X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbemain.c;h=07afa5d303a743351fa080abb1d2b46ff2bf2039;hb=6f068af98daa4725d60e5d23a8f98ec2841cfa44;hp=e2962e0a9ae18cc6a5d31834bea475440d92aa0b;hpb=622c00b8818a9848ece7e77fb3e6bfe5d984d4fc;p=libfirm diff --git a/ir/be/bemain.c b/ir/be/bemain.c index e2962e0a9..07afa5d30 100644 --- a/ir/be/bemain.c +++ b/ir/be/bemain.c @@ -1,18 +1,36 @@ +/* + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. + * + * This file is part of libFirm. + * + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. + */ + /** - * Backend driver. - * @author Sebastian Hack - * @date 25.11.2004 + * @file + * @brief Main Backend driver. + * @author Sebastian Hack + * @date 25.11.2004 + * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include +#include -#ifdef WITH_LIBCORE -#include -#include -#endif /* WITH_LIBCORE */ +#include "lc_opts.h" +#include "lc_opts_enum.h" #include "obst.h" #include "bitset.h" @@ -21,362 +39,847 @@ #include "irgopt.h" #include "irgraph.h" #include "irdump.h" -#include "phiclass.h" #include "irdom_t.h" #include "iredges_t.h" #include "irloop_t.h" #include "irtools.h" -#include "return.h" +#include "irverify.h" +#include "irprintf.h" +#include "iroptimize.h" +#include "firmstat.h" +#include "execfreq.h" +#include "irprofile.h" +#include "irpass_t.h" #include "bearch.h" -#include "firm/bearch_firm.h" -#include "ia32/bearch_ia32.h" -#include "arm/bearch_arm.h" -#include "ppc32/bearch_ppc32.h" -#include "mips/bearch_mips.h" - #include "be_t.h" -#include "benumb_t.h" +#include "bemodule.h" #include "beutil.h" -#include "benode_t.h" +#include "benode.h" #include "beirgmod.h" -#include "besched_t.h" +#include "besched.h" #include "belistsched.h" #include "belive_t.h" -#include "bespillilp.h" -#include "bespillbelady.h" #include "bera.h" -#include "beraextern.h" #include "bechordal_t.h" #include "beifg.h" -#include "beifg_impl.h" #include "becopyopt.h" #include "becopystat.h" #include "bessadestr.h" #include "beabi.h" #include "belower.h" +#include "bestat.h" +#include "beverify.h" +#include "be_dbgout.h" +#include "beirg.h" +#include "bestack.h" +#include "beemitter.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) +#define NEW_ID(s) new_id_from_chars(s, sizeof(s) - 1) /* 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 */ + false, /* profile_generate */ + false, /* profile_use */ + 0, /* try to omit frame pointer */ + 0, /* create PIC code */ + BE_VERIFY_WARN, /* verification level: warn */ + "i44pc52.info.uni-karlsruhe.de", /* ilp server */ + "cplex", /* ilp solver */ + 0, /* enable statistic event dumping */ + "", /* print stat events */ }; -/* dump flags */ -static unsigned dump_flags = 2 * DUMP_FINAL - 1; - -/* 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; +static const arch_isa_if_t *isa_if = 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 }, - { "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 }, - { "arm", &arm_isa_if }, - { "ppc", &ppc32_isa_if }, - { "mips", &mips_isa_if }, - { NULL, NULL } +/* verify options. */ +static const lc_opt_enum_int_items_t verify_items[] = { + { "off", BE_VERIFY_OFF }, + { "warn", BE_VERIFY_WARN }, + { "assert", BE_VERIFY_ASSERT }, + { 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 = { - (const void **) &ra, ra_items -}; - -static lc_opt_enum_const_ptr_var_t isa_var = { - (const void **) &isa_if, isa_items +static lc_opt_enum_int_var_t verify_var = { + &be_options.verify_option, verify_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_ENUM_MASK("dump", "dump irg on several occasions", &dump_var), + LC_OPT_ENT_BOOL ("omitfp", "omit frame