nsz Git - cparser/blob - driver/firm_opt.c

   1 /**
   2  *
   3  * @file firm_opt.c -- Firm-generating back end optimizations.
   4  *
   5  * (C) 2005-2010  Michael Beck   beck@ipd.info.uni-karlsruhe.de
   6  *
   7  * $Id$
   8  */
   9
  10 #include <config.h>
  11
  12 #include <stdbool.h>
  13 #include <stdlib.h>
  14 #include <string.h>
  15 #include <stdbool.h>
  16 #include <assert.h>
  17 #include <libfirm/firm.h>
  18
  19 #include "firm_opt.h"
  20 #include "firm_codegen.h"
  21 #include "firm_cmdline.h"
  22 #include "firm_timing.h"
  23 #include "ast2firm.h"
  24
  25 static ir_timer_t *t_vcg_dump;
  26 static ir_timer_t *t_verify;
  27 static ir_timer_t *t_all_opt;
  28 static bool do_irg_opt(ir_graph *irg, const char *name);
  29
  30 /** dump all the graphs depending on cond */
  31
  32 static void dump_all(const char *suffix)
  33 {
  34         if (!firm_dump.ir_graph)
  35                 return;
  36
  37         timer_push(t_vcg_dump);
  38         dump_all_ir_graphs(suffix);
  39         timer_pop(t_vcg_dump);
  40 }
  41
  42 /* entities of runtime functions */
  43 ir_entity_ptr rts_entities[rts_max];
  44
  45 /**
  46  * Map runtime functions.
  47  */
  48 static void rts_map(void)
  49 {
  50         static const struct {
  51                 ir_entity_ptr *ent; /**< address of the rts entity */
  52                 i_mapper_func func; /**< mapper function. */
  53         } mapper[] = {
  54                 /* integer */
  55                 { &rts_entities[rts_abs],     i_mapper_abs },
  56                 { &rts_entities[rts_labs],    i_mapper_abs },
  57                 { &rts_entities[rts_llabs],   i_mapper_abs },
  58                 { &rts_entities[rts_imaxabs], i_mapper_abs },
  59
  60                 /* double -> double */
  61                 { &rts_entities[rts_fabs],    i_mapper_abs },
  62                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
  63                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
  64                 { &rts_entities[rts_pow],     i_mapper_pow },
  65                 { &rts_entities[rts_exp],     i_mapper_exp },
  66                 { &rts_entities[rts_exp2],    i_mapper_exp },
  67                 { &rts_entities[rts_exp10],   i_mapper_exp },
  68                 { &rts_entities[rts_log],     i_mapper_log },
  69                 { &rts_entities[rts_log2],    i_mapper_log2 },
  70                 { &rts_entities[rts_log10],   i_mapper_log10 },
  71                 { &rts_entities[rts_sin],     i_mapper_sin },
  72                 { &rts_entities[rts_cos],     i_mapper_cos },
  73                 { &rts_entities[rts_tan],     i_mapper_tan },
  74                 { &rts_entities[rts_asin],    i_mapper_asin },
  75                 { &rts_entities[rts_acos],    i_mapper_acos },
  76                 { &rts_entities[rts_atan],    i_mapper_atan },
  77                 { &rts_entities[rts_sinh],    i_mapper_sinh },
  78                 { &rts_entities[rts_cosh],    i_mapper_cosh },
  79                 { &rts_entities[rts_tanh],    i_mapper_tanh },
  80
  81                 /* float -> float */
  82                 { &rts_entities[rts_fabsf],   i_mapper_abs },
  83                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
  84                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
  85                 { &rts_entities[rts_powf],    i_mapper_pow },
  86                 { &rts_entities[rts_expf],    i_mapper_exp },
  87                 { &rts_entities[rts_exp2f],   i_mapper_exp },
  88                 { &rts_entities[rts_exp10f],  i_mapper_exp },
  89                 { &rts_entities[rts_logf],    i_mapper_log },
  90                 { &rts_entities[rts_log2f],   i_mapper_log2 },
  91                 { &rts_entities[rts_log10f],  i_mapper_log10 },
  92                 { &rts_entities[rts_sinf],    i_mapper_sin },
  93                 { &rts_entities[rts_cosf],    i_mapper_cos },
  94                 { &rts_entities[rts_tanf],    i_mapper_tan },
  95                 { &rts_entities[rts_asinf],   i_mapper_asin },
  96                 { &rts_entities[rts_acosf],   i_mapper_acos },
  97                 { &rts_entities[rts_atanf],   i_mapper_atan },
  98                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
  99                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 100                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 101
 102                 /* long double -> long double */
 103                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 104                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 105                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 106                 { &rts_entities[rts_powl],    i_mapper_pow },
 107                 { &rts_entities[rts_expl],    i_mapper_exp },
 108                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 109                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 110                 { &rts_entities[rts_logl],    i_mapper_log },
 111                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 112                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 113                 { &rts_entities[rts_sinl],    i_mapper_sin },
 114                 { &rts_entities[rts_cosl],    i_mapper_cos },
