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                 timer_push(t_vcg_dump);
  36                 if (firm_dump.no_blocks)
  37                         dump_all_ir_graphs(dump_ir_graph, suffix);
  38                 else if (firm_dump.extbb)
  39                         dump_all_ir_graphs(dump_ir_extblock_graph, suffix);
  40                 else
  41                         dump_all_ir_graphs(dump_ir_block_graph, suffix);
  42                 timer_pop(t_vcg_dump);
  43         }
  44 }
  45
  46 /* set by the backend parameters */
  47 static const ir_settings_arch_dep_t *ad_param              = NULL;
  48 static create_intrinsic_fkt         *arch_create_intrinsic = NULL;
  49 static void                         *create_intrinsic_ctx  = NULL;
  50 static const ir_settings_if_conv_t  *if_conv_info          = NULL;
  51
  52 /* entities of runtime functions */
  53 ir_entity_ptr rts_entities[rts_max];
  54
  55 /**
  56  * factory for setting architecture dependent parameters
  57  */
  58 static const ir_settings_arch_dep_t *arch_factory(void)
  59 {
  60         static const ir_settings_arch_dep_t param = {
  61                 1,   /* also use subs */
  62                 4,   /* maximum shifts */
  63                 31,   /* maximum shift amount */
  64                 NULL, /* use default evaluator */
  65
  66                 1, /* allow Mulhs */
  67                 1, /* allow Mulus */
  68                 32  /* Mulh allowed up to 32 bit */
  69         };
  70
  71         return ad_param ? ad_param : &param;
  72 }
  73
  74 /**
  75  * Map runtime functions.
  76  */
  77 static void rts_map(void)
  78 {
  79         static const struct {
  80                 ir_entity_ptr *ent; /**< address of the rts entity */
  81                 i_mapper_func func; /**< mapper function. */
  82         } mapper[] = {
  83                 /* integer */
  84                 { &rts_entities[rts_abs],     i_mapper_abs },
  85                 { &rts_entities[rts_labs],    i_mapper_abs },
  86                 { &rts_entities[rts_llabs],   i_mapper_abs },
  87                 { &rts_entities[rts_imaxabs], i_mapper_abs },
  88
  89                 /* double -> double */
  90                 { &rts_entities[rts_fabs],    i_mapper_abs },
  91                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
  92                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
  93                 { &rts_entities[rts_pow],     i_mapper_pow },
  94                 { &rts_entities[rts_exp],     i_mapper_exp },
  95                 { &rts_entities[rts_exp2],    i_mapper_exp },
  96                 { &rts_entities[rts_exp10],   i_mapper_exp },
  97                 { &rts_entities[rts_log],     i_mapper_log },
  98                 { &rts_entities[rts_log2],    i_mapper_log2 },
  99                 { &rts_entities[rts_log10],   i_mapper_log10 },
 100                 { &rts_entities[rts_sin],     i_mapper_sin },
 101                 { &rts_entities[rts_cos],     i_mapper_cos },
 102                 { &rts_entities[rts_tan],     i_mapper_tan },
 103                 { &rts_entities[rts_asin],    i_mapper_asin },
 104                 { &rts_entities[rts_acos],    i_mapper_acos },
 105                 { &rts_entities[rts_atan],    i_mapper_atan },
 106                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 107                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 108                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 109
 110                 /* float -> float */
 111                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 112                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 113                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 114                 { &rts_entities[rts_powf],    i_mapper_pow },
 115                 { &rts_entities[rts_expf],    i_mapper_exp },
 116                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 117                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 118                 { &rts_entities[rts_logf],    i_mapper_log },
 119                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 120                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 121                 { &rts_entities[rts_sinf],    i_mapper_sin },
 122                 { &rts_entities[rts_cosf],    i_mapper_cos },
 123                 { &rts_entities[rts_tanf],    i_mapper_tan },
 124                 { &rts_entities[rts_asinf],   i_mapper_asin },
 125                 { &rts_entities[rts_acosf],   i_mapper_acos },
 126                 { &rts_entities[rts_atanf],   i_mapper_atan },
 127                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 128                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 129                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 130
 131                 /* long double -> long double */
 132                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 133                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 134                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 135                 { &rts_entities[rts_powl],    i_mapper_pow },
 136                 { &rts_entities[rts_expl],    i_mapper_exp },
 137                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 138                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 139                 { &rts_entities[rts_logl],    i_mapper_log },
