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
  24 /* optimization settings */
  25 struct a_firm_opt {
  26         bool     const_folding;   /**< enable constant folding */
  27         bool     cse;             /**< enable common-subexpression elimination */
  28         bool     confirm;         /**< enable Confirm optimization */
  29         bool     muls;            /**< enable architecture dependent mul optimization */
  30         bool     divs;            /**< enable architecture dependent div optimization */
  31         bool     mods;            /**< enable architecture dependent mod optimization */
  32         bool     alias_analysis;  /**< enable Alias Analysis */
  33         bool     strict_alias;    /**< enable strict Alias Analysis (using type based AA) */
  34         bool     no_alias;        /**< no aliasing possible. */
  35         bool     verify;          /**< Firm verifier setting */
  36         bool     check_all;       /**< enable checking all Firm phases */
  37         int      clone_threshold; /**< The threshold value for procedure cloning. */
  38         unsigned inline_maxsize;  /**< Maximum function size for inlining. */
  39         unsigned inline_threshold;/**< Inlining benefice threshold. */
  40         bool     verify_edges;    /**< verify edges */
  41 };
  42
  43 /** statistic options */
  44 typedef enum a_firmstat_selection_tag {
  45         STAT_NONE        = 0x00000000,
  46         STAT_BEFORE_OPT  = 0x00000001,
  47         STAT_AFTER_OPT   = 0x00000002,
  48         STAT_AFTER_LOWER = 0x00000004,
  49         STAT_FINAL_IR    = 0x00000008,
  50         STAT_FINAL       = 0x00000010,
  51 } a_firmstat_selection;
  52
  53 /* dumping options */
  54 struct a_firm_dump {
  55         bool debug_print;   /**< enable debug print */
  56         bool all_types;     /**< dump the All_types graph */
  57         bool no_blocks;     /**< dump non-blocked graph */
  58         bool extbb;         /**< dumps extended basic blocks */
  59         bool ir_graph;      /**< dump all graphs */
  60         bool all_phases;    /**< dump the IR graph after all phases */
  61         bool statistic;     /**< Firm statistic setting */
  62         bool stat_pattern;  /**< enable Firm statistic pattern */
  63         bool stat_dag;      /**< enable Firm DAG statistic */
  64 };
  65
  66 struct a_firm_be_opt {
  67         bool selection;
  68         bool node_stat;
  69 };
  70
  71 /* optimization settings */
  72 static struct a_firm_opt firm_opt = {
  73         .const_folding    =  true,
  74         .cse              =  true,
  75         .confirm          =  true,
  76         .muls             =  true,
  77         .divs             =  true,
  78         .mods             =  true,
  79         .alias_analysis   =  true,
  80         .strict_alias     =  false,
  81         .no_alias         =  false,
  82         .verify           =  FIRM_VERIFICATION_ON,
  83         .check_all        =  true,
  84         .clone_threshold  =  DEFAULT_CLONE_THRESHOLD,
  85         .inline_maxsize   =  750,
  86         .inline_threshold =  0,
  87         .verify_edges     =  false,
  88 };
  89
  90 /* dumping options */
  91 static struct a_firm_dump firm_dump = {
  92         .debug_print  = false,
  93         .all_types    = false,
  94         .no_blocks    = false,
  95         .extbb        = false,
  96         .ir_graph     = false,
  97         .all_phases   = false,
  98         .statistic    = STAT_NONE,
  99         .stat_pattern = 0,
 100         .stat_dag     = 0,
 101 };
 102
 103 #define X(a)  a, sizeof(a)-1
 104
 105 /** Parameter description structure */
 106 static const struct params {
 107   const char *option;      /**< name of the option */
 108   size_t     opt_len;      /**< length of the option string */
 109   bool       *flag;        /**< address of variable to set/reset */
 110   bool       set;          /**< iff true, variable will be set, else reset */
 111   const char *description; /**< description of this option */
 112 } firm_options[] = {
 113   /* firm optimization options */
 114   { X("no-opt"),                 NULL,                       0, "disable all FIRM optimizations" },
 115   { X("cse"),                    &firm_opt.cse,              1, "enable common subexpression elimination" },
 116   { X("no-cse"),                 &firm_opt.cse,              0, "disable common subexpression elimination" },
 117   { X("const-fold"),             &firm_opt.const_folding,    1, "enable constant folding" },
 118   { X("no-const-fold"),          &firm_opt.const_folding,    0, "disable constant folding" },