pointer", &be_options.omit_fp), + LC_OPT_ENT_BOOL ("pic", "create PIC code", &be_options.pic), + LC_OPT_ENT_ENUM_PTR ("verify", "verify the backend irg", &verify_var), + LC_OPT_ENT_BOOL ("time", "get backend timing statistics", &be_options.timing), + LC_OPT_ENT_BOOL ("profilegenerate", "instrument the code for execution count profiling", &be_options.opt_profile_generate), + LC_OPT_ENT_BOOL ("profileuse", "use existing profile data", &be_options.opt_profile_use), + LC_OPT_ENT_BOOL ("statev", "dump statistic events", &be_options.statev), + LC_OPT_ENT_STR ("filtev", "filter for stat events (regex if support is active", &be_options.filtev, sizeof(be_options.filtev)), #ifdef WITH_ILP LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)), LC_OPT_ENT_STR ("ilp.solver", "the ilp solver name", be_options.ilp_solver, sizeof(be_options.ilp_solver)), #endif /* WITH_ILP */ - { NULL } + LC_OPT_LAST }; -#endif /* WITH_LIBCORE */ +static be_module_list_entry_t *isa_ifs = NULL; + + +asm_constraint_flags_t asm_constraint_flags[256]; + +void be_init_default_asm_constraint_flags(void) +{ + asm_constraint_flags['?'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['!'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['&'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT + | ASM_CONSTRAINT_FLAG_MODIFIER_EARLYCLOBBER; + asm_constraint_flags['%'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT + | ASM_CONSTRAINT_FLAG_MODIFIER_COMMUTATIVE; + asm_constraint_flags['!'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + + asm_constraint_flags['='] = ASM_CONSTRAINT_FLAG_MODIFIER_WRITE + | ASM_CONSTRAINT_FLAG_MODIFIER_NO_READ; + asm_constraint_flags['+'] = ASM_CONSTRAINT_FLAG_MODIFIER_READ + | ASM_CONSTRAINT_FLAG_MODIFIER_WRITE; + + asm_constraint_flags['i'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['s'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['E'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['F'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['G'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['H'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['I'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['J'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['K'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['L'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['M'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['N'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['O'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + asm_constraint_flags['P'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + + asm_constraint_flags['m'] = ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP; + asm_constraint_flags['o'] = ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP; + asm_constraint_flags['V'] = ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP; + asm_constraint_flags['<'] = ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP; + asm_constraint_flags['>'] = ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP; + + asm_constraint_flags['p'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['0'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['1'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['2'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['3'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['4'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['5'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['6'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['7'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['8'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + asm_constraint_flags['9'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER; + + asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER + | ASM_CONSTRAINT_FLAG_SUPPORTS_MEMOP + | ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE; + + /* these should have been catched by the parsing code already */ + asm_constraint_flags['#'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['*'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags[' '] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['\t'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['\n'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; + asm_constraint_flags['\r'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT; +} + +asm_constraint_flags_t be_parse_asm_constraints(const char *constraint) +{ + asm_constraint_flags_t flags = ASM_CONSTRAINT_FLAG_NONE; + const char *c; + asm_constraint_flags_t tflags; + + for (c = constraint; *c != '\0'; ++c) { + switch (*c) { + case '#': + /* 'comment' stuff */ + while (*c != 0 && *c != ',') + ++c; + break; + case '*': + /* 'comment' character */ + ++c; + break; + case ' ': + case '\t': + case '\n': + case '\r': + break; + default: + tflags = asm_constraint_flags[(int) *c]; + if (tflags != 0) { + flags |= tflags; + } else { + flags |= isa_if->parse_asm_constraint(&c); + } + break; + } + } + + if (( + flags & ASM_CONSTRAINT_FLAG_MODIFIER_WRITE && + flags & ASM_CONSTRAINT_FLAG_MODIFIER_NO_WRITE + ) || ( + flags & ASM_CONSTRAINT_FLAG_MODIFIER_READ && + flags & ASM_CONSTRAINT_FLAG_MODIFIER_NO_READ + )) { + flags |= ASM_CONSTRAINT_FLAG_INVALID; + } + if (!(flags & (ASM_CONSTRAINT_FLAG_MODIFIER_READ | + ASM_CONSTRAINT_FLAG_MODIFIER_WRITE | + ASM_CONSTRAINT_FLAG_MODIFIER_NO_WRITE | + ASM_CONSTRAINT_FLAG_MODIFIER_NO_READ) + )) { + flags |= ASM_CONSTRAINT_FLAG_MODIFIER_READ; + } + + return flags; +} + +int be_is_valid_clobber(const char *clobber) +{ + /* memory is a valid clobber. (the frontend has to detect this case too, + * because it has to add memory edges to the asm) */ + if (strcmp(clobber, "memory") == 0) + return 1; + /* cc (condition code) is always valid */ + if (strcmp(clobber, "cc") == 0) + return 1; + + return isa_if->is_valid_clobber(clobber); +} + +void be_register_isa_if(const char *name, const arch_isa_if_t *isa) +{ + if (isa_if == NULL) + isa_if = isa; + + be_add_module_to_list(&isa_ifs, name, (void*) isa); +} void be_opt_register(void) { -#ifdef WITH_LIBCORE - int i; - lc_opt_entry_t *be_grp_ra; + lc_opt_entry_t *be_grp; static int run_once = 0; - 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"); + if (run_once) + return; + run_once = 1; - lc_opt_add_table(be_grp_root, be_main_options); + be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + lc_opt_add_table(be_grp, 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); - } + be_add_module_list_opt(be_grp, "isa", "the instruction set architecture", + &isa_ifs, (void**) &isa_if); - /* 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); - } - } -#endif /* WITH_LIBCORE */ + be_init_modules(); } /* Parse one argument. */ -int be_parse_arg(const char *arg) { -#ifdef WITH_LIBCORE - return lc_opt_from_single_arg(be_grp_root, NULL, arg, NULL); -#endif /* WITH_LIBCORE */ +int be_parse_arg(const char *arg) +{ + lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + if (strcmp(arg, "help") == 0 || (arg[0] == '?' && arg[1] == '\0')) { + lc_opt_print_help_for_entry(be_grp, '-', stdout); + return -1; + } + return lc_opt_from_single_arg(be_grp, NULL, arg, NULL); } +/* Perform schedule verification if requested. */ +static void be_sched_verify(ir_graph *irg, int verify_opt) +{ + if (verify_opt == BE_VERIFY_WARN) { + be_verify_schedule(irg); + } else if (verify_opt == BE_VERIFY_ASSERT) { + assert(be_verify_schedule(irg) && "Schedule verification failed."); + } +} -void be_init(void) +/* Initialize the Firm backend. Must be run first in init_firm()! */ +void firm_be_init(void) { be_opt_register(); + be_init_modules(); +} + +/* Finalize the Firm backend. */ +void firm_be_finish(void) +{ + be_quit_modules(); +} - be_sched_init(); - be_liveness_init(); - be_numbering_init(); - be_copy_opt_init(); - copystat_init(); - phi_class_init(); +/* Returns the backend parameter */ +const backend_params *be_get_backend_param(void) +{ + return isa_if->get_params(); } +/** + * 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); - env->arch_env = obstack_alloc(&env->obst, sizeof(env->arch_env[0])); - env->options = &be_options; - FIRM_DBG_REGISTER(env->dbg, "be.main"); + env->options = &be_options; + env->ent_trampoline_map = pmap_create(); + env->pic_trampolines_type = new_type_class(NEW_ID("$PIC_TRAMPOLINE_TYPE")); + env->ent_pic_symbol_map = pmap_create(); + env->pic_symbols_type = new_type_struct(NEW_ID("$PIC_SYMBOLS_TYPE")); - arch_env_init(env->arch_env, isa_if, file_handle); + remove_irp_type(env->pic_trampolines_type); + remove_irp_type(env->pic_symbols_type); + set_class_final(env->pic_trampolines_type, 1); - /* Register the irn handler of the architecture */ - if (arch_isa_get_irn_handler(env->arch_env->isa)) - arch_env_push_irn_handler(env->arch_env, arch_isa_get_irn_handler(env->arch_env->isa)); - - /* - * Register the node handler of the back end infrastructure. - * This irn handler takes care of the platform independent - * spill, reload and perm nodes. - */ - arch_env_push_irn_handler(env->arch_env, &be_node_irn_handler); - env->phi_handler = be_phi_handler_new(env->arch_env); - arch_env_push_irn_handler(env->arch_env, env->phi_handler); + memset(asm_constraint_flags, 0, sizeof(asm_constraint_flags)); + env->arch_env = arch_env_init(isa_if, file_handle, env); return env; } +/** + * Called when the be_main_env_t can be destroyed. + */ static void be_done_env(be_main_env_t *env) { - env->arch_env->isa->impl->done(env->arch_env->isa); - be_phi_handler_free(env->phi_handler); - obstack_free(&env->obst, NULL); + pmap_destroy(env->ent_trampoline_map); + pmap_destroy(env->ent_pic_symbol_map); + free_type(env->pic_trampolines_type); + free_type(env->pic_symbols_type); } -static void dump(int mask, ir_graph *irg, const char *suffix, - void (*dumper)(ir_graph *, const char *)) +/** + * 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) { - if(dump_flags & mask) - be_dump(irg, suffix, dumper); + if (be_options.dump_flags & mask) + dump_ir_graph(irg, suffix); } /** - * Prepare a backend graph for code generation. + * Prepare a backend graph for code generation and initialize its irg */ -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; + /* don't duplicate locals in backend when dumping... */ + ir_remove_dump_flags(ir_dump_flag_consts_local); - /* Normalize proj nodes. */ - normalize_proj_nodes(irg); + dump(DUMP_INITIAL, irg, "begin"); - /* Make just one return node. */ - normalize_one_return(irg); + irg->be_data = birg; - /* Remove critical edges */ - remove_critical_cf_edges(irg); + memset(birg, 0, sizeof(*birg)); + birg->irg = irg; + birg->main_env = env; + obstack_init(&birg->obst); + + edges_deactivate_kind(irg, EDGE_KIND_DEP); + edges_activate_kind(irg, EDGE_KIND_DEP); + + /* set the current graph (this is important for several firm functions) */ + current_ir_graph = irg; - /* Compute the dominance information. */ - free_dom(irg); - compute_doms(irg); + /* we do this before critical