void NeonVectorBackend::visit(const Nodecl::VectorAdd& n)
{
walk(n.get_lhs());
walk(n.get_rhs());
TL::Type t = n.get_type();
ERROR_CONDITION(!t.is_vector(), "Invalid type", 0);
TL::Type element = t.vector_element();
ERROR_CONDITION(!element.is_float(), "Not implemented: %s", print_declarator(element.get_internal_type()));
TL::Symbol builtin_fun = TL::Scope::get_global_scope().get_symbol_from_name("vaddq_f32");
ERROR_CONDITION(!builtin_fun.is_valid(), "Symbol not found", 0);
n.replace(
Nodecl::FunctionCall::make(
builtin_fun.make_nodecl(/* set_ref_type */ true),
Nodecl::List::make(
n.get_lhs(),
n.get_rhs()),
/* alternate-name */ Nodecl::NodeclBase::null(),
/* function-form */ Nodecl::NodeclBase::null(),
n.get_type(),
n.get_locus()
)
);
}
Nodecl::Symbol Symbol::make_nodecl(bool set_ref_type, const locus_t* locus) const
{
Nodecl::Symbol sym = Nodecl::Symbol::make(*this, locus);
if (set_ref_type)
{
TL::Type t = this->get_type();
if (!t.is_any_reference())
t = t.get_lvalue_reference_to();
sym.set_type(t);
}
else
{
sym.set_type(this->get_type());
}
// Set constant (currently only for variables)
if (this->is_variable()
&& this->get_type().is_const()
&& !this->is_parameter() // avoid 'void f(const int n = 3)'
&& !this->get_value().is_null()
&& this->get_value().is_constant())
{
sym.set_constant(this->get_value().get_constant());
}
return sym;
}
void KNCModuleVisitor::visit(const Nodecl::VectorBitwiseXor& node)
{
TL::Type type = node.get_type().basic_type();
// Intrinsic name
file << "_mm512_xor";
// Postfix
if (type.is_float())
{
file << "_ps";
}
else if (type.is_integral_type())
{
file << "_si128";
}
else
{
running_error("KNC Codegen: Node %s at %s has an unsupported type.",
ast_print_node_type(node.get_kind()),
node.get_locus_str().c_str());
}
file << "(";
walk(node.get_lhs());
file << ", ";
walk(node.get_rhs());
file << ")";
}
TL::Type get_array_of_vector(TL::Type t)
{
ERROR_CONDITION(!t.is_vector(), "Invalid type", 0);
return t.vector_element().get_array_to(
const_value_to_nodecl(
const_value_get_signed_int(t.vector_num_elements())
),
TL::Scope::get_global_scope());
}
Nodecl::Symbol Symbol::make_nodecl(bool set_ref_type, const locus_t* locus) const
{
Nodecl::Symbol sym = Nodecl::Symbol::make(*this, locus);
if (set_ref_type)
{
TL::Type t = this->get_type();
if (!t.is_any_reference())
t = t.get_lvalue_reference_to();
sym.set_type(t);
}
return sym;
}
TL::Type get_array_of_mask(TL::Type t)
{
ERROR_CONDITION(!t.is_mask(), "Invalid type", 0);
int n = (t.get_mask_num_elements() / 8)
+ !!(t.get_mask_num_elements() % 8);
return TL::Type::get_unsigned_char_type().get_array_to(
const_value_to_nodecl(
const_value_get_signed_int(n)
),
TL::Scope::get_global_scope());
}
void LoweringVisitor::create_reduction_function(OpenMP::Reduction* red,
Nodecl::NodeclBase construct,
TL::Type reduction_type,
TL::Symbol& basic_reduction_function,
TL::Symbol& vector_reduction_function)
{
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
basic_reduction_function = create_basic_reduction_function_c(red, construct);
// This is not yet well supported in Nanos++
if (!reduction_type.is_array())
{
vector_reduction_function = create_vector_reduction_function_c(red, construct);
}
}
else if (IS_FORTRAN_LANGUAGE)
{
basic_reduction_function = create_basic_reduction_function_fortran(red, construct);
}
else
{
internal_error("Code unreachable", 0);
}
}
void KNCModuleVisitor::visit(const Nodecl::VectorMinus& node)
{
TL::Type type = node.get_type().basic_type();
// Intrinsic name
file << "_mm512_sub";
// Postfix
if (type.is_float())
{
file << "_ps";
}
else if (type.is_double())
{
file << "_pd";
}
else if (type.is_signed_int() ||
type.is_unsigned_int())
{
file << "_epi32";
}
else if (type.is_signed_short_int() ||
type.is_unsigned_short_int())
{
file << "_epi16";
}
else if (type.is_char() ||
type.is_signed_char() ||
type.is_unsigned_char())
{
file << "_epi8";
}
else
{
running_error("KNC Codegen: Node %s at %s has an unsupported type.",
ast_print_node_type(node.get_kind()),
node.get_locus_str().c_str());
}
file << "(";
walk(node.get_lhs());
file << ", ";
walk(node.get_rhs());
file << ")";
}
bool VectorizerVectorReduction::is_supported_reduction(bool is_builtin,
const std::string& reduction_name,
const TL::Type& reduction_type)
{
if(is_builtin)