 140                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 141                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 142                 { &rts_entities[rts_sinl],    i_mapper_sin },
 143                 { &rts_entities[rts_cosl],    i_mapper_cos },
 144                 { &rts_entities[rts_tanl],    i_mapper_tan },
 145                 { &rts_entities[rts_asinl],   i_mapper_asin },
 146                 { &rts_entities[rts_acosl],   i_mapper_acos },
 147                 { &rts_entities[rts_atanl],   i_mapper_atan },
 148                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 149                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 150                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 151
 152                 /* string */
 153                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 154                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 155                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 156                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 157                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 158                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 159                 { &rts_entities[rts_memmove], i_mapper_memmove },
 160                 { &rts_entities[rts_memset],  i_mapper_memset },
 161                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 162         };
 163         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
 164         unsigned i, n_map;
 165
 166         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
 167                 if (*mapper[i].ent != NULL) {
 168                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 169                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 170                         rec[n_map].i_call.i_mapper = mapper[i].func;
 171                         rec[n_map].i_call.ctx      = NULL;
 172                         rec[n_map].i_call.link     = NULL;
 173                         ++n_map;
 174                 }
 175         }
 176
 177         if (n_map > 0)
 178                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 179 }
 180
 181 static int *irg_dump_no;
 182
 183 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
 184 {
 185         char name[64];
 186         snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++,
 187                  suffix);
 188
 189         timer_push(t_vcg_dump);
 190         if (firm_dump.no_blocks)
 191                 dump_ir_graph(irg, name);
 192         else if (firm_dump.extbb)
 193                 dump_ir_extblock_graph(irg, name);
 194         else
 195                 dump_ir_block_graph(irg, name);
 196         timer_pop(t_vcg_dump);
 197 }
 198
 199 static int firm_const_exists;
 200
 201 static void do_optimize_funccalls(void)
 202 {
 203         optimize_funccalls(firm_const_exists, NULL);
 204 }
 205
 206 static void do_gcse(ir_graph *irg)
 207 {
 208         set_opt_global_cse(1);
 209         optimize_graph_df(irg);
 210         place_code(irg);
 211         set_opt_global_cse(0);
 212 }
 213
 214 static void do_lower_highlevel(ir_graph *irg)
 215 {
 216         lower_highlevel_graph(irg, firm_opt.lower_bitfields);
 217 }
 218
 219 static void do_if_conv(ir_graph *irg)
 220 {
 221         opt_if_conv(irg, if_conv_info);
 222 }
 223
 224 static void do_stred(ir_graph *irg)
 225 {
 226         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 227 }
 228
 229 static void after_inline_opt(ir_graph *irg)
 230 {
 231         do_irg_opt(irg, "scalar-replace");
 232         do_irg_opt(irg, "local");
 233         do_irg_opt(irg, "control-flow");
 234         do_irg_opt(irg, "combo");
 235 }
 236
 237 static void do_inline(void)
 238 {
 239         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 240                          after_inline_opt);
 241 }
 242
 243 static void do_cloning(void)
 244 {
 245         proc_cloning((float) firm_opt.clone_threshold);
 246 }
 247
 248 static void do_lower_switch(ir_graph *irg)
 249 {
 250         lower_switch(irg, firm_opt.spare_size);
 251 }
 252
 253 static void do_lower_mux(ir_graph *irg)
 254 {
 255         lower_mux(irg, NULL);
 256 }
 257
 258 static void do_lower_dw_ops(void)
 259 {
 260         lwrdw_param_t init = {
 261                 1,
 262                 1,
 263                 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
 264                 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
 265                 get_atomic_mode(ATOMIC_TYPE_INT),
 266                 get_atomic_mode(ATOMIC_TYPE_UINT),
 267                 def_create_intrinsic_fkt,
 268                 NULL
 269         };
 270
 271         if (arch_create_intrinsic) {
 272                 init.create_intrinsic = arch_create_intrinsic;
 273                 init.ctx              = create_intrinsic_ctx;
 274         }
 275         lower_dw_ops(&init);
 276 }
 277
 278 static void do_vrp(ir_graph *irg)
 279 {
 280         set_vrp_data(irg);
 281 }
 282
 283 typedef enum opt_target {
 284         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 285         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 286 } opt_target_t;
 287
 288 typedef enum opt_flags {
 289         OPT_FLAG_NONE         = 0,
