCheck.Pointers[j].PointerValue; | ||||
856 | |||||
857 | unsigned ASi = PtrI->getType()->getPointerAddressSpace(); | ||||
858 | unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); | ||||
859 | if (ASi != ASj) { | ||||
860 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false) | ||||
861 | dbgs() << "LAA: Runtime check would require comparison between"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false) | ||||
862 | " different address spaces\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false); | ||||
863 | return false; | ||||
864 | } | ||||
865 | } | ||||
866 | } | ||||
867 | |||||
868 | if (MayNeedRTCheck && CanDoRT) | ||||
869 | RtCheck.generateChecks(DepCands, IsDepCheckNeeded); | ||||
870 | |||||
871 | LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() << " pointer comparisons.\n" ; } } while (false) | ||||
872 | << " pointer comparisons.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() << " pointer comparisons.\n" ; } } while (false); | ||||
873 | |||||
874 | // If we can do run-time checks, but there are no checks, no runtime checks | ||||
875 | // are needed. This can happen when all pointers point to the same underlying | ||||
876 | // object for example. | ||||
877 | RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck; | ||||
878 | |||||
879 | bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT; | ||||
880 | if (!CanDoRTIfNeeded) | ||||
881 | RtCheck.reset(); | ||||
882 | return CanDoRTIfNeeded; | ||||
883 | } | ||||
884 | |||||
885 | void AccessAnalysis::processMemAccesses() { | ||||
886 | // We process the set twice: first we process read-write pointers, last we | ||||
887 | // process read-only pointers. This allows us to skip dependence tests for | ||||
888 | // read-only pointers. | ||||
889 | |||||
890 | LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Processing memory accesses...\n" ; } } while (false); | ||||
891 | LLVM_DEBUG(dbgs() << " AST: "; AST.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " AST: "; AST.dump(); } } while (false); | ||||
892 | LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n"; } } while (false); | ||||
893 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
894 | for (auto A : Accesses)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
895 | dbgs() << "\t" << *A.first.getPointer() << " ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
896 | << (A.first.getInt()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
897 | ? "write"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
898 | : (ReadOnlyPtr.count(A.first.getPointer()) ? "read-only"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
899 | : "read"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
900 | << ")\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ) | ||||
901 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.first.getPointer() << " (" << ( A.first.getInt() ? "write" : (ReadOnlyPtr.count(A.first.getPointer ()) ? "read-only" : "read")) << ")\n"; }; } } while (false ); | ||||
902 | |||||
903 | // The AliasSetTracker has nicely partitioned our pointers by metadata | ||||
904 | // compatibility and potential for underlying-object overlap. As a result, we | ||||
905 | // only need to check for potential pointer dependencies within each alias | ||||
906 | // set. | ||||
907 | for (const auto &AS : AST) { | ||||
908 | // Note that both the alias-set tracker and the alias sets themselves used | ||||
909 | // linked lists internally and so the iteration order here is deterministic | ||||
910 | // (matching the original instruction order within each set). | ||||
911 | |||||
912 | bool SetHasWrite = false; | ||||
913 | |||||
914 | // Map of pointers to last access encountered. | ||||
915 | typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap; | ||||
916 | UnderlyingObjToAccessMap ObjToLastAccess; | ||||
917 | |||||
918 | // Set of access to check after all writes have been processed. | ||||
919 | PtrAccessMap DeferredAccesses; | ||||
920 | |||||
921 | // Iterate over each alias set twice, once to process read/write pointers, | ||||
922 | // and then to process read-only pointers. | ||||
923 | for (int SetIteration = 0; SetIteration < 2; ++SetIteration) { | ||||
924 | bool UseDeferred = SetIteration > 0; | ||||
925 | PtrAccessMap &S = UseDeferred
| ||||
926 | |||||
927 | for (const auto &AV : AS) { | ||||
928 | Value *Ptr = AV.getValue(); | ||||
929 | |||||
930 | // For a single memory access in AliasSetTracker, Accesses may contain | ||||
931 | // both read and write, and they both need to be handled for CheckDeps. | ||||
932 | for (const auto &AC : S) { | ||||
933 | if (AC.first.getPointer() != Ptr) | ||||
934 | continue; | ||||
935 | |||||
936 | bool IsWrite = AC.first.getInt(); | ||||
937 | |||||
938 | // If we're using the deferred access set, then it contains only | ||||
939 | // reads. | ||||
940 | bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; | ||||
941 | if (UseDeferred
| ||||
942 | continue; | ||||
943 | // Otherwise, the pointer must be in the PtrAccessSet, either as a | ||||
944 | // read or a write. | ||||
945 | assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||(static_cast <bool> (((IsReadOnlyPtr && UseDeferred ) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?") ? void (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 947, __extension__ __PRETTY_FUNCTION__)) | ||||
946 | S.count(MemAccessInfo(Ptr, false))) &&(static_cast <bool> (((IsReadOnlyPtr && UseDeferred ) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?") ? void (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 947, __extension__ __PRETTY_FUNCTION__)) | ||||
947 | "Alias-set pointer not in the access set?")(static_cast <bool> (((IsReadOnlyPtr && UseDeferred ) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?") ? void (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 947, __extension__ __PRETTY_FUNCTION__)); | ||||
948 | |||||
949 | MemAccessInfo Access(Ptr, IsWrite); | ||||
950 | DepCands.insert(Access); | ||||
951 | |||||
952 | // Memorize read-only pointers for later processing and skip them in | ||||
953 | // the first round (they need to be checked after we have seen all | ||||
954 | // write pointers). Note: we also mark pointer that are not | ||||
955 | // consecutive as "read-only" pointers (so that we check | ||||
956 | // "a[b[i]] +="). Hence, we need the second check for "!IsWrite". | ||||
957 | if (!UseDeferred && IsReadOnlyPtr) { | ||||
958 | // We only use the pointer keys, the types vector values don't | ||||
959 | // matter. | ||||
960 | DeferredAccesses.insert({Access, {}}); | ||||
961 | continue; | ||||
962 | } | ||||
963 | |||||
964 | // If this is a write - check other reads and writes for conflicts. If | ||||
965 | // this is a read only check other writes for conflicts (but only if | ||||
966 | // there is no other write to the ptr - this is an optimization to | ||||
967 | // catch "a[i] = a[i] + " without having to do a dependence check). | ||||
968 | if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) { | ||||
969 | CheckDeps.push_back(Access); | ||||
970 | IsRTCheckAnalysisNeeded = true; | ||||
971 | } | ||||
972 | |||||
973 | if (IsWrite) | ||||
974 | SetHasWrite = true; | ||||
975 | |||||
976 | // Create sets of pointers connected by a shared alias set and | ||||
977 | // underlying object. | ||||
978 | typedef SmallVector<const Value *, 16> ValueVector; | ||||
979 | ValueVector TempObjects; | ||||
980 | |||||
981 | getUnderlyingObjects(Ptr, TempObjects, LI); | ||||
982 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Underlying objects for pointer " << *Ptr << "\n"; } } while (false) | ||||
983 | << "Underlying objects for pointer " << *Ptr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Underlying objects for pointer " << *Ptr << "\n"; } } while (false); | ||||
984 | for (const Value *UnderlyingObj : TempObjects) { | ||||
985 | // nullptr never alias, don't join sets for pointer that have "null" | ||||
986 | // in their UnderlyingObjects list. | ||||
987 | if (isa<ConstantPointerNull>(UnderlyingObj) && | ||||
988 | !NullPointerIsDefined( | ||||
989 | TheLoop->getHeader()->getParent(), | ||||
990 | UnderlyingObj->getType()->getPointerAddressSpace())) | ||||
991 | continue; | ||||
992 | |||||
993 | UnderlyingObjToAccessMap::iterator Prev = | ||||
994 | ObjToLastAccess.find(UnderlyingObj); | ||||
995 | if (Prev != ObjToLastAccess.end()) | ||||
996 | DepCands.unionSets(Access, Prev->second); | ||||
997 | |||||
998 | ObjToLastAccess[UnderlyingObj] = Access; | ||||
999 | LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " " << *UnderlyingObj << "\n"; } } while (false); | ||||
1000 | } | ||||
1001 | } | ||||
1002 | } | ||||
1003 | } | ||||
1004 | } | ||||
1005 | } | ||||
1006 | |||||
1007 | static bool isInBoundsGep(Value *Ptr) { | ||||
1008 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) | ||||
1009 | return GEP->isInBounds(); | ||||
1010 | return false; | ||||
1011 | } | ||||
1012 | |||||
1013 | /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping, | ||||
1014 | /// i.e. monotonically increasing/decreasing. | ||||
1015 | static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, | ||||
1016 | PredicatedScalarEvolution &PSE, const Loop *L) { | ||||
1017 | // FIXME: This should probably only return true for NUW. | ||||
1018 | if (AR->getNoWrapFlags(SCEV::NoWrapMask)) | ||||
1019 | return true; | ||||
1020 | |||||
1021 | // Scalar evolution does not propagate the non-wrapping flags to values that | ||||
1022 | // are derived from a non-wrapping induction variable because non-wrapping | ||||
1023 | // could be flow-sensitive. | ||||
1024 | // | ||||
1025 | // Look through the potentially overflowing instruction to try to prove | ||||
1026 | // non-wrapping for the *specific* value of Ptr. | ||||
1027 | |||||
1028 | // The arithmetic implied by an inbounds GEP can't overflow. | ||||
1029 | auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); | ||||
1030 | if (!GEP || !GEP->isInBounds()) | ||||
1031 | return false; | ||||
1032 | |||||
1033 | // Make sure there is only one non-const index and analyze that. | ||||
1034 | Value *NonConstIndex = nullptr; | ||||
1035 | for (Value *Index : GEP->indices()) | ||||
1036 | if (!isa<ConstantInt>(Index)) { | ||||
1037 | if (NonConstIndex) | ||||
1038 | return false; | ||||
1039 | NonConstIndex = Index; | ||||
1040 | } | ||||
1041 | if (!NonConstIndex) | ||||
1042 | // The recurrence is on the pointer, ignore for now. | ||||
1043 | return false; | ||||
1044 | |||||
1045 | // The index in GEP is signed. It is non-wrapping if it's derived from a NSW | ||||
1046 | // AddRec using a NSW operation. | ||||
1047 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex)) | ||||
1048 | if (OBO->hasNoSignedWrap() && | ||||
1049 | // Assume constant for other the operand so that the AddRec can be | ||||
1050 | // easily found. | ||||
1051 | isa<ConstantInt>(OBO->getOperand(1))) { | ||||
1052 | auto *OpScev = PSE.getSCEV(OBO->getOperand(0)); | ||||
1053 | |||||
1054 | if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev)) | ||||
1055 | return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW); | ||||
1056 | } | ||||
1057 | |||||
1058 | return false; | ||||
1059 | } | ||||
1060 | |||||
1061 | /// Check whether the access through \p Ptr has a constant stride. | ||||
1062 | int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy, | ||||
1063 | Value *Ptr, const Loop *Lp, | ||||
1064 | const ValueToValueMap &StridesMap, bool Assume, | ||||
1065 | bool ShouldCheckWrap) { | ||||
1066 | Type *Ty = Ptr->getType(); | ||||
1067 | assert(Ty->isPointerTy() && "Unexpected non-ptr")(static_cast <bool> (Ty->isPointerTy() && "Unexpected non-ptr" ) ? void (0) : __assert_fail ("Ty->isPointerTy() && \"Unexpected non-ptr\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1067, __extension__ __PRETTY_FUNCTION__)); | ||||
1068 | |||||
1069 | if (isa<ScalableVectorType>(AccessTy)) { | ||||
1070 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Scalable object: " << *AccessTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Scalable object: " << *AccessTy << "\n"; } } while (false) | ||||
