3 * @author Sebastian Hack
15 #include <libcore/lc_opts.h>
16 #include <libcore/lc_opts_enum.h>
17 #include <libcore/lc_timing.h>
18 #endif /* WITH_LIBCORE */
29 #include "iredges_t.h"
39 #include "firm/bearch_firm.h"
40 #include "ia32/bearch_ia32.h"
41 #include "arm/bearch_arm.h"
42 #include "ppc32/bearch_ppc32.h"
43 #include "mips/bearch_mips.h"
50 #include "besched_t.h"
51 #include "belistsched.h"
53 #include "bespillbelady.h"
55 #include "beraextern.h"
56 #include "bechordal_t.h"
58 #include "beifg_impl.h"
59 #include "becopyopt.h"
60 #include "becopystat.h"
61 #include "bessadestr.h"
64 #include "beschedmris.h"
67 #include "beprofile.h"
68 #include "be_dbgout.h"
70 /* options visible for anyone */
71 static be_options_t be_options = {
72 DUMP_NONE, /* dump flags */
73 BE_TIME_OFF, /* no timing */
74 0, /* no opt profile */
75 1, /* try to omit frame pointer */
76 0, /* always stabs debugging output */
77 BE_VRFY_WARN, /* verification level: warn */
78 "i44pc52.info.uni-karlsruhe.de", /* ilp server */
79 "cplex" /* ilp solver */
83 static char config_file[256] = { 0 };
85 /* register allocator to use. */
86 static const be_ra_t *ra = &be_ra_chordal_allocator;
88 /* back end instruction set architecture to use */
89 static const arch_isa_if_t *isa_if = &ia32_isa_if;
93 static lc_opt_entry_t *be_grp_root = NULL;
95 /* possible dumping options */
96 static const lc_opt_enum_mask_items_t dump_items[] = {
97 { "none", DUMP_NONE },
98 { "initial", DUMP_INITIAL },
100 { "sched", DUMP_SCHED },
101 { "prepared", DUMP_PREPARED },
102 { "regalloc", DUMP_RA },
103 { "final", DUMP_FINAL },
105 { "all", 2 * DUMP_BE - 1 },
109 /* register allocators */
110 static const lc_opt_enum_const_ptr_items_t ra_items[] = {
111 { "chordal", &be_ra_chordal_allocator },
112 { "external", &be_ra_external_allocator },
116 /* instruction set architectures. */
117 static const lc_opt_enum_const_ptr_items_t isa_items[] = {
118 { "ia32", &ia32_isa_if },
120 { "arm", &arm_isa_if },
121 { "ppc32", &ppc32_isa_if },
122 { "mips", &mips_isa_if },
127 /* verify options. */
128 static const lc_opt_enum_int_items_t vrfy_items[] = {
129 { "off", BE_VRFY_OFF },
130 { "warn", BE_VRFY_WARN },
131 { "assert", BE_VRFY_ASSERT },
135 static lc_opt_enum_mask_var_t dump_var = {
136 &be_options.dump_flags, dump_items
139 static lc_opt_enum_const_ptr_var_t ra_var = {
140 (const void **) &ra, ra_items
143 static lc_opt_enum_const_ptr_var_t isa_var = {
144 (const void **) &isa_if, isa_items
147 static lc_opt_enum_int_var_t vrfy_var = {
148 &be_options.vrfy_option, vrfy_items
151 static const lc_opt_table_entry_t be_main_options[] = {
152 LC_OPT_ENT_STR ("config", "read another config file containing backend options", config_file, sizeof(config_file)),
153 LC_OPT_ENT_ENUM_MASK("dump", "dump irg on several occasions", &dump_var),
154 LC_OPT_ENT_ENUM_PTR ("ra", "register allocator", &ra_var),
155 LC_OPT_ENT_ENUM_PTR ("isa", "the instruction set architecture", &isa_var),
156 LC_OPT_ENT_NEGBOOL ("noomitfp", "do not omit frame pointer", &be_options.omit_fp),
