/*
$Id$
Copyright 1989-2016 MINES ParisTech
This file is part of PIPS.
PIPS is free software: you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
any later version.
PIPS is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE.
See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with PIPS. If not, see .
*/
// do not compile unless required
#include "phases.h"
#if defined(BUILDER_SCALARIZATION) || \
defined(BUILDER_CONSTANT_ARRAY_SCALARIZATION) || \
defined(BUILDER_QUICK_SCALARIZATION)
#ifdef HAVE_CONFIG_H
#include "pips_config.h"
#endif
/* package scalarization
*
* Substitute array references by scalar references when legal and
* profitable in loop bodies. The key function is loop_scalarization().
*
* The legality condition is based on a linear algebra function to
* decide if a region function defined a function from any one
* iteration towards an array element.
*
* Because the C3 linear library cannot create new entities, i.e. new
* dimensions, the c_functional_graph_p() cannot have a clean
* interface.
*
* Nevertheless, some functions should be moved in linear.
*/
#include
#include
#include "genC.h"
#include "linear.h"
#include "misc.h"
#include "pipsdbm.h"
#include "properties.h"
#include "ri.h"
#include "ri-util.h"
#include "prettyprint.h" // for debugging
#include "effects-util.h" // used
#include "effects-simple.h" // used
#include "control.h" // module_reorder
#include "text-util.h" // used
#include "transformer.h" // used
#include "semantics.h" // used
#include "effects-generic.h" // used
#include "effects-convex.h" // used
#include "dg.h"
/* instantiation of the dependence graph */
typedef dg_arc_label arc_label;
typedef dg_vertex_label vertex_label;
#include "graph.h"
//////////////////// Properties
static int scalarization_across_control_test_level = 0;
static bool scalarization_across_control_test_is_exactness_p()
{
return (scalarization_across_control_test_level == 0);
}
static bool scalarization_across_control_test_is_strict_test_p()
{
return (scalarization_across_control_test_level == 1);
}
static void scalarization_across_control_test_level_init()
{
const char * test_s =
get_string_property("SCALARIZATION_ACROSS_CONTROL_TEST");
if (strstr(test_s, "strict"))
scalarization_across_control_test_level = 1;
else if (strstr(test_s, "cheap"))
scalarization_across_control_test_level = 2;
else
scalarization_across_control_test_level = 0;
}
static void scalarization_across_control_test_level_reset()
{
scalarization_across_control_test_level = 0;
}
////////////////////
/*
* WARNING ! This function introduces side-effects on g
*
* MA: It probably shouldn't be located here (seems very "linear" oriented)
*/
static Psysteme sc_add_offset_variables(Psysteme g, Pbase b, Pbase db) {
Pbase gb = sc_base(g);
/* check validity conditions:
- b must appear in g's basis but db must not, and
- b and db must have the same dimension.
*/
if (base_included_p(b,gb)
&& !base_included_p(db,gb)
&& base_dimension(b)==base_dimension(db)
) {
Pcontrainte eqs = sc_egalites(g);
Pcontrainte ineqs = sc_inegalites(g);
Pcontrainte c = CONTRAINTE_UNDEFINED;
/* update g's basis */
Pbase tb = sc_base(g);
sc_base(g) = base_union(tb, db);
sc_dimension(g) += base_dimension(db);
base_rm(tb);
for(c = eqs ; !CONTRAINTE_UNDEFINED_P(c) ; c = contrainte_succ(c)) {
Pbase cb = BASE_NULLE;
Pbase cdb = BASE_NULLE;
for (cb=b,cdb=db ;
!BASE_NULLE_P(cb) ;
cb = vecteur_succ(cb), cdb = vecteur_succ(cdb)) {
Variable v = vecteur_var(cb);
Variable dv = vecteur_var(cdb);
Value coeff = vect_coeff(v, contrainte_vecteur(c));
vect_add_elem(&(contrainte_vecteur(c)), dv, coeff);
}
}
for(c = ineqs ; !CONTRAINTE_UNDEFINED_P(c) ; c= contrainte_succ(c)) {
Pbase cb = BASE_NULLE;
Pbase cdb = BASE_NULLE;
for (cb=b,cdb=db ;
!BASE_NULLE_P(cb) ;
cb = vecteur_succ(cb), cdb = vecteur_succ(cdb)) {
Variable v = vecteur_var(cb);
Variable dv = vecteur_var(cdb);
Value coeff = vect_coeff(v, contrainte_vecteur(c));
vect_add_elem(&(contrainte_vecteur(c)), dv, coeff);
}
}
}
return g;
}
/*
This function checks that graph g is a function graph from domain d
to range r.
Return value :
- true if the graph is certainly functional,
- false if it *might* be not functional.
NOTE: parameter dr could be derived from r within this function,
but we do not know how to create generic variables in Linear. The
initial implementation used a PIPS function to generate the new
variables.
*
* MA: It probably shouldn't be located here (seems very "linear" oriented)
*/
static bool sc_functional_graph_p(Psysteme g, Pbase d, Pbase r, Pbase dr)
{
bool functional_p = true;
Psysteme g1, g2;
// check validity conditions: d and r should be included in g's basis.
Pbase gb = sc_base(g);
if (!(base_included_p(d, gb) && base_included_p(r, gb))) {
// illegal arguments
functional_p = false; //TODO
}
else {
Pbase cr = BASE_UNDEFINED;
// Create two copies of g
g1 = sc_copy(g);
g2 = sc_copy(g);
// Substitute r by r + dr in g2 (indexed by dimension)
g2 = sc_add_offset_variables(g2, r, dr);
// Merge g1 and g2 into a unique system g1, eliminating redundencies.
g1 = sc_append(g1, g2);
g1 = sc_elim_redond(g1);
// Project g1 on dr space. If projection fails, return FALSE.
Pbase dr4 = BASE_NULLE;
// dr4 := list of variables of g1's basis which are not in dr
for ( cr = sc_base(g1) ; !BASE_UNDEFINED_P(cr) ; cr = vecteur_succ(cr) ) {
Variable vr = vecteur_var(cr);
if (!base_find_variable(dr, vr))
dr4 = base_add_variable(dr4, vr);
}
sc_projection_along_variables_ofl_ctrl(&g1, dr4, OFL_CTRL);
base_rm(dr4);
/*
ifdebug(1) {
pips_debug(1, "g1 =\n");
sc_print(g1, (get_variable_name_t)entity_local_name);
}
*/
if (SC_EMPTY_P(g1) || sc_empty_p(g1)) {
functional_p = false;
}
else {
// Check that all r_b_i variables are null, using sc_minmax_of_variables()
for ( cr = dr ; !BASE_UNDEFINED_P(cr) ; cr = vecteur_succ(cr) ) {
Psysteme g1b = sc_copy(g1);
Variable dv = vecteur_var(cr);
Value pmin, pmax;
bool feasible_p = sc_minmax_of_variable(g1b,dv, &pmin, &pmax);
if (!(feasible_p
&& value_eq(VALUE_ZERO, pmin)
&& value_eq(VALUE_ZERO, pmax))) {
functional_p = false;
break;
}
}
}
// Cleanup
sc_rm(g1);
sc_rm(g2);
}
return functional_p;
}
/*
This function checks that graph g is a total function graph, i.e. a
mapping graph,from domain d to range r.
*
* MA: It probably shouldn't be located here (seems very "linear" oriented)
*/
static bool sc_totally_functional_graph_p( Psysteme g, // function graph
Pbase d, // domain's basis
Psysteme D, // membership predicate for functional domain
Pbase r, // range's predicate
Pbase dr // difference variable
)
{
bool totally_functional_p = false;
if (sc_functional_graph_p(g, d, r, dr)) {
// Check args coherence : d should be included in D's basis.
if (base_included_p(d, sc_base(D))) {
// Project g on d along r.
Psysteme g1 = sc_copy(g);
sc_projection_along_variables_ofl_ctrl(&g1, r, OFL_CTRL);
// By definition of a total function, D must be included in g1.
totally_functional_p = sc_inclusion_p_ofl_ctrl(D, g1, OFL_CTRL);
}
}
return totally_functional_p;
}
// We need a Pbase to accumulate loop indices during loop traversal
static Pbase loop_indices_b = BASE_NULLE;
// These are needed for callback function reference_substitute
struct subst_ref_ctx {
reference scalarized_reference;
entity scalarized_replacement_variable;
};
// gen_recurse callback function for
// statement_substitute_scalarized_array_references
static bool reference_substitute(reference r, struct subst_ref_ctx *ctx) {
bool result = true;
entity v = reference_variable(r);
entity scalarized_v = reference_variable(ctx->scalarized_reference);
// This test proved to strong by conditional04 because the region
// computation does too good a job...
