3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
17 #include <libfirm/firm.h>
18 #include <libfirm/be.h>
22 #include "firm_codegen.h"
23 #include "firm_cmdline.h"
24 #include "firm_timing.h"
27 #if defined(_DEBUG) || defined(FIRM_DEBUG)
28 #define DBG(x) dbg_printf x
30 #define DBG(x) ((void)0)
31 #endif /* _DEBUG || FIRM_DEBUG */
34 /** dump all the graphs depending on cond */
35 #define DUMP_ALL(cond, suffix) \
38 timer_push(TV_VCG_DUMP); \
39 if (firm_dump.no_blocks) \
40 dump_all_ir_graphs(dump_ir_graph, suffix); \
41 else if (firm_dump.extbb) \
42 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
44 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
49 /** dump all control flow graphs depending on cond */
50 #define DUMP_ALL_CFG(cond, suffix) \
53 timer_push(TV_VCG_DUMP); \
54 dump_all_ir_graphs(dump_cfg, suffix); \
59 /** check all graphs depending on cond */
60 #define CHECK_ALL(cond) \
64 timer_push(TV_VERIFY); \
65 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
66 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
73 /** dump graphs irg depending on cond */
74 #define DUMP_ONE(cond, irg, suffix) \
77 timer_push(TV_VCG_DUMP); \
78 if (firm_dump.no_blocks) \
79 dump_ir_graph(irg, suffix); \
80 else if (firm_dump.extbb) \
81 dump_ir_extblock_graph(irg, suffix); \
83 dump_ir_block_graph(irg, suffix); \
88 /** dump control flow graph irg depending on cond */
89 #define DUMP_ONE_CFG(cond, irg, suffix) \
92 timer_push(TV_VCG_DUMP); \
93 dump_cfg(irg, suffix); \
98 /** check a graph irg depending on cond */
99 #define CHECK_ONE(cond, irg) \
102 timer_push(TV_VERIFY); \
103 irg_verify(irg, VRFY_ENFORCE_SSA); \
109 /* set by the backend parameters */
110 static const ir_settings_arch_dep_t *ad_param = NULL;
111 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
112 static void *create_intrinsic_ctx = NULL;
113 static const ir_settings_if_conv_t *if_conv_info = NULL;
115 /* entities of runtime functions */
116 ir_entity_ptr rts_entities[rts_max];
119 * factory for setting architecture dependent parameters
121 static const ir_settings_arch_dep_t *arch_factory(void)
123 static const ir_settings_arch_dep_t param = {
124 1, /* also use subs */
125 4, /* maximum shifts */
126 31, /* maximum shift amount */
127 NULL, /* use default evaluator */
131 32 /* Mulh allowed up to 32 bit */
134 return ad_param ? ad_param : ¶m;
138 * Map runtime functions.
