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 static int firm_const_exists;
 154
 155 static void do_optimize_funccalls(void)
 156 {
 157         optimize_funccalls(firm_const_exists, NULL);
 158 }
 159
 160 static void do_gcse(ir_graph *irg)
 161 {
 162         set_opt_global_cse(1);
 163         optimize_graph_df(irg);
 164         place_code(irg);
 165         set_opt_global_cse(0);
 166 }
 167
 168 static void do_lower_highlevel(ir_graph *irg)
 169 {
 170         lower_highlevel_graph(irg, firm_opt.lower_bitfields);
 171 }
 172
 173 static void do_stred(ir_graph *irg)
 174 {
 175         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 176 }
 177
 178 static void after_inline_opt(ir_graph *irg)
 179 {
 180         do_irg_opt(irg, "scalar-replace");
 181         do_irg_opt(irg, "local");
 182         do_irg_opt(irg, "control-flow");
 183         do_irg_opt(irg, "combo");
 184 }
 185
 186 static void do_inline(void)
 187 {
 188         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 189                          after_inline_opt);
 190 }
 191
 192 static void do_cloning(void)
 193 {
 194         proc_cloning((float) firm_opt.clone_threshold);
 195 }
 196
 197 static void do_lower_mux(ir_graph *irg)
 198 {
 199         lower_mux(irg, NULL);
 200 }
 201
 202 static void do_vrp(ir_graph *irg)
 203 {
 204         set_vrp_data(irg);
 205 }
 206
 207 typedef enum opt_target {
 208         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 209         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 210 } opt_target_t;
 211
 212 typedef enum opt_flags {
 213         OPT_FLAG_NONE         = 0,
 214         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 215         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 216         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 217         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 218                                              -foptions for this transformation */
 219         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 220                                              so we need it even with -O0 */
 221 } opt_flags_t;
 222
 223 typedef void (*transform_irg_func)(ir_graph *irg);
 224 typedef void (*transform_irp_func)(void);
 225
 226 typedef struct {
 227         opt_target_t  target;
 228         const char   *name;
 229         union {
 230                 transform_irg_func transform_irg;
 231                 transform_irp_func transform_irp;
 232         } u;
 233         const char   *description;
 234         opt_flags_t   flags;
 235 } opt_config_t;
 236
 237 static opt_config_t opts[] = {
 238 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 239 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 240         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 241         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 242         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
 243         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 244         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 245         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 246         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 247         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 248         IRG("gcse",              do_gcse,                  "global common subexpression elimination",               OPT_FLAG_NONE),
 249         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 250         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 251         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 252         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 253         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 254         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 255         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 256         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 257         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 258         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 259         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 260         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 261         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 262         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 263         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 264         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 265         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 266         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 267         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
 268         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 269         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),
 270         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 271         IRP("opt-func-call",     do_optimize_funccalls,    "function call optimization",                            OPT_FLAG_NONE),
