3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
14 #include <libfirm/firm.h>
15 #include <libfirm/be.h>
19 #include "firm_codegen.h"
20 #include "firm_cmdline.h"
21 #include "firm_timing.h"
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 ir_mode *firm_imm_fp_mode = NULL;
117 /* entities of runtime functions */
118 ir_entity_ptr rts_entities[rts_max];
121 * factory for setting architecture dependent parameters
123 static const ir_settings_arch_dep_t *arch_factory(void)
125 static const ir_settings_arch_dep_t param = {
126 1, /* also use subs */
127 4, /* maximum shifts */
128 31, /* maximum shift amount */
129 NULL, /* use default evaluator */
133 32 /* Mulh allowed up to 32 bit */
136 return ad_param ? ad_param : ¶m;
140 * Map runtime functions.
142 static void rts_map(void) {
143 static const struct {
144 ir_entity_ptr *ent; /**< address of the rts entity */
145 i_mapper_func func; /**< mapper function. */
148 { &rts_entities[rts_abs], i_mapper_abs },
149 { &rts_entities[rts_labs], i_mapper_abs },
150 { &rts_entities[rts_llabs], i_mapper_abs },
151 { &rts_entities[rts_imaxabs], i_mapper_abs },
153 /* double -> double */
154 { &rts_entities[rts_fabs], i_mapper_abs },
155 { &rts_entities[rts_sqrt], i_mapper_sqrt },
156 { &rts_entities[rts_cbrt], i_mapper_cbrt },
157 { &rts_entities[rts_pow], i_mapper_pow },
158 { &rts_entities[rts_exp], i_mapper_exp },
159 { &rts_entities[rts_exp2], i_mapper_exp },
160 { &rts_entities[rts_exp10], i_mapper_exp },
161 { &rts_entities[rts_log], i_mapper_log },
162 { &rts_entities[rts_log2], i_mapper_log2 },
163 { &rts_entities[rts_log10], i_mapper_log10 },
164 { &rts_entities[rts_sin], i_mapper_sin },
165 { &rts_entities[rts_cos], i_mapper_cos },
166 { &rts_entities[rts_tan], i_mapper_tan },
167 { &rts_entities[rts_asin], i_mapper_asin },
168 { &rts_entities[rts_acos], i_mapper_acos },
169 { &rts_entities[rts_atan], i_mapper_atan },
170 { &rts_entities[rts_sinh], i_mapper_sinh },
171 { &rts_entities[rts_cosh], i_mapper_cosh },
172 { &rts_entities[rts_tanh], i_mapper_tanh },
175 { &rts_entities[rts_fabsf], i_mapper_abs },
176 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
177 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
178 { &rts_entities[rts_powf], i_mapper_pow },
179 { &rts_entities[rts_expf], i_mapper_exp },
180 { &rts_entities[rts_exp2f], i_mapper_exp },
181 { &rts_entities[rts_exp10f], i_mapper_exp },
182 { &rts_entities[rts_logf], i_mapper_log },
183 { &rts_entities[rts_log2f], i_mapper_log2 },
184 { &rts_entities[rts_log10f], i_mapper_log10 },
185 { &rts_entities[rts_sinf], i_mapper_sin },
186 { &rts_entities[rts_cosf], i_mapper_cos },
187 { &rts_entities[rts_tanf], i_mapper_tan },
188 { &rts_entities[rts_asinf], i_mapper_asin },
189 { &rts_entities[rts_acosf], i_mapper_acos },
190 { &rts_entities[rts_atanf], i_mapper_atan },
191 { &rts_entities[rts_sinhf], i_mapper_sinh },
192 { &rts_entities[rts_coshf], i_mapper_cosh },
193 { &rts_entities[rts_tanhf], i_mapper_tanh },
195 /* long double -> long double */
196 { &rts_entities[rts_fabsl], i_mapper_abs },
197 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
198 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
199 { &rts_entities[rts_powl], i_mapper_pow },
200 { &rts_entities[rts_expl], i_mapper_exp },
201 { &rts_entities[rts_exp2l], i_mapper_exp },
202 { &rts_entities[rts_exp10l], i_mapper_exp },
203 { &rts_entities[rts_logl], i_mapper_log },
204 { &rts_entities[rts_log2l], i_mapper_log2 },
205 { &rts_entities[rts_log10l], i_mapper_log10 },
206 { &rts_entities[rts_sinl], i_mapper_sin },
207 { &rts_entities[rts_cosl], i_mapper_cos },
208 { &rts_entities[rts_tanl], i_mapper_tan },
209 { &rts_entities[rts_asinl], i_mapper_asin },
210 { &rts_entities[rts_acosl], i_mapper_acos },
211 { &rts_entities[rts_atanl], i_mapper_atan },
212 { &rts_entities[rts_sinhl], i_mapper_sinh },
213 { &rts_entities[rts_coshl], i_mapper_cosh },
214 { &rts_entities[rts_tanhl], i_mapper_tanh },
217 { &rts_entities[rts_memcpy], i_mapper_memcpy },
