3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2010 Michael Beck beck@ipd.info.uni-karlsruhe.de
17 #include <libfirm/firm.h>
20 #include "firm_codegen.h"
21 #include "firm_cmdline.h"
22 #include "firm_timing.h"
25 static ir_timer_t *t_vcg_dump;
26 static ir_timer_t *t_verify;
27 static ir_timer_t *t_all_opt;
28 static bool do_irg_opt(ir_graph *irg, const char *name);
30 /** dump all the graphs depending on cond */
32 static void dump_all(const char *suffix)
34 if (!firm_dump.ir_graph)
37 timer_push(t_vcg_dump);
38 dump_all_ir_graphs(suffix);
39 timer_pop(t_vcg_dump);
42 /* entities of runtime functions */
43 ir_entity_ptr rts_entities[rts_max];
46 * Map runtime functions.
48 static void rts_map(void)
51 ir_entity_ptr *ent; /**< address of the rts entity */
52 i_mapper_func func; /**< mapper function. */
55 { &rts_entities[rts_abs], i_mapper_abs },
56 { &rts_entities[rts_labs], i_mapper_abs },
57 { &rts_entities[rts_llabs], i_mapper_abs },
58 { &rts_entities[rts_imaxabs], i_mapper_abs },
60 /* double -> double */
61 { &rts_entities[rts_fabs], i_mapper_abs },
62 { &rts_entities[rts_sqrt], i_mapper_sqrt },
63 { &rts_entities[rts_cbrt], i_mapper_cbrt },
64 { &rts_entities[rts_pow], i_mapper_pow },
65 { &rts_entities[rts_exp], i_mapper_exp },
66 { &rts_entities[rts_exp2], i_mapper_exp },
67 { &rts_entities[rts_exp10], i_mapper_exp },
68 { &rts_entities[rts_log], i_mapper_log },
69 { &rts_entities[rts_log2], i_mapper_log2 },
70 { &rts_entities[rts_log10], i_mapper_log10 },
71 { &rts_entities[rts_sin], i_mapper_sin },
72 { &rts_entities[rts_cos], i_mapper_cos },
73 { &rts_entities[rts_tan], i_mapper_tan },
74 { &rts_entities[rts_asin], i_mapper_asin },
75 { &rts_entities[rts_acos], i_mapper_acos },
76 { &rts_entities[rts_atan], i_mapper_atan },
77 { &rts_entities[rts_sinh], i_mapper_sinh },
78 { &rts_entities[rts_cosh], i_mapper_cosh },
79 { &rts_entities[rts_tanh], i_mapper_tanh },
82 { &rts_entities[rts_fabsf], i_mapper_abs },
83 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
84 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
85 { &rts_entities[rts_powf], i_mapper_pow },
86 { &rts_entities[rts_expf], i_mapper_exp },
87 { &rts_entities[rts_exp2f], i_mapper_exp },
88 { &rts_entities[rts_exp10f], i_mapper_exp },
89 { &rts_entities[rts_logf], i_mapper_log },
90 { &rts_entities[rts_log2f], i_mapper_log2 },
91 { &rts_entities[rts_log10f], i_mapper_log10 },
92 { &rts_entities[rts_sinf], i_mapper_sin },
93 { &rts_entities[rts_cosf], i_mapper_cos },
94 { &rts_entities[rts_tanf], i_mapper_tan },
95 { &rts_entities[rts_asinf], i_mapper_asin },
96 { &rts_entities[rts_acosf], i_mapper_acos },
97 { &rts_entities[rts_atanf], i_mapper_atan },
98 { &rts_entities[rts_sinhf], i_mapper_sinh },
99 { &rts_entities[rts_coshf], i_mapper_cosh },
100 { &rts_entities[rts_tanhf], i_mapper_tanh },
102 /* long double -> long double */
103 { &rts_entities[rts_fabsl], i_mapper_abs },
104 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
105 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
106 { &rts_entities[rts_powl], i_mapper_pow },
107 { &rts_entities[rts_expl], i_mapper_exp },
108 { &rts_entities[rts_exp2l], i_mapper_exp },
109 { &rts_entities[rts_exp10l], i_mapper_exp },
110 { &rts_entities[rts_logl], i_mapper_log },
