3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2007 Michael Beck beck@ipd.info.uni-karlsruhe.de
12 #include <libfirm/firm.h>
13 #include <libfirm/be.h>
17 #include "firm_codegen.h"
18 #include "firm_cmdline.h"
19 #include "firm_timing.h"
22 #define snprintf _snprintf
25 #if defined(_DEBUG) || defined(FIRM_DEBUG)
26 #define DBG(x) dbg_printf x
29 #endif /* _DEBUG || FIRM_DEBUG */
32 /** dump all the graphs depending on cond */
33 #define DUMP_ALL(cond, suffix) \
36 timer_push(TV_VCG_DUMP); \
37 if (firm_dump.no_blocks) \
38 dump_all_ir_graphs(dump_ir_graph, suffix); \
39 else if (firm_dump.extbb) \
40 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);\
42 dump_all_ir_graphs(dump_ir_block_graph, suffix); \
47 /** dump all control flow graphs depending on cond */
48 #define DUMP_ALL_CFG(cond, suffix) \
51 timer_push(TV_VCG_DUMP); \
52 dump_all_ir_graphs(dump_cfg, suffix); \
57 /** check all graphs depending on cond */
58 #define CHECK_ALL(cond) \
62 timer_push(TV_VERIFY); \
63 for (ii = get_irp_n_irgs() - 1; ii >= 0; --ii) \
64 irg_verify(get_irp_irg(ii), VRFY_ENFORCE_SSA); \
71 /** dump graphs irg depending on cond */
72 #define DUMP_ONE(cond, irg, suffix) \
75 timer_push(TV_VCG_DUMP); \
76 if (firm_dump.no_blocks) \
77 dump_ir_graph(irg, suffix); \
78 else if (firm_dump.extbb) \
79 dump_ir_extblock_graph(irg, suffix); \
81 dump_ir_block_graph(irg, suffix); \
86 /** dump control flow graph irg depending on cond */
87 #define DUMP_ONE_CFG(cond, irg, suffix) \
90 timer_push(TV_VCG_DUMP); \
91 dump_cfg(irg, suffix); \
96 /** check a graph irg depending on cond */
97 #define CHECK_ONE(cond, irg) \
100 timer_push(TV_VERIFY); \
101 irg_verify(irg, VRFY_ENFORCE_SSA); \
107 /* set by the backend parameters */
108 static const ir_settings_arch_dep_t *ad_param = NULL;
109 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
110 static void *create_intrinsic_ctx = NULL;
111 static const ir_settings_if_conv_t *if_conv_info = NULL;
113 ir_mode *firm_imm_fp_mode = NULL;
115 /* entities of runtime functions */
116 ir_entity_ptr rts_entities[rts_max];
119 * factory for setting architecture dependent parameters
121 static const ir_settings_arch_dep_t *arch_factory(void)
123 static const ir_settings_arch_dep_t param = {
124 1, /* also use subs */
125 4, /* maximum shifts */
126 31, /* maximum shift amount */
127 NULL, /* use default evaluator */
131 32 /* Mulh allowed up to 32 bit */
134 return ad_param ? ad_param : ¶m;
138 * Map runtime functions.