218 { &rts_entities[rts_memset], i_mapper_memset },
219 { &rts_entities[rts_strcmp], i_mapper_strcmp },
220 { &rts_entities[rts_strncmp], i_mapper_strncmp },
221 { &rts_entities[rts_strlen], i_mapper_strlen }
223 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
226 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
227 if (*mapper[i].ent != NULL) {
228 rec[n_map].i_call.kind = INTRINSIC_CALL;
229 rec[n_map].i_call.i_ent = *mapper[i].ent;
230 rec[n_map].i_call.i_mapper = mapper[i].func;
231 rec[n_map].i_call.ctx = NULL;
232 rec[n_map].i_call.link = NULL;
236 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
239 static int *irg_dump_no;
241 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
244 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
245 DUMP_ONE(1, irg, name);
249 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
252 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
253 DUMP_ONE_CFG(1, irg, name);
257 static void dump_all_count(const char *const suffix)
259 const int n_irgs = get_irp_n_irgs();
262 for (i = 0; i < n_irgs; ++i)
263 dump_graph_count(get_irp_irg(i), suffix);
266 #define DUMP_ONE_C(cond, irg, suffix) \
269 dump_graph_count((irg), (suffix)); \
273 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
276 dump_graph_cfg_count((irg), (suffix)); \
280 #define DUMP_ALL_C(cond, suffix) \
283 dump_all_count((suffix)); \
287 static void remove_unused_functions(void)
289 ir_entity **keep_methods;
292 /* Analysis that finds the free methods,
293 i.e. methods that are dereferenced.
294 Optimizes polymorphic calls :-). */
295 cgana(&arr_len, &keep_methods);
297 /* Remove methods that are never called. */
298 gc_irgs(arr_len, keep_methods);
302 static int firm_const_exists;
304 static void do_optimize_funccalls(void)
306 optimize_funccalls(firm_const_exists, NULL);
309 static void do_gcse(ir_graph *irg)
311 set_opt_global_cse(1);
312 optimize_graph_df(irg);
314 set_opt_global_cse(0);
317 static void do_lower_highlevel(ir_graph *irg)
319 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
322 static void do_if_conv(ir_graph *irg)
324 opt_if_conv(irg, if_conv_info);
327 static void do_stred(ir_graph *irg)
329 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
332 static void do_inline(void)
334 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold);
337 static void do_cloning(void)
339 proc_cloning((float) firm_opt.clone_threshold);
342 static void do_lower_switch(ir_graph *irg)
344 lower_switch(irg, firm_opt.spare_size);
347 typedef enum opt_target {
352 typedef void (*transform_irg_func)(ir_graph *irg);
353 typedef void (*transform_irp_func)(void);
354 typedef void (*func_ptr_t)(void);
366 static opt_config_t opts[] = {
367 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, true, true, true, -1 },
368 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, true, true, true, TV_COMBO },
369 { OPT_TARGET_IRG, "controlflow", (func_ptr_t) optimize_cf, true, true, true, TV_CF_OPT },
370 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, true, true, true, TV_LOCAL_OPT },
371 { OPT_TARGET_IRP, "gc_irgs", (func_ptr_t) remove_unused_functions, true, false, false, TV_CGANA },
372 { OPT_TARGET_IRP, "tailrec", (func_ptr_t) opt_tail_recursion, true, true, true, TV_TAIL_REC },
373 { OPT_TARGET_IRP, "funccalls", (func_ptr_t) do_optimize_funccalls, true, true, true, TV_REAL_FUNC_CALL },
374 { OPT_TARGET_IRP, "lowerconst", (func_ptr_t) lower_const_code, true, false, false, -1 },
375 { OPT_TARGET_IRG, "onereturn", (func_ptr_t) normalize_one_return, true, false, false, -1 },
376 { OPT_TARGET_IRG, "scalar", (func_ptr_t) scalar_replacement_opt, true, true, true, TV_SCALAR_REPLACE },
377 { OPT_TARGET_IRG, "reassoc", (func_ptr_t) optimize_reassociation, true, true, true, TV_REASSOCIATION },
378 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, true, true, true, TV_CODE_PLACE },
379 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, true, true, true, TV_CODE_PLACE },
380 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, true, true, true, TV_CONFIRM_CREATE },
