3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
11 #include <libfirm/firm.h>
12 #include <libfirm/be.h>
16 #include "firm_codegen.h"
17 #include "firm_cmdline.h"
18 #include "firm_timing.h"
21 #define snprintf _snprintf
24 #if defined(_DEBUG) || defined(FIRM_DEBUG)
25 #define DBG(x) dbg_printf x
28 #endif /* _DEBUG || FIRM_DEBUG */
31 /** dump all the graphs depending on cond */
32 #define DUMP_ALL(cond, suffix) \
35 timer_push(TV_VCG_DUMP); \
36 if (firm_dump.no_blocks) \
37 dump_all_ir_graphs(dump_ir_graph, suffix); \
38 else if (firm_dump.extbb) \
39 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
41 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
46 /** dump all control flow graphs depending on cond */
47 #define DUMP_ALL_CFG(cond, suffix) \
50 timer_push(TV_VCG_DUMP); \
51 dump_all_ir_graphs(dump_cfg, suffix); \
56 /** check all graphs depending on cond */
57 #define CHECK_ALL(cond) \
61 timer_push(TV_VERIFY); \
62 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
63 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
70 /** dump graphs irg depending on cond */
71 #define DUMP_ONE(cond, irg, suffix) \
74 timer_push(TV_VCG_DUMP); \
75 if (firm_dump.no_blocks) \
76 dump_ir_graph(irg, suffix); \
77 else if (firm_dump.extbb) \
78 dump_ir_extblock_graph(irg, suffix); \
80 dump_ir_block_graph(irg, suffix); \
85 /** dump control flow graph irg depending on cond */
86 #define DUMP_ONE_CFG(cond, irg, suffix) \
89 timer_push(TV_VCG_DUMP); \
90 dump_cfg(irg, suffix); \
95 /** check a graph irg depending on cond */
96 #define CHECK_ONE(cond, irg) \
99 timer_push(TV_VERIFY); \
100 irg_verify(irg, VRFY_ENFORCE_SSA); \
106 /* set by the backend parameters */
107 static const ir_settings_arch_dep_t *ad_param = NULL;
108 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
109 static void *create_intrinsic_ctx = NULL;
110 static const ir_settings_if_conv_t *if_conv_info = NULL;
112 ir_mode *firm_imm_fp_mode = NULL;
114 /* entities of runtime functions */
115 ir_entity_ptr rts_entities[rts_max];
118 * factory for setting architecture dependent parameters
120 static const ir_settings_arch_dep_t *arch_factory(void)
122 static const ir_settings_arch_dep_t param = {
123 1, /* also use subs */
124 4, /* maximum shifts */
125 31, /* maximum shift amount */
126 NULL, /* use default evaluator */
130 32 /* Mulh allowed up to 32 bit */
133 return ad_param ? ad_param : ¶m;
137 * Map runtime functions.