 119   { X("inline-max-size=<size>"), NULL,                       0, "set maximum size for function inlining" },
 120   { X("inline-threshold=<size>"),NULL,                       0, "set benefice threshold for function inlining" },
 121   { X("confirm"),                &firm_opt.confirm,          1, "enable Confirm optimization" },
 122   { X("no-confirm"),             &firm_opt.confirm,          0, "disable Confirm optimization" },
 123   { X("opt-mul"),                &firm_opt.muls,             0, "enable multiplication optimization" },
 124   { X("no-opt-mul"),             &firm_opt.muls,             0, "disable multiplication optimization" },
 125   { X("opt-div"),                &firm_opt.divs,             0, "enable division optimization" },
 126   { X("no-opt-div"),             &firm_opt.divs,             0, "disable division optimization" },
 127   { X("opt-mod"),                &firm_opt.mods,             0, "enable remainder optimization" },
 128   { X("no-opt-mod"),             &firm_opt.mods,             0, "disable remainder optimization" },
 129   { X("opt-alias"),              &firm_opt.alias_analysis,   1, "enable alias analysis" },
 130   { X("no-opt-alias"),           &firm_opt.alias_analysis,   0, "disable alias analysis" },
 131   { X("alias"),                  &firm_opt.no_alias,         0, "aliasing occurs" },
 132   { X("no-alias"),               &firm_opt.no_alias,         1, "no aliasing occurs" },
 133   { X("strict-aliasing"),        &firm_opt.strict_alias,     1, "strict alias rules" },
 134   { X("no-strict-aliasing"),     &firm_opt.strict_alias,     0, "strict alias rules" },
 135   { X("clone-threshold=<value>"),NULL,                       0, "set clone threshold to <value>" },
 136
 137   /* other firm regarding options */
 138   { X("verify-off"),             &firm_opt.verify,           FIRM_VERIFICATION_OFF,    "disable node verification" },
 139   { X("verify-on"),              &firm_opt.verify,           FIRM_VERIFICATION_ON,     "enable node verification" },
 140   { X("verify-report"),          &firm_opt.verify,           FIRM_VERIFICATION_REPORT, "node verification, report only" },
 141   { X("check-all"),              &firm_opt.check_all,        1, "enable checking all Firm phases" },
 142   { X("no-check-all"),           &firm_opt.check_all,        0, "disable checking all Firm phases" },
 143   { X("verify-edges-on"),        &firm_opt.verify_edges,     1, "enable out edge verification" },
 144   { X("verify-edges-off"),       &firm_opt.verify_edges,     0, "disable out edge verification" },
 145
 146   /* dumping */
 147   { X("dump-ir"),                &firm_dump.ir_graph,        1, "dump IR graph" },
 148   { X("dump-all-types"),         &firm_dump.all_types,       1, "dump graph of all types" },
 149   { X("dump-no-blocks"),         &firm_dump.no_blocks,       1, "dump non-blocked graph" },
 150   { X("dump-extbb"),             &firm_dump.extbb,           1, "dump extended basic blocks" },
 151   { X("dump-all-phases"),        &firm_dump.all_phases,      1, "dump graphs for all optimization phases" },
 152   { X("dump-filter=<string>"),   NULL,                       0, "set dumper filter" },
 153
 154   /* misc */
 155   { X("stat-before-opt"),        &firm_dump.statistic,       STAT_BEFORE_OPT,  "Firm statistic output before optimizations" },
 156   { X("stat-after-opt"),         &firm_dump.statistic,       STAT_AFTER_OPT,   "Firm statistic output after optimizations" },
 157   { X("stat-after-lower"),       &firm_dump.statistic,       STAT_AFTER_LOWER, "Firm statistic output after lowering" },
 158   { X("stat-final-ir"),          &firm_dump.statistic,       STAT_FINAL_IR,    "Firm statistic after final optimization" },
 159   { X("stat-final"),             &firm_dump.statistic,       STAT_FINAL,       "Firm statistic after code generation" },
 160   { X("stat-pattern"),           &firm_dump.stat_pattern,    1, "Firm statistic calculates most used pattern" },
 161   { X("stat-dag"),               &firm_dump.stat_dag,        1, "Firm calculates DAG statistics" },
 162 };
 163
 164 #undef X
 165
 166 static ir_timer_t *t_vcg_dump;
 167 static ir_timer_t *t_verify;
 168 static ir_timer_t *t_all_opt;
 169 static bool do_irg_opt(ir_graph *irg, const char *name);
 170
 171 /** dump all the graphs depending on cond */
 172
 173 static void dump_all(const char *suffix)
 174 {
 175         if (!firm_dump.ir_graph)
 176                 return;
 177
 178         timer_push(t_vcg_dump);
 179         dump_all_ir_graphs(suffix);
 180         timer_pop(t_vcg_dump);
 181 }
 182
 183 /* entities of runtime functions */
 184 ir_entity_ptr rts_entities[rts_max];
 185
 186 /**
 187  * Map runtime functions.
