2 * Copyright (C) 1995-2007 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 Driver for the chordal register allocator.
23 * @author Sebastian Hack
39 #include "firm_config.h"
43 #include "irgraph_t.h"
44 #include "irprintf_t.h"
56 #include "iredges_t.h"
58 #include "bechordal_t.h"
60 #include "bejavacoal.h"
63 #include "besched_t.h"
67 #include "beifg_impl.h"
69 #include "bestatevent.h"
76 #include "bespillbelady.h"
77 #include "bespillmorgan.h"
78 #include "bespillslots.h"
79 #include "bespilloptions.h"
83 #include "bespillremat.h"
86 #include "bejavacoal.h"
87 #include "becopystat.h"
88 #include "becopyopt.h"
89 #include "bessadestr.h"
93 static be_ra_chordal_opts_t options = {
95 BE_CH_LOWER_PERM_SWAP,
101 typedef struct _post_spill_env_t {
102 be_chordal_env_t cenv;
104 const arch_register_class_t *cls;
105 double pre_spill_cost;
108 static be_ra_timer_t ra_timer = {
123 #include <libcore/lc_opts.h>
124 #include <libcore/lc_opts_enum.h>
125 #include <libcore/lc_timing.h>
127 static const lc_opt_enum_int_items_t lower_perm_items[] = {
128 { "copy", BE_CH_LOWER_PERM_COPY },
129 { "swap", BE_CH_LOWER_PERM_SWAP },
133 static const lc_opt_enum_int_items_t lower_perm_stat_items[] = {
137 static const lc_opt_enum_int_items_t dump_items[] = {
138 { "none", BE_CH_DUMP_NONE },
139 { "spill", BE_CH_DUMP_SPILL },
140 { "live", BE_CH_DUMP_LIVE },
141 { "color", BE_CH_DUMP_COLOR },
142 { "copymin", BE_CH_DUMP_COPYMIN },
143 { "ssadestr", BE_CH_DUMP_SSADESTR },
144 { "tree", BE_CH_DUMP_TREE_INTV },
145 { "constr", BE_CH_DUMP_CONSTR },
146 { "lower", BE_CH_DUMP_LOWER },
147 { "spillslots", BE_CH_DUMP_SPILLSLOTS },
148 { "appel", BE_CH_DUMP_APPEL },
149 { "all", BE_CH_DUMP_ALL },
153 static const lc_opt_enum_int_items_t be_ch_vrfy_items[] = {
154 { "off", BE_CH_VRFY_OFF },
155 { "warn", BE_CH_VRFY_WARN },
156 { "assert", BE_CH_VRFY_ASSERT },
160 static lc_opt_enum_int_var_t lower_perm_var = {
161 &options.lower_perm_opt, lower_perm_items
164 static lc_opt_enum_int_var_t dump_var = {
165 &options.dump_flags, dump_items
168 static lc_opt_enum_int_var_t be_ch_vrfy_var = {
169 &options.vrfy_option, be_ch_vrfy_items
172 static const lc_opt_table_entry_t be_chordal_options[] = {
173 LC_OPT_ENT_ENUM_PTR ("perm", "perm lowering options", &lower_perm_var),
174 LC_OPT_ENT_ENUM_MASK("dump", "select dump phases", &dump_var),
175 LC_OPT_ENT_ENUM_PTR ("vrfy", "verify options", &be_ch_vrfy_var),
178 #endif /* WITH_LIBCORE */
180 static void dump(unsigned mask, ir_graph *irg,
181 const arch_register_class_t *cls,
183 void (*dump_func)(ir_graph *, const char *))
185 if((options.dump_flags & mask) == mask) {
188 snprintf(buf, sizeof(buf), "-%s%s", cls->name, suffix);
189 be_dump(irg, buf, dump_func);
192 be_dump(irg, suffix, dump_func);
197 * Checks for every reload if it's user can perform the load on itself.