140 static void rts_map(void) {
141 static const struct {
142 ir_entity_ptr *ent; /**< address of the rts entity */
143 i_mapper_func func; /**< mapper function. */
146 { &rts_entities[rts_abs], i_mapper_abs },
147 { &rts_entities[rts_labs], i_mapper_abs },
148 { &rts_entities[rts_llabs], i_mapper_abs },
149 { &rts_entities[rts_imaxabs], i_mapper_abs },
151 /* double -> double */
152 { &rts_entities[rts_fabs], i_mapper_abs },
153 { &rts_entities[rts_sqrt], i_mapper_sqrt },
154 { &rts_entities[rts_cbrt], i_mapper_cbrt },
155 { &rts_entities[rts_pow], i_mapper_pow },
156 { &rts_entities[rts_exp], i_mapper_exp },
157 { &rts_entities[rts_exp2], i_mapper_exp },
158 { &rts_entities[rts_exp10], i_mapper_exp },
159 { &rts_entities[rts_log], i_mapper_log },
160 { &rts_entities[rts_log2], i_mapper_log2 },
161 { &rts_entities[rts_log10], i_mapper_log10 },
162 { &rts_entities[rts_sin], i_mapper_sin },
163 { &rts_entities[rts_cos], i_mapper_cos },
164 { &rts_entities[rts_tan], i_mapper_tan },
165 { &rts_entities[rts_asin], i_mapper_asin },
166 { &rts_entities[rts_acos], i_mapper_acos },
167 { &rts_entities[rts_atan], i_mapper_atan },
168 { &rts_entities[rts_sinh], i_mapper_sinh },
169 { &rts_entities[rts_cosh], i_mapper_cosh },
170 { &rts_entities[rts_tanh], i_mapper_tanh },
173 { &rts_entities[rts_fabsf], i_mapper_abs },
174 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
175 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
176 { &rts_entities[rts_powf], i_mapper_pow },
177 { &rts_entities[rts_expf], i_mapper_exp },
178 { &rts_entities[rts_exp2f], i_mapper_exp },
179 { &rts_entities[rts_exp10f], i_mapper_exp },
180 { &rts_entities[rts_logf], i_mapper_log },
181 { &rts_entities[rts_log2f], i_mapper_log2 },
182 { &rts_entities[rts_log10f], i_mapper_log10 },
183 { &rts_entities[rts_sinf], i_mapper_sin },
184 { &rts_entities[rts_cosf], i_mapper_cos },
185 { &rts_entities[rts_tanf], i_mapper_tan },
186 { &rts_entities[rts_asinf], i_mapper_asin },
187 { &rts_entities[rts_acosf], i_mapper_acos },
188 { &rts_entities[rts_atanf], i_mapper_atan },
189 { &rts_entities[rts_sinhf], i_mapper_sinh },
190 { &rts_entities[rts_coshf], i_mapper_cosh },
191 { &rts_entities[rts_tanhf], i_mapper_tanh },
193 /* long double -> long double */
194 { &rts_entities[rts_fabsl], i_mapper_abs },
195 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
196 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
197 { &rts_entities[rts_powl], i_mapper_pow },
198 { &rts_entities[rts_expl], i_mapper_exp },
199 { &rts_entities[rts_exp2l], i_mapper_exp },
200 { &rts_entities[rts_exp10l], i_mapper_exp },
201 { &rts_entities[rts_logl], i_mapper_log },
202 { &rts_entities[rts_log2l], i_mapper_log2 },
203 { &rts_entities[rts_log10l], i_mapper_log10 },
204 { &rts_entities[rts_sinl], i_mapper_sin },
205 { &rts_entities[rts_cosl], i_mapper_cos },
206 { &rts_entities[rts_tanl], i_mapper_tan },
207 { &rts_entities[rts_asinl], i_mapper_asin },
208 { &rts_entities[rts_acosl], i_mapper_acos },
209 { &rts_entities[rts_atanl], i_mapper_atan },
210 { &rts_entities[rts_sinhl], i_mapper_sinh },
211 { &rts_entities[rts_coshl], i_mapper_cosh },
212 { &rts_entities[rts_tanhl], i_mapper_tanh },
215 { &rts_entities[rts_memcpy], i_mapper_memcpy },
216 { &rts_entities[rts_memset], i_mapper_memset },
217 { &rts_entities[rts_strcmp], i_mapper_strcmp },
218 { &rts_entities[rts_strncmp], i_mapper_strncmp },
219 { &rts_entities[rts_strlen], i_mapper_strlen }
221 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
224 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
225 if (*mapper[i].ent != NULL) {
226 rec[n_map].i_call.kind = INTRINSIC_CALL;
227 rec[n_map].i_call.i_ent = *mapper[i].ent;
228 rec[n_map].i_call.i_mapper = mapper[i].func;
229 rec[n_map].i_call.ctx = NULL;
230 rec[n_map].i_call.link = NULL;
234 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
237 static int *irg_dump_no;
239 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
242 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
243 DUMP_ONE(1, irg, name);
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);
255 static void dump_all_count(const char *const suffix)
257 const int n_irgs = get_irp_n_irgs();
260 for (i = 0; i < n_irgs; ++i)
261 dump_graph_count(get_irp_irg(i), suffix);
264 #define DUMP_ONE_C(cond, irg, suffix) \
267 dump_graph_count((irg), (suffix)); \
271 #define DUMP_ONE_CFG_C(cond, irg, suffix) \
274 dump_graph_cfg_count((irg), (suffix)); \
278 #define DUMP_ALL_C(cond, suffix) \
281 dump_all_count((suffix)); \
285 static void remove_unused_functions(void)
287 ir_entity **keep_methods;
290 /* Analysis that finds the free methods,
291 i.e. methods that are dereferenced.
