71#define DEBUG_TYPE "loop-accesses"
75 cl::desc(
"Sets the SIMD width. Zero is autoselect."),
81 cl::desc(
"Sets the vectorization interleave count. "
82 "Zero is autoselect."),
89 cl::desc(
"When performing memory disambiguation checks at runtime do not "
90 "generate more than this number of comparisons (default = 8)."),
97 cl::desc(
"Maximum number of comparisons done when trying to merge "
98 "runtime memory checks. (default = 100)"),
107 cl::desc(
"Maximum number of dependences collected by "
108 "loop-access analysis (default = 100)"),
124 cl::desc(
"Enable symbolic stride memory access versioning"));
129 "store-to-load-forwarding-conflict-detection",
cl::Hidden,
130 cl::desc(
"Enable conflict detection in loop-access analysis"),
135 cl::desc(
"Maximum recursion depth when finding forked SCEVs (default = 5)"),
140 cl::desc(
"Speculate that non-constant strides are unit in LAA"),
146 "Hoist inner loop runtime memory checks to outer loop if possible"),
151 return ::VectorizationInterleave.getNumOccurrences() > 0;
162 if (SI == PtrToStride.
end())
166 const SCEV *StrideSCEV = SI->second;
171 assert(isa<SCEVUnknown>(StrideSCEV) &&
"shouldn't be in map");
179 <<
" by: " << *Expr <<
"\n");
189 NeedsFreeze(RtCheck.Pointers[
Index].NeedsFreeze) {
209 DenseMap<std::pair<const SCEV *, Type *>,
214 {{PtrExpr, AccessTy},
223 ScStart = ScEnd = PtrExpr;
224 }
else if (
auto *AR = dyn_cast<SCEVAddRecExpr>(PtrExpr)) {
227 ScStart = AR->getStart();
228 ScEnd = AR->evaluateAtIteration(Ex, *SE);
229 const SCEV *Step = AR->getStepRecurrence(*SE);
233 if (
const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
234 if (CStep->getValue()->isNegative())
255 Iter->second = {ScStart, ScEnd};
262 Type *AccessTy,
bool WritePtr,
263 unsigned DepSetId,
unsigned ASId,
268 assert(!isa<SCEVCouldNotCompute>(ScStart) &&
269 !isa<SCEVCouldNotCompute>(ScEnd) &&
270 "must be able to compute both start and end expressions");
271 Pointers.emplace_back(
Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, PtrExpr,
275bool RuntimePointerChecking::tryToCreateDiffCheck(
298 if (AccSrc.
size() != 1 || AccSink.
size() != 1)
302 if (AccSink[0] < AccSrc[0])
305 auto *SrcAR = dyn_cast<SCEVAddRecExpr>(Src->Expr);
306 auto *SinkAR = dyn_cast<SCEVAddRecExpr>(Sink->Expr);
317 if (isa<ScalableVectorType>(SrcTy) || isa<ScalableVectorType>(DstTy))
321 SinkAR->getLoop()->getHeader()->getDataLayout();
323 std::max(
DL.getTypeAllocSize(SrcTy),
DL.getTypeAllocSize(DstTy));
328 auto *Step = dyn_cast<SCEVConstant>(SinkAR->getStepRecurrence(*SE));
329 if (!Step || Step != SrcAR->getStepRecurrence(*SE) ||
330 Step->getAPInt().abs() != AllocSize)
338 if (Step->getValue()->isNegative())
343 if (isa<SCEVCouldNotCompute>(SinkStartInt) ||
344 isa<SCEVCouldNotCompute>(SrcStartInt))
347 const Loop *InnerLoop = SrcAR->getLoop();
353 isa<SCEVAddRecExpr>(SinkStartInt) && isa<SCEVAddRecExpr>(SrcStartInt)) {
354 auto *SrcStartAR = cast<SCEVAddRecExpr>(SrcStartInt);
355 auto *SinkStartAR = cast<SCEVAddRecExpr>(SinkStartInt);
356 const Loop *StartARLoop = SrcStartAR->getLoop();
357 if (StartARLoop == SinkStartAR->getLoop() &&
362 SrcStartAR->getStepRecurrence(*SE) !=
363 SinkStartAR->getStepRecurrence(*SE)) {
364 LLVM_DEBUG(
dbgs() <<
"LAA: Not creating diff runtime check, since these "
365 "cannot be hoisted out of the outer loop\n");
371 <<
"SrcStart: " << *SrcStartInt <<
'\n'
372 <<
"SinkStartInt: " << *SinkStartInt <<
'\n');
373 DiffChecks.emplace_back(SrcStartInt, SinkStartInt, AllocSize,
374 Src->NeedsFreeze ||
Sink->NeedsFreeze);
387 CanUseDiffCheck = CanUseDiffCheck && tryToCreateDiffCheck(CGI, CGJ);
388 Checks.
push_back(std::make_pair(&CGI, &CGJ));
395void RuntimePointerChecking::generateChecks(
398 groupChecks(DepCands, UseDependencies);
404 for (
const auto &
I : M.Members)
405 for (
const auto &J :
N.Members)
420 return C->getValue()->isNegative() ? J :
I;
427 RtCheck.
Pointers[
Index].PointerValue->getType()->getPointerAddressSpace(),
436 "all pointers in a checking group must be in the same address space");
462void RuntimePointerChecking::groupChecks(
508 if (!UseDependencies) {
514 unsigned TotalComparisons = 0;
519 It->second.push_back(
Index);
548 auto PointerI = PositionMap.
find(
MI->getPointer());
550 "pointer in equivalence class not found in PositionMap");
551 for (
unsigned Pointer : PointerI->second) {
568 if (Group.addPointer(Pointer, *
this)) {
591 return (PtrToPartition[PtrIdx1] != -1 &&
592 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
616 unsigned Depth)
const {
618 for (
const auto &[Check1, Check2] : Checks) {
619 const auto &
First = Check1->Members, &Second = Check2->Members;
623 OS.
indent(
Depth + 2) <<
"Comparing group (" << Check1 <<
"):\n";
624 for (
unsigned K :
First)
627 OS.
indent(
Depth + 2) <<
"Against group (" << Check2 <<
"):\n";
628 for (
unsigned K : Second)
641 OS.
indent(
Depth + 4) <<
"(Low: " << *CG.Low <<
" High: " << *CG.High
643 for (
unsigned Member : CG.Members) {
655class AccessAnalysis {
665 : TheLoop(TheLoop), BAA(*AA), AST(BAA), LI(LI), DepCands(DA), PSE(PSE),
666 LoopAliasScopes(LoopAliasScopes) {
668 BAA.enableCrossIterationMode();
674 AST.add(adjustLoc(Loc));
675 Accesses[MemAccessInfo(
Ptr,
false)].insert(AccessTy);
677 ReadOnlyPtr.insert(
Ptr);
683 AST.add(adjustLoc(Loc));
684 Accesses[MemAccessInfo(
Ptr,
true)].insert(AccessTy);
695 MemAccessInfo Access,
Type *AccessTy,
698 Loop *TheLoop,
unsigned &RunningDepId,
699 unsigned ASId,
bool ShouldCheckStride,
bool Assume);
708 Value *&UncomputablePtr,
bool ShouldCheckWrap =
false);
712 void buildDependenceSets() {
713 processMemAccesses();
721 bool isDependencyCheckNeeded() {
return !CheckDeps.empty(); }
729 MemAccessInfoList &getDependenciesToCheck() {
return CheckDeps; }
733 return UnderlyingObjects;
758 return LoopAliasScopes.contains(cast<MDNode>(Scope));
767 void processMemAccesses();
771 PtrAccessMap Accesses;
777 MemAccessInfoList CheckDeps;
803 bool IsRTCheckAnalysisNeeded =
false;
821 const SCEV *PtrScev,
Loop *L,
bool Assume) {
845 int64_t Stride =
getPtrStride(PSE, AccessTy,
Ptr, L, Strides).value_or(0);
858 while (!WorkList.
empty()) {
862 auto *PN = dyn_cast<PHINode>(
Ptr);
866 if (PN && InnermostLoop.
contains(PN->getParent()) &&
867 PN->getParent() != InnermostLoop.
getHeader()) {
868 for (
const Use &Inc : PN->incoming_values())
901 if (isa<SCEVAddRecExpr>(Scev) || L->isLoopInvariant(
Ptr) ||
902 !isa<Instruction>(
Ptr) ||
Depth == 0) {
913 auto GetBinOpExpr = [&SE](
unsigned Opcode,
const SCEV *L,
const SCEV *R) {
915 case Instruction::Add:
917 case Instruction::Sub:
925 unsigned Opcode =
I->getOpcode();
927 case Instruction::GetElementPtr: {
929 Type *SourceTy =
GEP->getSourceElementType();
932 if (
I->getNumOperands() != 2 || SourceTy->
isVectorTy()) {
942 bool NeedsFreeze =
any_of(BaseScevs, UndefPoisonCheck) ||
943 any_of(OffsetScevs, UndefPoisonCheck);
948 if (OffsetScevs.
