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 #if defined(_DEBUG) || defined(FIRM_DEBUG)
  26 #define DBG(x)  dbg_printf x
  27 #else
  28 #define DBG(x) ((void)0)
  29 #endif /* _DEBUG || FIRM_DEBUG */
  30
  31 static ir_timer_t *t_vcg_dump;
  32 static ir_timer_t *t_verify;
  33 static ir_timer_t *t_all_opt;
  34 static bool do_irg_opt(ir_graph *irg, const char *name);
  35
  36 /** dump all the graphs depending on cond */
  37
  38 static void dump_all(const char *suffix) {
  39         if (firm_dump.ir_graph) {
  40                 timer_push(t_vcg_dump);
  41                 if (firm_dump.no_blocks)
  42                         dump_all_ir_graphs(dump_ir_graph, suffix);
  43                 else if (firm_dump.extbb)
  44                         dump_all_ir_graphs(dump_ir_extblock_graph, suffix);
  45                 else
  46                         dump_all_ir_graphs(dump_ir_block_graph, suffix);
  47                 timer_pop(t_vcg_dump);
  48         }
  49 }
  50
  51 /* set by the backend parameters */
  52 static const ir_settings_arch_dep_t *ad_param              = NULL;
  53 static create_intrinsic_fkt         *arch_create_intrinsic = NULL;
  54 static void                         *create_intrinsic_ctx  = NULL;
  55 static const ir_settings_if_conv_t  *if_conv_info          = NULL;
  56
  57 /* entities of runtime functions */
  58 ir_entity_ptr rts_entities[rts_max];
  59
  60 /**
  61  * factory for setting architecture dependent parameters
  62  */
  63 static const ir_settings_arch_dep_t *arch_factory(void)
  64 {
  65         static const ir_settings_arch_dep_t param = {
  66                 1,   /* also use subs */
  67                 4,   /* maximum shifts */
  68                 31,   /* maximum shift amount */
  69                 NULL, /* use default evaluator */
  70
  71                 1, /* allow Mulhs */
  72                 1, /* allow Mulus */
  73                 32  /* Mulh allowed up to 32 bit */
  74         };
  75
  76         return ad_param ? ad_param : &param;
  77 }
  78
  79 /**
  80  * Map runtime functions.
  81  */
  82 static void rts_map(void) {
  83         static const struct {
  84                 ir_entity_ptr *ent; /**< address of the rts entity */
  85                 i_mapper_func func; /**< mapper function. */
  86         } mapper[] = {
  87                 /* integer */
  88                 { &rts_entities[rts_abs],     i_mapper_abs },
  89                 { &rts_entities[rts_labs],    i_mapper_abs },
  90                 { &rts_entities[rts_llabs],   i_mapper_abs },
  91                 { &rts_entities[rts_imaxabs], i_mapper_abs },
  92
  93                 /* double -> double */
  94                 { &rts_entities[rts_fabs],    i_mapper_abs },
  95                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
  96                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
  97                 { &rts_entities[rts_pow],     i_mapper_pow },
  98                 { &rts_entities[rts_exp],     i_mapper_exp },
  99                 { &rts_entities[rts_exp2],    i_mapper_exp },
 100                 { &rts_entities[rts_exp10],   i_mapper_exp },
 101                 { &rts_entities[rts_log],     i_mapper_log },
 102                 { &rts_entities[rts_log2],    i_mapper_log2 },
 103                 { &rts_entities[rts_log10],   i_mapper_log10 },
 104                 { &rts_entities[rts_sin],     i_mapper_sin },
 105                 { &rts_entities[rts_cos],     i_mapper_cos },
 106                 { &rts_entities[rts_tan],     i_mapper_tan },
 107                 { &rts_entities[rts_asin],    i_mapper_asin },
 108                 { &rts_entities[rts_acos],    i_mapper_acos },
 109                 { &rts_entities[rts_atan],    i_mapper_atan },
 110                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 111                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 112                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 113
 114                 /* float -> float */
 115                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 116                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 117                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 118                 { &rts_entities[rts_powf],    i_mapper_pow },
 119                 { &rts_entities[rts_expf],    i_mapper_exp },
 120                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 121                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 122                 { &rts_entities[rts_logf],    i_mapper_log },
