Bug Summary

File:llvm/include/llvm/ADT/PointerIntPair.h
Warning:line 189, 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'

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name LoopAccessAnalysis.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/build-llvm -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Analysis -I /build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/llvm/lib/Analysis -I include -I /build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/= -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-01-25-232935-20746-1 -x c++ /build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/llvm/lib/Analysis/LoopAccessAnalysis.cpp

/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/llvm/lib/Analysis/LoopAccessAnalysis.cpp

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 <cstdlib>
64#include <iterator>
65#include <utility>
66#include <vector>
67
68using namespace llvm;
69
70#define DEBUG_TYPE"loop-accesses" "loop-accesses"
71
72static cl::opt<unsigned, true>
73VectorizationFactor("force-vector-width", cl::Hidden,
74 cl::desc("Sets the SIMD width. Zero is autoselect."),
75 cl::location(VectorizerParams::VectorizationFactor));
76unsigned VectorizerParams::VectorizationFactor;
77
78static cl::opt<unsigned, true>
79VectorizationInterleave("force-vector-interleave", cl::Hidden,
80 cl::desc("Sets the vectorization interleave count. "
81 "Zero is autoselect."),
82 cl::location(
83 VectorizerParams::VectorizationInterleave));
84unsigned VectorizerParams::VectorizationInterleave;
85
86static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
87 "runtime-memory-check-threshold", cl::Hidden,
88 cl::desc("When performing memory disambiguation checks at runtime do not "
89 "generate more than this number of comparisons (default = 8)."),
90 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
91unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
92
93/// The maximum iterations used to merge memory checks
94static cl::opt<unsigned> MemoryCheckMergeThreshold(
95 "memory-check-merge-threshold", cl::Hidden,
96 cl::desc("Maximum number of comparisons done when trying to merge "
97 "runtime memory checks. (default = 100)"),
98 cl::init(100));
99
100/// Maximum SIMD width.
101const unsigned VectorizerParams::MaxVectorWidth = 64;
102
103/// We collect dependences up to this threshold.
104static cl::opt<unsigned>
105 MaxDependences("max-dependences", cl::Hidden,
106 cl::desc("Maximum number of dependences collected by "
107 "loop-access analysis (default = 100)"),
108 cl::init(100));
109
110/// This enables versioning on the strides of symbolically striding memory
111/// accesses in code like the following.
112/// for (i = 0; i < N; ++i)
113/// A[i * Stride1] += B[i * Stride2] ...
114///
115/// Will be roughly translated to
116/// if (Stride1 == 1 && Stride2 == 1) {
117/// for (i = 0; i < N; i+=4)
118/// A[i:i+3] += ...
119/// } else
120/// ...
121static cl::opt<bool> EnableMemAccessVersioning(
122 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
123 cl::desc("Enable symbolic stride memory access versioning"));
124
125/// Enable store-to-load forwarding conflict detection. This option can
126/// be disabled for correctness testing.
127static cl::opt<bool> EnableForwardingConflictDetection(
128 "store-to-load-forwarding-conflict-detection", cl::Hidden,
129 cl::desc("Enable conflict detection in loop-access analysis"),
130 cl::init(true));
131
132bool VectorizerParams::isInterleaveForced() {
133 return ::VectorizationInterleave.getNumOccurrences() > 0;
134}
135
136Value *llvm::stripIntegerCast(Value *V) {
137 if (auto *CI = dyn_cast<CastInst>(V))
138 if (CI->getOperand(0)->getType()->isIntegerTy())
139 return CI->getOperand(0);
140 return V;
141}
142
143const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
144 const ValueToValueMap &PtrToStride,
145 Value *Ptr) {
146 const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
147
148 // If there is an entry in the map return the SCEV of the pointer with the
149 // symbolic stride replaced by one.
150 ValueToValueMap::const_iterator SI = PtrToStride.find(Ptr);
151 if (SI == PtrToStride.end())
152 // For a non-symbolic stride, just return the original expression.
153 return OrigSCEV;
154
155 Value *StrideVal = stripIntegerCast(SI->second);
156
157 ScalarEvolution *SE = PSE.getSE();
158 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
159 const auto *CT =
160 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
161
162 PSE.addPredicate(*SE->getEqualPredicate(U, CT));
163 auto *Expr = PSE.getSCEV(Ptr);
164
165 LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEVdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " <<
*OrigSCEV << " by: " << *Expr << "\n"; } }
while (false)
166 << " by: " << *Expr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " <<
*OrigSCEV << " by: " << *Expr << "\n"; } }
while (false)
;
167 return Expr;
168}
169
170RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
171 unsigned Index, RuntimePointerChecking &RtCheck)
172 : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start),
173 AddressSpace(RtCheck.Pointers[Index]
174 .PointerValue->getType()
175 ->getPointerAddressSpace()) {
176 Members.push_back(Index);
177}
178
179/// Calculate Start and End points of memory access.
180/// Let's assume A is the first access and B is a memory access on N-th loop
181/// iteration. Then B is calculated as:
182/// B = A + Step*N .
183/// Step value may be positive or negative.
184/// N is a calculated back-edge taken count:
185/// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
186/// Start and End points are calculated in the following way:
187/// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
188/// where SizeOfElt is the size of single memory access in bytes.
189///
190/// There is no conflict when the intervals are disjoint:
191/// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
192void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
193 unsigned DepSetId, 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 =
231 SE->getStoreSizeOfExpr(IdxTy, Ptr->getType()->getPointerElementType());
232 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
233
234 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
235}
236
237SmallVector<RuntimePointerCheck, 4>
238RuntimePointerChecking::generateChecks() const {
239 SmallVector<RuntimePointerCheck, 4> Checks;
240
241 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
242 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
243 const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I];
244 const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J];
245
246 if (needsChecking(CGI, CGJ))
247 Checks.push_back(std::make_pair(&CGI, &CGJ));
248 }
249 }
250 return Checks;
251}
252
253void RuntimePointerChecking::generateChecks(
254 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
255 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", 255, __extension__
__PRETTY_FUNCTION__))
;
256 groupChecks(DepCands, UseDependencies);
257 Checks = generateChecks();
258}
259
260bool RuntimePointerChecking::needsChecking(
261 const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const {
262 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
263 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
264 if (needsChecking(M.Members[I], N.Members[J]))
265 return true;
266 return false;
267}
268
269/// Compare \p I and \p J and return the minimum.
270/// Return nullptr in case we couldn't find an answer.
271static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
272 ScalarEvolution *SE) {
273 const SCEV *Diff = SE->getMinusSCEV(J, I);
274 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
275
276 if (!C)
277 return nullptr;
278 if (C->getValue()->isNegative())
279 return J;
280 return I;
281}
282
283bool RuntimeCheckingPtrGroup::addPointer(unsigned Index,
284 RuntimePointerChecking &RtCheck) {
285 return addPointer(
286 Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End,
287 RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(),
288 *RtCheck.SE);
289}
290
291bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start,
292 const SCEV *End, unsigned AS,
293 ScalarEvolution &SE) {
294 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", 295, __extension__
__PRETTY_FUNCTION__))
295 "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", 295, __extension__
__PRETTY_FUNCTION__))
;
296
297 // Compare the starts and ends with the known minimum and maximum
298 // of this set. We need to know how we compare against the min/max
299 // of the set in order to be able to emit memchecks.
300 const SCEV *Min0 = getMinFromExprs(Start, Low, &SE);
301 if (!Min0)
302 return false;
303
304 const SCEV *Min1 = getMinFromExprs(End, High, &SE);
305 if (!Min1)
306 return false;
307
308 // Update the low bound expression if we've found a new min value.
309 if (Min0 == Start)
310 Low = Start;
311
312 // Update the high bound expression if we've found a new max value.
313 if (Min1 != End)
314 High = End;
315
316 Members.push_back(Index);
317 return true;
318}
319
320void RuntimePointerChecking::groupChecks(
321 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
322 // We build the groups from dependency candidates equivalence classes
323 // because:
324 // - We know that pointers in the same equivalence class share
325 // the same underlying object and therefore there is a chance
326 // that we can compare pointers
327 // - We wouldn't be able to merge two pointers for which we need
328 // to emit a memcheck. The classes in DepCands are already
329 // conveniently built such that no two pointers in the same
330 // class need checking against each other.
331
332 // We use the following (greedy) algorithm to construct the groups
333 // For every pointer in the equivalence class:
334 // For each existing group:
335 // - if the difference between this pointer and the min/max bounds
336 // of the group is a constant, then make the pointer part of the
337 // group and update the min/max bounds of that group as required.
338
339 CheckingGroups.clear();
340
341 // If we need to check two pointers to the same underlying object
342 // with a non-constant difference, we shouldn't perform any pointer
343 // grouping with those pointers. This is because we can easily get
344 // into cases where the resulting check would return false, even when
345 // the accesses are safe.
346 //
347 // The following example shows this:
348 // for (i = 0; i < 1000; ++i)
349 // a[5000 + i * m] = a[i] + a[i + 9000]
350 //
351 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
352 // (0, 10000) which is always false. However, if m is 1, there is no
353 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
354 // us to perform an accurate check in this case.
355 //
356 // The above case requires that we have an UnknownDependence between
357 // accesses to the same underlying object. This cannot happen unless
358 // FoundNonConstantDistanceDependence is set, and therefore UseDependencies
359 // is also false. In this case we will use the fallback path and create
360 // separate checking groups for all pointers.
361
362 // If we don't have the dependency partitions, construct a new
363 // checking pointer group for each pointer. This is also required
364 // for correctness, because in this case we can have checking between
365 // pointers to the same underlying object.
366 if (!UseDependencies) {
367 for (unsigned I = 0; I < Pointers.size(); ++I)
368 CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this));
369 return;
370 }
371
372 unsigned TotalComparisons = 0;
373
374 DenseMap<Value *, unsigned> PositionMap;
375 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
376 PositionMap[Pointers[Index].PointerValue] = Index;
377
378 // We need to keep track of what pointers we've already seen so we
379 // don't process them twice.
380 SmallSet<unsigned, 2> Seen;
381
382 // Go through all equivalence classes, get the "pointer check groups"
383 // and add them to the overall solution. We use the order in which accesses
384 // appear in 'Pointers' to enforce determinism.
385 for (unsigned I = 0; I < Pointers.size(); ++I) {
386 // We've seen this pointer before, and therefore already processed
387 // its equivalence class.
388 if (Seen.count(I))
389 continue;
390
391 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
392 Pointers[I].IsWritePtr);
393
394 SmallVector<RuntimeCheckingPtrGroup, 2> Groups;
395 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
396
397 // Because DepCands is constructed by visiting accesses in the order in
398 // which they appear in alias sets (which is deterministic) and the
399 // iteration order within an equivalence class member is only dependent on
400 // the order in which unions and insertions are performed on the
401 // equivalence class, the iteration order is deterministic.
402 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
403 MI != ME; ++MI) {
404 auto PointerI = PositionMap.find(MI->getPointer());
405 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", 406, __extension__
__PRETTY_FUNCTION__))
406 "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", 406, __extension__
__PRETTY_FUNCTION__))
;
407 unsigned Pointer = PointerI->second;
408 bool Merged = false;
409 // Mark this pointer as seen.
410 Seen.insert(Pointer);
411
412 // Go through all the existing sets and see if we can find one
413 // which can include this pointer.
414 for (RuntimeCheckingPtrGroup &Group : Groups) {
415 // Don't perform more than a certain amount of comparisons.
