void goto_symext::symex_free(const expr2tc &expr)
{
const code_free2t &code = to_code_free2t(expr);
// Trigger 'free'-mode dereference of this pointer. Should generate various
// dereference failure callbacks.
expr2tc tmp = code.operand;
dereference(tmp, false, true);
address_of2tc addrof(code.operand->type, tmp);
pointer_offset2tc ptr_offs(pointer_type2(), addrof);
equality2tc eq(ptr_offs, zero_ulong);
claim(eq, "Operand of free must have zero pointer offset");
// Clear the alloc bit, and set the deallocated bit.
guardt guard;
type2tc sym_type = type2tc(new array_type2t(get_bool_type(),
expr2tc(), true));
pointer_object2tc ptr_obj(pointer_type2(), code.operand);
symbol2tc dealloc_sym(sym_type, deallocd_arr_name);
index2tc dealloc_index_expr(get_bool_type(), dealloc_sym, ptr_obj);
expr2tc truth = true_expr;
symex_assign_rec(dealloc_index_expr, truth, guard);
symbol2tc valid_sym(sym_type, valid_ptr_arr_name);
index2tc valid_index_expr(get_bool_type(), valid_sym, ptr_obj);
expr2tc falsity = false_expr;
symex_assign_rec(valid_index_expr, falsity, guard);
}
void goto_symext::track_new_pointer(
const expr2tc &ptr_obj,
const type2tc &new_type,
const expr2tc &size)
{
// Also update all the accounting data.
// Mark that object as being dynamic, in the __ESBMC_is_dynamic array
type2tc sym_type =
type2tc(new array_type2t(get_bool_type(), expr2tc(), true));
symbol2tc sym(sym_type, dyn_info_arr_name);
index2tc idx(get_bool_type(), sym, ptr_obj);
expr2tc truth = gen_true_expr();
symex_assign(code_assign2tc(idx, truth), true);
symbol2tc valid_sym(sym_type, valid_ptr_arr_name);
index2tc valid_index_expr(get_bool_type(), valid_sym, ptr_obj);
truth = gen_true_expr();
symex_assign(code_assign2tc(valid_index_expr, truth), true);
symbol2tc dealloc_sym(sym_type, deallocd_arr_name);
index2tc dealloc_index_expr(get_bool_type(), dealloc_sym, ptr_obj);
expr2tc falseity = gen_false_expr();
symex_assign(code_assign2tc(dealloc_index_expr, falseity), true);
type2tc sz_sym_type =
type2tc(new array_type2t(pointer_type2(), expr2tc(), true));
symbol2tc sz_sym(sz_sym_type, alloc_size_arr_name);
index2tc sz_index_expr(get_bool_type(), sz_sym, ptr_obj);
expr2tc object_size_exp;
if(is_nil_expr(size))
{
try
{
mp_integer object_size = type_byte_size(new_type);
object_size_exp =
constant_int2tc(pointer_type2(), object_size.to_ulong());
}
catch(array_type2t::dyn_sized_array_excp *e)
{
object_size_exp = typecast2tc(pointer_type2(), e->size);
}
}
else
{
object_size_exp = size;
}
symex_assign(code_assign2tc(sz_index_expr, object_size_exp), true);
}
void goto_symext::symex_free(const expr2tc &expr)
{
const code_free2t &code = to_code_free2t(expr);
// Trigger 'free'-mode dereference of this pointer. Should generate various
// dereference failure callbacks.
expr2tc tmp = code.operand;
dereference(tmp, dereferencet::FREE);
// Don't rely on the output of dereference in free mode; instead fetch all
// the internal dereference state for pointed at objects, and creates claims
// that if pointed at, their offset is zero.
internal_deref_items.clear();
tmp = code.operand;
// Create temporary, dummy, dereference
tmp = dereference2tc(get_uint8_type(), tmp);
dereference(tmp, dereferencet::INTERNAL);
// Only add assertions to check pointer offset if pointer check is enabled
if(!options.get_bool_option("no-pointer-check"))
{
for(auto const &item : internal_deref_items)
{
guardt g = cur_state->guard;
g.add(item.guard);
expr2tc offset = item.offset;
expr2tc eq = equality2tc(offset, gen_ulong(0));
g.guard_expr(eq);
claim(eq, "Operand of free must have zero pointer offset");
}
}
// Clear the alloc bit, and set the deallocated bit.
