62#define DEBUG_TYPE "openmp-ir-builder"
69 cl::desc(
"Use optimistic attributes describing "
70 "'as-if' properties of runtime calls."),
74 "openmp-ir-builder-unroll-threshold-factor",
cl::Hidden,
75 cl::desc(
"Factor for the unroll threshold to account for code "
76 "simplifications still taking place"),
87 if (!IP1.isSet() || !IP2.isSet())
89 return IP1.getBlock() == IP2.getBlock() && IP1.getPoint() == IP2.getPoint();
94 switch (SchedType & ~OMPScheduleType::MonotonicityMask) {
95 case OMPScheduleType::UnorderedStaticChunked:
96 case OMPScheduleType::UnorderedStatic:
97 case OMPScheduleType::UnorderedDynamicChunked:
98 case OMPScheduleType::UnorderedGuidedChunked:
99 case OMPScheduleType::UnorderedRuntime:
100 case OMPScheduleType::UnorderedAuto:
101 case OMPScheduleType::UnorderedTrapezoidal:
102 case OMPScheduleType::UnorderedGreedy:
103 case OMPScheduleType::UnorderedBalanced:
104 case OMPScheduleType::UnorderedGuidedIterativeChunked:
105 case OMPScheduleType::UnorderedGuidedAnalyticalChunked:
106 case OMPScheduleType::UnorderedSteal:
107 case OMPScheduleType::UnorderedStaticBalancedChunked:
108 case OMPScheduleType::UnorderedGuidedSimd:
109 case OMPScheduleType::UnorderedRuntimeSimd:
110 case OMPScheduleType::OrderedStaticChunked:
111 case OMPScheduleType::OrderedStatic:
112 case OMPScheduleType::OrderedDynamicChunked:
113 case OMPScheduleType::OrderedGuidedChunked:
114 case OMPScheduleType::OrderedRuntime:
115 case OMPScheduleType::OrderedAuto:
116 case OMPScheduleType::OrderdTrapezoidal:
117 case OMPScheduleType::NomergeUnorderedStaticChunked:
118 case OMPScheduleType::NomergeUnorderedStatic:
119 case OMPScheduleType::NomergeUnorderedDynamicChunked:
120 case OMPScheduleType::NomergeUnorderedGuidedChunked:
121 case OMPScheduleType::NomergeUnorderedRuntime:
122 case OMPScheduleType::NomergeUnorderedAuto:
123 case OMPScheduleType::NomergeUnorderedTrapezoidal:
124 case OMPScheduleType::NomergeUnorderedGreedy:
125 case OMPScheduleType::NomergeUnorderedBalanced:
126 case OMPScheduleType::NomergeUnorderedGuidedIterativeChunked:
127 case OMPScheduleType::NomergeUnorderedGuidedAnalyticalChunked:
128 case OMPScheduleType::NomergeUnorderedSteal:
129 case OMPScheduleType::NomergeOrderedStaticChunked:
130 case OMPScheduleType::NomergeOrderedStatic:
131 case OMPScheduleType::NomergeOrderedDynamicChunked:
132 case OMPScheduleType::NomergeOrderedGuidedChunked:
133 case OMPScheduleType::NomergeOrderedRuntime:
134 case OMPScheduleType::NomergeOrderedAuto:
135 case OMPScheduleType::NomergeOrderedTrapezoidal:
143 SchedType & OMPScheduleType::MonotonicityMask;
144 if (MonotonicityFlags == OMPScheduleType::MonotonicityMask)
155 if (Features.
count(
"+wavefrontsize64"))
156 return omp::getAMDGPUGridValues<64>();
157 return omp::getAMDGPUGridValues<32>();
168 bool HasSimdModifier) {
170 switch (ClauseKind) {
171 case OMP_SCHEDULE_Default:
172 case OMP_SCHEDULE_Static:
173 return HasChunks ? OMPScheduleType::BaseStaticChunked
174 : OMPScheduleType::BaseStatic;
175 case OMP_SCHEDULE_Dynamic:
176 return OMPScheduleType::BaseDynamicChunked;
177 case OMP_SCHEDULE_Guided:
178 return HasSimdModifier ? OMPScheduleType::BaseGuidedSimd
179 : OMPScheduleType::BaseGuidedChunked;
180 case OMP_SCHEDULE_Auto:
182 case OMP_SCHEDULE_Runtime:
183 return HasSimdModifier ? OMPScheduleType::BaseRuntimeSimd
184 : OMPScheduleType::BaseRuntime;
192 bool HasOrderedClause) {
193 assert((BaseScheduleType & OMPScheduleType::ModifierMask) ==
194 OMPScheduleType::None &&
195 "Must not have ordering nor monotonicity flags already set");
198 ? OMPScheduleType::ModifierOrdered
199 : OMPScheduleType::ModifierUnordered;
200 OMPScheduleType OrderingScheduleType = BaseScheduleType | OrderingModifier;
203 if (OrderingScheduleType ==
204 (OMPScheduleType::BaseGuidedSimd | OMPScheduleType::ModifierOrdered))
205 return OMPScheduleType::OrderedGuidedChunked;
206 else if (OrderingScheduleType == (OMPScheduleType::BaseRuntimeSimd |
207 OMPScheduleType::ModifierOrdered))
208 return OMPScheduleType::OrderedRuntime;
210 return OrderingScheduleType;
216 bool HasSimdModifier,
bool HasMonotonic,
217 bool HasNonmonotonic,
bool HasOrderedClause) {
218 assert((ScheduleType & OMPScheduleType::MonotonicityMask) ==
219 OMPScheduleType::None &&
220 "Must not have monotonicity flags already set");
221 assert((!HasMonotonic || !HasNonmonotonic) &&
222 "Monotonic and Nonmonotonic are contradicting each other");
225 return ScheduleType | OMPScheduleType::ModifierMonotonic;
226 }
else if (HasNonmonotonic) {
227 return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
237 if ((BaseScheduleType == OMPScheduleType::BaseStatic) ||
238 (BaseScheduleType == OMPScheduleType::BaseStaticChunked) ||
244 return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
252 bool HasSimdModifier,
bool HasMonotonicModifier,
253 bool HasNonmonotonicModifier,
bool HasOrderedClause) {
259 OrderedSchedule, HasSimdModifier, HasMonotonicModifier,
260 HasNonmonotonicModifier, HasOrderedClause);
274 auto *Br = cast<BranchInst>(Term);
275 assert(!Br->isConditional() &&
276 "BB's terminator must be an unconditional branch (or degenerate)");
279 Br->setSuccessor(0,
Target);
284 NewBr->setDebugLoc(
DL);
289 assert(New->getFirstInsertionPt() == New->begin() &&
290 "Target BB must not have PHI nodes");
294 New->splice(New->begin(), Old, IP.
getPoint(), Old->
end());
322 New->replaceSuccessorsPhiUsesWith(Old, New);
365 std::stack<Instruction *> &ToBeDeleted,
367 const Twine &
Name =
"",
bool AsPtr =
true) {
372 ToBeDeleted.push(FakeValAddr);
375 FakeVal = FakeValAddr;
379 ToBeDeleted.push(FakeVal);
392 ToBeDeleted.push(UseFakeVal);
403enum OpenMPOffloadingRequiresDirFlags {
405 OMP_REQ_UNDEFINED = 0x000,
407 OMP_REQ_NONE = 0x001,
409 OMP_REQ_REVERSE_OFFLOAD = 0x002,
411 OMP_REQ_UNIFIED_ADDRESS = 0x004,
413 OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008,
415 OMP_REQ_DYNAMIC_ALLOCATORS = 0x010,
422 : RequiresFlags(OMP_REQ_UNDEFINED) {}
425 bool IsTargetDevice,
bool IsGPU,
bool OpenMPOffloadMandatory,
426 bool HasRequiresReverseOffload,
bool HasRequiresUnifiedAddress,
427 bool HasRequiresUnifiedSharedMemory,
bool HasRequiresDynamicAllocators)
428 : IsTargetDevice(IsTargetDevice), IsGPU(IsGPU),
429 OpenMPOffloadMandatory(OpenMPOffloadMandatory),
430 RequiresFlags(OMP_REQ_UNDEFINED) {
431 if (HasRequiresReverseOffload)
432 RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD;
433 if (HasRequiresUnifiedAddress)
434 RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS;
435 if (HasRequiresUnifiedSharedMemory)
436 RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY;
437 if (HasRequiresDynamicAllocators)
438 RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS;
442 return RequiresFlags & OMP_REQ_REVERSE_OFFLOAD;
446 return RequiresFlags & OMP_REQ_UNIFIED_ADDRESS;
450 return RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY;
454 return RequiresFlags & OMP_REQ_DYNAMIC_ALLOCATORS;
459 :
static_cast<int64_t
>(OMP_REQ_NONE);
464 RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD;
466 RequiresFlags &= ~OMP_REQ_REVERSE_OFFLOAD;
471 RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS;
473 RequiresFlags &= ~OMP_REQ_UNIFIED_ADDRESS;
478 RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY;
480 RequiresFlags &= ~OMP_REQ_UNIFIED_SHARED_MEMORY;
485 RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS;
487 RequiresFlags &= ~OMP_REQ_DYNAMIC_ALLOCATORS;
505 Value *NumThreads3D =
528 auto FnAttrs = Attrs.getFnAttrs();
529 auto RetAttrs = Attrs.getRetAttrs();
531 for (
size_t ArgNo = 0; ArgNo < Fn.
arg_size(); ++ArgNo)
536 bool Param =
true) ->
void {
537 bool HasSignExt = AS.hasAttribute(Attribute::SExt);
538 bool HasZeroExt = AS.hasAttribute(Attribute::ZExt);
539 if (HasSignExt || HasZeroExt) {
540 assert(AS.getNumAttributes() == 1 &&
541 "Currently not handling extension attr combined with others.");
543 if (
auto AK = TargetLibraryInfo::getExtAttrForI32Param(
T, HasSignExt))
546 TargetLibraryInfo::getExtAttrForI32Return(
T, HasSignExt))
553#define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet;
554#include "llvm/Frontend/OpenMP/OMPKinds.def"
558#define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets) \
560 FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet); \
561 addAttrSet(RetAttrs, RetAttrSet, false); \
562 for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo) \
563 addAttrSet(ArgAttrs[ArgNo], ArgAttrSets[ArgNo]); \
564 Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs)); \
566#include "llvm/Frontend/OpenMP/OMPKinds.def"
580#define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...) \
582 FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__}, \
584 Fn = M.getFunction(Str); \
586#include "llvm/Frontend/OpenMP/OMPKinds.def"
592#define OMP_RTL(Enum, Str, ...) \
594 Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M); \
596#include "llvm/Frontend/OpenMP/OMPKinds.def"
600 if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) {
610 LLVMContext::MD_callback,
612 2, {-1, -1},
true)}));
625 assert(Fn &&
"Failed to create OpenMP runtime function");
632 auto *Fn = dyn_cast<llvm::Function>(RTLFn.
getCallee());
633 assert(Fn &&
"Failed to create OpenMP runtime function pointer");
648 for (
auto Inst =
Block->getReverseIterator()->begin();
649 Inst !=
Block->getReverseIterator()->end();) {
650 if (
auto *
AllocaInst = dyn_cast_if_present<llvm::AllocaInst>(Inst)) {
674 ParallelRegionBlockSet.
clear();
676 OI.collectBlocks(ParallelRegionBlockSet,
Blocks);
695 ".omp_par", ArgsInZeroAddressSpace);
699 <<
" Exit: " << OI.ExitBB->getName() <<
"\n");
701 "Expected OpenMP outlining to be possible!");
703 for (
auto *V : OI.ExcludeArgsFromAggregate)
710 if (TargetCpuAttr.isStringAttribute())
713 auto TargetFeaturesAttr = OuterFn->
getFnAttribute(
"target-features");
714 if (TargetFeaturesAttr.isStringAttribute())
715 OutlinedFn->
addFnAttr(TargetFeaturesAttr);
718 LLVM_DEBUG(
dbgs() <<
" Outlined function: " << *OutlinedFn <<
"\n");
720 "OpenMP outlined functions should not return a value!");
732 assert(OI.EntryBB->getUniquePredecessor() == &ArtificialEntry);
739 "Expected instructions to add in the outlined region entry");
746 if (
I.isTerminator())
749 I.moveBeforePreserving(*OI.EntryBB, OI.EntryBB->getFirstInsertionPt());
752 OI.EntryBB->moveBefore(&ArtificialEntry);
759 if (OI.PostOutlineCB)
760 OI.PostOutlineCB(*OutlinedFn);
791 errs() <<
"Error of kind: " << Kind
792 <<
" when emitting offload entries and metadata during "
793 "OMPIRBuilder finalization \n";
800 std::vector<WeakTrackingVH> LLVMCompilerUsed = {
802 emitUsed(
"llvm.compiler.used", LLVMCompilerUsed);
824 unsigned Reserve2Flags) {
826 LocFlags |= OMP_IDENT_FLAG_KMPC;
834 ConstantInt::get(
Int32, Reserve2Flags),
835 ConstantInt::get(
Int32, SrcLocStrSize), SrcLocStr};
842 if (
GV.getValueType() == OpenMPIRBuilder::Ident &&
GV.hasInitializer())
843 if (
GV.getInitializer() == Initializer)
848 M, OpenMPIRBuilder::Ident,
863 SrcLocStrSize = LocStr.
size();
872 if (
GV.isConstant() &&
GV.hasInitializer() &&
873 GV.getInitializer() == Initializer)
884 unsigned Line,
unsigned Column,
890 Buffer.
append(FunctionName);
892 Buffer.
append(std::to_string(Line));
894 Buffer.
append(std::to_string(Column));
902 StringRef UnknownLoc =
";unknown;unknown;0;0;;";
913 if (
DIFile *DIF = DIL->getFile())
914 if (std::optional<StringRef> Source = DIF->getSource())
920 DIL->getColumn(), SrcLocStrSize);
932 "omp_global_thread_num");
937 bool ForceSimpleCall,
bool CheckCancelFlag) {
947 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR;
950 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS;
953 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE;
956 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL;
959 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL;
972 bool UseCancelBarrier =
977 UseCancelBarrier ? OMPRTL___kmpc_cancel_barrier
978 : OMPRTL___kmpc_barrier),
981 if (UseCancelBarrier && CheckCancelFlag)
990 omp::Directive CanceledDirective) {
1002 Value *CancelKind =
nullptr;
1003 switch (CanceledDirective) {
1004#define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value) \
1005 case DirectiveEnum: \
1006 CancelKind = Builder.getInt32(Value); \
1008#include "llvm/Frontend/OpenMP/OMPKinds.def"
1019 auto ExitCB = [
this, CanceledDirective, Loc](
InsertPointTy IP) {
1020 if (CanceledDirective == OMPD_parallel) {
1024 omp::Directive::OMPD_unknown,
false,
1034 UI->eraseFromParent();
1047 auto *KernelArgsPtr =
1060 NumThreads, HostPtr, KernelArgsPtr};
1088 assert(OutlinedFnID &&
"Invalid outlined function ID!");
1092 Value *Return =
nullptr;
1112 Args.NumTeams, Args.NumThreads,
1113 OutlinedFnID, ArgsVector));
1126 emitBlock(OffloadContBlock, CurFn,
true);
1131 omp::Directive CanceledDirective,
1134 "Unexpected cancellation!");
1184 OutlinedFn.
addFnAttr(Attribute::NoUnwind);
1187 "Expected at least tid and bounded tid as arguments");
1188 unsigned NumCapturedVars = OutlinedFn.
arg_size() - 2;
1191 assert(CI &&
"Expected call instruction to outlined function");
1192 CI->
getParent()->setName(
"omp_parallel");
1195 Type *PtrTy = OMPIRBuilder->VoidPtr;
1199 OpenMPIRBuilder ::InsertPointTy CurrentIP = Builder.
saveIP();
1203 Value *Args = ArgsAlloca;
1211 for (
unsigned Idx = 0;
Idx < NumCapturedVars;
Idx++) {
1223 Value *Parallel51CallArgs[] = {
1227 NumThreads ? NumThreads : Builder.
getInt32(-1),
1230 Builder.
CreateBitCast(&OutlinedFn, OMPIRBuilder->ParallelTaskPtr),
1233 Builder.
getInt64(NumCapturedVars)};
1238 Builder.
CreateCall(RTLFn, Parallel51CallArgs);
1253 I->eraseFromParent();
1275 if (
auto *
F = dyn_cast<Function>(RTLFn.
getCallee())) {
1276 if (!
F->hasMetadata(LLVMContext::MD_callback)) {
1284 F->addMetadata(LLVMContext::MD_callback,
1293 OutlinedFn.
addFnAttr(Attribute::NoUnwind);
1296 "Expected at least tid and bounded tid as arguments");
1297 unsigned NumCapturedVars = OutlinedFn.
arg_size() - 2;
1300 CI->
getParent()->setName(
"omp_parallel");
1304 Value *ForkCallArgs[] = {
1305 Ident, Builder.
getInt32(NumCapturedVars),
1306 Builder.
CreateBitCast(&OutlinedFn, OMPIRBuilder->ParallelTaskPtr)};
1309 RealArgs.
append(std::begin(ForkCallArgs), std::end(ForkCallArgs));
1318 auto PtrTy = OMPIRBuilder->VoidPtr;
1319 if (IfCondition && NumCapturedVars == 0) {
1323 if (IfCondition && RealArgs.
back()->getType() != PtrTy)
1341 I->eraseFromParent();
1349 omp::ProcBindKind ProcBind,
bool IsCancellable) {
1376 if (ProcBind != OMP_PROC_BIND_default) {
1380 ConstantInt::get(
Int32,
unsigned(ProcBind),
true)};
1408 TIDAddrAlloca, PointerType ::get(
M.
getContext(), 0),
"tid.addr.ascast");
1413 "zero.addr.ascast");
1437 if (IP.getBlock()->end() == IP.getPoint()) {
1443 assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 &&
1444 IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB &&
1445 "Unexpected insertion point for finalization call!");
1481 LLVM_DEBUG(
dbgs() <<
"Before body codegen: " << *OuterFn <<
"\n");
1484 assert(BodyGenCB &&
"Expected body generation callback!");
1486 BodyGenCB(InnerAllocaIP, CodeGenIP);
1488 LLVM_DEBUG(
dbgs() <<
"After body codegen: " << *OuterFn <<
"\n");
1494 std::move(ToBeDeleted)](
Function &OutlinedFn) {
1496 IfCondition, NumThreads, PrivTID, PrivTIDAddr,
1497 ThreadID, ToBeDeletedVec);
1502 std::move(ToBeDeleted)](
Function &OutlinedFn) {
1504 PrivTID, PrivTIDAddr, ToBeDeletedVec);
1521 PRegOutlinedExitBB->
setName(
"omp.par.outlined.exit");
1522 Blocks.push_back(PRegOutlinedExitBB);
1533 ".omp_par", ArgsInZeroAddressSpace);
1538 Extractor.
findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
1541 LLVM_DEBUG(
dbgs() <<
"Before privatization: " << *OuterFn <<
"\n");
1546 auto PrivHelper = [&](
Value &V) {
1547 if (&V == TIDAddr || &V == ZeroAddr) {
1553 for (
Use &U : V.uses())
1554 if (
auto *UserI = dyn_cast<Instruction>(U.getUser()))
1555 if (ParallelRegionBlockSet.
count(UserI->getParent()))
1565 if (!V.getType()->isPointerTy()) {
1584 Value *ReplacementValue =
nullptr;
1585 CallInst *CI = dyn_cast<CallInst>(&V);
1587 ReplacementValue = PrivTID;
1590 PrivCB(InnerAllocaIP,
Builder.
saveIP(), V, *Inner, ReplacementValue));
1595 assert(ReplacementValue &&
1596 "Expected copy/create callback to set replacement value!");
1597 if (ReplacementValue == &V)
1602 UPtr->set(ReplacementValue);
1619 for (
Value *Input : Inputs) {
1624 for (
Value *Output : Outputs)
1628 "OpenMP outlining should not produce live-out values!");
1630 LLVM_DEBUG(
dbgs() <<
"After privatization: " << *OuterFn <<
"\n");
1633 dbgs() <<
" PBR: " << BB->getName() <<
"\n";
1641 assert(FiniInfo.DK == OMPD_parallel &&
1642 "Unexpected finalization stack state!");
1652 InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end());
1653 UI->eraseFromParent();
1746 BodyGenCB(TaskAllocaIP, TaskBodyIP);
1754 std::stack<Instruction *> ToBeDeleted;
1756 Builder, AllocaIP, ToBeDeleted, TaskAllocaIP,
"global.tid",
false));
1758 OI.
PostOutlineCB = [
this, Ident, Tied, Final, IfCondition, Dependencies,
1759 TaskAllocaBB, ToBeDeleted](
Function &OutlinedFn)
mutable {
1761 assert(OutlinedFn.getNumUses() == 1 &&
1762 "there must be a single user for the outlined function");
1763 CallInst *StaleCI = cast<CallInst>(OutlinedFn.user_back());
1767 bool HasShareds = StaleCI->
arg_size() > 1;
1806 assert(ArgStructAlloca &&
1807 "Unable to find the alloca instruction corresponding to arguments "
1808 "for extracted function");
1811 assert(ArgStructType &&
"Unable to find struct type corresponding to "
1812 "arguments for extracted function");
1820 TaskAllocFn, {Ident, ThreadID,
Flags,
1821 TaskSize, SharedsSize,
1833 Value *DepArray =
nullptr;
1834 if (Dependencies.
size()) {
1849 static_cast<unsigned int>(RTLDependInfoFields::BaseAddr));
1856 static_cast<unsigned int>(RTLDependInfoFields::Len));
1863 static_cast<unsigned int>(RTLDependInfoFields::Flags));
1866 static_cast<unsigned int>(Dep.DepKind)),
1897 Instruction *ThenTI = IfTerminator, *ElseTI =
nullptr;
1903 if (Dependencies.
size()) {
1927 if (Dependencies.
size()) {
1948 Shareds, [Shareds](
Use &U) {
return U.getUser() != Shareds; });
1951 while (!ToBeDeleted.empty()) {
1952 ToBeDeleted.top()->eraseFromParent();
2002 if (IP.getBlock()->end() != IP.getPoint())
2013 auto *CaseBB = IP.getBlock()->getSinglePredecessor();
2014 auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
2015 auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
2046 unsigned CaseNumber = 0;
2047 for (
auto SectionCB : SectionCBs) {
2063 Value *LB = ConstantInt::get(I32Ty, 0);
2064 Value *UB = ConstantInt::get(I32Ty, SectionCBs.
size());
2065 Value *ST = ConstantInt::get(I32Ty, 1);
2067 Loc, LoopBodyGenCB, LB, UB, ST,
true,
false, AllocaIP,
"section_loop");
2069 applyStaticWorkshareLoop(Loc.