edge split. As this produces less returns, + because sometimes (= 164.gzip) multiple returns are slower */ + normalize_n_returns(irg); + + /* Remove critical edges */ + remove_critical_cf_edges_ex(irg, /*ignore_exception_edges=*/0); /* Ensure, that the ir_edges are computed. */ - edges_activate(irg); + edges_assure(irg); + + set_irg_phase_state(irg, phase_backend); + be_info_init_irg(irg); + + dump(DUMP_INITIAL, irg, "prepared"); +} + +int be_timing; + +static const char *get_timer_name(be_timer_id_t id) +{ + switch (id) { + case T_ABI: return "abi"; + case T_CODEGEN: return "codegen"; + case T_RA_PREPARATION: return "ra_preparation"; + case T_SCHED: return "sched"; + case T_CONSTR: return "constr"; + case T_FINISH: return "finish"; + case T_EMIT: return "emit"; + case T_VERIFY: return "verify"; + case T_OTHER: return "other"; + case T_HEIGHTS: return "heights"; + case T_LIVE: return "live"; + case T_EXECFREQ: return "execfreq"; + case T_SSA_CONSTR: return "ssa_constr"; + case T_RA_PROLOG: return "ra_prolog"; + case T_RA_EPILOG: return "ra_epilog"; + case T_RA_CONSTR: return "ra_constr"; + case T_RA_SPILL: return "ra_spill"; + case T_RA_SPILL_APPLY: return "ra_spill_apply"; + case T_RA_COLOR: return "ra_color"; + case T_RA_IFG: return "ra_ifg"; + case T_RA_COPYMIN: return "ra_copymin"; + case T_RA_SSA: return "ra_ssa"; + case T_RA_OTHER: return "ra_other"; + } + return "unknown"; +} +ir_timer_t *be_timers[T_LAST+1]; - /* check, if the dominance property is fulfilled. */ - be_check_dominance(irg); +void be_lower_for_target(void) +{ + size_t i; - /* reset the phi handler. */ - be_phi_handler_reset(birg->main_env->phi_handler); + isa_if->lower_for_target(); + /* set the phase to low */ + for (i = get_irp_n_irgs(); i > 0;) { + ir_graph *irg = get_irp_irg(--i); + set_irg_phase_state(irg, phase_low); + } + set_irp_phase_state(phase_low); +} + +static void emit_global_asms(void) +{ + size_t n = get_irp_n_asms(); + size_t i; + for (i = 0; i < n; ++i) { + be_emit_cstring("#APP\n"); + be_emit_ident(get_irp_asm(i)); + be_emit_cstring("\n#NO_APP\n"); + } } /** * 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; - arch_isa_t *isa; + static const char suffix[] = ".prof"; + + size_t i, num_birgs; + int stat_active = 0; be_main_env_t env; + char prof_filename[256]; + be_irg_t *birgs; + ir_graph **irg_list, **backend_irg_list; + arch_env_t *arch_env; + + be_timing = (be_options.timing == BE_TIME_ON); + + if (be_timing) { + for (i = 0; i < T_LAST+1; ++i) { + be_timers[i] = ir_timer_new(); + } + } be_init_env(&env, file_handle); + env.cup_name = cup_name; - isa = arch_env_get_isa(env.arch_env); + be_dbg_open(); + be_dbg_unit_begin(cup_name); + be_dbg_types(); - /* For all graphs */ - for (i = 0, n = get_irp_n_irgs(); i < n; ++i) { - ir_graph *irg = get_irp_irg(i); - const arch_code_generator_if_t *cg_if; - be_irg_t birg; + emit_global_asms(); + + arch_env = env.arch_env; - birg.irg = irg; - birg.main_env = &env; + /* backend may provide an ordered list of irgs where code should be + * generated for */ + irg_list = NEW_ARR_F(ir_graph *, 0); + backend_irg_list = arch_env_get_backend_irg_list(arch_env, &irg_list); - DBG((env.dbg, LEVEL_2, "====> IRG: %F\n", irg)); - dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph); + /* we might need 1 birg more for instrumentation constructor */ + num_birgs = backend_irg_list ? ARR_LEN(backend_irg_list) : get_irp_n_irgs(); + birgs = ALLOCAN(be_irg_t, num_birgs + 1); + + be_info_init(); + + /* First: initialize all birgs */ + for (i = 0; i < num_birgs; ++i) { + ir_graph *irg = backend_irg_list ? backend_irg_list[i] : get_irp_irg(i); + initialize_birg(&birgs[i], irg, &env); + } + arch_env_handle_intrinsics(arch_env); + DEL_ARR_F(irg_list); + + /* + Get the filename for the profiling data. + Beware: '\0' is already included in sizeof(suffix) + */ + sprintf(prof_filename, "%.