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_timing.h"
  21 #include "ast2firm.h"
  22 #include "adt/strutil.h"
  23 #include "adt/util.h"
  24
  25 /* optimization settings */
  26 struct a_firm_opt {
  27         bool     const_folding;   /**< enable constant folding */
  28         bool     cse;             /**< enable common-subexpression elimination */
  29         bool     confirm;         /**< enable Confirm optimization */
  30         bool     muls;            /**< enable architecture dependent mul optimization */
  31         bool     divs;            /**< enable architecture dependent div optimization */
  32         bool     mods;            /**< enable architecture dependent mod optimization */
  33         bool     alias_analysis;  /**< enable Alias Analysis */
  34         bool     strict_alias;    /**< enable strict Alias Analysis (using type based AA) */
  35         bool     no_alias;        /**< no aliasing possible. */
  36         bool     verify;          /**< Firm verifier setting */
  37         bool     check_all;       /**< enable checking all Firm phases */
  38         int      clone_threshold; /**< The threshold value for procedure cloning. */
  39         unsigned inline_maxsize;  /**< Maximum function size for inlining. */
  40         unsigned inline_threshold;/**< Inlining benefice threshold. */
  41         bool     verify_edges;    /**< verify edges */
  42 };
  43
  44 /** statistic options */
  45 typedef enum a_firmstat_selection_tag {
  46         STAT_NONE        = 0x00000000,
  47         STAT_BEFORE_OPT  = 0x00000001,
  48         STAT_AFTER_OPT   = 0x00000002,
  49         STAT_AFTER_LOWER = 0x00000004,
  50         STAT_FINAL_IR    = 0x00000008,
  51         STAT_FINAL       = 0x00000010,
  52 } a_firmstat_selection;
  53
  54 /* dumping options */
  55 struct a_firm_dump {
  56         bool debug_print;   /**< enable debug print */
  57         bool all_types;     /**< dump the All_types graph */
  58         bool no_blocks;     /**< dump non-blocked graph */
  59         bool extbb;         /**< dumps extended basic blocks */
  60         bool ir_graph;      /**< dump all graphs */
  61         bool all_phases;    /**< dump the IR graph after all phases */
  62         bool statistic;     /**< Firm statistic setting */
  63         bool stat_pattern;  /**< enable Firm statistic pattern */
  64         bool stat_dag;      /**< enable Firm DAG statistic */
  65 };
  66
  67 struct a_firm_be_opt {
  68         bool selection;
  69         bool node_stat;
  70 };
  71
  72 /* optimization settings */
  73 static struct a_firm_opt firm_opt = {
  74         .const_folding    =  true,
  75         .cse              =  true,
  76         .confirm          =  true,
  77         .muls             =  true,
  78         .divs             =  true,
  79         .mods             =  true,
  80         .alias_analysis   =  true,
  81         .strict_alias     =  false,
  82         .no_alias         =  false,
  83         .verify           =  FIRM_VERIFICATION_ON,
  84         .check_all        =  true,
  85         .clone_threshold  =  DEFAULT_CLONE_THRESHOLD,
  86         .inline_maxsize   =  750,
  87         .inline_threshold =  0,
  88         .verify_edges     =  false,
  89 };
  90
  91 /* dumping options */
  92 static struct a_firm_dump firm_dump = {
  93         .debug_print  = false,
  94         .all_types    = false,
  95         .no_blocks    = false,
  96         .extbb        = false,
  97         .ir_graph     = false,
  98         .all_phases   = false,
  99         .statistic    = STAT_NONE,
 100         .stat_pattern = 0,
 101         .stat_dag     = 0,
 102 };
 103
 104 #define X(a)  a, sizeof(a)-1
 105
 106 /** Parameter description structure */
 107 static const struct params {
 108   const char *option;      /**< name of the option */
 109   size_t     opt_len;      /**< length of the option string */
 110   bool       *flag;        /**< address of variable to set/reset */
 111   bool       set;          /**< iff true, variable will be set, else reset */
 112   const char *description; /**< description of this option */
 113 } firm_options[] = {
 114   /* firm optimization options */
 115   { X("no-opt"),                 NULL,                       0, "disable all FIRM optimizations" },
 116   { X("cse"),                    &firm_opt.cse,              1, "enable common subexpression elimination" },
 117   { X("no-cse"),                 &firm_opt.cse,              0, "disable common subexpression elimination" },
 118   { X("const-fold"),             &firm_opt.const_folding,    1, "enable constant folding" },
 119   { X("no-const-fold"),          &firm_opt.const_folding,    0, "disable constant folding" },
