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"
23 #if defined(_DEBUG) || defined(FIRM_DEBUG)
24 #define DBG(x) dbg_printf x
27 #endif /* _DEBUG || FIRM_DEBUG */
30 /** dump all the graphs depending on cond */
31 #define DUMP_ALL(cond, suffix) \
34 timer_push(TV_VCG_DUMP); \
35 if (firm_dump.no_blocks) \
36 dump_all_ir_graphs(dump_ir_graph, suffix); \
37 else if (firm_dump.extbb) \
38 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
40 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
45 /** dump all control flow graphs depending on cond */
46 #define DUMP_ALL_CFG(cond, suffix) \
49 timer_push(TV_VCG_DUMP); \
50 dump_all_ir_graphs(dump_cfg, suffix); \
55 /** check all graphs depending on cond */
56 #define CHECK_ALL(cond) \
60 timer_push(TV_VERIFY); \
61 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
62 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
69 /** dump graphs irg depending on cond */
70 #define DUMP_ONE(cond, irg, suffix) \
73 timer_push(TV_VCG_DUMP); \
74 if (firm_dump.no_blocks) \
75 dump_ir_graph(irg, suffix); \
76 else if (firm_dump.extbb) \
77 dump_ir_extblock_graph(irg, suffix); \
79 dump_ir_block_graph(irg, suffix); \
84 /** dump control flow graph irg depending on cond */
85 #define DUMP_ONE_CFG(cond, irg, suffix) \
88 timer_push(TV_VCG_DUMP); \
89 dump_cfg(irg, suffix); \
94 /** check a graph irg depending on cond */
95 #define CHECK_ONE(cond, irg) \
98 timer_push(TV_VERIFY); \
99 irg_verify(irg, VRFY_ENFORCE_SSA); \
105 /* set by the backend parameters */
106 static const ir_settings_arch_dep_t *ad_param = NULL;
107 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
108 static void *create_intrinsic_ctx = NULL;
109 static const ir_settings_if_conv_t *if_conv_info = NULL;
111 /* entities of runtime functions */
112 ir_entity_ptr rts_entities[rts_max];
115 * factory for setting architecture dependent parameters
117 static const ir_settings_arch_dep_t *arch_factory(void)
119 static const ir_settings_arch_dep_t param = {
120 1, /* also use subs */
121 4, /* maximum shifts */
122 31, /* maximum shift amount */
123 NULL, /* use default evaluator */
127 32 /* Mulh allowed up to 32 bit */
130 return ad_param ? ad_param : ¶m;
134 * Map runtime functions.
136 static void rts_map(void) {
137 static const struct {
138 ir_entity_ptr *ent; /**< address of the rts entity */
139 i_mapper_func func; /**< mapper function. */
142 { &rts_entities[rts_abs], i_mapper_abs },
143 { &rts_entities[rts_labs], i_mapper_abs },
144 { &rts_entities[rts_llabs], i_mapper_abs },
145 { &rts_entities[rts_imaxabs], i_mapper_abs },
147 /* double -> double */
148 { &rts_entities[rts_fabs], i_mapper_abs },
149 { &rts_entities[rts_sqrt], i_mapper_sqrt },
150 { &rts_entities[rts_cbrt], i_mapper_cbrt },
151 { &rts_entities[rts_pow], i_mapper_pow },
152 { &rts_entities[rts_exp], i_mapper_exp },
153 { &rts_entities[rts_exp2], i_mapper_exp },
154 { &rts_entities[rts_exp10], i_mapper_exp },
155 { &rts_entities[rts_log], i_mapper_log },
156 { &rts_entities[rts_log2], i_mapper_log2 },
157 { &rts_entities[rts_log10], i_mapper_log10 },
158 { &rts_entities[rts_sin], i_mapper_sin },
159 { &rts_entities[rts_cos], i_mapper_cos },
160 { &rts_entities[rts_tan], i_mapper_tan },
161 { &rts_entities[rts_asin], i_mapper_asin },
162 { &rts_entities[rts_acos], i_mapper_acos },
163 { &rts_entities[rts_atan], i_mapper_atan },
164 { &rts_entities[rts_sinh], i_mapper_sinh },
165 { &rts_entities[rts_cosh], i_mapper_cosh },
166 { &rts_entities[rts_tanh], i_mapper_tanh },
169 { &rts_entities[rts_fabsf], i_mapper_abs },
170 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
171 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
172 { &rts_entities[rts_powf], i_mapper_pow },
173 { &rts_entities[rts_expf], i_mapper_exp },
174 { &rts_entities[rts_exp2f], i_mapper_exp },
175 { &rts_entities[rts_exp10f], i_mapper_exp },