157 LC_OPT_ENT_BOOL ("stabs", "enable stabs debug support", &be_options.stabs_debug_support),
158 LC_OPT_ENT_ENUM_PTR ("vrfy", "verify the backend irg (off, warn, assert)", &vrfy_var),
159 LC_OPT_ENT_BOOL ("time", "get backend timing statistics", &be_options.timing),
160 LC_OPT_ENT_BOOL ("profile", "instrument the code for execution count profiling", &be_options.opt_profile),
163 LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)),
164 LC_OPT_ENT_STR ("ilp.solver", "the ilp solver name", be_options.ilp_solver, sizeof(be_options.ilp_solver)),
165 #endif /* WITH_ILP */
169 #endif /* WITH_LIBCORE */
171 void be_opt_register(void)
175 lc_opt_entry_t *be_grp_ra;
176 static int run_once = 0;
180 be_grp_root = lc_opt_get_grp(firm_opt_get_root(), "be");
181 be_grp_ra = lc_opt_get_grp(be_grp_root, "ra");
183 lc_opt_add_table(be_grp_root, be_main_options);
185 /* register allocator options */
186 for(i = 0; ra_items[i].name != NULL; ++i) {
187 const be_ra_t *ra = ra_items[i].value;
188 ra->register_options(be_grp_ra);
191 /* register isa options */
192 for(i = 0; isa_items[i].name != NULL; ++i) {
193 const arch_isa_if_t *isa = isa_items[i].value;
194 isa->register_options(be_grp_root);
197 /* scheduler register options */
198 list_sched_register_options(be_grp_root);
200 #endif /* WITH_LIBCORE */
203 /* Parse one argument. */
204 int be_parse_arg(const char *arg) {
206 if (strcmp(arg, "help") == 0 || (arg[0] == '?' && arg[1] == '\0')) {
207 lc_opt_print_help(be_grp_root, stdout);
210 return lc_opt_from_single_arg(be_grp_root, NULL, arg, NULL);
213 #endif /* WITH_LIBCORE */
216 /** The be parameters returned by default, all off. */
217 const static backend_params be_params = {
224 /* Perform schedule verification if requested. */
225 static void be_sched_vrfy(ir_graph *irg, int vrfy_opt) {
226 if (vrfy_opt == BE_VRFY_WARN) {
227 be_verify_schedule(irg);
229 else if (vrfy_opt == BE_VRFY_ASSERT) {
230 assert(be_verify_schedule(irg) && "Schedule verification failed.");
234 /* Initialize the Firm backend. Must be run BEFORE init_firm()! */
235 const backend_params *be_init(void)
245 if (isa_if->get_params)
246 return isa_if->get_params();
251 * Initializes the main environment for the backend.
253 * @param env an empty environment
254 * @param file_handle the file handle where the output will be written to
256 static be_main_env_t *be_init_env(be_main_env_t *env, FILE *file_handle)
258 memset(env, 0, sizeof(*env));
259 obstack_init(&env->obst);
260 env->arch_env = obstack_alloc(&env->obst, sizeof(env->arch_env[0]));
261 env->options = &be_options;
262 FIRM_DBG_REGISTER(env->dbg, "be.main");
264 arch_env_init(env->arch_env, isa_if, file_handle, env);
266 /* Register the irn handler of the architecture */
267 if (arch_isa_get_irn_handler(env->arch_env->isa))
268 arch_env_push_irn_handler(env->arch_env, arch_isa_get_irn_handler(env->arch_env->isa));
271 * Register the node handler of the back end infrastructure.
272 * This irn handler takes care of the platform independent
273 * spill, reload and perm nodes.