 123                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 124                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 125                 { &rts_entities[rts_sinf],    i_mapper_sin },
 126                 { &rts_entities[rts_cosf],    i_mapper_cos },
 127                 { &rts_entities[rts_tanf],    i_mapper_tan },
 128                 { &rts_entities[rts_asinf],   i_mapper_asin },
 129                 { &rts_entities[rts_acosf],   i_mapper_acos },
 130                 { &rts_entities[rts_atanf],   i_mapper_atan },
 131                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 132                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 133                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 134
 135                 /* long double -> long double */
 136                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 137                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 138                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 139                 { &rts_entities[rts_powl],    i_mapper_pow },
 140                 { &rts_entities[rts_expl],    i_mapper_exp },
 141                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 142                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 143                 { &rts_entities[rts_logl],    i_mapper_log },
 144                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 145                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 146                 { &rts_entities[rts_sinl],    i_mapper_sin },
 147                 { &rts_entities[rts_cosl],    i_mapper_cos },
 148                 { &rts_entities[rts_tanl],    i_mapper_tan },
 149                 { &rts_entities[rts_asinl],   i_mapper_asin },
 150                 { &rts_entities[rts_acosl],   i_mapper_acos },
 151                 { &rts_entities[rts_atanl],   i_mapper_atan },
 152                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 153                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 154                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 155
 156                 /* string */
 157                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 158                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 159                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 160                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 161                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 162                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 163                 { &rts_entities[rts_memmove], i_mapper_memmove },
 164                 { &rts_entities[rts_memset],  i_mapper_memset },
 165                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 166         };
 167         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
 168         unsigned i, n_map;
 169
 170         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
 171                 if (*mapper[i].ent != NULL) {
 172                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 173                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 174                         rec[n_map].i_call.i_mapper = mapper[i].func;
 175                         rec[n_map].i_call.ctx      = NULL;
 176                         rec[n_map].i_call.link     = NULL;
 177                         ++n_map;
 178                 }
 179         }
 180
 181         if (n_map > 0)
 182                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 183 }
 184
 185 static int *irg_dump_no;
 186
 187 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
 188 {
 189         char name[64];
 190         snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++,
 191                  suffix);
 192
 193         timer_push(t_vcg_dump);
 194         if (firm_dump.no_blocks)
 195                 dump_ir_graph(irg, name);
 196         else if (firm_dump.extbb)
 197                 dump_ir_extblock_graph(irg, name);
 198         else
 199                 dump_ir_block_graph(irg, name);
 200         timer_pop(t_vcg_dump);
 201 }
 202
 203 static int firm_const_exists;
 204
 205 static void do_optimize_funccalls(void)
 206 {
 207         optimize_funccalls(firm_const_exists, NULL);
 208 }
 209
 210 static void do_gcse(ir_graph *irg)
 211 {
 212         set_opt_global_cse(1);
 213         optimize_graph_df(irg);
 214         place_code(irg);
 215         set_opt_global_cse(0);
 216 }
 217
 218 static void do_lower_highlevel(ir_graph *irg)
 219 {
 220         lower_highlevel_graph(irg, firm_opt.lower_bitfields);
 221 }
 222
 223 static void do_if_conv(ir_graph *irg)
 224 {
 225         opt_if_conv(irg, if_conv_info);
 226 }
 227
