aboutsummaryrefslogtreecommitdiff
path: root/flang/lib/Lower
diff options
context:
space:
mode:
Diffstat (limited to 'flang/lib/Lower')
-rw-r--r--flang/lib/Lower/Allocatable.cpp13
-rw-r--r--flang/lib/Lower/Bridge.cpp5
-rw-r--r--flang/lib/Lower/CUDA.cpp27
-rw-r--r--flang/lib/Lower/IO.cpp3
-rw-r--r--flang/lib/Lower/OpenMP/Atomic.cpp97
-rw-r--r--flang/lib/Lower/OpenMP/ClauseProcessor.cpp6
-rw-r--r--flang/lib/Lower/OpenMP/Clauses.cpp85
-rw-r--r--flang/lib/Lower/OpenMP/OpenMP.cpp15
8 files changed, 205 insertions, 46 deletions
diff --git a/flang/lib/Lower/Allocatable.cpp b/flang/lib/Lower/Allocatable.cpp
index 53239cb..e7a6c4d 100644
--- a/flang/lib/Lower/Allocatable.cpp
+++ b/flang/lib/Lower/Allocatable.cpp
@@ -629,6 +629,10 @@ private:
unsigned allocatorIdx = Fortran::lower::getAllocatorIdx(alloc.getSymbol());
fir::ExtendedValue exv = isSource ? sourceExv : moldExv;
+ if (const Fortran::semantics::Symbol *sym{GetLastSymbol(sourceExpr)})
+ if (Fortran::semantics::IsCUDADevice(*sym))
+ TODO(loc, "CUDA Fortran: allocate with device source");
+
// Generate a sequence of runtime calls.
errorManager.genStatCheck(builder, loc);
genAllocateObjectInit(box, allocatorIdx);
@@ -767,6 +771,15 @@ private:
const fir::MutableBoxValue &box,
ErrorManager &errorManager,
const Fortran::semantics::Symbol &sym) {
+
+ if (const Fortran::semantics::DeclTypeSpec *declTypeSpec = sym.GetType())
+ if (const Fortran::semantics::DerivedTypeSpec *derivedTypeSpec =
+ declTypeSpec->AsDerived())
+ if (derivedTypeSpec->HasDefaultInitialization(
+ /*ignoreAllocatable=*/true, /*ignorePointer=*/true))
+ TODO(loc,
+ "CUDA Fortran: allocate on device with default initialization");
+
Fortran::lower::StatementContext stmtCtx;
cuf::DataAttributeAttr cudaAttr =
Fortran::lower::translateSymbolCUFDataAttribute(builder.getContext(),
diff --git a/flang/lib/Lower/Bridge.cpp b/flang/lib/Lower/Bridge.cpp
index 68adf34..525fb0e 100644
--- a/flang/lib/Lower/Bridge.cpp
+++ b/flang/lib/Lower/Bridge.cpp
@@ -4987,11 +4987,8 @@ private:
// host = device
if (!lhsIsDevice && rhsIsDevice) {
- if (Fortran::lower::isTransferWithConversion(rhs)) {
+ if (auto elementalOp = Fortran::lower::isTransferWithConversion(rhs)) {
mlir::OpBuilder::InsertionGuard insertionGuard(builder);
- auto elementalOp =
- mlir::dyn_cast<hlfir::ElementalOp>(rhs.getDefiningOp());
- assert(elementalOp && "expect elemental op");
auto designateOp =
*elementalOp.getBody()->getOps<hlfir::DesignateOp>().begin();
builder.setInsertionPoint(elementalOp);
diff --git a/flang/lib/Lower/CUDA.cpp b/flang/lib/Lower/CUDA.cpp
index bb4bdee..9501b0e 100644
--- a/flang/lib/Lower/CUDA.cpp
+++ b/flang/lib/Lower/CUDA.cpp
@@ -68,11 +68,26 @@ cuf::DataAttributeAttr Fortran::lower::translateSymbolCUFDataAttribute(