 120   { X("inline-max-size=<size>"), NULL,                       0, "set maximum size for function inlining" },
 121   { X("inline-threshold=<size>"),NULL,                       0, "set benefice threshold for function inlining" },
 122   { X("confirm"),                &firm_opt.confirm,          1, "enable Confirm optimization" },
 123   { X("no-confirm"),             &firm_opt.confirm,          0, "disable Confirm optimization" },
 124   { X("opt-mul"),                &firm_opt.muls,             0, "enable multiplication optimization" },
 125   { X("no-opt-mul"),             &firm_opt.muls,             0, "disable multiplication optimization" },
 126   { X("opt-div"),                &firm_opt.divs,             0, "enable division optimization" },
 127   { X("no-opt-div"),             &firm_opt.divs,             0, "disable division optimization" },
 128   { X("opt-mod"),                &firm_opt.mods,             0, "enable remainder optimization" },
 129   { X("no-opt-mod"),             &firm_opt.mods,             0, "disable remainder optimization" },
 130   { X("opt-alias"),              &firm_opt.alias_analysis,   1, "enable alias analysis" },
 131   { X("no-opt-alias"),           &firm_opt.alias_analysis,   0, "disable alias analysis" },
 132   { X("alias"),                  &firm_opt.no_alias,         0, "aliasing occurs" },
 133   { X("no-alias"),               &firm_opt.no_alias,         1, "no aliasing occurs" },
 134   { X("strict-aliasing"),        &firm_opt.strict_alias,     1, "strict alias rules" },
 135   { X("no-strict-aliasing"),     &firm_opt.strict_alias,     0, "strict alias rules" },
 136   { X("clone-threshold=<value>"),NULL,                       0, "set clone threshold to <value>" },
 137
 138   /* other firm regarding options */
 139   { X("verify-off"),             &firm_opt.verify,           FIRM_VERIFICATION_OFF,    "disable node verification" },
 140   { X("verify-on"),              &firm_opt.verify,           FIRM_VERIFICATION_ON,     "enable node verification" },
 141   { X("verify-report"),          &firm_opt.verify,           FIRM_VERIFICATION_REPORT, "node verification, report only" },
 142   { X("check-all"),              &firm_opt.check_all,        1, "enable checking all Firm phases" },
 143   { X("no-check-all"),           &firm_opt.check_all,        0, "disable checking all Firm phases" },
 144   { X("verify-edges-on"),        &firm_opt.verify_edges,     1, "enable out edge verification" },
 145   { X("verify-edges-off"),       &firm_opt.verify_edges,     0, "disable out edge verification" },
 146
 147   /* dumping */
 148   { X("dump-ir"),                &firm_dump.ir_graph,        1, "dump IR graph" },
 149   { X("dump-all-types"),         &firm_dump.all_types,       1, "dump graph of all types" },
 150   { X("dump-no-blocks"),         &firm_dump.no_blocks,       1, "dump non-blocked graph" },
 151   { X("dump-extbb"),             &firm_dump.extbb,           1, "dump extended basic blocks" },
 152   { X("dump-all-phases"),        &firm_dump.all_phases,      1, "dump graphs for all optimization phases" },
 153   { X("dump-filter=<string>"),   NULL,                       0, "set dumper filter" },
 154
 155   /* misc */
 156   { X("stat-before-opt"),        &firm_dump.statistic,       STAT_BEFORE_OPT,  "Firm statistic output before optimizations" },
 157   { X("stat-after-opt"),         &firm_dump.statistic,       STAT_AFTER_OPT,   "Firm statistic output after optimizations" },
 158   { X("stat-after-lower"),       &firm_dump.statistic,       STAT_AFTER_LOWER, "Firm statistic output after lowering" },
 159   { X("stat-final-ir"),          &firm_dump.statistic,       STAT_FINAL_IR,    "Firm statistic after final optimization" },
 160   { X("stat-final"),             &firm_dump.statistic,       STAT_FINAL,       "Firm statistic after code generation" },
 161   { X("stat-pattern"),           &firm_dump.stat_pattern,    1, "Firm statistic calculates most used pattern" },
 162   { X("stat-dag"),               &firm_dump.stat_dag,        1, "Firm calculates DAG statistics" },
 163 };
 164
 165 #undef X
 166
 167 static ir_timer_t *t_vcg_dump;
 168 static ir_timer_t *t_verify;
 169 static ir_timer_t *t_all_opt;
 170 static bool do_irg_opt(ir_graph *irg, const char *name);
 171
 172 /** dump all the graphs depending on cond */
 173
 174 static void dump_all(const char *suffix)
 175 {
 176         if (!firm_dump.ir_graph)
 177                 return;
 178
 179         timer_push(t_vcg_dump);
 180         dump_all_ir_graphs(suffix);
 181         timer_pop(t_vcg_dump);
 182 }
 183
 184 /* entities of runtime functions */
 185 ir_entity_ptr rts_entities[rts_max];
 186
 187 /**
 188  * Map runtime functions.