 228 static void do_stred(ir_graph *irg)
 229 {
 230         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 231 }
 232
 233 static void after_inline_opt(ir_graph *irg)
 234 {
 235         do_irg_opt(irg, "scalar-replace");
 236         do_irg_opt(irg, "local");
 237         do_irg_opt(irg, "control-flow");
 238         do_irg_opt(irg, "combo");
 239 }
 240
 241 static void do_inline(void)
 242 {
 243         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 244                          after_inline_opt);
 245 }
 246
 247 static void do_cloning(void)
 248 {
 249         proc_cloning((float) firm_opt.clone_threshold);
 250 }
 251
 252 static void do_lower_switch(ir_graph *irg)
 253 {
 254         lower_switch(irg, firm_opt.spare_size);
 255 }
 256
 257 static void do_lower_mux(ir_graph *irg)
 258 {
 259         lower_mux(irg, NULL);
 260 }
 261
 262 static void do_lower_dw_ops(void)
 263 {
 264         lwrdw_param_t init = {
 265                 1,
 266                 1,
 267                 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
 268                 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
 269                 get_atomic_mode(ATOMIC_TYPE_INT),
 270                 get_atomic_mode(ATOMIC_TYPE_UINT),
 271                 def_create_intrinsic_fkt,
 272                 NULL
 273         };
 274
 275         if (arch_create_intrinsic) {
 276                 init.create_intrinsic = arch_create_intrinsic;
 277                 init.ctx              = create_intrinsic_ctx;
 278         }
 279         lower_dw_ops(&init);
 280 }
 281
 282 static void do_vrp(ir_graph *irg)
 283 {
 284         set_vrp_data(irg);
 285 }
 286
 287 typedef enum opt_target {
 288         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 289         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 290 } opt_target_t;
 291
 292 typedef enum opt_flags {
 293         OPT_FLAG_NONE         = 0,
 294         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 295         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 296         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 297         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 298                                              -foptions for this transformation */
 299         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 300                                              so we need it even with -O0 */
 301 } opt_flags_t;
 302
 303 typedef void (*transform_irg_func)(ir_graph *irg);
 304 typedef void (*transform_irp_func)(void);
 305
 306 typedef struct {
 307         opt_target_t  target;
 308         const char   *name;
 309         union {
 310                 transform_irg_func transform_irg;
 311                 transform_irp_func transform_irp;
 312         } u;
 313         const char   *description;
 314         opt_flags_t   flags;
 315 } opt_config_t;
 316
 317 static opt_config_t opts[] = {
 318 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 319 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 320         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 321         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 322         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
 323         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 324         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 325         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 326         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 327         IRG("gcse",              do_gcse,                  "global common subexpression elimination",               OPT_FLAG_NONE),
 328         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 329         IRG("if-conversion",     do_if_conv,               "if-conversion",                                         OPT_FLAG_NONE),
 330         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 331         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 332         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 333         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 334         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 335         IRG("lower-switch",      do_lower_switch,          "switch lowering",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 336         IRG("one-return",        normalize_one_return,     "normalisation to 1 return",                             OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 337         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 338         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 339         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 340         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 341         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 342         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 343         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 344         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 345         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 346         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 347         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 348         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