1071 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Scalable object: " << *AccessTy << "\n"; } } while (false); | ||||
1072 | return 0; | ||||
1073 | } | ||||
1074 | |||||
1075 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr); | ||||
1076 | |||||
1077 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); | ||||
1078 | if (Assume && !AR) | ||||
1079 | AR = PSE.getAsAddRec(Ptr); | ||||
1080 | |||||
1081 | if (!AR) { | ||||
1082 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr << " SCEV: " << *PtrScev << "\n" ; } } while (false) | ||||
1083 | << " SCEV: " << *PtrScev << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr << " SCEV: " << *PtrScev << "\n" ; } } while (false); | ||||
1084 | return 0; | ||||
1085 | } | ||||
1086 | |||||
1087 | // The access function must stride over the innermost loop. | ||||
1088 | if (Lp != AR->getLoop()) { | ||||
1089 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not striding over innermost loop " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) | ||||
1090 | << *Ptr << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not striding over innermost loop " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); | ||||
1091 | return 0; | ||||
1092 | } | ||||
1093 | |||||
1094 | // The address calculation must not wrap. Otherwise, a dependence could be | ||||
1095 | // inverted. | ||||
1096 | // An inbounds getelementptr that is a AddRec with a unit stride | ||||
1097 | // cannot wrap per definition. The unit stride requirement is checked later. | ||||
1098 | // An getelementptr without an inbounds attribute and unit stride would have | ||||
1099 | // to access the pointer value "0" which is undefined behavior in address | ||||
1100 | // space 0, therefore we can also vectorize this case. | ||||
1101 | unsigned AddrSpace = Ty->getPointerAddressSpace(); | ||||
1102 | bool IsInBoundsGEP = isInBoundsGep(Ptr); | ||||
1103 | bool IsNoWrapAddRec = !ShouldCheckWrap || | ||||
1104 | PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) || | ||||
1105 | isNoWrapAddRec(Ptr, AR, PSE, Lp); | ||||
1106 | if (!IsNoWrapAddRec && !IsInBoundsGEP && | ||||
1107 | NullPointerIsDefined(Lp->getHeader()->getParent(), AddrSpace)) { | ||||
1108 | if (Assume) { | ||||
1109 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
1110 | IsNoWrapAddRec = true; | ||||
1111 | LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1112 | << "LAA: Pointer: " << *Ptr << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1113 | << "LAA: SCEV: " << *AR << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1114 | << "LAA: Added an overflow assumption\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false); | ||||
1115 | } else { | ||||
1116 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) | ||||
1117 | dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) | ||||
1118 | << *Ptr << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); | ||||
1119 | return 0; | ||||
1120 | } | ||||
1121 | } | ||||
1122 | |||||
1123 | // Check the step is constant. | ||||
1124 | const SCEV *Step = AR->getStepRecurrence(*PSE.getSE()); | ||||
1125 | |||||
1126 | // Calculate the pointer stride and check if it is constant. | ||||
1127 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); | ||||
1128 | if (!C) { | ||||
1129 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) | ||||
1130 | << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); | ||||
1131 | return 0; | ||||
1132 | } | ||||
1133 | |||||
1134 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); | ||||
1135 | TypeSize AllocSize = DL.getTypeAllocSize(AccessTy); | ||||
1136 | int64_t Size = AllocSize.getFixedSize(); | ||||
1137 | const APInt &APStepVal = C->getAPInt(); | ||||
1138 | |||||
1139 | // Huge step value - give up. | ||||
1140 | if (APStepVal.getBitWidth() > 64) | ||||
1141 | return 0; | ||||
1142 | |||||
1143 | int64_t StepVal = APStepVal.getSExtValue(); | ||||
1144 | |||||
1145 | // Strided access. | ||||
1146 | int64_t Stride = StepVal / Size; | ||||
1147 | int64_t Rem = StepVal % Size; | ||||
1148 | if (Rem) | ||||
1149 | return 0; | ||||
1150 | |||||
1151 | // If the SCEV could wrap but we have an inbounds gep with a unit stride we | ||||
1152 | // know we can't "wrap around the address space". In case of address space | ||||
1153 | // zero we know that this won't happen without triggering undefined behavior. | ||||
1154 | if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 && | ||||
1155 | (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(), | ||||
1156 | AddrSpace))) { | ||||
1157 | if (Assume) { | ||||
1158 | // We can avoid this case by adding a run-time check. | ||||
1159 | LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1160 | << "inbounds or in address space 0 may wrap:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1161 | << "LAA: Pointer: " << *Ptr << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1162 | << "LAA: SCEV: " << *AR << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) | ||||
1163 | << "LAA: Added an overflow assumption\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false); | ||||
1164 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
1165 | } else | ||||
1166 | return 0; | ||||
1167 | } | ||||
1168 | |||||
1169 | return Stride; | ||||
1170 | } | ||||
1171 | |||||
1172 | Optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB, | ||||
1173 | Value *PtrB, const DataLayout &DL, | ||||
1174 | ScalarEvolution &SE, bool StrictCheck, | ||||
1175 | bool CheckType) { | ||||
1176 | assert(PtrA && PtrB && "Expected non-nullptr pointers.")(static_cast <bool> (PtrA && PtrB && "Expected non-nullptr pointers." ) ? void (0) : __assert_fail ("PtrA && PtrB && \"Expected non-nullptr pointers.\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1176, __extension__ __PRETTY_FUNCTION__)); | ||||
1177 | assert(cast<PointerType>(PtrA->getType())(static_cast <bool> (cast<PointerType>(PtrA->getType ()) ->isOpaqueOrPointeeTypeMatches(ElemTyA) && "Wrong PtrA type" ) ? void (0) : __assert_fail ("cast<PointerType>(PtrA->getType()) ->isOpaqueOrPointeeTypeMatches(ElemTyA) && \"Wrong PtrA type\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1178, __extension__ __PRETTY_FUNCTION__)) | ||||
1178 | ->isOpaqueOrPointeeTypeMatches(ElemTyA) && "Wrong PtrA type")(static_cast <bool> (cast<PointerType>(PtrA->getType ()) ->isOpaqueOrPointeeTypeMatches(ElemTyA) && "Wrong PtrA type" ) ? void (0) : __assert_fail ("cast<PointerType>(PtrA->getType()) ->isOpaqueOrPointeeTypeMatches(ElemTyA) && \"Wrong PtrA type\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1178, __extension__ __PRETTY_FUNCTION__)); | ||||
1179 | assert(cast<PointerType>(PtrB->getType())(static_cast <bool> (cast<PointerType>(PtrB->getType ()) ->isOpaqueOrPointeeTypeMatches(ElemTyB) && "Wrong PtrB type" ) ? void (0) : __assert_fail ("cast<PointerType>(PtrB->getType()) ->isOpaqueOrPointeeTypeMatches(ElemTyB) && \"Wrong PtrB type\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1180, __extension__ __PRETTY_FUNCTION__)) | ||||
1180 | ->isOpaqueOrPointeeTypeMatches(ElemTyB) && "Wrong PtrB type")(static_cast <bool> (cast<PointerType>(PtrB->getType ()) ->isOpaqueOrPointeeTypeMatches(ElemTyB) && "Wrong PtrB type" ) ? void (0) : __assert_fail ("cast<PointerType>(PtrB->getType()) ->isOpaqueOrPointeeTypeMatches(ElemTyB) && \"Wrong PtrB type\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1180, __extension__ __PRETTY_FUNCTION__)); | ||||
1181 | |||||
1182 | // Make sure that A and B are different pointers. | ||||
1183 | if (PtrA == PtrB) | ||||
1184 | return 0; | ||||
1185 | |||||
1186 | // Make sure that the element types are the same if required. | ||||
1187 | if (CheckType && ElemTyA != ElemTyB) | ||||
1188 | return None; | ||||
1189 | |||||
1190 | unsigned ASA = PtrA->getType()->getPointerAddressSpace(); | ||||
1191 | unsigned ASB = PtrB->getType()->getPointerAddressSpace(); | ||||
1192 | |||||
1193 | // Check that the address spaces match. | ||||
1194 | if (ASA != ASB) | ||||
1195 | return None; | ||||
1196 | unsigned IdxWidth = DL.getIndexSizeInBits(ASA); | ||||
1197 | |||||
1198 | APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0); | ||||
1199 | Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); | ||||
1200 | Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); | ||||
1201 | |||||
1202 | int Val; | ||||
1203 | if (PtrA1 == PtrB1) { | ||||
1204 | // Retrieve the address space again as pointer stripping now tracks through | ||||
1205 | // `addrspacecast`. | ||||
1206 | ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace(); | ||||
1207 | ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace(); | ||||
1208 | // Check that the address spaces match and that the pointers are valid. | ||||
1209 | if (ASA != ASB) | ||||
1210 | return None; | ||||
1211 | |||||
1212 | IdxWidth = DL.getIndexSizeInBits(ASA); | ||||
1213 | OffsetA = OffsetA.sextOrTrunc(IdxWidth); | ||||
1214 | OffsetB = OffsetB.sextOrTrunc(IdxWidth); | ||||
1215 | |||||
1216 | OffsetB -= OffsetA; | ||||
1217 | Val = OffsetB.getSExtValue(); | ||||
1218 | } else { | ||||
1219 | // Otherwise compute the distance with SCEV between the base pointers. | ||||
1220 | const SCEV *PtrSCEVA = SE.getSCEV(PtrA); | ||||
1221 | const SCEV *PtrSCEVB = SE.getSCEV(PtrB); | ||||
1222 | const auto *Diff = | ||||
1223 | dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA)); | ||||
1224 | if (!Diff) | ||||
1225 | return None; | ||||
1226 | Val = Diff->getAPInt().getSExtValue(); | ||||
1227 | } | ||||
1228 | int Size = DL.getTypeStoreSize(ElemTyA); | ||||
1229 | int Dist = Val / Size; | ||||
1230 | |||||
1231 | // Ensure that the calculated distance matches the type-based one after all | ||||
1232 | // the bitcasts removal in the provided pointers. | ||||
1233 | if (!StrictCheck || Dist * Size == Val) | ||||
1234 | return Dist; | ||||
1235 | return None; | ||||
1236 | } | ||||
1237 | |||||
1238 | bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, | ||||
1239 | const DataLayout &DL, ScalarEvolution &SE, | ||||
1240 | SmallVectorImpl<unsigned> &SortedIndices) { | ||||
1241 | assert(llvm::all_of((static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1243, __extension__ __PRETTY_FUNCTION__)) | ||||
1242 | VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&(static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1243, __extension__ __PRETTY_FUNCTION__)) | ||||
1243 | "Expected list of pointer operands.")(static_cast <bool> (llvm::all_of( VL, [](const Value * V) { return V->getType()->isPointerTy(); }) && "Expected list of pointer operands." ) ? void (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1243, __extension__ __PRETTY_FUNCTION__)); | ||||
1244 | // Walk over the pointers, and map each of them to an offset relative to | ||||
1245 | // first pointer in the array. | ||||
1246 | Value *Ptr0 = VL[0]; | ||||
1247 | |||||
1248 | using DistOrdPair = std::pair<int64_t, int>; | ||||
1249 | auto Compare = [](const DistOrdPair &L, const DistOrdPair &R) { | ||||
1250 | return L.first < R.first; | ||||
1251 | }; | ||||
1252 | std::set<DistOrdPair, decltype(Compare)> Offsets(Compare); | ||||
1253 | Offsets.emplace(0, 0); | ||||
1254 | int Cnt = 1; | ||||
1255 | bool IsConsecutive = true; | ||||
1256 | for (auto *Ptr : VL.drop_front()) { | ||||
1257 | Optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE, | ||||
1258 | /*StrictCheck=*/true); | ||||
1259 | if (!Diff) | ||||
1260 | return false; | ||||
1261 | |||||
1262 | // Check if the pointer with the same offset is found. | ||||
1263 | int64_t Offset = *Diff; | ||||
1264 | auto Res = Offsets.emplace(Offset, Cnt); | ||||
1265 | if (!Res.second) | ||||
1266 | return false; | ||||
1267 | // Consecutive order if the inserted element is the last one. | ||||
1268 | IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end(); | ||||
1269 | ++Cnt; | ||||
1270 | } | ||||
1271 | SortedIndices.clear(); | ||||
1272 | if (!IsConsecutive) { | ||||
1273 | // Fill SortedIndices array only if it is non-consecutive. | ||||
1274 | SortedIndices.resize(VL.size()); | ||||
1275 | Cnt = 0; | ||||
1276 | for (const std::pair<int64_t, int> &Pair : Offsets) { | ||||
1277 | SortedIndices[Cnt] = Pair.second; | ||||
1278 | ++Cnt; | ||||
