3 * @author Sebastian Hack
14 #include <libcore/lc_opts.h>
15 #include <libcore/lc_opts_enum.h>
16 #include <libcore/lc_timing.h>
17 #endif /* WITH_LIBCORE */
28 #include "iredges_t.h"
36 #include "firm/bearch_firm.h"
37 #include "ia32/bearch_ia32.h"
38 #include "arm/bearch_arm.h"
39 #include "ppc32/bearch_ppc32.h"
40 #include "mips/bearch_mips.h"
47 #include "besched_t.h"
48 #include "belistsched.h"
50 #include "bespillbelady.h"
52 #include "beraextern.h"
53 #include "bechordal_t.h"
55 #include "beifg_impl.h"
56 #include "becopyopt.h"
57 #include "becopystat.h"
58 #include "bessadestr.h"
61 #include "beschedmris.h"
65 /* options visible for anyone */
66 static be_options_t be_options = {
67 DUMP_NONE, /* dump options */
68 BE_TIME_OFF, /* no timing */
69 BE_SCHED_SELECT_HEUR, /* mueller heuristic selector */
71 "i44pc52.info.uni-karlsruhe.de", /* ilp server */
72 "cplex" /* ilp solver */
76 static char config_file[256] = { 0 };
79 static unsigned dump_flags = 0;
82 static int vrfy_option = BE_VRFY_WARN;
84 /* register allocator to use. */
85 static const be_ra_t *ra = &be_ra_chordal_allocator;
87 /* back end instruction set architecture to use */
88 static const arch_isa_if_t *isa_if = &ia32_isa_if;
92 static lc_opt_entry_t *be_grp_root = NULL;
94 /* possible dumping options */
95 static const lc_opt_enum_mask_items_t dump_items[] = {
96 { "none", DUMP_NONE },
97 { "initial", DUMP_INITIAL },
99 { "sched", DUMP_SCHED },
100 { "prepared", DUMP_PREPARED },
101 { "regalloc", DUMP_RA },
102 { "final", DUMP_FINAL },
104 { "all", 2 * DUMP_BE - 1 },
108 /* register allocators */
109 static const lc_opt_enum_const_ptr_items_t ra_items[] = {
110 { "chordal", &be_ra_chordal_allocator },
111 { "external", &be_ra_external_allocator },
115 /* instruction set architectures. */
116 static const lc_opt_enum_const_ptr_items_t isa_items[] = {
117 { "ia32", &ia32_isa_if },
119 { "arm", &arm_isa_if },
120 { "ppc32", &ppc32_isa_if },
121 { "mips", &mips_isa_if },
126 /* verify options. */
127 static const lc_opt_enum_int_items_t vrfy_items[] = {
128 { "off", BE_VRFY_OFF },
129 { "warn", BE_VRFY_WARN },
130 { "assert", BE_VRFY_ASSERT },
134 /* schedule selector options. */
135 static const lc_opt_enum_int_items_t sched_select_items[] = {
136 { "trivial", BE_SCHED_SELECT_TRIVIAL },
137 { "regpress", BE_SCHED_SELECT_REGPRESS },
138 { "muchnik", BE_SCHED_SELECT_MUCHNIK },
139 { "heur", BE_SCHED_SELECT_HEUR },
140 { "hmuchnik", BE_SCHED_SELECT_HMUCHNIK },
144 static lc_opt_enum_mask_var_t dump_var = {
145 &dump_flags, dump_items
148 static lc_opt_enum_const_ptr_var_t ra_var = {
149 (const void **) &ra, ra_items
152 static lc_opt_enum_const_ptr_var_t isa_var = {
153 (const void **) &isa_if, isa_items
156 static lc_opt_enum_int_var_t vrfy_var = {
157 &vrfy_option, vrfy_items
160 static lc_opt_enum_int_var_t sched_select_var = {
161 &be_options.sched_select, sched_select_items
164 static const lc_opt_table_entry_t be_main_options[] = {
165 LC_OPT_ENT_STR ("config", "read another config file containing backend options", config_file, sizeof(config_file)),
166 LC_OPT_ENT_ENUM_MASK("dump", "dump irg on several occasions", &dump_var),
