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 */
37 #define DUMP_ALL(cond, suffix) \
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 /** dump all control flow graphs depending on cond */
52 #define DUMP_ALL_CFG(cond, suffix) \
55 timer_push(t_vcg_dump); \
56 dump_all_ir_graphs(dump_cfg, suffix); \
57 timer_pop(t_vcg_dump); \
61 /** dump graphs irg depending on cond */
62 #define DUMP_ONE(cond, irg, suffix) \
65 timer_push(t_vcg_dump); \
66 if (firm_dump.no_blocks) \
67 dump_ir_graph(irg, suffix); \
68 else if (firm_dump.extbb) \
69 dump_ir_extblock_graph(irg, suffix); \
71 dump_ir_block_graph(irg, suffix); \
72 timer_pop(t_vcg_dump); \
76 /** dump control flow graph irg depending on cond */
77 #define DUMP_ONE_CFG(cond, irg, suffix) \
80 timer_push(t_vcg_dump); \
81 dump_cfg(irg, suffix); \
82 timer_pop(t_vcg_dump); \
86 /* set by the backend parameters */
87 static const ir_settings_arch_dep_t *ad_param = NULL;
88 static create_intrinsic_fkt *arch_create_intrinsic = NULL;
89 static void *create_intrinsic_ctx = NULL;
90 static const ir_settings_if_conv_t *if_conv_info = NULL;
92 /* entities of runtime functions */
93 ir_entity_ptr rts_entities[rts_max];
96 * factory for setting architecture dependent parameters
98 static const ir_settings_arch_dep_t *arch_factory(void)
100 static const ir_settings_arch_dep_t param = {
101 1, /* also use subs */
102 4, /* maximum shifts */
103 31, /* maximum shift amount */
104 NULL, /* use default evaluator */
108 32 /* Mulh allowed up to 32 bit */
111 return ad_param ? ad_param : ¶m;
115 * Map runtime functions.
117 static void rts_map(void) {
118 static const struct {
119 ir_entity_ptr *ent; /**< address of the rts entity */
120 i_mapper_func func; /**< mapper function. */
123 { &rts_entities[rts_abs], i_mapper_abs },
124 { &rts_entities[rts_labs], i_mapper_abs },
125 { &rts_entities[rts_llabs], i_mapper_abs },
126 { &rts_entities[rts_imaxabs], i_mapper_abs },
128 /* double -> double */
129 { &rts_entities[rts_fabs], i_mapper_abs },
130 { &rts_entities[rts_sqrt], i_mapper_sqrt },
131 { &rts_entities[rts_cbrt], i_mapper_cbrt },
132 { &rts_entities[rts_pow], i_mapper_pow },
133 { &rts_entities[rts_exp], i_mapper_exp },
134 { &rts_entities[rts_exp2], i_mapper_exp },
135 { &rts_entities[rts_exp10], i_mapper_exp },
136 { &rts_entities[rts_log], i_mapper_log },
137 { &rts_entities[rts_log2], i_mapper_log2 },
138 { &rts_entities[rts_log10], i_mapper_log10 },
139 { &rts_entities[rts_sin], i_mapper_sin },
140 { &rts_entities[rts_cos], i_mapper_cos },
141 { &rts_entities[rts_tan], i_mapper_tan },
142 { &rts_entities[rts_asin], i_mapper_asin },
143 { &rts_entities[rts_acos], i_mapper_acos },
144 { &rts_entities[rts_atan], i_mapper_atan },
145 { &rts_entities[rts_sinh], i_mapper_sinh },
146 { &rts_entities[rts_cosh], i_mapper_cosh },
147 { &rts_entities[rts_tanh], i_mapper_tanh },
150 { &rts_entities[rts_fabsf], i_mapper_abs },
151 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
152 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
153 { &rts_entities[rts_powf], i_mapper_pow },
154 { &rts_entities[rts_expf], i_mapper_exp },
155 { &rts_entities[rts_exp2f], i_mapper_exp },