199 static void memory_operand_walker(ir_node *irn, void *env) {
200 be_chordal_env_t *cenv = env;
201 const arch_env_t *aenv = cenv->birg->main_env->arch_env;
202 const ir_edge_t *edge, *ne;
206 if (! be_is_Reload(irn))
209 /* only use memory operands, if the reload is only used by 1 node */
210 if(get_irn_n_edges(irn) > 1)
213 spill = be_get_Reload_mem(irn);
214 block = get_nodes_block(irn);
216 foreach_out_edge_safe(irn, edge, ne) {
217 ir_node *src = get_edge_src_irn(edge);
218 int pos = get_edge_src_pos(edge);
220 assert(src && "outedges broken!");
222 if (get_nodes_block(src) == block && arch_possible_memory_operand(aenv, src, pos)) {
223 arch_perform_memory_operand(aenv, src, spill, pos);
227 /* kill the Reload */
228 if (get_irn_n_edges(irn) == 0) {
230 set_irn_n(irn, be_pos_Reload_mem, new_Bad());
231 set_irn_n(irn, be_pos_Reload_frame, new_Bad());
236 * Starts a walk for memory operands if supported by the backend.
238 static INLINE void check_for_memory_operands(be_chordal_env_t *chordal_env) {
239 irg_walk_graph(chordal_env->irg, NULL, memory_operand_walker, chordal_env);
243 * Sorry for doing stats again...
245 typedef struct _node_stat_t {
246 unsigned int n_phis; /**< Phis of the current register class. */
247 unsigned int n_mem_phis; /**< Memory Phis (Phis with spill operands). */
248 unsigned int n_copies; /**< Copies */
249 unsigned int n_perms; /**< Perms */
250 unsigned int n_spills; /**< Spill nodes */
251 unsigned int n_reloads; /**< Reloads */
254 struct node_stat_walker {
256 const arch_env_t *arch_env;
258 const arch_register_class_t *cls;
261 static void node_stat_walker(ir_node *irn, void *data)
263 struct node_stat_walker *env = data;
264 const arch_env_t *aenv = env->arch_env;
266 if (arch_irn_consider_in_reg_alloc(aenv, env->cls, irn)) {
268 /* if the node is a normal phi */
272 else if(arch_irn_classify(aenv, irn) & arch_irn_class_spill)
273 ++env->stat->n_spills;
275 else if(arch_irn_classify(aenv, irn) & arch_irn_class_reload)
276 ++env->stat->n_reloads;
278 else if(arch_irn_classify(aenv, irn) & arch_irn_class_copy)
279 ++env->stat->n_copies;
281 else if(arch_irn_classify(aenv, irn) & arch_irn_class_perm)
282 ++env->stat->n_perms;
285 /* a mem phi is a PhiM with a mem phi operand or a Spill operand */
286 else if(is_Phi(irn) && get_irn_mode(irn) == mode_M) {
289 for(i = get_irn_arity(irn) - 1; i >= 0; --i) {
290 ir_node *op = get_irn_n(irn, i);
292 if((is_Phi(op) && bitset_contains_irn(env->mem_phis, op)) || (arch_irn_classify(aenv, op) & arch_irn_class_spill)) {
293 bitset_add_irn(env->mem_phis, irn);
294 env->stat->n_mem_phis++;
301 static void node_stats(be_irg_t *birg, const arch_register_class_t *cls, node_stat_t *stat)
303 struct node_stat_walker env;
305 memset(stat, 0, sizeof(stat[0]));
306 env.arch_env = birg->main_env->arch_env;
307 env.mem_phis = bitset_irg_malloc(birg->irg);
310 irg_walk_graph(birg->irg, NULL, node_stat_walker, &env);
311 bitset_free(env.mem_phis);
314 static void insn_count_walker(ir_node *irn, void *data)
318 switch(get_irn_opcode(irn)) {
329 static unsigned int count_insns(ir_graph *irg)
332 irg_walk_graph(irg, insn_count_walker, NULL, &cnt);
337 * Initialize all timers.