176 { &rts_entities[rts_logf], i_mapper_log },
177 { &rts_entities[rts_log2f], i_mapper_log2 },
178 { &rts_entities[rts_log10f], i_mapper_log10 },
179 { &rts_entities[rts_sinf], i_mapper_sin },
180 { &rts_entities[rts_cosf], i_mapper_cos },
181 { &rts_entities[rts_tanf], i_mapper_tan },
182 { &rts_entities[rts_asinf], i_mapper_asin },
183 { &rts_entities[rts_acosf], i_mapper_acos },
184 { &rts_entities[rts_atanf], i_mapper_atan },
185 { &rts_entities[rts_sinhf], i_mapper_sinh },
186 { &rts_entities[rts_coshf], i_mapper_cosh },
187 { &rts_entities[rts_tanhf], i_mapper_tanh },
189 /* long double -> long double */
190 { &rts_entities[rts_fabsl], i_mapper_abs },
191 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
192 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
193 { &rts_entities[rts_powl], i_mapper_pow },
194 { &rts_entities[rts_expl], i_mapper_exp },
195 { &rts_entities[rts_exp2l], i_mapper_exp },
196 { &rts_entities[rts_exp10l], i_mapper_exp },
197 { &rts_entities[rts_logl], i_mapper_log },
198 { &rts_entities[rts_log2l], i_mapper_log2 },
199 { &rts_entities[rts_log10l], i_mapper_log10 },
200 { &rts_entities[rts_sinl], i_mapper_sin },
201 { &rts_entities[rts_cosl], i_mapper_cos },
202 { &rts_entities[rts_tanl], i_mapper_tan },
203 { &rts_entities[rts_asinl], i_mapper_asin },
204 { &rts_entities[rts_acosl], i_mapper_acos },
205 { &rts_entities[rts_atanl], i_mapper_atan },
206 { &rts_entities[rts_sinhl], i_mapper_sinh },
207 { &rts_entities[rts_coshl], i_mapper_cosh },
208 { &rts_entities[rts_tanhl], i_mapper_tanh },
211 { &rts_entities[rts_memcpy], i_mapper_memcpy },
212 { &rts_entities[rts_memset], i_mapper_memset },
213 { &rts_entities[rts_strcmp], i_mapper_strcmp },
214 { &rts_entities[rts_strncmp], i_mapper_strncmp },
215 { &rts_entities[rts_strlen], i_mapper_strlen }
217 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
220 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
221 if (*mapper[i].ent != NULL) {
222 rec[n_map].i_call.kind = INTRINSIC_CALL;
223 rec[n_map].i_call.i_ent = *mapper[i].ent;
224 rec[n_map].i_call.i_mapper = mapper[i].func;
225 rec[n_map].i_call.ctx = NULL;
226 rec[n_map].i_call.link = NULL;
230 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
233 static int *irg_dump_no;
235 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
238 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
239 DUMP_ONE(1, irg, name);
242 static void dump_graph_cfg_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_CFG(1, irg, name);
249 static void dump_all_count(const char *const suffix)
251 const int n_irgs = get_irp_n_irgs();
254 for (i = 0; i < n_irgs; ++i)
255 dump_graph_count(get_irp_irg(i), suffix);
258 #define DUMP_ONE_C(cond, irg, suffix) \
261 dump_graph_count((irg), (suffix)); \
265 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
268 dump_graph_cfg_count((irg), (suffix)); \
272 #define DUMP_ALL_C(cond, suffix) \
275 dump_all_count((suffix)); \
280 * run all the Firm optimizations
282 * @param input_filename the name of the (main) source file
283 * @param firm_const_exists non-zero, if the const attribute was used on functions
285 static void do_firm_optimizations(const char *input_filename, int firm_const_exists)
291 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
293 set_opt_strength_red(firm_opt.strength_red);
294 set_opt_scalar_replacement(firm_opt.scalar_replace);
295 set_opt_auto_create_sync(firm_opt.auto_sync);
296 set_opt_alias_analysis(firm_opt.alias_analysis);
298 aa_opt = aa_opt_no_opt;
299 if (firm_opt.strict_alias)
300 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
301 if (firm_opt.no_alias)
302 aa_opt = aa_opt_no_alias;
304 set_irp_memory_disambiguator_options(aa_opt);
306 timer_start(TV_ALL_OPT);
308 if (firm_opt.remove_unused) {
309 ir_entity **keep_methods;
312 /* Analysis that finds the free methods,
313 i.e. methods that are dereferenced.