416 // This should limit the cost of grouping the pointers to something
417 // reasonable. If we do end up hitting this threshold, the algorithm
418 // will create separate groups for all remaining pointers.
419 if (TotalComparisons > MemoryCheckMergeThreshold)
420 break;
421
422 TotalComparisons++;
423
424 if (Group.addPointer(Pointer, *this)) {
425 Merged = true;
426 break;
427 }
428 }
429
430 if (!Merged)
431 // We couldn't add this pointer to any existing set or the threshold
432 // for the number of comparisons has been reached. Create a new group
433 // to hold the current pointer.
434 Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this));
435 }
436
437 // We've computed the grouped checks for this partition.
438 // Save the results and continue with the next one.
439 llvm::copy(Groups, std::back_inserter(CheckingGroups));
440 }
441}
442
443bool RuntimePointerChecking::arePointersInSamePartition(
444 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
445 unsigned PtrIdx2) {
446 return (PtrToPartition[PtrIdx1] != -1 &&
447 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
448}
449
450bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
451 const PointerInfo &PointerI = Pointers[I];
452 const PointerInfo &PointerJ = Pointers[J];
453
454 // No need to check if two readonly pointers intersect.
455 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
456 return false;
457
458 // Only need to check pointers between two different dependency sets.
459 if (PointerI.DependencySetId == PointerJ.DependencySetId)
460 return false;
461
462 // Only need to check pointers in the same alias set.
463 if (PointerI.AliasSetId != PointerJ.AliasSetId)
464 return false;
465
466 return true;
467}
468
469void RuntimePointerChecking::printChecks(
470 raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks,
471 unsigned Depth) const {
472 unsigned N = 0;
473 for (const auto &Check : Checks) {
474 const auto &First = Check.first->Members, &Second = Check.second->Members;
475
476 OS.indent(Depth) << "Check " << N++ << ":\n";
477
478 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
479 for (unsigned K = 0; K < First.size(); ++K)
480 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
481
482 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
483 for (unsigned K = 0; K < Second.size(); ++K)
484 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
485 }
486}
487
488void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
489
490 OS.indent(Depth) << "Run-time memory checks:\n";
491 printChecks(OS, Checks, Depth);
492
493 OS.indent(Depth) << "Grouped accesses:\n";
494 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
495 const auto &CG = CheckingGroups[I];
496
497 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
498 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
499 << ")\n";
500 for (unsigned J = 0; J < CG.Members.size(); ++J) {
501 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
502 << "\n";
503 }
504 }
505}
506
507namespace {
508
509/// Analyses memory accesses in a loop.
510///
511/// Checks whether run time pointer checks are needed and builds sets for data
512/// dependence checking.
513class AccessAnalysis {
514public:
515 /// Read or write access location.
516 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
517 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList;
518
519 AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI,
520 MemoryDepChecker::DepCandidates &DA,
521 PredicatedScalarEvolution &PSE)
522 : TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA), PSE(PSE) {}
523
524 /// Register a load and whether it is only read from.
525 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
526 Value *Ptr = const_cast<Value*>(Loc.Ptr);
527 AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags);
528 Accesses.insert(MemAccessInfo(Ptr, false));
529 if (IsReadOnly)
530 ReadOnlyPtr.insert(Ptr);
531 }
532
533 /// Register a store.
534 void addStore(MemoryLocation &Loc) {
535 Value *Ptr = const_cast<Value*>(Loc.Ptr);
536 AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags);
537 Accesses.insert(MemAccessInfo(Ptr, true));
538 }
539
540 /// Check if we can emit a run-time no-alias check for \p Access.
541 ///
542 /// Returns true if we can emit a run-time no alias check for \p Access.
543 /// If we can check this access, this also adds it to a dependence set and
544 /// adds a run-time to check for it to \p RtCheck. If \p Assume is true,
545 /// we will attempt to use additional run-time checks in order to get
546 /// the bounds of the pointer.
547 bool createCheckForAccess(RuntimePointerChecking &RtCheck,
548 MemAccessInfo Access,
549 const ValueToValueMap &Strides,
550 DenseMap<Value *, unsigned> &DepSetId,
551 Loop *TheLoop, unsigned &RunningDepId,
552 unsigned ASId, bool ShouldCheckStride,
553 bool Assume);
554
555 /// Check whether we can check the pointers at runtime for
556 /// non-intersection.
557 ///
558 /// Returns true if we need no check or if we do and we can generate them
559 /// (i.e. the pointers have computable bounds).
560 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
561 Loop *TheLoop, const ValueToValueMap &Strides,
562 bool ShouldCheckWrap = false);
563
564 /// Goes over all memory accesses, checks whether a RT check is needed
565 /// and builds sets of dependent accesses.
566 void buildDependenceSets() {
567 processMemAccesses();
11
Calling 'AccessAnalysis::processMemAccesses'
568 }
569
570 /// Initial processing of memory accesses determined that we need to
571 /// perform dependency checking.
572 ///
573 /// Note that this can later be cleared if we retry memcheck analysis without
574 /// dependency checking (i.e. FoundNonConstantDistanceDependence).
575 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
576
577 /// We decided that no dependence analysis would be used. Reset the state.
578 void resetDepChecks(MemoryDepChecker &DepChecker) {
579 CheckDeps.clear();
580 DepChecker.clearDependences();
581 }
582
583 MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; }
584
585private:
586 typedef SetVector<MemAccessInfo> PtrAccessSet;
587
588 /// Go over all memory access and check whether runtime pointer checks
589 /// are needed and build sets of dependency check candidates.
590 void processMemAccesses();
591
592 /// Set of all accesses.
593 PtrAccessSet Accesses;
594
595 /// The loop being checked.
596 const Loop *TheLoop;
597
598 /// List of accesses that need a further dependence check.
599 MemAccessInfoList CheckDeps;
600
601 /// Set of pointers that are read only.
602 SmallPtrSet<Value*, 16> ReadOnlyPtr;
603
604 /// An alias set tracker to partition the access set by underlying object and
605 //intrinsic property (such as TBAA metadata).
606 AliasSetTracker AST;
607
608 LoopInfo *LI;
609
610 /// Sets of potentially dependent accesses - members of one set share an
611 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
612 /// dependence check.
613 MemoryDepChecker::DepCandidates &DepCands;
614
615 /// Initial processing of memory accesses determined that we may need
616 /// to add memchecks. Perform the analysis to determine the necessary checks.
617 ///
618 /// Note that, this is different from isDependencyCheckNeeded. When we retry
619 /// memcheck analysis without dependency checking
620 /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is
621 /// cleared while this remains set if we have potentially dependent accesses.
622 bool IsRTCheckAnalysisNeeded = false;
623
624 /// The SCEV predicate containing all the SCEV-related assumptions.
625 PredicatedScalarEvolution &PSE;
626};
627
628} // end anonymous namespace
629
630/// Check whether a pointer can participate in a runtime bounds check.
631/// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr
632/// by adding run-time checks (overflow checks) if necessary.
633static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
634 const ValueToValueMap &Strides, Value *Ptr,
635 Loop *L, bool Assume) {
636 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
637
638 // The bounds for loop-invariant pointer is trivial.
639 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
640 return true;
641
642 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
643
644 if (!AR && Assume)
645 AR = PSE.getAsAddRec(Ptr);
646
647 if (!AR)
648 return false;
649
650 return AR->isAffine();
651}
652
653/// Check whether a pointer address cannot wrap.
654static bool isNoWrap(PredicatedScalarEvolution &PSE,
655 const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
656 const SCEV *PtrScev = PSE.getSCEV(Ptr);
657 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
658 return true;
659
660 Type *AccessTy = Ptr->getType()->getPointerElementType();
661 int64_t Stride = getPtrStride(PSE, AccessTy, Ptr, L, Strides);
662 if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW))
663 return true;
664
665 return false;
666}
667
668static void visitPointers(Value *StartPtr, const Loop &InnermostLoop,
669 function_ref<void(Value *)> AddPointer) {
670 SmallPtrSet<Value *, 8> Visited;
671 SmallVector<Value *> WorkList;
672 WorkList.push_back(StartPtr);
673
674 while (!WorkList.empty()) {
675 Value *Ptr = WorkList.pop_back_val();
676 if (!Visited.insert(Ptr).second)
677 continue;
678 auto *PN = dyn_cast<PHINode>(Ptr);
679 // SCEV does not look through non-header PHIs inside the loop. Such phis
680 // can be analyzed by adding separate accesses for each incoming pointer
681 // value.
682 if (PN && InnermostLoop.contains(PN->getParent()) &&
683 PN->getParent() != InnermostLoop.getHeader()) {
684 for (const Use &Inc : PN->incoming_values())
685 WorkList.push_back(Inc);
686 } else
687 AddPointer(Ptr);
688 }
689}
690
691bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck,
692 MemAccessInfo Access,
693 const ValueToValueMap &StridesMap,
694 DenseMap<Value *, unsigned> &DepSetId,
695 Loop *TheLoop, unsigned &RunningDepId,
696 unsigned ASId, bool ShouldCheckWrap,
697 bool Assume) {
698 Value *Ptr = Access.getPointer();
699
700 if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume))
701 return false;
702
703 // When we run after a failing dependency check we have to make sure
704 // we don't have wrapping pointers.
705 if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, TheLoop)) {
706 auto *Expr = PSE.getSCEV(Ptr);
707 if (!Assume || !isa<SCEVAddRecExpr>(Expr))
708 return false;
709 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
710 }
711
712 // The id of the dependence set.
713 unsigned DepId;
714
715 if (isDependencyCheckNeeded()) {
716 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
717 unsigned &LeaderId = DepSetId[Leader];
718 if (!LeaderId)
719 LeaderId = RunningDepId++;
720 DepId = LeaderId;
721 } else
722 // Each access has its own dependence set.
723 DepId = RunningDepId++;
724
725 bool IsWrite = Access.getInt();
726 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
727 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)
;
728
729 return true;
730 }
731
732bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
733 ScalarEvolution *SE, Loop *TheLoop,
734 const ValueToValueMap &StridesMap,
735 bool ShouldCheckWrap) {
736 // Find pointers with computable bounds. We are going to use this information
737 // to place a runtime bound check.
738 bool CanDoRT = true;
739
740 bool MayNeedRTCheck = false;
741 if (!IsRTCheckAnalysisNeeded) return true;
742
743 bool IsDepCheckNeeded = isDependencyCheckNeeded();
744
745 // We assign a consecutive id to access from different alias sets.
746 // Accesses between different groups doesn't need to be checked.
747 unsigned ASId = 0;
748 for (auto &AS : AST) {
749 int NumReadPtrChecks = 0;
750 int NumWritePtrChecks = 0;
751 bool CanDoAliasSetRT = true;
752 ++ASId;
753
754 // We assign consecutive id to access from different dependence sets.
755 // Accesses within the same set don't need a runtime check.
756 unsigned RunningDepId = 1;
757 DenseMap<Value *, unsigned> DepSetId;
758
759 SmallVector<MemAccessInfo, 4> Retries;
760
761 // First, count how many write and read accesses are in the alias set. Also
762 // collect MemAccessInfos for later.
763 SmallVector<MemAccessInfo, 4> AccessInfos;
764 for (const auto &A : AS) {
765 Value *Ptr = A.getValue();
766 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
767
768 if (IsWrite)
769 ++NumWritePtrChecks;
770 else
771 ++NumReadPtrChecks;
772 AccessInfos.emplace_back(Ptr, IsWrite);
773 }
774
775 // We do not need runtime checks for this alias set, if there are no writes
776 // or a single write and no reads.