111 { &rts_entities[rts_log2l], i_mapper_log2 },
112 { &rts_entities[rts_log10l], i_mapper_log10 },
113 { &rts_entities[rts_sinl], i_mapper_sin },
114 { &rts_entities[rts_cosl], i_mapper_cos },
115 { &rts_entities[rts_tanl], i_mapper_tan },
116 { &rts_entities[rts_asinl], i_mapper_asin },
117 { &rts_entities[rts_acosl], i_mapper_acos },
118 { &rts_entities[rts_atanl], i_mapper_atan },
119 { &rts_entities[rts_sinhl], i_mapper_sinh },
120 { &rts_entities[rts_coshl], i_mapper_cosh },
121 { &rts_entities[rts_tanhl], i_mapper_tanh },
124 { &rts_entities[rts_strcmp], i_mapper_strcmp },
125 { &rts_entities[rts_strncmp], i_mapper_strncmp },
126 { &rts_entities[rts_strcpy], i_mapper_strcpy },
127 { &rts_entities[rts_strlen], i_mapper_strlen },
128 { &rts_entities[rts_memcpy], i_mapper_memcpy },
129 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
130 { &rts_entities[rts_memmove], i_mapper_memmove },
131 { &rts_entities[rts_memset], i_mapper_memset },
132 { &rts_entities[rts_memcmp], i_mapper_memcmp }
134 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
137 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
138 if (*mapper[i].ent != NULL) {
139 rec[n_map].i_call.kind = INTRINSIC_CALL;
140 rec[n_map].i_call.i_ent = *mapper[i].ent;
141 rec[n_map].i_call.i_mapper = mapper[i].func;
142 rec[n_map].i_call.ctx = NULL;
143 rec[n_map].i_call.link = NULL;
149 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
152 static int *irg_dump_no;
154 static void do_lower_highlevel(ir_graph *irg)
156 lower_highlevel_graph(irg, false);
159 static void do_stred(ir_graph *irg)
161 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
164 static void after_inline_opt(ir_graph *irg)
166 do_irg_opt(irg, "scalar-replace");
167 do_irg_opt(irg, "local");
168 do_irg_opt(irg, "control-flow");
169 do_irg_opt(irg, "combo");
172 static void do_inline(void)
174 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
178 static void do_cloning(void)
180 proc_cloning((float) firm_opt.clone_threshold);
183 static void do_lower_mux(ir_graph *irg)
185 lower_mux(irg, NULL);
188 static void do_gcse(ir_graph *irg)
190 set_opt_global_cse(1);
191 optimize_graph_df(irg);
192 set_opt_global_cse(0);
195 static void lower_blockcopy(ir_graph *irg)
197 lower_CopyB(irg, 128, 4);
200 typedef enum opt_target {
201 OPT_TARGET_IRG, /**< optimization function works on a single graph */
202 OPT_TARGET_IRP /**< optimization function works on the complete program */
205 typedef enum opt_flags {
207 OPT_FLAG_ENABLED = 1 << 0, /**< enable the optimization */
208 OPT_FLAG_NO_DUMP = 1 << 1, /**< don't dump after transformation */
209 OPT_FLAG_NO_VERIFY = 1 << 2, /**< don't verify after transformation */
210 OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
211 -foptions for this transformation */
212 OPT_FLAG_ESSENTIAL = 1 << 4, /**< output won't work without this pass
213 so we need it even with -O0 */
216 typedef void (*transform_irg_func)(ir_graph *irg);
217 typedef void (*transform_irp_func)(void);
223 transform_irg_func transform_irg;
224 transform_irp_func transform_irp;
226 const char *description;
230 static opt_config_t opts[] = {
231 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
232 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b, c, d }
