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

   1 /**
   2  *
   3  * @file firm_opt.c -- Firm-generating back end optimizations.
   4  *
   5  * (C) 2005-2010  Michael Beck   beck@ipd.info.uni-karlsruhe.de
   6  *
   7  * $Id$
   8  */
   9
  10 #include <config.h>
  11
  12 #include <stdbool.h>
  13 #include <stdlib.h>
  14 #include <string.h>
  15 #include <stdbool.h>
  16 #include <assert.h>
  17 #include <libfirm/firm.h>
  18
  19 #include "firm_opt.h"
  20 #include "firm_codegen.h"
  21 #include "firm_cmdline.h"
  22 #include "firm_timing.h"
  23 #include "ast2firm.h"
  24
  25 #if defined(_DEBUG) || defined(FIRM_DEBUG)
  26 #define DBG(x)  dbg_printf x
  27 #else
  28 #define DBG(x) ((void)0)
  29 #endif /* _DEBUG || FIRM_DEBUG */
  30
  31 static ir_timer_t *t_vcg_dump;
  32 static ir_timer_t *t_verify;
  33 static ir_timer_t *t_all_opt;
  34 static bool do_irg_opt(ir_graph *irg, const char *name);
  35
  36 /** dump all the graphs depending on cond */
  37
  38 static void dump_all(const char *suffix) {
  39         if (firm_dump.ir_graph) {
  40                 timer_push(t_vcg_dump);
  41                 if (firm_dump.no_blocks)
  42                         dump_all_ir_graphs(dump_ir_graph, suffix);
  43                 else if (firm_dump.extbb)
  44                         dump_all_ir_graphs(dump_ir_extblock_graph, suffix);
  45                 else
  46                         dump_all_ir_graphs(dump_ir_block_graph, suffix);
  47                 timer_pop(t_vcg_dump);
  48         }
  49 }
  50
  51 /* set by the backend parameters */
  52 static const ir_settings_arch_dep_t *ad_param              = NULL;
  53 static create_intrinsic_fkt         *arch_create_intrinsic = NULL;
  54 static void                         *create_intrinsic_ctx  = NULL;
  55 static const ir_settings_if_conv_t  *if_conv_info          = NULL;
  56
  57 /* entities of runtime functions */
  58 ir_entity_ptr rts_entities[rts_max];
  59
  60 /**
  61  * factory for setting architecture dependent parameters
  62  */
  63 static const ir_settings_arch_dep_t *arch_factory(void)
  64 {
  65         static const ir_settings_arch_dep_t param = {
  66                 1,   /* also use subs */
  67                 4,   /* maximum shifts */
  68                 31,   /* maximum shift amount */
  69                 NULL, /* use default evaluator */
  70
  71                 1, /* allow Mulhs */
  72                 1, /* allow Mulus */
  73                 32  /* Mulh allowed up to 32 bit */
  74         };
  75
  76         return ad_param ? ad_param : &param;
  77 }
  78
  79 /**
  80  * Map runtime functions.
  81  */
  82 static void rts_map(void) {
  83         static const struct {
  84                 ir_entity_ptr *ent; /**< address of the rts entity */
  85                 i_mapper_func func; /**< mapper function. */
  86         } mapper[] = {
  87                 /* integer */
  88                 { &rts_entities[rts_abs],     i_mapper_abs },
  89                 { &rts_entities[rts_labs],    i_mapper_abs },
  90                 { &rts_entities[rts_llabs],   i_mapper_abs },
  91                 { &rts_entities[rts_imaxabs], i_mapper_abs },
  92
  93                 /* double -> double */
  94                 { &rts_entities[rts_fabs],    i_mapper_abs },
  95                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
  96                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
  97                 { &rts_entities[rts_pow],     i_mapper_pow },
  98                 { &rts_entities[rts_exp],     i_mapper_exp },
  99                 { &rts_entities[rts_exp2],    i_mapper_exp },
 100                 { &rts_entities[rts_exp10],   i_mapper_exp },
 101                 { &rts_entities[rts_log],     i_mapper_log },
 102                 { &rts_entities[rts_log2],    i_mapper_log2 },
 103                 { &rts_entities[rts_log10],   i_mapper_log10 },
 104                 { &rts_entities[rts_sin],     i_mapper_sin },
 105                 { &rts_entities[rts_cos],     i_mapper_cos },
 106                 { &rts_entities[rts_tan],     i_mapper_tan },
 107                 { &rts_entities[rts_asin],    i_mapper_asin },
 108                 { &rts_entities[rts_acos],    i_mapper_acos },
 109                 { &rts_entities[rts_atan],    i_mapper_atan },
 110                 { &rts_entities[rts_sinh],    i_mapper_sinh },
 111                 { &rts_entities[rts_cosh],    i_mapper_cosh },
 112                 { &rts_entities[rts_tanh],    i_mapper_tanh },
 113
 114                 /* float -> float */
 115                 { &rts_entities[rts_fabsf],   i_mapper_abs },
 116                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
 117                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
 118                 { &rts_entities[rts_powf],    i_mapper_pow },
 119                 { &rts_entities[rts_expf],    i_mapper_exp },
 120                 { &rts_entities[rts_exp2f],   i_mapper_exp },
 121                 { &rts_entities[rts_exp10f],  i_mapper_exp },