 189  */
 190 static void rts_map(void)
 191 {
 192         static const struct {
 193                 ir_entity_ptr *ent; /**< address of the rts entity */
 194                 i_mapper_func func; /**< mapper function. */
 195         } mapper[] = {
 196                 /* integer */
 197                 { &rts_entities[rts_abs],     i_mapper_abs },
 198                 { &rts_entities[rts_labs],    i_mapper_abs },
 199                 { &rts_entities[rts_llabs],   i_mapper_abs },
 200                 { &rts_entities[rts_imaxabs], i_mapper_abs },
 201
 202                 /* double -> double */
 203                 { &rts_entities[rts_fabs],    i_mapper_abs },
 204                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
 205                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
 206                 { &rts_entities[rts_pow],     i_mapper_pow },
 207                 { &rts_entities[rts_exp],     i_mapper_exp },
 208                 { &rts_entities[rts_exp2],    i_mapper_exp },
 209                 { &rts_entities[rts_exp10],   i_mapper_exp },
 210                 { &rts_entities[rts_log],     i_mapper_log },
 211                 { &rts_entities[rts_log2],    i_mapper_log2 },
 212                 { &rts_entities[rts_log10],   i_mapper_log10 },
 213                 { &rts_entities[rts_sin],     i_mapper_sin },
 214                 { &rts_entities[rts_cos],     i_mapper_cos },
 215                 { &rts_entities[rts_tan],     i_mapper_tan },
 216                 { &rts_entities[rts_asin],    i_mapper_asin },
 217                 { &rts_entities[rts_acos],    i_mapper_acos },
 218                 { &rts_entities[rts_atan],    i_mapper_atan },
 219                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 220                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 221                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 222
 223                 /* float -> float */
 224                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 225                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 226                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 227                 { &rts_entities[rts_powf],    i_mapper_pow },
 228                 { &rts_entities[rts_expf],    i_mapper_exp },
 229                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 230                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 231                 { &rts_entities[rts_logf],    i_mapper_log },
 232                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 233                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 234                 { &rts_entities[rts_sinf],    i_mapper_sin },
 235                 { &rts_entities[rts_cosf],    i_mapper_cos },
 236                 { &rts_entities[rts_tanf],    i_mapper_tan },
 237                 { &rts_entities[rts_asinf],   i_mapper_asin },
 238                 { &rts_entities[rts_acosf],   i_mapper_acos },
 239                 { &rts_entities[rts_atanf],   i_mapper_atan },
 240                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 241                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 242                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 243
 244                 /* long double -> long double */
 245                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 246                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 247                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 248                 { &rts_entities[rts_powl],    i_mapper_pow },
 249                 { &rts_entities[rts_expl],    i_mapper_exp },
 250                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 251                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 252                 { &rts_entities[rts_logl],    i_mapper_log },
 253                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 254                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 255                 { &rts_entities[rts_sinl],    i_mapper_sin },
 256                 { &rts_entities[rts_cosl],    i_mapper_cos },
 257                 { &rts_entities[rts_tanl],    i_mapper_tan },
 258                 { &rts_entities[rts_asinl],   i_mapper_asin },
 259                 { &rts_entities[rts_acosl],   i_mapper_acos },
 260                 { &rts_entities[rts_atanl],   i_mapper_atan },
 261                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 262                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 263                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 264
 265                 /* string */
 266                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 267                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 268                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 269                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 270                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 271                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 272                 { &rts_entities[rts_memmove], i_mapper_memmove },
 273                 { &rts_entities[rts_memset],  i_mapper_memset },
 274                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 275         };
 276         i_record rec[lengthof(mapper)];
 277         size_t   n_map = 0;
 278
 279         for (size_t i = 0; i != lengthof(mapper); ++i) {
 280                 if (*mapper[i].ent != NULL) {
 281                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 282                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 283                         rec[n_map].i_call.i_mapper = mapper[i].func;
 284                         rec[n_map].i_call.ctx      = NULL;
 285                         rec[n_map].i_call.link     = NULL;
 286                         ++n_map;
 287                 }
 288         }
 289
 290         if (n_map > 0)
