3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2009 Michael Beck beck@ipd.info.uni-karlsruhe.de
17 #include <libfirm/firm.h>
20 #include "firm_codegen.h"
21 #include "firm_cmdline.h"
22 #include "firm_timing.h"
25 #if defined(_DEBUG) || defined(FIRM_DEBUG)
26 #define DBG(x) dbg_printf x
28 #define DBG(x) ((void)0)
29 #endif /* _DEBUG || FIRM_DEBUG */
31 static ir_timer_t *t_vcg_dump;
32 static ir_timer_t *t_verify;
33 static ir_timer_t *t_all_opt;
34 static bool do_irg_opt(ir_graph *irg, const char *name);
36 /** dump all the graphs depending on cond */
38 static void dump_all(const char *suffix) {
39 if (firm_dump.ir_graph) {
40 timer_push(t_vcg_dump);
41 if (firm_dump.no_blocks)
42 dump_all_ir_graphs(dump_ir_graph, suffix);
43 else if (firm_dump.extbb)
44 dump_all_ir_graphs(dump_ir_extblock_graph, suffix);
46 dump_all_ir_graphs(dump_ir_block_graph, suffix);
47 timer_pop(t_vcg_dump);
51 /* set by the backend parameters */
52 static const ir_settings_arch_dep_t *ad_param = NULL;
53 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
54 static void *create_intrinsic_ctx = NULL;
55 static const ir_settings_if_conv_t *if_conv_info = NULL;
57 /* entities of runtime functions */
58 ir_entity_ptr rts_entities[rts_max];
61 * factory for setting architecture dependent parameters
63 static const ir_settings_arch_dep_t *arch_factory(void)
65 static const ir_settings_arch_dep_t param = {
66 1, /* also use subs */
67 4, /* maximum shifts */
68 31, /* maximum shift amount */
69 NULL, /* use default evaluator */
73 32 /* Mulh allowed up to 32 bit */
76 return ad_param ? ad_param : ¶m;
80 * Map runtime functions.
82 static void rts_map(void) {
84 ir_entity_ptr *ent; /**< address of the rts entity */
85 i_mapper_func func; /**< mapper function. */
88 { &rts_entities[rts_abs], i_mapper_abs },
89 { &rts_entities[rts_labs], i_mapper_abs },
90 { &rts_entities[rts_llabs], i_mapper_abs },
91 { &rts_entities[rts_imaxabs], i_mapper_abs },
93 /* double -> double */
94 { &rts_entities[rts_fabs], i_mapper_abs },
95 { &rts_entities[rts_sqrt], i_mapper_sqrt },
96 { &rts_entities[rts_cbrt], i_mapper_cbrt },
97 { &rts_entities[rts_pow], i_mapper_pow },
98 { &rts_entities[rts_exp], i_mapper_exp },
99 { &rts_entities[rts_exp2], i_mapper_exp },
100 { &rts_entities[rts_exp10], i_mapper_exp },
101 { &rts_entities[rts_log], i_mapper_log },
102 { &rts_entities[rts_log2], i_mapper_log2 },
103 { &rts_entities[rts_log10], i_mapper_log10 },
104 { &rts_entities[rts_sin], i_mapper_sin },
105 { &rts_entities[rts_cos], i_mapper_cos },
106 { &rts_entities[rts_tan], i_mapper_tan },
107 { &rts_entities[rts_asin], i_mapper_asin },
108 { &rts_entities[rts_acos], i_mapper_acos },
109 { &rts_entities[rts_atan], i_mapper_atan },
110 { &rts_entities[rts_sinh], i_mapper_sinh },
111 { &rts_entities[rts_cosh], i_mapper_cosh },
112 { &rts_entities[rts_tanh], i_mapper_tanh },
115 { &rts_entities[rts_fabsf], i_mapper_abs },
116 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
117 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
118 { &rts_entities[rts_powf], i_mapper_pow },