 122                 { &rts_entities[rts_logf],    i_mapper_log },
 123                 { &rts_entities[rts_log2f],   i_mapper_log2 },
 124                 { &rts_entities[rts_log10f],  i_mapper_log10 },
 125                 { &rts_entities[rts_sinf],    i_mapper_sin },
 126                 { &rts_entities[rts_cosf],    i_mapper_cos },
 127                 { &rts_entities[rts_tanf],    i_mapper_tan },
 128                 { &rts_entities[rts_asinf],   i_mapper_asin },
 129                 { &rts_entities[rts_acosf],   i_mapper_acos },
 130                 { &rts_entities[rts_atanf],   i_mapper_atan },
 131                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
 132                 { &rts_entities[rts_coshf],   i_mapper_cosh },
 133                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
 134
 135                 /* long double -> long double */
 136                 { &rts_entities[rts_fabsl],   i_mapper_abs },
 137                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
 138                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
 139                 { &rts_entities[rts_powl],    i_mapper_pow },
 140                 { &rts_entities[rts_expl],    i_mapper_exp },
 141                 { &rts_entities[rts_exp2l],   i_mapper_exp },
 142                 { &rts_entities[rts_exp10l],  i_mapper_exp },
 143                 { &rts_entities[rts_logl],    i_mapper_log },
 144                 { &rts_entities[rts_log2l],   i_mapper_log2 },
 145                 { &rts_entities[rts_log10l],  i_mapper_log10 },
 146                 { &rts_entities[rts_sinl],    i_mapper_sin },
 147                 { &rts_entities[rts_cosl],    i_mapper_cos },
 148                 { &rts_entities[rts_tanl],    i_mapper_tan },
 149                 { &rts_entities[rts_asinl],   i_mapper_asin },
 150                 { &rts_entities[rts_acosl],   i_mapper_acos },
 151                 { &rts_entities[rts_atanl],   i_mapper_atan },
 152                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
 153                 { &rts_entities[rts_coshl],   i_mapper_cosh },
 154                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
 155
 156                 /* string */
 157                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
 158                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
 159                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
 160                 { &rts_entities[rts_strlen],  i_mapper_strlen },
 161                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
 162                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
 163                 { &rts_entities[rts_memmove], i_mapper_memmove },
 164                 { &rts_entities[rts_memset],  i_mapper_memset },
 165                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
 166         };
 167         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
 168         unsigned i, n_map;
 169
 170         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
 171                 if (*mapper[i].ent != NULL) {
 172                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
 173                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
 174                         rec[n_map].i_call.i_mapper = mapper[i].func;
 175                         rec[n_map].i_call.ctx      = NULL;
 176                         rec[n_map].i_call.link     = NULL;
 177                         ++n_map;
 178                 }
 179         }
 180
 181         if (n_map > 0)
 182                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
 183 }
 184
 185 static int *irg_dump_no;
 186
 187 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
 188 {
 189         char name[64];
 190         snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++,
 191                  suffix);
 192
 193         timer_push(t_vcg_dump);
 194         if (firm_dump.no_blocks)
 195                 dump_ir_graph(irg, name);
 196         else if (firm_dump.extbb)
 197                 dump_ir_extblock_graph(irg, name);
 198         else
 199                 dump_ir_block_graph(irg, name);
 200         timer_pop(t_vcg_dump);
 201 }
 202
 203 static int firm_const_exists;
 204
 205 static void do_optimize_funccalls(void)
 206 {
 207         optimize_funccalls(firm_const_exists, NULL);
 208 }
 209
 210 static void do_gcse(ir_graph *irg)
 211 {
 212         set_opt_global_cse(1);
 213         optimize_graph_df(irg);
 214         place_code(irg);
 215         set_opt_global_cse(0);
 216 }
 217
 218 static void do_lower_highlevel(ir_graph *irg)
 219 {
 220         lower_highlevel_graph(irg, firm_opt.lower_bitfields);
 221 }
 222
 223 static void do_if_conv(ir_graph *irg)
 224 {
 225         opt_if_conv(irg, if_conv_info);
 226 }
 227
 228 static void do_stred(ir_graph *irg)
 229 {
 230         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
 231 }
 232
 233 static void after_inline_opt(ir_graph *irg)
 234 {
 235         do_irg_opt(irg, "scalar-replace");