*s%s", + (int)(sizeof(prof_filename) - sizeof(suffix)), cup_name, suffix); + + if (be_options.opt_profile_use) { + bool res = ir_profile_read(prof_filename); + if (!res) { + fprintf(stderr, "Warning: Couldn't read profile data '%s'\n", + prof_filename); + } + } + if (be_options.opt_profile_generate) { + ir_graph *prof_init_irg + = ir_profile_instrument(prof_filename); + initialize_birg(&birgs[num_birgs], prof_init_irg, &env); + num_birgs++; + } + + stat_active = stat_is_active(); + + /* 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; /* set the current graph (this is important for several firm functions) */ - current_ir_graph = birg.irg; + current_ir_graph = irg; - /* Get the code generator interface. */ - cg_if = isa->impl->get_code_generator_if(isa); + stat_ev_if { + stat_ev_ctx_push_fobj("bemain_irg", irg); + be_stat_ev("bemain_insns_start", be_count_insns(irg)); + be_stat_ev("bemain_blocks_start", be_count_blocks(irg)); + } - /* get a code generator for this graph. */ - birg.cg = cg_if->init(&birg); + /* stop and reset timers */ + be_timer_push(T_OTHER); + + /* Verify the initial graph */ + be_timer_push(T_VERIFY); + if (be_options.verify_option == BE_VERIFY_WARN) { + irg_verify(irg, VERIFY_ENFORCE_SSA); + be_check_dominance(irg); + } else if (be_options.verify_option == BE_VERIFY_ASSERT) { + assert(irg_verify(irg, VERIFY_ENFORCE_SSA) && "irg verification failed"); + assert(be_check_dominance(irg) && "Dominance verification failed"); + } + be_timer_pop(T_VERIFY); - /* create the code generator and generate code. */ - prepare_graph(&birg); + /* get a code generator for this graph. */ + arch_env->impl->init_graph(irg); /* some transformations need to be done before abi introduce */ - arch_code_generator_before_abi(birg.cg); + if (arch_env->impl->before_abi != NULL) + arch_env->impl->before_abi(irg); /* implement the ABI conventions. */ - birg.abi = be_abi_introduce(&birg); - dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph); + if (!arch_env->custom_abi) { + be_timer_push(T_ABI); + be_abi_introduce(irg); + be_timer_pop(T_ABI); + dump(DUMP_ABI, irg, "abi"); + } - /* generate code */ - arch_code_generator_prepare_graph(birg.cg); + /* we have to do cfopt+remove_critical_edges as we can't have Bad-blocks + * or critical edges in the backend */ + optimize_cf(irg); + remove_critical_cf_edges(irg); - /* - * 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. - */ + /* We often have dead code reachable through out-edges here. So for + * now we rebuild edges (as we need correct user count for code + * selection) */ 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, "before-code-selection"); - dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph); + if (be_options.verify_option == BE_VERIFY_WARN) { + be_check_dominance(irg); + } else if (be_options.verify_option == BE_VERIFY_ASSERT) { + assert(be_check_dominance(irg) && "Dominance verification failed"); + } - /* add Keeps for should_be_different constrained nodes */ - assure_constraints(&birg); - dump(DUMP_PREPARED, irg, "-assured", dump_ir_block_graph); + /* perform codeselection */ + be_timer_push(T_CODEGEN); + if (arch_env->impl->prepare_graph != NULL) + arch_env->impl->prepare_graph(irg); + be_timer_pop(T_CODEGEN); - /* Schedule the graphs. */ - arch_code_generator_before_sched(birg.cg); - list_sched(env.arch_env, irg); + if (be_options.verify_option == BE_VERIFY_WARN) { + be_check_dominance(irg); + } else if (be_options.verify_option == BE_VERIFY_ASSERT) { + assert(be_check_dominance(irg) && "Dominance verification failed"); + } - /* 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); + dump(DUMP_PREPARED, irg, "code-selection"); + + be_timer_push(T_EXECFREQ); + /** + * Create execution frequencies from profile data or estimate some + */ + if (ir_profile_has_data()) + birg->exec_freq = ir_create_execfreqs_from_profile(irg); + else { + /* TODO: edges are corrupt for EDGE_KIND_BLOCK after the local + * optimize graph phase merges blocks in the x86 backend */ + edges_deactivate(irg); + birg->exec_freq = compute_execfreq(irg, 10); + } + be_timer_pop(T_EXECFREQ); + + + /* disabled for now, fails for EmptyFor.c and XXEndless.c */ + /* be_live_chk_compare(irg); */ + + /* schedule the irg */ + be_timer_push(T_SCHED); + be_schedule_graph(irg); + be_timer_pop(T_SCHED); + + dump(DUMP_SCHED, irg, "sched"); + + /* check schedule */ + be_timer_push(T_VERIFY); + be_sched_verify(irg, be_options.verify_option); + be_timer_pop(T_VERIFY); + + /* 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_cse(0); + + /* add Keeps for should_be_different constrained nodes */ + /* beware: needs schedule due to usage of be_ssa_constr */ + assure_constraints(irg); + be_timer_pop(T_CONSTR); + + dump(DUMP_SCHED, irg, "assured"); + + /* stuff needs to be done after scheduling but before register allocation */ + be_timer_push(T_RA_PREPARATION); + if (arch_env->impl->before_ra != NULL) + arch_env->impl->before_ra(irg); + be_timer_pop(T_RA_PREPARATION); - /* Verify the schedule */ - sched_verify_irg(irg); + /* connect all stack modifying nodes together (see beabi.c) */ + be_timer_push(T_ABI); + be_abi_fix_stack_nodes(irg); + be_timer_pop(T_ABI); + + dump(DUMP_SCHED, irg, "fix_stack"); + + /* check schedule */ + be_timer_push(T_VERIFY); + be_sched_verify(irg, be_options.verify_option); + be_timer_pop(T_VERIFY); + + stat_ev_if { + stat_ev_dbl("bemain_costs_before_ra", + be_estimate_irg_costs(irg, birg->exec_freq)); + be_stat_ev("bemain_insns_before_ra", be_count_insns(irg)); + be_stat_ev("bemain_blocks_before_ra", be_count_blocks(irg)); + } /* Do register allocation */ - arch_code_generator_before_ra(birg.cg); - ra->allocate(&birg); - dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched); + be_allocate_registers(irg); + + stat_ev_dbl("bemain_costs_before_ra", be_estimate_irg_costs(irg, birg->exec_freq)); + + dump(DUMP_RA, irg, "ra"); + + /* let the code generator prepare the graph for emitter */ + be_timer_push(T_FINISH); + if (arch_env->impl->after_ra != NULL) + arch_env->impl->after_ra(irg); + be_timer_pop(T_FINISH); - arch_code_generator_after_ra(birg.cg); - be_abi_fix_stack_bias(birg.abi); + /* fix stack offsets */ + be_timer_push(T_ABI); + be_abi_fix_stack_nodes(irg); + be_remove_dead_nodes_from_schedule(irg); + be_abi_fix_stack_bias(irg); + be_timer_pop(T_ABI); + + dump(DUMP_SCHED, irg, "fix_stack_after_ra"); + + be_timer_push(T_FINISH); + if (arch_env->impl->finish != NULL) + arch_env->impl->finish(irg); + be_timer_pop(T_FINISH); + + dump(DUMP_FINAL, irg, "finish"); + + stat_ev_if { + be_stat_ev("bemain_insns_finish", be_count_insns(irg)); + be_stat_ev("bemain_blocks_finish", be_count_blocks(irg)); + } - arch_code_generator_done(birg.cg); - dump(DUMP_FINAL, irg, "-end", dump_ir_block_graph_sched); - be_abi_free(birg.abi); + /* check schedule and register allocation */ + be_timer_push(T_VERIFY); + if (be_options.verify_option == BE_VERIFY_WARN) { + irg_verify(irg, VERIFY_ENFORCE_SSA); + be_check_dominance(irg); + be_verify_schedule(irg); + be_verify_register_allocation(irg); + } else