// if(reference_equal_p(r,ctx->scalarized_reference)) {
if(v==scalarized_v) {
// Scalarize only if r refers to an array element and not to a slice
list inds = reference_indices(r);
size_t d = type_depth(ultimate_type(entity_type(v)));
if (gen_length(inds) == d) {
reference_variable(r) = ctx->scalarized_replacement_variable;
reference_indices(r) = NIL; // TODO: add missing gen_full_free_list(reference_indices(r))
result = false; /* do not recurse in reference indices which are discarded */
}
}
return result;
}
static bool declarations_reference_substitute(statement st,
struct subst_ref_ctx *ctx) {
if (declaration_statement_p(st)) {
FOREACH(ENTITY, decl, statement_declarations(st)) {
value init_val = entity_initial(decl);
if (! value_undefined_p(init_val)) {
gen_context_recurse(init_val, ctx,
reference_domain, reference_substitute, gen_null2);
}
}
}
return true;
}
static void statement_substitute_scalarized_array_references(statement st,
reference pvr,
entity s) {
//TODO: create context ansd use gen_multi_recurse_with_context
struct subst_ref_ctx ctx;
ctx.scalarized_reference = pvr;
ctx.scalarized_replacement_variable = s;
gen_context_multi_recurse(st,&ctx,
statement_domain, declarations_reference_substitute, gen_null,
reference_domain, reference_substitute, gen_null,
NULL);
}
static void * car_effect_to_variable(gen_chunk car) {
return effect_variable(EFFECT(car)); // type 'entity'
}
static Pbase make_phi_base(int phi_min, int phi_max) {
Pbase phi_b = BASE_NULLE;
int i;
for(i=phi_min; i<=phi_max; i++) {
phi_b = base_add_variable(phi_b, (Variable) make_phi_entity(i));
}
return(phi_b);
}
static effect effect_write_or_read_on_variable(list el,
entity v,
bool write_p) {
effect result = effect_undefined;
if(v) {
FOREACH(EFFECT, e, el) {
if (store_effect_p(e)) {
action a = effect_action(e);
entity ev = effect_entity(e);
if(same_entity_p(ev, v)
&& (write_p ? action_write_p(a) : action_read_p(a))) {
return(e);
}
}
}
}
return result;
}
/* Check that the region function r in pr is constant when the store
* sigma is modified by transformer t. You want for all sigma,
*
* r(sigma) = r(t(sigma))
*
* which can be rewritten
*
* {dphi | \exists sigma \exists sigma' \exists phi t.q.
* t(sigma,sigma')
* ^ r(sigma,phi)
* ^ r(sigma',phi+dphi)} = {0}
*
* This criterion is for constant singleton region in context...
*/
static bool __attribute__ ((__unused__)) constant_region_in_context_p(effect pr,
transformer t __attribute__ ((__unused__)))
{
entity v = effect_variable(pr);
int nd = type_depth(entity_type(v));
bool constant_p = (nd==0);
if(!constant_p) {
descriptor d = effect_descriptor(pr);
if (descriptor_convex_p(d)) {
ifdebug(2) {
pips_debug(2,"Considering regions : ");
print_region(pr);
}
/* Try a simple test first: none of the arguments in t appears
* in a constraint of pr */
Psysteme sc = descriptor_convex(d);
list args = transformer_arguments(t);
Pbase eb = sc_to_minimal_basis(sc);// effective base of sc
constant_p = true;
FOREACH(ENTITY, a, args) {
Variable vp = base_find_variable(eb, (Variable) a);
if(vp == (Variable) a) {
constant_p = false;
break;
}
}
if(!constant_p) {
/* Here, we should have a more sophisticated test... */
;
// Clean-up
//base_rm(phi_b);
//base_rm(d_phi_b);
}
pips_debug(2, "returning: %s\n", bool_to_string(constant_p));
}
}
return constant_p;
}
/* Check that region r has a unique element with respect to the subset
of the space defined by basis loop_indices_p and by the predicate
of precondition p. In other words, the relationship between the
iterations i and the array element phi is a function, by definition
of a function.
This cannot be checked without auxiliary variables. It is not
possible to create PIPS compatible variables in Linear. Hence, they
are created here.
The proof is based on r(i)=phi and r(i)=phi+dphi and r is a
function implies dphi=0. If dphi!=0, then r cannot be a function.
But we also want that each iteration uses at least one element. The
function must be a total function. FI: I do not remember why we
need a total function.
*/
static bool region_totally_functional_graph_p(effect pr,
Pbase loop_indices_b,
transformer p) {
bool rtfg_p = false;
descriptor d= effect_descriptor(pr);
pips_assert("d is a convex descriptot", descriptor_convex_p(d));
Psysteme sc_union = descriptor_convex(d);
entity pv = effect_variable(pr);
int nd = type_depth(entity_type(pv));
Psysteme D = predicate_system(transformer_relation(p));
Pbase phi_b = make_phi_base(1, nd);
Pbase d_phi_b = BASE_NULLE;
Pbase cr = BASE_NULLE;
/* Build base d_phi_b using
* make_local_temporary_integer_value_entity(void) */
for ( cr = phi_b ; !BASE_UNDEFINED_P(cr) ; cr = vecteur_succ(cr) ) {
entity e_d_phi_b = make_local_temporary_integer_value_entity();
d_phi_b = base_add_variable(d_phi_b, (Variable) e_d_phi_b);
}
rtfg_p = sc_totally_functional_graph_p(sc_union, loop_indices_b,
D, phi_b, d_phi_b);
base_rm(phi_b);
base_rm(d_phi_b);
return rtfg_p;
}
static bool singleton_region_in_context_p(effect pr,
transformer prec, // a range
Pbase loop_indices_b, // domain
bool strict_p) {
bool singleton_p = false;
// Check if the referenced variable is scalar or array
// D is not needed if we are interested in a partial function even
// over D
// Psysteme D = predicate_system(transformer_relation(prec));
entity pv = effect_variable(pr);
int nd = type_depth(entity_type(pv));
descriptor d = effect_descriptor(pr);
Psysteme sc_union = SC_UNDEFINED;
if (descriptor_convex_p(d)) {
sc_union = descriptor_convex(d);
} else {
pips_internal_error("convex region expected\n");
}
if(!sc_empty_p(sc_union)) {
Pbase phi_b = make_phi_base(1, nd);
Pbase d_phi_b = BASE_NULLE;
Pbase cr = BASE_NULLE;
/* Build base d_phi_b using
* make_local_temporary_integer_value_entity(void) */
for ( cr = phi_b ; !BASE_UNDEFINED_P(cr) ; cr = vecteur_succ(cr) ) {
entity e_d_phi_b = make_local_temporary_integer_value_entity();
d_phi_b = base_add_variable(d_phi_b, (Variable) e_d_phi_b);
}
if(strict_p) {
Psysteme D = predicate_system(transformer_relation(prec));
singleton_p = sc_totally_functional_graph_p(sc_union, loop_indices_b,
D, phi_b, d_phi_b);
} else {
// A loop may be entered or not, so the region is not a total
//function, Maybe the name of this function should be changed
singleton_p = sc_functional_graph_p(sc_union, loop_indices_b,
phi_b, d_phi_b);
}
// Clean-up
base_rm(phi_b);
base_rm(d_phi_b);
}
return singleton_p;
}
/* Generate a scalar variable sv to replace all references to pv in
s. If iv or ov are defined, generate a copy-in and a copy-out
statements. */
static void scalarize_variable_in_statement(entity pv,
statement s,
entity iv,
entity ov) {
// Create new temp var of same type as the array element pv
type pvt = ultimate_type(entity_type(pv));
variable pvtv = type_variable(pvt);
basic pvb = variable_basic(pvtv);
basic svb = copy_basic(pvb);
// Copy the a reference to pv, just in case we need it later
reference pvr = copy_reference(find_reference_to_variable(s, pv));
if (reference_undefined_p(pvr)) { /* may happen with quick_scalarization */
return;
}
ifdebug(3){
pips_debug(3, "begin for entity %s and statement:\n",
entity_name(pv));
print_statement(s);
}
// Create a new variable and add
// its declaration to the current module
// If no default prefix is defined, use the variable name
// as prefix
const char* dpref = get_string_property("SCALARIZATION_PREFIX");
const char* epref = strlen(dpref)==0?
concatenate(entity_user_name(pv), "_", NULL)
: dpref;
entity sv = make_new_scalar_variable_with_prefix(epref,
get_current_module_entity(),
svb);
/* FI: the language could be checked, but Fortran prettyprinter
ignores the qualifier and the qualifier benefits the complexity
analyzer. */
if (get_bool_property("SCALARIZATION_USE_REGISTERS"))
set_register_qualifier(sv);
entity m = get_current_module_entity();
if(get_bool_property("SCALARIZATION_PRESERVE_PERFECT_LOOP_NEST")) {
// declaration at module level
AddEntityToCurrentModule(sv);
} else {
// Declare the scalar variable as local as possible
// but at the beginning of the current statement if it is a block,
// since scalarized references may happen in declarations initializations
if (statement_sequence_p(s)) {
if (c_module_p(m)) {
add_declaration_statement_at_beginning(s, sv);
}
AddLocalEntityToDeclarationsOnly(sv, m, s);
} else {
AddLocalEntityToDeclarations(sv, m, s);
}
}
pips_debug(1,"Creating variable %s for variable %s\n",
entity_name(sv), entity_name(pv));
// Substitute all references to pv with references to new variable
statement_substitute_scalarized_array_references(s, pvr, sv);
ifdebug(3){
pips_debug(3, "statement after substitution by sv (%s):\n",
entity_name(sv));
print_statement(s);
}
// Take care of copy-in and copy-out code if necessary
if (get_bool_property("SCALARIZATION_FORCE_OUT")
|| !entity_undefined_p(ov)) {
// Generate copy-out code
statement co_s = make_assign_statement(reference_to_expression(pvr),
entity_to_expression(sv));
insert_statement(s, co_s, false);
} else {
//free_reference(pvr);
}
if (!entity_undefined_p(iv)) {
if(fortran_language_module_p(m)
|| get_bool_property("SCALARIZATION_PRESERVE_PERFECT_LOOP_NEST")) {
// Generate copy-in code with an assignment statement:
statement ci_s =
make_assign_statement(entity_to_expression(sv),
reference_to_expression(copy_reference(pvr)));
insert_statement(s, ci_s, true);
} else if(c_language_module_p(m)) {
// in a less intrusive way?
expression ie = reference_to_expression(copy_reference(pvr));
entity_initial(sv) = make_value_expression(ie);
// The initialization is not necessarily generated at the right
//spot when dealing with loops, or, looking at the problem
//differently, the declaration of the scalarized variable is not
//located properly...
} else {
pips_user_error("Module \"%s\" is not written in Fortran or C.\n"
"Other languages are not supported by "
"the scalarization pass.\n", entity_user_name(m));
}
} else {
//free_reference(pvr);
}
}
/* Union of all effects on pv in list crwl
*
* Note: the usual assumption is made, no more than two effects on pv
* in crwl...
* No, the assumption is that there is no more than two effects
* on a memory access path. BC.
