3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
13 #include <libfirm/firm.h>
14 #include <libfirm/be.h>
18 #include "firm_codegen.h"
19 #include "firm_cmdline.h"
20 #include "firm_timing.h"
23 #define snprintf _snprintf
26 #if defined(_DEBUG) || defined(FIRM_DEBUG)
27 #define DBG(x) dbg_printf x
30 #endif /* _DEBUG || FIRM_DEBUG */
33 /** dump all the graphs depending on cond */
34 #define DUMP_ALL(cond, suffix) \
37 timer_push(TV_VCG_DUMP); \
38 if (firm_dump.no_blocks) \
39 dump_all_ir_graphs(dump_ir_graph, suffix); \
40 else if (firm_dump.extbb) \
41 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
43 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
48 /** dump all control flow graphs depending on cond */
49 #define DUMP_ALL_CFG(cond, suffix) \
52 timer_push(TV_VCG_DUMP); \
53 dump_all_ir_graphs(dump_cfg, suffix); \
58 /** check all graphs depending on cond */
59 #define CHECK_ALL(cond) \
63 timer_push(TV_VERIFY); \
64 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
65 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
72 /** dump graphs irg depending on cond */
73 #define DUMP_ONE(cond, irg, suffix) \
76 timer_push(TV_VCG_DUMP); \
77 if (firm_dump.no_blocks) \
78 dump_ir_graph(irg, suffix); \
79 else if (firm_dump.extbb) \
80 dump_ir_extblock_graph(irg, suffix); \
82 dump_ir_block_graph(irg, suffix); \
87 /** dump control flow graph irg depending on cond */
88 #define DUMP_ONE_CFG(cond, irg, suffix) \
91 timer_push(TV_VCG_DUMP); \
92 dump_cfg(irg, suffix); \
97 /** check a graph irg depending on cond */
98 #define CHECK_ONE(cond, irg) \
101 timer_push(TV_VERIFY); \
102 irg_verify(irg, VRFY_ENFORCE_SSA); \
108 /* set by the backend parameters */
109 static const ir_settings_arch_dep_t *ad_param = NULL;
110 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
111 static void *create_intrinsic_ctx = NULL;
112 static const ir_settings_if_conv_t *if_conv_info = NULL;
114 ir_mode *firm_imm_fp_mode = NULL;
116 /* entities of runtime functions */
117 ir_entity_ptr rts_entities[rts_max];
120 * factory for setting architecture dependent parameters
122 static const ir_settings_arch_dep_t *arch_factory(void)
124 static const ir_settings_arch_dep_t param = {
125 1, /* also use subs */
126 4, /* maximum shifts */
127 31, /* maximum shift amount */
128 NULL, /* use default evaluator */
132 32 /* Mulh allowed up to 32 bit */
135 return ad_param ? ad_param : ¶m;
139 * Map runtime functions.
141 static void rts_map(void) {
142 static const struct {
143 ir_entity_ptr *ent; /**< address of the rts entity */
144 i_mapper_func func; /**< mapper function. */
147 { &rts_entities[rts_abs], i_mapper_abs },
148 { &rts_entities[rts_labs], i_mapper_abs },
149 { &rts_entities[rts_llabs], i_mapper_abs },
150 { &rts_entities[rts_imaxabs], i_mapper_abs },
152 /* double -> double */
153 { &rts_entities[rts_fabs], i_mapper_abs },
154 { &rts_entities[rts_sqrt], i_mapper_sqrt },
155 { &rts_entities[rts_cbrt], i_mapper_cbrt },
156 { &rts_entities[rts_pow], i_mapper_pow },
157 { &rts_entities[rts_exp], i_mapper_exp },
158 { &rts_entities[rts_exp2], i_mapper_exp },
159 { &rts_entities[rts_exp10], i_mapper_exp },
160 { &rts_entities[rts_log], i_mapper_log },
161 { &rts_entities[rts_log2], i_mapper_log2 },
162 { &rts_entities[rts_log10], i_mapper_log10 },
163 { &rts_entities[rts_sin], i_mapper_sin },
164 { &rts_entities[rts_cos], i_mapper_cos },
165 { &rts_entities[rts_tan], i_mapper_tan },
166 { &rts_entities[rts_asin], i_mapper_asin },
167 { &rts_entities[rts_acos], i_mapper_acos },
