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_lower_for_target(void)
 203 {
 204         int i;
 205         const backend_params *be_params = be_get_backend_param();
 206         be_params->lower_for_target();
 207
 208         /* set the phase to low */
 209         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 210                 ir_graph *irg = get_irp_irg(i);
 211                 set_irg_phase_state(irg, phase_low);
 212         }
 213         set_irp_phase_state(phase_low);
 214 }
 215
 216 static void do_vrp(ir_graph *irg)
 217 {
 218         set_vrp_data(irg);
 219 }
 220
 221 typedef enum opt_target {
 222         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 223         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 224 } opt_target_t;
 225
 226 typedef enum opt_flags {
 227         OPT_FLAG_NONE         = 0,
 228         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 229         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 230         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 231         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 232                                              -foptions for this transformation */
 233         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 234                                              so we need it even with -O0 */
 235 } opt_flags_t;
 236
 237 typedef void (*transform_irg_func)(ir_graph *irg);
 238 typedef void (*transform_irp_func)(void);
 239
 240 typedef struct {
 241         opt_target_t  target;
 242         const char   *name;
 243         union {
 244                 transform_irg_func transform_irg;
 245                 transform_irp_func transform_irp;
 246         } u;
 247         const char   *description;
 248         opt_flags_t   flags;
 249 } opt_config_t;
 250
 251 static opt_config_t opts[] = {
 252 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 253 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 254         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 255         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 256         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
 257         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 258         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 259         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 260         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 261         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 262         IRG("gcse",              do_gcse,                  "global common subexpression elimination",               OPT_FLAG_NONE),
 263         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 264         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 265         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 266         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 267         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 268         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 269         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 270         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 271         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 272         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 273         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 274         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 275         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 276         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 277         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 278         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 279         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 280         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 281         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
 282         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 283         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),
 284         IRP("target-lowering",   do_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 285         IRP("opt-func-call",     do_optimize_funccalls,    "function call optimization",                            OPT_FLAG_NONE),
 286         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 287         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 288         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
 289 #undef IRP
 290 #undef IRG
 291 };
 292 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 293 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 294
 295 static opt_config_t *get_opt(const char *name)
 296 {
 297         int i;
 298         for (i = 0; i < n_opts; ++i) {
 299                 opt_config_t *config = &opts[i];
 300                 if (strcmp(config->name, name) == 0)
 301                         return config;
 302         }
 303
 304         return NULL;
 305 }
 306
 307 static void set_opt_enabled(const char *name, bool enabled)
 308 {
 309         opt_config_t *config = get_opt(name);
 310         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 311                 | (enabled ? OPT_FLAG_ENABLED : 0);
 312 }
 313
 314 static bool get_opt_enabled(const char *name)
 315 {
 316         opt_config_t *config = get_opt(name);
 317         return (config->flags & OPT_FLAG_ENABLED) != 0;
 318 }
 319
 320 /**
 321  * perform an optimisation on a single graph
 322  *
 323  * @return  true if something changed, false otherwise
 324  */
 325 static bool do_irg_opt(ir_graph *irg, const char *name)
 326 {
 327         ir_graph     *old_irg;
 328         opt_config_t *config = get_opt(name);
 329         size_t        n      = config - opts;