275 arch_env_push_irn_handler(env->arch_env, &be_node_irn_handler);
276 env->phi_handler = be_phi_handler_new(env->arch_env);
277 arch_env_push_irn_handler(env->arch_env, env->phi_handler);
279 env->db_handle = be_options.stabs_debug_support ? be_stabs_open(file_handle) : be_nulldbg_open();
283 static void be_done_env(be_main_env_t *env)
285 env->arch_env->isa->impl->done(env->arch_env->isa);
286 be_dbg_close(env->db_handle);
287 be_phi_handler_free(env->phi_handler);
288 obstack_free(&env->obst, NULL);
292 * A wrapper around a firm dumper. Dumps only, if
295 * @param mask a bitmask containing the reason what will be dumped
296 * @param irg the IR graph to dump
297 * @param suffix the suffix for the dumper
298 * @param dumper the dumper to be called
300 static void dump(int mask, ir_graph *irg, const char *suffix,
301 void (*dumper)(ir_graph *, const char *))
303 if(be_options.dump_flags & mask)
304 be_dump(irg, suffix, dumper);
308 * Prepare a backend graph for code generation and initialize its birg
310 static void initialize_birg(be_irg_t *birg, ir_graph *irg, be_main_env_t *env)
312 memset(birg, 0, sizeof(*birg));
314 birg->main_env = env;
316 edges_deactivate_kind(irg, EDGE_KIND_DEP);
317 edges_activate_kind(irg, EDGE_KIND_DEP);
319 DBG((env->dbg, LEVEL_2, "====> IRG: %F\n", irg));
320 dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph);
322 be_stat_init_irg(env->arch_env, irg);
323 be_do_stat_nodes(irg, "01 Begin");
325 /* set the current graph (this is important for several firm functions) */
326 current_ir_graph = irg;
328 /* Normalize proj nodes. */
329 normalize_proj_nodes(irg);
331 /* Make just one return node. */
332 normalize_one_return(irg);
334 /* Remove critical edges */
335 remove_critical_cf_edges(irg);
337 /* Compute the dominance information. */
341 /* Ensure, that the ir_edges are computed. */
344 /* check, if the dominance property is fulfilled. */
345 be_check_dominance(irg);
347 /* reset the phi handler. */
348 be_phi_handler_reset(env->phi_handler);
353 #define BE_TIMER_PUSH(timer) \
354 if (be_options.timing == BE_TIME_ON) { \
355 int res = lc_timer_push(timer); \
356 if (be_options.vrfy_option == BE_VRFY_ASSERT) \
357 assert(res && "Timer already on stack, cannot be pushed twice."); \
358 else if (be_options.vrfy_option == BE_VRFY_WARN && ! res) \
359 fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \
360 lc_timer_get_name(timer)); \
362 #define BE_TIMER_POP(timer) \
363 if (be_options.timing == BE_TIME_ON) { \
364 lc_timer_t *tmp = lc_timer_pop(); \
365 if (be_options.vrfy_option == BE_VRFY_ASSERT) \
366 assert(tmp == timer && "Attempt to pop wrong timer."); \
367 else if (be_options.vrfy_option == BE_VRFY_WARN && tmp != timer) \
368 fprintf(stderr, "Attempt to pop wrong timer. %s is on stack, trying to pop %s.\n", \
369 lc_timer_get_name(tmp), lc_timer_get_name(timer)); \
373 #define BE_TIMER_ONLY(code) do { if (be_options.timing == BE_TIME_ON) { code; } } while(0)
377 #define BE_TIMER_PUSH(timer)
378 #define BE_TIMER_POP(timer)
379 #define BE_TIMER_ONLY(code)
381 #endif /* WITH_LIBCORE */
385 * The Firm backend main loop.
386 * Do architecture specific lowering for all graphs
387 * and call the architecture specific code generator.