 290         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 291         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 292         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 293         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 294                                              -foptions for this transformation */
 295         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 296                                              so we need it even with -O0 */
 297 } opt_flags_t;
 298
 299 typedef void (*transform_irg_func)(ir_graph *irg);
 300 typedef void (*transform_irp_func)(void);
 301
 302 typedef struct {
 303         opt_target_t  target;
 304         const char   *name;
 305         union {
 306                 transform_irg_func transform_irg;
 307                 transform_irp_func transform_irp;
 308         } u;
 309         const char   *description;
 310         opt_flags_t   flags;
 311 } opt_config_t;
 312
 313 static opt_config_t opts[] = {
 314 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 315 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 316         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 317         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 318         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
 319         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 320         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 321         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 322         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 323         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 324         IRG("gcse",              do_gcse,                  "global common subexpression elimination",               OPT_FLAG_NONE),
 325         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 326         IRG("if-conversion",     do_if_conv,               "if-conversion",                                         OPT_FLAG_NONE),
 327         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 328         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 329         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 330         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 331         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 332         IRG("lower-switch",      do_lower_switch,          "switch lowering",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 333         IRG("one-return",        normalize_one_return,     "normalisation to 1 return",                             OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 334         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 335         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 336         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 337         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 338         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 339         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 340         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 341         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 342         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 343         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 344         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 345         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
 346         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 347         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),
 348         IRP("lower-dw",          do_lower_dw_ops,          "lowering of doubleword operations",                     OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 349         IRP("opt-func-call",     do_optimize_funccalls,    "function call optimization",                            OPT_FLAG_NONE),
 350         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 351         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 352         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
 353 #undef IRP
 354 #undef IRG
 355 };
 356 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 357 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 358
 359 static opt_config_t *get_opt(const char *name)
 360 {
 361         int i;
 362         for (i = 0; i < n_opts; ++i) {
 363                 opt_config_t *config = &opts[i];
 364                 if (strcmp(config->name, name) == 0)
 365                         return config;
 366         }
 367
 368         return NULL;
 369 }
 370
 371 static void set_opt_enabled(const char *name, bool enabled)
 372 {
 373         opt_config_t *config = get_opt(name);
 374         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 375                 | (enabled ? OPT_FLAG_ENABLED : 0);
 376 }
 377
 378 static bool get_opt_enabled(const char *name)
 379 {
 380         opt_config_t *config = get_opt(name);
 381         return (config->flags & OPT_FLAG_ENABLED) != 0;
 382 }
 383
 384 /**
 385  * perform an optimisation on a single graph
 386  *
 387  * @return  true if something changed, false otherwise
 388  */
 389 static bool do_irg_opt(ir_graph *irg, const char *name)
 390 {
 391         ir_graph     *old_irg;
 392         opt_config_t *config = get_opt(name);
 393         size_t        n      = config - opts;
 394         assert(config != NULL);
 395         assert(config->target == OPT_TARGET_IRG);
 396         if (! (config->flags & OPT_FLAG_ENABLED))
 397                 return false;
 398
 399         old_irg          = current_ir_graph;