{
if(_environment._device.compare("smp") == 0)
{
if(reduction_name.compare("+") == 0)
{
if(reduction_type.is_signed_int())
{
return true;
}
}
}
else if(_environment._device.compare("knc") == 0)
{
if((reduction_name.compare("+") == 0) ||
(reduction_name.compare("-") == 0))
{
if(reduction_type.is_signed_int())
{
return true;
}
else if(reduction_type.is_float())
{
return true;
}
else if (reduction_type.is_double())
{
return true;
}
}
}
}
return false;
}
static TL::Symbol create_initializer_function_c(
OpenMP::Reduction* red,
TL::Type reduction_type,
Nodecl::NodeclBase construct)
{
std::string fun_name;
{
std::stringstream ss;
ss << "nanos_ini_" << red << "_" << reduction_type.get_internal_type() << "_" << simple_hash_str(construct.get_filename().c_str());
fun_name = ss.str();
}
Nodecl::NodeclBase function_body;
Source src;
src << "void " << fun_name << "("
<< as_type(reduction_type.no_ref().get_lvalue_reference_to()) << " omp_priv,"
<< as_type(reduction_type.no_ref().get_lvalue_reference_to()) << " omp_orig)"
<< "{"
<< statement_placeholder(function_body)
<< "}"
;
Nodecl::NodeclBase function_code = src.parse_global(construct.retrieve_context().get_global_scope());
TL::Scope inside_function = ReferenceScope(function_body).get_scope();
TL::Symbol param_omp_priv = inside_function.get_symbol_from_name("omp_priv");
ERROR_CONDITION(!param_omp_priv.is_valid(), "Symbol omp_priv not found", 0);
TL::Symbol param_omp_orig = inside_function.get_symbol_from_name("omp_orig");
ERROR_CONDITION(!param_omp_orig.is_valid(), "Symbol omp_orig not found", 0);
TL::Symbol function_sym = inside_function.get_symbol_from_name(fun_name);
ERROR_CONDITION(!function_sym.is_valid(), "Symbol %s not found", fun_name.c_str());
Nodecl::NodeclBase initializer = red->get_initializer().shallow_copy();
if (initializer.is<Nodecl::StructuredValue>())
{
Nodecl::StructuredValue structured_value = initializer.as<Nodecl::StructuredValue>();
if (structured_value.get_form().is<Nodecl::StructuredValueBracedImplicit>())
{
structured_value.set_form(Nodecl::StructuredValueCompoundLiteral::make());
}
}
Nodecl::Utils::SimpleSymbolMap translation_map;
translation_map.add_map(red->get_omp_priv(), param_omp_priv);
translation_map.add_map(red->get_omp_orig(), param_omp_orig);
Nodecl::NodeclBase new_initializer = Nodecl::Utils::deep_copy(initializer, inside_function, translation_map);
if (red->get_is_initialization())
{
// The original initializer was something like 'omp_priv = expr1', but the
// new_initializer only represents the lhs expression (in our example, expr1).
// For this reason we create manually an assignment expression.
Nodecl::NodeclBase param_omp_priv_ref = Nodecl::Symbol::make(param_omp_priv);
param_omp_priv_ref.set_type(param_omp_priv.get_type());
function_body.replace(
Nodecl::List::make(
Nodecl::ExpressionStatement::make(
Nodecl::Assignment::make(
param_omp_priv_ref,
new_initializer,
param_omp_priv_ref.get_type().no_ref())
)));
}
else
{
function_body.replace(
Nodecl::List::make(Nodecl::ExpressionStatement::make(new_initializer)));
}
// As the initializer function is needed during the instantiation of
// the task, this function should be inserted before the construct
Nodecl::Utils::prepend_to_enclosing_top_level_location(construct,
function_code);
return function_sym;
}
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);
}
std::string as_type(TL::Type t)
{
return type_to_source(t.get_internal_type());
}
bool casting_needs_reinterpr_or_pack(TL::Type& casted_type, TL::Type& cast_type)
{
if (casted_type.is_float())
{
if (cast_type.is_float())
return false;
else
return true;
}
else if (casted_type.is_double())
{
if (cast_type.is_double())
return false;
else
return true;
}
else if (casted_type.is_signed_int()
|| casted_type.is_unsigned_int())
{
if (cast_type.is_signed_int()
|| cast_type.is_unsigned_int())
return false;
else
return true;
}
else if (casted_type.is_signed_short_int()
|| casted_type.is_unsigned_short_int())
{
if (cast_type.is_signed_short_int()
|| cast_type.is_unsigned_short_int())
return false;
else
return true;
}
else if (casted_type.is_signed_char()
|| casted_type.is_unsigned_char()
|| casted_type.is_char())
{
if (cast_type.is_signed_char()
|| cast_type.is_unsigned_char()
|| cast_type.is_char())
return false;
else
