static void handle_ompss_opencl_deallocate_intrinsic(
Nodecl::FunctionCall function_call,
Nodecl::NodeclBase expr_stmt)
{
Nodecl::List arguments = function_call.get_arguments().as<Nodecl::List>();
ERROR_CONDITION(arguments.size() != 1, "More than one argument in ompss_opencl_deallocate call", 0);
Nodecl::NodeclBase actual_argument = arguments[0];
ERROR_CONDITION(!actual_argument.is<Nodecl::FortranActualArgument>(), "Unexpected tree", 0);
Nodecl::NodeclBase arg = actual_argument.as<Nodecl::FortranActualArgument>().get_argument();
TL::Symbol array_sym = ::fortran_data_ref_get_symbol(arg.get_internal_nodecl());
ERROR_CONDITION(
!(array_sym.get_type().is_fortran_array()
&& array_sym.is_allocatable())
&&
!(array_sym.get_type().is_pointer()
&& array_sym.get_type().points_to().is_fortran_array()),
"The argument of 'ompss_opencl_deallocate' intrinsic must be "
"an allocatable array or a pointer to an array\n", 0);
// Replace the current intrinsic call by a call to the Nanos++ API
TL::Symbol ptr_of_arr_sym = get_function_ptr_of(array_sym, expr_stmt.retrieve_context());
TL::Source new_function_call;
new_function_call
<< "CALL NANOS_OPENCL_DEALLOCATE_FORTRAN("
<< ptr_of_arr_sym.get_name() << "("<< as_expression(arg) << "))\n"
;
expr_stmt.replace(new_function_call.parse_statement(expr_stmt));
}
void VectorizerVisitorExpression::visit(const Nodecl::ArraySubscript& n)
{
// Computing new vector type
TL::Type vector_type = n.get_type();
if (vector_type.is_lvalue_reference())
{
vector_type = vector_type.references_to();
}
vector_type = get_qualified_vector_to(vector_type, _vector_length);
TL::Type basic_type = n.get_type();
if (basic_type.is_lvalue_reference())
{
basic_type = basic_type.references_to();
}
// Vector Load
if (Vectorizer::_analysis_info->is_adjacent_access(
Vectorizer::_analysis_scopes->back(),
n))
{
const Nodecl::VectorLoad vector_load =
Nodecl::VectorLoad::make(
Nodecl::Reference::make(
Nodecl::ParenthesizedExpression::make(
n.shallow_copy(),
basic_type,
n.get_locus()),
basic_type.get_pointer_to(),
n.get_locus()),
vector_type,
n.get_locus());
n.replace(vector_load);
}
else // Vector Gather
{
const Nodecl::NodeclBase base = n.get_subscripted();
const Nodecl::List subscripts = n.get_subscripts().as<Nodecl::List>();
ERROR_CONDITION(subscripts.size() > 1,
"Vectorizer: Gather on multidimensional array is not supported yet!", 0);
std::cerr << "Gather: " << n.prettyprint() << "\n";
Nodecl::NodeclBase strides = *subscripts.begin();
walk(strides);
const Nodecl::VectorGather vector_gather =
Nodecl::VectorGather::make(
base.shallow_copy(),
strides,
vector_type,
n.get_locus());
n.replace(vector_gather);
}
}
void ExtensibleGraph::propagate_arguments_in_function_graph( Nodecl::NodeclBase arguments )
{
ERROR_CONDITION( _nodecl.is_null( ), "Found a null nodecl for a graph that is supposed to contain a FunctionCode", 0 );
ERROR_CONDITION( !_nodecl.is<Nodecl::FunctionCode>( ), "Expected FunctionCode but '%s' found",
ast_print_node_type( _nodecl.get_kind( ) ) );
Symbol func_sym( _nodecl.get_symbol( ) );
ERROR_CONDITION( !func_sym.is_valid( ), "Invalid symbol for a nodecl that is supposed to contain a FunctionCode", 0 );
Nodecl::List args = arguments.as<Nodecl::List>( );
ObjectList<Symbol> params = func_sym.get_function_parameters( );
int n_common_params = std::max( args.size( ), params.size( ) );
sym_to_nodecl_map rename_map;
for( int i = 0; i < n_common_params; ++i )
{
rename_map[params[i]] = args[i];
}
RenameVisitor rv( rename_map );
Node* graph_entry = _graph->get_graph_entry_node( );
propagate_argument_rec( graph_entry, &rv );
ExtensibleGraph::clear_visits( graph_entry );
}
void PointerSize::compute_pointer_vars_size_rec(Node* current)
{
