X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbemain.c;h=4b109caedefd6fcb133d1ed0abed4ea7f8149f91;hb=505e4b0f5056fdc1553842e138c23572172ce565;hp=15c811870df82ed22d0adf43982a0c8b707a842c;hpb=71dc2f5649c2c7024756568d87e8aa19a0136cf3;p=libfirm diff --git a/ir/be/bemain.c b/ir/be/bemain.c index 15c811870..4b109caed 100644 --- a/ir/be/bemain.c +++ b/ir/be/bemain.c @@ -1,17 +1,20 @@ /** * Backend driver. * @author Sebastian Hack - * @date 25.11.2004 + * @date 25.11.2004 + * @cvsid $Id$ */ #ifdef HAVE_CONFIG_H #include "config.h" #endif #include +#include #ifdef WITH_LIBCORE #include #include +#include #endif /* WITH_LIBCORE */ #include "obst.h" @@ -26,7 +29,12 @@ #include "iredges_t.h" #include "irloop_t.h" #include "irtools.h" +#include "irvrfy.h" +#include "irprintf.h" #include "return.h" +#include "firmstat.h" +#include "cfopt.h" +#include "execfreq.h" #include "bearch.h" #include "firm/bearch_firm.h" @@ -34,6 +42,7 @@ #include "arm/bearch_arm.h" #include "ppc32/bearch_ppc32.h" #include "mips/bearch_mips.h" +// #include "sta/bearch_sta.h" #include "be_t.h" #include "benumb_t.h" @@ -43,7 +52,6 @@ #include "besched_t.h" #include "belistsched.h" #include "belive_t.h" -#include "bespillilp.h" #include "bespillbelady.h" #include "bera.h" #include "beraextern.h" @@ -55,67 +63,95 @@ #include "bessadestr.h" #include "beabi.h" #include "belower.h" +#include "beschedmris.h" +#include "bestat.h" +#include "beverify.h" +#include "beprofile.h" +#include "beblocksched.h" +#include "be_dbgout.h" -#define DUMP_INITIAL (1 << 0) -#define DUMP_ABI (1 << 1) -#define DUMP_SCHED (1 << 2) -#define DUMP_PREPARED (1 << 3) -#define DUMP_RA (1 << 4) -#define DUMP_FINAL (1 << 5) +#ifdef WITH_ILP +#include "beilpsched.h" +#endif /* WITH_ILP */ /* options visible for anyone */ static be_options_t be_options = { - /* ilp server */ - "i44pc52.info.uni-karlsruhe.de", - - /* ilp solver */ - "cplex" + DUMP_NONE, /* dump flags */ + BE_TIME_OFF, /* no timing */ + 0, /* no opt profile */ + 1, /* try to omit frame pointer */ + 0, /* always stabs debugging output */ + BE_VRFY_WARN, /* verification level: warn */ + BE_SCHED_LIST, /* scheduler: list scheduler */ + "i44pc52.info.uni-karlsruhe.de", /* ilp server */ + "cplex" /* ilp solver */ }; -/* dump flags */ -static unsigned dump_flags = 2 * DUMP_FINAL - 1; +/* config file. */ +static char config_file[256] = { 0 }; /* register allocator to use. */ static const be_ra_t *ra = &be_ra_chordal_allocator; /* back end instruction set architecture to use */ static const arch_isa_if_t *isa_if = &ia32_isa_if; + #ifdef WITH_LIBCORE static lc_opt_entry_t *be_grp_root = NULL; /* possible dumping options */ static const lc_opt_enum_mask_items_t dump_items[] = { - { "none", 0 }, + { "none", DUMP_NONE }, { "initial", DUMP_INITIAL }, { "abi", DUMP_ABI }, { "sched", DUMP_SCHED }, { "prepared", DUMP_PREPARED }, { "regalloc", DUMP_RA }, { "final", DUMP_FINAL }, - { "all", 2 * DUMP_FINAL - 1 }, + { "be", DUMP_BE }, + { "all", 2 * DUMP_BE - 1 }, { NULL, 0 } }; /* register allocators */ static const lc_opt_enum_const_ptr_items_t ra_items[] = { - { "chordal", &be_ra_chordal_allocator }, + { "chordal", &be_ra_chordal_allocator }, { "external", &be_ra_external_allocator }, { NULL, NULL } }; /* instruction set architectures. */ static const lc_opt_enum_const_ptr_items_t isa_items[] = { - { "firm", &firm_isa }, { "ia32", &ia32_isa_if }, +#if 0 + { "sta", &sta_isa_if }, { "arm", &arm_isa_if }, - { "ppc", &ppc32_isa_if }, + { "ppc32", &ppc32_isa_if }, { "mips", &mips_isa_if }, +#endif { NULL, NULL } }; +/* verify options. */ +static const lc_opt_enum_int_items_t vrfy_items[] = { + { "off", BE_VRFY_OFF }, + { "warn", BE_VRFY_WARN }, + { "assert", BE_VRFY_ASSERT }, + { NULL, 0 } +}; + +/* scheduling options. */ +static const lc_opt_enum_int_items_t sched_items[] = { + { "list", BE_SCHED_LIST }, +#ifdef WITH_ILP + { "ilp", BE_SCHED_ILP }, +#endif /* WITH_ILP */ + { NULL, 0 } +}; + static lc_opt_enum_mask_var_t dump_var = { - &dump_flags, dump_items + &be_options.dump_flags, dump_items }; static lc_opt_enum_const_ptr_var_t ra_var = { @@ -126,10 +162,26 @@ static lc_opt_enum_const_ptr_var_t isa_var = { (const void **) &isa_if, isa_items }; +static lc_opt_enum_int_var_t vrfy_var = { + &be_options.vrfy_option, vrfy_items +}; + +static lc_opt_enum_int_var_t sched_var = { + &be_options.scheduler, sched_items +}; + static const lc_opt_table_entry_t be_main_options[] = { - LC_OPT_ENT_ENUM_MASK("dump", "dump irg on several occasions", &dump_var), - LC_OPT_ENT_ENUM_PTR("ra", "register allocator", &ra_var), - LC_OPT_ENT_ENUM_PTR("isa", "the instruction set architecture", &isa_var), + LC_OPT_ENT_STR ("config", "read another config file containing backend options", config_file, sizeof(config_file)), + LC_OPT_ENT_ENUM_MASK("dump", "dump irg on several occasions", &dump_var), + LC_OPT_ENT_ENUM_PTR ("ra", "register allocator", &ra_var), + LC_OPT_ENT_ENUM_PTR ("isa", "the instruction set architecture", &isa_var), + LC_OPT_ENT_NEGBOOL ("noomitfp", "do not omit frame pointer", &be_options.omit_fp), + LC_OPT_ENT_BOOL ("stabs", "enable stabs debug support", &be_options.stabs_debug_support), + LC_OPT_ENT_ENUM_PTR ("vrfy", "verify the backend irg", &vrfy_var), + LC_OPT_ENT_BOOL ("time", "get backend timing statistics", &be_options.timing), + LC_OPT_ENT_BOOL ("profile", "instrument the code for execution count profiling", &be_options.opt_profile), + LC_OPT_ENT_ENUM_PTR ("sched", "select a scheduler", &sched_var), + LC_OPT_ENT_STR ("statfile", "append statistics to file statfile", &be_options.stat_file_name, sizeof(be_options.stat_file_name)), #ifdef WITH_ILP LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)), @@ -145,40 +197,94 @@ void be_opt_register(void) #ifdef WITH_LIBCORE int i; lc_opt_entry_t *be_grp_ra; + static int run_once = 0; - be_grp_root = lc_opt_get_grp(firm_opt_get_root(), "be"); - be_grp_ra = lc_opt_get_grp(be_grp_root, "ra"); + if (! run_once) { + run_once = 1; + be_grp_root = lc_opt_get_grp(firm_opt_get_root(), "be"); + be_grp_ra = lc_opt_get_grp(be_grp_root, "ra"); - lc_opt_add_table(be_grp_root, be_main_options); + lc_opt_add_table(be_grp_root, be_main_options); - /* register allocator options */ - for(i = 0; ra_items[i].name != NULL; ++i) { - const be_ra_t *ra = ra_items[i].value; - ra->register_options(be_grp_ra); + /* register allocator options */ + for(i = 0; ra_items[i].name != NULL; ++i) { + const be_ra_t *ra = ra_items[i].value; + ra->register_options(be_grp_ra); + } + + /* register isa options */ + for(i = 0; isa_items[i].name != NULL; ++i) { + const arch_isa_if_t *isa = isa_items[i].value; + isa->register_options(be_grp_root); + } + + /* list scheduler register options */ + list_sched_register_options(be_grp_root); + +#ifdef WITH_ILP + /* ilp scheduler register options */ + ilpsched_register_options(be_grp_root); +#endif /* WITH_ILP */ + + be_block_schedule_register_options(be_grp_root); } +#endif /* WITH_LIBCORE */ +} - /* register isa options */ - for(i = 0; isa_items[i].name != NULL; ++i) { - const arch_isa_if_t *isa = isa_items[i].value; - isa->register_options(be_grp_root); +/* Parse one argument. */ +int be_parse_arg(const char *arg) { +#ifdef WITH_LIBCORE + if (strcmp(arg, "help") == 0 || (arg[0] == '?' && arg[1] == '\0')) { + lc_opt_print_help(be_grp_root, stdout); + return -1; } + return lc_opt_from_single_arg(be_grp_root, NULL, arg, NULL); +#else + return 0; #endif /* WITH_LIBCORE */ } +/** The be parameters returned by default, all off. */ +const static backend_params be_params = { + NULL, + NULL, + 0, + NULL, +}; -void be_init(void) +/* Perform schedule verification if requested. */ +static void be_sched_vrfy(ir_graph *irg, int vrfy_opt) { + if (vrfy_opt == BE_VRFY_WARN) { + be_verify_schedule(irg); + } + else if (vrfy_opt == BE_VRFY_ASSERT) { + assert(be_verify_schedule(irg) && "Schedule verification failed."); + } +} + +/* Initialize the Firm backend. Must be run BEFORE init_firm()! */ +const backend_params *be_init(void) { be_opt_register(); be_sched_init(); - be_liveness_init(); be_numbering_init(); be_copy_opt_init(); copystat_init(); phi_class_init(); + + if (isa_if->get_params) + return isa_if->get_params(); + return &be_params; } -static be_main_env_t *be_init_env(be_main_env_t *env) +/** + * Initializes the main environment for the backend. + * + * @param env an empty environment + * @param file_handle the file handle where the output will be written to + */ +static be_main_env_t *be_init_env(be_main_env_t *env, FILE *file_handle) { memset(env, 0, sizeof(*env)); obstack_init(&env->obst); @@ -186,7 +292,7 @@ static be_main_env_t *be_init_env(be_main_env_t *env) env->options = &be_options; FIRM_DBG_REGISTER(env->dbg, "be.main"); - arch_env_init(env->arch_env, isa_if); + arch_env_init(env->arch_env, isa_if, file_handle, env); /* Register the irn handler of the architecture */ if (arch_isa_get_irn_handler(env->arch_env->isa)) @@ -201,29 +307,54 @@ static be_main_env_t *be_init_env(be_main_env_t *env) env->phi_handler = be_phi_handler_new(env->arch_env); arch_env_push_irn_handler(env->arch_env, env->phi_handler); + env->db_handle = be_options.stabs_debug_support ? be_stabs_open(file_handle) : be_nulldbg_open(); return env; } static void be_done_env(be_main_env_t *env) { env->arch_env->isa->impl->done(env->arch_env->isa); + be_dbg_close(env->db_handle); be_phi_handler_free(env->phi_handler); obstack_free(&env->obst, NULL); } +/** + * A wrapper around a firm dumper. Dumps only, if + * flags are enabled. + * + * @param mask a bitmask containing the reason what will be dumped + * @param irg the IR graph to dump + * @param suffix the suffix for the dumper + * @param dumper the dumper to be called + */ static void dump(int mask, ir_graph *irg, const char *suffix, - void (*dumper)(ir_graph *, const char *)) + void (*dumper)(ir_graph *, const char *)) { - if(dump_flags & mask) + if(be_options.dump_flags & mask) be_dump(irg, suffix, dumper); } /** - * Prepare a backend graph for code generation. + * Prepare a backend graph for code generation