*
* The return effect is given a read action on memory. A read-or-write
* action would be more meaningful.
*/
static effect unified_rw_effect_of_variable(entity pv, list crwl) {
effect pru = effect_undefined;
//int nd = type_depth(entity_type(pv));
Psysteme sc_union = SC_UNDEFINED;
Psysteme sc1 = SC_UNDEFINED;
effect pr1 = effect_undefined;
Psysteme sc2 = SC_UNDEFINED;
effect pr2 = effect_undefined;
tag app1, app2;
// action_write_p(effect_action(pr))))
if ((pr1 = effect_write_or_read_on_variable(crwl, pv, true))
!= effect_undefined) {
descriptor d1 = effect_descriptor(pr1);
app1 = effect_approximation_tag(pr1);
if (descriptor_convex_p(d1)) {
sc1 = descriptor_convex(d1);
}
} else {
app1 = is_approximation_may;
}
//!(action_write_p(effect_action(pr)))))
/* Search Effect with the opposite action */
if ((pr2 = effect_write_or_read_on_variable(crwl, pv, false))
!= effect_undefined) {
descriptor d2= effect_descriptor(pr2);
app2 = effect_approximation_tag(pr2);
if (descriptor_convex_p(d2)) {
sc2 = descriptor_convex(d2);
}
} else {
app2 = is_approximation_may;
}
tag app_t = approximation_or(app1, app2);
// we enforce that oneof the effect is exact to be sure that no accesses
// are unduely moved out of control
if (!scalarization_across_control_test_is_exactness_p()
|| app_t == is_approximation_exact) {
/* Merge Read and Write Effects constraints on pv */
if (!SC_UNDEFINED_P(sc1) && sc_dimension(sc1) !=0) {
if (!SC_UNDEFINED_P(sc2) && sc_dimension(sc2) !=0) {
sc_union=sc_cute_convex_hull(sc2,sc1);
} else {
sc_union=sc_dup(sc1);
}
} else {
if (!SC_UNDEFINED_P(sc2) && sc_dimension(sc2) !=0) {
sc_union=sc_dup(sc2);
} else {
sc_union = SC_UNDEFINED;
}
}
if(!SC_UNDEFINED_P(sc_union)) {
cell c = make_cell_reference(make_reference(pv,NIL));
action a = make_action_read(make_action_kind_store()); // arbitrary choice
approximation ap = make_approximation_may(); // safe default
descriptor d = make_descriptor_convex(sc_union);
pru = make_effect(c, a, ap, d);
}
}
return pru;
}
struct scalarization_ctx {
// To keep track of already scalarized variables
list scalarized_variables;
// To associate a list of privatized variables to each statement
hash_table statement_scalarized_variables;
// Flag that indicate if we should keep perfect parallel loop nest
bool keep_perfect_parallel;
// Loops that should not be scalarized (according to previous property)
set blacklisted_loops;
// Do not scalarize "&vars..."
bool skip_address_of_variables;
};
static void blacklist_perfectly_nested_parallel_loop(statement s,
set blacklisted_loops) {
if(statement_loop(s)) {
int depth = depth_of_parallel_perfect_loop_nest(s);
for(int d=2;d<=depth;d++) {
s=get_first_inner_perfectly_nested_loop(loop_body(statement_loop(s)));
pips_assert("Perfectly nested loops coherency !",
!statement_undefined_p(s));
set_add_element(blacklisted_loops,blacklisted_loops,s);
}
}
}
/* This function has been replaced by statement_scalarization() */
static bool loop_scalarization(loop l, struct scalarization_ctx *ctx)
{
entity i = loop_index(l);
statement s = loop_body(l);
ifdebug(1) {
pips_debug(1, "Statement:\n");
print_statement(s);
}
transformer prec = load_statement_precondition(s);
transformer prec_r = transformer_range(prec);
//Psysteme D = predicate_system(transformer_relation(transformer_range(prec)));
effects ie = load_in_effects(s);
effects oe = load_out_effects(s);
effects crwe = load_cumulated_rw_effects(s);
bool memory_effects_only_p = get_bool_property("MEMORY_EFFECTS_ONLY");
bool memory_in_out_regions_only_p =
get_bool_property("MEMORY_IN_OUT_EFFECTS_ONLY");
list irl = effects_effects(ie);
list orl = effects_effects(oe);
list crwl = effects_effects(crwe);
if (!memory_effects_only_p) {
crwl = effects_store_effects(crwl);
}
if (!memory_in_out_regions_only_p) {
irl = effects_store_effects(irl);
orl = effects_store_effects(orl);
}
ifdebug(1) {
pips_debug(1, "Entering function...\n");
pips_debug(1,
"Entering level-%d loop, index=%s\n",
base_dimension(loop_indices_b), entity_name(i));
pips_debug(1, "OUT regions:\n");
print_regions(orl);
pips_debug(1, "CUMULATED RW regions:\n");
print_regions(crwl);
}
Pvecteur var_already_seen = VECTEUR_NUL;
// Each variable in effects is a candidate for scalarization
FOREACH (EFFECT, pr, crwl) {
// Now we determine which read/write effects are not copied out.
entity pv = effect_variable(pr);
// Does the current variable appear in the in effect?
entity iv = (entity) gen_find(pv,
irl,
(bool(*)()) gen_eq,
car_effect_to_variable);
// Does the current variable appear in the out effect?
entity ov = (entity) gen_find(pv,
orl,
(bool(*)()) gen_eq,
car_effect_to_variable);
descriptor d = effect_descriptor(pr);
if (descriptor_convex_p(d) &&
!entity_is_argument_p(pv, ctx->scalarized_variables)
&& !vect_coeff(pv,var_already_seen)) {
ifdebug(2) {
pips_debug(2,"Considering regions : ");
print_region(pr);
}
vect_add_elem(&var_already_seen,pv,1);
effect pru = unified_rw_effect_of_variable(pv, crwl);
int nd = type_depth(entity_type(pv));
if (!effect_undefined_p(pru)) {
ifdebug(2) {
pips_debug(2,"pru not undefined: ");
print_region(pru);
}
// Estimate the dynamic number of *element* and *variable*
// occurrences in the loop body
int neo = count_references_to_variable_element(s, pv);
int nvo = count_references_to_variable(s, pv);
/* Legality criterion:
if nvo is greater than neo, there must be hidden references
to the array, due for instance to a function call, and the
substitution might break dependence arcs.
So, we go on only if the two are equal.
We assume that array variables cannot be declared volatile
*/
if (nvo != neo) {
ifdebug(2) {
pips_debug(2,
"Legality criterion not met: %d!=%d (nvo!=neo)\n",
nvo, neo);
}
} else {
bool read_pv = effects_read_variable_p(crwl, pv);
bool written_pv = effects_write_variable_p(crwl, pv);
bool read_and_written_pv = read_pv && written_pv;
/* Profitability criterion:
- nd > 0: it's an array (do not scalarize scalar variables)
- neo > 2: if the number of references if greater than 2, the
copy-in and copy-out code overhead is assumed to be small
enough to make scalarization profitable.
- neo > 1: if the number of references is 2, the copy-in *xor*
the copy-out overhead meets the above criterion.
- else: if there is neither copy-in nor copy-out,
privatization is always useful.
FI: we should also check if the reference is loop invariant
and then decide that it is always profitable to scalarize
it... See scalarization34 in Transformations. We should
check if the region is a constant function.
*/
if (nd <= 0 // Scalar
|| (neo <= 2 && // Two or less references
(neo < 2 || read_and_written_pv) && // ...
(!entity_undefined_p(iv) || !entity_undefined_p(ov)) // At least copy in or copy out !
)
) {
ifdebug(2) {
pips_debug(2,"Profitability criterion not met: (nd) %d>0 (scalar) "
"or not one of the following: (%d&&%d&&%d) "
"(neo) %d <= 2 and "
"((neo) %d <= 1 || %d read_and_written_pv) and "
"(%d (!entity_undefined_p(iv)) || %d (!entity_undefined_p(ov)))\n",
nd,
neo <= 2,
(neo <= 1 || read_and_written_pv),
(!entity_undefined_p(iv) || !entity_undefined_p(ov)),
neo,
neo,
!read_and_written_pv,
entity_undefined_p(iv),
entity_undefined_p(ov));
}
} else {
if(region_totally_functional_graph_p(pru,
loop_indices_b,
prec_r)) {
/* The array references can be replaced by references to a
scalar */
scalarize_variable_in_statement(pv, s, iv, ov);
ctx->scalarized_variables = arguments_add_entity(ctx->scalarized_variables,
pv);
}
// Clean-up
reset_temporary_value_counter();
}
}
}
//sc_rm(sc_union);
free_effect(pru);
}
}
vect_rm(var_already_seen);
// Do not leak
if (!memory_in_out_regions_only_p) {
gen_free_list(orl);
gen_free_list(irl);
}
if (!memory_effects_only_p) {
gen_free_list(crwl);
}
return true;
}
typedef struct {
bool constant_p;
entity e;
list trans_args;
} reference_testing_ctxt;
static bool reference_constant_wrt_ctxt_p(reference ref,
reference_testing_ctxt *ctxt) {
bool continue_p = true;
if (reference_variable(ref) == ctxt->e) {
FOREACH(EXPRESSION, exp, reference_indices(ref)) {
list l_eff_exp = proper_effects_of_expression(exp);
FOREACH(EFFECT, eff_exp, l_eff_exp) {
entity e_exp = effect_entity(eff_exp);
/* if the indice contains a reference to an array element, we assume it
* is not constant
* we could use the cumulated effects of the current statement for more
* precision but as we could not scalarize the array because the region
* would not contain a single element, it's of no use here.
