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

   1 /*
   2  * This file is part of cparser.
   3  * Copyright (C) 2012 Michael Beck <mm.beck@gmx.net>
   4  */
   5
   6 /**
   7  * @file
   8  * @author Michael Beck, Matthias Braun
   9  * @brief Firm-generating back end optimizations.
  10  */
  11 #include <config.h>
  12
  13 #include <stdbool.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <stdbool.h>
  17 #include <assert.h>
  18 #include <libfirm/firm.h>
  19
  20 #include "firm_opt.h"
  21 #include "firm_timing.h"
  22 #include "ast2firm.h"
  23 #include "adt/strutil.h"
  24 #include "adt/util.h"
  25
  26 /* optimization settings */
  27 struct a_firm_opt {
  28         bool     const_folding;   /**< enable constant folding */
  29         bool     cse;             /**< enable common-subexpression elimination */
  30         bool     confirm;         /**< enable Confirm optimization */
  31         bool     muls;            /**< enable architecture dependent mul optimization */
  32         bool     divs;            /**< enable architecture dependent div optimization */
  33         bool     mods;            /**< enable architecture dependent mod optimization */
  34         bool     alias_analysis;  /**< enable Alias Analysis */
  35         bool     strict_alias;    /**< enable strict Alias Analysis (using type based AA) */
  36         bool     no_alias;        /**< no aliasing possible. */
  37         bool     verify;          /**< Firm verifier setting */
  38         bool     check_all;       /**< enable checking all Firm phases */
  39         int      clone_threshold; /**< The threshold value for procedure cloning. */
  40         unsigned inline_maxsize;  /**< Maximum function size for inlining. */
  41         unsigned inline_threshold;/**< Inlining benefice threshold. */
  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 ir_graph;      /**< dump all graphs */
  59         bool all_phases;    /**< dump the IR graph after all phases */
  60         bool statistic;     /**< Firm statistic setting */
  61 };
  62
  63 struct a_firm_be_opt {
  64         bool selection;
  65         bool node_stat;
  66 };
  67
  68 /* optimization settings */
  69 static struct a_firm_opt firm_opt = {
  70         .const_folding    =  true,
  71         .cse              =  true,
  72         .confirm          =  true,
  73         .muls             =  true,
  74         .divs             =  true,
  75         .mods             =  true,
  76         .alias_analysis   =  true,
  77         .strict_alias     =  false,
  78         .no_alias         =  false,
  79         .verify           =  FIRM_VERIFICATION_ON,
  80         .check_all        =  true,
  81         .clone_threshold  =  DEFAULT_CLONE_THRESHOLD,
  82         .inline_maxsize   =  750,
  83         .inline_threshold =  0,
  84 };
  85
  86 /* dumping options */
  87 static struct a_firm_dump firm_dump = {
  88         .debug_print  = false,
  89         .all_types    = false,
  90         .ir_graph     = false,
  91         .all_phases   = false,
  92         .statistic    = STAT_NONE,
  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
 134   /* dumping */
 135   { X("dump-ir"),                &firm_dump.ir_graph,        1, "dump IR graph" },
 136   { X("dump-all-types"),         &firm_dump.all_types,       1, "dump graph of all types" },
 137   { X("dump-all-phases"),        &firm_dump.all_phases,      1, "dump graphs for all optimization phases" },
 138   { X("dump-filter=<string>"),   NULL,                       0, "set dumper filter" },
 139
 140   /* misc */
 141   { X("stat-before-opt"),        &firm_dump.statistic,       STAT_BEFORE_OPT,  "Firm statistic output before optimizations" },
 142   { X("stat-after-opt"),         &firm_dump.statistic,       STAT_AFTER_OPT,   "Firm statistic output after optimizations" },
 143   { X("stat-after-lower"),       &firm_dump.statistic,       STAT_AFTER_LOWER, "Firm statistic output after lowering" },
 144   { X("stat-final-ir"),          &firm_dump.statistic,       STAT_FINAL_IR,    "Firm statistic after final optimization" },
 145   { X("stat-final"),             &firm_dump.statistic,       STAT_FINAL,       "Firm statistic after code generation" },
 146 };
 147
 148 #undef X
 149
 150 static ir_timer_t *t_vcg_dump;
 151 static ir_timer_t *t_verify;
 152 static ir_timer_t *t_all_opt;
 153 static ir_timer_t *t_backend;
 154 static bool do_irg_opt(ir_graph *irg, const char *name);
 155
 156 /** dump all the graphs depending on cond */
 157
 158 static void dump_all(const char *suffix)
 159 {
 160         if (!firm_dump.ir_graph)
 161                 return;
 162
 163         timer_push(t_vcg_dump);
 164         dump_all_ir_graphs(suffix);
 165         timer_pop(t_vcg_dump);
 166 }
 167
 168 /* entities of runtime functions */
 169 ir_entity *rts_entities[rts_max];
 170
 171 /**
 172  * Map runtime functions.
