3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
14 #include <libfirm/firm.h>
15 #include <libfirm/be.h>
19 #include "firm_codegen.h"
20 #include "firm_cmdline.h"
21 #include "firm_timing.h"
25 #define snprintf _snprintf
28 #if defined(_DEBUG) || defined(FIRM_DEBUG)
29 #define DBG(x) dbg_printf x
32 #endif /* _DEBUG || FIRM_DEBUG */
35 /** dump all the graphs depending on cond */
36 #define DUMP_ALL(cond, suffix) \
39 timer_push(TV_VCG_DUMP); \
40 if (firm_dump.no_blocks) \
41 dump_all_ir_graphs(dump_ir_graph, suffix); \
42 else if (firm_dump.extbb) \
43 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
45 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
50 /** dump all control flow graphs depending on cond */
51 #define DUMP_ALL_CFG(cond, suffix) \
54 timer_push(TV_VCG_DUMP); \
55 dump_all_ir_graphs(dump_cfg, suffix); \
60 /** check all graphs depending on cond */
61 #define CHECK_ALL(cond) \
65 timer_push(TV_VERIFY); \
66 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
67 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
74 /** dump graphs irg depending on cond */
75 #define DUMP_ONE(cond, irg, suffix) \
78 timer_push(TV_VCG_DUMP); \
79 if (firm_dump.no_blocks) \
80 dump_ir_graph(irg, suffix); \
81 else if (firm_dump.extbb) \
82 dump_ir_extblock_graph(irg, suffix); \
84 dump_ir_block_graph(irg, suffix); \
89 /** dump control flow graph irg depending on cond */
90 #define DUMP_ONE_CFG(cond, irg, suffix) \
93 timer_push(TV_VCG_DUMP); \
94 dump_cfg(irg, suffix); \
99 /** check a graph irg depending on cond */
100 #define CHECK_ONE(cond, irg) \
103 timer_push(TV_VERIFY); \
104 irg_verify(irg, VRFY_ENFORCE_SSA); \
110 /* set by the backend parameters */
111 static const ir_settings_arch_dep_t *ad_param = NULL;
112 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
113 static void *create_intrinsic_ctx = NULL;
114 static const ir_settings_if_conv_t *if_conv_info = NULL;
116 ir_mode *firm_imm_fp_mode = NULL;
118 /* entities of runtime functions */
119 ir_entity_ptr rts_entities[rts_max];
122 * factory for setting architecture dependent parameters
124 static const ir_settings_arch_dep_t *arch_factory(void)
126 static const ir_settings_arch_dep_t param = {
127 1, /* also use subs */
128 4, /* maximum shifts */
129 31, /* maximum shift amount */
130 NULL, /* use default evaluator */
134 32 /* Mulh allowed up to 32 bit */
137 return ad_param ? ad_param : ¶m;
141 * Map runtime functions.
143 static void rts_map(void) {
144 static const struct {
145 ir_entity_ptr *ent; /**< address of the rts entity */
146 i_mapper_func func; /**< mapper function. */
149 { &rts_entities[rts_abs], i_mapper_abs },
150 { &rts_entities[rts_labs], i_mapper_abs },
151 { &rts_entities[rts_llabs], i_mapper_abs },
152 { &rts_entities[rts_imaxabs], i_mapper_abs },
154 /* double -> double */
155 { &rts_entities[rts_fabs], i_mapper_abs },
156 { &rts_entities[rts_sqrt], i_mapper_sqrt },
157 { &rts_entities[rts_cbrt], i_mapper_cbrt },
158 { &rts_entities[rts_pow], i_mapper_pow },
159 { &rts_entities[rts_exp], i_mapper_exp },
160 { &rts_entities[rts_exp2], i_mapper_exp },
161 { &rts_entities[rts_exp10], i_mapper_exp },
162 { &rts_entities[rts_log], i_mapper_log },
163 { &rts_entities[rts_log2], i_mapper_log2 },
164 { &rts_entities[rts_log10], i_mapper_log10 },
165 { &rts_entities[rts_sin], i_mapper_sin },
166 { &rts_entities[rts_cos], i_mapper_cos },
167 { &rts_entities[rts_tan], i_mapper_tan },
168 { &rts_entities[rts_asin], i_mapper_asin },
169 { &rts_entities[rts_acos], i_mapper_acos },
170 { &rts_entities[rts_atan], i_mapper_atan },
171 { &rts_entities[rts_sinh], i_mapper_sinh },