size() == 2 && BaseScevs.
size() == 1)
950 else if (BaseScevs.
size() == 2 && OffsetScevs.
size() == 1)
953 ScevList.emplace_back(Scev, NeedsFreeze);
971 ScevList.emplace_back(SE->
getAddExpr(get<0>(BaseScevs[0]), Scaled1),
973 ScevList.emplace_back(SE->
getAddExpr(get<0>(BaseScevs[1]), Scaled2),
977 case Instruction::Select: {
984 if (ChildScevs.
size() == 2) {
985 ScevList.push_back(ChildScevs[0]);
986 ScevList.push_back(ChildScevs[1]);
991 case Instruction::PHI: {
996 if (
I->getNumOperands() == 2) {
1000 if (ChildScevs.
size() == 2) {
1001 ScevList.push_back(ChildScevs[0]);
1002 ScevList.push_back(ChildScevs[1]);
1007 case Instruction::Add:
1008 case Instruction::Sub: {
1016 any_of(LScevs, UndefPoisonCheck) ||
any_of(RScevs, UndefPoisonCheck);
1021 if (LScevs.
size() == 2 && RScevs.
size() == 1)
1023 else if (RScevs.
size() == 2 && LScevs.
size() == 1)
1026 ScevList.emplace_back(Scev, NeedsFreeze);
1030 ScevList.emplace_back(
1031 GetBinOpExpr(Opcode, get<0>(LScevs[0]), get<0>(RScevs[0])),
1033 ScevList.emplace_back(
1034 GetBinOpExpr(Opcode, get<0>(LScevs[1]), get<0>(RScevs[1])),
1040 LLVM_DEBUG(
dbgs() <<
"ForkedPtr unhandled instruction: " << *
I <<
"\n");
1057 if (Scevs.
size() == 2 &&
1058 (isa<SCEVAddRecExpr>(get<0>(Scevs[0])) ||
1060 (isa<SCEVAddRecExpr>(get<0>(Scevs[1])) ||
1072 MemAccessInfo Access,
Type *AccessTy,
1075 Loop *TheLoop,
unsigned &RunningDepId,
1076 unsigned ASId,
bool ShouldCheckWrap,
1083 for (
auto &
P : TranslatedPtrs) {
1084 const SCEV *PtrExpr = get<0>(
P);
1090 if (ShouldCheckWrap) {
1092 if (TranslatedPtrs.size() > 1)
1095 if (!
isNoWrap(PSE, StridesMap,
Ptr, AccessTy, TheLoop)) {
1097 if (!Assume || !isa<SCEVAddRecExpr>(Expr))
1104 if (TranslatedPtrs.size() == 1)
1109 for (
auto [PtrExpr, NeedsFreeze] : TranslatedPtrs) {
1113 if (isDependencyCheckNeeded()) {
1115 unsigned &LeaderId = DepSetId[Leader];
1117 LeaderId = RunningDepId++;
1121 DepId = RunningDepId++;
1123 bool IsWrite = Access.getInt();
1124 RtCheck.
insert(TheLoop,
Ptr, PtrExpr, AccessTy, IsWrite, DepId, ASId, PSE,
1135 Value *&UncomputablePtr,
bool ShouldCheckWrap) {
1138 bool CanDoRT =
true;
1140 bool MayNeedRTCheck =
false;
1141 if (!IsRTCheckAnalysisNeeded)
return true;
1143 bool IsDepCheckNeeded = isDependencyCheckNeeded();
1148 for (
auto &AS : AST) {
1149 int NumReadPtrChecks = 0;
1150 int NumWritePtrChecks = 0;
1151 bool CanDoAliasSetRT =
true;
1153 auto ASPointers = AS.getPointers();
1157 unsigned RunningDepId = 1;
1165 for (
const Value *ConstPtr : ASPointers) {
1167 bool IsWrite = Accesses.count(MemAccessInfo(
Ptr,
true));
1169 ++NumWritePtrChecks;
1177 if (NumWritePtrChecks == 0 ||
1178 (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) {
1179 assert((ASPointers.size() <= 1 ||
1182 MemAccessInfo AccessWrite(
const_cast<Value *
>(
Ptr),
1184 return DepCands.
findValue(AccessWrite) == DepCands.
end();
1186 "Can only skip updating CanDoRT below, if all entries in AS "
1187 "are reads or there is at most 1 entry");
1191 for (
auto &Access : AccessInfos) {
1192 for (
const auto &AccessTy : Accesses[Access]) {
1193 if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap,
1194 DepSetId, TheLoop, RunningDepId, ASId,
1195 ShouldCheckWrap,
false)) {
1197 << *Access.getPointer() <<
'\n');
1199 CanDoAliasSetRT =
false;
1213 bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.
empty();
1217 if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) {
1221 CanDoAliasSetRT =
true;
1222 for (
const auto &[Access, AccessTy] : Retries) {
1223 if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap,
1224 DepSetId, TheLoop, RunningDepId, ASId,
1225 ShouldCheckWrap,
true)) {
1226 CanDoAliasSetRT =
false;
1227 UncomputablePtr = Access.getPointer();
1233 CanDoRT &= CanDoAliasSetRT;
1234 MayNeedRTCheck |= NeedsAliasSetRTCheck;
1243 unsigned NumPointers = RtCheck.
Pointers.size();
1244 for (
unsigned i = 0; i < NumPointers; ++i) {
1245 for (
unsigned j = i + 1;
j < NumPointers; ++
j) {
1247 if (RtCheck.
Pointers[i].DependencySetId ==
1248 RtCheck.
Pointers[j].DependencySetId)
1261 dbgs() <<
"LAA: Runtime check would require comparison between"
1262 " different address spaces\n");
1268 if (MayNeedRTCheck && CanDoRT)
1272 <<
" pointer comparisons.\n");
1279 bool CanDoRTIfNeeded = !RtCheck.
Need || CanDoRT;
1280 if (!CanDoRTIfNeeded)
1282 return CanDoRTIfNeeded;
1285void AccessAnalysis::processMemAccesses() {
1292 LLVM_DEBUG(
dbgs() <<
"LAA: Accesses(" << Accesses.size() <<
"):\n");
1294 for (
const auto &[
A,
_] : Accesses)
1295 dbgs() <<
"\t" << *
A.getPointer() <<
" ("
1296 << (
A.getInt() ?
"write"
1297 : (ReadOnlyPtr.count(
A.getPointer()) ?
"read-only"
1306 for (
const auto &AS : AST) {
1310 auto ASPointers = AS.getPointers();
1312 bool SetHasWrite =
false;
1316 UnderlyingObjToAccessMap ObjToLastAccess;
1319 PtrAccessMap DeferredAccesses;
1323 for (
int SetIteration = 0; SetIteration < 2; ++SetIteration) {
1324 bool UseDeferred = SetIteration > 0;
1325 PtrAccessMap &S = UseDeferred ? DeferredAccesses : Accesses;
1327 for (
const Value *ConstPtr : ASPointers) {
1332 for (
const auto &[AC,
_] : S) {
1333 if (AC.getPointer() !=
Ptr)
1336 bool IsWrite = AC.getInt();
1340 bool IsReadOnlyPtr = ReadOnlyPtr.count(
Ptr) && !IsWrite;
1341 if (UseDeferred && !IsReadOnlyPtr)
1345 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
1346 S.count(MemAccessInfo(
Ptr,
false))) &&
1347 "Alias-set pointer not in the access set?");
1349 MemAccessInfo Access(
Ptr, IsWrite);
1357 if (!UseDeferred && IsReadOnlyPtr) {
1360 DeferredAccesses.insert({Access, {}});
1368 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
1369 CheckDeps.push_back(Access);
1370 IsRTCheckAnalysisNeeded =
true;
1379 ValueVector TempObjects;
1381 UnderlyingObjects[
Ptr] = {};
1385 <<
"Underlying objects for pointer " << *
Ptr <<
"\n");
1386 for (
const Value *UnderlyingObj : UOs) {
1389 if (isa<ConstantPointerNull>(UnderlyingObj) &&
1395 UnderlyingObjToAccessMap::iterator Prev =
1396 ObjToLastAccess.find(UnderlyingObj);
1397 if (Prev != ObjToLastAccess.end())
1398 DepCands.
unionSets(Access, Prev->second);
1400 ObjToLastAccess[UnderlyingObj] = Access;
1429 auto *
GEP = dyn_cast<GetElementPtrInst>(
Ptr);
1430 if (!