119 { &rts_entities[rts_expf], i_mapper_exp },
120 { &rts_entities[rts_exp2f], i_mapper_exp },
121 { &rts_entities[rts_exp10f], i_mapper_exp },
122 { &rts_entities[rts_logf], i_mapper_log },
123 { &rts_entities[rts_log2f], i_mapper_log2 },
124 { &rts_entities[rts_log10f], i_mapper_log10 },
125 { &rts_entities[rts_sinf], i_mapper_sin },
126 { &rts_entities[rts_cosf], i_mapper_cos },
127 { &rts_entities[rts_tanf], i_mapper_tan },
128 { &rts_entities[rts_asinf], i_mapper_asin },
129 { &rts_entities[rts_acosf], i_mapper_acos },
130 { &rts_entities[rts_atanf], i_mapper_atan },
131 { &rts_entities[rts_sinhf], i_mapper_sinh },
132 { &rts_entities[rts_coshf], i_mapper_cosh },
133 { &rts_entities[rts_tanhf], i_mapper_tanh },
135 /* long double -> long double */
136 { &rts_entities[rts_fabsl], i_mapper_abs },
137 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
138 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
139 { &rts_entities[rts_powl], i_mapper_pow },
140 { &rts_entities[rts_expl], i_mapper_exp },
141 { &rts_entities[rts_exp2l], i_mapper_exp },
142 { &rts_entities[rts_exp10l], i_mapper_exp },
143 { &rts_entities[rts_logl], i_mapper_log },
144 { &rts_entities[rts_log2l], i_mapper_log2 },
145 { &rts_entities[rts_log10l], i_mapper_log10 },
146 { &rts_entities[rts_sinl], i_mapper_sin },
147 { &rts_entities[rts_cosl], i_mapper_cos },
148 { &rts_entities[rts_tanl], i_mapper_tan },
149 { &rts_entities[rts_asinl], i_mapper_asin },
150 { &rts_entities[rts_acosl], i_mapper_acos },
151 { &rts_entities[rts_atanl], i_mapper_atan },
152 { &rts_entities[rts_sinhl], i_mapper_sinh },
153 { &rts_entities[rts_coshl], i_mapper_cosh },
154 { &rts_entities[rts_tanhl], i_mapper_tanh },
157 { &rts_entities[rts_strcmp], i_mapper_strcmp },
158 { &rts_entities[rts_strncmp], i_mapper_strncmp },
159 { &rts_entities[rts_strcpy], i_mapper_strcpy },
160 { &rts_entities[rts_strlen], i_mapper_strlen },
161 { &rts_entities[rts_memcpy], i_mapper_memcpy },
162 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
163 { &rts_entities[rts_memmove], i_mapper_memmove },
164 { &rts_entities[rts_memset], i_mapper_memset },
165 { &rts_entities[rts_memcmp], i_mapper_memcmp }
167 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
170 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
171 if (*mapper[i].ent != NULL) {
172 rec[n_map].i_call.kind = INTRINSIC_CALL;
173 rec[n_map].i_call.i_ent = *mapper[i].ent;
174 rec[n_map].i_call.i_mapper = mapper[i].func;
175 rec[n_map].i_call.ctx = NULL;
176 rec[n_map].i_call.link = NULL;
182 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
185 static int *irg_dump_no;
187 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
190 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++,
193 timer_push(t_vcg_dump);
194 if (firm_dump.no_blocks)
195 dump_ir_graph(irg, name);
196 else if (firm_dump.extbb)
197 dump_ir_extblock_graph(irg, name);
199 dump_ir_block_graph(irg, name);
200 timer_pop(t_vcg_dump);
203 static void remove_unused_functions(void)
205 ir_entity **keep_methods;
208 /* Analysis that finds the free methods,
209 i.e. methods that are dereferenced.