172 { &rts_entities[rts_cosh], i_mapper_cosh },
173 { &rts_entities[rts_tanh], i_mapper_tanh },
176 { &rts_entities[rts_fabsf], i_mapper_abs },
177 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
178 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
179 { &rts_entities[rts_powf], i_mapper_pow },
180 { &rts_entities[rts_expf], i_mapper_exp },
181 { &rts_entities[rts_exp2f], i_mapper_exp },
182 { &rts_entities[rts_exp10f], i_mapper_exp },
183 { &rts_entities[rts_logf], i_mapper_log },
184 { &rts_entities[rts_log2f], i_mapper_log2 },
185 { &rts_entities[rts_log10f], i_mapper_log10 },
186 { &rts_entities[rts_sinf], i_mapper_sin },
187 { &rts_entities[rts_cosf], i_mapper_cos },
188 { &rts_entities[rts_tanf], i_mapper_tan },
189 { &rts_entities[rts_asinf], i_mapper_asin },
190 { &rts_entities[rts_acosf], i_mapper_acos },
191 { &rts_entities[rts_atanf], i_mapper_atan },
192 { &rts_entities[rts_sinhf], i_mapper_sinh },
193 { &rts_entities[rts_coshf], i_mapper_cosh },
194 { &rts_entities[rts_tanhf], i_mapper_tanh },
196 /* long double -> long double */
197 { &rts_entities[rts_fabsl], i_mapper_abs },
198 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
199 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
200 { &rts_entities[rts_powl], i_mapper_pow },
201 { &rts_entities[rts_expl], i_mapper_exp },
202 { &rts_entities[rts_exp2l], i_mapper_exp },
203 { &rts_entities[rts_exp10l], i_mapper_exp },
204 { &rts_entities[rts_logl], i_mapper_log },
205 { &rts_entities[rts_log2l], i_mapper_log2 },
206 { &rts_entities[rts_log10l], i_mapper_log10 },
207 { &rts_entities[rts_sinl], i_mapper_sin },
208 { &rts_entities[rts_cosl], i_mapper_cos },
209 { &rts_entities[rts_tanl], i_mapper_tan },
210 { &rts_entities[rts_asinl], i_mapper_asin },
211 { &rts_entities[rts_acosl], i_mapper_acos },
212 { &rts_entities[rts_atanl], i_mapper_atan },
213 { &rts_entities[rts_sinhl], i_mapper_sinh },
214 { &rts_entities[rts_coshl], i_mapper_cosh },
215 { &rts_entities[rts_tanhl], i_mapper_tanh },
218 { &rts_entities[rts_memcpy], i_mapper_memcpy },
219 { &rts_entities[rts_memset], i_mapper_memset },
220 { &rts_entities[rts_strcmp], i_mapper_strcmp },
221 { &rts_entities[rts_strncmp], i_mapper_strncmp },
222 { &rts_entities[rts_strlen], i_mapper_strlen }
224 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
227 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
228 if (*mapper[i].ent != NULL) {
229 rec[n_map].i_call.kind = INTRINSIC_CALL;
230 rec[n_map].i_call.i_ent = *mapper[i].ent;
231 rec[n_map].i_call.i_mapper = mapper[i].func;
232 rec[n_map].i_call.ctx = NULL;
233 rec[n_map].i_call.link = NULL;
237 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
240 static int *irg_dump_no;
242 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
245 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
246 DUMP_ONE(1, irg, name);
250 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
253 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
254 DUMP_ONE_CFG(1, irg, name);
258 static void dump_all_count(const char *const suffix)
260 const int n_irgs = get_irp_n_irgs();
263 for (i = 0; i < n_irgs; ++i)
264 dump_graph_count(get_irp_irg(i), suffix);
267 #define DUMP_ONE_C(cond, irg, suffix) \
270 dump_graph_count((irg), (suffix)); \
274 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
277 dump_graph_cfg_count((irg), (suffix)); \
281 #define DUMP_ALL_C(cond, suffix) \
284 dump_all_count((suffix)); \
288 static void remove_unused_functions(void)
290 ir_entity **keep_methods;
293 /* Analysis that finds the free methods,
294 i.e. methods that are dereferenced.