if(current->is_visited())
return;
current->set_visited(true);
if(current->is_graph_node())
{
compute_pointer_vars_size_rec(current->get_graph_entry_node());
}
else
{
if(current->has_statements())
{
NBase s;
NBase value;
TL::Type t;
NodeclList stmts = current->get_statements();
for(NodeclList::iterator it = stmts.begin(); it != stmts.end(); ++it)
{
// If assignment (or object init) check whether its for is a dynamic allocation of resources for a pointer type
if(it->is<Nodecl::ObjectInit>() || it->is<Nodecl::Assignment>())
{
// Get the variable assigned and the value used for the assignment
if(it->is<Nodecl::ObjectInit>())
{
Symbol tmp(it->get_symbol());
s = Nodecl::Symbol::make(tmp);
t = tmp.get_type();
s.set_type(t);
value = tmp.get_value().no_conv();
}
else if(it->is<Nodecl::Assignment>())
{
s = it->as<Nodecl::Assignment>().get_lhs().no_conv();
t = s.get_type().no_ref();
if(!s.is<Nodecl::Symbol>() && !s.is<Nodecl::ClassMemberAccess>() && !s.is<Nodecl::ArraySubscript>())
continue;
value = it->as<Nodecl::Assignment>().get_rhs().no_conv();
}
// Check whether this is a pointer and the assignment is a recognized memory operation
if(t.is_pointer() && !value.is_null()) // This can be null if uninitialized ObjectInit
{
if(value.is<Nodecl::FunctionCall>())
{
Symbol called_sym = value.as<Nodecl::FunctionCall>().get_called().get_symbol();
Type return_t = called_sym.get_type().returns();
Nodecl::List args = value.as<Nodecl::FunctionCall>().get_arguments().as<Nodecl::List>();
std::string sym_name = called_sym.get_name();
NBase size = NBase::null();
if((sym_name == "malloc") && (args.size() == 1))
{ // void* malloc (size_t size);
Type arg0_t = args[0].get_type();
if(return_t.is_pointer() && return_t.points_to().is_void() && arg0_t.is_same_type(get_size_t_type()))
{ // We recognize the form 'sizeof(base_type) * n_elemes' and 'n_elemes * sizeof(base_type)'
if(args[0].is<Nodecl::Mul>())
{
NBase lhs = args[0].as<Nodecl::Mul>().get_lhs().no_conv();
NBase rhs = args[0].as<Nodecl::Mul>().get_rhs().no_conv();
if(lhs.is<Nodecl::Sizeof>() && (rhs.is<Nodecl::IntegerLiteral>() || rhs.is<Nodecl::Symbol>()))
size = rhs;
else if(rhs.is<Nodecl::Sizeof>() && (lhs.is<Nodecl::IntegerLiteral>() || lhs.is<Nodecl::Symbol>()))
size = lhs;
}
}
}
else if((sym_name == "calloc") && (args.size() == 2))
{ // void* calloc (size_t num, size_t size);
Type arg0_t = args[0].get_type();
Type arg1_t = args[1].get_type();
if(return_t.is_pointer() && return_t.points_to().is_void()
&& arg0_t.is_same_type(get_size_t_type())
&& arg1_t.is_same_type(get_size_t_type()))
{
size = args[0];
}
}
if(!size.is_null())
_pcfg->set_pointer_n_elems(s, size);
}
}
// Clear up the common variables s, value and t
s = NBase::null();
value = NBase::null();
t = Type();
}
}
}
}
// Keep iterating over the children
ObjectList<Node*> children = current->get_children();
for(ObjectList<Node*>::iterator it = children.begin(); it != children.end(); ++it)
compute_pointer_vars_size_rec(*it);
}
int SuitableAlignmentVisitor::visit( const Nodecl::ArraySubscript& n )
{
if( _nesting_level == 0 ) // Target access
{
_nesting_level++;
int i;
int alignment = 0;
Nodecl::NodeclBase subscripted = n.get_subscripted( );
TL::Type element_type = subscripted.get_type( );
// TODO: subscript is aligned
Nodecl::List subscripts = n.get_subscripts( ).as<Nodecl::List>( );
int num_subscripts = subscripts.size( );
// Get dimension sizes
int *dimension_sizes = (int *)malloc( ( num_subscripts-1 ) * sizeof( int ) );
for( i = 0; i < (num_subscripts-1); i++ ) // Skip the first one. It does not have size
{
// Iterate on array subscript type
if( element_type.is_array( ) )
{
element_type = element_type.array_element( );
}
else if( element_type.is_pointer( ) )
{