 188  */
 189 static void rts_map(void)
 190 {
 191         static const struct {
 192                 ir_entity_ptr *ent; /**< address of the rts entity */
 193                 i_mapper_func func; /**< mapper function. */
 194         } mapper[] = {
 195                 /* integer */
 196                 { &rts_entities[rts_abs],     i_mapper_abs },
 197                 { &rts_entities[rts_labs],    i_mapper_abs },
 198                 { &rts_entities[rts_llabs],   i_mapper_abs },
 199                 { &rts_entities[rts_imaxabs], i_mapper_abs },
 200
 201                 /* double -> double */
 202                 { &rts_entities[rts_fabs],    i_mapper_abs },
 203                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
 204                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
 205                 { &rts_entities[rts_pow],     i_mapper_pow },
 206                 { &rts_entities[rts_exp],     i_mapper_exp },
 207                 { &rts_entities[rts_exp2],    i_mapper_exp },
 208                 { &rts_entities[rts_exp10],   i_mapper_exp },
 209                 { &rts_entities[rts_log],     i_mapper_log },
 210                 { &rts_entities[rts_log2],    i_mapper_log2 },
 211                 { &rts_entities[rts_log10],   i_mapper_log10 },
 212                 { &rts_entities[rts_sin],     i_mapper_sin },
 213                 { &rts_entities[rts_cos],     i_mapper_cos },
 214                 { &rts_entities[rts_tan],     i_mapper_tan },
 215                 { &rts_entities[rts_asin],    i_mapper_asin },
 216                 { &rts_entities[rts_acos],    i_mapper_acos },
 217                 { &rts_entities[rts_atan],    i_mapper_atan },
 218                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 219                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 220                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 221
 222                 /* float -> float */
 223                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 224                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 225                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 226                 { &rts_entities[rts_powf],    i_mapper_pow },
 227                 { &rts_entities[rts_expf],    i_mapper_exp },
 228                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 229                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 230                 { &rts_entities[rts_logf],    i_mapper_log },
 231                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 232                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 233                 { &rts_entities[rts_sinf],    i_mapper_sin },
 234                 { &rts_entities[rts_cosf],    i_mapper_cos },
 235                 { &rts_entities[rts_tanf],    i_mapper_tan },
 236                 { &rts_entities[rts_asinf],   i_mapper_asin },
 237                 { &rts_entities[rts_acosf],   i_mapper_acos },
 238                 { &rts_entities[rts_atanf],   i_mapper_atan },
 239                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 240                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 241                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 242
 243                 /* long double -> long double */
 244                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 245                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 246                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 247                 { &rts_entities[rts_powl],    i_mapper_pow },
 248                 { &rts_entities[rts_expl],    i_mapper_exp },
 249                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 250                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 251                 { &rts_entities[rts_logl],    i_mapper_log },
 252                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 253                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 254                 { &rts_entities[rts_sinl],    i_mapper_sin },
 255                 { &rts_entities[rts_cosl],    i_mapper_cos },
 256                 { &rts_entities[rts_tanl],    i_mapper_tan },
 257                 { &rts_entities[rts_asinl],   i_mapper_asin },
 258                 { &rts_entities[rts_acosl],   i_mapper_acos },
 259                 { &rts_entities[rts_atanl],   i_mapper_atan },
 260                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 261                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 262                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 263
 264                 /* string */
 265                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 266                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 267                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 268                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 269                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 270                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 271                 { &rts_entities[rts_memmove], i_mapper_memmove },
 272                 { &rts_entities[rts_memset],  i_mapper_memset },
 273                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 274         };
 275         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
 276         unsigned i, n_map;
 277
 278         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
 279                 if (*mapper[i].ent != NULL) {
 280                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 281                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 282                         rec[n_map].i_call.i_mapper = mapper[i].func;
 283                         rec[n_map].i_call.ctx      = NULL;
 284                         rec[n_map].i_call.link     = NULL;
 285                         ++n_map;
 286                 }
 287         }
 288
 289         if (n_map > 0)
 290                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 291 }
 292
 293 static int *irg_dump_no;
 294
 295 static void do_lower_highlevel(ir_graph *irg)
 296 {