210 Optimizes polymorphic calls :-). */
211 cgana(&arr_len, &keep_methods);
213 /* Remove methods that are never called. */
214 gc_irgs(arr_len, keep_methods);
218 static int firm_const_exists;
220 static void do_optimize_funccalls(void)
222 optimize_funccalls(firm_const_exists, NULL);
225 static void do_gcse(ir_graph *irg)
227 set_opt_global_cse(1);
228 optimize_graph_df(irg);
230 set_opt_global_cse(0);
233 static void do_lower_highlevel(ir_graph *irg)
235 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
238 static void do_if_conv(ir_graph *irg)
240 opt_if_conv(irg, if_conv_info);
243 static void do_stred(ir_graph *irg)
245 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
248 static void after_inline_opt(ir_graph *irg)
250 do_irg_opt(irg, "scalar-replace");
251 do_irg_opt(irg, "local");
252 do_irg_opt(irg, "control-flow");
253 do_irg_opt(irg, "combo");
256 static void do_inline(void)
258 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
262 static void do_cloning(void)
264 proc_cloning((float) firm_opt.clone_threshold);
267 static void do_lower_switch(ir_graph *irg)
269 lower_switch(irg, firm_opt.spare_size);
272 static void do_lower_dw_ops(void)
274 lwrdw_param_t init = {
277 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
278 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
279 get_atomic_mode(ATOMIC_TYPE_INT),
280 get_atomic_mode(ATOMIC_TYPE_UINT),
281 def_create_intrinsic_fkt,
285 if (arch_create_intrinsic) {
286 init.create_intrinsic = arch_create_intrinsic;
287 init.ctx = create_intrinsic_ctx;
292 typedef enum opt_target {
293 OPT_TARGET_IRG, /**< optimization function works on a single graph */
294 OPT_TARGET_IRP /**< optimization function works on the complete program */
297 typedef enum opt_flags {
299 OPT_FLAG_ENABLED = 1 << 0, /**< enable the optimization */
300 OPT_FLAG_NO_DUMP = 1 << 1, /**< don't dump after transformation */
301 OPT_FLAG_NO_VERIFY = 1 << 2, /**< don't verify after transformation */
302 OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
303 -foptions for this transformation */
306 typedef void (*transform_irg_func)(ir_graph *irg);
307 typedef void (*transform_irp_func)(void);
308 typedef void (*func_ptr_t)(void);
314 const char *description;
318 static opt_config_t opts[] = {
319 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, "optimization of known library functions", OPT_FLAG_HIDE_OPTIONS },
320 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, "combined CCE, UCE and GVN", OPT_FLAG_NONE},
321 { OPT_TARGET_IRG, "control-flow", (func_ptr_t) optimize_cf, "optimization of control-flow", OPT_FLAG_HIDE_OPTIONS },
322 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, "local graph optimizations", OPT_FLAG_HIDE_OPTIONS },
323 { OPT_TARGET_IRP, "remove-unused", (func_ptr_t) remove_unused_functions, "removal of unused functions", OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
324 { OPT_TARGET_IRP, "opt-tail-rec", (func_ptr_t) opt_tail_recursion, "tail-recursion eliminiation", OPT_FLAG_NONE },
325 { OPT_TARGET_IRP, "opt-func-call", (func_ptr_t) do_optimize_funccalls, "function call optimization", OPT_FLAG_NONE },
326 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, "lowering", OPT_FLAG_HIDE_OPTIONS },
327 { OPT_TARGET_IRP, "lower-const", (func_ptr_t) lower_const_code, "lowering of constant code", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
328 { OPT_TARGET_IRP, "lower-dw", (func_ptr_t) do_lower_dw_ops, "lowering of doubleword operations", OPT_FLAG_HIDE_OPTIONS },