139 static void rts_map(void) {
140 static const struct {
141 ir_entity_ptr *ent; /**< address of the rts entity */
142 i_mapper_func func; /**< mapper function. */
145 { &rts_entities[rts_abs], i_mapper_abs },
146 { &rts_entities[rts_labs], i_mapper_abs },
147 { &rts_entities[rts_llabs], i_mapper_abs },
148 { &rts_entities[rts_imaxabs], i_mapper_abs },
149 { &rts_entities[rts_alloca], i_mapper_alloca },
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);
246 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
249 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
250 DUMP_ONE_CFG(1, irg, name);
253 static void dump_all_count(const char *const suffix)
255 const int n_irgs = get_irp_n_irgs();
258 for (i = 0; i < n_irgs; ++i)
259 dump_graph_count(get_irp_irg(i), suffix);
262 #define DUMP_ONE_C(cond, irg, suffix) \
265 dump_graph_count((irg), (suffix)); \
269 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
272 dump_graph_cfg_count((irg), (suffix)); \
276 #define DUMP_ALL_C(cond, suffix) \
279 dump_all_count((suffix)); \
284 * run all the Firm optimizations
286 * @param input_filename the name of the (main) source file
287 * @param firm_const_exists non-zero, if the const attribute was used on functions
289 static void do_firm_optimizations(const char *input_filename, int firm_const_exists)
295 /* FIXME: cloning might ADD new graphs. */
296 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
298 set_opt_strength_red(firm_opt.strength_red);
299 set_opt_scalar_replacement(firm_opt.scalar_replace);
300 set_opt_auto_create_sync(firm_opt.auto_sync);
301 set_opt_alias_analysis(firm_opt.alias_analysis);
303 aa_opt = aa_opt_no_opt;
304 if (firm_opt.strict_alias)
305 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
306 if (firm_opt.no_alias)
307 aa_opt = aa_opt_no_alias;
309 set_irp_memory_disambiguator_options(aa_opt);
311 timer_start(TV_ALL_OPT);
313 if (firm_opt.remove_unused) {
314 ir_entity **keep_methods;
317 /* Analysis that finds the free methods,
318 i.e. methods that are dereferenced.
319 Optimizes polymorphic calls :-). */
320 cgana(&arr_len, &keep_methods);
322 /* Remove methods that are never called. */
323 gc_irgs(arr_len, keep_methods);
328 if (! firm_opt.freestanding) {
330 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "rts");
331 CHECK_ALL(firm_opt.check_all);
334 /* first step: kill dead code */
335 for (i = 0; i < get_irp_n_irgs(); i++) {
336 irg = current_ir_graph = get_irp_irg(i);
337 /* Confirm construction currently can only handle blocks with only one control
338 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
339 the optimization of switch constructs. */
340 timer_push(TV_CF_OPT);
341 optimize_graph_df(irg);
346 if (firm_opt.func_calls) {
347 timer_push(TV_REAL_FUNC_CALL);
348 optimize_funccalls(firm_const_exists, NULL);
350 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "func_call");
351 CHECK_ALL(firm_opt.check_all);
354 /* do lowering on the const code irg */
357 for (i = 0; i < get_irp_n_irgs(); i++) {
358 irg = current_ir_graph = get_irp_irg(i);
361 /* If SIMD optimization is on, make sure we have only 1 return */
362 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
363 normalize_one_return(irg);
367 if (firm_opt.modes) {
368 /* convert all modes into integer if possible */
369 arch_mode_conversion(irg, predefs.mode_uint);
373 timer_push(TV_SCALAR_REPLACE);
374 scalar_replacement_opt(irg);
376 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "scalar");
377 CHECK_ONE(firm_opt.check_all, irg);
379 timer_push(TV_LOCAL_OPT);
380 optimize_graph_df(irg);
382 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
383 CHECK_ONE(firm_opt.check_all, irg);
385 timer_push(TV_REASSOCIATION);
386 optimize_reassociation(irg);
388 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "reassoc");
389 CHECK_ONE(firm_opt.check_all, irg);
392 CHECK_ONE(firm_opt.check_all, irg);
394 if (firm_opt.code_place) {
395 timer_push(TV_CODE_PLACE);
396 set_opt_global_cse(1);
397 optimize_graph_df(irg);
399 set_opt_global_cse(0);
401 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "place");
402 CHECK_ONE(firm_opt.check_all, irg);
405 if (firm_opt.confirm) {
406 /* Confirm construction currently can only handle blocks with only one control
407 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
408 the optimization of switch constructs. */
409 timer_push(TV_CF_OPT);
412 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
413 CHECK_ONE(firm_opt.check_all, irg);
414 timer_push(TV_CONFIRM_CREATE);