 330         assert(config != NULL);
 331         assert(config->target == OPT_TARGET_IRG);
 332         if (! (config->flags & OPT_FLAG_ENABLED))
 333                 return false;
 334
 335         old_irg          = current_ir_graph;
 336         current_ir_graph = irg;
 337
 338         timer_push(timers[n]);
 339         config->u.transform_irg(irg);
 340         timer_pop(timers[n]);
 341
 342         if (firm_dump.all_phases && firm_dump.ir_graph) {
 343                 dump_ir_graph(irg, name);
 344         }
 345
 346         if (firm_opt.check_all) {
 347                 timer_push(t_verify);
 348                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 349                 timer_pop(t_verify);
 350         }
 351
 352         current_ir_graph = old_irg;
 353         return true;
 354 }
 355
 356 static void do_irp_opt(const char *name)
 357 {
 358         opt_config_t *config = get_opt(name);
 359         size_t        n      = config - opts;
 360         assert(config->target == OPT_TARGET_IRP);
 361         if (! (config->flags & OPT_FLAG_ENABLED))
 362                 return;
 363
 364         timer_push(timers[n]);
 365         config->u.transform_irp();
 366         timer_pop(timers[n]);
 367
 368         if (firm_dump.ir_graph && firm_dump.all_phases) {
 369                 int i;
 370                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 371                         ir_graph *irg = get_irp_irg(i);
 372                         dump_ir_graph(irg, name);
 373                 }
 374         }
 375
 376         if (firm_opt.check_all) {
 377                 int i;
 378                 timer_push(t_verify);
 379                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 380                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 381                 }
 382                 timer_pop(t_verify);
 383         }
 384 }
 385
 386 /**
 387  * Enable transformations which should be always safe (and cheap) to perform
 388  */
 389 static void enable_safe_defaults(void)
 390 {
 391         set_opt_enabled("remove-unused", true);
 392         set_opt_enabled("opt-tail-rec", true);
 393         set_opt_enabled("opt-func-call", true);
 394         set_opt_enabled("reassociation", true);
 395         set_opt_enabled("control-flow", true);
 396         set_opt_enabled("local", true);
 397         set_opt_enabled("lower-const", true);
 398         set_opt_enabled("scalar-replace", true);
 399         set_opt_enabled("place", true);
 400         set_opt_enabled("confirm", true);
 401         set_opt_enabled("opt-load-store", true);
 402         set_opt_enabled("lower", true);
 403         set_opt_enabled("deconv", true);
 404         set_opt_enabled("remove-confirms", true);
 405         set_opt_enabled("ivopts", true);
 406         set_opt_enabled("dead", true);
 407         set_opt_enabled("remove-phi-cycles", true);
 408         set_opt_enabled("frame", true);
 409         set_opt_enabled("combo", true);
 410         set_opt_enabled("invert-loops", true);
 411         set_opt_enabled("target-lowering", true);
 412         set_opt_enabled("rts", true);
 413 }
 414
 415 /**
 416  * run all the Firm optimizations
 417  *
 418  * @param input_filename     the name of the (main) source file
 419  */
 420 static void do_firm_optimizations(const char *input_filename)
 421 {
 422         int      i;
 423         unsigned aa_opt;
 424
 425         set_opt_alias_analysis(firm_opt.alias_analysis);
 426
 427         aa_opt = aa_opt_no_opt;
 428         if (firm_opt.strict_alias)
 429                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 430         if (firm_opt.no_alias)
 431                 aa_opt = aa_opt_no_alias;
 432
 433         set_irp_memory_disambiguator_options(aa_opt);
 434
 435         /* parameter passing code should set them directly sometime... */
 436         set_opt_enabled("gcse", firm_opt.gcse);
 437         set_opt_enabled("place", !firm_opt.gcse);
 438         set_opt_enabled("confirm", firm_opt.confirm);
 439         set_opt_enabled("remove-confirms", firm_opt.confirm);
 440
 441         /* osr supersedes remove_phi_cycles */
 442         if (get_opt_enabled("ivopts"))
 443                 set_opt_enabled("remove-phi-cycles", false);
 444
 445         timer_start(t_all_opt);
 446
 447         do_irp_opt("rts");
 448
 449         /* first step: kill dead code */
 450         for (i = 0; i < get_irp_n_irgs(); i++) {
 451                 ir_graph *irg = get_irp_irg(i);
 452                 do_irg_opt(irg, "combo");
 453                 do_irg_opt(irg, "local");
 454                 do_irg_opt(irg, "control-flow");
 455         }
 456
 457         do_irp_opt("remove-unused");
 458         for (i = 0; i < get_irp_n_irgs(); ++i) {
 459                 ir_graph *irg = get_irp_irg(i);
 460                 do_irg_opt(irg, "opt-tail-rec");
 461         }
 462         do_irp_opt("opt-func-call");
 463         do_irp_opt("lower-const");
 464
 465         for (i = 0; i < get_irp_n_irgs(); i++) {
 466                 ir_graph *irg = get_irp_irg(i);
 467
 468                 do_irg_opt(irg, "scalar-replace");
 469                 do_irg_opt(irg, "invert-loops");
 470                 do_irg_opt(irg, "unroll-loops");
 471                 do_irg_opt(irg, "local");
 472                 do_irg_opt(irg, "reassociation");
 473                 do_irg_opt(irg, "local");
 474                 do_irg_opt(irg, "gcse");
 475
 476                 if (firm_opt.confirm) {
 477                         /* Confirm construction currently can only handle blocks with only
 478                            one control flow predecessor. Calling optimize_cf here removes
