nsz Git - cparser/blob - driver/firm_opt.c

   1 /**
   2  * (C) 2005-2010
   3  * @file
   4  * @author Michael Beck, Matthias Braun
   5  * @brief Firm-generating back end optimizations.
   6  */
   7 #include <config.h>
   8
   9 #include <stdbool.h>
  10 #include <stdlib.h>
  11 #include <string.h>
  12 #include <stdbool.h>
  13 #include <assert.h>
  14 #include <libfirm/firm.h>
  15
  16 #include "firm_opt.h"
  17 #include "firm_timing.h"
  18 #include "ast2firm.h"
  19 #include "adt/strutil.h"
  20 #include "adt/util.h"
  21
  22 /* optimization settings */
  23 struct a_firm_opt {
  24         bool     const_folding;   /**< enable constant folding */
  25         bool     cse;             /**< enable common-subexpression elimination */
  26         bool     confirm;         /**< enable Confirm optimization */
  27         bool     muls;            /**< enable architecture dependent mul optimization */
  28         bool     divs;            /**< enable architecture dependent div optimization */
  29         bool     mods;            /**< enable architecture dependent mod optimization */
  30         bool     alias_analysis;  /**< enable Alias Analysis */
  31         bool     strict_alias;    /**< enable strict Alias Analysis (using type based AA) */
  32         bool     no_alias;        /**< no aliasing possible. */
  33         bool     verify;          /**< Firm verifier setting */
  34         bool     check_all;       /**< enable checking all Firm phases */
  35         int      clone_threshold; /**< The threshold value for procedure cloning. */
  36         unsigned inline_maxsize;  /**< Maximum function size for inlining. */
  37         unsigned inline_threshold;/**< Inlining benefice threshold. */
  38 };
  39
  40 /** statistic options */
  41 typedef enum a_firmstat_selection_tag {
  42         STAT_NONE        = 0x00000000,
  43         STAT_BEFORE_OPT  = 0x00000001,
  44         STAT_AFTER_OPT   = 0x00000002,
  45         STAT_AFTER_LOWER = 0x00000004,
  46         STAT_FINAL_IR    = 0x00000008,
  47         STAT_FINAL       = 0x00000010,
  48 } a_firmstat_selection;
  49
  50 /* dumping options */
  51 struct a_firm_dump {
  52         bool debug_print;   /**< enable debug print */
  53         bool all_types;     /**< dump the All_types graph */
  54         bool ir_graph;      /**< dump all graphs */
  55         bool all_phases;    /**< dump the IR graph after all phases */
  56         bool statistic;     /**< Firm statistic setting */
  57         bool stat_pattern;  /**< enable Firm statistic pattern */
  58         bool stat_dag;      /**< enable Firm DAG statistic */
  59 };
  60
  61 struct a_firm_be_opt {
  62         bool selection;
  63         bool node_stat;
  64 };
  65
  66 /* optimization settings */
  67 static struct a_firm_opt firm_opt = {
  68         .const_folding    =  true,
  69         .cse              =  true,
  70         .confirm          =  true,
  71         .muls             =  true,
  72         .divs             =  true,
  73         .mods             =  true,
  74         .alias_analysis   =  true,
  75         .strict_alias     =  false,
  76         .no_alias         =  false,
  77         .verify           =  FIRM_VERIFICATION_ON,
  78         .check_all        =  true,
  79         .clone_threshold  =  DEFAULT_CLONE_THRESHOLD,
  80         .inline_maxsize   =  750,
  81         .inline_threshold =  0,
  82 };
  83
  84 /* dumping options */
  85 static struct a_firm_dump firm_dump = {
  86         .debug_print  = false,
  87         .all_types    = false,
  88         .ir_graph     = false,
  89         .all_phases   = false,
  90         .statistic    = STAT_NONE,
  91         .stat_pattern = 0,
  92         .stat_dag     = 0,
  93 };
  94
  95 #define X(a)  a, sizeof(a)-1
  96
  97 /** Parameter description structure */
  98 static const struct params {
  99   const char *option;      /**< name of the option */
 100   size_t     opt_len;      /**< length of the option string */
 101   bool       *flag;        /**< address of variable to set/reset */
 102   bool       set;          /**< iff true, variable will be set, else reset */
 103   const char *description; /**< description of this option */
 104 } firm_options[] = {
 105   /* firm optimization options */
 106   { X("no-opt"),                 NULL,                       0, "disable all FIRM optimizations" },
 107   { X("cse"),                    &firm_opt.cse,              1, "enable common subexpression elimination" },
 108   { X("no-cse"),                 &firm_opt.cse,              0, "disable common subexpression elimination" },