295 Optimizes polymorphic calls :-). */
296 cgana(&arr_len, &keep_methods);
298 /* Remove methods that are never called. */
299 gc_irgs(arr_len, keep_methods);
303 static int firm_const_exists;
305 static void do_optimize_funccalls(void)
307 optimize_funccalls(firm_const_exists, NULL);
310 static void do_gcse(ir_graph *irg)
312 set_opt_global_cse(1);
313 optimize_graph_df(irg);
315 set_opt_global_cse(0);
318 static void do_lower_highlevel(ir_graph *irg)
320 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
323 static void do_if_conv(ir_graph *irg)
325 opt_if_conv(irg, if_conv_info);
328 static void do_stred(ir_graph *irg)
330 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
333 static void do_inline(void)
335 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold);
338 static void do_cloning(void)
340 proc_cloning((float) firm_opt.clone_threshold);
343 static void do_lower_switch(ir_graph *irg)
345 lower_switch(irg, firm_opt.spare_size);
348 typedef enum opt_target {
353 typedef void (*transform_irg_func)(ir_graph *irg);
354 typedef void (*transform_irp_func)(void);
355 typedef void (*func_ptr_t)(void);
367 static opt_config_t opts[] = {
368 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, true, true, true, -1 },
369 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, true, true, true, TV_COMBO },
370 { OPT_TARGET_IRG, "controlflow", (func_ptr_t) optimize_cf, true, true, true, TV_CF_OPT },
371 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, true, true, true, TV_LOCAL_OPT },
372 { OPT_TARGET_IRP, "gc_irgs", (func_ptr_t) remove_unused_functions, true, false, false, TV_CGANA },
373 { OPT_TARGET_IRP, "tailrec", (func_ptr_t) opt_tail_recursion, true, true, true, TV_TAIL_REC },
374 { OPT_TARGET_IRP, "funccalls", (func_ptr_t) do_optimize_funccalls, true, true, true, TV_REAL_FUNC_CALL },
375 { OPT_TARGET_IRP, "lowerconst", (func_ptr_t) lower_const_code, true, false, false, -1 },
376 { OPT_TARGET_IRG, "onereturn", (func_ptr_t) normalize_one_return, true, false, false, -1 },
377 { OPT_TARGET_IRG, "scalar", (func_ptr_t) scalar_replacement_opt, true, true, true, TV_SCALAR_REPLACE },
378 { OPT_TARGET_IRG, "reassoc", (func_ptr_t) optimize_reassociation, true, true, true, TV_REASSOCIATION },
379 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, true, true, true, TV_CODE_PLACE },
380 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, true, true, true, TV_CODE_PLACE },
381 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, true, true, true, TV_CONFIRM_CREATE },
382 { OPT_TARGET_IRG, "ldst", (func_ptr_t) optimize_load_store, true, true, true, TV_LOAD_STORE },
383 { OPT_TARGET_IRG, "sync", (func_ptr_t) opt_sync, true, true, true, -1 },
384 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, true, true, true, -1 },
385 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, true, true, true, TV_DECONV },
386 { OPT_TARGET_IRG, "condeval", (func_ptr_t) opt_cond_eval, true, true, true, TV_COND_EVAL },
387 { OPT_TARGET_IRG, "remove_confirms", (func_ptr_t) remove_confirms, true, false, false, TV_CONFIRM_CREATE },
388 { OPT_TARGET_IRG, "gvnpre", (func_ptr_t) do_gvn_pre, true, true, true, -1 },
389 { OPT_TARGET_IRG, "ifconv", (func_ptr_t) do_if_conv, true, true, true, TV_IF_CONV },
390 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, true, true, true, -1 },
391 { OPT_TARGET_IRG, "shape-blocks",(func_ptr_t) shape_blocks, true, true, true, TV_END_MELT },
392 { OPT_TARGET_IRG, "stred", (func_ptr_t) do_stred, true, true, true, TV_OSR },
393 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, true, false, true, TV_DEAD_NODE },
394 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, true, true, true, -1 },
395 { OPT_TARGET_IRP, "clone", (func_ptr_t) do_cloning, true, true, true, -1 },