and initialize its birg */ -static void prepare_graph(be_irg_t *birg) +static void initialize_birg(be_irg_t *birg, ir_graph *irg, be_main_env_t *env) { - ir_graph *irg = birg->irg; + memset(birg, 0, sizeof(*birg)); + birg->irg = irg; + birg->main_env = env; + + edges_deactivate_kind(irg, EDGE_KIND_DEP); + edges_activate_kind(irg, EDGE_KIND_DEP); + + DBG((env->dbg, LEVEL_2, "====> IRG: %F\n", irg)); + dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph); + + be_stat_init_irg(env->arch_env, irg); + be_do_stat_nodes(irg, "01 Begin"); + + /* set the current graph (this is important for several firm functions) */ + current_ir_graph = irg; /* Normalize proj nodes. */ normalize_proj_nodes(irg); @@ -239,119 +370,468 @@ static void prepare_graph(be_irg_t *birg) compute_doms(irg); /* Ensure, that the ir_edges are computed. */ - edges_activate(irg); + edges_assure(irg); /* check, if the dominance property is fulfilled. */ be_check_dominance(irg); /* reset the phi handler. */ - be_phi_handler_reset(birg->main_env->phi_handler); + be_phi_handler_reset(env->phi_handler); } +#ifdef WITH_LIBCORE + +#define BE_TIMER_PUSH(timer) \ + if (be_options.timing == BE_TIME_ON) { \ + int res = lc_timer_push(timer); \ + if (be_options.vrfy_option == BE_VRFY_ASSERT) \ + assert(res && "Timer already on stack, cannot be pushed twice."); \ + else if (be_options.vrfy_option == BE_VRFY_WARN && ! res) \ + fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \ + lc_timer_get_name(timer)); \ + } +#define BE_TIMER_POP(timer) \ + if (be_options.timing == BE_TIME_ON) { \ + lc_timer_t *tmp = lc_timer_pop(); \ + if (be_options.vrfy_option == BE_VRFY_ASSERT) \ + assert(tmp == timer && "Attempt to pop wrong timer."); \ + else if (be_options.vrfy_option == BE_VRFY_WARN && tmp != timer) \ + fprintf(stderr, "Attempt to pop wrong timer. %s is on stack, trying to pop %s.\n", \ + lc_timer_get_name(tmp), lc_timer_get_name(timer)); \ + timer = tmp; \ + } + +#define BE_TIMER_ONLY(code) do { if (be_options.timing == BE_TIME_ON) { code; } } while(0) + +#else + +#define BE_TIMER_PUSH(timer) +#define BE_TIMER_POP(timer) +#define BE_TIMER_ONLY(code) + +#endif /* WITH_LIBCORE */ + + /** * The Firm backend main loop. * Do architecture specific lowering for all graphs * and call the architecture specific code generator. + * + * @param file_handle the file handle the output will be written to + * @param cup_name name of the compilation unit */ -static void be_main_loop(FILE *file_handle) +static void be_main_loop(FILE *file_handle, const char *cup_name) { - int i, n; + int i; arch_isa_t *isa; be_main_env_t env; + unsigned num_nodes_b = 0; + unsigned num_nodes_a = 0; + unsigned num_nodes_r = 0; + char prof_filename[256]; + static const char suffix[] = ".prof"; + be_irg_t *birgs; + unsigned num_birgs; + + be_ra_timer_t *ra_timer; + +#ifdef WITH_LIBCORE + lc_timer_t *t_abi = NULL; + lc_timer_t *t_codegen = NULL; + lc_timer_t *t_sched = NULL; + lc_timer_t *t_constr = NULL; + lc_timer_t *t_regalloc = NULL; + lc_timer_t *t_finish = NULL; + lc_timer_t *t_emit = NULL; + lc_timer_t *t_other = NULL; + lc_timer_t *t_verify = NULL; + + if (be_options.timing == BE_TIME_ON) { + t_abi = lc_timer_register("beabi", "be abi introduction"); + t_codegen = lc_timer_register("codegen", "codegeneration"); + t_sched = lc_timer_register("sched", "scheduling"); + t_constr = lc_timer_register("constr", "assure constraints"); + t_regalloc = lc_timer_register("regalloc", "register allocation"); + t_finish = lc_timer_register("finish", "graph finish"); + t_emit = lc_timer_register("emiter", "code emiter"); + t_verify = lc_timer_register("verify", "graph verification"); + t_other = lc_timer_register("other", "other"); + } +#endif /* WITH_LIBCORE */ - be_init_env(&env); + be_init_env(&env, file_handle); isa = arch_env_get_isa(env.arch_env); - /* For all graphs */ - for(i = 0, n = get_irp_n_irgs(); i < n; ++i) { + be_dbg_so(env.db_handle, cup_name); + be_dbg_types(env.db_handle); + + /* we might need 1 birg more for instrumentation constructor */ + num_birgs = get_irp_n_irgs(); + birgs = alloca(sizeof(birgs[0]) * (num_birgs + 1)); + + /* First: initialize all birgs */ + for(i = 0; i < get_irp_n_irgs(); ++i) { ir_graph *irg = get_irp_irg(i); - const arch_code_generator_if_t *cg_if; - be_irg_t birg; - birg.irg = irg; - birg.main_env = &env; + initialize_birg(&birgs[i], irg, &env); + } + + /* + Get the filename for the profiling data. + Beware: '\0' is already included in sizeof(suffix) + */ + memset(prof_filename, 0, sizeof(prof_filename)); + strncpy(prof_filename, cup_name, sizeof(prof_filename) - sizeof(suffix)); + strcat(prof_filename, suffix); - DBG((env.dbg, LEVEL_2, "====> IRG: %F\n", irg)); - dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph); + /* + Next: Either instruments all irgs with profiling code + or try to read in profile data for current translation unit. + */ + if (be_options.opt_profile) { + ir_graph *prof_init_irg = be_profile_instrument(prof_filename, profile_default); + initialize_birg(&birgs[num_birgs], prof_init_irg, &env); + num_birgs++; + set_method_img_section(get_irg_entity(prof_init_irg), section_constructors); + } + else { + be_profile_read(prof_filename); + } + + /* For all graphs */ + for (i = 0; i < num_birgs; ++i) { + be_irg_t *birg = & birgs[i]; + ir_graph *irg = birg->irg; + optimization_state_t state; + const arch_code_generator_if_t *cg_if; + char irg_name[128]; /* set the current graph (this is important for several firm functions) */ - current_ir_graph = birg.irg; + current_ir_graph = irg; + + if(be_stat_ev_is_active()) { + ir_snprintf(irg_name, sizeof(irg_name), "%F", irg); + be_stat_tags[STAT_TAG_CLS] = ""; + be_stat_tags[STAT_TAG_IRG] = irg_name; + be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file); + } + + /* stop and reset timers */ + BE_TIMER_ONLY( + LC_STOP_AND_RESET_TIMER(t_abi); + LC_STOP_AND_RESET_TIMER(t_codegen); + LC_STOP_AND_RESET_TIMER(t_sched); + LC_STOP_AND_RESET_TIMER(t_constr); + LC_STOP_AND_RESET_TIMER(t_regalloc); + LC_STOP_AND_RESET_TIMER(t_finish); + LC_STOP_AND_RESET_TIMER(t_emit); + LC_STOP_AND_RESET_TIMER(t_verify); + LC_STOP_AND_RESET_TIMER(t_other); + ); + BE_TIMER_PUSH(t_other); /* t_other */ + + /** + * Create execution frequencies from profile data or estimate some + */ + if (be_profile_has_data()) { + birg->execfreqs = be_create_execfreqs_from_profile(irg); + } else { + birg->execfreqs = compute_execfreq(irg, 10); + } + + BE_TIMER_ONLY(num_nodes_b = get_num_reachable_nodes(irg)); /* Get the code generator interface. */ cg_if = isa->impl->get_code_generator_if(isa); /* get a code generator for this graph. */ - birg.cg = cg_if->init(file_handle, &birg); - - /* create