140 static void rts_map(void) {
141 static const struct {
142 ir_entity_ptr *ent; /**< address of the rts entity */
143 i_mapper_func func; /**< mapper function. */
146 { &rts_entities[rts_abs], i_mapper_abs },
147 { &rts_entities[rts_labs], i_mapper_abs },
148 { &rts_entities[rts_llabs], i_mapper_abs },
149 { &rts_entities[rts_imaxabs], i_mapper_abs },
151 /* double -> double */
152 { &rts_entities[rts_fabs], i_mapper_abs },
153 { &rts_entities[rts_sqrt], i_mapper_sqrt },
154 { &rts_entities[rts_cbrt], i_mapper_cbrt },
155 { &rts_entities[rts_pow], i_mapper_pow },
156 { &rts_entities[rts_exp], i_mapper_exp },
157 { &rts_entities[rts_exp2], i_mapper_exp },
158 { &rts_entities[rts_exp10], i_mapper_exp },
159 { &rts_entities[rts_log], i_mapper_log },
160 { &rts_entities[rts_log2], i_mapper_log2 },
161 { &rts_entities[rts_log10], i_mapper_log10 },
162 { &rts_entities[rts_sin], i_mapper_sin },
163 { &rts_entities[rts_cos], i_mapper_cos },
164 { &rts_entities[rts_tan], i_mapper_tan },
165 { &rts_entities[rts_asin], i_mapper_asin },
166 { &rts_entities[rts_acos], i_mapper_acos },
167 { &rts_entities[rts_atan], i_mapper_atan },
168 { &rts_entities[rts_sinh], i_mapper_sinh },
169 { &rts_entities[rts_cosh], i_mapper_cosh },
170 { &rts_entities[rts_tanh], i_mapper_tanh },
173 { &rts_entities[rts_fabsf], i_mapper_abs },
174 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
175 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
176 { &rts_entities[rts_powf], i_mapper_pow },
177 { &rts_entities[rts_expf], i_mapper_exp },
178 { &rts_entities[rts_exp2f], i_mapper_exp },
179 { &rts_entities[rts_exp10f], i_mapper_exp },
180 { &rts_entities[rts_logf], i_mapper_log },
181 { &rts_entities[rts_log2f], i_mapper_log2 },
182 { &rts_entities[rts_log10f], i_mapper_log10 },
183 { &rts_entities[rts_sinf], i_mapper_sin },
184 { &rts_entities[rts_cosf], i_mapper_cos },
185 { &rts_entities[rts_tanf], i_mapper_tan },
186 { &rts_entities[rts_asinf], i_mapper_asin },
187 { &rts_entities[rts_acosf], i_mapper_acos },
188 { &rts_entities[rts_atanf], i_mapper_atan },
189 { &rts_entities[rts_sinhf], i_mapper_sinh },
190 { &rts_entities[rts_coshf], i_mapper_cosh },
191 { &rts_entities[rts_tanhf], i_mapper_tanh },
193 /* long double -> long double */
194 { &rts_entities[rts_fabsl], i_mapper_abs },
195 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
196 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
197 { &rts_entities[rts_powl], i_mapper_pow },
198 { &rts_entities[rts_expl], i_mapper_exp },
199 { &rts_entities[rts_exp2l], i_mapper_exp },
200 { &rts_entities[rts_exp10l], i_mapper_exp },
201 { &rts_entities[rts_logl], i_mapper_log },
202 { &rts_entities[rts_log2l], i_mapper_log2 },
203 { &rts_entities[rts_log10l], i_mapper_log10 },
204 { &rts_entities[rts_sinl], i_mapper_sin },
205 { &rts_entities[rts_cosl], i_mapper_cos },
206 { &rts_entities[rts_tanl], i_mapper_tan },
207 { &rts_entities[rts_asinl], i_mapper_asin },
208 { &rts_entities[rts_acosl], i_mapper_acos },
209 { &rts_entities[rts_atanl], i_mapper_atan },
210 { &rts_entities[rts_sinhl], i_mapper_sinh },
211 { &rts_entities[rts_coshl], i_mapper_cosh },
212 { &rts_entities[rts_tanhl], i_mapper_tanh },
215 { &rts_entities[rts_memcpy], i_mapper_memcpy },
216 { &rts_entities[rts_memset], i_mapper_memset },
217 { &rts_entities[rts_strcmp], i_mapper_strcmp },
218 { &rts_entities[rts_strncmp], i_mapper_strncmp },
219 { &rts_entities[rts_strlen], i_mapper_strlen }
221 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
224 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
225 if (*mapper[i].ent != NULL) {
226 rec[n_map].i_call.kind = INTRINSIC_CALL;
227 rec[n_map].i_call.i_ent = *mapper[i].ent;
228 rec[n_map].i_call.i_mapper = mapper[i].func;
229 rec[n_map].i_call.ctx = NULL;
230 rec[n_map].i_call.link = NULL;
234 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
237 static int *irg_dump_no;
239 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
242 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
243 DUMP_ONE(1, irg, name);
247 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
250 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
251 DUMP_ONE_CFG(1, irg, name);
255 static void dump_all_count(const char *const suffix)
257 const int n_irgs = get_irp_n_irgs();
260 for (i = 0; i < n_irgs; ++i)
261 dump_graph_count(get_irp_irg(i), suffix);
264 #define DUMP_ONE_C(cond, irg, suffix) \
267 dump_graph_count((irg), (suffix)); \
271 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
274 dump_graph_cfg_count((irg), (suffix)); \
278 #define DUMP_ALL_C(cond, suffix) \
281 dump_all_count((suffix)); \
285 static void remove_unused_functions(void)
287 ir_entity **keep_methods;
290 /* Analysis that finds the free methods,
291 i.e. methods that are dereferenced.