314 Optimizes polymorphic calls :-). */
315 cgana(&arr_len, &keep_methods);
317 /* Remove methods that are never called. */
318 gc_irgs(arr_len, keep_methods);
323 if (! firm_opt.freestanding) {
325 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "rts");
326 CHECK_ALL(firm_opt.check_all);
329 if (firm_opt.tail_rec) {
330 timer_push(TV_TAIL_REC);
331 opt_tail_recursion();
334 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
335 CHECK_ALL(firm_opt.check_all);
338 if (firm_opt.func_calls) {
339 timer_push(TV_REAL_FUNC_CALL);
340 optimize_funccalls(firm_const_exists);
342 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "func_call");
343 CHECK_ALL(firm_opt.check_all);
346 if (firm_opt.do_inline) {
347 timer_push(TV_INLINE);
348 inline_leave_functions(500, 80, 30, FALSE);
350 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
351 CHECK_ALL(firm_opt.check_all);
354 for (i = 0; i < get_irp_n_irgs(); i++) {
355 irg = current_ir_graph = get_irp_irg(i);
359 /* If SIMD optimization is on, make sure we have only 1 return */
360 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
361 normalize_one_return(irg);
366 if (firm_opt.modes) {
367 /* convert all modes into integer if possible */
368 arch_mode_conversion(irg, predefs.mode_uint);
371 timer_push(TV_SCALAR_REPLACE);
372 scalar_replacement_opt(irg);
374 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "scalar");
375 CHECK_ONE(firm_opt.check_all, irg);
377 timer_push(TV_LOCAL_OPT);
378 optimize_graph_df(irg);
380 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
381 CHECK_ONE(firm_opt.check_all, irg);
383 timer_push(TV_REASSOCIATION);
384 optimize_reassociation(irg);
386 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "reassoc");
387 CHECK_ONE(firm_opt.check_all, irg);
389 if (firm_opt.confirm) {
390 /* Confirm construction currently can only handle blocks with only one control
391 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
392 the optimization of switch constructs. */
393 timer_push(TV_CF_OPT);
396 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
397 CHECK_ONE(firm_opt.check_all, irg);
398 timer_push(TV_CONFIRM_CREATE);
399 construct_confirms(irg);
401 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "confirms");
402 CHECK_ONE(firm_opt.check_all, irg);
405 timer_push(TV_LOCAL_OPT);
406 optimize_graph_df(irg);
408 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
409 CHECK_ONE(firm_opt.check_all, irg);
412 CHECK_ONE(firm_opt.check_all, irg);
414 if (firm_opt.code_place) {
415 timer_push(TV_CODE_PLACE);
416 set_opt_global_cse(1);
417 optimize_graph_df(irg);
419 set_opt_global_cse(0);
421 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "place");
422 CHECK_ONE(firm_opt.check_all, irg);
425 if (firm_opt.luffig) {
427 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst2");
430 timer_push(TV_CF_OPT);
433 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
434 CHECK_ONE(firm_opt.check_all, irg);
436 /* should we really remove the Confirm here? */
437 if (firm_opt.confirm) {
438 timer_push(TV_CONFIRM_CREATE);
439 remove_confirms(irg);
443 irg_verify(irg, VRFY_ENFORCE_SSA);
444 if (firm_opt.gvn_pre) {
446 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "pre");
447 CHECK_ONE(firm_opt.check_all, irg);
448 irg_verify(irg, VRFY_ENFORCE_SSA);
451 if (firm_opt.loop_unrolling) {
452 timer_push(TV_LOOP_UNROLL);
453 optimize_loop_unrolling(irg);
455 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "loop");
456 CHECK_ONE(firm_opt.check_all, irg);
459 if (firm_opt.load_store) {
460 timer_push(TV_LOAD_STORE);
461 optimize_load_store(irg);
463 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst");
464 CHECK_ONE(firm_opt.check_all, irg);
467 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
469 if (firm_opt.deconv) {
470 timer_push(TV_DECONV);
473 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "deconv");
474 CHECK_ONE(firm_opt.check_all, irg);
477 if (firm_opt.cond_eval) {
478 timer_push(TV_COND_EVAL);
481 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