396 { OPT_TARGET_IRG, "lower_switch", (func_ptr_t) do_lower_switch, true, true, true, -1 },
398 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
400 static opt_config_t *get_opt(const char *name)
403 for (i = 0; i < n_opts; ++i) {
404 opt_config_t *config = &opts[i];
405 if (strcmp(config->name, name) == 0)
413 static void set_opt_enabled(const char *name, bool enabled)
415 opt_config_t *config = get_opt(name);
416 config->enabled = enabled;
419 static void do_irg_opt(ir_graph *irg, const char *name)
421 transform_irg_func func;
423 opt_config_t *config = get_opt(name);
424 assert(config->target == OPT_TARGET_IRG);
425 if (!config->enabled)
428 if (config->timer != -1)
429 timer_push(config->timer);
431 old_irg = current_ir_graph;
432 current_ir_graph = irg;
434 func = (transform_irg_func) config->func;
437 if (config->timer != -1)
440 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
441 CHECK_ONE(firm_opt.check_all, irg);
443 current_ir_graph = old_irg;
446 static void do_irp_opt(const char *name)
448 transform_irp_func func;
449 opt_config_t *config = get_opt(name);
450 assert(config->target == OPT_TARGET_IRP);
451 if (!config->enabled)
454 if (config->timer != -1)
455 timer_push(config->timer);
457 func = (transform_irp_func) config->func;
460 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
461 CHECK_ALL(firm_opt.check_all);
463 if (config->timer != -1)
468 * run all the Firm optimizations
470 * @param input_filename the name of the (main) source file
472 static void do_firm_optimizations(const char *input_filename)
477 /* FIXME: cloning might ADD new graphs. */
478 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
480 set_opt_strength_red(firm_opt.strength_red);
481 set_opt_scalar_replacement(firm_opt.scalar_replace);
482 set_opt_auto_create_sync(firm_opt.auto_sync);
483 set_opt_alias_analysis(firm_opt.alias_analysis);
484 set_opt_combo(firm_opt.combo);
486 aa_opt = aa_opt_no_opt;
487 if (firm_opt.strict_alias)
488 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
489 if (firm_opt.no_alias)
490 aa_opt = aa_opt_no_alias;
492 set_irp_memory_disambiguator_options(aa_opt);
494 /* parameter passing code should set them directly sometime... */
495 set_opt_enabled("rts", firm_opt.freestanding);
496 set_opt_enabled("gc_irgs", firm_opt.remove_unused);
497 set_opt_enabled("tailrec", firm_opt.tail_rec);
498 set_opt_enabled("funccalls", firm_opt.func_calls);
499 set_opt_enabled("gcse", firm_opt.gcse);
500 set_opt_enabled("place", !firm_opt.gcse);
501 set_opt_enabled("confirm", firm_opt.confirm);
502 set_opt_enabled("remove_confirms", firm_opt.confirm);
503 set_opt_enabled("sync", firm_opt.sync);
504 set_opt_enabled("ldst", firm_opt.load_store);
505 set_opt_enabled("deconv", firm_opt.deconv);
506 set_opt_enabled("condeval", firm_opt.cond_eval);
507 set_opt_enabled("gvnpre", firm_opt.gvn_pre);
508 set_opt_enabled("ifconv", firm_opt.if_conversion);
509 set_opt_enabled("bool", firm_opt.bool_opt);
510 set_opt_enabled("shape-blocks", firm_opt.shape_blocks);
511 set_opt_enabled("inline", firm_opt.do_inline);
512 set_opt_enabled("clone", firm_opt.cloning);
513 set_opt_enabled("combo", firm_opt.combo);
515 timer_start(TV_ALL_OPT);
519 /* first step: kill dead code */
520 if (firm_opt.combo) {
521 for (i = 0; i < get_irp_n_irgs(); i++) {
522 ir_graph *irg = get_irp_irg(i);
523 do_irg_opt(irg, "combo");
526 for (i = 0; i < get_irp_n_irgs(); i++) {
527 ir_graph *irg = get_irp_irg(i);
528 do_irg_opt(irg, "local");
529 do_irg_opt(irg, "controlflow");
533 do_irp_opt("gc_irgs");
534 do_irp_opt("tailrec");
535 do_irp_opt("funccalls");
536 do_irp_opt("lowerconst");
538 for (i = 0; i < get_irp_n_irgs(); i++) {
539 ir_graph *irg = get_irp_irg(i);