 236         do_irg_opt(irg, "local");
 237         do_irg_opt(irg, "control-flow");
 238         do_irg_opt(irg, "combo");
 239 }
 240
 241 static void do_inline(void)
 242 {
 243         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
 244                          after_inline_opt);
 245 }
 246
 247 static void do_cloning(void)
 248 {
 249         proc_cloning((float) firm_opt.clone_threshold);
 250 }
 251
 252 static void do_lower_switch(ir_graph *irg)
 253 {
 254         lower_switch(irg, firm_opt.spare_size);
 255 }
 256
 257 static void do_lower_mux(ir_graph *irg)
 258 {
 259         lower_mux(irg, NULL);
 260 }
 261
 262 static void do_lower_dw_ops(void)
 263 {
 264         lwrdw_param_t init = {
 265                 1,
 266                 1,
 267                 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
 268                 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
 269                 get_atomic_mode(ATOMIC_TYPE_INT),
 270                 get_atomic_mode(ATOMIC_TYPE_UINT),
 271                 def_create_intrinsic_fkt,
 272                 NULL
 273         };
 274
 275         if (arch_create_intrinsic) {
 276                 init.create_intrinsic = arch_create_intrinsic;
 277                 init.ctx              = create_intrinsic_ctx;
 278         }
 279         lower_dw_ops(&init);
 280 }
 281
 282 static void do_vrp(ir_graph *irg)
 283 {
 284         set_vrp_data(irg);
 285 }
 286
 287 typedef enum opt_target {
 288         OPT_TARGET_IRG, /**< optimization function works on a single graph */
 289         OPT_TARGET_IRP  /**< optimization function works on the complete program */
 290 } opt_target_t;
 291
 292 typedef enum opt_flags {
 293         OPT_FLAG_NONE         = 0,
 294         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
 295         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
 296         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
 297         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
 298                                              -foptions for this transformation */
 299         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
 300                                              so we need it even with -O0 */
 301 } opt_flags_t;
 302
 303 typedef void (*transform_irg_func)(ir_graph *irg);
 304 typedef void (*transform_irp_func)(void);
 305 typedef void (*func_ptr_t)(void);
 306
 307 typedef struct {
 308         opt_target_t  target;
 309         const char   *name;
 310         func_ptr_t    func;
 311         const char   *description;
 312         opt_flags_t   flags;
 313 } opt_config_t;
 314
 315 static opt_config_t opts[] = {
 316         { OPT_TARGET_IRP, "rts",             (func_ptr_t) rts_map,                 "optimization of known library functions", OPT_FLAG_HIDE_OPTIONS },
 317         { OPT_TARGET_IRG, "combo",           (func_ptr_t) combo,                   "combined CCE, UCE and GVN",               OPT_FLAG_NONE},
 318         { OPT_TARGET_IRG, "control-flow",    (func_ptr_t) optimize_cf,             "optimization of control-flow",            OPT_FLAG_HIDE_OPTIONS },
 319         { OPT_TARGET_IRG, "local",           (func_ptr_t) optimize_graph_df,       "local graph optimizations",               OPT_FLAG_HIDE_OPTIONS },
 320         { OPT_TARGET_IRP, "remove-unused",   (func_ptr_t) garbage_collect_entities,"removal of unused functions/variables",   OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
 321         { OPT_TARGET_IRP, "opt-tail-rec",    (func_ptr_t) opt_tail_recursion,      "tail-recursion eliminiation",             OPT_FLAG_NONE },
 322         { OPT_TARGET_IRP, "opt-func-call",   (func_ptr_t) do_optimize_funccalls,   "function call optimization",              OPT_FLAG_NONE },
 323         { OPT_TARGET_IRG, "lower",           (func_ptr_t) do_lower_highlevel,      "lowering",                                OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL },
 324         { OPT_TARGET_IRP, "lower-const",     (func_ptr_t) lower_const_code,        "lowering of constant code",               OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY | OPT_FLAG_ESSENTIAL },
 325         { OPT_TARGET_IRP, "lower-dw",        (func_ptr_t) do_lower_dw_ops,         "lowering of doubleword operations",       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL },
 326         { OPT_TARGET_IRG, "lower-switch",    (func_ptr_t) do_lower_switch,         "switch lowering",                         OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL },
 327         { OPT_TARGET_IRG, "one-return",      (func_ptr_t) normalize_one_return,    "normalisation to 1 return",               OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
 328         { OPT_TARGET_IRG, "scalar-replace",  (func_ptr_t) scalar_replacement_opt,  "scalar replacement",                      OPT_FLAG_NONE },
 329         { OPT_TARGET_IRG, "reassociation",   (func_ptr_t) optimize_reassociation,  "reassociation",                           OPT_FLAG_NONE },
