[flang][CUDA] Apply intrinsic operator overrides (#151018)

Fortran's intrinsic numeric and relational operators can be overridden
with explicit interfaces so long as one or more of the dummy arguments
have the DEVICE attribute. Semantics already allows this without
complaint, but fails to replace the operations with the defined specific
procedure calls when analyzing expressions.
This commit is contained in:
Peter Klausler 2025-07-30 11:41:40 -07:00 committed by GitHub
parent 0d6a67c1ad
commit b01ab5318e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
6 changed files with 126 additions and 16 deletions

View File

@ -162,7 +162,6 @@ public:
warningsAreErrors_ = x;
return *this;
}
SemanticsContext &set_debugModuleWriter(bool x) {
debugModuleWriter_ = x;
return *this;

View File

@ -761,14 +761,13 @@ void CUDAChecker::Enter(const parser::AssignmentStmt &x) {
// legal.
if (nbLhs == 0 && nbRhs > 1) {
context_.Say(lhsLoc,
"More than one reference to a CUDA object on the right hand side of the assigment"_err_en_US);
"More than one reference to a CUDA object on the right hand side of the assignment"_err_en_US);
}
if (Fortran::evaluate::HasCUDADeviceAttrs(assign->lhs) &&
Fortran::evaluate::HasCUDAImplicitTransfer(assign->rhs)) {
if (evaluate::HasCUDADeviceAttrs(assign->lhs) &&
evaluate::HasCUDAImplicitTransfer(assign->rhs)) {
if (GetNbOfCUDAManagedOrUnifiedSymbols(assign->lhs) == 1 &&
GetNbOfCUDAManagedOrUnifiedSymbols(assign->rhs) == 1 &&
GetNbOfCUDADeviceSymbols(assign->rhs) == 1) {
GetNbOfCUDAManagedOrUnifiedSymbols(assign->rhs) == 1 && nbRhs == 1) {
return; // This is a special case handled on the host.
}
context_.Say(lhsLoc, "Unsupported CUDA data transfer"_err_en_US);

View File

@ -2081,7 +2081,7 @@ static bool ConflictsWithIntrinsicAssignment(const Procedure &proc) {
}
static bool ConflictsWithIntrinsicOperator(
const GenericKind &kind, const Procedure &proc) {
const GenericKind &kind, const Procedure &proc, SemanticsContext &context) {
if (!kind.IsIntrinsicOperator()) {
return false;
}
@ -2167,7 +2167,7 @@ bool CheckHelper::CheckDefinedOperator(SourceName opName, GenericKind kind,
}
} else if (!checkDefinedOperatorArgs(opName, specific, proc)) {
return false; // error was reported
} else if (ConflictsWithIntrinsicOperator(kind, proc)) {
} else if (ConflictsWithIntrinsicOperator(kind, proc, context_)) {
msg = "%s function '%s' conflicts with intrinsic operator"_err_en_US;
}
if (msg) {