777 if (NumWritePtrChecks == 0 ||
778 (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) {
779 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", 786, __extension__
__PRETTY_FUNCTION__))
780 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", 786, __extension__
__PRETTY_FUNCTION__))
781 [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", 786, __extension__
__PRETTY_FUNCTION__))
782 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", 786, __extension__
__PRETTY_FUNCTION__))
783 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", 786, __extension__
__PRETTY_FUNCTION__))
784 })) &&(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", 786, __extension__
__PRETTY_FUNCTION__))
785 "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", 786, __extension__
__PRETTY_FUNCTION__))
786 "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", 786, __extension__
__PRETTY_FUNCTION__))
;
787 continue;
788 }
789
790 for (auto &Access : AccessInfos) {
791 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop,
792 RunningDepId, ASId, ShouldCheckWrap, false)) {
793 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
)
794 << *Access.getPointer() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Can't find bounds for ptr:"
<< *Access.getPointer() << '\n'; } } while (false
)
;
795 Retries.push_back(Access);
796 CanDoAliasSetRT = false;
797 }
798 }
799
800 // Note that this function computes CanDoRT and MayNeedRTCheck
801 // independently. For example CanDoRT=false, MayNeedRTCheck=false means that
802 // we have a pointer for which we couldn't find the bounds but we don't
803 // actually need to emit any checks so it does not matter.
804 //
805 // We need runtime checks for this alias set, if there are at least 2
806 // dependence sets (in which case RunningDepId > 2) or if we need to re-try
807 // any bound checks (because in that case the number of dependence sets is
808 // incomplete).
809 bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty();
810
811 // We need to perform run-time alias checks, but some pointers had bounds
812 // that couldn't be checked.
813 if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) {
814 // Reset the CanDoSetRt flag and retry all accesses that have failed.
815 // We know that we need these checks, so we can now be more aggressive
816 // and add further checks if required (overflow checks).
817 CanDoAliasSetRT = true;
818 for (auto Access : Retries)
819 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId,
820 TheLoop, RunningDepId, ASId,
821 ShouldCheckWrap, /*Assume=*/true)) {
822 CanDoAliasSetRT = false;
823 break;
824 }
825 }
826
827 CanDoRT &= CanDoAliasSetRT;
828 MayNeedRTCheck |= NeedsAliasSetRTCheck;
829 ++ASId;
830 }
831
832 // If the pointers that we would use for the bounds comparison have different
833 // address spaces, assume the values aren't directly comparable, so we can't
834 // use them for the runtime check. We also have to assume they could
835 // overlap. In the future there should be metadata for whether address spaces
836 // are disjoint.
837 unsigned NumPointers = RtCheck.Pointers.size();
838 for (unsigned i = 0; i < NumPointers; ++i) {
839 for (unsigned j = i + 1; j < NumPointers; ++j) {
840 // Only need to check pointers between two different dependency sets.
841 if (RtCheck.Pointers[i].DependencySetId ==
842 RtCheck.Pointers[j].DependencySetId)
843 continue;
844 // Only need to check pointers in the same alias set.
845 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
846 continue;
847
848 Value *PtrI = RtCheck.Pointers[i].PointerValue;
849 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
850
851 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
852 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
853 if (ASi != ASj) {
854 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between"
" different address spaces\n"; } } while (false)
855 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)
856 " 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)
;
857 return false;
858 }
859 }
860 }
861
862 if (MayNeedRTCheck && CanDoRT)
863 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
864
865 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)
866 << " pointer comparisons.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: We need to do " <<
RtCheck.getNumberOfChecks() << " pointer comparisons.\n"
; } } while (false)
;
867
868 // If we can do run-time checks, but there are no checks, no runtime checks
869 // are needed. This can happen when all pointers point to the same underlying
870 // object for example.
871 RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck;
872
873 bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT;
874 if (!CanDoRTIfNeeded)
875 RtCheck.reset();
876 return CanDoRTIfNeeded;
877}
878
879void AccessAnalysis::processMemAccesses() {
880 // We process the set twice: first we process read-write pointers, last we
881 // process read-only pointers. This allows us to skip dependence tests for
882 // read-only pointers.
883
884 LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Processing memory accesses...\n"
; } } while (false)
;
12
Assuming 'DebugFlag' is false
885 LLVM_DEBUG(dbgs() << " AST: "; AST.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << " AST: "; AST.dump(); }
} while (false)
;
13
Loop condition is false. Exiting loop
886 LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Accesses(" <<
Accesses.size() << "):\n"; } } while (false)
;
14
Loop condition is false. Exiting loop
887 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
15
Loop condition is false. Exiting loop
16
Loop condition is false. Exiting loop
888 for (auto A : Accesses)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
889 dbgs() << "\t" << *A.getPointer() << " (" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
890 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
891 "read-only" : "read")) << ")\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
892 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { { for (auto A : Accesses) dbgs() <<
"\t" << *A.getPointer() << " (" << (A.getInt
() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only"
: "read")) << ")\n"; }; } } while (false)
;
893
894 // The AliasSetTracker has nicely partitioned our pointers by metadata
895 // compatibility and potential for underlying-object overlap. As a result, we
896 // only need to check for potential pointer dependencies within each alias
897 // set.
898 for (const auto &AS : AST) {
899 // Note that both the alias-set tracker and the alias sets themselves used
900 // linked lists internally and so the iteration order here is deterministic
901 // (matching the original instruction order within each set).
902
903 bool SetHasWrite = false;
904
905 // Map of pointers to last access encountered.
906 typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap;
907 UnderlyingObjToAccessMap ObjToLastAccess;
908
909 // Set of access to check after all writes have been processed.
910 PtrAccessSet DeferredAccesses;
911
912 // Iterate over each alias set twice, once to process read/write pointers,
913 // and then to process read-only pointers.
914 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
17
Loop condition is true. Entering loop body
915 bool UseDeferred = SetIteration > 0;
916 PtrAccessSet &S = UseDeferred
17.1
'UseDeferred' is false
17.1
'UseDeferred' is false
? DeferredAccesses : Accesses;
18
'?' condition is false
917
918 for (const auto &AV : AS) {
919 Value *Ptr = AV.getValue();
920
921 // For a single memory access in AliasSetTracker, Accesses may contain
922 // both read and write, and they both need to be handled for CheckDeps.
923 for (const auto &AC : S) {
924 if (AC.getPointer() != Ptr)
19
Assuming the condition is false
20
Taking false branch
925 continue;
926
927 bool IsWrite = AC.getInt();
21
'IsWrite' initialized here
928
929 // If we're using the deferred access set, then it contains only
930 // reads.
931 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
22
Assuming the condition is false
932 if (UseDeferred
22.1
'UseDeferred' is false
22.1
'UseDeferred' is false
&& !IsReadOnlyPtr)
933 continue;
934 // Otherwise, the pointer must be in the PtrAccessSet, either as a
935 // read or a write.
936 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", 938, __extension__
__PRETTY_FUNCTION__))
23
Assuming 'IsWrite' is false
24
Assuming the condition is true
25
'?' condition is true
937 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", 938, __extension__
__PRETTY_FUNCTION__))
938 "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", 938, __extension__
__PRETTY_FUNCTION__))
;
939
940 MemAccessInfo Access(Ptr, IsWrite);
26
Passing value via 2nd parameter 'IntVal'
27
Calling constructor for 'PointerIntPair<llvm::Value *, 1U, bool, llvm::PointerLikeTypeTraits<llvm::Value *>, llvm::PointerIntPairInfo<llvm::Value *, 1, llvm::PointerLikeTypeTraits<llvm::Value *>>>'
941 DepCands.insert(Access);
942
943 // Memorize read-only pointers for later processing and skip them in
944 // the first round (they need to be checked after we have seen all
945 // write pointers). Note: we also mark pointer that are not
946 // consecutive as "read-only" pointers (so that we check
947 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
948 if (!UseDeferred && IsReadOnlyPtr) {
949 DeferredAccesses.insert(Access);
950 continue;
951 }
952
953 // If this is a write - check other reads and writes for conflicts. If
954 // this is a read only check other writes for conflicts (but only if
955 // there is no other write to the ptr - this is an optimization to
956 // catch "a[i] = a[i] + " without having to do a dependence check).
957 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
958 CheckDeps.push_back(Access);
959 IsRTCheckAnalysisNeeded = true;
960 }
961
962 if (IsWrite)
963 SetHasWrite = true;
964
965 // Create sets of pointers connected by a shared alias set and
966 // underlying object.
967 typedef SmallVector<const Value *, 16> ValueVector;
968 ValueVector TempObjects;
969
970 getUnderlyingObjects(Ptr, TempObjects, LI);
971 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "Underlying objects for pointer "
<< *Ptr << "\n"; } } while (false)
972 << "Underlying objects for pointer " << *Ptr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "Underlying objects for pointer "
<< *Ptr << "\n"; } } while (false)
;
973 for (const Value *UnderlyingObj : TempObjects) {
974 // nullptr never alias, don't join sets for pointer that have "null"
975 // in their UnderlyingObjects list.
976 if (isa<ConstantPointerNull>(UnderlyingObj) &&
977 !NullPointerIsDefined(
978 TheLoop->getHeader()->getParent(),
979 UnderlyingObj->getType()->getPointerAddressSpace()))
980 continue;
981
982 UnderlyingObjToAccessMap::iterator Prev =
983 ObjToLastAccess.find(UnderlyingObj);
984 if (Prev != ObjToLastAccess.end())
985 DepCands.unionSets(Access, Prev->second);
986
987 ObjToLastAccess[UnderlyingObj] = Access;
988 LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << " " << *UnderlyingObj
<< "\n"; } } while (false)
;
989 }
990 }
991 }
992 }
993 }
994}
995
996static bool isInBoundsGep(Value *Ptr) {
997 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
998 return GEP->isInBounds();
999 return false;
1000}
1001
1002/// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
1003/// i.e. monotonically increasing/decreasing.
1004static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
1005 PredicatedScalarEvolution &PSE, const Loop *L) {
1006 // FIXME: This should probably only return true for NUW.
1007 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
1008 return true;
1009
1010 // Scalar evolution does not propagate the non-wrapping flags to values that
1011 // are derived from a non-wrapping induction variable because non-wrapping
1012 // could be flow-sensitive.
1013 //
1014 // Look through the potentially overflowing instruction to try to prove
1015 // non-wrapping for the *specific* value of Ptr.
1016
1017 // The arithmetic implied by an inbounds GEP can't overflow.
1018 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
1019 if (!GEP || !GEP->isInBounds())
1020 return false;
1021
1022 // Make sure there is only one non-const index and analyze that.
1023 Value *NonConstIndex = nullptr;
1024 for (Value *Index : GEP->indices())
1025 if (!isa<ConstantInt>(Index)) {
1026 if (NonConstIndex)
1027 return false;
1028 NonConstIndex = Index;
1029 }
1030 if (!NonConstIndex)
1031 // The recurrence is on the pointer, ignore for now.
1032 return false;
1033
1034 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
1035 // AddRec using a NSW operation.
1036 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
1037 if (OBO->hasNoSignedWrap() &&
1038 // Assume constant for other the operand so that the AddRec can be
1039 // easily found.
1040 isa<ConstantInt>(OBO->getOperand(1))) {
1041 auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
1042
1043 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
1044 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
1045 }
1046
1047 return false;
1048}
1049
1050/// Check whether the access through \p Ptr has a constant stride.