return cuf::getDataAttribute(mlirContext, cudaAttr);
}
-bool Fortran::lower::isTransferWithConversion(mlir::Value rhs) {
+hlfir::ElementalOp Fortran::lower::isTransferWithConversion(mlir::Value rhs) {
+ auto isConversionElementalOp = [](hlfir::ElementalOp elOp) {
+ return llvm::hasSingleElement(
+ elOp.getBody()->getOps<hlfir::DesignateOp>()) &&
+ llvm::hasSingleElement(elOp.getBody()->getOps<fir::LoadOp>()) == 1 &&
+ llvm::hasSingleElement(elOp.getBody()->getOps<fir::ConvertOp>()) ==
+ 1;
+ };
+ if (auto declOp = mlir::dyn_cast<hlfir::DeclareOp>(rhs.getDefiningOp())) {
+ if (!declOp.getMemref().getDefiningOp())
+ return {};
+ if (auto associateOp = mlir::dyn_cast<hlfir::AssociateOp>(
+ declOp.getMemref().getDefiningOp()))
+ if (auto elOp = mlir::dyn_cast<hlfir::ElementalOp>(
+ associateOp.getSource().getDefiningOp()))
+ if (isConversionElementalOp(elOp))
+ return elOp;
+ }
if (auto elOp = mlir::dyn_cast<hlfir::ElementalOp>(rhs.getDefiningOp()))
- if (llvm::hasSingleElement(elOp.getBody()->getOps<hlfir::DesignateOp>()) &&
- llvm::hasSingleElement(elOp.getBody()->getOps<fir::LoadOp>()) == 1 &&
- llvm::hasSingleElement(elOp.getBody()->getOps<fir::ConvertOp>()) == 1)
- return true;
- return false;
+ if (isConversionElementalOp(elOp))
+ return elOp;
+ return {};
}
diff --git a/flang/lib/Lower/IO.cpp b/flang/lib/Lower/IO.cpp
index 604b137..cd53dc9 100644
--- a/flang/lib/Lower/IO.cpp
+++ b/flang/lib/Lower/IO.cpp
@@ -950,7 +950,8 @@ static void genIoLoop(Fortran::lower::AbstractConverter &converter,
makeNextConditionalOn(builder, loc, checkResult, ok, inLoop);
const auto &itemList = std::get<0>(ioImpliedDo.t);
const auto &control = std::get<1>(ioImpliedDo.t);
- const auto &loopSym = *control.name.thing.thing.symbol;
+ const auto &loopSym =
+ *Fortran::parser::UnwrapRef<Fortran::parser::Name>(control.name).symbol;
mlir::Value loopVar = fir::getBase(converter.genExprAddr(
Fortran::evaluate::AsGenericExpr(loopSym).value(), stmtCtx));
auto genControlValue = [&](const Fortran::parser::ScalarIntExpr &expr) {
diff --git a/flang/lib/Lower/OpenMP/Atomic.cpp b/flang/lib/Lower/OpenMP/Atomic.cpp
index ff82a36..3ab8a58 100644
--- a/flang/lib/Lower/OpenMP/Atomic.cpp
+++ b/flang/lib/Lower/OpenMP/Atomic.cpp
@@ -20,6 +20,7 @@
#include "flang/Optimizer/Builder/FIRBuilder.h"
#include "flang/Optimizer/Builder/Todo.h"
#include "flang/Parser/parse-tree.h"
+#include "flang/Semantics/openmp-utils.h"
#include "flang/Semantics/semantics.h"
#include "flang/Semantics/type.h"
#include "flang/Support/Fortran.h"
@@ -183,12 +184,8 @@ getMemoryOrderFromRequires(const semantics::Scope &scope) {
// scope.
// For safety, traverse all enclosing scopes and check if their symbol
// contains REQUIRES.