292 Optimizes polymorphic calls :-). */
293 cgana(&arr_len, &keep_methods);
295 /* Remove methods that are never called. */
296 gc_irgs(arr_len, keep_methods);
300 static int firm_const_exists;
302 static void do_optimize_funccalls(void)
304 optimize_funccalls(firm_const_exists, NULL);
307 static void do_gcse(ir_graph *irg)
309 set_opt_global_cse(1);
311 set_opt_global_cse(0);
314 static void do_lower_highlevel(ir_graph *irg)
316 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
319 static void do_if_conv(ir_graph *irg)
321 opt_if_conv(irg, if_conv_info);
324 static void do_stred(ir_graph *irg)
326 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
329 static void do_inline(void)
331 inline_functions(500, 50);
334 static void do_cloning(void)
336 proc_cloning((float) firm_opt.clone_threshold);
339 static void do_lower_switch(ir_graph *irg)
341 lower_switch(irg, 128);
344 typedef enum opt_target {
349 typedef void (*transform_irg_func)(ir_graph *irg);
350 typedef void (*transform_irp_func)(void);
351 typedef void (*func_ptr_t)(void);
363 static opt_config_t opts[] = {
364 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, true, true, true, -1 },
365 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, true, true, true, TV_COMBO },
366 { OPT_TARGET_IRG, "controlflow", (func_ptr_t) optimize_cf, true, true, true, TV_CF_OPT },
367 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, true, true, true, TV_LOCAL_OPT },
368 { OPT_TARGET_IRP, "gc_irgs", (func_ptr_t) remove_unused_functions, true, false, false, TV_CGANA },
369 { OPT_TARGET_IRP, "tailrec", (func_ptr_t) opt_tail_recursion, true, true, true, TV_TAIL_REC },
370 { OPT_TARGET_IRP, "funccalls", (func_ptr_t) do_optimize_funccalls, true, true, true, TV_REAL_FUNC_CALL },
371 { OPT_TARGET_IRP, "lowerconst", (func_ptr_t) lower_const_code, true, false, false, -1 },
372 { OPT_TARGET_IRG, "onereturn", (func_ptr_t) normalize_one_return, true, false, false, -1 },
373 { OPT_TARGET_IRG, "scalar", (func_ptr_t) scalar_replacement_opt, true, true, true, TV_SCALAR_REPLACE },
374 { OPT_TARGET_IRG, "reassoc", (func_ptr_t) optimize_reassociation, true, true, true, TV_REASSOCIATION },
375 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, true, true, true, TV_CODE_PLACE },
376 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, true, true, true, TV_CODE_PLACE },
377 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, true, true, true, TV_CONFIRM_CREATE },
378 { OPT_TARGET_IRG, "ldst", (func_ptr_t) optimize_load_store, true, true, true, TV_LOAD_STORE },
379 { OPT_TARGET_IRG, "ldst2", (func_ptr_t) opt_ldst2, true, true, true, -1 },
380 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, true, true, true, -1 },
381 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, true, true, true, TV_DECONV },
382 { OPT_TARGET_IRG, "condeval", (func_ptr_t) opt_cond_eval, true, true, true, TV_COND_EVAL },
383 { OPT_TARGET_IRG, "remove_confirms", (func_ptr_t) remove_confirms, true, false, false, TV_CONFIRM_CREATE },
384 { OPT_TARGET_IRG, "gvnpre", (func_ptr_t) do_gvn_pre, true, true, true, -1 },
385 { OPT_TARGET_IRG, "ifconv", (func_ptr_t) do_if_conv, true, true, true, TV_IF_CONV },
386 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, true, true, true, -1 },
387 { OPT_TARGET_IRG, "stred", (func_ptr_t) do_stred, true, true, true, TV_OSR },
388 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, true, false, true, TV_DEAD_NODE },
389 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, true, true, true, -1 },
390 { OPT_TARGET_IRP, "clone", (func_ptr_t) do_cloning, true, true, true, -1 },
391 { OPT_TARGET_IRG, "lower_switch", (func_ptr_t) do_lower_switch, true, true, true, -1 },