DL,
LoopInfo, AllocaIP, !IsNowait);
2073 assert(FiniInfo.DK == OMPD_sections &&
2074 "Unexpected finalization stack state!");
2080 AfterIP = {FiniBB, FiniBB->
begin()};
2094 if (IP.getBlock()->end() != IP.getPoint())
2113 Directive OMPD = Directive::OMPD_sections;
2116 return EmitOMPInlinedRegion(OMPD,
nullptr,
nullptr, BodyGenCB, FiniCBWrapper,
2128 std::vector<WeakTrackingVH> &
List) {
2135 for (
unsigned I = 0, E =
List.size();
I != E; ++
I)
2139 if (UsedArray.
empty())
2146 GV->setSection(
"llvm.metadata");
2149Value *OpenMPIRBuilder::getGPUThreadID() {
2152 OMPRTL___kmpc_get_hardware_thread_id_in_block),
2156Value *OpenMPIRBuilder::getGPUWarpSize() {
2161Value *OpenMPIRBuilder::getNVPTXWarpID() {
2166Value *OpenMPIRBuilder::getNVPTXLaneID() {
2168 assert(LaneIDBits < 32 &&
"Invalid LaneIDBits size in NVPTX device.");
2169 unsigned LaneIDMask = ~0
u >> (32u - LaneIDBits);
2174Value *OpenMPIRBuilder::castValueToType(InsertPointTy AllocaIP,
Value *
From,
2179 assert(FromSize > 0 &&
"From size must be greater than zero");
2180 assert(ToSize > 0 &&
"To size must be greater than zero");
2181 if (FromType == ToType)
2183 if (FromSize == ToSize)
2193 CastItem,
FromType->getPointerTo());
2198Value *OpenMPIRBuilder::createRuntimeShuffleFunction(InsertPointTy AllocaIP,
2203 assert(
Size <= 8 &&
"Unsupported bitwidth in shuffle instruction");
2207 Value *ElemCast = castValueToType(AllocaIP, Element, CastTy);
2211 Size <= 4 ? RuntimeFunction::OMPRTL___kmpc_shuffle_int32
2212 : RuntimeFunction::OMPRTL___kmpc_shuffle_int64);
2213 Value *WarpSizeCast =
2215 Value *ShuffleCall =
2217 return castValueToType(AllocaIP, ShuffleCall, CastTy);
2220void OpenMPIRBuilder::shuffleAndStore(InsertPointTy AllocaIP,
Value *SrcAddr,
2236 Value *ElemPtr = DstAddr;
2238 for (
unsigned IntSize = 8; IntSize >= 1; IntSize /= 2) {
2250 if ((
Size / IntSize) > 1) {
2274 Value *Res = createRuntimeShuffleFunction(
2283 Value *LocalElemPtr =
2290 Value *Res = createRuntimeShuffleFunction(
2304void OpenMPIRBuilder::emitReductionListCopy(
2305 InsertPointTy AllocaIP, CopyAction Action,
Type *ReductionArrayTy,
2307 CopyOptionsTy CopyOptions) {
2310 Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
2314 for (
auto En :
enumerate(ReductionInfos)) {
2315 const ReductionInfo &RI = En.value();
2316 Value *SrcElementAddr =
nullptr;
2317 Value *DestElementAddr =
nullptr;
2318 Value *DestElementPtrAddr =
nullptr;
2320 bool ShuffleInElement =
false;
2323 bool UpdateDestListPtr =
false;
2327 ReductionArrayTy, SrcBase,
2328 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
2334 ReductionArrayTy, DestBase,
2335 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
2341 ".omp.reduction.element");
2344 DestElementAddr = DestAlloca;
2347 DestElementAddr->
getName() +
".ascast");
2349 ShuffleInElement =
true;
2350 UpdateDestListPtr =
true;
2362 if (ShuffleInElement) {
2363 shuffleAndStore(AllocaIP, SrcElementAddr, DestElementAddr, RI.ElementType,
2364 RemoteLaneOffset, ReductionArrayTy);
2366 switch (RI.EvaluationKind) {
2375 RI.ElementType, SrcElementAddr, 0, 0,
".realp");
2377 RI.ElementType->getStructElementType(0), SrcRealPtr,
".real");
2379 RI.ElementType, SrcElementAddr, 0, 1,
".imagp");
2381 RI.ElementType->getStructElementType(1), SrcImgPtr,
".imag");
2384 RI.ElementType, DestElementAddr, 0, 0,
".realp");
2386 RI.ElementType, DestElementAddr, 0, 1,
".imagp");
2408 if (UpdateDestListPtr) {
2411 DestElementAddr->
getName() +
".ascast");
2417Function *OpenMPIRBuilder::emitInterWarpCopyFunction(
2427 "_omp_reduction_inter_warp_copy_func", &
M);
2450 "__openmp_nvptx_data_transfer_temporary_storage";
2454 if (!TransferMedium) {
2463 Value *GPUThreadID = getGPUThreadID();
2465 Value *LaneID = getNVPTXLaneID();
2467 Value *WarpID = getNVPTXWarpID();
2476 Arg0Type,
nullptr, ReduceListArg->
getName() +
".addr");
2480 ReduceListAlloca, Arg0Type, ReduceListAlloca->
getName() +
".ascast");
2483 NumWarpsAlloca->
getName() +
".ascast");
2494 for (
auto En :
enumerate(ReductionInfos)) {
2499 const ReductionInfo &RI = En.value();
2501 for (
unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /= 2) {
2504 unsigned NumIters = RealTySize / TySize;
2507 Value *Cnt =
nullptr;
2508 Value *CntAddr =
nullptr;
2518 CntAddr->
getName() +
".ascast");
2537 omp::Directive::OMPD_unknown,
2550 auto *RedListArrayTy =
2556 {ConstantInt::get(IndexTy, 0),
2557 ConstantInt::get(IndexTy, En.index())});
2582 omp::Directive::OMPD_unknown,
2591 Value *NumWarpsVal =
2594 Value *IsActiveThread =
2605 Value *TargetElemPtrPtr =
2607 {ConstantInt::get(IndexTy, 0),
2608 ConstantInt::get(IndexTy, En.index())});
2609 Value *TargetElemPtrVal =
2611 Value *TargetElemPtr = TargetElemPtrVal;
2617 Value *SrcMediumValue =
2636 RealTySize %= TySize;
2646Function *OpenMPIRBuilder::emitShuffleAndReduceFunction(
2652 {Builder.getPtrTy(), Builder.getInt16Ty(),
2653 Builder.getInt16Ty(), Builder.getInt16Ty()},
2657 "_omp_reduction_shuffle_and_reduce_func", &
M);
2678 Type *ReduceListArgType = ReduceListArg->
getType();
2682 ReduceListArgType,
nullptr, ReduceListArg->
getName() +
".addr");
2684 LaneIDArg->
getName() +
".addr");
2686 LaneIDArgType,
nullptr, RemoteLaneOffsetArg->
getName() +
".addr");
2688 AlgoVerArg->
getName() +
".addr");
2695 RedListArrayTy,
nullptr,
".omp.reduction.remote_reduce_list");
2698 ReduceListAlloca, ReduceListArgType,
2699 ReduceListAlloca->
getName() +
".ascast");
2701 LaneIdAlloca, LaneIDArgPtrType, LaneIdAlloca->
getName() +
".ascast");
2703 RemoteLaneOffsetAlloca, LaneIDArgPtrType,
2704 RemoteLaneOffsetAlloca->
getName() +
".ascast");
2706 AlgoVerAlloca, LaneIDArgPtrType, AlgoVerAlloca->
getName() +
".ascast");
2709 RemoteReductionListAlloca->
getName() +
".ascast");
2718 Value *RemoteLaneOffset =
2727 emitReductionListCopy(
2729 ReduceList, RemoteListAddrCast, {RemoteLaneOffset,
nullptr,
nullptr});
2760 Value *RemoteOffsetComp =
2777 ->addFnAttr(Attribute::NoUnwind);
2798 ReductionInfos, RemoteListAddrCast, ReduceList);
2811Function *OpenMPIRBuilder::emitListToGlobalCopyFunction(
2818 {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
2822 "_omp_reduction_list_to_global_copy_func", &
M);
2839 BufferArg->
getName() +
".addr");
2846 BufferArgAlloca->
getName() +
".ascast");
2851 ReduceListArgAlloca->
getName() +
".ascast");
2857 Value *LocalReduceList =
2859 Value *BufferArgVal =
2864 for (
auto En :
enumerate(ReductionInfos)) {
2865 const ReductionInfo &RI = En.value();
2866 auto *RedListArrayTy =
2870 RedListArrayTy, LocalReduceList,
2871 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
2879 ReductionsBufferTy, BufferVD, 0, En.index());
2881 switch (RI.EvaluationKind) {
2889 RI.ElementType, ElemPtr, 0, 0,
".realp");
2891 RI.ElementType->getStructElementType(0), SrcRealPtr,
".real");
2893 RI.ElementType, ElemPtr, 0, 1,
".imagp");
2895 RI.ElementType->getStructElementType(1), SrcImgPtr,
".imag");
2898 RI.ElementType, GlobVal, 0, 0,
".realp");
2900 RI.ElementType, GlobVal, 0, 1,
".imagp");
2921Function *OpenMPIRBuilder::emitListToGlobalReduceFunction(
2928 {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
2932 "_omp_reduction_list_to_global_reduce_func", &
M);
2949 BufferArg->
getName() +
".addr");
2954 auto *RedListArrayTy =
2959 Value *LocalReduceList =
2964 BufferArgAlloca->
getName() +
".ascast");
2969 ReduceListArgAlloca->
getName() +
".ascast");
2972 LocalReduceList->
getName() +
".ascast");
2982 for (
auto En :
enumerate(ReductionInfos)) {
2984 RedListArrayTy, LocalReduceListAddrCast,
2985 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
2990 ReductionsBufferTy, BufferVD, 0, En.index());
2998 ->addFnAttr(Attribute::NoUnwind);
3004Function *OpenMPIRBuilder::emitGlobalToListCopyFunction(
3011 {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
3015 "_omp_reduction_global_to_list_copy_func", &
M);
3032 BufferArg->
getName() +
".addr");
3039 BufferArgAlloca->
getName() +
".ascast");
3044 ReduceListArgAlloca->
getName() +
".ascast");
3049 Value *LocalReduceList =
3055 for (
auto En :
enumerate(ReductionInfos)) {
3057 auto *RedListArrayTy =
3061 RedListArrayTy, LocalReduceList,
3062 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
3069 ReductionsBufferTy, BufferVD, 0, En.index());
3112Function *OpenMPIRBuilder::emitGlobalToListReduceFunction(
3119 {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
3123 "_omp_reduction_global_to_list_reduce_func", &
M);
3140 BufferArg->
getName() +
".addr");
3150 Value *LocalReduceList =
3155 BufferArgAlloca->
getName() +
".ascast");
3160 ReduceListArgAlloca->
getName() +
".ascast");
3163 LocalReduceList->
getName() +
".ascast");
3173 for (
auto En :
enumerate(ReductionInfos)) {
3175 RedListArrayTy, ReductionList,
3176 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
3181 ReductionsBufferTy, BufferVD, 0, En.index());
3189 ->addFnAttr(Attribute::NoUnwind);
3195std::string OpenMPIRBuilder::getReductionFuncName(
StringRef Name)
const {
3196 std::string Suffix =
3198 return (
Name + Suffix).str();
3201Function *OpenMPIRBuilder::createReductionFunction(
3203 ReductionGenCBKind ReductionGenCBKind,
AttributeList FuncAttrs) {
3205 {Builder.getPtrTy(), Builder.getPtrTy()},
3207 std::string
Name = getReductionFuncName(ReducerName);
3219 Value *LHSArrayPtr =
nullptr;
3220 Value *RHSArrayPtr =
nullptr;
3231 LHSAlloca, Arg0Type, LHSAlloca->
getName() +
".ascast");
3233 RHSAlloca, Arg1Type, RHSAlloca->
getName() +
".ascast");
3243 for (
auto En :
enumerate(ReductionInfos)) {
3244 const ReductionInfo &RI = En.value();
3246 RedArrayTy, RHSArrayPtr,
3247 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
3250 RHSI8Ptr, RI.PrivateVariable->getType(),
3251 RHSI8Ptr->
getName() +
".ascast");
3254 RedArrayTy, LHSArrayPtr,
3255 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
3258 LHSI8Ptr, RI.Variable->getType(), LHSI8Ptr->
getName() +
".ascast");
3269 return ReductionFunc;
3275 for (
auto En :
enumerate(ReductionInfos)) {
3276 unsigned Index = En.index();
3277 const ReductionInfo &RI = En.value();
3278 Value *LHSFixupPtr, *RHSFixupPtr;
3285 LHSPtrs[
Index], [ReductionFunc](
const Use &U) {
3286 return cast<Instruction>(
U.getUser())->getParent()->getParent() ==
3290 RHSPtrs[
Index], [ReductionFunc](
const Use &U) {
3291 return cast<Instruction>(
U.getUser())->getParent()->getParent() ==
3297 return ReductionFunc;
3305 assert(RI.Variable &&
"expected non-null variable");
3306 assert(RI.PrivateVariable &&
"expected non-null private variable");
3307 assert((RI.ReductionGen || RI.ReductionGenClang) &&
3308 "expected non-null reduction generator callback");
3311 RI.Variable->getType() == RI.PrivateVariable->getType() &&
3312 "expected variables and their private equivalents to have the same "
3315 assert(RI.Variable->getType()->isPointerTy() &&
3316 "expected variables to be pointers");
3323 bool IsNoWait,
bool IsTeamsReduction,
bool HasDistribute,
3325 unsigned ReductionBufNum,
Value *SrcLocInfo) {
3339 if (ReductionInfos.
size() == 0)
3358 if (GridValue.has_value())
3381 Value *ReductionListAlloca =
3384 ReductionListAlloca, PtrTy, ReductionListAlloca->
getName() +
".ascast");
3388 for (
auto En :
enumerate(ReductionInfos)) {
3391 RedArrayTy, ReductionList,
3392 {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
3399 emitShuffleAndReduceFunction(ReductionInfos, ReductionFunc, FuncAttrs);
3400 Function *WcFunc = emitInterWarpCopyFunction(Loc, ReductionInfos, FuncAttrs);
3405 unsigned MaxDataSize = 0;
3407 for (
auto En :
enumerate(ReductionInfos)) {
3409 if (
Size > MaxDataSize)
3411 ReductionTypeArgs.
emplace_back(En.value().ElementType);
3413 Value *ReductionDataSize =
3415 if (!IsTeamsReduction) {
3416 Value *SarFuncCast =
3420 Value *Args[] = {RTLoc, ReductionDataSize, RL, SarFuncCast, WcFuncCast};
3422 RuntimeFunction::OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2);
3427 Ctx, ReductionTypeArgs,
"struct._globalized_locals_ty");
3429 RuntimeFunction::OMPRTL___kmpc_reduction_get_fixed_buffer);
3430 Function *LtGCFunc = emitListToGlobalCopyFunction(
3431 ReductionInfos, ReductionsBufferTy, FuncAttrs);
3432 Function *LtGRFunc = emitListToGlobalReduceFunction(
3433 ReductionInfos, ReductionFunc, ReductionsBufferTy, FuncAttrs);
3434 Function *GtLCFunc = emitGlobalToListCopyFunction(
3435 ReductionInfos, ReductionsBufferTy, FuncAttrs);
3436 Function *GtLRFunc = emitGlobalToListReduceFunction(
3437 ReductionInfos, ReductionFunc, ReductionsBufferTy, FuncAttrs);
3441 RedFixedBuferFn, {},
"_openmp_teams_reductions_buffer_$_$ptr");
3443 Value *Args3[] = {RTLoc,
3444 KernelTeamsReductionPtr,
3456 RuntimeFunction::OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2);
3473 for (
auto En :
enumerate(ReductionInfos)) {
3480 Value *LHSPtr, *RHSPtr;
3482 &LHSPtr, &RHSPtr, CurFunc));
3487 return cast<Instruction>(U.getUser())->getParent()->getParent() ==
3491 return cast<Instruction>(U.getUser())->getParent()->getParent() ==
3495 assert(
false &&
"Unhandled ReductionGenCBKind");
3511 ".omp.reduction.func", &M);
3522 assert(RI.Variable &&
"expected non-null variable");
3523 assert(RI.PrivateVariable &&
"expected non-null private variable");
3524 assert(RI.ReductionGen &&
"expected non-null reduction generator callback");
3525 assert(RI.Variable->getType() == RI.PrivateVariable->getType() &&
3526 "expected variables and their private equivalents to have the same "
3528 assert(RI.Variable->getType()->isPointerTy() &&
3529 "expected variables to be pointers");
3542 unsigned NumReductions = ReductionInfos.