542 /* If SIMD optimization is on, make sure we have only 1 return */
543 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
544 do_irg_opt("onereturn");
547 do_irg_opt(irg, "scalar");
548 do_irg_opt(irg, "local");
549 do_irg_opt(irg, "reassoc");
550 do_irg_opt(irg, "local");
551 do_irg_opt(irg, "gcse");
553 if (firm_opt.confirm) {
554 /* Confirm construction currently can only handle blocks with only one control
555 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
556 the optimization of switch constructs. */
557 do_irg_opt(irg, "controlflow");
558 do_irg_opt(irg, "confirm");
559 do_irg_opt(irg, "local");
562 do_irg_opt(irg, "controlflow");
563 do_irg_opt(irg, "ldst");
564 do_irg_opt(irg, "lower");
565 do_irg_opt(irg, "deconv");
566 do_irg_opt(irg, "condeval");
567 do_irg_opt(irg, "remove_confirms");
568 do_irg_opt(irg, "gvnpre");
569 do_irg_opt(irg, "place");
570 do_irg_opt(irg, "controlflow");
572 if (firm_opt.if_conversion) {
573 do_irg_opt(irg, "ifconv");
574 do_irg_opt(irg, "local");
575 do_irg_opt(irg, "controlflow");
578 do_irg_opt(irg, "bool");
579 do_irg_opt(irg, "shape-blocks");
580 do_irg_opt(irg, "lower_switch");
581 do_irg_opt(irg, "stred");
582 do_irg_opt(irg, "local");
583 do_irg_opt(irg, "dead");
586 do_irp_opt("inline");
589 for (i = 0; i < get_irp_n_irgs(); i++) {
590 ir_graph *irg = get_irp_irg(i);
591 do_irg_opt(irg, "local");
592 do_irg_opt(irg, "controlflow");
593 do_irg_opt(irg, "condeval");
594 do_irg_opt(irg, "local");
595 do_irg_opt(irg, "controlflow");
598 if (firm_dump.ir_graph) {
599 /* recompute backedges for nicer dumps */
600 for (i = 0; i < get_irp_n_irgs(); i++)
601 construct_cf_backedges(get_irp_irg(i));
604 do_irp_opt("gc_irgs");
606 DUMP_ALL(firm_dump.ir_graph, "-opt");
607 /* verify optimized graphs */
608 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
609 ir_graph *irg = get_irp_irg(i);
610 irg_verify(irg, VRFY_ENFORCE_SSA);
613 if (firm_dump.statistic & STAT_AFTER_OPT)
614 stat_dump_snapshot(input_filename, "opt");
616 timer_stop(TV_ALL_OPT);
617 } /* do_firm_optimizations */
620 * compute the size of a type (do implicit lowering)
622 * @param ty a Firm type
624 static int compute_type_size(ir_type *ty)
626 optimization_state_t state;
627 unsigned align_all = 1;
628 int n, size = 0, set = 0;
631 if (get_type_state(ty) == layout_fixed) {
632 /* do not layout already layouted types again */
636 if (is_Method_type(ty) || ty == get_glob_type()) {
637 /* no need for size calculation for method types or the global type */
641 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
643 switch (get_type_tpop_code(ty)) {
646 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
647 ir_entity *ent = get_compound_member(ty, i);
648 ir_type *ent_ty = get_entity_type(ent);
649 unsigned align, misalign;
651 /* compute member types */
652 if (! compute_type_size(ent_ty))
655 align = get_type_alignment_bytes(ent_ty);
656 align_all = align > align_all ? align : align_all;
657 misalign = (align ? size % align : 0);
658 size += (misalign ? align - misalign : 0);
660 set_entity_offset(ent, size);
661 size += get_type_size_bytes(ent_ty);
663 DBG((" member %s %s -> (size: %u, align: %u)\n",
664 get_type_name(ent_ty), get_entity_name(ent),
665 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
667 if (align_all > 0 && size % align_all) {
668 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
669 size += align_all - (size % align_all);
670 DBG(("correcting type-size to %d\n", size));
672 set_type_alignment_bytes(ty, align_all);
677 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
678 ir_entity *ent = get_union_member(ty, i);
680 if (! compute_type_size(get_entity_type(ent)))
682 s = get_type_size_bytes(get_entity_type(ent));
