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>
14 #include <libfirm/be.h>
15 #endif /* FIRM_BACKEND */
19 #include "firm_codegen.h"
20 #include "firm_cmdline.h"
21 #include "firm_timing.h"
24 #define snprintf _snprintf
27 #if defined(_DEBUG) || defined(FIRM_DEBUG)
28 #define DBG(x) dbg_printf x
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);
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 if (firm_opt.tail_rec) {
335 timer_push(TV_TAIL_REC);
336 opt_tail_recursion();
339 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
340 CHECK_ALL(firm_opt.check_all);
343 if (firm_opt.func_calls) {
344 timer_push(TV_REAL_FUNC_CALL);
345 optimize_funccalls(firm_const_exists, NULL);
347 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "func_call");
348 CHECK_ALL(firm_opt.check_all);
351 if (firm_opt.do_inline) {
352 timer_push(TV_INLINE);
353 inline_leave_functions(500, 80, 30, FALSE);
355 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
356 CHECK_ALL(firm_opt.check_all);
359 /* do lowering on the const code irg */
362 for (i = 0; i < get_irp_n_irgs(); i++) {
363 irg = current_ir_graph = get_irp_irg(i);
367 /* If SIMD optimization is on, make sure we have only 1 return */
368 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
369 normalize_one_return(irg);
374 if (firm_opt.modes) {
375 /* convert all modes into integer if possible */
376 arch_mode_conversion(irg, predefs.mode_uint);
379 timer_push(TV_SCALAR_REPLACE);
380 scalar_replacement_opt(irg);
382 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "scalar");
383 CHECK_ONE(firm_opt.check_all, irg);
385 timer_push(TV_LOCAL_OPT);
386 optimize_graph_df(irg);
388 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
389 CHECK_ONE(firm_opt.check_all, irg);
391 timer_push(TV_REASSOCIATION);
392 optimize_reassociation(irg);
394 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "reassoc");
395 CHECK_ONE(firm_opt.check_all, irg);
397 if (firm_opt.confirm) {
398 /* Confirm construction currently can only handle blocks with only one control
399 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
400 the optimization of switch constructs. */
401 timer_push(TV_CF_OPT);
404 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
405 CHECK_ONE(firm_opt.check_all, irg);
406 timer_push(TV_CONFIRM_CREATE);
407 construct_confirms(irg);
409 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "confirms");
410 CHECK_ONE(firm_opt.check_all, irg);
413 timer_push(TV_LOCAL_OPT);
414 optimize_graph_df(irg);
416 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
417 CHECK_ONE(firm_opt.check_all, irg);
420 CHECK_ONE(firm_opt.check_all, irg);
422 if (firm_opt.code_place) {
423 timer_push(TV_CODE_PLACE);
424 set_opt_global_cse(1);
425 optimize_graph_df(irg);
427 set_opt_global_cse(0);
429 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "place");
430 CHECK_ONE(firm_opt.check_all, irg);
433 if (firm_opt.luffig) {
435 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst2");
436 CHECK_ONE(firm_opt.check_all, irg);
439 timer_push(TV_CF_OPT);
442 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
443 CHECK_ONE(firm_opt.check_all, irg);
445 /* should we really remove the Confirm here? */
446 if (firm_opt.confirm) {
447 timer_push(TV_CONFIRM_CREATE);
448 remove_confirms(irg);
452 irg_verify(irg, VRFY_ENFORCE_SSA);
453 if (firm_opt.gvn_pre) {
455 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "pre");
456 CHECK_ONE(firm_opt.check_all, irg);
457 irg_verify(irg, VRFY_ENFORCE_SSA);
460 if (firm_opt.load_store) {
461 timer_push(TV_LOAD_STORE);
462 optimize_load_store(irg);
464 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst");
465 CHECK_ONE(firm_opt.check_all, irg);
468 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
470 if (firm_opt.deconv) {
471 timer_push(TV_DECONV);
474 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "deconv");