- for (const auto *sc{&scope}; sc->kind() != semantics::Scope::Kind::Global;
- sc = &sc->parent()) {
- const semantics::Symbol *sym = sc->symbol();
- if (!sym)
- continue;
-
+ const semantics::Scope &unitScope = semantics::omp::GetProgramUnit(scope);
+ if (auto *symbol = unitScope.symbol()) {
const common::OmpMemoryOrderType *admo = common::visit(
[](auto &&s) {
using WithOmpDeclarative = semantics::WithOmpDeclarative;
@@ -198,7 +195,8 @@ getMemoryOrderFromRequires(const semantics::Scope &scope) {
}
return static_cast<const common::OmpMemoryOrderType *>(nullptr);
},
- sym->details());
+ symbol->details());
+
if (admo)
return getMemoryOrderKind(*admo);
}
@@ -214,19 +212,83 @@ getDefaultAtomicMemOrder(semantics::SemanticsContext &semaCtx) {
return std::nullopt;
}
-static std::optional<mlir::omp::ClauseMemoryOrderKind>
+static std::pair<std::optional<mlir::omp::ClauseMemoryOrderKind>, bool>
getAtomicMemoryOrder(semantics::SemanticsContext &semaCtx,
const omp::List<omp::Clause> &clauses,
const semantics::Scope &scope) {
for (const omp::Clause &clause : clauses) {
if (auto maybeKind = getMemoryOrderKind(clause.id))
- return *maybeKind;
+ return std::make_pair(*maybeKind, /*canOverride=*/false);
}
if (auto maybeKind = getMemoryOrderFromRequires(scope))
- return *maybeKind;
+ return std::make_pair(*maybeKind, /*canOverride=*/true);
- return getDefaultAtomicMemOrder(semaCtx);
+ return std::make_pair(getDefaultAtomicMemOrder(semaCtx),
+ /*canOverride=*/false);
+}
+
+static std::optional<mlir::omp::ClauseMemoryOrderKind>
+makeValidForAction(std::optional<mlir::omp::ClauseMemoryOrderKind> memOrder,
+ int action0, int action1, unsigned version) {
+ // When the atomic default memory order specified on a REQUIRES directive is
+ // disallowed on a given ATOMIC operation, and it's not ACQ_REL, the order
+ // reverts to RELAXED. ACQ_REL decays to either ACQUIRE or RELEASE, depending
+ // on the operation.
+
+ if (!memOrder) {
+ return memOrder;
+ }
+
+ using Analysis = parser::OpenMPAtomicConstruct::Analysis;
+ // Figure out the main action (i.e. disregard a potential capture operation)
+ int action = action0;
+ if (action1 != Analysis::None)
+ action = action0 == Analysis::Read ? action1 : action0;
+
+ // Avaliable orderings: acquire, acq_rel, relaxed, release, seq_cst
+
+ if (action == Analysis::Read) {
+ // "acq_rel" decays to "acquire"
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Acq_rel)
+ return mlir::omp::ClauseMemoryOrderKind::Acquire;
+ } else if (action == Analysis::Write) {
+ // "acq_rel" decays to "release"
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Acq_rel)
+ return mlir::omp::ClauseMemoryOrderKind::Release;
+ }
+
+ if (version > 50) {
+ if (action == Analysis::Read) {
+ // "release" prohibited
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Release)
+ return mlir::omp::ClauseMemoryOrderKind::Relaxed;
+ }
+ if (action == Analysis::Write) {
+ // "acquire" prohibited
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Acquire)
+ return mlir::omp::ClauseMemoryOrderKind::Relaxed;
+ }
+ } else {
+ if (action == Analysis::Read) {
+ // "release" prohibited
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Release)