482 CHECK_ONE(firm_opt.check_all, irg);
486 compute_postdoms(irg);
487 DUMP_ONE_CFG_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dom");
488 CHECK_ONE(firm_opt.check_all, irg);
490 construct_backedges(irg);
492 timer_push(TV_CF_OPT);
495 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
496 CHECK_ONE(firm_opt.check_all, irg);
498 if (firm_opt.if_conversion) {
499 timer_push(TV_IF_CONV);
500 opt_if_conv(current_ir_graph, if_conv_info);
502 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "if");
503 CHECK_ONE(firm_opt.check_all, current_ir_graph);
505 timer_push(TV_LOCAL_OPT);
506 optimize_graph_df(current_ir_graph);
508 timer_push(TV_CF_OPT);
509 optimize_cf(current_ir_graph);
511 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "after_if");
512 CHECK_ONE(firm_opt.check_all, current_ir_graph);
515 if (firm_opt.bool_opt) {
517 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "bool");
518 CHECK_ONE(firm_opt.check_all, irg);
522 opt_osr(current_ir_graph, osr_flag_default /*| osr_flag_ignore_x86_shift*/);
524 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "stred");
525 CHECK_ONE(firm_opt.check_all, irg);
527 timer_push(TV_LOCAL_OPT);
528 optimize_graph_df(irg);
530 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
531 CHECK_ONE(firm_opt.check_all, irg);
533 edges_deactivate(irg);
534 timer_push(TV_DEAD_NODE);
535 dead_node_elimination(irg);
537 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dead");
538 CHECK_ONE(firm_opt.check_all, irg);
541 if (firm_opt.cloning) {
542 proc_cloning((float)firm_opt.clone_threshold);
543 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "clone");
546 if (firm_dump.ir_graph) {
547 /* recompute backedges for nicer dumps */
548 for (i = 0; i < get_irp_n_irgs(); i++)
549 construct_cf_backedges(get_irp_irg(i));
552 if (firm_opt.remove_unused) {
553 ir_entity **keep_methods;
556 /* Analysis that finds the free methods,
557 i.e. methods that are dereferenced.
558 Optimizes polymorphic calls :-). */
559 cgana(&arr_len, &keep_methods);
561 /* Remove methods that are never called. */
562 gc_irgs(arr_len, keep_methods);
568 DUMP_ALL(firm_dump.ir_graph, "-opt");
570 /* verify optimized graphs */
571 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
572 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
574 if (firm_dump.statistic & STAT_AFTER_OPT)
575 stat_dump_snapshot(input_filename, "opt");
577 timer_stop(TV_ALL_OPT);
578 } /* do_firm_optimizations */
581 * compute the size of a type (do implicit lowering)
583 * @param ty a Firm type
585 static int compute_type_size(ir_type *ty)
587 optimization_state_t state;
588 unsigned align_all = 1;
589 int n, size = 0, set = 0;
592 if (get_type_state(ty) == layout_fixed) {
593 /* do not layout already layouted types again */
597 if (is_Method_type(ty) || ty == get_glob_type()) {
598 /* no need for size calculation for method types or the global type */
602 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
604 switch (get_type_tpop_code(ty)) {
607 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
608 ir_entity *ent = get_compound_member(ty, i);
609 ir_type *ent_ty = get_entity_type(ent);
610 unsigned align, misalign;
612 /* compute member types */
613 if (! compute_type_size(ent_ty))
616 align = get_type_alignment_bytes(ent_ty);
617 align_all = align > align_all ? align : align_all;
618 misalign = (align ? size % align : 0);
619 size += (misalign ? align - misalign : 0);
621 set_entity_offset(ent, size);
622 size += get_type_size_bytes(ent_ty);
624 DBG((" member %s %s -> (size: %u, align: %u)\n",
625 get_type_name(ent_ty), get_entity_name(ent),
626 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
628 if (align_all > 0 && size % align_all) {
629 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
630 size += align_all - (size % align_all);
631 DBG(("correcting type-size to %d\n", size));
633 set_type_alignment_bytes(ty, align_all);
638 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
639 ir_entity *ent = get_union_member(ty, i);
