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 #if defined(_DEBUG) || defined(FIRM_DEBUG)
26 #define DBG(x) dbg_printf x
28 #define DBG(x) ((void)0)
29 #endif /* _DEBUG || FIRM_DEBUG */
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);
36 /** dump all the graphs depending on cond */
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);
46 dump_all_ir_graphs(dump_ir_block_graph, suffix);
47 timer_pop(t_vcg_dump);
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;
57 /* entities of runtime functions */
58 ir_entity_ptr rts_entities[rts_max];
61 * factory for setting architecture dependent parameters
63 static const ir_settings_arch_dep_t *arch_factory(void)
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 */
73 32 /* Mulh allowed up to 32 bit */
76 return ad_param ? ad_param : ¶m;
80 * Map runtime functions.
82 static void rts_map(void) {
84 ir_entity_ptr *ent; /**< address of the rts entity */
85 i_mapper_func func; /**< mapper function. */
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 },
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 },
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 },
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 },
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 }
167 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
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;
182 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
185 static int *irg_dump_no;
187 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
190 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++,
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);
199 dump_ir_block_graph(irg, name);
200 timer_pop(t_vcg_dump);
203 static int firm_const_exists;
205 static void do_optimize_funccalls(void)
207 optimize_funccalls(firm_const_exists, NULL);
210 static void do_gcse(ir_graph *irg)
212 set_opt_global_cse(1);
213 optimize_graph_df(irg);
215 set_opt_global_cse(0);
218 static void do_lower_highlevel(ir_graph *irg)
220 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
223 static void do_if_conv(ir_graph *irg)
225 opt_if_conv(irg, if_conv_info);
228 static void do_stred(ir_graph *irg)
230 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
233 static void after_inline_opt(ir_graph *irg)
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");
241 static void do_inline(void)
243 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
247 static void do_cloning(void)
249 proc_cloning((float) firm_opt.clone_threshold);
252 static void do_lower_switch(ir_graph *irg)
254 lower_switch(irg, firm_opt.spare_size);
257 static void do_lower_mux(ir_graph *irg)
259 lower_mux(irg, NULL);
262 static void do_lower_dw_ops(void)
264 lwrdw_param_t init = {
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,
275 if (arch_create_intrinsic) {
276 init.create_intrinsic = arch_create_intrinsic;
277 init.ctx = create_intrinsic_ctx;
282 static void do_vrp(ir_graph *irg)
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 */
292 typedef enum opt_flags {
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 */
303 typedef void (*transform_irg_func)(ir_graph *irg);
304 typedef void (*transform_irp_func)(void);
310 transform_irg_func transform_irg;
311 transform_irp_func transform_irp;
313 const char *description;
317 static opt_config_t opts[] = {
318 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
319 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b, c, d }
320 IRG("bool", opt_bool, "bool simplification", OPT_FLAG_NONE),
321 IRG("combo", combo, "combined CCE, UCE and GVN", OPT_FLAG_NONE),
322 IRG("confirm", construct_confirms, "confirm optimisation", OPT_FLAG_HIDE_OPTIONS),
323 IRG("control-flow", optimize_cf, "optimization of control-flow", OPT_FLAG_HIDE_OPTIONS),
324 IRG("dead", dead_node_elimination, "dead node elimination", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
325 IRG("deconv", conv_opt, "conv node elimination", OPT_FLAG_NONE),
326 IRG("fp-vrp", fixpoint_vrp, "fixpoint value range propagation", OPT_FLAG_NONE),
327 IRG("frame", opt_frame_irg, "remove unused frame entities", OPT_FLAG_NONE),
328 IRG("gcse", do_gcse, "global common subexpression elimination", OPT_FLAG_NONE),
329 IRG("gvn-pre", do_gvn_pre, "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
330 IRG("if-conversion", do_if_conv, "if-conversion", OPT_FLAG_NONE),