329 { OPT_TARGET_IRG, "lower-switch", (func_ptr_t) do_lower_switch, "switch lowering", OPT_FLAG_HIDE_OPTIONS },
330 { OPT_TARGET_IRG, "one-return", (func_ptr_t) normalize_one_return, "normalisation to 1 return", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
331 { OPT_TARGET_IRG, "scalar-replace", (func_ptr_t) scalar_replacement_opt, "scalar replacement", OPT_FLAG_NONE },
332 { OPT_TARGET_IRG, "reassociation", (func_ptr_t) optimize_reassociation, "reassociation", OPT_FLAG_NONE },
333 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, "global common subexpression elimination", OPT_FLAG_NONE },
334 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, "code placement", OPT_FLAG_NONE },
335 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, "confirm optimisation", OPT_FLAG_HIDE_OPTIONS },
336 { OPT_TARGET_IRG, "opt-load-store", (func_ptr_t) optimize_load_store, "load store optimization", OPT_FLAG_NONE },
337 { OPT_TARGET_IRG, "parallelize-mem", (func_ptr_t) opt_parallelize_mem, "parallelize memory", OPT_FLAG_NONE },
338 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, "conv node elimination", OPT_FLAG_NONE },
339 { OPT_TARGET_IRG, "thread-jumps", (func_ptr_t) opt_jumpthreading, "path-sensitive jumpthreading", OPT_FLAG_NONE },
340 { OPT_TARGET_IRG, "remove-confirms", (func_ptr_t) remove_confirms, "confirm removal", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
341 { OPT_TARGET_IRG, "gvn-pre", (func_ptr_t) do_gvn_pre, "global value numbering partial redundancy elimination", OPT_FLAG_NONE },
342 { OPT_TARGET_IRG, "if-conversion", (func_ptr_t) do_if_conv, "if-conversion", OPT_FLAG_NONE },
343 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, "bool simplification", OPT_FLAG_NONE },
344 { OPT_TARGET_IRG, "shape-blocks", (func_ptr_t) shape_blocks, "block shaping", OPT_FLAG_NONE },
345 { OPT_TARGET_IRG, "ivopts", (func_ptr_t) do_stred, "induction variable strength reduction", OPT_FLAG_NONE },
346 { OPT_TARGET_IRG, "remove-phi-cycles", (func_ptr_t) remove_phi_cycles, "removal of phi cycles", OPT_FLAG_HIDE_OPTIONS },
347 { OPT_TARGET_IRG, "dead", (func_ptr_t) dead_node_elimination, "dead node elimination", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
348 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, "inlining", OPT_FLAG_NONE },
349 { OPT_TARGET_IRP, "opt-proc-clone", (func_ptr_t) do_cloning, "procedure cloning", OPT_FLAG_NONE },
350 { OPT_TARGET_IRG, "invert-loops", (func_ptr_t) do_loop_inversion, "loop inversion", OPT_FLAG_NONE },
351 { OPT_TARGET_IRG, "peel-loops", (func_ptr_t) do_loop_peeling, "loop peeling", OPT_FLAG_NONE },
353 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
354 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
356 static opt_config_t *get_opt(const char *name)
359 for (i = 0; i < n_opts; ++i) {
360 opt_config_t *config = &opts[i];
361 if (strcmp(config->name, name) == 0)
368 static void set_opt_enabled(const char *name, bool enabled)
370 opt_config_t *config = get_opt(name);
371 config->flags = (config->flags & ~OPT_FLAG_ENABLED)
372 | (enabled ? OPT_FLAG_ENABLED : 0);
375 static bool get_opt_enabled(const char *name)
377 opt_config_t *config = get_opt(name);
378 return (config->flags & OPT_FLAG_ENABLED) != 0;
382 * perform an optimisation on a single graph
384 * @return true if something changed, false otherwise
386 static bool do_irg_opt(ir_graph *irg, const char *name)
388 transform_irg_func func;
390 opt_config_t *config = get_opt(name);
391 size_t n = config - opts;
392 assert(config != NULL);
393 assert(config->target == OPT_TARGET_IRG);