type2tc sym_type =
type2tc(new array_type2t(get_bool_type(), expr2tc(), true));
expr2tc ptr_obj = pointer_object2tc(pointer_type2(), code.operand);
dereference(ptr_obj, dereferencet::READ);
symbol2tc dealloc_sym(sym_type, deallocd_arr_name);
index2tc dealloc_index_expr(get_bool_type(), dealloc_sym, ptr_obj);
expr2tc truth = gen_true_expr();
symex_assign(code_assign2tc(dealloc_index_expr, truth), true);
symbol2tc valid_sym(sym_type, valid_ptr_arr_name);
index2tc valid_index_expr(get_bool_type(), valid_sym, ptr_obj);
expr2tc falsity = gen_false_expr();
symex_assign(code_assign2tc(valid_index_expr, falsity), true);
}
void goto_symext::symex_free(const expr2tc &expr)
{
const code_free2t &code = to_code_free2t(expr);
// Trigger 'free'-mode dereference of this pointer. Should generate various
// dereference failure callbacks.
expr2tc tmp = code.operand;
dereference(tmp, false, true);
// Don't rely on the output of dereference in free mode; instead fetch all
// the internal dereference state for pointed at objects, and creates claims
// that if pointed at, their offset is zero.
internal_deref_items.clear();
tmp = code.operand;
// Create temporary, dummy, dereference
tmp = dereference2tc(get_uint8_type(), tmp);
dereference(tmp, false, false, true); // 'internal' dereference
for (const auto &item : internal_deref_items) {
guardt g = cur_state->guard;
g.add(item.guard);
expr2tc offset = item.offset;
expr2tc eq = equality2tc(offset, zero_ulong);
g.guard_expr(eq);
claim(eq, "Operand of free must have zero pointer offset");
}
// Clear the alloc bit, and set the deallocated bit.
guardt guard;
type2tc sym_type = type2tc(new array_type2t(get_bool_type(),
expr2tc(), true));
pointer_object2tc ptr_obj(pointer_type2(), code.operand);
symbol2tc dealloc_sym(sym_type, deallocd_arr_name);
index2tc dealloc_index_expr(get_bool_type(), dealloc_sym, ptr_obj);
expr2tc truth = true_expr;
symex_assign_rec(dealloc_index_expr, truth, guard);
symbol2tc valid_sym(sym_type, valid_ptr_arr_name);
index2tc valid_index_expr(get_bool_type(), valid_sym, ptr_obj);
expr2tc falsity = false_expr;
symex_assign_rec(valid_index_expr, falsity, guard);
}
Type* type_check_not(Expr const* e1)
{
if (is_bool_type(e1->type()))
return get_bool_type();
error("Expression not of bool type found in not expression: ");
print(e1);
print("\n");
return nullptr;
}
Type* type_check_equality(Expr const* e1, Expr const* e2)
{
if (e1->type() == e2->type())
return get_bool_type();
else {
error("Expressions of mismatched types found in equality expressions: ");
print(e1);
print(" and ");
print(e2);
print("\n");
}
return nullptr;
}
// Returns the type of a unary expression.
//
// The operand of a unary arithmetic expression (-e, +e, and ~e)
// shall have integer type. The result type the expression is `int`.
//
// The operand of the unary logical expression (!e) shall have
// boolean type. The result type the expression is `bool`.
Type const*
get_type(Unary_op op, Expr const* e)
{
Type const* z = get_int_type();
Type const* b = get_bool_type();
switch (op) {
case num_neg_op:
case num_pos_op:
case bit_not_op:
return expect_type(e, z, z);
case log_not_op:
return expect_type(e, b, b);
default:
break;
}
lingo_unreachable();
}
Type* type_check_ordering(Expr const* e1, Expr const* e2)
{
if (is_reals_type(e1->type())) {
if (is_reals_type(e2->type()))
return get_bool_type();
else {
error("Expression not of integer or double type found in ordering expression: ");
print(e2);
print("\n");
}
}
else {
error("Expression not of integer or double type found in ordering expression: ");
print(e1);
print("\n");
}
return nullptr;
}
Type* type_check_logical(Expr const* e1, Expr const* e2)
{
if (is_bool_type(e1->type())) {
if (is_bool_type(e2->type()))
return get_bool_type();
else {
error("Expression not of bool type found in logical expression: ");
print(e2);
print("\n");
}
}
else {
error("Expression not of bool type found in logical expression: ");
print(e1);
print("\n");
}
return nullptr;
}
// Returns the type of a binary expression.