1279 | } | ||||
1280 | } | ||||
1281 | return true; | ||||
1282 | } | ||||
1283 | |||||
1284 | /// Returns true if the memory operations \p A and \p B are consecutive. | ||||
1285 | bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, | ||||
1286 | ScalarEvolution &SE, bool CheckType) { | ||||
1287 | Value *PtrA = getLoadStorePointerOperand(A); | ||||
1288 | Value *PtrB = getLoadStorePointerOperand(B); | ||||
1289 | if (!PtrA || !PtrB) | ||||
1290 | return false; | ||||
1291 | Type *ElemTyA = getLoadStoreType(A); | ||||
1292 | Type *ElemTyB = getLoadStoreType(B); | ||||
1293 | Optional<int> Diff = getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE, | ||||
1294 | /*StrictCheck=*/true, CheckType); | ||||
1295 | return Diff && *Diff == 1; | ||||
1296 | } | ||||
1297 | |||||
1298 | void MemoryDepChecker::addAccess(StoreInst *SI) { | ||||
1299 | visitPointers(SI->getPointerOperand(), *InnermostLoop, | ||||
1300 | [this, SI](Value *Ptr) { | ||||
1301 | Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx); | ||||
1302 | InstMap.push_back(SI); | ||||
1303 | ++AccessIdx; | ||||
1304 | }); | ||||
1305 | } | ||||
1306 | |||||
1307 | void MemoryDepChecker::addAccess(LoadInst *LI) { | ||||
1308 | visitPointers(LI->getPointerOperand(), *InnermostLoop, | ||||
1309 | [this, LI](Value *Ptr) { | ||||
1310 | Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx); | ||||
1311 | InstMap.push_back(LI); | ||||
1312 | ++AccessIdx; | ||||
1313 | }); | ||||
1314 | } | ||||
1315 | |||||
1316 | MemoryDepChecker::VectorizationSafetyStatus | ||||
1317 | MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) { | ||||
1318 | switch (Type) { | ||||
1319 | case NoDep: | ||||
1320 | case Forward: | ||||
1321 | case BackwardVectorizable: | ||||
1322 | return VectorizationSafetyStatus::Safe; | ||||
1323 | |||||
1324 | case Unknown: | ||||
1325 | return VectorizationSafetyStatus::PossiblySafeWithRtChecks; | ||||
1326 | case ForwardButPreventsForwarding: | ||||
1327 | case Backward: | ||||
1328 | case BackwardVectorizableButPreventsForwarding: | ||||
1329 | return VectorizationSafetyStatus::Unsafe; | ||||
1330 | } | ||||
1331 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1331); | ||||
1332 | } | ||||
1333 | |||||
1334 | bool MemoryDepChecker::Dependence::isBackward() const { | ||||
1335 | switch (Type) { | ||||
1336 | case NoDep: | ||||
1337 | case Forward: | ||||
1338 | case ForwardButPreventsForwarding: | ||||
1339 | case Unknown: | ||||
1340 | return false; | ||||
1341 | |||||
1342 | case BackwardVectorizable: | ||||
1343 | case Backward: | ||||
1344 | case BackwardVectorizableButPreventsForwarding: | ||||
1345 | return true; | ||||
1346 | } | ||||
1347 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1347); | ||||
1348 | } | ||||
1349 | |||||
1350 | bool MemoryDepChecker::Dependence::isPossiblyBackward() const { | ||||
1351 | return isBackward() || Type == Unknown; | ||||
1352 | } | ||||
1353 | |||||
1354 | bool MemoryDepChecker::Dependence::isForward() const { | ||||
1355 | switch (Type) { | ||||
1356 | case Forward: | ||||
1357 | case ForwardButPreventsForwarding: | ||||
1358 | return true; | ||||
1359 | |||||
1360 | case NoDep: | ||||
1361 | case Unknown: | ||||
1362 | case BackwardVectorizable: | ||||
1363 | case Backward: | ||||
1364 | case BackwardVectorizableButPreventsForwarding: | ||||
1365 | return false; | ||||
1366 | } | ||||
1367 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1367); | ||||
1368 | } | ||||
1369 | |||||
1370 | bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance, | ||||
1371 | uint64_t TypeByteSize) { | ||||
1372 | // If loads occur at a distance that is not a multiple of a feasible vector | ||||
1373 | // factor store-load forwarding does not take place. | ||||
1374 | // Positive dependences might cause troubles because vectorizing them might | ||||
1375 | // prevent store-load forwarding making vectorized code run a lot slower. | ||||
1376 | // a[i] = a[i-3] ^ a[i-8]; | ||||
1377 | // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and | ||||
1378 | // hence on your typical architecture store-load forwarding does not take | ||||
1379 | // place. Vectorizing in such cases does not make sense. | ||||
1380 | // Store-load forwarding distance. | ||||
1381 | |||||
1382 | // After this many iterations store-to-load forwarding conflicts should not | ||||
1383 | // cause any slowdowns. | ||||
1384 | const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize; | ||||
1385 | // Maximum vector factor. | ||||
1386 | uint64_t MaxVFWithoutSLForwardIssues = std::min( | ||||
1387 | VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes); | ||||
1388 | |||||
1389 | // Compute the smallest VF at which the store and load would be misaligned. | ||||
1390 | for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues; | ||||
1391 | VF *= 2) { | ||||
1392 | // If the number of vector iteration between the store and the load are | ||||
1393 | // small we could incur conflicts. | ||||
1394 | if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) { | ||||
1395 | MaxVFWithoutSLForwardIssues = (VF >> 1); | ||||
1396 | break; | ||||
1397 | } | ||||
1398 | } | ||||
1399 | |||||
1400 | if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) { | ||||
1401 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false) | ||||
1402 | dbgs() << "LAA: Distance " << Distancedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false) | ||||
1403 | << " that could cause a store-load forwarding conflict\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false); | ||||
1404 | return true; | ||||
1405 | } | ||||
1406 | |||||
1407 | if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && | ||||
1408 | MaxVFWithoutSLForwardIssues != | ||||
1409 | VectorizerParams::MaxVectorWidth * TypeByteSize) | ||||
1410 | MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; | ||||
1411 | return false; | ||||
1412 | } | ||||
1413 | |||||
1414 | void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) { | ||||
1415 | if (Status < S) | ||||
1416 | Status = S; | ||||
1417 | } | ||||
1418 | |||||
1419 | /// Given a non-constant (unknown) dependence-distance \p Dist between two | ||||
1420 | /// memory accesses, that have the same stride whose absolute value is given | ||||
1421 | /// in \p Stride, and that have the same type size \p TypeByteSize, | ||||
1422 | /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is | ||||
1423 | /// possible to prove statically that the dependence distance is larger | ||||
1424 | /// than the range that the accesses will travel through the execution of | ||||
1425 | /// the loop. If so, return true; false otherwise. This is useful for | ||||
1426 | /// example in loops such as the following (PR31098): | ||||
1427 | /// for (i = 0; i < D; ++i) { | ||||
1428 | /// = out[i]; | ||||
1429 | /// out[i+D] = | ||||
1430 | /// } | ||||
1431 | static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, | ||||
1432 | const SCEV &BackedgeTakenCount, | ||||
1433 | const SCEV &Dist, uint64_t Stride, | ||||
1434 | uint64_t TypeByteSize) { | ||||
1435 | |||||
1436 | // If we can prove that | ||||
1437 | // (**) |Dist| > BackedgeTakenCount * Step | ||||
1438 | // where Step is the absolute stride of the memory accesses in bytes, | ||||
1439 | // then there is no dependence. | ||||
1440 | // | ||||
1441 | // Rationale: | ||||
1442 | // We basically want to check if the absolute distance (|Dist/Step|) | ||||
1443 | // is >= the loop iteration count (or > BackedgeTakenCount). | ||||
1444 | // This is equivalent to the Strong SIV Test (Practical Dependence Testing, | ||||
1445 | // Section 4.2.1); Note, that for vectorization it is sufficient to prove | ||||
1446 | // that the dependence distance is >= VF; This is checked elsewhere. | ||||
1447 | // But in some cases we can prune unknown dependence distances early, and | ||||
1448 | // even before selecting the VF, and without a runtime test, by comparing | ||||
1449 | // the distance against the loop iteration count. Since the vectorized code | ||||
1450 | // will be executed only if LoopCount >= VF, proving distance >= LoopCount | ||||
1451 | // also guarantees that distance >= VF. | ||||
1452 | // | ||||
1453 | const uint64_t ByteStride = Stride * TypeByteSize; | ||||
1454 | const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride); | ||||
1455 | const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step); | ||||
1456 | |||||
1457 | const SCEV *CastedDist = &Dist; | ||||
1458 | const SCEV *CastedProduct = Product; | ||||
1459 | uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType()); | ||||
1460 | uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType()); | ||||
1461 | |||||
1462 | // The dependence distance can be positive/negative, so we sign extend Dist; | ||||
1463 | // The multiplication of the absolute stride in bytes and the | ||||
1464 | // backedgeTakenCount is non-negative, so we zero extend Product. | ||||
1465 | if (DistTypeSize > ProductTypeSize) | ||||
1466 | CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType()); | ||||
1467 | else | ||||
1468 | CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType()); | ||||
1469 | |||||
1470 | // Is Dist - (BackedgeTakenCount * Step) > 0 ? | ||||
1471 | // (If so, then we have proven (**) because |Dist| >= Dist) | ||||
1472 | const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct); | ||||
1473 | if (SE.isKnownPositive(Minus)) | ||||
1474 | return true; | ||||
1475 | |||||
1476 | // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ? | ||||
1477 | // (If so, then we have proven (**) because |Dist| >= -1*Dist) | ||||
1478 | const SCEV *NegDist = SE.getNegativeSCEV(CastedDist); | ||||
1479 | Minus = SE.getMinusSCEV(NegDist, CastedProduct); | ||||
1480 | if (SE.isKnownPositive(Minus)) | ||||
1481 | return true; | ||||
1482 | |||||
1483 | return false; | ||||
1484 | } | ||||
1485 | |||||
1486 | /// Check the dependence for two accesses with the same stride \p Stride. | ||||
1487 | /// \p Distance is the positive distance and \p TypeByteSize is type size in | ||||
1488 | /// bytes. | ||||
1489 | /// | ||||
1490 | /// \returns true if they are independent. | ||||
1491 | static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, | ||||
1492 | uint64_t TypeByteSize) { | ||||
1493 | assert(Stride > 1 && "The stride must be greater than 1")(static_cast <bool> (Stride > 1 && "The stride must be greater than 1" ) ? void (0) : __assert_fail ("Stride > 1 && \"The stride must be greater than 1\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1493, __extension__ __PRETTY_FUNCTION__)); | ||||
1494 | assert(TypeByteSize > 0 && "The type size in byte must be non-zero")(static_cast <bool> (TypeByteSize > 0 && "The type size in byte must be non-zero" ) ? void (0) : __assert_fail ("TypeByteSize > 0 && \"The type size in byte must be non-zero\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1494, __extension__ __PRETTY_FUNCTION__)); | ||||
1495 | assert(Distance > 0 && "The distance must be non-zero")(static_cast <bool> (Distance > 0 && "The distance must be non-zero" ) ? void (0) : __assert_fail ("Distance > 0 && \"The distance must be non-zero\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1495, __extension__ __PRETTY_FUNCTION__)); | ||||
1496 | |||||
1497 | // Skip if the distance is not multiple of type byte size. | ||||
1498 | if (Distance % TypeByteSize) | ||||
1499 | return false; | ||||
1500 | |||||
1501 | uint64_t ScaledDist = Distance / TypeByteSize; | ||||
1502 | |||||
1503 | // No dependence if the scaled distance is not multiple of the stride. | ||||
1504 | // E.g. | ||||
1505 | // for (i = 0; i < 1024 ; i += 4) | ||||
1506 | // A[i+2] = A[i] + 1; | ||||
1507 | // | ||||
1508 | // Two accesses in memory (scaled distance is 2, stride is 4): | ||||
1509 | // | A[0] | | | | A[4] | | | | | ||||
1510 | // | | | A[2] | | | | A[6] | | | ||||
1511 | // | ||||
1512 | // E.g. | ||||
1513 | // for (i = 0; i < 1024 ; i += 3) | ||||
1514 | // A[i+4] = A[i] + 1; | ||||
1515 | // | ||||
1516 | // Two accesses in memory (scaled distance is 4, stride is 3): | ||||
1517 | // | A[0] | | | A[3] | | | A[6] | | | | ||||
1518 | // | | | | | A[4] | | | A[7] | | | ||||