641 if (! compute_type_size(get_entity_type(ent)))
643 s = get_type_size_bytes(get_entity_type(ent));
645 set_entity_offset(ent, 0);
646 size = (s > size ? s : size);
652 dims = get_array_n_dimensions(ty);
654 if (! compute_type_size(get_array_element_type(ty)))
659 save_optimization_state(&state);
661 set_opt_constant_folding(1);
663 for (i = 0; i < dims; ++i) {
664 ir_node *lower = get_array_lower_bound(ty, i);
665 ir_node *upper = get_array_upper_bound(ty, i);
666 ir_graph *rem = current_ir_graph;
667 tarval *tv_lower, *tv_upper;
668 long val_lower, val_upper;
670 current_ir_graph = get_const_code_irg();
671 local_optimize_node(lower);
672 local_optimize_node(upper);
673 current_ir_graph = rem;
675 tv_lower = computed_value(lower);
676 tv_upper = computed_value(upper);
678 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
680 * we cannot calculate the size of this array yet, it
681 * even might be unknown until the end, like argv[]
683 restore_optimization_state(&state);
687 val_upper = get_tarval_long(tv_upper);
688 val_lower = get_tarval_long(tv_lower);
689 size *= val_upper - val_lower;
691 restore_optimization_state(&state);
693 DBG(("array %s -> (elements: %d, element type size: %d)\n",
695 size, get_type_size_bytes(get_array_element_type(ty))));
696 size *= get_type_size_bytes(get_array_element_type(ty));
705 set_type_size_bytes(ty, size);
706 set_type_state(ty, layout_fixed);
709 DBG(("size: %d\n", get_type_size_bytes(ty)));
712 } /* compute_type_size */
715 * layout all types of the Firm graph
717 static void compute_type_sizes(void)
723 /* all frame types */
724 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
725 irg = get_irp_irg(i);
726 /* do not optimize away variables in debug mode */
727 if (firm_opt.debug_mode == DBG_MODE_NONE)
729 compute_type_size(get_irg_frame_type(irg));
732 /* all other types */
733 for (i = get_irp_n_types() - 1; i >= 0; --i) {
734 tp = get_irp_type(i);
735 compute_type_size(tp);
737 if (is_Method_type(tp)) {
738 tp = get_method_value_res_type(tp);
741 /* we have a value result type for this method, lower */
742 compute_type_size(tp);
746 } /* compute_type_sizes */
751 * @param input_filename the name of the (main) source file
753 static void do_firm_lowering(const char *input_filename)
757 /* do class lowering and vtbl creation */
758 // lower_classes_to_struct("vtbl", "m");
761 timer_push(TV_LOWER);
766 if (firm_opt.lower_ll) {
767 lwrdw_param_t init = {
772 def_create_intrinsic_fkt,
777 if (arch_create_intrinsic) {
778 init.create_intrinsic = arch_create_intrinsic;
779 init.ctx = create_intrinsic_ctx;
781 timer_push(TV_DW_LOWER);
783 DUMP_ALL(firm_dump.ir_graph, "-dw");
787 if (firm_dump.statistic & STAT_AFTER_LOWER)
788 stat_dump_snapshot(input_filename, "low");
790 /* verify lowered graphs */
791 timer_push(TV_VERIFY);
792 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
793 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
796 DUMP_ALL(firm_dump.ir_graph, "-low");
798 if (firm_opt.enabled) {
799 timer_start(TV_ALL_OPT);
801 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
803 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
804 current_ir_graph = get_irp_irg(i);
806 timer_push(TV_REASSOCIATION);
807 optimize_reassociation(current_ir_graph);
809 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
810 CHECK_ONE(firm_opt.check_all, current_ir_graph);
813 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
814 current_ir_graph = get_irp_irg(i);
816 if (firm_opt.code_place)
817 set_opt_global_cse(1);
819 timer_push(TV_LOCAL_OPT);
820 optimize_graph_df(current_ir_graph);
822 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
823 if (! firm_opt.code_place)
824 CHECK_ONE(firm_opt.check_all, current_ir_graph);
826 if (firm_opt.code_place) {
827 timer_push(TV_CODE_PLACE);
828 place_code(current_ir_graph);
829 set_opt_global_cse(0);
831 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
832 CHECK_ONE(firm_opt.check_all, current_ir_graph);
835 // set_opt_global_cse(0);
836 timer_push(TV_LOAD_STORE);