156 { &rts_entities[rts_exp10f], i_mapper_exp },
157 { &rts_entities[rts_logf], i_mapper_log },
158 { &rts_entities[rts_log2f], i_mapper_log2 },
159 { &rts_entities[rts_log10f], i_mapper_log10 },
160 { &rts_entities[rts_sinf], i_mapper_sin },
161 { &rts_entities[rts_cosf], i_mapper_cos },
162 { &rts_entities[rts_tanf], i_mapper_tan },
163 { &rts_entities[rts_asinf], i_mapper_asin },
164 { &rts_entities[rts_acosf], i_mapper_acos },
165 { &rts_entities[rts_atanf], i_mapper_atan },
166 { &rts_entities[rts_sinhf], i_mapper_sinh },
167 { &rts_entities[rts_coshf], i_mapper_cosh },
168 { &rts_entities[rts_tanhf], i_mapper_tanh },
170 /* long double -> long double */
171 { &rts_entities[rts_fabsl], i_mapper_abs },
172 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
173 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
174 { &rts_entities[rts_powl], i_mapper_pow },
175 { &rts_entities[rts_expl], i_mapper_exp },
176 { &rts_entities[rts_exp2l], i_mapper_exp },
177 { &rts_entities[rts_exp10l], i_mapper_exp },
178 { &rts_entities[rts_logl], i_mapper_log },
179 { &rts_entities[rts_log2l], i_mapper_log2 },
180 { &rts_entities[rts_log10l], i_mapper_log10 },
181 { &rts_entities[rts_sinl], i_mapper_sin },
182 { &rts_entities[rts_cosl], i_mapper_cos },
183 { &rts_entities[rts_tanl], i_mapper_tan },
184 { &rts_entities[rts_asinl], i_mapper_asin },
185 { &rts_entities[rts_acosl], i_mapper_acos },
186 { &rts_entities[rts_atanl], i_mapper_atan },
187 { &rts_entities[rts_sinhl], i_mapper_sinh },
188 { &rts_entities[rts_coshl], i_mapper_cosh },
189 { &rts_entities[rts_tanhl], i_mapper_tanh },
192 { &rts_entities[rts_strcmp], i_mapper_strcmp },
193 { &rts_entities[rts_strncmp], i_mapper_strncmp },
194 { &rts_entities[rts_strcpy], i_mapper_strcpy },
195 { &rts_entities[rts_strlen], i_mapper_strlen },
196 { &rts_entities[rts_memcpy], i_mapper_memcpy },
197 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
198 { &rts_entities[rts_memmove], i_mapper_memmove },
199 { &rts_entities[rts_memset], i_mapper_memset },
200 { &rts_entities[rts_memcmp], i_mapper_memcmp }
202 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
205 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i)
206 if (*mapper[i].ent != NULL) {
207 rec[n_map].i_call.kind = INTRINSIC_CALL;
208 rec[n_map].i_call.i_ent = *mapper[i].ent;
209 rec[n_map].i_call.i_mapper = mapper[i].func;
210 rec[n_map].i_call.ctx = NULL;
211 rec[n_map].i_call.link = NULL;
215 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
218 static int *irg_dump_no;
220 static void dump_graph_count(ir_graph *const irg, const char *const suffix)
223 snprintf(name, sizeof(name), "-%02d_%s", irg_dump_no[get_irg_idx(irg)]++, suffix);
224 DUMP_ONE(1, irg, name);
227 static void dump_all_count(const char *const suffix)
229 const int n_irgs = get_irp_n_irgs();
232 for (i = 0; i < n_irgs; ++i)
233 dump_graph_count(get_irp_irg(i), suffix);
236 #define DUMP_ONE_C(cond, irg, suffix) \
239 dump_graph_count((irg), (suffix)); \
243 #define DUMP_ALL_C(cond, suffix) \
246 dump_all_count((suffix)); \
250 static void remove_unused_functions(void)
252 ir_entity **keep_methods;
255 /* Analysis that finds the free methods,
256 i.e. methods that are dereferenced.