 297         lower_highlevel_graph(irg, false);
 298 }
 299
 300 static void do_stred(ir_graph *irg)
 301 {
 302         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 303 }
 304
 305 static void after_inline_opt(ir_graph *irg)
 306 {
 307         do_irg_opt(irg, "scalar-replace");
 308         do_irg_opt(irg, "local");
 309         do_irg_opt(irg, "control-flow");
 310         do_irg_opt(irg, "combo");
 311 }
 312
 313 static void do_inline(void)
 314 {
 315         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 316                          after_inline_opt);
 317 }
 318
 319 static void do_cloning(void)
 320 {
 321         proc_cloning((float) firm_opt.clone_threshold);
 322 }
 323
 324 static void do_lower_mux(ir_graph *irg)
 325 {
 326         lower_mux(irg, NULL);
 327 }
 328
 329 static void do_gcse(ir_graph *irg)
 330 {
 331         set_opt_global_cse(1);
 332         optimize_graph_df(irg);
 333         set_opt_global_cse(0);
 334 }
 335
 336 static void lower_blockcopy(ir_graph *irg)
 337 {
 338         lower_CopyB(irg, 128, 4);
 339 }
 340
 341 typedef enum opt_target {
 342         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 343         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 344 } opt_target_t;
 345
 346 typedef enum opt_flags {
 347         OPT_FLAG_NONE         = 0,
 348         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 349         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 350         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 351         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 352                                              -foptions for this transformation */
 353         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 354                                              so we need it even with -O0 */
 355 } opt_flags_t;
 356
 357 typedef void (*transform_irg_func)(ir_graph *irg);
 358 typedef void (*transform_irp_func)(void);
 359
 360 typedef struct {
 361         opt_target_t  target;
 362         const char   *name;
 363         union {
 364                 transform_irg_func transform_irg;
 365                 transform_irp_func transform_irp;
 366         } u;
 367         const char   *description;
 368         opt_flags_t   flags;
 369 } opt_config_t;
 370
 371 static opt_config_t opts[] = {
 372 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
 373 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
 374         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
 375         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
 376         IRG("confirm",           construct_confirms,       "confirm optimization",                                  OPT_FLAG_HIDE_OPTIONS),
 377         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
 378         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 379         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
 380         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
 381         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
 382         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
 383         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
 384         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
 385         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
 386         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
 387         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 388         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
 389         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
 390         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
 391         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
 392         IRG("gcse",              do_gcse,                  "global common subexpression eliminiation",              OPT_FLAG_NONE),
 393         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
 394         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
 395         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 396         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
 397         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
 398         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
 399         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
 400         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
 401         IRG("vrp",               set_vrp_data,             "value range propagation",                               OPT_FLAG_NONE),
 402         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
 403         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),
 404         IRG("lower-blockcopy",   lower_blockcopy,          "replace small block copies with load/store sequences",  OPT_FLAG_NO_DUMP),
 405         IRP("target-lowering",   be_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
 406         IRP("opt-func-call",     optimize_funccalls,       "function call optimization",                            OPT_FLAG_NONE),
 407         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
 408         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
 409         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_NONE),
 410         IRP("opt-cc",            mark_private_methods,     "calling conventions optimization",                      OPT_FLAG_NONE),
 411 #undef IRP
 412 #undef IRG
 413 };
 414 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 415 static ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 416
 417 static opt_config_t *get_opt(const char *name)
 418 {
 419         int i;
 420         for (i = 0; i < n_opts; ++i) {
 421                 opt_config_t *config = &opts[i];
 422                 if (strcmp(config->name, name) == 0)
 423                         return config;
 424         }
 425
 426         return NULL;
 427 }
 428
 429 static void set_opt_enabled(const char *name, bool enabled)
 430 {
 431         opt_config_t *config = get_opt(name);
 432         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 433                 | (enabled ? OPT_FLAG_ENABLED : 0);
 434 }
 435
 436 static bool get_opt_enabled(const char *name)
 437 {
 438         opt_config_t *config = get_opt(name);