1051int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy,
1052 Value *Ptr, const Loop *Lp,
1053 const ValueToValueMap &StridesMap, bool Assume,
1054 bool ShouldCheckWrap) {
1055 Type *Ty = Ptr->getType();
1056 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", 1056, __extension__
__PRETTY_FUNCTION__))
;
1057
1058 if (isa<ScalableVectorType>(AccessTy)) {
1059 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)
1060 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Bad stride - Scalable object: "
<< *AccessTy << "\n"; } } while (false)
;
1061 return 0;
1062 }
1063
1064 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
1065
1066 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
1067 if (Assume && !AR)
1068 AR = PSE.getAsAddRec(Ptr);
1069
1070 if (!AR) {
1071 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)
1072 << " 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)
;
1073 return 0;
1074 }
1075
1076 // The access function must stride over the innermost loop.
1077 if (Lp != AR->getLoop()) {
1078 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)
1079 << *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)
;
1080 return 0;
1081 }
1082
1083 // The address calculation must not wrap. Otherwise, a dependence could be
1084 // inverted.
1085 // An inbounds getelementptr that is a AddRec with a unit stride
1086 // cannot wrap per definition. The unit stride requirement is checked later.
1087 // An getelementptr without an inbounds attribute and unit stride would have
1088 // to access the pointer value "0" which is undefined behavior in address
1089 // space 0, therefore we can also vectorize this case.
1090 unsigned AddrSpace = Ty->getPointerAddressSpace();
1091 bool IsInBoundsGEP = isInBoundsGep(Ptr);
1092 bool IsNoWrapAddRec = !ShouldCheckWrap ||
1093 PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
1094 isNoWrapAddRec(Ptr, AR, PSE, Lp);
1095 if (!IsNoWrapAddRec && !IsInBoundsGEP &&
1096 NullPointerIsDefined(Lp->getHeader()->getParent(), AddrSpace)) {
1097 if (Assume) {
1098 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1099 IsNoWrapAddRec = true;
1100 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)
1101 << "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)
1102 << "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)
1103 << "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)
;
1104 } else {
1105 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)
1106 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)
1107 << *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)
;
1108 return 0;
1109 }
1110 }
1111
1112 // Check the step is constant.
1113 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
1114
1115 // Calculate the pointer stride and check if it is constant.
1116 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1117 if (!C) {
1118 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)
1119 << " 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)
;
1120 return 0;
1121 }
1122
1123 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
1124 TypeSize AllocSize = DL.getTypeAllocSize(AccessTy);
1125 int64_t Size = AllocSize.getFixedSize();
1126 const APInt &APStepVal = C->getAPInt();
1127
1128 // Huge step value - give up.
1129 if (APStepVal.getBitWidth() > 64)
1130 return 0;
1131
1132 int64_t StepVal = APStepVal.getSExtValue();
1133
1134 // Strided access.
1135 int64_t Stride = StepVal / Size;
1136 int64_t Rem = StepVal % Size;
1137 if (Rem)
1138 return 0;
1139
1140 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
1141 // know we can't "wrap around the address space". In case of address space
1142 // zero we know that this won't happen without triggering undefined behavior.
1143 if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 &&
1144 (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(),
1145 AddrSpace))) {
1146 if (Assume) {
1147 // We can avoid this case by adding a run-time check.
1148 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)
1149 << "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)
1150 << "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)
1151 << "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)
1152 << "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)
;
1153 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1154 } else
1155 return 0;
1156 }
1157
1158 return Stride;
1159}
1160
1161Optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB,
1162 Value *PtrB, const DataLayout &DL,
1163 ScalarEvolution &SE, bool StrictCheck,
1164 bool CheckType) {
1165 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", 1165, __extension__
__PRETTY_FUNCTION__))
;
1166 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", 1167, __extension__
__PRETTY_FUNCTION__))
1167 ->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", 1167, __extension__
__PRETTY_FUNCTION__))
;
1168 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", 1169, __extension__
__PRETTY_FUNCTION__))
1169 ->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", 1169, __extension__
__PRETTY_FUNCTION__))
;
1170
1171 // Make sure that A and B are different pointers.
1172 if (PtrA == PtrB)
1173 return 0;
1174
1175 // Make sure that the element types are the same if required.
1176 if (CheckType && ElemTyA != ElemTyB)
1177 return None;
1178
1179 unsigned ASA = PtrA->getType()->getPointerAddressSpace();
1180 unsigned ASB = PtrB->getType()->getPointerAddressSpace();
1181
1182 // Check that the address spaces match.
1183 if (ASA != ASB)
1184 return None;
1185 unsigned IdxWidth = DL.getIndexSizeInBits(ASA);
1186
1187 APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
1188 Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1189 Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1190
1191 int Val;
1192 if (PtrA1 == PtrB1) {
1193 // Retrieve the address space again as pointer stripping now tracks through
1194 // `addrspacecast`.
1195 ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace();
1196 ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace();
1197 // Check that the address spaces match and that the pointers are valid.
1198 if (ASA != ASB)
1199 return None;
1200
1201 IdxWidth = DL.getIndexSizeInBits(ASA);
1202 OffsetA = OffsetA.sextOrTrunc(IdxWidth);
1203 OffsetB = OffsetB.sextOrTrunc(IdxWidth);
1204
1205 OffsetB -= OffsetA;
1206 Val = OffsetB.getSExtValue();
1207 } else {
1208 // Otherwise compute the distance with SCEV between the base pointers.
1209 const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1210 const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1211 const auto *Diff =
1212 dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA));
1213 if (!Diff)
1214 return None;
1215 Val = Diff->getAPInt().getSExtValue();
1216 }
1217 int Size = DL.getTypeStoreSize(ElemTyA);
1218 int Dist = Val / Size;
1219
1220 // Ensure that the calculated distance matches the type-based one after all
1221 // the bitcasts removal in the provided pointers.
1222 if (!StrictCheck || Dist * Size == Val)
1223 return Dist;
1224 return None;
1225}
1226
1227bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy,
1228 const DataLayout &DL, ScalarEvolution &SE,
1229 SmallVectorImpl<unsigned> &SortedIndices) {
1230 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", 1232, __extension__
__PRETTY_FUNCTION__))
1231 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", 1232, __extension__
__PRETTY_FUNCTION__))
1232 "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", 1232, __extension__
__PRETTY_FUNCTION__))
;
1233 // Walk over the pointers, and map each of them to an offset relative to
1234 // first pointer in the array.
1235 Value *Ptr0 = VL[0];
1236
1237 using DistOrdPair = std::pair<int64_t, int>;
1238 auto Compare = [](const DistOrdPair &L, const DistOrdPair &R) {
1239 return L.first < R.first;
1240 };
1241 std::set<DistOrdPair, decltype(Compare)> Offsets(Compare);
1242 Offsets.emplace(0, 0);
1243 int Cnt = 1;
1244 bool IsConsecutive = true;
1245 for (auto *Ptr : VL.drop_front()) {
1246 Optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE,
1247 /*StrictCheck=*/true);
1248 if (!Diff)
1249 return false;
1250
1251 // Check if the pointer with the same offset is found.
1252 int64_t Offset = *Diff;
1253 auto Res = Offsets.emplace(Offset, Cnt);
1254 if (!Res.second)
1255 return false;
1256 // Consecutive order if the inserted element is the last one.
1257 IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end();
1258 ++Cnt;
1259 }
1260 SortedIndices.clear();
1261 if (!IsConsecutive) {
1262 // Fill SortedIndices array only if it is non-consecutive.
1263 SortedIndices.resize(VL.size());
1264 Cnt = 0;
1265 for (const std::pair<int64_t, int> &Pair : Offsets) {
1266 SortedIndices[Cnt] = Pair.second;
1267 ++Cnt;
1268 }
1269 }
1270 return true;
1271}
1272
1273/// Returns true if the memory operations \p A and \p B are consecutive.
1274bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1275 ScalarEvolution &SE, bool CheckType) {
1276 Value *PtrA = getLoadStorePointerOperand(A);
1277 Value *PtrB = getLoadStorePointerOperand(B);
1278 if (!PtrA || !PtrB)
1279 return false;
1280 Type *ElemTyA = getLoadStoreType(A);
1281 Type *ElemTyB = getLoadStoreType(B);
1282 Optional<int> Diff = getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE,
1283 /*StrictCheck=*/true, CheckType);
1284 return Diff && *Diff == 1;
1285}
1286
1287void MemoryDepChecker::addAccess(StoreInst *SI) {
1288 visitPointers(SI->getPointerOperand(), *InnermostLoop,
1289 [this, SI](Value *Ptr) {
1290 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
1291 InstMap.push_back(SI);
1292 ++AccessIdx;
1293 });
1294}
1295
1296void MemoryDepChecker::addAccess(LoadInst *LI) {
1297 visitPointers(LI->getPointerOperand(), *InnermostLoop,
1298 [this, LI](Value *Ptr) {
1299 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
1300 InstMap.push_back(LI);
1301 ++AccessIdx;
1302 });
1303}
1304
1305MemoryDepChecker::VectorizationSafetyStatus
1306MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1307 switch (Type) {
1308 case NoDep:
1309 case Forward:
1310 case BackwardVectorizable:
1311 return VectorizationSafetyStatus::Safe;
1312
1313 case Unknown:
1314 return VectorizationSafetyStatus::PossiblySafeWithRtChecks;
1315 case ForwardButPreventsForwarding:
1316 case Backward:
1317 case BackwardVectorizableButPreventsForwarding:
1318 return VectorizationSafetyStatus::Unsafe;
1319 }
1320 llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp"
, 1320)
;
1321}
1322
1323bool MemoryDepChecker::Dependence::isBackward() const {
1324 switch (Type) {
1325 case NoDep:
1326 case Forward:
1327 case ForwardButPreventsForwarding:
1328 case Unknown:
1329 return false;
1330
1331 case BackwardVectorizable:
1332 case Backward:
1333 case BackwardVectorizableButPreventsForwarding:
1334 return true;
1335 }
1336 llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp"
, 1336)
;
1337}
1338
1339bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1340 return isBackward() || Type == Unknown;
1341}
1342
1343bool MemoryDepChecker::Dependence::isForward() const {
1344 switch (Type) {
1345 case Forward:
1346 case ForwardButPreventsForwarding:
1347 return true;
1348
1349 case NoDep:
1350 case Unknown:
1351 case BackwardVectorizable:
1352 case Backward:
1353 case BackwardVectorizableButPreventsForwarding:
1354 return false;
1355 }
1356 llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "llvm/lib/Analysis/LoopAccessAnalysis.cpp"
, 1356)
;
1357}
1358
1359bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1360 uint64_t TypeByteSize) {
1361 // If loads occur at a distance that is not a multiple of a feasible vector
1362 // factor store-load forwarding does not take place.
1363 // Positive dependences might cause troubles because vectorizing them might
1364 // prevent store-load forwarding making vectorized code run a lot slower.
1365 // a[i] = a[i-3] ^ a[i-8];
1366 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1367 // hence on your typical architecture store-load forwarding does not take
1368 // place. Vectorizing in such cases does not make sense.
1369 // Store-load forwarding distance.
1370
1371 // After this many iterations store-to-load forwarding conflicts should not
1372 // cause any slowdowns.
1373 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1374 // Maximum vector factor.
1375 uint64_t MaxVFWithoutSLForwardIssues = std::min(
1376 VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1377
1378 // Compute the smallest VF at which the store and load would be misaligned.