257 Optimizes polymorphic calls :-). */
258 cgana(&arr_len, &keep_methods);
260 /* Remove methods that are never called. */
261 gc_irgs(arr_len, keep_methods);
265 static int firm_const_exists;
267 static void do_optimize_funccalls(void)
269 optimize_funccalls(firm_const_exists, NULL);
272 static void do_gcse(ir_graph *irg)
274 set_opt_global_cse(1);
275 optimize_graph_df(irg);
277 set_opt_global_cse(0);
280 static void do_lower_highlevel(ir_graph *irg)
282 lower_highlevel_graph(irg, firm_opt.lower_bitfields);
285 static void do_if_conv(ir_graph *irg)
287 opt_if_conv(irg, if_conv_info);
290 static void do_stred(ir_graph *irg)
292 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
295 static void after_inline_opt(ir_graph *irg)
297 do_irg_opt(irg, "scalar-replace");
298 do_irg_opt(irg, "local");
299 do_irg_opt(irg, "control-flow");
300 do_irg_opt(irg, "combo");
303 static void do_inline(void)
305 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
309 static void do_cloning(void)
311 proc_cloning((float) firm_opt.clone_threshold);
314 static void do_lower_switch(ir_graph *irg)
316 lower_switch(irg, firm_opt.spare_size);
319 static void do_lower_dw_ops(void)
321 lwrdw_param_t init = {
324 get_atomic_mode(ATOMIC_TYPE_LONGLONG),
325 get_atomic_mode(ATOMIC_TYPE_ULONGLONG),
326 get_atomic_mode(ATOMIC_TYPE_INT),
327 get_atomic_mode(ATOMIC_TYPE_UINT),
328 def_create_intrinsic_fkt,
332 if (arch_create_intrinsic) {
333 init.create_intrinsic = arch_create_intrinsic;
334 init.ctx = create_intrinsic_ctx;
339 typedef enum opt_target {
340 OPT_TARGET_IRG, /**< optimization function works on a single graph */
341 OPT_TARGET_IRP /**< optimization function works on the complete program */
344 typedef enum opt_flags {
346 OPT_FLAG_ENABLED = 1 << 0, /**< enable the optimization */
347 OPT_FLAG_NO_DUMP = 1 << 1, /**< don't dump after transformation */
348 OPT_FLAG_NO_VERIFY = 1 << 2, /**< don't verify after transformation */
349 OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
350 -foptions for this transformation */
353 typedef void (*transform_irg_func)(ir_graph *irg);
354 typedef void (*transform_irp_func)(void);
355 typedef void (*func_ptr_t)(void);
361 const char *description;
365 static opt_config_t opts[] = {
366 { OPT_TARGET_IRP, "rts", (func_ptr_t) rts_map, "optimization of known library functions", OPT_FLAG_HIDE_OPTIONS },
367 { OPT_TARGET_IRG, "combo", (func_ptr_t) combo, "combined CCE, UCE and GVN", OPT_FLAG_NONE},
368 { OPT_TARGET_IRG, "control-flow", (func_ptr_t) optimize_cf, "optimization of control-flow", OPT_FLAG_HIDE_OPTIONS },
369 { OPT_TARGET_IRG, "local", (func_ptr_t) optimize_graph_df, "local graph optimizations", OPT_FLAG_HIDE_OPTIONS },
370 { OPT_TARGET_IRP, "remove-unused", (func_ptr_t) remove_unused_functions, "removal of unused functions", OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
371 { OPT_TARGET_IRP, "opt-tail-rec", (func_ptr_t) opt_tail_recursion, "tail-recursion eliminiation", OPT_FLAG_NONE },
372 { OPT_TARGET_IRP, "opt-func-call", (func_ptr_t) do_optimize_funccalls, "function call optimization", OPT_FLAG_NONE },
373 { OPT_TARGET_IRG, "lower", (func_ptr_t) do_lower_highlevel, "lowering", OPT_FLAG_HIDE_OPTIONS },
374 { 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 },
375 { OPT_TARGET_IRP, "lower-dw", (func_ptr_t) do_lower_dw_ops, "lowering of doubleword operations", OPT_FLAG_HIDE_OPTIONS },
376 { OPT_TARGET_IRG, "lower-switch", (func_ptr_t) do_lower_switch, "switch lowering", OPT_FLAG_HIDE_OPTIONS },
377 { 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 },
378 { OPT_TARGET_IRG, "scalar-replace", (func_ptr_t) scalar_replacement_opt, "scalar replacement", OPT_FLAG_NONE },
379 { OPT_TARGET_IRG, "reassociation", (func_ptr_t) optimize_reassociation, "reassociation", OPT_FLAG_NONE },
380 { OPT_TARGET_IRG, "gcse", (func_ptr_t) do_gcse, "global common subexpression elimination", OPT_FLAG_NONE },