 115                 { &rts_entities[rts_tanl],    i_mapper_tan },
 116                 { &rts_entities[rts_asinl],   i_mapper_asin },
 117                 { &rts_entities[rts_acosl],   i_mapper_acos },
 118                 { &rts_entities[rts_atanl],   i_mapper_atan },
 119                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 120                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 121                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 122
 123                 /* string */
 124                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 125                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 126                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 127                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 128                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 129                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 130                 { &rts_entities[rts_memmove], i_mapper_memmove },
 131                 { &rts_entities[rts_memset],  i_mapper_memset },
 132                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 133         };
 134         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
 135         unsigned i, n_map;
 136
 137         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
 138                 if (*mapper[i].ent != NULL) {
 139                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 140                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 141                         rec[n_map].i_call.i_mapper = mapper[i].func;
 142                         rec[n_map].i_call.ctx      = NULL;
 143                         rec[n_map].i_call.link     = NULL;
 144                         ++n_map;
 145                 }
 146         }
 147
 148         if (n_map > 0)
 149                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 150 }
 151
 152 static int *irg_dump_no;
 153
 154 static void do_lower_highlevel(ir_graph *irg)
 155 {
 156         lower_highlevel_graph(irg, false);
 157 }
 158
 159 static void do_stred(ir_graph *irg)
 160 {
 161         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 162 }
 163
 164 static void after_inline_opt(ir_graph *irg)
 165 {
 166         do_irg_opt(irg, "scalar-replace");
 167         do_irg_opt(irg, "local");
 168         do_irg_opt(irg, "control-flow");
 169         do_irg_opt(irg, "combo");
 170 }
 171
 172 static void do_inline(void)
 173 {
 174         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 175                          after_inline_opt);
 176 }
 177
 178 static void do_cloning(void)
 179 {
 180         proc_cloning((float) firm_opt.clone_threshold);
 181 }
 182
 183 static void do_lower_mux(ir_graph *irg)
 184 {
 185         lower_mux(irg, NULL);
 186 }
 187
 188 static void do_vrp(ir_graph *irg)
 189 {
 190         set_vrp_data(irg);
 191 }
 192
 193 typedef enum opt_target {
 194         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 195         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 196 } opt_target_t;
 197
 198 typedef enum opt_flags {
 199         OPT_FLAG_NONE         = 0,
 200         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 201         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 202         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 203         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 204                                              -foptions for this transformation */
 205         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 206                                              so we need it even with -O0 */
 207 } opt_flags_t;
 208
 209 typedef void (*transform_irg_func)(ir_graph *irg);
 210 typedef void (*transform_irp_func)(void);
 211
 212 typedef struct {
 213         opt_target_t  target;
 214         const char   *name;
 215         union {
 216                 transform_irg_func transform_irg;
 217                 transform_irp_func transform_irp;
 218         } u;
 219         const char   *description;
 220         opt_flags_t   flags;
 221 } opt_config_t;
 222
 223 static opt_config_t opts[] = {
 224 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 225 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 226         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 227         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 228         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
 229         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 230         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 231         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 232         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 233         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 234         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 235         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 236         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 237         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 238         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 239         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 240         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 241         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 242         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 243         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 244         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 245         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 246         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 247         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 248         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 249         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 250         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 251         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 252         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