1379 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1380 VF *= 2) {
1381 // If the number of vector iteration between the store and the load are
1382 // small we could incur conflicts.
1383 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1384 MaxVFWithoutSLForwardIssues = (VF >> 1);
1385 break;
1386 }
1387 }
1388
1389 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1390 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)
1391 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)
1392 << " 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)
;
1393 return true;
1394 }
1395
1396 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1397 MaxVFWithoutSLForwardIssues !=
1398 VectorizerParams::MaxVectorWidth * TypeByteSize)
1399 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1400 return false;
1401}
1402
1403void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) {
1404 if (Status < S)
1405 Status = S;
1406}
1407
1408/// Given a non-constant (unknown) dependence-distance \p Dist between two
1409/// memory accesses, that have the same stride whose absolute value is given
1410/// in \p Stride, and that have the same type size \p TypeByteSize,
1411/// in a loop whose takenCount is \p BackedgeTakenCount, check if it is
1412/// possible to prove statically that the dependence distance is larger
1413/// than the range that the accesses will travel through the execution of
1414/// the loop. If so, return true; false otherwise. This is useful for
1415/// example in loops such as the following (PR31098):
1416/// for (i = 0; i < D; ++i) {
1417/// = out[i];
1418/// out[i+D] =
1419/// }
1420static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE,
1421 const SCEV &BackedgeTakenCount,
1422 const SCEV &Dist, uint64_t Stride,
1423 uint64_t TypeByteSize) {
1424
1425 // If we can prove that
1426 // (**) |Dist| > BackedgeTakenCount * Step
1427 // where Step is the absolute stride of the memory accesses in bytes,
1428 // then there is no dependence.
1429 //
1430 // Rationale:
1431 // We basically want to check if the absolute distance (|Dist/Step|)
1432 // is >= the loop iteration count (or > BackedgeTakenCount).
1433 // This is equivalent to the Strong SIV Test (Practical Dependence Testing,
1434 // Section 4.2.1); Note, that for vectorization it is sufficient to prove
1435 // that the dependence distance is >= VF; This is checked elsewhere.
1436 // But in some cases we can prune unknown dependence distances early, and
1437 // even before selecting the VF, and without a runtime test, by comparing
1438 // the distance against the loop iteration count. Since the vectorized code
1439 // will be executed only if LoopCount >= VF, proving distance >= LoopCount
1440 // also guarantees that distance >= VF.
1441 //
1442 const uint64_t ByteStride = Stride * TypeByteSize;
1443 const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride);
1444 const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step);
1445
1446 const SCEV *CastedDist = &Dist;
1447 const SCEV *CastedProduct = Product;
1448 uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType());
1449 uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType());
1450
1451 // The dependence distance can be positive/negative, so we sign extend Dist;
1452 // The multiplication of the absolute stride in bytes and the
1453 // backedgeTakenCount is non-negative, so we zero extend Product.
1454 if (DistTypeSize > ProductTypeSize)
1455 CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType());
1456 else
1457 CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType());
1458
1459 // Is Dist - (BackedgeTakenCount * Step) > 0 ?
1460 // (If so, then we have proven (**) because |Dist| >= Dist)
1461 const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct);
1462 if (SE.isKnownPositive(Minus))
1463 return true;
1464
1465 // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ?
1466 // (If so, then we have proven (**) because |Dist| >= -1*Dist)
1467 const SCEV *NegDist = SE.getNegativeSCEV(CastedDist);
1468 Minus = SE.getMinusSCEV(NegDist, CastedProduct);
1469 if (SE.isKnownPositive(Minus))
1470 return true;
1471
1472 return false;
1473}
1474
1475/// Check the dependence for two accesses with the same stride \p Stride.
1476/// \p Distance is the positive distance and \p TypeByteSize is type size in
1477/// bytes.
1478///
1479/// \returns true if they are independent.
1480static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1481 uint64_t TypeByteSize) {
1482 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", 1482, __extension__
__PRETTY_FUNCTION__))
;
1483 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", 1483, __extension__
__PRETTY_FUNCTION__))
;
1484 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", 1484, __extension__
__PRETTY_FUNCTION__))
;
1485
1486 // Skip if the distance is not multiple of type byte size.
1487 if (Distance % TypeByteSize)
1488 return false;
1489
1490 uint64_t ScaledDist = Distance / TypeByteSize;
1491
1492 // No dependence if the scaled distance is not multiple of the stride.
1493 // E.g.
1494 // for (i = 0; i < 1024 ; i += 4)
1495 // A[i+2] = A[i] + 1;
1496 //
1497 // Two accesses in memory (scaled distance is 2, stride is 4):
1498 // | A[0] | | | | A[4] | | | |
1499 // | | | A[2] | | | | A[6] | |
1500 //
1501 // E.g.
1502 // for (i = 0; i < 1024 ; i += 3)
1503 // A[i+4] = A[i] + 1;
1504 //
1505 // Two accesses in memory (scaled distance is 4, stride is 3):
1506 // | A[0] | | | A[3] | | | A[6] | | |
1507 // | | | | | A[4] | | | A[7] | |
1508 return ScaledDist % Stride;
1509}
1510
1511MemoryDepChecker::Dependence::DepType
1512MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1513 const MemAccessInfo &B, unsigned BIdx,
1514 const ValueToValueMap &Strides) {
1515 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", 1515, __extension__
__PRETTY_FUNCTION__))
;
1516
1517 Value *APtr = A.getPointer();
1518 Value *BPtr = B.getPointer();
1519 bool AIsWrite = A.getInt();
1520 bool BIsWrite = B.getInt();
1521 Type *ATy = APtr->getType()->getPointerElementType();
1522 Type *BTy = BPtr->getType()->getPointerElementType();
1523
1524 // Two reads are independent.
1525 if (!AIsWrite && !BIsWrite)
1526 return Dependence::NoDep;
1527
1528 // We cannot check pointers in different address spaces.
1529 if (APtr->getType()->getPointerAddressSpace() !=
1530 BPtr->getType()->getPointerAddressSpace())
1531 return Dependence::Unknown;
1532
1533 int64_t StrideAPtr =
1534 getPtrStride(PSE, ATy, APtr, InnermostLoop, Strides, true);
1535 int64_t StrideBPtr =
1536 getPtrStride(PSE, BTy, BPtr, InnermostLoop, Strides, true);
1537
1538 const SCEV *Src = PSE.getSCEV(APtr);
1539 const SCEV *Sink = PSE.getSCEV(BPtr);
1540
1541 // If the induction step is negative we have to invert source and sink of the
1542 // dependence.
1543 if (StrideAPtr < 0) {
1544 std::swap(APtr, BPtr);
1545 std::swap(ATy, BTy);
1546 std::swap(Src, Sink);
1547 std::swap(AIsWrite, BIsWrite);
1548 std::swap(AIdx, BIdx);
1549 std::swap(StrideAPtr, StrideBPtr);
1550 }
1551
1552 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1553
1554 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)
1555 << "(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)
;
1556 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)
1557 << *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)
;
1558
1559 // Need accesses with constant stride. We don't want to vectorize
1560 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1561 // the address space.
1562 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1563 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)
;
1564 return Dependence::Unknown;
1565 }
1566
1567 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1568 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1569 bool HasSameSize =
1570 DL.getTypeStoreSizeInBits(ATy) == DL.getTypeStoreSizeInBits(BTy);
1571 uint64_t Stride = std::abs(StrideAPtr);
1572 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1573 if (!C) {
1574 if (!isa<SCEVCouldNotCompute>(Dist) && HasSameSize &&
1575 isSafeDependenceDistance(DL, *(PSE.getSE()),
1576 *(PSE.getBackedgeTakenCount()), *Dist, Stride,
1577 TypeByteSize))
1578 return Dependence::NoDep;
1579
1580 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)
;
1581 FoundNonConstantDistanceDependence = true;
1582 return Dependence::Unknown;
1583 }
1584
1585 const APInt &Val = C->getAPInt();
1586 int64_t Distance = Val.getSExtValue();
1587
1588 // Attempt to prove strided accesses independent.
1589 if (std::abs(Distance) > 0 && Stride > 1 && HasSameSize &&
1590 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1591 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)
;
1592 return Dependence::NoDep;
1593 }
1594
1595 // Negative distances are not plausible dependencies.
1596 if (Val.isNegative()) {
1597 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1598 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1599 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1600 !HasSameSize)) {
1601 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)
;
1602 return Dependence::ForwardButPreventsForwarding;
1603 }
1604
1605 LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Dependence is negative\n"
; } } while (false)
;
1606 return Dependence::Forward;
1607 }
1608
1609 // Write to the same location with the same size.
1610 if (Val == 0) {
1611 if (HasSameSize)
1612 return Dependence::Forward;
1613 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Zero dependence difference but different type sizes\n"
; } } while (false)
1614 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)
;
1615 return Dependence::Unknown;
1616 }
1617
1618 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", 1618, __extension__
__PRETTY_FUNCTION__))
;
1619
1620 if (!HasSameSize) {
1621 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)
1622 "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)
;
1623 return Dependence::Unknown;
1624 }
1625
1626 // Bail out early if passed-in parameters make vectorization not feasible.
1627 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1628 VectorizerParams::VectorizationFactor : 1);
1629 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1630 VectorizerParams::VectorizationInterleave : 1);
1631 // The minimum number of iterations for a vectorized/unrolled version.
1632 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1633
1634 // It's not vectorizable if the distance is smaller than the minimum distance
1635 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1636 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1637 // TypeByteSize (No need to plus the last gap distance).
1638 //
1639 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1640 // foo(int *A) {
1641 // int *B = (int *)((char *)A + 14);
1642 // for (i = 0 ; i < 1024 ; i += 2)
1643 // B[i] = A[i] + 1;
1644 // }
1645 //
1646 // Two accesses in memory (stride is 2):
1647 // | A[0] | | A[2] | | A[4] | | A[6] | |
1648 // | B[0] | | B[2] | | B[4] |
1649 //
1650 // Distance needs for vectorizing iterations except the last iteration:
1651 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1652 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1653 //
1654 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1655 // 12, which is less than distance.
1656 //
1657 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1658 // the minimum distance needed is 28, which is greater than distance. It is
1659 // not safe to do vectorization.
1660 uint64_t MinDistanceNeeded =
1661 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1662 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1663 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)
1664 << Distance << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Failure because of positive distance "
<< Distance << '\n'; } } while (false)
;
1665 return Dependence::Backward;
1666 }
1667
1668 // Unsafe if the minimum distance needed is greater than max safe distance.
1669 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1670 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)
1671 << 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)
;
1672 return Dependence::Backward;
1673 }
1674
1675 // Positive distance bigger than max vectorization factor.
1676 // FIXME: Should use max factor instead of max distance in bytes, which could
1677 // not handle different types.
1678 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1679 // void foo (int *A, char *B) {
1680 // for (unsigned i = 0; i < 1024; i++) {
1681 // A[i+2] = A[i] + 1;
1682 // B[i+2] = B[i] + 1;
1683 // }
1684 // }
1685 //
1686 // This case is currently unsafe according to the max safe distance. If we
1687 // analyze the two accesses on array B, the max safe dependence distance
1688 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1689 // is 8, which is less than 2 and forbidden vectorization, But actually
1690 // both A and B could be vectorized by 2 iterations.