return true;
}
else
{
fatal_error("error: casting_needs_reinterpr_or_pack does not support this casting in HLT SIMD.\n");
}
}
TL::Source LoopBlocking::do_blocking()
{
Source result, block_loops;
result
<< block_loops
;
ObjectList<ForStatement> nest_loops = _for_nest_info.get_nest_list();
_nesting = std::min(_nest_factors.size(), nest_loops.size());
TL::Source *current_innermost_part = &block_loops;
// For every loop declare its block loop variable and the inter-block loop
ObjectList<TL::Expression>::iterator current_factor = _nest_factors.begin();
ObjectList<TL::ForStatement>::iterator current_for = nest_loops.begin();
for (int current_nest = 0;
current_nest < _nesting;
current_nest++, current_for++, current_factor++)
{
TL::IdExpression induction_var = current_for->get_induction_variable();
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
std::string var = "_blk_" + sym.get_name();
TL::Source *new_innermost_part = new TL::Source();
(*current_innermost_part)
<< "for(" << type.get_declaration(sym.get_scope(), var) << " = " << current_for->get_lower_bound() << ";"
<< var << current_for->get_bound_operator() << current_for->get_upper_bound() << ";"
<< var << "+= ( " << current_for->get_step() << ") * " << current_factor->prettyprint() << ")"
<< (*new_innermost_part)
;
current_innermost_part = new_innermost_part;
}
// Now for every loop, declare the intra-loop
current_factor = _nest_factors.begin();
current_for = nest_loops.begin();
for (int current_nest = 0;
current_nest < _nesting;
current_nest++, current_for++, current_factor++)
{
TL::IdExpression induction_var = current_for->get_induction_variable();
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
std::string var = induction_var.prettyprint();
std::string init_var = var;
// If the loop declares the iterator in the for statement
// declare it again
AST_t loop_init = current_for->get_iterating_init();
if (Declaration::predicate(loop_init))
{
// Fix init_var to be a declaration
init_var = type.get_declaration(sym.get_scope(), var);
}
std::string blk_var = "_blk_" + sym.get_name();
TL::Source min_code;
TL::Source *new_innermost_part = new TL::Source();
(*current_innermost_part)
<< "for(" << init_var << " = " << blk_var << ";"
<< var << current_for->get_bound_operator() << min_code << ";"
<< var << "+= ( " << current_for->get_step() << "))"
<< (*new_innermost_part)
;
TL::Source a, b;
min_code
<< "((" << a << ") < (" << b << ") ? (" << a << ") : (" << b << "))"
;
a << blk_var << " + (" << current_for->get_step() << ") * (" << current_factor->prettyprint() << " - 1 )";
b << current_for->get_upper_bound();
current_innermost_part = new_innermost_part;
}
// And now the innermost loop
(*current_innermost_part)
<< nest_loops[_nesting - 1].get_loop_body()
;
return result;
}
void LoweringVisitor::reduction_initialization_code(
OutlineInfo& outline_info,
Nodecl::NodeclBase ref_tree,
Nodecl::NodeclBase construct)
{
ERROR_CONDITION(ref_tree.is_null(), "Invalid tree", 0);
if (!Nanos::Version::interface_is_at_least("master", 5023))
{
running_error("%s: error: a newer version of Nanos++ (>=5023) is required for reductions support\n",
construct.get_locus_str().c_str());
}
TL::ObjectList<OutlineDataItem*> reduction_items = outline_info.get_data_items().filter(
predicate(lift_pointer(functor(&OutlineDataItem::is_reduction))));
ERROR_CONDITION (reduction_items.empty(), "No reductions to process", 0);
Source result;
Source reduction_declaration,
thread_initializing_reduction_info,
thread_fetching_reduction_info;
result
<< reduction_declaration
<< "{"
<< as_type(get_bool_type()) << " red_single_guard;"
<< "nanos_err_t err;"
<< "err = nanos_enter_sync_init(&red_single_guard);"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< "if (red_single_guard)"
<< "{"
<< "int nanos_num_threads = nanos_omp_get_num_threads();"
<< thread_initializing_reduction_info
<< "err = nanos_release_sync_init();"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< "}"
<< "else"
<< "{"
<< "err = nanos_wait_sync_init();"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< thread_fetching_reduction_info
<< "}"
<< "}"
;
for (TL::ObjectList<OutlineDataItem*>::iterator it = reduction_items.begin();
it != reduction_items.end();
it++)
{
std::string nanos_red_name = "nanos_red_" + (*it)->get_symbol().get_name();