415 construct_confirms(irg);
417 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "confirms");
418 CHECK_ONE(firm_opt.check_all, irg);
420 timer_push(TV_LOCAL_OPT);
421 optimize_graph_df(irg);
423 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
424 CHECK_ONE(firm_opt.check_all, irg);
428 if (firm_opt.luffig) {
430 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst2");
431 CHECK_ONE(firm_opt.check_all, irg);
435 timer_push(TV_CF_OPT);
438 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
439 CHECK_ONE(firm_opt.check_all, irg);
441 irg_verify(irg, VRFY_ENFORCE_SSA);
442 if (firm_opt.gvn_pre) {
444 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "pre");
445 CHECK_ONE(firm_opt.check_all, irg);
446 irg_verify(irg, VRFY_ENFORCE_SSA);
449 if (firm_opt.load_store) {
450 timer_push(TV_LOAD_STORE);
451 optimize_load_store(irg);
453 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst");
454 CHECK_ONE(firm_opt.check_all, irg);
457 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
459 if (firm_opt.deconv) {
460 timer_push(TV_DECONV);
463 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "deconv");
464 CHECK_ONE(firm_opt.check_all, irg);
467 if (firm_opt.cond_eval) {
468 timer_push(TV_COND_EVAL);
471 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
472 CHECK_ONE(firm_opt.check_all, irg);
475 /* should we really remove the Confirm here? */
476 if (firm_opt.confirm) {
477 timer_push(TV_CONFIRM_CREATE);
478 remove_confirms(irg);
483 compute_postdoms(irg);
484 DUMP_ONE_CFG_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dom");
485 CHECK_ONE(firm_opt.check_all, irg);
487 construct_backedges(irg);
489 timer_push(TV_CF_OPT);
492 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
493 CHECK_ONE(firm_opt.check_all, irg);
495 if (firm_opt.if_conversion) {
496 timer_push(TV_IF_CONV);
497 opt_if_conv(current_ir_graph, if_conv_info);
499 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "if");
500 CHECK_ONE(firm_opt.check_all, current_ir_graph);
502 timer_push(TV_LOCAL_OPT);
503 optimize_graph_df(current_ir_graph);
505 timer_push(TV_CF_OPT);
506 optimize_cf(current_ir_graph);
508 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "after_if");
509 CHECK_ONE(firm_opt.check_all, current_ir_graph);
512 if (firm_opt.bool_opt) {
514 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "bool");
515 CHECK_ONE(firm_opt.check_all, irg);
519 opt_osr(current_ir_graph, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
521 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "stred");
522 CHECK_ONE(firm_opt.check_all, irg);
524 timer_push(TV_LOCAL_OPT);
525 optimize_graph_df(irg);
527 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
528 CHECK_ONE(firm_opt.check_all, irg);
530 edges_deactivate(irg);
531 timer_push(TV_DEAD_NODE);
532 dead_node_elimination(irg);
534 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dead");
535 CHECK_ONE(firm_opt.check_all, irg);
538 if (firm_opt.do_inline) {
539 inline_functions(5000, 500);
540 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
541 CHECK_ALL(firm_opt.check_all);
543 if (firm_opt.cloning) {
544 proc_cloning((float)firm_opt.clone_threshold);
545 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "clone");
546 CHECK_ALL(firm_opt.check_all);
548 if (firm_opt.tail_rec) {
549 timer_push(TV_TAIL_REC);
550 opt_tail_recursion();
553 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
554 CHECK_ALL(firm_opt.check_all);
557 if (firm_opt.cond_eval) {
558 for (i = 0; i < get_irp_n_irgs(); i++) {
559 if (firm_opt.luffig) {
561 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst2");
562 CHECK_ONE(firm_opt.check_all, irg);
565 irg = get_irp_irg(i);
566 timer_push(TV_COND_EVAL);
569 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
570 CHECK_ONE(firm_opt.check_all, irg);
574 /* final run of local optimizations */
575 for (i = 0; i < get_irp_n_irgs(); i++) {
576 irg = get_irp_irg(i);
577 timer_push(TV_LOCAL_OPT);
578 optimize_graph_df(irg);
580 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "local_opt");
581 CHECK_ONE(firm_opt.check_all, irg);
583 timer_push(TV_CF_OPT);
586 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
587 CHECK_ONE(firm_opt.check_all, irg);
590 if (firm_dump.ir_graph) {
591 /* recompute backedges for nicer dumps */
592 for (i = 0; i < get_irp_n_irgs(); i++)
593 construct_cf_backedges(get_irp_irg(i));
596 if (firm_opt.remove_unused) {
597 ir_entity **keep_methods;
600 /* Analysis that finds the free methods,
601 i.e. methods that are dereferenced.