292 Optimizes polymorphic calls :-). */
293 cgana(&arr_len, &keep_methods);
295 /* Remove methods that are never called. */
296 gc_irgs(arr_len, keep_methods);
300 static int firm_const_exists;
302 static void do_optimize_funccalls(void)
304 optimize_funccalls(firm_const_exists, NULL);
307 static void do_gcse(ir_graph *irg)
309 set_opt_global_cse(1);
310 optimize_graph_df(irg);
312 set_opt_global_cse(0);
315 static void do_lower_highlevel(ir_graph *irg)
317 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
320 static void do_if_conv(ir_graph *irg)
322 opt_if_conv(irg, if_conv_info);
325 static void do_stred(ir_graph *irg)
327 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
330 static void do_inline(void)
332 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold);
335 static void do_cloning(void)
337 proc_cloning((float) firm_opt.clone_threshold);
340 static void do_lower_switch(ir_graph *irg)
342 lower_switch(irg, firm_opt.spare_size);
345 typedef enum opt_target {
350 typedef void (*transform_irg_func)(ir_graph *irg);
351 typedef void (*transform_irp_func)(void);
352 typedef void (*func_ptr_t)(void);
364 static opt_config_t opts[] = {
365 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, true, true, true, -1 },
366 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, true, true, true, TV_COMBO },
367 { OPT_TARGET_IRG, "controlflow", (func_ptr_t) optimize_cf, true, true, true, TV_CF_OPT },
368 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, true, true, true, TV_LOCAL_OPT },
369 { OPT_TARGET_IRP, "gc_irgs", (func_ptr_t) remove_unused_functions, true, false, false, TV_CGANA },
370 { OPT_TARGET_IRP, "tailrec", (func_ptr_t) opt_tail_recursion, true, true, true, TV_TAIL_REC },
371 { OPT_TARGET_IRP, "funccalls", (func_ptr_t) do_optimize_funccalls, true, true, true, TV_REAL_FUNC_CALL },
372 { OPT_TARGET_IRP, "lowerconst", (func_ptr_t) lower_const_code, true, false, false, -1 },
373 { OPT_TARGET_IRG, "onereturn", (func_ptr_t) normalize_one_return, true, false, false, -1 },
374 { OPT_TARGET_IRG, "scalar", (func_ptr_t) scalar_replacement_opt, true, true, true, TV_SCALAR_REPLACE },
375 { OPT_TARGET_IRG, "reassoc", (func_ptr_t) optimize_reassociation, true, true, true, TV_REASSOCIATION },
376 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, true, true, true, TV_CODE_PLACE },
377 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, true, true, true, TV_CODE_PLACE },
378 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, true, true, true, TV_CONFIRM_CREATE },
379 { OPT_TARGET_IRG, "ldst", (func_ptr_t) optimize_load_store, true, true, true, TV_LOAD_STORE },
380 { OPT_TARGET_IRG, "sync", (func_ptr_t) opt_sync, true, true, true, -1 },
381 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, true, true, true, -1 },
382 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, true, true, true, TV_DECONV },
383 { OPT_TARGET_IRG, "condeval", (func_ptr_t) opt_cond_eval, true, true, true, TV_COND_EVAL },
384 { OPT_TARGET_IRG, "remove_confirms", (func_ptr_t) remove_confirms, true, false, false, TV_CONFIRM_CREATE },
385 { OPT_TARGET_IRG, "gvnpre", (func_ptr_t) do_gvn_pre, true, true, true, -1 },
386 { OPT_TARGET_IRG, "ifconv", (func_ptr_t) do_if_conv, true, true, true, TV_IF_CONV },
387 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, true, true, true, -1 },
388 { OPT_TARGET_IRG, "shape-blocks",(func_ptr_t) shape_blocks, true, true, true, TV_END_MELT },