 330         { OPT_TARGET_IRG, "gcse",            (func_ptr_t) do_gcse,                 "global common subexpression elimination", OPT_FLAG_NONE },
 331         { OPT_TARGET_IRG, "place",           (func_ptr_t) place_code,              "code placement",                          OPT_FLAG_NONE },
 332         { OPT_TARGET_IRG, "confirm",         (func_ptr_t) construct_confirms,      "confirm optimisation",                    OPT_FLAG_HIDE_OPTIONS },
 333         { OPT_TARGET_IRG, "opt-load-store",  (func_ptr_t) optimize_load_store,     "load store optimization",                 OPT_FLAG_NONE },
 334         { OPT_TARGET_IRG, "parallelize-mem", (func_ptr_t) opt_parallelize_mem,     "parallelize memory",                      OPT_FLAG_NONE },
 335         { OPT_TARGET_IRG, "deconv",          (func_ptr_t) conv_opt,                "conv node elimination",                   OPT_FLAG_NONE },
 336         { OPT_TARGET_IRG, "thread-jumps",    (func_ptr_t) opt_jumpthreading,       "path-sensitive jumpthreading",            OPT_FLAG_NONE },
 337         { OPT_TARGET_IRG, "remove-confirms", (func_ptr_t) remove_confirms,         "confirm removal",                         OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
 338         { OPT_TARGET_IRG, "gvn-pre",         (func_ptr_t) do_gvn_pre,              "global value numbering partial redundancy elimination", OPT_FLAG_NONE },
 339         { OPT_TARGET_IRG, "if-conversion",   (func_ptr_t) do_if_conv,              "if-conversion",                           OPT_FLAG_NONE },
 340         { OPT_TARGET_IRG, "bool",            (func_ptr_t) opt_bool,                "bool simplification",                     OPT_FLAG_NONE },
 341         { OPT_TARGET_IRG, "shape-blocks",    (func_ptr_t) shape_blocks,            "block shaping",                           OPT_FLAG_NONE },
 342         { OPT_TARGET_IRG, "ivopts",          (func_ptr_t) do_stred,                "induction variable strength reduction",   OPT_FLAG_NONE },
 343         { OPT_TARGET_IRG, "remove-phi-cycles", (func_ptr_t) remove_phi_cycles,     "removal of phi cycles",                   OPT_FLAG_HIDE_OPTIONS },
 344         { OPT_TARGET_IRG, "dead",            (func_ptr_t) dead_node_elimination,   "dead node elimination",                   OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
 345         { OPT_TARGET_IRP, "inline",          (func_ptr_t) do_inline,               "inlining",                                OPT_FLAG_NONE },
 346         { OPT_TARGET_IRP, "opt-proc-clone",  (func_ptr_t) do_cloning,              "procedure cloning",                       OPT_FLAG_NONE },
 347         { OPT_TARGET_IRG, "invert-loops",    (func_ptr_t) do_loop_inversion,       "loop inversion",                          OPT_FLAG_NONE },
 348         { OPT_TARGET_IRG, "unroll-loops",    (func_ptr_t) do_loop_unrolling,       "loop unrolling",                          OPT_FLAG_NONE },
 349         { OPT_TARGET_IRG, "lower-mux",       (func_ptr_t) do_lower_mux,            "mux lowering",                            OPT_FLAG_NONE },
 350         { OPT_TARGET_IRG, "vrp",                         (func_ptr_t) do_vrp,                              "value range propagation",                             OPT_FLAG_NONE },
 351
 352 };
 353 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
 354 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
 355
 356 static opt_config_t *get_opt(const char *name)
 357 {
 358         int i;
 359         for (i = 0; i < n_opts; ++i) {
 360                 opt_config_t *config = &opts[i];
 361                 if (strcmp(config->name, name) == 0)
 362                         return config;
 363         }
 364
 365         return NULL;
 366 }
 367
 368 static void set_opt_enabled(const char *name, bool enabled)
 369 {
 370         opt_config_t *config = get_opt(name);
 371         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
 372                 | (enabled ? OPT_FLAG_ENABLED : 0);
 373 }
 374
 375 static bool get_opt_enabled(const char *name)
 376 {
 377         opt_config_t *config = get_opt(name);
 378         return (config->flags & OPT_FLAG_ENABLED) != 0;
 379 }
 380
 381 /**
 382  * perform an optimisation on a single graph
 383  *
 384  * @return  true if something changed, false otherwise
 385  */
 386 static bool do_irg_opt(ir_graph *irg, const char *name)
 387 {
 388         transform_irg_func  func;
 389         ir_graph           *old_irg;
 390         opt_config_t       *config = get_opt(name);
 391         size_t              n      = config - opts;
 392         assert(config != NULL);
 393         assert(config->target == OPT_TARGET_IRG);
 394         if (! (config->flags & OPT_FLAG_ENABLED))
 395                 return false;
 396
 397         old_irg          = current_ir_graph;
 398         current_ir_graph = irg;
 399
 400         func = (transform_irg_func) config->func;
 401
 402         timer_push(timers[n]);
 403         func(irg);
 404         timer_pop(timers[n]);
 405
 406         if (firm_dump.all_phases && firm_dump.ir_graph) {