1519 | return ScaledDist % Stride; | ||||
1520 | } | ||||
1521 | |||||
1522 | MemoryDepChecker::Dependence::DepType | ||||
1523 | MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, | ||||
1524 | const MemAccessInfo &B, unsigned BIdx, | ||||
1525 | const ValueToValueMap &Strides) { | ||||
1526 | assert (AIdx < BIdx && "Must pass arguments in program order")(static_cast <bool> (AIdx < BIdx && "Must pass arguments in program order" ) ? void (0) : __assert_fail ("AIdx < BIdx && \"Must pass arguments in program order\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1526, __extension__ __PRETTY_FUNCTION__)); | ||||
1527 | |||||
1528 | Value *APtr = A.getPointer(); | ||||
1529 | Value *BPtr = B.getPointer(); | ||||
1530 | bool AIsWrite = A.getInt(); | ||||
1531 | bool BIsWrite = B.getInt(); | ||||
1532 | Type *ATy = getLoadStoreType(InstMap[AIdx]); | ||||
1533 | Type *BTy = getLoadStoreType(InstMap[BIdx]); | ||||
1534 | |||||
1535 | // Two reads are independent. | ||||
1536 | if (!AIsWrite && !BIsWrite) | ||||
1537 | return Dependence::NoDep; | ||||
1538 | |||||
1539 | // We cannot check pointers in different address spaces. | ||||
1540 | if (APtr->getType()->getPointerAddressSpace() != | ||||
1541 | BPtr->getType()->getPointerAddressSpace()) | ||||
1542 | return Dependence::Unknown; | ||||
1543 | |||||
1544 | int64_t StrideAPtr = | ||||
1545 | getPtrStride(PSE, ATy, APtr, InnermostLoop, Strides, true); | ||||
1546 | int64_t StrideBPtr = | ||||
1547 | getPtrStride(PSE, BTy, BPtr, InnermostLoop, Strides, true); | ||||
1548 | |||||
1549 | const SCEV *Src = PSE.getSCEV(APtr); | ||||
1550 | const SCEV *Sink = PSE.getSCEV(BPtr); | ||||
1551 | |||||
1552 | // If the induction step is negative we have to invert source and sink of the | ||||
1553 | // dependence. | ||||
1554 | if (StrideAPtr < 0) { | ||||
1555 | std::swap(APtr, BPtr); | ||||
1556 | std::swap(ATy, BTy); | ||||
1557 | std::swap(Src, Sink); | ||||
1558 | std::swap(AIsWrite, BIsWrite); | ||||
1559 | std::swap(AIdx, BIdx); | ||||
1560 | std::swap(StrideAPtr, StrideBPtr); | ||||
1561 | } | ||||
1562 | |||||
1563 | const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src); | ||||
1564 | |||||
1565 | LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sinkdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink << "(Induction step: " << StrideAPtr << ")\n"; } } while (false) | ||||
1566 | << "(Induction step: " << StrideAPtr << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink << "(Induction step: " << StrideAPtr << ")\n"; } } while (false); | ||||
1567 | LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to " << *InstMap[BIdx] << ": " << *Dist << "\n"; } } while (false) | ||||
1568 | << *InstMap[BIdx] << ": " << *Dist << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to " << *InstMap[BIdx] << ": " << *Dist << "\n"; } } while (false); | ||||
1569 | |||||
1570 | // Need accesses with constant stride. We don't want to vectorize | ||||
1571 | // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in | ||||
1572 | // the address space. | ||||
1573 | if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ | ||||
1574 | LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Pointer access with non-constant stride\n" ; } } while (false); | ||||
1575 | return Dependence::Unknown; | ||||
1576 | } | ||||
1577 | |||||
1578 | auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); | ||||
1579 | uint64_t TypeByteSize = DL.getTypeAllocSize(ATy); | ||||
1580 | bool HasSameSize = | ||||
1581 | DL.getTypeStoreSizeInBits(ATy) == DL.getTypeStoreSizeInBits(BTy); | ||||
1582 | uint64_t Stride = std::abs(StrideAPtr); | ||||
1583 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); | ||||
1584 | if (!C) { | ||||
1585 | if (!isa<SCEVCouldNotCompute>(Dist) && HasSameSize && | ||||
1586 | isSafeDependenceDistance(DL, *(PSE.getSE()), | ||||
1587 | *(PSE.getBackedgeTakenCount()), *Dist, Stride, | ||||
1588 | TypeByteSize)) | ||||
1589 | return Dependence::NoDep; | ||||
1590 | |||||
1591 | LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Dependence because of non-constant distance\n" ; } } while (false); | ||||
1592 | FoundNonConstantDistanceDependence = true; | ||||
1593 | return Dependence::Unknown; | ||||
1594 | } | ||||
1595 | |||||
1596 | const APInt &Val = C->getAPInt(); | ||||
1597 | int64_t Distance = Val.getSExtValue(); | ||||
1598 | |||||
1599 | // Attempt to prove strided accesses independent. | ||||
1600 | if (std::abs(Distance) > 0 && Stride > 1 && HasSameSize && | ||||
1601 | areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) { | ||||
1602 | LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Strided accesses are independent\n" ; } } while (false); | ||||
1603 | return Dependence::NoDep; | ||||
1604 | } | ||||
1605 | |||||
1606 | // Negative distances are not plausible dependencies. | ||||
1607 | if (Val.isNegative()) { | ||||
1608 | bool IsTrueDataDependence = (AIsWrite && !BIsWrite); | ||||
1609 | if (IsTrueDataDependence && EnableForwardingConflictDetection && | ||||
1610 | (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || | ||||
1611 | !HasSameSize)) { | ||||
1612 | LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Forward but may prevent st->ld forwarding\n" ; } } while (false); | ||||
1613 | return Dependence::ForwardButPreventsForwarding; | ||||
1614 | } | ||||
1615 | |||||
1616 | LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Dependence is negative\n" ; } } while (false); | ||||
1617 | return Dependence::Forward; | ||||
1618 | } | ||||
1619 | |||||
1620 | // Write to the same location with the same size. | ||||
1621 | if (Val == 0) { | ||||
1622 | if (HasSameSize) | ||||
1623 | return Dependence::Forward; | ||||
1624 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Zero dependence difference but different type sizes\n" ; } } while (false) | ||||
1625 | dbgs() << "LAA: Zero dependence difference but different type sizes\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Zero dependence difference but different type sizes\n" ; } } while (false); | ||||
1626 | return Dependence::Unknown; | ||||
1627 | } | ||||
1628 | |||||
1629 | assert(Val.isStrictlyPositive() && "Expect a positive value")(static_cast <bool> (Val.isStrictlyPositive() && "Expect a positive value") ? void (0) : __assert_fail ("Val.isStrictlyPositive() && \"Expect a positive value\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1629, __extension__ __PRETTY_FUNCTION__)); | ||||
1630 | |||||
1631 | if (!HasSameSize) { | ||||
1632 | LLVM_DEBUG(dbgs() << "LAA: ReadWrite-Write positive dependency with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: ReadWrite-Write positive dependency with " "different type sizes\n"; } } while (false) | ||||
1633 | "different type sizes\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: ReadWrite-Write positive dependency with " "different type sizes\n"; } } while (false); | ||||
1634 | return Dependence::Unknown; | ||||
1635 | } | ||||
1636 | |||||
1637 | // Bail out early if passed-in parameters make vectorization not feasible. | ||||
1638 | unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ? | ||||
1639 | VectorizerParams::VectorizationFactor : 1); | ||||
1640 | unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ? | ||||
1641 | VectorizerParams::VectorizationInterleave : 1); | ||||
1642 | // The minimum number of iterations for a vectorized/unrolled version. | ||||
1643 | unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U); | ||||
1644 | |||||
1645 | // It's not vectorizable if the distance is smaller than the minimum distance | ||||
1646 | // needed for a vectroized/unrolled version. Vectorizing one iteration in | ||||
1647 | // front needs TypeByteSize * Stride. Vectorizing the last iteration needs | ||||
1648 | // TypeByteSize (No need to plus the last gap distance). | ||||
1649 | // | ||||
1650 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. | ||||
1651 | // foo(int *A) { | ||||
1652 | // int *B = (int *)((char *)A + 14); | ||||
1653 | // for (i = 0 ; i < 1024 ; i += 2) | ||||
1654 | // B[i] = A[i] + 1; | ||||
1655 | // } | ||||
1656 | // | ||||
1657 | // Two accesses in memory (stride is 2): | ||||
1658 | // | A[0] | | A[2] | | A[4] | | A[6] | | | ||||
1659 | // | B[0] | | B[2] | | B[4] | | ||||
1660 | // | ||||
1661 | // Distance needs for vectorizing iterations except the last iteration: | ||||
1662 | // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4. | ||||
1663 | // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4. | ||||
1664 | // | ||||
1665 | // If MinNumIter is 2, it is vectorizable as the minimum distance needed is | ||||
1666 | // 12, which is less than distance. | ||||
1667 | // | ||||
1668 | // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4), | ||||
1669 | // the minimum distance needed is 28, which is greater than distance. It is | ||||
1670 | // not safe to do vectorization. | ||||
1671 | uint64_t MinDistanceNeeded = | ||||
1672 | TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize; | ||||
1673 | if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) { | ||||
1674 | LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because of positive distance " << Distance << '\n'; } } while (false) | ||||
1675 | << Distance << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because of positive distance " << Distance << '\n'; } } while (false); | ||||
1676 | return Dependence::Backward; | ||||
1677 | } | ||||
1678 | |||||
1679 | // Unsafe if the minimum distance needed is greater than max safe distance. | ||||
1680 | if (MinDistanceNeeded > MaxSafeDepDistBytes) { | ||||
1681 | LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because it needs at least " << MinDistanceNeeded << " size in bytes"; } } while (false) | ||||
1682 | << MinDistanceNeeded << " size in bytes")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because it needs at least " << MinDistanceNeeded << " size in bytes"; } } while (false); | ||||
1683 | return Dependence::Backward; | ||||
1684 | } | ||||
1685 | |||||
1686 | // Positive distance bigger than max vectorization factor. | ||||
1687 | // FIXME: Should use max factor instead of max distance in bytes, which could | ||||
1688 | // not handle different types. | ||||
1689 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. | ||||
1690 | // void foo (int *A, char *B) { | ||||
1691 | // for (unsigned i = 0; i < 1024; i++) { | ||||
1692 | // A[i+2] = A[i] + 1; | ||||
1693 | // B[i+2] = B[i] + 1; | ||||
1694 | // } | ||||
1695 | // } | ||||
1696 | // | ||||
1697 | // This case is currently unsafe according to the max safe distance. If we | ||||
1698 | // analyze the two accesses on array B, the max safe dependence distance | ||||
1699 | // is 2. Then we analyze the accesses on array A, the minimum distance needed | ||||
1700 | // is 8, which is less than 2 and forbidden vectorization, But actually | ||||
1701 | // both A and B could be vectorized by 2 iterations. | ||||
1702 | MaxSafeDepDistBytes = | ||||
1703 | std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes); | ||||
1704 | |||||
1705 | bool IsTrueDataDependence = (!AIsWrite && BIsWrite); | ||||
1706 | if (IsTrueDataDependence && EnableForwardingConflictDetection && | ||||
1707 | couldPreventStoreLoadForward(Distance, TypeByteSize)) | ||||
1708 | return Dependence::BackwardVectorizableButPreventsForwarding; | ||||
1709 | |||||
1710 | uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride); | ||||
1711 | LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Positive distance " << Val.getSExtValue() << " with max VF = " << MaxVF << '\n'; } } while (false) | ||||
1712 | << " with max VF = " << MaxVF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Positive distance " << Val.getSExtValue() << " with max VF = " << MaxVF << '\n'; } } while (false); | ||||
1713 | uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8; | ||||
1714 | MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits); | ||||
1715 | return Dependence::BackwardVectorizable; | ||||
1716 | } | ||||
1717 | |||||
1718 | bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets, | ||||
1719 | MemAccessInfoList &CheckDeps, | ||||
1720 | const ValueToValueMap &Strides) { | ||||