 173  */
 174 static void rts_map(void)
 175 {
 176         static const struct {
 177                 ir_entity   **ent; /**< address of the rts entity */
 178                 i_mapper_func func; /**< mapper function. */
 179         } mapper[] = {
 180                 /* integer */
 181                 { &rts_entities[rts_abs],     i_mapper_abs },
 182                 { &rts_entities[rts_labs],    i_mapper_abs },
 183                 { &rts_entities[rts_llabs],   i_mapper_abs },
 184                 { &rts_entities[rts_imaxabs], i_mapper_abs },
 185
 186                 /* double -> double */
 187                 { &rts_entities[rts_fabs],    i_mapper_abs },
 188                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
 189                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
 190                 { &rts_entities[rts_pow],     i_mapper_pow },
 191                 { &rts_entities[rts_exp],     i_mapper_exp },
 192                 { &rts_entities[rts_exp2],    i_mapper_exp },
 193                 { &rts_entities[rts_exp10],   i_mapper_exp },
 194                 { &rts_entities[rts_log],     i_mapper_log },
 195                 { &rts_entities[rts_log2],    i_mapper_log2 },
 196                 { &rts_entities[rts_log10],   i_mapper_log10 },
 197                 { &rts_entities[rts_sin],     i_mapper_sin },
 198                 { &rts_entities[rts_cos],     i_mapper_cos },
 199                 { &rts_entities[rts_tan],     i_mapper_tan },
 200                 { &rts_entities[rts_asin],    i_mapper_asin },
 201                 { &rts_entities[rts_acos],    i_mapper_acos },
 202                 { &rts_entities[rts_atan],    i_mapper_atan },
 203                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 204                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 205                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 206
 207                 /* float -> float */
 208                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 209                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 210                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 211                 { &rts_entities[rts_powf],    i_mapper_pow },
 212                 { &rts_entities[rts_expf],    i_mapper_exp },
 213                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 214                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 215                 { &rts_entities[rts_logf],    i_mapper_log },
 216                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 217                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 218                 { &rts_entities[rts_sinf],    i_mapper_sin },
 219                 { &rts_entities[rts_cosf],    i_mapper_cos },
 220                 { &rts_entities[rts_tanf],    i_mapper_tan },
 221                 { &rts_entities[rts_asinf],   i_mapper_asin },
 222                 { &rts_entities[rts_acosf],   i_mapper_acos },
 223                 { &rts_entities[rts_atanf],   i_mapper_atan },
 224                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 225                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 226                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 227
 228                 /* long double -> long double */
 229                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 230                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 231                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 232                 { &rts_entities[rts_powl],    i_mapper_pow },
 233                 { &rts_entities[rts_expl],    i_mapper_exp },
 234                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 235                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 236                 { &rts_entities[rts_logl],    i_mapper_log },
 237                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 238                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 239                 { &rts_entities[rts_sinl],    i_mapper_sin },
 240                 { &rts_entities[rts_cosl],    i_mapper_cos },
 241                 { &rts_entities[rts_tanl],    i_mapper_tan },
 242                 { &rts_entities[rts_asinl],   i_mapper_asin },
 243                 { &rts_entities[rts_acosl],   i_mapper_acos },
 244                 { &rts_entities[rts_atanl],   i_mapper_atan },
 245                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 246                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 247                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 248