 291                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 292 }
 293
 294 static int *irg_dump_no;
 295
 296 static void do_stred(ir_graph *irg)
 297 {
 298         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 299 }
 300
 301 static void after_inline_opt(ir_graph *irg)
 302 {
 303         do_irg_opt(irg, "scalar-replace");
 304         do_irg_opt(irg, "local");
 305         do_irg_opt(irg, "control-flow");
 306         do_irg_opt(irg, "combo");
 307 }
 308
 309 static void do_inline(void)
 310 {
 311         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 312                          after_inline_opt);
 313 }
 314
 315 static void do_cloning(void)
 316 {
 317         proc_cloning((float) firm_opt.clone_threshold);
 318 }
 319
 320 static void do_lower_mux(ir_graph *irg)
 321 {
 322         lower_mux(irg, NULL);
 323 }
 324
 325 static void do_gcse(ir_graph *irg)
 326 {
 327         set_opt_global_cse(1);
 328         optimize_graph_df(irg);
 329         set_opt_global_cse(0);
 330 }
 331
 332 typedef enum opt_target {
 333         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 334         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 335 } opt_target_t;
 336
 337 typedef enum opt_flags {
 338         OPT_FLAG_NONE         = 0,
 339         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 340         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 341         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 342         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 343                                              -foptions for this transformation */
 344         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 345                                              so we need it even with -O0 */
 346 } opt_flags_t;
 347
 348 typedef void (*transform_irg_func)(ir_graph *irg);
 349 typedef void (*transform_irp_func)(void);
 350
 351 typedef struct {
 352         opt_target_t  target;
 353         const char   *name;
 354         union {
 355                 transform_irg_func transform_irg;
 356                 transform_irp_func transform_irp;
 357         } u;
 358         const char   *description;
 359         opt_flags_t   flags;
 360         ir_timer_t   *timer;
 361 } opt_config_t;
 362
 363 static opt_config_t opts[] = {
 364 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 365 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 366         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 367         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 368         IRG("confirm",           construct_confirms,       "confirm optimization",                                  OPT_FLAG_HIDE_OPTIONS),
 369         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 370         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 371         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 372         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 373         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 374         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 375         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 376         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 377         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 378         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 379         IRG("lower",             lower_highlevel_graph,    "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 380         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 381         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 382         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 383         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 384         IRG("gcse",              do_gcse,                  "global common subexpression eliminiation",              OPT_FLAG_NONE),
 385         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 386         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 387         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 388         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 389         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 390         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 391         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 392         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 393         IRG("vrp",               set_vrp_data,             "value range propagation",                               OPT_FLAG_NONE),
 394         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 395         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),
 396         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 397         IRP("opt-func-call",     optimize_funccalls,       "function call optimization",                            OPT_FLAG_NONE),
 398         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 399         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 400         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_NONE),
 401         IRP("opt-cc",            mark_private_methods,     "calling conventions optimization",                      OPT_FLAG_NONE),
 402 #undef IRP
 403 #undef IRG
 404 };
 405
 406 #define FOR_EACH_OPT(i) for (opt_config_t *i = opts; i != endof(opts); ++i)
 407
 408 static opt_config_t *get_opt(const char *name)
 409 {
 410         FOR_EACH_OPT(config) {
 411                 if (streq(config->name, name))
 412                         return config;
 413         }
 414
 415         return NULL;
 416 }
 417
 418 static void set_opt_enabled(const char *name, bool enabled)
 419 {
 420         opt_config_t *config = get_opt(name);
 421         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 422                 | (enabled ? OPT_FLAG_ENABLED : 0);
 423 }
 424
 425 static bool get_opt_enabled(const char *name)
 426 {
 427         opt_config_t *config = get_opt(name);
 428         return (config->flags & OPT_FLAG_ENABLED) != 0;