1691 MaxSafeDepDistBytes =
1692 std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1693
1694 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1695 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1696 couldPreventStoreLoadForward(Distance, TypeByteSize))
1697 return Dependence::BackwardVectorizableButPreventsForwarding;
1698
1699 uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride);
1700 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)
1701 << " 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)
;
1702 uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
1703 MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits);
1704 return Dependence::BackwardVectorizable;
1705}
1706
1707bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1708 MemAccessInfoList &CheckDeps,
1709 const ValueToValueMap &Strides) {
1710
1711 MaxSafeDepDistBytes = -1;
1712 SmallPtrSet<MemAccessInfo, 8> Visited;
1713 for (MemAccessInfo CurAccess : CheckDeps) {
1714 if (Visited.count(CurAccess))
1715 continue;
1716
1717 // Get the relevant memory access set.
1718 EquivalenceClasses<MemAccessInfo>::iterator I =
1719 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1720
1721 // Check accesses within this set.
1722 EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1723 AccessSets.member_begin(I);
1724 EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1725 AccessSets.member_end();
1726
1727 // Check every access pair.
1728 while (AI != AE) {
1729 Visited.insert(*AI);
1730 bool AIIsWrite = AI->getInt();
1731 // Check loads only against next equivalent class, but stores also against
1732 // other stores in the same equivalence class - to the same address.
1733 EquivalenceClasses<MemAccessInfo>::member_iterator OI =
1734 (AIIsWrite ? AI : std::next(AI));
1735 while (OI != AE) {
1736 // Check every accessing instruction pair in program order.
1737 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1738 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1739 // Scan all accesses of another equivalence class, but only the next
1740 // accesses of the same equivalent class.
1741 for (std::vector<unsigned>::iterator
1742 I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()),
1743 I2E = (OI == AI ? I1E : Accesses[*OI].end());
1744 I2 != I2E; ++I2) {
1745 auto A = std::make_pair(&*AI, *I1);
1746 auto B = std::make_pair(&*OI, *I2);
1747
1748 assert(*I1 != *I2)(static_cast <bool> (*I1 != *I2) ? void (0) : __assert_fail
("*I1 != *I2", "llvm/lib/Analysis/LoopAccessAnalysis.cpp", 1748
, __extension__ __PRETTY_FUNCTION__))
;
1749 if (*I1 > *I2)
1750 std::swap(A, B);
1751
1752 Dependence::DepType Type =
1753 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1754 mergeInStatus(Dependence::isSafeForVectorization(Type));
1755
1756 // Gather dependences unless we accumulated MaxDependences
1757 // dependences. In that case return as soon as we find the first
1758 // unsafe dependence. This puts a limit on this quadratic
1759 // algorithm.
1760 if (RecordDependences) {
1761 if (Type != Dependence::NoDep)
1762 Dependences.push_back(Dependence(A.second, B.second, Type));
1763
1764 if (Dependences.size() >= MaxDependences) {
1765 RecordDependences = false;
1766 Dependences.clear();
1767 LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n"
; } } while (false)
1768 << "Too many dependences, stopped recording\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n"
; } } while (false)
;
1769 }
1770 }
1771 if (!RecordDependences && !isSafeForVectorization())
1772 return false;
1773 }
1774 ++OI;
1775 }
1776 AI++;
1777 }
1778 }
1779
1780 LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "Total Dependences: " <<
Dependences.size() << "\n"; } } while (false)
;
1781 return isSafeForVectorization();
1782}
1783
1784SmallVector<Instruction *, 4>
1785MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1786 MemAccessInfo Access(Ptr, isWrite);
1787 auto &IndexVector = Accesses.find(Access)->second;
1788
1789 SmallVector<Instruction *, 4> Insts;
1790 transform(IndexVector,
1791 std::back_inserter(Insts),
1792 [&](unsigned Idx) { return this->InstMap[Idx]; });
1793 return Insts;
1794}
1795
1796const char *MemoryDepChecker::Dependence::DepName[] = {
1797 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1798 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1799
1800void MemoryDepChecker::Dependence::print(
1801 raw_ostream &OS, unsigned Depth,
1802 const SmallVectorImpl<Instruction *> &Instrs) const {
1803 OS.indent(Depth) << DepName[Type] << ":\n";
1804 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1805 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1806}
1807
1808bool LoopAccessInfo::canAnalyzeLoop() {
1809 // We need to have a loop header.
1810 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)
1811 << 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)
1812 << 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)
;
1813
1814 // We can only analyze innermost loops.
1815 if (!TheLoop->isInnermost()) {
1816 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)
;
1817 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
1818 return false;
1819 }
1820
1821 // We must have a single backedge.
1822 if (TheLoop->getNumBackEdges() != 1) {
1823 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: loop control flow is not understood by analyzer\n"
; } } while (false)
1824 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)
;
1825 recordAnalysis("CFGNotUnderstood")
1826 << "loop control flow is not understood by analyzer";
1827 return false;
1828 }
1829
1830 // ScalarEvolution needs to be able to find the exit count.
1831 const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1832 if (isa<SCEVCouldNotCompute>(ExitCount)) {
1833 recordAnalysis("CantComputeNumberOfIterations")
1834 << "could not determine number of loop iterations";
1835 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)
;
1836 return false;
1837 }
1838
1839 return true;
1840}
1841
1842void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
1843 const TargetLibraryInfo *TLI,
1844 DominatorTree *DT) {
1845 typedef SmallPtrSet<Value*, 16> ValueSet;
1846
1847 // Holds the Load and Store instructions.
1848 SmallVector<LoadInst *, 16> Loads;
1849 SmallVector<StoreInst *, 16> Stores;
1850
1851 // Holds all the different accesses in the loop.
1852 unsigned NumReads = 0;
1853 unsigned NumReadWrites = 0;
1854
1855 bool HasComplexMemInst = false;
1856
1857 // A runtime check is only legal to insert if there are no convergent calls.
1858 HasConvergentOp = false;
1859
1860 PtrRtChecking->Pointers.clear();
1861 PtrRtChecking->Need = false;
1862
1863 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1864
1865 const bool EnableMemAccessVersioningOfLoop =
1866 EnableMemAccessVersioning &&
1
Assuming the condition is false
1867 !TheLoop->getHeader()->getParent()->hasOptSize();
1868
1869 // For each block.
1870 for (BasicBlock *BB : TheLoop->blocks()) {
2
Assuming '__begin1' is equal to '__end1'
1871 // Scan the BB and collect legal loads and stores. Also detect any
1872 // convergent instructions.
1873 for (Instruction &I : *BB) {
1874 if (auto *Call = dyn_cast<CallBase>(&I)) {
1875 if (Call->isConvergent())
1876 HasConvergentOp = true;
1877 }
1878
1879 // With both a non-vectorizable memory instruction and a convergent
1880 // operation, found in this loop, no reason to continue the search.
1881 if (HasComplexMemInst && HasConvergentOp) {
1882 CanVecMem = false;
1883 return;
1884 }
1885
1886 // Avoid hitting recordAnalysis multiple times.
1887 if (HasComplexMemInst)
1888 continue;
1889
1890 // If this is a load, save it. If this instruction can read from memory
1891 // but is not a load, then we quit. Notice that we don't handle function
1892 // calls that read or write.
1893 if (I.mayReadFromMemory()) {
1894 // Many math library functions read the rounding mode. We will only
1895 // vectorize a loop if it contains known function calls that don't set
1896 // the flag. Therefore, it is safe to ignore this read from memory.
1897 auto *Call = dyn_cast<CallInst>(&I);
1898 if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1899 continue;
1900
1901 // If the function has an explicit vectorized counterpart, we can safely
1902 // assume that it can be vectorized.
1903 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1904 !VFDatabase::getMappings(*Call).empty())
1905 continue;
1906
1907 auto *Ld = dyn_cast<LoadInst>(&I);
1908 if (!Ld) {
1909 recordAnalysis("CantVectorizeInstruction", Ld)
1910 << "instruction cannot be vectorized";
1911 HasComplexMemInst = true;
1912 continue;
1913 }
1914 if (!Ld->isSimple() && !IsAnnotatedParallel) {
1915 recordAnalysis("NonSimpleLoad", Ld)
1916 << "read with atomic ordering or volatile read";
1917 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)
;
1918 HasComplexMemInst = true;
1919 continue;
1920 }
1921 NumLoads++;
1922 Loads.push_back(Ld);
1923 DepChecker->addAccess(Ld);
1924 if (EnableMemAccessVersioningOfLoop)
1925 collectStridedAccess(Ld);
1926 continue;
1927 }
1928
1929 // Save 'store' instructions. Abort if other instructions write to memory.
1930 if (I.mayWriteToMemory()) {
1931 auto *St = dyn_cast<StoreInst>(&I);
1932 if (!St) {
1933 recordAnalysis("CantVectorizeInstruction", St)
1934 << "instruction cannot be vectorized";
1935 HasComplexMemInst = true;
1936 continue;
1937 }
1938 if (!St->isSimple() && !IsAnnotatedParallel) {
1939 recordAnalysis("NonSimpleStore", St)
1940 << "write with atomic ordering or volatile write";
1941 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)
;
1942 HasComplexMemInst = true;
1943 continue;
1944 }
1945 NumStores++;
1946 Stores.push_back(St);
1947 DepChecker->addAccess(St);
1948 if (EnableMemAccessVersioningOfLoop)
1949 collectStridedAccess(St);
1950 }
1951 } // Next instr.
1952 } // Next block.
1953
1954 if (HasComplexMemInst
2.1
'HasComplexMemInst' is false
2.1
'HasComplexMemInst' is false
) {
3
Taking false branch
1955 CanVecMem = false;
1956 return;
1957 }
1958
1959 // Now we have two lists that hold the loads and the stores.
1960 // Next, we find the pointers that they use.
1961
1962 // Check if we see any stores. If there are no stores, then we don't
1963 // care if the pointers are *restrict*.
1964 if (!Stores.size()) {
4
Assuming the condition is false
5
Taking false branch
1965 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)
;
1966 CanVecMem = true;
1967 return;
1968 }
1969
1970 MemoryDepChecker::DepCandidates DependentAccesses;
1971 AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE);
1972
1973 // Holds the analyzed pointers. We don't want to call getUnderlyingObjects
1974 // multiple times on the same object. If the ptr is accessed twice, once
1975 // for read and once for write, it will only appear once (on the write
1976 // list). This is okay, since we are going to check for conflicts between
1977 // writes and between reads and writes, but not between reads and reads.
1978 ValueSet Seen;
1979
1980 // Record uniform store addresses to identify if we have multiple stores
1981 // to the same address.
1982 ValueSet UniformStores;
1983
1984 for (StoreInst *ST : Stores) {
6
Assuming '__begin1' is equal to '__end1'
1985 Value *Ptr = ST->getPointerOperand();
1986
1987 if (isUniform(Ptr))
1988 HasDependenceInvolvingLoopInvariantAddress |=
1989 !UniformStores.insert(Ptr).second;
1990
1991 // If we did *not* see this pointer before, insert it to the read-write
1992 // list. At this phase it is only a 'write' list.
1993 if (Seen.insert(Ptr).second) {
1994 ++NumReadWrites;
1995
1996 MemoryLocation Loc = MemoryLocation::get(ST);
1997 // The TBAA metadata could have a control dependency on the predication
1998 // condition, so we cannot rely on it when determining whether or not we
1999 // need runtime pointer checks.