389 * @param file_handle the file handle the output will be written to
390 * @param cup_name name of the compilation unit
392 static void be_main_loop(FILE *file_handle, const char *cup_name)
397 unsigned num_nodes_b = 0;
398 unsigned num_nodes_a = 0;
399 unsigned num_nodes_r = 0;
400 char prof_filename[256];
401 static const char suffix[] = ".prof";
405 be_ra_timer_t *ra_timer;
408 lc_timer_t *t_abi = NULL;
409 lc_timer_t *t_codegen = NULL;
410 lc_timer_t *t_sched = NULL;
411 lc_timer_t *t_constr = NULL;
412 lc_timer_t *t_regalloc = NULL;
413 lc_timer_t *t_finish = NULL;
414 lc_timer_t *t_emit = NULL;
415 lc_timer_t *t_other = NULL;
416 lc_timer_t *t_verify = NULL;
418 if (be_options.timing == BE_TIME_ON) {
419 t_abi = lc_timer_register("beabi", "be abi introduction");
420 t_codegen = lc_timer_register("codegen", "codegeneration");
421 t_sched = lc_timer_register("sched", "scheduling");
422 t_constr = lc_timer_register("constr", "assure constraints");
423 t_regalloc = lc_timer_register("regalloc", "register allocation");
424 t_finish = lc_timer_register("finish", "graph finish");
425 t_emit = lc_timer_register("emiter", "code emiter");
426 t_verify = lc_timer_register("verify", "graph verification");
427 t_other = lc_timer_register("other", "other");
429 #endif /* WITH_LIBCORE */
431 be_init_env(&env, file_handle);
433 isa = arch_env_get_isa(env.arch_env);
435 be_dbg_so(env.db_handle, cup_name);
436 be_dbg_types(env.db_handle);
438 /* we might need 1 birg more for instrumentation constructor */
439 num_birgs = get_irp_n_irgs();
440 birgs = alloca(sizeof(birgs[0]) * (num_birgs + 1));
442 /* First: initialize all birgs */
443 for(i = 0; i < get_irp_n_irgs(); ++i) {
444 ir_graph *irg = get_irp_irg(i);
446 initialize_birg(&birgs[i], irg, &env);
450 Get the filename for the profiling data.
451 Beware: '\0' is already included in sizeof(suffix)
453 memset(prof_filename, 0, sizeof(prof_filename));
454 strncpy(prof_filename, cup_name, sizeof(prof_filename) - sizeof(suffix));
455 strcat(prof_filename, suffix);
458 Next: Either instruments all irgs with profiling code
459 or try to read in profile data for current translation unit.
461 if (be_options.opt_profile) {
462 ir_graph *prof_init_irg = be_profile_instrument(prof_filename, profile_default);
463 initialize_birg(&birgs[num_birgs], prof_init_irg, &env);
465 set_method_img_section(get_irg_entity(prof_init_irg), section_constructors);
468 be_profile_read(prof_filename);
472 for (i = 0; i < num_birgs; ++i) {
473 be_irg_t *birg = & birgs[i];
474 ir_graph *irg = birg->irg;
475 optimization_state_t state;
476 const arch_code_generator_if_t *cg_if;
478 /* stop and reset timers */
480 LC_STOP_AND_RESET_TIMER(t_abi);
481 LC_STOP_AND_RESET_TIMER(t_codegen);
482 LC_STOP_AND_RESET_TIMER(t_sched);
483 LC_STOP_AND_RESET_TIMER(t_constr);
484 LC_STOP_AND_RESET_TIMER(t_regalloc);
485 LC_STOP_AND_RESET_TIMER(t_finish);
486 LC_STOP_AND_RESET_TIMER(t_emit);
487 LC_STOP_AND_RESET_TIMER(t_verify);
488 LC_STOP_AND_RESET_TIMER(t_other);
490 BE_TIMER_PUSH(t_other); /* t_other */
493 * Create execution frequencies from profile data or estimate some
495 if(be_profile_has_data()) {
496 birg->execfreqs = be_create_execfreqs_from_profile(irg);
498 birg->execfreqs = compute_execfreq(irg, 10);
501 BE_TIMER_ONLY(num_nodes_b = get_num_reachable_nodes(irg));
503 /* Get the code generator interface. */
504 cg_if = isa->impl->get_code_generator_if(isa);
506 /* get a code generator for this graph. */
507 birg->cg = cg_if->init(birg);
509 /* some transformations need to be done before abi introduce */
510 arch_code_generator_before_abi(birg->cg);
512 /* set the current graph (this is important for several firm functions) */
513 current_ir_graph = irg;
515 /* reset the phi handler. */
516 be_phi_handler_reset(env.phi_handler);
518 /* implement the ABI conventions. */
519 BE_TIMER_PUSH(t_abi);
520 birg->abi = be_abi_introduce(birg);
523 dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph);
524 be_do_stat_nodes(irg, "02 Abi");
527 BE_TIMER_PUSH(t_codegen);
528 arch_code_generator_prepare_graph(birg->cg);
529 BE_TIMER_POP(t_codegen);
531 be_do_stat_nodes(irg, "03 Prepare");
534 Since the code generator made a lot of new nodes and skipped
535 a lot of old ones, we should do dead node elimination here.