168 { &rts_entities[rts_atan], i_mapper_atan },
169 { &rts_entities[rts_sinh], i_mapper_sinh },
170 { &rts_entities[rts_cosh], i_mapper_cosh },
171 { &rts_entities[rts_tanh], i_mapper_tanh },
174 { &rts_entities[rts_fabsf], i_mapper_abs },
175 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
176 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
177 { &rts_entities[rts_powf], i_mapper_pow },
178 { &rts_entities[rts_expf], i_mapper_exp },
179 { &rts_entities[rts_exp2f], i_mapper_exp },
180 { &rts_entities[rts_exp10f], i_mapper_exp },
181 { &rts_entities[rts_logf], i_mapper_log },
182 { &rts_entities[rts_log2f], i_mapper_log2 },
183 { &rts_entities[rts_log10f], i_mapper_log10 },
184 { &rts_entities[rts_sinf], i_mapper_sin },
185 { &rts_entities[rts_cosf], i_mapper_cos },
186 { &rts_entities[rts_tanf], i_mapper_tan },
187 { &rts_entities[rts_asinf], i_mapper_asin },
188 { &rts_entities[rts_acosf], i_mapper_acos },
189 { &rts_entities[rts_atanf], i_mapper_atan },
190 { &rts_entities[rts_sinhf], i_mapper_sinh },
191 { &rts_entities[rts_coshf], i_mapper_cosh },
192 { &rts_entities[rts_tanhf], i_mapper_tanh },
194 /* long double -> long double */
195 { &rts_entities[rts_fabsl], i_mapper_abs },
196 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
197 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
198 { &rts_entities[rts_powl], i_mapper_pow },
199 { &rts_entities[rts_expl], i_mapper_exp },
200 { &rts_entities[rts_exp2l], i_mapper_exp },
201 { &rts_entities[rts_exp10l], i_mapper_exp },
202 { &rts_entities[rts_logl], i_mapper_log },
203 { &rts_entities[rts_log2l], i_mapper_log2 },
204 { &rts_entities[rts_log10l], i_mapper_log10 },
205 { &rts_entities[rts_sinl], i_mapper_sin },
206 { &rts_entities[rts_cosl], i_mapper_cos },
207 { &rts_entities[rts_tanl], i_mapper_tan },
208 { &rts_entities[rts_asinl], i_mapper_asin },
209 { &rts_entities[rts_acosl], i_mapper_acos },
210 { &rts_entities[rts_atanl], i_mapper_atan },
211 { &rts_entities[rts_sinhl], i_mapper_sinh },
212 { &rts_entities[rts_coshl], i_mapper_cosh },
213 { &rts_entities[rts_tanhl], i_mapper_tanh },
216 { &rts_entities[rts_memcpy], i_mapper_memcpy },
217 { &rts_entities[rts_memset], i_mapper_memset },
218 { &rts_entities[rts_strcmp], i_mapper_strcmp },
219 { &rts_entities[rts_strncmp], i_mapper_strncmp },
220 { &rts_entities[rts_strlen], i_mapper_strlen }
222 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
225 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
226 if (*mapper[i].ent != NULL) {
227 rec[n_map].i_call.kind = INTRINSIC_CALL;
228 rec[n_map].i_call.i_ent = *mapper[i].ent;
229 rec[n_map].i_call.i_mapper = mapper[i].func;
230 rec[n_map].i_call.ctx = NULL;
231 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);
248 static void dump_graph_cfg_count(ir_graph *const irg, const char *const suffix)
251 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
252 DUMP_ONE_CFG(1, irg, name);
256 static void dump_all_count(const char *const suffix)
258 const int n_irgs = get_irp_n_irgs();
261 for (i = 0; i < n_irgs; ++i)
262 dump_graph_count(get_irp_irg(i), suffix);
265 #define DUMP_ONE_C(cond, irg, suffix) \
268 dump_graph_count((irg), (suffix)); \
272 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
275 dump_graph_cfg_count((irg), (suffix)); \
279 #define DUMP_ALL_C(cond, suffix) \
282 dump_all_count((suffix)); \
286 static void remove_unused_functions(void)
288 ir_entity **keep_methods;
291 /* Analysis that finds the free methods,
292 i.e. methods that are dereferenced.
