2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief ILP based copy minimization.
23 * @author Daniel Grund
27 * ILP formalization using G=(V, E, Q):
28 * - 2 class of variables: coloring vars x_ic and equal color vars y_ij
30 * - Clique-star constraints
33 * \min \sum_{ (i,j) \in Q } w_ij y_ij
35 * \sum_c x_nc = 1 n \in N, c \in C
37 * x_nc = 0 n \in N, c \not\in C(n)
39 * \sum x_nc <= 1 x_nc \in Clique \in AllCliques, c \in C
41 * \sum_{e \in p} y_e >= 1 p \in P path constraints
43 * \sum_{e \in cs} y_e >= |cs| - 1 cs \in CP clique-star constraints
45 * x_nc, y_ij \in N, w_ij \in R^+
49 #endif /* HAVE_CONFIG_H */
51 #include "firm_config.h"
56 #include "raw_bitset.h"
61 #include "becopyilp_t.h"
63 #include "besched_t.h"
67 typedef struct _local_env_t {
69 int first_x_var, last_x_var;
71 bitset_t *normal_colors;
73 DEBUG_ONLY(firm_dbg_module_t *dbg;)
76 static void build_coloring_cstr(ilp_env_t *ienv) {
77 be_ifg_t *ifg = ienv->co->cenv->ifg;
78 void *iter = be_ifg_nodes_iter_alloca(ifg);
83 colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
85 be_ifg_foreach_node(ifg, iter, irn)
86 if (!sr_is_removed(ienv->sr, irn)) {
89 const arch_register_req_t *req;
90 int curr_node_color = get_irn_col(ienv->co, irn);
91 int node_nr = (int)get_irn_idx(irn);
92 local_env_t *lenv = ienv->env;
94 pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
96 req = arch_get_register_req(ienv->co->aenv, irn, -1);
98 bitset_clear_all(colors);
100 /* get assignable colors */
101 if (arch_register_req_is(req, limited)) {
102 rbitset_copy_to_bitset(req->limited, colors);
104 bitset_copy(colors, lenv->normal_colors);
107 /* add the coloring constraint */
108 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
110 bitset_foreach(colors, col) {
111 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
112 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
113 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
115 lenv->last_x_var = var_idx;
116 if (lenv->first_x_var == -1)
117 lenv->first_x_var = var_idx;
120 /* add register constraint constraints */
121 bitset_foreach_clear(colors, col) {
122 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
123 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
124 lpp_set_start_value(ienv->lp, var_idx, 0.0);
125 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
127 lenv->last_x_var = var_idx;
132 static void build_interference_cstr(ilp_env_t *ienv) {
133 lpp_t *lpp = ienv->lp;
134 local_env_t *lenv = ienv->env;
135 be_ifg_t *ifg = ienv->co->cenv->ifg;
136 int n_colors = lenv->n_colors;
139 void *iter = be_ifg_cliques_iter_alloca(ifg);
140 ir_node **clique = alloca(sizeof(*clique) * n_colors);
145 /* for each maximal clique */
146 be_ifg_foreach_clique(ifg, iter, clique, &size) {
149 for (i=0; i<size; ++i)
150 if (!sr_is_removed(ienv->sr, clique[i]))
157 for (col=0; col<n_colors; ++col) {
158 int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
160 /* for each member of this clique */
161 for (i=0; i<size; ++i) {
162 ir_node *irn = clique[i];
164 if (!sr_is_removed(ienv->sr, irn)) {
165 int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
166 lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
175 * TODO: Remove the dependency of the opt-units data structure
176 * by walking over all affinity edges. Graph structure
177 * does not provide this walker, yet.
179 static void build_affinity_cstr(ilp_env_t *ienv) {
180 local_env_t *lenv = ienv->env;
181 int n_colors = lenv->n_colors;
184 /* for all optimization units */
185 list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
187 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
189 int root_col, arg_col;
191 root = curr->nodes[0];
192 root_nr = (int) get_irn_idx(root);
193 root_col = get_irn_col(ienv->co, root);
195 for (i = 1; i < curr->node_count; ++i) {
196 arg = curr->nodes[i];
197 arg_nr = (int) get_irn_idx(arg);
198 arg_col = get_irn_col(ienv->co, arg);
200 /* add a new affinity variable */
201 y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
202 lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
204 /* add constraints relating the affinity var to the color vars */
205 for (col=0; col<n_colors; ++col) {
206 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
207 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
208 arg_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr, col));
210 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx, 1.0);
211 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx, -1.0);
212 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
219 * Helping stuff for build_clique_star_cstr
221 typedef struct _edge_t {
225 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
226 const edge_t *e1 = k1;
227 const edge_t *e2 = k2;
230 return ! (e1->n1 == e2->n1 && e1->n2 == e2->n2);
233 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
235 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
238 if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
246 return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
249 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
252 if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
259 return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
262 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
265 if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
272 e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
280 #define pset_foreach(pset, irn) for(irn=pset_first(pset); irn; irn=pset_next(pset))
283 * Search for an interference clique and an external node
284 * with affinity edges to all nodes of the clique.
285 * At most 1 node of the clique can be colored equally with the external node.