size();
3549 for (
auto En :
enumerate(ReductionInfos)) {
3550 unsigned Index = En.index();
3568 ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE
3573 unsigned RedArrayByteSize =
DL.getTypeStoreSize(RedArrayTy);
3576 Value *Lock = getOMPCriticalRegionLock(
".reduction");
3578 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait
3579 : RuntimeFunction::OMPRTL___kmpc_reduce);
3582 {Ident, ThreadId, NumVariables, RedArraySize, RedArray,
3583 ReductionFunc, Lock},
3602 for (
auto En :
enumerate(ReductionInfos)) {
3607 Value *RedValue =
nullptr;
3608 if (!IsByRef[En.index()]) {
3610 "red.value." +
Twine(En.index()));
3612 Value *PrivateRedValue =
3614 "red.private.value." +
Twine(En.index()));
3616 if (IsByRef[En.index()]) {
3618 PrivateRedValue, Reduced));
3621 PrivateRedValue, Reduced));
3626 if (!IsByRef[En.index()])
3630 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait
3631 : RuntimeFunction::OMPRTL___kmpc_end_reduce);
3639 if (CanGenerateAtomic &&
llvm::none_of(IsByRef, [](
bool P) {
return P; })) {
3660 for (
auto En :
enumerate(ReductionInfos)) {
3663 RedArrayTy, LHSArrayPtr, 0, En.index());
3668 RedArrayTy, RHSArrayPtr, 0, En.index());
3678 if (!IsByRef[En.index()])
3695 Directive OMPD = Directive::OMPD_master;
3700 Value *Args[] = {Ident, ThreadId};
3708 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
3719 Directive OMPD = Directive::OMPD_masked;
3725 Value *ArgsEnd[] = {Ident, ThreadId};
3733 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
3768 IndVarPHI->
addIncoming(ConstantInt::get(IndVarTy, 0), Preheader);
3781 "omp_" +
Name +
".next",
true);
3792 CL->Header = Header;
3811 NextBB, NextBB,
Name);
3835 Value *Start,
Value *Stop,
Value *Step,
bool IsSigned,
bool InclusiveStop,
3845 auto *IndVarTy = cast<IntegerType>(Start->getType());
3846 assert(IndVarTy == Stop->
getType() &&
"Stop type mismatch");
3847 assert(IndVarTy == Step->
getType() &&
"Step type mismatch");
3853 ConstantInt *Zero = ConstantInt::get(IndVarTy, 0);
3881 Value *CountIfLooping;
3882 if (InclusiveStop) {
3892 "omp_" +
Name +
".tripcount");
3913 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u);
3916 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u);
3922 InsertPointTy AllocaIP,
3923 bool NeedsBarrier) {
3924 assert(CLI->
isValid() &&
"Requires a valid canonical loop");
3926 "Require dedicated allocate IP");
3938 Type *IVTy =
IV->getType();
3958 Constant *One = ConstantInt::get(IVTy, 1);
3966 Constant *SchedulingType = ConstantInt::get(
3967 I32Type,
static_cast<int>(OMPScheduleType::UnorderedStatic));
3972 {SrcLoc, ThreadNum, SchedulingType, PLastIter, PLowerBound,
3973 PUpperBound, PStride, One,
Zero});
3978 CLI->setTripCount(TripCount);
3999 omp::Directive::OMPD_for,
false,
4010 bool NeedsBarrier,
Value *ChunkSize) {
4011 assert(CLI->
isValid() &&
"Requires a valid canonical loop");
4012 assert(ChunkSize &&
"Chunk size is required");
4017 Type *IVTy =
IV->getType();
4019 "Max supported tripcount bitwidth is 64 bits");
4021 :
Type::getInt64Ty(Ctx);
4024 Constant *One = ConstantInt::get(InternalIVTy, 1);
4036 Value *PLowerBound =
4038 Value *PUpperBound =
4047 Value *CastedChunkSize =
4049 Value *CastedTripCount =
4052 Constant *SchedulingType = ConstantInt::get(
4053 I32Type,
static_cast<int>(OMPScheduleType::UnorderedStaticChunked));
4067 SchedulingType, PLastIter,
4068 PLowerBound, PUpperBound,
4073 Value *FirstChunkStart =
4075 Value *FirstChunkStop =
4080 Value *NextChunkStride =
4085 Value *DispatchCounter;
4089 FirstChunkStart, CastedTripCount, NextChunkStride,
4113 Value *IsLastChunk =
4115 Value *CountUntilOrigTripCount =
4118 IsLastChunk, CountUntilOrigTripCount, ChunkRange,
"omp_chunk.tripcount");
4119 Value *BackcastedChunkTC =
4121 CLI->setTripCount(BackcastedChunkTC);
4126 Value *BackcastedDispatchCounter =
4160 case WorksharingLoopType::ForStaticLoop:
4163 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_4u);
4166 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_8u);
4168 case WorksharingLoopType::DistributeStaticLoop:
4171 M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_4u);
4174 M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_8u);
4176 case WorksharingLoopType::DistributeForStaticLoop:
4179 M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_4u);
4182 M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_8u);
4185 if (Bitwidth != 32 && Bitwidth != 64) {
4207 if (LoopType == WorksharingLoopType::DistributeStaticLoop) {
4208 RealArgs.
push_back(ConstantInt::get(TripCountTy, 0));
4213 M, omp::RuntimeFunction::OMPRTL_omp_get_num_threads);
4214 Builder.
restoreIP({InsertBlock, std::prev(InsertBlock->
end())});
4219 RealArgs.
push_back(ConstantInt::get(TripCountTy, 0));
4220 if (LoopType == WorksharingLoopType::DistributeForStaticLoop) {
4221 RealArgs.
push_back(ConstantInt::get(TripCountTy, 0));
4257 CleanUpInfo.
collectBlocks(RegionBlockSet, BlocksToBeRemoved);
4265 "Expected unique undroppable user of outlined function");
4266 CallInst *OutlinedFnCallInstruction = dyn_cast<CallInst>(OutlinedFnUser);
4267 assert(OutlinedFnCallInstruction &&
"Expected outlined function call");
4269 "Expected outlined function call to be located in loop preheader");
4271 if (OutlinedFnCallInstruction->
arg_size() > 1)
4278 LoopBodyArg, ParallelTaskPtr, TripCount,
4281 for (
auto &ToBeDeletedItem : ToBeDeleted)
4282 ToBeDeletedItem->eraseFromParent();
4288 InsertPointTy AllocaIP,
4301 OI.OuterAllocaBB = AllocaIP.getBlock();
4306 "omp.prelatch",
true);
4326 OI.collectBlocks(ParallelRegionBlockSet,
Blocks);
4328 ParallelRegionBlockSet.
end());
4348 Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
4357 if (ParallelRegionBlockSet.
count(Inst->getParent())) {
4358 Inst->replaceUsesOfWith(CLI->
getIndVar(), NewLoopCntLoad);
4364 OI.ExcludeArgsFromAggregate.push_back(NewLoopCntLoad);
4371 OI.PostOutlineCB = [=, ToBeDeletedVec =
4372 std::move(ToBeDeleted)](
Function &OutlinedFn) {
4374 ToBeDeletedVec, LoopType);
4382 bool NeedsBarrier, omp::ScheduleKind SchedKind,
Value *ChunkSize,
4383 bool HasSimdModifier,
bool HasMonotonicModifier,
4384 bool HasNonmonotonicModifier,
bool HasOrderedClause,
4387 return applyWorkshareLoopTarget(
DL, CLI, AllocaIP, LoopType);
4389 SchedKind, ChunkSize, HasSimdModifier, HasMonotonicModifier,
4390 HasNonmonotonicModifier, HasOrderedClause);
4392 bool IsOrdered = (EffectiveScheduleType & OMPScheduleType::ModifierOrdered) ==
4393 OMPScheduleType::ModifierOrdered;
4394 switch (EffectiveScheduleType & ~OMPScheduleType::ModifierMask) {
4395 case OMPScheduleType::BaseStatic:
4396 assert(!ChunkSize &&
"No chunk size with static-chunked schedule");
4398 return applyDynamicWorkshareLoop(
DL, CLI, AllocaIP, EffectiveScheduleType,
4399 NeedsBarrier, ChunkSize);
4401 return applyStaticWorkshareLoop(
DL, CLI, AllocaIP, NeedsBarrier);
4403 case OMPScheduleType::BaseStaticChunked:
4405 return applyDynamicWorkshareLoop(
DL, CLI, AllocaIP, EffectiveScheduleType,
4406 NeedsBarrier, ChunkSize);
4408 return applyStaticChunkedWorkshareLoop(
DL, CLI, AllocaIP, NeedsBarrier,
4411 case OMPScheduleType::BaseRuntime:
4412 case OMPScheduleType::BaseAuto:
4413 case OMPScheduleType::BaseGreedy:
4414 case OMPScheduleType::BaseBalanced:
4415 case OMPScheduleType::BaseSteal:
4416 case OMPScheduleType::BaseGuidedSimd:
4417 case OMPScheduleType::BaseRuntimeSimd:
4419 "schedule type does not support user-defined chunk sizes");
4421 case OMPScheduleType::BaseDynamicChunked:
4422 case OMPScheduleType::BaseGuidedChunked:
4423 case OMPScheduleType::BaseGuidedIterativeChunked:
4424 case OMPScheduleType::BaseGuidedAnalyticalChunked:
4425 case OMPScheduleType::BaseStaticBalancedChunked:
4426 return applyDynamicWorkshareLoop(
DL, CLI, AllocaIP, EffectiveScheduleType,
4427 NeedsBarrier, ChunkSize);
4443 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u);
4446 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u);
4459 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u);
4462 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u);
4474 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_4u);
4477 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_8u);
4484 assert(CLI->
isValid() &&
"Requires a valid canonical loop");
4486 "Require dedicated allocate IP");
4488 "Require valid schedule type");
4490 bool Ordered = (SchedType & OMPScheduleType::ModifierOrdered) ==
4491 OMPScheduleType::ModifierOrdered;
4502 Type *IVTy =
IV->getType();
4520 Constant *One = ConstantInt::get(IVTy, 1);
4541 ConstantInt::get(I32Type,
static_cast<int>(SchedType));
4545 {SrcLoc, ThreadNum, SchedulingType, One,
4546 UpperBound, One, Chunk});
4556 PLowerBound, PUpperBound, PStride});
4557 Constant *Zero32 = ConstantInt::get(I32Type, 0);
4566 auto *PI = cast<PHINode>(Phi);
4567 PI->setIncomingBlock(0, OuterCond);
4568 PI->setIncomingValue(0, LowerBound);
4572 auto *Br = cast<BranchInst>(Term);
4573 Br->setSuccessor(0, OuterCond);
4581 auto *CI = cast<CmpInst>(Comp);
4582 CI->setOperand(1, UpperBound);
4585 auto *BI = cast<BranchInst>(Branch);
4586 assert(BI->getSuccessor(1) == Exit);
4587 BI->setSuccessor(1, OuterCond);
4600 omp::Directive::OMPD_for,
false,
4620 auto HasRemainingUses = [&BBsToErase](
BasicBlock *BB) {
4621 for (
Use &U : BB->uses()) {
4622 auto *UseInst = dyn_cast<Instruction>(U.getUser());
4625 if (BBsToErase.count(UseInst->getParent()))
4632 while (BBsToErase.remove_if(HasRemainingUses)) {
4643 assert(
Loops.size() >= 1 &&
"At least one loop required");
4644 size_t NumLoops =
Loops.size();
4648 return Loops.front();
4660 Loop->collectControlBlocks(OldControlBBs);
4664 if (ComputeIP.
isSet())
4671 Value *CollapsedTripCount =
nullptr;
4674 "All loops to collapse must be valid canonical loops");
4675 Value *OrigTripCount = L->getTripCount();
4676 if (!CollapsedTripCount) {
4677 CollapsedTripCount = OrigTripCount;
4689 OrigPreheader->
getNextNode(), OrigAfter,
"collapsed");
4697 Value *Leftover = Result->getIndVar();
4699 NewIndVars.
resize(NumLoops);
4700 for (
int i = NumLoops - 1; i >= 1; --i) {
4701 Value *OrigTripCount =
Loops[i]->getTripCount();
4704 NewIndVars[i] = NewIndVar;
4709 NewIndVars[0] = Leftover;
4718 BasicBlock *ContinueBlock = Result->getBody();
4720 auto ContinueWith = [&ContinueBlock, &ContinuePred,
DL](
BasicBlock *Dest,
4727 ContinueBlock =
nullptr;
4728 ContinuePred = NextSrc;
4735 for (
size_t i = 0; i < NumLoops - 1; ++i)
4736 ContinueWith(
Loops[i]->getBody(),
Loops[i + 1]->getHeader());
4742 for (
size_t i = NumLoops - 1; i > 0; --i)
4743 ContinueWith(
Loops[i]->getAfter(),
Loops[i - 1]->getLatch());
4746 ContinueWith(Result->getLatch(),
nullptr);
4753 for (
size_t i = 0; i < NumLoops; ++i)
4754 Loops[i]->getIndVar()->replaceAllUsesWith(NewIndVars[i]);
4768std::vector<CanonicalLoopInfo *>
4772 "Must pass as many tile sizes as there are loops");
4773 int NumLoops =
Loops.size();
4774 assert(NumLoops >= 1 &&
"At least one loop to tile required");
4786 Loop->collectControlBlocks(OldControlBBs);
4794 assert(L->isValid() &&
"All input loops must be valid canonical loops");
4795 OrigTripCounts.
push_back(L->getTripCount());
4806 for (
int i = 0; i < NumLoops - 1; ++i) {
4819 for (
int i = 0; i < NumLoops; ++i) {
4821 Value *OrigTripCount = OrigTripCounts[i];
4834 Value *FloorTripOverflow =
4840 "omp_floor" +
Twine(i) +
".tripcount",
true);
4848 std::vector<CanonicalLoopInfo *> Result;
4849 Result.reserve(NumLoops * 2);
4862 auto EmbeddNewLoop =
4863 [
this,
DL,
F, InnerEnter, &Enter, &
Continue, &OutroInsertBefore](
4866 DL, TripCount,
F, InnerEnter, OutroInsertBefore,
Name);
4871 Enter = EmbeddedLoop->
getBody();
4873 OutroInsertBefore = EmbeddedLoop->
getLatch();
4874 return EmbeddedLoop;
4878 const Twine &NameBase) {
4881 EmbeddNewLoop(
P.value(), NameBase +
Twine(
P.index()));
4882 Result.push_back(EmbeddedLoop);
4886 EmbeddNewLoops(FloorCount,
"floor");
4892 for (
int i = 0; i < NumLoops; ++i) {
4896 Value *FloorIsEpilogue =
4898 Value *TileTripCount =
4905 EmbeddNewLoops(TileCounts,
"tile");
4910 for (std::pair<BasicBlock *, BasicBlock *>
P : InbetweenCode) {
4919 BodyEnter =
nullptr;
4920 BodyEntered = ExitBB;
4933 for (
int i = 0; i < NumLoops; ++i) {
4936 Value *OrigIndVar = OrigIndVars[i];
4964 if (Properties.
empty())
4987 assert(
Loop->isValid() &&
"Expecting a valid CanonicalLoopInfo");
4991 assert(Latch &&
"A valid CanonicalLoopInfo must have a unique latch");
4999 if (
I.mayReadOrWriteMemory()) {
5003 I.setMetadata(LLVMContext::MD_access_group, AccessGroup);
5025 const Twine &NamePrefix) {
5031 SplitBefore = dyn_cast<Instruction>(IfCond);
5077 VMap[
Block] = NewBB;
5087 if (TargetTriple.
isX86()) {
5088 if (Features.
lookup(
"avx512f"))
5090 else if (Features.
lookup(
"avx"))
5094 if (TargetTriple.
isPPC())
5096 if (TargetTriple.
isWasm())
5103 Value *IfCond, OrderKind Order,
5122 if (AlignedVars.
size()) {
5125 for (
auto &AlignedItem : AlignedVars) {
5126 Value *AlignedPtr = AlignedItem.first;
5127 Value *Alignment = AlignedItem.second;
5129 AlignedPtr, Alignment);
5136 createIfVersion(CanonicalLoop, IfCond, VMap,
"simd");
5140 "Cannot find value which corresponds to original loop latch");
5141 assert(isa<BasicBlock>(MappedLatch) &&
5142 "Cannot cast mapped latch block value to BasicBlock");
5143 BasicBlock *NewLatchBlock = dyn_cast<BasicBlock>(MappedLatch);
5172 if ((Safelen ==
nullptr) || (Order == OrderKind::OMP_ORDER_concurrent)) {
5180 Ctx, {
MDString::get(Ctx,
"llvm.loop.parallel_accesses"), AccessGroup}));
5188 Ctx, {
MDString::get(Ctx,
"llvm.loop.vectorize.enable"), BoolConst}));
5190 if (Simdlen || Safelen) {
5194 ConstantInt *VectorizeWidth = Simdlen ==
nullptr ? Safelen : Simdlen;
5220static std::unique_ptr<TargetMachine>
5224 StringRef CPU =
F->getFnAttribute(
"target-cpu").getValueAsString();
5225 StringRef Features =
F->getFnAttribute(
"target-features").getValueAsString();
5226 const std::string &
Triple = M->getTargetTriple();
5236 std::nullopt, OptLevel));
5260 [&](
const Function &
F) {
return TM->getTargetTransformInfo(
F); });
5275 assert(L &&
"Expecting CanonicalLoopInfo to be recognized as a loop");
5280 nullptr, ORE,
static_cast<int>(OptLevel),
5301 <<
" Threshold=" << UP.
Threshold <<
"\n"
5304 <<
" PartialOptSizeThreshold="
5323 if (
auto *Load = dyn_cast<LoadInst>(&
I)) {
5324 Ptr = Load->getPointerOperand();
5325 }
else if (
auto *Store = dyn_cast<StoreInst>(&
I)) {
5326 Ptr = Store->getPointerOperand();
5330 Ptr =
Ptr->stripPointerCasts();
5332 if (
auto *Alloca = dyn_cast<AllocaInst>(
Ptr)) {
5333 if (Alloca->getParent() == &
F->getEntryBlock())
5353 int MaxTripCount = 0;
5354 bool MaxOrZero =
false;
5355 unsigned TripMultiple = 0;
5357 bool UseUpperBound =
false;
5359 MaxTripCount, MaxOrZero, TripMultiple, UCE, UP, PP,
5361 unsigned Factor = UP.
Count;
5362 LLVM_DEBUG(
dbgs() <<
"Suggesting unroll factor of " << Factor <<
"\n");
5373 assert(Factor >= 0 &&
"Unroll factor must not be negative");
5389 Ctx, {
MDString::get(Ctx,
"llvm.loop.unroll.count"), FactorConst}));
5402 *UnrolledCLI =
Loop;
5407 "unrolling only makes sense with a factor of 2 or larger");
5409 Type *IndVarTy =
Loop->getIndVarType();
5416 std::vector<CanonicalLoopInfo *>
LoopNest =
5431 Ctx, {
MDString::get(Ctx,
"llvm.loop.unroll.count"), FactorConst})});
5434 (*UnrolledCLI)->assertOK();
5452 Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD};
5471 if (!CPVars.
empty()) {
5476 Directive OMPD = Directive::OMPD_single;
5481 Value *Args[] = {Ident, ThreadId};
5507 EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCBWrapper,
5512 for (
size_t I = 0, E = CPVars.
size();
I < E; ++
I)
5515 ConstantInt::get(
Int64, 0), CPVars[
I],
5518 }
else if (!IsNowait)
5520 omp::Directive::OMPD_unknown,
false,
5532 Directive OMPD = Directive::OMPD_critical;
5537 Value *LockVar = getOMPCriticalRegionLock(CriticalName);
5538 Value *Args[] = {Ident, ThreadId, LockVar};
5555 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
5563 const Twine &
Name,
bool IsDependSource) {
5566 [](
Value *SV) {
return SV->
getType()->isIntegerTy(64); }) &&
5567 "OpenMP runtime requires depend vec with i64 type");
5580 for (
unsigned I = 0;
I < NumLoops; ++
I) {
5594 Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP};
5612 Directive OMPD = Directive::OMPD_ordered;
5621 Value *Args[] = {Ident, ThreadId};
5631 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
5637 BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB,
bool Conditional,
5638 bool HasFinalize,
bool IsCancellable) {
5647 if (!isa_and_nonnull<BranchInst>(SplitPos))
5654 emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional);
5664 "Unexpected control flow graph state!!");
5665 emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize);
5667 "Unexpected Control Flow State!");
5673 "Unexpected Insertion point location!");
5676 auto InsertBB = merged ? ExitPredBB : ExitBB;
5677 if (!isa_and_nonnull<BranchInst>(SplitPos))
5687 if (!Conditional || !EntryCall)
5707 UI->eraseFromParent();
5715 omp::Directive OMPD, InsertPointTy FinIP,
Instruction *ExitCall,
5723 "Unexpected finalization stack state!");
5726 assert(Fi.DK == OMPD &&
"Unexpected Directive for Finalization call!");
5776 if (isa_and_nonnull<BranchInst>(OMP_Entry->
getTerminator())) {
5778 "copyin.not.master.end");
5833 Value *DependenceAddress,
bool HaveNowaitClause) {
5841 if (Device ==
nullptr)
5842 Device = ConstantInt::get(
Int32, -1);
5843 Constant *InteropTypeVal = ConstantInt::get(
Int32, (
int)InteropType);
5844 if (NumDependences ==
nullptr) {
5845 NumDependences = ConstantInt::get(
Int32, 0);
5849 Value *HaveNowaitClauseVal = ConstantInt::get(
Int32, HaveNowaitClause);
5851 Ident, ThreadId, InteropVar, InteropTypeVal,
5852 Device, NumDependences, DependenceAddress, HaveNowaitClauseVal};
5861 Value *NumDependences,
Value *DependenceAddress,
bool HaveNowaitClause) {
5869 if (Device ==
nullptr)
5870 Device = ConstantInt::get(
Int32, -1);
5871 if (NumDependences ==
nullptr) {
5872 NumDependences = ConstantInt::get(
Int32, 0);
5876 Value *HaveNowaitClauseVal = ConstantInt::get(
Int32, HaveNowaitClause);
5878 Ident, ThreadId, InteropVar, Device,
5879 NumDependences, DependenceAddress, HaveNowaitClauseVal};
5888 Value *NumDependences,
5889 Value *DependenceAddress,
5890 bool HaveNowaitClause) {
5897 if (Device ==
nullptr)
5898 Device = ConstantInt::get(
Int32, -1);
5899 if (NumDependences ==
nullptr) {
5900 NumDependences = ConstantInt::get(
Int32, 0);
5904 Value *HaveNowaitClauseVal = ConstantInt::get(
Int32, HaveNowaitClause);
5906 Ident, ThreadId, InteropVar, Device,
5907 NumDependences, DependenceAddress, HaveNowaitClauseVal};
5936 int32_t MinThreadsVal, int32_t MaxThreadsVal,
5937 int32_t MinTeamsVal, int32_t MaxTeamsVal) {
5954 if (MinTeamsVal > 1 || MaxTeamsVal > 0)
5958 if (MaxThreadsVal < 0)
5959 MaxThreadsVal = std::max(
5962 if (MaxThreadsVal > 0)
5974 const std::string DebugPrefix =
"_debug__";