 479                            Bad predecessors and help the optimization of switch constructs.
 480                          */
 481                         do_irg_opt(irg, "control-flow");
 482                         do_irg_opt(irg, "confirm");
 483                         do_irg_opt(irg, "vrp");
 484                         do_irg_opt(irg, "local");
 485                 }
 486
 487                 do_irg_opt(irg, "control-flow");
 488                 do_irg_opt(irg, "opt-load-store");
 489                 do_irg_opt(irg, "fp-vrp");
 490                 do_irg_opt(irg, "lower");
 491                 do_irg_opt(irg, "deconv");
 492                 do_irg_opt(irg, "thread-jumps");
 493                 do_irg_opt(irg, "remove-confirms");
 494                 do_irg_opt(irg, "gvn-pre");
 495                 do_irg_opt(irg, "place");
 496                 do_irg_opt(irg, "control-flow");
 497
 498                 if (do_irg_opt(irg, "if-conversion")) {
 499                         do_irg_opt(irg, "local");
 500                         do_irg_opt(irg, "control-flow");
 501                 }
 502                 /* this doesn't make too much sense but tests the mux destruction... */
 503                 do_irg_opt(irg, "lower-mux");
 504
 505                 do_irg_opt(irg, "bool");
 506                 do_irg_opt(irg, "shape-blocks");
 507                 do_irg_opt(irg, "ivopts");
 508                 do_irg_opt(irg, "local");
 509                 do_irg_opt(irg, "dead");
 510                 do_irg_opt(irg, "frame");
 511         }
 512
 513         do_irp_opt("inline");
 514         do_irp_opt("opt-proc-clone");
 515
 516         for (i = 0; i < get_irp_n_irgs(); i++) {
 517                 ir_graph *irg = get_irp_irg(i);
 518                 do_irg_opt(irg, "local");
 519                 do_irg_opt(irg, "control-flow");
 520                 do_irg_opt(irg, "thread-jumps");
 521                 do_irg_opt(irg, "local");
 522                 do_irg_opt(irg, "control-flow");
 523
 524                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 525                         do_irg_opt(irg, "local");
 526                         do_irg_opt(irg, "vrp");
 527                         do_irg_opt(irg, "local");
 528                         do_irg_opt(irg, "vrp");
 529                 }
 530         }
 531
 532         if (firm_dump.ir_graph) {
 533                 /* recompute backedges for nicer dumps */
 534                 for (i = 0; i < get_irp_n_irgs(); i++)
 535                         construct_cf_backedges(get_irp_irg(i));
 536         }
 537
 538         dump_all("opt");
 539
 540         if (firm_dump.statistic & STAT_AFTER_OPT)
 541                 stat_dump_snapshot(input_filename, "opt");
 542
 543         timer_stop(t_all_opt);
 544 }
 545
 546 /**
 547  * do Firm lowering
 548  *
 549  * @param input_filename  the name of the (main) source file
 550  */
 551 static void do_firm_lowering(const char *input_filename)
 552 {
 553         int i;
 554
 555         do_irp_opt("target-lowering");
 556
 557         if (firm_dump.statistic & STAT_AFTER_LOWER)
 558                 stat_dump_snapshot(input_filename, "low");
 559
 560         if (firm_opt.enabled) {
 561                 timer_start(t_all_opt);
 562
 563                 /* run reassociation first on all graphs BEFORE the architecture
 564                    dependent optimizations are enabled */
 565                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 566                         ir_graph *irg = get_irp_irg(i);
 567                         do_irg_opt(irg, "reassociation");
 568                 }
 569
 570                 /* enable architecture dependent optimizations */
 571                 arch_dep_set_opts((arch_dep_opts_t)
 572                                 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 573                                  (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 574                                  (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 575