393 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
395 static opt_config_t *get_opt(const char *name)
398 for (i = 0; i < n_opts; ++i) {
399 opt_config_t *config = &opts[i];
400 if (strcmp(config->name, name) == 0)
408 static void set_opt_enabled(const char *name, bool enabled)
410 opt_config_t *config = get_opt(name);
411 config->enabled = enabled;
414 static void do_irg_opt(ir_graph *irg, const char *name)
416 transform_irg_func func;
418 opt_config_t *config = get_opt(name);
419 assert(config->target == OPT_TARGET_IRG);
420 if (!config->enabled)
423 if (config->timer != -1)
424 timer_push(config->timer);
426 old_irg = current_ir_graph;
427 current_ir_graph = irg;
429 func = (transform_irg_func) config->func;
432 if (config->timer != -1)
435 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
436 CHECK_ONE(firm_opt.check_all, irg);
438 current_ir_graph = old_irg;
441 static void do_irp_opt(const char *name)
443 transform_irp_func func;
444 opt_config_t *config = get_opt(name);
445 assert(config->target == OPT_TARGET_IRP);
446 if (!config->enabled)
449 if (config->timer != -1)
450 timer_push(config->timer);
452 func = (transform_irp_func) config->func;
455 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
456 CHECK_ALL(firm_opt.check_all);
458 if (config->timer != -1)
463 * run all the Firm optimizations
465 * @param input_filename the name of the (main) source file
467 static void do_firm_optimizations(const char *input_filename)
472 /* FIXME: cloning might ADD new graphs. */
473 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
475 set_opt_strength_red(firm_opt.strength_red);
476 set_opt_scalar_replacement(firm_opt.scalar_replace);
477 set_opt_auto_create_sync(firm_opt.auto_sync);
478 set_opt_alias_analysis(firm_opt.alias_analysis);
480 aa_opt = aa_opt_no_opt;
481 if (firm_opt.strict_alias)
482 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
483 if (firm_opt.no_alias)
484 aa_opt = aa_opt_no_alias;
486 set_irp_memory_disambiguator_options(aa_opt);
488 /* parameter passing code should set them directly sometime... */
489 set_opt_enabled("rts", firm_opt.freestanding);
490 set_opt_enabled("gc_irgs", firm_opt.remove_unused);
491 set_opt_enabled("tailrec", firm_opt.tail_rec);
492 set_opt_enabled("funccalls", firm_opt.func_calls);
493 set_opt_enabled("gcse", firm_opt.gcse);
494 set_opt_enabled("place", !firm_opt.gcse);
495 set_opt_enabled("confirm", firm_opt.confirm);
496 set_opt_enabled("remove_confirms", firm_opt.confirm);
497 set_opt_enabled("ldst2", firm_opt.luffig);
498 set_opt_enabled("ldst", firm_opt.load_store);
499 set_opt_enabled("deconv", firm_opt.deconv);
500 set_opt_enabled("condeval", firm_opt.cond_eval);
501 set_opt_enabled("gvnpre", firm_opt.gvn_pre);
502 set_opt_enabled("ifconv", firm_opt.if_conversion);
503 set_opt_enabled("bool", firm_opt.bool_opt);
504 set_opt_enabled("inline", firm_opt.do_inline);
505 set_opt_enabled("clone", firm_opt.cloning);
506 set_opt_enabled("combo", firm_opt.combo);
508 timer_start(TV_ALL_OPT);
512 for (i = 0; i < get_irp_n_irgs(); i++) {
513 ir_graph *irg = get_irp_irg(i);
514 do_irg_opt(irg, "combo");
515 do_irg_opt(irg, "local");
517 /* Confirm construction currently can only handle blocks with only one control
518 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
519 the optimization of switch constructs. */
520 do_irg_opt(irg, "controlflow");
523 do_irp_opt("gc_irgs");
524 do_irp_opt("tailrec");
525 do_irp_opt("funccalls");
526 do_irp_opt("lowerconst");
528 for (i = 0; i < get_irp_n_irgs(); i++) {
529 ir_graph *irg = get_irp_irg(i);
532 /* If SIMD optimization is on, make sure we have only 1 return */