GEP || !
GEP->isInBounds())
1434 Value *NonConstIndex =
nullptr;
1436 if (!isa<ConstantInt>(
Index)) {
1439 NonConstIndex =
Index;
1447 if (
auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
1448 if (OBO->hasNoSignedWrap() &&
1451 isa<ConstantInt>(OBO->getOperand(1))) {
1452 auto *OpScev = PSE.
getSCEV(OBO->getOperand(0));
1454 if (
auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
1455 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(
SCEV::FlagNSW);
1466 bool Assume,
bool ShouldCheckWrap) {
1470 if (isa<ScalableVectorType>(AccessTy)) {
1471 LLVM_DEBUG(
dbgs() <<
"LAA: Bad stride - Scalable object: " << *AccessTy
1473 return std::nullopt;
1484 <<
" SCEV: " << *PtrScev <<
"\n");
1485 return std::nullopt;
1490 LLVM_DEBUG(
dbgs() <<
"LAA: Bad stride - Not striding over innermost loop "
1491 << *
Ptr <<
" SCEV: " << *AR <<
"\n");
1492 return std::nullopt;
1502 <<
" SCEV: " << *AR <<
"\n");
1503 return std::nullopt;
1507 TypeSize AllocSize =
DL.getTypeAllocSize(AccessTy);
1509 const APInt &APStepVal =
C->getAPInt();
1513 return std::nullopt;
1518 int64_t Stride = StepVal /
Size;
1519 int64_t Rem = StepVal %
Size;
1521 return std::nullopt;
1523 if (!ShouldCheckWrap)
1535 if (
auto *
GEP = dyn_cast<GetElementPtrInst>(
Ptr);
1536 GEP &&
GEP->isInBounds() && (Stride == 1 || Stride == -1))
1544 (Stride == 1 || Stride == -1))
1550 <<
"LAA: Pointer: " << *
Ptr <<
"\n"
1551 <<
"LAA: SCEV: " << *AR <<
"\n"
1552 <<
"LAA: Added an overflow assumption\n");
1556 dbgs() <<
"LAA: Bad stride - Pointer may wrap in the address space "
1557 << *
Ptr <<
" SCEV: " << *AR <<
"\n");
1558 return std::nullopt;
1566 assert(PtrA && PtrB &&
"Expected non-nullptr pointers.");
1574 return std::nullopt;
1581 return std::nullopt;
1582 unsigned IdxWidth =
DL.getIndexSizeInBits(ASA);
1584 APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
1589 if (PtrA1 == PtrB1) {
1592 ASA = cast<PointerType>(PtrA1->
getType())->getAddressSpace();
1593 ASB = cast<PointerType>(PtrB1->
getType())->getAddressSpace();
1596 return std::nullopt;
1598 IdxWidth =
DL.getIndexSizeInBits(ASA);
1599 OffsetA = OffsetA.sextOrTrunc(IdxWidth);
1609 dyn_cast<SCEVConstant>(SE.
getMinusSCEV(PtrSCEVB, PtrSCEVA));
1611 return std::nullopt;
1612 Val = Diff->getAPInt().getSExtValue();
1614 int Size =
DL.getTypeStoreSize(ElemTyA);
1615 int Dist = Val /
Size;
1619 if (!StrictCheck || Dist *
Size == Val)
1621 return std::nullopt;
1629 "Expected list of pointer operands.");
1632 Value *Ptr0 = VL[0];
1634 using DistOrdPair = std::pair<int64_t, int>;
1636 std::set<DistOrdPair,
decltype(Compare)> Offsets(Compare);
1637 Offsets.emplace(0, 0);
1638 bool IsConsecutive =
true;
1647 auto [It, IsInserted] = Offsets.emplace(
Offset,
Idx);
1651 IsConsecutive &= std::next(It) == Offsets.end();
1653 SortedIndices.
clear();
1654 if (!IsConsecutive) {
1658 SortedIndices[
Idx] = Off.second;
1672 std::optional<int> Diff =
1675 return Diff && *Diff == 1;
1681 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
1682 InstMap.push_back(SI);
1690 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
1691 InstMap.push_back(LI);
1719 case ForwardButPreventsForwarding:
1721 case IndirectUnsafe:
1724 case BackwardVectorizable:
1726 case BackwardVectorizableButPreventsForwarding:
1739 case ForwardButPreventsForwarding:
1744 case BackwardVectorizable:
1746 case BackwardVectorizableButPreventsForwarding:
1747 case IndirectUnsafe:
1753bool MemoryDepChecker::couldPreventStoreLoadForward(
uint64_t Distance,
1767 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1769 uint64_t MaxVFWithoutSLForwardIssues = std::min(
1773 for (
uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1777 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1778 MaxVFWithoutSLForwardIssues = (VF >> 1);
1783 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1785 dbgs() <<
"LAA: Distance " << Distance
1786 <<
" that could cause a store-load forwarding conflict\n");
1790 if (MaxVFWithoutSLForwardIssues < MinDepDistBytes &&
1791 MaxVFWithoutSLForwardIssues !=
1793 MinDepDistBytes = MaxVFWithoutSLForwardIssues;
1815 const SCEV &MaxBTC,
const SCEV &Dist,
1836 const uint64_t ByteStride = MaxStride * TypeByteSize;
1840 const SCEV *CastedDist = &Dist;
1841 const SCEV *CastedProduct = Product;
1848 if (DistTypeSizeBits > ProductTypeSizeBits)
1873 assert(Stride > 1 &&
"The stride must be greater than 1");
1874 assert(TypeByteSize > 0 &&
"The type size in byte must be non-zero");
1875 assert(Distance > 0 &&
"The distance must be non-zero");
1878 if (Distance % TypeByteSize)
1881 uint64_t ScaledDist = Distance / TypeByteSize;
1899 return ScaledDist % Stride;
1907 return any_of(UnderlyingObjects, [&SE, L](
const Value *UO) {
1913 MemoryDepChecker::DepDistanceStrideAndSizeInfo>
1914MemoryDepChecker::getDependenceDistanceStrideAndSize(
1918 &UnderlyingObjects) {
1920 auto &SE = *PSE.