std::pair<OpenMP::Reduction*, TL::Type> reduction_info = (*it)->get_reduction_info();
OpenMP::Reduction* reduction = reduction_info.first;
TL::Type reduction_type = reduction_info.second;
if (reduction_type.is_any_reference())
reduction_type = reduction_type.references_to();
TL::Type reduction_element_type = reduction_type;
if (IS_FORTRAN_LANGUAGE)
{
while (reduction_element_type.is_fortran_array())
reduction_element_type = reduction_element_type.array_element();
}
else
{
while (reduction_element_type.is_array())
reduction_element_type = reduction_element_type.array_element();
}
Source element_size;
if (IS_FORTRAN_LANGUAGE)
{
if (reduction_type.is_fortran_array())
{
// We need to parse this bit in Fortran
Source number_of_bytes;
number_of_bytes << "SIZE(" << (*it)->get_symbol().get_name() << ") * " << reduction_element_type.get_size();
element_size << as_expression(number_of_bytes.parse_expression(construct));
}
else
{
element_size << "sizeof(" << as_type(reduction_type) << ")";
}
}
else
{
element_size << "sizeof(" << as_type(reduction_type) << ")";
}
reduction_declaration
<< "nanos_reduction_t* " << nanos_red_name << ";"
;
Source allocate_private_buffer, cleanup_code;
Source num_scalars;
TL::Symbol basic_reduction_function, vector_reduction_function;
create_reduction_function(reduction, construct, reduction_type, basic_reduction_function, vector_reduction_function);
(*it)->reduction_set_basic_function(basic_reduction_function);
thread_initializing_reduction_info
<< "err = nanos_malloc((void**)&" << nanos_red_name << ", sizeof(nanos_reduction_t), "
<< "\"" << construct.get_filename() << "\", " << construct.get_line() << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< nanos_red_name << "->original = (void*)"
<< (reduction_type.is_array() ? "" : "&") << (*it)->get_symbol().get_name() << ";"
<< allocate_private_buffer
<< nanos_red_name << "->vop = "
<< (vector_reduction_function.is_valid() ? as_symbol(vector_reduction_function) : "0") << ";"
<< nanos_red_name << "->bop = (void(*)(void*,void*,int))" << as_symbol(basic_reduction_function) << ";"
<< nanos_red_name << "->element_size = " << element_size << ";"
<< nanos_red_name << "->num_scalars = " << num_scalars << ";"
<< cleanup_code
<< "err = nanos_register_reduction(" << nanos_red_name << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
;
if (IS_C_LANGUAGE
|| IS_CXX_LANGUAGE)
{
if (reduction_type.is_array())
{
num_scalars << "sizeof(" << as_type(reduction_type) << ") / sizeof(" << as_type(reduction_element_type) <<")";
}
else
{
num_scalars << "1";
}
allocate_private_buffer
<< "err = nanos_malloc(&" << nanos_red_name << "->privates, sizeof(" << as_type(reduction_type) << ") * nanos_num_threads, "
<< "\"" << construct.get_filename() << "\", " << construct.get_line() << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< nanos_red_name << "->descriptor = " << nanos_red_name << "->privates;"
<< "rdv_" << (*it)->get_field_name() << " = (" << as_type( (*it)->get_private_type().get_pointer_to() ) << ")" << nanos_red_name << "->privates;"
;
thread_fetching_reduction_info
<< "err = nanos_reduction_get(&" << nanos_red_name << ", "
<< (reduction_type.is_array() ? "" : "&") << (*it)->get_symbol().get_name() << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< "rdv_" << (*it)->get_field_name() << " = (" << as_type( (*it)->get_private_type().get_pointer_to() ) << ")" << nanos_red_name << "->privates;"
;
cleanup_code
<< nanos_red_name << "->cleanup = nanos_free0;"
;
}
else if (IS_FORTRAN_LANGUAGE)
{
Type private_reduction_vector_type;
Source extra_dims;
{
TL::Type t = (*it)->get_symbol().get_type().no_ref();
int rank = 0;
if (t.is_fortran_array())
{
rank = t.fortran_rank();
}
if (rank != 0)
{
// We need to parse this bit in Fortran
Source size_call;
size_call << "SIZE(" << (*it)->get_symbol().get_name() << ")";
num_scalars << as_expression(size_call.parse_expression(construct));
}
else
{
num_scalars << "1";
}
private_reduction_vector_type = fortran_get_n_ranked_type_with_descriptor(
get_void_type(), rank + 1, construct.retrieve_context().get_decl_context());
int i;
for (i = 0; i < rank; i++)
{
Source lbound_src;
lbound_src << "LBOUND(" << (*it)->get_symbol().get_name() << ", DIM = " << (rank - i) << ")";
Source ubound_src;