5976 KernelName = KernelName.
drop_back(DebugPrefix.length());
5979 omp::RuntimeFunction::OMPRTL___kmpc_target_init);
5982 Twine DynamicEnvironmentName = KernelName +
"_dynamic_environment";
5983 Constant *DynamicEnvironmentInitializer =
5987 DynamicEnvironmentInitializer, DynamicEnvironmentName,
5989 DL.getDefaultGlobalsAddressSpace());
5993 DynamicEnvironmentGV->
getType() == DynamicEnvironmentPtr
5994 ? DynamicEnvironmentGV
5996 DynamicEnvironmentPtr);
5999 ConfigurationEnvironment, {
6000 UseGenericStateMachineVal,
6001 MayUseNestedParallelismVal,
6008 ReductionBufferLength,
6011 KernelEnvironment, {
6012 ConfigurationEnvironmentInitializer,
6016 std::string KernelEnvironmentName =
6017 (KernelName +
"_kernel_environment").str();
6020 KernelEnvironmentInitializer, KernelEnvironmentName,
6022 DL.getDefaultGlobalsAddressSpace());
6026 KernelEnvironmentGV->
getType() == KernelEnvironmentPtr
6027 ? KernelEnvironmentGV
6029 KernelEnvironmentPtr);
6035 ThreadKind, ConstantInt::get(ThreadKind->
getType(), -1),
6058 UI->eraseFromParent();
6066 int32_t TeamsReductionDataSize,
6067 int32_t TeamsReductionBufferLength) {
6072 omp::RuntimeFunction::OMPRTL___kmpc_target_deinit);
6076 if (!TeamsReductionBufferLength || !TeamsReductionDataSize)
6082 const std::string DebugPrefix =
"_debug__";
6084 KernelName = KernelName.
drop_back(DebugPrefix.length());
6085 auto *KernelEnvironmentGV =
6087 assert(KernelEnvironmentGV &&
"Expected kernel environment global\n");
6088 auto *KernelEnvironmentInitializer = KernelEnvironmentGV->getInitializer();
6090 KernelEnvironmentInitializer,
6091 ConstantInt::get(
Int32, TeamsReductionDataSize), {0, 7});
6093 NewInitializer, ConstantInt::get(
Int32, TeamsReductionBufferLength),
6095 KernelEnvironmentGV->setInitializer(NewInitializer);
6100 NamedMDNode *MD = M.getOrInsertNamedMetadata(
"nvvm.annotations");
6104 auto *KernelOp = dyn_cast<ConstantAsMetadata>(
Op->getOperand(0));
6105 if (!KernelOp || KernelOp->getValue() != &
Kernel)
6107 auto *Prop = dyn_cast<MDString>(
Op->getOperand(1));
6108 if (!Prop || Prop->getString() !=
Name)
6120 auto *OldVal = cast<ConstantAsMetadata>(ExistingOp->
getOperand(2));
6121 int32_t OldLimit = cast<ConstantInt>(OldVal->getValue())->getZExtValue();
6124 OldVal->getValue()->getType(),
6125 Min ? std::min(OldLimit,
Value) : std::max(OldLimit,
Value))));
6134 NamedMDNode *MD = M.getOrInsertNamedMetadata(
"nvvm.annotations");
6139std::pair<int32_t, int32_t>
6141 int32_t ThreadLimit =
6146 if (!Attr.isValid() || !Attr.isStringAttribute())
6147 return {0, ThreadLimit};
6150 if (!llvm::to_integer(UBStr, UB, 10))
6151 return {0, ThreadLimit};
6152 UB = ThreadLimit ? std::min(ThreadLimit, UB) : UB;
6153 if (!llvm::to_integer(LBStr, LB, 10))
6159 auto *OldVal = cast<ConstantAsMetadata>(ExistingOp->getOperand(2));
6160 int32_t UB = cast<ConstantInt>(OldVal->getValue())->getZExtValue();
6161 return {0, ThreadLimit ? std::min(ThreadLimit, UB) : UB};
6163 return {0, ThreadLimit};
6173 llvm::utostr(LB) +
"," + llvm::utostr(UB));
6180std::pair<int32_t, int32_t>
6187 int32_t LB, int32_t UB) {
6192 Kernel.
addFnAttr(
"amdgpu-max-num-workgroups", llvm::utostr(LB) +
",1,1");
6197void OpenMPIRBuilder::setOutlinedTargetRegionFunctionAttributes(
6212 assert(OutlinedFn &&
"The outlined function must exist if embedded");
6221Constant *OpenMPIRBuilder::createTargetRegionEntryAddr(
Function *OutlinedFn,
6227 "Named kernel already exists?");
6242 ? GenerateFunctionCallback(EntryFnName)
6248 if (!IsOffloadEntry)
6251 std::string EntryFnIDName =
6253 ? std::string(EntryFnName)
6257 EntryFnName, EntryFnIDName);
6264 setOutlinedTargetRegionFunctionAttributes(OutlinedFn);
6265 auto OutlinedFnID = createOutlinedFunctionID(OutlinedFn, EntryFnIDName);
6266 auto EntryAddr = createTargetRegionEntryAddr(OutlinedFn, EntryFnName);
6268 EntryInfo, EntryAddr, OutlinedFnID,
6270 return OutlinedFnID;
6293 bool IsStandAlone = !BodyGenCB;
6318 Value *OffloadingArgs[] = {SrcLocInfo, DeviceID,
6325 assert(MapperFunc &&
"MapperFunc missing for standalone target data");
6330 omp::OMPRTL___tgt_target_data_begin_mapper);
6334 for (
auto DeviceMap :
Info.DevicePtrInfoMap) {
6335 if (isa<AllocaInst>(DeviceMap.second.second)) {
6372 Value *OffloadingArgs[] = {SrcLocInfo, DeviceID,
6389 emitIfClause(IfCond, BeginThenGen, BeginElseGen, AllocaIP);
6399 emitIfClause(IfCond, EndThenGen, EndElseGen, AllocaIP);
6405 emitIfClause(IfCond, BeginThenGen, EndElseGen, AllocaIP);
6416 bool IsGPUDistribute) {
6417 assert((IVSize == 32 || IVSize == 64) &&
6418 "IV size is not compatible with the omp runtime");
6420 if (IsGPUDistribute)
6422 ? (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_4
6423 : omp::OMPRTL___kmpc_distribute_static_init_4u)
6424 : (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_8
6425 : omp::OMPRTL___kmpc_distribute_static_init_8u);
6427 Name = IVSize == 32 ? (IVSigned ? omp::OMPRTL___kmpc_for_static_init_4
6428 : omp::OMPRTL___kmpc_for_static_init_4u)
6429 : (IVSigned ? omp::OMPRTL___kmpc_for_static_init_8
6430 : omp::OMPRTL___kmpc_for_static_init_8u);
6437 assert((IVSize == 32 || IVSize == 64) &&
6438 "IV size is not compatible with the omp runtime");
6440 ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_4
6441 : omp::OMPRTL___kmpc_dispatch_init_4u)
6442 : (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_8
6443 : omp::OMPRTL___kmpc_dispatch_init_8u);
6450 assert((IVSize == 32 || IVSize == 64) &&
6451 "IV size is not compatible with the omp runtime");
6453 ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_4
6454 : omp::OMPRTL___kmpc_dispatch_next_4u)
6455 : (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_8
6456 : omp::OMPRTL___kmpc_dispatch_next_8u);
6463 assert((IVSize == 32 || IVSize == 64) &&
6464 "IV size is not compatible with the omp runtime");
6466 ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_4
6467 : omp::OMPRTL___kmpc_dispatch_fini_4u)
6468 : (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_8
6469 : omp::OMPRTL___kmpc_dispatch_fini_8u);
6488 for (
auto &Arg : Inputs)
6489 ParameterTypes.
push_back(Arg->getType()->isPointerTy()
6493 for (
auto &Arg : Inputs)
6494 ParameterTypes.
push_back(Arg->getType());
6503 auto OldInsertPoint = Builder.
saveIP();
6531 auto AllocaIP = Builder.
saveIP();
6536 const auto &ArgRange =
6538 ?
make_range(Func->arg_begin() + 1, Func->arg_end())
6559 if (
auto *Const = dyn_cast<Constant>(Input))
6564 if (
auto *Instr = dyn_cast<Instruction>(
User))
6565 if (Instr->getFunction() == Func)
6566 Instr->replaceUsesOfWith(Input, InputCopy);
6572 for (
auto InArg :
zip(Inputs, ArgRange)) {
6573 Value *Input = std::get<0>(InArg);
6574 Argument &Arg = std::get<1>(InArg);
6575 Value *InputCopy =
nullptr;
6578 ArgAccessorFuncCB(Arg, Input, InputCopy, AllocaIP, Builder.
saveIP()));
6596 if (llvm::isa<llvm::GlobalValue>(std::get<0>(InArg)) ||
6597 llvm::isa<llvm::GlobalObject>(std::get<0>(InArg)) ||
6598 llvm::isa<llvm::GlobalVariable>(std::get<0>(InArg))) {
6599 DeferredReplacement.
push_back(std::make_pair(Input, InputCopy));
6603 ReplaceValue(Input, InputCopy, Func);
6607 for (
auto Deferred : DeferredReplacement)
6608 ReplaceValue(std::get<0>(Deferred), std::get<1>(Deferred), Func);
6624 [&OMPBuilder, &Builder, &Inputs, &CBFunc,
6625 &ArgAccessorFuncCB](
StringRef EntryFnName) {
6627 CBFunc, ArgAccessorFuncCB);
6631 OutlinedFn, OutlinedFnID);
6637 int32_t NumTeams, int32_t NumThreads,
6654 auto &&EmitTargetCallFallbackCB =
6675 bool HasNoWait =
false;
6678 NumTeamsVal, NumThreadsVal,
6679 DynCGGroupMem, HasNoWait);
6682 Builder, OutlinedFn, OutlinedFnID, EmitTargetCallFallbackCB, KArgs,
6683 DeviceID, RTLoc, AllocaIP));
6701 OutlinedFnID, Args, CBFunc, ArgAccessorFuncCB);
6704 NumThreads, Args, GenMapInfoCB);
6719 return OS.str().str();
6733 assert(Elem.second->getValueType() == Ty &&
6734 "OMP internal variable has different type than requested");
6750 GV->setAlignment(std::max(TypeAlign, PtrAlign));
6757Value *OpenMPIRBuilder::getOMPCriticalRegionLock(
StringRef CriticalName) {
6758 std::string Prefix =
Twine(
"gomp_critical_user_", CriticalName).
str();
6759 std::string
Name = getNameWithSeparators({Prefix,
"var"},
".",
".");
6770 return SizePtrToInt;
6775 std::string VarName) {
6783 return MaptypesArrayGlobal;
6788 unsigned NumOperands,
6797 ArrI8PtrTy,
nullptr,
".offload_baseptrs");
6801 ArrI64Ty,
nullptr,
".offload_sizes");
6812 int64_t DeviceID,
unsigned NumOperands) {
6818 Value *ArgsBaseGEP =
6820 {Builder.getInt32(0), Builder.getInt32(0)});
6823 {Builder.getInt32(0), Builder.getInt32(0)});
6824 Value *ArgSizesGEP =
6826 {Builder.getInt32(0), Builder.getInt32(0)});
6832 ArgSizesGEP, MaptypesArg, MapnamesArg, NullPtr});
6840 assert((!ForEndCall ||
Info.separateBeginEndCalls()) &&
6841 "expected region end call to runtime only when end call is separate");
6843 auto VoidPtrTy = UnqualPtrTy;
6844 auto VoidPtrPtrTy = UnqualPtrTy;
6846 auto Int64PtrTy = UnqualPtrTy;
6848 if (!
Info.NumberOfPtrs) {
6860 Info.RTArgs.BasePointersArray,
6871 ForEndCall &&
Info.RTArgs.MapTypesArrayEnd ?
Info.RTArgs.MapTypesArrayEnd
6872 :
Info.RTArgs.MapTypesArray,
6887 if (!
Info.HasMapper)
6912 "struct.descriptor_dim");
6914 enum { OffsetFD = 0, CountFD, StrideFD };
6918 for (
unsigned I = 0, L = 0, E = NonContigInfo.
Dims.
size();
I < E; ++
I) {
6921 if (NonContigInfo.
Dims[
I] == 1)
6928 for (
unsigned II = 0, EE = NonContigInfo.
Dims[
I];
II < EE; ++
II) {
6929 unsigned RevIdx = EE -
II - 1;
6932 {Builder.getInt64(0), Builder.getInt64(II)});
6936 NonContigInfo.
Offsets[L][RevIdx], OffsetLVal,
6941 NonContigInfo.
Counts[L][RevIdx], CountLVal,
6946 NonContigInfo.
Strides[L][RevIdx], StrideLVal,
6955 Info.RTArgs.PointersArray, 0,
I);
6969 Info.clearArrayInfo();
6972 if (
Info.NumberOfPtrs == 0)
6982 PointerArrayType,
nullptr,
".offload_baseptrs");
6985 PointerArrayType,
nullptr,
".offload_ptrs");
6987 PointerArrayType,
nullptr,
".offload_mappers");
6988 Info.RTArgs.MappersArray = MappersArray;
6995 ConstantInt::get(Int64Ty, 0));
6997 for (
unsigned I = 0, E = CombinedInfo.
Sizes.
size();
I < E; ++
I) {
6998 if (
auto *CI = dyn_cast<Constant>(CombinedInfo.
Sizes[
I])) {
6999 if (!isa<ConstantExpr>(CI) && !isa<GlobalValue>(CI)) {
7000 if (IsNonContiguous &&
7001 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>
>(
7003 OpenMPOffloadMappingFlags::OMP_MAP_NON_CONTIG))
7011 RuntimeSizes.
set(
I);
7014 if (RuntimeSizes.
all()) {
7017 SizeArrayType,
nullptr,
".offload_sizes");
7023 auto *SizesArrayGbl =
7028 if (!RuntimeSizes.
any()) {
7029 Info.RTArgs.SizesArray = SizesArrayGbl;
7035 SizeArrayType,
nullptr,
".offload_sizes");
7040 SizesArrayGbl, OffloadSizeAlign,
7045 Info.RTArgs.SizesArray = Buffer;
7053 for (
auto mapFlag : CombinedInfo.
Types)
7055 static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>
>(
7059 Info.RTArgs.MapTypesArray = MapTypesArrayGbl;
7064 auto *MapNamesArrayGbl =
7066 Info.RTArgs.MapNamesArray = MapNamesArrayGbl;
7068 Info.RTArgs.MapNamesArray =
7074 if (
Info.separateBeginEndCalls()) {
7075 bool EndMapTypesDiffer =
false;
7077 if (
Type &
static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>
>(
7078 OpenMPOffloadMappingFlags::OMP_MAP_PRESENT)) {
7079 Type &= ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>(
7080 OpenMPOffloadMappingFlags::OMP_MAP_PRESENT);
7081 EndMapTypesDiffer =
true;
7084 if (EndMapTypesDiffer) {
7086 Info.RTArgs.MapTypesArrayEnd = MapTypesArrayGbl;
7091 for (
unsigned I = 0;
I <
Info.NumberOfPtrs; ++
I) {
7099 if (
Info.requiresDevicePointerInfo()) {
7106 DeviceAddrCB(
I,
Info.DevicePtrInfoMap[BPVal].second);
7108 Info.DevicePtrInfoMap[BPVal] = {BP, BP};
7110 DeviceAddrCB(
I, BP);
7122 if (RuntimeSizes.
test(
I)) {
7136 if (
Value *CustomMFunc = CustomMapperCB(
I))
7140 {Builder.getIntN(IndexSize, 0), Builder.getIntN(IndexSize, I)});
7146 Info.NumberOfPtrs == 0)
7191 if (
auto *CI = dyn_cast<ConstantInt>(
Cond)) {
7192 auto CondConstant = CI->getSExtValue();
7222bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic(
7226 "Unexpected Atomic Ordering.");
7290 assert(
X.Var->getType()->isPointerTy() &&
7291 "OMP Atomic expects a pointer to target memory");
7292 Type *XElemTy =
X.ElemTy;
7295 "OMP atomic read expected a scalar type");
7297 Value *XRead =
nullptr;
7303 XRead = cast<Value>(XLD);
7317 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Read);
7329 assert(
X.Var->getType()->isPointerTy() &&
7330 "OMP Atomic expects a pointer to target memory");
7331 Type *XElemTy =
X.ElemTy;
7334 "OMP atomic write expected a scalar type");
7349 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Write);
7362 Type *XTy =
X.Var->getType();
7364 "OMP Atomic expects a pointer to target memory");
7365 Type *XElemTy =
X.ElemTy;
7368 "OMP atomic update expected a scalar type");
7371 "OpenMP atomic does not support LT or GT operations");
7374 emitAtomicUpdate(AllocaIP,
X.Var,
X.ElemTy, Expr, AO, RMWOp, UpdateOp,
7375 X.IsVolatile, IsXBinopExpr);
7376 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Update);
7381Value *OpenMPIRBuilder::emitRMWOpAsInstruction(
Value *Src1,
Value *Src2,
7413std::pair<Value *, Value *> OpenMPIRBuilder::emitAtomicUpdate(
7416 AtomicUpdateCallbackTy &UpdateOp,
bool VolatileX,
bool IsXBinopExpr) {
7419 bool emitRMWOp =
false;
7427 emitRMWOp = XElemTy;
7430 emitRMWOp = (IsXBinopExpr && XElemTy);
7437 std::pair<Value *, Value *> Res;
7444 Res.second = Res.first;
7446 Res.second = emitRMWOpAsInstruction(Res.first, Expr, RMWOp);
7464 X->getName() +
".atomic.cont");
7468 NewAtomicAddr->
setName(
X->getName() +
"x.new.val");
7471 PHI->addIncoming(OldVal, CurBB);
7477 X->getName() +
".atomic.fltCast");
7480 X->getName() +
".atomic.ptrCast");
7491 Result->setVolatile(VolatileX);
7497 Res.first = OldExprVal;
7517 bool UpdateExpr,
bool IsPostfixUpdate,
bool IsXBinopExpr) {
7522 Type *XTy =
X.Var->getType();
7524 "OMP Atomic expects a pointer to target memory");
7525 Type *XElemTy =
X.ElemTy;
7528 "OMP atomic capture expected a scalar type");
7530 "OpenMP atomic does not support LT or GT operations");
7536 std::pair<Value *, Value *> Result =
7537 emitAtomicUpdate(AllocaIP,
X.Var,
X.ElemTy, Expr, AO, AtomicOp, UpdateOp,
7538 X.IsVolatile, IsXBinopExpr);
7540 Value *CapturedVal = (IsPostfixUpdate ? Result.first : Result.second);
7543 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Capture);
7555 IsPostfixUpdate, IsFailOnly, Failure);
7567 assert(
X.Var->getType()->isPointerTy() &&
7568 "OMP atomic expects a pointer to target memory");
7571 assert(V.Var->getType()->isPointerTy() &&
"v.var must be of pointer type");
7572 assert(V.ElemTy ==
X.ElemTy &&
"x and v must be of same type");
7577 if (
Op == OMPAtomicCompareOp::EQ) {
7596 "OldValue and V must be of same type");
7597 if (IsPostfixUpdate) {
7615 CurBBTI,
X.Var->getName() +
".atomic.exit");
7635 Value *CapturedValue =
7643 assert(R.Var->getType()->isPointerTy() &&
7644 "r.var must be of pointer type");
7645 assert(R.ElemTy->isIntegerTy() &&
"r must be of integral type");
7648 Value *ResultCast = R.IsSigned
7654 assert((
Op == OMPAtomicCompareOp::MAX ||
Op == OMPAtomicCompareOp::MIN) &&
7655 "Op should be either max or min at this point");
7656 assert(!IsFailOnly &&
"IsFailOnly is only valid when the comparison is ==");
7696 Value *CapturedValue =
nullptr;
7697 if (IsPostfixUpdate) {
7698 CapturedValue = OldValue;
7730 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Compare);
7777 bool SubClausesPresent =
7778 (NumTeamsLower || NumTeamsUpper || ThreadLimit || IfExpr);
7781 assert((NumTeamsLower ==
nullptr || NumTeamsUpper !=
nullptr) &&
7782 "if lowerbound is non-null, then upperbound must also be non-null "
7783 "for bounds on num_teams");
7785 if (NumTeamsUpper ==
nullptr)
7788 if (NumTeamsLower ==
nullptr)
7789 NumTeamsLower = NumTeamsUpper;
7793 "argument to if clause must be an integer value");
7798 ConstantInt::get(IfExpr->
getType(), 0));
7807 if (ThreadLimit ==
nullptr)
7813 {Ident, ThreadNum, NumTeamsLower, NumTeamsUpper, ThreadLimit});
7818 BodyGenCB(AllocaIP, CodeGenIP);
7826 std::stack<Instruction *> ToBeDeleted;
7829 Builder, OuterAllocaIP, ToBeDeleted, AllocaIP,
"gid",
true));
7831 Builder, OuterAllocaIP, ToBeDeleted, AllocaIP,
"tid",
true));
7833 auto HostPostOutlineCB = [
this, Ident,
7834 ToBeDeleted](
Function &OutlinedFn)
mutable {
7839 "there must be a single user for the outlined function");
7841 ToBeDeleted.push(StaleCI);
7844 "Outlined function must have two or three arguments only");
7846 bool HasShared = OutlinedFn.