167 LC_OPT_ENT_ENUM_PTR ("ra", "register allocator", &ra_var),
168 LC_OPT_ENT_ENUM_PTR ("isa", "the instruction set architecture", &isa_var),
169 LC_OPT_ENT_NEGBOOL ("noomitfp", "do not omit frame pointer", &be_omit_fp),
170 LC_OPT_ENT_ENUM_PTR ("vrfy", "verify the backend irg (off, warn, assert)", &vrfy_var),
171 LC_OPT_ENT_BOOL ("time", "get backend timing statistics", &be_options.timing),
172 LC_OPT_ENT_BOOL ("sched.mris", "enable mris schedule preparation", &be_options.mris),
173 LC_OPT_ENT_ENUM_PTR ("sched.select", "schedule node selector (trivial, regpress, muchnik, heur, hmuchnik)", &sched_select_var),
176 LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)),
177 LC_OPT_ENT_STR ("ilp.solver", "the ilp solver name", be_options.ilp_solver, sizeof(be_options.ilp_solver)),
178 #endif /* WITH_ILP */
182 #endif /* WITH_LIBCORE */
184 void be_opt_register(void)
188 lc_opt_entry_t *be_grp_ra;
189 static int run_once = 0;
193 be_grp_root = lc_opt_get_grp(firm_opt_get_root(), "be");
194 be_grp_ra = lc_opt_get_grp(be_grp_root, "ra");
196 lc_opt_add_table(be_grp_root, be_main_options);
198 /* register allocator options */
199 for(i = 0; ra_items[i].name != NULL; ++i) {
200 const be_ra_t *ra = ra_items[i].value;
201 ra->register_options(be_grp_ra);
204 /* register isa options */
205 for(i = 0; isa_items[i].name != NULL; ++i) {
206 const arch_isa_if_t *isa = isa_items[i].value;
207 isa->register_options(be_grp_root);
210 #endif /* WITH_LIBCORE */
213 /* Parse one argument. */
214 int be_parse_arg(const char *arg) {
216 if (strcmp(arg, "help") == 0 || (arg[0] == '?' && arg[1] == '\0')) {
217 lc_opt_print_help(be_grp_root, stdout);
220 return lc_opt_from_single_arg(be_grp_root, NULL, arg, NULL);
223 #endif /* WITH_LIBCORE */
226 /** The be parameters returned by default, all off. */
227 const static backend_params be_params = {
234 /* Perform schedule verification if requested. */
235 static void be_sched_vrfy(ir_graph *irg, int vrfy_opt) {
236 if (vrfy_opt == BE_VRFY_WARN) {
237 be_verify_schedule(irg);
239 else if (vrfy_opt == BE_VRFY_ASSERT) {
240 assert(be_verify_schedule(irg) && "Schedule verification failed.");
244 /* Initialize the Firm backend. Must be run BEFORE init_firm()! */
245 const backend_params *be_init(void)
255 if (isa_if->get_params)
256 return isa_if->get_params();
261 * Initializes the main environment for the backend.
263 * @param env an empty environment
264 * @param file_handle the file handle where the output will be written to
266 static be_main_env_t *be_init_env(be_main_env_t *env, FILE *file_handle)
268 memset(env, 0, sizeof(*env));
269 obstack_init(&env->obst);
270 env->arch_env = obstack_alloc(&env->obst, sizeof(env->arch_env[0]));
271 env->options = &be_options;
272 env->options->dump_flags = dump_flags;
273 FIRM_DBG_REGISTER(env->dbg, "be.main");
275 arch_env_init(env->arch_env, isa_if, file_handle);
277 /* Register the irn handler of the architecture */
278 if (arch_isa_get_irn_handler(env->arch_env->isa))
279 arch_env_push_irn_handler(env->arch_env, arch_isa_get_irn_handler(env->arch_env->isa));
282 * Register the node handler of the back end infrastructure.