 109   { X("const-fold"),             &firm_opt.const_folding,    1, "enable constant folding" },
 110   { X("no-const-fold"),          &firm_opt.const_folding,    0, "disable constant folding" },
 111   { X("inline-max-size=<size>"), NULL,                       0, "set maximum size for function inlining" },
 112   { X("inline-threshold=<size>"),NULL,                       0, "set benefice threshold for function inlining" },
 113   { X("confirm"),                &firm_opt.confirm,          1, "enable Confirm optimization" },
 114   { X("no-confirm"),             &firm_opt.confirm,          0, "disable Confirm optimization" },
 115   { X("opt-mul"),                &firm_opt.muls,             0, "enable multiplication optimization" },
 116   { X("no-opt-mul"),             &firm_opt.muls,             0, "disable multiplication optimization" },
 117   { X("opt-div"),                &firm_opt.divs,             0, "enable division optimization" },
 118   { X("no-opt-div"),             &firm_opt.divs,             0, "disable division optimization" },
 119   { X("opt-mod"),                &firm_opt.mods,             0, "enable remainder optimization" },
 120   { X("no-opt-mod"),             &firm_opt.mods,             0, "disable remainder optimization" },
 121   { X("opt-alias"),              &firm_opt.alias_analysis,   1, "enable alias analysis" },
 122   { X("no-opt-alias"),           &firm_opt.alias_analysis,   0, "disable alias analysis" },
 123   { X("alias"),                  &firm_opt.no_alias,         0, "aliasing occurs" },
 124   { X("no-alias"),               &firm_opt.no_alias,         1, "no aliasing occurs" },
 125   { X("strict-aliasing"),        &firm_opt.strict_alias,     1, "strict alias rules" },
 126   { X("no-strict-aliasing"),     &firm_opt.strict_alias,     0, "strict alias rules" },
 127   { X("clone-threshold=<value>"),NULL,                       0, "set clone threshold to <value>" },
 128
 129   /* other firm regarding options */
 130   { X("verify-off"),             &firm_opt.verify,           FIRM_VERIFICATION_OFF,    "disable node verification" },
 131   { X("verify-on"),              &firm_opt.verify,           FIRM_VERIFICATION_ON,     "enable node verification" },
 132   { X("verify-report"),          &firm_opt.verify,           FIRM_VERIFICATION_REPORT, "node verification, report only" },
 133   { X("check-all"),              &firm_opt.check_all,        1, "enable checking all Firm phases" },
 134   { X("no-check-all"),           &firm_opt.check_all,        0, "disable checking all Firm phases" },
 135
 136   /* dumping */
 137   { X("dump-ir"),                &firm_dump.ir_graph,        1, "dump IR graph" },
 138   { X("dump-all-types"),         &firm_dump.all_types,       1, "dump graph of all types" },
 139   { X("dump-all-phases"),        &firm_dump.all_phases,      1, "dump graphs for all optimization phases" },
 140   { X("dump-filter=<string>"),   NULL,                       0, "set dumper filter" },
 141
 142   /* misc */
 143   { X("stat-before-opt"),        &firm_dump.statistic,       STAT_BEFORE_OPT,  "Firm statistic output before optimizations" },
 144   { X("stat-after-opt"),         &firm_dump.statistic,       STAT_AFTER_OPT,   "Firm statistic output after optimizations" },
 145   { X("stat-after-lower"),       &firm_dump.statistic,       STAT_AFTER_LOWER, "Firm statistic output after lowering" },
 146   { X("stat-final-ir"),          &firm_dump.statistic,       STAT_FINAL_IR,    "Firm statistic after final optimization" },
 147   { X("stat-final"),             &firm_dump.statistic,       STAT_FINAL,       "Firm statistic after code generation" },
 148   { X("stat-pattern"),           &firm_dump.stat_pattern,    1, "Firm statistic calculates most used pattern" },
 149   { X("stat-dag"),               &firm_dump.stat_dag,        1, "Firm calculates DAG statistics" },
 150 };
 151
 152 #undef X
 153
 154 static ir_timer_t *t_vcg_dump;
 155 static ir_timer_t *t_verify;
 156 static ir_timer_t *t_all_opt;
 157 static ir_timer_t *t_backend;
 158 static bool do_irg_opt(ir_graph *irg, const char *name);
 159
 160 /** dump all the graphs depending on cond */
 161
 162 static void dump_all(const char *suffix)
 163 {
 164         if (!firm_dump.ir_graph)
 165                 return;
 166
 167         timer_push(t_vcg_dump);
 168         dump_all_ir_graphs(suffix);
 169         timer_pop(t_vcg_dump);
 170 }
 171
 172 /* entities of runtime functions */
 173 ir_entity *rts_entities[rts_max];
 174
 175 /**
 176  * Map runtime functions.