+ return mlir::omp::ClauseMemoryOrderKind::Relaxed;
+ } else {
+ if (action & Analysis::Write) { // include "update"
+ // "acquire" prohibited
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Acquire)
+ return mlir::omp::ClauseMemoryOrderKind::Relaxed;
+ if (action == Analysis::Update) {
+ // "acq_rel" prohibited
+ if (*memOrder == mlir::omp::ClauseMemoryOrderKind::Acq_rel)
+ return mlir::omp::ClauseMemoryOrderKind::Relaxed;
+ }
+ }
+ }
+ }
+
+ return memOrder;
}
static mlir::omp::ClauseMemoryOrderKindAttr
@@ -449,16 +511,19 @@ void Fortran::lower::omp::lowerAtomic(
mlir::Value atomAddr =
fir::getBase(converter.genExprAddr(atom, stmtCtx, &loc));
mlir::IntegerAttr hint = getAtomicHint(converter, clauses);
- std::optional<mlir::omp::ClauseMemoryOrderKind> memOrder =
- getAtomicMemoryOrder(semaCtx, clauses,
- semaCtx.FindScope(construct.source));
+ auto [memOrder, canOverride] = getAtomicMemoryOrder(
+ semaCtx, clauses, semaCtx.FindScope(construct.source));
+
+ unsigned version = semaCtx.langOptions().OpenMPVersion;
+ int action0 = analysis.op0.what & analysis.Action;
+ int action1 = analysis.op1.what & analysis.Action;
+ if (canOverride)
+ memOrder = makeValidForAction(memOrder, action0, action1, version);
if (auto *cond = get(analysis.cond)) {
(void)cond;
TODO(loc, "OpenMP ATOMIC COMPARE");
} else {
- int action0 = analysis.op0.what & analysis.Action;
- int action1 = analysis.op1.what & analysis.Action;
mlir::Operation *captureOp = nullptr;
fir::FirOpBuilder::InsertPoint preAt = builder.saveInsertionPoint();
fir::FirOpBuilder::InsertPoint atomicAt, postAt;
diff --git a/flang/lib/Lower/OpenMP/ClauseProcessor.cpp b/flang/lib/Lower/OpenMP/ClauseProcessor.cpp
index 55eda7e..85398be 100644
--- a/flang/lib/Lower/OpenMP/ClauseProcessor.cpp
+++ b/flang/lib/Lower/OpenMP/ClauseProcessor.cpp
@@ -1343,8 +1343,10 @@ bool ClauseProcessor::processMap(
const parser::CharBlock &source) {
using Map = omp::clause::Map;
mlir::Location clauseLocation = converter.genLocation(source);
- const auto &[mapType, typeMods, refMod, mappers, iterator, objects] =
- clause.t;
+ const auto &[mapType, typeMods, attachMod, refMod, mappers, iterator,
+ objects] = clause.t;
+ if (attachMod)
+ TODO(currentLocation, "ATTACH modifier is not implemented yet");
llvm::omp::OpenMPOffloadMappingFlags mapTypeBits =
llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_NONE;
std::string mapperIdName = "__implicit_mapper";
diff --git a/flang/lib/Lower/OpenMP/Clauses.cpp b/flang/lib/Lower/OpenMP/Clauses.cpp
index fac37a3..ba34212 100644
--- a/flang/lib/Lower/OpenMP/Clauses.cpp
+++ b/flang/lib/Lower/OpenMP/Clauses.cpp
@@ -219,7 +219,6 @@ MAKE_EMPTY_CLASS(AcqRel, AcqRel);
MAKE_EMPTY_CLASS(Acquire, Acquire);
MAKE_EMPTY_CLASS(Capture, Capture);
MAKE_EMPTY_CLASS(Compare, Compare);
-MAKE_EMPTY_CLASS(DynamicAllocators, DynamicAllocators);
MAKE_EMPTY_CLASS(Full, Full);
MAKE_EMPTY_CLASS(Inbranch, Inbranch);
MAKE_EMPTY_CLASS(Mergeable, Mergeable);
@@ -235,13 +234,9 @@ MAKE_EMPTY_CLASS(OmpxBare, OmpxBare);
MAKE_EMPTY_CLASS(Read, Read);
MAKE_EMPTY_CLASS(Relaxed, Relaxed);
MAKE_EMPTY_CLASS(Release, Release);
-MAKE_EMPTY_CLASS(ReverseOffload, ReverseOffload);
MAKE_EMPTY_CLASS(SeqCst, SeqCst);
-MAKE_EMPTY_CLASS(SelfMaps, SelfMaps);