1721 | |||||
1722 | MaxSafeDepDistBytes = -1; | ||||
1723 | SmallPtrSet<MemAccessInfo, 8> Visited; | ||||
1724 | for (MemAccessInfo CurAccess : CheckDeps) { | ||||
1725 | if (Visited.count(CurAccess)) | ||||
1726 | continue; | ||||
1727 | |||||
1728 | // Get the relevant memory access set. | ||||
1729 | EquivalenceClasses<MemAccessInfo>::iterator I = | ||||
1730 | AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); | ||||
1731 | |||||
1732 | // Check accesses within this set. | ||||
1733 | EquivalenceClasses<MemAccessInfo>::member_iterator AI = | ||||
1734 | AccessSets.member_begin(I); | ||||
1735 | EquivalenceClasses<MemAccessInfo>::member_iterator AE = | ||||
1736 | AccessSets.member_end(); | ||||
1737 | |||||
1738 | // Check every access pair. | ||||
1739 | while (AI != AE) { | ||||
1740 | Visited.insert(*AI); | ||||
1741 | bool AIIsWrite = AI->getInt(); | ||||
1742 | // Check loads only against next equivalent class, but stores also against | ||||
1743 | // other stores in the same equivalence class - to the same address. | ||||
1744 | EquivalenceClasses<MemAccessInfo>::member_iterator OI = | ||||
1745 | (AIIsWrite ? AI : std::next(AI)); | ||||
1746 | while (OI != AE) { | ||||
1747 | // Check every accessing instruction pair in program order. | ||||
1748 | for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), | ||||
1749 | I1E = Accesses[*AI].end(); I1 != I1E; ++I1) | ||||
1750 | // Scan all accesses of another equivalence class, but only the next | ||||
1751 | // accesses of the same equivalent class. | ||||
1752 | for (std::vector<unsigned>::iterator | ||||
1753 | I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()), | ||||
1754 | I2E = (OI == AI ? I1E : Accesses[*OI].end()); | ||||
1755 | I2 != I2E; ++I2) { | ||||
1756 | auto A = std::make_pair(&*AI, *I1); | ||||
1757 | auto B = std::make_pair(&*OI, *I2); | ||||
1758 | |||||
1759 | assert(*I1 != *I2)(static_cast <bool> (*I1 != *I2) ? void (0) : __assert_fail ("*I1 != *I2", "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1759 , __extension__ __PRETTY_FUNCTION__)); | ||||
1760 | if (*I1 > *I2) | ||||
1761 | std::swap(A, B); | ||||
1762 | |||||
1763 | Dependence::DepType Type = | ||||
1764 | isDependent(*A.first, A.second, *B.first, B.second, Strides); | ||||
1765 | mergeInStatus(Dependence::isSafeForVectorization(Type)); | ||||
1766 | |||||
1767 | // Gather dependences unless we accumulated MaxDependences | ||||
1768 | // dependences. In that case return as soon as we find the first | ||||
1769 | // unsafe dependence. This puts a limit on this quadratic | ||||
1770 | // algorithm. | ||||
1771 | if (RecordDependences) { | ||||
1772 | if (Type != Dependence::NoDep) | ||||
1773 | Dependences.push_back(Dependence(A.second, B.second, Type)); | ||||
1774 | |||||
1775 | if (Dependences.size() >= MaxDependences) { | ||||
1776 | RecordDependences = false; | ||||
1777 | Dependences.clear(); | ||||
1778 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n" ; } } while (false) | ||||
1779 | << "Too many dependences, stopped recording\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n" ; } } while (false); | ||||
1780 | } | ||||
1781 | } | ||||
1782 | if (!RecordDependences && !isSafeForVectorization()) | ||||
1783 | return false; | ||||
1784 | } | ||||
1785 | ++OI; | ||||
1786 | } | ||||
1787 | AI++; | ||||
1788 | } | ||||
1789 | } | ||||
1790 | |||||
1791 | LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Total Dependences: " << Dependences.size() << "\n"; } } while (false); | ||||
1792 | return isSafeForVectorization(); | ||||
1793 | } | ||||
1794 | |||||
1795 | SmallVector<Instruction *, 4> | ||||
1796 | MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const { | ||||
1797 | MemAccessInfo Access(Ptr, isWrite); | ||||
1798 | auto &IndexVector = Accesses.find(Access)->second; | ||||
1799 | |||||
1800 | SmallVector<Instruction *, 4> Insts; | ||||
1801 | transform(IndexVector, | ||||
1802 | std::back_inserter(Insts), | ||||
1803 | [&](unsigned Idx) { return this->InstMap[Idx]; }); | ||||
1804 | return Insts; | ||||
1805 | } | ||||
1806 | |||||
1807 | const char *MemoryDepChecker::Dependence::DepName[] = { | ||||
1808 | "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward", | ||||
1809 | "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"}; | ||||
1810 | |||||
1811 | void MemoryDepChecker::Dependence::print( | ||||
1812 | raw_ostream &OS, unsigned Depth, | ||||
1813 | const SmallVectorImpl<Instruction *> &Instrs) const { | ||||
1814 | OS.indent(Depth) << DepName[Type] << ":\n"; | ||||
1815 | OS.indent(Depth + 2) << *Instrs[Source] << " -> \n"; | ||||
1816 | OS.indent(Depth + 2) << *Instrs[Destination] << "\n"; | ||||
1817 | } | ||||
1818 | |||||
1819 | bool LoopAccessInfo::canAnalyzeLoop() { | ||||
1820 | // We need to have a loop header. | ||||
1821 | LLVM_DEBUG(dbgs() << "LAA: Found a loop in "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false) | ||||
1822 | << TheLoop->getHeader()->getParent()->getName() << ": "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false) | ||||
1823 | << TheLoop->getHeader()->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false); | ||||
1824 | |||||
1825 | // We can only analyze innermost loops. | ||||
1826 | if (!TheLoop->isInnermost()) { | ||||
1827 | LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop is not the innermost loop\n" ; } } while (false); | ||||
1828 | recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop"; | ||||
1829 | return false; | ||||
1830 | } | ||||
1831 | |||||
1832 | // We must have a single backedge. | ||||
1833 | if (TheLoop->getNumBackEdges() != 1) { | ||||
1834 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop control flow is not understood by analyzer\n" ; } } while (false) | ||||
1835 | dbgs() << "LAA: loop control flow is not understood by analyzer\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop control flow is not understood by analyzer\n" ; } } while (false); | ||||
1836 | recordAnalysis("CFGNotUnderstood") | ||||
1837 | << "loop control flow is not understood by analyzer"; | ||||
1838 | return false; | ||||
1839 | } | ||||
1840 | |||||
1841 | // ScalarEvolution needs to be able to find the exit count. | ||||
1842 | const SCEV *ExitCount = PSE->getBackedgeTakenCount(); | ||||
1843 | if (isa<SCEVCouldNotCompute>(ExitCount)) { | ||||
1844 | recordAnalysis("CantComputeNumberOfIterations") | ||||
1845 | << "could not determine number of loop iterations"; | ||||
1846 | LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: SCEV could not compute the loop exit count.\n" ; } } while (false); | ||||
1847 | return false; | ||||
1848 | } | ||||
1849 | |||||
1850 | return true; | ||||
1851 | } | ||||
1852 | |||||
1853 | void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI, | ||||
1854 | const TargetLibraryInfo *TLI, | ||||
1855 | DominatorTree *DT) { | ||||
1856 | // Holds the Load and Store instructions. | ||||
1857 | SmallVector<LoadInst *, 16> Loads; | ||||
1858 | SmallVector<StoreInst *, 16> Stores; | ||||
1859 | |||||
1860 | // Holds all the different accesses in the loop. | ||||
1861 | unsigned NumReads = 0; | ||||
1862 | unsigned NumReadWrites = 0; | ||||
1863 | |||||
1864 | bool HasComplexMemInst = false; | ||||
1865 | |||||
1866 | // A runtime check is only legal to insert if there are no convergent calls. | ||||
1867 | HasConvergentOp = false; | ||||
1868 | |||||
1869 | PtrRtChecking->Pointers.clear(); | ||||
1870 | PtrRtChecking->Need = false; | ||||
1871 | |||||
1872 | const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); | ||||
1873 | |||||
1874 | const bool EnableMemAccessVersioningOfLoop = | ||||
1875 | EnableMemAccessVersioning && | ||||
| |||||
1876 | !TheLoop->getHeader()->getParent()->hasOptSize(); | ||||
1877 | |||||
1878 | // For each block. | ||||
1879 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||
1880 | // Scan the BB and collect legal loads and stores. Also detect any | ||||
1881 | // convergent instructions. | ||||
1882 | for (Instruction &I : *BB) { | ||||
1883 | if (auto *Call = dyn_cast<CallBase>(&I)) { | ||||
1884 | if (Call->isConvergent()) | ||||
1885 | HasConvergentOp = true; | ||||
1886 | } | ||||
1887 | |||||
1888 | // With both a non-vectorizable memory instruction and a convergent | ||||
1889 | // operation, found in this loop, no reason to continue the search. | ||||
1890 | if (HasComplexMemInst && HasConvergentOp) { | ||||
1891 | CanVecMem = false; | ||||
1892 | return; | ||||
1893 | } | ||||
1894 | |||||
1895 | // Avoid hitting recordAnalysis multiple times. | ||||
1896 | if (HasComplexMemInst) | ||||
1897 | continue; | ||||
1898 | |||||
1899 | // If this is a load, save it. If this instruction can read from memory | ||||
1900 | // but is not a load, then we quit. Notice that we don't handle function | ||||
1901 | // calls that read or write. | ||||
1902 | if (I.mayReadFromMemory()) { | ||||
1903 | // Many math library functions read the rounding mode. We will only | ||||
1904 | // vectorize a loop if it contains known function calls that don't set | ||||
1905 | // the flag. Therefore, it is safe to ignore this read from memory. | ||||
1906 | auto *Call = dyn_cast<CallInst>(&I); | ||||
1907 | if (Call && getVectorIntrinsicIDForCall(Call, TLI)) | ||||
1908 | continue; | ||||
1909 | |||||
1910 | // If the function has an explicit vectorized counterpart, we can safely | ||||
1911 | // assume that it can be vectorized. | ||||
1912 | if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() && | ||||
1913 | !VFDatabase::getMappings(*Call).empty()) | ||||
1914 | continue; | ||||
1915 | |||||
1916 | auto *Ld = dyn_cast<LoadInst>(&I); | ||||
1917 | if (!Ld) { | ||||
1918 | recordAnalysis("CantVectorizeInstruction", Ld) | ||||
1919 | << "instruction cannot be vectorized"; | ||||
1920 | HasComplexMemInst = true; | ||||
1921 | continue; | ||||
1922 | } | ||||
1923 | if (!Ld->isSimple() && !IsAnnotatedParallel) { | ||||
1924 | recordAnalysis("NonSimpleLoad", Ld) | ||||
1925 | << "read with atomic ordering or volatile read"; | ||||
1926 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a non-simple load.\n" ; } } while (false); | ||||
1927 | HasComplexMemInst = true; | ||||
1928 | continue; | ||||
1929 | } | ||||
1930 | NumLoads++; | ||||
1931 | Loads.push_back(Ld); | ||||
1932 | DepChecker->addAccess(Ld); | ||||
1933 | if (EnableMemAccessVersioningOfLoop) | ||||
1934 | collectStridedAccess(Ld); | ||||
1935 | continue; | ||||
1936 | } | ||||
1937 | |||||
1938 | // Save 'store' instructions. Abort if other instructions write to memory. | ||||
1939 | if (I.mayWriteToMemory()) { | ||||
1940 | auto *St = dyn_cast<StoreInst>(&I); | ||||
1941 | if (!St) { | ||||
1942 | recordAnalysis("CantVectorizeInstruction", St) | ||||
1943 | << "instruction cannot be vectorized"; | ||||
1944 | HasComplexMemInst = true; | ||||
1945 | continue; | ||||
1946 | } | ||||
1947 | if (!St->isSimple() && !IsAnnotatedParallel) { | ||||
1948 | recordAnalysis("NonSimpleStore", St) | ||||
1949 | << "write with atomic ordering or volatile write"; | ||||
1950 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a non-simple store.\n" ; } } while (false); | ||||
1951 | HasComplexMemInst = true; | ||||
1952 | continue; | ||||
1953 | } | ||||
1954 | NumStores++; | ||||
1955 | Stores.push_back(St); | ||||
1956 | DepChecker->addAccess(St); | ||||
1957 | if (EnableMemAccessVersioningOfLoop) | ||||
1958 | collectStridedAccess(St); | ||||
1959 | } | ||||
1960 | } // Next instr. | ||||
1961 | } // Next block. | ||||
1962 | |||||
1963 | if (HasComplexMemInst
| ||||
1964 | CanVecMem = false; | ||||
1965 | return; | ||||
1966 | } | ||||
1967 | |||||
1968 | // Now we have two lists that hold the loads and the stores. | ||||
1969 | // Next, we find the pointers that they use. | ||||
1970 | |||||
1971 | // Check if we see any stores. If there are no stores, then we don't | ||||
1972 | // care if the pointers are *restrict*. | ||||
1973 | if (!Stores.size()) { | ||||
1974 | LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a read-only loop!\n" ; } } while (false); | ||||
1975 | CanVecMem = true; | ||||
1976 | return; | ||||
1977 | } | ||||
1978 | |||||
1979 | MemoryDepChecker::DepCandidates DependentAccesses; | ||||