233 IRG("bool", opt_bool, "bool simplification", OPT_FLAG_NONE),
234 IRG("combo", combo, "combined CCE, UCE and GVN", OPT_FLAG_NONE),
235 IRG("confirm", construct_confirms, "confirm optimisation", OPT_FLAG_HIDE_OPTIONS),
236 IRG("control-flow", optimize_cf, "optimization of control-flow", OPT_FLAG_HIDE_OPTIONS),
237 IRG("dead", dead_node_elimination, "dead node elimination", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
238 IRG("deconv", conv_opt, "conv node elimination", OPT_FLAG_NONE),
239 IRG("fp-vrp", fixpoint_vrp, "fixpoint value range propagation", OPT_FLAG_NONE),
240 IRG("frame", opt_frame_irg, "remove unused frame entities", OPT_FLAG_NONE),
241 IRG("gvn-pre", do_gvn_pre, "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
242 IRG("if-conversion", opt_if_conv, "if-conversion", OPT_FLAG_NONE),
243 IRG("invert-loops", do_loop_inversion, "loop inversion", OPT_FLAG_NONE),
244 IRG("ivopts", do_stred, "induction variable strength reduction", OPT_FLAG_NONE),
245 IRG("local", optimize_graph_df, "local graph optimizations", OPT_FLAG_HIDE_OPTIONS),
246 IRG("lower", do_lower_highlevel, "lowering", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
247 IRG("lower-mux", do_lower_mux, "mux lowering", OPT_FLAG_NONE),
248 IRG("opt-load-store", optimize_load_store, "load store optimization", OPT_FLAG_NONE),
249 IRG("opt-tail-rec", opt_tail_rec_irg, "tail-recursion eliminiation", OPT_FLAG_NONE),
250 IRG("parallelize-mem", opt_parallelize_mem, "parallelize memory", OPT_FLAG_NONE),
251 IRG("gcse", do_gcse, "global common subexpression eliminiation", OPT_FLAG_NONE),
252 IRG("place", place_code, "code placement", OPT_FLAG_NONE),
253 IRG("reassociation", optimize_reassociation, "reassociation", OPT_FLAG_NONE),
254 IRG("remove-confirms", remove_confirms, "confirm removal", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
255 IRG("remove-phi-cycles", remove_phi_cycles, "removal of phi cycles", OPT_FLAG_HIDE_OPTIONS),
256 IRG("scalar-replace", scalar_replacement_opt, "scalar replacement", OPT_FLAG_NONE),
257 IRG("shape-blocks", shape_blocks, "block shaping", OPT_FLAG_NONE),
258 IRG("thread-jumps", opt_jumpthreading, "path-sensitive jumpthreading", OPT_FLAG_NONE),
259 IRG("unroll-loops", do_loop_unrolling, "loop unrolling", OPT_FLAG_NONE),
260 IRG("vrp", set_vrp_data, "value range propagation", OPT_FLAG_NONE),
261 IRP("inline", do_inline, "inlining", OPT_FLAG_NONE),
262 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),
263 IRG("lower-blockcopy", lower_blockcopy, "replace small block copies with load/store sequences", OPT_FLAG_NO_DUMP),
264 IRP("target-lowering", be_lower_for_target, "lowering necessary for target architecture", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
265 IRP("opt-func-call", optimize_funccalls, "function call optimization", OPT_FLAG_NONE),
266 IRP("opt-proc-clone", do_cloning, "procedure cloning", OPT_FLAG_NONE),
267 IRP("remove-unused", garbage_collect_entities, "removal of unused functions/variables", OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
268 IRP("rts", rts_map, "optimization of known library functions", OPT_FLAG_NONE),
269 IRP("opt-cc", mark_private_methods, "calling conventions optimization", OPT_FLAG_NONE),
273 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