 349         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 350         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),
 351         IRP("lower-dw",          do_lower_dw_ops,          "lowering of doubleword operations",                     OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 352         IRP("opt-func-call",     do_optimize_funccalls,    "function call optimization",                            OPT_FLAG_NONE),
 353         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 354         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 355         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
 356 #undef IRP
 357 #undef IRG
 358 };
 359 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 360 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 361
 362 static opt_config_t *get_opt(const char *name)
 363 {
 364         int i;
 365         for (i = 0; i < n_opts; ++i) {
 366                 opt_config_t *config = &opts[i];
 367                 if (strcmp(config->name, name) == 0)
 368                         return config;
 369         }
 370
 371         return NULL;
 372 }
 373
 374 static void set_opt_enabled(const char *name, bool enabled)
 375 {
 376         opt_config_t *config = get_opt(name);
 377         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 378                 | (enabled ? OPT_FLAG_ENABLED : 0);
 379 }
 380
 381 static bool get_opt_enabled(const char *name)
 382 {
 383         opt_config_t *config = get_opt(name);
 384         return (config->flags & OPT_FLAG_ENABLED) != 0;
 385 }
 386
 387 /**
 388  * perform an optimisation on a single graph
 389  *
 390  * @return  true if something changed, false otherwise
 391  */
 392 static bool do_irg_opt(ir_graph *irg, const char *name)
 393 {
 394         ir_graph     *old_irg;
 395         opt_config_t *config = get_opt(name);
 396         size_t        n      = config - opts;
 397         assert(config != NULL);
 398         assert(config->target == OPT_TARGET_IRG);
 399         if (! (config->flags & OPT_FLAG_ENABLED))
 400                 return false;
 401
 402         old_irg          = current_ir_graph;
 403         current_ir_graph = irg;
 404
 405         timer_push(timers[n]);
 406         config->u.transform_irg(irg);
 407         timer_pop(timers[n]);
 408
 409         if (firm_dump.all_phases && firm_dump.ir_graph) {
 410                 dump_graph_count(irg, name);
 411         }
 412
 413         if (firm_opt.check_all) {
 414                 timer_push(t_verify);
 415                 irg_verify(irg, VRFY_ENFORCE_SSA);
 416                 timer_pop(t_verify);
 417         }
 418
 419         current_ir_graph = old_irg;
 420         return true;
 421 }
 422
 423 static void do_irp_opt(const char *name)
 424 {
 425         opt_config_t *config = get_opt(name);
 426         size_t        n      = config - opts;
 427         assert(config->target == OPT_TARGET_IRP);
 428         if (! (config->flags & OPT_FLAG_ENABLED))
 429                 return;
 430
 431         timer_push(timers[n]);
 432         config->u.transform_irp();
 433         timer_pop(timers[n]);
 434
 435         if (firm_dump.ir_graph && firm_dump.all_phases) {
 436                 int i;
 437                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 438                         ir_graph *irg = get_irp_irg(i);
 439                         dump_graph_count(irg, name);
 440                 }
 441         }
 442
 443         if (firm_opt.check_all) {
 444                 int i;
 445                 timer_push(t_verify);
 446                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 447                         irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
 448                 }
 449                 timer_pop(t_verify);
 450         }
 451 }
 452
 453 /**
 454  * Enable transformations which should be always safe (and cheap) to perform
 455  */
 456 static void enable_safe_defaults(void)
 457 {
 458         set_opt_enabled("remove-unused", true);
 459         set_opt_enabled("opt-tail-rec", true);
 460         set_opt_enabled("opt-func-call", true);