MAKE_EMPTY_CLASS(Simd, Simd);
MAKE_EMPTY_CLASS(Threads, Threads);
-MAKE_EMPTY_CLASS(UnifiedAddress, UnifiedAddress);
-MAKE_EMPTY_CLASS(UnifiedSharedMemory, UnifiedSharedMemory);
MAKE_EMPTY_CLASS(Unknown, Unknown);
MAKE_EMPTY_CLASS(Untied, Untied);
MAKE_EMPTY_CLASS(Weak, Weak);
@@ -775,7 +770,18 @@ Doacross make(const parser::OmpClause::Doacross &inp,
return makeDoacross(inp.v.v, semaCtx);
}
-// DynamicAllocators: empty
+DynamicAllocators make(const parser::OmpClause::DynamicAllocators &inp,
+ semantics::SemanticsContext &semaCtx) {
+ // inp.v -> td::optional<arser::OmpDynamicAllocatorsClause>
+ auto &&maybeRequired = maybeApply(
+ [&](const parser::OmpDynamicAllocatorsClause &c) {
+ return makeExpr(c.v, semaCtx);
+ },
+ inp.v);
+
+ return DynamicAllocators{/*Required=*/std::move(maybeRequired)};
+}
+
DynGroupprivate make(const parser::OmpClause::DynGroupprivate &inp,
semantics::SemanticsContext &semaCtx) {
@@ -1069,6 +1075,15 @@ Map make(const parser::OmpClause::Map &inp,
);
CLAUSET_ENUM_CONVERT( //
+ convertAttachMod, parser::OmpAttachModifier::Value, Map::AttachModifier,
+ // clang-format off
+ MS(Always, Always)
+ MS(Auto, Auto)
+ MS(Never, Never)
+ // clang-format on
+ );
+
+ CLAUSET_ENUM_CONVERT( //
convertRefMod, parser::OmpRefModifier::Value, Map::RefModifier,
// clang-format off
MS(Ref_Ptee, RefPtee)
@@ -1115,6 +1130,13 @@ Map make(const parser::OmpClause::Map &inp,
if (!modSet.empty())
maybeTypeMods = Map::MapTypeModifiers(modSet.begin(), modSet.end());
+ auto attachMod = [&]() -> std::optional<Map::AttachModifier> {
+ if (auto *t =
+ semantics::OmpGetUniqueModifier<parser::OmpAttachModifier>(mods))
+ return convertAttachMod(t->v);
+ return std::nullopt;
+ }();
+
auto refMod = [&]() -> std::optional<Map::RefModifier> {
if (auto *t = semantics::OmpGetUniqueModifier<parser::OmpRefModifier>(mods))
return convertRefMod(t->v);
@@ -1135,6 +1157,7 @@ Map make(const parser::OmpClause::Map &inp,
return Map{{/*MapType=*/std::move(type),
/*MapTypeModifiers=*/std::move(maybeTypeMods),
+ /*AttachModifier=*/std::move(attachMod),
/*RefModifier=*/std::move(refMod), /*Mapper=*/std::move(mappers),
/*Iterator=*/std::move(iterator),
/*LocatorList=*/makeObjects(t2, semaCtx)}};
@@ -1321,7 +1344,18 @@ Reduction make(const parser::OmpClause::Reduction &inp,
// Relaxed: empty
// Release: empty
-// ReverseOffload: empty
+
+ReverseOffload make(const parser::OmpClause::ReverseOffload &inp,
+ semantics::SemanticsContext &semaCtx) {
+ // inp.v -> std::optional<parser::OmpReverseOffloadClause>
+ auto &&maybeRequired = maybeApply(
+ [&](const parser::OmpReverseOffloadClause &c) {
+ return makeExpr(c.v, semaCtx);
+ },
+ inp.v);
+
+ return ReverseOffload{/*Required=*/std::move(maybeRequired)};
+}
Safelen make(const parser::OmpClause::Safelen &inp,
semantics::SemanticsContext &semaCtx) {
@@ -1374,6 +1408,18 @@ Schedule make(const parser::OmpClause::Schedule &inp,
// SeqCst: empty
+SelfMaps make(const parser::OmpClause::SelfMaps &inp,
+ semantics::SemanticsContext &semaCtx) {
+ // inp.v -> std::optional<parser::OmpSelfMapsClause>
+ auto &&maybeRequired = maybeApply(
+ [&](const parser::OmpSelfMapsClause &c) {
+ return makeExpr(c.v, semaCtx);
+ },
+ inp.v);
+
