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_alloca], i_mapper_alloca },
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;
235 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
238 static int *irg_dump_no;
240 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
243 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
244 DUMP_ONE(1, irg, name);
247 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
250 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
251 DUMP_ONE_CFG(1, irg, name);
254 static void dump_all_count(const char *const suffix)
256 const int n_irgs = get_irp_n_irgs();
259 for (i = 0; i < n_irgs; ++i)
260 dump_graph_count(get_irp_irg(i), suffix);
263 #define DUMP_ONE_C(cond, irg, suffix) \
266 dump_graph_count((irg), (suffix)); \
270 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
273 dump_graph_cfg_count((irg), (suffix)); \
277 #define DUMP_ALL_C(cond, suffix) \
280 dump_all_count((suffix)); \
285 * run all the Firm optimizations
287 * @param input_filename the name of the (main) source file
288 * @param firm_const_exists non-zero, if the const attribute was used on functions
290 static void do_firm_optimizations(const char *input_filename, int firm_const_exists)
296 /* FIXME: cloning might ADD new graphs. */
297 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
299 set_opt_strength_red(firm_opt.strength_red);
300 set_opt_scalar_replacement(firm_opt.scalar_replace);
301 set_opt_auto_create_sync(firm_opt.auto_sync);
302 set_opt_alias_analysis(firm_opt.alias_analysis);
304 aa_opt = aa_opt_no_opt;
305 if (firm_opt.strict_alias)
306 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
307 if (firm_opt.no_alias)
308 aa_opt = aa_opt_no_alias;
310 set_irp_memory_disambiguator_options(aa_opt);
312 timer_start(TV_ALL_OPT);
314 if (firm_opt.remove_unused) {
315 ir_entity **keep_methods;
318 /* Analysis that finds the free methods,
319 i.e. methods that are dereferenced.
320 Optimizes polymorphic calls :-). */
321 cgana(&arr_len, &keep_methods);
323 /* Remove methods that are never called. */
324 gc_irgs(arr_len, keep_methods);
329 if (! firm_opt.freestanding) {
331 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "rts");
332 CHECK_ALL(firm_opt.check_all);
335 /* first step: kill dead code */
336 for (i = 0; i < get_irp_n_irgs(); i++) {
337 irg = current_ir_graph = get_irp_irg(i);
338 /* Confirm construction currently can only handle blocks with only one control
339 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
340 the optimization of switch constructs. */
341 timer_push(TV_CF_OPT);
342 optimize_graph_df(irg);
347 if (firm_opt.func_calls) {
348 timer_push(TV_REAL_FUNC_CALL);
349 optimize_funccalls(firm_const_exists, NULL);
351 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "func_call");
352 CHECK_ALL(firm_opt.check_all);
355 /* do lowering on the const code irg */
358 for (i = 0; i < get_irp_n_irgs(); i++) {
359 irg = current_ir_graph = get_irp_irg(i);
362 /* If SIMD optimization is on, make sure we have only 1 return */
363 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
364 normalize_one_return(irg);
369 if (firm_opt.modes) {
370 /* convert all modes into integer if possible */
371 arch_mode_conversion(irg, predefs.mode_uint);
374 timer_push(TV_SCALAR_REPLACE);
375 scalar_replacement_opt(irg);
377 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "scalar");
378 CHECK_ONE(firm_opt.check_all, irg);
380 timer_push(TV_LOCAL_OPT);
381 optimize_graph_df(irg);
383 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
384 CHECK_ONE(firm_opt.check_all, irg);
386 timer_push(TV_REASSOCIATION);
387 optimize_reassociation(irg);
389 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "reassoc");
390 CHECK_ONE(firm_opt.check_all, irg);
393 CHECK_ONE(firm_opt.check_all, irg);
395 if (firm_opt.code_place) {
396 timer_push(TV_CODE_PLACE);
397 set_opt_global_cse(1);
398 optimize_graph_df(irg);
400 set_opt_global_cse(0);
402 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "place");
403 CHECK_ONE(firm_opt.check_all, irg);
406 if (firm_opt.confirm) {
407 /* Confirm construction currently can only handle blocks with only one control