 439         return (config->flags & OPT_FLAG_ENABLED) != 0;
 440 }
 441
 442 /**
 443  * perform an optimization on a single graph
 444  *
 445  * @return  true if something changed, false otherwise
 446  */
 447 static bool do_irg_opt(ir_graph *irg, const char *name)
 448 {
 449         ir_graph     *old_irg;
 450         opt_config_t *config = get_opt(name);
 451         size_t        n      = config - opts;
 452         assert(config != NULL);
 453         assert(config->target == OPT_TARGET_IRG);
 454         if (! (config->flags & OPT_FLAG_ENABLED))
 455                 return false;
 456
 457         old_irg          = current_ir_graph;
 458         current_ir_graph = irg;
 459
 460         timer_push(timers[n]);
 461         config->u.transform_irg(irg);
 462         timer_pop(timers[n]);
 463
 464         if (firm_dump.all_phases && firm_dump.ir_graph) {
 465                 dump_ir_graph(irg, name);
 466         }
 467
 468         if (firm_opt.check_all) {
 469                 timer_push(t_verify);
 470                 irg_verify(irg, VERIFY_ENFORCE_SSA);
 471                 timer_pop(t_verify);
 472         }
 473
 474         current_ir_graph = old_irg;
 475         return true;
 476 }
 477
 478 static void do_irp_opt(const char *name)
 479 {
 480         opt_config_t *config = get_opt(name);
 481         size_t        n      = config - opts;
 482         assert(config->target == OPT_TARGET_IRP);
 483         if (! (config->flags & OPT_FLAG_ENABLED))
 484                 return;
 485
 486         timer_push(timers[n]);
 487         config->u.transform_irp();
 488         timer_pop(timers[n]);
 489
 490         if (firm_dump.ir_graph && firm_dump.all_phases) {
 491                 int i;
 492                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 493                         ir_graph *irg = get_irp_irg(i);
 494                         dump_ir_graph(irg, name);
 495                 }
 496         }
 497
 498         if (firm_opt.check_all) {
 499                 int i;
 500                 timer_push(t_verify);
 501                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 502                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
 503                 }
 504                 timer_pop(t_verify);
 505         }
 506 }
 507
 508 /**
 509  * Enable transformations which should be always safe (and cheap) to perform
 510  */
 511 static void enable_safe_defaults(void)
 512 {
 513         set_opt_enabled("remove-unused", true);
 514         set_opt_enabled("opt-tail-rec", true);
 515         set_opt_enabled("opt-func-call", true);
 516         set_opt_enabled("reassociation", true);
 517         set_opt_enabled("control-flow", true);
 518         set_opt_enabled("local", true);
 519         set_opt_enabled("lower-const", true);
 520         set_opt_enabled("scalar-replace", true);
 521         set_opt_enabled("place", true);
 522         set_opt_enabled("gcse", true);
 523         set_opt_enabled("confirm", true);
 524         set_opt_enabled("opt-load-store", true);
 525         set_opt_enabled("lower", true);
 526         set_opt_enabled("lower-blockcopy", true);
 527         set_opt_enabled("deconv", true);
 528         set_opt_enabled("remove-confirms", true);
 529         set_opt_enabled("ivopts", true);
 530         set_opt_enabled("dead", true);
 531         set_opt_enabled("remove-phi-cycles", true);
 532         set_opt_enabled("frame", true);
 533         set_opt_enabled("combo", true);
 534         set_opt_enabled("invert-loops", true);
 535         set_opt_enabled("target-lowering", true);
 536         set_opt_enabled("rts", true);
 537         set_opt_enabled("parallelize-mem", true);
 538         set_opt_enabled("opt-cc", true);
 539 }
 540
 541 /**
 542  * run all the Firm optimizations
 543  *
 544  * @param input_filename     the name of the (main) source file
 545  */
 546 static void do_firm_optimizations(const char *input_filename)
 547 {
 548         size_t   i;
 549         unsigned aa_opt;
 550
 551         set_opt_alias_analysis(firm_opt.alias_analysis);
 552
 553         aa_opt = aa_opt_no_opt;
 554         if (firm_opt.strict_alias)
 555                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 556         if (firm_opt.no_alias)
 557                 aa_opt = aa_opt_no_alias;
 558
 559         set_irp_memory_disambiguator_options(aa_opt);
 560
 561         /* parameter passing code should set them directly sometime... */
 562         set_opt_enabled("confirm", firm_opt.confirm);
 563         set_opt_enabled("remove-confirms", firm_opt.confirm);
 564
 565         /* osr supersedes remove_phi_cycles */
 566         if (get_opt_enabled("ivopts"))
 567                 set_opt_enabled("remove-phi-cycles", false);
 568
 569         timer_start(t_all_opt);
 570
 571         do_irp_opt("rts");
 572
 573         /* first step: kill dead code */
 574         for (i = 0; i < get_irp_n_irgs(); i++) {
 575                 ir_graph *irg = get_irp_irg(i);
 576                 do_irg_opt(irg, "combo");
 577                 do_irg_opt(irg, "local");
 578                 do_irg_opt(irg, "control-flow");
 579         }
 580
 581         do_irp_opt("remove-unused");
 582         for (i = 0; i < get_irp_n_irgs(); ++i) {
 583                 ir_graph *irg = get_irp_irg(i);
 584                 do_irg_opt(irg, "opt-tail-rec");
 585         }
 586         do_irp_opt("opt-func-call");
 587         do_irp_opt("lower-const");
 588
 589         for (i = 0; i < get_irp_n_irgs(); i++) {
 590                 ir_graph *irg = get_irp_irg(i);
 591
 592                 do_irg_opt(irg, "scalar-replace");
 593                 do_irg_opt(irg, "invert-loops");
 594                 do_irg_opt(irg, "unroll-loops");
 595                 do_irg_opt(irg, "local");
 596                 do_irg_opt(irg, "reassociation");
 597                 do_irg_opt(irg, "local");
 598                 do_irg_opt(irg, "gcse");
 599                 do_irg_opt(irg, "place");
 600
 601                 if (firm_opt.confirm) {
 602                         /* Confirm construction currently can only handle blocks with only
 603                            one control flow predecessor. Calling optimize_cf here removes