 400         current_ir_graph = irg;
 401
 402         timer_push(timers[n]);
 403         config->u.transform_irg(irg);
 404         timer_pop(timers[n]);
 405
 406         if (firm_dump.all_phases && firm_dump.ir_graph) {
 407                 dump_graph_count(irg, name);
 408         }
 409
 410         if (firm_opt.check_all) {
 411                 timer_push(t_verify);
 412                 irg_verify(irg, VRFY_ENFORCE_SSA);
 413                 timer_pop(t_verify);
 414         }
 415
 416         current_ir_graph = old_irg;
 417         return true;
 418 }
 419
 420 static void do_irp_opt(const char *name)
 421 {
 422         opt_config_t *config = get_opt(name);
 423         size_t        n      = config - opts;
 424         assert(config->target == OPT_TARGET_IRP);
 425         if (! (config->flags & OPT_FLAG_ENABLED))
 426                 return;
 427
 428         timer_push(timers[n]);
 429         config->u.transform_irp();
 430         timer_pop(timers[n]);
 431
 432         if (firm_dump.ir_graph && firm_dump.all_phases) {
 433                 int i;
 434                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 435                         ir_graph *irg = get_irp_irg(i);
 436                         dump_graph_count(irg, name);
 437                 }
 438         }
 439
 440         if (firm_opt.check_all) {
 441                 int i;
 442                 timer_push(t_verify);
 443                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 444                         irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
 445                 }
 446                 timer_pop(t_verify);
 447         }
 448 }
 449
 450 /**
 451  * Enable transformations which should be always safe (and cheap) to perform
 452  */
 453 static void enable_safe_defaults(void)
 454 {
 455         set_opt_enabled("remove-unused", true);
 456         set_opt_enabled("opt-tail-rec", true);
 457         set_opt_enabled("opt-func-call", true);
 458         set_opt_enabled("reassociation", true);
 459         set_opt_enabled("control-flow", true);
 460         set_opt_enabled("local", true);
 461         set_opt_enabled("lower-const", true);
 462         set_opt_enabled("scalar-replace", true);
 463         set_opt_enabled("place", true);
 464         set_opt_enabled("confirm", true);
 465         set_opt_enabled("opt-load-store", true);
 466         set_opt_enabled("lower", true);
 467         set_opt_enabled("deconv", true);
 468         set_opt_enabled("remove-confirms", true);
 469         set_opt_enabled("ivopts", true);
 470         set_opt_enabled("dead", true);
 471         set_opt_enabled("lower-switch", true);
 472         set_opt_enabled("remove-phi-cycles", true);
 473         set_opt_enabled("frame", true);
 474         set_opt_enabled("combo", true);
 475         set_opt_enabled("invert-loops", true);
 476 }
 477
 478 /**
 479  * run all the Firm optimizations
 480  *
 481  * @param input_filename     the name of the (main) source file
 482  */
 483 static void do_firm_optimizations(const char *input_filename)
 484 {
 485         int      i;
 486         unsigned aa_opt;
 487
 488         set_opt_alias_analysis(firm_opt.alias_analysis);
 489
 490         aa_opt = aa_opt_no_opt;
 491         if (firm_opt.strict_alias)
 492                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 493         if (firm_opt.no_alias)
 494                 aa_opt = aa_opt_no_alias;
 495
 496         set_irp_memory_disambiguator_options(aa_opt);
 497
 498         /* parameter passing code should set them directly sometime... */
 499         set_opt_enabled("rts", !firm_opt.freestanding);
 500         set_opt_enabled("gcse", firm_opt.gcse);
 501         set_opt_enabled("place", !firm_opt.gcse);
 502         set_opt_enabled("confirm", firm_opt.confirm);
 503         set_opt_enabled("remove-confirms", firm_opt.confirm);
 504
 505         /* osr supersedes remove_phi_cycles */
 506         if (get_opt_enabled("ivopts"))
 507                 set_opt_enabled("remove-phi-cycles", false);
 508
 509         timer_start(t_all_opt);
 510
 511         do_irp_opt("rts");
 512
 513         /* first step: kill dead code */
 514         for (i = 0; i < get_irp_n_irgs(); i++) {
 515                 ir_graph *irg = get_irp_irg(i);
 516                 do_irg_opt(irg, "combo");
 517                 do_irg_opt(irg, "local");
 518                 do_irg_opt(irg, "control-flow");
 519         }
 520
 521         do_irp_opt("remove-unused");
 522         for (i = 0; i < get_irp_n_irgs(); ++i) {
 523                 ir_graph *irg = get_irp_irg(i);
 524                 do_irg_opt(irg, "opt-tail-rec");
 525         }
 526         do_irp_opt("opt-func-call");
 527         do_irp_opt("lower-const");
 528
 529         for (i = 0; i < get_irp_n_irgs(); i++) {
 530                 ir_graph *irg = get_irp_irg(i);
 531
 532                 do_irg_opt(irg, "scalar-replace");
 533                 do_irg_opt(irg, "invert-loops");
 534                 do_irg_opt(irg, "unroll-loops");
 535                 do_irg_opt(irg, "local");
 536                 do_irg_opt(irg, "reassociation");
 537                 do_irg_opt(irg, "local");
 538                 do_irg_opt(irg, "gcse");
 539
 540                 if (firm_opt.confirm) {
 541                         /* Confirm construction currently can only handle blocks with only
 542                            one control flow predecessor. Calling optimize_cf here removes