536 Note that this requires disabling the edges here.
538 edges_deactivate(irg);
539 //dead_node_elimination(irg);
542 /* Compute loop nesting information (for weighting copies) */
543 construct_cf_backedges(irg);
544 dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph);
545 BE_TIMER_ONLY(num_nodes_r = get_num_reachable_nodes(irg));
547 /* let backend prepare scheduling */
548 BE_TIMER_PUSH(t_codegen);
549 arch_code_generator_before_sched(birg->cg);
550 BE_TIMER_POP(t_codegen);
552 /* schedule the irg */
553 BE_TIMER_PUSH(t_sched);
554 list_sched(birg, &be_options);
555 BE_TIMER_POP(t_sched);
557 dump(DUMP_SCHED, irg, "-sched", dump_ir_block_graph_sched);
560 BE_TIMER_PUSH(t_verify);
561 be_sched_vrfy(irg, be_options.vrfy_option);
562 BE_TIMER_POP(t_verify);
564 be_do_stat_nodes(irg, "04 Schedule");
566 /* introduce patterns to assure constraints */
567 BE_TIMER_PUSH(t_constr);
568 /* we switch off optimizations here, because they might cause trouble */
569 save_optimization_state(&state);
571 set_opt_normalize(0);
573 /* add Keeps for should_be_different constrained nodes */
574 /* beware: needs schedule due to usage of be_ssa_constr */
575 assure_constraints(birg);
576 BE_TIMER_POP(t_constr);
578 dump(DUMP_SCHED, irg, "-assured", dump_ir_block_graph_sched);
579 be_do_stat_nodes(irg, "05 Constraints");
581 /* connect all stack modifying nodes together (see beabi.c) */
582 BE_TIMER_PUSH(t_abi);
583 be_abi_fix_stack_nodes(birg->abi, NULL);
586 dump(DUMP_SCHED, irg, "-fix_stack", dump_ir_block_graph_sched);
589 BE_TIMER_PUSH(t_verify);
590 be_sched_vrfy(irg, be_options.vrfy_option);
591 BE_TIMER_POP(t_verify);
593 /* do some statistics */
594 be_do_stat_reg_pressure(birg);
596 /* stuff needs to be done after scheduling but before register allocation */
597 BE_TIMER_PUSH(t_codegen);
598 arch_code_generator_before_ra(birg->cg);
599 BE_TIMER_POP(t_codegen);
601 /* Do register allocation */
602 BE_TIMER_PUSH(t_regalloc);
603 ra_timer = ra->allocate(birg);
604 BE_TIMER_POP(t_regalloc);
606 dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched);
607 be_do_stat_nodes(irg, "06 Register Allocation");
609 /* let the code generator prepare the graph for emitter */
610 BE_TIMER_PUSH(t_finish);
611 arch_code_generator_after_ra(birg->cg);
612 BE_TIMER_POP(t_finish);
614 /* fix stack offsets */
615 BE_TIMER_PUSH(t_abi);
616 be_abi_fix_stack_nodes(birg->abi, NULL);
617 be_remove_dead_nodes_from_schedule(irg);
618 be_abi_fix_stack_bias(birg->abi);
621 BE_TIMER_PUSH(t_finish);
622 arch_code_generator_finish(birg->cg);
623 BE_TIMER_POP(t_finish);
625 dump(DUMP_FINAL, irg, "-finish", dump_ir_block_graph_sched);
627 /* check schedule and register allocation */
628 BE_TIMER_PUSH(t_verify);
629 if (be_options.vrfy_option == BE_VRFY_WARN) {
630 //irg_verify(irg, VRFY_ENFORCE_SSA);
631 be_check_dominance(irg);
632 be_verify_out_edges(irg);
633 be_verify_schedule(irg);
634 be_verify_register_allocation(env.arch_env, irg);
636 else if (be_options.vrfy_option == BE_VRFY_ASSERT) {
637 //assert(irg_verify(irg, VRFY_ENFORCE_SSA) && "irg verification failed");
638 assert(be_verify_out_edges(irg));
639 assert(be_check_dominance(irg) && "Dominance verification failed");