 576                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 577                         ir_graph *irg = get_irp_irg(i);
 578
 579                         do_irg_opt(irg, "local");
 580                         do_irg_opt(irg, "gcse");
 581                         do_irg_opt(irg, "control-flow");
 582                         do_irg_opt(irg, "opt-load-store");
 583                         do_irg_opt(irg, "local");
 584                         do_irg_opt(irg, "control-flow");
 585
 586                         if (do_irg_opt(irg, "vrp")) {
 587                                 do_irg_opt(irg, "local");
 588                                 do_irg_opt(irg, "control-flow");
 589                                 do_irg_opt(irg, "vrp");
 590                                 do_irg_opt(irg, "local");
 591                                 do_irg_opt(irg, "control-flow");
 592                         }
 593
 594                         if (do_irg_opt(irg, "if-conversion")) {
 595                                 do_irg_opt(irg, "local");
 596                                 do_irg_opt(irg, "control-flow");
 597                         }
 598
 599                         do_irg_opt(irg, "parallelize-mem");
 600                 }
 601                 timer_stop(t_all_opt);
 602
 603                 do_irp_opt("remove-unused");
 604
 605                 dump_all("low-opt");
 606         }
 607
 608         if (firm_opt.cc_opt)
 609                 mark_private_methods();
 610
 611         if (firm_dump.statistic & STAT_FINAL) {
 612                 stat_dump_snapshot(input_filename, "final");
 613         }
 614 }
 615
 616 /**
 617  * Initialize for the Firm-generating back end.
 618  */
 619 void gen_firm_init(void)
 620 {
 621         unsigned pattern = 0;
 622         int      i;
 623
 624         for (i = 0; i < n_opts; ++i) {
 625                 timers[i] = ir_timer_new();
 626                 timer_register(timers[i], opts[i].description);
 627         }
 628         t_verify = ir_timer_new();
 629         timer_register(t_verify, "Firm: verify pass");
 630         t_vcg_dump = ir_timer_new();
 631         timer_register(t_vcg_dump, "Firm: vcg dumping");
 632         t_all_opt = ir_timer_new();
 633         timer_register(t_all_opt, "Firm: all optimizations");
 634
 635         if (firm_dump.stat_pattern)
 636                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 637
 638         if (firm_dump.stat_dag)
 639                 pattern |= FIRMSTAT_COUNT_DAG;
 640
 641         ir_init(NULL);
 642         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 643                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 644                         | FIRMSTAT_COUNT_CONSTS | pattern);
 645
 646         edges_init_dbg(firm_opt.verify_edges);
 647
 648         /* Sel node cannot produce NULL pointers */
 649         set_opt_sel_based_null_check_elim(1);
 650
 651         /* dynamic dispatch works currently only if whole world scenarios */
 652         set_opt_dyn_meth_dispatch(0);
 653
 654         /* do not run architecture dependent optimizations in building phase */
 655         arch_dep_set_opts(arch_dep_none);
 656
 657         do_node_verification((firm_verification_t) firm_opt.verify);
 658         if (firm_dump.filter != NULL)
 659                 ir_set_dump_filter(firm_dump.filter);
 660         if (firm_dump.extbb)
 661                 ir_add_dump_flags(ir_dump_flag_group_extbb);
 662         if (firm_dump.no_blocks)
 663                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
 664
 665         if (firm_opt.enabled) {
 666                 set_optimize(1);
 667                 set_opt_constant_folding(firm_opt.const_folding);
 668                 set_opt_algebraic_simplification(firm_opt.const_folding);
 669                 set_opt_cse(firm_opt.cse);
 670                 set_opt_global_cse(0);
 671                 set_opt_unreachable_code(1);
 672         } else {
 673                 set_optimize(0);
 674         }
 675 }
 676
 677 /**
 678  * Called, after the Firm generation is completed,
 679  * do all optimizations and backend call here.