293 Optimizes polymorphic calls :-). */
294 cgana(&arr_len, &keep_methods);
296 /* Remove methods that are never called. */
297 gc_irgs(arr_len, keep_methods);
301 static int firm_const_exists;
303 static void do_optimize_funccalls(void)
305 optimize_funccalls(firm_const_exists, NULL);
308 static void do_gcse(ir_graph *irg)
310 set_opt_global_cse(1);
312 set_opt_global_cse(0);
315 static void do_lower_highlevel(ir_graph *irg)
317 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
320 static void do_if_conv(ir_graph *irg)
322 opt_if_conv(irg, if_conv_info);
325 static void do_stred(ir_graph *irg)
327 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
330 static void do_inline(void)
332 inline_functions(500, 50);
335 static void do_cloning(void)
337 proc_cloning((float) firm_opt.clone_threshold);
340 static void do_lower_switch(ir_graph *irg)
342 lower_switch(irg, 128);
345 typedef enum opt_target {
350 typedef void (*transform_irg_func)(ir_graph *irg);
351 typedef void (*transform_irp_func)(void);
352 typedef void (*func_ptr_t)(void);
364 static opt_config_t opts[] = {
365 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, true, true, true, -1 },
366 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, true, true, true, TV_COMBO },
367 { OPT_TARGET_IRG, "controlflow", (func_ptr_t) optimize_cf, true, true, true, TV_CF_OPT },
368 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, true, true, true, TV_LOCAL_OPT },
369 { OPT_TARGET_IRP, "gc_irgs", (func_ptr_t) remove_unused_functions, true, false, false, TV_CGANA },
370 { OPT_TARGET_IRP, "tailrec", (func_ptr_t) opt_tail_recursion, true, true, true, TV_TAIL_REC },
371 { OPT_TARGET_IRP, "funccalls", (func_ptr_t) do_optimize_funccalls, true, true, true, TV_REAL_FUNC_CALL },
372 { OPT_TARGET_IRP, "lowerconst", (func_ptr_t) lower_const_code, true, false, false, -1 },
373 { OPT_TARGET_IRG, "onereturn", (func_ptr_t) normalize_one_return, true, false, false, -1 },
374 { OPT_TARGET_IRG, "scalar", (func_ptr_t) scalar_replacement_opt, true, true, true, TV_SCALAR_REPLACE },
375 { OPT_TARGET_IRG, "reassoc", (func_ptr_t) optimize_reassociation, true, true, true, TV_REASSOCIATION },
376 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, true, true, true, TV_CODE_PLACE },
377 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, true, true, true, TV_CODE_PLACE },
378 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, true, true, true, TV_CONFIRM_CREATE },
379 { OPT_TARGET_IRG, "ldst", (func_ptr_t) optimize_load_store, true, true, true, TV_LOAD_STORE },
380 { OPT_TARGET_IRG, "ldst2", (func_ptr_t) opt_ldst2, true, true, true, -1 },
381 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, true, true, true, -1 },
382 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, true, true, true, TV_DECONV },
383 { OPT_TARGET_IRG, "condeval", (func_ptr_t) opt_cond_eval, true, true, true, TV_COND_EVAL },
384 { OPT_TARGET_IRG, "remove_confirms", (func_ptr_t) remove_confirms, true, false, false, TV_CONFIRM_CREATE },
385 { OPT_TARGET_IRG, "gvnpre", (func_ptr_t) do_gvn_pre, true, true, true, -1 },
386 { OPT_TARGET_IRG, "ifconv", (func_ptr_t) do_if_conv, true, true, true, TV_IF_CONV },
387 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, true, true, true, -1 },
388 { OPT_TARGET_IRG, "stred", (func_ptr_t) do_stred, true, true, true, TV_OSR },
389 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, true, false, true, TV_DEAD_NODE },
390 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, true, true, true, -1 },
391 { OPT_TARGET_IRP, "clone", (func_ptr_t) do_cloning, true, true, true, -1 },
392 { OPT_TARGET_IRG, "lower_switch", (func_ptr_t) do_lower_switch, true, true, true, -1 },
394 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
396 static opt_config_t *get_opt(const char *name)
399 for (i = 0; i < n_opts; ++i) {
400 opt_config_t *config = &opts[i];
401 if (strcmp(config->name, name) == 0)