389 { OPT_TARGET_IRG, "stred", (func_ptr_t) do_stred, true, true, true, TV_OSR },
390 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, true, false, true, TV_DEAD_NODE },
391 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, true, true, true, -1 },
392 { OPT_TARGET_IRP, "clone", (func_ptr_t) do_cloning, true, true, true, -1 },
393 { OPT_TARGET_IRG, "lower_switch", (func_ptr_t) do_lower_switch, true, true, true, -1 },
395 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
397 static opt_config_t *get_opt(const char *name)
400 for (i = 0; i < n_opts; ++i) {
401 opt_config_t *config = &opts[i];
402 if (strcmp(config->name, name) == 0)
410 static void set_opt_enabled(const char *name, bool enabled)
412 opt_config_t *config = get_opt(name);
413 config->enabled = enabled;
416 static void do_irg_opt(ir_graph *irg, const char *name)
418 transform_irg_func func;
420 opt_config_t *config = get_opt(name);
421 assert(config->target == OPT_TARGET_IRG);
422 if (!config->enabled)
425 if (config->timer != -1)
426 timer_push(config->timer);
428 old_irg = current_ir_graph;
429 current_ir_graph = irg;
431 func = (transform_irg_func) config->func;
434 if (config->timer != -1)
437 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
438 CHECK_ONE(firm_opt.check_all, irg);
440 current_ir_graph = old_irg;
443 static void do_irp_opt(const char *name)
445 transform_irp_func func;
446 opt_config_t *config = get_opt(name);
447 assert(config->target == OPT_TARGET_IRP);
448 if (!config->enabled)
451 if (config->timer != -1)
452 timer_push(config->timer);
454 func = (transform_irp_func) config->func;
457 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
458 CHECK_ALL(firm_opt.check_all);
460 if (config->timer != -1)
465 * run all the Firm optimizations
467 * @param input_filename the name of the (main) source file
469 static void do_firm_optimizations(const char *input_filename)
474 /* FIXME: cloning might ADD new graphs. */
475 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
477 set_opt_strength_red(firm_opt.strength_red);
478 set_opt_scalar_replacement(firm_opt.scalar_replace);
479 set_opt_auto_create_sync(firm_opt.auto_sync);
480 set_opt_alias_analysis(firm_opt.alias_analysis);
481 set_opt_combo(firm_opt.combo);
483 aa_opt = aa_opt_no_opt;
484 if (firm_opt.strict_alias)
485 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
486 if (firm_opt.no_alias)
487 aa_opt = aa_opt_no_alias;
489 set_irp_memory_disambiguator_options(aa_opt);
491 /* parameter passing code should set them directly sometime... */
492 set_opt_enabled("rts", firm_opt.freestanding);
493 set_opt_enabled("gc_irgs", firm_opt.remove_unused);
494 set_opt_enabled("tailrec", firm_opt.tail_rec);
495 set_opt_enabled("funccalls", firm_opt.func_calls);
496 set_opt_enabled("gcse", firm_opt.gcse);
497 set_opt_enabled("place", !firm_opt.gcse);
498 set_opt_enabled("confirm", firm_opt.confirm);
499 set_opt_enabled("remove_confirms", firm_opt.confirm);
500 set_opt_enabled("sync", firm_opt.sync);
501 set_opt_enabled("ldst", firm_opt.load_store);
502 set_opt_enabled("deconv", firm_opt.deconv);
503 set_opt_enabled("condeval", firm_opt.cond_eval);
504 set_opt_enabled("gvnpre", firm_opt.gvn_pre);
505 set_opt_enabled("ifconv", firm_opt.if_conversion);
506 set_opt_enabled("bool", firm_opt.bool_opt);
507 set_opt_enabled("shape-blocks", firm_opt.shape_blocks);
508 set_opt_enabled("inline", firm_opt.do_inline);