533 if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
534 do_irg_opt("onereturn");
537 do_irg_opt(irg, "scalar");
538 do_irg_opt(irg, "local");
539 do_irg_opt(irg, "reassoc");
540 do_irg_opt(irg, "local");
541 do_irg_opt(irg, "gcse");
543 if (firm_opt.confirm) {
544 /* Confirm construction currently can only handle blocks with only one control
545 flow predecessor. Calling optimize_cf here removes Bad predecessors and help
546 the optimization of switch constructs. */
547 do_irg_opt(irg, "controlflow");
548 do_irg_opt(irg, "confirm");
549 do_irg_opt(irg, "local");
552 do_irg_opt(irg, "ldst2");
553 do_irg_opt(irg, "controlflow");
554 do_irg_opt(irg, "ldst");
555 do_irg_opt(irg, "lower");
556 do_irg_opt(irg, "deconv");
557 do_irg_opt(irg, "condeval");
558 do_irg_opt(irg, "lower_switch");
559 do_irg_opt(irg, "remove_confirms");
560 do_irg_opt(irg, "gvnpre");
561 do_irg_opt(irg, "place");
562 do_irg_opt(irg, "controlflow");
564 if (firm_opt.if_conversion) {
565 do_irg_opt(irg, "ifconv");
566 do_irg_opt(irg, "local");
567 do_irg_opt(irg, "controlflow");
570 do_irg_opt(irg, "bool");
571 do_irg_opt(irg, "stred");
572 do_irg_opt(irg, "local");
573 do_irg_opt(irg, "dead");
576 do_irp_opt("inline");
579 for (i = 0; i < get_irp_n_irgs(); i++) {
580 ir_graph *irg = get_irp_irg(i);
581 do_irg_opt(irg, "local");
582 do_irg_opt(irg, "controlflow");
583 do_irg_opt(irg, "condeval");
584 do_irg_opt(irg, "local");
585 do_irg_opt(irg, "controlflow");
588 if (firm_dump.ir_graph) {
589 /* recompute backedges for nicer dumps */
590 for (i = 0; i < get_irp_n_irgs(); i++)
591 construct_cf_backedges(get_irp_irg(i));
594 do_irp_opt("gc_irgs");
596 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);
691 set_opt_algebraic_simplification(1);
693 for (i = 0; i < dims; ++i) {
694 ir_node *lower = get_array_lower_bound(ty, i);
695 ir_node *upper = get_array_upper_bound(ty, i);
696 ir_graph *rem = current_ir_graph;
697 tarval *tv_lower, *tv_upper;
698 long val_lower, val_upper;
700 current_ir_graph = get_const_code_irg();
701 local_optimize_node(lower);
702 local_optimize_node(upper);
703 current_ir_graph = rem;
705 tv_lower = computed_value(lower);
706 tv_upper = computed_value(upper);
708 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
710 * we cannot calculate the size of this array yet, it
711 * even might be unknown until the end, like argv[]
713 restore_optimization_state(&state);
717 val_upper = get_tarval_long(tv_upper);
718 val_lower = get_tarval_long(tv_lower);
719 size *= val_upper - val_lower;
721 restore_optimization_state(&state);
723 DBG(("array %s -> (elements: %d, element type size: %d)\n",
725 size, get_type_size_bytes(get_array_element_type(ty))));
726 size *= get_type_size_bytes(get_array_element_type(ty));
735 set_type_size_bytes(ty, size);
736 set_type_state(ty, layout_fixed);
739 DBG(("size: %d\n", get_type_size_bytes(ty)));
742 } /* compute_type_size */
745 * layout all non-frame types of the Firm graph
747 static void compute_type_sizes(void)
752 /* all non-frame other types */
753 for (i = get_irp_n_types() - 1; i >= 0; --i) {
754 tp = get_irp_type(i);
755 compute_type_size(tp);
757 if (is_Method_type(tp)) {
758 tp = get_method_value_res_type(tp);
761 /* we have a value result type for this method, lower */
762 compute_type_size(tp);
766 } /* compute_type_sizes */
769 * layout all frame-types of the Firm graph
771 static void compute_frame_type_sizes(void)
776 /* all frame types */
777 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
778 irg = get_irp_irg(i);
779 /* do not optimize away variables in debug mode */
780 if (firm_opt.debug_mode == DBG_MODE_NONE)