409 static void set_opt_enabled(const char *name, bool enabled)
411 opt_config_t *config = get_opt(name);
412 config->enabled = enabled;
415 static void do_irg_opt(ir_graph *irg, const char *name)
417 transform_irg_func func;
419 opt_config_t *config = get_opt(name);
420 assert(config->target == OPT_TARGET_IRG);
421 if (!config->enabled)
424 if (config->timer != -1)
425 timer_push(config->timer);
427 old_irg = current_ir_graph;
428 current_ir_graph = irg;
430 func = (transform_irg_func) config->func;
433 if (config->timer != -1)
436 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
437 CHECK_ONE(firm_opt.check_all, irg);
439 current_ir_graph = old_irg;
442 static void do_irp_opt(const char *name)
444 transform_irp_func func;
445 opt_config_t *config = get_opt(name);
446 assert(config->target == OPT_TARGET_IRP);
447 if (!config->enabled)
450 if (config->timer != -1)
451 timer_push(config->timer);
453 func = (transform_irp_func) config->func;
456 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
457 CHECK_ALL(firm_opt.check_all);
459 if (config->timer != -1)
464 * run all the Firm optimizations
466 * @param input_filename the name of the (main) source file
468 static void do_firm_optimizations(const char *input_filename)
473 /* FIXME: cloning might ADD new graphs. */
474 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
476 set_opt_strength_red(firm_opt.strength_red);
477 set_opt_scalar_replacement(firm_opt.scalar_replace);
478 set_opt_auto_create_sync(firm_opt.auto_sync);
479 set_opt_alias_analysis(firm_opt.alias_analysis);
481 aa_opt = aa_opt_no_opt;
482 if (firm_opt.strict_alias)
483 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
484 if (firm_opt.no_alias)
485 aa_opt = aa_opt_no_alias;
487 set_irp_memory_disambiguator_options(aa_opt);
489 /* parameter passing code should set them directly sometime... */
490 set_opt_enabled("rts", firm_opt.freestanding);
491 set_opt_enabled("gc_irgs", firm_opt.remove_unused);
492 set_opt_enabled("tailrec", firm_opt.tail_rec);
493 set_opt_enabled("funccalls", firm_opt.func_calls);
494 set_opt_enabled("gcse", firm_opt.gcse);
495 set_opt_enabled("place", !firm_opt.gcse);
496 set_opt_enabled("confirm", firm_opt.confirm);
497 set_opt_enabled("remove_confirms", firm_opt.confirm);
498 set_opt_enabled("ldst2", firm_opt.luffig);
499 set_opt_enabled("ldst", firm_opt.load_store);
500 set_opt_enabled("deconv", firm_opt.deconv);
501 set_opt_enabled("condeval", firm_opt.cond_eval);
502 set_opt_enabled("gvnpre", firm_opt.gvn_pre);
503 set_opt_enabled("ifconv", firm_opt.if_conversion);
504 set_opt_enabled("bool", firm_opt.bool_opt);
505 set_opt_enabled("inline", firm_opt.do_inline);
506 set_opt_enabled("clone", firm_opt.cloning);
507 set_opt_enabled("combo", firm_opt.combo);
509 timer_start(TV_ALL_OPT);
513 for (i = 0; i < get_irp_n_irgs(); i++) {
514 ir_graph *irg = get_irp_irg(i);
515 do_irg_opt(irg, "combo");
516 do_irg_opt(irg, "local");
518 /* Confirm construction currently can only handle blocks with only one control
519 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
520 the optimization of switch constructs. */
521 do_irg_opt(irg, "controlflow");
524 do_irp_opt("gc_irgs");
525 do_irp_opt("tailrec");
526 do_irp_opt("funccalls");
527 do_irp_opt("lowerconst");
529 for (i = 0; i < get_irp_n_irgs(); i++) {
530 ir_graph *irg = get_irp_irg(i);
533 /* If SIMD optimization is on, make sure we have only 1 return */
534 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
535 do_irg_opt("onereturn");
538 do_irg_opt(irg, "scalar");
539 do_irg_opt(irg, "local");
540 do_irg_opt(irg, "reassoc");
541 do_irg_opt(irg, "local");
542 do_irg_opt(irg, "gcse");
544 if (firm_opt.confirm) {
545 /* Confirm construction currently can only handle blocks with only one control
546 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
547 the optimization of switch constructs. */
548 do_irg_opt(irg, "controlflow");