408 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
409 the optimization of switch constructs. */
410 timer_push(TV_CF_OPT);
413 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
414 CHECK_ONE(firm_opt.check_all, irg);
415 timer_push(TV_CONFIRM_CREATE);
416 construct_confirms(irg);
418 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "confirms");
419 CHECK_ONE(firm_opt.check_all, irg);
421 timer_push(TV_LOCAL_OPT);
422 optimize_graph_df(irg);
424 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
425 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);
434 timer_push(TV_CF_OPT);
437 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
438 CHECK_ONE(firm_opt.check_all, irg);
440 irg_verify(irg, VRFY_ENFORCE_SSA);
441 if (firm_opt.gvn_pre) {
443 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "pre");
444 CHECK_ONE(firm_opt.check_all, irg);
445 irg_verify(irg, VRFY_ENFORCE_SSA);
448 if (firm_opt.load_store) {
449 timer_push(TV_LOAD_STORE);
450 optimize_load_store(irg);
452 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "ldst");
453 CHECK_ONE(firm_opt.check_all, irg);
456 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
458 if (firm_opt.deconv) {
459 timer_push(TV_DECONV);
462 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "deconv");
463 CHECK_ONE(firm_opt.check_all, irg);
466 if (firm_opt.cond_eval) {
467 timer_push(TV_COND_EVAL);
470 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
471 CHECK_ONE(firm_opt.check_all, irg);
474 /* should we really remove the Confirm here? */
475 if (firm_opt.confirm) {
476 timer_push(TV_CONFIRM_CREATE);
477 remove_confirms(irg);
482 compute_postdoms(irg);
483 DUMP_ONE_CFG_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dom");
484 CHECK_ONE(firm_opt.check_all, irg);
486 construct_backedges(irg);
488 timer_push(TV_CF_OPT);
491 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
492 CHECK_ONE(firm_opt.check_all, irg);
494 if (firm_opt.if_conversion) {
495 timer_push(TV_IF_CONV);
496 opt_if_conv(current_ir_graph, if_conv_info);
498 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "if");
499 CHECK_ONE(firm_opt.check_all, current_ir_graph);
501 timer_push(TV_LOCAL_OPT);
502 optimize_graph_df(current_ir_graph);
504 timer_push(TV_CF_OPT);
505 optimize_cf(current_ir_graph);
507 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "after_if");
508 CHECK_ONE(firm_opt.check_all, current_ir_graph);
511 if (firm_opt.bool_opt) {
513 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "bool");
514 CHECK_ONE(firm_opt.check_all, irg);
518 opt_osr(current_ir_graph, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
520 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "stred");
521 CHECK_ONE(firm_opt.check_all, irg);
523 timer_push(TV_LOCAL_OPT);
524 optimize_graph_df(irg);
526 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "lopt");
527 CHECK_ONE(firm_opt.check_all, irg);
529 edges_deactivate(irg);
530 timer_push(TV_DEAD_NODE);
531 dead_node_elimination(irg);
533 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "dead");
534 CHECK_ONE(firm_opt.check_all, irg);
537 if (firm_opt.do_inline) {
538 inline_functions(500, 50);
539 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
540 CHECK_ALL(firm_opt.check_all);
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);
547 if (firm_opt.tail_rec) {
548 timer_push(TV_TAIL_REC);
549 opt_tail_recursion();
552 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
553 CHECK_ALL(firm_opt.check_all);
556 if (firm_opt.cond_eval) {
557 for (i = 0; i < get_irp_n_irgs(); i++) {
558 irg = get_irp_irg(i);
559 timer_push(TV_COND_EVAL);
562 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cond_eval");
563 CHECK_ONE(firm_opt.check_all, irg);
565 timer_push(TV_CF_OPT);
568 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "cfopt");
569 CHECK_ONE(firm_opt.check_all, irg);
573 if (firm_dump.ir_graph) {
574 /* recompute backedges for nicer dumps */
575 for (i = 0; i < get_irp_n_irgs(); i++)
576 construct_cf_backedges(get_irp_irg(i));
579 if (firm_opt.remove_unused) {
580 ir_entity **keep_methods;
583 /* Analysis that finds the free methods,
584 i.e. methods that are dereferenced.