331 IRG("invert-loops", do_loop_inversion, "loop inversion", OPT_FLAG_NONE),
332 IRG("ivopts", do_stred, "induction variable strength reduction", OPT_FLAG_NONE),
333 IRG("local", optimize_graph_df, "local graph optimizations", OPT_FLAG_HIDE_OPTIONS),
334 IRG("lower", do_lower_highlevel, "lowering", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
335 IRG("lower-mux", do_lower_mux, "mux lowering", OPT_FLAG_NONE),
336 IRG("lower-switch", do_lower_switch, "switch lowering", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
337 IRG("one-return", normalize_one_return, "normalisation to 1 return", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
338 IRG("opt-load-store", optimize_load_store, "load store optimization", OPT_FLAG_NONE),
339 IRG("opt-tail-rec", opt_tail_rec_irg, "tail-recursion eliminiation", OPT_FLAG_NONE),
340 IRG("parallelize-mem", opt_parallelize_mem, "parallelize memory", OPT_FLAG_NONE),
341 IRG("place", place_code, "code placement", OPT_FLAG_NONE),
342 IRG("reassociation", optimize_reassociation, "reassociation", OPT_FLAG_NONE),
343 IRG("remove-confirms", remove_confirms, "confirm removal", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
344 IRG("remove-phi-cycles", remove_phi_cycles, "removal of phi cycles", OPT_FLAG_HIDE_OPTIONS),
345 IRG("scalar-replace", scalar_replacement_opt, "scalar replacement", OPT_FLAG_NONE),
346 IRG("shape-blocks", shape_blocks, "block shaping", OPT_FLAG_NONE),
347 IRG("thread-jumps", opt_jumpthreading, "path-sensitive jumpthreading", OPT_FLAG_NONE),
348 IRG("unroll-loops", do_loop_unrolling, "loop unrolling", OPT_FLAG_NONE),
349 IRG("vrp", do_vrp, "value range propagation", OPT_FLAG_NONE),
350 IRP("inline", do_inline, "inlining", OPT_FLAG_NONE),
351 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),
352 IRP("lower-dw", do_lower_dw_ops, "lowering of doubleword operations", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
353 IRP("opt-func-call", do_optimize_funccalls, "function call optimization", OPT_FLAG_NONE),
354 IRP("opt-proc-clone", do_cloning, "procedure cloning", OPT_FLAG_NONE),
355 IRP("remove-unused", garbage_collect_entities, "removal of unused functions/variables", OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
356 IRP("rts", rts_map, "optimization of known library functions", OPT_FLAG_HIDE_OPTIONS),
360 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
361 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
363 static opt_config_t *get_opt(const char *name)
366 for (i = 0; i < n_opts; ++i) {
367 opt_config_t *config = &opts[i];
368 if (strcmp(config->name, name) == 0)
375 static void set_opt_enabled(const char *name, bool enabled)
377 opt_config_t *config = get_opt(name);
378 config->flags = (config->flags & ~OPT_FLAG_ENABLED)
379 | (enabled ? OPT_FLAG_ENABLED : 0);
382 static bool get_opt_enabled(const char *name)
384 opt_config_t *config = get_opt(name);
385 return (config->flags & OPT_FLAG_ENABLED) != 0;
389 * perform an optimisation on a single graph
391 * @return true if something changed, false otherwise
393 static bool do_irg_opt(ir_graph *irg, const char *name)
396 opt_config_t *config = get_opt(name);
397 size_t n = config - opts;
398 assert(config != NULL);
399 assert(config->target == OPT_TARGET_IRG);
400 if (! (config->flags & OPT_FLAG_ENABLED))
403 old_irg = current_ir_graph;
404 current_ir_graph = irg;
406 timer_push(timers[n]);
407 config->u.transform_irg(irg);
408 timer_pop(timers[n]);
410 if (firm_dump.all_phases && firm_dump.ir_graph) {
411 dump_graph_count(irg, name);
414 if (firm_opt.check_all) {
415 timer_push(t_verify);
416 irg_verify(irg, VRFY_ENFORCE_SSA);
420 current_ir_graph = old_irg;
424 static void do_irp_opt(const char *name)
426 opt_config_t *config = get_opt(name);
427 size_t n = config - opts;
428 assert(config->target == OPT_TARGET_IRP);
429 if (! (config->flags & OPT_FLAG_ENABLED))
432 timer_push(timers[n]);
433 config->u.transform_irp();
434 timer_pop(timers[n]);
436 if (firm_dump.ir_graph && firm_dump.all_phases) {
438 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
439 ir_graph *irg = get_irp_irg(i);
440 dump_graph_count(irg, name);
444 if (firm_opt.check_all) {
446 timer_push(t_verify);
447 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
448 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