549 do_irg_opt(irg, "confirm");
550 do_irg_opt(irg, "local");
553 do_irg_opt(irg, "ldst2");
554 do_irg_opt(irg, "controlflow");
555 do_irg_opt(irg, "ldst");
556 do_irg_opt(irg, "lower");
557 do_irg_opt(irg, "deconv");
558 do_irg_opt(irg, "condeval");
559 do_irg_opt(irg, "lower_switch");
560 do_irg_opt(irg, "remove_confirms");
561 do_irg_opt(irg, "gvnpre");
562 do_irg_opt(irg, "place");
563 do_irg_opt(irg, "controlflow");
565 if (firm_opt.if_conversion) {
566 do_irg_opt(irg, "ifconv");
567 do_irg_opt(irg, "local");
568 do_irg_opt(irg, "controlflow");
571 do_irg_opt(irg, "bool");
572 do_irg_opt(irg, "stred");
573 do_irg_opt(irg, "local");
574 do_irg_opt(irg, "dead");
577 do_irp_opt("inline");
580 for (i = 0; i < get_irp_n_irgs(); i++) {
581 ir_graph *irg = get_irp_irg(i);
582 do_irg_opt(irg, "local");
583 do_irg_opt(irg, "controlflow");
584 do_irg_opt(irg, "condeval");
585 do_irg_opt(irg, "local");
586 do_irg_opt(irg, "controlflow");
589 if (firm_dump.ir_graph) {
590 /* recompute backedges for nicer dumps */
591 for (i = 0; i < get_irp_n_irgs(); i++)
592 construct_cf_backedges(get_irp_irg(i));
595 do_irp_opt("gc_irgs");
597 DUMP_ALL(firm_dump.ir_graph, "-opt");
598 /* verify optimized graphs */
599 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
600 ir_graph *irg = get_irp_irg(i);
601 irg_verify(irg, VRFY_ENFORCE_SSA);
604 if (firm_dump.statistic & STAT_AFTER_OPT)
605 stat_dump_snapshot(input_filename, "opt");
607 timer_stop(TV_ALL_OPT);
608 } /* do_firm_optimizations */
611 * compute the size of a type (do implicit lowering)
613 * @param ty a Firm type
615 static int compute_type_size(ir_type *ty)
617 optimization_state_t state;
618 unsigned align_all = 1;
619 int n, size = 0, set = 0;
622 if (get_type_state(ty) == layout_fixed) {
623 /* do not layout already layouted types again */
627 if (is_Method_type(ty) || ty == get_glob_type()) {
628 /* no need for size calculation for method types or the global type */
632 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
634 switch (get_type_tpop_code(ty)) {
637 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
638 ir_entity *ent = get_compound_member(ty, i);
639 ir_type *ent_ty = get_entity_type(ent);
640 unsigned align, misalign;
642 /* compute member types */
643 if (! compute_type_size(ent_ty))
646 align = get_type_alignment_bytes(ent_ty);
647 align_all = align > align_all ? align : align_all;
648 misalign = (align ? size % align : 0);
649 size += (misalign ? align - misalign : 0);
651 set_entity_offset(ent, size);
652 size += get_type_size_bytes(ent_ty);
654 DBG((" member %s %s -> (size: %u, align: %u)\n",
655 get_type_name(ent_ty), get_entity_name(ent),
656 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
658 if (align_all > 0 && size % align_all) {
659 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
660 size += align_all - (size % align_all);
661 DBG(("correcting type-size to %d\n", size));
663 set_type_alignment_bytes(ty, align_all);
668 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
669 ir_entity *ent = get_union_member(ty, i);
671 if (! compute_type_size(get_entity_type(ent)))
673 s = get_type_size_bytes(get_entity_type(ent));
675 set_entity_offset(ent, 0);
676 size = (s > size ? s : size);
682 dims = get_array_n_dimensions(ty);
684 if (! compute_type_size(get_array_element_type(ty)))
689 save_optimization_state(&state);
691 set_opt_constant_folding(1);
692 set_opt_algebraic_simplification(1);
694 for (i = 0; i < dims; ++i) {
695 ir_node *lower = get_array_lower_bound(ty, i);
696 ir_node *upper = get_array_upper_bound(ty, i);
697 ir_graph *rem = current_ir_graph;
698 tarval *tv_lower, *tv_upper;
699 long val_lower, val_upper;
701 current_ir_graph = get_const_code_irg();
702 local_optimize_node(lower);
703 local_optimize_node(upper);
704 current_ir_graph = rem;
706 tv_lower = computed_value(lower);
707 tv_upper = computed_value(upper);
709 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