ubound_src << "UBOUND(" << (*it)->get_symbol().get_name() << ", DIM = " << (rank - i) << ")";
extra_dims
<< "["
<< as_expression(lbound_src.parse_expression(construct))
<< ":"
<< as_expression(ubound_src.parse_expression(construct))
<< "]";
t = t.array_element();
}
}
allocate_private_buffer
<< "@FORTRAN_ALLOCATE@((*rdv_" << (*it)->get_field_name() << ")[0:(nanos_num_threads-1)]" << extra_dims <<");"
<< nanos_red_name << "->privates = &(*rdv_" << (*it)->get_field_name() << ");"
<< "err = nanos_malloc(&" << nanos_red_name << "->descriptor, sizeof(" << as_type(private_reduction_vector_type) << "), "
<< "\"" << construct.get_filename() << "\", " << construct.get_line() << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< "err = nanos_memcpy(" << nanos_red_name << "->descriptor, "
"&rdv_" << (*it)->get_field_name() << ", sizeof(" << as_type(private_reduction_vector_type) << "));"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
;
thread_fetching_reduction_info
<< "err = nanos_reduction_get(&" << nanos_red_name << ", &" << (*it)->get_symbol().get_name() << ");"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
<< "err = nanos_memcpy("
<< "&rdv_" << (*it)->get_field_name() << ","
<< nanos_red_name << "->descriptor, "
<< "sizeof(" << as_type(private_reduction_vector_type) << "));"
<< "if (err != NANOS_OK)"
<< "nanos_handle_error(err);"
;
TL::Symbol reduction_cleanup = create_reduction_cleanup_function(reduction, construct);
cleanup_code
<< nanos_red_name << "->cleanup = " << as_symbol(reduction_cleanup) << ";"
;
}
else
{
internal_error("Code unreachable", 0);
}
}
FORTRAN_LANGUAGE()
{
Source::source_language = SourceLanguage::C;
}
ref_tree.replace(result.parse_statement(ref_tree));
FORTRAN_LANGUAGE()
{
Source::source_language = SourceLanguage::Current;
}
}
TL::Symbol new_function_symbol(
TL::Symbol current_function,
const std::string& name,
TL::Type return_type,
TL::ObjectList<std::string> parameter_names,
TL::ObjectList<TL::Type> parameter_types)
{
if (IS_FORTRAN_LANGUAGE && current_function.is_nested_function())
{
// Get the enclosing function
current_function = current_function.get_scope().get_related_symbol();
}
decl_context_t decl_context = current_function.get_scope().get_decl_context();
ERROR_CONDITION(parameter_names.size() != parameter_types.size(), "Mismatch between names and types", 0);
decl_context_t function_context;
if (IS_FORTRAN_LANGUAGE)
{
function_context = new_program_unit_context(decl_context);
}
else
{
function_context = new_function_context(decl_context);
function_context = new_block_context(function_context);
}
// Build the function type
int num_parameters = 0;
scope_entry_t** parameter_list = NULL;
parameter_info_t* p_types = new parameter_info_t[parameter_types.size()];
parameter_info_t* it_ptypes = &(p_types[0]);
TL::ObjectList<TL::Type>::iterator type_it = parameter_types.begin();
for (TL::ObjectList<std::string>::iterator it = parameter_names.begin();
it != parameter_names.end();
it++, it_ptypes++, type_it++)
{
scope_entry_t* param = new_symbol(function_context, function_context.current_scope, it->c_str());
param->entity_specs.is_user_declared = 1;
param->kind = SK_VARIABLE;
param->locus = make_locus("", 0, 0);
param->defined = 1;
param->type_information = get_unqualified_type(type_it->get_internal_type());
P_LIST_ADD(parameter_list, num_parameters, param);
it_ptypes->is_ellipsis = 0;
it_ptypes->nonadjusted_type_info = NULL;
it_ptypes->type_info = get_indirect_type(param);
}
type_t *function_type = get_new_function_type(
return_type.get_internal_type(),
p_types,
parameter_types.size());
delete[] p_types;
// Now, we can create the new function symbol
scope_entry_t* new_function_sym = NULL;
if (!current_function.get_type().is_template_specialized_type())
{
new_function_sym = new_symbol(decl_context, decl_context.current_scope, name.c_str());
new_function_sym->entity_specs.is_user_declared = 1;
new_function_sym->kind = SK_FUNCTION;
new_function_sym->locus = make_locus("", 0, 0);
new_function_sym->type_information = function_type;
}
else
{
scope_entry_t* new_template_sym = new_symbol(
decl_context, decl_context.current_scope, name.c_str());
new_template_sym->kind = SK_TEMPLATE;
new_template_sym->locus = make_locus("", 0, 0);
new_template_sym->type_information = get_new_template_type(