 177  */
 178 static void rts_map(void)
 179 {
 180         static const struct {
 181                 ir_entity   **ent; /**< address of the rts entity */
 182                 i_mapper_func func; /**< mapper function. */
 183         } mapper[] = {
 184                 /* integer */
 185                 { &rts_entities[rts_abs],     i_mapper_abs },
 186                 { &rts_entities[rts_labs],    i_mapper_abs },
 187                 { &rts_entities[rts_llabs],   i_mapper_abs },
 188                 { &rts_entities[rts_imaxabs], i_mapper_abs },
 189
 190                 /* double -> double */
 191                 { &rts_entities[rts_fabs],    i_mapper_abs },
 192                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
 193                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
 194                 { &rts_entities[rts_pow],     i_mapper_pow },
 195                 { &rts_entities[rts_exp],     i_mapper_exp },
 196                 { &rts_entities[rts_exp2],    i_mapper_exp },
 197                 { &rts_entities[rts_exp10],   i_mapper_exp },
 198                 { &rts_entities[rts_log],     i_mapper_log },
 199                 { &rts_entities[rts_log2],    i_mapper_log2 },
 200                 { &rts_entities[rts_log10],   i_mapper_log10 },
 201                 { &rts_entities[rts_sin],     i_mapper_sin },
 202                 { &rts_entities[rts_cos],     i_mapper_cos },
 203                 { &rts_entities[rts_tan],     i_mapper_tan },
 204                 { &rts_entities[rts_asin],    i_mapper_asin },
 205                 { &rts_entities[rts_acos],    i_mapper_acos },
 206                 { &rts_entities[rts_atan],    i_mapper_atan },
 207                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 208                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 209                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 210
 211                 /* float -> float */
 212                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 213                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 214                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 215                 { &rts_entities[rts_powf],    i_mapper_pow },
 216                 { &rts_entities[rts_expf],    i_mapper_exp },
 217                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 218                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 219                 { &rts_entities[rts_logf],    i_mapper_log },
 220                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 221                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 222                 { &rts_entities[rts_sinf],    i_mapper_sin },
 223                 { &rts_entities[rts_cosf],    i_mapper_cos },
 224                 { &rts_entities[rts_tanf],    i_mapper_tan },
 225                 { &rts_entities[rts_asinf],   i_mapper_asin },
 226                 { &rts_entities[rts_acosf],   i_mapper_acos },
 227                 { &rts_entities[rts_atanf],   i_mapper_atan },
 228                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 229                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 230                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 231
 232                 /* long double -> long double */
 233                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 234                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 235                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 236                 { &rts_entities[rts_powl],    i_mapper_pow },
 237                 { &rts_entities[rts_expl],    i_mapper_exp },
 238                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 239                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 240                 { &rts_entities[rts_logl],    i_mapper_log },
 241                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 242                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 243                 { &rts_entities[rts_sinl],    i_mapper_sin },
 244                 { &rts_entities[rts_cosl],    i_mapper_cos },
 245                 { &rts_entities[rts_tanl],    i_mapper_tan },
 246                 { &rts_entities[rts_asinl],   i_mapper_asin },
 247                 { &rts_entities[rts_acosl],   i_mapper_acos },
 248                 { &rts_entities[rts_atanl],   i_mapper_atan },
 249                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 250                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 251                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 252
 253                 /* string */
 254                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 255                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 256                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 257                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 258                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 259                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 260                 { &rts_entities[rts_memmove], i_mapper_memmove },
 261                 { &rts_entities[rts_memset],  i_mapper_memset },
 262                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 263         };
 264         i_record rec[lengthof(mapper)];
 265         size_t   n_map = 0;
 266
 267         for (size_t i = 0; i != lengthof(mapper); ++i) {
 268                 if (*mapper[i].ent != NULL) {
 269                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 270                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 271                         rec[n_map].i_call.i_mapper = mapper[i].func;
 272                         rec[n_map].i_call.ctx      = NULL;
 273                         rec[n_map].i_call.link     = NULL;
 274                         ++n_map;
 275                 }
 276         }
 277
 278         if (n_map > 0)