 253         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 254         IRP("lower-const",       lower_const_code,         "lowering of constant code",                             OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY | OPT_FLAG_ESSENTIAL),
 255         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 256         IRP("opt-func-call",     optimize_funccalls,       "function call optimization",                            OPT_FLAG_NONE),
 257         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 258         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 259         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
 260         IRP("opt-cc",            mark_private_methods,     "calling conventions optimization",                      OPT_FLAG_NONE),
 261 #undef IRP
 262 #undef IRG
 263 };
 264 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 265 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 266
 267 static opt_config_t *get_opt(const char *name)
 268 {
 269         int i;
 270         for (i = 0; i < n_opts; ++i) {
 271                 opt_config_t *config = &opts[i];
 272                 if (strcmp(config->name, name) == 0)
 273                         return config;
 274         }
 275
 276         return NULL;
 277 }
 278
 279 static void set_opt_enabled(const char *name, bool enabled)
 280 {
 281         opt_config_t *config = get_opt(name);
 282         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 283                 | (enabled ? OPT_FLAG_ENABLED : 0);
 284 }
 285
 286 static bool get_opt_enabled(const char *name)
 287 {
 288         opt_config_t *config = get_opt(name);
 289         return (config->flags & OPT_FLAG_ENABLED) != 0;
 290 }
 291
 292 /**
 293  * perform an optimisation on a single graph
 294  *
 295  * @return  true if something changed, false otherwise
 296  */
 297 static bool do_irg_opt(ir_graph *irg, const char *name)
 298 {
 299         ir_graph     *old_irg;
 300         opt_config_t *config = get_opt(name);
 301         size_t        n      = config - opts;
 302         assert(config != NULL);
 303         assert(config->target == OPT_TARGET_IRG);
 304         if (! (config->flags & OPT_FLAG_ENABLED))
 305                 return false;
 306
 307         old_irg          = current_ir_graph;
 308         current_ir_graph = irg;
 309
 310         timer_push(timers[n]);
 311         config->u.transform_irg(irg);
 312         timer_pop(timers[n]);
 313
 314         if (firm_dump.all_phases && firm_dump.ir_graph) {
 315                 dump_ir_graph(irg, name);
 316         }
 317
 318         if (firm_opt.check_all) {
 319                 timer_push(t_verify);
 320                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 321                 timer_pop(t_verify);
 322         }
 323
 324         current_ir_graph = old_irg;
 325         return true;
 326 }
 327
 328 static void do_irp_opt(const char *name)
 329 {
 330         opt_config_t *config = get_opt(name);
 331         size_t        n      = config - opts;
 332         assert(config->target == OPT_TARGET_IRP);
 333         if (! (config->flags & OPT_FLAG_ENABLED))
 334                 return;
 335
 336         timer_push(timers[n]);
 337         config->u.transform_irp();
 338         timer_pop(timers[n]);
 339
 340         if (firm_dump.ir_graph && firm_dump.all_phases) {
 341                 int i;
 342                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 343                         ir_graph *irg = get_irp_irg(i);
 344                         dump_ir_graph(irg, name);
 345                 }
 346         }
 347
 348         if (firm_opt.check_all) {
 349                 int i;
 350                 timer_push(t_verify);
 351                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 352                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 353                 }
 354                 timer_pop(t_verify);
 355         }
 356 }
 357
 358 /**
 359  * Enable transformations which should be always safe (and cheap) to perform
 360  */
 361 static void enable_safe_defaults(void)
 362 {
 363         set_opt_enabled("remove-unused", true);
 364         set_opt_enabled("opt-tail-rec", true);
 365         set_opt_enabled("opt-func-call", true);
 366         set_opt_enabled("reassociation", true);
 367         set_opt_enabled("control-flow", true);
 368         set_opt_enabled("local", true);
 369         set_opt_enabled("lower-const", true);
 370         set_opt_enabled("scalar-replace", true);
 371         set_opt_enabled("place", true);