782 compute_type_size(get_irg_frame_type(irg));
784 } /* compute_frame_type_sizes */
789 * @param input_filename the name of the (main) source file
791 static void do_firm_lowering(const char *input_filename)
795 /* do class lowering and vtbl creation */
796 // lower_classes_to_struct("vtbl", "m");
799 timer_push(TV_LOWER);
804 if (firm_opt.lower_ll) {
805 lwrdw_param_t init = {
810 def_create_intrinsic_fkt,
815 if (arch_create_intrinsic) {
816 init.create_intrinsic = arch_create_intrinsic;
817 init.ctx = create_intrinsic_ctx;
819 timer_push(TV_DW_LOWER);
824 if (firm_dump.statistic & STAT_AFTER_LOWER)
825 stat_dump_snapshot(input_filename, "low");
827 /* verify lowered graphs */
828 timer_push(TV_VERIFY);
829 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
830 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
833 DUMP_ALL(firm_dump.ir_graph, "-low");
835 if (firm_opt.enabled) {
836 timer_start(TV_ALL_OPT);
838 /* run reassociation first on all graphs BEFORE the architecture dependent optimizations
840 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
841 current_ir_graph = get_irp_irg(i);
843 timer_push(TV_REASSOCIATION);
844 optimize_reassociation(current_ir_graph);
846 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "reassoc");
847 CHECK_ONE(firm_opt.check_all, current_ir_graph);
850 /* enable architecture dependent optimizations */
851 arch_dep_set_opts((arch_dep_opts_t)
852 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
853 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
854 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
856 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
857 current_ir_graph = get_irp_irg(i);
860 set_opt_global_cse(1);
862 timer_push(TV_LOCAL_OPT);
863 optimize_graph_df(current_ir_graph);
865 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
867 CHECK_ONE(firm_opt.check_all, current_ir_graph);
870 timer_push(TV_CODE_PLACE);
871 place_code(current_ir_graph);
872 set_opt_global_cse(0);
874 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "place");
875 CHECK_ONE(firm_opt.check_all, current_ir_graph);
878 timer_push(TV_LOAD_STORE);
879 optimize_load_store(current_ir_graph);
881 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "ldst");
882 CHECK_ONE(firm_opt.check_all, current_ir_graph);
885 timer_push(TV_LOCAL_OPT);
886 optimize_graph_df(current_ir_graph);
888 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "lopt");
890 timer_push(TV_CF_OPT);
891 optimize_cf(current_ir_graph);
893 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "cf");
894 CHECK_ONE(firm_opt.check_all, current_ir_graph);
896 if (firm_opt.if_conversion) {
897 timer_push(TV_IF_CONV);
898 opt_if_conv(current_ir_graph, if_conv_info);
900 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "if");
901 CHECK_ONE(firm_opt.check_all, current_ir_graph);
903 timer_push(TV_LOCAL_OPT);
904 optimize_graph_df(current_ir_graph);
906 timer_push(TV_CF_OPT);
907 optimize_cf(current_ir_graph);
909 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, current_ir_graph, "after_if");
910 CHECK_ONE(firm_opt.check_all, current_ir_graph);
913 timer_stop(TV_ALL_OPT);
915 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
919 mark_private_methods();
921 /* set the phase to low */
922 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
923 set_irg_phase_low(get_irp_irg(i));
925 /* all graphs are lowered, set the irp phase to low */
926 set_irp_phase_state(phase_low);
928 if (firm_dump.statistic & STAT_FINAL) {
929 stat_dump_snapshot(input_filename, "final");
931 } /* do_firm_lowering */
934 * Initialize for the Firm-generating back end.