element_type = element_type.points_to( );
}
else
{
WARNING_MESSAGE( "Array subscript does not have array type or pointer to array type", 0 );
return -1;
}
if( !element_type.array_has_size( ) )
{
WARNING_MESSAGE( "Array type does not have size", 0 );
return -1;
}
// Compute dimension alignment
Nodecl::NodeclBase dimension_size_node = element_type.array_get_size( );
// If VLA, get the actual size
if(dimension_size_node.is<Nodecl::Symbol>() &&
dimension_size_node.get_symbol().is_saved_expression())
{
dimension_size_node = dimension_size_node.get_symbol().get_value();
}
int dimension_size = -1;
if( dimension_size_node.is_constant( ) )
{
dimension_size = const_value_cast_to_signed_int( dimension_size_node.get_constant( ) );
if( is_suitable_constant( dimension_size * _type_size ) )
dimension_size = 0;
}
// If dimension size is suitable
else if( is_suitable_expression( dimension_size_node ) )
{
dimension_size = 0;
}
if( VERBOSE )
printf( "Dim %d, size %d\n", i, dimension_size );
dimension_sizes[i] = dimension_size;
}
int it_alignment;
Nodecl::List::iterator it = subscripts.begin( );
// Multiply dimension sizes by indexes
for( i=0; it != subscripts.end( ); i++ )
{
it_alignment = walk( *it );
it++;
if( it == subscripts.end( ) ) break; // Last dimmension does not have to be multiplied
// a[i][j][k] -> i -> i*J*K
for( int j = i; j < (num_subscripts-1); j++ )
{
if( ( dimension_sizes[j] == 0 ) || ( it_alignment == 0 ) )
{
it_alignment = 0;
}
else if( ( dimension_sizes[j] < 0 ) || ( it_alignment < 0 ) )
{
it_alignment = -1;
}
else
{
it_alignment *= dimension_sizes[j];
}
}
if( it_alignment < 0 )
{
return -1;
}
alignment += it_alignment;
}
// Add adjacent dimension
alignment += it_alignment;
free(dimension_sizes);
_nesting_level--;
return alignment;
}
// Nested array subscript
else
{
if (is_suitable_expression(n))
{
return 0;
}
return -1;
}
}
static void handle_ompss_opencl_allocate_intrinsic(
Nodecl::FunctionCall function_call,
std::map<std::pair<TL::Type, std::pair<int, bool> > , Symbol> &declared_ocl_allocate_functions,
Nodecl::NodeclBase expr_stmt)
{
Nodecl::List arguments = function_call.get_arguments().as<Nodecl::List>();
ERROR_CONDITION(arguments.size() != 1, "More than one argument in 'ompss_opencl_allocate' call\n", 0);
Nodecl::NodeclBase actual_argument = arguments[0];
ERROR_CONDITION(!actual_argument.is<Nodecl::FortranActualArgument>(), "Unexpected tree\n", 0);
Nodecl::NodeclBase arg = actual_argument.as<Nodecl::FortranActualArgument>().get_argument();
ERROR_CONDITION(!arg.is<Nodecl::ArraySubscript>(), "Unreachable code\n", 0);
Nodecl::NodeclBase subscripted = arg.as<Nodecl::ArraySubscript>().get_subscripted();
TL::Symbol subscripted_symbol = ::fortran_data_ref_get_symbol(subscripted.get_internal_nodecl());
ERROR_CONDITION(
!(subscripted_symbol.get_type().is_fortran_array()
&& subscripted_symbol.is_allocatable())
&&
!(subscripted_symbol.get_type().is_pointer()
&& subscripted_symbol.get_type().points_to().is_fortran_array()),
"The argument of 'ompss_opencl_allocate' intrinsic must be "
"an allocatable array or a pointer to an array with all its bounds specified\n", 0);
TL::Type array_type;
int num_dimensions;
bool is_allocatable;
if (subscripted_symbol.is_allocatable())
{
array_type = subscripted_symbol.get_type();
num_dimensions = subscripted_symbol.get_type().get_num_dimensions();
is_allocatable = true;
}
else
{
array_type = subscripted_symbol.get_type().points_to();
num_dimensions = array_type.get_num_dimensions();
is_allocatable = false;
}
TL::Type element_type = array_type;
while (element_type.is_array())
{
element_type = element_type.array_element();
}
ERROR_CONDITION(!array_type.is_array(), "This type should be an array type", 0);