 461         set_opt_enabled("reassociation", true);
 462         set_opt_enabled("control-flow", true);
 463         set_opt_enabled("local", true);
 464         set_opt_enabled("lower-const", true);
 465         set_opt_enabled("scalar-replace", true);
 466         set_opt_enabled("place", true);
 467         set_opt_enabled("confirm", true);
 468         set_opt_enabled("opt-load-store", true);
 469         set_opt_enabled("lower", true);
 470         set_opt_enabled("deconv", true);
 471         set_opt_enabled("remove-confirms", true);
 472         set_opt_enabled("ivopts", true);
 473         set_opt_enabled("dead", true);
 474         set_opt_enabled("lower-switch", true);
 475         set_opt_enabled("remove-phi-cycles", true);
 476         set_opt_enabled("frame", true);
 477         set_opt_enabled("combo", true);
 478 }
 479
 480 /**
 481  * run all the Firm optimizations
 482  *
 483  * @param input_filename     the name of the (main) source file
 484  */
 485 static void do_firm_optimizations(const char *input_filename)
 486 {
 487         int      i;
 488         unsigned aa_opt;
 489
 490         set_opt_alias_analysis(firm_opt.alias_analysis);
 491
 492         aa_opt = aa_opt_no_opt;
 493         if (firm_opt.strict_alias)
 494                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 495         if (firm_opt.no_alias)
 496                 aa_opt = aa_opt_no_alias;
 497
 498         set_irp_memory_disambiguator_options(aa_opt);
 499
 500         /* parameter passing code should set them directly sometime... */
 501         set_opt_enabled("rts", !firm_opt.freestanding);
 502         set_opt_enabled("gcse", firm_opt.gcse);
 503         set_opt_enabled("place", !firm_opt.gcse);
 504         set_opt_enabled("confirm", firm_opt.confirm);
 505         set_opt_enabled("remove-confirms", firm_opt.confirm);
 506
 507         /* osr supersedes remove_phi_cycles */
 508         if (get_opt_enabled("ivopts"))
 509                 set_opt_enabled("remove-phi-cycles", false);
 510
 511         timer_start(t_all_opt);
 512
 513         do_irp_opt("rts");
 514
 515         /* first step: kill dead code */
 516         for (i = 0; i < get_irp_n_irgs(); i++) {
 517                 ir_graph *irg = get_irp_irg(i);
 518                 do_irg_opt(irg, "combo");
 519                 do_irg_opt(irg, "local");
 520                 do_irg_opt(irg, "control-flow");
 521         }
 522
 523         do_irp_opt("remove-unused");
 524         for (i = 0; i < get_irp_n_irgs(); ++i) {
 525                 ir_graph *irg = get_irp_irg(i);
 526                 do_irg_opt(irg, "opt-tail-rec");
 527         }
 528         do_irp_opt("opt-func-call");
 529         do_irp_opt("lower-const");
 530
 531         for (i = 0; i < get_irp_n_irgs(); i++) {
 532                 ir_graph *irg = get_irp_irg(i);
 533
 534                 do_irg_opt(irg, "scalar-replace");
 535                 do_irg_opt(irg, "invert-loops");
 536                 do_irg_opt(irg, "unroll-loops");
 537                 do_irg_opt(irg, "local");
 538                 do_irg_opt(irg, "reassociation");
 539                 do_irg_opt(irg, "local");
 540                 do_irg_opt(irg, "gcse");
 541
 542                 if (firm_opt.confirm) {
 543                         /* Confirm construction currently can only handle blocks with only
 544                            one control flow predecessor. Calling optimize_cf here removes
 545                            Bad predecessors and help the optimization of switch constructs.
 546                          */
 547                         do_irg_opt(irg, "control-flow");
 548                         do_irg_opt(irg, "confirm");
 549                         do_irg_opt(irg, "vrp");
 550                         do_irg_opt(irg, "local");
 551                 }
 552
 553                 do_irg_opt(irg, "control-flow");
 554                 do_irg_opt(irg, "opt-load-store");
 555                 do_irg_opt(irg, "lower");
 556                 do_irg_opt(irg, "deconv");
 557                 do_irg_opt(irg, "thread-jumps");
 558                 do_irg_opt(irg, "remove-confirms");
 559                 do_irg_opt(irg, "gvn-pre");
 560                 do_irg_opt(irg, "place");
 561                 do_irg_opt(irg, "control-flow");
 562
 563                 if (do_irg_opt(irg, "if-conversion")) {
 564                         do_irg_opt(irg, "local");
 565                         do_irg_opt(irg, "control-flow");
 566                 }
 567                 /* this doesn't make too much sense but tests the mux destruction... */
 568                 do_irg_opt(irg, "lower-mux");
 569
 570                 do_irg_opt(irg, "bool");
 571                 do_irg_opt(irg, "shape-blocks");