602 Optimizes polymorphic calls :-). */
603 cgana(&arr_len, &keep_methods);
605 /* Remove methods that are never called. */
606 gc_irgs(arr_len, keep_methods);
612 DUMP_ALL(firm_dump.ir_graph, "-opt");
614 /* verify optimized graphs */
615 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
616 ir_graph *irg = get_irp_irg(i);
617 irg_verify(irg, VRFY_ENFORCE_SSA);
620 if (firm_dump.statistic & STAT_AFTER_OPT)
621 stat_dump_snapshot(input_filename, "opt");
623 timer_stop(TV_ALL_OPT);
624 } /* do_firm_optimizations */
627 * compute the size of a type (do implicit lowering)
629 * @param ty a Firm type
631 static int compute_type_size(ir_type *ty)
633 optimization_state_t state;
634 unsigned align_all = 1;
635 int n, size = 0, set = 0;
638 if (get_type_state(ty) == layout_fixed) {
639 /* do not layout already layouted types again */
643 if (is_Method_type(ty) || ty == get_glob_type()) {
644 /* no need for size calculation for method types or the global type */
648 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
650 switch (get_type_tpop_code(ty)) {
653 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
654 ir_entity *ent = get_compound_member(ty, i);
655 ir_type *ent_ty = get_entity_type(ent);
656 unsigned align, misalign;
658 /* compute member types */
659 if (! compute_type_size(ent_ty))
662 align = get_type_alignment_bytes(ent_ty);
663 align_all = align > align_all ? align : align_all;
664 misalign = (align ? size % align : 0);
665 size += (misalign ? align - misalign : 0);
667 set_entity_offset(ent, size);
668 size += get_type_size_bytes(ent_ty);
670 DBG((" member %s %s -> (size: %u, align: %u)\n",
671 get_type_name(ent_ty), get_entity_name(ent),
672 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
674 if (align_all > 0 && size % align_all) {
675 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
676 size += align_all - (size % align_all);
677 DBG(("correcting type-size to %d\n", size));
679 set_type_alignment_bytes(ty, align_all);
684 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
685 ir_entity *ent = get_union_member(ty, i);
687 if (! compute_type_size(get_entity_type(ent)))
689 s = get_type_size_bytes(get_entity_type(ent));
691 set_entity_offset(ent, 0);
692 size = (s > size ? s : size);
698 dims = get_array_n_dimensions(ty);
700 if (! compute_type_size(get_array_element_type(ty)))
705 save_optimization_state(&state);
707 set_opt_constant_folding(1);
708 set_opt_algebraic_simplification(1);
710 for (i = 0; i < dims; ++i) {
711 ir_node *lower = get_array_lower_bound(ty, i);
712 ir_node *upper = get_array_upper_bound(ty, i);
713 ir_graph *rem = current_ir_graph;
714 tarval *tv_lower, *tv_upper;
715 long val_lower, val_upper;
717 current_ir_graph = get_const_code_irg();
718 local_optimize_node(lower);
719 local_optimize_node(upper);
720 current_ir_graph = rem;
722 tv_lower = computed_value(lower);
723 tv_upper = computed_value(upper);
725 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
727 * we cannot calculate the size of this array yet, it
728 * even might be unknown until the end, like argv[]
730 restore_optimization_state(&state);
734 val_upper = get_tarval_long(tv_upper);
735 val_lower = get_tarval_long(tv_lower);
736 size *= val_upper - val_lower;
738 restore_optimization_state(&state);
740 DBG(("array %s -> (elements: %d, element type size: %d)\n",
742 size, get_type_size_bytes(get_array_element_type(ty))));
743 size *= get_type_size_bytes(get_array_element_type(ty));
752 set_type_size_bytes(ty, size);
753 set_type_state(ty, layout_fixed);
756 DBG(("size: %d\n", get_type_size_bytes(ty)));
759 } /* compute_type_size */
762 * layout all non-frame types of the Firm graph
764 static void compute_type_sizes(void)
769 /* all non-frame other types */
770 for (i = get_irp_n_types() - 1; i >= 0; --i) {
771 tp = get_irp_type(i);
772 compute_type_size(tp);
774 if (is_Method_type(tp)) {
775 tp = get_method_value_res_type(tp);
778 /* we have a value result type for this method, lower */