381 { OPT_TARGET_IRG, "ldst", (func_ptr_t) optimize_load_store, true, true, true, TV_LOAD_STORE },
382 { OPT_TARGET_IRG, "sync", (func_ptr_t) opt_sync, true, true, true, -1 },
383 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, true, true, true, -1 },
384 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, true, true, true, TV_DECONV },
385 { OPT_TARGET_IRG, "condeval", (func_ptr_t) opt_cond_eval, true, true, true, TV_COND_EVAL },
386 { OPT_TARGET_IRG, "remove_confirms", (func_ptr_t) remove_confirms, true, false, false, TV_CONFIRM_CREATE },
387 { OPT_TARGET_IRG, "gvnpre", (func_ptr_t) do_gvn_pre, true, true, true, -1 },
388 { OPT_TARGET_IRG, "ifconv", (func_ptr_t) do_if_conv, true, true, true, TV_IF_CONV },
389 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, true, true, true, -1 },
390 { OPT_TARGET_IRG, "shape-blocks",(func_ptr_t) shape_blocks, true, true, true, TV_END_MELT },
391 { OPT_TARGET_IRG, "stred", (func_ptr_t) do_stred, true, true, true, TV_OSR },
392 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, true, false, true, TV_DEAD_NODE },
393 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, true, true, true, -1 },
394 { OPT_TARGET_IRP, "clone", (func_ptr_t) do_cloning, true, true, true, -1 },
395 { OPT_TARGET_IRG, "lower_switch", (func_ptr_t) do_lower_switch, true, true, true, -1 },
397 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
399 static opt_config_t *get_opt(const char *name)
402 for (i = 0; i < n_opts; ++i) {
403 opt_config_t *config = &opts[i];
404 if (strcmp(config->name, name) == 0)
412 static void set_opt_enabled(const char *name, bool enabled)
414 opt_config_t *config = get_opt(name);
415 config->enabled = enabled;
418 static void do_irg_opt(ir_graph *irg, const char *name)
420 transform_irg_func func;
422 opt_config_t *config = get_opt(name);
423 assert(config->target == OPT_TARGET_IRG);
424 if (!config->enabled)
427 if (config->timer != -1)
428 timer_push(config->timer);
430 old_irg = current_ir_graph;
431 current_ir_graph = irg;
433 func = (transform_irg_func) config->func;
436 if (config->timer != -1)
439 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
440 CHECK_ONE(firm_opt.check_all, irg);
442 current_ir_graph = old_irg;
445 static void do_irp_opt(const char *name)
447 transform_irp_func func;
448 opt_config_t *config = get_opt(name);
449 assert(config->target == OPT_TARGET_IRP);
450 if (!config->enabled)
453 if (config->timer != -1)
454 timer_push(config->timer);
456 func = (transform_irp_func) config->func;
459 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
460 CHECK_ALL(firm_opt.check_all);
462 if (config->timer != -1)
467 * run all the Firm optimizations
469 * @param input_filename the name of the (main) source file
471 static void do_firm_optimizations(const char *input_filename)
476 /* FIXME: cloning might ADD new graphs. */
477 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
479 set_opt_strength_red(firm_opt.strength_red);
480 set_opt_scalar_replacement(firm_opt.scalar_replace);
481 set_opt_auto_create_sync(firm_opt.auto_sync);
482 set_opt_alias_analysis(firm_opt.alias_analysis);
483 set_opt_combo(firm_opt.combo);
485 aa_opt = aa_opt_no_opt;
486 if (firm_opt.strict_alias)
487 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
488 if (firm_opt.no_alias)
489 aa_opt = aa_opt_no_alias;
491 set_irp_memory_disambiguator_options(aa_opt);
493 /* parameter passing code should set them directly sometime... */
494 set_opt_enabled("rts", firm_opt.freestanding);
495 set_opt_enabled("gc_irgs", firm_opt.remove_unused);
496 set_opt_enabled("tailrec", firm_opt.tail_rec);
497 set_opt_enabled("funccalls", firm_opt.func_calls);
498 set_opt_enabled("gcse", firm_opt.gcse);
499 set_opt_enabled("place", !firm_opt.gcse);
500 set_opt_enabled("confirm", firm_opt.confirm);
501 set_opt_enabled("remove_confirms", firm_opt.confirm);
502 set_opt_enabled("sync", firm_opt.sync);
503 set_opt_enabled("ldst", firm_opt.load_store);
504 set_opt_enabled("deconv", firm_opt.deconv);