274 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
276 static opt_config_t *get_opt(const char *name)
279 for (i = 0; i < n_opts; ++i) {
280 opt_config_t *config = &opts[i];
281 if (strcmp(config->name, name) == 0)
288 static void set_opt_enabled(const char *name, bool enabled)
290 opt_config_t *config = get_opt(name);
291 config->flags = (config->flags & ~OPT_FLAG_ENABLED)
292 | (enabled ? OPT_FLAG_ENABLED : 0);
295 static bool get_opt_enabled(const char *name)
297 opt_config_t *config = get_opt(name);
298 return (config->flags & OPT_FLAG_ENABLED) != 0;
302 * perform an optimisation on a single graph
304 * @return true if something changed, false otherwise
306 static bool do_irg_opt(ir_graph *irg, const char *name)
309 opt_config_t *config = get_opt(name);
310 size_t n = config - opts;
311 assert(config != NULL);
312 assert(config->target == OPT_TARGET_IRG);
313 if (! (config->flags & OPT_FLAG_ENABLED))
316 old_irg = current_ir_graph;
317 current_ir_graph = irg;
319 timer_push(timers[n]);
320 config->u.transform_irg(irg);
321 timer_pop(timers[n]);
323 if (firm_dump.all_phases && firm_dump.ir_graph) {
324 dump_ir_graph(irg, name);
327 if (firm_opt.check_all) {
328 timer_push(t_verify);
329 irg_verify(irg, VERIFY_ENFORCE_SSA);
333 current_ir_graph = old_irg;
337 static void do_irp_opt(const char *name)
339 opt_config_t *config = get_opt(name);
340 size_t n = config - opts;
341 assert(config->target == OPT_TARGET_IRP);
342 if (! (config->flags & OPT_FLAG_ENABLED))
345 timer_push(timers[n]);
346 config->u.transform_irp();
347 timer_pop(timers[n]);
349 if (firm_dump.ir_graph && firm_dump.all_phases) {
351 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
352 ir_graph *irg = get_irp_irg(i);
353 dump_ir_graph(irg, name);
357 if (firm_opt.check_all) {
359 timer_push(t_verify);
360 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
361 irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
368 * Enable transformations which should be always safe (and cheap) to perform
370 static void enable_safe_defaults(void)
372 set_opt_enabled("remove-unused", true);
373 set_opt_enabled("opt-tail-rec", true);
374 set_opt_enabled("opt-func-call", true);
375 set_opt_enabled("reassociation", true);
376 set_opt_enabled("control-flow", true);
377 set_opt_enabled("local", true);
378 set_opt_enabled("lower-const", true);
379 set_opt_enabled("scalar-replace", true);
380 set_opt_enabled("place", true);
381 set_opt_enabled("gcse", true);
382 set_opt_enabled("confirm", true);
383 set_opt_enabled("opt-load-store", true);
384 set_opt_enabled("lower", true);
385 set_opt_enabled("lower-blockcopy", true);
386 set_opt_enabled("deconv", true);
387 set_opt_enabled("remove-confirms", true);
388 set_opt_enabled("ivopts", true);
389 set_opt_enabled("dead", true);
390 set_opt_enabled("remove-phi-cycles", true);
391 set_opt_enabled("frame", true);
392 set_opt_enabled("combo", true);
393 set_opt_enabled("invert-loops", true);
394 set_opt_enabled("target-lowering", true);
395 set_opt_enabled("rts", true);
396 set_opt_enabled("parallelize-mem", true);
397 set_opt_enabled("opt-cc", true);
401 * run all the Firm optimizations
403 * @param input_filename the name of the (main) source file
405 static void do_firm_optimizations(const char *input_filename)