640 assert(be_verify_schedule(irg) && "Schedule verification failed");
641 assert(be_verify_register_allocation(env.arch_env, irg)
642 && "register allocation verification failed");
644 BE_TIMER_POP(t_verify);
646 /* emit assembler code */
647 BE_TIMER_PUSH(t_emit);
648 arch_code_generator_done(birg->cg);
649 BE_TIMER_POP(t_emit);
651 dump(DUMP_FINAL, irg, "-end", dump_ir_extblock_graph_sched);
653 BE_TIMER_PUSH(t_abi);
654 be_abi_free(birg->abi);
657 be_do_stat_nodes(irg, "07 Final");
658 restore_optimization_state(&state);
660 BE_TIMER_ONLY(num_nodes_a = get_num_reachable_nodes(irg));
661 BE_TIMER_POP(t_other);
663 #define LC_EMIT(timer) printf("%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
664 #define LC_EMIT_RA(timer) printf("\t%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
666 printf("==>> IRG %s <<==\n", get_entity_name(get_irg_entity(irg)));
667 printf("# nodes at begin: %u\n", num_nodes_b);
668 printf("# nodes before ra: %u\n", num_nodes_r);
669 printf("# nodes at end: %u\n\n", num_nodes_a);
675 LC_EMIT_RA(ra_timer->t_prolog);
676 LC_EMIT_RA(ra_timer->t_live);
677 LC_EMIT_RA(ra_timer->t_spill);
678 LC_EMIT_RA(ra_timer->t_spillslots);
679 LC_EMIT_RA(ra_timer->t_color);
680 LC_EMIT_RA(ra_timer->t_ifg);
681 LC_EMIT_RA(ra_timer->t_copymin);
682 LC_EMIT_RA(ra_timer->t_ssa);
683 LC_EMIT_RA(ra_timer->t_epilog);
684 LC_EMIT_RA(ra_timer->t_verify);
685 LC_EMIT_RA(ra_timer->t_other);
694 free_execfreq(birg->execfreqs);
696 /* switched off due to statistics (statistic module needs all irgs) */
697 if (! stat_is_active())
708 /* Main interface to the frontend. */
709 void be_main(FILE *file_handle, const char *cup_name)
712 lc_timer_t *t = NULL;
713 #endif /* WITH_LIBCORE */
716 /* The user specified another config file to read. do that now. */
717 if(strlen(config_file) > 0) {
720 if((f = fopen(config_file, "rt")) != NULL) {
721 lc_opt_from_file(config_file, f, NULL);
726 if (be_options.timing == BE_TIME_ON) {
727 t = lc_timer_register("bemain", "measure complete bemain loop");
729 if (lc_timer_enter_high_priority()) {
730 fprintf(stderr, "Warning: Could not enter high priority mode.\n");
733 lc_timer_reset_and_start(t);
735 #endif /* WITH_LIBCORE */
737 /* never build code for pseudo irgs */
738 set_visit_pseudo_irgs(0);
741 be_main_loop(file_handle, cup_name);
744 if (be_options.timing == BE_TIME_ON) {
746 lc_timer_leave_high_priority();
747 printf("%-20s: %lu msec\n", "BEMAINLOOP", lc_timer_elapsed_msec(t));
749 #endif /* WITH_LIBCORE */
752 /** The debug info retriever function. */
753 static retrieve_dbg_func retrieve_dbg = NULL;
755 /* Sets a debug info retriever. */
756 void be_set_debug_retrieve(retrieve_dbg_func func) {
760 /* Retrieve the debug info. */
761 const char *be_retrieve_dbg_info(const dbg_info *dbg, unsigned *line) {
763 return retrieve_dbg(dbg, line);
768 int be_put_ignore_regs(const be_irg_t *birg, const arch_register_class_t *cls, bitset_t *bs)
771 bs = bitset_alloca(cls->n_regs);
773 bitset_clear_all(bs);
775 assert(bitset_size(bs) == (unsigned) cls->n_regs);
776 arch_put_non_ignore_regs(birg->main_env->arch_env, cls, bs);
778 be_abi_put_ignore_regs(birg->abi, cls, bs);
779 return bitset_popcnt(bs);