 572                 do_irg_opt(irg, "lower-switch");
 573                 do_irg_opt(irg, "ivopts");
 574                 do_irg_opt(irg, "local");
 575                 do_irg_opt(irg, "dead");
 576                 do_irg_opt(irg, "frame");
 577         }
 578
 579         do_irp_opt("inline");
 580         do_irp_opt("opt-proc-clone");
 581
 582         for (i = 0; i < get_irp_n_irgs(); i++) {
 583                 ir_graph *irg = get_irp_irg(i);
 584                 do_irg_opt(irg, "local");
 585                 do_irg_opt(irg, "control-flow");
 586                 do_irg_opt(irg, "thread-jumps");
 587                 do_irg_opt(irg, "local");
 588                 do_irg_opt(irg, "control-flow");
 589
 590                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 591                         do_irg_opt(irg, "local");
 592                         do_irg_opt(irg, "vrp");
 593                         do_irg_opt(irg, "local");
 594                         do_irg_opt(irg, "vrp");
 595                 }
 596         }
 597
 598         if (firm_dump.ir_graph) {
 599                 /* recompute backedges for nicer dumps */
 600                 for (i = 0; i < get_irp_n_irgs(); i++)
 601                         construct_cf_backedges(get_irp_irg(i));
 602         }
 603
 604         dump_all("-opt");
 605
 606         if (firm_dump.statistic & STAT_AFTER_OPT)
 607                 stat_dump_snapshot(input_filename, "opt");
 608
 609         timer_stop(t_all_opt);
 610 }
 611
 612 /**
 613  * do Firm lowering
 614  *
 615  * @param input_filename  the name of the (main) source file
 616  */
 617 static void do_firm_lowering(const char *input_filename)
 618 {
 619         int i;
 620
 621         do_irp_opt("lower-dw");
 622
 623         if (firm_dump.statistic & STAT_AFTER_LOWER)
 624                 stat_dump_snapshot(input_filename, "low");
 625
 626         dump_all("-low");
 627
 628         if (firm_opt.enabled) {
 629                 timer_start(t_all_opt);
 630
 631                 /* run reassociation first on all graphs BEFORE the architecture
 632                    dependent optimizations are enabled */
 633                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 634                         ir_graph *irg = get_irp_irg(i);
 635                         do_irg_opt(irg, "reassociation");
 636                 }
 637
 638                 /* enable architecture dependent optimizations */
 639                 arch_dep_set_opts((arch_dep_opts_t)
 640                                 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 641                                  (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 642                                  (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 643
 644                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 645                         ir_graph *irg = get_irp_irg(i);
 646
 647                         do_irg_opt(irg, "local");
 648                         do_irg_opt(irg, "gcse");
 649                         do_irg_opt(irg, "opt-load-store");
 650                         do_irg_opt(irg, "local");
 651                         do_irg_opt(irg, "control-flow");
 652
 653                         if (do_irg_opt(irg, "vrp")) {
 654                                 do_irg_opt(irg, "local");
 655                                 do_irg_opt(irg, "control-flow");
 656                                 do_irg_opt(irg, "vrp");
 657                                 do_irg_opt(irg, "local");
 658                                 do_irg_opt(irg, "control-flow");
 659                         }
 660
 661                         if (do_irg_opt(irg, "if-conversion")) {
 662                                 do_irg_opt(irg, "local");
 663                                 do_irg_opt(irg, "control-flow");
 664                         }
 665
 666                         do_irg_opt(irg, "parallelize-mem");
 667                 }
 668                 timer_stop(t_all_opt);
 669
 670                 do_irp_opt("remove-unused");
 671
 672                 dump_all("-low-opt");
 673         }
 674
 675         if (firm_opt.cc_opt)
 676                 mark_private_methods();
 677
 678         /* set the phase to low */
 679         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 680                 set_irg_phase_low(get_irp_irg(i));
 681
 682         /* all graphs are lowered, set the irp phase to low */
 683         set_irp_phase_state(phase_low);
 684
 685         if (firm_dump.statistic & STAT_FINAL) {
 686                 stat_dump_snapshot(input_filename, "final");
 687         }
 688 }
 689
 690 /**
 691  * Initialize for the Firm-generating back end.