 543                            Bad predecessors and help the optimization of switch constructs.
 544                          */
 545                         do_irg_opt(irg, "control-flow");
 546                         do_irg_opt(irg, "confirm");
 547                         do_irg_opt(irg, "vrp");
 548                         do_irg_opt(irg, "local");
 549                 }
 550
 551                 do_irg_opt(irg, "control-flow");
 552                 do_irg_opt(irg, "opt-load-store");
 553                 do_irg_opt(irg, "fp-vrp");
 554                 do_irg_opt(irg, "lower");
 555                 do_irg_opt(irg, "deconv");
 556                 do_irg_opt(irg, "thread-jumps");
 557                 do_irg_opt(irg, "remove-confirms");
 558                 do_irg_opt(irg, "gvn-pre");
 559                 do_irg_opt(irg, "place");
 560                 do_irg_opt(irg, "control-flow");
 561
 562                 if (do_irg_opt(irg, "if-conversion")) {
 563                         do_irg_opt(irg, "local");
 564                         do_irg_opt(irg, "control-flow");
 565                 }
 566                 /* this doesn't make too much sense but tests the mux destruction... */
 567                 do_irg_opt(irg, "lower-mux");
 568
 569                 do_irg_opt(irg, "bool");
 570                 do_irg_opt(irg, "shape-blocks");
 571                 do_irg_opt(irg, "lower-switch");
 572                 do_irg_opt(irg, "ivopts");
 573                 do_irg_opt(irg, "local");
 574                 do_irg_opt(irg, "dead");
 575                 do_irg_opt(irg, "frame");
 576         }
 577
 578         do_irp_opt("inline");
 579         do_irp_opt("opt-proc-clone");
 580
 581         for (i = 0; i < get_irp_n_irgs(); i++) {
 582                 ir_graph *irg = get_irp_irg(i);
 583                 do_irg_opt(irg, "local");
 584                 do_irg_opt(irg, "control-flow");
 585                 do_irg_opt(irg, "thread-jumps");
 586                 do_irg_opt(irg, "local");
 587                 do_irg_opt(irg, "control-flow");
 588
 589                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 590                         do_irg_opt(irg, "local");
 591                         do_irg_opt(irg, "vrp");
 592                         do_irg_opt(irg, "local");
 593                         do_irg_opt(irg, "vrp");
 594                 }
 595         }
 596
 597         if (firm_dump.ir_graph) {
 598                 /* recompute backedges for nicer dumps */
 599                 for (i = 0; i < get_irp_n_irgs(); i++)
 600                         construct_cf_backedges(get_irp_irg(i));
 601         }
 602
 603         dump_all("-opt");
 604
 605         if (firm_dump.statistic & STAT_AFTER_OPT)
 606                 stat_dump_snapshot(input_filename, "opt");
 607
 608         timer_stop(t_all_opt);
 609 }
 610
 611 /**
 612  * do Firm lowering
 613  *
 614  * @param input_filename  the name of the (main) source file
 615  */
 616 static void do_firm_lowering(const char *input_filename)
 617 {
 618         int i;
 619
 620         do_irp_opt("lower-dw");
 621
 622         if (firm_dump.statistic & STAT_AFTER_LOWER)
 623                 stat_dump_snapshot(input_filename, "low");
 624
 625         dump_all("-low");
 626
 627         if (firm_opt.enabled) {
 628                 timer_start(t_all_opt);
 629
 630                 /* run reassociation first on all graphs BEFORE the architecture
 631                    dependent optimizations are enabled */
 632                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 633                         ir_graph *irg = get_irp_irg(i);
 634                         do_irg_opt(irg, "reassociation");
 635                 }
 636
 637                 /* enable architecture dependent optimizations */
 638                 arch_dep_set_opts((arch_dep_opts_t)
 639                                 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 640                                  (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 641                                  (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 642
 643                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 644                         ir_graph *irg = get_irp_irg(i);
 645
 646                         do_irg_opt(irg, "local");
 647                         do_irg_opt(irg, "gcse");
 648                         do_irg_opt(irg, "opt-load-store");
 649                         do_irg_opt(irg, "local");
 650                         do_irg_opt(irg, "control-flow");
 651
 652                         if (do_irg_opt(irg, "vrp")) {
 653                                 do_irg_opt(irg, "local");
 654                                 do_irg_opt(irg, "control-flow");
 655                                 do_irg_opt(irg, "vrp");
 656                                 do_irg_opt(irg, "local");
 657                                 do_irg_opt(irg, "control-flow");
 658                         }
 659
 660                         if (do_irg_opt(irg, "if-conversion")) {
 661                                 do_irg_opt(irg, "local");
 662                                 do_irg_opt(irg, "control-flow");
 663                         }
 664
 665                         do_irg_opt(irg, "parallelize-mem");
 666                 }
 667                 timer_stop(t_all_opt);
 668
 669                 do_irp_opt("remove-unused");
 670
 671                 dump_all("-low-opt");
 672         }
 673
 674         if (firm_opt.cc_opt)
 675                 mark_private_methods();
 676
 677         /* set the phase to low */
 678         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 679                 set_irg_phase_low(get_irp_irg(i));
 680
 681         /* all graphs are lowered, set the irp phase to low */
 682         set_irp_phase_state(phase_low);
 683
 684         if (firm_dump.statistic & STAT_FINAL) {
 685                 stat_dump_snapshot(input_filename, "final");
 686         }
 687 }
 688
 689 /**
 690  * Initialize for the Firm-generating back end.