 429 }
 430
 431 /**
 432  * perform an optimization on a single graph
 433  *
 434  * @return  true if something changed, false otherwise
 435  */
 436 static bool do_irg_opt(ir_graph *irg, const char *name)
 437 {
 438         opt_config_t *const config = get_opt(name);
 439         assert(config != NULL);
 440         assert(config->target == OPT_TARGET_IRG);
 441         if (! (config->flags & OPT_FLAG_ENABLED))
 442                 return false;
 443
 444         ir_graph *const old_irg = current_ir_graph;
 445         current_ir_graph = irg;
 446
 447         timer_push(config->timer);
 448         config->u.transform_irg(irg);
 449         timer_pop(config->timer);
 450
 451         if (firm_dump.all_phases && firm_dump.ir_graph) {
 452                 dump_ir_graph(irg, name);
 453         }
 454
 455         if (firm_opt.check_all) {
 456                 timer_push(t_verify);
 457                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 458                 timer_pop(t_verify);
 459         }
 460
 461         current_ir_graph = old_irg;
 462         return true;
 463 }
 464
 465 static void do_irp_opt(const char *name)
 466 {
 467         opt_config_t *const config = get_opt(name);
 468         assert(config->target == OPT_TARGET_IRP);
 469         if (! (config->flags & OPT_FLAG_ENABLED))
 470                 return;
 471
 472         timer_push(config->timer);
 473         config->u.transform_irp();
 474         timer_pop(config->timer);
 475
 476         if (firm_dump.ir_graph && firm_dump.all_phases) {
 477                 int i;
 478                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 479                         ir_graph *irg = get_irp_irg(i);
 480                         dump_ir_graph(irg, name);
 481                 }
 482         }
 483
 484         if (firm_opt.check_all) {
 485                 int i;
 486                 timer_push(t_verify);
 487                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 488                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 489                 }
 490                 timer_pop(t_verify);
 491         }
 492 }
 493
 494 /**
 495  * Enable transformations which should be always safe (and cheap) to perform
 496  */
 497 static void enable_safe_defaults(void)
 498 {
 499         set_opt_enabled("remove-unused", true);
 500         set_opt_enabled("opt-tail-rec", true);
 501         set_opt_enabled("opt-func-call", true);
 502         set_opt_enabled("reassociation", true);
 503         set_opt_enabled("control-flow", true);
 504         set_opt_enabled("local", true);
 505         set_opt_enabled("lower-const", true);
 506         set_opt_enabled("scalar-replace", true);
 507         set_opt_enabled("place", true);
 508         set_opt_enabled("gcse", true);
 509         set_opt_enabled("confirm", true);
 510         set_opt_enabled("opt-load-store", true);
 511         set_opt_enabled("lower", true);
 512         set_opt_enabled("deconv", true);
 513         set_opt_enabled("remove-confirms", true);
 514         set_opt_enabled("ivopts", true);
 515         set_opt_enabled("dead", true);
 516         set_opt_enabled("remove-phi-cycles", true);
 517         set_opt_enabled("frame", true);
 518         set_opt_enabled("combo", true);
 519         set_opt_enabled("invert-loops", true);
 520         set_opt_enabled("target-lowering", true);
 521         set_opt_enabled("rts", true);
 522         set_opt_enabled("parallelize-mem", true);
 523         set_opt_enabled("opt-cc", true);
 524 }
 525
 526 /**
 527  * run all the Firm optimizations
 528  *
 529  * @param input_filename     the name of the (main) source file
 530  */
 531 static void do_firm_optimizations(const char *input_filename)
 532 {
 533         size_t   i;
 534         unsigned aa_opt;
 535
 536         set_opt_alias_analysis(firm_opt.alias_analysis);
 537
 538         aa_opt = aa_opt_no_opt;
 539         if (firm_opt.strict_alias)
 540                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 541         if (firm_opt.no_alias)
 542                 aa_opt = aa_opt_no_alias;
 543
 544         set_irp_memory_disambiguator_options(aa_opt);
 545
 546         /* parameter passing code should set them directly sometime... */
 547         set_opt_enabled("confirm", firm_opt.confirm);
 548         set_opt_enabled("remove-confirms", firm_opt.confirm);
 549
 550         /* osr supersedes remove_phi_cycles */
 551         if (get_opt_enabled("ivopts"))
 552                 set_opt_enabled("remove-phi-cycles", false);
 553
 554         timer_start(t_all_opt);
 555
 556         do_irp_opt("rts");
 557
 558         /* first step: kill dead code */
 559         for (i = 0; i < get_irp_n_irgs(); i++) {
 560                 ir_graph *irg = get_irp_irg(i);
 561                 do_irg_opt(irg, "combo");
 562                 do_irg_opt(irg, "local");
 563                 do_irg_opt(irg, "control-flow");
 564         }
 565
 566         do_irp_opt("remove-unused");
 567         for (i = 0; i < get_irp_n_irgs(); ++i) {
 568                 ir_graph *irg = get_irp_irg(i);
 569                 do_irg_opt(irg, "opt-tail-rec");
 570         }
 571         do_irp_opt("opt-func-call");
 572         do_irp_opt("lower-const");
 573
 574         for (i = 0; i < get_irp_n_irgs(); i++) {
 575                 ir_graph *irg = get_irp_irg(i);
 576
 577                 do_irg_opt(irg, "scalar-replace");
 578                 do_irg_opt(irg, "invert-loops");
 579                 do_irg_opt(irg, "unroll-loops");
 580                 do_irg_opt(irg, "local");
 581                 do_irg_opt(irg, "reassociation");
 582                 do_irg_opt(irg, "local");
 583                 do_irg_opt(irg, "gcse");
 584                 do_irg_opt(irg, "place");
 585
 586                 if (firm_opt.confirm) {
 587                         /* Confirm construction currently can only handle blocks with only
 588                            one control flow predecessor. Calling optimize_cf here removes