455 * Enable transformations which should be always safe (and cheap) to perform
457 static void enable_safe_defaults(void)
459 set_opt_enabled("remove-unused", true);
460 set_opt_enabled("opt-tail-rec", true);
461 set_opt_enabled("opt-func-call", true);
462 set_opt_enabled("reassociation", true);
463 set_opt_enabled("control-flow", true);
464 set_opt_enabled("local", true);
465 set_opt_enabled("lower-const", true);
466 set_opt_enabled("scalar-replace", true);
467 set_opt_enabled("place", true);
468 set_opt_enabled("confirm", true);
469 set_opt_enabled("opt-load-store", true);
470 set_opt_enabled("lower", true);
471 set_opt_enabled("deconv", true);
472 set_opt_enabled("remove-confirms", true);
473 set_opt_enabled("ivopts", true);
474 set_opt_enabled("dead", true);
475 set_opt_enabled("lower-switch", true);
476 set_opt_enabled("remove-phi-cycles", true);
477 set_opt_enabled("frame", true);
478 set_opt_enabled("combo", true);
482 * run all the Firm optimizations
484 * @param input_filename the name of the (main) source file
486 static void do_firm_optimizations(const char *input_filename)
491 set_opt_alias_analysis(firm_opt.alias_analysis);
493 aa_opt = aa_opt_no_opt;
494 if (firm_opt.strict_alias)
495 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
496 if (firm_opt.no_alias)
497 aa_opt = aa_opt_no_alias;
499 set_irp_memory_disambiguator_options(aa_opt);
501 /* parameter passing code should set them directly sometime... */
502 set_opt_enabled("rts", !firm_opt.freestanding);
503 set_opt_enabled("gcse", firm_opt.gcse);
504 set_opt_enabled("place", !firm_opt.gcse);
505 set_opt_enabled("confirm", firm_opt.confirm);
506 set_opt_enabled("remove-confirms", firm_opt.confirm);
508 /* osr supersedes remove_phi_cycles */
509 if (get_opt_enabled("ivopts"))
510 set_opt_enabled("remove-phi-cycles", false);
512 timer_start(t_all_opt);
516 /* first step: kill dead code */
517 for (i = 0; i < get_irp_n_irgs(); i++) {
518 ir_graph *irg = get_irp_irg(i);
519 do_irg_opt(irg, "combo");
520 do_irg_opt(irg, "local");
521 do_irg_opt(irg, "control-flow");
524 do_irp_opt("remove-unused");
525 for (i = 0; i < get_irp_n_irgs(); ++i) {
526 ir_graph *irg = get_irp_irg(i);
527 do_irg_opt(irg, "opt-tail-rec");
529 do_irp_opt("opt-func-call");
530 do_irp_opt("lower-const");
532 for (i = 0; i < get_irp_n_irgs(); i++) {
533 ir_graph *irg = get_irp_irg(i);
535 do_irg_opt(irg, "scalar-replace");
536 do_irg_opt(irg, "invert-loops");
537 do_irg_opt(irg, "unroll-loops");
538 do_irg_opt(irg, "local");
539 do_irg_opt(irg, "reassociation");
540 do_irg_opt(irg, "local");
541 do_irg_opt(irg, "gcse");
543 if (firm_opt.confirm) {
544 /* Confirm construction currently can only handle blocks with only
545 one control flow predecessor. Calling optimize_cf here removes
546 Bad predecessors and help the optimization of switch constructs.
548 do_irg_opt(irg, "control-flow");
549 do_irg_opt(irg, "confirm");
550 do_irg_opt(irg, "vrp");
551 do_irg_opt(irg, "local");
554 do_irg_opt(irg, "control-flow");
555 do_irg_opt(irg, "opt-load-store");
556 do_irg_opt(irg, "fp-vrp");
557 do_irg_opt(irg, "lower");
558 do_irg_opt(irg, "deconv");
559 do_irg_opt(irg, "thread-jumps");
560 do_irg_opt(irg, "remove-confirms");
561 do_irg_opt(irg, "gvn-pre");
562 do_irg_opt(irg, "place");
563 do_irg_opt(irg, "control-flow");
565 if (do_irg_opt(irg, "if-conversion")) {
566 do_irg_opt(irg, "local");
567 do_irg_opt(irg, "control-flow");
569 /* this doesn't make too much sense but tests the mux destruction... */
570 do_irg_opt(irg, "lower-mux");
572 do_irg_opt(irg, "bool");
573 do_irg_opt(irg, "shape-blocks");
574 do_irg_opt(irg, "lower-switch");
575 do_irg_opt(irg, "ivopts");
576 do_irg_opt(irg, "local");
577 do_irg_opt(irg, "dead");
578 do_irg_opt(irg, "frame");
581 do_irp_opt("inline");
582 do_irp_opt("opt-proc-clone");
584 for (i = 0; i < get_irp_n_irgs(); i++) {
585 ir_graph *irg = get_irp_irg(i);
586 do_irg_opt(irg, "local");
587 do_irg_opt(irg, "control-flow");
588 do_irg_opt(irg, "thread-jumps");
589 do_irg_opt(irg, "local");
590 do_irg_opt(irg, "control-flow");
592 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
593 do_irg_opt(irg, "local");
594 do_irg_opt(irg, "vrp");
595 do_irg_opt(irg, "local");
596 do_irg_opt(irg, "vrp");
600 if (firm_dump.ir_graph) {
601 /* recompute backedges for nicer dumps */
602 for (i = 0; i < get_irp_n_irgs(); i++)
603 construct_cf_backedges(get_irp_irg(i));