*/
if (!entity_scalar_p(e_exp)) {
ctxt->constant_p = false;
} else if (entity_in_list_p(e_exp, ctxt->trans_args)) {
ctxt->constant_p = false;
}
}
gen_full_free_list(l_eff_exp);
if (!ctxt->constant_p) {
break;
}
}
continue_p = ctxt->constant_p;
}
return continue_p;
}
static bool declarations_reference_constant_wrt_ctxt_p(statement st,
reference_testing_ctxt *p_ctxt) {
if (declaration_statement_p(st)) {
FOREACH(ENTITY, decl, statement_declarations(st)) {
value init_val = entity_initial(decl);
if (! value_undefined_p(init_val)) {
gen_context_recurse(init_val, p_ctxt,
reference_domain, reference_constant_wrt_ctxt_p, gen_null2);
}
}
}
return true;
}
static bool statement_entity_references_constant_in_context_p(statement s,
entity e,
list l_modified_variables) {
int nd = type_depth(entity_type(e));
bool constant_p = (nd==0);
if(!constant_p) {
pips_debug(2,"Considering entity : %s\n", entity_name(e));
/* none of the modified variables appears in an index of a reference from entity e */
reference_testing_ctxt ctxt;
ctxt.constant_p = true;
ctxt.e = e;
ctxt.trans_args = l_modified_variables;
gen_context_multi_recurse(s, &ctxt,
statement_domain, declarations_reference_constant_wrt_ctxt_p, gen_null,
reference_domain, reference_constant_wrt_ctxt_p, gen_null,
NULL);
constant_p = ctxt.constant_p;
pips_debug(2, "returning: %s\n", bool_to_string(constant_p));
}
return constant_p;
}
typedef struct {
entity e;
bool used_through_external_calls;
} calls_test_ctxt;
static bool entity_accessed_in_call_in(call c, calls_test_ctxt *ctxt)
{
entity func = call_function(c);
type uet = ultimate_type(entity_type(func));
if(type_functional_p(uet) && value_code_p(entity_initial(func)))
{
// get the summary effects of the function
const char *func_name = module_local_name(func);
list func_eff = effects_to_list((effects) db_get_memory_resource(DBR_SUMMARY_REGIONS,
func_name,
true));
/* tests if the function may refer to the global variable */
list l_conflicts = effects_entities_which_may_conflict_with_scalar_entity(func_eff, ctxt->e);
if (!ENDP(l_conflicts))
{
ctxt->used_through_external_calls = true;
gen_free_list(l_conflicts);
}
}
return (!ctxt->used_through_external_calls);
}
static bool entity_accessed_in_statement_declarations_calls_in(statement s, calls_test_ctxt * ctxt)
{
if (declaration_statement_p(s))
{
list l_decls = statement_declarations(s);
FOREACH(ENTITY, e, l_decls)
{
gen_context_recurse(entity_initial(e), &ctxt,
call_domain, entity_accessed_in_call_in, gen_null2);
}
}
return (!ctxt->used_through_external_calls);
}
static bool entity_accessed_through_calls_in_statement_p(entity e, statement s)
{
calls_test_ctxt ctxt;
ctxt.e = e;
ctxt.used_through_external_calls = false;
gen_context_multi_recurse(s, &ctxt,
statement_domain, entity_accessed_in_statement_declarations_calls_in, gen_null,
call_domain, entity_accessed_in_call_in, gen_null, NULL);
return (ctxt.used_through_external_calls);
}
/* Scalarize array references in any kind of statement
*
* The first cut is a cut-and-paste of loop_scalarization()
*/
static bool statement_scalarization(statement s,
struct scalarization_ctx *ctx) {
ifdebug(1) {
pips_debug(1, "Statement:\n");
print_statement(s);
}
transformer tran = load_statement_transformer(s);
transformer prec = load_statement_precondition(s);
// Psysteme D = predicate_system(transformer_relation(transformer_range(prec)));
bool memory_effects_only_p = get_bool_property("MEMORY_EFFECTS_ONLY");
bool memory_in_out_regions_only_p =
get_bool_property("MEMORY_IN_OUT_EFFECTS_ONLY");
effects ie = load_in_effects(s);
effects oe = load_out_effects(s);
effects crwe = load_cumulated_rw_effects(s);
list irl = effects_effects(ie);
list orl = effects_effects(oe);
list crwl = effects_effects(crwe);
if (!memory_effects_only_p) {
crwl = effects_store_effects(crwl);
}
if (!memory_in_out_regions_only_p) {
irl = effects_store_effects(irl);
orl = effects_store_effects(orl);
}
/* List of entities forbidden in an array subscript,
* because modified in the statement
* or because declared deeper in the statement.
*/
set declared_entities = get_declared_entities(s);
FOREACH(entity, e, transformer_arguments(tran)) {
set_add_element(declared_entities,declared_entities,e);
}
list forbiden_variables_in_subscript = set_to_list(declared_entities);
set_free(declared_entities);
/* List of variables than shoud be privatized in s */
list local_scalarized_variables = NIL;
ifdebug(1) {
pips_debug(1, "Entering function...\n");
pips_debug(1, "With statement...\n");
print_statement(s);
pips_debug(1, "OUT regions:\n");
print_regions(orl);
pips_debug(1, "CUMULATED RW regions:\n");
print_regions(crwl);
}
Pvecteur var_already_seen = VECTEUR_NUL;
// Each variable in effects is a candidate for scalarization
FOREACH (EFFECT, pr, crwl) {
// Now we determine which read/write effects are not copied out.
// I'm not sure this works for arrays of structures for instance
// because there may be several paths with different lengths
// from a same entity.
// All the computation is done in reference to the entity pv
// whereas it should be done in reference to a memory access path.
// BC (02/20/2012).
entity pv = effect_variable(pr); // Private variable
int nd = type_depth(entity_type(pv));
if(!entity_is_argument_p(pv, ctx->scalarized_variables) // sclarized at
// a higher level?
&& !vect_coeff(pv,var_already_seen) // Each variable may appear
// several times in the
// effect list
&& nd > 0 // Only array references can be scalarized
&& !entity_volatile_variable_p(pv) // Volatile arrays cannot be scalarized
&& (!top_level_entity_p(pv) || !entity_accessed_through_calls_in_statement_p(pv, s)) // global arrays accessed through calls cannot yet be scalarized
) {
// Does the current variable appear in the in effect?
entity iv = (entity) gen_find(pv,
irl,
(bool(*)()) gen_eq,
car_effect_to_variable);
// Does the current variable appear in the out effect?
entity ov = (entity) gen_find(pv,
orl,
(bool(*)()) gen_eq,
car_effect_to_variable);
descriptor d = effect_descriptor(pr);
if (descriptor_convex_p(d)) {
ifdebug(2) {
pips_debug(2,"Considering regions : ");
print_region(pr);
}
vect_add_elem(&var_already_seen,pv,1);
// scalarization_across_control_test_level is checked inside
// if it is equal to 0 and there is not at least one exact effect
// then an undefined effect is returned.
effect pru = unified_rw_effect_of_variable(pv, crwl);
if (!effect_undefined_p(pru)) { // FI: this should be an assert?
// Estimate the dynamic number of *element* and *variable*
// occurrences in the loop body
int neo = count_references_to_variable_element(s, pv);
int nvo = count_references_to_variable(s, pv);
/* Legality criterion:
if nvo is greater than neo, there must be hidden references
to the array, due for instance to a function call, and the
substitution might break dependence arcs.
So, we go on only if the two are equal.
*/
if (nvo != neo) {
ifdebug(2) {
pips_debug(2,
"Legality criterion not met : %d!=%d (nvo!=neo)\n",
nvo, neo);
}
} else {
bool read_pv = effects_read_variable_p(crwl, pv);
bool written_pv = effects_write_variable_p(crwl, pv);
bool read_and_written_pv = read_pv && written_pv;
int threshold = get_int_property("SCALARIZATION_THRESHOLD");
if(threshold<2) {
pips_user_error("The scalarization threshold should be at "
"least 2\nCheck property SCALARIZATION_THRESHOLD\n");
}
/* Profitability criterion:
- nd > 0: it's an array (do not scalarize scalar variables)
- neo > 2: if the number of references if greater than 2, the
copy-in and copy-out code overhead is assumed to be small
enough to make scalarization profitable.
- neo > 1: if the number of references is 2, the copy-in *xor*
the copy-out overhead meets the above criterion.
- else: if there is neither copy-in nor copy-out,
privatization is always useful.
FI: we should also check if the reference is loop invariant
and then decide that it is always profitable to scalarize
it... See scalarization34 in Transformations. We should
check if the region is a constant function.
*/
if (nd <= 0 // Scalar -> FI: this is already tested above...
|| (neo <= threshold && // Two or fewer references
(neo < threshold || read_and_written_pv) && // ...
(!entity_undefined_p(iv) || !entity_undefined_p(ov)) // At least copy in or copy out !
)
) {
ifdebug(2) {
pips_debug(2,"Profitability criterion not met: (nd) %d>0 (scalar) "
"or not one of the following : (%d&&%d&&%d) "
"(neo) %d <= 2 and "
"((neo) %d <= 1 || %d read_and_written_pv) and "
"(%d (!entity_undefined_p(iv)) || %d (!entity_undefined_p(ov)))\n",
nd,
neo <= 2,
(neo <= 1 || read_and_written_pv),
(!entity_undefined_p(iv) || !entity_undefined_p(ov)),
neo,
neo,
!read_and_written_pv,
entity_undefined_p(iv),
entity_undefined_p(ov));
}
} else {
pips_debug(1,"Profitability criterion met \n");
bool strict_p = scalarization_across_control_test_is_strict_test_p();
/* Check that a unique element of pv is used. Its region
must be constant within the statement, i.e. wrt to the
statement transformer tran.
This last criterion is not strong enough, because variables
modified in the statement have already been eliminated from
regions. We enforce a stronger criterion implemented by
statement_entity_references_constant_in_context_p which checks
that each reference from pru entity has constant indices wrt
to the statement transformer tran and has no complex indices
(array elements, structs). BC.