the code generator and generate code. */ - prepare_graph(&birg); + birg->cg = cg_if->init(birg); /* some transformations need to be done before abi introduce */ - arch_code_generator_before_abi(birg.cg); + arch_code_generator_before_abi(birg->cg); + + /* reset the phi handler. */ + be_phi_handler_reset(env.phi_handler); /* implement the ABI conventions. */ - birg.abi = be_abi_introduce(&birg); + BE_TIMER_PUSH(t_abi); + birg->abi = be_abi_introduce(birg); + BE_TIMER_POP(t_abi); + dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph); + be_do_stat_nodes(irg, "02 Abi"); /* generate code */ - arch_code_generator_prepare_graph(birg.cg); + BE_TIMER_PUSH(t_codegen); + arch_code_generator_prepare_graph(birg->cg); + BE_TIMER_POP(t_codegen); + + be_do_stat_nodes(irg, "03 Prepare"); /* - * Since the code generator made a lot of new nodes and skipped - * a lot of old ones, we should do dead node elimination here. - * Note that this requires disabling the edges here. - */ + Since the code generator made a lot of new nodes and skipped + a lot of old ones, we should do dead node elimination here. + Note that this requires disabling the edges here. + */ edges_deactivate(irg); //dead_node_elimination(irg); edges_activate(irg); /* Compute loop nesting information (for weighting copies) */ construct_cf_backedges(irg); - dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph); + BE_TIMER_ONLY(num_nodes_r = get_num_reachable_nodes(irg)); + + /* let backend prepare scheduling */ + BE_TIMER_PUSH(t_codegen); + arch_code_generator_before_sched(birg->cg); + BE_TIMER_POP(t_codegen); + + /* schedule the irg */ + BE_TIMER_PUSH(t_sched); + switch (be_options.scheduler) { + default: + fprintf(stderr, "Warning: invalid scheduler (%d) selected, falling back to list scheduler.\n", be_options.scheduler); + case BE_SCHED_LIST: + list_sched(birg, &be_options); + break; +#ifdef WITH_ILP + case BE_SCHED_ILP: + be_ilp_sched(birg); + //fprintf(stderr, "Warning: ILP scheduler not yet fully implemented, falling back to list scheduler.\n"); + //list_sched(birg, &be_options); + break; +#endif /* WITH_ILP */ + }; + BE_TIMER_POP(t_sched); + + dump(DUMP_SCHED, irg, "-sched", dump_ir_block_graph_sched); + + /* check schedule */ + BE_TIMER_PUSH(t_verify); + be_sched_vrfy(irg, be_options.vrfy_option); + BE_TIMER_POP(t_verify); + + be_do_stat_nodes(irg, "04 Schedule"); - /* add Keeps for should_be_different constrained nodes */ - assure_constraints(&birg); - dump(DUMP_PREPARED, irg, "-assured", dump_ir_block_graph); + /* introduce patterns to assure constraints */ + BE_TIMER_PUSH(t_constr); + /* we switch off optimizations here, because they might cause trouble */ + save_optimization_state(&state); + set_optimize(0); + set_opt_normalize(0); - /* Schedule the graphs. */ - arch_code_generator_before_sched(birg.cg); - list_sched(env.arch_env, irg); + /* add Keeps for should_be_different constrained nodes */ + /* beware: needs schedule due to usage of be_ssa_constr */ + assure_constraints(birg); + BE_TIMER_POP(t_constr); + + dump(DUMP_SCHED, irg, "-assured", dump_ir_block_graph_sched); + be_do_stat_nodes(irg, "05 Constraints"); /* connect all stack modifying nodes together (see beabi.c) */ - be_abi_fix_stack_nodes(birg.abi); - dump(DUMP_SCHED, irg, "-sched", dump_ir_block_graph_sched); + BE_TIMER_PUSH(t_abi); + be_abi_fix_stack_nodes(birg->abi, NULL); + BE_TIMER_POP(t_abi); - /* Verify