 604                            Bad predecessors and help the optimization of switch constructs.
 605                          */
 606                         do_irg_opt(irg, "control-flow");
 607                         do_irg_opt(irg, "confirm");
 608                         do_irg_opt(irg, "vrp");
 609                         do_irg_opt(irg, "local");
 610                 }
 611
 612                 do_irg_opt(irg, "control-flow");
 613                 do_irg_opt(irg, "opt-load-store");
 614                 do_irg_opt(irg, "fp-vrp");
 615                 do_irg_opt(irg, "lower");
 616                 do_irg_opt(irg, "deconv");
 617                 do_irg_opt(irg, "thread-jumps");
 618                 do_irg_opt(irg, "remove-confirms");
 619                 do_irg_opt(irg, "gvn-pre");
 620                 do_irg_opt(irg, "gcse");
 621                 do_irg_opt(irg, "place");
 622                 do_irg_opt(irg, "control-flow");
 623
 624                 if (do_irg_opt(irg, "if-conversion")) {
 625                         do_irg_opt(irg, "local");
 626                         do_irg_opt(irg, "control-flow");
 627                 }
 628                 /* this doesn't make too much sense but tests the mux destruction... */
 629                 do_irg_opt(irg, "lower-mux");
 630
 631                 do_irg_opt(irg, "bool");
 632                 do_irg_opt(irg, "shape-blocks");
 633                 do_irg_opt(irg, "ivopts");
 634                 do_irg_opt(irg, "local");
 635                 do_irg_opt(irg, "dead");
 636         }
 637
 638         do_irp_opt("inline");
 639         do_irp_opt("opt-proc-clone");
 640
 641         for (i = 0; i < get_irp_n_irgs(); i++) {
 642                 ir_graph *irg = get_irp_irg(i);
 643                 do_irg_opt(irg, "local");
 644                 do_irg_opt(irg, "control-flow");
 645                 do_irg_opt(irg, "thread-jumps");
 646                 do_irg_opt(irg, "local");
 647                 do_irg_opt(irg, "control-flow");
 648
 649                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 650                         do_irg_opt(irg, "local");
 651                         do_irg_opt(irg, "vrp");
 652                         do_irg_opt(irg, "local");
 653                         do_irg_opt(irg, "vrp");
 654                 }
 655         }
 656
 657         if (firm_dump.ir_graph) {
 658                 /* recompute backedges for nicer dumps */
 659                 for (i = 0; i < get_irp_n_irgs(); i++)
 660                         construct_cf_backedges(get_irp_irg(i));
 661         }
 662
 663         dump_all("opt");
 664
 665         if (firm_dump.statistic & STAT_AFTER_OPT)
 666                 stat_dump_snapshot(input_filename, "opt");
 667
 668         timer_stop(t_all_opt);
 669 }
 670
 671 /**
 672  * do Firm lowering
 673  *
 674  * @param input_filename  the name of the (main) source file
 675  */
 676 static void do_firm_lowering(const char *input_filename)
 677 {
 678         int i;
 679
 680         /* enable architecture dependent optimizations */
 681         arch_dep_set_opts((arch_dep_opts_t)
 682                         ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 683                          (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 684                          (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 685         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 686                 ir_graph *irg = get_irp_irg(i);
 687                 do_irg_opt(irg, "reassociation");
 688                 do_irg_opt(irg, "local");
 689         }
 690
 691         do_irp_opt("target-lowering");
 692
 693         if (firm_dump.statistic & STAT_AFTER_LOWER)
 694                 stat_dump_snapshot(input_filename, "low");
 695
 696         timer_start(t_all_opt);
 697
 698         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 699                 ir_graph *irg = get_irp_irg(i);
 700
 701                 do_irg_opt(irg, "local");
 702                 do_irg_opt(irg, "deconv");
 703                 do_irg_opt(irg, "control-flow");
 704                 do_irg_opt(irg, "opt-load-store");
 705                 do_irg_opt(irg, "gcse");
 706                 do_irg_opt(irg, "place");
 707                 do_irg_opt(irg, "control-flow");
 708
 709                 if (do_irg_opt(irg, "vrp")) {
 710                         do_irg_opt(irg, "local");
 711                         do_irg_opt(irg, "control-flow");
 712                         do_irg_opt(irg, "vrp");
 713                         do_irg_opt(irg, "local");
 714                         do_irg_opt(irg, "control-flow");
 715                 }
 716
 717                 if (do_irg_opt(irg, "if-conversion")) {
 718                         do_irg_opt(irg, "local");
 719                         do_irg_opt(irg, "control-flow");
 720                 }
 721
 722                 set_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2);
 723                 do_irg_opt(irg, "local");
 724
 725                 do_irg_opt(irg, "parallelize-mem");
 726                 do_irg_opt(irg, "frame");
 727         }
 728         do_irp_opt("remove-unused");
 729         do_irp_opt("opt-cc");
 730         timer_stop(t_all_opt);
 731         dump_all("low-opt");
 732
 733         if (firm_dump.statistic & STAT_FINAL) {
 734                 stat_dump_snapshot(input_filename, "final");
 735         }
 736 }
 737
 738 /**
 739  * Initialize for the Firm-generating back end.