 691  */
 692 void gen_firm_init(void)
 693 {
 694         unsigned pattern = 0;
 695         int      i;
 696
 697         for (i = 0; i < n_opts; ++i) {
 698                 timers[i] = ir_timer_new();
 699                 timer_register(timers[i], opts[i].description);
 700         }
 701         t_verify = ir_timer_new();
 702         timer_register(t_verify, "Firm: verify pass");
 703         t_vcg_dump = ir_timer_new();
 704         timer_register(t_vcg_dump, "Firm: vcg dumping");
 705         t_all_opt = ir_timer_new();
 706         timer_register(t_all_opt, "Firm: all optimizations");
 707
 708         if (firm_dump.stat_pattern)
 709                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 710
 711         if (firm_dump.stat_dag)
 712                 pattern |= FIRMSTAT_COUNT_DAG;
 713
 714         ir_init(NULL);
 715         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 716                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 717                         | FIRMSTAT_COUNT_CONSTS | pattern);
 718
 719         if (firm_be_opt.selection == BE_FIRM_BE) {
 720                 const backend_params *be_params = be_get_backend_param();
 721
 722                 if (be_params->do_dw_lowering)
 723                         set_opt_enabled("lower-dw", true);
 724
 725                 arch_create_intrinsic   = be_params->arch_create_intrinsic_fkt;
 726                 create_intrinsic_ctx    = be_params->create_intrinsic_ctx;
 727
 728                 ad_param                = be_params->dep_param;
 729                 if_conv_info            = be_params->if_conv_info;
 730         }
 731
 732         edges_init_dbg(firm_opt.vrfy_edges);
 733
 734         /* Sel node cannot produce NULL pointers */
 735         set_opt_sel_based_null_check_elim(1);
 736
 737         /* dynamic dispatch works currently only if whole world scenarios */
 738         set_opt_dyn_meth_dispatch(0);
 739
 740         arch_dep_init(arch_factory);
 741
 742         /* do not run architecture dependent optimizations in building phase */
 743         arch_dep_set_opts(arch_dep_none);
 744
 745         do_node_verification((firm_verification_t) firm_opt.vrfy);
 746         if (firm_dump.filter)
 747                 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
 748
 749         if (firm_opt.enabled) {
 750                 set_optimize(1);
 751                 set_opt_constant_folding(firm_opt.const_folding);
 752                 set_opt_algebraic_simplification(firm_opt.const_folding);
 753                 set_opt_cse(firm_opt.cse);
 754                 set_opt_global_cse(0);
 755                 set_opt_unreachable_code(1);
 756                 set_opt_control_flow(firm_opt.control_flow);
 757                 set_opt_control_flow_weak_simplification(1);
 758                 set_opt_control_flow_strong_simplification(1);
 759         } else {
 760                 set_optimize(0);
 761         }
 762
 763         /* do not dump entity ld names */
 764         dump_ld_names(0);
 765 }
 766
 767 /**
 768  * Called, after the Firm generation is completed,
 769  * do all optimizations and backend call here.