339 static void be_init_timer(be_options_t *main_opts)
341 if (main_opts->timing == BE_TIME_ON) {
342 ra_timer.t_prolog = lc_timer_register("ra_prolog", "regalloc prolog");
343 ra_timer.t_epilog = lc_timer_register("ra_epilog", "regalloc epilog");
344 ra_timer.t_live = lc_timer_register("ra_liveness", "be liveness");
345 ra_timer.t_spill = lc_timer_register("ra_spill", "spiller");
346 ra_timer.t_spillslots = lc_timer_register("ra_spillslots", "spillslots");
347 ra_timer.t_color = lc_timer_register("ra_color", "graph coloring");
348 ra_timer.t_ifg = lc_timer_register("ra_ifg", "interference graph");
349 ra_timer.t_copymin = lc_timer_register("ra_copymin", "copy minimization");
350 ra_timer.t_ssa = lc_timer_register("ra_ssadestr", "ssa destruction");
351 ra_timer.t_verify = lc_timer_register("ra_verify", "graph verification");
352 ra_timer.t_other = lc_timer_register("ra_other", "other time");
354 LC_STOP_AND_RESET_TIMER(ra_timer.t_prolog);
355 LC_STOP_AND_RESET_TIMER(ra_timer.t_epilog);
356 LC_STOP_AND_RESET_TIMER(ra_timer.t_live);
357 LC_STOP_AND_RESET_TIMER(ra_timer.t_spill);
358 LC_STOP_AND_RESET_TIMER(ra_timer.t_spillslots);
359 LC_STOP_AND_RESET_TIMER(ra_timer.t_color);
360 LC_STOP_AND_RESET_TIMER(ra_timer.t_ifg);
361 LC_STOP_AND_RESET_TIMER(ra_timer.t_copymin);
362 LC_STOP_AND_RESET_TIMER(ra_timer.t_ssa);
363 LC_STOP_AND_RESET_TIMER(ra_timer.t_verify);
364 LC_STOP_AND_RESET_TIMER(ra_timer.t_other);
366 global_ra_timer = &ra_timer;
370 #define BE_TIMER_INIT(main_opts) be_init_timer(main_opts)
372 #define BE_TIMER_PUSH(timer) \
373 if (main_opts->timing == BE_TIME_ON) { \
374 if (! lc_timer_push(timer)) { \
375 if (options.vrfy_option == BE_CH_VRFY_ASSERT) \
376 assert(!"Timer already on stack, cannot be pushed twice."); \
377 else if (options.vrfy_option == BE_CH_VRFY_WARN) \
378 fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \
379 lc_timer_get_name(timer)); \
382 #define BE_TIMER_POP(timer) \
383 if (main_opts->timing == BE_TIME_ON) { \
384 lc_timer_t *tmp = lc_timer_pop(); \
385 if (options.vrfy_option == BE_CH_VRFY_ASSERT) \
386 assert(tmp == timer && "Attempt to pop wrong timer."); \
387 else if (options.vrfy_option == BE_CH_VRFY_WARN && tmp != timer) \
388 fprintf(stderr, "Attempt to pop wrong timer. %s is on stack, trying to pop %s.\n", \
389 lc_timer_get_name(tmp), lc_timer_get_name(timer)); \
394 * Perform things which need to be done per register class before spilling.
396 static void pre_spill(const arch_isa_t *isa, int cls_idx, post_spill_env_t *pse) {
397 be_chordal_env_t *chordal_env = &pse->cenv;
398 be_irg_t *birg = pse->birg;
399 node_stat_t node_stat;
401 pse->cls = arch_isa_get_reg_class(isa, cls_idx);
402 chordal_env->cls = pse->cls;
403 chordal_env->border_heads = pmap_create();
404 chordal_env->ignore_colors = bitset_malloc(chordal_env->cls->n_regs);
406 be_assure_liveness(birg);
407 be_liveness_assure_chk(be_get_birg_liveness(birg));
408 stat_ev_ctx_push("cls", pse->cls->name);
409 stat_ev_dbl("phis_before_spill", node_stat.n_phis);
410 stat_ev_do(node_stats(birg, pse->cls, &node_stat));
412 /* put all ignore registers into the ignore register set. */
413 be_put_ignore_regs(birg, pse->cls, chordal_env->ignore_colors);
415 be_pre_spill_prepare_constr(chordal_env);
416 dump(BE_CH_DUMP_CONSTR, birg->irg, pse->cls, "-constr-pre", dump_ir_block_graph_sched);
422 * Perform things which need to be done per register class after spilling.