*/
/* prec might have to be replaced by prec_r, its range. */
/* It is not clear if a unique element of pv should
always be used, wasting some opportunities but
making sure that no non-existing reference is added,
or if at most a unique element is used, which may to
spurious out of bounds accesses. See scalarization30
for an example and explanations. */
if (//constant_region_in_context_p(pru, tran)
statement_entity_references_constant_in_context_p(s,
effect_entity(pru),
forbiden_variables_in_subscript )
&& singleton_region_in_context_p(pru, prec,
loop_indices_b, strict_p)) {
/* The array references can be replaced a references to a
scalar */
// FI: this must be postponed to the bottom up phase
//scalarize_variable_in_statement(pv, s, iv, ov);
local_scalarized_variables =
arguments_add_entity(local_scalarized_variables, pv);
ctx->scalarized_variables =
arguments_add_entity(ctx->scalarized_variables, pv);
}
reset_temporary_value_counter();
}
}
}
//sc_rm(sc_union);
free_effect(pru);
}
}
}
vect_rm(var_already_seen);
/* Associate the current list of privatized variables to statement s */
hash_put(ctx->statement_scalarized_variables,
(void *) s,
(void *) local_scalarized_variables);
// Do not leak
if (!memory_in_out_regions_only_p) {
gen_free_list(orl);
gen_free_list(irl);
}
if (!memory_effects_only_p) {
gen_free_list(crwl);
}
gen_free_list(forbiden_variables_in_subscript);
return true;
}
/* gen_recurse callback on entering a statement. If it's a loop,
* process it. Stack the loop index to build the iteration
* space. Perform loop scalarization.
*/
static bool scalarization_loop_statement_in(statement ls,
struct scalarization_ctx *ctx) {
bool result_p = true;
if (statement_loop_p(ls)) {
loop l = statement_loop(ls);
/* Insert a marker to keep track of the privatized variables
inside loop l, lest they are privatized a second time in an inner
loop.
*/
entity i = loop_index(l);
ctx->scalarized_variables = arguments_add_entity(ctx->scalarized_variables,
i);
// Accumulate new index in "domain" basis
loop_indices_b = base_add_variable(loop_indices_b, (Variable) i);
result_p = loop_scalarization(l,ctx);
}
return result_p;
}
/* gen_recurse callback on exiting a statement. If it's a loop, process it.
* unstack all variables related to loops internal wrt the current
* loop to restore the current iteration space.
*/
static void scalarization_loop_statement_out(statement s,
struct scalarization_ctx *ctx) {
if (statement_loop_p(s)) {
loop l = statement_loop(s);
entity i = loop_index(l);
list nl = NIL;
pips_debug(1, "Exiting loop with index %s, size=%d\n",
entity_name(i), base_dimension(loop_indices_b));
/* Remove variables privatized in the current loop, so that
successive loops do not interfere with each other.
*/
for (list el=ctx->scalarized_variables; !ENDP(el); POP(el)) {
entity e = ENTITY(CAR(el));
if (e == i) {
break;
} else {
nl = CONS(ENTITY, e, nl);
}
}
gen_free_list(ctx->scalarized_variables);
ctx->scalarized_variables = gen_nreverse(nl);
loop_indices_b = base_remove_variable(loop_indices_b, (Variable) i);
}
}
/* gen_recurse callback on entering a statement. If it is not a
* declaration statement process it: keep track of scalarized
* variables.
*/
static bool scalarization_statement_in(statement s,
struct scalarization_ctx *ctx) {
bool result_p = true;
// In fact, only sequences and loops are good candidates... although
// the comma and the assignment operator could also lead to
// some scalarization...
//
// Call statements are not as appropriate for the profitability
// criterion... but this has been fixed with a profitability threshold
//
// This test must be exactly replicated in scalaration_statement_out
if(!declaration_statement_p(s) /*&& !statement_call_p(s)*/
&& !set_belong_p(ctx->blacklisted_loops, s)) {
result_p = statement_scalarization(s,ctx);
if (statement_loop_p(s)) {
loop l = statement_loop(s);
entity i = loop_index(l);
pips_debug(1, "Exiting loop with index %s, size=%d\n",
entity_name(i), base_dimension(loop_indices_b));
loop_indices_b = base_add_variable(loop_indices_b, (Variable) i);
if(ctx->keep_perfect_parallel) {
blacklist_perfectly_nested_parallel_loop(s, ctx->blacklisted_loops);
}
}
}
return result_p;
}
typedef struct {
entity var;
bool found;
set seen;
} aov_ctx;
static void aov_call(call c, aov_ctx * ctx)
{
entity called = call_function(c);
if (ENTITY_ADDRESS_OF_P(called))
{
list args = call_arguments(c);
pips_assert("1 arg to &", gen_length(args) == 1);
if (set_belong_p(ctx->seen, EXPRESSION(CAR(args))))
{
ctx->found = true;
gen_recurse_stop(NULL);
}
}
}
static void aov_cast(cast c, aov_ctx * ctx)
{
expression e = cast_expression(c);
if (set_belong_p(ctx->seen, e))
{
expression enclosing =
(expression) gen_get_current_ancestor(expression_domain);
if (enclosing)
set_add_element(ctx->seen, ctx->seen, enclosing);
}
}
static void aov_ref(reference r, aov_ctx * ctx)
{
if (reference_variable(r) == ctx->var)
{
expression enclosing =
(expression) gen_get_current_ancestor(expression_domain);
if (enclosing)
set_add_element(ctx->seen, ctx->seen, enclosing);
}
}
// return whether "&v" exists in statement s
// this is not really efficient, maybe some data structure could be used
// to keep track of that... but we want in statement s *only*...
static bool address_of_variable_in_statement(statement s, entity v)
{
aov_ctx ctx;
ctx.var = v;
ctx.found = false;
ctx.seen = set_make(set_pointer);
gen_context_multi_recurse(s, &ctx,
call_domain, gen_true, aov_call,
cast_domain, gen_true, aov_cast,
reference_domain, gen_true, aov_ref,
NULL);
set_free(ctx.seen);
return ctx.found;
}
/* gen_recurse callback on exiting a statement. Privatize the
variables collected during the top-down phase */
static void scalarization_statement_out(statement s,
struct scalarization_ctx *ctx) {
// This test must be exactly replicated in scalaration_statement_in
if(!declaration_statement_p(s) /*&& !statement_call_p(s)*/
&& !set_belong_p(ctx->blacklisted_loops,s)) {
/* If statement s is a loop, remove its loop index from basis
loop_indices_b which keep track of loop nesting. Do it before
s is updated by privatization: the added declarations may
require s to be changed into a sequence and the loop moved
down in the AST */
if (statement_loop_p(s)) {
loop l = statement_loop(s);
entity i = loop_index(l);
pips_debug(1, "Exiting loop with index %s, size=%d\n",
entity_name(i), base_dimension(loop_indices_b));
loop_indices_b = base_remove_variable(loop_indices_b, (Variable) i);
}
/* Retrieve the variables to privatize for statement s */
list local_scalarized_variables =
(list) hash_get(ctx->statement_scalarized_variables, (void *) s);
if(!ENDP(local_scalarized_variables)) {
/* The ENDP test is not necessary but it saves time, especially
when using gdb... */
/* Privatize each of them, with copy_in or copy_out when necessary */
bool memory_in_out_regions_only_p =
get_bool_property("MEMORY_IN_OUT_EFFECTS_ONLY");
effects ie = load_in_effects(s);
effects oe = load_out_effects(s);
list irl = effects_effects(ie);
list orl = effects_effects(oe);
if (!memory_in_out_regions_only_p) {
irl = effects_store_effects(irl);
orl = effects_store_effects(orl);
}
ifdebug(3) {
pips_debug(3, "statement before replacements:\n");
print_statement(s);
}
FOREACH(ENTITY, pv, local_scalarized_variables)
{
if (ctx->skip_address_of_variables &&
address_of_variable_in_statement(s, pv))
{
pips_debug(4, "skipping variable %s\n", entity_name(pv));
continue;
}
entity iv = (entity)
gen_find(pv, irl, (bool (*)())gen_eq, car_effect_to_variable);
// Does the current variable appear in the out effect?