 407                 dump_graph_count(irg, name);
 408         }
 409
 410         if (firm_opt.check_all) {
 411                 timer_push(t_verify);
 412                 irg_verify(irg, VRFY_ENFORCE_SSA);
 413                 timer_pop(t_verify);
 414         }
 415
 416         current_ir_graph = old_irg;
 417         return true;
 418 }
 419
 420 static void do_irp_opt(const char *name)
 421 {
 422         transform_irp_func  func;
 423         opt_config_t       *config = get_opt(name);
 424         size_t              n      = config - opts;
 425         assert(config->target == OPT_TARGET_IRP);
 426         if (! (config->flags & OPT_FLAG_ENABLED))
 427                 return;
 428
 429         func = (transform_irp_func) config->func;
 430
 431         timer_push(timers[n]);
 432         func();
 433         timer_pop(timers[n]);
 434
 435         if (firm_dump.ir_graph && firm_dump.all_phases) {
 436                 int i;
 437                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 438                         ir_graph *irg = get_irp_irg(i);
 439                         dump_graph_count(irg, name);
 440                 }
 441         }
 442
 443         if (firm_opt.check_all) {
 444                 int i;
 445                 timer_push(t_verify);
 446                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 447                         irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
 448                 }
 449                 timer_pop(t_verify);
 450         }
 451 }
 452
 453 /**
 454  * Enable transformations which should be always safe (and cheap) to perform
 455  */
 456 static void enable_safe_defaults(void)
 457 {
 458         set_opt_enabled("remove-unused", true);
 459         set_opt_enabled("opt-tail-rec", true);
 460         set_opt_enabled("opt-func-call", true);
 461         set_opt_enabled("reassociation", true);
 462         set_opt_enabled("control-flow", true);
 463         set_opt_enabled("local", true);
 464         set_opt_enabled("lower-const", true);
 465         set_opt_enabled("scalar-replace", true);
 466         set_opt_enabled("place", true);
 467         set_opt_enabled("confirm", true);
 468         set_opt_enabled("opt-load-store", true);
 469         set_opt_enabled("lower", true);
 470         set_opt_enabled("deconv", true);
 471         set_opt_enabled("remove-confirms", true);
 472         set_opt_enabled("ivopts", true);
 473         set_opt_enabled("dead", true);
 474         set_opt_enabled("lower-switch", true);
 475         set_opt_enabled("remove-phi-cycles", true);
 476 }
 477
 478 /**
 479  * run all the Firm optimizations
 480  *
 481  * @param input_filename     the name of the (main) source file
 482  */
 483 static void do_firm_optimizations(const char *input_filename)
 484 {
 485         int      i;
 486         unsigned aa_opt;
 487
 488         set_opt_alias_analysis(firm_opt.alias_analysis);
 489
 490         aa_opt = aa_opt_no_opt;
 491         if (firm_opt.strict_alias)
 492                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
 493         if (firm_opt.no_alias)
 494                 aa_opt = aa_opt_no_alias;
 495
 496         set_irp_memory_disambiguator_options(aa_opt);
 497
 498         /* parameter passing code should set them directly sometime... */
 499         set_opt_enabled("rts", !firm_opt.freestanding);
 500         set_opt_enabled("gcse", firm_opt.gcse);
 501         set_opt_enabled("place", !firm_opt.gcse);
 502         set_opt_enabled("confirm", firm_opt.confirm);
 503         set_opt_enabled("remove-confirms", firm_opt.confirm);
 504
 505         /* osr supersedes remove_phi_cycles */
 506         if (get_opt_enabled("ivopts"))
 507                 set_opt_enabled("remove-phi-cycles", false);
 508
 509         timer_start(t_all_opt);
 510
 511         do_irp_opt("rts");
 512
 513         /* first step: kill dead code */
 514         for (i = 0; i < get_irp_n_irgs(); i++) {
 515                 ir_graph *irg = get_irp_irg(i);
 516                 do_irg_opt(irg, "combo");
 517                 do_irg_opt(irg, "local");
 518                 do_irg_opt(irg, "control-flow");
 519         }
 520
 521         do_irp_opt("remove-unused");
 522         do_irp_opt("opt-tail-rec");
 523         do_irp_opt("opt-func-call");
 524         do_irp_opt("lower-const");
 525
 526         for (i = 0; i < get_irp_n_irgs(); i++) {
 527                 ir_graph *irg = get_irp_irg(i);
 528
 529                 do_irg_opt(irg, "scalar-replace");
 530                 do_irg_opt(irg, "invert-loops");
 531                 do_irg_opt(irg, "unroll-loops");
 532                 do_irg_opt(irg, "local");
 533                 do_irg_opt(irg, "reassociation");
 534                 do_irg_opt(irg, "local");
 535                 do_irg_opt(irg, "gcse");
 536
 537                 if (firm_opt.confirm) {
 538                         /* Confirm construction currently can only handle blocks with only
 539                            one control flow predecessor. Calling optimize_cf here removes