View File

@ -165,10 +165,17 @@ public:
bool CheckForNullPointer(const char *where = "as an operand here");
bool CheckForAssumedRank(const char *where = "as an operand here");
bool AnyCUDADeviceData() const;
// Returns true if an interface has been defined for an intrinsic operator
// with one or more device operands.
bool HasDeviceDefinedIntrinsicOpOverride(const char *) const;
template <typename E> bool HasDeviceDefinedIntrinsicOpOverride(E opr) const {
return HasDeviceDefinedIntrinsicOpOverride(
context_.context().languageFeatures().GetNames(opr));
}
// Find and return a user-defined operator or report an error.
// The provided message is used if there is no such operator.
// If a definedOpSymbolPtr is provided, the caller must check
// for its accessibility.
MaybeExpr TryDefinedOp(
const char *, parser::MessageFixedText, bool isUserOp = false);
template <typename E>
@ -183,6 +190,8 @@ public:
void Dump(llvm::raw_ostream &);
private:
bool HasDeviceDefinedIntrinsicOpOverride(
const std::vector<const char *> &) const;
MaybeExpr TryDefinedOp(
const std::vector<const char *> &, parser::MessageFixedText);
MaybeExpr TryBoundOp(const Symbol &, int passIndex);
@ -202,7 +211,7 @@ private:
void SayNoMatch(
const std::string &, bool isAssignment = false, bool isAmbiguous = false);
std::string TypeAsFortran(std::size_t);
bool AnyUntypedOrMissingOperand();
bool AnyUntypedOrMissingOperand() const;
ExpressionAnalyzer &context_;
ActualArguments actuals_;
@ -4497,13 +4506,20 @@ void ArgumentAnalyzer::Analyze(
bool ArgumentAnalyzer::IsIntrinsicRelational(RelationalOperator opr,
const DynamicType &leftType, const DynamicType &rightType) const {
CHECK(actuals_.size() == 2);
return semantics::IsIntrinsicRelational(
opr, leftType, GetRank(0), rightType, GetRank(1));
return !(context_.context().languageFeatures().IsEnabled(
common::LanguageFeature::CUDA) &&
HasDeviceDefinedIntrinsicOpOverride(opr)) &&
semantics::IsIntrinsicRelational(
opr, leftType, GetRank(0), rightType, GetRank(1));
}
bool ArgumentAnalyzer::IsIntrinsicNumeric(NumericOperator opr) const {
std::optional<DynamicType> leftType{GetType(0)};
if (actuals_.size() == 1) {
if (context_.context().languageFeatures().IsEnabled(
common::LanguageFeature::CUDA) &&
HasDeviceDefinedIntrinsicOpOverride(AsFortran(opr))) {
return false;
} else if (actuals_.size() == 1) {
if (IsBOZLiteral(0)) {
return opr == NumericOperator::Add; // unary '+'
} else {
@ -4617,6 +4633,53 @@ bool ArgumentAnalyzer::CheckForAssumedRank(const char *where) {
return true;
}
bool ArgumentAnalyzer::AnyCUDADeviceData() const {
for (const std::optional<ActualArgument> &arg : actuals_) {
if (arg) {
if (const Expr<SomeType> *expr{arg->UnwrapExpr()}) {
if (HasCUDADeviceAttrs(*expr)) {
return true;
}
}
}
}
return false;
}
// Some operations can be defined with explicit non-type-bound interfaces
// that would erroneously conflict with intrinsic operations in their
// types and ranks but have one or more dummy arguments with the DEVICE
// attribute.
bool ArgumentAnalyzer::HasDeviceDefinedIntrinsicOpOverride(
const char *opr) const {
if (AnyCUDADeviceData() && !AnyUntypedOrMissingOperand()) {
std::string oprNameString{"operator("s + opr + ')'};
parser::CharBlock oprName{oprNameString};
parser::Messages buffer;
auto restorer{context_.GetContextualMessages().SetMessages(buffer)};
const auto &scope{context_.context().FindScope(source_)};
if (Symbol * generic{scope.FindSymbol(oprName)}) {
parser::Name name{generic->name(), generic};
const Symbol *resultSymbol{nullptr};
if (context_.AnalyzeDefinedOp(
name, ActualArguments{actuals_}, resultSymbol)) {
return true;
}
}
}
return false;
}
bool ArgumentAnalyzer::HasDeviceDefinedIntrinsicOpOverride(
const std::vector<const char *> &oprNames) const {
for (const char *opr : oprNames) {
if (HasDeviceDefinedIntrinsicOpOverride(opr)) {
return true;
}
}
return false;
}
MaybeExpr ArgumentAnalyzer::TryDefinedOp(
const char *opr, parser::MessageFixedText error, bool isUserOp) {
if (AnyUntypedOrMissingOperand()) {
@ -5135,7 +5198,7 @@ std::string ArgumentAnalyzer::TypeAsFortran(std::size_t i) {
}
}
bool ArgumentAnalyzer::AnyUntypedOrMissingOperand() {
bool ArgumentAnalyzer::AnyUntypedOrMissingOperand() const {
for (const auto &actual : actuals_) {
if (!actual ||
(!actual->GetType() && !IsBareNullPointer(actual->UnwrapExpr()))) {

View File

@ -0,0 +1,49 @@
! RUN: %flang_fc1 -fdebug-unparse %s 2>&1 | FileCheck %s
module overrides
type realResult
real a
end type
interface operator(*)
procedure :: multHostDevice, multDeviceHost
end interface
interface assignment(=)
procedure :: assignHostResult, assignDeviceResult
end interface
contains
elemental function multHostDevice(x, y) result(result)
real, intent(in) :: x
real, intent(in), device :: y
type(realResult) result
result%a = x * y
end
elemental function multDeviceHost(x, y) result(result)
real, intent(in), device :: x
real, intent(in) :: y
type(realResult) result
result%a = x * y
end
elemental subroutine assignHostResult(lhs, rhs)
real, intent(out) :: lhs
type(realResult), intent(in) :: rhs
lhs = rhs%a
end
elemental subroutine assignDeviceResult(lhs, rhs)
real, intent(out), device :: lhs
type(realResult), intent(in) :: rhs
lhs = rhs%a
end
end
program p
use overrides
real, device :: da, db
real :: ha, hb
!CHECK: CALL assigndeviceresult(db,multhostdevice(2._4,da))
db = 2. * da
!CHECK: CALL assigndeviceresult(db,multdevicehost(da,2._4))
db = da * 2.
!CHECK: CALL assignhostresult(ha,multhostdevice(2._4,da))
ha = 2. * da
!CHECK: CALL assignhostresult(ha,multdevicehost(da,2._4))
ha = da * 2.
end

View File

@ -16,7 +16,7 @@ subroutine sub1()
real, device :: adev(10), bdev(10)
real :: ahost(10)
!ERROR: More than one reference to a CUDA object on the right hand side of the assigment
!ERROR: More than one reference to a CUDA object on the right hand side of the assignment
ahost = adev + bdev
ahost = adev + adev