283 * This irn handler takes care of the platform independent
284 * spill, reload and perm nodes.
286 arch_env_push_irn_handler(env->arch_env, &be_node_irn_handler);
287 env->phi_handler = be_phi_handler_new(env->arch_env);
288 arch_env_push_irn_handler(env->arch_env, env->phi_handler);
293 static void be_done_env(be_main_env_t *env)
295 env->arch_env->isa->impl->done(env->arch_env->isa);
296 be_phi_handler_free(env->phi_handler);
297 obstack_free(&env->obst, NULL);
301 * A wrapper around a firm dumper. Dumps only, if
304 * @param mask a bitmask containing the reason what will be dumped
305 * @param irg the IR graph to dump
306 * @param suffix the suffix for the dumper
307 * @param dumper the dumper to be called
309 static void dump(int mask, ir_graph *irg, const char *suffix,
310 void (*dumper)(ir_graph *, const char *))
312 if(dump_flags & mask)
313 be_dump(irg, suffix, dumper);
317 * Prepare a backend graph for code generation.
319 static void prepare_graph(be_irg_t *birg)
321 ir_graph *irg = birg->irg;
323 /* Normalize proj nodes. */
324 normalize_proj_nodes(irg);
326 /* Make just one return node. */
327 normalize_one_return(irg);
329 /* Remove critical edges */
330 remove_critical_cf_edges(irg);
332 /* Compute the dominance information. */
336 /* Ensure, that the ir_edges are computed. */
339 /* check, if the dominance property is fulfilled. */
340 be_check_dominance(irg);
342 /* reset the phi handler. */
343 be_phi_handler_reset(birg->main_env->phi_handler);
348 #define BE_TIMER_PUSH(timer) \
349 if (be_options.timing == BE_TIME_ON) { \
350 int res = lc_timer_push(timer); \
351 if (vrfy_option == BE_VRFY_ASSERT) \
352 assert(res && "Timer already on stack, cannot be pushed twice."); \
353 else if (vrfy_option == BE_VRFY_WARN && ! res) \
354 fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \
355 lc_timer_get_name(timer)); \
357 #define BE_TIMER_POP(timer) \
358 if (be_options.timing == BE_TIME_ON) { \
359 lc_timer_t *tmp = lc_timer_pop(); \
360 if (vrfy_option == BE_VRFY_ASSERT) \
361 assert(tmp == timer && "Attempt to pop wrong timer."); \
362 else if (vrfy_option == BE_VRFY_WARN && tmp != timer) \
363 fprintf(stderr, "Attempt to pop wrong timer. %s is on stack, trying to pop %s.\n", \
364 lc_timer_get_name(tmp), lc_timer_get_name(timer)); \
368 #define BE_TIMER_ONLY(code) do { if (be_options.timing == BE_TIME_ON) { code; } } while(0)
372 #define BE_TIMER_PUSH(timer)
373 #define BE_TIMER_POP(timer)
374 #define BE_TIMER_ONLY(code)
376 #endif /* WITH_LIBCORE */
380 * The Firm backend main loop.
381 * Do architecture specific lowering for all graphs
382 * and call the architecture specific code generator.