 279                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 280 }
 281
 282 static int *irg_dump_no;
 283
 284 static void do_stred(ir_graph *irg)
 285 {
 286         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 287 }
 288
 289 static void after_inline_opt(ir_graph *irg)
 290 {
 291         do_irg_opt(irg, "scalar-replace");
 292         do_irg_opt(irg, "local");
 293         do_irg_opt(irg, "control-flow");
 294         do_irg_opt(irg, "combo");
 295 }
 296
 297 static void do_inline(void)
 298 {
 299         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 300                          after_inline_opt);
 301 }
 302
 303 static void do_cloning(void)
 304 {
 305         proc_cloning((float) firm_opt.clone_threshold);
 306 }
 307
 308 static void do_lower_mux(ir_graph *irg)
 309 {
 310         lower_mux(irg, NULL);
 311 }
 312
 313 static void do_gcse(ir_graph *irg)
 314 {
 315         set_opt_global_cse(1);
 316         optimize_graph_df(irg);
 317         set_opt_global_cse(0);
 318 }
 319
 320 typedef enum opt_target {
 321         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 322         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 323 } opt_target_t;
 324
 325 typedef enum opt_flags {
 326         OPT_FLAG_NONE         = 0,
 327         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 328         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 329         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 330         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 331                                              -foptions for this transformation */
 332         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 333                                              so we need it even with -O0 */
 334 } opt_flags_t;
 335
 336 typedef void (*transform_irg_func)(ir_graph *irg);
 337 typedef void (*transform_irp_func)(void);
 338
 339 typedef struct {
 340         opt_target_t  target;
 341         const char   *name;
 342         union {
 343                 transform_irg_func transform_irg;
 344                 transform_irp_func transform_irp;
 345         } u;
 346         const char   *description;
 347         opt_flags_t   flags;
 348         ir_timer_t   *timer;
 349 } opt_config_t;
 350
 351 static opt_config_t opts[] = {
 352 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 353 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 354         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 355         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 356         IRG("confirm",           construct_confirms,       "confirm optimization",                                  OPT_FLAG_HIDE_OPTIONS),
 357         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 358         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 359         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 360         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 361         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 362         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 363         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 364         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 365         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 366         IRG("local",             local_opts,               "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 367         IRG("lower",             lower_highlevel_graph,    "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 368         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 369         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 370         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 371         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 372         IRG("gcse",              do_gcse,                  "global common subexpression eliminiation",              OPT_FLAG_NONE),
 373         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 374         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 375         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 376         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 377         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 378         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 379         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 380         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 381         IRG("vrp",               set_vrp_data,             "value range propagation",                               OPT_FLAG_NONE),
 382         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 383         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),
 384         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 385         IRP("opt-func-call",     optimize_funccalls,       "function call optimization",                            OPT_FLAG_NONE),
 386         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 387         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 388         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_NONE),
 389         IRP("opt-cc",            mark_private_methods,     "calling conventions optimization",                      OPT_FLAG_NONE),
 390 #undef IRP
 391 #undef IRG
 392 };
 393
 394 #define FOR_EACH_OPT(i) for (opt_config_t *i = opts; i != endof(opts); ++i)
 395
 396 static opt_config_t *get_opt(const char *name)
 397 {
 398         FOR_EACH_OPT(config) {
 399                 if (streq(config->name, name))
 400                         return config;
 401         }
 402
 403         return NULL;
 404 }
 405
 406 static void set_opt_enabled(const char *name, bool enabled)
 407 {
 408         opt_config_t *config = get_opt(name);
 409         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 410                 | (enabled ? OPT_FLAG_ENABLED : 0);
 411 }
 412
 413 static bool get_opt_enabled(const char *name)
 414 {
 415         opt_config_t *config = get_opt(name);
 416         return (config->flags & OPT_FLAG_ENABLED) != 0;