the schedule */ - sched_verify_irg(irg); + dump(DUMP_SCHED, irg, "-fix_stack", dump_ir_block_graph_sched); - /* Do register allocation */ - arch_code_generator_before_ra(birg.cg); - ra->allocate(&birg); - dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched); + /* check schedule */ + BE_TIMER_PUSH(t_verify); + be_sched_vrfy(irg, be_options.vrfy_option); + BE_TIMER_POP(t_verify); - arch_code_generator_after_ra(birg.cg); - be_abi_fix_stack_bias(birg.abi); + /* do some statistics */ + be_do_stat_reg_pressure(birg); - arch_code_generator_done(birg.cg); - dump(DUMP_FINAL, irg, "-end", dump_ir_block_graph_sched); - be_abi_free(birg.abi); + /* stuff needs to be done after scheduling but before register allocation */ + BE_TIMER_PUSH(t_codegen); + arch_code_generator_before_ra(birg->cg); + BE_TIMER_POP(t_codegen); - free_ir_graph(irg); - } + if(be_stat_ev_is_active()) { + be_stat_ev_l("costs_before_ra", + (long) be_estimate_irg_costs(irg, env.arch_env, birg->execfreqs)); + } + + /* Do register allocation */ + BE_TIMER_PUSH(t_regalloc); + ra_timer = ra->allocate(birg); + BE_TIMER_POP(t_regalloc); + if(be_stat_ev_is_active()) { + be_stat_ev_l("costs_after_ra", + (long) be_estimate_irg_costs(irg, env.arch_env, birg->execfreqs)); + } + + dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched); + be_do_stat_nodes(irg, "06 Register Allocation"); + + /* let the code generator prepare the graph for emitter */ + BE_TIMER_PUSH(t_finish); + arch_code_generator_after_ra(birg->cg); + BE_TIMER_POP(t_finish); + + /* fix stack offsets */ + BE_TIMER_PUSH(t_abi); + be_abi_fix_stack_nodes(birg->abi, NULL); + be_remove_dead_nodes_from_schedule(irg); + be_abi_fix_stack_bias(birg->abi); + BE_TIMER_POP(t_abi); + + BE_TIMER_PUSH(t_finish); + arch_code_generator_finish(birg->cg); + BE_TIMER_POP(t_finish); + + dump(DUMP_FINAL, irg, "-finish", dump_ir_block_graph_sched); + + /* check schedule and register allocation */ + BE_TIMER_PUSH(t_verify); + if (be_options.vrfy_option == BE_VRFY_WARN) { + //irg_verify(irg, VRFY_ENFORCE_SSA); + be_check_dominance(irg); + be_verify_out_edges(irg); + be_verify_schedule(irg); + be_verify_register_allocation(env.arch_env, irg); + } + else if (be_options.vrfy_option == BE_VRFY_ASSERT) { + //assert(irg_verify(irg, VRFY_ENFORCE_SSA) && "irg verification failed"); + assert(be_verify_out_edges(irg)); + assert(be_check_dominance(irg) && "Dominance verification failed"); + assert(be_verify_schedule(irg) && "Schedule verification failed"); + assert(be_verify_register_allocation(env.arch_env, irg) + && "register allocation verification failed"); + } + BE_TIMER_POP(t_verify); + + /* emit assembler code */ + BE_TIMER_PUSH(t_emit); + arch_code_generator_done(birg->cg); + BE_TIMER_POP(t_emit); + + dump(DUMP_FINAL, irg, "-end", dump_ir_extblock_graph_sched); + + BE_TIMER_PUSH(t_abi); + be_abi_free(birg->abi); + BE_TIMER_POP(t_abi); + + be_do_stat_nodes(irg, "07 Final"); + restore_optimization_state(&state); + + BE_TIMER_ONLY(num_nodes_a = get_num_reachable_nodes(irg)); + BE_TIMER_POP(t_other); + +#define LC_EMIT(timer) \ + if(!be_stat_ev_is_active()) { \ + printf("%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0); \ + } else { \ + be_stat_ev_l(lc_timer_get_name(timer), lc_timer_elapsed_msec(timer)); \ + } +#define