decl_context.template_parameters,
function_type,
uniquestr(name.c_str()),
decl_context, make_locus("", 0, 0));
template_type_set_related_symbol(new_template_sym->type_information, new_template_sym);
// The new function is the primary template specialization
new_function_sym = named_type_get_symbol(
template_type_get_primary_type(
new_template_sym->type_information));
}
function_context.function_scope->related_entry = new_function_sym;
function_context.block_scope->related_entry = new_function_sym;
new_function_sym->related_decl_context = function_context;
new_function_sym->entity_specs.related_symbols = parameter_list;
new_function_sym->entity_specs.num_related_symbols = num_parameters;
for (int i = 0; i < new_function_sym->entity_specs.num_related_symbols; ++i)
{
symbol_set_as_parameter_of_function(
new_function_sym->entity_specs.related_symbols[i], new_function_sym, /* parameter position */ i);
}
// Make it static
new_function_sym->entity_specs.is_static = 1;
// Make it member if the enclosing function is member
if (current_function.is_member())
{
new_function_sym->entity_specs.is_member = 1;
new_function_sym->entity_specs.class_type = current_function.get_class_type().get_internal_type();
new_function_sym->entity_specs.access = AS_PUBLIC;
::class_type_add_member(new_function_sym->entity_specs.class_type, new_function_sym);
}
if (current_function.is_inline())
new_function_sym->entity_specs.is_inline = 1;
// new_function_sym->entity_specs.is_defined_inside_class_specifier =
// current_function.get_internal_symbol()->entity_specs.is_defined_inside_class_specifier;
if (IS_FORTRAN_LANGUAGE && current_function.is_in_module())
{
scope_entry_t* module_sym = current_function.in_module().get_internal_symbol();
new_function_sym->entity_specs.in_module = module_sym;
P_LIST_ADD(
module_sym->entity_specs.related_symbols,
module_sym->entity_specs.num_related_symbols,
new_function_sym);
new_function_sym->entity_specs.is_module_procedure = 1;
}
return new_function_sym;
}
TL::Source LoopDistribution::do_distribution()
{
TL::Statement loop_body = _for_stmt.get_loop_body();
if (!loop_body.is_compound_statement())
{
return _for_stmt.prettyprint();
}
TL::Source result;
TL::Source distributed_loops, expanded_scalars;
result
<< "{"
<< expanded_scalars
<< distributed_loops
<< "}"
;
TL::ReplaceIdExpression escalar_expansion;
if (!_expand.empty())
{
if (!_for_stmt.is_regular_loop())
{
return _for_stmt.prettyprint();
}
expanded_scalars
<< "int __loop_trip = (" << _for_stmt.get_upper_bound() << ") - (" << _for_stmt.get_lower_bound() << " + 1);"
// Absolute value (calculated this way allows for conditional move instructions)
<< "__loop_trip = (__loop_trip < 0) ? (-__loop_trip) : __loop_trip;"
;
// We need this because of array nature
TL::AST_t array_expr;
{
TL::AST_t array_expr_ref_tree;
Source temporal_parse;
temporal_parse
<< "{"
<< expanded_scalars
<< statement_placeholder(array_expr_ref_tree)
<< "}"
;
temporal_parse.parse_statement(_for_stmt.get_ast(),
_for_stmt.get_scope_link());
array_expr = Source("__loop_trip").parse_expression(array_expr_ref_tree,
_for_stmt.get_scope_link());
}
for (TL::ObjectList<TL::Symbol>::iterator it = _expand.begin();
it != _expand.end();
it++)
{
TL::Symbol &sym(*it);
std::string expanded_scalar_name = "_" + sym.get_name();
TL::Type type = sym.get_type();
TL::Type array_type = type.get_array_to(array_expr,
_for_stmt.get_scope_link().get_scope(array_expr));
expanded_scalars
<< array_type.get_declaration(it->get_scope(), expanded_scalar_name) << ";";
escalar_expansion.add_replacement(sym,
expanded_scalar_name + "[" + _for_stmt.get_induction_variable().prettyprint() + "]" );
}
}
TL::ObjectList<TL::Statement> statement_list = loop_body.get_inner_statements();
for (TL::ObjectList<TL::Statement>::iterator it = statement_list.begin();
it != statement_list.end();
it++)
{
Statement &stmt(*it);
distributed_loops
<< "for ( " << _for_stmt.get_iterating_init().prettyprint() // This one already includes ';'
<< _for_stmt.get_iterating_condition() << ";"
<< _for_stmt.get_iterating_expression() << ")"
<< "{"
<< escalar_expansion.replace(stmt)
<< "}"
;
}
return result;
}
void LoweringVisitor::perform_partial_reduction(OutlineInfo& outline_info, Nodecl::NodeclBase ref_tree)
{