 249                 /* string */
 250                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 251                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 252                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 253                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 254                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 255                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 256                 { &rts_entities[rts_memmove], i_mapper_memmove },
 257                 { &rts_entities[rts_memset],  i_mapper_memset },
 258                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 259         };
 260         i_record rec[lengthof(mapper)];
 261         size_t   n_map = 0;
 262
 263         for (size_t i = 0; i != lengthof(mapper); ++i) {
 264                 if (*mapper[i].ent != NULL) {
 265                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 266                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 267                         rec[n_map].i_call.i_mapper = mapper[i].func;
 268                         rec[n_map].i_call.ctx      = NULL;
 269                         rec[n_map].i_call.link     = NULL;
 270                         ++n_map;
 271                 }
 272         }
 273
 274         if (n_map > 0)
 275                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 276 }
 277
 278 static int *irg_dump_no;
 279
 280 typedef enum opt_target {
 281         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 282         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 283 } opt_target_t;
 284
 285 typedef enum opt_flags {
 286         OPT_FLAG_NONE         = 0,
 287         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 288         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 289         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 290         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 291                                              -foptions for this transformation */
 292         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 293                                              so we need it even with -O0 */
 294 } opt_flags_t;
 295
 296 typedef void (*transform_irg_func)(ir_graph *irg);
 297 typedef void (*transform_irp_func)(void);
 298
 299 typedef struct {
 300         opt_target_t  target;
 301         const char   *name;
 302         union {
 303                 transform_irg_func transform_irg;
 304                 transform_irp_func transform_irp;
 305         } u;
 306         const char   *description;
 307         opt_flags_t   flags;
 308         ir_timer_t   *timer;
 309 } opt_config_t;
 310
 311 static opt_config_t *get_opt(const char *name);
 312
 313 static void do_stred(ir_graph *irg)
 314 {
 315         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 316 }
 317
 318 static void after_inline_opt(ir_graph *irg)
 319 {
 320         opt_config_t *const config = get_opt("inline");
 321         timer_stop(config->timer);
 322
 323         do_irg_opt(irg, "scalar-replace");
 324         do_irg_opt(irg, "local");
 325         do_irg_opt(irg, "control-flow");
 326         do_irg_opt(irg, "combo");
 327
 328         timer_start(config->timer);
 329 }
 330
 331 static void do_inline(void)
 332 {
 333         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 334                          after_inline_opt);
 335 }
 336
 337 static void do_cloning(void)
 338 {
 339         proc_cloning((float) firm_opt.clone_threshold);
 340 }
 341
 342 static void do_lower_mux(ir_graph *irg)
 343 {
 344         lower_mux(irg, NULL);
 345 }
 346
 347 static void do_gcse(ir_graph *irg)
 348 {
 349         set_opt_global_cse(1);
 350         optimize_graph_df(irg);
 351         set_opt_global_cse(0);
 352 }
 353
 354 static opt_config_t opts[] = {
 355 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 356 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 357         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 358         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 359         IRG("confirm",           construct_confirms,       "confirm optimization",                                  OPT_FLAG_HIDE_OPTIONS),