505 set_opt_enabled("condeval", firm_opt.cond_eval);
506 set_opt_enabled("gvnpre", firm_opt.gvn_pre);
507 set_opt_enabled("ifconv", firm_opt.if_conversion);
508 set_opt_enabled("bool", firm_opt.bool_opt);
509 set_opt_enabled("shape-blocks", firm_opt.shape_blocks);
510 set_opt_enabled("inline", firm_opt.do_inline);
511 set_opt_enabled("clone", firm_opt.cloning);
512 set_opt_enabled("combo", firm_opt.combo);
514 timer_start(TV_ALL_OPT);
518 /* first step: kill dead code */
519 if (firm_opt.combo) {
520 for (i = 0; i < get_irp_n_irgs(); i++) {
521 ir_graph *irg = get_irp_irg(i);
522 do_irg_opt(irg, "combo");
525 for (i = 0; i < get_irp_n_irgs(); i++) {
526 ir_graph *irg = get_irp_irg(i);
527 do_irg_opt(irg, "local");
528 do_irg_opt(irg, "controlflow");
532 do_irp_opt("gc_irgs");
533 do_irp_opt("tailrec");
534 do_irp_opt("funccalls");
535 do_irp_opt("lowerconst");
537 for (i = 0; i < get_irp_n_irgs(); i++) {
538 ir_graph *irg = get_irp_irg(i);
541 /* If SIMD optimization is on, make sure we have only 1 return */
542 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
543 do_irg_opt("onereturn");
546 do_irg_opt(irg, "scalar");
547 do_irg_opt(irg, "local");
548 do_irg_opt(irg, "reassoc");
549 do_irg_opt(irg, "local");
550 do_irg_opt(irg, "gcse");
552 if (firm_opt.confirm) {
553 /* Confirm construction currently can only handle blocks with only one control
554 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
555 the optimization of switch constructs. */
556 do_irg_opt(irg, "controlflow");
557 do_irg_opt(irg, "confirm");
558 do_irg_opt(irg, "local");
561 do_irg_opt(irg, "controlflow");
562 do_irg_opt(irg, "ldst");
563 do_irg_opt(irg, "lower");
564 do_irg_opt(irg, "deconv");
565 do_irg_opt(irg, "condeval");
566 do_irg_opt(irg, "remove_confirms");
567 do_irg_opt(irg, "gvnpre");
568 do_irg_opt(irg, "place");
569 do_irg_opt(irg, "controlflow");
571 if (firm_opt.if_conversion) {
572 do_irg_opt(irg, "ifconv");
573 do_irg_opt(irg, "local");
574 do_irg_opt(irg, "controlflow");
577 do_irg_opt(irg, "bool");
578 do_irg_opt(irg, "shape-blocks");
579 do_irg_opt(irg, "lower_switch");
580 do_irg_opt(irg, "stred");
581 do_irg_opt(irg, "local");
582 do_irg_opt(irg, "dead");
585 do_irp_opt("inline");
588 for (i = 0; i < get_irp_n_irgs(); i++) {
589 ir_graph *irg = get_irp_irg(i);
590 do_irg_opt(irg, "local");
591 do_irg_opt(irg, "controlflow");
592 do_irg_opt(irg, "condeval");
593 do_irg_opt(irg, "local");
594 do_irg_opt(irg, "controlflow");
597 if (firm_dump.ir_graph) {
598 /* recompute backedges for nicer dumps */
599 for (i = 0; i < get_irp_n_irgs(); i++)
600 construct_cf_backedges(get_irp_irg(i));
603 do_irp_opt("gc_irgs");
605 DUMP_ALL(firm_dump.ir_graph, "-opt");
606 /* verify optimized graphs */
607 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
608 ir_graph *irg = get_irp_irg(i);
609 irg_verify(irg, VRFY_ENFORCE_SSA);
612 if (firm_dump.statistic & STAT_AFTER_OPT)
613 stat_dump_snapshot(input_filename, "opt");
615 timer_stop(TV_ALL_OPT);
616 } /* do_firm_optimizations */
619 * compute the size of a type (do implicit lowering)
621 * @param ty a Firm type
623 static int compute_type_size(ir_type *ty)
625 optimization_state_t state;
626 unsigned align_all = 1;
627 int n, size = 0, set = 0;
630 if (get_type_state(ty) == layout_fixed) {
631 /* do not layout already layouted types again */
635 if (is_Method_type(ty) || ty == get_glob_type()) {
636 /* no need for size calculation for method types or the global type */
640 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
642 switch (get_type_tpop_code(ty)) {
645 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
646 ir_entity *ent = get_compound_member(ty, i);
647 ir_type *ent_ty = get_entity_type(ent);
648 unsigned align, misalign;
650 /* compute member types */
651 if (! compute_type_size(ent_ty))
654 align = get_type_alignment_bytes(ent_ty);
655 align_all = align > align_all ? align : align_all;
656 misalign = (align ? size % align : 0);
657 size += (misalign ? align - misalign : 0);