arg_size() == 3;
7854 assert(StaleCI &&
"Error while outlining - no CallInst user found for the "
7855 "outlined function.");
7862 omp::RuntimeFunction::OMPRTL___kmpc_fork_teams),
7865 while (!ToBeDeleted.empty()) {
7883 std::string VarName) {
7892 return MapNamesArrayGlobal;
7897void OpenMPIRBuilder::initializeTypes(
Module &M) {
7900#define OMP_TYPE(VarName, InitValue) VarName = InitValue;
7901#define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize) \
7902 VarName##Ty = ArrayType::get(ElemTy, ArraySize); \
7903 VarName##PtrTy = PointerType::getUnqual(VarName##Ty);
7904#define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...) \
7905 VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg); \
7906 VarName##Ptr = PointerType::getUnqual(VarName);
7907#define OMP_STRUCT_TYPE(VarName, StructName, Packed, ...) \
7908 T = StructType::getTypeByName(Ctx, StructName); \
7910 T = StructType::create(Ctx, {__VA_ARGS__}, StructName, Packed); \
7912 VarName##Ptr = PointerType::getUnqual(T);
7913#include "llvm/Frontend/OpenMP/OMPKinds.def"
7924 while (!Worklist.
empty()) {
7928 if (BlockSet.
insert(SuccBB).second)
7940 "omp_offloading_entries");
7964 Fn->
addFnAttr(
"uniform-work-group-size",
"true");
7983 auto &&GetMDInt = [
this](
unsigned V) {
7991 auto &&TargetRegionMetadataEmitter =
7992 [&
C, MD, &OrderedEntries, &GetMDInt, &GetMDString](
8007 GetMDInt(E.getKind()), GetMDInt(EntryInfo.DeviceID),
8008 GetMDInt(EntryInfo.FileID), GetMDString(EntryInfo.ParentName),
8009 GetMDInt(EntryInfo.Line), GetMDInt(EntryInfo.Count),
8010 GetMDInt(E.getOrder())};
8013 OrderedEntries[E.getOrder()] = std::make_pair(&E, EntryInfo);
8022 auto &&DeviceGlobalVarMetadataEmitter =
8023 [&
C, &OrderedEntries, &GetMDInt, &GetMDString, MD](
8033 Metadata *Ops[] = {GetMDInt(E.getKind()), GetMDString(MangledName),
8034 GetMDInt(E.getFlags()), GetMDInt(E.getOrder())};
8038 OrderedEntries[E.getOrder()] = std::make_pair(&E, varInfo);
8045 DeviceGlobalVarMetadataEmitter);
8047 for (
const auto &E : OrderedEntries) {
8048 assert(E.first &&
"All ordered entries must exist!");
8049 if (
const auto *CE =
8050 dyn_cast<OffloadEntriesInfoManager::OffloadEntryInfoTargetRegion>(
8052 if (!CE->getID() || !CE->getAddress()) {
8064 }
else if (
const auto *CE =
dyn_cast<
8075 if (!CE->getAddress()) {
8080 if (CE->getVarSize() == 0)
8086 "Declaret target link address is set.");
8089 if (!CE->getAddress()) {
8101 if (
auto *
GV = dyn_cast<GlobalValue>(CE->getAddress()))
8102 if ((
GV->hasLocalLinkage() ||
GV->hasHiddenVisibility()) &&
8110 Flags, CE->getLinkage(), CE->getVarName());
8113 Flags, CE->getLinkage());
8134 unsigned FileID,
unsigned Line,
unsigned Count) {
8137 <<
llvm::format(
"_%x_", FileID) << ParentName <<
"_l" << Line;
8144 unsigned NewCount = getTargetRegionEntryInfoCount(EntryInfo);
8147 EntryInfo.
Line, NewCount);
8154 auto FileIDInfo = CallBack();
8157 "getTargetEntryUniqueInfo, error message: " +
8163 std::get<1>(FileIDInfo));
8169 static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>
>(
8171 !(Remain & 1); Remain = Remain >> 1)
8189 if (
static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>
>(
8191 static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>
>(
8198 Flags &= ~omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF;
8199 Flags |= MemberOfFlag;
8205 bool IsDeclaration,
bool IsExternallyVisible,
8207 std::vector<GlobalVariable *> &GeneratedRefs,
bool OpenMPSIMD,
8208 std::vector<Triple> TargetTriple,
Type *LlvmPtrTy,
8209 std::function<
Constant *()> GlobalInitializer,
8225 if (!IsExternallyVisible)
8227 OS <<
"_decl_tgt_ref_ptr";
8236 auto *
GV = cast<GlobalVariable>(
Ptr);
8240 if (GlobalInitializer)
8241 GV->setInitializer(GlobalInitializer());
8247 CaptureClause, DeviceClause, IsDeclaration, IsExternallyVisible,
8248 EntryInfo, MangledName, GeneratedRefs, OpenMPSIMD, TargetTriple,
8249 GlobalInitializer, VariableLinkage, LlvmPtrTy, cast<Constant>(
Ptr));
8252 return cast<Constant>(
Ptr);
8261 bool IsDeclaration,
bool IsExternallyVisible,
8263 std::vector<GlobalVariable *> &GeneratedRefs,
bool OpenMPSIMD,
8264 std::vector<Triple> TargetTriple,
8265 std::function<
Constant *()> GlobalInitializer,
8282 VarName = MangledName;
8290 Linkage = (VariableLinkage) ? VariableLinkage() : LlvmVal->
getLinkage();
8306 auto *GvAddrRef = cast<GlobalVariable>(AddrRef);
8307 GvAddrRef->setConstant(
true);
8309 GvAddrRef->setInitializer(
Addr);
8310 GeneratedRefs.push_back(GvAddrRef);
8320 VarName = (
Addr) ?
Addr->getName() :
"";
8324 CaptureClause, DeviceClause, IsDeclaration, IsExternallyVisible,
8325 EntryInfo, MangledName, GeneratedRefs, OpenMPSIMD, TargetTriple,
8326 LlvmPtrTy, GlobalInitializer, VariableLinkage);
8327 VarName = (
Addr) ?
Addr->getName() :
"";
8348 auto &&GetMDInt = [MN](
unsigned Idx) {
8349 auto *V = cast<ConstantAsMetadata>(MN->getOperand(
Idx));
8350 return cast<ConstantInt>(V->getValue())->getZExtValue();
8353 auto &&GetMDString = [MN](
unsigned Idx) {
8354 auto *V = cast<MDString>(MN->getOperand(
Idx));
8355 return V->getString();
8358 switch (GetMDInt(0)) {
8386 if (HostFilePath.
empty())
8390 if (std::error_code Err = Buf.getError()) {
8392 "OpenMPIRBuilder: " +
8400 if (std::error_code Err =
M.getError()) {
8402 (
"error parsing host file inside of OpenMPIRBuilder: " + Err.message())
8414 return OffloadEntriesTargetRegion.empty() &&
8415 OffloadEntriesDeviceGlobalVar.empty();
8418unsigned OffloadEntriesInfoManager::getTargetRegionEntryInfoCount(
8420 auto It = OffloadEntriesTargetRegionCount.find(
8421 getTargetRegionEntryCountKey(EntryInfo));
8422 if (It == OffloadEntriesTargetRegionCount.end())
8427void OffloadEntriesInfoManager::incrementTargetRegionEntryInfoCount(
8429 OffloadEntriesTargetRegionCount[getTargetRegionEntryCountKey(EntryInfo)] =
8430 EntryInfo.
Count + 1;
8436 OffloadEntriesTargetRegion[EntryInfo] =
8438 OMPTargetRegionEntryTargetRegion);
8439 ++OffloadingEntriesNum;
8445 assert(EntryInfo.
Count == 0 &&
"expected default EntryInfo");
8448 EntryInfo.
Count = getTargetRegionEntryInfoCount(EntryInfo);
8452 if (OMPBuilder->Config.isTargetDevice()) {
8454 if (!hasTargetRegionEntryInfo(EntryInfo)) {
8457 auto &Entry = OffloadEntriesTargetRegion[EntryInfo];
8458 Entry.setAddress(
Addr);
8460 Entry.setFlags(
Flags);
8463 hasTargetRegionEntryInfo(EntryInfo,
true))
8465 assert(!hasTargetRegionEntryInfo(EntryInfo) &&
8466 "Target region entry already registered!");
8468 OffloadEntriesTargetRegion[EntryInfo] = Entry;
8469 ++OffloadingEntriesNum;
8471 incrementTargetRegionEntryInfoCount(EntryInfo);
8478 EntryInfo.
Count = getTargetRegionEntryInfoCount(EntryInfo);
8480 auto It = OffloadEntriesTargetRegion.find(EntryInfo);
8481 if (It == OffloadEntriesTargetRegion.end()) {
8485 if (!IgnoreAddressId && (It->second.getAddress() || It->second.getID()))
8493 for (
const auto &It : OffloadEntriesTargetRegion) {
8494 Action(It.first, It.second);
8500 OffloadEntriesDeviceGlobalVar.try_emplace(
Name, Order,
Flags);
8501 ++OffloadingEntriesNum;
8507 if (OMPBuilder->Config.isTargetDevice()) {
8509 if (!hasDeviceGlobalVarEntryInfo(VarName))
8511 auto &Entry = OffloadEntriesDeviceGlobalVar[VarName];
8512 if (Entry.getAddress() && hasDeviceGlobalVarEntryInfo(VarName)) {
8513 if (Entry.getVarSize() == 0) {
8514 Entry.setVarSize(VarSize);
8515 Entry.setLinkage(Linkage);
8519 Entry.setVarSize(VarSize);
8520 Entry.setLinkage(Linkage);
8521 Entry.setAddress(
Addr);
8523 if (hasDeviceGlobalVarEntryInfo(VarName)) {
8524 auto &Entry = OffloadEntriesDeviceGlobalVar[VarName];
8525 assert(Entry.isValid() && Entry.getFlags() ==
Flags &&
8526 "Entry not initialized!");
8527 if (Entry.getVarSize() == 0) {
8528 Entry.setVarSize(VarSize);
8529 Entry.setLinkage(Linkage);
8534 OffloadEntriesDeviceGlobalVar.try_emplace(VarName, OffloadingEntriesNum,
8538 OffloadEntriesDeviceGlobalVar.try_emplace(
8539 VarName, OffloadingEntriesNum,
Addr, VarSize,
Flags, Linkage,
"");
8540 ++OffloadingEntriesNum;
8547 for (
const auto &E : OffloadEntriesDeviceGlobalVar)
8548 Action(E.getKey(), E.getValue());
8555void CanonicalLoopInfo::collectControlBlocks(
8562 BBs.
append({getPreheader(), Header,
Cond, Latch, Exit, getAfter()});
8574void CanonicalLoopInfo::setTripCount(
Value *TripCount) {
8578 assert(isa<CmpInst>(CmpI) &&
"First inst must compare IV with TripCount");
8586void CanonicalLoopInfo::mapIndVar(
8596 for (
Use &U : OldIV->
uses()) {
8597 auto *
User = dyn_cast<Instruction>(U.getUser());
8600 if (
User->getParent() == getCond())
8602 if (
User->getParent() == getLatch())
8608 Value *NewIV = Updater(OldIV);
8611 for (
Use *U : ReplacableUses)
8632 "Preheader must terminate with unconditional branch");
8634 "Preheader must jump to header");
8637 assert(isa<BranchInst>(Header->getTerminator()) &&
8638 "Header must terminate with unconditional branch");
8639 assert(Header->getSingleSuccessor() ==
Cond &&
8640 "Header must jump to exiting block");
8643 assert(
Cond->getSinglePredecessor() == Header &&
8644 "Exiting block only reachable from header");
8646 assert(isa<BranchInst>(
Cond->getTerminator()) &&
8647 "Exiting block must terminate with conditional branch");
8649 "Exiting block must have two successors");
8650 assert(cast<BranchInst>(
Cond->getTerminator())->getSuccessor(0) == Body &&
8651 "Exiting block's first successor jump to the body");
8652 assert(cast<BranchInst>(
Cond->getTerminator())->getSuccessor(1) == Exit &&
8653 "Exiting block's second successor must exit the loop");
8657 "Body only reachable from exiting block");
8662 "Latch must terminate with unconditional branch");
8670 assert(isa<BranchInst>(Exit->getTerminator()) &&
8671 "Exit block must terminate with unconditional branch");
8673 "Exit block must jump to after block");
8677 "After block only reachable from exit block");
8681 assert(IndVar &&
"Canonical induction variable not found?");
8683 "Induction variable must be an integer");
8685 "Induction variable must be a PHI in the loop header");
8686 assert(cast<PHINode>(IndVar)->getIncomingBlock(0) == Preheader);
8688 cast<ConstantInt>(cast<PHINode>(IndVar)->getIncomingValue(0))->
isZero());
8689 assert(cast<PHINode>(IndVar)->getIncomingBlock(1) == Latch);
8691 auto *NextIndVar = cast<PHINode>(IndVar)->getIncomingValue(1);
8693 assert(cast<BinaryOperator>(NextIndVar)->
getOpcode() == BinaryOperator::Add);
8694 assert(cast<BinaryOperator>(NextIndVar)->getOperand(0) == IndVar);
8695 assert(cast<ConstantInt>(cast<BinaryOperator>(NextIndVar)->getOperand(1))
8698 Value *TripCount = getTripCount();
8699 assert(TripCount &&
"Loop trip count not found?");
8701 "Trip count and induction variable must have the same type");
8703 auto *CmpI = cast<CmpInst>(&
Cond->front());
8705 "Exit condition must be a signed less-than comparison");
8707 "Exit condition must compare the induction variable");
8709 "Exit condition must compare with the trip count");
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static cl::opt< ITMode > IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT), cl::values(clEnumValN(DefaultIT, "arm-default-it", "Generate any type of IT block"), clEnumValN(RestrictedIT, "arm-restrict-it", "Disallow complex IT blocks")))
This file contains the simple types necessary to represent the attributes associated with functions a...
static const Function * getParent(const Value *V)
#define LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()
LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE() pulls the operator overloads used by LLVM_MARK_AS_BITMASK_EN...
#define LLVM_MARK_AS_BITMASK_ENUM(LargestValue)
LLVM_MARK_AS_BITMASK_ENUM lets you opt in an individual enum type so you can perform bitwise operatio...
BlockVerifier::State From
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
DenseMap< Block *, BlockRelaxAux > Blocks
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
Rewrite Partial Register Uses
This file provides various utilities for inspecting and working with the control flow graph in LLVM I...
iv Induction Variable Users
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
static cl::opt< unsigned > TileSize("fuse-matrix-tile-size", cl::init(4), cl::Hidden, cl::desc("Tile size for matrix instruction fusion using square-shaped tiles."))
uint64_t IntrinsicInst * II
#define OMP_KERNEL_ARG_VERSION
Provides definitions for Target specific Grid Values.
static OMPScheduleType getOpenMPBaseScheduleType(llvm::omp::ScheduleKind ClauseKind, bool HasChunks, bool HasSimdModifier)
Determine which scheduling algorithm to use, determined from schedule clause arguments.
static void emitTargetOutlinedFunction(OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, TargetRegionEntryInfo &EntryInfo, Function *&OutlinedFn, Constant *&OutlinedFnID, SmallVectorImpl< Value * > &Inputs, OpenMPIRBuilder::TargetBodyGenCallbackTy &CBFunc, OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy &ArgAccessorFuncCB)
static Function * createOutlinedFunction(OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, StringRef FuncName, SmallVectorImpl< Value * > &Inputs, OpenMPIRBuilder::TargetBodyGenCallbackTy &CBFunc, OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy &ArgAccessorFuncCB)
static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL)
Make Source branch to Target.
static void emitTargetCall(OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, OpenMPIRBuilder::InsertPointTy AllocaIP, Function *OutlinedFn, Constant *OutlinedFnID, int32_t NumTeams, int32_t NumThreads, SmallVectorImpl< Value * > &Args, OpenMPIRBuilder::GenMapInfoCallbackTy GenMapInfoCB)
static FunctionCallee getKmpcForDynamicFiniForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder)
Returns an LLVM function to call for finalizing the dynamic loop using depending on type.
static void updateNVPTXMetadata(Function &Kernel, StringRef Name, int32_t Value, bool Min)
static OMPScheduleType getOpenMPOrderingScheduleType(OMPScheduleType BaseScheduleType, bool HasOrderedClause)
Adds ordering modifier flags to schedule type.
static OMPScheduleType getOpenMPMonotonicityScheduleType(OMPScheduleType ScheduleType, bool HasSimdModifier, bool HasMonotonic, bool HasNonmonotonic, bool HasOrderedClause)
Adds monotonicity modifier flags to schedule type.
static void addSimdMetadata(BasicBlock *Block, MDNode *AccessGroup, LoopInfo &LI)
Attach llvm.access.group metadata to the memref instructions of Block.
static OMPScheduleType computeOpenMPScheduleType(ScheduleKind ClauseKind, bool HasChunks, bool HasSimdModifier, bool HasMonotonicModifier, bool HasNonmonotonicModifier, bool HasOrderedClause)
Determine the schedule type using schedule and ordering clause arguments.
static bool isValidWorkshareLoopScheduleType(OMPScheduleType SchedType)
static Function * getFreshReductionFunc(Module &M)
static void raiseUserConstantDataAllocasToEntryBlock(IRBuilderBase &Builder, Function *Function)
static MDNode * getNVPTXMDNode(Function &Kernel, StringRef Name)
static FunctionCallee getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder)
Returns an LLVM function to call for updating the next loop using OpenMP dynamic scheduling depending...
static bool isConflictIP(IRBuilder<>::InsertPoint IP1, IRBuilder<>::InsertPoint IP2)
Return whether IP1 and IP2 are ambiguous, i.e.
static void checkReductionInfos(ArrayRef< OpenMPIRBuilder::ReductionInfo > ReductionInfos, bool IsGPU)
static FunctionCallee getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder)
Returns an LLVM function to call for initializing loop bounds using OpenMP dynamic scheduling dependi...
static cl::opt< double > UnrollThresholdFactor("openmp-ir-builder-unroll-threshold-factor", cl::Hidden, cl::desc("Factor for the unroll threshold to account for code " "simplifications still taking place"), cl::init(1.5))
static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI)
Heuristically determine the best-performant unroll factor for CLI.
static void workshareLoopTargetCallback(OpenMPIRBuilder *OMPIRBuilder, CanonicalLoopInfo *CLI, Value *Ident, Function &OutlinedFn, Type *ParallelTaskPtr, const SmallVector< Instruction *, 4 > &ToBeDeleted, WorksharingLoopType LoopType)
static OpenMPIRBuilder::InsertPointTy getInsertPointAfterInstr(Instruction *I)
static void redirectAllPredecessorsTo(BasicBlock *OldTarget, BasicBlock *NewTarget, DebugLoc DL)
Redirect all edges that branch to OldTarget to NewTarget.
static std::unique_ptr< TargetMachine > createTargetMachine(Function *F, CodeGenOptLevel OptLevel)
Create the TargetMachine object to query the backend for optimization preferences.
static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder)
static void addBasicBlockMetadata(BasicBlock *BB, ArrayRef< Metadata * > Properties)
Attach metadata Properties to the basic block described by BB.
static cl::opt< bool > OptimisticAttributes("openmp-ir-builder-optimistic-attributes", cl::Hidden, cl::desc("Use optimistic attributes describing " "'as-if' properties of runtime calls."), cl::init(false))
static FunctionCallee getKmpcForStaticLoopForType(Type *Ty, OpenMPIRBuilder *OMPBuilder, WorksharingLoopType LoopType)
static void createTargetLoopWorkshareCall(OpenMPIRBuilder *OMPBuilder, WorksharingLoopType LoopType, BasicBlock *InsertBlock, Value *Ident, Value *LoopBodyArg, Type *ParallelTaskPtr, Value *TripCount, Function &LoopBodyFn)
static const omp::GV & getGridValue(const Triple &T, Function *Kernel)
Value * createFakeIntVal(IRBuilder<> &Builder, OpenMPIRBuilder::InsertPointTy OuterAllocaIP, std::stack< Instruction * > &ToBeDeleted, OpenMPIRBuilder::InsertPointTy InnerAllocaIP, const Twine &Name="", bool AsPtr=true)
static void addLoopMetadata(CanonicalLoopInfo *Loop, ArrayRef< Metadata * > Properties)
Attach loop metadata Properties to the loop described by Loop.
static void removeUnusedBlocksFromParent(ArrayRef< BasicBlock * > BBs)
Determine which blocks in BBs are reachable from outside and remove the ones that are not reachable f...
static void targetParallelCallback(OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn, Function *OuterFn, BasicBlock *OuterAllocaBB, Value *Ident, Value *IfCondition, Value *NumThreads, Instruction *PrivTID, AllocaInst *PrivTIDAddr, Value *ThreadID, const SmallVector< Instruction *, 4 > &ToBeDeleted)
static void hostParallelCallback(OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn, Function *OuterFn, Value *Ident, Value *IfCondition, Instruction *PrivTID, AllocaInst *PrivTIDAddr, const SmallVector< Instruction *, 4 > &ToBeDeleted)
FunctionAnalysisManager FAM
const char LLVMTargetMachineRef TM
This file defines the Pass Instrumentation classes that provide instrumentation points into the pass ...