394 if (! (config->flags & OPT_FLAG_ENABLED))
398 old_irg = current_ir_graph;
399 current_ir_graph = irg;
401 func = (transform_irg_func) config->func;
403 timer_push(timers[n]);
405 timer_pop(timers[n]);
407 if (firm_dump.all_phases && firm_dump.ir_graph) {
408 dump_graph_count(irg, name);
411 if (firm_opt.check_all) {
412 timer_push(t_verify);
413 irg_verify(irg, VRFY_ENFORCE_SSA);
417 current_ir_graph = old_irg;
421 static void do_irp_opt(const char *name)
423 transform_irp_func func;
424 opt_config_t *config = get_opt(name);
425 size_t n = config - opts;
426 assert(config->target == OPT_TARGET_IRP);
427 if (! (config->flags & OPT_FLAG_ENABLED))
430 func = (transform_irp_func) config->func;
432 timer_push(timers[n]);
434 timer_pop(timers[n]);
436 if (firm_dump.ir_graph && firm_dump.all_phases) {
438 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
439 ir_graph *irg = get_irp_irg(i);
440 dump_graph_count(irg, name);
444 if (firm_opt.check_all) {
446 timer_push(t_verify);
447 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
448 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
455 * Enable transformations which should be always safe (and cheap) to perform
457 static void enable_safe_defaults(void)
459 set_opt_enabled("remove-unused", true);
460 set_opt_enabled("opt-tail-rec", true);
461 set_opt_enabled("opt-func-call", true);
462 set_opt_enabled("reassociation", true);
463 set_opt_enabled("control-flow", true);
464 set_opt_enabled("local", true);
465 set_opt_enabled("lower-const", true);
466 set_opt_enabled("scalar-replace", true);
467 set_opt_enabled("place", true);
468 set_opt_enabled("confirm", true);
469 set_opt_enabled("opt-load-store", true);
470 set_opt_enabled("lower", true);
471 set_opt_enabled("deconv", true);
472 set_opt_enabled("remove-confirms", true);
473 set_opt_enabled("ivopts", true);
474 set_opt_enabled("dead", true);
475 set_opt_enabled("lower-switch", true);
476 set_opt_enabled("remove-phi-cycles", true);
480 * run all the Firm optimizations
482 * @param input_filename the name of the (main) source file
484 static void do_firm_optimizations(const char *input_filename)
489 set_opt_alias_analysis(firm_opt.alias_analysis);
491 aa_opt = aa_opt_no_opt;
492 if (firm_opt.strict_alias)
493 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
494 if (firm_opt.no_alias)
495 aa_opt = aa_opt_no_alias;
497 set_irp_memory_disambiguator_options(aa_opt);
499 /* parameter passing code should set them directly sometime... */
500 set_opt_enabled("rts", !firm_opt.freestanding);
501 set_opt_enabled("gcse", firm_opt.gcse);
502 set_opt_enabled("place", !firm_opt.gcse);
503 set_opt_enabled("confirm", firm_opt.confirm);
504 set_opt_enabled("remove-confirms", firm_opt.confirm);
506 /* osr supersedes remove_phi_cycles */
507 if (get_opt_enabled("ivopts"))
508 set_opt_enabled("remove-phi-cycles", false);
510 timer_start(t_all_opt);
514 /* first step: kill dead code */
515 for (i = 0; i < get_irp_n_irgs(); i++) {
516 ir_graph *irg = get_irp_irg(i);
517 do_irg_opt(irg, "combo");
518 do_irg_opt(irg, "local");
519 do_irg_opt(irg, "control-flow");
522 do_irp_opt("remove-unused");
523 do_irp_opt("opt-tail-rec");
524 do_irp_opt("opt-func-call");
525 do_irp_opt("lower-const");
527 for (i = 0; i < get_irp_n_irgs(); i++) {
528 ir_graph *irg = get_irp_irg(i);
530 do_irg_opt(irg, "scalar-replace");
531 do_irg_opt(irg, "invert-loops");
532 do_irg_opt(irg, "local");
533 do_irg_opt(irg, "reassociation");
534 do_irg_opt(irg, "local");
535 do_irg_opt(irg, "gcse");