getSE();
1921 auto [APtr, AIsWrite] =
A;
1922 auto [BPtr, BIsWrite] =
B;
1925 if (!AIsWrite && !BIsWrite)
1932 if (APtr->getType()->getPointerAddressSpace() !=
1933 BPtr->getType()->getPointerAddressSpace())
1936 int64_t StrideAPtr =
1937 getPtrStride(PSE, ATy, APtr, InnermostLoop, SymbolicStrides,
true)
1939 int64_t StrideBPtr =
1940 getPtrStride(PSE, BTy, BPtr, InnermostLoop, SymbolicStrides,
true)
1949 if (StrideAPtr < 0) {
1956 LLVM_DEBUG(
dbgs() <<
"LAA: Src Scev: " << *Src <<
"Sink Scev: " << *Sink
1957 <<
"(Induction step: " << StrideAPtr <<
")\n");
1958 LLVM_DEBUG(
dbgs() <<
"LAA: Distance for " << *AInst <<
" to " << *BInst
1959 <<
": " << *Dist <<
"\n");
1975 const auto &[SrcStart, SrcEnd] =
1977 const auto &[SinkStart, SinkEnd] =
1979 if (!isa<SCEVCouldNotCompute>(SrcStart) &&
1980 !isa<SCEVCouldNotCompute>(SrcEnd) &&
1981 !isa<SCEVCouldNotCompute>(SinkStart) &&
1982 !isa<SCEVCouldNotCompute>(SinkEnd)) {
1993 if (!StrideAPtr || !StrideBPtr || (StrideAPtr > 0 && StrideBPtr < 0) ||
1994 (StrideAPtr < 0 && StrideBPtr > 0)) {
1995 LLVM_DEBUG(
dbgs() <<
"Pointer access with non-constant stride\n");
1999 uint64_t TypeByteSize =
DL.getTypeAllocSize(ATy);
2001 DL.getTypeStoreSizeInBits(ATy) ==
DL.getTypeStoreSizeInBits(BTy);
2004 return DepDistanceStrideAndSizeInfo(Dist, std::abs(StrideAPtr),
2005 std::abs(StrideBPtr), TypeByteSize,
2006 AIsWrite, BIsWrite);
2013 &UnderlyingObjects) {
2014 assert(AIdx < BIdx &&
"Must pass arguments in program order");
2018 auto Res = getDependenceDistanceStrideAndSize(
2019 A, InstMap[AIdx],
B, InstMap[BIdx], UnderlyingObjects);
2020 if (std::holds_alternative<Dependence::DepType>(Res))
2021 return std::get<Dependence::DepType>(Res);
2023 auto &[Dist, StrideA, StrideB, TypeByteSize, AIsWrite, BIsWrite] =
2024 std::get<DepDistanceStrideAndSizeInfo>(Res);
2025 bool HasSameSize = TypeByteSize > 0;
2027 std::optional<uint64_t> CommonStride =
2028 StrideA == StrideB ? std::make_optional(StrideA) :
std::nullopt;
2029 if (isa<SCEVCouldNotCompute>(Dist)) {
2032 FoundNonConstantDistanceDependence |= CommonStride.has_value();
2033 LLVM_DEBUG(
dbgs() <<
"LAA: Dependence because of uncomputable distance.\n");
2039 uint64_t MaxStride = std::max(StrideA, StrideB);
2048 *Dist, MaxStride, TypeByteSize))
2055 const APInt &Val =
C->getAPInt();
2060 if (std::abs(Distance) > 0 && CommonStride && *CommonStride > 1 &&
2077 LLVM_DEBUG(
dbgs() <<
"LAA: possibly zero dependence difference but "
2078 "different type sizes\n");
2082 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
2097 FoundNonConstantDistanceDependence |= CommonStride.has_value();
2101 couldPreventStoreLoadForward(
C->getAPInt().abs().getZExtValue(),
2104 dbgs() <<
"LAA: Forward but may prevent st->ld forwarding\n");
2115 if (MinDistance <= 0) {
2116 FoundNonConstantDistanceDependence |= CommonStride.has_value();
2120 if (!isa<SCEVConstant>(Dist)) {
2129 FoundNonConstantDistanceDependence |= CommonStride.has_value();
2133 LLVM_DEBUG(
dbgs() <<
"LAA: ReadWrite-Write positive dependency with "
2134 "different type sizes\n");
2147 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
2180 TypeByteSize * *CommonStride * (MinNumIter - 1) + TypeByteSize;
2181 if (MinDistanceNeeded >
static_cast<uint64_t>(MinDistance)) {
2182 if (!isa<SCEVConstant>(Dist)) {
2189 LLVM_DEBUG(
dbgs() <<
"LAA: Failure because of positive minimum distance "
2190 << MinDistance <<
'\n');
2196 if (MinDistanceNeeded > MinDepDistBytes) {
2198 << MinDistanceNeeded <<
" size in bytes\n");
2219 std::min(
static_cast<uint64_t>(MinDistance), MinDepDistBytes);
2221 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
2222 uint64_t MinDepDistBytesOld = MinDepDistBytes;
2224 isa<SCEVConstant>(Dist) &&
2225 couldPreventStoreLoadForward(MinDistance, TypeByteSize)) {
2228 assert(MinDepDistBytes == MinDepDistBytesOld &&
2229 "An update to MinDepDistBytes requires an update to "
2230 "MaxSafeVectorWidthInBits");
2231 (void)MinDepDistBytesOld;
2237 uint64_t MaxVF = MinDepDistBytes / (TypeByteSize * *CommonStride);
2238 LLVM_DEBUG(
dbgs() <<
"LAA: Positive min distance " << MinDistance
2239 <<
" with max VF = " << MaxVF <<
'\n');
2241 uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
2242 if (!isa<SCEVConstant>(Dist) && MaxVFInBits < MaxTargetVectorWidthInBits) {
2249 MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits);
2256 &UnderlyingObjects) {
2258 MinDepDistBytes = -1;
2261 if (Visited.
count(CurAccess))
2277 bool AIIsWrite = AI->getInt();
2281 (AIIsWrite ? AI : std::next(AI));
2284 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
2285 I1E = Accesses[*AI].
end(); I1 != I1E; ++I1)
2288 for (std::vector<unsigned>::iterator
2289 I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()),
2290 I2E = (OI == AI ? I1E : Accesses[*OI].end());
2292 auto A = std::make_pair(&*AI, *I1);
2293 auto B = std::make_pair(&*OI, *I2);
2300 B.second, UnderlyingObjects);
2307 if (RecordDependences) {
2312 RecordDependences =
false;
2313 Dependences.clear();
2315 <<
"Too many dependences, stopped recording\n");
2327 LLVM_DEBUG(
dbgs() <<
"Total Dependences: " << Dependences.size() <<
"\n");
2334 auto &IndexVector = Accesses.find(Access)->second;
2338 std::back_inserter(Insts),
2339 [&](
unsigned Idx) {
return this->InstMap[
Idx]; });
2348 "ForwardButPreventsForwarding",
2350 "BackwardVectorizable",
2351 "BackwardVectorizableButPreventsForwarding"};
2361bool LoopAccessInfo::canAnalyzeLoop() {
2370 recordAnalysis(
"NotInnerMostLoop") <<
"loop is not the innermost loop";
2377 dbgs() <<
"LAA: loop control flow is not understood by analyzer\n");
2378 recordAnalysis(
"CFGNotUnderstood")
2379 <<
"loop control flow is not understood by analyzer";
2387 if (isa<SCEVCouldNotCompute>(ExitCount)) {
2388 recordAnalysis(
"CantComputeNumberOfIterations")
2389 <<
"could not determine number of loop iterations";
2390 LLVM_DEBUG(
dbgs() <<
"LAA: SCEV could not compute the loop exit count.\n");
2408 unsigned NumReads = 0;
2409 unsigned NumReadWrites = 0;
2411 bool HasComplexMemInst =
false;
2414 HasConvergentOp =
false;
2416 PtrRtChecking->Pointers.
clear();
2417 PtrRtChecking->Need =
false;
2421 const bool EnableMemAccessVersioningOfLoop =
2433 if (
auto *Call = dyn_cast<CallBase>(&
I)) {
2434 if (
Call->isConvergent())
2435 HasConvergentOp =
true;
2440 if (HasComplexMemInst && HasConvergentOp)
2444 if (HasComplexMemInst)
2448 if (
auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&
I))
2449 for (
Metadata *
Op : Decl->getScopeList()->operands())
2450 LoopAliasScopes.
insert(cast<MDNode>(
Op));
2455 auto *
Call = dyn_cast<CallInst>(&
I);
2462 if (
I.mayReadFromMemory()) {
2465 if (Call && !
Call->isNoBuiltin() &&
Call->getCalledFunction() &&
2469 auto *Ld = dyn_cast<LoadInst>(&
I);
2471 recordAnalysis(
"CantVectorizeInstruction", Ld)
2472 <<
"instruction cannot be vectorized";
2473 HasComplexMemInst =
true;
2476 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2477 recordAnalysis(
"NonSimpleLoad", Ld)
2478 <<
"read with atomic ordering or volatile read";
2480 HasComplexMemInst =
true;
2486 if (EnableMemAccessVersioningOfLoop)
2487 collectStridedAccess(Ld);
2492 if (
I.mayWriteToMemory()) {
2493 auto *St = dyn_cast<StoreInst>(&
I);
2495 recordAnalysis(
"CantVectorizeInstruction", St)
2496 <<
"instruction cannot be vectorized";
2497 HasComplexMemInst =
true;
2500 if (!St->isSimple() && !IsAnnotatedParallel) {
2501 recordAnalysis(
"NonSimpleStore", St)
2502 <<
"write with atomic ordering or volatile write";
2504 HasComplexMemInst =
true;
2510 if (EnableMemAccessVersioningOfLoop)
2511 collectStridedAccess(St);
2516 if (HasComplexMemInst)
2524 if (!Stores.
size()) {
2530 AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE,
2547 if (isInvariant(
Ptr)) {
2549 StoresToInvariantAddresses.push_back(ST);
2550 HasStoreStoreDependenceInvolvingLoopInvariantAddress |=
2557 if (Seen.
insert({Ptr, AccessTy}).second) {
2564 if (blockNeedsPredication(
ST->getParent(), TheLoop, DT))
2568 [&Accesses, AccessTy, Loc](
Value *
Ptr) {
2569 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2570 Accesses.addStore(NewLoc, AccessTy);
2575 if (IsAnnotatedParallel) {
2577 dbgs() <<
"LAA: A loop annotated parallel, ignore memory dependency "
2592 bool IsReadOnlyPtr =
false;
2594 if (Seen.
insert({Ptr, AccessTy}).second ||
2595 !