 272         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 273         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 274         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
 275 #undef IRP
 276 #undef IRG
 277 };
 278 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 279 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 280
 281 static opt_config_t *get_opt(const char *name)
 282 {
 283         int i;
 284         for (i = 0; i < n_opts; ++i) {
 285                 opt_config_t *config = &opts[i];
 286                 if (strcmp(config->name, name) == 0)
 287                         return config;
 288         }
 289
 290         return NULL;
 291 }
 292
 293 static void set_opt_enabled(const char *name, bool enabled)
 294 {
 295         opt_config_t *config = get_opt(name);
 296         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 297                 | (enabled ? OPT_FLAG_ENABLED : 0);
 298 }
 299
 300 static bool get_opt_enabled(const char *name)
 301 {
 302         opt_config_t *config = get_opt(name);
 303         return (config->flags & OPT_FLAG_ENABLED) != 0;
 304 }
 305
 306 /**
 307  * perform an optimisation on a single graph
 308  *
 309  * @return  true if something changed, false otherwise
 310  */
 311 static bool do_irg_opt(ir_graph *irg, const char *name)
 312 {
 313         ir_graph     *old_irg;
 314         opt_config_t *config = get_opt(name);
 315         size_t        n      = config - opts;
 316         assert(config != NULL);
 317         assert(config->target == OPT_TARGET_IRG);
 318         if (! (config->flags & OPT_FLAG_ENABLED))
 319                 return false;
 320
 321         old_irg          = current_ir_graph;
 322         current_ir_graph = irg;
 323
 324         timer_push(timers[n]);
 325         config->u.transform_irg(irg);
 326         timer_pop(timers[n]);
 327
 328         if (firm_dump.all_phases && firm_dump.ir_graph) {
 329                 dump_ir_graph(irg, name);
 330         }
 331
 332         if (firm_opt.check_all) {
 333                 timer_push(t_verify);
 334                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 335                 timer_pop(t_verify);
 336         }
 337
 338         current_ir_graph = old_irg;
 339         return true;
 340 }
 341
 342 static void do_irp_opt(const char *name)
 343 {
 344         opt_config_t *config = get_opt(name);
 345         size_t        n      = config - opts;
 346         assert(config->target == OPT_TARGET_IRP);
 347         if (! (config->flags & OPT_FLAG_ENABLED))
 348                 return;
 349
 350         timer_push(timers[n]);
 351         config->u.transform_irp();
 352         timer_pop(timers[n]);
 353
 354         if (firm_dump.ir_graph && firm_dump.all_phases) {
 355                 int i;
 356                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 357                         ir_graph *irg = get_irp_irg(i);
 358                         dump_ir_graph(irg, name);
 359                 }
 360         }
 361
 362         if (firm_opt.check_all) {
 363                 int i;
 364                 timer_push(t_verify);
 365                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 366                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 367                 }
 368                 timer_pop(t_verify);
 369         }
 370 }
 371
 372 /**
 373  * Enable transformations which should be always safe (and cheap) to perform
 374  */
 375 static void enable_safe_defaults(void)
 376 {
 377         set_opt_enabled("remove-unused", true);
 378         set_opt_enabled("opt-tail-rec", true);
 379         set_opt_enabled("opt-func-call", true);
 380         set_opt_enabled("reassociation", true);
 381         set_opt_enabled("control-flow", true);
 382         set_opt_enabled("local", true);
 383         set_opt_enabled("lower-const", true);
 384         set_opt_enabled("scalar-replace", true);
 385         set_opt_enabled("place", true);
 386         set_opt_enabled("confirm", true);
 387         set_opt_enabled("opt-load-store", true);
 388         set_opt_enabled("lower", true);
 389         set_opt_enabled("deconv", true);
 390         set_opt_enabled("remove-confirms", true);
 391         set_opt_enabled("ivopts", true);
 392         set_opt_enabled("dead", true);
 393         set_opt_enabled("remove-phi-cycles", true);
 394         set_opt_enabled("frame", true);
 395         set_opt_enabled("combo", true);
 396         set_opt_enabled("invert-loops", true);
 397         set_opt_enabled("target-lowering", true);
 398         set_opt_enabled("rts", true);
 399         set_opt_enabled("parallelize-mem", true);
 400 }
 401
 402 /**
 403  * run all the Firm optimizations
 404  *
 405  * @param input_filename     the name of the (main) source file
 406  */
 407 static void do_firm_optimizations(const char *input_filename)
 408 {
 409         size_t   i;
 410         unsigned aa_opt;
 411
 412         set_opt_alias_analysis(firm_opt.alias_analysis);
 413
 414         aa_opt = aa_opt_no_opt;
 415         if (firm_opt.strict_alias)