659 set_entity_offset(ent, size);
660 size += get_type_size_bytes(ent_ty);
662 DBG((" member %s %s -> (size: %u, align: %u)\n",
663 get_type_name(ent_ty), get_entity_name(ent),
664 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
666 if (align_all > 0 && size % align_all) {
667 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
668 size += align_all - (size % align_all);
669 DBG(("correcting type-size to %d\n", size));
671 set_type_alignment_bytes(ty, align_all);
676 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
677 ir_entity *ent = get_union_member(ty, i);
679 if (! compute_type_size(get_entity_type(ent)))
681 s = get_type_size_bytes(get_entity_type(ent));
683 set_entity_offset(ent, 0);
684 size = (s > size ? s : size);
690 dims = get_array_n_dimensions(ty);
692 if (! compute_type_size(get_array_element_type(ty)))
697 save_optimization_state(&state);
699 set_opt_constant_folding(1);
700 set_opt_algebraic_simplification(1);
702 for (i = 0; i < dims; ++i) {
703 ir_node *lower = get_array_lower_bound(ty, i);
704 ir_node *upper = get_array_upper_bound(ty, i);
705 ir_graph *rem = current_ir_graph;
706 tarval *tv_lower, *tv_upper;
707 long val_lower, val_upper;
709 current_ir_graph = get_const_code_irg();
710 local_optimize_node(lower);
711 local_optimize_node(upper);
712 current_ir_graph = rem;
714 tv_lower = computed_value(lower);
715 tv_upper = computed_value(upper);
717 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
719 * we cannot calculate the size of this array yet, it
720 * even might be unknown until the end, like argv[]
722 restore_optimization_state(&state);
726 val_upper = get_tarval_long(tv_upper);
727 val_lower = get_tarval_long(tv_lower);
728 size *= val_upper - val_lower;
730 restore_optimization_state(&state);
732 DBG(("array %s -> (elements: %d, element type size: %d)\n",
734 size, get_type_size_bytes(get_array_element_type(ty))));
735 size *= get_type_size_bytes(get_array_element_type(ty));
744 set_type_size_bytes(ty, size);
745 set_type_state(ty, layout_fixed);
748 DBG(("size: %d\n", get_type_size_bytes(ty)));
751 } /* compute_type_size */
754 * layout all non-frame types of the Firm graph
756 static void compute_type_sizes(void)
761 /* all non-frame other types */
762 for (i = get_irp_n_types() - 1; i >= 0; --i) {
763 tp = get_irp_type(i);
764 compute_type_size(tp);
766 if (is_Method_type(tp)) {
767 tp = get_method_value_res_type(tp);
770 /* we have a value result type for this method, lower */
771 compute_type_size(tp);
775 } /* compute_type_sizes */
778 * layout all frame-types of the Firm graph
780 static void compute_frame_type_sizes(void)
785 /* all frame types */
786 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
787 irg = get_irp_irg(i);
788 /* do not optimize away variables in debug mode */
789 if (firm_opt.debug_mode == DBG_MODE_NONE)
791 compute_type_size(get_irg_frame_type(irg));
793 } /* compute_frame_type_sizes */
798 * @param input_filename the name of the (main) source file
800 static void do_firm_lowering(const char *input_filename)
804 /* do class lowering and vtbl creation */
805 // lower_classes_to_struct("vtbl", "m");
808 timer_push(TV_LOWER);
813 if (firm_opt.lower_ll) {
814 lwrdw_param_t init = {
819 def_create_intrinsic_fkt,
824 if (arch_create_intrinsic) {
825 init.create_intrinsic = arch_create_intrinsic;
826 init.ctx = create_intrinsic_ctx;
828 timer_push(TV_DW_LOWER);
833 if (firm_dump.statistic & STAT_AFTER_LOWER)
834 stat_dump_snapshot(input_filename, "low");
836 /* verify lowered graphs */
837 timer_push(TV_VERIFY);
838 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
839 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
842 DUMP_ALL(firm_dump.ir_graph, "-low");
844 if (firm_opt.enabled) {
845 timer_start(TV_ALL_OPT);
847 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
849 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
850 ir_graph *irg = get_irp_irg(i);
851 do_irg_opt(irg, "reassoc");
854 /* enable architecture dependent optimizations */
855 arch_dep_set_opts((arch_dep_opts_t)