getPtrStride(*PSE,
LD->getType(),
Ptr, TheLoop, SymbolicStrides).value_or(0)) {
2597 IsReadOnlyPtr =
true;
2603 LLVM_DEBUG(
dbgs() <<
"LAA: Found an unsafe dependency between a uniform "
2604 "load and uniform store to the same address!\n");
2605 HasLoadStoreDependenceInvolvingLoopInvariantAddress =
true;
2612 if (blockNeedsPredication(
LD->getParent(), TheLoop, DT))
2616 [&Accesses, AccessTy, Loc, IsReadOnlyPtr](
Value *
Ptr) {
2617 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2618 Accesses.addLoad(NewLoc, AccessTy, IsReadOnlyPtr);
2624 if (NumReadWrites == 1 && NumReads == 0) {
2631 Accesses.buildDependenceSets();
2635 Value *UncomputablePtr =
nullptr;
2636 bool CanDoRTIfNeeded =
2637 Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->
getSE(), TheLoop,
2638 SymbolicStrides, UncomputablePtr,
false);
2639 if (!CanDoRTIfNeeded) {
2640 auto *
I = dyn_cast_or_null<Instruction>(UncomputablePtr);
2641 recordAnalysis(
"CantIdentifyArrayBounds",
I)
2642 <<
"cannot identify array bounds";
2643 LLVM_DEBUG(
dbgs() <<
"LAA: We can't vectorize because we can't find "
2644 <<
"the array bounds.\n");
2649 dbgs() <<
"LAA: May be able to perform a memory runtime check if needed.\n");
2651 bool DepsAreSafe =
true;
2652 if (Accesses.isDependencyCheckNeeded()) {
2654 DepsAreSafe = DepChecker->
areDepsSafe(DependentAccesses,
2655 Accesses.getDependenciesToCheck(),
2656 Accesses.getUnderlyingObjects());
2662 Accesses.resetDepChecks(*DepChecker);
2664 PtrRtChecking->reset();
2665 PtrRtChecking->Need =
true;
2667 auto *SE = PSE->
getSE();
2668 UncomputablePtr =
nullptr;
2669 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(
2670 *PtrRtChecking, SE, TheLoop, SymbolicStrides, UncomputablePtr,
true);
2673 if (!CanDoRTIfNeeded) {
2674 auto *
I = dyn_cast_or_null<Instruction>(UncomputablePtr);
2675 recordAnalysis(
"CantCheckMemDepsAtRunTime",
I)
2676 <<
"cannot check memory dependencies at runtime";
2677 LLVM_DEBUG(
dbgs() <<
"LAA: Can't vectorize with memory checks\n");
2684 if (HasConvergentOp) {
2685 recordAnalysis(
"CantInsertRuntimeCheckWithConvergent")
2686 <<
"cannot add control dependency to convergent operation";
2687 LLVM_DEBUG(
dbgs() <<
"LAA: We can't vectorize because a runtime check "
2688 "would be needed with a convergent operation\n");
2694 dbgs() <<
"LAA: No unsafe dependent memory operations in loop. We"
2695 << (PtrRtChecking->Need ?
"" :
" don't")
2696 <<
" need runtime memory checks.\n");
2700 emitUnsafeDependenceRemark();
2704void LoopAccessInfo::emitUnsafeDependenceRemark() {
2705 const auto *Deps = getDepChecker().getDependences();
2713 if (Found == Deps->end())
2717 LLVM_DEBUG(
dbgs() <<
"LAA: unsafe dependent memory operations in loop\n");
2720 bool HasForcedDistribution =
false;
2721 std::optional<const MDOperand *>
Value =
2725 assert(
Op && mdconst::hasa<ConstantInt>(*
Op) &&
"invalid metadata");
2726 HasForcedDistribution = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
2729 const std::string
Info =
2730 HasForcedDistribution
2731 ?
"unsafe dependent memory operations in loop."
2732 :
"unsafe dependent memory operations in loop. Use "
2733 "#pragma clang loop distribute(enable) to allow loop distribution "
2734 "to attempt to isolate the offending operations into a separate "
2745 R <<
"\nBackward loop carried data dependence.";
2748 R <<
"\nForward loop carried data dependence that prevents "
2749 "store-to-load forwarding.";
2752 R <<
"\nBackward loop carried data dependence that prevents "
2753 "store-to-load forwarding.";
2756 R <<
"\nUnsafe indirect dependence.";
2759 R <<
"\nUnknown data dependence.";
2766 SourceLoc = DD->getDebugLoc();
2768 R <<
" Memory location is the same as accessed at "
2769 <<
ore::NV(
"Location", SourceLoc);
2784 assert(!Report &&
"Multiple reports generated");
2790 CodeRegion =
I->getParent();
2793 if (
I->getDebugLoc())
2794 DL =
I->getDebugLoc();
2797 Report = std::make_unique<OptimizationRemarkAnalysis>(
DEBUG_TYPE, RemarkName,
DL,
2803 auto *SE = PSE->
getSE();
2824 std::advance(GEPTI, LastOperand - 2);
2831 if (ElemSize != GEPAllocSize)
2851 for (
unsigned I = 0, E =
GEP->getNumOperands();
I != E; ++
I)
2852 if (
I != InductionOperand &&
2855 return GEP->getOperand(InductionOperand);
2861 auto *PtrTy = dyn_cast<PointerType>(
Ptr->getType());
2862 if (!PtrTy || PtrTy->isAggregateType())
2871 int64_t PtrAccessSize = 1;
2879 V =
C->getOperand();
2896 if (OrigPtr ==
Ptr) {
2897 if (
const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
2898 if (M->getOperand(0)->getSCEVType() !=
scConstant)
2901 const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt();
2908 if (PtrAccessSize != StepVal)
2910 V = M->getOperand(1);
2920 if (isa<SCEVUnknown>(V))
2923 if (
const auto *
C = dyn_cast<SCEVIntegralCastExpr>(V))
2924 if (isa<SCEVUnknown>(
C->getOperand()))
2930void LoopAccessInfo::collectStridedAccess(
Value *MemAccess) {
2945 LLVM_DEBUG(
dbgs() <<
"LAA: Found a strided access that is a candidate for "
2950 LLVM_DEBUG(
dbgs() <<
" Chose not to due to -laa-speculate-unit-stride\n");
2975 const SCEV *CastedStride = StrideExpr;
2976 const SCEV *CastedBECount = MaxBTC;
2978 if (BETypeSizeBits >= StrideTypeSizeBits)
2982 const SCEV *StrideMinusBETaken = SE->
getMinusSCEV(CastedStride, CastedBECount);
2988 dbgs() <<
"LAA: Stride>=TripCount; No point in versioning as the "
2989 "Stride==1 predicate will imply that the loop executes "
2993 LLVM_DEBUG(
dbgs() <<
"LAA: Found a strided access that we can version.\n");
2997 const SCEV *StrideBase = StrideExpr;
2998 if (
const auto *
C = dyn_cast<SCEVIntegralCastExpr>(StrideBase))
2999 StrideBase =
C->getOperand();
3000 SymbolicStrides[
Ptr] = cast<SCEVUnknown>(StrideBase);
3008 PtrRtChecking(nullptr), TheLoop(L) {
3009 unsigned MaxTargetVectorWidthInBits = std::numeric_limits<unsigned>::max();
3016 MaxTargetVectorWidthInBits = FixedWidth.
getFixedValue() * 2;
3022 MaxTargetVectorWidthInBits = std::numeric_limits<unsigned>::max();
3024 DepChecker = std::make_unique<MemoryDepChecker>(*PSE, L, SymbolicStrides,
3025 MaxTargetVectorWidthInBits);
3026 PtrRtChecking = std::make_unique<RuntimePointerChecking>(*DepChecker, SE);
3027 if (canAnalyzeLoop())
3028 CanVecMem = analyzeLoop(AA, LI, TLI, DT);
3036 OS <<
" with a maximum safe vector width of "
3038 if (PtrRtChecking->Need)
3039 OS <<
" with run-time checks";
3043 if (HasConvergentOp)
3051 for (
const auto &Dep : *Dependences) {
3059 PtrRtChecking->print(
OS,
Depth);
3063 <<
"Non vectorizable stores to invariant address were "
3064 << (HasStoreStoreDependenceInvolvingLoopInvariantAddress ||
3065 HasLoadStoreDependenceInvolvingLoopInvariantAddress
3068 <<
"found in loop.\n";
3080 auto [It, Inserted] = LoopAccessInfoMap.insert({&L,
nullptr});
3084 std::make_unique<LoopAccessInfo>(&L, &SE,
TTI, TLI, &AA, &DT, &LI);
3094 for (
const auto &[L, LAI] : LoopAccessInfoMap) {
3095 if (LAI->getRuntimePointerChecking()->getChecks().empty() &&
3096 LAI->getPSE().getPredicate().isAlwaysTrue())
3102 LoopAccessInfoMap.erase(L);
This file implements a class to represent arbitrary precision integral constant values and operations...
ReachingDefAnalysis InstSet & ToRemove
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
Analysis containing CSE Info
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
This file defines the DenseMap class.