779 compute_type_size(tp);
783 } /* compute_type_sizes */
786 * layout all frame-types of the Firm graph
788 static void compute_frame_type_sizes(void)
793 /* all frame types */
794 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
795 irg = get_irp_irg(i);
796 /* do not optimize away variables in debug mode */
797 if (firm_opt.debug_mode == DBG_MODE_NONE)
799 compute_type_size(get_irg_frame_type(irg));
801 } /* compute_frame_type_sizes */
806 * @param input_filename the name of the (main) source file
808 static void do_firm_lowering(const char *input_filename)
812 /* do class lowering and vtbl creation */
813 // lower_classes_to_struct("vtbl", "m");
816 timer_push(TV_LOWER);
821 if (firm_opt.lower_ll) {
822 lwrdw_param_t init = {
827 def_create_intrinsic_fkt,
832 if (arch_create_intrinsic) {
833 init.create_intrinsic = arch_create_intrinsic;
834 init.ctx = create_intrinsic_ctx;
836 timer_push(TV_DW_LOWER);
841 if (firm_dump.statistic & STAT_AFTER_LOWER)
842 stat_dump_snapshot(input_filename, "low");
844 /* verify lowered graphs */
845 timer_push(TV_VERIFY);
846 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
847 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
850 DUMP_ALL(firm_dump.ir_graph, "-low");
852 if (firm_opt.enabled) {
853 timer_start(TV_ALL_OPT);
855 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
857 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
858 current_ir_graph = get_irp_irg(i);
860 timer_push(TV_REASSOCIATION);
861 optimize_reassociation(current_ir_graph);
863 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
864 CHECK_ONE(firm_opt.check_all, current_ir_graph);
867 /* enable architecture dependent optimizations */
868 arch_dep_set_opts((arch_dep_opts_t)
869 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
870 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
871 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
873 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
874 current_ir_graph = get_irp_irg(i);
876 if (firm_opt.code_place)
877 set_opt_global_cse(1);
879 timer_push(TV_LOCAL_OPT);
880 optimize_graph_df(current_ir_graph);
882 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
883 if (! firm_opt.code_place)
884 CHECK_ONE(firm_opt.check_all, current_ir_graph);
886 if (firm_opt.code_place) {
887 timer_push(TV_CODE_PLACE);
888 place_code(current_ir_graph);
889 set_opt_global_cse(0);
891 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
892 CHECK_ONE(firm_opt.check_all, current_ir_graph);
895 timer_push(TV_LOAD_STORE);
896 optimize_load_store(current_ir_graph);
898 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
899 CHECK_ONE(firm_opt.check_all, current_ir_graph);
902 timer_push(TV_LOCAL_OPT);
903 optimize_graph_df(current_ir_graph);
905 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
907 timer_push(TV_CF_OPT);
908 optimize_cf(current_ir_graph);
910 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
911 CHECK_ONE(firm_opt.check_all, current_ir_graph);
913 if (firm_opt.if_conversion) {
914 timer_push(TV_IF_CONV);
915 opt_if_conv(current_ir_graph, if_conv_info);
917 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
918 CHECK_ONE(firm_opt.check_all, current_ir_graph);
920 timer_push(TV_LOCAL_OPT);
921 optimize_graph_df(current_ir_graph);
923 timer_push(TV_CF_OPT);
924 optimize_cf(current_ir_graph);
926 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
927 CHECK_ONE(firm_opt.check_all, current_ir_graph);
930 timer_stop(TV_ALL_OPT);
932 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
936 mark_private_methods();
938 /* set the phase to low */
939 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
940 set_irg_phase_low(get_irp_irg(i));
942 /* all graphs are lowered, set the irp phase to low */
943 set_irp_phase_state(phase_low);
945 if (firm_dump.statistic & STAT_FINAL) {
946 stat_dump_snapshot(input_filename, "final");
948 } /* do_firm_lowering */
951 * Initialize for the Firm-generating back end.