 417 }
 418
 419 /**
 420  * perform an optimization on a single graph
 421  *
 422  * @return  true if something changed, false otherwise
 423  */
 424 static bool do_irg_opt(ir_graph *irg, const char *name)
 425 {
 426         opt_config_t *const config = get_opt(name);
 427         assert(config != NULL);
 428         assert(config->target == OPT_TARGET_IRG);
 429         if (! (config->flags & OPT_FLAG_ENABLED))
 430                 return false;
 431
 432         ir_graph *const old_irg = current_ir_graph;
 433         current_ir_graph = irg;
 434
 435         timer_push(config->timer);
 436         config->u.transform_irg(irg);
 437         timer_pop(config->timer);
 438
 439         if (firm_dump.all_phases && firm_dump.ir_graph) {
 440                 dump_ir_graph(irg, name);
 441         }
 442
 443         if (firm_opt.check_all) {
 444                 timer_push(t_verify);
 445                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 446                 timer_pop(t_verify);
 447         }
 448
 449         current_ir_graph = old_irg;
 450         return true;
 451 }
 452
 453 static void do_irp_opt(const char *name)
 454 {
 455         opt_config_t *const config = get_opt(name);
 456         assert(config->target == OPT_TARGET_IRP);
 457         if (! (config->flags & OPT_FLAG_ENABLED))
 458                 return;
 459
 460         timer_push(config->timer);
 461         config->u.transform_irp();
 462         timer_pop(config->timer);
 463
 464         if (firm_dump.ir_graph && firm_dump.all_phases) {
 465                 int i;
 466                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 467                         ir_graph *irg = get_irp_irg(i);
 468                         dump_ir_graph(irg, name);
 469                 }
 470         }
 471
 472         if (firm_opt.check_all) {
 473                 int i;
 474                 timer_push(t_verify);
 475                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 476                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 477                 }
 478                 timer_pop(t_verify);
 479         }
 480 }
 481
 482 /**
 483  * Enable transformations which should be always safe (and cheap) to perform
 484  */
 485 static void enable_safe_defaults(void)
 486 {
 487         set_opt_enabled("remove-unused", true);
 488         set_opt_enabled("opt-tail-rec", true);
 489         set_opt_enabled("opt-func-call", true);
 490         set_opt_enabled("reassociation", true);
 491         set_opt_enabled("control-flow", true);
 492         set_opt_enabled("local", true);
 493         set_opt_enabled("lower-const", true);
 494         set_opt_enabled("scalar-replace", true);
 495         set_opt_enabled("place", true);
 496         set_opt_enabled("gcse", true);
 497         set_opt_enabled("confirm", true);
 498         set_opt_enabled("opt-load-store", true);
 499         set_opt_enabled("lower", true);
 500         set_opt_enabled("deconv", true);
 501         set_opt_enabled("remove-confirms", true);
 502         set_opt_enabled("ivopts", true);
 503         set_opt_enabled("dead", true);
 504         set_opt_enabled("remove-phi-cycles", true);
 505         set_opt_enabled("frame", true);
 506         set_opt_enabled("combo", true);
 507         set_opt_enabled("invert-loops", true);
 508         set_opt_enabled("target-lowering", true);
 509         set_opt_enabled("rts", true);
 510         set_opt_enabled("parallelize-mem", true);
 511         set_opt_enabled("opt-cc", true);
 512 }
 513
 514 /**
 515  * run all the Firm optimizations
 516  *
 517  * @param input_filename     the name of the (main) source file
 518  */
 519 static void do_firm_optimizations(const char *input_filename)
 520 {
 521         size_t   i;
 522         unsigned aa_opt;
 523
 524         set_opt_alias_analysis(firm_opt.alias_analysis);
 525
 526         aa_opt = aa_opt_no_opt;
 527         if (firm_opt.strict_alias)
 528                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 529         if (firm_opt.no_alias)
 530                 aa_opt = aa_opt_no_alias;
 531
 532         set_irp_memory_disambiguator_options(aa_opt);
 533
 534         /* parameter passing code should set them directly sometime... */
 535         set_opt_enabled("confirm", firm_opt.confirm);
 536         set_opt_enabled("remove-confirms", firm_opt.confirm);
 537
 538         /* osr supersedes remove_phi_cycles */
 539         if (get_opt_enabled("ivopts"))
 540                 set_opt_enabled("remove-phi-cycles", false);
 541
 542         timer_start(t_all_opt);
 543
 544         do_irp_opt("rts");
 545
 546         /* first step: kill dead code */
 547         for (i = 0; i < get_irp_n_irgs(); i++) {
 548                 ir_graph *irg = get_irp_irg(i);
 549                 do_irg_opt(irg, "combo");
 550                 do_irg_opt(irg, "local");
 551                 do_irg_opt(irg, "control-flow");
 552         }
 553
 554         do_irp_opt("remove-unused");
 555         for (i = 0; i < get_irp_n_irgs(); ++i) {
 556                 ir_graph *irg = get_irp_irg(i);
 557                 do_irg_opt(irg, "opt-tail-rec");
 558         }
 559         do_irp_opt("opt-func-call");
 560         do_irp_opt("lower-const");
 561
 562         for (i = 0; i < get_irp_n_irgs(); i++) {
 563                 ir_graph *irg = get_irp_irg(i);
 564
 565                 do_irg_opt(irg, "scalar-replace");
 566                 do_irg_opt(irg, "invert-loops");
 567                 do_irg_opt(irg, "unroll-loops");
 568                 do_irg_opt(irg, "local");
 569                 do_irg_opt(irg, "reassociation");
 570                 do_irg_opt(irg, "local");
 571                 do_irg_opt(irg, "gcse");
 572                 do_irg_opt(irg, "place");
 573
 574                 if (firm_opt.confirm) {
 575                         /* Confirm construction currently can only handle blocks with only
 576                            one control flow predecessor. Calling optimize_cf here removes