entity ov = (entity)
gen_find(pv, orl, (bool (*)())gen_eq, car_effect_to_variable);
scalarize_variable_in_statement(pv, s, iv, ov);
}
ifdebug(3) {
pips_debug(3, "statement on exit:\n");
print_statement(s);
}
/* Remove variables scalarized in s from the list of scalarized
variables so that
successive statements do not interfere with each other. */
ctx->scalarized_variables =
arguments_difference(ctx->scalarized_variables,
local_scalarized_variables);
/* This list is no longer useful. It is still accessible via the
hashtable statement_scalarized_variables but s should not be
visited again. If the lists are not freed one by one, they
should be freed later when the hash_tabe itself is freed */
gen_free_list(local_scalarized_variables);
// Do not leak
if (!memory_in_out_regions_only_p) {
gen_free_list(orl);
gen_free_list(irl);
}
}
}
return;
}
bool scalarization(const string module_name)
{
entity module;
statement module_stat;
set_current_module_entity(module_name_to_entity(module_name));
module = get_current_module_entity();
set_current_module_statement((statement) db_get_memory_resource(DBR_CODE,
module_name,
true));
module_stat = get_current_module_statement();
set_cumulated_rw_effects((statement_effects) db_get_memory_resource(DBR_REGIONS,
module_name,
true));
set_proper_rw_effects((statement_effects) db_get_memory_resource(DBR_REGIONS,
module_name,
true));
module_to_value_mappings(module);
// Used for statement scalarization
set_transformer_map((statement_mapping) db_get_memory_resource(DBR_TRANSFORMERS,
module_name,
true));
// Used for loop scalarization
set_precondition_map((statement_mapping) db_get_memory_resource(DBR_PRECONDITIONS,
module_name,
true));
set_in_effects((statement_effects)
db_get_memory_resource(DBR_IN_REGIONS, module_name, true));
set_out_effects((statement_effects)
db_get_memory_resource(DBR_OUT_REGIONS, module_name, true));
debug_on("SCALARIZATION_DEBUG_LEVEL");
pips_debug(1, "begin\n");
// Context for gen_context_recurse
struct scalarization_ctx ctx;
ctx.keep_perfect_parallel =
get_bool_property("SCALARIZATION_KEEP_PERFECT_PARALLEL_LOOP_NESTS");
ctx.blacklisted_loops = set_make(set_pointer);
ctx.skip_address_of_variables =
get_bool_property("SCALARIZATION_SKIP_ADDRESS_OF_VARIABLES");
scalarization_across_control_test_level_init();
if (false) {
// DEAD CODE
/* We now traverse our module's statements looking for loop statements. */
loop_indices_b = BASE_NULLE;
ctx.scalarized_variables = NIL;
gen_context_recurse(module_stat,
&ctx,
statement_domain,
scalarization_loop_statement_in,
scalarization_loop_statement_out);
} else {
/* We now traverse our module's statements looking for all
statements. We look for constant array references to scalarize
on the way down. The effective scalarization is performed during
the bottom-up phase. */
loop_indices_b = BASE_NULLE;
ctx.scalarized_variables = NIL;
ctx.statement_scalarized_variables = hash_table_make(hash_pointer, 0);
gen_context_recurse(module_stat,
&ctx,
statement_domain,
scalarization_statement_in,
scalarization_statement_out);
/* Not enough the lists should be freed too... but they are freed
on the way by scalarization_statement_out() */
hash_table_free(ctx.statement_scalarized_variables);
ifdebug(1) {
pips_assert("scalarized_variables is empty",
ENDP(ctx.scalarized_variables));
pips_assert("loop_indices_b is empty", BASE_NULLE_P(loop_indices_b));
}
}
set_free(ctx.blacklisted_loops); // No leak
scalarization_across_control_test_level_reset();
pips_debug(1, "end\n");
debug_off();
/* Save modified code to database */
module_reorder(module_stat);
DB_PUT_MEMORY_RESOURCE(DBR_CODE, strdup(module_name), module_stat);
/* TODO: Cleanup after scalarization */
pips_assert("Loop index Pbase is empty", BASE_NULLE_P(loop_indices_b));
reset_current_module_entity();
reset_current_module_statement();
reset_cumulated_rw_effects();
reset_proper_rw_effects();
reset_precondition_map();
reset_transformer_map();
reset_in_effects();
reset_out_effects();
free_value_mappings();
/* Return value */
bool good_result_p = true;
return (good_result_p);
}
/*
* Scalarization of constant array references: constant_array_scalarization
*
* For instance, a[0] is replaced by a_0.
*/
typedef struct {
entity array;
bool constant_p;
} references_constant_param;
static void all_array_references_constant_walker(reference ref,
references_constant_param* p) {
if(same_entity_p(p->array,reference_variable(ref))) {
expression offset = reference_offset(ref);
p->constant_p &= extended_expression_constant_p(offset)
&& (!ENDP(reference_indices(ref)));
free_expression(offset);
}
}
static bool all_array_references_constant_p(statement in, entity array) {
references_constant_param p = { array,true};
gen_context_recurse(in, &p,
reference_domain,
gen_true2,
all_array_references_constant_walker);
return p.constant_p;
}
typedef struct {
entity array;
hash_table mapping;
} replace_references_constant_param;
static void replace_constant_array_references_walker(reference ref,
replace_references_constant_param *p) {
if(same_entity_p(p->array,reference_variable(ref))) {
/* we know for sure all indices are constant */
expression offset = reference_offset(ref);
intptr_t value;
if(!expression_integer_value(offset,&value))
pips_internal_error("reference index should be constants");
/* add one to the value, because 0 seems reserved */
entity var = (entity)hash_get(p->mapping,(void*)(1+value));
if(var == HASH_UNDEFINED_VALUE) {
var = make_new_scalar_variable_with_prefix(entity_user_name(p->array),
get_current_module_entity(),
basic_of_reference(ref));
/* FI: let's try to maintain the storage section */
storage as = entity_storage(p->array);
if(storage_ram_p(as)) {
ram ar = storage_ram(as);
storage vs = entity_storage(var);
ram vr = storage_ram(vs);
ram_section(vr) = ram_section(ar);
}
else {
pips_internal_error("Arrays should be allocated in RAM.\n");
}
hash_put(p->mapping,(void*)(1+value),var);
AddEntityToCurrentModule(var);
}
gen_full_free_list(reference_indices(ref));
free_expression(offset);
reference_indices(ref)=NIL;
reference_variable(ref)=var;
}
}
static void replace_constant_array_references(statement in, entity array) {
replace_references_constant_param p = { array,
hash_table_make(hash_int, HASH_DEFAULT_SIZE) };
gen_context_recurse(in, &p,
reference_domain,
gen_true2,
replace_constant_array_references_walker);
// FI: hos is the hash-table freed?
}
/* Pass constant_array_scalarization
*/
static int inverse_compare_entities(const entity *pe1, const entity *pe2)
{
return -compare_entities(pe1, pe2);
}
bool constant_array_scalarization(const string module_name)
{
set_current_module_entity(module_name_to_entity(module_name));
set_current_module_statement((statement) db_get_memory_resource(DBR_CODE,
module_name,
true));
statement cms = get_current_module_statement();
set sreferenced_entities = get_referenced_entities(cms);
// Use a sorted list of entities to generate code deterministically
list rel = set_to_sorted_list(sreferenced_entities,
(gen_cmp_func_t) inverse_compare_entities);
list del = NIL; // dynamic entity list
list sel = NIL; // static entity list
FOREACH(entity, e, rel) {
if((entity_array_p(e)||entity_pointer_p(e)) &&
all_array_references_constant_p(cms,e)) {
// Check legality...
if(same_string_p(entity_module_name(e), module_name)
&& !formal_parameter_p(e)) {
/* Do we have to cope with some array initialization? */
if(value_unknown_p(entity_initial(e))) {
if(entity_static_variable_p(e))
sel = CONS(ENTITY, e, sel);
else
del = CONS(ENTITY, e, del);
}
else {
/* FI: another option would be to split the initialization
* as part of the constant array scalarization pass. It is
* not possible to call brace_expression_to_statements()
* here because we do not iterate over the statements but
* over the variables. Another option would be to add a
* property to split_initialization to restrict the list of
* variables to process.
*/
pips_user_warning("Array \"%s\" is not scalarized, "
"because its initialization should be split. "
"\nApply pass split_initialization first.\n",
entity_user_name(e));
}
// FI: should the initial declaration be removed?
// This could be left to another pass cleaning unused declarations:
// CLEAN_DECLARATIONS?
}
else {
pips_user_warning("Array \"%s\" is not scalarized "
"because the resulting code is likely to be wrong\n",
entity_user_name(e));
}
}
}
/* Perform the replacement for non-static arrays */
FOREACH(entity, e, del) {
replace_constant_array_references(cms,e);
}
/* Perform the replacement for static arrays */
FOREACH(entity, e, sel) {
replace_constant_array_references(cms,e);
}
/* Perform the replacement for static arrays... once you know how to
generate the declarations of the elements properly... */
/* FOREACH(entity, e, del) { */
/* replace_constant_array_references(cms,e); */
/* } */
set_free(sreferenced_entities);
gen_free_list(rel); gen_free_list(del); gen_free_list(sel);
module_reorder(get_current_module_statement());
DB_PUT_MEMORY_RESOURCE(DBR_CODE, module_name, get_current_module_statement());
reset_current_module_entity();
reset_current_module_statement();
return true;
}
/***********************************************************************/
/* QUICK SCALARIZATION */
/***********************************************************************/
/*
scalarizes arrays in already parallel loops. uses only the dg and
simple effects (both proper and cumulated).
*/
/** checks whether an effect memory path describes a set of array elements
*/
static bool array_effect_p(effect eff) {
bool result = false;
entity e = effect_entity(eff);
if (!entity_scalar_p(e) && ! effects_package_entity_p(e)) {
type t = entity_basic_concrete_type(e);
if (type_variable_p(t)) {
list inds = reference_indices(effect_any_reference(eff));
list dims = variable_dimensions(type_variable(t));
result = (gen_length(inds) == gen_length(dims));
}
}
return result;
}
/** returns true if entity e belongs to C module "main"
could be moved to ri-util/entity.c, but this version is specific
because we assume that entity_module_main_p cannot return true if
the current module is not the "main" module. This is not valid in
the general case, for instance when performing interprocedural
translations. However, this choice has been made for performance
reasons.
*/
static bool entity_module_main_p(entity e) {
static entity current_module = entity_undefined;
static bool current_module_main_p = false;
bool result = false;
if (get_current_module_entity() != current_module) {
current_module = get_current_module_entity();
if (c_module_p(current_module)) {
const char* current_module_name= entity_local_name(current_module);
if (strcmp(current_module_name, "main") == 0) {
current_module_main_p = true;
}
pips_debug(8, "module name: %s, local_name: %s, returning %s\n",
entity_name(current_module), entity_local_name(current_module),
bool_to_string(current_module_main_p));
}
}
/* in the context of scalarization, an entity which is scalarizable
is an entity declared in the current module - so the current module
must be the "main" function.
*/
if (current_module_main_p) {
result = local_entity_of_module_p(e, current_module);
pips_debug(8, "does entity %s belong to module main? %s\n ",
entity_name(e), bool_to_string(result));
} else {
result = false;
}
return result;
}
/** checks if an entity meets the scalarizability criteria.