 360         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 361         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 362         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 363         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 364         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 365         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 366         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 367         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 368         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 369         IRG("local",             local_opts,               "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 370         IRG("lower",             lower_highlevel_graph,    "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 371         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 372         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 373         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 374         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 375         IRG("gcse",              do_gcse,                  "global common subexpression eliminiation",              OPT_FLAG_NONE),
 376         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 377         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 378         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 379         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 380         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 381         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 382         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 383         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 384         IRG("vrp",               set_vrp_data,             "value range propagation",                               OPT_FLAG_NONE),
 385         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 386         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),
 387         IRP("local-const",       local_opts_const_code,    "local optimisation of constant initializers",
 388                                     OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY | OPT_FLAG_ESSENTIAL),
 389         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 390         IRP("opt-func-call",     optimize_funccalls,       "function call optimization",                            OPT_FLAG_NONE),
 391         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 392         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 393         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_NONE),
 394         IRP("opt-cc",            mark_private_methods,     "calling conventions optimization",                      OPT_FLAG_NONE),
 395 #undef IRP
 396 #undef IRG
 397 };
 398
 399 #define FOR_EACH_OPT(i) for (opt_config_t *i = opts; i != endof(opts); ++i)
 400
 401 static opt_config_t *get_opt(const char *name)
 402 {
 403         FOR_EACH_OPT(config) {
 404                 if (streq(config->name, name))
 405                         return config;
 406         }
 407
 408         return NULL;
 409 }
 410
 411 static void set_opt_enabled(const char *name, bool enabled)
 412 {
 413         opt_config_t *config = get_opt(name);
 414         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 415                 | (enabled ? OPT_FLAG_ENABLED : 0);
 416 }
 417
 418 static bool get_opt_enabled(const char *name)
 419 {
 420         opt_config_t *config = get_opt(name);
 421         return (config->flags & OPT_FLAG_ENABLED) != 0;
 422 }
 423
 424 /**
 425  * perform an optimization on a single graph
 426  *
 427  * @return  true if something changed, false otherwise
 428  */
 429 static bool do_irg_opt(ir_graph *irg, const char *name)
 430 {
 431         opt_config_t *const config = get_opt(name);
 432         assert(config != NULL);
 433         assert(config->target == OPT_TARGET_IRG);
 434         if (! (config->flags & OPT_FLAG_ENABLED))
 435                 return false;
 436
 437         ir_graph *const old_irg = current_ir_graph;
 438         current_ir_graph = irg;
 439
 440         timer_start(config->timer);
 441         config->u.transform_irg(irg);
 442         timer_stop(config->timer);
 443
 444         if (firm_dump.all_phases && firm_dump.ir_graph) {
 445                 dump_ir_graph(irg, name);
 446         }
 447
 448         if (firm_opt.verify) {
 449                 timer_push(t_verify);
 450                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 451                 timer_pop(t_verify);
 452         }
 453
 454         current_ir_graph = old_irg;
 455         return true;
 456 }
 457
 458 static void do_irp_opt(const char *name)
 459 {
 460         opt_config_t *const config = get_opt(name);
 461         assert(config->target == OPT_TARGET_IRP);
 462         if (! (config->flags & OPT_FLAG_ENABLED))
 463                 return;
 464