711 * we cannot calculate the size of this array yet, it
712 * even might be unknown until the end, like argv[]
714 restore_optimization_state(&state);
718 val_upper = get_tarval_long(tv_upper);
719 val_lower = get_tarval_long(tv_lower);
720 size *= val_upper - val_lower;
722 restore_optimization_state(&state);
724 DBG(("array %s -> (elements: %d, element type size: %d)\n",
726 size, get_type_size_bytes(get_array_element_type(ty))));
727 size *= get_type_size_bytes(get_array_element_type(ty));
736 set_type_size_bytes(ty, size);
737 set_type_state(ty, layout_fixed);
740 DBG(("size: %d\n", get_type_size_bytes(ty)));
743 } /* compute_type_size */
746 * layout all non-frame types of the Firm graph
748 static void compute_type_sizes(void)
753 /* all non-frame other types */
754 for (i = get_irp_n_types() - 1; i >= 0; --i) {
755 tp = get_irp_type(i);
756 compute_type_size(tp);
758 if (is_Method_type(tp)) {
759 tp = get_method_value_res_type(tp);
762 /* we have a value result type for this method, lower */
763 compute_type_size(tp);
767 } /* compute_type_sizes */
770 * layout all frame-types of the Firm graph
772 static void compute_frame_type_sizes(void)
777 /* all frame types */
778 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
779 irg = get_irp_irg(i);
780 /* do not optimize away variables in debug mode */
781 if (firm_opt.debug_mode == DBG_MODE_NONE)
783 compute_type_size(get_irg_frame_type(irg));
785 } /* compute_frame_type_sizes */
790 * @param input_filename the name of the (main) source file
792 static void do_firm_lowering(const char *input_filename)
796 /* do class lowering and vtbl creation */
797 // lower_classes_to_struct("vtbl", "m");
800 timer_push(TV_LOWER);
805 if (firm_opt.lower_ll) {
806 lwrdw_param_t init = {
811 def_create_intrinsic_fkt,
816 if (arch_create_intrinsic) {
817 init.create_intrinsic = arch_create_intrinsic;
818 init.ctx = create_intrinsic_ctx;
820 timer_push(TV_DW_LOWER);
825 if (firm_dump.statistic & STAT_AFTER_LOWER)
826 stat_dump_snapshot(input_filename, "low");
828 /* verify lowered graphs */
829 timer_push(TV_VERIFY);
830 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
831 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
834 DUMP_ALL(firm_dump.ir_graph, "-low");
836 if (firm_opt.enabled) {
837 timer_start(TV_ALL_OPT);
839 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
841 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
842 current_ir_graph = get_irp_irg(i);
844 timer_push(TV_REASSOCIATION);
845 optimize_reassociation(current_ir_graph);
847 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
848 CHECK_ONE(firm_opt.check_all, current_ir_graph);
851 /* enable architecture dependent optimizations */
852 arch_dep_set_opts((arch_dep_opts_t)
853 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
854 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
855 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
857 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
858 current_ir_graph = get_irp_irg(i);
861 set_opt_global_cse(1);
863 timer_push(TV_LOCAL_OPT);
864 optimize_graph_df(current_ir_graph);
866 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
868 CHECK_ONE(firm_opt.check_all, current_ir_graph);
871 timer_push(TV_CODE_PLACE);
872 place_code(current_ir_graph);
873 set_opt_global_cse(0);
875 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
876 CHECK_ONE(firm_opt.check_all, current_ir_graph);
879 timer_push(TV_LOAD_STORE);
880 optimize_load_store(current_ir_graph);
882 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
883 CHECK_ONE(firm_opt.check_all, current_ir_graph);
886 timer_push(TV_LOCAL_OPT);
887 optimize_graph_df(current_ir_graph);
889 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
891 timer_push(TV_CF_OPT);
892 optimize_cf(current_ir_graph);
894 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
895 CHECK_ONE(firm_opt.check_all, current_ir_graph);