608 if (firm_dump.statistic & STAT_AFTER_OPT)
609 stat_dump_snapshot(input_filename, "opt");
611 timer_stop(t_all_opt);
617 * @param input_filename the name of the (main) source file
619 static void do_firm_lowering(const char *input_filename)
623 do_irp_opt("lower-dw");
625 if (firm_dump.statistic & STAT_AFTER_LOWER)
626 stat_dump_snapshot(input_filename, "low");
630 if (firm_opt.enabled) {
631 timer_start(t_all_opt);
633 /* run reassociation first on all graphs BEFORE the architecture
634 dependent optimizations are enabled */
635 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
636 ir_graph *irg = get_irp_irg(i);
637 do_irg_opt(irg, "reassociation");
640 /* enable architecture dependent optimizations */
641 arch_dep_set_opts((arch_dep_opts_t)
642 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
643 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
644 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
646 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
647 ir_graph *irg = get_irp_irg(i);
649 do_irg_opt(irg, "local");
650 do_irg_opt(irg, "gcse");
651 do_irg_opt(irg, "opt-load-store");
652 do_irg_opt(irg, "local");
653 do_irg_opt(irg, "control-flow");
655 if (do_irg_opt(irg, "vrp")) {
656 do_irg_opt(irg, "local");
657 do_irg_opt(irg, "control-flow");
658 do_irg_opt(irg, "vrp");
659 do_irg_opt(irg, "local");
660 do_irg_opt(irg, "control-flow");
663 if (do_irg_opt(irg, "if-conversion")) {
664 do_irg_opt(irg, "local");
665 do_irg_opt(irg, "control-flow");
668 do_irg_opt(irg, "parallelize-mem");
670 timer_stop(t_all_opt);
672 do_irp_opt("remove-unused");
674 dump_all("-low-opt");
678 mark_private_methods();
680 /* set the phase to low */
681 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
682 set_irg_phase_low(get_irp_irg(i));
684 /* all graphs are lowered, set the irp phase to low */
685 set_irp_phase_state(phase_low);
687 if (firm_dump.statistic & STAT_FINAL) {
688 stat_dump_snapshot(input_filename, "final");
693 * Initialize for the Firm-generating back end.
695 void gen_firm_init(void)
697 firm_parameter_t params;
698 unsigned pattern = 0;
701 for (i = 0; i < n_opts; ++i) {
702 timers[i] = ir_timer_new();
703 timer_register(timers[i], opts[i].description);
705 t_verify = ir_timer_new();
706 timer_register(t_verify, "Firm: verify pass");
707 t_vcg_dump = ir_timer_new();
708 timer_register(t_vcg_dump, "Firm: vcg dumping");
709 t_all_opt = ir_timer_new();
710 timer_register(t_all_opt, "Firm: all optimizations");
712 if (firm_dump.stat_pattern)
713 pattern |= FIRMSTAT_PATTERN_ENABLED;
715 if (firm_dump.stat_dag)
716 pattern |= FIRMSTAT_COUNT_DAG;
718 memset(¶ms, 0, sizeof(params));
719 params.size = sizeof(params);
720 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
721 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS
723 params.initialize_local_func = uninitialized_local_var;
724 params.cc_mask = 0; /* no regparam, cdecl */
728 if (firm_be_opt.selection == BE_FIRM_BE) {
729 const backend_params *be_params = be_get_backend_param();
731 if (be_params->do_dw_lowering)
732 set_opt_enabled("lower-dw", true);
734 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
735 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
737 ad_param = be_params->dep_param;
738 if_conv_info = be_params->if_conv_info;
741 edges_init_dbg(firm_opt.vrfy_edges);
743 /* Sel node cannot produce NULL pointers */
744 set_opt_sel_based_null_check_elim(1);
746 /* dynamic dispatch works currently only if whole world scenarios */
747 set_opt_dyn_meth_dispatch(0);
749 arch_dep_init(arch_factory);
751 /* do not run architecture dependent optimizations in building phase */
752 arch_dep_set_opts(arch_dep_none);
754 do_node_verification((firm_verification_t) firm_opt.vrfy);
755 if (firm_dump.filter)
756 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
758 if (firm_opt.enabled) {
760 set_opt_constant_folding(firm_opt.const_folding);
761 set_opt_algebraic_simplification(firm_opt.const_folding);
762 set_opt_cse(firm_opt.cse);
763 set_opt_global_cse(0);
764 set_opt_unreachable_code(1);
765 set_opt_control_flow(firm_opt.control_flow);
766 set_opt_control_flow_weak_simplification(1);
767 set_opt_control_flow_strong_simplification(1);
772 /* do not dump entity ld names */
777 * Called, after the Firm generation is completed,
778 * do all optimizations and backend call here.