std::pair<TL::Type, std::pair<int, bool> > key =
std::make_pair(element_type, std::make_pair(num_dimensions, is_allocatable));
std::map<std::pair<TL::Type, std::pair<int, bool> > , Symbol>::iterator it_new_fun =
declared_ocl_allocate_functions.find(key);
// Reuse the auxiliar function if it already exists
Symbol new_function_sym;
if (it_new_fun != declared_ocl_allocate_functions.end())
{
new_function_sym = it_new_fun->second;
}
else
{
new_function_sym = create_new_function_opencl_allocate(
expr_stmt, subscripted_symbol, element_type, num_dimensions, is_allocatable);
declared_ocl_allocate_functions[key] = new_function_sym;
}
// Replace the current intrinsic call by a call to the new function
TL::Source actual_arg_array;
Nodecl::NodeclBase subscripted_lvalue = subscripted.shallow_copy();
subscripted_lvalue.set_type(subscripted_symbol.get_type().no_ref().get_lvalue_reference_to());
actual_arg_array << as_expression(subscripted_lvalue);
TL::Source actual_arg_bounds;
Nodecl::List subscripts = arg.as<Nodecl::ArraySubscript>().get_subscripts().as<Nodecl::List>();
for (Nodecl::List::reverse_iterator it = subscripts.rbegin();
it != subscripts.rend();
it++)
{
Nodecl::NodeclBase subscript = *it, lower, upper;
if (it != subscripts.rbegin())
actual_arg_bounds << ", ";
if (subscript.is<Nodecl::Range>())
{
lower = subscript.as<Nodecl::Range>().get_lower();
upper = subscript.as<Nodecl::Range>().get_upper();
}
else
{
lower = nodecl_make_integer_literal(
fortran_get_default_integer_type(),
const_value_get_signed_int(1), make_locus("", 0, 0));
upper = subscript;
}
actual_arg_bounds << as_expression(lower) << "," << as_expression(upper);
}
TL::Source new_function_call;
new_function_call
<< "CALL " << as_symbol(new_function_sym) << "("
<< actual_arg_array << ", "
<< actual_arg_bounds << ")\n"
;
expr_stmt.replace(new_function_call.parse_statement(expr_stmt));
}
void VectorizerVisitorExpression::visit(const Nodecl::Assignment& n)
{
Nodecl::NodeclBase lhs = n.get_lhs();
walk(n.get_rhs());
// Computing new vector type
TL::Type vector_type = n.get_type();
/*
if (vector_type.is_lvalue_reference())
{
vector_type = vector_type.references_to();
}
*/
vector_type = get_qualified_vector_to(vector_type, _vector_length);
if(lhs.is<Nodecl::ArraySubscript>())
{
// Vector Store
if(Vectorizer::_analysis_info->is_adjacent_access(
Vectorizer::_analysis_scopes->back(),
lhs))
{
TL::Type basic_type = lhs.get_type();
if (basic_type.is_lvalue_reference())
{
basic_type = basic_type.references_to();
}
const Nodecl::VectorStore vector_store =
Nodecl::VectorStore::make(
Nodecl::Reference::make(
Nodecl::ParenthesizedExpression::make(
lhs.shallow_copy(),
basic_type,
n.get_locus()),
basic_type.get_pointer_to(),
n.get_locus()),
n.get_rhs().shallow_copy(),
vector_type,
n.get_locus());
n.replace(vector_store);
}
else // Vector Scatter
{
const Nodecl::ArraySubscript lhs_array = lhs.as<Nodecl::ArraySubscript>();
const Nodecl::NodeclBase base = lhs_array.get_subscripted();
const Nodecl::List subscripts = lhs_array.get_subscripts().as<Nodecl::List>();
std::cerr << "Scatter: " << lhs_array.prettyprint() << "\n";
ERROR_CONDITION(subscripts.size() > 1,
"Vectorizer: Scatter on multidimensional array is not supported yet!", 0);
Nodecl::NodeclBase strides = *subscripts.begin();
walk(strides);
const Nodecl::VectorScatter vector_scatter =
Nodecl::VectorScatter::make(
base.shallow_copy(),
strides,
n.get_rhs().shallow_copy(),
vector_type,
n.get_locus());
n.replace(vector_scatter);
}
}
else // Register
{
walk(lhs);
const Nodecl::VectorAssignment vector_assignment =
Nodecl::VectorAssignment::make(
lhs.shallow_copy(),
n.get_rhs().shallow_copy(),
vector_type,
n.get_locus());
n.replace(vector_assignment);
}
}