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