 416                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 417         if (firm_opt.no_alias)
 418                 aa_opt = aa_opt_no_alias;
 419
 420         set_irp_memory_disambiguator_options(aa_opt);
 421
 422         /* parameter passing code should set them directly sometime... */
 423         set_opt_enabled("gcse", firm_opt.gcse);
 424         set_opt_enabled("place", !firm_opt.gcse);
 425         set_opt_enabled("confirm", firm_opt.confirm);
 426         set_opt_enabled("remove-confirms", firm_opt.confirm);
 427
 428         /* osr supersedes remove_phi_cycles */
 429         if (get_opt_enabled("ivopts"))
 430                 set_opt_enabled("remove-phi-cycles", false);
 431
 432         timer_start(t_all_opt);
 433
 434         do_irp_opt("rts");
 435
 436         /* first step: kill dead code */
 437         for (i = 0; i < get_irp_n_irgs(); i++) {
 438                 ir_graph *irg = get_irp_irg(i);
 439                 do_irg_opt(irg, "combo");
 440                 do_irg_opt(irg, "local");
 441                 do_irg_opt(irg, "control-flow");
 442         }
 443
 444         do_irp_opt("remove-unused");
 445         for (i = 0; i < get_irp_n_irgs(); ++i) {
 446                 ir_graph *irg = get_irp_irg(i);
 447                 do_irg_opt(irg, "opt-tail-rec");
 448         }
 449         do_irp_opt("opt-func-call");
 450         do_irp_opt("lower-const");
 451
 452         for (i = 0; i < get_irp_n_irgs(); i++) {
 453                 ir_graph *irg = get_irp_irg(i);
 454
 455                 do_irg_opt(irg, "scalar-replace");
 456                 do_irg_opt(irg, "invert-loops");
 457                 do_irg_opt(irg, "unroll-loops");
 458                 do_irg_opt(irg, "local");
 459                 do_irg_opt(irg, "reassociation");
 460                 do_irg_opt(irg, "local");
 461                 do_irg_opt(irg, "gcse");
 462
 463                 if (firm_opt.confirm) {
 464                         /* Confirm construction currently can only handle blocks with only
 465                            one control flow predecessor. Calling optimize_cf here removes
 466                            Bad predecessors and help the optimization of switch constructs.
 467                          */
 468                         do_irg_opt(irg, "control-flow");
 469                         do_irg_opt(irg, "confirm");
 470                         do_irg_opt(irg, "vrp");
 471                         do_irg_opt(irg, "local");
 472                 }
 473
 474                 do_irg_opt(irg, "control-flow");
 475                 do_irg_opt(irg, "opt-load-store");
 476                 do_irg_opt(irg, "fp-vrp");
 477                 do_irg_opt(irg, "lower");
 478                 do_irg_opt(irg, "deconv");
 479                 do_irg_opt(irg, "thread-jumps");
 480                 do_irg_opt(irg, "remove-confirms");
 481                 do_irg_opt(irg, "gvn-pre");
 482                 do_irg_opt(irg, "place");
 483                 do_irg_opt(irg, "control-flow");
 484
 485                 if (do_irg_opt(irg, "if-conversion")) {
 486                         do_irg_opt(irg, "local");
 487                         do_irg_opt(irg, "control-flow");
 488                 }
 489                 /* this doesn't make too much sense but tests the mux destruction... */
 490                 do_irg_opt(irg, "lower-mux");
 491
 492                 do_irg_opt(irg, "bool");
 493                 do_irg_opt(irg, "shape-blocks");
 494                 do_irg_opt(irg, "ivopts");
 495                 do_irg_opt(irg, "local");
 496                 do_irg_opt(irg, "dead");
 497                 do_irg_opt(irg, "frame");
 498         }
 499
 500         do_irp_opt("inline");
 501         do_irp_opt("opt-proc-clone");
 502
 503         for (i = 0; i < get_irp_n_irgs(); i++) {
 504                 ir_graph *irg = get_irp_irg(i);
 505                 do_irg_opt(irg, "local");
 506                 do_irg_opt(irg, "control-flow");
 507                 do_irg_opt(irg, "thread-jumps");
 508                 do_irg_opt(irg, "local");
 509                 do_irg_opt(irg, "control-flow");
 510
 511                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 512                         do_irg_opt(irg, "local");
 513                         do_irg_opt(irg, "vrp");
 514                         do_irg_opt(irg, "local");
 515                         do_irg_opt(irg, "vrp");
 516                 }
 517         }
 518
 519         if (firm_dump.ir_graph) {
 520                 /* recompute backedges for nicer dumps */
 521                 for (i = 0; i < get_irp_n_irgs(); i++)
 522                         construct_cf_backedges(get_irp_irg(i));
 523         }
 524
 525         dump_all("opt");
 526
 527         if (firm_dump.statistic & STAT_AFTER_OPT)
 528                 stat_dump_snapshot(input_filename, "opt");
 529
 530         timer_stop(t_all_opt);
 531 }
 532
 533 /**
 534  * do Firm lowering
 535  *
 536  * @param input_filename  the name of the (main) source file
 537  */