381 { OPT_TARGET_IRG, "place", (func_ptr_t) place_code, "code placement", OPT_FLAG_NONE },
382 { OPT_TARGET_IRG, "confirm", (func_ptr_t) construct_confirms, "confirm optimisation", OPT_FLAG_HIDE_OPTIONS },
383 { OPT_TARGET_IRG, "opt-load-store", (func_ptr_t) optimize_load_store, "load store optimization", OPT_FLAG_NONE },
384 { OPT_TARGET_IRG, "parallelize-mem", (func_ptr_t) opt_parallelize_mem, "parallelize memory", OPT_FLAG_NONE },
385 { OPT_TARGET_IRG, "deconv", (func_ptr_t) conv_opt, "conv node elimination", OPT_FLAG_NONE },
386 { OPT_TARGET_IRG, "thread-jumps", (func_ptr_t) opt_jumpthreading, "path-sensitive jumpthreading", OPT_FLAG_NONE },
387 { OPT_TARGET_IRG, "remove-confirms", (func_ptr_t) remove_confirms, "confirm removal", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY },
388 { OPT_TARGET_IRG, "gvn-pre", (func_ptr_t) do_gvn_pre, "global value numbering partial redundancy elimination", OPT_FLAG_NONE },
389 { OPT_TARGET_IRG, "if-conversion", (func_ptr_t) do_if_conv, "if-conversion", OPT_FLAG_NONE },
390 { OPT_TARGET_IRG, "bool", (func_ptr_t) opt_bool, "bool simplification", OPT_FLAG_NONE },
391 { OPT_TARGET_IRG, "shape-blocks", (func_ptr_t) shape_blocks, "block shaping", OPT_FLAG_NONE },
392 { OPT_TARGET_IRG, "ivopts", (func_ptr_t) do_stred, "induction variable strength reduction", OPT_FLAG_NONE },
393 { OPT_TARGET_IRG, "remove-phi-cycles", (func_ptr_t) remove_phi_cycles, "removal of phi cycles", OPT_FLAG_HIDE_OPTIONS },
394 { 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 },
395 { OPT_TARGET_IRP, "inline", (func_ptr_t) do_inline, "inlining", OPT_FLAG_NONE },
396 { OPT_TARGET_IRP, "opt-proc-clone", (func_ptr_t) do_cloning, "procedure cloning", OPT_FLAG_NONE },
397 { OPT_TARGET_IRG, "invert-loops", (func_ptr_t) do_loop_inversion, "loop inversion", OPT_FLAG_NONE },
398 { OPT_TARGET_IRG, "peel-loops", (func_ptr_t) do_loop_peeling, "loop peeling", OPT_FLAG_NONE },
400 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
401 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
403 static opt_config_t *get_opt(const char *name)
406 for (i = 0; i < n_opts; ++i) {
407 opt_config_t *config = &opts[i];
408 if (strcmp(config->name, name) == 0)
415 static void set_opt_enabled(const char *name, bool enabled)
417 opt_config_t *config = get_opt(name);
418 config->flags = (config->flags & ~OPT_FLAG_ENABLED)
419 | (enabled ? OPT_FLAG_ENABLED : 0);
422 static bool get_opt_enabled(const char *name)
424 opt_config_t *config = get_opt(name);
425 return (config->flags & OPT_FLAG_ENABLED) != 0;
429 * perform an optimisation on a single graph
431 * @return true if something changed, false otherwise
433 static bool do_irg_opt(ir_graph *irg, const char *name)
435 transform_irg_func func;
437 opt_config_t *config = get_opt(name);
438 size_t n = config - opts;
439 assert(config != NULL);
440 assert(config->target == OPT_TARGET_IRG);
441 if (! (config->flags & OPT_FLAG_ENABLED))
445 old_irg = current_ir_graph;
446 current_ir_graph = irg;
448 func = (transform_irg_func) config->func;
450 timer_push(timers[n]);
452 timer_pop(timers[n]);
454 DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, config->name);
456 if (firm_opt.check_all) {
457 timer_push(t_verify);
458 irg_verify(irg, VRFY_ENFORCE_SSA);
462 current_ir_graph = old_irg;
466 static void do_irp_opt(const char *name)
468 transform_irp_func func;
469 opt_config_t *config = get_opt(name);
470 size_t n = config - opts;
471 assert(config->target == OPT_TARGET_IRP);
472 if (! (config->flags & OPT_FLAG_ENABLED))
475 func = (transform_irp_func) config->func;
477 timer_push(timers[n]);
479 timer_pop(timers[n]);
481 DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, config->name);
483 if (firm_opt.check_all) {
485 timer_push(t_verify);
486 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
487 irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
494 * Enable transformations which should be always safe (and cheap) to perform