 770  *
 771  * @param out                a file handle for the output, may be NULL
 772  * @param input_filename     the name of the (main) source file
 773  * @param c_mode             non-zero if "C" was compiled
 774  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
 775  */
 776 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode,
 777                      int new_firm_const_exists)
 778 {
 779         int i;
 780
 781 #if 0
 782         if (firm_opt.enable_statev) {
 783                 char buf[1024];
 784                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
 785                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
 786                 ir_stat_ev_compilation_unit(input_filename);
 787         }
 788 #endif
 789
 790         firm_const_exists = new_firm_const_exists;
 791
 792         /* the general for dumping option must be set, or the others will not work*/
 793         firm_dump.ir_graph
 794                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 795
 796         dump_keepalive_edges(1);
 797         dump_consts_local(1);
 798         dump_dominator_information(1);
 799         dump_loop_information(0);
 800
 801         if (!firm_dump.edge_labels)
 802                 turn_off_edge_labels();
 803
 804         /* FIXME: cloning might ADD new graphs. */
 805         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 806
 807         if (firm_dump.all_types) {
 808                 dump_all_types("");
 809                 if (! c_mode) {
 810                         dump_class_hierarchy(0, "");
 811                         dump_class_hierarchy(1, "-with-entities");
 812                 }
 813         }
 814
 815         /* finalize all graphs */
 816         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 817                 ir_graph *irg = get_irp_irg(i);
 818                 irg_finalize_cons(irg);
 819         }
 820         dump_all("");
 821
 822         timer_push(t_verify);
 823         tr_vrfy();
 824         timer_pop(t_verify);
 825
 826         /* all graphs are finalized, set the irp phase to high */
 827         set_irp_phase_state(phase_high);
 828
 829         /* BEWARE: kill unreachable code before doing compound lowering */
 830         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 831                 ir_graph *irg = get_irp_irg(i);
 832                 do_irg_opt(irg, "control-flow");
 833         }
 834
 835         /* lower all compound call return values */
 836         lower_compound_params();
 837
 838         /* lower copyb nodes */
 839         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 840                 ir_graph *irg = get_irp_irg(i);
 841                 lower_CopyB(irg, 128, 4);
 842         }
 843
 844         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 845                 stat_dump_snapshot(input_filename, "noopt");
 846         }
 847
 848         if (firm_opt.enabled)
 849                 do_firm_optimizations(input_filename);
 850
 851         if (firm_opt.lower)
 852                 do_firm_lowering(input_filename);
 853
 854         /* set the phase to low */
 855         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 856                 set_irg_phase_low(get_irp_irg(i));
 857
 858         if (firm_dump.statistic & STAT_FINAL_IR)
 859                 stat_dump_snapshot(input_filename, "final-ir");
 860
 861         /* run the code generator */
 862         if (firm_be_opt.selection != BE_NONE)
 863                 do_codegen(out, input_filename);
 864
 865         if (firm_dump.statistic & STAT_FINAL)
 866                 stat_dump_snapshot(input_filename, "final");
 867 }
 868
 869 void disable_all_opts(void)
 870 {
 871         for (int i = 0; i < n_opts; ++i) {
 872                 opt_config_t *config = &opts[i];
 873                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 874                         config->flags |= OPT_FLAG_ENABLED;
 875                 } else {
 876                         config->flags &= ~OPT_FLAG_ENABLED;
 877                 }
 878         }
 879 }
 880
 881 int firm_opt_option(const char *opt)
 882 {
 883         bool enable = true;
 884         if (strncmp(opt, "no-", 3) == 0) {
 885                 enable = false;
 886                 opt = opt + 3;
 887         }
 888
 889         opt_config_t *config = get_opt(opt);
 890         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 891                 return 0;
 892
 893         config->flags &= ~OPT_FLAG_ENABLED;
 894         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 895         return 1;
 896 }
 897
 898 void firm_opt_option_help(void)
 899 {
 900         int i;
 901
 902         for (i = 0; i < n_opts; ++i) {
 903                 char buf[1024];
 904                 char buf2[1024];
 905
 906                 const opt_config_t *config = &opts[i];
 907                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 908                         continue;
 909
 910                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 911                 print_option_help(config->name, buf2);
 912                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 913                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 914                 print_option_help(buf, buf2);
 915         }
 916 }
 917
 918 /**
 919  * Do very early initializations
 920  */
 921 void firm_early_init(void)
 922 {
 923         /* arg: need this here for command line options */
 924         be_opt_register();
 925
 926         enable_safe_defaults();
 927 }