 538 static void do_firm_lowering(const char *input_filename)
 539 {
 540         int i;
 541
 542         /* enable architecture dependent optimizations */
 543         arch_dep_set_opts((arch_dep_opts_t)
 544                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 545                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 546                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 547         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 548                 ir_graph *irg = get_irp_irg(i);
 549                 do_irg_opt(irg, "reassociation");
 550                 do_irg_opt(irg, "local");
 551         }
 552
 553         do_irp_opt("target-lowering");
 554
 555         if (firm_dump.statistic & STAT_AFTER_LOWER)
 556                 stat_dump_snapshot(input_filename, "low");
 557
 558         if (firm_opt.enabled) {
 559                 timer_start(t_all_opt);
 560
 561                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 562                         ir_graph *irg = get_irp_irg(i);
 563
 564                         do_irg_opt(irg, "local");
 565                         do_irg_opt(irg, "gcse");
 566                         do_irg_opt(irg, "control-flow");
 567                         do_irg_opt(irg, "opt-load-store");
 568                         do_irg_opt(irg, "local");
 569                         do_irg_opt(irg, "control-flow");
 570
 571                         if (do_irg_opt(irg, "vrp")) {
 572                                 do_irg_opt(irg, "local");
 573                                 do_irg_opt(irg, "control-flow");
 574                                 do_irg_opt(irg, "vrp");
 575                                 do_irg_opt(irg, "local");
 576                                 do_irg_opt(irg, "control-flow");
 577                         }
 578
 579                         if (do_irg_opt(irg, "if-conversion")) {
 580                                 do_irg_opt(irg, "local");
 581                                 do_irg_opt(irg, "control-flow");
 582                         }
 583
 584                         do_irg_opt(irg, "parallelize-mem");
 585                 }
 586                 timer_stop(t_all_opt);
 587
 588                 do_irp_opt("remove-unused");
 589
 590                 dump_all("low-opt");
 591         }
 592
 593         if (firm_opt.cc_opt)
 594                 mark_private_methods();
 595
 596         if (firm_dump.statistic & STAT_FINAL) {
 597                 stat_dump_snapshot(input_filename, "final");
 598         }
 599 }
 600
 601 /**
 602  * Initialize for the Firm-generating back end.
 603  */
 604 void gen_firm_init(void)
 605 {
 606         unsigned pattern = 0;
 607         int      i;
 608
 609         for (i = 0; i < n_opts; ++i) {
 610                 timers[i] = ir_timer_new();
 611                 timer_register(timers[i], opts[i].description);
 612         }
 613         t_verify = ir_timer_new();
 614         timer_register(t_verify, "Firm: verify pass");
 615         t_vcg_dump = ir_timer_new();
 616         timer_register(t_vcg_dump, "Firm: vcg dumping");
 617         t_all_opt = ir_timer_new();
 618         timer_register(t_all_opt, "Firm: all optimizations");
 619
 620         if (firm_dump.stat_pattern)
 621                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 622
 623         if (firm_dump.stat_dag)
 624                 pattern |= FIRMSTAT_COUNT_DAG;
 625
 626         ir_init(NULL);
 627         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 628                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 629                         | FIRMSTAT_COUNT_CONSTS | pattern);
 630
 631         edges_init_dbg(firm_opt.verify_edges);
 632
 633         /* Sel node cannot produce NULL pointers */
 634         set_opt_sel_based_null_check_elim(1);
 635
 636         /* dynamic dispatch works currently only if whole world scenarios */
 637         set_opt_dyn_meth_dispatch(0);
 638
 639         /* do not run architecture dependent optimizations in building phase */
 640         arch_dep_set_opts(arch_dep_none);
 641
 642         do_node_verification((firm_verification_t) firm_opt.verify);
 643         if (firm_dump.filter != NULL)
 644                 ir_set_dump_filter(firm_dump.filter);
 645         if (firm_dump.extbb)
 646                 ir_add_dump_flags(ir_dump_flag_group_extbb);
 647         if (firm_dump.no_blocks)
 648                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
 649
 650         if (firm_opt.enabled) {
 651                 set_optimize(1);
 652                 set_opt_constant_folding(firm_opt.const_folding);
 653                 set_opt_algebraic_simplification(firm_opt.const_folding);
 654                 set_opt_cse(firm_opt.cse);
 655                 set_opt_global_cse(0);
 656                 set_opt_unreachable_code(1);
 657         } else {
 658                 set_optimize(0);
 659         }
 660 }
 661
 662 /**
 663  * Called, after the Firm generation is completed,
 664  * do all optimizations and backend call here.