 589                            Bad predecessors and help the optimization of switch constructs.
 590                          */
 591                         do_irg_opt(irg, "control-flow");
 592                         do_irg_opt(irg, "confirm");
 593                         do_irg_opt(irg, "vrp");
 594                         do_irg_opt(irg, "local");
 595                 }
 596
 597                 do_irg_opt(irg, "control-flow");
 598                 do_irg_opt(irg, "opt-load-store");
 599                 do_irg_opt(irg, "fp-vrp");
 600                 do_irg_opt(irg, "lower");
 601                 do_irg_opt(irg, "deconv");
 602                 do_irg_opt(irg, "thread-jumps");
 603                 do_irg_opt(irg, "remove-confirms");
 604                 do_irg_opt(irg, "gvn-pre");
 605                 do_irg_opt(irg, "gcse");
 606                 do_irg_opt(irg, "place");
 607                 do_irg_opt(irg, "control-flow");
 608
 609                 if (do_irg_opt(irg, "if-conversion")) {
 610                         do_irg_opt(irg, "local");
 611                         do_irg_opt(irg, "control-flow");
 612                 }
 613                 /* this doesn't make too much sense but tests the mux destruction... */
 614                 do_irg_opt(irg, "lower-mux");
 615
 616                 do_irg_opt(irg, "bool");
 617                 do_irg_opt(irg, "shape-blocks");
 618                 do_irg_opt(irg, "ivopts");
 619                 do_irg_opt(irg, "local");
 620                 do_irg_opt(irg, "dead");
 621         }
 622
 623         do_irp_opt("inline");
 624         do_irp_opt("opt-proc-clone");
 625
 626         for (i = 0; i < get_irp_n_irgs(); i++) {
 627                 ir_graph *irg = get_irp_irg(i);
 628                 do_irg_opt(irg, "local");
 629                 do_irg_opt(irg, "control-flow");
 630                 do_irg_opt(irg, "thread-jumps");
 631                 do_irg_opt(irg, "local");
 632                 do_irg_opt(irg, "control-flow");
 633
 634                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 635                         do_irg_opt(irg, "local");
 636                         do_irg_opt(irg, "vrp");
 637                         do_irg_opt(irg, "local");
 638                         do_irg_opt(irg, "vrp");
 639                 }
 640         }
 641
 642         if (firm_dump.ir_graph) {
 643                 /* recompute backedges for nicer dumps */
 644                 for (i = 0; i < get_irp_n_irgs(); i++)
 645                         construct_cf_backedges(get_irp_irg(i));
 646         }
 647
 648         dump_all("opt");
 649
 650         if (firm_dump.statistic & STAT_AFTER_OPT)
 651                 stat_dump_snapshot(input_filename, "opt");
 652
 653         timer_stop(t_all_opt);
 654 }
 655
 656 /**
 657  * do Firm lowering
 658  *
 659  * @param input_filename  the name of the (main) source file
 660  */
 661 static void do_firm_lowering(const char *input_filename)
 662 {
 663         int i;
 664
 665         /* enable architecture dependent optimizations */
 666         arch_dep_set_opts((arch_dep_opts_t)
 667                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 668                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 669                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 670         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 671                 ir_graph *irg = get_irp_irg(i);
 672                 do_irg_opt(irg, "reassociation");
 673                 do_irg_opt(irg, "local");
 674         }
 675
 676         do_irp_opt("target-lowering");
 677
 678         if (firm_dump.statistic & STAT_AFTER_LOWER)
 679                 stat_dump_snapshot(input_filename, "low");
 680
 681         timer_start(t_all_opt);
 682
 683         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 684                 ir_graph *irg = get_irp_irg(i);
 685
 686                 do_irg_opt(irg, "local");
 687                 do_irg_opt(irg, "deconv");
 688                 do_irg_opt(irg, "control-flow");
 689                 do_irg_opt(irg, "opt-load-store");
 690                 do_irg_opt(irg, "gcse");
 691                 do_irg_opt(irg, "place");
 692                 do_irg_opt(irg, "control-flow");
 693
 694                 if (do_irg_opt(irg, "vrp")) {
 695                         do_irg_opt(irg, "local");
 696                         do_irg_opt(irg, "control-flow");
 697                         do_irg_opt(irg, "vrp");
 698                         do_irg_opt(irg, "local");
 699                         do_irg_opt(irg, "control-flow");
 700                 }
 701
 702                 if (do_irg_opt(irg, "if-conversion")) {
 703                         do_irg_opt(irg, "local");
 704                         do_irg_opt(irg, "control-flow");
 705                 }
 706
 707                 set_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2);
 708                 do_irg_opt(irg, "local");
 709
 710                 do_irg_opt(irg, "parallelize-mem");
 711                 do_irg_opt(irg, "frame");
 712         }
 713         do_irp_opt("remove-unused");
 714         do_irp_opt("opt-cc");
 715         timer_stop(t_all_opt);
 716         dump_all("low-opt");
 717
 718         if (firm_dump.statistic & STAT_FINAL) {
 719                 stat_dump_snapshot(input_filename, "final");
 720         }
 721 }
 722
 723 /**
 724  * Initialize for the Firm-generating back end.