 540                            Bad predecessors and help the optimization of switch constructs.
 541                          */
 542                         do_irg_opt(irg, "control-flow");
 543                         do_irg_opt(irg, "confirm");
 544                         do_irg_opt(irg, "vrp");
 545                         do_irg_opt(irg, "local");
 546                 }
 547
 548                 do_irg_opt(irg, "control-flow");
 549                 do_irg_opt(irg, "opt-load-store");
 550                 do_irg_opt(irg, "lower");
 551                 do_irg_opt(irg, "deconv");
 552                 do_irg_opt(irg, "thread-jumps");
 553                 do_irg_opt(irg, "remove-confirms");
 554                 do_irg_opt(irg, "gvn-pre");
 555                 do_irg_opt(irg, "place");
 556                 do_irg_opt(irg, "control-flow");
 557
 558                 if (do_irg_opt(irg, "if-conversion")) {
 559                         do_irg_opt(irg, "local");
 560                         do_irg_opt(irg, "control-flow");
 561                 }
 562                 /* this doesn't make too much sense but tests the mux destruction... */
 563                 do_irg_opt(irg, "lower-mux");
 564
 565                 do_irg_opt(irg, "bool");
 566                 do_irg_opt(irg, "shape-blocks");
 567                 do_irg_opt(irg, "lower-switch");
 568                 do_irg_opt(irg, "ivopts");
 569                 do_irg_opt(irg, "local");
 570                 do_irg_opt(irg, "dead");
 571         }
 572
 573         do_irp_opt("inline");
 574         do_irp_opt("opt-proc-clone");
 575
 576         for (i = 0; i < get_irp_n_irgs(); i++) {
 577                 ir_graph *irg = get_irp_irg(i);
 578                 do_irg_opt(irg, "local");
 579                 do_irg_opt(irg, "control-flow");
 580                 do_irg_opt(irg, "thread-jumps");
 581                 do_irg_opt(irg, "local");
 582                 do_irg_opt(irg, "control-flow");
 583
 584                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
 585                         do_irg_opt(irg, "local");
 586                         do_irg_opt(irg, "vrp");
 587                         do_irg_opt(irg, "local");
 588                         do_irg_opt(irg, "vrp");
 589                 }
 590         }
 591
 592         if (firm_dump.ir_graph) {
 593                 /* recompute backedges for nicer dumps */
 594                 for (i = 0; i < get_irp_n_irgs(); i++)
 595                         construct_cf_backedges(get_irp_irg(i));
 596         }
 597
 598         dump_all("-opt");
 599
 600         if (firm_dump.statistic & STAT_AFTER_OPT)
 601                 stat_dump_snapshot(input_filename, "opt");
 602
 603         timer_stop(t_all_opt);
 604 }
 605
 606 /**
 607  * do Firm lowering
 608  *
 609  * @param input_filename  the name of the (main) source file
 610  */
 611 static void do_firm_lowering(const char *input_filename)
 612 {
 613         int i;
 614
 615         do_irp_opt("lower-dw");
 616
 617         if (firm_dump.statistic & STAT_AFTER_LOWER)
 618                 stat_dump_snapshot(input_filename, "low");
 619
 620         dump_all("-low");
 621
 622         if (firm_opt.enabled) {
 623                 timer_start(t_all_opt);
 624
 625                 /* run reassociation first on all graphs BEFORE the architecture
 626                    dependent optimizations are enabled */
 627                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 628                         ir_graph *irg = get_irp_irg(i);
 629                         do_irg_opt(irg, "reassociation");
 630                 }
 631
 632                 /* enable architecture dependent optimizations */
 633                 arch_dep_set_opts((arch_dep_opts_t)
 634                                 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
 635                                  (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
 636                                  (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
 637
 638                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 639                         ir_graph *irg = get_irp_irg(i);
 640
 641                         do_irg_opt(irg, "local");
 642                         do_irg_opt(irg, "gcse");
 643                         do_irg_opt(irg, "opt-load-store");
 644                         do_irg_opt(irg, "local");
 645                         do_irg_opt(irg, "control-flow");
 646
 647                         if (do_irg_opt(irg, "vrp")) {
 648                                 do_irg_opt(irg, "local");
 649                                 do_irg_opt(irg, "control-flow");
 650                                 do_irg_opt(irg, "vrp");
 651                                 do_irg_opt(irg, "local");
 652                                 do_irg_opt(irg, "control-flow");
 653                         }
 654
 655                         if (do_irg_opt(irg, "if-conversion")) {
 656                                 do_irg_opt(irg, "local");
 657                                 do_irg_opt(irg, "control-flow");
 658                         }
 659
 660                         do_irg_opt(irg, "parallelize-mem");
 661                 }
 662                 timer_stop(t_all_opt);
 663
 664                 do_irp_opt("remove-unused");
 665
 666                 dump_all("-low-opt");
 667         }
 668
 669         if (firm_opt.cc_opt)
 670                 mark_private_methods();
 671
 672         /* set the phase to low */
 673         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 674                 set_irg_phase_low(get_irp_irg(i));
 675
 676         /* all graphs are lowered, set the irp phase to low */
 677         set_irp_phase_state(phase_low);
 678
 679         if (firm_dump.statistic & STAT_FINAL) {
 680                 stat_dump_snapshot(input_filename, "final");
 681         }
 682 }
 683
 684 /**
 685  * Initialize for the Firm-generating back end.