 665  *
 666  * @param out                a file handle for the output, may be NULL
 667  * @param input_filename     the name of the (main) source file
 668  * @param c_mode             non-zero if "C" was compiled
 669  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
 670  */
 671 void gen_firm_finish(FILE *out, const char *input_filename,
 672                      int new_firm_const_exists)
 673 {
 674         int i;
 675
 676 #if 0
 677         if (firm_opt.enable_statev) {
 678                 char buf[1024];
 679                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
 680                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
 681                 ir_stat_ev_compilation_unit(input_filename);
 682         }
 683 #endif
 684
 685         firm_const_exists = new_firm_const_exists;
 686
 687         /* the general for dumping option must be set, or the others will not work*/
 688         firm_dump.ir_graph
 689                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 690
 691         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 692                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 693         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 694
 695         /* FIXME: cloning might ADD new graphs. */
 696         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 697
 698         if (firm_dump.all_types) {
 699                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 700         }
 701
 702         /* finalize all graphs */
 703         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 704                 ir_graph *irg = get_irp_irg(i);
 705                 irg_finalize_cons(irg);
 706         }
 707         dump_all("");
 708
 709         timer_push(t_verify);
 710         tr_verify();
 711         timer_pop(t_verify);
 712
 713         /* all graphs are finalized, set the irp phase to high */
 714         set_irp_phase_state(phase_high);
 715
 716         /* BEWARE: kill unreachable code before doing compound lowering */
 717         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 718                 ir_graph *irg = get_irp_irg(i);
 719                 do_irg_opt(irg, "control-flow");
 720         }
 721
 722         /* lower copyb nodes */
 723         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 724                 ir_graph *irg = get_irp_irg(i);
 725                 lower_CopyB(irg, 128, 4);
 726         }
 727
 728         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 729                 stat_dump_snapshot(input_filename, "noopt");
 730         }
 731
 732         if (firm_opt.enabled)
 733                 do_firm_optimizations(input_filename);
 734
 735         if (firm_opt.lower)
 736                 do_firm_lowering(input_filename);
 737
 738         /* set the phase to low */
 739         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 740                 set_irg_phase_state(get_irp_irg(i), phase_low);
 741
 742         if (firm_dump.statistic & STAT_FINAL_IR)
 743                 stat_dump_snapshot(input_filename, "final-ir");
 744
 745         /* run the code generator */
 746         if (firm_be_opt.selection != BE_NONE)
 747                 do_codegen(out, input_filename);
 748
 749         if (firm_dump.statistic & STAT_FINAL)
 750                 stat_dump_snapshot(input_filename, "final");
 751 }
 752
 753 void disable_all_opts(void)
 754 {
 755         for (int i = 0; i < n_opts; ++i) {
 756                 opt_config_t *config = &opts[i];
 757                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 758                         config->flags |= OPT_FLAG_ENABLED;
 759                 } else {
 760                         config->flags &= ~OPT_FLAG_ENABLED;
 761                 }
 762         }
 763 }
 764
 765 int firm_opt_option(const char *opt)
 766 {
 767         bool enable = true;
 768         if (strncmp(opt, "no-", 3) == 0) {
 769                 enable = false;
 770                 opt = opt + 3;
 771         }
 772
 773         opt_config_t *config = get_opt(opt);
 774         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 775                 return 0;
 776
 777         config->flags &= ~OPT_FLAG_ENABLED;
 778         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 779         return 1;
 780 }
 781
 782 void firm_opt_option_help(void)
 783 {
 784         int i;
 785
 786         for (i = 0; i < n_opts; ++i) {
 787                 char buf[1024];
 788                 char buf2[1024];
 789
 790                 const opt_config_t *config = &opts[i];
 791                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 792                         continue;
 793
 794                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 795                 print_option_help(config->name, buf2);
 796                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 797                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 798                 print_option_help(buf, buf2);
 799         }
 800 }
 801
 802 /**
 803  * Do very early initializations
 804  */
 805 void firm_early_init(void)
 806 {
 807         /* arg: need this here for command line options */
 808         be_opt_register();
 809
 810         enable_safe_defaults();
 811 }