void OpenMPTransform::parallel_for_postorder(PragmaCustomConstruct parallel_for_construct)
{
// One more parallel seen
num_parallels++;
// Decrease the parallel nesting level
parallel_nesting--;
// Get the enclosing function definition
FunctionDefinition function_definition = parallel_for_construct.get_enclosing_function();
Scope function_scope = function_definition.get_scope();
IdExpression function_name = function_definition.get_function_name();
// This was computed in the preorder
ObjectList<Symbol>& shared_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("shared_references");
ObjectList<Symbol>& private_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("private_references");
ObjectList<Symbol>& firstprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("firstprivate_references");
ObjectList<Symbol>& lastprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("lastprivate_references");
ObjectList<OpenMP::ReductionSymbol>& reduction_references =
parallel_for_construct.get_data<ObjectList<OpenMP::ReductionSymbol> >("reduction_references");
ObjectList<Symbol>& copyin_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("copyin_references");
ObjectList<Symbol>& copyprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("copyprivate_references");
// Get the construct_body of the statement
Statement construct_body = parallel_for_construct.get_statement();
// The construct is in fact a ForStatement in a #pragma omp parallel do
ForStatement for_statement(construct_body);
Statement loop_body = for_statement.get_loop_body();
// Create the replacement map and the pass_by_pointer set
ObjectList<ParameterInfo> parameter_info_list;
ReplaceIdExpression replace_references =
set_replacements(function_definition,
loop_body,
shared_references,
private_references,
firstprivate_references,
lastprivate_references,
reduction_references,
inner_reductions_stack.top(),
copyin_references,
copyprivate_references,
parameter_info_list);
// Get the outline function name
Source outlined_function_name = get_outlined_function_name(function_name);
// Create the outline for parallel for
AST_t outline_code = get_outline_parallel_for(
parallel_for_construct,
function_definition,
outlined_function_name,
for_statement,
loop_body,
replace_references,
parameter_info_list,
private_references,
firstprivate_references,
lastprivate_references,
reduction_references,
copyin_references,
copyprivate_references);
// Now prepend the outline
function_definition.get_ast().prepend_sibling_function(outline_code);
PragmaCustomClause if_clause = parallel_for_construct.get_clause("if");
PragmaCustomClause num_threads = parallel_for_construct.get_clause("num_threads");
PragmaCustomClause groups_clause = parallel_for_construct.get_clause("groups");
Source instrument_code_before;
Source instrument_code_after;
AST_t spawn_code = get_parallel_spawn_code(
parallel_for_construct.get_ast(),
function_definition,
parallel_for_construct.get_scope(),
parallel_for_construct.get_scope_link(),
parameter_info_list,
reduction_references,
if_clause,
num_threads,
groups_clause,
instrument_code_before,
instrument_code_after
);
// Discard inner reduction information
inner_reductions_stack.pop();
// Replace all the whole construct with spawn_code
parallel_for_construct.get_ast().replace(spawn_code);
}
void OpenMPTransform::parallel_for_postorder(OpenMP::ParallelForConstruct parallel_for_construct)
{
// One more parallel seen
num_parallels++;
// Remove the induction var from the stack
induction_var_stack.pop();
// Decrease the parallel nesting level
parallel_nesting--;
// Get the directive
OpenMP::Directive directive = parallel_for_construct.directive();
// Get the enclosing function definition
FunctionDefinition function_definition = parallel_for_construct.get_enclosing_function();
Scope function_scope = function_definition.get_scope();
IdExpression function_name = function_definition.get_function_name();
// This was computed in the preorder
ObjectList<Symbol>& shared_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("shared_references");
ObjectList<Symbol>& private_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("private_references");
ObjectList<Symbol>& firstprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("firstprivate_references");
ObjectList<Symbol>& lastprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("lastprivate_references");
ObjectList<OpenMP::ReductionSymbol>& reduction_references =
parallel_for_construct.get_data<ObjectList<OpenMP::ReductionSymbol> >("reduction_references");
ObjectList<Symbol>& copyin_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("copyin_references");
ObjectList<Symbol>& copyprivate_references =
parallel_for_construct.get_data<ObjectList<Symbol> >("copyprivate_references");
// Get the construct_body of the statement
Statement construct_body = parallel_for_construct.body();
// The construct is in fact a ForStatement in a #pragma omp parallel do
ForStatement for_statement(construct_body);
Statement loop_body = for_statement.get_loop_body();
// Create the replacement map and the pass_by_pointer set
ObjectList<ParameterInfo> parameter_info_list;
ReplaceIdExpression replace_references =
set_replacements(function_definition,
directive,
loop_body,
shared_references,
private_references,
firstprivate_references,
lastprivate_references,
reduction_references,
inner_reductions_stack.top(),
copyin_references,
copyprivate_references,
parameter_info_list);
// Get the outline function name
Source outlined_function_name = get_outlined_function_name(function_name);
// Create the outline for parallel for
AST_t outline_code = get_outline_parallel_for(
parallel_for_construct,
function_definition,
outlined_function_name,
for_statement,
loop_body,
replace_references,
parameter_info_list,
private_references,
firstprivate_references,
lastprivate_references,
reduction_references,
copyin_references,
copyprivate_references,
directive);
// Now prepend the outline
function_definition.get_ast().prepend_sibling_function(outline_code);
OpenMP::Clause num_threads = directive.num_threads_clause();
OpenMP::CustomClause groups_clause = directive.custom_clause("groups");
Source instrument_code_before;
Source instrument_code_after;
if (instrumentation_requested())
{
instrument_code_before
<< "const int EVENT_PARALLEL = 60000001;"
<< "const int VALUE_PARALLEL_FOR = 1;"
<< "mintaka_event(EVENT_PARALLEL, VALUE_PARALLEL_FOR);"
<< "mintaka_state_schedule();"
;
instrument_code_after
<< "const int VALUE_PARALLEL_CLOSE = 0;"
<< "mintaka_event(EVENT_PARALLEL, VALUE_PARALLEL_CLOSE);"
<< "mintaka_state_run();"
;
}
AST_t spawn_code = get_parallel_spawn_code(
parallel_for_construct.get_ast(),
function_definition,
parallel_for_construct.get_scope(),
parallel_for_construct.get_scope_link(),
parameter_info_list,
reduction_references,
num_threads,
groups_clause,
instrument_code_before,
instrument_code_after
);
// Discard inner reduction information
inner_reductions_stack.pop();
// Replace all the whole construct with spawn_code
parallel_for_construct.get_ast().replace(spawn_code);
}