509 set_opt_enabled("clone", firm_opt.cloning);
510 set_opt_enabled("combo", firm_opt.combo);
512 timer_start(TV_ALL_OPT);
516 /* first step: kill dead code */
517 if (firm_opt.combo) {
518 for (i = 0; i < get_irp_n_irgs(); i++) {
519 ir_graph *irg = get_irp_irg(i);
520 do_irg_opt(irg, "combo");
523 for (i = 0; i < get_irp_n_irgs(); i++) {
524 ir_graph *irg = get_irp_irg(i);
525 do_irg_opt(irg, "local");
526 do_irg_opt(irg, "controlflow");
530 do_irp_opt("gc_irgs");
531 do_irp_opt("tailrec");
532 do_irp_opt("funccalls");
533 do_irp_opt("lowerconst");
535 for (i = 0; i < get_irp_n_irgs(); i++) {
536 ir_graph *irg = get_irp_irg(i);
538 do_irg_opt(irg, "scalar");
539 do_irg_opt(irg, "local");
540 do_irg_opt(irg, "reassoc");
541 do_irg_opt(irg, "local");
542 do_irg_opt(irg, "gcse");
544 if (firm_opt.confirm) {
545 /* Confirm construction currently can only handle blocks with only one control
546 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
547 the optimization of switch constructs. */
548 do_irg_opt(irg, "controlflow");
549 do_irg_opt(irg, "confirm");
550 do_irg_opt(irg, "local");
553 do_irg_opt(irg, "controlflow");
554 do_irg_opt(irg, "ldst");
555 do_irg_opt(irg, "lower");
556 do_irg_opt(irg, "deconv");
557 do_irg_opt(irg, "condeval");
558 do_irg_opt(irg, "remove_confirms");
559 do_irg_opt(irg, "gvnpre");
560 do_irg_opt(irg, "place");
561 do_irg_opt(irg, "controlflow");
563 if (firm_opt.if_conversion) {
564 do_irg_opt(irg, "ifconv");
565 do_irg_opt(irg, "local");
566 do_irg_opt(irg, "controlflow");
569 do_irg_opt(irg, "bool");
570 do_irg_opt(irg, "shape-blocks");
571 do_irg_opt(irg, "lower_switch");
572 do_irg_opt(irg, "stred");
573 do_irg_opt(irg, "local");
574 do_irg_opt(irg, "dead");
577 do_irp_opt("inline");
580 for (i = 0; i < get_irp_n_irgs(); i++) {
581 ir_graph *irg = get_irp_irg(i);
582 do_irg_opt(irg, "local");
583 do_irg_opt(irg, "controlflow");
584 do_irg_opt(irg, "condeval");
585 do_irg_opt(irg, "local");
586 do_irg_opt(irg, "controlflow");
589 if (firm_dump.ir_graph) {
590 /* recompute backedges for nicer dumps */
591 for (i = 0; i < get_irp_n_irgs(); i++)
592 construct_cf_backedges(get_irp_irg(i));
595 do_irp_opt("gc_irgs");
597 DUMP_ALL(firm_dump.ir_graph, "-opt");
598 /* verify optimized graphs */
599 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
600 ir_graph *irg = get_irp_irg(i);
601 irg_verify(irg, VRFY_ENFORCE_SSA);
604 if (firm_dump.statistic & STAT_AFTER_OPT)
605 stat_dump_snapshot(input_filename, "opt");
607 timer_stop(TV_ALL_OPT);
608 } /* do_firm_optimizations */
611 * compute the size of a type (do implicit lowering)
613 * @param ty a Firm type
615 static int compute_type_size(ir_type *ty)
617 optimization_state_t state;
618 unsigned align_all = 1;
619 int n, size = 0, set = 0;
622 if (get_type_state(ty) == layout_fixed) {
623 /* do not layout already layouted types again */
627 if (is_Method_type(ty) || ty == get_glob_type()) {
628 /* no need for size calculation for method types or the global type */
632 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
634 switch (get_type_tpop_code(ty)) {
637 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
638 ir_entity *ent = get_compound_member(ty, i);
639 ir_type *ent_ty = get_entity_type(ent);
640 unsigned align, misalign;
642 /* compute member types */
643 if (! compute_type_size(ent_ty))
646 align = get_type_alignment_bytes(ent_ty);
647 align_all = align > align_all ? align : align_all;