684 set_entity_offset(ent, 0);
685 size = (s > size ? s : size);
691 dims = get_array_n_dimensions(ty);
693 if (! compute_type_size(get_array_element_type(ty)))
698 save_optimization_state(&state);
700 set_opt_constant_folding(1);
701 set_opt_algebraic_simplification(1);
703 for (i = 0; i < dims; ++i) {
704 ir_node *lower = get_array_lower_bound(ty, i);
705 ir_node *upper = get_array_upper_bound(ty, i);
706 ir_graph *rem = current_ir_graph;
707 tarval *tv_lower, *tv_upper;
708 long val_lower, val_upper;
710 current_ir_graph = get_const_code_irg();
711 local_optimize_node(lower);
712 local_optimize_node(upper);
713 current_ir_graph = rem;
715 tv_lower = computed_value(lower);
716 tv_upper = computed_value(upper);
718 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
720 * we cannot calculate the size of this array yet, it
721 * even might be unknown until the end, like argv[]
723 restore_optimization_state(&state);
727 val_upper = get_tarval_long(tv_upper);
728 val_lower = get_tarval_long(tv_lower);
729 size *= val_upper - val_lower;
731 restore_optimization_state(&state);
733 DBG(("array %s -> (elements: %d, element type size: %d)\n",
735 size, get_type_size_bytes(get_array_element_type(ty))));
736 size *= get_type_size_bytes(get_array_element_type(ty));
745 set_type_size_bytes(ty, size);
746 set_type_state(ty, layout_fixed);
749 DBG(("size: %d\n", get_type_size_bytes(ty)));
752 } /* compute_type_size */
755 * layout all non-frame types of the Firm graph
757 static void compute_type_sizes(void)
762 /* all non-frame other types */
763 for (i = get_irp_n_types() - 1; i >= 0; --i) {
764 tp = get_irp_type(i);
765 compute_type_size(tp);
767 if (is_Method_type(tp)) {
768 tp = get_method_value_res_type(tp);
771 /* we have a value result type for this method, lower */
772 compute_type_size(tp);
776 } /* compute_type_sizes */
779 * layout all frame-types of the Firm graph
781 static void compute_frame_type_sizes(void)
786 /* all frame types */
787 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
788 irg = get_irp_irg(i);
789 /* do not optimize away variables in debug mode */
790 if (firm_opt.debug_mode == DBG_MODE_NONE)
792 compute_type_size(get_irg_frame_type(irg));
794 } /* compute_frame_type_sizes */
799 * @param input_filename the name of the (main) source file
801 static void do_firm_lowering(const char *input_filename)
805 if (firm_opt.lower_ll) {
806 lwrdw_param_t init = {
809 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
810 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
811 get_atomic_mode(ATOMIC_TYPE_INT),
812 get_atomic_mode(ATOMIC_TYPE_UINT),
813 def_create_intrinsic_fkt,
818 if (arch_create_intrinsic) {
819 init.create_intrinsic = arch_create_intrinsic;
820 init.ctx = create_intrinsic_ctx;
822 timer_push(TV_DW_LOWER);
827 if (firm_dump.statistic & STAT_AFTER_LOWER)
828 stat_dump_snapshot(input_filename, "low");
830 /* verify lowered graphs */
831 timer_push(TV_VERIFY);
832 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
833 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
836 DUMP_ALL(firm_dump.ir_graph, "-low");
838 if (firm_opt.enabled) {
839 timer_start(TV_ALL_OPT);
841 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
843 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
844 ir_graph *irg = get_irp_irg(i);
845 do_irg_opt(irg, "reassoc");
848 /* enable architecture dependent optimizations */
849 arch_dep_set_opts((arch_dep_opts_t)
850 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
851 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
852 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
854 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
855 ir_graph *irg = get_irp_irg(i);
857 current_ir_graph = irg;