 577                            Bad predecessors and help the optimization of switch constructs.
 578                          */
 579                         do_irg_opt(irg, "control-flow");
 580                         do_irg_opt(irg, "confirm");
 581                         do_irg_opt(irg, "vrp");
 582                         do_irg_opt(irg, "local");
 583                 }
 584
 585                 do_irg_opt(irg, "control-flow");
 586                 do_irg_opt(irg, "opt-load-store");
 587                 do_irg_opt(irg, "fp-vrp");
 588                 do_irg_opt(irg, "lower");
 589                 do_irg_opt(irg, "deconv");
 590                 do_irg_opt(irg, "thread-jumps");
 591                 do_irg_opt(irg, "remove-confirms");
 592                 do_irg_opt(irg, "gvn-pre");
 593                 do_irg_opt(irg, "gcse");
 594                 do_irg_opt(irg, "place");
 595                 do_irg_opt(irg, "control-flow");
 596
 597                 if (do_irg_opt(irg, "if-conversion")) {
 598                         do_irg_opt(irg, "local");
 599                         do_irg_opt(irg, "control-flow");
 600                 }
 601                 /* this doesn't make too much sense but tests the mux destruction... */
 602                 do_irg_opt(irg, "lower-mux");
 603
 604                 do_irg_opt(irg, "bool");
 605                 do_irg_opt(irg, "shape-blocks");
 606                 do_irg_opt(irg, "ivopts");
 607                 do_irg_opt(irg, "local");
 608                 do_irg_opt(irg, "dead");
 609         }
 610
 611         do_irp_opt("inline");
 612         do_irp_opt("opt-proc-clone");
 613
 614         for (i = 0; i < get_irp_n_irgs(); i++) {
 615                 ir_graph *irg = get_irp_irg(i);
 616                 do_irg_opt(irg, "local");
 617                 do_irg_opt(irg, "control-flow");
 618                 do_irg_opt(irg, "thread-jumps");
 619                 do_irg_opt(irg, "local");
 620                 do_irg_opt(irg, "control-flow");
 621
 622                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 623                         do_irg_opt(irg, "local");
 624                         do_irg_opt(irg, "vrp");
 625                         do_irg_opt(irg, "local");
 626                         do_irg_opt(irg, "vrp");
 627                 }
 628         }
 629
 630         if (firm_dump.ir_graph) {
 631                 /* recompute backedges for nicer dumps */
 632                 for (i = 0; i < get_irp_n_irgs(); i++)
 633                         construct_cf_backedges(get_irp_irg(i));
 634         }
 635
 636         dump_all("opt");
 637
 638         if (firm_dump.statistic & STAT_AFTER_OPT)
 639                 stat_dump_snapshot(input_filename, "opt");
 640
 641         timer_stop(t_all_opt);
 642 }
 643
 644 /**
 645  * do Firm lowering
 646  *
 647  * @param input_filename  the name of the (main) source file
 648  */
 649 static void do_firm_lowering(const char *input_filename)
 650 {
 651         int i;
 652
 653         /* enable architecture dependent optimizations */
 654         arch_dep_set_opts((arch_dep_opts_t)
 655                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 656                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 657                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 658         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 659                 ir_graph *irg = get_irp_irg(i);
 660                 do_irg_opt(irg, "reassociation");
 661                 do_irg_opt(irg, "local");
 662         }
 663
 664         do_irp_opt("target-lowering");
 665
 666         if (firm_dump.statistic & STAT_AFTER_LOWER)
 667                 stat_dump_snapshot(input_filename, "low");
 668
 669         timer_start(t_all_opt);
 670
 671         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 672                 ir_graph *irg = get_irp_irg(i);
 673
 674                 do_irg_opt(irg, "local");
 675                 do_irg_opt(irg, "deconv");
 676                 do_irg_opt(irg, "control-flow");
 677                 do_irg_opt(irg, "opt-load-store");
 678                 do_irg_opt(irg, "gcse");
 679                 do_irg_opt(irg, "place");
 680                 do_irg_opt(irg, "control-flow");
 681
 682                 if (do_irg_opt(irg, "vrp")) {
 683                         do_irg_opt(irg, "local");
 684                         do_irg_opt(irg, "control-flow");
 685                         do_irg_opt(irg, "vrp");
 686                         do_irg_opt(irg, "local");
 687                         do_irg_opt(irg, "control-flow");
 688                 }
 689
 690                 if (do_irg_opt(irg, "if-conversion")) {
 691                         do_irg_opt(irg, "local");
 692                         do_irg_opt(irg, "control-flow");
 693                 }
 694
 695                 set_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2);
 696                 do_irg_opt(irg, "local");
 697
 698                 do_irg_opt(irg, "parallelize-mem");
 699                 do_irg_opt(irg, "frame");
 700         }
 701         do_irp_opt("remove-unused");
 702         do_irp_opt("opt-cc");
 703         timer_stop(t_all_opt);
 704         dump_all("low-opt");
 705
 706         if (firm_dump.statistic & STAT_FINAL) {
 707                 stat_dump_snapshot(input_filename, "final");
 708         }
 709 }
 710
 711 /**
 712  * Initialize for the Firm-generating back end.