//
// The operands of a binary arithmetic expression (e1 + e2, e1 - e2,
// e1 * e2, e1 / e2, e1 % e2, e1 & e2, e1 | e2, e1 ^ e2, e1 << e2,
// and e1 >> e2) shall have integer type. The result type the
// expression is `int`.
//
// The operands of a binary relational expression (e1 < e2, e1 > e2,
// e1 <= e2, e1 >= e2, e1 == e2, and e1 != e2) shall have integer or
// boolean type. The result type the expression is `bool`.
//
// The operands of a binary logical expression (e1 && e2 and e1 || e2)
// shall have boolean type. The result type the expression is `bool`.
Type const*
get_type(Binary_op op, Expr const* e1, Expr const* e2)
{
Type const* z = get_int_type();
Type const* b = get_bool_type();
switch (op) {
case num_add_op:
case num_sub_op:
case num_mul_op:
case num_div_op:
case num_mod_op:
case bit_and_op:
case bit_or_op:
case bit_xor_op:
case bit_lsh_op:
case bit_rsh_op:
// Arithmetic expressions have integer opreands and results.
return expect_type(e1, e2, z, z);
case rel_eq_op:
case rel_ne_op:
case rel_lt_op:
case rel_gt_op:
case rel_le_op:
case rel_ge_op:
// Relational expressions have any type and the result
// is bool.
return b;
case log_and_op:
case log_or_op:
// Logical expressions have boolean operands and result.
return expect_type(e1, e2, b, b);
default:
break;
}
lingo_unreachable();
}
void LoweringVisitor::loop_spawn_worksharing(OutlineInfo& outline_info,
Nodecl::NodeclBase construct,
Nodecl::List distribute_environment,
Nodecl::RangeLoopControl& range,
const std::string& outline_name,
TL::Symbol structure_symbol,
TL::Symbol slicer_descriptor,
Nodecl::NodeclBase task_label)
{
Symbol enclosing_function = Nodecl::Utils::get_enclosing_function(construct);
Nodecl::OpenMP::Schedule schedule = distribute_environment.find_first<Nodecl::OpenMP::Schedule>();
ERROR_CONDITION(schedule.is_null(), "Schedule tree is missing", 0);
Nodecl::NodeclBase lower = range.get_lower();
Nodecl::NodeclBase upper = range.get_upper();
Nodecl::NodeclBase step = range.get_step();
Source struct_size, dynamic_size, struct_arg_type_name;
struct_arg_type_name
<< ((structure_symbol.get_type().is_template_specialized_type()
&& structure_symbol.get_type().is_dependent()) ? "typename " : "")
<< structure_symbol.get_qualified_name(enclosing_function.get_scope())
;
struct_size << "sizeof( " << struct_arg_type_name << " )" << dynamic_size;
Source immediate_decl;
allocate_immediate_structure(
structure_symbol.get_user_defined_type(),
outline_info,
struct_arg_type_name,
struct_size,
// out
immediate_decl,
dynamic_size);
Source call_outline_function;
Source schedule_setup;
schedule_setup
<< "int nanos_chunk;"
;
if (schedule.get_text() == "runtime")
{
schedule_setup
<< "nanos_omp_sched_t nanos_runtime_sched;"
<< "nanos_err = nanos_omp_get_schedule(&nanos_runtime_sched, &nanos_chunk);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
<< "nanos_ws_t current_ws_policy = nanos_omp_find_worksharing(nanos_runtime_sched);"
;
}
else
{
Source schedule_name;
if (Nanos::Version::interface_is_at_least("openmp", 8))
{
schedule_name << "nanos_omp_sched_" << schedule.get_text();
}
else
{
// We used nanos_omp_sched in versions prior to 8
schedule_name << "omp_sched_" << schedule.get_text();
}
schedule_setup
<< "nanos_ws_t current_ws_policy = nanos_omp_find_worksharing(" << schedule_name << ");"
<< "if (current_ws_policy == 0)"
<< "nanos_handle_error(NANOS_UNIMPLEMENTED);"
<< "nanos_chunk = " << as_expression(schedule.get_chunk()) << ";"
;
}
Source worksharing_creation;
if (IS_CXX_LANGUAGE)
{
worksharing_creation
<< as_statement(Nodecl::CxxDef::make(Nodecl::NodeclBase::null(), slicer_descriptor));
}
worksharing_creation
<< "nanos_err = nanos_worksharing_create("
<< "&" << as_symbol(slicer_descriptor) << ","
<< "current_ws_policy,"
<< "(void**)&nanos_setup_info_loop,"
<< "&single_guard);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
;
Nodecl::NodeclBase fill_outline_arguments_tree, fill_immediate_arguments_tree;
TL::Source pm_specific_code;
if (!_lowering->in_ompss_mode())
{
// OpenMP
pm_specific_code
<< immediate_decl
<< statement_placeholder(fill_immediate_arguments_tree)