 372         set_opt_enabled("confirm", true);
 373         set_opt_enabled("opt-load-store", true);
 374         set_opt_enabled("lower", true);
 375         set_opt_enabled("deconv", true);
 376         set_opt_enabled("remove-confirms", true);
 377         set_opt_enabled("ivopts", true);
 378         set_opt_enabled("dead", true);
 379         set_opt_enabled("remove-phi-cycles", true);
 380         set_opt_enabled("frame", true);
 381         set_opt_enabled("combo", true);
 382         set_opt_enabled("invert-loops", true);
 383         set_opt_enabled("target-lowering", true);
 384         set_opt_enabled("rts", true);
 385         set_opt_enabled("parallelize-mem", true);
 386         set_opt_enabled("opt-cc", true);
 387 }
 388
 389 /**
 390  * run all the Firm optimizations
 391  *
 392  * @param input_filename     the name of the (main) source file
 393  */
 394 static void do_firm_optimizations(const char *input_filename)
 395 {
 396         size_t   i;
 397         unsigned aa_opt;
 398
 399         set_opt_alias_analysis(firm_opt.alias_analysis);
 400
 401         aa_opt = aa_opt_no_opt;
 402         if (firm_opt.strict_alias)
 403                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 404         if (firm_opt.no_alias)
 405                 aa_opt = aa_opt_no_alias;
 406
 407         set_irp_memory_disambiguator_options(aa_opt);
 408
 409         /* parameter passing code should set them directly sometime... */
 410         set_opt_enabled("confirm", firm_opt.confirm);
 411         set_opt_enabled("remove-confirms", firm_opt.confirm);
 412
 413         /* osr supersedes remove_phi_cycles */
 414         if (get_opt_enabled("ivopts"))
 415                 set_opt_enabled("remove-phi-cycles", false);
 416
 417         timer_start(t_all_opt);
 418
 419         do_irp_opt("rts");
 420
 421         /* first step: kill dead code */
 422         for (i = 0; i < get_irp_n_irgs(); i++) {
 423                 ir_graph *irg = get_irp_irg(i);
 424                 do_irg_opt(irg, "combo");
 425                 do_irg_opt(irg, "local");
 426                 do_irg_opt(irg, "control-flow");
 427         }
 428
 429         do_irp_opt("remove-unused");
 430         for (i = 0; i < get_irp_n_irgs(); ++i) {
 431                 ir_graph *irg = get_irp_irg(i);
 432                 do_irg_opt(irg, "opt-tail-rec");
 433         }
 434         do_irp_opt("opt-func-call");
 435         do_irp_opt("lower-const");
 436
 437         for (i = 0; i < get_irp_n_irgs(); i++) {
 438                 ir_graph *irg = get_irp_irg(i);
 439
 440                 do_irg_opt(irg, "scalar-replace");
 441                 do_irg_opt(irg, "invert-loops");
 442                 do_irg_opt(irg, "unroll-loops");
 443                 do_irg_opt(irg, "local");
 444                 do_irg_opt(irg, "reassociation");
 445                 do_irg_opt(irg, "local");
 446                 do_irg_opt(irg, "place");
 447
 448                 if (firm_opt.confirm) {
 449                         /* Confirm construction currently can only handle blocks with only
 450                            one control flow predecessor. Calling optimize_cf here removes
 451                            Bad predecessors and help the optimization of switch constructs.
 452                          */
 453                         do_irg_opt(irg, "control-flow");
 454                         do_irg_opt(irg, "confirm");
 455                         do_irg_opt(irg, "vrp");
 456                         do_irg_opt(irg, "local");
 457                 }
 458
 459                 do_irg_opt(irg, "control-flow");
 460                 do_irg_opt(irg, "opt-load-store");
 461                 do_irg_opt(irg, "fp-vrp");
 462                 do_irg_opt(irg, "lower");
 463                 do_irg_opt(irg, "deconv");
 464                 do_irg_opt(irg, "thread-jumps");
 465                 do_irg_opt(irg, "remove-confirms");
 466                 do_irg_opt(irg, "gvn-pre");
 467                 do_irg_opt(irg, "place");
 468                 do_irg_opt(irg, "control-flow");
 469
 470                 if (do_irg_opt(irg, "if-conversion")) {
 471                         do_irg_opt(irg, "local");
 472                         do_irg_opt(irg, "control-flow");
 473                 }
 474                 /* this doesn't make too much sense but tests the mux destruction... */
 475                 do_irg_opt(irg, "lower-mux");
 476
 477                 do_irg_opt(irg, "bool");
 478                 do_irg_opt(irg, "shape-blocks");
 479                 do_irg_opt(irg, "ivopts");
 480                 do_irg_opt(irg, "local");
 481                 do_irg_opt(irg, "dead");
 482                 do_irg_opt(irg, "frame");
 483         }
 484
 485         do_irp_opt("inline");
 486         do_irp_opt("opt-proc-clone");
 487
 488         for (i = 0; i < get_irp_n_irgs(); i++) {
 489                 ir_graph *irg = get_irp_irg(i);
 490                 do_irg_opt(irg, "local");
 491                 do_irg_opt(irg, "control-flow");
 492                 do_irg_opt(irg, "thread-jumps");
 493                 do_irg_opt(irg, "local");
 494                 do_irg_opt(irg, "control-flow");
 495
 496                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 497                         do_irg_opt(irg, "local");
 498                         do_irg_opt(irg, "vrp");