Generic implementation of equivalence classes through the use Tarjan's efficient union-find algorithm...
static std::pair< const SCEV *, const SCEV * > getStartAndEndForAccess(const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy, PredicatedScalarEvolution &PSE, DenseMap< std::pair< const SCEV *, Type * >, std::pair< const SCEV *, const SCEV * > > &PointerBounds)
Calculate Start and End points of memory access.
static cl::opt< unsigned > MaxDependences("max-dependences", cl::Hidden, cl::desc("Maximum number of dependences collected by " "loop-access analysis (default = 100)"), cl::init(100))
We collect dependences up to this threshold.
static cl::opt< bool > EnableForwardingConflictDetection("store-to-load-forwarding-conflict-detection", cl::Hidden, cl::desc("Enable conflict detection in loop-access analysis"), cl::init(true))
Enable store-to-load forwarding conflict detection.
static void findForkedSCEVs(ScalarEvolution *SE, const Loop *L, Value *Ptr, SmallVectorImpl< PointerIntPair< const SCEV *, 1, bool > > &ScevList, unsigned Depth)
static bool hasComputableBounds(PredicatedScalarEvolution &PSE, Value *Ptr, const SCEV *PtrScev, Loop *L, bool Assume)
Check whether a pointer can participate in a runtime bounds check.
static cl::opt< unsigned > MemoryCheckMergeThreshold("memory-check-merge-threshold", cl::Hidden, cl::desc("Maximum number of comparisons done when trying to merge " "runtime memory checks. (default = 100)"), cl::init(100))
The maximum iterations used to merge memory checks.
static bool isNoWrap(PredicatedScalarEvolution &PSE, const DenseMap< Value *, const SCEV * > &Strides, Value *Ptr, Type *AccessTy, Loop *L)
Check whether a pointer address cannot wrap.
static const SCEV * getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp)
Get the stride of a pointer access in a loop.
static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep)
Find the operand of the GEP that should be checked for consecutive stores.
static cl::opt< unsigned, true > VectorizationInterleave("force-vector-interleave", cl::Hidden, cl::desc("Sets the vectorization interleave count. " "Zero is autoselect."), cl::location(VectorizerParams::VectorizationInterleave))
static bool isLoopVariantIndirectAddress(ArrayRef< const Value * > UnderlyingObjects, ScalarEvolution &SE, const Loop *L)
Returns true if any of the underlying objects has a loop varying address, i.e.
static cl::opt< bool, true > HoistRuntimeChecks("hoist-runtime-checks", cl::Hidden, cl::desc("Hoist inner loop runtime memory checks to outer loop if possible"), cl::location(VectorizerParams::HoistRuntimeChecks), cl::init(true))
static cl::opt< unsigned, true > VectorizationFactor("force-vector-width", cl::Hidden, cl::desc("Sets the SIMD width. Zero is autoselect."), cl::location(VectorizerParams::VectorizationFactor))
static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, const SCEV &MaxBTC, const SCEV &Dist, uint64_t MaxStride, uint64_t TypeByteSize)
Given a dependence-distance Dist between two memory accesses, that have strides in the same direction...
static cl::opt< unsigned, true > RuntimeMemoryCheckThreshold("runtime-memory-check-threshold", cl::Hidden, cl::desc("When performing memory disambiguation checks at runtime do not " "generate more than this number of comparisons (default = 8)."), cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8))
static void visitPointers(Value *StartPtr, const Loop &InnermostLoop, function_ref< void(Value *)> AddPointer)
static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, PredicatedScalarEvolution &PSE, const Loop *L)
Return true if an AddRec pointer Ptr is unsigned non-wrapping, i.e.
static Value * stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp)
If the argument is a GEP, then returns the operand identified by getGEPInductionOperand.
static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, uint64_t TypeByteSize)
Check the dependence for two accesses with the same stride Stride.
static const SCEV * getMinFromExprs(const SCEV *I, const SCEV *J, ScalarEvolution *SE)
Compare I and J and return the minimum.
static cl::opt< unsigned > MaxForkedSCEVDepth("max-forked-scev-depth", cl::Hidden, cl::desc("Maximum recursion depth when finding forked SCEVs (default = 5)"), cl::init(5))
static cl::opt< bool > SpeculateUnitStride("laa-speculate-unit-stride", cl::Hidden, cl::desc("Speculate that non-constant strides are unit in LAA"), cl::init(true))
static SmallVector< PointerIntPair< const SCEV *, 1, bool > > findForkedPointer(PredicatedScalarEvolution &PSE, const DenseMap< Value *, const SCEV * > &StridesMap, Value *Ptr, const Loop *L)
static cl::opt< bool > EnableMemAccessVersioning("enable-mem-access-versioning", cl::init(true), cl::Hidden, cl::desc("Enable symbolic stride memory access versioning"))
This enables versioning on the strides of symbolically striding memory accesses in code like the foll...
This header provides classes for managing per-loop analyses.
This file provides utility analysis objects describing memory locations.
FunctionAnalysisManager FAM
This header defines various interfaces for pass management in LLVM.
This file defines the PointerIntPair class.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static LLVM_ATTRIBUTE_ALWAYS_INLINE bool CheckType(MVT::SimpleValueType VT, SDValue N, const TargetLowering *TLI, const DataLayout &DL)
This file implements a set that has insertion order iteration characteristics.
This file defines the SmallPtrSet class.
This file defines the SmallSet class.
This file defines the SmallVector class.
static SymbolRef::Type getType(const Symbol *Sym)
static const X86InstrFMA3Group Groups[]
A manager for alias analyses.
Class for arbitrary precision integers.
unsigned getBitWidth() const
Return the number of bits in the APInt.
APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
int64_t getSExtValue() const
Get sign extended value.
This templated class represents "all analyses that operate over <a particular IR unit>" (e....
API to communicate dependencies between analyses during invalidation.
bool invalidate(IRUnitT &IR, const PreservedAnalyses &PA)
Trigger the invalidation of some other analysis pass if not already handled and return whether it was...
A container for analyses that lazily runs them and caches their results.
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
size_t size() const
size - Get the array size.
bool empty() const
empty - Check if the array is empty.
LLVM Basic Block Representation.
const Function * getParent() const
Return the enclosing method, or null if none.
const DataLayout & getDataLayout() const
Get the data layout of the module this basic block belongs to.
This class is a wrapper over an AAResults, and it is intended to be used only when there are no IR ch...
@ ICMP_ULE
unsigned less or equal
This class represents an Operation in the Expression.
A parsed version of the target data layout string in and methods for querying it.
iterator find(const_arg_type_t< KeyT > Val)
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Analysis pass which computes a DominatorTree.
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
bool dominates(const BasicBlock *BB, const Use &U) const
Return true if the (end of the) basic block BB dominates the use U.
EquivalenceClasses - This represents a collection of equivalence classes and supports three efficient...
iterator findValue(const ElemTy &V) const
findValue - Return an iterator to the specified value.
iterator insert(const ElemTy &Data)
insert - Insert a new value into the union/find set, ignoring the request if the value already exists...
member_iterator member_end() const
typename std::set< ECValue, ECValueComparator >::const_iterator iterator
iterator* - Provides a way to iterate over all values in the set.
member_iterator member_begin(iterator I) const
member_iterator unionSets(const ElemTy &V1, const ElemTy &V2)
union - Merge the two equivalence sets for the specified values, inserting them if they do not alread...
const ElemTy & getLeaderValue(const ElemTy &V) const
getLeaderValue - Return the leader for the specified value that is in the set.
bool hasOptSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
an instruction for type-safe pointer arithmetic to access elements of arrays and structs
Type * getResultElementType() const
PointerType * getType() const
Global values are always pointers.
const DataLayout & getDataLayout() const
Get the data layout of the module this instruction belongs to.
Class to represent integer types.
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
An instruction for reading from memory.
Value * getPointerOperand()
static constexpr LocationSize beforeOrAfterPointer()
Any location before or after the base pointer (but still within the underlying object).
This analysis provides dependence information for the memory accesses of a loop.
Result run(Function &F, FunctionAnalysisManager &AM)
bool invalidate(Function &F, const PreservedAnalyses &PA, FunctionAnalysisManager::Invalidator &Inv)
const LoopAccessInfo & getInfo(Loop &L)
Drive the analysis of memory accesses in the loop.
const MemoryDepChecker & getDepChecker() const
the Memory Dependence Checker which can determine the loop-independent and loop-carried dependences b...
bool isInvariant(Value *V) const
Returns true if value V is loop invariant.
void print(raw_ostream &OS, unsigned Depth=0) const
Print the information about the memory accesses in the loop.
static bool blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, DominatorTree *DT)
Return true if the block BB needs to be predicated in order for the loop to be vectorized.
LoopAccessInfo(Loop *L, ScalarEvolution *SE, const TargetTransformInfo *TTI, const TargetLibraryInfo *TLI, AAResults *AA, DominatorTree *DT, LoopInfo *LI)
Analysis pass that exposes the LoopInfo for a function.