 692  */
 693 void gen_firm_init(void)
 694 {
 695         firm_parameter_t params;
 696         unsigned         pattern = 0;
 697         int              i;
 698
 699         for (i = 0; i < n_opts; ++i) {
 700                 timers[i] = ir_timer_new();
 701                 timer_register(timers[i], opts[i].description);
 702         }
 703         t_verify = ir_timer_new();
 704         timer_register(t_verify, "Firm: verify pass");
 705         t_vcg_dump = ir_timer_new();
 706         timer_register(t_vcg_dump, "Firm: vcg dumping");
 707         t_all_opt = ir_timer_new();
 708         timer_register(t_all_opt, "Firm: all optimizations");
 709
 710         if (firm_dump.stat_pattern)
 711                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 712
 713         if (firm_dump.stat_dag)
 714                 pattern |= FIRMSTAT_COUNT_DAG;
 715
 716         memset(&params, 0, sizeof(params));
 717         params.size                  = sizeof(params);
 718         params.enable_statistics     = firm_dump.statistic == STAT_NONE ? 0 :
 719                 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS
 720                 | pattern;
 721         params.initialize_local_func = uninitialized_local_var;
 722         params.cc_mask               = 0; /* no regparam, cdecl */
 723
 724         ir_init(&params);
 725
 726         if (firm_be_opt.selection == BE_FIRM_BE) {
 727                 const backend_params *be_params = be_get_backend_param();
 728
 729                 if (be_params->do_dw_lowering)
 730                         set_opt_enabled("lower-dw", true);
 731
 732                 arch_create_intrinsic   = be_params->arch_create_intrinsic_fkt;
 733                 create_intrinsic_ctx    = be_params->create_intrinsic_ctx;
 734
 735                 ad_param                = be_params->dep_param;
 736                 if_conv_info            = be_params->if_conv_info;
 737         }
 738
 739         edges_init_dbg(firm_opt.vrfy_edges);
 740
 741         /* Sel node cannot produce NULL pointers */
 742         set_opt_sel_based_null_check_elim(1);
 743
 744         /* dynamic dispatch works currently only if whole world scenarios */
 745         set_opt_dyn_meth_dispatch(0);
 746
 747         arch_dep_init(arch_factory);
 748
 749         /* do not run architecture dependent optimizations in building phase */
 750         arch_dep_set_opts(arch_dep_none);
 751
 752         do_node_verification((firm_verification_t) firm_opt.vrfy);
 753         if (firm_dump.filter)
 754                 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
 755
 756         if (firm_opt.enabled) {
 757                 set_optimize(1);
 758                 set_opt_constant_folding(firm_opt.const_folding);
 759                 set_opt_algebraic_simplification(firm_opt.const_folding);
 760                 set_opt_cse(firm_opt.cse);
 761                 set_opt_global_cse(0);
 762                 set_opt_unreachable_code(1);
 763                 set_opt_control_flow(firm_opt.control_flow);
 764                 set_opt_control_flow_weak_simplification(1);
 765                 set_opt_control_flow_strong_simplification(1);
 766         } else {
 767                 set_optimize(0);
 768         }
 769
 770         /* do not dump entity ld names */
 771         dump_ld_names(0);
 772 }
 773
 774 /**
 775  * Called, after the Firm generation is completed,
 776  * do all optimizations and backend call here.