859 do_irg_opt(irg, "local");
860 do_irg_opt(irg, "gcse");
861 do_irg_opt(irg, "ldst");
862 do_irg_opt(irg, "local");
863 do_irg_opt(irg, "controlflow");
865 if (firm_opt.if_conversion) {
866 do_irg_opt(irg, "ifconv");
867 do_irg_opt(irg, "local");
868 do_irg_opt(irg, "controlflow");
871 do_irg_opt(current_ir_graph, "sync");
873 timer_stop(TV_ALL_OPT);
875 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
879 mark_private_methods();
881 /* set the phase to low */
882 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
883 set_irg_phase_low(get_irp_irg(i));
885 /* all graphs are lowered, set the irp phase to low */
886 set_irp_phase_state(phase_low);
888 if (firm_dump.statistic & STAT_FINAL) {
889 stat_dump_snapshot(input_filename, "final");
891 } /* do_firm_lowering */
894 * Initialize for the Firm-generating back end.
896 void gen_firm_init(void)
898 firm_parameter_t params;
899 unsigned pattern = 0;
901 /* the automatic state is only set if inlining is enabled */
902 firm_opt.auto_inline = firm_opt.do_inline;
904 if (firm_dump.stat_pattern)
905 pattern |= FIRMSTAT_PATTERN_ENABLED;
907 if (firm_dump.stat_dag)
908 pattern |= FIRMSTAT_COUNT_DAG;
910 memset(¶ms, 0, sizeof(params));
911 params.size = sizeof(params);
912 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
913 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
914 params.initialize_local_func = uninitialized_local_var;
915 params.cc_mask = 0; /* no regparam, cdecl */
916 params.builtin_dbg = NULL;
919 /* Activate Graph rewriting if SIMD optimization is turned on */
920 /* This has to be done before init_firm() is called! */
921 if (firm_ext_grs.simd_opt)
927 if (firm_be_opt.selection == BE_FIRM_BE) {
928 const backend_params *be_params = be_get_backend_param();
930 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
931 params.arch_op_settings = be_params->arch_op_settings;
933 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
934 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
936 ad_param = be_params->dep_param;
937 if_conv_info = be_params->if_conv_info;
939 if (be_params->has_imm_fp_mode)
940 firm_imm_fp_mode = be_params->imm_fp_mode;
942 /* OS option must be set to the backend */
943 switch (firm_opt.os_support) {
944 case OS_SUPPORT_MINGW:
945 firm_be_option("ia32-gasmode=mingw");
947 case OS_SUPPORT_MACHO:
948 firm_be_option("ia32-gasmode=macho");
950 case OS_SUPPORT_LINUX:
952 firm_be_option("ia32-gasmode=linux");
956 dbg_init(NULL, NULL, dbg_snprint);
957 edges_init_dbg(firm_opt.vrfy_edges);
958 //cbackend_set_debug_retrieve(dbg_retrieve);
960 set_opt_precise_exc_context(firm_opt.precise_exc);
961 set_opt_fragile_ops(firm_opt.fragile_ops);
963 /* Sel node cannot produce NULL pointers */
964 set_opt_sel_based_null_check_elim(1);
966 /* dynamic dispatch works currently only if whole world scenarios */
967 set_opt_dyn_meth_dispatch(0);
969 arch_dep_init(arch_factory);
971 /* do not run architecture dependent optimizations in building phase */
972 arch_dep_set_opts(arch_dep_none);
974 do_node_verification((firm_verification_t) firm_opt.vrfy);
975 if (firm_dump.filter)
976 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
978 if (firm_opt.enabled) {
980 set_opt_constant_folding(firm_opt.const_folding);
981 set_opt_algebraic_simplification(firm_opt.const_folding);
982 set_opt_cse(firm_opt.cse);
983 set_opt_global_cse(0);
984 set_opt_unreachable_code(1);
985 set_opt_control_flow(firm_opt.control_flow);
986 set_opt_control_flow_weak_simplification(1);
987 set_opt_control_flow_strong_simplification(1);
992 /* do not dump entity ld names */
994 } /* gen_firm_init */
997 * Called, after the Firm generation is completed,
998 * do all optimizations and backend call here.