410 set_opt_alias_analysis(firm_opt.alias_analysis);
412 aa_opt = aa_opt_no_opt;
413 if (firm_opt.strict_alias)
414 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
415 if (firm_opt.no_alias)
416 aa_opt = aa_opt_no_alias;
418 set_irp_memory_disambiguator_options(aa_opt);
420 /* parameter passing code should set them directly sometime... */
421 set_opt_enabled("confirm", firm_opt.confirm);
422 set_opt_enabled("remove-confirms", firm_opt.confirm);
424 /* osr supersedes remove_phi_cycles */
425 if (get_opt_enabled("ivopts"))
426 set_opt_enabled("remove-phi-cycles", false);
428 timer_start(t_all_opt);
432 /* first step: kill dead code */
433 for (i = 0; i < get_irp_n_irgs(); i++) {
434 ir_graph *irg = get_irp_irg(i);
435 do_irg_opt(irg, "combo");
436 do_irg_opt(irg, "local");
437 do_irg_opt(irg, "control-flow");
440 do_irp_opt("remove-unused");
441 for (i = 0; i < get_irp_n_irgs(); ++i) {
442 ir_graph *irg = get_irp_irg(i);
443 do_irg_opt(irg, "opt-tail-rec");
445 do_irp_opt("opt-func-call");
446 do_irp_opt("lower-const");
448 for (i = 0; i < get_irp_n_irgs(); i++) {
449 ir_graph *irg = get_irp_irg(i);
451 do_irg_opt(irg, "scalar-replace");
452 do_irg_opt(irg, "invert-loops");
453 do_irg_opt(irg, "unroll-loops");
454 do_irg_opt(irg, "local");
455 do_irg_opt(irg, "reassociation");
456 do_irg_opt(irg, "local");
457 do_irg_opt(irg, "gcse");
458 do_irg_opt(irg, "place");
460 if (firm_opt.confirm) {
461 /* Confirm construction currently can only handle blocks with only
462 one control flow predecessor. Calling optimize_cf here removes
463 Bad predecessors and help the optimization of switch constructs.
465 do_irg_opt(irg, "control-flow");
466 do_irg_opt(irg, "confirm");
467 do_irg_opt(irg, "vrp");
468 do_irg_opt(irg, "local");
471 do_irg_opt(irg, "control-flow");
472 do_irg_opt(irg, "opt-load-store");
473 do_irg_opt(irg, "fp-vrp");
474 do_irg_opt(irg, "lower");
475 do_irg_opt(irg, "deconv");
476 do_irg_opt(irg, "thread-jumps");
477 do_irg_opt(irg, "remove-confirms");
478 do_irg_opt(irg, "gvn-pre");
479 do_irg_opt(irg, "gcse");
480 do_irg_opt(irg, "place");
481 do_irg_opt(irg, "control-flow");
483 if (do_irg_opt(irg, "if-conversion")) {
484 do_irg_opt(irg, "local");
485 do_irg_opt(irg, "control-flow");
487 /* this doesn't make too much sense but tests the mux destruction... */
488 do_irg_opt(irg, "lower-mux");
490 do_irg_opt(irg, "bool");
491 do_irg_opt(irg, "shape-blocks");
492 do_irg_opt(irg, "ivopts");
493 do_irg_opt(irg, "local");
494 do_irg_opt(irg, "dead");
497 do_irp_opt("inline");
498 do_irp_opt("opt-proc-clone");
500 for (i = 0; i < get_irp_n_irgs(); i++) {
501 ir_graph *irg = get_irp_irg(i);
502 do_irg_opt(irg, "local");
503 do_irg_opt(irg, "control-flow");
504 do_irg_opt(irg, "thread-jumps");
505 do_irg_opt(irg, "local");
506 do_irg_opt(irg, "control-flow");
508 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
509 do_irg_opt(irg, "local");
510 do_irg_opt(irg, "vrp");
511 do_irg_opt(irg, "local");
512 do_irg_opt(irg, "vrp");
516 if (firm_dump.ir_graph) {
517 /* recompute backedges for nicer dumps */
518 for (i = 0; i < get_irp_n_irgs(); i++)
519 construct_cf_backedges(get_irp_irg(i));
524 if (firm_dump.statistic & STAT_AFTER_OPT)
525 stat_dump_snapshot(input_filename, "opt");
527 timer_stop(t_all_opt);