287 static void build_clique_star_cstr(ilp_env_t *ienv) {
288 affinity_node_t *aff;
290 /* for each node with affinity edges */
291 co_gs_foreach_aff_node(ienv->co, aff) {
294 const ir_node *center = aff->irn;
297 int i, o, n_nodes, n_edges;
299 if (arch_irn_is(ienv->co->aenv, aff->irn, ignore))
303 edges = new_set(compare_edge_t, 8);
305 /* get all affinity neighbours */
307 co_gs_foreach_neighb(aff, nbr) {
308 if (!arch_irn_is(ienv->co->aenv, nbr->irn, ignore)) {
309 obstack_ptr_grow(&ob, nbr->irn);
313 nodes = obstack_finish(&ob);
315 /* get all interference edges between these */
317 for (i=0; i<n_nodes; ++i)
319 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
320 add_edge(edges, nodes[i], nodes[o], &n_edges);
322 /* cover all these interference edges with maximal cliques */
325 pset *clique = pset_new_ptr(8);
328 /* get 2 starting nodes to form a clique */
329 for (e=set_first(edges); !e->n1; e=set_next(edges))
332 /* we could be stepped out of the loop before the set iterated to the end */
335 pset_insert_ptr(clique, e->n1);
336 pset_insert_ptr(clique, e->n2);
337 remove_edge(edges, e->n1, e->n2, &n_edges);
339 /* while the clique is growing */
343 /* search for a candidate to extend the clique */
344 for (i=0; i<n_nodes; ++i) {
345 ir_node *member, *cand = nodes[i];
348 /* if its already in the clique try the next */
349 if (pset_find_ptr(clique, cand))
352 /* are there all necessary interferences? */
354 pset_foreach(clique, member) {
355 if (!find_edge(edges, cand, member)) {
362 /* now we know if we have a clique extender */
364 /* first remove all covered edges */
365 pset_foreach(clique, member)
366 remove_edge(edges, cand, member, &n_edges);
368 /* insert into clique */
369 pset_insert_ptr(clique, cand);
376 /* now the clique is maximal. Finally add the constraint */
379 int var_idx, cst_idx, center_nr, member_nr;
382 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
383 center_nr = get_irn_idx(center);
385 pset_foreach(clique, member) {
386 member_nr = get_irn_idx(member);
387 var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
388 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
396 obstack_free(&ob, NULL);
401 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
402 be_ifg_t *ifg = ienv->co->cenv->ifg;
405 affinity_node_t *aff;
408 /* do not walk backwards or in circles */
409 if (pdeq_contains(path, irn))
412 if (arch_irn_is(ienv->co->aenv, irn, ignore))
415 /* insert the new irn */
416 pdeq_putr(path, irn);
420 /* check for forbidden interferences */
421 len = pdeq_len(path);
422 curr_path = alloca(len * sizeof(*curr_path));
423 pdeq_copyl(path, (const void **)curr_path);
425 for (i=1; i<len; ++i)
426 if (be_ifg_connected(ifg, irn, curr_path[i]))
431 /* check for terminating interference */
432 if (be_ifg_connected(ifg, irn, curr_path[0])) {
434 /* One node is not a path. */
435 /* And a path of length 2 is covered by a clique star constraint. */
437 /* finally build the constraint */
438 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
439 for (i=1; i<len; ++i) {
441 int nr_1 = get_irn_idx(curr_path[i-1]);
442 int nr_2 = get_irn_idx(curr_path[i]);
443 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
444 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
448 /* this path cannot be extended anymore */
454 /* recursively extend the path */
455 aff = get_affinity_info(ienv->co, irn);
456 co_gs_foreach_neighb(aff, nbr)
457 extend_path(ienv, path, nbr->irn);
467 * Search a path of affinity edges, whose ends are connected
468 * by an interference edge and there are no other interference
470 * Then at least one of these affinity edges must break.
472 static void build_path_cstr(ilp_env_t *ienv) {
473 affinity_node_t *aff_info;
475 /* for each node with affinity edges */
476 co_gs_foreach_aff_node(ienv->co, aff_info) {
477 pdeq *path = new_pdeq();
479 extend_path(ienv, path, aff_info->irn);
485 static void ilp2_build(ilp_env_t *ienv) {
486 local_env_t *lenv = ienv->env;
489 ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
490 build_coloring_cstr(ienv);
491 build_interference_cstr(ienv);
492 build_affinity_cstr(ienv);
493 build_clique_star_cstr(ienv);
494 build_path_cstr(ienv);
496 lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
497 lpp_set_bound(ienv->lp, lower_bound);
498 lpp_set_time_limit(ienv->lp, lenv->time_limit);
501 static void ilp2_apply(ilp_env_t *ienv) {
502 local_env_t *lenv = ienv->env;
504 lpp_sol_state_t state;
507 /* first check if there was sth. to optimize */
508 if (lenv->first_x_var >= 0) {
510 count = lenv->last_x_var - lenv->first_x_var + 1;
511 sol = xmalloc(count * sizeof(sol[0]));
512 state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
513 if (state != lpp_optimal) {
514 printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
515 assert(state >= lpp_feasible && "The solution should at least be feasible!");
518 for (i=0; i<count; ++i) {
522 if (sol[i] > 1-EPSILON) { /* split variable name into components */
523 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
525 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
526 ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
527 assert(irn && "This node number must be present in the map");
529 set_irn_col(ienv->co, irn, color);
531 assert(0 && "This should be a x-var");
537 /* TODO adapt to multiple possible ILPs */
538 copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp))); //now we have ms
539 copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
540 copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
541 copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
545 int co_solve_ilp2(copy_opt_t *co) {
546 lpp_sol_state_t sol_state;
550 ASSERT_OU_AVAIL(co); //See build_clique_st
556 my.nr_2_irn = pmap_create();
557 FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
559 my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
560 bitset_clear_all(my.normal_colors);
561 arch_put_non_ignore_regs(co->aenv, co->cls, my.normal_colors);
562 my.n_colors = bitset_popcnt(my.normal_colors);
564 ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
566 sol_state = ilp_go(ienv);
568 pmap_destroy(my.nr_2_irn);
571 return sol_state == lpp_optimal;
576 static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
579 #endif /* WITH_ILP */