 499                         do_irg_opt(irg, "local");
 500                         do_irg_opt(irg, "vrp");
 501                 }
 502         }
 503
 504         if (firm_dump.ir_graph) {
 505                 /* recompute backedges for nicer dumps */
 506                 for (i = 0; i < get_irp_n_irgs(); i++)
 507                         construct_cf_backedges(get_irp_irg(i));
 508         }
 509
 510         dump_all("opt");
 511
 512         if (firm_dump.statistic & STAT_AFTER_OPT)
 513                 stat_dump_snapshot(input_filename, "opt");
 514
 515         timer_stop(t_all_opt);
 516 }
 517
 518 /**
 519  * do Firm lowering
 520  *
 521  * @param input_filename  the name of the (main) source file
 522  */
 523 static void do_firm_lowering(const char *input_filename)
 524 {
 525         int i;
 526
 527         /* enable architecture dependent optimizations */
 528         arch_dep_set_opts((arch_dep_opts_t)
 529                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 530                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 531                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 532         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 533                 ir_graph *irg = get_irp_irg(i);
 534                 do_irg_opt(irg, "reassociation");
 535                 do_irg_opt(irg, "local");
 536         }
 537
 538         do_irp_opt("target-lowering");
 539
 540         if (firm_dump.statistic & STAT_AFTER_LOWER)
 541                 stat_dump_snapshot(input_filename, "low");
 542
 543         timer_start(t_all_opt);
 544
 545         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 546                 ir_graph *irg = get_irp_irg(i);
 547
 548                 do_irg_opt(irg, "local");
 549                 do_irg_opt(irg, "place");
 550                 do_irg_opt(irg, "control-flow");
 551                 do_irg_opt(irg, "opt-load-store");
 552                 do_irg_opt(irg, "local");
 553                 do_irg_opt(irg, "control-flow");
 554
 555                 if (do_irg_opt(irg, "vrp")) {
 556                         do_irg_opt(irg, "local");
 557                         do_irg_opt(irg, "control-flow");
 558                         do_irg_opt(irg, "vrp");
 559                         do_irg_opt(irg, "local");
 560                         do_irg_opt(irg, "control-flow");
 561                 }
 562
 563                 if (do_irg_opt(irg, "if-conversion")) {
 564                         do_irg_opt(irg, "local");
 565                         do_irg_opt(irg, "control-flow");
 566                 }
 567
 568                 do_irg_opt(irg, "parallelize-mem");
 569         }
 570         do_irp_opt("remove-unused");
 571         do_irp_opt("opt-cc");
 572         timer_stop(t_all_opt);
 573         dump_all("low-opt");
 574
 575         if (firm_dump.statistic & STAT_FINAL) {
 576                 stat_dump_snapshot(input_filename, "final");
 577         }
 578 }
 579
 580 /**
 581  * Initialize for the Firm-generating back end.
 582  */
 583 void gen_firm_init(void)
 584 {
 585         unsigned pattern = 0;
 586         int      i;
 587
 588         for (i = 0; i < n_opts; ++i) {
 589                 timers[i] = ir_timer_new();
 590                 timer_register(timers[i], opts[i].description);
 591         }
 592         t_verify = ir_timer_new();
 593         timer_register(t_verify, "Firm: verify pass");
 594         t_vcg_dump = ir_timer_new();
 595         timer_register(t_vcg_dump, "Firm: vcg dumping");
 596         t_all_opt = ir_timer_new();
 597         timer_register(t_all_opt, "Firm: all optimizations");
 598
 599         if (firm_dump.stat_pattern)
 600                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 601
 602         if (firm_dump.stat_dag)
 603                 pattern |= FIRMSTAT_COUNT_DAG;
 604
 605         ir_init(NULL);
 606         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 607                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 608                         | FIRMSTAT_COUNT_CONSTS | pattern);
 609
 610         edges_init_dbg(firm_opt.verify_edges);
 611
 612         /* Sel node cannot produce NULL pointers */
 613         set_opt_sel_based_null_check_elim(1);
 614
 615         /* dynamic dispatch works currently only if whole world scenarios */
 616         set_opt_dyn_meth_dispatch(0);
 617
 618         /* do not run architecture dependent optimizations in building phase */
 619         arch_dep_set_opts(arch_dep_none);
 620
 621         do_node_verification((firm_verification_t) firm_opt.verify);
 622         if (firm_dump.filter != NULL)
 623                 ir_set_dump_filter(firm_dump.filter);
 624         if (firm_dump.extbb)
 625                 ir_add_dump_flags(ir_dump_flag_group_extbb);
 626         if (firm_dump.no_blocks)
 627                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
 628
 629         set_optimize(1);
 630         set_opt_constant_folding(firm_opt.const_folding);
 631         set_opt_algebraic_simplification(firm_opt.const_folding);
 632         set_opt_cse(firm_opt.cse);
 633         set_opt_global_cse(0);
 634         set_opt_unreachable_code(1);
 635 }
 636
 637 /**
 638  * Called, after the Firm generation is completed,
 639  * do all optimizations and backend call here.