585 Optimizes polymorphic calls :-). */
586 cgana(&arr_len, &keep_methods);
588 /* Remove methods that are never called. */
589 gc_irgs(arr_len, keep_methods);
595 DUMP_ALL(firm_dump.ir_graph, "-opt");
597 /* verify optimized graphs */
598 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
599 ir_graph *irg = get_irp_irg(i);
600 irg_verify(irg, VRFY_ENFORCE_SSA);
603 if (firm_dump.statistic & STAT_AFTER_OPT)
604 stat_dump_snapshot(input_filename, "opt");
606 timer_stop(TV_ALL_OPT);
607 } /* do_firm_optimizations */
610 * compute the size of a type (do implicit lowering)
612 * @param ty a Firm type
614 static int compute_type_size(ir_type *ty)
616 optimization_state_t state;
617 unsigned align_all = 1;
618 int n, size = 0, set = 0;
621 if (get_type_state(ty) == layout_fixed) {
622 /* do not layout already layouted types again */
626 if (is_Method_type(ty) || ty == get_glob_type()) {
627 /* no need for size calculation for method types or the global type */
631 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
633 switch (get_type_tpop_code(ty)) {
636 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
637 ir_entity *ent = get_compound_member(ty, i);
638 ir_type *ent_ty = get_entity_type(ent);
639 unsigned align, misalign;
641 /* compute member types */
642 if (! compute_type_size(ent_ty))
645 align = get_type_alignment_bytes(ent_ty);
646 align_all = align > align_all ? align : align_all;
647 misalign = (align ? size % align : 0);
648 size += (misalign ? align - misalign : 0);
650 set_entity_offset(ent, size);
651 size += get_type_size_bytes(ent_ty);
653 DBG((" member %s %s -> (size: %u, align: %u)\n",
654 get_type_name(ent_ty), get_entity_name(ent),
655 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
657 if (align_all > 0 && size % align_all) {
658 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
659 size += align_all - (size % align_all);
660 DBG(("correcting type-size to %d\n", size));
662 set_type_alignment_bytes(ty, align_all);
667 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
668 ir_entity *ent = get_union_member(ty, i);
670 if (! compute_type_size(get_entity_type(ent)))
672 s = get_type_size_bytes(get_entity_type(ent));
674 set_entity_offset(ent, 0);
675 size = (s > size ? s : size);
681 dims = get_array_n_dimensions(ty);
683 if (! compute_type_size(get_array_element_type(ty)))
688 save_optimization_state(&state);
690 set_opt_constant_folding(1);
692 for (i = 0; i < dims; ++i) {
693 ir_node *lower = get_array_lower_bound(ty, i);
694 ir_node *upper = get_array_upper_bound(ty, i);
695 ir_graph *rem = current_ir_graph;
696 tarval *tv_lower, *tv_upper;
697 long val_lower, val_upper;
699 current_ir_graph = get_const_code_irg();
700 local_optimize_node(lower);
701 local_optimize_node(upper);
702 current_ir_graph = rem;
704 tv_lower = computed_value(lower);
705 tv_upper = computed_value(upper);
707 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
709 * we cannot calculate the size of this array yet, it
710 * even might be unknown until the end, like argv[]
712 restore_optimization_state(&state);
716 val_upper = get_tarval_long(tv_upper);
717 val_lower = get_tarval_long(tv_lower);
718 size *= val_upper - val_lower;
720 restore_optimization_state(&state);
722 DBG(("array %s -> (elements: %d, element type size: %d)\n",
724 size, get_type_size_bytes(get_array_element_type(ty))));
725 size *= get_type_size_bytes(get_array_element_type(ty));
734 set_type_size_bytes(ty, size);
735 set_type_state(ty, layout_fixed);
738 DBG(("size: %d\n", get_type_size_bytes(ty)));
741 } /* compute_type_size */
744 * layout all non-frame types of the Firm graph
746 static void compute_type_sizes(void)
751 /* all non-frame other types */
752 for (i = get_irp_n_types() - 1; i >= 0; --i) {
753 tp = get_irp_type(i);
754 compute_type_size(tp);
756 if (is_Method_type(tp)) {
757 tp = get_method_value_res_type(tp);
760 /* we have a value result type for this method, lower */
761 compute_type_size(tp);
765 } /* compute_type_sizes */
768 * layout all frame-types of the Firm graph
770 static void compute_frame_type_sizes(void)
775 /* all frame types */
776 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
777 irg = get_irp_irg(i);
778 /* do not optimize away variables in debug mode */
779 if (firm_opt.debug_mode == DBG_MODE_NONE)
781 compute_type_size(get_irg_frame_type(irg));
783 } /* compute_frame_type_sizes */
788 * @param input_filename the name of the (main) source file
790 static void do_firm_lowering(const char *input_filename)
794 /* do class lowering and vtbl creation */
795 // lower_classes_to_struct("vtbl", "m");
798 timer_push(TV_LOWER);
803 if (firm_opt.lower_ll) {
804 lwrdw_param_t init = {
809 def_create_intrinsic_fkt,
814 if (arch_create_intrinsic) {
815 init.create_intrinsic = arch_create_intrinsic;
816 init.ctx = create_intrinsic_ctx;
818 timer_push(TV_DW_LOWER);
823 if (firm_dump.statistic & STAT_AFTER_LOWER)
824 stat_dump_snapshot(input_filename, "low");
826 /* verify lowered graphs */
827 timer_push(TV_VERIFY);
828 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
829 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
832 DUMP_ALL(firm_dump.ir_graph, "-low");
834 if (firm_opt.enabled) {
835 timer_start(TV_ALL_OPT);
837 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
839 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