424 static void post_spill(post_spill_env_t *pse, int iteration) {
425 be_chordal_env_t *chordal_env = &pse->cenv;
426 be_irg_t *birg = pse->birg;
427 ir_graph *irg = birg->irg;
428 const be_main_env_t *main_env = birg->main_env;
429 be_options_t *main_opts = main_env->options;
430 node_stat_t node_stat;
431 int colors_n = arch_register_class_n_regs(chordal_env->cls);
432 int allocatable_regs = colors_n - be_put_ignore_regs(birg, chordal_env->cls, NULL);
434 /* some special classes contain only ignore regs, no work to be done */
435 if (allocatable_regs > 0) {
437 stat_ev_ctx_push("cls", pse->cls->name);
438 stat_ev_do(node_stats(birg, pse->cls, &node_stat));
439 stat_ev_dbl("spillcosts", be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq) - pse->pre_spill_cost);
440 stat_ev_dbl("phis_after_spill", node_stat.n_phis);
441 stat_ev_dbl("mem_phis", node_stat.n_mem_phis);
442 stat_ev_dbl("reloads", node_stat.n_reloads);
443 stat_ev_dbl("spills", node_stat.n_spills);
446 If we have a backend provided spiller, post spill is
447 called in a loop after spilling for each register class.
448 But we only need to fix stack nodes once in this case.
450 if (iteration == 0) {
451 check_for_memory_operands(chordal_env);
452 be_abi_fix_stack_nodes(birg->abi);
455 BE_TIMER_PUSH(ra_timer.t_verify);
457 /* verify schedule and register pressure */
458 if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
459 be_verify_schedule(birg);
460 be_verify_register_pressure(birg, pse->cls, irg);
462 else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
463 assert(be_verify_schedule(birg) && "Schedule verification failed");
464 assert(be_verify_register_pressure(birg, pse->cls, irg)
465 && "Register pressure verification failed");
467 BE_TIMER_POP(ra_timer.t_verify);
469 /* Color the graph. */
470 BE_TIMER_PUSH(ra_timer.t_color);
471 be_ra_chordal_color(chordal_env);
472 BE_TIMER_POP(ra_timer.t_color);
474 dump(BE_CH_DUMP_CONSTR, irg, pse->cls, "-color", dump_ir_block_graph_sched);
476 /* Create the ifg with the selected flavor */
477 BE_TIMER_PUSH(ra_timer.t_ifg);
478 chordal_env->ifg = be_create_ifg(chordal_env);
479 BE_TIMER_POP(ra_timer.t_ifg);
484 stat_ev_do(be_ifg_stat(birg, chordal_env->ifg, &stat));
485 stat_ev_dbl("ifg_nodes", stat.n_nodes);
486 stat_ev_dbl("ifg_edges", stat.n_edges);
487 stat_ev_dbl("ifg_comps", stat.n_comps);
489 stat_ev_do(node_stats(birg, pse->cls, &node_stat));
490 stat_ev_dbl("perms_before_coal", node_stat.n_perms);
491 stat_ev_dbl("copies_before_coal", node_stat.n_copies);
494 /* copy minimization */
495 BE_TIMER_PUSH(ra_timer.t_copymin);
496 co_driver(chordal_env);
497 BE_TIMER_POP(ra_timer.t_copymin);
499 dump(BE_CH_DUMP_COPYMIN, irg, pse->cls, "-copymin", dump_ir_block_graph_sched);
501 BE_TIMER_PUSH(ra_timer.t_ssa);
503 /* ssa destruction */
504 be_ssa_destruction(chordal_env);
506 BE_TIMER_POP(ra_timer.t_ssa);
508 dump(BE_CH_DUMP_SSADESTR, irg, pse->cls, "-ssadestr", dump_ir_block_graph_sched);
510 BE_TIMER_PUSH(ra_timer.t_verify);
511 if (chordal_env->opts->vrfy_option != BE_CH_VRFY_OFF) {
512 be_ssa_destruction_check(chordal_env);
514 BE_TIMER_POP(ra_timer.t_verify);
516 stat_ev_do(node_stats(birg, pse->cls, &node_stat));
517 stat_ev_dbl("perms_after_coal", node_stat.n_perms);
518 stat_ev_dbl("copies_after_coal", node_stat.n_copies);
521 /* the ifg exists only if there are allocatable regs */
522 be_ifg_free(chordal_env->ifg);
525 /* free some always allocated data structures */
526 pmap_destroy(chordal_env->border_heads);
527 bitset_free(chordal_env->ignore_colors);
531 * Performs chordal register allocation for each register class on given irg.