 713  */
 714 void gen_firm_init(void)
 715 {
 716         ir_init();
 717         enable_safe_defaults();
 718
 719         FOR_EACH_OPT(i) {
 720                 i->timer = ir_timer_new();
 721                 timer_register(i->timer, i->description);
 722         }
 723         t_verify = ir_timer_new();
 724         timer_register(t_verify, "Firm: verify pass");
 725         t_vcg_dump = ir_timer_new();
 726         timer_register(t_vcg_dump, "Firm: vcg dumping");
 727         t_all_opt = ir_timer_new();
 728         timer_register(t_all_opt, "Firm: all optimizations");
 729         t_backend = ir_timer_new();
 730         timer_register(t_backend, "Firm: backend");
 731 }
 732
 733 static void init_statistics(void)
 734 {
 735         unsigned pattern = 0;
 736
 737         if (firm_dump.stat_pattern)
 738                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 739
 740         if (firm_dump.stat_dag)
 741                 pattern |= FIRMSTAT_COUNT_DAG;
 742
 743         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 744                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 745                         | FIRMSTAT_COUNT_CONSTS | pattern);
 746 }
 747
 748 /**
 749  * Called, after the Firm generation is completed,
 750  * do all optimizations and backend call here.
 751  *
 752  * @param out                a file handle for the output, may be NULL
 753  * @param input_filename     the name of the (main) source file
 754  */
 755 void generate_code(FILE *out, const char *input_filename)
 756 {
 757         int i;
 758
 759         set_optimize(1);
 760         set_opt_constant_folding(firm_opt.const_folding);
 761         set_opt_algebraic_simplification(firm_opt.const_folding);
 762         set_opt_cse(firm_opt.cse);
 763         set_opt_global_cse(0);
 764
 765         init_statistics();
 766
 767         do_node_verification((firm_verification_t) firm_opt.verify);
 768
 769         /* the general for dumping option must be set, or the others will not work*/
 770         firm_dump.ir_graph = (bool) (firm_dump.ir_graph | firm_dump.all_phases);
 771
 772         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 773                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 774         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 775
 776         /* FIXME: cloning might ADD new graphs. */
 777         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 778
 779         if (firm_dump.all_types) {
 780                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 781         }
 782
 783         dump_all("");
 784
 785         timer_push(t_verify);
 786         tr_verify();
 787         timer_pop(t_verify);
 788
 789         /* all graphs are finalized, set the irp phase to high */
 790         set_irp_phase_state(phase_high);
 791
 792         /* BEWARE: kill unreachable code before doing compound lowering */
 793         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 794                 ir_graph *irg = get_irp_irg(i);
 795                 do_irg_opt(irg, "control-flow");
 796         }
 797
 798         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 799                 stat_dump_snapshot(input_filename, "noopt");
 800         }
 801
 802         do_firm_optimizations(input_filename);
 803         do_firm_lowering(input_filename);
 804
 805         /* set the phase to low */
 806         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 807                 set_irg_phase_state(get_irp_irg(i), phase_low);
 808
 809         if (firm_dump.statistic & STAT_FINAL_IR)
 810                 stat_dump_snapshot(input_filename, "final-ir");
 811
 812         /* run the code generator */
 813         timer_start(t_backend);
 814         be_main(out, input_filename);
 815         timer_stop(t_backend);
 816
 817         if (firm_dump.statistic & STAT_FINAL)
 818                 stat_dump_snapshot(input_filename, "final");
 819 }
 820
 821 void gen_firm_finish(void)
 822 {
 823         ir_finish();
 824 }
 825
 826 static void disable_all_opts(void)
 827 {
 828         firm_opt.cse             = false;
 829         firm_opt.confirm         = false;
 830         firm_opt.muls            = false;
 831         firm_opt.divs            = false;
 832         firm_opt.mods            = false;
 833         firm_opt.alias_analysis  = false;
 834         firm_opt.strict_alias    = false;
 835         firm_opt.no_alias        = false;
 836
 837         FOR_EACH_OPT(config) {
 838                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 839                         config->flags |= OPT_FLAG_ENABLED;
 840                 } else {