 465         timer_start(config->timer);
 466         config->u.transform_irp();
 467         timer_stop(config->timer);
 468
 469         if (firm_dump.ir_graph && firm_dump.all_phases) {
 470                 int i;
 471                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 472                         ir_graph *irg = get_irp_irg(i);
 473                         dump_ir_graph(irg, name);
 474                 }
 475         }
 476
 477         if (firm_opt.verify) {
 478                 int i;
 479                 timer_push(t_verify);
 480                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 481                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 482                 }
 483                 timer_pop(t_verify);
 484         }
 485 }
 486
 487 /**
 488  * Enable transformations which should be always safe (and cheap) to perform
 489  */
 490 static void enable_safe_defaults(void)
 491 {
 492         set_opt_enabled("remove-unused", true);
 493         set_opt_enabled("opt-tail-rec", true);
 494         set_opt_enabled("opt-func-call", true);
 495         set_opt_enabled("reassociation", true);
 496         set_opt_enabled("control-flow", true);
 497         set_opt_enabled("local", true);
 498         set_opt_enabled("lower-const", true);
 499         set_opt_enabled("local-const", true);
 500         set_opt_enabled("scalar-replace", true);
 501         set_opt_enabled("place", true);
 502         set_opt_enabled("gcse", true);
 503         set_opt_enabled("confirm", true);
 504         set_opt_enabled("opt-load-store", true);
 505         set_opt_enabled("lower", true);
 506         set_opt_enabled("deconv", true);
 507         set_opt_enabled("remove-confirms", true);
 508         set_opt_enabled("ivopts", true);
 509         set_opt_enabled("dead", true);
 510         set_opt_enabled("remove-phi-cycles", true);
 511         set_opt_enabled("frame", true);
 512         set_opt_enabled("combo", true);
 513         set_opt_enabled("invert-loops", true);
 514         set_opt_enabled("target-lowering", true);
 515         set_opt_enabled("rts", true);
 516         set_opt_enabled("parallelize-mem", true);
 517         set_opt_enabled("opt-cc", true);
 518 }
 519
 520 /**
 521  * run all the Firm optimizations
 522  *
 523  * @param input_filename     the name of the (main) source file
 524  */
 525 static void do_firm_optimizations(const char *input_filename)
 526 {
 527         size_t   i;
 528         unsigned aa_opt;
 529
 530         set_opt_alias_analysis(firm_opt.alias_analysis);
 531
 532         aa_opt = aa_opt_no_opt;
 533         if (firm_opt.strict_alias)
 534                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 535         if (firm_opt.no_alias)
 536                 aa_opt = aa_opt_no_alias;
 537
 538         set_irp_memory_disambiguator_options(aa_opt);
 539
 540         /* parameter passing code should set them directly sometime... */
 541         set_opt_enabled("confirm", firm_opt.confirm);
 542         set_opt_enabled("remove-confirms", firm_opt.confirm);
 543
 544         /* osr supersedes remove_phi_cycles */
 545         if (get_opt_enabled("ivopts"))
 546                 set_opt_enabled("remove-phi-cycles", false);
 547
 548         do_irp_opt("rts");
 549
 550         /* first step: kill dead code */
 551         for (i = 0; i < get_irp_n_irgs(); i++) {
 552                 ir_graph *irg = get_irp_irg(i);
 553                 do_irg_opt(irg, "combo");
 554                 do_irg_opt(irg, "local");
 555                 do_irg_opt(irg, "control-flow");
 556         }
 557
 558         do_irp_opt("remove-unused");
 559         for (i = 0; i < get_irp_n_irgs(); ++i) {
 560                 ir_graph *irg = get_irp_irg(i);
 561                 do_irg_opt(irg, "opt-tail-rec");
 562         }
 563         do_irp_opt("opt-func-call");
 564         do_irp_opt("lower-const");
 565
 566         for (i = 0; i < get_irp_n_irgs(); i++) {
 567                 ir_graph *irg = get_irp_irg(i);
 568
 569                 do_irg_opt(irg, "scalar-replace");
 570                 do_irg_opt(irg, "invert-loops");
 571                 do_irg_opt(irg, "unroll-loops");
 572                 do_irg_opt(irg, "local");
 573                 do_irg_opt(irg, "reassociation");
 574                 do_irg_opt(irg, "local");
 575                 do_irg_opt(irg, "gcse");
 576                 do_irg_opt(irg, "place");
 577
 578                 if (firm_opt.confirm) {
 579                         /* Confirm construction currently can only handle blocks with only
 580                            one control flow predecessor. Calling optimize_cf here removes