384 * @param file_handle the file handle the output will be written to
386 static void be_main_loop(FILE *file_handle)
391 unsigned num_nodes_b = 0;
392 unsigned num_nodes_a = 0;
393 unsigned num_nodes_r = 0;
396 lc_timer_t *t_prolog, *t_abi, *t_codegen, *t_sched, *t_constr, *t_regalloc, *t_finish, *t_emit, *t_other, *t_verify;
397 be_ra_timer_t *ra_timer;
399 if (be_options.timing == BE_TIME_ON) {
400 t_prolog = lc_timer_register("prolog", "prolog");
401 t_abi = lc_timer_register("beabi", "be abi introduction");
402 t_codegen = lc_timer_register("codegen", "codegeneration");
403 t_sched = lc_timer_register("sched", "scheduling");
404 t_constr = lc_timer_register("constr", "assure constraints");
405 t_regalloc = lc_timer_register("regalloc", "register allocation");
406 t_finish = lc_timer_register("finish", "graph finish");
407 t_emit = lc_timer_register("emiter", "code emiter");
408 t_verify = lc_timer_register("verify", "graph verification");
409 t_other = lc_timer_register("other", "other");
411 #endif /* WITH_LIBCORE */
413 be_init_env(&env, file_handle);
415 isa = arch_env_get_isa(env.arch_env);
417 /* for debugging, anchors helps */
418 // dump_all_anchors(1);
421 for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
422 ir_graph *irg = get_irp_irg(i);
423 const arch_code_generator_if_t *cg_if;
425 optimization_state_t state;
427 /* stop and reset timers */
429 LC_STOP_AND_RESET_TIMER(t_prolog);
430 LC_STOP_AND_RESET_TIMER(t_abi);
431 LC_STOP_AND_RESET_TIMER(t_codegen);
432 LC_STOP_AND_RESET_TIMER(t_sched);
433 LC_STOP_AND_RESET_TIMER(t_constr);
434 LC_STOP_AND_RESET_TIMER(t_regalloc);
435 LC_STOP_AND_RESET_TIMER(t_finish);
436 LC_STOP_AND_RESET_TIMER(t_emit);
437 LC_STOP_AND_RESET_TIMER(t_verify);
438 LC_STOP_AND_RESET_TIMER(t_other);
440 BE_TIMER_PUSH(t_other); /* t_other */
442 BE_TIMER_ONLY(num_nodes_b = get_num_reachable_nodes(irg));
445 birg.main_env = &env;
447 edges_deactivate_kind(irg, EDGE_KIND_DEP);
448 edges_activate_kind(irg, EDGE_KIND_DEP);
450 DBG((env.dbg, LEVEL_2, "====> IRG: %F\n", irg));
451 dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph);
453 BE_TIMER_PUSH(t_prolog);
455 be_stat_init_irg(env.arch_env, irg);
456 be_do_stat_nodes(irg, "01 Begin");
458 /* set the current graph (this is important for several firm functions) */
459 current_ir_graph = birg.irg;
461 /* Get the code generator interface. */
462 cg_if = isa->impl->get_code_generator_if(isa);
464 /* get a code generator for this graph. */
465 birg.cg = cg_if->init(&birg);
467 /* create the code generator and generate code. */
468 prepare_graph(&birg);
470 BE_TIMER_POP(t_prolog);
472 /* some transformations need to be done before abi introduce */
473 BE_TIMER_PUSH(t_codegen);
474 arch_code_generator_before_abi(birg.cg);
475 BE_TIMER_POP(t_codegen);
477 /* implement the ABI conventions. */
478 BE_TIMER_PUSH(t_abi);
479 birg.abi = be_abi_introduce(&birg);
482 dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph);
483 be_do_stat_nodes(irg, "02 Abi");
486 BE_TIMER_PUSH(t_codegen);
487 arch_code_generator_prepare_graph(birg.cg);
488 BE_TIMER_POP(t_codegen);
490 be_do_stat_nodes(irg, "03 Prepare");
493 * Since the code generator made a lot of new nodes and skipped
494 * a lot of old ones, we should do dead node elimination here.
495 * Note that this requires disabling the edges here.