475 CHECK_ONE(firm_opt.check_all, irg);
478 if (firm_opt.cond_eval) {
479 timer_push(TV_COND_EVAL);
482 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
483 CHECK_ONE(firm_opt.check_all, irg);
487 compute_postdoms(irg);
488 DUMP_ONE_CFG_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dom");
489 CHECK_ONE(firm_opt.check_all, irg);
491 construct_backedges(irg);
493 timer_push(TV_CF_OPT);
496 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
497 CHECK_ONE(firm_opt.check_all, irg);
499 if (firm_opt.if_conversion) {
500 timer_push(TV_IF_CONV);
501 opt_if_conv(current_ir_graph, if_conv_info);
503 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "if");
504 CHECK_ONE(firm_opt.check_all, current_ir_graph);
506 timer_push(TV_LOCAL_OPT);
507 optimize_graph_df(current_ir_graph);
509 timer_push(TV_CF_OPT);
510 optimize_cf(current_ir_graph);
512 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "after_if");
513 CHECK_ONE(firm_opt.check_all, current_ir_graph);
516 if (firm_opt.bool_opt) {
518 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "bool");
519 CHECK_ONE(firm_opt.check_all, irg);
523 opt_osr(current_ir_graph, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
525 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "stred");
526 CHECK_ONE(firm_opt.check_all, irg);
528 timer_push(TV_LOCAL_OPT);
529 optimize_graph_df(irg);
531 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
532 CHECK_ONE(firm_opt.check_all, irg);
534 edges_deactivate(irg);
535 timer_push(TV_DEAD_NODE);
536 dead_node_elimination(irg);
538 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dead");
539 CHECK_ONE(firm_opt.check_all, irg);
542 if (firm_opt.cloning) {
543 proc_cloning((float)firm_opt.clone_threshold);
544 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "clone");
545 CHECK_ALL(firm_opt.check_all);
548 if (firm_dump.ir_graph) {
549 /* recompute backedges for nicer dumps */
550 for (i = 0; i < get_irp_n_irgs(); i++)
551 construct_cf_backedges(get_irp_irg(i));
554 if (firm_opt.remove_unused) {
555 ir_entity **keep_methods;
558 /* Analysis that finds the free methods,
559 i.e. methods that are dereferenced.
560 Optimizes polymorphic calls :-). */
561 cgana(&arr_len, &keep_methods);
563 /* Remove methods that are never called. */
564 gc_irgs(arr_len, keep_methods);
570 DUMP_ALL(firm_dump.ir_graph, "-opt");
572 /* verify optimized graphs */
573 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
574 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
576 if (firm_dump.statistic & STAT_AFTER_OPT)
577 stat_dump_snapshot(input_filename, "opt");
579 timer_stop(TV_ALL_OPT);
580 } /* do_firm_optimizations */
583 * compute the size of a type (do implicit lowering)
585 * @param ty a Firm type
587 static int compute_type_size(ir_type *ty)
589 optimization_state_t state;
590 unsigned align_all = 1;
591 int n, size = 0, set = 0;
594 if (get_type_state(ty) == layout_fixed) {
595 /* do not layout already layouted types again */
599 if (is_Method_type(ty) || ty == get_glob_type()) {
600 /* no need for size calculation for method types or the global type */
604 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
606 switch (get_type_tpop_code(ty)) {
609 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
610 ir_entity *ent = get_compound_member(ty, i);
611 ir_type *ent_ty = get_entity_type(ent);
612 unsigned align, misalign;
614 /* compute member types */
615 if (! compute_type_size(ent_ty))
618 align = get_type_alignment_bytes(ent_ty);
619 align_all = align > align_all ? align : align_all;
620 misalign = (align ? size % align : 0);
621 size += (misalign ? align - misalign : 0);
623 set_entity_offset(ent, size);
624 size += get_type_size_bytes(ent_ty);
626 DBG((" member %s %s -> (size: %u, align: %u)\n",
627 get_type_name(ent_ty), get_entity_name(ent),
628 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