496 static void enable_safe_defaults(void)
498 set_opt_enabled("remove-unused", true);
499 set_opt_enabled("opt-tail-rec", true);
500 set_opt_enabled("opt-func-call", true);
501 set_opt_enabled("reassociation", true);
502 set_opt_enabled("control-flow", true);
503 set_opt_enabled("local", true);
504 set_opt_enabled("lower-const", true);
505 set_opt_enabled("scalar-replace", true);
506 set_opt_enabled("place", true);
507 set_opt_enabled("confirm", true);
508 set_opt_enabled("opt-load-store", true);
509 set_opt_enabled("lower", true);
510 set_opt_enabled("deconv", true);
511 set_opt_enabled("remove-confirms", true);
512 set_opt_enabled("ivopts", true);
513 set_opt_enabled("dead", true);
514 set_opt_enabled("lower-switch", true);
515 set_opt_enabled("remove-phi-cycles", true);
519 * run all the Firm optimizations
521 * @param input_filename the name of the (main) source file
523 static void do_firm_optimizations(const char *input_filename)
528 /* FIXME: cloning might ADD new graphs. */
529 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
531 set_opt_alias_analysis(firm_opt.alias_analysis);
533 aa_opt = aa_opt_no_opt;
534 if (firm_opt.strict_alias)
535 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
536 if (firm_opt.no_alias)
537 aa_opt = aa_opt_no_alias;
539 set_irp_memory_disambiguator_options(aa_opt);
541 /* parameter passing code should set them directly sometime... */
542 set_opt_enabled("rts", !firm_opt.freestanding);
543 set_opt_enabled("gcse", firm_opt.gcse);
544 set_opt_enabled("place", !firm_opt.gcse);
545 set_opt_enabled("confirm", firm_opt.confirm);
546 set_opt_enabled("remove-confirms", firm_opt.confirm);
548 /* osr supersedes remove_phi_cycles */
549 if (get_opt_enabled("ivopts"))
550 set_opt_enabled("remove-phi-cycles", false);
552 timer_start(t_all_opt);
556 /* first step: kill dead code */
557 for (i = 0; i < get_irp_n_irgs(); i++) {
558 ir_graph *irg = get_irp_irg(i);
559 do_irg_opt(irg, "combo");
560 do_irg_opt(irg, "local");
561 do_irg_opt(irg, "control-flow");
564 do_irp_opt("remove-unused");
565 do_irp_opt("opt-tail-rec");
566 do_irp_opt("opt-func-call");
567 do_irp_opt("lower-const");
569 for (i = 0; i < get_irp_n_irgs(); i++) {
570 ir_graph *irg = get_irp_irg(i);
572 do_irg_opt(irg, "scalar-replace");
573 do_irg_opt(irg, "invert-loops");
574 do_irg_opt(irg, "local");
575 do_irg_opt(irg, "reassociation");
576 do_irg_opt(irg, "local");
577 do_irg_opt(irg, "gcse");
579 if (firm_opt.confirm) {
580 /* Confirm construction currently can only handle blocks with only one
581 control flow predecessor. Calling optimize_cf here removes Bad
582 predecessors and help the optimization of switch constructs. */
583 do_irg_opt(irg, "control-flow");
584 do_irg_opt(irg, "confirm");
585 do_irg_opt(irg, "local");
588 do_irg_opt(irg, "control-flow");
589 do_irg_opt(irg, "opt-load-store");
590 do_irg_opt(irg, "lower");
591 do_irg_opt(irg, "deconv");
592 do_irg_opt(irg, "thread-jumps");
593 do_irg_opt(irg, "remove-confirms");
594 do_irg_opt(irg, "gvn-pre");
595 do_irg_opt(irg, "place");
596 do_irg_opt(irg, "control-flow");
598 if (do_irg_opt(irg, "if-conversion")) {
599 do_irg_opt(irg, "local");
600 do_irg_opt(irg, "control-flow");
603 do_irg_opt(irg, "bool");
604 do_irg_opt(irg, "shape-blocks");
605 do_irg_opt(irg, "lower-switch");
606 do_irg_opt(irg, "ivopts");
607 do_irg_opt(irg, "local");
608 do_irg_opt(irg, "dead");
611 do_irp_opt("inline");
612 do_irp_opt("opt-proc-clone");
614 for (i = 0; i < get_irp_n_irgs(); i++) {
615 ir_graph *irg = get_irp_irg(i);
616 do_irg_opt(irg, "local");
617 do_irg_opt(irg, "control-flow");
618 do_irg_opt(irg, "thread-jumps");
619 do_irg_opt(irg, "local");
620 do_irg_opt(irg, "control-flow");
623 if (firm_dump.ir_graph) {
624 /* recompute backedges for nicer dumps */
625 for (i = 0; i < get_irp_n_irgs(); i++)
626 construct_cf_backedges(get_irp_irg(i));
629 do_irp_opt("remove-unused");