897 if (firm_opt.if_conversion) {
898 timer_push(TV_IF_CONV);
899 opt_if_conv(current_ir_graph, if_conv_info);
901 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
902 CHECK_ONE(firm_opt.check_all, current_ir_graph);
904 timer_push(TV_LOCAL_OPT);
905 optimize_graph_df(current_ir_graph);
907 timer_push(TV_CF_OPT);
908 optimize_cf(current_ir_graph);
910 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
911 CHECK_ONE(firm_opt.check_all, current_ir_graph);
914 timer_stop(TV_ALL_OPT);
916 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
920 mark_private_methods();
922 /* set the phase to low */
923 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
924 set_irg_phase_low(get_irp_irg(i));
926 /* all graphs are lowered, set the irp phase to low */
927 set_irp_phase_state(phase_low);
929 if (firm_dump.statistic & STAT_FINAL) {
930 stat_dump_snapshot(input_filename, "final");
932 } /* do_firm_lowering */
935 * Initialize for the Firm-generating back end.
937 void gen_firm_init(void)
939 firm_parameter_t params;
940 unsigned pattern = 0;
942 /* the automatic state is only set if inlining is enabled */
943 firm_opt.auto_inline = firm_opt.do_inline;
945 if (firm_dump.stat_pattern)
946 pattern |= FIRMSTAT_PATTERN_ENABLED;
948 if (firm_dump.stat_dag)
949 pattern |= FIRMSTAT_COUNT_DAG;
951 memset(¶ms, 0, sizeof(params));
952 params.size = sizeof(params);
953 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
954 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS | pattern;
955 params.initialize_local_func = uninitialized_local_var;
956 params.cc_mask = 0; /* no regparam, cdecl */
957 params.builtin_dbg = NULL;
960 /* Activate Graph rewriting if SIMD optimization is turned on */
961 /* This has to be done before init_firm() is called! */
962 if (firm_ext_grs.simd_opt)
968 if (firm_be_opt.selection == BE_FIRM_BE) {
969 const backend_params *be_params = be_get_backend_param();
971 firm_opt.lower_ll = (a_byte) be_params->do_dw_lowering;
972 params.arch_op_settings = be_params->arch_op_settings;
974 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
975 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
977 ad_param = be_params->dep_param;
978 if_conv_info = be_params->if_conv_info;
980 if (be_params->has_imm_fp_mode)
981 firm_imm_fp_mode = be_params->imm_fp_mode;
983 /* OS option must be set to the backend */
984 switch (firm_opt.os_support) {
985 case OS_SUPPORT_MINGW:
986 firm_be_option("ia32-gasmode=mingw");
988 case OS_SUPPORT_MACHO:
989 firm_be_option("ia32-gasmode=macho");
991 case OS_SUPPORT_LINUX:
993 firm_be_option("ia32-gasmode=linux");
997 dbg_init(NULL, NULL, dbg_snprint);
998 edges_init_dbg(firm_opt.vrfy_edges);
999 //cbackend_set_debug_retrieve(dbg_retrieve);
1001 set_opt_precise_exc_context(firm_opt.precise_exc);
1002 set_opt_fragile_ops(firm_opt.fragile_ops);
1004 /* Sel node cannot produce NULL pointers */
1005 set_opt_sel_based_null_check_elim(1);
1007 /* dynamic dispatch works currently only if whole world scenarios */
1008 set_opt_dyn_meth_dispatch(0);
1010 arch_dep_init(arch_factory);
1012 /* do not run architecture dependent optimizations in building phase */
1013 arch_dep_set_opts(arch_dep_none);
1015 do_node_verification((firm_verification_t) firm_opt.vrfy);
1016 if (firm_dump.filter)
1017 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
1019 if (firm_opt.enabled) {
1021 set_opt_constant_folding(firm_opt.const_folding);
1022 set_opt_algebraic_simplification(firm_opt.const_folding);
1023 set_opt_cse(firm_opt.cse);
1024 set_opt_global_cse(0);
1025 set_opt_unreachable_code(1);
1026 set_opt_control_flow(firm_opt.control_flow);
1027 set_opt_control_flow_weak_simplification(1);
1028 set_opt_control_flow_strong_simplification(1);
1033 /* do not dump entity ld names */
1035 } /* gen_firm_init */
1038 * Called, after the Firm generation is completed,
1039 * do all optimizations and backend call here.