630 if (align_all > 0 && size % align_all) {
631 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
632 size += align_all - (size % align_all);
633 DBG(("correcting type-size to %d\n", size));
635 set_type_alignment_bytes(ty, align_all);
640 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
641 ir_entity *ent = get_union_member(ty, i);
643 if (! compute_type_size(get_entity_type(ent)))
645 s = get_type_size_bytes(get_entity_type(ent));
647 set_entity_offset(ent, 0);
648 size = (s > size ? s : size);
654 dims = get_array_n_dimensions(ty);
656 if (! compute_type_size(get_array_element_type(ty)))
661 save_optimization_state(&state);
663 set_opt_constant_folding(1);
665 for (i = 0; i < dims; ++i) {
666 ir_node *lower = get_array_lower_bound(ty, i);
667 ir_node *upper = get_array_upper_bound(ty, i);
668 ir_graph *rem = current_ir_graph;
669 tarval *tv_lower, *tv_upper;
670 long val_lower, val_upper;
672 current_ir_graph = get_const_code_irg();
673 local_optimize_node(lower);
674 local_optimize_node(upper);
675 current_ir_graph = rem;
677 tv_lower = computed_value(lower);
678 tv_upper = computed_value(upper);
680 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
682 * we cannot calculate the size of this array yet, it
683 * even might be unknown until the end, like argv[]
685 restore_optimization_state(&state);
689 val_upper = get_tarval_long(tv_upper);
690 val_lower = get_tarval_long(tv_lower);
691 size *= val_upper - val_lower;
693 restore_optimization_state(&state);
695 DBG(("array %s -> (elements: %d, element type size: %d)\n",
697 size, get_type_size_bytes(get_array_element_type(ty))));
698 size *= get_type_size_bytes(get_array_element_type(ty));
707 set_type_size_bytes(ty, size);
708 set_type_state(ty, layout_fixed);
711 DBG(("size: %d\n", get_type_size_bytes(ty)));
714 } /* compute_type_size */
717 * layout all non-frame types of the Firm graph
719 static void compute_type_sizes(void)
724 /* all non-frame other types */
725 for (i = get_irp_n_types() - 1; i >= 0; --i) {
726 tp = get_irp_type(i);
727 compute_type_size(tp);
729 if (is_Method_type(tp)) {
730 tp = get_method_value_res_type(tp);
733 /* we have a value result type for this method, lower */
734 compute_type_size(tp);
738 } /* compute_type_sizes */
741 * layout all frame-types of the Firm graph
743 static void compute_frame_type_sizes(void)
748 /* all frame types */
749 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
750 irg = get_irp_irg(i);
751 /* do not optimize away variables in debug mode */
752 if (firm_opt.debug_mode == DBG_MODE_NONE)
754 compute_type_size(get_irg_frame_type(irg));
756 } /* compute_frame_type_sizes */
761 * @param input_filename the name of the (main) source file
763 static void do_firm_lowering(const char *input_filename)
767 /* do class lowering and vtbl creation */
768 // lower_classes_to_struct("vtbl", "m");
771 timer_push(TV_LOWER);
776 if (firm_opt.lower_ll) {
777 lwrdw_param_t init = {
782 def_create_intrinsic_fkt,
787 if (arch_create_intrinsic) {
788 init.create_intrinsic = arch_create_intrinsic;
789 init.ctx = create_intrinsic_ctx;
791 timer_push(TV_DW_LOWER);
793 DUMP_ALL(firm_dump.ir_graph, "-dw");
797 if (firm_dump.statistic & STAT_AFTER_LOWER)
798 stat_dump_snapshot(input_filename, "low");
800 /* verify lowered graphs */
801 timer_push(TV_VERIFY);
802 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
803 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
806 DUMP_ALL(firm_dump.ir_graph, "-low");
808 if (firm_opt.enabled) {
809 timer_start(TV_ALL_OPT);
811 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
813 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
814 current_ir_graph = get_irp_irg(i);
816 timer_push(TV_REASSOCIATION);
817 optimize_reassociation(current_ir_graph);
819 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