 640  *
 641  * @param out                a file handle for the output, may be NULL
 642  * @param input_filename     the name of the (main) source file
 643  * @param c_mode             non-zero if "C" was compiled
 644  */
 645 void gen_firm_finish(FILE *out, const char *input_filename)
 646 {
 647         int i;
 648
 649         /* the general for dumping option must be set, or the others will not work*/
 650         firm_dump.ir_graph
 651                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 652
 653         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 654                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 655         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 656
 657         /* FIXME: cloning might ADD new graphs. */
 658         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 659
 660         if (firm_dump.all_types) {
 661                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 662         }
 663
 664         /* finalize all graphs */
 665         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 666                 ir_graph *irg = get_irp_irg(i);
 667                 irg_finalize_cons(irg);
 668         }
 669         dump_all("");
 670
 671         timer_push(t_verify);
 672         tr_verify();
 673         timer_pop(t_verify);
 674
 675         /* all graphs are finalized, set the irp phase to high */
 676         set_irp_phase_state(phase_high);
 677
 678         /* BEWARE: kill unreachable code before doing compound lowering */
 679         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 680                 ir_graph *irg = get_irp_irg(i);
 681                 do_irg_opt(irg, "control-flow");
 682         }
 683
 684         /* lower copyb nodes */
 685         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 686                 ir_graph *irg = get_irp_irg(i);
 687                 lower_CopyB(irg, 128, 4);
 688         }
 689
 690         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 691                 stat_dump_snapshot(input_filename, "noopt");
 692         }
 693
 694         do_firm_optimizations(input_filename);
 695         do_firm_lowering(input_filename);
 696
 697         /* set the phase to low */
 698         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 699                 set_irg_phase_state(get_irp_irg(i), phase_low);
 700
 701         if (firm_dump.statistic & STAT_FINAL_IR)
 702                 stat_dump_snapshot(input_filename, "final-ir");
 703
 704         /* run the code generator */
 705         if (firm_be_opt.selection != BE_NONE)
 706                 do_codegen(out, input_filename);
 707
 708         if (firm_dump.statistic & STAT_FINAL)
 709                 stat_dump_snapshot(input_filename, "final");
 710 }
 711
 712 void disable_all_opts(void)
 713 {
 714         for (int i = 0; i < n_opts; ++i) {
 715                 opt_config_t *config = &opts[i];
 716                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 717                         config->flags |= OPT_FLAG_ENABLED;
 718                 } else {
 719                         config->flags &= ~OPT_FLAG_ENABLED;
 720                 }
 721         }
 722 }
 723
 724 int firm_opt_option(const char *opt)
 725 {
 726         bool enable = true;
 727         if (strncmp(opt, "no-", 3) == 0) {
 728                 enable = false;
 729                 opt = opt + 3;
 730         }
 731
 732         opt_config_t *config = get_opt(opt);
 733         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 734                 return 0;
 735
 736         config->flags &= ~OPT_FLAG_ENABLED;
 737         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 738         return 1;
 739 }
 740
 741 void firm_opt_option_help(void)
 742 {
 743         int i;
 744
 745         for (i = 0; i < n_opts; ++i) {
 746                 char buf[1024];
 747                 char buf2[1024];
 748
 749                 const opt_config_t *config = &opts[i];
 750                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 751                         continue;
 752
 753                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 754                 print_option_help(config->name, buf2);
 755                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 756                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 757                 print_option_help(buf, buf2);
 758         }
 759 }
 760
 761 /**
 762  * Do very early initializations
 763  */
 764 void firm_early_init(void)
 765 {
 766         /* arg: need this here for command line options */
 767         be_opt_register();
 768
 769         enable_safe_defaults();
 770 }