533 * @param birg Backend irg object
534 * @return Structure containing timer for the single phases or NULL if no timing requested.
536 static void be_ra_chordal_main(be_irg_t *birg)
538 const be_main_env_t *main_env = birg->main_env;
539 const arch_isa_t *isa = arch_env_get_isa(main_env->arch_env);
540 ir_graph *irg = birg->irg;
541 be_options_t *main_opts = main_env->options;
543 be_chordal_env_t chordal_env;
546 BE_TIMER_INIT(main_opts);
547 BE_TIMER_PUSH(ra_timer.t_other);
548 BE_TIMER_PUSH(ra_timer.t_prolog);
550 be_assure_dom_front(birg);
551 be_assure_liveness(birg);
553 chordal_env.obst = &obst;
554 chordal_env.opts = &options;
555 chordal_env.irg = irg;
556 chordal_env.birg = birg;
557 chordal_env.border_heads = NULL;
558 chordal_env.ifg = NULL;
559 chordal_env.ignore_colors = NULL;
563 BE_TIMER_POP(ra_timer.t_prolog);
565 be_stat_ev("insns_before", count_insns(irg));
567 if (! arch_code_generator_has_spiller(birg->cg)) {
568 /* use one of the generic spiller */
570 /* Perform the following for each register class. */
571 for (j = 0, m = arch_isa_get_n_reg_class(isa); j < m; ++j) {
572 post_spill_env_t pse;
574 memcpy(&pse.cenv, &chordal_env, sizeof(chordal_env));
576 pre_spill(isa, j, &pse);
578 BE_TIMER_PUSH(ra_timer.t_spill);
579 be_do_spill(birg, pse.cls);
580 BE_TIMER_POP(ra_timer.t_spill);
582 dump(BE_CH_DUMP_SPILL, irg, pse.cls, "-spill", dump_ir_block_graph_sched);
587 post_spill_env_t *pse;
589 /* the backend has it's own spiller */
590 m = arch_isa_get_n_reg_class(isa);
592 pse = alloca(m * sizeof(pse[0]));
594 for (j = 0; j < m; ++j) {
595 memcpy(&pse[j].cenv, &chordal_env, sizeof(chordal_env));
597 pre_spill(isa, j, &pse[j]);
600 BE_TIMER_PUSH(ra_timer.t_spill);
601 arch_code_generator_spill(birg->cg, birg);
602 BE_TIMER_POP(ra_timer.t_spill);
603 dump(BE_CH_DUMP_SPILL, irg, NULL, "-spill", dump_ir_block_graph_sched);
605 for (j = 0; j < m; ++j) {
606 post_spill(&pse[j], j);
610 be_verify_register_allocation(main_env->arch_env, irg);
613 BE_TIMER_PUSH(ra_timer.t_epilog);
614 lower_nodes_after_ra(birg, options.lower_perm_opt & BE_CH_LOWER_PERM_COPY ? 1 : 0);
615 dump(BE_CH_DUMP_LOWER, irg, NULL, "-belower-after-ra", dump_ir_block_graph_sched);
617 obstack_free(&obst, NULL);
618 be_liveness_invalidate(be_get_birg_liveness(birg));
619 BE_TIMER_POP(ra_timer.t_epilog);
621 BE_TIMER_POP(ra_timer.t_other);
623 be_stat_ev("insns_after", count_insns(irg));
628 static be_ra_t be_ra_chordal_allocator = {
632 void be_init_chordal_main(void)
635 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
636 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
637 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
639 lc_opt_add_table(chordal_grp, be_chordal_options);
641 be_register_allocator("chordal", &be_ra_chordal_allocator);
644 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal_main);