837 optimize_load_store(current_ir_graph);
839 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
840 CHECK_ONE(firm_opt.check_all, current_ir_graph);
843 timer_push(TV_LOCAL_OPT);
844 optimize_graph_df(current_ir_graph);
846 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
848 timer_push(TV_CF_OPT);
849 optimize_cf(current_ir_graph);
851 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
852 CHECK_ONE(firm_opt.check_all, current_ir_graph);
854 if (firm_opt.if_conversion) {
855 timer_push(TV_IF_CONV);
856 opt_if_conv(current_ir_graph, if_conv_info);
858 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
859 CHECK_ONE(firm_opt.check_all, current_ir_graph);
861 timer_push(TV_LOCAL_OPT);
862 optimize_graph_df(current_ir_graph);
864 timer_push(TV_CF_OPT);
865 optimize_cf(current_ir_graph);
867 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
868 CHECK_ONE(firm_opt.check_all, current_ir_graph);
871 timer_stop(TV_ALL_OPT);
873 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
877 mark_private_methods();
879 /* set the phase to low */
880 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
881 set_irg_phase_low(get_irp_irg(i));
883 /* all graphs are lowered, set the irp phase to low */
884 set_irp_phase_state(phase_low);
886 if (firm_dump.statistic & STAT_FINAL) {
887 stat_dump_snapshot(input_filename, "final");
889 } /* do_firm_lowering */
892 * Initialize for the Firm-generating back end.
894 void gen_firm_init(void)
896 firm_parameter_t params;
898 unsigned pattern = 0;
900 /* the automatic state is only set if inlining is enabled */
901 firm_opt.auto_inline = firm_opt.do_inline;
903 if (firm_dump.stat_pattern)
904 pattern |= FIRMSTAT_PATTERN_ENABLED;
906 if (firm_dump.stat_dag)
907 pattern |= FIRMSTAT_COUNT_DAG;
909 memset(¶ms, 0, sizeof(params));
910 params.size = sizeof(params);
911 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
912 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
913 params.initialize_local_func = uninitialized_local_var;
914 params.cc_mask = 0; /* no regparam, cdecl */
915 params.builtin_dbg = NULL;
918 if (firm_be_opt.selection == BE_FIRM_BE) {
919 const backend_params *be_params = be_init();
921 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
922 params.arch_op_settings = be_params->arch_op_settings;
924 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
925 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
927 ad_param = be_params->dep_param;
928 if_conv_info = be_params->if_conv_info;
930 #endif /* FIRM_BACKEND */
933 /* Activate Graph rewriting if SIMD optimization is turned on */
934 /* This has to be done before init_firm() is called! */
935 if (firm_ext_grs.simd_opt)
940 dbg_init(NULL, NULL, dbg_snprint);
941 edges_init_dbg(firm_opt.vrfy_edges);
942 //cbackend_set_debug_retrieve(dbg_retrieve);
944 set_opt_precise_exc_context(firm_opt.precise_exc);
945 set_opt_fragile_ops(firm_opt.fragile_ops);
947 /* dynamic dispatch works currently only if whole world scenarios */
948 set_opt_dyn_meth_dispatch(0);
950 arch_dep_init(arch_factory);
952 /* do not run architecture dependent optimizations in building phase */
953 arch_dep_set_opts(arch_dep_none);
955 do_node_verification((firm_verification_t) firm_opt.vrfy);
956 if (firm_dump.filter)
957 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
959 if (firm_opt.enabled) {
961 set_opt_constant_folding(firm_opt.const_folding);
962 set_opt_cse(firm_opt.cse);
963 set_opt_global_cse (0);
964 set_opt_unreachable_code(1);
965 set_opt_control_flow(firm_opt.control_flow);
966 set_opt_control_flow_weak_simplification(1);
967 set_opt_control_flow_strong_simplification(1);
972 dump_filter = getenv("FIRM_DUMP_FILTER");
974 only_dump_method_with_name(new_id_from_str(dump_filter));
976 /* do not dump entity ld names */
980 /* init the ycomp debugger extension */
981 if (firm_opt.ycomp_dbg)
982 firm_init_ycomp_debugger(firm_opt.ycomp_host, firm_opt.ycomp_port);
984 } /* gen_firm_init */
987 * Called, after the Firm generation is completed,
988 * do all optimizations and backend call here.