780 * @param out a file handle for the output, may be NULL
781 * @param input_filename the name of the (main) source file
782 * @param c_mode non-zero if "C" was compiled
783 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
785 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode,
786 int new_firm_const_exists)
791 if (firm_opt.enable_statev) {
793 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
794 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
795 ir_stat_ev_compilation_unit(input_filename);
799 firm_const_exists = new_firm_const_exists;
801 /* the general for dumping option must be set, or the others will not work*/
803 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
805 dump_keepalive_edges(1);
806 dump_consts_local(1);
807 dump_dominator_information(1);
808 dump_loop_information(0);
810 if (!firm_dump.edge_labels)
811 turn_off_edge_labels();
813 /* FIXME: cloning might ADD new graphs. */
814 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
816 if (firm_dump.all_types) {
819 dump_class_hierarchy(0, "");
820 dump_class_hierarchy(1, "-with-entities");
824 /* finalize all graphs */
825 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
826 ir_graph *irg = get_irp_irg(i);
827 irg_finalize_cons(irg);
831 timer_push(t_verify);
835 /* all graphs are finalized, set the irp phase to high */
836 set_irp_phase_state(phase_high);
838 /* BEWARE: kill unreachable code before doing compound lowering */
839 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
840 ir_graph *irg = get_irp_irg(i);
841 do_irg_opt(irg, "control-flow");
844 /* lower all compound call return values */
845 lower_compound_params();
847 /* lower copyb nodes */
848 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
849 ir_graph *irg = get_irp_irg(i);
850 lower_CopyB(irg, 128, 4);
853 if (firm_dump.statistic & STAT_BEFORE_OPT) {
854 stat_dump_snapshot(input_filename, "noopt");
857 if (firm_opt.enabled)
858 do_firm_optimizations(input_filename);
861 do_firm_lowering(input_filename);
863 /* set the phase to low */
864 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
865 set_irg_phase_low(get_irp_irg(i));
867 if (firm_dump.statistic & STAT_FINAL_IR)
868 stat_dump_snapshot(input_filename, "final-ir");
870 /* run the code generator */
871 if (firm_be_opt.selection != BE_NONE)
872 do_codegen(out, input_filename);
874 if (firm_dump.statistic & STAT_FINAL)
875 stat_dump_snapshot(input_filename, "final");
878 void disable_all_opts(void)
880 for (int i = 0; i < n_opts; ++i) {
881 opt_config_t *config = &opts[i];
882 if (config->flags & OPT_FLAG_ESSENTIAL) {
883 config->flags |= OPT_FLAG_ENABLED;
885 config->flags &= ~OPT_FLAG_ENABLED;
890 int firm_opt_option(const char *opt)
893 if (strncmp(opt, "no-", 3) == 0) {
898 opt_config_t *config = get_opt(opt);
899 if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
902 config->flags &= ~OPT_FLAG_ENABLED;
903 config->flags |= enable ? OPT_FLAG_ENABLED : 0;
907 void firm_opt_option_help(void)
911 for (i = 0; i < n_opts; ++i) {
915 const opt_config_t *config = &opts[i];
916 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
919 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
920 print_option_help(config->name, buf2);
921 snprintf(buf, sizeof(buf), "no-%s", config->name);
922 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
923 print_option_help(buf, buf2);
928 * Do very early initializations
930 void firm_early_init(void)
932 /* arg: need this here for command line options */
935 enable_safe_defaults();