 680  *
 681  * @param out                a file handle for the output, may be NULL
 682  * @param input_filename     the name of the (main) source file
 683  * @param c_mode             non-zero if "C" was compiled
 684  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
 685  */
 686 void gen_firm_finish(FILE *out, const char *input_filename,
 687                      int new_firm_const_exists)
 688 {
 689         int i;
 690
 691 #if 0
 692         if (firm_opt.enable_statev) {
 693                 char buf[1024];
 694                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
 695                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
 696                 ir_stat_ev_compilation_unit(input_filename);
 697         }
 698 #endif
 699
 700         firm_const_exists = new_firm_const_exists;
 701
 702         /* the general for dumping option must be set, or the others will not work*/
 703         firm_dump.ir_graph
 704                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 705
 706         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 707                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 708         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 709
 710         /* FIXME: cloning might ADD new graphs. */
 711         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 712
 713         if (firm_dump.all_types) {
 714                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 715         }
 716
 717         /* finalize all graphs */
 718         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 719                 ir_graph *irg = get_irp_irg(i);
 720                 irg_finalize_cons(irg);
 721         }
 722         dump_all("");
 723
 724         timer_push(t_verify);
 725         tr_verify();
 726         timer_pop(t_verify);
 727
 728         /* all graphs are finalized, set the irp phase to high */
 729         set_irp_phase_state(phase_high);
 730
 731         /* BEWARE: kill unreachable code before doing compound lowering */
 732         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 733                 ir_graph *irg = get_irp_irg(i);
 734                 do_irg_opt(irg, "control-flow");
 735         }
 736
 737         /* lower all compound call return values */
 738         lower_compound_params();
 739
 740         /* lower copyb nodes */
 741         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 742                 ir_graph *irg = get_irp_irg(i);
 743                 lower_CopyB(irg, 128, 4);
 744         }
 745
 746         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 747                 stat_dump_snapshot(input_filename, "noopt");
 748         }
 749
 750         if (firm_opt.enabled)
 751                 do_firm_optimizations(input_filename);
 752
 753         if (firm_opt.lower)
 754                 do_firm_lowering(input_filename);
 755
 756         /* set the phase to low */
 757         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 758                 set_irg_phase_state(get_irp_irg(i), phase_low);
 759
 760         if (firm_dump.statistic & STAT_FINAL_IR)
 761                 stat_dump_snapshot(input_filename, "final-ir");
 762
 763         /* run the code generator */
 764         if (firm_be_opt.selection != BE_NONE)
 765                 do_codegen(out, input_filename);
 766
 767         if (firm_dump.statistic & STAT_FINAL)
 768                 stat_dump_snapshot(input_filename, "final");
 769 }
 770
 771 void disable_all_opts(void)
 772 {
 773         for (int i = 0; i < n_opts; ++i) {
 774                 opt_config_t *config = &opts[i];
 775                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 776                         config->flags |= OPT_FLAG_ENABLED;
 777                 } else {
 778                         config->flags &= ~OPT_FLAG_ENABLED;
 779                 }
 780         }
 781 }
 782
 783 int firm_opt_option(const char *opt)
 784 {
 785         bool enable = true;
 786         if (strncmp(opt, "no-", 3) == 0) {
 787                 enable = false;
 788                 opt = opt + 3;
 789         }
 790
 791         opt_config_t *config = get_opt(opt);
 792         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 793                 return 0;
 794
 795         config->flags &= ~OPT_FLAG_ENABLED;
 796         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 797         return 1;
 798 }
 799
 800 void firm_opt_option_help(void)
 801 {
 802         int i;
 803
 804         for (i = 0; i < n_opts; ++i) {
 805                 char buf[1024];
 806                 char buf2[1024];
 807
 808                 const opt_config_t *config = &opts[i];
 809                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 810                         continue;
 811
 812                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 813                 print_option_help(config->name, buf2);
 814                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 815                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 816                 print_option_help(buf, buf2);
 817         }
 818 }
 819
 820 /**
 821  * Do very early initializations
 822  */
 823 void firm_early_init(void)
 824 {
 825         /* arg: need this here for command line options */
 826         be_opt_register();
 827
 828         enable_safe_defaults();
 829 }