 740  */
 741 void gen_firm_init(void)
 742 {
 743         unsigned pattern = 0;
 744         int      i;
 745
 746         for (i = 0; i < n_opts; ++i) {
 747                 timers[i] = ir_timer_new();
 748                 timer_register(timers[i], opts[i].description);
 749         }
 750         t_verify = ir_timer_new();
 751         timer_register(t_verify, "Firm: verify pass");
 752         t_vcg_dump = ir_timer_new();
 753         timer_register(t_vcg_dump, "Firm: vcg dumping");
 754         t_all_opt = ir_timer_new();
 755         timer_register(t_all_opt, "Firm: all optimizations");
 756
 757         if (firm_dump.stat_pattern)
 758                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 759
 760         if (firm_dump.stat_dag)
 761                 pattern |= FIRMSTAT_COUNT_DAG;
 762
 763         ir_init(NULL);
 764         firm_init_stat(firm_dump.statistic == STAT_NONE ?
 765                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
 766                         | FIRMSTAT_COUNT_CONSTS | pattern);
 767
 768         edges_init_dbg(firm_opt.verify_edges);
 769
 770         /* Sel node cannot produce NULL pointers */
 771         set_opt_sel_based_null_check_elim(1);
 772
 773         /* dynamic dispatch works currently only if whole world scenarios */
 774         set_opt_dyn_meth_dispatch(0);
 775
 776         /* do not run architecture dependent optimizations in building phase */
 777         arch_dep_set_opts(arch_dep_none);
 778
 779         do_node_verification((firm_verification_t) firm_opt.verify);
 780         if (firm_dump.extbb)
 781                 ir_add_dump_flags(ir_dump_flag_group_extbb);
 782         if (firm_dump.no_blocks)
 783                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
 784
 785         set_optimize(1);
 786         set_opt_constant_folding(firm_opt.const_folding);
 787         set_opt_algebraic_simplification(firm_opt.const_folding);
 788         set_opt_cse(firm_opt.cse);
 789         set_opt_global_cse(0);
 790         set_opt_unreachable_code(1);
 791 }
 792
 793 /**
 794  * Called, after the Firm generation is completed,
 795  * do all optimizations and backend call here.
 796  *
 797  * @param out                a file handle for the output, may be NULL
 798  * @param input_filename     the name of the (main) source file
 799  * @param c_mode             non-zero if "C" was compiled
 800  */
 801 void gen_firm_finish(FILE *out, const char *input_filename)
 802 {
 803         int i;
 804
 805         /* the general for dumping option must be set, or the others will not work*/
 806         firm_dump.ir_graph
 807                 = (bool) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 808
 809         ir_add_dump_flags(ir_dump_flag_keepalive_edges
 810                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
 811         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
 812
 813         /* FIXME: cloning might ADD new graphs. */
 814         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 815
 816         if (firm_dump.all_types) {
 817                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
 818         }
 819
 820         /* finalize all graphs */
 821         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 822                 ir_graph *irg = get_irp_irg(i);
 823                 irg_finalize_cons(irg);
 824         }
 825         dump_all("");
 826
 827         timer_push(t_verify);
 828         tr_verify();
 829         timer_pop(t_verify);
 830
 831         /* all graphs are finalized, set the irp phase to high */
 832         set_irp_phase_state(phase_high);
 833
 834         /* BEWARE: kill unreachable code before doing compound lowering */
 835         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 836                 ir_graph *irg = get_irp_irg(i);
 837                 do_irg_opt(irg, "control-flow");
 838                 do_irg_opt(irg, "lower-blockcopy");
 839         }
 840
 841         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 842                 stat_dump_snapshot(input_filename, "noopt");
 843         }
 844
 845         do_firm_optimizations(input_filename);
 846         do_firm_lowering(input_filename);
 847
 848         /* set the phase to low */
 849         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 850                 set_irg_phase_state(get_irp_irg(i), phase_low);
 851
 852         if (firm_dump.statistic & STAT_FINAL_IR)
 853                 stat_dump_snapshot(input_filename, "final-ir");
 854
 855         /* run the code generator */
 856         ir_timer_t *timer = ir_timer_new();
 857         timer_register(timer, "Firm: backend");
 858         timer_start(timer);
 859         be_main(out, input_filename);
 860         timer_stop(timer);
 861
 862         if (firm_dump.statistic & STAT_FINAL)
 863                 stat_dump_snapshot(input_filename, "final");
 864 }
 865
 866 static void disable_all_opts(void)
 867 {
 868         firm_opt.cse             = false;
 869         firm_opt.confirm         = false;
 870         firm_opt.muls            = false;
 871         firm_opt.divs            = false;
 872         firm_opt.mods            = false;
 873         firm_opt.alias_analysis  = false;
 874         firm_opt.strict_alias    = false;
 875         firm_opt.no_alias        = false;
 876
 877         for (int i = 0; i < n_opts; ++i) {