 777  *
 778  * @param out                a file handle for the output, may be NULL
 779  * @param input_filename     the name of the (main) source file
 780  * @param c_mode             non-zero if "C" was compiled
 781  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
 782  */
 783 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode,
 784                      int new_firm_const_exists)
 785 {
 786         int i;
 787
 788 #if 0
 789         if (firm_opt.enable_statev) {
 790                 char buf[1024];
 791                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
 792                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
 793                 ir_stat_ev_compilation_unit(input_filename);
 794         }
 795 #endif
 796
 797         firm_const_exists = new_firm_const_exists;
 798
 799         /* the general for dumping option must be set, or the others will not work*/
 800         firm_dump.ir_graph
 801                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 802
 803         dump_keepalive_edges(1);
 804         dump_consts_local(1);
 805         dump_dominator_information(1);
 806         dump_loop_information(0);
 807
 808         if (!firm_dump.edge_labels)
 809                 turn_off_edge_labels();
 810
 811         /* FIXME: cloning might ADD new graphs. */
 812         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 813
 814         if (firm_dump.all_types) {
 815                 dump_all_types("");
 816                 if (! c_mode) {
 817                         dump_class_hierarchy(0, "");
 818                         dump_class_hierarchy(1, "-with-entities");
 819                 }
 820         }
 821
 822         /* finalize all graphs */
 823         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 824                 ir_graph *irg = get_irp_irg(i);
 825                 irg_finalize_cons(irg);
 826         }
 827         dump_all("");
 828
 829         timer_push(t_verify);
 830         tr_vrfy();
 831         timer_pop(t_verify);
 832
 833         /* all graphs are finalized, set the irp phase to high */
 834         set_irp_phase_state(phase_high);
 835
 836         /* BEWARE: kill unreachable code before doing compound lowering */
 837         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 838                 ir_graph *irg = get_irp_irg(i);
 839                 do_irg_opt(irg, "control-flow");
 840         }
 841
 842         /* lower all compound call return values */
 843         lower_compound_params();
 844
 845         /* lower copyb nodes */
 846         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 847                 ir_graph *irg = get_irp_irg(i);
 848                 lower_CopyB(irg, 128, 4);
 849         }
 850
 851         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 852                 stat_dump_snapshot(input_filename, "noopt");
 853         }
 854
 855         if (firm_opt.enabled)
 856                 do_firm_optimizations(input_filename);
 857
 858         if (firm_opt.lower)
 859                 do_firm_lowering(input_filename);
 860
 861         /* set the phase to low */
 862         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 863                 set_irg_phase_low(get_irp_irg(i));
 864
 865         if (firm_dump.statistic & STAT_FINAL_IR)
 866                 stat_dump_snapshot(input_filename, "final-ir");
 867
 868         /* run the code generator */
 869         if (firm_be_opt.selection != BE_NONE)
 870                 do_codegen(out, input_filename);
 871
 872         if (firm_dump.statistic & STAT_FINAL)
 873                 stat_dump_snapshot(input_filename, "final");
 874 }
 875
 876 void disable_all_opts(void)
 877 {
 878         for (int i = 0; i < n_opts; ++i) {
 879                 opt_config_t *config = &opts[i];
 880                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 881                         config->flags |= OPT_FLAG_ENABLED;
 882                 } else {
 883                         config->flags &= ~OPT_FLAG_ENABLED;
 884                 }
 885         }
 886 }
 887
 888 int firm_opt_option(const char *opt)
 889 {
 890         bool enable = true;
 891         if (strncmp(opt, "no-", 3) == 0) {
 892                 enable = false;
 893                 opt = opt + 3;
 894         }
 895
 896         opt_config_t *config = get_opt(opt);
 897         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 898                 return 0;
 899
 900         config->flags &= ~OPT_FLAG_ENABLED;
 901         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 902         return 1;
 903 }
 904
 905 void firm_opt_option_help(void)
 906 {
 907         int i;
 908
 909         for (i = 0; i < n_opts; ++i) {
 910                 char buf[1024];
 911                 char buf2[1024];
 912
 913                 const opt_config_t *config = &opts[i];
 914                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 915                         continue;
 916
 917                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 918                 print_option_help(config->name, buf2);
 919                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 920                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 921                 print_option_help(buf, buf2);
 922         }
 923 }
 924
 925 /**
 926  * Do very early initializations
 927  */
 928 void firm_early_init(void)
 929 {
 930         /* arg: need this here for command line options */
 931         be_opt_register();
 932
 933         enable_safe_defaults();
 934 }