+ return SelfMaps{/*Required=*/std::move(maybeRequired)};
+}
+
Severity make(const parser::OmpClause::Severity &inp,
semantics::SemanticsContext &semaCtx) {
// inp -> empty
@@ -1463,8 +1509,29 @@ To make(const parser::OmpClause::To &inp,
/*LocatorList=*/makeObjects(t3, semaCtx)}};
}
-// UnifiedAddress: empty
-// UnifiedSharedMemory: empty
+UnifiedAddress make(const parser::OmpClause::UnifiedAddress &inp,
+ semantics::SemanticsContext &semaCtx) {
+ // inp.v -> std::optional<parser::OmpUnifiedAddressClause>
+ auto &&maybeRequired = maybeApply(
+ [&](const parser::OmpUnifiedAddressClause &c) {
+ return makeExpr(c.v, semaCtx);
+ },
+ inp.v);
+
+ return UnifiedAddress{/*Required=*/std::move(maybeRequired)};
+}
+
+UnifiedSharedMemory make(const parser::OmpClause::UnifiedSharedMemory &inp,
+ semantics::SemanticsContext &semaCtx) {
+ // inp.v -> std::optional<parser::OmpUnifiedSharedMemoryClause>
+ auto &&maybeRequired = maybeApply(
+ [&](const parser::OmpUnifiedSharedMemoryClause &c) {
+ return makeExpr(c.v, semaCtx);
+ },
+ inp.v);
+
+ return UnifiedSharedMemory{/*Required=*/std::move(maybeRequired)};
+}
Uniform make(const parser::OmpClause::Uniform &inp,
semantics::SemanticsContext &semaCtx) {
diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index 444f274..f86ee01 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -4208,18 +4208,17 @@ bool Fortran::lower::markOpenMPDeferredDeclareTargetFunctions(
void Fortran::lower::genOpenMPRequires(mlir::Operation *mod,
const semantics::Symbol *symbol) {
using MlirRequires = mlir::omp::ClauseRequires;
- using SemaRequires = semantics::WithOmpDeclarative::RequiresFlag;
if (auto offloadMod =
llvm::dyn_cast<mlir::omp::OffloadModuleInterface>(mod)) {
- semantics::WithOmpDeclarative::RequiresFlags semaFlags;
+ semantics::WithOmpDeclarative::RequiresClauses reqs;
if (symbol) {
common::visit(
[&](const auto &details) {
if constexpr (std::is_base_of_v<semantics::WithOmpDeclarative,
std::decay_t<decltype(details)>>) {
if (details.has_ompRequires())
- semaFlags = *details.ompRequires();
+ reqs = *details.ompRequires();
}
},
symbol->details());
@@ -4228,14 +4227,14 @@ void Fortran::lower::genOpenMPRequires(mlir::Operation *mod,
// Use pre-populated omp.requires module attribute if it was set, so that
// the "-fopenmp-force-usm" compiler option is honored.
MlirRequires mlirFlags = offloadMod.getRequires();
- if (semaFlags.test(SemaRequires::ReverseOffload))
+ if (reqs.test(llvm::omp::Clause::OMPC_dynamic_allocators))
+ mlirFlags = mlirFlags | MlirRequires::dynamic_allocators;
+ if (reqs.test(llvm::omp::Clause::OMPC_reverse_offload))
mlirFlags = mlirFlags | MlirRequires::reverse_offload;
- if (semaFlags.test(SemaRequires::UnifiedAddress))
+ if (reqs.test(llvm::omp::Clause::OMPC_unified_address))
mlirFlags = mlirFlags | MlirRequires::unified_address;
- if (semaFlags.test(SemaRequires::UnifiedSharedMemory))
+ if (reqs.test(llvm::omp::Clause::OMPC_unified_shared_memory))
mlirFlags = mlirFlags | MlirRequires::unified_shared_memory;
- if (semaFlags.test(SemaRequires::DynamicAllocators))
- mlirFlags = mlirFlags | MlirRequires::dynamic_allocators;
offloadMod.setRequires(mlirFlags);
}
generated by cgit v1.1 at 2025-10-17 14:41:04 +0000