1980 | AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE); | ||||
1981 | |||||
1982 | // Holds the analyzed pointers. We don't want to call getUnderlyingObjects | ||||
1983 | // multiple times on the same object. If the ptr is accessed twice, once | ||||
1984 | // for read and once for write, it will only appear once (on the write | ||||
1985 | // list). This is okay, since we are going to check for conflicts between | ||||
1986 | // writes and between reads and writes, but not between reads and reads. | ||||
1987 | SmallSet<std::pair<Value *, Type *>, 16> Seen; | ||||
1988 | |||||
1989 | // Record uniform store addresses to identify if we have multiple stores | ||||
1990 | // to the same address. | ||||
1991 | SmallPtrSet<Value *, 16> UniformStores; | ||||
1992 | |||||
1993 | for (StoreInst *ST : Stores) { | ||||
1994 | Value *Ptr = ST->getPointerOperand(); | ||||
1995 | |||||
1996 | if (isUniform(Ptr)) | ||||
1997 | HasDependenceInvolvingLoopInvariantAddress |= | ||||
1998 | !UniformStores.insert(Ptr).second; | ||||
1999 | |||||
2000 | // If we did *not* see this pointer before, insert it to the read-write | ||||
2001 | // list. At this phase it is only a 'write' list. | ||||
2002 | Type *AccessTy = getLoadStoreType(ST); | ||||
2003 | if (Seen.insert({Ptr, AccessTy}).second) { | ||||
2004 | ++NumReadWrites; | ||||
2005 | |||||
2006 | MemoryLocation Loc = MemoryLocation::get(ST); | ||||
2007 | // The TBAA metadata could have a control dependency on the predication | ||||
2008 | // condition, so we cannot rely on it when determining whether or not we | ||||
2009 | // need runtime pointer checks. | ||||
2010 | if (blockNeedsPredication(ST->getParent(), TheLoop, DT)) | ||||
2011 | Loc.AATags.TBAA = nullptr; | ||||
2012 | |||||
2013 | visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop, | ||||
2014 | [&Accesses, AccessTy, Loc](Value *Ptr) { | ||||
2015 | MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr); | ||||
2016 | Accesses.addStore(NewLoc, AccessTy); | ||||
2017 | }); | ||||
2018 | } | ||||
2019 | } | ||||
2020 | |||||
2021 | if (IsAnnotatedParallel) { | ||||
2022 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false) | ||||
2023 | dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false) | ||||
2024 | << "checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false); | ||||
2025 | CanVecMem = true; | ||||
2026 | return; | ||||
2027 | } | ||||
2028 | |||||
2029 | for (LoadInst *LD : Loads) { | ||||
2030 | Value *Ptr = LD->getPointerOperand(); | ||||
2031 | // If we did *not* see this pointer before, insert it to the | ||||
2032 | // read list. If we *did* see it before, then it is already in | ||||
2033 | // the read-write list. This allows us to vectorize expressions | ||||
2034 | // such as A[i] += x; Because the address of A[i] is a read-write | ||||
2035 | // pointer. This only works if the index of A[i] is consecutive. | ||||
2036 | // If the address of i is unknown (for example A[B[i]]) then we may | ||||
2037 | // read a few words, modify, and write a few words, and some of the | ||||
2038 | // words may be written to the same address. | ||||
2039 | bool IsReadOnlyPtr = false; | ||||
2040 | Type *AccessTy = getLoadStoreType(LD); | ||||
2041 | if (Seen.insert({Ptr, AccessTy}).second || | ||||
2042 | !getPtrStride(*PSE, LD->getType(), Ptr, TheLoop, SymbolicStrides)) { | ||||
2043 | ++NumReads; | ||||
2044 | IsReadOnlyPtr = true; | ||||
2045 | } | ||||
2046 | |||||
2047 | // See if there is an unsafe dependency between a load to a uniform address and | ||||
2048 | // store to the same uniform address. | ||||
2049 | if (UniformStores.count(Ptr)) { | ||||
2050 | LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found an unsafe dependency between a uniform " "load and uniform store to the same address!\n"; } } while ( false) | ||||
2051 | "load and uniform store to the same address!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found an unsafe dependency between a uniform " "load and uniform store to the same address!\n"; } } while ( false); | ||||
2052 | HasDependenceInvolvingLoopInvariantAddress = true; | ||||
2053 | } | ||||
2054 | |||||
2055 | MemoryLocation Loc = MemoryLocation::get(LD); | ||||
2056 | // The TBAA metadata could have a control dependency on the predication | ||||
2057 | // condition, so we cannot rely on it when determining whether or not we | ||||
2058 | // need runtime pointer checks. | ||||
2059 | if (blockNeedsPredication(LD->getParent(), TheLoop, DT)) | ||||
2060 | Loc.AATags.TBAA = nullptr; | ||||
2061 | |||||
2062 | visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop, | ||||
2063 | [&Accesses, AccessTy, Loc, IsReadOnlyPtr](Value *Ptr) { | ||||
2064 | MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr); | ||||
2065 | Accesses.addLoad(NewLoc, AccessTy, IsReadOnlyPtr); | ||||
2066 | }); | ||||
2067 | } | ||||
2068 | |||||
2069 | // If we write (or read-write) to a single destination and there are no | ||||
2070 | // other reads in this loop then is it safe to vectorize. | ||||
2071 | if (NumReadWrites
| ||||
2072 | LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a write-only loop!\n" ; } } while (false); | ||||
2073 | CanVecMem = true; | ||||
2074 | return; | ||||
2075 | } | ||||
2076 | |||||
2077 | // Build dependence sets and check whether we need a runtime pointer bounds | ||||
2078 | // check. | ||||
2079 | Accesses.buildDependenceSets(); | ||||
2080 | |||||
2081 | // Find pointers with computable bounds. We are going to use this information | ||||
2082 | // to place a runtime bound check. | ||||
2083 | Value *UncomputablePtr = nullptr; | ||||
2084 | bool CanDoRTIfNeeded = | ||||
2085 | Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), TheLoop, | ||||
2086 | SymbolicStrides, UncomputablePtr, false); | ||||
2087 | if (!CanDoRTIfNeeded) { | ||||
2088 | auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr); | ||||
2089 | recordAnalysis("CantIdentifyArrayBounds", I) | ||||
2090 | << "cannot identify array bounds"; | ||||
2091 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because we can't find " << "the array bounds.\n"; } } while (false) | ||||
2092 | << "the array bounds.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because we can't find " << "the array bounds.\n"; } } while (false); | ||||
2093 | CanVecMem = false; | ||||
2094 | return; | ||||
2095 | } | ||||
2096 | |||||
2097 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n" ; } } while (false) | ||||
2098 | dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n" ; } } while (false); | ||||
2099 | |||||
2100 | CanVecMem = true; | ||||
2101 | if (Accesses.isDependencyCheckNeeded()) { | ||||
2102 | LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Checking memory dependencies\n" ; } } while (false); | ||||
2103 | CanVecMem = DepChecker->areDepsSafe( | ||||
2104 | DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides); | ||||
2105 | MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes(); | ||||
2106 | |||||
2107 | if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) { | ||||
2108 | LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Retrying with memory checks\n" ; } } while (false); | ||||
2109 | |||||
2110 | // Clear the dependency checks. We assume they are not needed. | ||||
2111 | Accesses.resetDepChecks(*DepChecker); | ||||
2112 | |||||
2113 | PtrRtChecking->reset(); | ||||
2114 | PtrRtChecking->Need = true; | ||||
2115 | |||||
2116 | auto *SE = PSE->getSE(); | ||||
2117 | UncomputablePtr = nullptr; | ||||
2118 | CanDoRTIfNeeded = Accesses.canCheckPtrAtRT( | ||||
2119 | *PtrRtChecking, SE, TheLoop, SymbolicStrides, UncomputablePtr, true); | ||||
2120 | |||||
2121 | // Check that we found the bounds for the pointer. | ||||
2122 | if (!CanDoRTIfNeeded) { | ||||
2123 | auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr); | ||||
2124 | recordAnalysis("CantCheckMemDepsAtRunTime", I) | ||||
2125 | << "cannot check memory dependencies at runtime"; | ||||
2126 | LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't vectorize with memory checks\n" ; } } while (false); | ||||
2127 | CanVecMem = false; | ||||
2128 | return; | ||||
2129 | } | ||||
2130 | |||||
2131 | CanVecMem = true; | ||||
2132 | } | ||||
2133 | } | ||||
2134 | |||||
2135 | if (HasConvergentOp) { | ||||
2136 | recordAnalysis("CantInsertRuntimeCheckWithConvergent") | ||||
2137 | << "cannot add control dependency to convergent operation"; | ||||
2138 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because a runtime check " "would be needed with a convergent operation\n"; } } while ( false) | ||||
2139 | "would be needed with a convergent operation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because a runtime check " "would be needed with a convergent operation\n"; } } while ( false); | ||||
2140 | CanVecMem = false; | ||||
2141 | return; | ||||
2142 | } | ||||
2143 | |||||
2144 | if (CanVecMem) | ||||
2145 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) | ||||
2146 | dbgs() << "LAA: No unsafe dependent memory operations in loop. We"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) | ||||
2147 | << (PtrRtChecking->Need ? "" : " don't")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) | ||||
2148 | << " need runtime memory checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false); | ||||
2149 | else | ||||
2150 | emitUnsafeDependenceRemark(); | ||||
2151 | } | ||||
2152 | |||||
2153 | void LoopAccessInfo::emitUnsafeDependenceRemark() { | ||||
2154 | auto Deps = getDepChecker().getDependences(); | ||||
2155 | if (!Deps) | ||||
2156 | return; | ||||
2157 | auto Found = std::find_if( | ||||
2158 | Deps->begin(), Deps->end(), [](const MemoryDepChecker::Dependence &D) { | ||||
2159 | return MemoryDepChecker::Dependence::isSafeForVectorization(D.Type) != | ||||
2160 | MemoryDepChecker::VectorizationSafetyStatus::Safe; | ||||
2161 | }); | ||||
2162 | if (Found == Deps->end()) | ||||
2163 | return; | ||||
2164 | MemoryDepChecker::Dependence Dep = *Found; | ||||
2165 | |||||
2166 | LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: unsafe dependent memory operations in loop\n" ; } } while (false); | ||||
2167 | |||||
2168 | // Emit remark for first unsafe dependence | ||||
2169 | OptimizationRemarkAnalysis &R = | ||||
2170 | recordAnalysis("UnsafeDep", Dep.getDestination(*this)) | ||||
2171 | << "unsafe dependent memory operations in loop. Use " | ||||
2172 | "#pragma loop distribute(enable) to allow loop distribution " | ||||
2173 | "to attempt to isolate the offending operations into a separate " | ||||
2174 | "loop"; | ||||
2175 | |||||
2176 | switch (Dep.Type) { | ||||
2177 | case MemoryDepChecker::Dependence::NoDep: | ||||
2178 | case MemoryDepChecker::Dependence::Forward: | ||||
2179 | case MemoryDepChecker::Dependence::BackwardVectorizable: | ||||
2180 | llvm_unreachable("Unexpected dependence")::llvm::llvm_unreachable_internal("Unexpected dependence", "llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 2180); | ||||
2181 | case MemoryDepChecker::Dependence::Backward: | ||||
2182 | R << "\nBackward loop carried data dependence."; | ||||
2183 | break; | ||||
2184 | case MemoryDepChecker::Dependence::ForwardButPreventsForwarding: | ||||
2185 | R << "\nForward loop carried data dependence that prevents " | ||||
2186 | "store-to-load forwarding."; | ||||
2187 | break; | ||||
2188 | case MemoryDepChecker::Dependence::BackwardVectorizableButPreventsForwarding: | ||||
2189 | R << "\nBackward loop carried data dependence that prevents " | ||||
2190 | "store-to-load forwarding."; | ||||
2191 | break; | ||||
2192 | case MemoryDepChecker::Dependence::Unknown: | ||||
2193 | R << "\nUnknown data dependence."; | ||||
2194 | break; | ||||
2195 | } | ||||
2196 | |||||
2197 | if (Instruction *I = Dep.getSource(*this)) { | ||||
2198 | DebugLoc SourceLoc = I->getDebugLoc(); | ||||
2199 | if (auto *DD = dyn_cast_or_null<Instruction>(getPointerOperand(I))) | ||||
2200 | SourceLoc = DD->getDebugLoc(); | ||||
2201 | if (SourceLoc) | ||||
2202 | R << " Memory location is the same as accessed at " | ||||
2203 | << ore::NV("Location", SourceLoc); | ||||
2204 | } | ||||
2205 | } | ||||
2206 | |||||