936 void gen_firm_init(void)
938 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;
982 /* OS option must be set to the backend */
983 switch (firm_opt.os_support) {
984 case OS_SUPPORT_MINGW:
985 firm_be_option("ia32-gasmode=mingw");
987 case OS_SUPPORT_MACHO:
988 firm_be_option("ia32-gasmode=macho");
990 case OS_SUPPORT_LINUX:
992 firm_be_option("ia32-gasmode=linux");
996 dbg_init(NULL, NULL, dbg_snprint);
997 edges_init_dbg(firm_opt.vrfy_edges);
998 //cbackend_set_debug_retrieve(dbg_retrieve);
1000 set_opt_precise_exc_context(firm_opt.precise_exc);
1001 set_opt_fragile_ops(firm_opt.fragile_ops);
1003 /* Sel node cannot produce NULL pointers */
1004 set_opt_sel_based_null_check_elim(1);
1006 /* dynamic dispatch works currently only if whole world scenarios */
1007 set_opt_dyn_meth_dispatch(0);
1009 arch_dep_init(arch_factory);
1011 /* do not run architecture dependent optimizations in building phase */
1012 arch_dep_set_opts(arch_dep_none);
1014 do_node_verification((firm_verification_t) firm_opt.vrfy);
1015 if (firm_dump.filter)
1016 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
1018 if (firm_opt.enabled) {
1020 set_opt_constant_folding(firm_opt.const_folding);
1021 set_opt_algebraic_simplification(firm_opt.const_folding);
1022 set_opt_cse(firm_opt.cse);
1023 set_opt_global_cse(0);
1024 set_opt_unreachable_code(1);
1025 set_opt_control_flow(firm_opt.control_flow);
1026 set_opt_control_flow_weak_simplification(1);
1027 set_opt_control_flow_strong_simplification(1);
1032 /* do not dump entity ld names */
1034 } /* gen_firm_init */
1037 * Called, after the Firm generation is completed,
1038 * do all optimizations and backend call here.
1040 * @param out a file handle for the output, may be NULL
1041 * @param input_filename the name of the (main) source file
1042 * @param c_mode non-zero if "C" was compiled
1043 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
1045 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
1049 firm_const_exists = new_firm_const_exists;
1051 /* the general for dumping option must be set, or the others will not work */
1053 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1055 dump_keepalive_edges(1);
1056 dump_consts_local(1);
1057 dump_dominator_information(1);
1058 dump_loop_information(0);
1060 if (!firm_dump.edge_labels)
1061 turn_off_edge_labels();
1063 if (firm_dump.all_types) {
1066 dump_class_hierarchy(0, "");
1067 dump_class_hierarchy(1, "-with-entities");
1071 /* finalize all graphs */
1072 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1073 ir_graph *irg = get_irp_irg(i);
1075 irg_finalize_cons(irg);
1076 DUMP_ONE(firm_dump.ir_graph, irg, "");
1078 /* verify the graph */
1079 timer_push(TV_VERIFY);
1080 irg_verify(irg, VRFY_ENFORCE_SSA);
1084 timer_push(TV_VERIFY);
1088 /* all graphs are finalized, set the irp phase to high */
1089 set_irp_phase_state(phase_high);
1091 /* BEWARE: kill unreachable code before doing compound lowering */
1092 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1093 ir_graph *irg = get_irp_irg(i);
1097 /* lower all compound call return values */
1098 lower_compound_params();
1100 /* computes the sizes of all types that are still not computed */
1101 compute_type_sizes();
1103 /* lower copyb nodes */
1104 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1105 ir_graph *irg = get_irp_irg(i);
1106 lower_CopyB(irg, 128, 4);
1109 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1110 stat_dump_snapshot(input_filename, "noopt");
1113 if (firm_opt.enabled)
1114 do_firm_optimizations(input_filename);
1116 if (firm_dump.gen_firm_asm) {
1117 timer_push(TV_FIRM_ASM);
1118 gen_Firm_assembler(input_filename);
1124 do_firm_lowering(input_filename);
1126 /* computes the sizes of all frame types */
1127 compute_frame_type_sizes();
1129 /* set the phase to low */
1130 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1131 set_irg_phase_low(get_irp_irg(i));
1135 /** SIMD Optimization Extensions **/
1137 /* Pattern creation step. No code has to be generated, so
1138 exit after pattern creation */
1139 if (firm_ext_grs.create_pattern) {
1140 ext_grs_create_pattern();
1144 /* SIMD optimization step. Uses graph patterns to find
1145 rich instructions and rewrite */
1146 if (firm_ext_grs.simd_opt)
1150 if (firm_dump.statistic & STAT_FINAL_IR)
1151 stat_dump_snapshot(input_filename, "final-ir");
1153 /* run the code generator */
1154 if (firm_be_opt.selection != BE_NONE)
1155 do_codegen(out, input_filename);
1157 if (firm_dump.statistic & STAT_FINAL)
1158 stat_dump_snapshot(input_filename, "final");
1161 if (firm_opt.ycomp_dbg)
1162 firm_finish_ycomp_debugger();
1164 } /* gen_firm_finish */
1167 * Do very early initializations
1169 void firm_early_init(void) {
1170 /* arg: need this here for command line options */
1172 firm_init_options(NULL, 0, NULL);
1173 } /* firm_early_init */