533 * @param input_filename the name of the (main) source file
535 static void do_firm_lowering(const char *input_filename)
539 /* enable architecture dependent optimizations */
540 arch_dep_set_opts((arch_dep_opts_t)
541 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
542 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
543 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
544 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
545 ir_graph *irg = get_irp_irg(i);
546 do_irg_opt(irg, "reassociation");
547 do_irg_opt(irg, "local");
550 do_irp_opt("target-lowering");
552 if (firm_dump.statistic & STAT_AFTER_LOWER)
553 stat_dump_snapshot(input_filename, "low");
555 timer_start(t_all_opt);
557 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
558 ir_graph *irg = get_irp_irg(i);
560 do_irg_opt(irg, "local");
561 do_irg_opt(irg, "deconv");
562 do_irg_opt(irg, "control-flow");
563 do_irg_opt(irg, "opt-load-store");
564 do_irg_opt(irg, "gcse");
565 do_irg_opt(irg, "place");
566 do_irg_opt(irg, "control-flow");
568 if (do_irg_opt(irg, "vrp")) {
569 do_irg_opt(irg, "local");
570 do_irg_opt(irg, "control-flow");
571 do_irg_opt(irg, "vrp");
572 do_irg_opt(irg, "local");
573 do_irg_opt(irg, "control-flow");
576 if (do_irg_opt(irg, "if-conversion")) {
577 do_irg_opt(irg, "local");
578 do_irg_opt(irg, "control-flow");
581 do_irg_opt(irg, "parallelize-mem");
582 do_irg_opt(irg, "frame");
584 do_irp_opt("remove-unused");
585 do_irp_opt("opt-cc");
586 timer_stop(t_all_opt);
589 if (firm_dump.statistic & STAT_FINAL) {
590 stat_dump_snapshot(input_filename, "final");
595 * Initialize for the Firm-generating back end.
597 void gen_firm_init(void)
599 unsigned pattern = 0;
602 for (i = 0; i < n_opts; ++i) {
603 timers[i] = ir_timer_new();
604 timer_register(timers[i], opts[i].description);
606 t_verify = ir_timer_new();
607 timer_register(t_verify, "Firm: verify pass");
608 t_vcg_dump = ir_timer_new();
609 timer_register(t_vcg_dump, "Firm: vcg dumping");
610 t_all_opt = ir_timer_new();
611 timer_register(t_all_opt, "Firm: all optimizations");
613 if (firm_dump.stat_pattern)
614 pattern |= FIRMSTAT_PATTERN_ENABLED;
616 if (firm_dump.stat_dag)
617 pattern |= FIRMSTAT_COUNT_DAG;
620 firm_init_stat(firm_dump.statistic == STAT_NONE ?
621 0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
622 | FIRMSTAT_COUNT_CONSTS | pattern);
624 edges_init_dbg(firm_opt.verify_edges);
626 /* Sel node cannot produce NULL pointers */
627 set_opt_sel_based_null_check_elim(1);
629 /* dynamic dispatch works currently only if whole world scenarios */
630 set_opt_dyn_meth_dispatch(0);
632 /* do not run architecture dependent optimizations in building phase */
633 arch_dep_set_opts(arch_dep_none);
635 do_node_verification((firm_verification_t) firm_opt.verify);
636 if (firm_dump.filter != NULL)
637 ir_set_dump_filter(firm_dump.filter);
639 ir_add_dump_flags(ir_dump_flag_group_extbb);
640 if (firm_dump.no_blocks)
641 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
644 set_opt_constant_folding(firm_opt.const_folding);
645 set_opt_algebraic_simplification(firm_opt.const_folding);
646 set_opt_cse(firm_opt.cse);
647 set_opt_global_cse(0);
648 set_opt_unreachable_code(1);
652 * Called, after the Firm generation is completed,
653 * do all optimizations and backend call here.