648 misalign = (align ? size % align : 0);
649 size += (misalign ? align - misalign : 0);
651 set_entity_offset(ent, size);
652 size += get_type_size_bytes(ent_ty);
654 DBG((" member %s %s -> (size: %u, align: %u)\n",
655 get_type_name(ent_ty), get_entity_name(ent),
656 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
658 if (align_all > 0 && size % align_all) {
659 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
660 size += align_all - (size % align_all);
661 DBG(("correcting type-size to %d\n", size));
663 set_type_alignment_bytes(ty, align_all);
668 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
669 ir_entity *ent = get_union_member(ty, i);
671 if (! compute_type_size(get_entity_type(ent)))
673 s = get_type_size_bytes(get_entity_type(ent));
675 set_entity_offset(ent, 0);
676 size = (s > size ? s : size);
682 dims = get_array_n_dimensions(ty);
684 if (! compute_type_size(get_array_element_type(ty)))
689 save_optimization_state(&state);
691 set_opt_constant_folding(1);
692 set_opt_algebraic_simplification(1);
694 for (i = 0; i < dims; ++i) {
695 ir_node *lower = get_array_lower_bound(ty, i);
696 ir_node *upper = get_array_upper_bound(ty, i);
697 ir_graph *rem = current_ir_graph;
698 tarval *tv_lower, *tv_upper;
699 long val_lower, val_upper;
701 current_ir_graph = get_const_code_irg();
702 local_optimize_node(lower);
703 local_optimize_node(upper);
704 current_ir_graph = rem;
706 tv_lower = computed_value(lower);
707 tv_upper = computed_value(upper);
709 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
711 * we cannot calculate the size of this array yet, it
712 * even might be unknown until the end, like argv[]
714 restore_optimization_state(&state);
718 val_upper = get_tarval_long(tv_upper);
719 val_lower = get_tarval_long(tv_lower);
720 size *= val_upper - val_lower;
722 restore_optimization_state(&state);
724 DBG(("array %s -> (elements: %d, element type size: %d)\n",
726 size, get_type_size_bytes(get_array_element_type(ty))));
727 size *= get_type_size_bytes(get_array_element_type(ty));
736 set_type_size_bytes(ty, size);
737 set_type_state(ty, layout_fixed);
740 DBG(("size: %d\n", get_type_size_bytes(ty)));
743 } /* compute_type_size */
746 * layout all non-frame types of the Firm graph
748 static void compute_type_sizes(void)
753 /* all non-frame other types */
754 for (i = get_irp_n_types() - 1; i >= 0; --i) {
755 tp = get_irp_type(i);
756 compute_type_size(tp);
758 if (is_Method_type(tp)) {
759 tp = get_method_value_res_type(tp);
762 /* we have a value result type for this method, lower */
763 compute_type_size(tp);
767 } /* compute_type_sizes */
770 * layout all frame-types of the Firm graph
772 static void compute_frame_type_sizes(void)
777 /* all frame types */
778 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
779 irg = get_irp_irg(i);
780 /* do not optimize away variables in debug mode */
781 if (firm_opt.debug_mode == DBG_MODE_NONE)
783 compute_type_size(get_irg_frame_type(irg));
785 } /* compute_frame_type_sizes */
790 * @param input_filename the name of the (main) source file
792 static void do_firm_lowering(const char *input_filename)
796 if (firm_opt.lower_ll) {
797 lwrdw_param_t init = {
800 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
801 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
802 get_atomic_mode(ATOMIC_TYPE_INT),
803 get_atomic_mode(ATOMIC_TYPE_UINT),