 581                            Bad predecessors and help the optimization of switch constructs.
 582                          */
 583                         do_irg_opt(irg, "control-flow");
 584                         do_irg_opt(irg, "confirm");
 585                         do_irg_opt(irg, "vrp");
 586                         do_irg_opt(irg, "local");
 587                 }
 588
 589                 do_irg_opt(irg, "control-flow");
 590                 do_irg_opt(irg, "opt-load-store");
 591                 do_irg_opt(irg, "fp-vrp");
 592                 do_irg_opt(irg, "lower");
 593                 do_irg_opt(irg, "deconv");
 594                 do_irg_opt(irg, "thread-jumps");
 595                 do_irg_opt(irg, "remove-confirms");
 596                 do_irg_opt(irg, "gvn-pre");
 597                 do_irg_opt(irg, "gcse");
 598                 do_irg_opt(irg, "place");
 599                 do_irg_opt(irg, "control-flow");
 600
 601                 if (do_irg_opt(irg, "if-conversion")) {
 602                         do_irg_opt(irg, "local");
 603                         do_irg_opt(irg, "control-flow");
 604                 }
 605                 /* this doesn't make too much sense but tests the mux destruction... */
 606                 do_irg_opt(irg, "lower-mux");
 607
 608                 do_irg_opt(irg, "bool");
 609                 do_irg_opt(irg, "shape-blocks");
 610                 do_irg_opt(irg, "ivopts");
 611                 do_irg_opt(irg, "local");
 612                 do_irg_opt(irg, "dead");
 613         }
 614
 615         do_irp_opt("inline");
 616         do_irp_opt("opt-proc-clone");
 617
 618         for (i = 0; i < get_irp_n_irgs(); i++) {
 619                 ir_graph *irg = get_irp_irg(i);
 620                 do_irg_opt(irg, "local");
 621                 do_irg_opt(irg, "control-flow");
 622                 do_irg_opt(irg, "thread-jumps");
 623                 do_irg_opt(irg, "local");
 624                 do_irg_opt(irg, "control-flow");
 625
 626                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 627                         do_irg_opt(irg, "local");
 628                         do_irg_opt(irg, "vrp");
 629                         do_irg_opt(irg, "local");
 630                         do_irg_opt(irg, "vrp");
 631                 }
 632         }
 633
 634         if (firm_dump.ir_graph) {
 635                 /* recompute backedges for nicer dumps */
 636                 for (i = 0; i < get_irp_n_irgs(); i++)
 637                         construct_cf_backedges(get_irp_irg(i));
 638         }
 639
 640         dump_all("opt");
 641
 642         if (firm_dump.statistic & STAT_AFTER_OPT)
 643                 stat_dump_snapshot(input_filename, "opt");
 644 }
 645
 646 /**
 647  * do Firm lowering
 648  *
 649  * @param input_filename  the name of the (main) source file
 650  */
 651 static void do_firm_lowering(const char *input_filename)
 652 {
 653         int i;
 654
 655         /* enable architecture dependent optimizations */
 656         arch_dep_set_opts((arch_dep_opts_t)
 657                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 658                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 659                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 660         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 661                 ir_graph *irg = get_irp_irg(i);
 662                 do_irg_opt(irg, "reassociation");
 663                 do_irg_opt(irg, "local");
 664         }
 665
 666         do_irp_opt("target-lowering");
 667
 668         if (firm_dump.statistic & STAT_AFTER_LOWER)
 669                 stat_dump_snapshot(input_filename, "low");
 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                 add_irg_constraints(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2);
 696                 do_irg_opt(irg, "local");
 697
 698                 do_irg_opt(irg, "parallelize-mem");
 699                 do_irg_opt(irg, "frame");
 700         }
 701         /* hack so we get global initializers constant folded even at -O0 */
 702         set_opt_constant_folding(1);
 703         set_opt_algebraic_simplification(1);
 704         do_irp_opt("local-const");
 705         set_opt_constant_folding(firm_opt.const_folding);
 706         set_opt_algebraic_simplification(firm_opt.const_folding);
 707         do_irp_opt("remove-unused");
 708         do_irp_opt("opt-cc");
 709         dump_all("low-opt");
 710
 711         if (firm_dump.statistic & STAT_FINAL) {
 712                 stat_dump_snapshot(input_filename, "final");
 713         }
 714 }
 715
 716 /**
 717  * Initialize for the Firm-generating back end.