 725  */
 726 void gen_firm_init(void)
 727 {
 728         unsigned pattern = 0;
 729
 730         FOR_EACH_OPT(i) {
 731                 i->timer = ir_timer_new();
 732                 timer_register(i->timer, i->description);
 733         }
 734         t_verify = ir_timer_new();
 735         timer_register(t_verify, "Firm: verify pass");
 736         t_vcg_dump = ir_timer_new();
 737         timer_register(t_vcg_dump, "Firm: vcg dumping");
 738         t_all_opt = ir_timer_new();
 739         timer_register(t_all_opt, "Firm: all optimizations");
 740
 741         if (firm_dump.stat_pattern)
 742                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 743
 744         if (firm_dump.stat_dag)
 745                 pattern |= FIRMSTAT_COUNT_DAG;
 746
 747         ir_init(NULL);
 748         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 749                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 750                         | FIRMSTAT_COUNT_CONSTS | pattern);
 751
 752         edges_init_dbg(firm_opt.verify_edges);
 753
 754         /* Sel node cannot produce NULL pointers */
 755         set_opt_sel_based_null_check_elim(1);
 756
 757         /* dynamic dispatch works currently only if whole world scenarios */
 758         set_opt_dyn_meth_dispatch(0);
 759
 760         /* do not run architecture dependent optimizations in building phase */
 761         arch_dep_set_opts(arch_dep_none);
 762
 763         do_node_verification((firm_verification_t) firm_opt.verify);
 764         if (firm_dump.extbb)
 765                 ir_add_dump_flags(ir_dump_flag_group_extbb);
 766         if (firm_dump.no_blocks)
 767                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
 768
 769         set_optimize(1);
 770         set_opt_constant_folding(firm_opt.const_folding);
 771         set_opt_algebraic_simplification(firm_opt.const_folding);
 772         set_opt_cse(firm_opt.cse);
 773         set_opt_global_cse(0);
 774         set_opt_unreachable_code(1);
 775 }
 776
 777 /**
 778  * Called, after the Firm generation is completed,
 779  * do all optimizations and backend call here.
 780  *
 781  * @param out                a file handle for the output, may be NULL
 782  * @param input_filename     the name of the (main) source file
 783  * @param c_mode             non-zero if "C" was compiled
 784  */
 785 void gen_firm_finish(FILE *out, const char *input_filename)
 786 {
 787         int i;
 788
 789         /* the general for dumping option must be set, or the others will not work*/
 790         firm_dump.ir_graph
 791                 = (bool) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 792
 793         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 794                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 795         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 796
 797         /* FIXME: cloning might ADD new graphs. */
 798         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 799
 800         if (firm_dump.all_types) {
 801                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 802         }
 803
 804         /* finalize all graphs */
 805         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 806                 ir_graph *irg = get_irp_irg(i);
 807                 irg_finalize_cons(irg);
 808         }
 809         dump_all("");
 810
 811         timer_push(t_verify);
 812         tr_verify();
 813         timer_pop(t_verify);
 814
 815         /* all graphs are finalized, set the irp phase to high */
 816         set_irp_phase_state(phase_high);
 817
 818         /* BEWARE: kill unreachable code before doing compound lowering */
 819         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 820                 ir_graph *irg = get_irp_irg(i);
 821                 do_irg_opt(irg, "control-flow");
 822         }
 823
 824         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 825                 stat_dump_snapshot(input_filename, "noopt");
 826         }
 827
 828         do_firm_optimizations(input_filename);
 829         do_firm_lowering(input_filename);
 830
 831         /* set the phase to low */
 832         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 833                 set_irg_phase_state(get_irp_irg(i), phase_low);
 834
 835         if (firm_dump.statistic & STAT_FINAL_IR)
 836                 stat_dump_snapshot(input_filename, "final-ir");
 837
 838         /* run the code generator */
 839         ir_timer_t *timer = ir_timer_new();
 840         timer_register(timer, "Firm: backend");
 841         timer_start(timer);
 842         be_main(out, input_filename);
 843         timer_stop(timer);
 844
 845         if (firm_dump.statistic & STAT_FINAL)
 846                 stat_dump_snapshot(input_filename, "final");
 847 }
 848
 849 static void disable_all_opts(void)
 850 {
 851         firm_opt.cse             = false;
 852         firm_opt.confirm         = false;
 853         firm_opt.muls            = false;
 854         firm_opt.divs            = false;
 855         firm_opt.mods            = false;
 856         firm_opt.alias_analysis  = false;
 857         firm_opt.strict_alias    = false;