if (be_options.verify_option == BE_VERIFY_ASSERT) { + assert(irg_verify(irg, VERIFY_ENFORCE_SSA) && "irg verification failed"); + assert(be_check_dominance(irg) && "Dominance verification failed"); + assert(be_verify_schedule(irg) && "Schedule verification failed"); + assert(be_verify_register_allocation(irg) + && "register allocation verification failed"); - free_ir_graph(irg); + } + be_timer_pop(T_VERIFY); + + /* emit assembler code */ + be_timer_push(T_EMIT); + if (arch_env->impl->emit != NULL) + arch_env->impl->emit(irg); + be_timer_pop(T_EMIT); + + dump(DUMP_FINAL, irg, "end"); + + if (!arch_env->custom_abi) { + be_timer_push(T_ABI); + be_abi_free(irg); + be_timer_pop(T_ABI); + } + + restore_optimization_state(&state); + + be_timer_pop(T_OTHER); + + if (be_timing) { + be_timer_id_t t; + if (stat_ev_enabled) { + for (t = T_FIRST; t < T_LAST+1; ++t) { + char buf[128]; + snprintf(buf, sizeof(buf), "bemain_time_%s", + get_timer_name(t)); + stat_ev_dbl(buf, ir_timer_elapsed_usec(be_timers[i])); + } + } else { + printf("==>> IRG %s <<==\n", + get_entity_name(get_irg_entity(irg))); + for (t = T_FIRST; t < T_LAST+1; ++t) { + double val = ir_timer_elapsed_usec(be_timers[t]) / 1000.0; + printf("%-20s: %8.3lf msec\n", get_timer_name(t), val); + } + } + for (t = T_FIRST; t < T_LAST+1; ++t) { + ir_timer_stop(be_timers[t]); + ir_timer_reset(be_timers[t]); + } + } + + be_free_birg(irg); + stat_ev_ctx_pop("bemain_irg"); } + arch_env_done(arch_env); + + be_dbg_unit_end(); + be_dbg_close(); + + ir_profile_free(); be_done_env(&env); + + be_info_free(); } /* Main interface to the frontend. */ -void be_main(FILE *file_handle) +void be_main(FILE *file_handle, const char *cup_name) { - /* never build code for pseudo irgs */ - set_visit_pseudo_irgs(0); + ir_timer_t *t = NULL; - be_node_init(); - be_main_loop(file_handle); -} + if (be_options.timing == BE_TIME_ON) { + t = ir_timer_new(); + + if (ir_timer_enter_high_priority()) { + fprintf(stderr, "Warning: Could not enter high priority mode.\n"); + } + + ir_timer_reset_and_start(t); + } + + if (be_options.statev) { + const char *dot = strrchr(cup_name, '.'); + const char *pos = dot ? dot : cup_name + strlen(cup_name); + char *buf = ALLOCAN(char, pos - cup_name + 1); + strncpy(buf, cup_name, pos - cup_name); + buf[pos - cup_name] = '\0'; + + be_options.statev = 1; + stat_ev_begin(buf, be_options.filtev); + stat_ev_ctx_push_str("bemain_compilation_unit", cup_name); + } -/** The debug info retriever function. */ -static retrieve_dbg_func retrieve_dbg = NULL; + be_main_loop(file_handle, cup_name); -/* Sets a debug info retriever. */ -void be_set_debug_retrieve(retrieve_dbg_func func) { - retrieve_dbg = func; + if (be_options.timing == BE_TIME_ON) { + ir_timer_stop(t); + ir_timer_leave_high_priority(); + stat_ev_if { + stat_ev_dbl("bemain_backend_time", ir_timer_elapsed_msec(t)); + } else { + double val = ir_timer_elapsed_usec(t) / 1000.0; + printf("%-20s: %8.3lf msec\n", "BEMAINLOOP", val); + } + } + + if (be_options.statev) { + stat_ev_ctx_pop("bemain_compilation_unit"); + stat_ev_end(); + } } -/* Retrieve the debug info. */ -const char *be_retrieve_dbg_info(const dbg_info *dbg, unsigned *line) { - if (retrieve_dbg) - return retrieve_dbg(dbg, line); - *line = 0; - return NULL; +static int do_lower_for_target(ir_prog *irp, void *context) +{ + be_lower_for_target(); + (void) context; + (void) irp; + return 0; +} + +ir_prog_pass_t *lower_for_target_pass(const char *name) +{ + ir_prog_pass_t *pass = XMALLOCZ(ir_prog_pass_t); + return def_prog_pass_constructor(pass, + name ? name : "lower_for_target", + do_lower_for_target); }