1000 * @param out a file handle for the output, may be NULL
1001 * @param input_filename the name of the (main) source file
1002 * @param c_mode non-zero if "C" was compiled
1003 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
1005 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
1009 firm_const_exists = new_firm_const_exists;
1011 /* the general for dumping option must be set, or the others will not work */
1013 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1015 dump_keepalive_edges(1);
1016 dump_consts_local(1);
1017 dump_dominator_information(1);
1018 dump_loop_information(0);
1020 if (!firm_dump.edge_labels)
1021 turn_off_edge_labels();
1023 if (firm_dump.all_types) {
1026 dump_class_hierarchy(0, "");
1027 dump_class_hierarchy(1, "-with-entities");
1031 /* finalize all graphs */
1032 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1033 ir_graph *irg = get_irp_irg(i);
1035 irg_finalize_cons(irg);
1036 DUMP_ONE(firm_dump.ir_graph, irg, "");
1038 /* verify the graph */
1039 timer_push(TV_VERIFY);
1040 irg_verify(irg, VRFY_ENFORCE_SSA);
1044 timer_push(TV_VERIFY);
1048 /* all graphs are finalized, set the irp phase to high */
1049 set_irp_phase_state(phase_high);
1051 /* BEWARE: kill unreachable code before doing compound lowering */
1052 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1053 ir_graph *irg = get_irp_irg(i);
1057 /* lower all compound call return values */
1058 lower_compound_params();
1060 /* computes the sizes of all types that are still not computed */
1061 compute_type_sizes();
1063 /* lower copyb nodes */
1064 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1065 ir_graph *irg = get_irp_irg(i);
1066 lower_CopyB(irg, 128, 4);
1069 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1070 stat_dump_snapshot(input_filename, "noopt");
1073 if (firm_opt.enabled)
1074 do_firm_optimizations(input_filename);
1076 if (firm_dump.gen_firm_asm) {
1077 timer_push(TV_FIRM_ASM);
1078 gen_Firm_assembler(input_filename);
1084 do_firm_lowering(input_filename);
1086 /* computes the sizes of all frame types */
1087 compute_frame_type_sizes();
1089 /* set the phase to low */
1090 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1091 set_irg_phase_low(get_irp_irg(i));
1095 /** SIMD Optimization Extensions **/
1097 /* Pattern creation step. No code has to be generated, so
1098 exit after pattern creation */
1099 if (firm_ext_grs.create_pattern) {
1100 ext_grs_create_pattern();
1104 /* SIMD optimization step. Uses graph patterns to find
1105 rich instructions and rewrite */
1106 if (firm_ext_grs.simd_opt)
1110 if (firm_dump.statistic & STAT_FINAL_IR)
1111 stat_dump_snapshot(input_filename, "final-ir");
1113 /* run the code generator */
1114 if (firm_be_opt.selection != BE_NONE)
1115 do_codegen(out, input_filename);
1117 if (firm_dump.statistic & STAT_FINAL)
1118 stat_dump_snapshot(input_filename, "final");
1121 if (firm_opt.ycomp_dbg)
1122 firm_finish_ycomp_debugger();
1124 } /* gen_firm_finish */
1127 * Do very early initializations
1129 void firm_early_init(void) {
1130 /* arg: need this here for command line options */
1132 firm_init_options(NULL, 0, NULL);
1133 } /* firm_early_init */