 858         firm_opt.no_alias        = false;
 859
 860         FOR_EACH_OPT(config) {
 861                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 862                         config->flags |= OPT_FLAG_ENABLED;
 863                 } else {
 864                         config->flags &= ~OPT_FLAG_ENABLED;
 865                 }
 866         }
 867 }
 868
 869 static bool firm_opt_option(const char *opt)
 870 {
 871         char const* const rest   = strstart(opt, "no-");
 872         bool        const enable = rest ? opt = rest, false : true;
 873
 874         opt_config_t *config = get_opt(opt);
 875         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 876                 return false;
 877
 878         config->flags &= ~OPT_FLAG_ENABLED;
 879         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 880         return true;
 881 }
 882
 883 void firm_option_help(print_option_help_func print_option_help)
 884 {
 885         print_option_help(firm_options[0].option, firm_options[0].description);
 886
 887         FOR_EACH_OPT(config) {
 888                 char buf[1024];
 889                 char buf2[1024];
 890
 891                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 892                         continue;
 893
 894                 snprintf(buf, sizeof(buf), "-f%s", config->name);
 895                 snprintf(buf2, sizeof(buf2), "enable %s", config->description);
 896                 print_option_help(buf, buf2);
 897                 snprintf(buf, sizeof(buf), "-fno-%s", config->name);
 898                 snprintf(buf2, sizeof(buf2), "disable %s", config->description);
 899                 print_option_help(buf, buf2);
 900         }
 901
 902         for (size_t k = 0; k != lengthof(firm_options); ++k) {
 903                 char buf[1024];
 904                 char buf2[1024];
 905                 snprintf(buf, sizeof(buf), "-f%s", firm_options[k].option);
 906                 snprintf(buf2, sizeof(buf2), "%s", firm_options[k].description);
 907                 print_option_help(buf, buf2);
 908         }
 909 }
 910
 911 int firm_option(const char *const opt)
 912 {
 913         char const* val;
 914         if ((val = strstart(opt, "dump-filter="))) {
 915                 ir_set_dump_filter(val);
 916                 return 1;
 917         } else if ((val = strstart(opt, "clone-threshold="))) {
 918                 sscanf(val, "%d", &firm_opt.clone_threshold);
 919                 return 1;
 920         } else if ((val = strstart(opt, "inline-max-size="))) {
 921                 sscanf(val, "%u", &firm_opt.inline_maxsize);
 922                 return 1;
 923         } else if ((val = strstart(opt, "inline-threshold="))) {
 924                 sscanf(val, "%u", &firm_opt.inline_threshold);
 925                 return 1;
 926         } else if (streq(opt, "no-opt")) {
 927                 disable_all_opts();
 928                 return 1;
 929         }
 930
 931         size_t const len = strlen(opt);
 932         for (size_t i = lengthof(firm_options); i != 0;) {
 933                 struct params const* const o = &firm_options[--i];
 934                 if (len == o->opt_len && strncmp(opt, o->option, len) == 0) {
 935                         /* statistic options do accumulate */
 936                         if (o->flag == &firm_dump.statistic)
 937                                 *o->flag = (bool) (*o->flag | o->set);
 938                         else
 939                                 *o->flag = o->set;
 940
 941                         return 1;
 942                 }
 943         }
 944
 945         /* maybe this enables/disables optimizations */
 946         if (firm_opt_option(opt))
 947                 return 1;
 948
 949         return 0;
 950 }
 951
 952 static void set_be_option(const char *arg)
 953 {
 954         int res = be_parse_arg(arg);
 955         (void) res;
 956         assert(res);
 957 }
 958
 959 static void set_option(const char *arg)
 960 {
 961         int res = firm_option(arg);
 962         (void) res;
 963         assert(res);
 964 }
 965
 966 void choose_optimization_pack(int level)
 967 {
 968         /* apply optimization level */
 969         switch(level) {
 970         case 0:
 971                 set_option("no-opt");
 972                 break;
 973         case 1:
 974                 set_option("no-inline");
 975                 break;
 976         default:
 977         case 4:
 978                 /* use_builtins = true; */
 979                 /* fallthrough */
 980         case 3:
 981                 set_option("thread-jumps");
 982                 set_option("if-conversion");
 983                 /* fallthrough */
 984         case 2:
 985                 set_option("strict-aliasing");
 986                 set_option("inline");
 987                 set_option("fp-vrp");
 988                 set_option("deconv");
 989                 set_be_option("omitfp");
 990                 break;
 991         }
 992 }
 993
 994 /**
 995  * Do very early initializations
 996  */
 997 void firm_early_init(void)
 998 {
 999         /* arg: need this here for command line options */
1000         be_opt_register();
1001
1002         enable_safe_defaults();
1003 }