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
bool isInnermost() const
Return true if the loop does not contain any (natural) loops.
unsigned getNumBackEdges() const
Calculate the number of back edges to the loop header.
BlockT * getHeader() const
LoopT * getParentLoop() const
Return the parent loop if it exists or nullptr for top level loops.
Wrapper class to LoopBlocksDFS that provides a standard begin()/end() interface for the DFS reverse p...
Represents a single loop in the control flow graph.
std::string getLocStr() const
Return a string containing the debug location of the loop (file name + line number if present,...
bool isAnnotatedParallel() const
Returns true if the loop is annotated parallel.
DebugLoc getStartLoc() const
Return the debug location of the start of this loop.
ArrayRef< MDOperand > operands() const
Tracking metadata reference owned by Metadata.
This class implements a map that also provides access to all stored values in a deterministic order.
Checks memory dependences among accesses to the same underlying object to determine whether there vec...
ArrayRef< unsigned > getOrderForAccess(Value *Ptr, bool IsWrite) const
Return the program order indices for the access location (Ptr, IsWrite).
bool areDepsSafe(DepCandidates &AccessSets, MemAccessInfoList &CheckDeps, const DenseMap< Value *, SmallVector< const Value *, 16 > > &UnderlyingObjects)
Check whether the dependencies between the accesses are safe.
bool isSafeForAnyVectorWidth() const
Return true if the number of elements that are safe to operate on simultaneously is not bounded.
const SmallVectorImpl< Instruction * > & getMemoryInstructions() const
The vector of memory access instructions.
const Loop * getInnermostLoop() const
uint64_t getMaxSafeVectorWidthInBits() const
Return the number of elements that are safe to operate on simultaneously, multiplied by the size of t...
bool isSafeForVectorization() const
No memory dependence was encountered that would inhibit vectorization.
const SmallVectorImpl< Dependence > * getDependences() const
Returns the memory dependences.
DenseMap< std::pair< const SCEV *, Type * >, std::pair< const SCEV *, const SCEV * > > & getPointerBounds()
SmallVector< Instruction *, 4 > getInstructionsForAccess(Value *Ptr, bool isWrite) const
Find the set of instructions that read or write via Ptr.
VectorizationSafetyStatus
Type to keep track of the status of the dependence check.
@ PossiblySafeWithRtChecks
bool shouldRetryWithRuntimeCheck() const
In same cases when the dependency check fails we can still vectorize the loop with a dynamic array ac...
void addAccess(StoreInst *SI)
Register the location (instructions are given increasing numbers) of a write access.
PointerIntPair< Value *, 1, bool > MemAccessInfo
Representation for a specific memory location.
static MemoryLocation get(const LoadInst *LI)
Return a location with information about the memory reference by the given instruction.
LocationSize Size
The maximum size of the location, in address-units, or UnknownSize if the size is not known.
AAMDNodes AATags
The metadata nodes which describes the aliasing of the location (each member is null if that kind of ...
const Value * Ptr
The address of the start of the location.
An interface layer with SCEV used to manage how we see SCEV expressions for values in the context of ...
void addPredicate(const SCEVPredicate &Pred)
Adds a new predicate.
ScalarEvolution * getSE() const
Returns the ScalarEvolution analysis used.
const SCEVPredicate & getPredicate() const
bool hasNoOverflow(Value *V, SCEVWrapPredicate::IncrementWrapFlags Flags)
Returns true if we've proved that V doesn't wrap by means of a SCEV predicate.
void setNoOverflow(Value *V, SCEVWrapPredicate::IncrementWrapFlags Flags)
Proves that V doesn't overflow by adding SCEV predicate.
void print(raw_ostream &OS, unsigned Depth) const
Print the SCEV mappings done by the Predicated Scalar Evolution.
const SCEVAddRecExpr * getAsAddRec(Value *V)
Attempts to produce an AddRecExpr for V by adding additional SCEV predicates.
const SCEV * getSymbolicMaxBackedgeTakenCount()
Get the (predicated) symbolic max backedge count for the analyzed loop.
const SCEV * getSCEV(Value *V)
Returns the SCEV expression of V, in the context of the current SCEV predicate.
A set of analyses that are preserved following a run of a transformation pass.
PreservedAnalysisChecker getChecker() const
Build a checker for this PreservedAnalyses and the specified analysis type.
Holds information about the memory runtime legality checks to verify that a group of pointers do not ...
bool Need
This flag indicates if we need to add the runtime check.
void reset()
Reset the state of the pointer runtime information.
unsigned getNumberOfChecks() const
Returns the number of run-time checks required according to needsChecking.
void printChecks(raw_ostream &OS, const SmallVectorImpl< RuntimePointerCheck > &Checks, unsigned Depth=0) const
Print Checks.
bool needsChecking(const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const
Decide if we need to add a check between two groups of pointers, according to needsChecking.
void print(raw_ostream &OS, unsigned Depth=0) const
Print the list run-time memory checks necessary.
SmallVector< RuntimeCheckingPtrGroup, 2 > CheckingGroups
Holds a partitioning of pointers into "check groups".
void generateChecks(MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies)
Generate the checks and store it.
friend struct RuntimeCheckingPtrGroup
static bool arePointersInSamePartition(const SmallVectorImpl< int > &PtrToPartition, unsigned PtrIdx1, unsigned PtrIdx2)
Check if pointers are in the same partition.
SmallVector< PointerInfo, 2 > Pointers
Information about the pointers that may require checking.
void insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr, Type *AccessTy, bool WritePtr, unsigned DepSetId, unsigned ASId, PredicatedScalarEvolution &PSE, bool NeedsFreeze)
Insert a pointer and calculate the start and end SCEVs.
This node represents a polynomial recurrence on the trip count of the specified loop.
const SCEV * getStepRecurrence(ScalarEvolution &SE) const
Constructs and returns the recurrence indicating how much this expression steps by.
bool isAffine() const
Return true if this represents an expression A + B*x where A and B are loop invariant values.
const Loop * getLoop() const
This class represents a constant integer value.
This is the base class for unary integral cast operator classes.
This node represents multiplication of some number of SCEVs.
NoWrapFlags getNoWrapFlags(NoWrapFlags Mask=NoWrapMask) const
virtual void print(raw_ostream &OS, unsigned Depth=0) const =0
Prints a textual representation of this predicate with an indentation of Depth.
This class represents an analyzed expression in the program.
Type * getType() const
Return the LLVM type of this SCEV expression.
Analysis pass that exposes the ScalarEvolution for a function.
The main scalar evolution driver.
bool isKnownNonNegative(const SCEV *S)
Test if the given expression is known to be non-negative.
const SCEV * getNegativeSCEV(const SCEV *V, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap)
Return the SCEV object corresponding to -V.
bool isKnownNonPositive(const SCEV *S)
Test if the given expression is known to be non-positive.
const SCEV * getUMaxExpr(const SCEV *LHS, const SCEV *RHS)
const SCEVPredicate * getEqualPredicate(const SCEV *LHS, const SCEV *RHS)
const SCEV * getConstant(ConstantInt *V)
const SCEV * getSCEV(Value *V)
Return a SCEV expression for the full generality of the specified expression.
const SCEV * getNoopOrSignExtend(const SCEV *V, Type *Ty)
Return a SCEV corresponding to a conversion of the input value to the specified type.
const SCEV * getOne(Type *Ty)
Return a SCEV for the constant 1 of a specific type.
const SCEV * getPtrToIntExpr(const SCEV *Op, Type *Ty)
bool isLoopInvariant(const SCEV *S, const Loop *L)
Return true if the value of the given SCEV is unchanging in the specified loop.
bool isKnownPositive(const SCEV *S)
Test if the given expression is known to be positive.
bool isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS)
Test if the given expression is known to satisfy the condition described by Pred, LHS,...
const SCEV * getZeroExtendExpr(const SCEV *Op, Type *Ty, unsigned Depth=0)
bool isSCEVable(Type *Ty) const
Test if values of the given type are analyzable within the SCEV framework.