497 edges_deactivate(irg);
498 //dead_node_elimination(irg);
501 /* Compute loop nesting information (for weighting copies) */
502 construct_cf_backedges(irg);
503 dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph);
504 BE_TIMER_ONLY(num_nodes_r = get_num_reachable_nodes(irg));
506 /* let backend prepare scheduling */
507 BE_TIMER_PUSH(t_codegen);
508 arch_code_generator_before_sched(birg.cg);
509 BE_TIMER_POP(t_codegen);
511 /* schedule the irg */
512 BE_TIMER_PUSH(t_sched);
513 list_sched(&birg, &be_options);
514 BE_TIMER_POP(t_sched);
516 dump(DUMP_SCHED, irg, "-sched", dump_ir_block_graph_sched);
519 BE_TIMER_PUSH(t_verify);
520 be_sched_vrfy(birg.irg, vrfy_option);
521 BE_TIMER_POP(t_verify);
523 be_do_stat_nodes(irg, "04 Schedule");
525 /* introduce patterns to assure constraints */
526 BE_TIMER_PUSH(t_constr);
527 /* we switch off optimizations here, because they might cause trouble */
528 save_optimization_state(&state);
530 set_opt_normalize(0);
532 /* add Keeps for should_be_different constrained nodes */
533 /* beware: needs schedule due to usage of be_ssa_constr */
534 assure_constraints(&birg);
535 BE_TIMER_POP(t_constr);
537 dump(DUMP_SCHED, irg, "-assured", dump_ir_block_graph_sched);
538 be_do_stat_nodes(irg, "05 Constraints");
540 /* connect all stack modifying nodes together (see beabi.c) */
541 BE_TIMER_PUSH(t_abi);
542 be_abi_fix_stack_nodes(birg.abi, NULL);
545 dump(DUMP_SCHED, irg, "-fix_stack", dump_ir_block_graph_sched);
548 BE_TIMER_PUSH(t_verify);
549 be_sched_vrfy(birg.irg, vrfy_option);
550 BE_TIMER_POP(t_verify);
552 /* do some statistics */
553 be_do_stat_reg_pressure(&birg);
555 /* stuff needs to be done after scheduling but before register allocation */
556 BE_TIMER_PUSH(t_codegen);
557 arch_code_generator_before_ra(birg.cg);
558 BE_TIMER_POP(t_codegen);
560 /* Do register allocation */
561 BE_TIMER_PUSH(t_regalloc);
562 ra_timer = ra->allocate(&birg);
563 BE_TIMER_PUSH(t_regalloc);
565 dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched);
566 be_do_stat_nodes(irg, "06 Register Allocation");
568 /* let the codegenerator prepare the graph for emitter */
569 BE_TIMER_PUSH(t_finish);
570 arch_code_generator_after_ra(birg.cg);
571 BE_TIMER_POP(t_finish);
573 /* fix stack offsets */
574 BE_TIMER_PUSH(t_abi);
575 //be_abi_fix_stack_bias(birg.abi);
578 BE_TIMER_PUSH(t_finish);
579 arch_code_generator_finish(birg.cg);
580 BE_TIMER_POP(t_finish);
582 /* fix stack offsets */
583 BE_TIMER_PUSH(t_abi);
584 be_abi_fix_stack_nodes(birg.abi, NULL);
585 be_remove_dead_nodes_from_schedule(birg.irg);
586 be_abi_fix_stack_bias(birg.abi);
589 dump(DUMP_FINAL, irg, "-finish", dump_ir_block_graph_sched);
591 /* check schedule and register allocation */
592 BE_TIMER_PUSH(t_verify);
593 if (vrfy_option == BE_VRFY_WARN) {
594 //irg_verify(birg.irg, VRFY_ENFORCE_SSA);
595 be_check_dominance(birg.irg);
596 be_verify_schedule(birg.irg);
597 be_verify_register_allocation(env.arch_env, birg.irg);
599 else if (vrfy_option == BE_VRFY_ASSERT) {
600 //assert(irg_verify(birg.irg, VRFY_ENFORCE_SSA) && "irg verification failed");
601 assert(be_check_dominance(birg.irg) && "Dominance verification failed");