655 * @param out a file handle for the output, may be NULL
656 * @param input_filename the name of the (main) source file
657 * @param c_mode non-zero if "C" was compiled
659 void gen_firm_finish(FILE *out, const char *input_filename)
663 /* the general for dumping option must be set, or the others will not work*/
665 = (bool) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
667 ir_add_dump_flags(ir_dump_flag_keepalive_edges
668 | ir_dump_flag_consts_local | ir_dump_flag_dominance);
669 ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
671 /* FIXME: cloning might ADD new graphs. */
672 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
674 if (firm_dump.all_types) {
675 dump_ir_prog_ext(dump_typegraph, "types.vcg");
678 /* finalize all graphs */
679 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
680 ir_graph *irg = get_irp_irg(i);
681 irg_finalize_cons(irg);
685 timer_push(t_verify);
689 /* all graphs are finalized, set the irp phase to high */
690 set_irp_phase_state(phase_high);
692 /* BEWARE: kill unreachable code before doing compound lowering */
693 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
694 ir_graph *irg = get_irp_irg(i);
695 do_irg_opt(irg, "control-flow");
696 do_irg_opt(irg, "lower-blockcopy");
699 if (firm_dump.statistic & STAT_BEFORE_OPT) {
700 stat_dump_snapshot(input_filename, "noopt");
703 do_firm_optimizations(input_filename);
704 do_firm_lowering(input_filename);
706 /* set the phase to low */
707 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
708 set_irg_phase_state(get_irp_irg(i), phase_low);
710 if (firm_dump.statistic & STAT_FINAL_IR)
711 stat_dump_snapshot(input_filename, "final-ir");
713 /* run the code generator */
714 ir_timer_t *timer = ir_timer_new();
715 timer_register(timer, "Firm: backend");
717 be_main(out, input_filename);
720 if (firm_dump.statistic & STAT_FINAL)
721 stat_dump_snapshot(input_filename, "final");
724 void disable_all_opts(void)
726 for (int i = 0; i < n_opts; ++i) {
727 opt_config_t *config = &opts[i];
728 if (config->flags & OPT_FLAG_ESSENTIAL) {
729 config->flags |= OPT_FLAG_ENABLED;
731 config->flags &= ~OPT_FLAG_ENABLED;
736 int firm_opt_option(const char *opt)
739 if (strncmp(opt, "no-", 3) == 0) {
744 opt_config_t *config = get_opt(opt);
745 if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
748 config->flags &= ~OPT_FLAG_ENABLED;
749 config->flags |= enable ? OPT_FLAG_ENABLED : 0;
753 void firm_opt_option_help(void)
757 for (i = 0; i < n_opts; ++i) {
761 const opt_config_t *config = &opts[i];
762 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
765 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
766 print_option_help(config->name, buf2);
767 snprintf(buf, sizeof(buf), "no-%s", config->name);
768 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
769 print_option_help(buf, buf2);
773 static void set_be_option(const char *arg)
775 int res = be_parse_arg(arg);
780 static void set_option(const char *arg)
782 int res = firm_option(arg);
787 void choose_optimization_pack(int level)
789 /* apply optimisation level */
792 set_option("no-opt");
795 set_option("no-inline");
799 /* use_builtins = true; */
802 set_option("thread-jumps");
803 set_option("if-conversion");
806 set_option("strict-aliasing");
807 set_option("inline");
808 set_option("fp-vrp");
809 set_option("deconv");
810 set_be_option("omitfp");
816 * Do very early initializations
818 void firm_early_init(void)
820 /* arg: need this here for command line options */
823 enable_safe_defaults();