804 def_create_intrinsic_fkt,
809 if (arch_create_intrinsic) {
810 init.create_intrinsic = arch_create_intrinsic;
811 init.ctx = create_intrinsic_ctx;
813 timer_push(TV_DW_LOWER);
818 if (firm_dump.statistic & STAT_AFTER_LOWER)
819 stat_dump_snapshot(input_filename, "low");
821 /* verify lowered graphs */
822 timer_push(TV_VERIFY);
823 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
824 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
827 DUMP_ALL(firm_dump.ir_graph, "-low");
829 if (firm_opt.enabled) {
830 timer_start(TV_ALL_OPT);
832 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
834 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
835 ir_graph *irg = get_irp_irg(i);
836 do_irg_opt(irg, "reassoc");
839 /* enable architecture dependent optimizations */
840 arch_dep_set_opts((arch_dep_opts_t)
841 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
842 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
843 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
845 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
846 ir_graph *irg = get_irp_irg(i);
848 current_ir_graph = irg;
850 do_irg_opt(irg, "local");
851 do_irg_opt(irg, "gcse");
852 do_irg_opt(irg, "ldst");
853 do_irg_opt(irg, "local");
854 do_irg_opt(irg, "controlflow");
856 if (firm_opt.if_conversion) {
857 do_irg_opt(irg, "ifconv");
858 do_irg_opt(irg, "local");
859 do_irg_opt(irg, "controlflow");
862 do_irg_opt(current_ir_graph, "sync");
864 timer_stop(TV_ALL_OPT);
866 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
870 mark_private_methods();
872 /* set the phase to low */
873 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
874 set_irg_phase_low(get_irp_irg(i));
876 /* all graphs are lowered, set the irp phase to low */
877 set_irp_phase_state(phase_low);
879 if (firm_dump.statistic & STAT_FINAL) {
880 stat_dump_snapshot(input_filename, "final");
882 } /* do_firm_lowering */
885 * Initialize for the Firm-generating back end.
887 void gen_firm_init(void)
889 firm_parameter_t params;
890 unsigned pattern = 0;
892 /* the automatic state is only set if inlining is enabled */
893 firm_opt.auto_inline = firm_opt.do_inline;
895 if (firm_dump.stat_pattern)
896 pattern |= FIRMSTAT_PATTERN_ENABLED;
898 if (firm_dump.stat_dag)
899 pattern |= FIRMSTAT_COUNT_DAG;
901 memset(¶ms, 0, sizeof(params));
902 params.size = sizeof(params);
903 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
904 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
905 params.initialize_local_func = uninitialized_local_var;
906 params.cc_mask = 0; /* no regparam, cdecl */
907 params.builtin_dbg = NULL;
911 if (firm_be_opt.selection == BE_FIRM_BE) {
912 const backend_params *be_params = be_get_backend_param();
914 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
916 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
917 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
919 ad_param = be_params->dep_param;
920 if_conv_info = be_params->if_conv_info;
923 dbg_init(NULL, NULL, dbg_snprint);
924 edges_init_dbg(firm_opt.vrfy_edges);
926 set_opt_precise_exc_context(firm_opt.precise_exc);
927 set_opt_fragile_ops(firm_opt.fragile_ops);
929 /* Sel node cannot produce NULL pointers */
930 set_opt_sel_based_null_check_elim(1);
932 /* dynamic dispatch works currently only if whole world scenarios */
933 set_opt_dyn_meth_dispatch(0);
935 arch_dep_init(arch_factory);
937 /* do not run architecture dependent optimizations in building phase */