LC_EMIT_RA(timer) \ + if(!be_stat_ev_is_active()) { \ + printf("\t%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0); \ + } else { \ + be_stat_ev_l(lc_timer_get_name(timer), lc_timer_elapsed_msec(timer)); \ + } + BE_TIMER_ONLY( + if(!be_stat_ev_is_active()) { + printf("==>> IRG %s <<==\n", get_entity_name(get_irg_entity(irg))); + printf("# nodes at begin: %u\n", num_nodes_b); + printf("# nodes before ra: %u\n", num_nodes_r); + printf("# nodes at end: %u\n\n", num_nodes_a); + } + LC_EMIT(t_abi); + LC_EMIT(t_codegen); + LC_EMIT(t_sched); + LC_EMIT(t_constr); + LC_EMIT(t_regalloc); + LC_EMIT_RA(ra_timer->t_prolog); + LC_EMIT_RA(ra_timer->t_live); + LC_EMIT_RA(ra_timer->t_spill); + LC_EMIT_RA(ra_timer->t_spillslots); + LC_EMIT_RA(ra_timer->t_color); + LC_EMIT_RA(ra_timer->t_ifg); + LC_EMIT_RA(ra_timer->t_copymin); + LC_EMIT_RA(ra_timer->t_ssa); + LC_EMIT_RA(ra_timer->t_epilog); + LC_EMIT_RA(ra_timer->t_verify); + LC_EMIT_RA(ra_timer->t_other); + LC_EMIT(t_finish); + LC_EMIT(t_emit); + LC_EMIT(t_verify); + LC_EMIT(t_other); + ); +#undef LC_EMIT_RA +#undef LC_EMIT + + free_execfreq(birg->execfreqs); + + /* switched off due to statistics (statistic module needs all irgs) */ + if (! stat_is_active()) + free_ir_graph(irg); + + if(be_stat_ev_is_active()) { + be_stat_ev_pop(); + } + } + be_profile_free(); be_done_env(&env); + +#undef BE_TIMER_POP +#undef BE_TIMER_PUSH +#undef BE_TIMER_ONLY } /* Main interface to the frontend. */ -void be_main(FILE *file_handle) +void be_main(FILE *file_handle, const char *cup_name) { +#ifdef WITH_LIBCORE + lc_timer_t *t = NULL; + + /* The user specified another config file to read. do that now. */ + if(strlen(config_file) > 0) { + FILE *f; + + if((f = fopen(config_file, "rt")) != NULL) { + lc_opt_from_file(config_file, f, NULL); + fclose(f); + } + } + + if (be_options.timing == BE_TIME_ON) { + t = lc_timer_register("bemain", "measure complete bemain loop"); + + if (lc_timer_enter_high_priority()) { + fprintf(stderr, "Warning: Could not enter high priority mode.\n"); + } + + lc_timer_reset_and_start(t); + } + + be_init_stat_file(be_options.stat_file_name, cup_name); +#endif /* WITH_LIBCORE */ + /* never build code for pseudo irgs */ set_visit_pseudo_irgs(0); - be_node_init(); - be_main_loop(file_handle); + be_node_init(); + + be_main_loop(file_handle, cup_name); + +#ifdef WITH_LIBCORE + if (be_options.timing == BE_TIME_ON) { + lc_timer_stop(t); + lc_timer_leave_high_priority(); + if(be_stat_ev_is_active()) { + be_stat_ev_l("backend_time", lc_timer_elapsed_msec(t)); + } else { + printf("%-20s: %lu msec\n", "BEMAINLOOP", lc_timer_elapsed_msec(t)); + } + } + + be_close_stat_file(); +#endif /* WITH_LIBCORE */ } /** The debug info retriever function. */ @@ -366,6 +846,22 @@ void be_set_debug_retrieve(retrieve_dbg_func func) { const char *be_retrieve_dbg_info(const dbg_info *dbg, unsigned *line) { if (retrieve_dbg) return retrieve_dbg(dbg, line); + *line = 0; return NULL; } + +int be_put_ignore_regs(const be_irg_t *birg, const arch_register_class_t *cls, bitset_t *bs) +{ + if(bs == NULL) + bs = bitset_alloca(cls->n_regs); + else + bitset_clear_all(bs); + + assert(bitset_size(bs) == (unsigned) cls->n_regs); + arch_put_non_ignore_regs(birg->main_env->arch_env, cls, bs); + bitset_flip_all(bs); + be_abi_put_ignore_regs(birg->abi, cls, bs); + + return bitset_popcnt(bs); +}