820 CHECK_ONE(firm_opt.check_all, current_ir_graph);
823 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
824 current_ir_graph = get_irp_irg(i);
826 if (firm_opt.code_place)
827 set_opt_global_cse(1);
829 timer_push(TV_LOCAL_OPT);
830 optimize_graph_df(current_ir_graph);
832 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
833 if (! firm_opt.code_place)
834 CHECK_ONE(firm_opt.check_all, current_ir_graph);
836 if (firm_opt.code_place) {
837 timer_push(TV_CODE_PLACE);
838 place_code(current_ir_graph);
839 set_opt_global_cse(0);
841 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
842 CHECK_ONE(firm_opt.check_all, current_ir_graph);
845 // set_opt_global_cse(0);
846 timer_push(TV_LOAD_STORE);
847 optimize_load_store(current_ir_graph);
849 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
850 CHECK_ONE(firm_opt.check_all, current_ir_graph);
853 timer_push(TV_LOCAL_OPT);
854 optimize_graph_df(current_ir_graph);
856 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
858 timer_push(TV_CF_OPT);
859 optimize_cf(current_ir_graph);
861 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
862 CHECK_ONE(firm_opt.check_all, current_ir_graph);
864 if (firm_opt.if_conversion) {
865 timer_push(TV_IF_CONV);
866 opt_if_conv(current_ir_graph, if_conv_info);
868 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
869 CHECK_ONE(firm_opt.check_all, current_ir_graph);
871 timer_push(TV_LOCAL_OPT);
872 optimize_graph_df(current_ir_graph);
874 timer_push(TV_CF_OPT);
875 optimize_cf(current_ir_graph);
877 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
878 CHECK_ONE(firm_opt.check_all, current_ir_graph);
881 timer_stop(TV_ALL_OPT);
883 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
887 mark_private_methods();
889 /* set the phase to low */
890 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
891 set_irg_phase_low(get_irp_irg(i));
893 /* all graphs are lowered, set the irp phase to low */
894 set_irp_phase_state(phase_low);
896 if (firm_dump.statistic & STAT_FINAL) {
897 stat_dump_snapshot(input_filename, "final");
899 } /* do_firm_lowering */
902 * Initialize for the Firm-generating back end.
904 void gen_firm_init(void)
906 firm_parameter_t params;
908 unsigned pattern = 0;
910 /* the automatic state is only set if inlining is enabled */
911 firm_opt.auto_inline = firm_opt.do_inline;
913 if (firm_dump.stat_pattern)
914 pattern |= FIRMSTAT_PATTERN_ENABLED;
916 if (firm_dump.stat_dag)
917 pattern |= FIRMSTAT_COUNT_DAG;
919 memset(¶ms, 0, sizeof(params));
920 params.size = sizeof(params);
921 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
922 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
923 params.initialize_local_func = uninitialized_local_var;
924 params.cc_mask = 0; /* no regparam, cdecl */
925 params.builtin_dbg = NULL;
928 if (firm_be_opt.selection == BE_FIRM_BE) {
929 const backend_params *be_params = be_init();
931 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
932 params.arch_op_settings = be_params->arch_op_settings;
934 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
935 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
937 ad_param = be_params->dep_param;
938 if_conv_info = be_params->if_conv_info;
940 #endif /* FIRM_BACKEND */
943 /* Activate Graph rewriting if SIMD optimization is turned on */
944 /* This has to be done before init_firm() is called! */
945 if (firm_ext_grs.simd_opt)
950 dbg_init(NULL, NULL, dbg_snprint);
951 edges_init_dbg(firm_opt.vrfy_edges);
952 //cbackend_set_debug_retrieve(dbg_retrieve);
954 set_opt_precise_exc_context(firm_opt.precise_exc);
955 set_opt_fragile_ops(firm_opt.fragile_ops);
957 /* dynamic dispatch works currently only if whole world scenarios */
958 set_opt_dyn_meth_dispatch(0);
960 arch_dep_init(arch_factory);
962 /* do not run architecture dependent optimizations in building phase */
963 arch_dep_set_opts(arch_dep_none);