1041 * @param out a file handle for the output, may be NULL
1042 * @param input_filename the name of the (main) source file
1043 * @param c_mode non-zero if "C" was compiled
1044 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
1046 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
1050 firm_const_exists = new_firm_const_exists;
1052 /* the general for dumping option must be set, or the others will not work */
1054 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1056 dump_keepalive_edges(1);
1057 dump_consts_local(1);
1058 dump_dominator_information(1);
1059 dump_loop_information(0);
1061 if (!firm_dump.edge_labels)
1062 turn_off_edge_labels();
1064 if (firm_dump.all_types) {
1067 dump_class_hierarchy(0, "");
1068 dump_class_hierarchy(1, "-with-entities");
1072 /* finalize all graphs */
1073 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1074 ir_graph *irg = get_irp_irg(i);
1076 irg_finalize_cons(irg);
1077 DUMP_ONE(firm_dump.ir_graph, irg, "");
1079 /* verify the graph */
1080 timer_push(TV_VERIFY);
1081 irg_verify(irg, VRFY_ENFORCE_SSA);
1085 timer_push(TV_VERIFY);
1089 /* all graphs are finalized, set the irp phase to high */
1090 set_irp_phase_state(phase_high);
1092 /* BEWARE: kill unreachable code before doing compound lowering */
1093 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1094 ir_graph *irg = get_irp_irg(i);
1098 /* lower all compound call return values */
1099 lower_compound_params();
1101 /* computes the sizes of all types that are still not computed */
1102 compute_type_sizes();
1104 /* lower copyb nodes */
1105 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1106 ir_graph *irg = get_irp_irg(i);
1107 lower_CopyB(irg, 128, 4);
1110 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1111 stat_dump_snapshot(input_filename, "noopt");
1114 if (firm_opt.enabled)
1115 do_firm_optimizations(input_filename);
1117 if (firm_dump.gen_firm_asm) {
1118 timer_push(TV_FIRM_ASM);
1119 gen_Firm_assembler(input_filename);
1125 do_firm_lowering(input_filename);
1127 /* computes the sizes of all frame types */
1128 compute_frame_type_sizes();
1130 /* set the phase to low */
1131 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1132 set_irg_phase_low(get_irp_irg(i));
1136 /** SIMD Optimization Extensions **/
1138 /* Pattern creation step. No code has to be generated, so
1139 exit after pattern creation */
1140 if (firm_ext_grs.create_pattern) {
1141 ext_grs_create_pattern();
1145 /* SIMD optimization step. Uses graph patterns to find
1146 rich instructions and rewrite */
1147 if (firm_ext_grs.simd_opt)
1151 if (firm_dump.statistic & STAT_FINAL_IR)
1152 stat_dump_snapshot(input_filename, "final-ir");
1154 /* run the code generator */
1155 if (firm_be_opt.selection != BE_NONE)
1156 do_codegen(out, input_filename);
1158 if (firm_dump.statistic & STAT_FINAL)
1159 stat_dump_snapshot(input_filename, "final");
1162 if (firm_opt.ycomp_dbg)
1163 firm_finish_ycomp_debugger();
1165 } /* gen_firm_finish */
1168 * Do very early initializations
1170 void firm_early_init(void) {
1171 /* arg: need this here for command line options */
1173 firm_init_options(NULL, 0, NULL);
1174 } /* firm_early_init */