631 DUMP_ALL(firm_dump.ir_graph, "-opt");
633 if (firm_dump.statistic & STAT_AFTER_OPT)
634 stat_dump_snapshot(input_filename, "opt");
636 timer_stop(t_all_opt);
640 * compute the size of a type (do implicit lowering)
642 * @param ty a Firm type
644 static int compute_type_size(ir_type *ty)
646 optimization_state_t state;
647 unsigned align_all = 1;
648 int n, size = 0, set = 0;
651 if (get_type_state(ty) == layout_fixed) {
652 /* do not layout already layouted types again */
656 if (is_Method_type(ty) || ty == get_glob_type()) {
657 /* no need for size calculation for method types or the global type */
661 DBG(("compute type size visiting: %s\n", get_type_name(ty)));
663 switch (get_type_tpop_code(ty)) {
666 for (i = 0, n = get_compound_n_members(ty); i < n; ++i) {
667 ir_entity *ent = get_compound_member(ty, i);
668 ir_type *ent_ty = get_entity_type(ent);
669 unsigned align, misalign;
671 /* inner functions do not expand the frame */
672 if (is_Method_type(ent_ty) && is_frame_type(ty))
675 /* compute member types */
676 if (! compute_type_size(ent_ty))
679 align = get_type_alignment_bytes(ent_ty);
680 align_all = align > align_all ? align : align_all;
681 misalign = (align ? size % align : 0);
682 size += (misalign ? align - misalign : 0);
684 set_entity_offset(ent, size);
685 size += get_type_size_bytes(ent_ty);
687 DBG((" member %s %s -> (size: %u, align: %u)\n",
688 get_type_name(ent_ty), get_entity_name(ent),
689 get_type_size_bytes(ent_ty), get_type_alignment_bytes(ent_ty)));
691 if (align_all > 0 && size % align_all) {
692 DBG(("align of the struct member: %u, type size: %d\n", align_all, size));
693 size += align_all - (size % align_all);
694 DBG(("correcting type-size to %d\n", size));
696 set_type_alignment_bytes(ty, align_all);
701 for (i = 0, n = get_union_n_members(ty); i < n; ++i) {
702 ir_entity *ent = get_union_member(ty, i);
704 if (! compute_type_size(get_entity_type(ent)))
706 s = get_type_size_bytes(get_entity_type(ent));
708 set_entity_offset(ent, 0);
709 size = (s > size ? s : size);
715 dims = get_array_n_dimensions(ty);
717 if (! compute_type_size(get_array_element_type(ty)))
722 save_optimization_state(&state);
724 set_opt_constant_folding(1);
725 set_opt_algebraic_simplification(1);
727 for (i = 0; i < dims; ++i) {
728 ir_node *lower = get_array_lower_bound(ty, i);
729 ir_node *upper = get_array_upper_bound(ty, i);
730 ir_graph *rem = current_ir_graph;
731 tarval *tv_lower, *tv_upper;
732 long val_lower, val_upper;
734 current_ir_graph = get_const_code_irg();
735 local_optimize_node(lower);
736 local_optimize_node(upper);
737 current_ir_graph = rem;
739 tv_lower = computed_value(lower);
740 tv_upper = computed_value(upper);
742 if (tv_lower == tarval_bad || tv_upper == tarval_bad) {
744 * we cannot calculate the size of this array yet, it
745 * even might be unknown until the end, like argv[]
747 restore_optimization_state(&state);
751 val_upper = get_tarval_long(tv_upper);
752 val_lower = get_tarval_long(tv_lower);
753 size *= val_upper - val_lower;
755 restore_optimization_state(&state);
757 DBG(("array %s -> (elements: %d, element type size: %d)\n",
759 size, get_type_size_bytes(get_array_element_type(ty))));
760 size *= get_type_size_bytes(get_array_element_type(ty));
769 set_type_size_bytes(ty, size);
770 set_type_state(ty, layout_fixed);
773 DBG(("size: %d\n", get_type_size_bytes(ty)));
776 } /* compute_type_size */
779 * layout all non-frame types of the Firm graph
781 static void compute_type_sizes(void)
786 /* all non-frame other types */
787 for (i = get_irp_n_types() - 1; i >= 0; --i) {
788 tp = get_irp_type(i);
789 compute_type_size(tp);
791 if (is_Method_type(tp)) {
792 tp = get_method_value_res_type(tp);
795 /* we have a value result type for this method, lower */
796 compute_type_size(tp);
800 } /* compute_type_sizes */
803 * layout all frame-types of the Firm graph
805 static void compute_frame_type_sizes(void)
810 /* all frame types */