840 current_ir_graph = get_irp_irg(i);
842 timer_push(TV_REASSOCIATION);
843 optimize_reassociation(current_ir_graph);
845 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
846 CHECK_ONE(firm_opt.check_all, current_ir_graph);
849 /* enable architecture dependent optimizations */
850 arch_dep_set_opts((arch_dep_opts_t)
851 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
852 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
853 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
855 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
856 current_ir_graph = get_irp_irg(i);
858 if (firm_opt.code_place)
859 set_opt_global_cse(1);
861 timer_push(TV_LOCAL_OPT);
862 optimize_graph_df(current_ir_graph);
864 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
865 if (! firm_opt.code_place)
866 CHECK_ONE(firm_opt.check_all, current_ir_graph);
868 if (firm_opt.code_place) {
869 timer_push(TV_CODE_PLACE);
870 place_code(current_ir_graph);
871 set_opt_global_cse(0);
873 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
874 CHECK_ONE(firm_opt.check_all, current_ir_graph);
877 timer_push(TV_LOAD_STORE);
878 optimize_load_store(current_ir_graph);
880 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
881 CHECK_ONE(firm_opt.check_all, current_ir_graph);
884 timer_push(TV_LOCAL_OPT);
885 optimize_graph_df(current_ir_graph);
887 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
889 timer_push(TV_CF_OPT);
890 optimize_cf(current_ir_graph);
892 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
893 CHECK_ONE(firm_opt.check_all, current_ir_graph);
895 if (firm_opt.if_conversion) {
896 timer_push(TV_IF_CONV);
897 opt_if_conv(current_ir_graph, if_conv_info);
899 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
900 CHECK_ONE(firm_opt.check_all, current_ir_graph);
902 timer_push(TV_LOCAL_OPT);
903 optimize_graph_df(current_ir_graph);
905 timer_push(TV_CF_OPT);
906 optimize_cf(current_ir_graph);
908 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
909 CHECK_ONE(firm_opt.check_all, current_ir_graph);
912 timer_stop(TV_ALL_OPT);
914 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
918 mark_private_methods();
920 /* set the phase to low */
921 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
922 set_irg_phase_low(get_irp_irg(i));
924 /* all graphs are lowered, set the irp phase to low */
925 set_irp_phase_state(phase_low);
927 if (firm_dump.statistic & STAT_FINAL) {
928 stat_dump_snapshot(input_filename, "final");
930 } /* do_firm_lowering */
933 * Initialize for the Firm-generating back end.
935 void gen_firm_init(void)
937 firm_parameter_t params;
939 unsigned pattern = 0;
941 /* the automatic state is only set if inlining is enabled */
942 firm_opt.auto_inline = firm_opt.do_inline;
944 if (firm_dump.stat_pattern)
945 pattern |= FIRMSTAT_PATTERN_ENABLED;
947 if (firm_dump.stat_dag)
948 pattern |= FIRMSTAT_COUNT_DAG;
950 memset(¶ms, 0, sizeof(params));
951 params.size = sizeof(params);
952 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
953 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
954 params.initialize_local_func = uninitialized_local_var;
955 params.cc_mask = 0; /* no regparam, cdecl */
956 params.builtin_dbg = NULL;
959 /* Activate Graph rewriting if SIMD optimization is turned on */
960 /* This has to be done before init_firm() is called! */
961 if (firm_ext_grs.simd_opt)
967 if (firm_be_opt.selection == BE_FIRM_BE) {
968 const backend_params *be_params = be_get_backend_param();
970 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
971 params.arch_op_settings = be_params->arch_op_settings;
973 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
974 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
976 ad_param = be_params->dep_param;
977 if_conv_info = be_params->if_conv_info;
979 if (be_params->has_imm_fp_mode)
980 firm_imm_fp_mode = be_params->imm_fp_mode;
983 dbg_init(NULL, NULL, dbg_snprint);
984 edges_init_dbg(firm_opt.vrfy_edges);
985 //cbackend_set_debug_retrieve(dbg_retrieve);
987 set_opt_precise_exc_context(firm_opt.precise_exc);
988 set_opt_fragile_ops(firm_opt.fragile_ops);
990 /* Sel node cannot produce NULL pointers */
991 set_opt_sel_based_null_check_elim(1);
993 /* dynamic dispatch works currently only if whole world scenarios */
994 set_opt_dyn_meth_dispatch(0);
996 arch_dep_init(arch_factory);
998 /* do not run architecture dependent optimizations in building phase */
999 arch_dep_set_opts(arch_dep_none);
1001 do_node_verification((firm_verification_t) firm_opt.vrfy);
1002 if (firm_dump.filter)
1003 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
1005 if (firm_opt.enabled) {
1007 set_opt_constant_folding(firm_opt.const_folding);
1008 set_opt_cse(firm_opt.cse);
1009 set_opt_global_cse (0);
1010 set_opt_unreachable_code(1);
1011 set_opt_control_flow(firm_opt.control_flow);
1012 set_opt_control_flow_weak_simplification(1);
1013 set_opt_control_flow_strong_simplification(1);
1018 dump_filter = getenv("FIRM_DUMP_FILTER");
1020 only_dump_method_with_name(new_id_from_str(dump_filter));
1022 /* do not dump entity ld names */
1026 /* init the ycomp debugger extension */
1027 if (firm_opt.ycomp_dbg)
1028 firm_init_ycomp_debugger(firm_opt.ycomp_host, firm_opt.ycomp_port);
1030 } /* gen_firm_init */
1033 * Called, after the Firm generation is completed,
1034 * do all optimizations and backend call here.