537 if (firm_opt.confirm) {
538 /* Confirm construction currently can only handle blocks with only
539 one control flow predecessor. Calling optimize_cf here removes
540 Bad predecessors and help the optimization of switch constructs.
542 do_irg_opt(irg, "control-flow");
543 do_irg_opt(irg, "confirm");
544 do_irg_opt(irg, "local");
547 do_irg_opt(irg, "control-flow");
548 do_irg_opt(irg, "opt-load-store");
549 do_irg_opt(irg, "lower");
550 do_irg_opt(irg, "deconv");
551 do_irg_opt(irg, "thread-jumps");
552 do_irg_opt(irg, "remove-confirms");
553 do_irg_opt(irg, "gvn-pre");
554 do_irg_opt(irg, "place");
555 do_irg_opt(irg, "control-flow");
557 if (do_irg_opt(irg, "if-conversion")) {
558 do_irg_opt(irg, "local");
559 do_irg_opt(irg, "control-flow");
562 do_irg_opt(irg, "bool");
563 do_irg_opt(irg, "shape-blocks");
564 do_irg_opt(irg, "lower-switch");
565 do_irg_opt(irg, "ivopts");
566 do_irg_opt(irg, "local");
567 do_irg_opt(irg, "dead");
570 do_irp_opt("inline");
571 do_irp_opt("opt-proc-clone");
573 for (i = 0; i < get_irp_n_irgs(); i++) {
574 ir_graph *irg = get_irp_irg(i);
575 do_irg_opt(irg, "local");
576 do_irg_opt(irg, "control-flow");
577 do_irg_opt(irg, "thread-jumps");
578 do_irg_opt(irg, "local");
579 do_irg_opt(irg, "control-flow");
582 if (firm_dump.ir_graph) {
583 /* recompute backedges for nicer dumps */
584 for (i = 0; i < get_irp_n_irgs(); i++)
585 construct_cf_backedges(get_irp_irg(i));
588 do_irp_opt("remove-unused");
592 if (firm_dump.statistic & STAT_AFTER_OPT)
593 stat_dump_snapshot(input_filename, "opt");
595 timer_stop(t_all_opt);
599 * compute the size of a type (do implicit lowering)
601 * @param ty a Firm type
603 static int compute_type_size(ir_type *ty)
605 optimization_state_t state;
606 unsigned align_all = 1;
607 int n, size = 0, set = 0;
610 if (get_type_state(ty) == layout_fixed) {
611 /* do not layout already layouted types again */
615 if (is_Method_type(ty) || ty == get_glob_type()) {
616 /* no need for size calculation for method types or the global type */
620 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
622 switch (get_type_tpop_code(ty)) {
625 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
626 ir_entity *ent = get_compound_member(ty, i);
627 ir_type *ent_ty = get_entity_type(ent);
628 unsigned align, misalign;
630 /* inner functions do not expand the frame */
631 if (is_Method_type(ent_ty) && is_frame_type(ty))
634 /* compute member types */
635 if (! compute_type_size(ent_ty))
638 align = get_type_alignment_bytes(ent_ty);
639 align_all = align > align_all ? align : align_all;
640 misalign = (align ? size % align : 0);
641 size += (misalign ? align - misalign : 0);
643 set_entity_offset(ent, size);
644 size += get_type_size_bytes(ent_ty);
646 DBG((" member %s %s -> (size: %u, align: %u)\n",
647 get_type_name(ent_ty), get_entity_name(ent),
648 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
650 if (align_all > 0 && size % align_all) {
651 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
652 size += align_all - (size % align_all);
653 DBG(("correcting type-size to %d\n", size));
655 set_type_alignment_bytes(ty, align_all);
660 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
661 ir_entity *ent = get_union_member(ty, i);
663 if (! compute_type_size(get_entity_type(ent)))
665 s = get_type_size_bytes(get_entity_type(ent));
667 set_entity_offset(ent, 0);
668 size = (s > size ? s : size);
674 dims = get_array_n_dimensions(ty);
676 if (! compute_type_size(get_array_element_type(ty)))
681 save_optimization_state(&state);
683 set_opt_constant_folding(1);
684 set_opt_algebraic_simplification(1);
686 for (i = 0; i < dims; ++i) {
687 ir_node *lower = get_array_lower_bound(ty, i);
688 ir_node *upper = get_array_upper_bound(ty, i);
689 ir_graph *rem = current_ir_graph;
690 tarval *tv_lower, *tv_upper;
691 long val_lower, val_upper;
693 current_ir_graph = get_const_code_irg();
694 local_optimize_node(lower);
695 local_optimize_node(upper);
696 current_ir_graph = rem;
698 tv_lower = computed_value(lower);
699 tv_upper = computed_value(upper);
701 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
703 * we cannot calculate the size of this array yet, it
704 * even might be unknown until the end, like argv[]
706 restore_optimization_state(&state);
710 val_upper = get_tarval_long(tv_upper);
711 val_lower = get_tarval_long(tv_lower);
712 size *= val_upper - val_lower;
714 restore_optimization_state(&state);
716 DBG(("array %s -> (elements: %d, element type size: %d)\n",
718 size, get_type_size_bytes(get_array_element_type(ty))));
719 size *= get_type_size_bytes(get_array_element_type(ty));
728 set_type_size_bytes(ty, size);
729 set_type_state(ty, layout_fixed);
732 DBG(("size: %d\n", get_type_size_bytes(ty)));
738 * layout all non-frame types of the Firm graph