Type * getEffectiveSCEVType(Type *Ty) const
Return a type with the same bitwidth as the given type and which represents how SCEV will treat the g...
const SCEV * getUMinExpr(const SCEV *LHS, const SCEV *RHS, bool Sequential=false)
APInt getSignedRangeMin(const SCEV *S)
Determine the min of the signed range for a particular SCEV.
const SCEV * getStoreSizeOfExpr(Type *IntTy, Type *StoreTy)
Return an expression for the store size of StoreTy that is type IntTy.
const SCEV * getMinusSCEV(const SCEV *LHS, const SCEV *RHS, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Return LHS-RHS.
const SCEV * getCouldNotCompute()
const SCEV * applyLoopGuards(const SCEV *Expr, const Loop *L)
Try to apply information from loop guards for L to Expr.
const SCEV * getMulExpr(SmallVectorImpl< const SCEV * > &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical multiply expression, or something simpler if possible.
const SCEV * getSizeOfExpr(Type *IntTy, TypeSize Size)
Return an expression for a TypeSize.
const SCEV * getAddExpr(SmallVectorImpl< const SCEV * > &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical add expression, or something simpler if possible.
const SCEV * getTruncateOrSignExtend(const SCEV *V, Type *Ty, unsigned Depth=0)
Return a SCEV corresponding to a conversion of the input value to the specified type.
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
size_type count(const T &V) const
count - Return 1 if the element is in the set, 0 otherwise.
std::pair< const_iterator, bool > insert(const T &V)
insert - Insert an element into the set if it isn't already there.
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
reference emplace_back(ArgTypes &&... Args)
void push_back(const T &Elt)
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
An instruction for storing to memory.
StringRef - Represent a constant reference to a string, i.e.
Analysis pass providing the TargetTransformInfo.
Analysis pass providing the TargetLibraryInfo.
Provides information about what library functions are available for the current target.
The instances of the Type class are immutable: once they are created, they are never changed.
bool isVectorTy() const
True if this is an instance of VectorType.
bool isPointerTy() const
True if this is an instance of PointerType.
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
A Use represents the edge between a Value definition and its users.
Value * getOperand(unsigned i) const
unsigned getNumOperands() const
static SmallVector< VFInfo, 8 > getMappings(const CallInst &CI)
Retrieve all the VFInfo instances associated to the CallInst CI.
LLVM Value Representation.
Type * getType() const
All values are typed, get the type of this value.
const Value * stripAndAccumulateInBoundsConstantOffsets(const DataLayout &DL, APInt &Offset) const
This is a wrapper around stripAndAccumulateConstantOffsets with the in-bounds requirement set to fals...
StringRef getName() const
Return a constant reference to the value's name.
constexpr ScalarTy getFixedValue() const
constexpr bool isNonZero() const
An efficient, type-erasing, non-owning reference to a callable.
TypeSize getSequentialElementStride(const DataLayout &DL) const
Type * getIndexedType() const
This class implements an extremely fast bulk output stream that can only output to a stream.
raw_ostream & indent(unsigned NumSpaces)
indent - Insert 'NumSpaces' spaces.
friend const_iterator end(StringRef path)
Get end iterator over path.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
@ C
The default llvm calling convention, compatible with C.
bool match(Val *V, const Pattern &P)
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
initializer< Ty > init(const Ty &Val)
LocationClass< Ty > location(Ty &L)
DiagnosticInfoOptimizationBase::Argument NV
This is an optimization pass for GlobalISel generic memory operations.
auto drop_begin(T &&RangeOrContainer, size_t N=1)
Return a range covering RangeOrContainer with the first N elements excluded.
std::optional< int > getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB, Value *PtrB, const DataLayout &DL, ScalarEvolution &SE, bool StrictCheck=false, bool CheckType=true)
Returns the distance between the pointers PtrA and PtrB iff they are compatible and it is possible to...
@ Low
Lower the current thread's priority such that it does not affect foreground tasks significantly.
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
Intrinsic::ID getVectorIntrinsicIDForCall(const CallInst *CI, const TargetLibraryInfo *TLI)
Returns intrinsic ID for call.
auto enumerate(FirstRange &&First, RestRanges &&...Rest)
Given two or more input ranges, returns a new range whose values are are tuples (A,...
unsigned getPointerAddressSpace(const Type *T)
std::optional< const MDOperand * > findStringMetadataForLoop(const Loop *TheLoop, StringRef Name)
Find string metadata for loop.
const Value * getLoadStorePointerOperand(const Value *V)
A helper function that returns the pointer operand of a load or store instruction.
const Value * getPointerOperand(const Value *V)
A helper function that returns the pointer operand of a load, store or GEP instruction.
OutputIt transform(R &&Range, OutputIt d_first, UnaryFunction F)
Wrapper function around std::transform to apply a function to a range and store the result elsewhere.
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
bool NullPointerIsDefined(const Function *F, unsigned AS=0)
Check whether null pointer dereferencing is considered undefined behavior for a given function or an ...
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
bool isPointerTy(const Type *T)
std::optional< int64_t > getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy, Value *Ptr, const Loop *Lp, const DenseMap< Value *, const SCEV * > &StridesMap=DenseMap< Value *, const SCEV * >(), bool Assume=false, bool ShouldCheckWrap=true)
If the pointer has a constant stride return it in units of the access type size.
bool sortPtrAccesses(ArrayRef< Value * > VL, Type *ElemTy, const DataLayout &DL, ScalarEvolution &SE, SmallVectorImpl< unsigned > &SortedIndices)
Attempt to sort the pointers in VL and return the sorted indices in SortedIndices,...
@ First
Helpers to iterate all locations in the MemoryEffectsBase class.
void getUnderlyingObjects(const Value *V, SmallVectorImpl< const Value * > &Objects, LoopInfo *LI=nullptr, unsigned MaxLookup=6)
This method is similar to getUnderlyingObject except that it can look through phi and select instruct...
const SCEV * replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, const DenseMap< Value *, const SCEV * > &PtrToStride, Value *Ptr)
Return the SCEV corresponding to a pointer with the symbolic stride replaced with constant one,...
bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, ScalarEvolution &SE, bool CheckType=true)
Returns true if the memory operations A and B are consecutive.
bool isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Return true if this function can prove that V does not have undef bits and is never poison.
OutputIt copy(R &&Range, OutputIt Out)
auto find_if(R &&Range, UnaryPredicate P)
Provide wrappers to std::find_if which take ranges instead of having to pass begin/end explicitly.
gep_type_iterator gep_type_begin(const User *GEP)
Type * getLoadStoreType(Value *I)
A helper function that returns the type of a load or store instruction.
Implement std::hash so that hash_code can be used in STL containers.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
IR Values for the lower and upper bounds of a pointer evolution.
MDNode * Scope
The tag for alias scope specification (used with noalias).
MDNode * TBAA
The tag for type-based alias analysis.
MDNode * NoAlias
The tag specifying the noalias scope.
A special type used by analysis passes to provide an address that identifies that particular analysis...
Dependece between memory access instructions.
Instruction * getDestination(const MemoryDepChecker &DepChecker) const
Return the destination instruction of the dependence.
DepType Type
The type of the dependence.
bool isPossiblyBackward() const
May be a lexically backward dependence type (includes Unknown).
Instruction * getSource(const MemoryDepChecker &DepChecker) const
Return the source instruction of the dependence.
bool isForward() const
Lexically forward dependence.
bool isBackward() const
Lexically backward dependence.
void print(raw_ostream &OS, unsigned Depth, const SmallVectorImpl< Instruction * > &Instrs) const
Print the dependence.
DepType
The type of the dependence.
@ BackwardVectorizableButPreventsForwarding
@ ForwardButPreventsForwarding
static const char * DepName[]
String version of the types.
static VectorizationSafetyStatus isSafeForVectorization(DepType Type)
Dependence types that don't prevent vectorization.
unsigned AddressSpace
Address space of the involved pointers.
bool addPointer(unsigned Index, RuntimePointerChecking &RtCheck)
Tries to add the pointer recorded in RtCheck at index Index to this pointer checking group.
bool NeedsFreeze
Whether the pointer needs to be frozen after expansion, e.g.
const SCEV * High
The SCEV expression which represents the upper bound of all the pointers in this group.
SmallVector< unsigned, 2 > Members
Indices of all the pointers that constitute this grouping.
RuntimeCheckingPtrGroup(unsigned Index, RuntimePointerChecking &RtCheck)
Create a new pointer checking group containing a single pointer, with index Index in RtCheck.
const SCEV * Low
The SCEV expression which represents the lower bound of all the pointers in this group.
bool IsWritePtr
Holds the information if this pointer is used for writing to memory.
unsigned DependencySetId
Holds the id of the set of pointers that could be dependent because of a shared underlying object.
unsigned AliasSetId
Holds the id of the disjoint alias set to which this pointer belongs.
static const unsigned MaxVectorWidth
Maximum SIMD width.
static unsigned VectorizationFactor
VF as overridden by the user.
static unsigned RuntimeMemoryCheckThreshold
\When performing memory disambiguation checks at runtime do not make more than this number of compari...
static bool isInterleaveForced()
True if force-vector-interleave was specified by the user.
static unsigned VectorizationInterleave
Interleave factor as overridden by the user.
static bool HoistRuntimeChecks
Function object to check whether the first component of a container supported by std::get (like std::...