 718  */
 719 void gen_firm_init(void)
 720 {
 721         ir_init();
 722         enable_safe_defaults();
 723
 724         FOR_EACH_OPT(i) {
 725                 i->timer = ir_timer_new();
 726                 timer_register(i->timer, i->description);
 727         }
 728         t_verify = ir_timer_new();
 729         timer_register(t_verify, "Firm: verify pass");
 730         t_vcg_dump = ir_timer_new();
 731         timer_register(t_vcg_dump, "Firm: vcg dumping");
 732         t_all_opt = ir_timer_new();
 733         timer_register(t_all_opt, "Firm: all optimizations");
 734         t_backend = ir_timer_new();
 735         timer_register(t_backend, "Firm: backend");
 736 }
 737
 738 /**
 739  * Called, after the Firm generation is completed,
 740  * do all optimizations and backend call here.
 741  *
 742  * @param out                a file handle for the output, may be NULL
 743  * @param input_filename     the name of the (main) source file
 744  */
 745 void generate_code(FILE *out, const char *input_filename)
 746 {
 747         int i;
 748
 749         /* initialize implicit opts, just to be sure because really the frontend
 750          * should have called it already before starting graph construction */
 751         init_implicit_optimizations();
 752         firm_init_stat();
 753
 754         do_node_verification((firm_verification_t) firm_opt.verify);
 755
 756         /* the general for dumping option must be set, or the others will not work*/
 757         firm_dump.ir_graph = (bool) (firm_dump.ir_graph | firm_dump.all_phases);
 758
 759         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 760                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 761         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 762
 763         /* FIXME: cloning might ADD new graphs. */
 764         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 765
 766         ir_timer_init_parent(t_verify);
 767         ir_timer_init_parent(t_vcg_dump);
 768         timer_start(t_all_opt);
 769
 770         if (firm_dump.all_types) {
 771                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 772         }
 773
 774         dump_all("");
 775
 776         if (firm_opt.verify) {
 777                 timer_push(t_verify);
 778                 tr_verify();
 779                 timer_pop(t_verify);
 780         }
 781
 782         /* BEWARE: kill unreachable code before doing compound lowering */
 783         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 784                 ir_graph *irg = get_irp_irg(i);
 785                 do_irg_opt(irg, "control-flow");
 786         }
 787
 788         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 789                 stat_dump_snapshot(input_filename, "noopt");
 790         }
 791
 792         do_firm_optimizations(input_filename);
 793         do_firm_lowering(input_filename);
 794
 795         timer_stop(t_all_opt);
 796
 797         if (firm_dump.statistic & STAT_FINAL_IR)
 798                 stat_dump_snapshot(input_filename, "final-ir");
 799
 800         /* run the code generator */
 801         timer_start(t_backend);
 802         be_main(out, input_filename);
 803         timer_stop(t_backend);
 804
 805         if (firm_dump.statistic & STAT_FINAL)
 806                 stat_dump_snapshot(input_filename, "final");
 807 }
 808
 809 void gen_firm_finish(void)
 810 {
 811         ir_finish();
 812 }
 813
 814 static void disable_all_opts(void)
 815 {
 816         firm_opt.cse             = false;
 817         firm_opt.confirm         = false;
 818         firm_opt.muls            = false;
 819         firm_opt.divs            = false;
 820         firm_opt.mods            = false;
 821         firm_opt.alias_analysis  = false;
 822         firm_opt.strict_alias    = false;
 823         firm_opt.no_alias        = false;
 824         firm_opt.const_folding   = false;
 825
 826         FOR_EACH_OPT(config) {
 827                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 828                         config->flags |= OPT_FLAG_ENABLED;
 829                 } else {
 830                         config->flags &= ~OPT_FLAG_ENABLED;
 831                 }
 832         }
 833 }
 834
 835 static bool firm_opt_option(const char *opt)
 836 {
 837         char const* const rest   = strstart(opt, "no-");
 838         bool        const enable = rest ? opt = rest, false : true;
 839
 840         opt_config_t *config = get_opt(opt);