990 * @param out a file handle for the output, may be NULL
991 * @param input_filename the name of the (main) source file
992 * @param c_mode non-zero if "C" was compiled
993 * @param firm_const_exists non-zero, if the const attribute was used on functions
995 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int firm_const_exists)
999 /* the general for dumping option must be set, or the others will not work */
1001 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1003 dump_keepalive_edges(1);
1004 dump_consts_local(1);
1005 dump_dominator_information(1);
1006 dump_loop_information(0);
1008 if (!firm_dump.edge_labels)
1009 turn_off_edge_labels();
1011 if (firm_dump.all_types) {
1014 dump_class_hierarchy(0, "");
1015 dump_class_hierarchy(1, "-with-entities");
1019 /* finalize all graphs */
1020 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1021 ir_graph *irg = get_irp_irg(i);
1023 irg_finalize_cons(irg);
1024 DUMP_ONE(firm_dump.ir_graph, irg, "");
1026 /* verify the graph */
1027 timer_push(TV_VERIFY);
1028 irg_verify(irg, VRFY_ENFORCE_SSA);
1032 timer_push(TV_VERIFY);
1036 /* all graphs are finalized, set the irp phase to high */
1037 set_irp_phase_state(phase_high);
1039 /* lower all compound call return values */
1040 lower_compound_params();
1042 /* computes the sizes of all types that are still not computed */
1043 compute_type_sizes();
1045 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1046 stat_dump_snapshot(input_filename, "noopt");
1049 if (firm_opt.enabled)
1050 do_firm_optimizations(input_filename, firm_const_exists);
1052 if (firm_dump.gen_firm_asm) {
1053 timer_push(TV_FIRM_ASM);
1054 gen_Firm_assembler(input_filename);
1060 do_firm_lowering(input_filename);
1062 /* set the phase to low */
1063 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1064 set_irg_phase_low(get_irp_irg(i));
1068 /** SIMD Optimization Extensions **/
1070 /* Pattern creation step. No code has to be generated, so
1071 exit after pattern creation */
1072 if (firm_ext_grs.create_pattern) {
1073 ext_grs_create_pattern();
1077 /* SIMD optimization step. Uses graph patterns to find
1078 rich instructions and rewrite */
1079 if (firm_ext_grs.simd_opt)
1083 /* enable architecture dependent optimizations */
1084 arch_dep_set_opts((arch_dep_opts_t)
1085 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
1086 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
1087 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
1090 if (firm_dump.statistic & STAT_FINAL_IR)
1091 stat_dump_snapshot(input_filename, "final-ir");
1093 /* run the code generator */
1094 if (firm_be_opt.selection != BE_NONE)
1095 do_codegen(out, input_filename);
1097 if (firm_dump.statistic & STAT_FINAL)
1098 stat_dump_snapshot(input_filename, "final");
1101 if (firm_opt.ycomp_dbg)
1102 firm_finish_ycomp_debugger();
1104 } /* gen_firm_finish */
1107 * Do very early initializations
1109 void firm_early_init(void) {
1111 /* arg: need this here for command line options */
1114 firm_init_options(NULL, 0, NULL);
1115 } /* firm_early_init */