This header defines various interfaces for pass management in LLVM.
const SmallVectorImpl< MachineOperand > & Cond
static bool isValid(const char C)
Returns true if C is a valid mangled character: <0-9a-zA-Z_>.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the SmallSet class.
static std::optional< unsigned > getOpcode(ArrayRef< VPValue * > Values)
Returns the opcode of Values or ~0 if they do not all agree.
static cl::opt< unsigned > MaxThreads("xcore-max-threads", cl::Optional, cl::desc("Maximum number of threads (for emulation thread-local storage)"), cl::Hidden, cl::value_desc("number"), cl::init(8))
static const uint32_t IV[8]
Class for arbitrary precision integers.
This class represents a conversion between pointers from one address space to another.
an instruction to allocate memory on the stack
PointerType * getType() const
Overload to return most specific pointer type.
Type * getAllocatedType() const
Return the type that is being allocated by the instruction.
unsigned getAddressSpace() const
Return the address space for the allocation.
std::optional< TypeSize > getAllocationSize(const DataLayout &DL) const
Get allocation size in bytes.
void setAlignment(Align Align)
const Value * getArraySize() const
Get the number of elements allocated.
A container for analyses that lazily runs them and caches their results.
bool registerPass(PassBuilderT &&PassBuilder)
Register an analysis pass with the manager.
This class represents an incoming formal argument to a Function.
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.
Class to represent array types.
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
A function analysis which provides an AssumptionCache.
AssumptionCache run(Function &F, FunctionAnalysisManager &)
A cache of @llvm.assume calls within a function.
An instruction that atomically checks whether a specified value is in a memory location,...
static AtomicOrdering getStrongestFailureOrdering(AtomicOrdering SuccessOrdering)
Returns the strongest permitted ordering on failure, given the desired ordering on success.
an instruction that atomically reads a memory location, combines it with another value,...
BinOp
This enumeration lists the possible modifications atomicrmw can make.
@ Min
*p = old <signed v ? old : v
@ UIncWrap
Increment one up to a maximum value.
@ Max
*p = old >signed v ? old : v
@ UMin
*p = old <unsigned v ? old : v
@ FMin
*p = minnum(old, v) minnum matches the behavior of llvm.minnum.
@ UMax
*p = old >unsigned v ? old : v
@ FMax
*p = maxnum(old, v) maxnum matches the behavior of llvm.maxnum.
@ UDecWrap
Decrement one until a minimum value or zero.
AttrBuilder & addAttribute(Attribute::AttrKind Val)
Add an attribute to the builder.
AttrBuilder & removeAttribute(Attribute::AttrKind Val)
Remove an attribute from the builder.
AttributeSet getFnAttrs() const
The function attributes are returned.
AttributeList addFnAttributes(LLVMContext &C, const AttrBuilder &B) const
Add function attribute to the list.
AttributeSet addAttributes(LLVMContext &C, AttributeSet AS) const
Add attributes to the attribute set.
AttributeSet addAttribute(LLVMContext &C, Attribute::AttrKind Kind) const
Add an argument attribute.
static Attribute get(LLVMContext &Context, AttrKind Kind, uint64_t Val=0)
Return a uniquified Attribute object.
StringRef getValueAsString() const
Return the attribute's value as a string.
LLVM Basic Block Representation.
void replaceSuccessorsPhiUsesWith(BasicBlock *Old, BasicBlock *New)
Update all phi nodes in this basic block's successors to refer to basic block New instead of basic bl...
iterator begin()
Instruction iterator methods.
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
reverse_iterator rbegin()
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
const Instruction & front() const
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
BasicBlock * splitBasicBlock(iterator I, const Twine &BBName="", bool Before=false)
Split the basic block into two basic blocks at the specified instruction.
const BasicBlock * getUniqueSuccessor() const
Return the successor of this block if it has a unique successor.
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
InstListType::reverse_iterator reverse_iterator
const BasicBlock * getUniquePredecessor() const
Return the predecessor of this block if it has a unique predecessor block.
const BasicBlock * getSingleSuccessor() const
Return the successor of this block if it has a single successor.
const Function * getParent() const
Return the enclosing method, or null if none.
SymbolTableList< BasicBlock >::iterator eraseFromParent()
Unlink 'this' from the containing function and delete it.
const Instruction * getFirstNonPHIOrDbg(bool SkipPseudoOp=true) const
Returns a pointer to the first instruction in this block that is not a PHINode or a debug intrinsic,...
InstListType::iterator iterator
Instruction iterators...
LLVMContext & getContext() const
Get the context in which this basic block lives.
void moveBefore(BasicBlock *MovePos)
Unlink this basic block from its current function and insert it into the function that MovePos lives ...
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
void splice(BasicBlock::iterator ToIt, BasicBlock *FromBB)
Transfer all instructions from FromBB to this basic block at ToIt.
const Instruction & back() const
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr if the function does no...
void removePredecessor(BasicBlock *Pred, bool KeepOneInputPHIs=false)
Update PHI nodes in this BasicBlock before removal of predecessor Pred.
Conditional or Unconditional Branch instruction.
static BranchInst * Create(BasicBlock *IfTrue, InsertPosition InsertBefore=nullptr)
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation or the function signa...
User::op_iterator arg_begin()
Return the iterator pointing to the beginning of the argument list.
Value * getArgOperand(unsigned i) const
User::op_iterator arg_end()
Return the iterator pointing to the end of the argument list.
unsigned arg_size() const
This class represents a function call, abstracting a target machine's calling convention.
Class to represented the control flow structure of an OpenMP canonical loop.
Value * getTripCount() const
Returns the llvm::Value containing the number of loop iterations.
BasicBlock * getHeader() const
The header is the entry for each iteration.
void assertOK() const
Consistency self-check.
Type * getIndVarType() const
Return the type of the induction variable (and the trip count).
BasicBlock * getBody() const
The body block is the single entry for a loop iteration and not controlled by CanonicalLoopInfo.
bool isValid() const
Returns whether this object currently represents the IR of a loop.
OpenMPIRBuilder::InsertPointTy getAfterIP() const
Return the insertion point for user code after the loop.
OpenMPIRBuilder::InsertPointTy getBodyIP() const
Return the insertion point for user code in the body.
BasicBlock * getAfter() const
The after block is intended for clean-up code such as lifetime end markers.
Function * getFunction() const
void invalidate()
Invalidate this loop.
BasicBlock * getLatch() const
Reaching the latch indicates the end of the loop body code.
OpenMPIRBuilder::InsertPointTy getPreheaderIP() const
Return the insertion point for user code before the loop.
BasicBlock * getCond() const
The condition block computes whether there is another loop iteration.
BasicBlock * getExit() const
Reaching the exit indicates no more iterations are being executed.
BasicBlock * getPreheader() const
The preheader ensures that there is only a single edge entering the loop.
Instruction * getIndVar() const
Returns the instruction representing the current logical induction variable.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
@ ICMP_SLT
signed less than
@ ICMP_SLE
signed less or equal
@ FCMP_OLT
0 1 0 0 True if ordered and less than
@ FCMP_OGT
0 0 1 0 True if ordered and greater than
@ ICMP_UGT
unsigned greater than
@ ICMP_SGT
signed greater than
@ ICMP_ULT
unsigned less than
@ ICMP_ULE
unsigned less or equal
static Constant * get(ArrayType *T, ArrayRef< Constant * > V)
static Constant * getString(LLVMContext &Context, StringRef Initializer, bool AddNull=true)
This method constructs a CDS and initializes it with a text string.
static Constant * get(LLVMContext &Context, ArrayRef< ElementTy > Elts)
get() constructor - Return a constant with array type with an element count and element type matching...
static Constant * getPointerCast(Constant *C, Type *Ty)
Create a BitCast, AddrSpaceCast, or a PtrToInt cast constant expression.
static Constant * getPointerBitCastOrAddrSpaceCast(Constant *C, Type *Ty)
Create a BitCast or AddrSpaceCast for a pointer type depending on the address space.
static Constant * getAddrSpaceCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
This is the shared class of boolean and integer constants.
static ConstantInt * getTrue(LLVMContext &Context)
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
static ConstantInt * getFalse(LLVMContext &Context)
static ConstantPointerNull * get(PointerType *T)
Static factory methods - Return objects of the specified value.
static Constant * get(StructType *T, ArrayRef< Constant * > V)
This is an important base class in LLVM.
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
This class represents an Operation in the Expression.
uint64_t getNumOperands() const
A parsed version of the target data layout string in and methods for querying it.
unsigned getDefaultGlobalsAddressSpace() const
Align getABIIntegerTypeAlignment(unsigned BitWidth) const
Returns the minimum ABI-required alignment for an integer type of the specified bitwidth.
unsigned getAllocaAddrSpace() const
TypeSize getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
unsigned getPointerSize(unsigned AS=0) const
Layout pointer size in bytes, rounded up to a whole number of bytes.
unsigned getIndexSizeInBits(unsigned AS) const
Size in bits of index used for address calculation in getelementptr.
TypeSize getTypeSizeInBits(Type *Ty) const
Size examples:
TypeSize getTypeStoreSize(Type *Ty) const
Returns the maximum number of bytes that may be overwritten by storing the specified type.
Align getPrefTypeAlign(Type *Ty) const
Returns the preferred stack/global alignment for the specified type.
Analysis pass which computes a DominatorTree.
DominatorTree run(Function &F, FunctionAnalysisManager &)
Run the analysis pass over a function and produce a dominator tree.
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Lightweight error class with error context and mandatory checking.
A handy container for a FunctionType+Callee-pointer pair, which can be passed around as a single enti...
Class to represent function types.
static FunctionType * get(Type *Result, ArrayRef< Type * > Params, bool isVarArg)
This static method is the primary way of constructing a FunctionType.
void addFnAttr(Attribute::AttrKind Kind)
Add function attributes to this function.
static Function * Create(FunctionType *Ty, LinkageTypes Linkage, unsigned AddrSpace, const Twine &N="", Module *M=nullptr)
const BasicBlock & getEntryBlock() const
FunctionType * getFunctionType() const
Returns the FunctionType for me.
void removeFromParent()
removeFromParent - This method unlinks 'this' from the containing module, but does not delete it.
const DataLayout & getDataLayout() const
Get the data layout of the module this function belongs to.
Attribute getFnAttribute(Attribute::AttrKind Kind) const
Return the attribute for the given attribute kind.
uint64_t getFnAttributeAsParsedInteger(StringRef Kind, uint64_t Default=0) const
For a string attribute Kind, parse attribute as an integer.
AttributeList getAttributes() const
Return the attribute list for this Function.
const Function & getFunction() const
void setAttributes(AttributeList Attrs)
Set the attribute list for this Function.
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function.
void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind)
adds the attribute to the list of attributes for the given arg.
Function::iterator insert(Function::iterator Position, BasicBlock *BB)
Insert BB in the basic block list at Position.
Type * getReturnType() const
Returns the type of the ret val.
void setCallingConv(CallingConv::ID CC)
Argument * getArg(unsigned i) const
bool hasMetadata() const
Return true if this value has any metadata attached to it.
void addMetadata(unsigned KindID, MDNode &MD)
Add a metadata attachment.
LinkageTypes getLinkage() const
void setLinkage(LinkageTypes LT)
Module * getParent()
Get the module that this global value is contained inside of...
void setDSOLocal(bool Local)
PointerType * getType() const
Global values are always pointers.
@ HiddenVisibility
The GV is hidden.
@ ProtectedVisibility
The GV is protected.
void setVisibility(VisibilityTypes V)
LinkageTypes
An enumeration for the kinds of linkage for global values.
@ PrivateLinkage
Like Internal, but omit from symbol table.
@ CommonLinkage
Tentative definitions.
@ InternalLinkage
Rename collisions when linking (static functions).
@ WeakODRLinkage
Same, but only replaced by something equivalent.
@ ExternalLinkage
Externally visible function.
@ WeakAnyLinkage
Keep one copy of named function when linking (weak)
@ AppendingLinkage
Special purpose, only applies to global arrays.
@ LinkOnceODRLinkage
Same, but only replaced by something equivalent.
Type * getValueType() const
InsertPoint - A saved insertion point.
BasicBlock * getBlock() const
bool isSet() const
Returns true if this insert point is set.
BasicBlock::iterator getPoint() const
Common base class shared among various IRBuilders.
Value * CreatePtrDiff(Type *ElemTy, Value *LHS, Value *RHS, const Twine &Name="")
Return the i64 difference between two pointer values, dividing out the size of the pointed-to objects...
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
AtomicCmpXchgInst * CreateAtomicCmpXchg(Value *Ptr, Value *Cmp, Value *New, MaybeAlign Align, AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, SyncScope::ID SSID=SyncScope::System)
AllocaInst * CreateAlloca(Type *Ty, unsigned AddrSpace, Value *ArraySize=nullptr, const Twine &Name="")
Value * CreateInsertValue(Value *Agg, Value *Val, ArrayRef< unsigned > Idxs, const Twine &Name="")
IntegerType * getIntNTy(unsigned N)
Fetch the type representing an N-bit integer.
Value * CreateICmpSGT(Value *LHS, Value *RHS, const Twine &Name="")
LoadInst * CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, const char *Name)
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
UnreachableInst * CreateUnreachable()
Value * CreatePointerCast(Value *V, Type *DestTy, const Twine &Name="")
Value * CreateExtractValue(Value *Agg, ArrayRef< unsigned > Idxs, const Twine &Name="")
CallInst * CreateAlignmentAssumption(const DataLayout &DL, Value *PtrValue, unsigned Alignment, Value *OffsetValue=nullptr)
Create an assume intrinsic call that represents an alignment assumption on the provided pointer.
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
BasicBlock::iterator GetInsertPoint() const
Value * CreateStructGEP(Type *Ty, Value *Ptr, unsigned Idx, const Twine &Name="")
IntegerType * getIndexTy(const DataLayout &DL, unsigned AddrSpace)
Fetch the type of an integer that should be used to index GEP operations within AddressSpace.
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Value * CreateIntToPtr(Value *V, Type *DestTy, const Twine &Name="")
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Value * CreateNSWAdd(Value *LHS, Value *RHS, const Twine &Name="")
BasicBlock * GetInsertBlock() const
void SetCurrentDebugLocation(DebugLoc L)
Set location information used by debugging information.
IntegerType * getInt64Ty()
Fetch the type representing a 64-bit integer.
Value * CreateInBoundsGEP(Type *Ty, Value *Ptr, ArrayRef< Value * > IdxList, const Twine &Name="")
Value * CreatePointerBitCastOrAddrSpaceCast(Value *V, Type *DestTy, const Twine &Name="")
Value * CreateUDiv(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Value * CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name="")
IntegerType * getInt16Ty()
Fetch the type representing a 16-bit integer.
Value * CreateGEP(Type *Ty, Value *Ptr, ArrayRef< Value * > IdxList, const Twine &Name="", GEPNoWrapFlags NW=GEPNoWrapFlags::none())
ConstantInt * getInt64(uint64_t C)
Get a constant 64-bit value.
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNSW=false)
InsertPoint saveIP() const
Returns the current insert point.
Constant * CreateGlobalStringPtr(StringRef Str, const Twine &Name="", unsigned AddressSpace=0, Module *M=nullptr, bool AddNull=true)
Same as CreateGlobalString, but return a pointer with "i8*" type instead of a pointer to array of i8.
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Value * CreateCmp(CmpInst::Predicate Pred, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
PHINode * CreatePHI(Type *Ty, unsigned NumReservedValues, const Twine &Name="")
SwitchInst * CreateSwitch(Value *V, BasicBlock *Dest, unsigned NumCases=10, MDNode *BranchWeights=nullptr, MDNode *Unpredictable=nullptr)
Create a switch instruction with the specified value, default dest, and with a hint for the number of...
Value * CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name="")
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
DebugLoc getCurrentDebugLocation() const
Get location information used by debugging information.
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
ConstantInt * getIntN(unsigned N, uint64_t C)
Get a constant N-bit value, zero extended or truncated from a 64-bit value.
BranchInst * CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, MDNode *BranchWeights=nullptr, MDNode *Unpredictable=nullptr)
Create a conditional 'br Cond, TrueDest, FalseDest' instruction.
LoadInst * CreateLoad(Type *Ty, Value *Ptr, const char *Name)
Provided to resolve 'CreateLoad(Ty, Ptr, "...")' correctly, instead of converting the string to 'bool...
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
LLVMContext & getContext() const
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
ReturnInst * CreateRetVoid()
Create a 'ret void' instruction.
Value * CreateConstInBoundsGEP2_32(Type *Ty, Value *Ptr, unsigned Idx0, unsigned Idx1, const Twine &Name="")
Value * CreateConstInBoundsGEP2_64(Type *Ty, Value *Ptr, uint64_t Idx0, uint64_t Idx1, const Twine &Name="")
StoreInst * CreateStore(Value *Val, Value *Ptr, bool isVolatile=false)
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Value * CreatePtrToInt(Value *V, Type *DestTy, const Twine &Name="")
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
AtomicRMWInst * CreateAtomicRMW(AtomicRMWInst::BinOp Op, Value *Ptr, Value *Val, MaybeAlign Align, AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="", bool IsNUW=false, bool IsNSW=false)
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
PointerType * getPtrTy(unsigned AddrSpace=0)
Fetch the type representing a pointer.
BranchInst * CreateBr(BasicBlock *Dest)
Create an unconditional 'br label X' instruction.
void ClearInsertionPoint()
Clear the insertion point: created instructions will not be inserted into a block.
Value * CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name="")
ConstantInt * getInt16(uint16_t C)
Get a constant 16-bit value.
Value * CreateICmpUGE(Value *LHS, Value *RHS, const Twine &Name="")
Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")
void restoreIP(InsertPoint IP)
Sets the current insert point to a previously-saved location.
Value * CreateIsNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg == 0.
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block.
Type * getVoidTy()
Fetch the type representing void.
StoreInst * CreateAlignedStore(Value *Val, Value *Ptr, MaybeAlign Align, bool isVolatile=false)
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
IntegerType * getInt8Ty()
Fetch the type representing an 8-bit integer.
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
CallInst * CreateMemCpy(Value *Dst, MaybeAlign DstAlign, Value *Src, MaybeAlign SrcAlign, uint64_t Size, bool isVolatile=false, MDNode *TBAATag=nullptr, MDNode *TBAAStructTag=nullptr, MDNode *ScopeTag=nullptr, MDNode *NoAliasTag=nullptr)
Create and insert a memcpy between the specified pointers.
Value * CreateSExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a SExt or Trunc from the integer value V to DestTy.
Value * CreateAddrSpaceCast(Value *V, Type *DestTy, const Twine &Name="")
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
void removeFromParent()
This method unlinks 'this' from the containing basic block, but does not delete it.
void moveBeforePreserving(Instruction *MovePos)
Perform a moveBefore operation, while signalling that the caller intends to preserve the original ord...
unsigned getNumSuccessors() const LLVM_READONLY
Return the number of successors that this instruction has.
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
InstListType::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
static bool classof(const Value *V)
Methods for support type inquiry through isa, cast, and dyn_cast:
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
BasicBlock * getSuccessor(unsigned Idx) const LLVM_READONLY
Return the specified successor. This instruction must be a terminator.
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
void insertAfter(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately after the specified instruction.
Class to represent integer types.
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
This is an important class for using LLVM in a threaded context.
An instruction for reading from memory.
void setAtomic(AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Sets the ordering constraint and the synchronization scope ID of this load instruction.
Analysis pass that exposes the LoopInfo for a function.
LoopInfo run(Function &F, FunctionAnalysisManager &AM)
LoopT * getLoopFor(const BlockT *BB) const
Return the inner most loop that BB lives in.
This class represents a loop nest and can be used to query its properties.
Represents a single loop in the control flow graph.
MDNode * createCallbackEncoding(unsigned CalleeArgNo, ArrayRef< int > Arguments, bool VarArgsArePassed)
Return metadata describing a callback (see llvm::AbstractCallSite).
void replaceOperandWith(unsigned I, Metadata *New)
Replace a specific operand.
static MDTuple * getDistinct(LLVMContext &Context, ArrayRef< Metadata * > MDs)
const MDOperand & getOperand(unsigned I) const
ArrayRef< MDOperand > operands() const
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata * > MDs)
static MDString * get(LLVMContext &Context, StringRef Str)
This class implements a map that also provides access to all stored values in a deterministic order.
static ErrorOr< std::unique_ptr< MemoryBuffer > > getFile(const Twine &Filename, bool IsText=false, bool RequiresNullTerminator=true, bool IsVolatile=false, std::optional< Align > Alignment=std::nullopt)
Open the specified file as a MemoryBuffer, returning a new MemoryBuffer if successful,...
A Module instance is used to store all the information related to an LLVM module.
LLVMContext & getContext() const
Get the global data context.