602 assert(be_verify_schedule(birg.irg) && "Schedule verification failed");
603 assert(be_verify_register_allocation(env.arch_env, birg.irg)
604 && "register allocation verification failed");
606 BE_TIMER_POP(t_verify);
608 /* emit assembler code */
609 BE_TIMER_PUSH(t_emit);
610 arch_code_generator_done(birg.cg);
611 BE_TIMER_POP(t_emit);
613 dump(DUMP_FINAL, irg, "-end", dump_ir_extblock_graph_sched);
615 BE_TIMER_PUSH(t_abi);
616 be_abi_free(birg.abi);
619 be_do_stat_nodes(irg, "07 Final");
620 restore_optimization_state(&state);
622 BE_TIMER_ONLY(num_nodes_a = get_num_reachable_nodes(irg));
623 BE_TIMER_POP(t_other);
625 #define LC_EMIT(timer) printf("%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
626 #define LC_EMIT_RA(timer) printf("\t%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
628 printf("==>> IRG %s <<==\n", get_entity_name(get_irg_entity(irg)));
629 printf("# nodes at begin: %u\n", num_nodes_b);
630 printf("# nodes before ra: %u\n", num_nodes_r);
631 printf("# nodes at end: %u\n\n", num_nodes_a);
638 LC_EMIT_RA(ra_timer->t_prolog);
639 LC_EMIT_RA(ra_timer->t_live);
640 LC_EMIT_RA(ra_timer->t_spill);
641 LC_EMIT_RA(ra_timer->t_spillslots);
642 LC_EMIT_RA(ra_timer->t_color);
643 LC_EMIT_RA(ra_timer->t_ifg);
644 LC_EMIT_RA(ra_timer->t_copymin);
645 LC_EMIT_RA(ra_timer->t_ssa);
646 LC_EMIT_RA(ra_timer->t_epilog);
647 LC_EMIT_RA(ra_timer->t_verify);
648 LC_EMIT_RA(ra_timer->t_other);
657 /* switched off due to statistics (statistic module needs all irgs) */
658 if (! stat_is_active())
669 /* Main interface to the frontend. */
670 void be_main(FILE *file_handle)
674 #endif /* WITH_LIBCORE */
677 /* The user specified another config file to read. do that now. */
678 if(strlen(config_file) > 0) {
681 if((f = fopen(config_file, "rt")) != NULL) {
682 lc_opt_from_file(config_file, f, NULL);
687 if (be_options.timing == BE_TIME_ON) {
688 t = lc_timer_register("bemain", "measure complete bemain loop");
690 if (lc_timer_enter_high_priority()) {
691 fprintf(stderr, "Warning: Could not enter high priority mode.\n");
694 lc_timer_reset_and_start(t);
696 #endif /* WITH_LIBCORE */
698 /* never build code for pseudo irgs */
699 set_visit_pseudo_irgs(0);
702 be_main_loop(file_handle);
705 if (be_options.timing == BE_TIME_ON) {
707 lc_timer_leave_high_priority();
708 printf("%-20s: %lu msec\n", "BEMAINLOOP", lc_timer_elapsed_msec(t));
710 #endif /* WITH_LIBCORE */
713 /** The debug info retriever function. */
714 static retrieve_dbg_func retrieve_dbg = NULL;
716 /* Sets a debug info retriever. */
717 void be_set_debug_retrieve(retrieve_dbg_func func) {
721 /* Retrieve the debug info. */
722 const char *be_retrieve_dbg_info(const dbg_info *dbg, unsigned *line) {
724 return retrieve_dbg(dbg, line);
729 int be_put_ignore_regs(const be_irg_t *birg, const arch_register_class_t *cls, bitset_t *bs)
732 bs = bitset_alloca(cls->n_regs);
734 bitset_clear_all(bs);
736 assert(bitset_size(bs) == (unsigned) cls->n_regs);
737 arch_put_non_ignore_regs(birg->main_env->arch_env, cls, bs);
739 be_abi_put_ignore_regs(birg->abi, cls, bs);
740 return bitset_popcnt(bs);