2207 | bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, | ||||
2208 | DominatorTree *DT) { | ||||
2209 | assert(TheLoop->contains(BB) && "Unknown block used")(static_cast <bool> (TheLoop->contains(BB) && "Unknown block used") ? void (0) : __assert_fail ("TheLoop->contains(BB) && \"Unknown block used\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 2209, __extension__ __PRETTY_FUNCTION__)); | ||||
2210 | |||||
2211 | // Blocks that do not dominate the latch need predication. | ||||
2212 | BasicBlock* Latch = TheLoop->getLoopLatch(); | ||||
2213 | return !DT->dominates(BB, Latch); | ||||
2214 | } | ||||
2215 | |||||
2216 | OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName, | ||||
2217 | Instruction *I) { | ||||
2218 | assert(!Report && "Multiple reports generated")(static_cast <bool> (!Report && "Multiple reports generated" ) ? void (0) : __assert_fail ("!Report && \"Multiple reports generated\"" , "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 2218, __extension__ __PRETTY_FUNCTION__)); | ||||
2219 | |||||
2220 | Value *CodeRegion = TheLoop->getHeader(); | ||||
2221 | DebugLoc DL = TheLoop->getStartLoc(); | ||||
2222 | |||||
2223 | if (I) { | ||||
2224 | CodeRegion = I->getParent(); | ||||
2225 | // If there is no debug location attached to the instruction, revert back to | ||||
2226 | // using the loop's. | ||||
2227 | if (I->getDebugLoc()) | ||||
2228 | DL = I->getDebugLoc(); | ||||
2229 | } | ||||
2230 | |||||
2231 | Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE"loop-accesses", RemarkName, DL, | ||||
2232 | CodeRegion); | ||||
2233 | return *Report; | ||||
2234 | } | ||||
2235 | |||||
2236 | bool LoopAccessInfo::isUniform(Value *V) const { | ||||
2237 | auto *SE = PSE->getSE(); | ||||
2238 | // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is | ||||
2239 | // never considered uniform. | ||||
2240 | // TODO: Is this really what we want? Even without FP SCEV, we may want some | ||||
2241 | // trivially loop-invariant FP values to be considered uniform. | ||||
2242 | if (!SE->isSCEVable(V->getType())) | ||||
2243 | return false; | ||||
2244 | return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); | ||||
2245 | } | ||||
2246 | |||||
2247 | void LoopAccessInfo::collectStridedAccess(Value *MemAccess) { | ||||
2248 | Value *Ptr = getLoadStorePointerOperand(MemAccess); | ||||
2249 | if (!Ptr) | ||||
2250 | return; | ||||
2251 | |||||
2252 | Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop); | ||||
2253 | if (!Stride) | ||||
2254 | return; | ||||
2255 | |||||
2256 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that is a candidate for " "versioning:"; } } while (false) | ||||
2257 | "versioning:")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that is a candidate for " "versioning:"; } } while (false); | ||||
2258 | LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n"; } } while (false ); | ||||
2259 | |||||
2260 | // Avoid adding the "Stride == 1" predicate when we know that | ||||
2261 | // Stride >= Trip-Count. Such a predicate will effectively optimize a single | ||||
2262 | // or zero iteration loop, as Trip-Count <= Stride == 1. | ||||
2263 | // | ||||
2264 | // TODO: We are currently not making a very informed decision on when it is | ||||
2265 | // beneficial to apply stride versioning. It might make more sense that the | ||||
2266 | // users of this analysis (such as the vectorizer) will trigger it, based on | ||||
2267 | // their specific cost considerations; For example, in cases where stride | ||||
2268 | // versioning does not help resolving memory accesses/dependences, the | ||||
2269 | // vectorizer should evaluate the cost of the runtime test, and the benefit | ||||
2270 | // of various possible stride specializations, considering the alternatives | ||||
2271 | // of using gather/scatters (if available). | ||||
2272 | |||||
2273 | const SCEV *StrideExpr = PSE->getSCEV(Stride); | ||||
2274 | const SCEV *BETakenCount = PSE->getBackedgeTakenCount(); | ||||
2275 | |||||
2276 | // Match the types so we can compare the stride and the BETakenCount. | ||||
2277 | // The Stride can be positive/negative, so we sign extend Stride; | ||||
2278 | // The backedgeTakenCount is non-negative, so we zero extend BETakenCount. | ||||
2279 | const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout(); | ||||
2280 | uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType()); | ||||
2281 | uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType()); | ||||
2282 | const SCEV *CastedStride = StrideExpr; | ||||
2283 | const SCEV *CastedBECount = BETakenCount; | ||||
2284 | ScalarEvolution *SE = PSE->getSE(); | ||||
2285 | if (BETypeSize >= StrideTypeSize) | ||||
2286 | CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType()); | ||||
2287 | else | ||||
2288 | CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType()); | ||||
2289 | const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount); | ||||
2290 | // Since TripCount == BackEdgeTakenCount + 1, checking: | ||||
2291 | // "Stride >= TripCount" is equivalent to checking: | ||||
2292 | // Stride - BETakenCount > 0 | ||||
2293 | if (SE->isKnownPositive(StrideMinusBETaken)) { | ||||
2294 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) | ||||
2295 | dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) | ||||
2296 | "Stride==1 predicate will imply that the loop executes "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) | ||||
2297 | "at most once.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false); | ||||
2298 | return; | ||||
2299 | } | ||||
2300 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that we can version.\n" ; } } while (false); | ||||
2301 | |||||
2302 | SymbolicStrides[Ptr] = Stride; | ||||
2303 | StrideSet.insert(Stride); | ||||
2304 | } | ||||
2305 | |||||
2306 | LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, | ||||
2307 | const TargetLibraryInfo *TLI, AAResults *AA, | ||||
2308 | DominatorTree *DT, LoopInfo *LI) | ||||
2309 | : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)), | ||||
2310 | PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)), | ||||
2311 | DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L) { | ||||
2312 | if (canAnalyzeLoop()) | ||||
2313 | analyzeLoop(AA, LI, TLI, DT); | ||||
2314 | } | ||||
2315 | |||||
2316 | void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { | ||||
2317 | if (CanVecMem) { | ||||
2318 | OS.indent(Depth) << "Memory dependences are safe"; | ||||
2319 | if (MaxSafeDepDistBytes != -1ULL) | ||||
2320 | OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes | ||||
2321 | << " bytes"; | ||||
2322 | if (PtrRtChecking->Need) | ||||
2323 | OS << " with run-time checks"; | ||||
2324 | OS << "\n"; | ||||
2325 | } | ||||
2326 | |||||
2327 | if (HasConvergentOp) | ||||
2328 | OS.indent(Depth) << "Has convergent operation in loop\n"; | ||||
2329 | |||||
2330 | if (Report) | ||||
2331 | OS.indent(Depth) << "Report: " << Report->getMsg() << "\n"; | ||||
2332 | |||||
2333 | if (auto *Dependences = DepChecker->getDependences()) { | ||||
2334 | OS.indent(Depth) << "Dependences:\n"; | ||||
2335 | for (auto &Dep : *Dependences) { | ||||
2336 | Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions()); | ||||
2337 | OS << "\n"; | ||||
2338 | } | ||||
2339 | } else | ||||
2340 | OS.indent(Depth) << "Too many dependences, not recorded\n"; | ||||
2341 | |||||
2342 | // List the pair of accesses need run-time checks to prove independence. | ||||
2343 | PtrRtChecking->print(OS, Depth); | ||||
2344 | OS << "\n"; | ||||
2345 | |||||
2346 | OS.indent(Depth) << "Non vectorizable stores to invariant address were " | ||||
2347 | << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ") | ||||
2348 | << "found in loop.\n"; | ||||
2349 | |||||
2350 | OS.indent(Depth) << "SCEV assumptions:\n"; | ||||
2351 | PSE->getPredicate().print(OS, Depth); | ||||
2352 | |||||
2353 | OS << "\n"; | ||||
2354 | |||||
2355 | OS.indent(Depth) << "Expressions re-written:\n"; | ||||
2356 | PSE->print(OS, Depth); | ||||
2357 | } | ||||
2358 | |||||
2359 | LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) { | ||||
2360 | initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry()); | ||||
2361 | } | ||||
2362 | |||||
2363 | const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) { | ||||
2364 | auto &LAI = LoopAccessInfoMap[L]; | ||||
2365 | |||||
2366 | if (!LAI) | ||||
2367 | LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI); | ||||
2368 | |||||
2369 | return *LAI; | ||||
2370 | } | ||||
2371 | |||||
2372 | void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const { | ||||
2373 | LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this); | ||||
2374 | |||||
2375 | for (Loop *TopLevelLoop : *LI) | ||||
2376 | for (Loop *L : depth_first(TopLevelLoop)) { | ||||
2377 | OS.indent(2) << L->getHeader()->getName() << ":\n"; | ||||
2378 | auto &LAI = LAA.getInfo(L); | ||||
2379 | LAI.print(OS, 4); | ||||
2380 | } | ||||
2381 | } | ||||
2382 | |||||
2383 | bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) { | ||||
2384 | SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||
2385 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||
2386 | TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||
2387 | AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||
2388 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||
2389 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||
2390 | |||||
2391 | return false; | ||||
2392 | } | ||||
2393 | |||||
2394 | void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { | ||||
2395 | AU.addRequiredTransitive<ScalarEvolutionWrapperPass>(); | ||||
2396 | AU.addRequiredTransitive<AAResultsWrapperPass>(); | ||||
2397 | AU.addRequiredTransitive<DominatorTreeWrapperPass>(); | ||||
2398 | AU.addRequiredTransitive<LoopInfoWrapperPass>(); | ||||
2399 | |||||
2400 | AU.setPreservesAll(); | ||||
2401 | } | ||||
2402 | |||||
2403 | char LoopAccessLegacyAnalysis::ID = 0; | ||||
2404 | static const char laa_name[] = "Loop Access Analysis"; | ||||
2405 | #define LAA_NAME"loop-accesses" "loop-accesses" | ||||
2406 | |||||
2407 | INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)static void *initializeLoopAccessLegacyAnalysisPassOnce(PassRegistry &Registry) { | ||||
2408 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||
2409 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||
2410 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||
2411 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||
2412 | INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)PassInfo *PI = new PassInfo( laa_name, "loop-accesses", & LoopAccessLegacyAnalysis::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopAccessLegacyAnalysis>), false, true); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLoopAccessLegacyAnalysisPassFlag ; void llvm::initializeLoopAccessLegacyAnalysisPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopAccessLegacyAnalysisPassFlag , initializeLoopAccessLegacyAnalysisPassOnce, std::ref(Registry )); } | ||||
2413 | |||||
2414 | AnalysisKey LoopAccessAnalysis::Key; | ||||
2415 | |||||
2416 | LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM, | ||||
2417 | LoopStandardAnalysisResults &AR) { | ||||
2418 | return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI); | ||||
2419 | } | ||||
2420 | |||||
2421 | namespace llvm { | ||||
2422 | |||||
2423 | Pass *createLAAPass() { | ||||
2424 | return new LoopAccessLegacyAnalysis(); | ||||
2425 | } | ||||
2426 | |||||
2427 | } // end namespace llvm |
1 | //===- llvm/ADT/PointerIntPair.h - Pair for pointer and int -----*- C++ -*-===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | /// | |||
9 | /// \file | |||
10 | /// This file defines the PointerIntPair class. | |||
11 | /// | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #ifndef LLVM_ADT_POINTERINTPAIR_H | |||
15 | #define LLVM_ADT_POINTERINTPAIR_H | |||
16 | ||||
17 | #include "llvm/Support/Compiler.h" | |||
18 | #include "llvm/Support/PointerLikeTypeTraits.h" | |||
19 | #include "llvm/Support/type_traits.h" | |||
20 | #include <cassert> | |||
21 | #include <cstdint> | |||
22 | #include <limits> | |||
23 | ||||