NamedMDNode * getNamedMetadata(const Twine &Name) const
Return the first NamedMDNode in the module with the specified name.
StringRef getName() const
Get a short "name" for the module.
const std::string & getTargetTriple() const
Get the target triple which is a string describing the target host.
iterator_range< global_iterator > globals()
const FunctionListType & getFunctionList() const
Get the Module's list of functions (constant).
GlobalVariable * getGlobalVariable(StringRef Name) const
Look up the specified global variable in the module symbol table.
GlobalValue * getNamedValue(StringRef Name) const
Return the global value in the module with the specified name, of arbitrary type.
NamedMDNode * getOrInsertNamedMetadata(StringRef Name)
Return the named MDNode in the module with the specified name.
const GlobalVariable * getNamedGlobal(StringRef Name) const
Return the global variable in the module with the specified name, of arbitrary type.
const DataLayout & getDataLayout() const
Get the data layout for the module's target platform.
iterator_range< op_iterator > operands()
void addOperand(MDNode *M)
Device global variable entries info.
Target region entries info.
Base class of the entries info.
@ OffloadingEntryInfoTargetRegion
Entry is a target region.
@ OffloadingEntryInfoDeviceGlobalVar
Entry is a declare target variable.
OMPTargetDeviceClauseKind
Kind of device clause for declare target variables and functions NOTE: Currently not used as a part o...
@ OMPTargetDeviceClauseAny
The target is marked for all devices.
void registerDeviceGlobalVarEntryInfo(StringRef VarName, Constant *Addr, int64_t VarSize, OMPTargetGlobalVarEntryKind Flags, GlobalValue::LinkageTypes Linkage)
Register device global variable entry.
void initializeDeviceGlobalVarEntryInfo(StringRef Name, OMPTargetGlobalVarEntryKind Flags, unsigned Order)
Initialize device global variable entry.
void actOnDeviceGlobalVarEntriesInfo(const OffloadDeviceGlobalVarEntryInfoActTy &Action)
OMPTargetRegionEntryKind
Kind of the target registry entry.
@ OMPTargetRegionEntryTargetRegion
Mark the entry as target region.
void getTargetRegionEntryFnName(SmallVectorImpl< char > &Name, const TargetRegionEntryInfo &EntryInfo)
bool hasTargetRegionEntryInfo(TargetRegionEntryInfo EntryInfo, bool IgnoreAddressId=false) const
Return true if a target region entry with the provided information exists.
void registerTargetRegionEntryInfo(TargetRegionEntryInfo EntryInfo, Constant *Addr, Constant *ID, OMPTargetRegionEntryKind Flags)
Register target region entry.
void actOnTargetRegionEntriesInfo(const OffloadTargetRegionEntryInfoActTy &Action)
unsigned size() const
Return number of entries defined so far.
void initializeTargetRegionEntryInfo(const TargetRegionEntryInfo &EntryInfo, unsigned Order)
Initialize target region entry.
OMPTargetGlobalVarEntryKind
Kind of the global variable entry..
@ OMPTargetGlobalVarEntryEnter
Mark the entry as a declare target enter.
@ OMPTargetGlobalRegisterRequires
Mark the entry as a register requires global.
@ OMPTargetGlobalVarEntryIndirect
Mark the entry as a declare target indirect global.
@ OMPTargetGlobalVarEntryLink
Mark the entry as a to declare target link.
@ OMPTargetGlobalVarEntryTo
Mark the entry as a to declare target.
bool hasDeviceGlobalVarEntryInfo(StringRef VarName) const
Checks if the variable with the given name has been registered already.
bool empty() const
Return true if a there are no entries defined.
std::optional< bool > IsTargetDevice
Flag to define whether to generate code for the role of the OpenMP host (if set to false) or device (...
void setGridValue(omp::GV G)
StringRef separator() const
int64_t getRequiresFlags() const
Returns requires directive clauses as flags compatible with those expected by libomptarget.
StringRef firstSeparator() const
std::optional< bool > EmitLLVMUsedMetaInfo
Flag for specifying if LLVMUsed information should be emitted.
omp::GV getGridValue() const
void setHasRequiresReverseOffload(bool Value)
bool hasRequiresUnifiedSharedMemory() const
void setHasRequiresUnifiedSharedMemory(bool Value)
bool hasRequiresDynamicAllocators() const
bool openMPOffloadMandatory() const
void setHasRequiresUnifiedAddress(bool Value)
bool isTargetDevice() const
void setHasRequiresDynamicAllocators(bool Value)
void setEmitLLVMUsed(bool Value=true)
bool hasRequiresReverseOffload() const
bool hasRequiresFlags() const
bool hasRequiresUnifiedAddress() const
Struct that keeps the information that should be kept throughout a 'target data' region.
An interface to create LLVM-IR for OpenMP directives.
Constant * getOrCreateIdent(Constant *SrcLocStr, uint32_t SrcLocStrSize, omp::IdentFlag Flags=omp::IdentFlag(0), unsigned Reserve2Flags=0)
Return an ident_t* encoding the source location SrcLocStr and Flags.
FunctionCallee getOrCreateRuntimeFunction(Module &M, omp::RuntimeFunction FnID)
Return the function declaration for the runtime function with FnID.
std::function< void(InsertPointTy CodeGenIP)> FinalizeCallbackTy
Callback type for variable finalization (think destructors).
InsertPointTy createTargetInit(const LocationDescription &Loc, bool IsSPMD, int32_t MinThreadsVal=0, int32_t MaxThreadsVal=0, int32_t MinTeamsVal=0, int32_t MaxTeamsVal=0)
The omp target interface.
void emitIfClause(Value *Cond, BodyGenCallbackTy ThenGen, BodyGenCallbackTy ElseGen, InsertPointTy AllocaIP={})
Emits code for OpenMP 'if' clause using specified BodyGenCallbackTy Here is the logic: if (Cond) { Th...
ReductionGenCBKind
Enum class for the RedctionGen CallBack type to be used.
CanonicalLoopInfo * collapseLoops(DebugLoc DL, ArrayRef< CanonicalLoopInfo * > Loops, InsertPointTy ComputeIP)
Collapse a loop nest into a single loop.
void createTaskyield(const LocationDescription &Loc)
Generator for '#omp taskyield'.
void emitBranch(BasicBlock *Target)
InsertPointTy createAtomicWrite(const LocationDescription &Loc, AtomicOpValue &X, Value *Expr, AtomicOrdering AO)
Emit atomic write for : X = Expr — Only Scalar data types.
static void writeThreadBoundsForKernel(const Triple &T, Function &Kernel, int32_t LB, int32_t UB)
InsertPointTy createCritical(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst)
Generator for '#omp critical'.
static TargetRegionEntryInfo getTargetEntryUniqueInfo(FileIdentifierInfoCallbackTy CallBack, StringRef ParentName="")
Creates a unique info for a target entry when provided a filename and line number from.
void emitTaskwaitImpl(const LocationDescription &Loc)
Generate a taskwait runtime call.
Constant * registerTargetRegionFunction(TargetRegionEntryInfo &EntryInfo, Function *OutlinedFunction, StringRef EntryFnName, StringRef EntryFnIDName)
Registers the given function and sets up the attribtues of the function Returns the FunctionID.
InsertPointTy createAtomicCapture(const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, AtomicOpValue &V, Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp, bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr)
Emit atomic update for constructs: — Only Scalar data types V = X; X = X BinOp Expr ,...
void initialize()
Initialize the internal state, this will put structures types and potentially other helpers into the ...
void createTargetDeinit(const LocationDescription &Loc, int32_t TeamsReductionDataSize=0, int32_t TeamsReductionBufferLength=1024)
Create a runtime call for kmpc_target_deinit.
CanonicalLoopInfo * createCanonicalLoop(const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB, Value *TripCount, const Twine &Name="loop")
Generator for the control flow structure of an OpenMP canonical loop.
void loadOffloadInfoMetadata(Module &M)
Loads all the offload entries information from the host IR metadata.
InsertPointTy createAtomicUpdate(const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr)
Emit atomic update for constructs: X = X BinOp Expr ,or X = Expr BinOp X For complex Operations: X = ...
void unrollLoopFull(DebugLoc DL, CanonicalLoopInfo *Loop)
Fully unroll a loop.
void emitFlush(const LocationDescription &Loc)
Generate a flush runtime call.
InsertPointTy createBarrier(const LocationDescription &Loc, omp::Directive Kind, bool ForceSimpleCall=false, bool CheckCancelFlag=true)
Emitter methods for OpenMP directives.
InsertPointTy emitKernelLaunch(const LocationDescription &Loc, Function *OutlinedFn, Value *OutlinedFnID, EmitFallbackCallbackTy EmitTargetCallFallbackCB, TargetKernelArgs &Args, Value *DeviceID, Value *RTLoc, InsertPointTy AllocaIP)
Generate a target region entry call and host fallback call.
InsertPointTy createCancel(const LocationDescription &Loc, Value *IfCondition, omp::Directive CanceledDirective)
Generator for '#omp cancel'.
static std::pair< int32_t, int32_t > readThreadBoundsForKernel(const Triple &T, Function &Kernel)
}
OpenMPIRBuilderConfig Config
The OpenMPIRBuilder Configuration.
CallInst * createOMPInteropDestroy(const LocationDescription &Loc, Value *InteropVar, Value *Device, Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause)
Create a runtime call for __tgt_interop_destroy.
InsertPointTy createAtomicRead(const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V, AtomicOrdering AO)
Emit atomic Read for : V = X — Only Scalar data types.
std::function< void(EmitMetadataErrorKind, TargetRegionEntryInfo)> EmitMetadataErrorReportFunctionTy
Callback function type.
InsertPointTy createOrderedThreadsSimd(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool IsThreads)
Generator for '#omp ordered [threads | simd]'.
OpenMPIRBuilder::InsertPointTy createTargetData(const LocationDescription &Loc, InsertPointTy AllocaIP, InsertPointTy CodeGenIP, Value *DeviceID, Value *IfCond, TargetDataInfo &Info, GenMapInfoCallbackTy GenMapInfoCB, omp::RuntimeFunction *MapperFunc=nullptr, function_ref< InsertPointTy(InsertPointTy CodeGenIP, BodyGenTy BodyGenType)> BodyGenCB=nullptr, function_ref< void(unsigned int, Value *)> DeviceAddrCB=nullptr, function_ref< Value *(unsigned int)> CustomMapperCB=nullptr, Value *SrcLocInfo=nullptr)
Generator for '#omp target data'.
std::forward_list< CanonicalLoopInfo > LoopInfos
Collection of owned canonical loop objects that eventually need to be free'd.
void createTaskwait(const LocationDescription &Loc)
Generator for '#omp taskwait'.
CanonicalLoopInfo * createLoopSkeleton(DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore, BasicBlock *PostInsertBefore, const Twine &Name={})
Create the control flow structure of a canonical OpenMP loop.
std::string createPlatformSpecificName(ArrayRef< StringRef > Parts) const
Get the create a name using the platform specific separators.
FunctionCallee createDispatchNextFunction(unsigned IVSize, bool IVSigned)
Returns __kmpc_dispatch_next_* runtime function for the specified size IVSize and sign IVSigned.
static void getKernelArgsVector(TargetKernelArgs &KernelArgs, IRBuilderBase &Builder, SmallVector< Value * > &ArgsVector)
Create the kernel args vector used by emitTargetKernel.
void unrollLoopHeuristic(DebugLoc DL, CanonicalLoopInfo *Loop)
Fully or partially unroll a loop.
omp::OpenMPOffloadMappingFlags getMemberOfFlag(unsigned Position)
Get OMP_MAP_MEMBER_OF flag with extra bits reserved based on the position given.
InsertPointTy createReductionsGPU(const LocationDescription &Loc, InsertPointTy AllocaIP, InsertPointTy CodeGenIP, ArrayRef< ReductionInfo > ReductionInfos, bool IsNoWait=false, bool IsTeamsReduction=false, bool HasDistribute=false, ReductionGenCBKind ReductionGenCBKind=ReductionGenCBKind::MLIR, std::optional< omp::GV > GridValue={}, unsigned ReductionBufNum=1024, Value *SrcLocInfo=nullptr)
Design of OpenMP reductions on the GPU.
void addAttributes(omp::RuntimeFunction FnID, Function &Fn)
Add attributes known for FnID to Fn.
Module & M
The underlying LLVM-IR module.
StringMap< Constant * > SrcLocStrMap
Map to remember source location strings.
void createMapperAllocas(const LocationDescription &Loc, InsertPointTy AllocaIP, unsigned NumOperands, struct MapperAllocas &MapperAllocas)
Create the allocas instruction used in call to mapper functions.
Constant * getOrCreateSrcLocStr(StringRef LocStr, uint32_t &SrcLocStrSize)
Return the (LLVM-IR) string describing the source location LocStr.
void addOutlineInfo(OutlineInfo &&OI)
Add a new region that will be outlined later.
FunctionCallee createDispatchFiniFunction(unsigned IVSize, bool IVSigned)
Returns __kmpc_dispatch_fini_* runtime function for the specified size IVSize and sign IVSigned.
void emitOffloadingArraysArgument(IRBuilderBase &Builder, OpenMPIRBuilder::TargetDataRTArgs &RTArgs, OpenMPIRBuilder::TargetDataInfo &Info, bool EmitDebug=false, bool ForEndCall=false)
Emit the arguments to be passed to the runtime library based on the arrays of base pointers,...
void unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop, int32_t Factor, CanonicalLoopInfo **UnrolledCLI)
Partially unroll a loop.
InsertPointTy createSections(const LocationDescription &Loc, InsertPointTy AllocaIP, ArrayRef< StorableBodyGenCallbackTy > SectionCBs, PrivatizeCallbackTy PrivCB, FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait)
Generator for '#omp sections'.
InsertPointTy createTask(const LocationDescription &Loc, InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB, bool Tied=true, Value *Final=nullptr, Value *IfCondition=nullptr, SmallVector< DependData > Dependencies={})
Generator for #omp task
void emitTaskyieldImpl(const LocationDescription &Loc)
Generate a taskyield runtime call.
void emitMapperCall(const LocationDescription &Loc, Function *MapperFunc, Value *SrcLocInfo, Value *MaptypesArg, Value *MapnamesArg, struct MapperAllocas &MapperAllocas, int64_t DeviceID, unsigned NumOperands)
Create the call for the target mapper function.
InsertPointTy createAtomicCompare(const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V, AtomicOpValue &R, Value *E, Value *D, AtomicOrdering AO, omp::OMPAtomicCompareOp Op, bool IsXBinopExpr, bool IsPostfixUpdate, bool IsFailOnly)
Emit atomic compare for constructs: — Only scalar data types cond-expr-stmt: x = x ordop expr ?...
InsertPointTy createOrderedDepend(const LocationDescription &Loc, InsertPointTy AllocaIP, unsigned NumLoops, ArrayRef< llvm::Value * > StoreValues, const Twine &Name, bool IsDependSource)
Generator for '#omp ordered depend (source | sink)'.
InsertPointTy createCopyinClauseBlocks(InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr, llvm::IntegerType *IntPtrTy, bool BranchtoEnd=true)
Generate conditional branch and relevant BasicBlocks through which private threads copy the 'copyin' ...
void emitOffloadingArrays(InsertPointTy AllocaIP, InsertPointTy CodeGenIP, MapInfosTy &CombinedInfo, TargetDataInfo &Info, bool IsNonContiguous=false, function_ref< void(unsigned int, Value *)> DeviceAddrCB=nullptr, function_ref< Value *(unsigned int)> CustomMapperCB=nullptr)
Emit the arrays used to pass the captures and map information to the offloading runtime library.
SmallVector< FinalizationInfo, 8 > FinalizationStack
The finalization stack made up of finalize callbacks currently in-flight, wrapped into FinalizationIn...
std::vector< CanonicalLoopInfo * > tileLoops(DebugLoc DL, ArrayRef< CanonicalLoopInfo * > Loops, ArrayRef< Value * > TileSizes)
Tile a loop nest.
CallInst * createOMPInteropInit(const LocationDescription &Loc, Value *InteropVar, omp::OMPInteropType InteropType, Value *Device, Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause)
Create a runtime call for __tgt_interop_init.
void finalize(Function *Fn=nullptr)
Finalize the underlying module, e.g., by outlining regions.
SmallVector< OutlineInfo, 16 > OutlineInfos
Collection of regions that need to be outlined during finalization.
Function * getOrCreateRuntimeFunctionPtr(omp::RuntimeFunction FnID)
const Triple T
The target triple of the underlying module.
DenseMap< std::pair< Constant *, uint64_t >, Constant * > IdentMap
Map to remember existing ident_t*.
CallInst * createOMPFree(const LocationDescription &Loc, Value *Addr, Value *Allocator, std::string Name="")
Create a runtime call for kmpc_free.
FunctionCallee createForStaticInitFunction(unsigned IVSize, bool IVSigned, bool IsGPUDistribute)
Returns __kmpc_for_static_init_* runtime function for the specified size IVSize and sign IVSigned.
CallInst * createOMPAlloc(const LocationDescription &Loc, Value *Size, Value *Allocator, std::string Name="")
Create a runtime call for kmpc_Alloc.
void emitNonContiguousDescriptor(InsertPointTy AllocaIP, InsertPointTy CodeGenIP, MapInfosTy &CombinedInfo, TargetDataInfo &Info)
Emit an array of struct descriptors to be assigned to the offload args.
InsertPointTy createSection(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB)
Generator for '#omp section'.
InsertPointTy applyWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, bool NeedsBarrier, llvm::omp::ScheduleKind SchedKind=llvm::omp::OMP_SCHEDULE_Default, Value *ChunkSize=nullptr, bool HasSimdModifier=false, bool HasMonotonicModifier=false, bool HasNonmonotonicModifier=false, bool HasOrderedClause=false, omp::WorksharingLoopType LoopType=omp::WorksharingLoopType::ForStaticLoop)
Modifies the canonical loop to be a workshare loop.
void emitBlock(BasicBlock *BB, Function *CurFn, bool IsFinished=false)
Value * getOrCreateThreadID(Value *Ident)
Return the current thread ID.
InsertPointTy createMaster(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB)
Generator for '#omp master'.
IRBuilder ::InsertPoint createParallel(const LocationDescription &Loc, InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB, PrivatizeCallbackTy PrivCB, FinalizeCallbackTy FiniCB, Value *IfCondition, Value *NumThreads, omp::ProcBindKind ProcBind, bool IsCancellable)
Generator for '#omp parallel'.
StringMap< GlobalVariable *, BumpPtrAllocator > InternalVars
An ordered map of auto-generated variables to their unique names.
GlobalVariable * getOrCreateInternalVariable(Type *Ty, const StringRef &Name, unsigned AddressSpace=0)
Gets (if variable with the given name already exist) or creates internal global variable with the spe...
FunctionCallee createDispatchInitFunction(unsigned IVSize, bool IVSigned)
Returns __kmpc_dispatch_init_* runtime function for the specified size IVSize and sign IVSigned.
InsertPointTy createSingle(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool IsNowait, ArrayRef< llvm::Value * > CPVars={}, ArrayRef< llvm::Function * > CPFuncs={})
Generator for '#omp single'.
CallInst * createOMPInteropUse(const LocationDescription &Loc, Value *InteropVar, Value *Device, Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause)
Create a runtime call for __tgt_interop_use.
IRBuilder<>::InsertPoint InsertPointTy
Type used throughout for insertion points.
GlobalVariable * createOffloadMapnames(SmallVectorImpl< llvm::Constant * > &Names, std::string VarName)
Create the global variable holding the offload names information.
static void writeTeamsForKernel(const Triple &T, Function &Kernel, int32_t LB, int32_t UB)
std::function< Function *(StringRef FunctionName)> FunctionGenCallback
Functions used to generate a function with the given name.
void setCorrectMemberOfFlag(omp::OpenMPOffloadMappingFlags &Flags, omp::OpenMPOffloadMappingFlags MemberOfFlag)
Given an initial flag set, this function modifies it to contain the passed in MemberOfFlag generated ...
void emitCancelationCheckImpl(Value *CancelFlag, omp::Directive CanceledDirective, FinalizeCallbackTy ExitCB={})
Generate control flow and cleanup for cancellation.
Constant * getOrCreateDefaultSrcLocStr(uint32_t &SrcLocStrSize)
Return the (LLVM-IR) string describing the default source location.
InsertPointTy createMasked(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, Value *Filter)
Generator for '#omp masked'.
void createOffloadEntry(Constant *ID, Constant *Addr, uint64_t Size, int32_t Flags, GlobalValue::LinkageTypes, StringRef Name="")
Creates offloading entry for the provided entry ID ID, address Addr, size Size, and flags Flags.
static unsigned getOpenMPDefaultSimdAlign(const Triple &TargetTriple, const StringMap< bool > &Features)
Get the default alignment value for given target.
unsigned getFlagMemberOffset()
Get the offset of the OMP_MAP_MEMBER_OF field.