<< "smp_" << outline_name << "(imm_args);"
;
}
else
{
// OmpSs
std::string wd_description =
(!task_label.is_null()) ? task_label.get_text() : enclosing_function.get_name();
Source const_wd_info;
const_wd_info
<< fill_const_wd_info(struct_arg_type_name,
/* is_untied */ false,
/* mandatory_creation */ true,
/* is_function_task */ false,
wd_description,
outline_info,
construct);
std::string dyn_props_var = "nanos_wd_dyn_props";
Source dynamic_wd_info;
dynamic_wd_info << "nanos_wd_dyn_props_t " << dyn_props_var << ";";
fill_dynamic_properties(dyn_props_var,
/* priority_expr */ nodecl_null(), /* final_expr */ nodecl_null(), /* is_implicit */ 0, dynamic_wd_info);
pm_specific_code
<< struct_arg_type_name << " *ol_args = (" << struct_arg_type_name <<"*) 0;"
<< const_wd_info
<< "nanos_wd_t nanos_wd_ = (nanos_wd_t) 0;"
<< dynamic_wd_info
<< "static nanos_slicer_t replicate = (nanos_slicer_t)0;"
<< "if (replicate == (nanos_slicer_t)0)"
<< "replicate = nanos_find_slicer(\"replicate\");"
<< "if (replicate == (nanos_slicer_t)0)"
<< "nanos_handle_error(NANOS_UNIMPLEMENTED);"
<< "nanos_err = nanos_create_sliced_wd(&nanos_wd_, "
<< "nanos_wd_const_data.base.num_devices, nanos_wd_const_data.devices, "
<< "(size_t)" << struct_size << ", nanos_wd_const_data.base.data_alignment, "
<< "(void**)&ol_args, nanos_current_wd(), replicate,"
<< "&nanos_wd_const_data.base.props, &" << dyn_props_var << ", 0, (nanos_copy_data_t**)0,"
<< "0, (nanos_region_dimension_internal_t**)0"
<< ");"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
<< statement_placeholder(fill_outline_arguments_tree)
<< "nanos_err = nanos_submit(nanos_wd_, 0, (nanos_data_access_t *) 0, (nanos_team_t) 0);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
;
}
TL::Source implicit_barrier_or_tw;
if (!distribute_environment.find_first<Nodecl::OpenMP::BarrierAtEnd>().is_null())
{
implicit_barrier_or_tw << get_implicit_sync_end_construct_source();
}
Source spawn_code;
spawn_code
<< "{"
<< as_type(get_bool_type()) << " single_guard;"
<< "nanos_err_t nanos_err;"
<< schedule_setup
<< "nanos_ws_info_loop_t nanos_setup_info_loop;"
<< "nanos_setup_info_loop.lower_bound = " << as_expression(lower) << ";"
<< "nanos_setup_info_loop.upper_bound = " << as_expression(upper) << ";"
<< "nanos_setup_info_loop.loop_step = " << as_expression(step) << ";"
<< "nanos_setup_info_loop.chunk_size = nanos_chunk;"
<< worksharing_creation
<< pm_specific_code
<< implicit_barrier_or_tw
<< "}"
;
Source fill_outline_arguments, fill_immediate_arguments;
fill_arguments(construct, outline_info, fill_outline_arguments, fill_immediate_arguments);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase spawn_code_tree = spawn_code.parse_statement(construct);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::Current;
Nodecl::NodeclBase arguments_tree;
TL::Source *fill_arguments;
if (!_lowering->in_ompss_mode())
{
// OpenMP
arguments_tree = fill_immediate_arguments_tree;
fill_arguments = &fill_immediate_arguments;
}
else
{
// OmpSs
arguments_tree = fill_outline_arguments_tree;
fill_arguments = &fill_outline_arguments;
}
// Now attach the slicer symbol to its final scope (see tl-lower-for-worksharing.cpp)
const decl_context_t* spawn_inner_context = arguments_tree.retrieve_context().get_decl_context();
slicer_descriptor.get_internal_symbol()->decl_context = spawn_inner_context;
::insert_entry(spawn_inner_context->current_scope, slicer_descriptor.get_internal_symbol());
// Parse the arguments
Nodecl::NodeclBase new_tree = fill_arguments->parse_statement(arguments_tree);
arguments_tree.replace(new_tree);
// Finally, replace the construct by the tree that represents the spawn code
construct.replace(spawn_code_tree);
}
BinaryExpression* NodeBuilder::is_less_than_or_equal_to_exp(Expression* arg1,
Expression* arg2)
{
check_numptr_types(arg1->get_result_type(), arg2->get_result_type());
return binary_exp("is_less_than_or_equal_to", get_bool_type(), arg1, arg2);
}
Requires_expr&
Builder::make_requires(Decl_list const& tps, Decl_list const& ps, Req_list const& rs)
{
return make<Requires_expr>(get_bool_type(), tps, ps, rs);
}
// Make a concept check. The type is bool.