856 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
857 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
858 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
860 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
861 ir_graph *irg = get_irp_irg(i);
863 current_ir_graph = irg;
865 do_irg_opt(irg, "local");
866 do_irg_opt(irg, "gcse");
867 do_irg_opt(irg, "ldst");
868 do_irg_opt(irg, "local");
869 do_irg_opt(irg, "controlflow");
871 if (firm_opt.if_conversion) {
872 do_irg_opt(irg, "ifconv");
873 do_irg_opt(irg, "local");
874 do_irg_opt(irg, "controlflow");
877 do_irg_opt(current_ir_graph, "sync");
879 timer_stop(TV_ALL_OPT);
881 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
885 mark_private_methods();
887 /* set the phase to low */
888 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
889 set_irg_phase_low(get_irp_irg(i));
891 /* all graphs are lowered, set the irp phase to low */
892 set_irp_phase_state(phase_low);
894 if (firm_dump.statistic & STAT_FINAL) {
895 stat_dump_snapshot(input_filename, "final");
897 } /* do_firm_lowering */
900 * Initialize for the Firm-generating back end.
902 void gen_firm_init(void)
904 firm_parameter_t params;
905 unsigned pattern = 0;
907 /* the automatic state is only set if inlining is enabled */
908 firm_opt.auto_inline = firm_opt.do_inline;
910 if (firm_dump.stat_pattern)
911 pattern |= FIRMSTAT_PATTERN_ENABLED;
913 if (firm_dump.stat_dag)
914 pattern |= FIRMSTAT_COUNT_DAG;
916 memset(¶ms, 0, sizeof(params));
917 params.size = sizeof(params);
918 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
919 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
920 params.initialize_local_func = uninitialized_local_var;
921 params.cc_mask = 0; /* no regparam, cdecl */
922 params.builtin_dbg = NULL;
925 /* Activate Graph rewriting if SIMD optimization is turned on */
926 /* This has to be done before init_firm() is called! */
927 if (firm_ext_grs.simd_opt)
933 if (firm_be_opt.selection == BE_FIRM_BE) {
934 const backend_params *be_params = be_get_backend_param();
936 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
937 params.arch_op_settings = be_params->arch_op_settings;
939 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
940 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
942 ad_param = be_params->dep_param;
943 if_conv_info = be_params->if_conv_info;
945 if (be_params->has_imm_fp_mode)
946 firm_imm_fp_mode = be_params->imm_fp_mode;
948 /* OS option must be set to the backend */
949 switch (firm_opt.os_support) {
950 case OS_SUPPORT_MINGW:
951 firm_be_option("ia32-gasmode=mingw");
953 case OS_SUPPORT_MACHO:
954 firm_be_option("ia32-gasmode=macho");
956 case OS_SUPPORT_LINUX:
958 firm_be_option("ia32-gasmode=linux");
962 dbg_init(NULL, NULL, dbg_snprint);
963 edges_init_dbg(firm_opt.vrfy_edges);
964 //cbackend_set_debug_retrieve(dbg_retrieve);
966 set_opt_precise_exc_context(firm_opt.precise_exc);
967 set_opt_fragile_ops(firm_opt.fragile_ops);
969 /* Sel node cannot produce NULL pointers */
970 set_opt_sel_based_null_check_elim(1);
972 /* dynamic dispatch works currently only if whole world scenarios */
973 set_opt_dyn_meth_dispatch(0);
975 arch_dep_init(arch_factory);
977 /* do not run architecture dependent optimizations in building phase */
978 arch_dep_set_opts(arch_dep_none);
980 do_node_verification((firm_verification_t) firm_opt.vrfy);
981 if (firm_dump.filter)
982 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
984 if (firm_opt.enabled) {
986 set_opt_constant_folding(firm_opt.const_folding);
987 set_opt_algebraic_simplification(firm_opt.const_folding);
988 set_opt_cse(firm_opt.cse);
989 set_opt_global_cse(0);
990 set_opt_unreachable_code(1);
991 set_opt_control_flow(firm_opt.control_flow);
992 set_opt_control_flow_weak_simplification(1);
993 set_opt_control_flow_strong_simplification(1);
998 /* do not dump entity ld names */
1000 } /* gen_firm_init */
1003 * Called, after the Firm generation is completed,
1004 * do all optimizations and backend call here.