811 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
812 irg = get_irp_irg(i);
813 /* do not optimize away variables in debug mode */
814 if (firm_opt.debug_mode == DBG_MODE_NONE)
816 compute_type_size(get_irg_frame_type(irg));
818 } /* compute_frame_type_sizes */
823 * @param input_filename the name of the (main) source file
825 static void do_firm_lowering(const char *input_filename)
829 do_irp_opt("lower-dw");
831 if (firm_dump.statistic & STAT_AFTER_LOWER)
832 stat_dump_snapshot(input_filename, "low");
834 DUMP_ALL(firm_dump.ir_graph, "-low");
836 if (firm_opt.enabled) {
837 timer_start(t_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 ir_graph *irg = get_irp_irg(i);
843 do_irg_opt(irg, "reassociation");
846 /* enable architecture dependent optimizations */
847 arch_dep_set_opts((arch_dep_opts_t)
848 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
849 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
850 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
852 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
853 ir_graph *irg = get_irp_irg(i);
855 current_ir_graph = irg;
857 do_irg_opt(irg, "local");
858 do_irg_opt(irg, "gcse");
859 do_irg_opt(irg, "opt-load-store");
860 do_irg_opt(irg, "local");
861 do_irg_opt(irg, "control-flow");
863 if (do_irg_opt(irg, "if-conversion")) {
864 do_irg_opt(irg, "local");
865 do_irg_opt(irg, "control-flow");
868 do_irg_opt(irg, "parallelize-mem");
870 timer_stop(t_all_opt);
872 DUMP_ALL(firm_dump.ir_graph, "-low-opt");
876 mark_private_methods();
878 /* set the phase to low */
879 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
880 set_irg_phase_low(get_irp_irg(i));
882 /* all graphs are lowered, set the irp phase to low */
883 set_irp_phase_state(phase_low);
885 if (firm_dump.statistic & STAT_FINAL) {
886 stat_dump_snapshot(input_filename, "final");
891 * Initialize for the Firm-generating back end.
893 void gen_firm_init(void)
895 firm_parameter_t params;
896 unsigned pattern = 0;
899 for (i = 0; i < n_opts; ++i) {
900 timers[i] = ir_timer_new();
901 timer_register(timers[i], opts[i].description);
903 t_verify = ir_timer_new();
904 timer_register(t_verify, "Firm: verify pass");
905 t_vcg_dump = ir_timer_new();
906 timer_register(t_vcg_dump, "Firm: vcg dumping");
907 t_all_opt = ir_timer_new();
908 timer_register(t_all_opt, "Firm: all optimizations");
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;
926 if (firm_be_opt.selection == BE_FIRM_BE) {
927 const backend_params *be_params = be_get_backend_param();
929 if (be_params->do_dw_lowering)
930 set_opt_enabled("lower-dw", true);
932 arch_create_intrinsic = be_params->arch_create_intrinsic_fkt;
933 create_intrinsic_ctx = be_params->create_intrinsic_ctx;
935 ad_param = be_params->dep_param;
936 if_conv_info = be_params->if_conv_info;
939 dbg_init(NULL, NULL, dbg_snprint);
940 edges_init_dbg(firm_opt.vrfy_edges);
942 /* Sel node cannot produce NULL pointers */
943 set_opt_sel_based_null_check_elim(1);
945 /* dynamic dispatch works currently only if whole world scenarios */
946 set_opt_dyn_meth_dispatch(0);
948 arch_dep_init(arch_factory);
950 /* do not run architecture dependent optimizations in building phase */
951 arch_dep_set_opts(arch_dep_none);
953 do_node_verification((firm_verification_t) firm_opt.vrfy);
954 if (firm_dump.filter)
955 only_dump_method_with_name(new_id_from_str(firm_dump.filter));
957 if (firm_opt.enabled) {
959 set_opt_constant_folding(firm_opt.const_folding);
960 set_opt_algebraic_simplification(firm_opt.const_folding);
961 set_opt_cse(firm_opt.cse);
962 set_opt_global_cse(0);
963 set_opt_unreachable_code(1);
964 set_opt_control_flow(firm_opt.control_flow);
965 set_opt_control_flow_weak_simplification(1);
966 set_opt_control_flow_strong_simplification(1);
971 /* do not dump entity ld names */
976 * Called, after the Firm generation is completed,
977 * do all optimizations and backend call here.