 686  */
 687 void gen_firm_init(void)
 688 {
 689         firm_parameter_t params;
 690         unsigned         pattern = 0;
 691         int              i;
 692
 693         for (i = 0; i < n_opts; ++i) {
 694                 timers[i] = ir_timer_new();
 695                 timer_register(timers[i], opts[i].description);
 696         }
 697         t_verify = ir_timer_new();
 698         timer_register(t_verify, "Firm: verify pass");
 699         t_vcg_dump = ir_timer_new();
 700         timer_register(t_vcg_dump, "Firm: vcg dumping");
 701         t_all_opt = ir_timer_new();
 702         timer_register(t_all_opt, "Firm: all optimizations");
 703
 704         if (firm_dump.stat_pattern)
 705                 pattern |= FIRMSTAT_PATTERN_ENABLED;
 706
 707         if (firm_dump.stat_dag)
 708                 pattern |= FIRMSTAT_COUNT_DAG;
 709
 710         memset(&params, 0, sizeof(params));
 711         params.size                  = sizeof(params);
 712         params.enable_statistics     = firm_dump.statistic == STAT_NONE ? 0 :
 713                 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS
 714                 | pattern;
 715         params.initialize_local_func = uninitialized_local_var;
 716         params.cc_mask               = 0; /* no regparam, cdecl */
 717
 718         ir_init(&params);
 719
 720         if (firm_be_opt.selection == BE_FIRM_BE) {
 721                 const backend_params *be_params = be_get_backend_param();
 722
 723                 if (be_params->do_dw_lowering)
 724                         set_opt_enabled("lower-dw", true);
 725
 726                 arch_create_intrinsic   = be_params->arch_create_intrinsic_fkt;
 727                 create_intrinsic_ctx    = be_params->create_intrinsic_ctx;
 728
 729                 ad_param                = be_params->dep_param;
 730                 if_conv_info            = be_params->if_conv_info;
 731         }
 732
 733         edges_init_dbg(firm_opt.vrfy_edges);
 734
 735         /* Sel node cannot produce NULL pointers */
 736         set_opt_sel_based_null_check_elim(1);
 737
 738         /* dynamic dispatch works currently only if whole world scenarios */
 739         set_opt_dyn_meth_dispatch(0);
 740
 741         arch_dep_init(arch_factory);
 742
 743         /* do not run architecture dependent optimizations in building phase */
 744         arch_dep_set_opts(arch_dep_none);
 745
 746         do_node_verification((firm_verification_t) firm_opt.vrfy);
 747         if (firm_dump.filter)
 748                 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
 749
 750         if (firm_opt.enabled) {
 751                 set_optimize(1);
 752                 set_opt_constant_folding(firm_opt.const_folding);
 753                 set_opt_algebraic_simplification(firm_opt.const_folding);
 754                 set_opt_cse(firm_opt.cse);
 755                 set_opt_global_cse(0);
 756                 set_opt_unreachable_code(1);
 757                 set_opt_control_flow(firm_opt.control_flow);
 758                 set_opt_control_flow_weak_simplification(1);
 759                 set_opt_control_flow_strong_simplification(1);
 760         } else {
 761                 set_optimize(0);
 762         }
 763
 764         /* do not dump entity ld names */
 765         dump_ld_names(0);
 766 }
 767
 768 /**
 769  * Called, after the Firm generation is completed,
 770  * do all optimizations and backend call here.