740 static void compute_type_sizes(void)
745 /* all non-frame other types */
746 for (i = get_irp_n_types() - 1; i >= 0; --i) {
747 tp = get_irp_type(i);
748 compute_type_size(tp);
750 if (is_Method_type(tp)) {
751 tp = get_method_value_res_type(tp);
754 /* we have a value result type for this method, lower */
755 compute_type_size(tp);
762 * layout all frame-types of the Firm graph
764 static void compute_frame_type_sizes(void)
769 /* all frame types */
770 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
771 irg = get_irp_irg(i);
772 /* do not optimize away variables in debug mode */
773 if (firm_opt.debug_mode == DBG_MODE_NONE)
775 compute_type_size(get_irg_frame_type(irg));
782 * @param input_filename the name of the (main) source file
784 static void do_firm_lowering(const char *input_filename)
788 do_irp_opt("lower-dw");
790 if (firm_dump.statistic & STAT_AFTER_LOWER)
791 stat_dump_snapshot(input_filename, "low");
795 if (firm_opt.enabled) {
796 timer_start(t_all_opt);
798 /* run reassociation first on all graphs BEFORE the architecture
799 dependent optimizations are enabled */
800 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
801 ir_graph *irg = get_irp_irg(i);
802 do_irg_opt(irg, "reassociation");
805 /* enable architecture dependent optimizations */
806 arch_dep_set_opts((arch_dep_opts_t)
807 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
808 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
809 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
811 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
812 ir_graph *irg = get_irp_irg(i);
814 current_ir_graph = irg;
816 do_irg_opt(irg, "local");
817 do_irg_opt(irg, "gcse");
818 do_irg_opt(irg, "opt-load-store");
819 do_irg_opt(irg, "local");
820 do_irg_opt(irg, "control-flow");
822 if (do_irg_opt(irg, "if-conversion")) {
823 do_irg_opt(irg, "local");
824 do_irg_opt(irg, "control-flow");
827 do_irg_opt(irg, "parallelize-mem");
829 timer_stop(t_all_opt);
831 dump_all("-low-opt");
835 mark_private_methods();
837 /* set the phase to low */
838 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
839 set_irg_phase_low(get_irp_irg(i));
841 /* all graphs are lowered, set the irp phase to low */
842 set_irp_phase_state(phase_low);
844 if (firm_dump.statistic & STAT_FINAL) {
845 stat_dump_snapshot(input_filename, "final");
850 * Initialize for the Firm-generating back end.
852 void gen_firm_init(void)
854 firm_parameter_t params;
855 unsigned pattern = 0;
858 for (i = 0; i < n_opts; ++i) {
859 timers[i] = ir_timer_new();
860 timer_register(timers[i], opts[i].description);
862 t_verify = ir_timer_new();
863 timer_register(t_verify, "Firm: verify pass");
864 t_vcg_dump = ir_timer_new();
865 timer_register(t_vcg_dump, "Firm: vcg dumping");
866 t_all_opt = ir_timer_new();
867 timer_register(t_all_opt, "Firm: all optimizations");
869 if (firm_dump.stat_pattern)
870 pattern |= FIRMSTAT_PATTERN_ENABLED;
872 if (firm_dump.stat_dag)
873 pattern |= FIRMSTAT_COUNT_DAG;
875 memset(¶ms, 0, sizeof(params));
876 params.size = sizeof(params);
877 params.enable_statistics = firm_dump.statistic == STAT_NONE ? 0 :
878 FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP | FIRMSTAT_COUNT_CONSTS
880 params.initialize_local_func = uninitialized_local_var;
881 params.cc_mask = 0; /* no regparam, cdecl */
885 if (firm_be_opt.selection == BE_FIRM_BE) {
886 const backend_params *be_params = be_get_backend_param();
888 if (be_params->do_dw_lowering)
889 set_opt_enabled("lower-dw", true);
891 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
892 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
894 ad_param = be_params->dep_param;
895 if_conv_info = be_params->if_conv_info;
898 edges_init_dbg(firm_opt.vrfy_edges);
900 /* Sel node cannot produce NULL pointers */
901 set_opt_sel_based_null_check_elim(1);
903 /* dynamic dispatch works currently only if whole world scenarios */
904 set_opt_dyn_meth_dispatch(0);
906 arch_dep_init(arch_factory);
908 /* do not run architecture dependent optimizations in building phase */
909 arch_dep_set_opts(arch_dep_none);
911 do_node_verification((firm_verification_t) firm_opt.vrfy);
912 if (firm_dump.filter)
913 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
915 if (firm_opt.enabled) {
917 set_opt_constant_folding(firm_opt.const_folding);
918 set_opt_algebraic_simplification(firm_opt.const_folding);
919 set_opt_cse(firm_opt.cse);
920 set_opt_global_cse(0);
921 set_opt_unreachable_code(1);
922 set_opt_control_flow(firm_opt.control_flow);
923 set_opt_control_flow_weak_simplification(1);
924 set_opt_control_flow_strong_simplification(1);
929 /* do not dump entity ld names */
934 * Called, after the Firm generation is completed,
935 * do all optimizations and backend call here.