InsertPointTy createTaskgroup(const LocationDescription &Loc, InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB)
Generator for the taskgroup construct.
void createOffloadEntriesAndInfoMetadata(EmitMetadataErrorReportFunctionTy &ErrorReportFunction)
void applySimd(CanonicalLoopInfo *Loop, MapVector< Value *, Value * > AlignedVars, Value *IfCond, omp::OrderKind Order, ConstantInt *Simdlen, ConstantInt *Safelen)
Add metadata to simd-ize a loop.
void emitTargetRegionFunction(TargetRegionEntryInfo &EntryInfo, FunctionGenCallback &GenerateFunctionCallback, bool IsOffloadEntry, Function *&OutlinedFn, Constant *&OutlinedFnID)
Create a unique name for the entry function using the source location information of the current targ...
InsertPointTy createTarget(const LocationDescription &Loc, OpenMPIRBuilder::InsertPointTy AllocaIP, OpenMPIRBuilder::InsertPointTy CodeGenIP, TargetRegionEntryInfo &EntryInfo, int32_t NumTeams, int32_t NumThreads, SmallVectorImpl< Value * > &Inputs, GenMapInfoCallbackTy GenMapInfoCB, TargetBodyGenCallbackTy BodyGenCB, TargetGenArgAccessorsCallbackTy ArgAccessorFuncCB)
Generator for '#omp target'.
bool isLastFinalizationInfoCancellable(omp::Directive DK)
Return true if the last entry in the finalization stack is of kind DK and cancellable.
InsertPointTy emitTargetKernel(const LocationDescription &Loc, InsertPointTy AllocaIP, Value *&Return, Value *Ident, Value *DeviceID, Value *NumTeams, Value *NumThreads, Value *HostPtr, ArrayRef< Value * > KernelArgs)
Generate a target region entry call.
GlobalVariable * createOffloadMaptypes(SmallVectorImpl< uint64_t > &Mappings, std::string VarName)
Create the global variable holding the offload mappings information.
CallInst * createCachedThreadPrivate(const LocationDescription &Loc, llvm::Value *Pointer, llvm::ConstantInt *Size, const llvm::Twine &Name=Twine(""))
Create a runtime call for kmpc_threadprivate_cached.
IRBuilder Builder
The LLVM-IR Builder used to create IR.
GlobalValue * createGlobalFlag(unsigned Value, StringRef Name)
Create a hidden global flag Name in the module with initial value Value.
Value * getSizeInBytes(Value *BasePtr)
Computes the size of type in bytes.
void registerTargetGlobalVariable(OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind CaptureClause, OffloadEntriesInfoManager::OMPTargetDeviceClauseKind DeviceClause, bool IsDeclaration, bool IsExternallyVisible, TargetRegionEntryInfo EntryInfo, StringRef MangledName, std::vector< GlobalVariable * > &GeneratedRefs, bool OpenMPSIMD, std::vector< Triple > TargetTriple, std::function< Constant *()> GlobalInitializer, std::function< GlobalValue::LinkageTypes()> VariableLinkage, Type *LlvmPtrTy, Constant *Addr)
Registers a target variable for device or host.
InsertPointTy createTeams(const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, Value *NumTeamsLower=nullptr, Value *NumTeamsUpper=nullptr, Value *ThreadLimit=nullptr, Value *IfExpr=nullptr)
Generator for #omp teams
BodyGenTy
Type of BodyGen to use for region codegen.
SmallVector< llvm::Function *, 16 > ConstantAllocaRaiseCandidates
A collection of candidate target functions that's constant allocas will attempt to be raised on a cal...
OffloadEntriesInfoManager OffloadInfoManager
Info manager to keep track of target regions.
static std::pair< int32_t, int32_t > readTeamBoundsForKernel(const Triple &T, Function &Kernel)
Read/write a bounds on teams for Kernel.
std::function< std::tuple< std::string, uint64_t >()> FileIdentifierInfoCallbackTy
const std::string ompOffloadInfoName
OMP Offload Info Metadata name string.
InsertPointTy createCopyPrivate(const LocationDescription &Loc, llvm::Value *BufSize, llvm::Value *CpyBuf, llvm::Value *CpyFn, llvm::Value *DidIt)
Generator for __kmpc_copyprivate.
InsertPointTy createReductions(const LocationDescription &Loc, InsertPointTy AllocaIP, ArrayRef< ReductionInfo > ReductionInfos, ArrayRef< bool > IsByRef, bool IsNoWait=false)
Generator for '#omp reduction'.
bool updateToLocation(const LocationDescription &Loc)
Update the internal location to Loc.
void createFlush(const LocationDescription &Loc)
Generator for '#omp flush'.
Constant * getAddrOfDeclareTargetVar(OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind CaptureClause, OffloadEntriesInfoManager::OMPTargetDeviceClauseKind DeviceClause, bool IsDeclaration, bool IsExternallyVisible, TargetRegionEntryInfo EntryInfo, StringRef MangledName, std::vector< GlobalVariable * > &GeneratedRefs, bool OpenMPSIMD, std::vector< Triple > TargetTriple, Type *LlvmPtrTy, std::function< Constant *()> GlobalInitializer, std::function< GlobalValue::LinkageTypes()> VariableLinkage)
Retrieve (or create if non-existent) the address of a declare target variable, used in conjunction wi...
EmitMetadataErrorKind
The kind of errors that can occur when emitting the offload entries and metadata.
@ EMIT_MD_DECLARE_TARGET_ERROR
@ EMIT_MD_GLOBAL_VAR_LINK_ERROR
@ EMIT_MD_TARGET_REGION_ERROR
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
Pseudo-analysis pass that exposes the PassInstrumentation to pass managers.
Class to represent pointers.
static PointerType * getUnqual(Type *ElementType)
This constructs a pointer to an object of the specified type in the default address space (address sp...
Analysis pass that exposes the ScalarEvolution for a function.
ScalarEvolution run(Function &F, FunctionAnalysisManager &AM)
The main scalar evolution driver.
A vector that has set insertion semantics.
bool empty() const
Determine if the SetVector is empty or not.
This is a 'bitvector' (really, a variable-sized bit array), optimized for the case when the array is ...
bool test(unsigned Idx) const
bool all() const
Returns true if all bits are set.
bool any() const
Returns true if any bit is set.
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...
std::pair< const_iterator, bool > insert(const T &V)
insert - Insert an element into the set if it isn't already there.
SmallString - A SmallString is just a SmallVector with methods and accessors that make it work better...
void append(StringRef RHS)
Append from a StringRef.
StringRef str() const
Explicit conversion to StringRef.
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
reference emplace_back(ArgTypes &&... Args)
void reserve(size_type N)
void append(ItTy in_start, ItTy in_end)
Add the specified range to the end of the SmallVector.
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.
void setAlignment(Align Align)
void setAtomic(AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Sets the ordering constraint and the synchronization scope ID of this store instruction.
StringMap - This is an unconventional map that is specialized for handling keys that are "strings",...
ValueTy lookup(StringRef Key) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
StringRef - Represent a constant reference to a string, i.e.
std::pair< StringRef, StringRef > split(char Separator) const
Split into two substrings around the first occurrence of a separator character.
constexpr bool empty() const
empty - Check if the string is empty.
constexpr size_t size() const
size - Get the string size.
size_t count(char C) const
Return the number of occurrences of C in the string.
bool ends_with(StringRef Suffix) const
Check if this string ends with the given Suffix.
StringRef drop_back(size_t N=1) const
Return a StringRef equal to 'this' but with the last N elements dropped.
Class to represent struct types.
static StructType * create(LLVMContext &Context, StringRef Name)
This creates an identified struct.
void addCase(ConstantInt *OnVal, BasicBlock *Dest)
Add an entry to the switch instruction.
Analysis pass providing the TargetTransformInfo.
Result run(const Function &F, FunctionAnalysisManager &)
Analysis pass providing the TargetLibraryInfo.
Target - Wrapper for Target specific information.
TargetMachine * createTargetMachine(StringRef TT, StringRef CPU, StringRef Features, const TargetOptions &Options, std::optional< Reloc::Model > RM, std::optional< CodeModel::Model > CM=std::nullopt, CodeGenOptLevel OL=CodeGenOptLevel::Default, bool JIT=false) const
createTargetMachine - Create a target specific machine implementation for the specified Triple.
Triple - Helper class for working with autoconf configuration names.
bool isPPC() const
Tests whether the target is PowerPC (32- or 64-bit LE or BE).
bool isX86() const
Tests whether the target is x86 (32- or 64-bit).
bool isWasm() const
Tests whether the target is wasm (32- and 64-bit).
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
std::string str() const
Return the twine contents as a std::string.
The instances of the Type class are immutable: once they are created, they are never changed.
unsigned getIntegerBitWidth() const
Type * getStructElementType(unsigned N) const
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
bool isPointerTy() const
True if this is an instance of PointerType.
static IntegerType * getInt1Ty(LLVMContext &C)
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
static Type * getVoidTy(LLVMContext &C)
bool isFloatingPointTy() const
Return true if this is one of the floating-point types.
static IntegerType * getInt32Ty(LLVMContext &C)
static IntegerType * getInt64Ty(LLVMContext &C)
bool isIntegerTy() const
True if this is an instance of IntegerType.
bool isVoidTy() const
Return true if this is 'void'.
static UndefValue * get(Type *T)
Static factory methods - Return an 'undef' object of the specified type.
This function has undefined behavior.
Produce an estimate of the unrolled cost of the specified loop.
bool canUnroll() const
Whether it is legal to unroll this loop.
uint64_t getRolledLoopSize() const
A Use represents the edge between a Value definition and its users.
void setOperand(unsigned i, Value *Val)
Value * getOperand(unsigned i) const
ValueT lookup(const KeyT &Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
LLVM Value Representation.
Type * getType() const
All values are typed, get the type of this value.
user_iterator user_begin()
void setName(const Twine &Name)
Change the name of the value.
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
iterator_range< user_iterator > users()
Align getPointerAlignment(const DataLayout &DL) const
Returns an alignment of the pointer value.
void replaceUsesWithIf(Value *New, llvm::function_ref< bool(Use &U)> ShouldReplace)
Go through the uses list for this definition and make each use point to "V" if the callback ShouldRep...
User * getUniqueUndroppableUser()
Return true if there is exactly one unique user of this value that cannot be dropped (that user can h...
unsigned getNumUses() const
This method computes the number of uses of this Value.
iterator_range< use_iterator > uses()
StringRef getName() const
Return a constant reference to the value's name.
An efficient, type-erasing, non-owning reference to a callable.
const ParentTy * getParent() const
self_iterator getIterator()
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
iterator insertAfter(iterator where, pointer New)
A raw_ostream that writes to an SmallVector or SmallString.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
@ AMDGPU_KERNEL
Used for AMDGPU code object kernels.
@ C
The default llvm calling convention, compatible with C.
initializer< Ty > init(const Ty &Val)
void emitOffloadingEntry(Module &M, Constant *Addr, StringRef Name, uint64_t Size, int32_t Flags, int32_t Data, StringRef SectionName)
Create an offloading section struct used to register this global at runtime.
OpenMPOffloadMappingFlags
Values for bit flags used to specify the mapping type for offloading.
@ OMP_MAP_PTR_AND_OBJ
The element being mapped is a pointer-pointee pair; both the pointer and the pointee should be mapped...
@ OMP_MAP_MEMBER_OF
The 16 MSBs of the flags indicate whether the entry is member of some struct/class.
@ OMP_DEVICEID_UNDEF
Device ID if the device was not defined, runtime should get it from environment variables in the spec...
IdentFlag
IDs for all omp runtime library ident_t flag encodings (see their defintion in openmp/runtime/src/kmp...
RuntimeFunction
IDs for all omp runtime library (RTL) functions.
static constexpr GV NVPTXGridValues
For Nvidia GPUs.
@ OMP_TGT_EXEC_MODE_GENERIC
WorksharingLoopType
A type of worksharing loop construct.
OMPAtomicCompareOp
Atomic compare operations. Currently OpenMP only supports ==, >, and <.
NodeAddr< PhiNode * > Phi
std::error_code getUniqueID(const Twine Path, UniqueID &Result)
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.
BasicBlock * splitBBWithSuffix(IRBuilderBase &Builder, bool CreateBranch, llvm::Twine Suffix=".split")
Like splitBB, but reuses the current block's name for the new name.
detail::zippy< detail::zip_shortest, T, U, Args... > zip(T &&t, U &&u, Args &&...args)
zip iterator for two or more iteratable types.
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
auto size(R &&Range, std::enable_if_t< std::is_base_of< std::random_access_iterator_tag, typename std::iterator_traits< decltype(Range.begin())>::iterator_category >::value, void > *=nullptr)
Get the size of a range.
Expected< std::unique_ptr< Module > > parseBitcodeFile(MemoryBufferRef Buffer, LLVMContext &Context, ParserCallbacks Callbacks={})
Read the specified bitcode file, returning the module.
uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator)
Returns the integer ceil(Numerator / Denominator).
auto enumerate(FirstRange &&First, RestRanges &&...Rest)
Given two or more input ranges, returns a new range whose values are are tuples (A,...
decltype(auto) dyn_cast(const From &Val)
dyn_cast<X> - Return the argument parameter cast to the specified type.
auto successors(const MachineBasicBlock *BB)
testing::Matcher< const detail::ErrorHolder & > Failed()
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
void append_range(Container &C, Range &&R)
Wrapper function to append range R to container C.
iterator_range< early_inc_iterator_impl< detail::IterOfRange< RangeT > > > make_early_inc_range(RangeT &&Range)
Make a range that does early increment to allow mutation of the underlying range without disrupting i...
ErrorOr< T > expectedToErrorOrAndEmitErrors(LLVMContext &Ctx, Expected< T > Val)
bool convertUsersOfConstantsToInstructions(ArrayRef< Constant * > Consts, Function *RestrictToFunc=nullptr, bool RemoveDeadConstants=true, bool IncludeSelf=false)
Replace constant expressions users of the given constants with instructions.
unsigned Log2_32(uint32_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
BasicBlock * CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, const Twine &NameSuffix="", Function *F=nullptr, ClonedCodeInfo *CodeInfo=nullptr, DebugInfoFinder *DIFinder=nullptr)
Return a copy of the specified basic block, but without embedding the block into a particular functio...
TargetTransformInfo::PeelingPreferences gatherPeelingPreferences(Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, std::optional< bool > UserAllowPeeling, std::optional< bool > UserAllowProfileBasedPeeling, bool UnrollingSpecficValues=false)
void spliceBB(IRBuilderBase::InsertPoint IP, BasicBlock *New, bool CreateBranch)
Move the instruction after an InsertPoint to the beginning of another BasicBlock.
void SplitBlockAndInsertIfThenElse(Value *Cond, BasicBlock::iterator SplitBefore, Instruction **ThenTerm, Instruction **ElseTerm, MDNode *BranchWeights=nullptr, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr)
SplitBlockAndInsertIfThenElse is similar to SplitBlockAndInsertIfThen, but also creates the ElseBlock...
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
bool none_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::none_of which take ranges instead of having to pass begin/end explicitly.
void report_fatal_error(Error Err, bool gen_crash_diag=true)
Report a serious error, calling any installed error handler.
BasicBlock * splitBB(IRBuilderBase::InsertPoint IP, bool CreateBranch, llvm::Twine Name={})
Split a BasicBlock at an InsertPoint, even if the block is degenerate (missing the terminator).
CodeGenOptLevel
Code generation optimization level.
bool computeUnrollCount(Loop *L, const TargetTransformInfo &TTI, DominatorTree &DT, LoopInfo *LI, AssumptionCache *AC, ScalarEvolution &SE, const SmallPtrSetImpl< const Value * > &EphValues, OptimizationRemarkEmitter *ORE, unsigned TripCount, unsigned MaxTripCount, bool MaxOrZero, unsigned TripMultiple, const UnrollCostEstimator &UCE, TargetTransformInfo::UnrollingPreferences &UP, TargetTransformInfo::PeelingPreferences &PP, bool &UseUpperBound)
format_object< Ts... > format(const char *Fmt, const Ts &... Vals)
These are helper functions used to produce formatted output.
raw_fd_ostream & errs()
This returns a reference to a raw_ostream for standard error.
AtomicOrdering
Atomic ordering for LLVM's memory model.
bool MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, MemoryDependenceResults *MemDep=nullptr, bool PredecessorWithTwoSuccessors=false, DominatorTree *DT=nullptr)
Attempts to merge a block into its predecessor, if possible.
DWARFExpression::Operation Op
void remapInstructionsInBlocks(ArrayRef< BasicBlock * > Blocks, ValueToValueMapTy &VMap)
Remaps instructions in Blocks using the mapping in VMap.
TargetTransformInfo::UnrollingPreferences gatherUnrollingPreferences(Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, llvm::OptimizationRemarkEmitter &ORE, int OptLevel, std::optional< unsigned > UserThreshold, std::optional< unsigned > UserCount, std::optional< bool > UserAllowPartial, std::optional< bool > UserRuntime, std::optional< bool > UserUpperBound, std::optional< unsigned > UserFullUnrollMaxCount)
Gather the various unrolling parameters based on the defaults, compiler flags, TTI overrides and user...
BasicBlock * SplitBlock(BasicBlock *Old, BasicBlock::iterator SplitPt, DominatorTree *DT, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, const Twine &BBName="", bool Before=false)
Split the specified block at the specified instruction.
auto predecessors(const MachineBasicBlock *BB)
Constant * ConstantFoldInsertValueInstruction(Constant *Agg, Constant *Val, ArrayRef< unsigned > Idxs)
ConstantFoldInsertValueInstruction - Attempt to constant fold an insertvalue instruction with the spe...
void DeleteDeadBlocks(ArrayRef< BasicBlock * > BBs, DomTreeUpdater *DTU=nullptr, bool KeepOneInputPHIs=false)
Delete the specified blocks from BB.
This struct is a compact representation of a valid (non-zero power of two) alignment.
static void collectEphemeralValues(const Loop *L, AssumptionCache *AC, SmallPtrSetImpl< const Value * > &EphValues)
Collect a loop's ephemeral values (those used only by an assume or similar intrinsics in the loop).
This struct is a compact representation of a valid (power of two) or undefined (0) alignment.
a struct to pack relevant information while generating atomic Ops
A struct to pack the relevant information for an OpenMP depend clause.
Description of a LLVM-IR insertion point (IP) and a debug/source location (filename,...
MapNonContiguousArrayTy Offsets
MapNonContiguousArrayTy Counts
MapNonContiguousArrayTy Strides
This structure contains combined information generated for mappable clauses, including base pointers,...
MapDeviceInfoArrayTy DevicePointers
MapValuesArrayTy BasePointers
MapValuesArrayTy Pointers
StructNonContiguousInfo NonContigInfo
Helper that contains information about regions we need to outline during finalization.
PostOutlineCBTy PostOutlineCB
void collectBlocks(SmallPtrSetImpl< BasicBlock * > &BlockSet, SmallVectorImpl< BasicBlock * > &BlockVector)
Collect all blocks in between EntryBB and ExitBB in both the given vector and set.
SmallVector< Value *, 2 > ExcludeArgsFromAggregate
BasicBlock * OuterAllocaBB
Information about an OpenMP reduction.
EvalKind EvaluationKind
Reduction evaluation kind - scalar, complex or aggregate.
ReductionGenAtomicCBTy AtomicReductionGen
Callback for generating the atomic reduction body, may be null.
ReductionGenCBTy ReductionGen
Callback for generating the reduction body.
Value * Variable
Reduction variable of pointer type.
Value * PrivateVariable
Thread-private partial reduction variable.
ReductionGenClangCBTy ReductionGenClang
Clang callback for generating the reduction body.
Type * ElementType
Reduction element type, must match pointee type of variable.
Container for the arguments used to pass data to the runtime library.
Value * SizesArray
The array of sizes passed to the runtime library.
Value * PointersArray
The array of section pointers passed to the runtime library.
Value * MappersArray
The array of user-defined mappers passed to the runtime library.
Value * BasePointersArray
The array of base pointer passed to the runtime library.
Value * MapTypesArray
The array of map types passed to the runtime library for the beginning of the region or for the entir...
Value * MapNamesArray
The array of original declaration names of mapped pointers sent to the runtime library for debugging.
Data structure that contains the needed information to construct the kernel args vector.
Value * NumTeams
The number of teams.
Value * DynCGGroupMem
The size of the dynamic shared memory.
TargetDataRTArgs RTArgs
Arguments passed to the runtime library.
Value * NumIterations
The number of iterations.
unsigned NumTargetItems
Number of arguments passed to the runtime library.
bool HasNoWait
True if the kernel has 'no wait' clause.
Value * NumThreads
The number of threads.
Data structure to contain the information needed to uniquely identify a target entry.
static void getTargetRegionEntryFnName(SmallVectorImpl< char > &Name, StringRef ParentName, unsigned DeviceID, unsigned FileID, unsigned Line, unsigned Count)
static const Target * lookupTarget(StringRef Triple, std::string &Error)
lookupTarget - Lookup a target based on a target triple.
Defines various target-specific GPU grid values that must be consistent between host RTL (plugin),...
unsigned GV_Warp_Size
The default value of maximum number of threads in a worker warp.