953 void gen_firm_init(void)
955 firm_parameter_t params;
956 unsigned pattern = 0;
958 /* the automatic state is only set if inlining is enabled */
959 firm_opt.auto_inline = firm_opt.do_inline;
961 if (firm_dump.stat_pattern)
962 pattern |= FIRMSTAT_PATTERN_ENABLED;
964 if (firm_dump.stat_dag)
965 pattern |= FIRMSTAT_COUNT_DAG;
967 memset(¶ms, 0, sizeof(params));
968 params.size = sizeof(params);
969 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
970 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
971 params.initialize_local_func = uninitialized_local_var;
972 params.cc_mask = 0; /* no regparam, cdecl */
973 params.builtin_dbg = NULL;
976 /* Activate Graph rewriting if SIMD optimization is turned on */
977 /* This has to be done before init_firm() is called! */
978 if (firm_ext_grs.simd_opt)
984 if (firm_be_opt.selection == BE_FIRM_BE) {
985 const backend_params *be_params = be_get_backend_param();
987 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
988 params.arch_op_settings = be_params->arch_op_settings;
990 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
991 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
993 ad_param = be_params->dep_param;
994 if_conv_info = be_params->if_conv_info;
996 if (be_params->has_imm_fp_mode)
997 firm_imm_fp_mode = be_params->imm_fp_mode;
999 /* OS option must be set to the backend */
1000 switch (firm_opt.os_support) {
1001 case OS_SUPPORT_MINGW:
1002 firm_be_option("ia32-gasmode=mingw");
1004 case OS_SUPPORT_MACHO:
1005 firm_be_option("ia32-gasmode=macho");
1007 case OS_SUPPORT_LINUX:
1009 firm_be_option("ia32-gasmode=linux");
1013 dbg_init(NULL, NULL, dbg_snprint);
1014 edges_init_dbg(firm_opt.vrfy_edges);
1015 //cbackend_set_debug_retrieve(dbg_retrieve);
1017 set_opt_precise_exc_context(firm_opt.precise_exc);
1018 set_opt_fragile_ops(firm_opt.fragile_ops);
1020 /* Sel node cannot produce NULL pointers */
1021 set_opt_sel_based_null_check_elim(1);
1023 /* dynamic dispatch works currently only if whole world scenarios */
1024 set_opt_dyn_meth_dispatch(0);
1026 arch_dep_init(arch_factory);
1028 /* do not run architecture dependent optimizations in building phase */
1029 arch_dep_set_opts(arch_dep_none);
1031 do_node_verification((firm_verification_t) firm_opt.vrfy);
1032 if (firm_dump.filter)
1033 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
1035 if (firm_opt.enabled) {
1037 set_opt_constant_folding(firm_opt.const_folding);
1038 set_opt_algebraic_simplification(firm_opt.const_folding);
1039 set_opt_cse(firm_opt.cse);
1040 set_opt_global_cse(0);
1041 set_opt_unreachable_code(1);
1042 set_opt_control_flow(firm_opt.control_flow);
1043 set_opt_control_flow_weak_simplification(1);
1044 set_opt_control_flow_strong_simplification(1);
1049 /* do not dump entity ld names */
1053 /* init the ycomp debugger extension */
1054 if (firm_opt.ycomp_dbg)
1055 firm_init_ycomp_debugger(firm_opt.ycomp_host, firm_opt.ycomp_port);
1057 } /* gen_firm_init */
1060 * Called, after the Firm generation is completed,
1061 * do all optimizations and backend call here.