979 * @param out a file handle for the output, may be NULL
980 * @param input_filename the name of the (main) source file
981 * @param c_mode non-zero if "C" was compiled
982 * @param new_firm_const_exists non-zero, if the const attribute was used on functions
984 void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int new_firm_const_exists)
989 if (firm_opt.enable_statev) {
991 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
992 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
993 ir_stat_ev_compilation_unit(input_filename);
997 firm_const_exists = new_firm_const_exists;
999 /* the general for dumping option must be set, or the others will not work */
1001 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
1003 dump_keepalive_edges(1);
1004 dump_consts_local(1);
1005 dump_dominator_information(1);
1006 dump_loop_information(0);
1008 if (!firm_dump.edge_labels)
1009 turn_off_edge_labels();
1011 if (firm_dump.all_types) {
1014 dump_class_hierarchy(0, "");
1015 dump_class_hierarchy(1, "-with-entities");
1019 /* finalize all graphs */
1020 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1021 ir_graph *irg = get_irp_irg(i);
1023 irg_finalize_cons(irg);
1024 DUMP_ONE(firm_dump.ir_graph, irg, "");
1027 timer_push(t_verify);
1029 timer_pop(t_verify);
1031 /* all graphs are finalized, set the irp phase to high */
1032 set_irp_phase_state(phase_high);
1034 /* BEWARE: kill unreachable code before doing compound lowering */
1035 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1036 ir_graph *irg = get_irp_irg(i);
1037 do_irg_opt(irg, "control-flow");
1040 /* lower all compound call return values */
1041 lower_compound_params();
1043 /* computes the sizes of all types that are still not computed */
1044 compute_type_sizes();
1046 /* lower copyb nodes */
1047 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1048 ir_graph *irg = get_irp_irg(i);
1049 lower_CopyB(irg, 128, 4);
1052 if (firm_dump.statistic & STAT_BEFORE_OPT) {
1053 stat_dump_snapshot(input_filename, "noopt");
1056 if (firm_opt.enabled)
1057 do_firm_optimizations(input_filename);
1059 if (firm_dump.gen_firm_asm) {
1060 ir_timer_t *timer = ir_timer_new();
1061 timer_register(timer, "Firm: Firm assembler");
1063 gen_Firm_assembler(input_filename);
1069 do_firm_lowering(input_filename);
1071 /* computes the sizes of all frame types */
1072 compute_frame_type_sizes();
1074 /* set the phase to low */
1075 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
1076 set_irg_phase_low(get_irp_irg(i));
1078 if (firm_dump.statistic & STAT_FINAL_IR)
1079 stat_dump_snapshot(input_filename, "final-ir");
1081 /* run the code generator */
1082 if (firm_be_opt.selection != BE_NONE)
1083 do_codegen(out, input_filename);
1085 if (firm_dump.statistic & STAT_FINAL)
1086 stat_dump_snapshot(input_filename, "final");
1089 int firm_opt_option(const char *opt)
1092 if (strncmp(opt, "no-", 3) == 0) {
1097 opt_config_t *config = get_opt(opt);
1098 if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
1101 config->flags &= ~OPT_FLAG_ENABLED;
1102 config->flags |= enable ? OPT_FLAG_ENABLED : 0;
1106 void firm_opt_option_help(void)
1110 for (i = 0; i < n_opts; ++i) {
1114 const opt_config_t *config = &opts[i];
1115 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
1118 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
1119 print_option_help(config->name, buf2);
1120 snprintf(buf, sizeof(buf), "no-%s", config->name);
1121 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
1122 print_option_help(buf, buf2);
1127 * Do very early initializations
1129 void firm_early_init(void)
1131 /* arg: need this here for command line options */
1134 enable_safe_defaults();