2000 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
2001 Loc.AATags.TBAA = nullptr;
2002
2003 visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2004 [&Accesses, Loc](Value *Ptr) {
2005 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2006 Accesses.addStore(NewLoc);
2007 });
2008 }
2009 }
2010
2011 if (IsAnnotatedParallel) {
7
Assuming 'IsAnnotatedParallel' is false
8
Taking false branch
2012 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
<< "checks.\n"; } } while (false)
2013 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)
2014 << "checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
<< "checks.\n"; } } while (false)
;
2015 CanVecMem = true;
2016 return;
2017 }
2018
2019 for (LoadInst *LD : Loads) {
9
Assuming '__begin1' is equal to '__end1'
2020 Value *Ptr = LD->getPointerOperand();
2021 // If we did *not* see this pointer before, insert it to the
2022 // read list. If we *did* see it before, then it is already in
2023 // the read-write list. This allows us to vectorize expressions
2024 // such as A[i] += x; Because the address of A[i] is a read-write
2025 // pointer. This only works if the index of A[i] is consecutive.
2026 // If the address of i is unknown (for example A[B[i]]) then we may
2027 // read a few words, modify, and write a few words, and some of the
2028 // words may be written to the same address.
2029 bool IsReadOnlyPtr = false;
2030 if (Seen.insert(Ptr).second ||
2031 !getPtrStride(*PSE, LD->getType(), Ptr, TheLoop, SymbolicStrides)) {
2032 ++NumReads;
2033 IsReadOnlyPtr = true;
2034 }
2035
2036 // See if there is an unsafe dependency between a load to a uniform address and
2037 // store to the same uniform address.
2038 if (UniformStores.count(Ptr)) {
2039 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)
2040 "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)
;
2041 HasDependenceInvolvingLoopInvariantAddress = true;
2042 }
2043
2044 MemoryLocation Loc = MemoryLocation::get(LD);
2045 // The TBAA metadata could have a control dependency on the predication
2046 // condition, so we cannot rely on it when determining whether or not we
2047 // need runtime pointer checks.
2048 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
2049 Loc.AATags.TBAA = nullptr;
2050
2051 visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2052 [&Accesses, Loc, IsReadOnlyPtr](Value *Ptr) {
2053 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2054 Accesses.addLoad(NewLoc, IsReadOnlyPtr);
2055 });
2056 }
2057
2058 // If we write (or read-write) to a single destination and there are no
2059 // other reads in this loop then is it safe to vectorize.
2060 if (NumReadWrites
9.1
'NumReadWrites' is not equal to 1
9.1
'NumReadWrites' is not equal to 1
== 1 && NumReads == 0) {
2061 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)
;
2062 CanVecMem = true;
2063 return;
2064 }
2065
2066 // Build dependence sets and check whether we need a runtime pointer bounds
2067 // check.
2068 Accesses.buildDependenceSets();
10
Calling 'AccessAnalysis::buildDependenceSets'
2069
2070 // Find pointers with computable bounds. We are going to use this information
2071 // to place a runtime bound check.
2072 bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
2073 TheLoop, SymbolicStrides);
2074 if (!CanDoRTIfNeeded) {
2075 recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds";
2076 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)
2077 << "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)
;
2078 CanVecMem = false;
2079 return;
2080 }
2081
2082 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)
2083 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)
;
2084
2085 CanVecMem = true;
2086 if (Accesses.isDependencyCheckNeeded()) {
2087 LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Checking memory dependencies\n"
; } } while (false)
;
2088 CanVecMem = DepChecker->areDepsSafe(
2089 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
2090 MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
2091
2092 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
2093 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)
;
2094
2095 // Clear the dependency checks. We assume they are not needed.
2096 Accesses.resetDepChecks(*DepChecker);
2097
2098 PtrRtChecking->reset();
2099 PtrRtChecking->Need = true;
2100
2101 auto *SE = PSE->getSE();
2102 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
2103 SymbolicStrides, true);
2104
2105 // Check that we found the bounds for the pointer.
2106 if (!CanDoRTIfNeeded) {
2107 recordAnalysis("CantCheckMemDepsAtRunTime")
2108 << "cannot check memory dependencies at runtime";
2109 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)
;
2110 CanVecMem = false;
2111 return;
2112 }
2113
2114 CanVecMem = true;
2115 }
2116 }
2117
2118 if (HasConvergentOp) {
2119 recordAnalysis("CantInsertRuntimeCheckWithConvergent")
2120 << "cannot add control dependency to convergent operation";
2121 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)
2122 "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)
;
2123 CanVecMem = false;
2124 return;
2125 }
2126
2127 if (CanVecMem)
2128 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)
2129 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)
2130 << (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)
2131 << " 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)
;
2132 else {
2133 recordAnalysis("UnsafeMemDep")
2134 << "unsafe dependent memory operations in loop. Use "
2135 "#pragma loop distribute(enable) to allow loop distribution "
2136 "to attempt to isolate the offending operations into a separate "
2137 "loop";
2138 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)
;
2139 }
2140}
2141
2142bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
2143 DominatorTree *DT) {
2144 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", 2144, __extension__
__PRETTY_FUNCTION__))
;
2145
2146 // Blocks that do not dominate the latch need predication.
2147 BasicBlock* Latch = TheLoop->getLoopLatch();
2148 return !DT->dominates(BB, Latch);
2149}
2150
2151OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName,
2152 Instruction *I) {
2153 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", 2153, __extension__
__PRETTY_FUNCTION__))
;
2154
2155 Value *CodeRegion = TheLoop->getHeader();
2156 DebugLoc DL = TheLoop->getStartLoc();
2157
2158 if (I) {
2159 CodeRegion = I->getParent();
2160 // If there is no debug location attached to the instruction, revert back to
2161 // using the loop's.
2162 if (I->getDebugLoc())
2163 DL = I->getDebugLoc();
2164 }
2165
2166 Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE"loop-accesses", RemarkName, DL,
2167 CodeRegion);
2168 return *Report;
2169}
2170
2171bool LoopAccessInfo::isUniform(Value *V) const {
2172 auto *SE = PSE->getSE();
2173 // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is
2174 // never considered uniform.
2175 // TODO: Is this really what we want? Even without FP SCEV, we may want some
2176 // trivially loop-invariant FP values to be considered uniform.
2177 if (!SE->isSCEVable(V->getType()))
2178 return false;
2179 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
2180}
2181
2182void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
2183 Value *Ptr = getLoadStorePointerOperand(MemAccess);
2184 if (!Ptr)
2185 return;
2186
2187 Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
2188 if (!Stride)
2189 return;
2190
2191 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)
2192 "versioning:")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Found a strided access that is a candidate for "
"versioning:"; } } while (false)
;
2193 LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << " Ptr: " << *Ptr <<
" Stride: " << *Stride << "\n"; } } while (false
)
;
2194
2195 // Avoid adding the "Stride == 1" predicate when we know that
2196 // Stride >= Trip-Count. Such a predicate will effectively optimize a single
2197 // or zero iteration loop, as Trip-Count <= Stride == 1.
2198 //
2199 // TODO: We are currently not making a very informed decision on when it is
2200 // beneficial to apply stride versioning. It might make more sense that the
2201 // users of this analysis (such as the vectorizer) will trigger it, based on
2202 // their specific cost considerations; For example, in cases where stride
2203 // versioning does not help resolving memory accesses/dependences, the
2204 // vectorizer should evaluate the cost of the runtime test, and the benefit
2205 // of various possible stride specializations, considering the alternatives
2206 // of using gather/scatters (if available).
2207
2208 const SCEV *StrideExpr = PSE->getSCEV(Stride);
2209 const SCEV *BETakenCount = PSE->getBackedgeTakenCount();
2210
2211 // Match the types so we can compare the stride and the BETakenCount.
2212 // The Stride can be positive/negative, so we sign extend Stride;
2213 // The backedgeTakenCount is non-negative, so we zero extend BETakenCount.
2214 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
2215 uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType());
2216 uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType());
2217 const SCEV *CastedStride = StrideExpr;
2218 const SCEV *CastedBECount = BETakenCount;
2219 ScalarEvolution *SE = PSE->getSE();
2220 if (BETypeSize >= StrideTypeSize)
2221 CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType());
2222 else
2223 CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType());
2224 const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount);
2225 // Since TripCount == BackEdgeTakenCount + 1, checking:
2226 // "Stride >= TripCount" is equivalent to checking:
2227 // Stride - BETakenCount > 0
2228 if (SE->isKnownPositive(StrideMinusBETaken)) {
2229 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)
2230 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)
2231 "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)
2232 "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)
;
2233 return;
2234 }
2235 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-accesses")) { dbgs() << "LAA: Found a strided access that we can version."
; } } while (false)
;
2236
2237 SymbolicStrides[Ptr] = Stride;
2238 StrideSet.insert(Stride);
2239}
2240
2241LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
2242 const TargetLibraryInfo *TLI, AAResults *AA,
2243 DominatorTree *DT, LoopInfo *LI)
2244 : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)),
2245 PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)),
2246 DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L) {
2247 if (canAnalyzeLoop())
2248 analyzeLoop(AA, LI, TLI, DT);
2249}
2250
2251void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
2252 if (CanVecMem) {
2253 OS.indent(Depth) << "Memory dependences are safe";
2254 if (MaxSafeDepDistBytes != -1ULL)
2255 OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
2256 << " bytes";
2257 if (PtrRtChecking->Need)
2258 OS << " with run-time checks";
2259 OS << "\n";
2260 }
2261
2262 if (HasConvergentOp)
2263 OS.indent(Depth) << "Has convergent operation in loop\n";
2264
2265 if (Report)
2266 OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";
2267
2268 if (auto *Dependences = DepChecker->getDependences()) {
2269 OS.indent(Depth) << "Dependences:\n";
2270 for (auto &Dep : *Dependences) {
2271 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
2272 OS << "\n";
2273 }
2274 } else
2275 OS.indent(Depth) << "Too many dependences, not recorded\n";
2276
2277 // List the pair of accesses need run-time checks to prove independence.
2278 PtrRtChecking->print(OS, Depth);
2279 OS << "\n";
2280
2281 OS.indent(Depth) << "Non vectorizable stores to invariant address were "
2282 << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ")
2283 << "found in loop.\n";
2284
2285 OS.indent(Depth) << "SCEV assumptions:\n";
2286 PSE->getUnionPredicate().print(OS, Depth);
2287
2288 OS << "\n";
2289
2290 OS.indent(Depth) << "Expressions re-written:\n";
2291 PSE->print(OS, Depth);
2292}
2293
2294LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) {
2295 initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry());
2296}
2297
2298const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
2299 auto &LAI = LoopAccessInfoMap[L];
2300
2301 if (!LAI)
2302 LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
2303
2304 return *LAI.get();
2305}
2306
2307void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
2308 LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
2309
2310 for (Loop *TopLevelLoop : *LI)
2311 for (Loop *L : depth_first(TopLevelLoop)) {
2312 OS.indent(2) << L->getHeader()->getName() << ":\n";
2313 auto &LAI = LAA.getInfo(L);
2314 LAI.print(OS, 4);
2315 }
2316}
2317
2318bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
2319 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2320 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2321 TLI = TLIP ? &TLIP->getTLI(F) : nullptr;
2322 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2323 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2324 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2325
2326 return false;
2327}
2328
2329void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2330 AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
2331 AU.addRequiredTransitive<AAResultsWrapperPass>();
2332 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
2333 AU.addRequiredTransitive<LoopInfoWrapperPass>();
2334
2335 AU.setPreservesAll();
2336}
2337
2338char LoopAccessLegacyAnalysis::ID = 0;
2339static const char laa_name[] = "Loop Access Analysis";
2340#define LAA_NAME"loop-accesses" "loop-accesses"
2341
2342INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)static void *initializeLoopAccessLegacyAnalysisPassOnce(PassRegistry
&Registry) {
2343INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
2344INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
2345INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
2346INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
2347INITIALIZE_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
)); }
2348
2349AnalysisKey LoopAccessAnalysis::Key;
2350
2351LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM,
2352 LoopStandardAnalysisResults &AR) {
2353 return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI);
2354}
2355
2356namespace llvm {
2357
2358 Pass *createLAAPass() {
2359 return new LoopAccessLegacyAnalysis();
2360 }
2361
2362} // end namespace llvm

/build/llvm-toolchain-snapshot-14~++20220125101009+ceec4383681c/llvm/include/llvm/ADT/PointerIntPair.h

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// This file defines the PointerIntPair class.