Check_expr&
Builder::make_check(Concept_decl& d, Term_list const& as)
{
return make<Check_expr>(get_bool_type(), d, as);
}
Boolean_expr&
Builder::get_bool(bool b)
{
return make<Boolean_expr>(get_bool_type(), b);
}
expr2tc
goto_symext::symex_mem(
const bool is_malloc,
const expr2tc &lhs,
const sideeffect2t &code)
{
if (is_nil_expr(lhs))
return expr2tc(); // ignore
// size
type2tc type = code.alloctype;
expr2tc size = code.size;
bool size_is_one = false;
if (is_nil_expr(size))
size_is_one=true;
else
{
cur_state->rename(size);
mp_integer i;
if (is_constant_int2t(size) && to_constant_int2t(size).as_ulong() == 1)
size_is_one = true;
}
if (is_nil_type(type))
type = char_type2();
else if (is_union_type(type)) {
// Filter out creation of instantiated unions. They're now all byte arrays.
size_is_one = false;
type = char_type2();
}
unsigned int &dynamic_counter = get_dynamic_counter();
dynamic_counter++;
// value
symbolt symbol;
symbol.base_name = "dynamic_" + i2string(dynamic_counter) +
(size_is_one ? "_value" : "_array");
symbol.name = "symex_dynamic::" + id2string(symbol.base_name);
symbol.lvalue = true;
typet renamedtype = ns.follow(migrate_type_back(type));
if(size_is_one)
symbol.type=renamedtype;
else
{
symbol.type=typet(typet::t_array);
symbol.type.subtype()=renamedtype;
symbol.type.size(migrate_expr_back(size));
}
symbol.type.dynamic(true);
symbol.mode="C";
new_context.add(symbol);
type2tc new_type;
migrate_type(symbol.type, new_type);
address_of2tc rhs_addrof(get_empty_type(), expr2tc());
if(size_is_one)
{
rhs_addrof.get()->type = get_pointer_type(pointer_typet(symbol.type));
rhs_addrof.get()->ptr_obj = symbol2tc(new_type, symbol.name);
}
else
{
type2tc subtype;
migrate_type(symbol.type.subtype(), subtype);
expr2tc sym = symbol2tc(new_type, symbol.name);
expr2tc idx_val = zero_ulong;
expr2tc idx = index2tc(subtype, sym, idx_val);
rhs_addrof.get()->type =
get_pointer_type(pointer_typet(symbol.type.subtype()));
rhs_addrof.get()->ptr_obj = idx;
}
expr2tc rhs = rhs_addrof;
expr2tc ptr_rhs = rhs;
if (!options.get_bool_option("force-malloc-success")) {
symbol2tc null_sym(rhs->type, "NULL");
sideeffect2tc choice(get_bool_type(), expr2tc(), expr2tc(), std::vector<expr2tc>(), type2tc(), sideeffect2t::nondet);
rhs = if2tc(rhs->type, choice, rhs, null_sym);
replace_nondet(rhs);
ptr_rhs = rhs;
}
if (rhs->type != lhs->type)
rhs = typecast2tc(lhs->type, rhs);
cur_state->rename(rhs);
expr2tc rhs_copy(rhs);
guardt guard;
symex_assign_rec(lhs, rhs, guard);
pointer_object2tc ptr_obj(pointer_type2(), ptr_rhs);
track_new_pointer(ptr_obj, new_type);
dynamic_memory.push_back(allocated_obj(rhs_copy, cur_state->guard, !is_malloc));
return rhs_addrof->ptr_obj;
}
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;
}
}