1006 * @param out a file handle for the output, may be NULL
1007 * @param input_filename the name of the (main) source file
1008 * @param c_mode non-zero if "C" was compiled
1009 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
1011 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
1015 firm_const_exists = new_firm_const_exists;
1017 /* the general for dumping option must be set, or the others will not work */
1019 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1021 dump_keepalive_edges(1);
1022 dump_consts_local(1);
1023 dump_dominator_information(1);
1024 dump_loop_information(0);
1026 if (!firm_dump.edge_labels)
1027 turn_off_edge_labels();
1029 if (firm_dump.all_types) {
1032 dump_class_hierarchy(0, "");
1033 dump_class_hierarchy(1, "-with-entities");
1037 /* finalize all graphs */
1038 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1039 ir_graph *irg = get_irp_irg(i);
1041 irg_finalize_cons(irg);
1042 DUMP_ONE(firm_dump.ir_graph, irg, "");
1044 /* verify the graph */
1045 timer_push(TV_VERIFY);
1046 irg_verify(irg, VRFY_ENFORCE_SSA);
1050 timer_push(TV_VERIFY);
1054 /* all graphs are finalized, set the irp phase to high */
1055 set_irp_phase_state(phase_high);
1057 /* BEWARE: kill unreachable code before doing compound lowering */
1058 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1059 ir_graph *irg = get_irp_irg(i);
1063 /* lower all compound call return values */
1064 lower_compound_params();
1066 /* computes the sizes of all types that are still not computed */
1067 compute_type_sizes();
1069 /* lower copyb nodes */
1070 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1071 ir_graph *irg = get_irp_irg(i);
1072 lower_CopyB(irg, 128, 4);
1075 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1076 stat_dump_snapshot(input_filename, "noopt");
1079 if (firm_opt.enabled)
1080 do_firm_optimizations(input_filename);
1082 if (firm_dump.gen_firm_asm) {
1083 timer_push(TV_FIRM_ASM);
1084 gen_Firm_assembler(input_filename);
1090 do_firm_lowering(input_filename);
1092 /* computes the sizes of all frame types */
1093 compute_frame_type_sizes();
1095 /* set the phase to low */
1096 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1097 set_irg_phase_low(get_irp_irg(i));
1101 /** SIMD Optimization Extensions **/
1103 /* Pattern creation step. No code has to be generated, so
1104 exit after pattern creation */
1105 if (firm_ext_grs.create_pattern) {
1106 ext_grs_create_pattern();
1110 /* SIMD optimization step. Uses graph patterns to find
1111 rich instructions and rewrite */
1112 if (firm_ext_grs.simd_opt)
1116 if (firm_dump.statistic & STAT_FINAL_IR)
1117 stat_dump_snapshot(input_filename, "final-ir");
1119 /* run the code generator */
1120 if (firm_be_opt.selection != BE_NONE)
1121 do_codegen(out, input_filename);
1123 if (firm_dump.statistic & STAT_FINAL)
1124 stat_dump_snapshot(input_filename, "final");
1127 if (firm_opt.ycomp_dbg)
1128 firm_finish_ycomp_debugger();
1130 } /* gen_firm_finish */
1133 * Do very early initializations
1135 void firm_early_init(void) {
1136 /* arg: need this here for command line options */
1138 firm_init_options(NULL, 0, NULL);
1139 } /* firm_early_init */