24 | namespace llvm { | |||
25 | ||||
26 | template <typename T, typename Enable> struct DenseMapInfo; | |||
27 | template <typename PointerT, unsigned IntBits, typename PtrTraits> | |||
28 | struct PointerIntPairInfo; | |||
29 | ||||
30 | /// PointerIntPair - This class implements a pair of a pointer and small | |||
31 | /// integer. It is designed to represent this in the space required by one | |||
32 | /// pointer by bitmangling the integer into the low part of the pointer. This | |||
33 | /// can only be done for small integers: typically up to 3 bits, but it depends | |||
34 | /// on the number of bits available according to PointerLikeTypeTraits for the | |||
35 | /// type. | |||
36 | /// | |||
37 | /// Note that PointerIntPair always puts the IntVal part in the highest bits | |||
38 | /// possible. For example, PointerIntPair<void*, 1, bool> will put the bit for | |||
39 | /// the bool into bit #2, not bit #0, which allows the low two bits to be used | |||
40 | /// for something else. For example, this allows: | |||
41 | /// PointerIntPair<PointerIntPair<void*, 1, bool>, 1, bool> | |||
42 | /// ... and the two bools will land in different bits. | |||
43 | template <typename PointerTy, unsigned IntBits, typename IntType = unsigned, | |||
44 | typename PtrTraits = PointerLikeTypeTraits<PointerTy>, | |||
45 | typename Info = PointerIntPairInfo<PointerTy, IntBits, PtrTraits>> | |||
46 | class PointerIntPair { | |||
47 | // Used by MSVC visualizer and generally helpful for debugging/visualizing. | |||
48 | using InfoTy = Info; | |||
49 | intptr_t Value = 0; | |||
50 | ||||
51 | public: | |||
52 | constexpr PointerIntPair() = default; | |||
53 | ||||
54 | PointerIntPair(PointerTy PtrVal, IntType IntVal) { | |||
55 | setPointerAndInt(PtrVal, IntVal); | |||
56 | } | |||
57 | ||||
58 | explicit PointerIntPair(PointerTy PtrVal) { initWithPointer(PtrVal); } | |||
59 | ||||
60 | PointerTy getPointer() const { return Info::getPointer(Value); } | |||
61 | ||||
62 | IntType getInt() const { return (IntType)Info::getInt(Value); } | |||
63 | ||||
64 | void setPointer(PointerTy PtrVal) & { | |||
65 | Value = Info::updatePointer(Value, PtrVal); | |||
66 | } | |||
67 | ||||
68 | void setInt(IntType IntVal) & { | |||
69 | Value = Info::updateInt(Value, static_cast<intptr_t>(IntVal)); | |||
70 | } | |||
71 | ||||
72 | void initWithPointer(PointerTy PtrVal) & { | |||
73 | Value = Info::updatePointer(0, PtrVal); | |||
74 | } | |||
75 | ||||
76 | void setPointerAndInt(PointerTy PtrVal, IntType IntVal) & { | |||
77 | Value = Info::updateInt(Info::updatePointer(0, PtrVal), | |||
78 | static_cast<intptr_t>(IntVal)); | |||
79 | } | |||
80 | ||||
81 | PointerTy const *getAddrOfPointer() const { | |||
82 | return const_cast<PointerIntPair *>(this)->getAddrOfPointer(); | |||
83 | } | |||
84 | ||||
85 | PointerTy *getAddrOfPointer() { | |||
86 | assert(Value == reinterpret_cast<intptr_t>(getPointer()) &&(static_cast <bool> (Value == reinterpret_cast<intptr_t >(getPointer()) && "Can only return the address if IntBits is cleared and " "PtrTraits doesn't change the pointer") ? void (0) : __assert_fail ("Value == reinterpret_cast<intptr_t>(getPointer()) && \"Can only return the address if IntBits is cleared and \" \"PtrTraits doesn't change the pointer\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 88, __extension__ __PRETTY_FUNCTION__)) | |||
87 | "Can only return the address if IntBits is cleared and "(static_cast <bool> (Value == reinterpret_cast<intptr_t >(getPointer()) && "Can only return the address if IntBits is cleared and " "PtrTraits doesn't change the pointer") ? void (0) : __assert_fail ("Value == reinterpret_cast<intptr_t>(getPointer()) && \"Can only return the address if IntBits is cleared and \" \"PtrTraits doesn't change the pointer\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 88, __extension__ __PRETTY_FUNCTION__)) | |||
88 | "PtrTraits doesn't change the pointer")(static_cast <bool> (Value == reinterpret_cast<intptr_t >(getPointer()) && "Can only return the address if IntBits is cleared and " "PtrTraits doesn't change the pointer") ? void (0) : __assert_fail ("Value == reinterpret_cast<intptr_t>(getPointer()) && \"Can only return the address if IntBits is cleared and \" \"PtrTraits doesn't change the pointer\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 88, __extension__ __PRETTY_FUNCTION__)); | |||
89 | return reinterpret_cast<PointerTy *>(&Value); | |||
90 | } | |||
91 | ||||
92 | void *getOpaqueValue() const { return reinterpret_cast<void *>(Value); } | |||
93 | ||||
94 | void setFromOpaqueValue(void *Val) & { | |||
95 | Value = reinterpret_cast<intptr_t>(Val); | |||
96 | } | |||
97 | ||||
98 | static PointerIntPair getFromOpaqueValue(void *V) { | |||
99 | PointerIntPair P; | |||
100 | P.setFromOpaqueValue(V); | |||
101 | return P; | |||
102 | } | |||
103 | ||||
104 | // Allow PointerIntPairs to be created from const void * if and only if the | |||
105 | // pointer type could be created from a const void *. | |||
106 | static PointerIntPair getFromOpaqueValue(const void *V) { | |||
107 | (void)PtrTraits::getFromVoidPointer(V); | |||
108 | return getFromOpaqueValue(const_cast<void *>(V)); | |||
109 | } | |||
110 | ||||
111 | bool operator==(const PointerIntPair &RHS) const { | |||
112 | return Value == RHS.Value; | |||
113 | } | |||
114 | ||||
115 | bool operator!=(const PointerIntPair &RHS) const { | |||
116 | return Value != RHS.Value; | |||
117 | } | |||
118 | ||||
119 | bool operator<(const PointerIntPair &RHS) const { return Value < RHS.Value; } | |||
120 | bool operator>(const PointerIntPair &RHS) const { return Value > RHS.Value; } | |||
121 | ||||
122 | bool operator<=(const PointerIntPair &RHS) const { | |||
123 | return Value <= RHS.Value; | |||
124 | } | |||
125 | ||||
126 | bool operator>=(const PointerIntPair &RHS) const { | |||
127 | return Value >= RHS.Value; | |||
128 | } | |||
129 | }; | |||
130 | ||||
131 | // Specialize is_trivially_copyable to avoid limitation of llvm::is_trivially_copyable | |||
132 | // when compiled with gcc 4.9. | |||
133 | template <typename PointerTy, unsigned IntBits, typename IntType, | |||
134 | typename PtrTraits, | |||
135 | typename Info> | |||
136 | struct is_trivially_copyable<PointerIntPair<PointerTy, IntBits, IntType, PtrTraits, Info>> : std::true_type { | |||
137 | #ifdef HAVE_STD_IS_TRIVIALLY_COPYABLE | |||
138 | static_assert(std::is_trivially_copyable<PointerIntPair<PointerTy, IntBits, IntType, PtrTraits, Info>>::value, | |||
139 | "inconsistent behavior between llvm:: and std:: implementation of is_trivially_copyable"); | |||
140 | #endif | |||
141 | }; | |||
142 | ||||
143 | ||||
144 | template <typename PointerT, unsigned IntBits, typename PtrTraits> | |||
145 | struct PointerIntPairInfo { | |||
146 | static_assert(PtrTraits::NumLowBitsAvailable < | |||
147 | std::numeric_limits<uintptr_t>::digits, | |||
148 | "cannot use a pointer type that has all bits free"); | |||
149 | static_assert(IntBits <= PtrTraits::NumLowBitsAvailable, | |||
150 | "PointerIntPair with integer size too large for pointer"); | |||
151 | enum MaskAndShiftConstants : uintptr_t { | |||
152 | /// PointerBitMask - The bits that come from the pointer. | |||
153 | PointerBitMask = | |||
154 | ~(uintptr_t)(((intptr_t)1 << PtrTraits::NumLowBitsAvailable) - 1), | |||
155 | ||||
156 | /// IntShift - The number of low bits that we reserve for other uses, and | |||
157 | /// keep zero. | |||
158 | IntShift = (uintptr_t)PtrTraits::NumLowBitsAvailable - IntBits, | |||
159 | ||||
160 | /// IntMask - This is the unshifted mask for valid bits of the int type. | |||
161 | IntMask = (uintptr_t)(((intptr_t)1 << IntBits) - 1), | |||
162 | ||||
163 | // ShiftedIntMask - This is the bits for the integer shifted in place. | |||
164 | ShiftedIntMask = (uintptr_t)(IntMask << IntShift) | |||
165 | }; | |||
166 | ||||
167 | static PointerT getPointer(intptr_t Value) { | |||
168 | return PtrTraits::getFromVoidPointer( | |||
169 | reinterpret_cast<void *>(Value & PointerBitMask)); | |||
170 | } | |||
171 | ||||
172 | static intptr_t getInt(intptr_t Value) { | |||
173 | return (Value >> IntShift) & IntMask; | |||
174 | } | |||
175 | ||||
176 | static intptr_t updatePointer(intptr_t OrigValue, PointerT Ptr) { | |||
177 | intptr_t PtrWord = | |||
178 | reinterpret_cast<intptr_t>(PtrTraits::getAsVoidPointer(Ptr)); | |||
179 | assert((PtrWord & ~PointerBitMask) == 0 &&(static_cast <bool> ((PtrWord & ~PointerBitMask) == 0 && "Pointer is not sufficiently aligned") ? void ( 0) : __assert_fail ("(PtrWord & ~PointerBitMask) == 0 && \"Pointer is not sufficiently aligned\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 180, __extension__ __PRETTY_FUNCTION__)) | |||
180 | "Pointer is not sufficiently aligned")(static_cast <bool> ((PtrWord & ~PointerBitMask) == 0 && "Pointer is not sufficiently aligned") ? void ( 0) : __assert_fail ("(PtrWord & ~PointerBitMask) == 0 && \"Pointer is not sufficiently aligned\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 180, __extension__ __PRETTY_FUNCTION__)); | |||
181 | // Preserve all low bits, just update the pointer. | |||
182 | return PtrWord | (OrigValue & ~PointerBitMask); | |||
183 | } | |||
184 | ||||
185 | static intptr_t updateInt(intptr_t OrigValue, intptr_t Int) { | |||
186 | intptr_t IntWord = static_cast<intptr_t>(Int); | |||
187 | assert((IntWord & ~IntMask) == 0 && "Integer too large for field")(static_cast <bool> ((IntWord & ~IntMask) == 0 && "Integer too large for field") ? void (0) : __assert_fail ("(IntWord & ~IntMask) == 0 && \"Integer too large for field\"" , "llvm/include/llvm/ADT/PointerIntPair.h", 187, __extension__ __PRETTY_FUNCTION__)); | |||
188 | ||||
189 | // Preserve all bits other than the ones we are updating. | |||
190 | return (OrigValue & ~ShiftedIntMask) | IntWord << IntShift; | |||
| ||||
191 | } | |||
192 | }; | |||
193 | ||||
194 | // Provide specialization of DenseMapInfo for PointerIntPair. | |||
195 | template <typename PointerTy, unsigned IntBits, typename IntType> | |||
196 | struct DenseMapInfo<PointerIntPair<PointerTy, IntBits, IntType>, void> { | |||
197 | using Ty = PointerIntPair<PointerTy, IntBits, IntType>; | |||
198 | ||||
199 | static Ty getEmptyKey() { | |||
200 | uintptr_t Val = static_cast<uintptr_t>(-1); | |||
201 | Val <<= PointerLikeTypeTraits<Ty>::NumLowBitsAvailable; | |||
202 | return Ty::getFromOpaqueValue(reinterpret_cast<void *>(Val)); | |||
203 | } | |||
204 | ||||
205 | static Ty getTombstoneKey() { | |||
206 | uintptr_t Val = static_cast<uintptr_t>(-2); | |||
207 | Val <<= PointerLikeTypeTraits<PointerTy>::NumLowBitsAvailable; | |||
208 | return Ty::getFromOpaqueValue(reinterpret_cast<void *>(Val)); | |||
209 | } | |||
210 | ||||
211 | static unsigned getHashValue(Ty V) { | |||
212 | uintptr_t IV = reinterpret_cast<uintptr_t>(V.getOpaqueValue()); | |||
213 | return unsigned(IV) ^ unsigned(IV >> 9); | |||
214 | } | |||
215 | ||||
216 | static bool isEqual(const Ty &LHS, const Ty &RHS) { return LHS == RHS; } | |||
217 | }; | |||
218 | ||||
219 | // Teach SmallPtrSet that PointerIntPair is "basically a pointer". | |||
220 | template <typename PointerTy, unsigned IntBits, typename IntType, | |||
221 | typename PtrTraits> | |||
222 | struct PointerLikeTypeTraits< | |||
223 | PointerIntPair<PointerTy, IntBits, IntType, PtrTraits>> { | |||
224 | static inline void * | |||
225 | getAsVoidPointer(const PointerIntPair<PointerTy, IntBits, IntType> &P) { | |||
226 | return P.getOpaqueValue(); | |||
227 | } | |||
228 | ||||
229 | static inline PointerIntPair<PointerTy, IntBits, IntType> | |||
230 | getFromVoidPointer(void *P) { | |||
231 | return PointerIntPair<PointerTy, IntBits, IntType>::getFromOpaqueValue(P); | |||
232 | } | |||
233 | ||||
234 | static inline PointerIntPair<PointerTy, IntBits, IntType> | |||
235 | getFromVoidPointer(const void *P) { | |||
236 | return PointerIntPair<PointerTy, IntBits, IntType>::getFromOpaqueValue(P); | |||
237 | } | |||
238 | ||||
239 | static constexpr int NumLowBitsAvailable = | |||
240 | PtrTraits::NumLowBitsAvailable - IntBits; | |||
241 | }; | |||
242 | ||||
243 | } // end namespace llvm | |||
244 | ||||
245 | #endif // LLVM_ADT_POINTERINTPAIR_H |