*/
static bool scalarizable_entity_p(entity e) {
storage s = entity_storage( e ) ;
bool result = true;
pips_debug(3,
"checking entity %s, with storage %s \n",
entity_name(e), storage_to_string(s));
ifdebug(3) {
if (storage_ram_p(s))
fprintf(stderr,
" and section %s\n",
entity_name(ram_section(storage_ram(s))));
}
/* global arrays are not considered as scalarizable because we
solely rely on internal dependencies. OUT regions would be
necessary to handle them (OUT simple effects should not be
precise enough), but we want to keep this algorithm very cheap on
purpose.
we allow scalarization of arrays with "static" qualifier when
they are declared in the "main" module of a C application.
*/
result = entity_array_p( e ) && !entity_volatile_variable_p(e) &&
((storage_formal_p( s )
&& parameter_passing_by_value_p(get_current_module_entity()))
||
(storage_ram_p( s )
&& (dynamic_area_p(ram_section(storage_ram(s)))
|| stack_area_p(ram_section(storage_ram(s)))
|| (static_area_p(ram_section(storage_ram(s)))
&& entity_module_main_p(e)))
)
);
pips_debug(3, "returning %s\n", bool_to_string(result));
return(result);
}
/**
if the statement st is a loop, gather the scalarizable candidates
in loops_scalarizable_candidates.
The candidates are the scalarizable array entities for which there
are cumulated effects attached to the loop body, or which are
declared inside the loop body.
*/
static bool statement_compute_scalarizable_candidates(statement st,
hash_table loops_scalarizable_candidates) {
if (statement_loop_p(st)) {
loop l = statement_loop(st);
statement b = loop_body(l);
set scalarizable_candidates = set_make(set_pointer);
pips_debug(1, "Entering loop statement %td (ordering %03zd)\n",
statement_number( st ), statement_ordering(st)) ;
ifdebug(1) {
print_effects(load_cumulated_rw_effects_list(b));
}
/* first scan the cumulated effects in search of candidates */
FOREACH(EFFECT, eff, load_cumulated_rw_effects_list(b)) {
entity e = effect_entity( eff ) ;
if(!entity_abstract_location_p(e)
&& scalarizable_entity_p(e)
&& action_write_p(effect_action(eff))
&& array_effect_p(eff)
&& !set_belong_p(scalarizable_candidates, e)) {
pips_debug(2, "adding array: %s from effects.\n", entity_name(e));
set_add_element(scalarizable_candidates, scalarizable_candidates, e) ;
}
}
/* then scan the loop_body declarations because they do not belong to
* cumulated effects */
FOREACH(ENTITY, e, statement_declarations(b)) {
if(!entity_abstract_location_p(e)
&& scalarizable_entity_p(e)
&& !set_belong_p(scalarizable_candidates, e)) {
pips_debug(2, "adding array: %s from declarations.\n", entity_name(e));
set_add_element(scalarizable_candidates, scalarizable_candidates, e) ;
}
}
ifdebug(1) {
pips_debug(1, "candidates:");
SET_FOREACH(entity, e, scalarizable_candidates) {
fprintf(stderr, " %s", entity_local_name(e));
}
fprintf(stderr, "\n");
}
/* finally put the set of candidates in the input hash table */
hash_put(loops_scalarizable_candidates, l, scalarizable_candidates);
pips_debug(1, "leaving loop\n");
}
return true;
}
/** remove scalarizable candidate e from the
loops_scalarizable_candidates corresponding to the loops which are
in ls but not in prefix.
*/
static void remove_scalarizable_candidate_from_loops(list prefix,
list ls,
entity e,
hash_table loops_scalarizable_candidates) {
pips_debug(1, "Begin\n");
if(ENDP(prefix)) {
if(!ENDP(ls)) {
ifdebug(1) {
pips_debug(1, "Removing %s from locals of ", entity_name(e)) ;
FOREACH(STATEMENT, st, ls) {
pips_debug(1, "%td ", statement_number(st)) ;
}
pips_debug(1, "\n" ) ;
}
FOREACH(STATEMENT, st, ls) {
pips_assert( "instruction i is a loop", statement_loop_p(st)) ;
set scalarizable_candidates =
(set) hash_get(loops_scalarizable_candidates,
(char *) statement_loop(st));
set_del_element(scalarizable_candidates, scalarizable_candidates, e);
pips_debug(1, "Variable %s is removed from scalarizable candidates of "
"statement %td\n",
entity_name(e), statement_number(st));
}
} else {
pips_debug(1, "ls is empty, end of recursion\n");
}
} else {
pips_assert( "The first statements in prefix and in ls are the same statement",
STATEMENT( CAR( prefix )) == STATEMENT( CAR( ls ))) ;
pips_debug(1, "Recurse on common prefix\n");
remove_scalarizable_candidate_from_loops(CDR(prefix), CDR(ls), e,
loops_scalarizable_candidates);
}
pips_debug(1, "End\n");
}
/**
Main part of the scalarization process: if edges from vertex @param
v corresponding to statement @param st may prevent scalarization of
effect @param eff entity, then remove the latter from non common
enclosing loops in @param loops_scalarizable_candidates.
This is mostly inspired from scalar variable privatization algorithm
*/
static void update_scalarizable_candidates(vertex v, statement st,
effect eff,
hash_table loops_scalarizable_candidates) {
list ls = load_statement_enclosing_loops(st);
entity e = effect_entity(eff);
ifdebug(1) {
if(statement_loop_p(st)) {
pips_debug(1, "Trying to scalarize %s in loop statement %td "
"(ordering %03zd)",
entity_local_name(e), statement_number( st ), statement_ordering(st));
} else {
pips_debug(1, "Trying to scalarize %s in statement %td\n",
entity_local_name( e ), statement_number( st )) ;
}
}
FOREACH(SUCCESSOR, succ, vertex_successors(v)) {
vertex succ_v = successor_vertex( succ ) ;
dg_vertex_label succ_l = (dg_vertex_label)vertex_vertex_label( succ_v ) ;
dg_arc_label arc_l = (dg_arc_label)successor_arc_label( succ ) ;
statement succ_st = ordering_to_statement(dg_vertex_label_statement(succ_l));
instruction succ_i = statement_instruction( succ_st ) ;
list succ_ls = load_statement_enclosing_loops( succ_st ) ;
list prefix = gen_common_prefix( ls, succ_ls ) ;
FOREACH(CONFLICT, c, dg_arc_label_conflicts(arc_l)) {
effect sc_eff = conflict_source( c ) ;
effect sk_eff = conflict_sink( c ) ;
bool keep = true;
if(store_effect_p(sc_eff) && store_effect_p(sk_eff)
&& e == effect_entity(sc_eff)
&& e == effect_entity(sk_eff)) {
ifdebug(3) {
pips_debug(3, "source effect:");
print_effect(sc_eff);
pips_debug(3, "sink effect:");
print_effect(sk_eff);
}
/* Take into account def-def and use-def edges only
if they are on a single and same element
which is true if the indices are constants
or only depend on common enclosing loops indices.
*/
if(action_write_p(effect_action(sk_eff))) {
set common_loop_indices = set_make(set_pointer);
pips_debug(3, "common loop indices :");
FOREACH(statement, s, prefix) {
loop l = statement_loop(s);
ifdebug(3) {
fprintf(stderr, "%s ", entity_name(loop_index(l)));
}
set_add_element(common_loop_indices, common_loop_indices, loop_index(l));
}
ifdebug(3) {
fprintf(stderr, "\n");
}
list sc_inds = reference_indices(effect_any_reference(sc_eff));
list sk_inds = reference_indices(effect_any_reference(sk_eff));
for(;
keep && !ENDP(sc_inds) && !ENDP(sk_inds);
POP(sc_inds), POP(sk_inds)) {
expression sc_exp = EXPRESSION(CAR(sc_inds));
expression sk_exp = EXPRESSION(CAR(sk_inds));
if (unbounded_expression_p(sc_exp) ||
unbounded_expression_p(sk_exp)) {
keep = false;
} else {
list l_eff_sc_exp = proper_effects_of_expression(sc_exp);
FOREACH(EFFECT, eff_sc_exp, l_eff_sc_exp) {
entity e_sc_exp = effect_entity(eff_sc_exp);
if (!set_belong_p(common_loop_indices, e_sc_exp)) {
keep = false;
break;
}
}
gen_full_free_list(l_eff_sc_exp);
keep = keep && expression_equal_p(sc_exp, sk_exp);
}
}
set_free(common_loop_indices);
} /* if(action_write_p(effect_action(sk_eff))) */
/* PC dependance and the sink is a loop index - shouldn't be necessary here */
else if(action_read_p( effect_action( sk_eff )) &&
(instruction_loop_p( succ_i) ||
is_implied_do_index( e, succ_i))) {
keep = true;
} else {
pips_debug(5,"Conflict for %s between statements %td and %td\n",
entity_local_name(e),
statement_number(st),
statement_number(succ_st));
if (v==succ_v) {
/* No decision can be made from this couple of effects alone */
keep = true;
} else {
pips_debug(5,"remove %s from candidates in non common enclosing loops\n",
entity_local_name(e));
keep = false;
}
}
if (!keep) {
pips_debug(1, "cannot keep candidate\n");
/* e cannot be a local variable at a lower level than
the common prefix because of this dependence
arc. */
remove_scalarizable_candidate_from_loops(prefix,
ls,
e,
loops_scalarizable_candidates);
remove_scalarizable_candidate_from_loops(prefix,
succ_ls,
e,
loops_scalarizable_candidates);
}
}
} /* FOREACH(CONFLICT, c, dg_arc_label_conflicts(arc_l)) */
gen_free_list( prefix ) ;
} /* FOREACH(SUCCESSOR, succ, vertex_successors(v)) */
pips_debug(1, "End\n");
}
/** local context type for scalarizability post-tests
*/
typedef struct {
entity e;
bool result;
reference first_ref;
list l_inds_first_ref;
} scalarizability_test_ctxt;
/** checks that all scalarizable proper effects of statement @param s
on current @param s ctxt entity are on the same reference
currently stored in @param ctxt.