 841         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 842                 return false;
 843
 844         config->flags &= ~OPT_FLAG_ENABLED;
 845         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 846         return true;
 847 }
 848
 849 void firm_option_help(print_option_help_func print_option_help)
 850 {
 851         FOR_EACH_OPT(config) {
 852                 char buf[1024];
 853                 char buf2[1024];
 854
 855                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 856                         continue;
 857
 858                 snprintf(buf, sizeof(buf), "-f%s", config->name);
 859                 snprintf(buf2, sizeof(buf2), "enable %s", config->description);
 860                 print_option_help(buf, buf2);
 861                 snprintf(buf, sizeof(buf), "-fno-%s", config->name);
 862                 snprintf(buf2, sizeof(buf2), "disable %s", config->description);
 863                 print_option_help(buf, buf2);
 864         }
 865
 866         for (size_t k = 0; k != lengthof(firm_options); ++k) {
 867                 char buf[1024];
 868                 char buf2[1024];
 869                 snprintf(buf, sizeof(buf), "-f%s", firm_options[k].option);
 870                 snprintf(buf2, sizeof(buf2), "%s", firm_options[k].description);
 871                 print_option_help(buf, buf2);
 872         }
 873 }
 874
 875 int firm_option(const char *const opt)
 876 {
 877         char const* val;
 878         if ((val = strstart(opt, "dump-filter="))) {
 879                 ir_set_dump_filter(val);
 880                 return 1;
 881         } else if ((val = strstart(opt, "clone-threshold="))) {
 882                 sscanf(val, "%d", &firm_opt.clone_threshold);
 883                 return 1;
 884         } else if ((val = strstart(opt, "inline-max-size="))) {
 885                 sscanf(val, "%u", &firm_opt.inline_maxsize);
 886                 return 1;
 887         } else if ((val = strstart(opt, "inline-threshold="))) {
 888                 sscanf(val, "%u", &firm_opt.inline_threshold);
 889                 return 1;
 890         } else if (streq(opt, "no-opt")) {
 891                 disable_all_opts();
 892                 return 1;
 893         }
 894
 895         size_t const len = strlen(opt);
 896         for (size_t i = lengthof(firm_options); i != 0;) {
 897                 struct params const* const o = &firm_options[--i];
 898                 if (len == o->opt_len && memcmp(opt, o->option, len) == 0) {
 899                         /* statistic options do accumulate */
 900                         if (o->flag == &firm_dump.statistic)
 901                                 *o->flag = (bool) (*o->flag | o->set);
 902                         else
 903                                 *o->flag = o->set;
 904
 905                         return 1;
 906                 }
 907         }
 908
 909         /* maybe this enables/disables optimizations */
 910         if (firm_opt_option(opt))
 911                 return 1;
 912
 913         return 0;
 914 }
 915
 916 static void set_be_option(const char *arg)
 917 {
 918         int res = be_parse_arg(arg);
 919         (void) res;
 920         assert(res);
 921 }
 922
 923 static void set_option(const char *arg)
 924 {
 925         int res = firm_option(arg);
 926         (void) res;
 927         assert(res);
 928 }
 929
 930 void choose_optimization_pack(int level)
 931 {
 932         /* apply optimization level */
 933         switch(level) {
 934         case 0:
 935                 set_option("no-opt");
 936                 break;
 937         case 1:
 938                 set_option("no-inline");
 939                 break;
 940         default:
 941         case 4:
 942                 /* use_builtins = true; */
 943                 /* fallthrough */
 944         case 3:
 945                 set_option("thread-jumps");
 946                 set_option("if-conversion");
 947                 /* fallthrough */
 948         case 2:
 949                 set_option("strict-aliasing");
 950                 set_option("inline");
 951                 set_option("fp-vrp");
 952                 set_option("deconv");
 953                 set_be_option("omitfp");
 954                 break;
 955         }
 956 }
 957
 958 void init_implicit_optimizations(void)
 959 {
 960         set_optimize(1);
 961         set_opt_constant_folding(firm_opt.const_folding);
 962         set_opt_algebraic_simplification(firm_opt.const_folding);
 963         set_opt_cse(firm_opt.cse);
 964         set_opt_global_cse(0);
 965 }