ERROR_CONDITION(ref_tree.is_null(), "Invalid tree", 0);
Source reduction_code;
TL::ObjectList<OutlineDataItem*> reduction_items = outline_info.get_data_items().filter(
predicate(lift_pointer(functor(&OutlineDataItem::is_reduction))));
if (!reduction_items.empty())
{
for (TL::ObjectList<OutlineDataItem*>::iterator it = reduction_items.begin();
it != reduction_items.end();
it++)
{
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
if ((*it)->get_private_type().is_array())
{
reduction_code
<< "__builtin_memcpy(rdv_" << (*it)->get_field_name() << "[nanos_omp_get_thread_num()],"
<< "rdp_" << (*it)->get_symbol().get_name() << ","
<< " sizeof(" << as_type((*it)->get_private_type()) << "));"
;
}
else
{
reduction_code
<< "rdv_" << (*it)->get_field_name() << "[nanos_omp_get_thread_num()] "
<< "= rdp_" << (*it)->get_symbol().get_name() << ";"
;
}
}
else if (IS_FORTRAN_LANGUAGE)
{
Source extra_dims;
{
TL::Type t = (*it)->get_symbol().get_type().no_ref();
int rank = 0;
if (t.is_fortran_array())
{
rank = t.fortran_rank();
}
int i;
for (i = 0; i < rank; i++)
{
extra_dims << ":,";
}
}
reduction_code
<< "rdv_" << (*it)->get_field_name() << "( " << extra_dims << "nanos_omp_get_thread_num() ) = rdp_" << (*it)->get_symbol().get_name() << "\n"
;
}
else
{
internal_error("Code unreachable", 0);
}
}
}
ref_tree.replace(reduction_code.parse_statement(ref_tree));
}
static TL::Symbol create_initializer_function_fortran(
OpenMP::Reduction* red,
TL::Type reduction_type,
Nodecl::NodeclBase construct)
{
std::string fun_name;
{
std::stringstream ss;
ss << "nanos_ini_" << red << "_" << reduction_type.get_internal_type() << "_" << simple_hash_str(construct.get_filename().c_str());
fun_name = ss.str();
}
Nodecl::NodeclBase initializer = red->get_initializer().shallow_copy();
TL::Type omp_out_type = reduction_type,
omp_ori_type = reduction_type;
// These sources are only used in array reductions
TL::Source omp_out_extra_attributes,
extra_stuff_array_red;
if (reduction_type.is_array())
{
Source dims_descr;
TL::Type t = reduction_type;
int rank = 0;
if (t.is_fortran_array())
{
rank = t.fortran_rank();
}
dims_descr << "(";
omp_out_extra_attributes << ", POINTER, DIMENSION(";
int i;
for (i = 0; i < rank; i++)
{
if (i != 0)
{
dims_descr << ",";
omp_out_extra_attributes << ",";
}
dims_descr << "LBOUND(omp_orig, DIM = " << (rank - i) << ")"
<< ":"
<< "UBOUND(omp_orig, DIM = " << (rank - i) << ")"
;
omp_out_extra_attributes << ":";
t = t.array_element();
}
dims_descr << ")";
omp_out_extra_attributes << ")";
omp_out_type = t;
extra_stuff_array_red << "ALLOCATE(omp_out" << dims_descr <<")\n";
}
Source src;
src << "SUBROUTINE " << fun_name << "(omp_out, omp_orig)\n"
<< "IMPLICIT NONE\n"
<< as_type(omp_out_type) << omp_out_extra_attributes << " :: omp_out\n"
<< as_type(omp_ori_type) << " :: omp_orig\n"
<< extra_stuff_array_red
<< "omp_out = " << as_expression(initializer) << "\n"
<< "END SUBROUTINE " << fun_name << "\n"
;
TL::Scope global_scope = construct.retrieve_context().get_global_scope();
Nodecl::NodeclBase function_code = src.parse_global(global_scope);
TL::Symbol function_sym = global_scope.get_symbol_from_name(fun_name);
ERROR_CONDITION(!function_sym.is_valid(), "Symbol %s not found", fun_name.c_str());
// As the initializer function is needed during the instantiation of
// the task, this function should be inserted before the construct
Nodecl::Utils::prepend_to_enclosing_top_level_location(construct,
function_code);
return function_sym;
}
TL::Source LoopUnroll::do_unroll()
{
if (!_for_stmt.regular_loop())
{
return silly_unroll();
}
// Get parts of the loop
IdExpression induction_var = _for_stmt.get_induction_variable();
Expression lower_bound = _for_stmt.get_lower_bound();
Expression upper_bound = _for_stmt.get_upper_bound();
Expression step = _for_stmt.get_step();
TL::Source operator_bound = _for_stmt.get_bound_operator();
Statement loop_body = _for_stmt.get_loop_body();
TL::Source result, epilogue,
main, induction_var_decl,
before_main, after_main;
std::stringstream ss;
ss << _factor;
result
<< "{"
<< induction_var_decl
<< before_main
<< main
<< after_main
<< epilogue
<< "}"
;
Source replicated_body;
Source epilogue_body;
if (_factor > 1)
{
AST_t init = _for_stmt.get_iterating_init();
if (Declaration::predicate(init))
{