1036 * @param out a file handle for the output, may be NULL
1037 * @param input_filename the name of the (main) source file
1038 * @param c_mode non-zero if "C" was compiled
1039 * @param firm_const_exists non-zero, if the const attribute was used on functions
1041 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int firm_const_exists)
1045 /* the general for dumping option must be set, or the others will not work */
1047 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1049 dump_keepalive_edges(1);
1050 dump_consts_local(1);
1051 dump_dominator_information(1);
1052 dump_loop_information(0);
1054 if (!firm_dump.edge_labels)
1055 turn_off_edge_labels();
1057 if (firm_dump.all_types) {
1060 dump_class_hierarchy(0, "");
1061 dump_class_hierarchy(1, "-with-entities");
1065 /* finalize all graphs */
1066 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1067 ir_graph *irg = get_irp_irg(i);
1069 irg_finalize_cons(irg);
1070 DUMP_ONE(firm_dump.ir_graph, irg, "");
1072 /* verify the graph */
1073 timer_push(TV_VERIFY);
1074 irg_verify(irg, VRFY_ENFORCE_SSA);
1078 timer_push(TV_VERIFY);
1082 /* all graphs are finalized, set the irp phase to high */
1083 set_irp_phase_state(phase_high);
1085 /* lower all compound call return values */
1086 lower_compound_params();
1088 /* computes the sizes of all types that are still not computed */
1089 compute_type_sizes();
1091 /* lower copyb nodes */
1092 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1093 ir_graph *irg = get_irp_irg(i);
1094 lower_CopyB(irg, 128, 4);
1097 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1098 stat_dump_snapshot(input_filename, "noopt");
1101 if (firm_opt.enabled)
1102 do_firm_optimizations(input_filename, firm_const_exists);
1104 if (firm_dump.gen_firm_asm) {
1105 timer_push(TV_FIRM_ASM);
1106 gen_Firm_assembler(input_filename);
1112 do_firm_lowering(input_filename);
1114 /* computes the sizes of all frame types */
1115 compute_frame_type_sizes();
1117 /* set the phase to low */
1118 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1119 set_irg_phase_low(get_irp_irg(i));
1123 /** SIMD Optimization Extensions **/
1125 /* Pattern creation step. No code has to be generated, so
1126 exit after pattern creation */
1127 if (firm_ext_grs.create_pattern) {
1128 ext_grs_create_pattern();
1132 /* SIMD optimization step. Uses graph patterns to find
1133 rich instructions and rewrite */
1134 if (firm_ext_grs.simd_opt)
1138 if (firm_dump.statistic & STAT_FINAL_IR)
1139 stat_dump_snapshot(input_filename, "final-ir");
1141 /* run the code generator */
1142 if (firm_be_opt.selection != BE_NONE)
1143 do_codegen(out, input_filename);
1145 if (firm_dump.statistic & STAT_FINAL)
1146 stat_dump_snapshot(input_filename, "final");
1149 if (firm_opt.ycomp_dbg)
1150 firm_finish_ycomp_debugger();
1152 } /* gen_firm_finish */
1155 * Do very early initializations
1157 void firm_early_init(void) {
1158 /* arg: need this here for command line options */
1160 firm_init_options(NULL, 0, NULL);
1161 } /* firm_early_init */