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_ADT_POINTERINTPAIR_H
14#define LLVM_ADT_POINTERINTPAIR_H
15
16#include "llvm/Support/Compiler.h"
17#include "llvm/Support/PointerLikeTypeTraits.h"
18#include "llvm/Support/type_traits.h"
19#include <cassert>
20#include <cstdint>
21#include <limits>
22
23namespace llvm {
24
25template <typename T, typename Enable> struct DenseMapInfo;
26template <typename PointerT, unsigned IntBits, typename PtrTraits>
27struct PointerIntPairInfo;
28
29/// PointerIntPair - This class implements a pair of a pointer and small
30/// integer. It is designed to represent this in the space required by one
31/// pointer by bitmangling the integer into the low part of the pointer. This
32/// can only be done for small integers: typically up to 3 bits, but it depends
33/// on the number of bits available according to PointerLikeTypeTraits for the
34/// type.
35///
36/// Note that PointerIntPair always puts the IntVal part in the highest bits
37/// possible. For example, PointerIntPair<void*, 1, bool> will put the bit for
38/// the bool into bit #2, not bit #0, which allows the low two bits to be used
39/// for something else. For example, this allows:
40/// PointerIntPair<PointerIntPair<void*, 1, bool>, 1, bool>
41/// ... and the two bools will land in different bits.
42template <typename PointerTy, unsigned IntBits, typename IntType = unsigned,
43 typename PtrTraits = PointerLikeTypeTraits<PointerTy>,
44 typename Info = PointerIntPairInfo<PointerTy, IntBits, PtrTraits>>
45class PointerIntPair {
46 // Used by MSVC visualizer and generally helpful for debugging/visualizing.
47 using InfoTy = Info;
48 intptr_t Value = 0;
49
50public:
51 constexpr PointerIntPair() = default;
52
53 PointerIntPair(PointerTy PtrVal, IntType IntVal) {
54 setPointerAndInt(PtrVal, IntVal);
28
Passing value via 2nd parameter 'IntVal'
29
Calling 'PointerIntPair::setPointerAndInt'
55 }
56
57 explicit PointerIntPair(PointerTy PtrVal) { initWithPointer(PtrVal); }
58
59 PointerTy getPointer() const { return Info::getPointer(Value); }
60
61 IntType getInt() const { return (IntType)Info::getInt(Value); }
62
63 void setPointer(PointerTy PtrVal) LLVM_LVALUE_FUNCTION& {
64 Value = Info::updatePointer(Value, PtrVal);
65 }
66
67 void setInt(IntType IntVal) LLVM_LVALUE_FUNCTION& {
68 Value = Info::updateInt(Value, static_cast<intptr_t>(IntVal));
69 }
70
71 void initWithPointer(PointerTy PtrVal) LLVM_LVALUE_FUNCTION& {
72 Value = Info::updatePointer(0, PtrVal);
73 }
74
75 void setPointerAndInt(PointerTy PtrVal, IntType IntVal) LLVM_LVALUE_FUNCTION& {
76 Value = Info::updateInt(Info::updatePointer(0, PtrVal),
31
Calling 'PointerIntPairInfo::updateInt'
77 static_cast<intptr_t>(IntVal));
30
Passing value via 2nd parameter 'Int'
78 }
79
80 PointerTy const *getAddrOfPointer() const {
81 return const_cast<PointerIntPair *>(this)->getAddrOfPointer();
82 }
83
84 PointerTy *getAddrOfPointer() {
85 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", 87, __extension__
__PRETTY_FUNCTION__))
86 "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", 87, __extension__
__PRETTY_FUNCTION__))
87 "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", 87, __extension__
__PRETTY_FUNCTION__))
;
88 return reinterpret_cast<PointerTy *>(&Value);
89 }
90
91 void *getOpaqueValue() const { return reinterpret_cast<void *>(Value); }
92
93 void setFromOpaqueValue(void *Val) LLVM_LVALUE_FUNCTION& {
94 Value = reinterpret_cast<intptr_t>(Val);
95 }
96
97 static PointerIntPair getFromOpaqueValue(void *V) {
98 PointerIntPair P;
99 P.setFromOpaqueValue(V);
100 return P;
101 }
102
103 // Allow PointerIntPairs to be created from const void * if and only if the
104 // pointer type could be created from a const void *.
105 static PointerIntPair getFromOpaqueValue(const void *V) {
106 (void)PtrTraits::getFromVoidPointer(V);
107 return getFromOpaqueValue(const_cast<void *>(V));
108 }
109
110 bool operator==(const PointerIntPair &RHS) const {
111 return Value == RHS.Value;
112 }
113
114 bool operator!=(const PointerIntPair &RHS) const {
115 return Value != RHS.Value;
116 }
117
118 bool operator<(const PointerIntPair &RHS) const { return Value < RHS.Value; }
119 bool operator>(const PointerIntPair &RHS) const { return Value > RHS.Value; }
120
121 bool operator<=(const PointerIntPair &RHS) const {
122 return Value <= RHS.Value;
123 }
124
125 bool operator>=(const PointerIntPair &RHS) const {
126 return Value >= RHS.Value;
127 }
128};
129
130// Specialize is_trivially_copyable to avoid limitation of llvm::is_trivially_copyable
131// when compiled with gcc 4.9.
132template <typename PointerTy, unsigned IntBits, typename IntType,
133 typename PtrTraits,
134 typename Info>
135struct is_trivially_copyable<PointerIntPair<PointerTy, IntBits, IntType, PtrTraits, Info>> : std::true_type {
136#ifdef HAVE_STD_IS_TRIVIALLY_COPYABLE
137 static_assert(std::is_trivially_copyable<PointerIntPair<PointerTy, IntBits, IntType, PtrTraits, Info>>::value,
138 "inconsistent behavior between llvm:: and std:: implementation of is_trivially_copyable");
139#endif
140};
141
142
143template <typename PointerT, unsigned IntBits, typename PtrTraits>
144struct PointerIntPairInfo {
145 static_assert(PtrTraits::NumLowBitsAvailable <
146 std::numeric_limits<uintptr_t>::digits,
147 "cannot use a pointer type that has all bits free");
148 static_assert(IntBits <= PtrTraits::NumLowBitsAvailable,
149 "PointerIntPair with integer size too large for pointer");
150 enum MaskAndShiftConstants : uintptr_t {
151 /// PointerBitMask - The bits that come from the pointer.
152 PointerBitMask =
153 ~(uintptr_t)(((intptr_t)1 << PtrTraits::NumLowBitsAvailable) - 1),
154
155 /// IntShift - The number of low bits that we reserve for other uses, and
156 /// keep zero.
157 IntShift = (uintptr_t)PtrTraits::NumLowBitsAvailable - IntBits,
158
159 /// IntMask - This is the unshifted mask for valid bits of the int type.
160 IntMask = (uintptr_t)(((intptr_t)1 << IntBits) - 1),
161
162 // ShiftedIntMask - This is the bits for the integer shifted in place.
163 ShiftedIntMask = (uintptr_t)(IntMask << IntShift)
164 };
165
166 static PointerT getPointer(intptr_t Value) {
167 return PtrTraits::getFromVoidPointer(
168 reinterpret_cast<void *>(Value & PointerBitMask));
169 }
170
171 static intptr_t getInt(intptr_t Value) {
172 return (Value >> IntShift) & IntMask;
173 }
174
175 static intptr_t updatePointer(intptr_t OrigValue, PointerT Ptr) {
176 intptr_t PtrWord =
177 reinterpret_cast<intptr_t>(PtrTraits::getAsVoidPointer(Ptr));
178 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", 179, __extension__
__PRETTY_FUNCTION__))
179 "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", 179, __extension__
__PRETTY_FUNCTION__))
;
180 // Preserve all low bits, just update the pointer.
181 return PtrWord | (OrigValue & ~PointerBitMask);
182 }
183
184 static intptr_t updateInt(intptr_t OrigValue, intptr_t Int) {
185 intptr_t IntWord = static_cast<intptr_t>(Int);
32
'IntWord' initialized to the value of 'Int'
186 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", 186, __extension__
__PRETTY_FUNCTION__))
;
33
'?' condition is true
187
188 // Preserve all bits other than the ones we are updating.
189 return (OrigValue & ~ShiftedIntMask) | IntWord << IntShift;
34
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'
190 }
191};
192
193// Provide specialization of DenseMapInfo for PointerIntPair.
194template <typename PointerTy, unsigned IntBits, typename IntType>
195struct DenseMapInfo<PointerIntPair<PointerTy, IntBits, IntType>, void> {
196 using Ty = PointerIntPair<PointerTy, IntBits, IntType>;
197
198 static Ty getEmptyKey() {
199 uintptr_t Val = static_cast<uintptr_t>(-1);
200 Val <<= PointerLikeTypeTraits<Ty>::NumLowBitsAvailable;
201 return Ty::getFromOpaqueValue(reinterpret_cast<void *>(Val));
202 }
203
204 static Ty getTombstoneKey() {
205 uintptr_t Val = static_cast<uintptr_t>(-2);
206 Val <<= PointerLikeTypeTraits<PointerTy>::NumLowBitsAvailable;
207 return Ty::getFromOpaqueValue(reinterpret_cast<void *>(Val));
208 }
209
210 static unsigned getHashValue(Ty V) {
211 uintptr_t IV = reinterpret_cast<uintptr_t>(V.getOpaqueValue());
212 return unsigned(IV) ^ unsigned(IV >> 9);
213 }
214
215 static bool isEqual(const Ty &LHS, const Ty &RHS) { return LHS == RHS; }
216};
217
218// Teach SmallPtrSet that PointerIntPair is "basically a pointer".
219template <typename PointerTy, unsigned IntBits, typename IntType,
220 typename PtrTraits>
221struct PointerLikeTypeTraits<
222 PointerIntPair<PointerTy, IntBits, IntType, PtrTraits>> {
223 static inline void *
224 getAsVoidPointer(const PointerIntPair<PointerTy, IntBits, IntType> &P) {
225 return P.getOpaqueValue();
226 }
227
228 static inline PointerIntPair<PointerTy, IntBits, IntType>
229 getFromVoidPointer(void *P) {
230 return PointerIntPair<PointerTy, IntBits, IntType>::getFromOpaqueValue(P);
231 }
232
233 static inline PointerIntPair<PointerTy, IntBits, IntType>
234 getFromVoidPointer(const void *P) {
235 return PointerIntPair<PointerTy, IntBits, IntType>::getFromOpaqueValue(P);
236 }
237
238 static constexpr int NumLowBitsAvailable =
239 PtrTraits::NumLowBitsAvailable - IntBits;
240};
241
242} // end namespace llvm
243
244#endif // LLVM_ADT_POINTERINTPAIR_H