1063 * @param out a file handle for the output, may be NULL
1064 * @param input_filename the name of the (main) source file
1065 * @param c_mode non-zero if "C" was compiled
1066 * @param firm_const_exists non-zero, if the const attribute was used on functions
1068 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int firm_const_exists)
1072 /* the general for dumping option must be set, or the others will not work */
1074 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1076 dump_keepalive_edges(1);
1077 dump_consts_local(1);
1078 dump_dominator_information(1);
1079 dump_loop_information(0);
1081 if (!firm_dump.edge_labels)
1082 turn_off_edge_labels();
1084 if (firm_dump.all_types) {
1087 dump_class_hierarchy(0, "");
1088 dump_class_hierarchy(1, "-with-entities");
1092 /* finalize all graphs */
1093 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1094 ir_graph *irg = get_irp_irg(i);
1096 irg_finalize_cons(irg);
1097 DUMP_ONE(firm_dump.ir_graph, irg, "");
1099 /* verify the graph */
1100 timer_push(TV_VERIFY);
1101 irg_verify(irg, VRFY_ENFORCE_SSA);
1105 timer_push(TV_VERIFY);
1109 /* all graphs are finalized, set the irp phase to high */
1110 set_irp_phase_state(phase_high);
1112 /* BEWARE: kill unreachable code before doing compound lowering */
1113 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1114 ir_graph *irg = get_irp_irg(i);
1118 /* lower all compound call return values */
1119 lower_compound_params();
1121 /* computes the sizes of all types that are still not computed */
1122 compute_type_sizes();
1124 /* lower copyb nodes */
1125 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1126 ir_graph *irg = get_irp_irg(i);
1127 lower_CopyB(irg, 128, 4);
1130 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1131 stat_dump_snapshot(input_filename, "noopt");
1134 if (firm_opt.enabled)
1135 do_firm_optimizations(input_filename, firm_const_exists);
1137 if (firm_dump.gen_firm_asm) {
1138 timer_push(TV_FIRM_ASM);
1139 gen_Firm_assembler(input_filename);
1145 do_firm_lowering(input_filename);
1147 /* computes the sizes of all frame types */
1148 compute_frame_type_sizes();
1150 /* set the phase to low */
1151 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1152 set_irg_phase_low(get_irp_irg(i));
1156 /** SIMD Optimization Extensions **/
1158 /* Pattern creation step. No code has to be generated, so
1159 exit after pattern creation */
1160 if (firm_ext_grs.create_pattern) {
1161 ext_grs_create_pattern();
1165 /* SIMD optimization step. Uses graph patterns to find
1166 rich instructions and rewrite */
1167 if (firm_ext_grs.simd_opt)
1171 if (firm_dump.statistic & STAT_FINAL_IR)
1172 stat_dump_snapshot(input_filename, "final-ir");
1174 /* run the code generator */
1175 if (firm_be_opt.selection != BE_NONE)
1176 do_codegen(out, input_filename);
1178 if (firm_dump.statistic & STAT_FINAL)
1179 stat_dump_snapshot(input_filename, "final");
1182 if (firm_opt.ycomp_dbg)
1183 firm_finish_ycomp_debugger();
1185 } /* gen_firm_finish */
1188 * Do very early initializations
1190 void firm_early_init(void) {
1191 /* arg: need this here for command line options */
1193 firm_init_options(NULL, 0, NULL);
1194 } /* firm_early_init */