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
// Declare it since it will have local scope
induction_var_decl
<< type.get_declaration(sym.get_scope(), sym.get_name()) << ";"
;
}
main
<< "for (" << induction_var << " = " << lower_bound << ";"
<< induction_var << operator_bound << "((" << upper_bound << ") - (" << _factor << " - 1)* (" << step << "));"
<< induction_var << "+= (" << step << ") * " << _factor << ")"
<< "{"
<< replicated_body
<< "}"
;
// FIXME - It could help to initialize here another variable and make both loops independent
epilogue
<< "for ( ; " // No initialization, keep using the old induction var
<< induction_var << operator_bound << upper_bound << ";"
<< induction_var << "+= (" << step << "))"
<< epilogue_body
;
if (!_remove_tasks)
{
epilogue_body << loop_body;
}
else
{
std::cerr << "Do not create task " << __FILE__ << ":" << __LINE__ << std::endl;
running_error("Path not supported yet", 0);
// epilogue_body << loop_body.get_ast().prettyprint_with_callback(functor(ignore_tasks));
}
}
else
{
// Leave it as is
main << "for(" << _for_stmt.get_iterating_init().prettyprint()
<< _for_stmt.get_iterating_condition() << ";"
<< _for_stmt.get_iterating_expression() << ")"
<< "{"
<< replicated_body
<< "}"
;
}
// Replicate the body
bool consider_omp = false;
if (TaskAggregation::contains_relevant_openmp(loop_body))
{
consider_omp = true;
}
if (_ignore_omp || !consider_omp)
{
simple_replication(_factor, replicated_body, epilogue_body,
induction_var, loop_body);
}
else
{
omp_replication(_factor, replicated_body, epilogue_body,
induction_var, loop_body, before_main, after_main);
}
return result;
}
Type Type::fix_references_()
{
if ((IS_C_LANGUAGE && this->is_any_reference())
|| (IS_CXX_LANGUAGE && this->is_rebindable_reference()))
{
TL::Type ref = this->references_to();
if (ref.is_array())
{
// T (&a)[10] -> T * const
// T (&a)[10][20] -> T (* const)[20]
ref = ref.array_element();
}
// T &a -> T * const a
TL::Type ptr = ref.get_pointer_to();
if (!this->is_rebindable_reference())
{
ptr = ptr.get_const_type();
}
return ptr;
}
else if (IS_FORTRAN_LANGUAGE && this->is_any_reference())
{
return this->references_to();
}
else if (this->is_array())
{
if (this->array_is_region())
{
Nodecl::NodeclBase lb, reg_lb, ub, reg_ub;
this->array_get_bounds(lb, ub);
this->array_get_region_bounds(reg_lb, reg_ub);
TL::Scope sc = array_type_get_region_size_expr_context(this->get_internal_type());
return this->array_element().fix_references_().get_array_to_with_region(lb, ub, reg_lb, reg_ub, sc);
}
else
{
Nodecl::NodeclBase size = this->array_get_size();
TL::Scope sc = array_type_get_array_size_expr_context(this->get_internal_type());
return this->array_element().fix_references_().get_array_to(size, sc);
}
}
else if (this->is_pointer())
{
TL::Type fixed = this->points_to().fix_references_().get_pointer_to();
cv_qualifier_t cv_qualif = CV_NONE;
::advance_over_typedefs_with_cv_qualif(this->get_internal_type(), &cv_qualif);
fixed = ::get_cv_qualified_type(fixed.get_internal_type(), cv_qualif);
return fixed;
}
else if (this->is_function())
{
// Do not fix unprototyped functions
if (this->lacks_prototype())
return (*this);
cv_qualifier_t cv_qualif = CV_NONE;
::advance_over_typedefs_with_cv_qualif(this->get_internal_type(), &cv_qualif);
ref_qualifier_t ref_qualifier = function_type_get_ref_qualifier(this->get_internal_type());
TL::Type fixed_result = this->returns().fix_references_();
bool has_ellipsis = 0;
TL::ObjectList<TL::Type> fixed_parameters = this->parameters(has_ellipsis);
for (TL::ObjectList<TL::Type>::iterator it = fixed_parameters.begin();
it != fixed_parameters.end();
it++)
{
*it = it->fix_references_();
}
TL::ObjectList<TL::Type> nonadjusted_fixed_parameters = this->nonadjusted_parameters();
for (TL::ObjectList<TL::Type>::iterator it = nonadjusted_fixed_parameters.begin();
it != nonadjusted_fixed_parameters.end();
it++)
{
*it = it->fix_references_();
}
TL::Type fixed_function = fixed_result.get_function_returning(
fixed_parameters,
nonadjusted_fixed_parameters,
has_ellipsis,
ref_qualifier);
fixed_function = TL::Type(get_cv_qualified_type(fixed_function.get_internal_type(), cv_qualif));
return fixed_function;
}
// Note: we are not fixing classes
else
{
// Anything else must be left untouched
return (*this);
}
}