 841                         config->flags &= ~OPT_FLAG_ENABLED;
 842                 }
 843         }
 844 }
 845
 846 static bool firm_opt_option(const char *opt)
 847 {
 848         char const* const rest   = strstart(opt, "no-");
 849         bool        const enable = rest ? opt = rest, false : true;
 850
 851         opt_config_t *config = get_opt(opt);
 852         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 853                 return false;
 854
 855         config->flags &= ~OPT_FLAG_ENABLED;
 856         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 857         return true;
 858 }
 859
 860 void firm_option_help(print_option_help_func print_option_help)
 861 {
 862         print_option_help(firm_options[0].option, firm_options[0].description);
 863
 864         FOR_EACH_OPT(config) {
 865                 char buf[1024];
 866                 char buf2[1024];
 867
 868                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 869                         continue;
 870
 871                 snprintf(buf, sizeof(buf), "-f%s", config->name);
 872                 snprintf(buf2, sizeof(buf2), "enable %s", config->description);
 873                 print_option_help(buf, buf2);
 874                 snprintf(buf, sizeof(buf), "-fno-%s", config->name);
 875                 snprintf(buf2, sizeof(buf2), "disable %s", config->description);
 876                 print_option_help(buf, buf2);
 877         }
 878
 879         for (size_t k = 0; k != lengthof(firm_options); ++k) {
 880                 char buf[1024];
 881                 char buf2[1024];
 882                 snprintf(buf, sizeof(buf), "-f%s", firm_options[k].option);
 883                 snprintf(buf2, sizeof(buf2), "%s", firm_options[k].description);
 884                 print_option_help(buf, buf2);
 885         }
 886 }
 887
 888 int firm_option(const char *const opt)
 889 {
 890         char const* val;
 891         if ((val = strstart(opt, "dump-filter="))) {
 892                 ir_set_dump_filter(val);
 893                 return 1;
 894         } else if ((val = strstart(opt, "clone-threshold="))) {
 895                 sscanf(val, "%d", &firm_opt.clone_threshold);
 896                 return 1;
 897         } else if ((val = strstart(opt, "inline-max-size="))) {
 898                 sscanf(val, "%u", &firm_opt.inline_maxsize);
 899                 return 1;
 900         } else if ((val = strstart(opt, "inline-threshold="))) {
 901                 sscanf(val, "%u", &firm_opt.inline_threshold);
 902                 return 1;
 903         } else if (streq(opt, "no-opt")) {
 904                 disable_all_opts();
 905                 return 1;
 906         }
 907
 908         size_t const len = strlen(opt);
 909         for (size_t i = lengthof(firm_options); i != 0;) {
 910                 struct params const* const o = &firm_options[--i];
 911                 if (len == o->opt_len && memcmp(opt, o->option, len) == 0) {
 912                         /* statistic options do accumulate */
 913                         if (o->flag == &firm_dump.statistic)
 914                                 *o->flag = (bool) (*o->flag | o->set);
 915                         else
 916                                 *o->flag = o->set;
 917
 918                         return 1;
 919                 }
 920         }
 921
 922         /* maybe this enables/disables optimizations */
 923         if (firm_opt_option(opt))
 924                 return 1;
 925
 926         return 0;
 927 }
 928
 929 static void set_be_option(const char *arg)
 930 {
 931         int res = be_parse_arg(arg);
 932         (void) res;
 933         assert(res);
 934 }
 935
 936 static void set_option(const char *arg)
 937 {
 938         int res = firm_option(arg);
 939         (void) res;
 940         assert(res);
 941 }
 942
 943 void choose_optimization_pack(int level)
 944 {
 945         /* apply optimization level */
 946         switch(level) {
 947         case 0:
 948                 set_option("no-opt");
 949                 break;
 950         case 1:
 951                 set_option("no-inline");
 952                 break;
 953         default:
 954         case 4:
 955                 /* use_builtins = true; */
 956                 /* fallthrough */
 957         case 3:
 958                 set_option("thread-jumps");
 959                 set_option("if-conversion");
 960                 /* fallthrough */
 961         case 2:
 962                 set_option("strict-aliasing");
 963                 set_option("inline");
 964                 set_option("fp-vrp");
 965                 set_option("deconv");
 966                 set_be_option("omitfp");
 967                 break;
 968         }
 969 }