 878                 opt_config_t *config = &opts[i];
 879                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 880                         config->flags |= OPT_FLAG_ENABLED;
 881                 } else {
 882                         config->flags &= ~OPT_FLAG_ENABLED;
 883                 }
 884         }
 885 }
 886
 887 static bool firm_opt_option(const char *opt)
 888 {
 889         bool enable = true;
 890         if (strncmp(opt, "no-", 3) == 0) {
 891                 enable = false;
 892                 opt = opt + 3;
 893         }
 894
 895         opt_config_t *config = get_opt(opt);
 896         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 897                 return false;
 898
 899         config->flags &= ~OPT_FLAG_ENABLED;
 900         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 901         return true;
 902 }
 903
 904 void firm_option_help(print_option_help_func print_option_help)
 905 {
 906         print_option_help(firm_options[0].option, firm_options[0].description);
 907
 908         for (int i = 0; i < n_opts; ++i) {
 909                 char buf[1024];
 910                 char buf2[1024];
 911
 912                 const opt_config_t *config = &opts[i];
 913                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 914                         continue;
 915
 916                 snprintf(buf, sizeof(buf), "-f%s", config->name);
 917                 snprintf(buf2, sizeof(buf2), "enable %s", config->description);
 918                 print_option_help(buf, buf2);
 919                 snprintf(buf, sizeof(buf), "-fno-%s", config->name);
 920                 snprintf(buf2, sizeof(buf2), "disable %s", config->description);
 921                 print_option_help(buf, buf2);
 922         }
 923
 924         size_t const n_options = sizeof(firm_options)/sizeof(firm_options[0]);
 925         for (size_t k = 0; k < n_options; ++k) {
 926                 char buf[1024];
 927                 char buf2[1024];
 928                 snprintf(buf, sizeof(buf), "-f%s", firm_options[k].option);
 929                 snprintf(buf2, sizeof(buf2), "%s", firm_options[k].description);
 930                 print_option_help(buf, buf2);
 931         }
 932 }
 933
 934 int firm_option(const char *const opt)
 935 {
 936         char const* val;
 937         if ((val = strstart(opt, "dump-filter="))) {
 938                 ir_set_dump_filter(val);
 939                 return 1;
 940         } else if ((val = strstart(opt, "clone-threshold="))) {
 941                 sscanf(val, "%d", &firm_opt.clone_threshold);
 942                 return 1;
 943         } else if ((val = strstart(opt, "inline-max-size="))) {
 944                 sscanf(val, "%u", &firm_opt.inline_maxsize);
 945                 return 1;
 946         } else if ((val = strstart(opt, "inline-threshold="))) {
 947                 sscanf(val, "%u", &firm_opt.inline_threshold);
 948                 return 1;
 949         } else if (streq(opt, "no-opt")) {
 950                 disable_all_opts();
 951                 return 1;
 952         }
 953
 954         size_t const len = strlen(opt);
 955         size_t const n_options = sizeof(firm_options)/sizeof(firm_options[0]);
 956         for (size_t i = n_options; i != 0;) {
 957                 struct params const* const o = &firm_options[--i];
 958                 if (len == o->opt_len && strncmp(opt, o->option, len) == 0) {
 959                         /* statistic options do accumulate */
 960                         if (o->flag == &firm_dump.statistic)
 961                                 *o->flag = (bool) (*o->flag | o->set);
 962                         else
 963                                 *o->flag = o->set;
 964
 965                         return 1;
 966                 }
 967         }
 968
 969         /* maybe this enables/disables optimizations */
 970         if (firm_opt_option(opt))
 971                 return 1;
 972
 973         return 0;
 974 }
 975
 976 static void set_be_option(const char *arg)
 977 {
 978         int res = be_parse_arg(arg);
 979         (void) res;
 980         assert(res);
 981 }
 982
 983 static void set_option(const char *arg)
 984 {
 985         int res = firm_option(arg);
 986         (void) res;
 987         assert(res);
 988 }
 989
 990 void choose_optimization_pack(int level)
 991 {
 992         /* apply optimization level */
 993         switch(level) {
 994         case 0:
 995                 set_option("no-opt");
 996                 break;
 997         case 1:
 998                 set_option("no-inline");
 999                 break;
1000         default:
1001         case 4:
1002                 /* use_builtins = true; */
1003                 /* fallthrough */
1004         case 3:
1005                 set_option("thread-jumps");
1006                 set_option("if-conversion");
1007                 /* fallthrough */
1008         case 2:
1009                 set_option("strict-aliasing");
1010                 set_option("inline");
1011                 set_option("fp-vrp");
1012                 set_option("deconv");
1013                 set_be_option("omitfp");
1014                 break;
1015         }
1016 }
1017
1018 /**
1019  * Do very early initializations
1020  */
1021 void firm_early_init(void)
1022 {
1023         /* arg: need this here for command line options */
1024         be_opt_register();
1025
1026         enable_safe_defaults();
1027 }