 771  *
 772  * @param out                a file handle for the output, may be NULL
 773  * @param input_filename     the name of the (main) source file
 774  * @param c_mode             non-zero if "C" was compiled
 775  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
 776  */
 777 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode,
 778                      int new_firm_const_exists)
 779 {
 780         int i;
 781
 782 #if 0
 783         if (firm_opt.enable_statev) {
 784                 char buf[1024];
 785                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
 786                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
 787                 ir_stat_ev_compilation_unit(input_filename);
 788         }
 789 #endif
 790
 791         firm_const_exists = new_firm_const_exists;
 792
 793         /* the general for dumping option must be set, or the others will not work*/
 794         firm_dump.ir_graph
 795                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
 796
 797         dump_keepalive_edges(1);
 798         dump_consts_local(1);
 799         dump_dominator_information(1);
 800         dump_loop_information(0);
 801
 802         if (!firm_dump.edge_labels)
 803                 turn_off_edge_labels();
 804
 805         /* FIXME: cloning might ADD new graphs. */
 806         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
 807
 808         if (firm_dump.all_types) {
 809                 dump_all_types("");
 810                 if (! c_mode) {
 811                         dump_class_hierarchy(0, "");
 812                         dump_class_hierarchy(1, "-with-entities");
 813                 }
 814         }
 815
 816         /* finalize all graphs */
 817         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 818                 ir_graph *irg = get_irp_irg(i);
 819                 irg_finalize_cons(irg);
 820         }
 821         dump_all("");
 822
 823         timer_push(t_verify);
 824         tr_vrfy();
 825         timer_pop(t_verify);
 826
 827         /* all graphs are finalized, set the irp phase to high */
 828         set_irp_phase_state(phase_high);
 829
 830         /* BEWARE: kill unreachable code before doing compound lowering */
 831         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 832                 ir_graph *irg = get_irp_irg(i);
 833                 do_irg_opt(irg, "control-flow");
 834         }
 835
 836         /* lower all compound call return values */
 837         lower_compound_params();
 838
 839         /* lower copyb nodes */
 840         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
 841                 ir_graph *irg = get_irp_irg(i);
 842                 lower_CopyB(irg, 128, 4);
 843         }
 844
 845         if (firm_dump.statistic & STAT_BEFORE_OPT) {
 846                 stat_dump_snapshot(input_filename, "noopt");
 847         }
 848
 849         if (firm_opt.enabled)
 850                 do_firm_optimizations(input_filename);
 851
 852         if (firm_opt.lower)
 853                 do_firm_lowering(input_filename);
 854
 855         /* set the phase to low */
 856         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
 857                 set_irg_phase_low(get_irp_irg(i));
 858
 859         if (firm_dump.statistic & STAT_FINAL_IR)
 860                 stat_dump_snapshot(input_filename, "final-ir");
 861
 862         /* run the code generator */
 863         if (firm_be_opt.selection != BE_NONE)
 864                 do_codegen(out, input_filename);
 865
 866         if (firm_dump.statistic & STAT_FINAL)
 867                 stat_dump_snapshot(input_filename, "final");
 868 }
 869
 870 void disable_all_opts(void)
 871 {
 872         for (int i = 0; i < n_opts; ++i) {
 873                 opt_config_t *config = &opts[i];
 874                 if (config->flags & OPT_FLAG_ESSENTIAL) {
 875                         config->flags |= OPT_FLAG_ENABLED;
 876                 } else {
 877                         config->flags &= ~OPT_FLAG_ENABLED;
 878                 }
 879         }
 880 }
 881
 882 int firm_opt_option(const char *opt)
 883 {
 884         bool enable = true;
 885         if (strncmp(opt, "no-", 3) == 0) {
 886                 enable = false;
 887                 opt = opt + 3;
 888         }
 889
 890         opt_config_t *config = get_opt(opt);
 891         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
 892                 return 0;
 893
 894         config->flags &= ~OPT_FLAG_ENABLED;
 895         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
 896         return 1;
 897 }
 898
 899 void firm_opt_option_help(void)
 900 {
 901         int i;
 902
 903         for (i = 0; i < n_opts; ++i) {
 904                 char buf[1024];
 905                 char buf2[1024];
 906
 907                 const opt_config_t *config = &opts[i];
 908                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
 909                         continue;
 910
 911                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
 912                 print_option_help(config->name, buf2);
 913                 snprintf(buf, sizeof(buf), "no-%s", config->name);
 914                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
 915                 print_option_help(buf, buf2);
 916         }
 917 }
 918
 919 /**
 920  * Do very early initializations
 921  */
 922 void firm_early_init(void)
 923 {
 924         /* arg: need this here for command line options */
 925         be_opt_register();
 926
 927         enable_safe_defaults();
 928 }