965 do_node_verification((firm_verification_t) firm_opt.vrfy);
966 if (firm_dump.filter)
967 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
969 if (firm_opt.enabled) {
971 set_opt_constant_folding(firm_opt.const_folding);
972 set_opt_cse(firm_opt.cse);
973 set_opt_global_cse (0);
974 set_opt_unreachable_code(1);
975 set_opt_control_flow(firm_opt.control_flow);
976 set_opt_control_flow_weak_simplification(1);
977 set_opt_control_flow_strong_simplification(1);
982 dump_filter = getenv("FIRM_DUMP_FILTER");
984 only_dump_method_with_name(new_id_from_str(dump_filter));
986 /* do not dump entity ld names */
990 /* init the ycomp debugger extension */
991 if (firm_opt.ycomp_dbg)
992 firm_init_ycomp_debugger(firm_opt.ycomp_host, firm_opt.ycomp_port);
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 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 firm_const_exists)
1009 /* the general for dumping option must be set, or the others will not work */
1011 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1013 dump_keepalive_edges(1);
1014 dump_consts_local(1);
1015 dump_dominator_information(1);
1016 dump_loop_information(0);
1018 if (!firm_dump.edge_labels)
1019 turn_off_edge_labels();
1021 if (firm_dump.all_types) {
1024 dump_class_hierarchy(0, "");
1025 dump_class_hierarchy(1, "-with-entities");
1029 /* finalize all graphs */
1030 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1031 ir_graph *irg = get_irp_irg(i);
1033 irg_finalize_cons(irg);
1034 DUMP_ONE(firm_dump.ir_graph, irg, "");
1036 /* verify the graph */
1037 timer_push(TV_VERIFY);
1038 irg_verify(irg, VRFY_ENFORCE_SSA);
1042 timer_push(TV_VERIFY);
1046 /* all graphs are finalized, set the irp phase to high */
1047 set_irp_phase_state(phase_high);
1049 /* lower all compound call return values */
1050 lower_compound_params();
1052 /* computes the sizes of all types that are still not computed */
1053 compute_type_sizes();
1055 /* lower copyb nodes */
1056 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1057 ir_graph *irg = get_irp_irg(i);
1058 lower_CopyB(irg, 128, 4);
1061 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1062 stat_dump_snapshot(input_filename, "noopt");
1065 if (firm_opt.enabled)
1066 do_firm_optimizations(input_filename, firm_const_exists);
1068 if (firm_dump.gen_firm_asm) {
1069 timer_push(TV_FIRM_ASM);
1070 gen_Firm_assembler(input_filename);
1076 do_firm_lowering(input_filename);
1078 /* computes the sizes of all frame types */
1079 compute_frame_type_sizes();
1081 /* set the phase to low */
1082 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1083 set_irg_phase_low(get_irp_irg(i));
1087 /** SIMD Optimization Extensions **/
1089 /* Pattern creation step. No code has to be generated, so
1090 exit after pattern creation */
1091 if (firm_ext_grs.create_pattern) {
1092 ext_grs_create_pattern();
1096 /* SIMD optimization step. Uses graph patterns to find
1097 rich instructions and rewrite */
1098 if (firm_ext_grs.simd_opt)
1102 /* enable architecture dependent optimizations */
1103 arch_dep_set_opts((arch_dep_opts_t)
1104 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
1105 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
1106 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
1109 if (firm_dump.statistic & STAT_FINAL_IR)
1110 stat_dump_snapshot(input_filename, "final-ir");
1112 /* run the code generator */
1113 if (firm_be_opt.selection != BE_NONE)
1114 do_codegen(out, input_filename);
1116 if (firm_dump.statistic & STAT_FINAL)
1117 stat_dump_snapshot(input_filename, "final");
1120 if (firm_opt.ycomp_dbg)
1121 firm_finish_ycomp_debugger();
1123 } /* gen_firm_finish */
1126 * Do very early initializations
1128 void firm_early_init(void) {
1130 /* arg: need this here for command line options */
1133 firm_init_options(NULL, 0, NULL);
1134 } /* firm_early_init */