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