938 arch_dep_set_opts(arch_dep_none);
940 do_node_verification((firm_verification_t) firm_opt.vrfy);
941 if (firm_dump.filter)
942 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
944 if (firm_opt.enabled) {
946 set_opt_constant_folding(firm_opt.const_folding);
947 set_opt_algebraic_simplification(firm_opt.const_folding);
948 set_opt_cse(firm_opt.cse);
949 set_opt_global_cse(0);
950 set_opt_unreachable_code(1);
951 set_opt_control_flow(firm_opt.control_flow);
952 set_opt_control_flow_weak_simplification(1);
953 set_opt_control_flow_strong_simplification(1);
958 /* do not dump entity ld names */
960 } /* gen_firm_init */
963 * Called, after the Firm generation is completed,
964 * do all optimizations and backend call here.
966 * @param out a file handle for the output, may be NULL
967 * @param input_filename the name of the (main) source file
968 * @param c_mode non-zero if "C" was compiled
969 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
971 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
975 firm_const_exists = new_firm_const_exists;
977 /* the general for dumping option must be set, or the others will not work */
979 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
981 dump_keepalive_edges(1);
982 dump_consts_local(1);
983 dump_dominator_information(1);
984 dump_loop_information(0);
986 if (!firm_dump.edge_labels)
987 turn_off_edge_labels();
989 if (firm_dump.all_types) {
992 dump_class_hierarchy(0, "");
993 dump_class_hierarchy(1, "-with-entities");
997 /* finalize all graphs */
998 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
999 ir_graph *irg = get_irp_irg(i);
1001 irg_finalize_cons(irg);
1002 DUMP_ONE(firm_dump.ir_graph, irg, "");
1004 /* verify the graph */
1005 timer_push(TV_VERIFY);
1006 irg_verify(irg, VRFY_ENFORCE_SSA);
1010 timer_push(TV_VERIFY);
1014 /* all graphs are finalized, set the irp phase to high */
1015 set_irp_phase_state(phase_high);
1017 /* BEWARE: kill unreachable code before doing compound lowering */
1018 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1019 ir_graph *irg = get_irp_irg(i);
1023 /* lower all compound call return values */
1024 lower_compound_params();
1026 /* computes the sizes of all types that are still not computed */
1027 compute_type_sizes();
1029 /* lower copyb nodes */
1030 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1031 ir_graph *irg = get_irp_irg(i);
1032 lower_CopyB(irg, 128, 4);
1035 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1036 stat_dump_snapshot(input_filename, "noopt");
1039 if (firm_opt.enabled)
1040 do_firm_optimizations(input_filename);
1042 if (firm_dump.gen_firm_asm) {
1043 timer_push(TV_FIRM_ASM);
1044 gen_Firm_assembler(input_filename);
1050 do_firm_lowering(input_filename);
1052 /* computes the sizes of all frame types */
1053 compute_frame_type_sizes();
1055 /* set the phase to low */
1056 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1057 set_irg_phase_low(get_irp_irg(i));
1059 if (firm_dump.statistic & STAT_FINAL_IR)
1060 stat_dump_snapshot(input_filename, "final-ir");
1062 /* run the code generator */
1063 if (firm_be_opt.selection != BE_NONE)
1064 do_codegen(out, input_filename);
1066 if (firm_dump.statistic & STAT_FINAL)
1067 stat_dump_snapshot(input_filename, "final");
1071 * Do very early initializations
1073 void firm_early_init(void)
1075 /* arg: need this here for command line options */