*/
static bool entity_can_be_scalarized_in_statement_in(statement s,
scalarizability_test_ctxt *ctxt) {
pips_debug(1, "entering statement %td\n", statement_number(s));
ifdebug(1){print_statement(s);}
list l_eff = load_cumulated_rw_effects_list(s);
bool continue_p = true;
if (!effects_may_read_or_write_memory_paths_from_entity_p(l_eff, ctxt->e)
&& !entity_in_list_p(ctxt->e, statement_declarations(s))) {
/* the statement has no cumulated effect from entity e, and e is
not declared inside: it cannot prevent scalarization, nor it's
inner statements.
*/
continue_p = false;
} else {
/* the statement has cumulated effects from entity e; check its
proper effects before going to inner statements
*/
l_eff = load_proper_rw_effects_list(s);
FOREACH(EFFECT, eff, l_eff) {
reference eff_ref = effect_any_reference(eff);
entity eff_e = reference_variable(eff_ref);
if (store_effect_p(eff) && eff_e == ctxt->e) {
if (reference_undefined_p(ctxt->first_ref)) {
ctxt->first_ref = eff_ref;
ctxt->l_inds_first_ref = reference_indices(ctxt->first_ref);
} else {
/* check that current ref is similar to first ref */
list l_inds_eff_ref = reference_indices(eff_ref);
pips_assert("all scalarizable references have the same number "
"of indices",
gen_length(l_inds_eff_ref) == gen_length(ctxt->l_inds_first_ref));
list l_first = ctxt->l_inds_first_ref;
FOREACH(EXPRESSION, eff_exp, l_inds_eff_ref) {
expression first_exp = EXPRESSION(CAR(l_first));
if (!expression_equal_p(first_exp, eff_exp)) {
ctxt->result = false;
continue_p = false; /* no need to go on recursing on statements */
break;
}
POP(l_first);
}
if (!ctxt->result)
break;
}
} /* if (eff_e == e) */
} /* FOREACH(EFFECT, eff, l_eff) */
}
return continue_p;
}
/** after gathering scalarizable candidates, checks that entity @param
e can be scalarized in statement @param s according to 2 criteria:
1- there are no hidden references to the array due for
instance to a function call
2- all the accessed references are strictly the same;
3- the privatizable reference indices do not depend on variables modified
inside the loop body.
*/
static bool entity_can_be_scalarized_in_statement_p(entity e, statement s,
list l_modified_variables)
{
bool result = true;
pips_debug(2, "checking for entity %s in statement %td\n",
entity_name(e), statement_number(s));
/* First check that there are no hidden references to the array due for
instance to a function call.
*/
// Estimate the dynamic number of *element* and *variable*
// occurrences in the loop body
int neo = count_references_to_variable_element(s, e);
int nvo = count_references_to_variable(s, e);
if (nvo != neo) {
pips_debug(2,
"First legality criterion not met: %d!=%d (nvo!=neo)\n",
nvo, neo);
result = false;
} else {
scalarizability_test_ctxt ctxt;
ctxt.result = true;
ctxt.e = e;
ctxt.first_ref = reference_undefined;
ctxt.l_inds_first_ref = NIL;
gen_context_recurse(s, &ctxt,
statement_domain, entity_can_be_scalarized_in_statement_in, gen_null2);
result = ctxt.result;
if (!result) {
pips_debug(2,"Second legality criterion not met\n");
} else if (reference_undefined_p(ctxt.first_ref)) {
pips_debug(2,"No remaining reference found\n");
result = false;
} else {
// check that the reference indices do not depend on variables
// modified in the statement
ifdebug(3) {
pips_debug(3, "checking reference %s wrt entities:\n",
words_to_string(effect_words_reference(ctxt.first_ref)));
print_entities(l_modified_variables);
}
reference_testing_ctxt ref_test_ctxt;
ref_test_ctxt.constant_p = true;
ref_test_ctxt.e = e;
ref_test_ctxt.trans_args = l_modified_variables;
(void) reference_constant_wrt_ctxt_p(ctxt.first_ref, &ref_test_ctxt);
ref_test_ctxt.trans_args = NIL;
result = ref_test_ctxt.constant_p;
pips_debug(2,"Third legality criterion %s met\n", result? "" : "not");
}
}
return result;
}
/** check if loop scalarizable candidates are legal candidates
legacy criteria:
1- we only scalarize e if all references are on the same
memory location. we could scalarize different references if we had
a scalarize_reference_in_statement function.
2- references cannot be scalarized if accesses are performed through calls
*/
static void check_loop_scalarizable_candidates(loop l,
hash_table loops_scalarizable_candidates) {
set loop_scalarizable_candidates = hash_get(loops_scalarizable_candidates, l);
statement body = loop_body(l);
list l_loop =
CONS(STATEMENT, (statement) gen_get_ancestor(statement_domain, l),NIL);
list l_modified_variables =
effects_to_written_scalar_entities(load_cumulated_rw_effects_list(body));
FOREACH(ENTITY, e, statement_declarations(body)) {
if (entity_scalar_p(e)) {
l_modified_variables = CONS(ENTITY, e, l_modified_variables);
}
}
SET_FOREACH(entity, e, loop_scalarizable_candidates) {
if (!entity_can_be_scalarized_in_statement_p(e,
body,
l_modified_variables)) {
remove_scalarizable_candidate_from_loops(NIL,
l_loop,
e,
loops_scalarizable_candidates);
}
}
gen_free_list(l_loop);
gen_free_list(l_modified_variables);
}
/** check if scalarizable candidates are legal candidates
legacy criteria:
1- we only scalarize e if all references are on the same
memory location. we could scalarize different references if we had
a scalarize_reference_in_statement function.
2- references cannot be scalarized if accesses are performed through calls
*/
static void check_scalarizable_candidates(statement s,
hash_table loops_scalarizable_candidates) {
gen_context_recurse(s, loops_scalarizable_candidates,
loop_domain, gen_true2, check_loop_scalarizable_candidates);
}
/** Once scalarizable candidates have been discovered, and checked,
actually perform the scalarization on the loop @param l.
*/
static void loop_scalarize_candidates(loop l,
hash_table loops_scalarizable_candidates) {
set loop_scalarizable_candidates = hash_get(loops_scalarizable_candidates, l);
/* be deterministic! */
list l_loop_scalarizable_candidates =
set_to_sorted_list(loop_scalarizable_candidates,
(gen_cmp_func_t)compare_entities);
FOREACH(entity, e, l_loop_scalarizable_candidates) {
pips_debug(3, "entity %s is going to be scalarized in loop with index %s\n",
entity_name(e), entity_name(loop_index(l)));
scalarize_variable_in_statement(e,
loop_body(l),
entity_undefined,
entity_undefined);
}
gen_free_list(l_loop_scalarizable_candidates);
}
/** Once scalarizable candidates have been discovered, and checked,
actually perform the scalarization on the loops reachable from
statement @param s.
*/
static void scalarize_candidates(statement s,
hash_table loops_scalarizable_candidates) {
gen_context_recurse(s, loops_scalarizable_candidates,
loop_domain, gen_true2, loop_scalarize_candidates);
}
bool quick_scalarization(const string module_name)
{
entity module;
statement module_stat;
set_current_module_entity(module_name_to_entity(module_name));
module = get_current_module_entity();
set_current_module_statement( (statement)db_get_memory_resource(DBR_CODE,
module_name,
true) );
module_stat = get_current_module_statement();
set_proper_rw_effects((statement_effects)
db_get_memory_resource(DBR_PROPER_EFFECTS, module_name, true));
set_cumulated_rw_effects((statement_effects)
db_get_memory_resource(DBR_CUMULATED_EFFECTS, module_name, true) );
module_to_value_mappings(module);
/* Get the data dependence graph (chains) : */
graph dependence_graph = (graph)db_get_memory_resource(DBR_DG,
module_name,
true);
set_enclosing_loops_map( loops_mapping_of_statement( module_stat ) );
set_ordering_to_statement(module_stat);
debug_on("SCALARIZATION_DEBUG_LEVEL");
pips_debug(1, "begin\n");
/* Build maximal lists of scalarizable candidates */
hash_table loops_scalarizable_candidates = hash_table_make(hash_pointer, 0);
gen_context_recurse(module_stat, loops_scalarizable_candidates,
statement_domain, statement_compute_scalarizable_candidates, gen_null2);
/* remove non private variables from locals */
FOREACH(VERTEX, v, graph_vertices( dependence_graph )) {
dg_vertex_label vl = (dg_vertex_label) vertex_vertex_label( v ) ;
statement st = ordering_to_statement(dg_vertex_label_statement(vl));
pips_debug(1, "Entering statement %03zd :\n", statement_ordering(st));
ifdebug(4) {
print_statement(st);
}
FOREACH(EFFECT, eff, load_proper_rw_effects_list( st )) {
ifdebug(4) {
pips_debug(1, "effect :");
print_effect(eff);
}
if( action_write_p(effect_action(eff)) && array_effect_p(eff)) {
update_scalarizable_candidates( v, st, eff,
loops_scalarizable_candidates) ;
}
}
}
/* modify code with scalarized variables
*/
/* first check if candidates are really scalarizable
* with scalarize_variable_in_statement*/
check_scalarizable_candidates(module_stat, loops_scalarizable_candidates);
scalarize_candidates(module_stat, loops_scalarizable_candidates);
/* free hash_table of scalarizable candidates */
HASH_FOREACH(loop, l, set, s_candidates, loops_scalarizable_candidates) {
set_free(s_candidates);
}
hash_table_free(loops_scalarizable_candidates);
pips_debug(1, "end\n");
debug_off();
/* Save modified code to database */
module_reorder(module_stat);
DB_PUT_MEMORY_RESOURCE(DBR_CODE, strdup(module_name), module_stat);
clean_enclosing_loops( );
reset_ordering_to_statement();
reset_current_module_entity();
reset_current_module_statement();
reset_proper_rw_effects();
reset_cumulated_rw_effects();
free_value_mappings();
return (true);
}
#endif // BUILDER_ * _ SCALARIZATION