937 * @param out a file handle for the output, may be NULL
938 * @param input_filename the name of the (main) source file
939 * @param c_mode non-zero if "C" was compiled
940 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
942 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode,
943 int new_firm_const_exists)
948 if (firm_opt.enable_statev) {
950 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
951 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
952 ir_stat_ev_compilation_unit(input_filename);
956 firm_const_exists = new_firm_const_exists;
958 /* the general for dumping option must be set, or the others will not work*/
960 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
962 dump_keepalive_edges(1);
963 dump_consts_local(1);
964 dump_dominator_information(1);
965 dump_loop_information(0);
967 if (!firm_dump.edge_labels)
968 turn_off_edge_labels();
970 /* FIXME: cloning might ADD new graphs. */
971 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
973 if (firm_dump.all_types) {
976 dump_class_hierarchy(0, "");
977 dump_class_hierarchy(1, "-with-entities");
981 /* finalize all graphs */
982 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
983 ir_graph *irg = get_irp_irg(i);
984 irg_finalize_cons(irg);
988 timer_push(t_verify);
992 /* all graphs are finalized, set the irp phase to high */
993 set_irp_phase_state(phase_high);
995 /* BEWARE: kill unreachable code before doing compound lowering */
996 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
997 ir_graph *irg = get_irp_irg(i);
998 do_irg_opt(irg, "control-flow");
1001 /* lower all compound call return values */
1002 lower_compound_params();
1004 /* computes the sizes of all types that are still not computed */
1005 compute_type_sizes();
1007 /* lower copyb nodes */
1008 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1009 ir_graph *irg = get_irp_irg(i);
1010 lower_CopyB(irg, 128, 4);
1013 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1014 stat_dump_snapshot(input_filename, "noopt");
1017 if (firm_opt.enabled)
1018 do_firm_optimizations(input_filename);
1021 do_firm_lowering(input_filename);
1023 /* computes the sizes of all frame types */
1024 compute_frame_type_sizes();
1026 /* set the phase to low */
1027 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1028 set_irg_phase_low(get_irp_irg(i));
1030 if (firm_dump.statistic & STAT_FINAL_IR)
1031 stat_dump_snapshot(input_filename, "final-ir");
1033 /* run the code generator */
1034 if (firm_be_opt.selection != BE_NONE)
1035 do_codegen(out, input_filename);
1037 if (firm_dump.statistic & STAT_FINAL)
1038 stat_dump_snapshot(input_filename, "final");
1041 int firm_opt_option(const char *opt)
1044 if (strncmp(opt, "no-", 3) == 0) {
1049 opt_config_t *config = get_opt(opt);
1050 if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
1053 config->flags &= ~OPT_FLAG_ENABLED;
1054 config->flags |= enable ? OPT_FLAG_ENABLED : 0;
1058 void firm_opt_option_help(void)
1062 for (i = 0; i < n_opts; ++i) {
1066 const opt_config_t *config = &opts[i];
1067 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
1070 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
1071 print_option_help(config->name, buf2);
1072 snprintf(buf, sizeof(buf), "no-%s", config->name);
1073 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
1074 print_option_help(buf, buf2);
1079 * Do very early initializations
1081 void firm_early_init(void)
1083 /* arg: need this here for command line options */
1086 enable_safe_defaults();