3 * @file firm_opt.c -- Firm-generating back end optimizations.
5 * (C) 2005-2010 Michael Beck beck@ipd.info.uni-karlsruhe.de
17 #include <libfirm/firm.h>
20 #include "firm_timing.h"
22 #include "adt/strutil.h"
24 /* optimization settings */
26 bool const_folding; /**< enable constant folding */
27 bool cse; /**< enable common-subexpression elimination */
28 bool confirm; /**< enable Confirm optimization */
29 bool muls; /**< enable architecture dependent mul optimization */
30 bool divs; /**< enable architecture dependent div optimization */
31 bool mods; /**< enable architecture dependent mod optimization */
32 bool alias_analysis; /**< enable Alias Analysis */
33 bool strict_alias; /**< enable strict Alias Analysis (using type based AA) */
34 bool no_alias; /**< no aliasing possible. */
35 bool verify; /**< Firm verifier setting */
36 bool check_all; /**< enable checking all Firm phases */
37 int clone_threshold; /**< The threshold value for procedure cloning. */
38 unsigned inline_maxsize; /**< Maximum function size for inlining. */
39 unsigned inline_threshold;/**< Inlining benefice threshold. */
40 bool verify_edges; /**< verify edges */
43 /** statistic options */
44 typedef enum a_firmstat_selection_tag {
45 STAT_NONE = 0x00000000,
46 STAT_BEFORE_OPT = 0x00000001,
47 STAT_AFTER_OPT = 0x00000002,
48 STAT_AFTER_LOWER = 0x00000004,
49 STAT_FINAL_IR = 0x00000008,
50 STAT_FINAL = 0x00000010,
51 } a_firmstat_selection;
55 bool debug_print; /**< enable debug print */
56 bool all_types; /**< dump the All_types graph */
57 bool no_blocks; /**< dump non-blocked graph */
58 bool extbb; /**< dumps extended basic blocks */
59 bool ir_graph; /**< dump all graphs */
60 bool all_phases; /**< dump the IR graph after all phases */
61 bool statistic; /**< Firm statistic setting */
62 bool stat_pattern; /**< enable Firm statistic pattern */
63 bool stat_dag; /**< enable Firm DAG statistic */
66 struct a_firm_be_opt {
71 /* optimization settings */
72 static struct a_firm_opt firm_opt = {
73 .const_folding = true,
79 .alias_analysis = true,
80 .strict_alias = false,
82 .verify = FIRM_VERIFICATION_ON,
84 .clone_threshold = DEFAULT_CLONE_THRESHOLD,
85 .inline_maxsize = 750,
86 .inline_threshold = 0,
87 .verify_edges = false,
91 static struct a_firm_dump firm_dump = {
98 .statistic = STAT_NONE,
103 #define X(a) a, sizeof(a)-1
105 /** Parameter description structure */
106 static const struct params {
107 const char *option; /**< name of the option */
108 size_t opt_len; /**< length of the option string */
109 bool *flag; /**< address of variable to set/reset */
110 bool set; /**< iff true, variable will be set, else reset */
111 const char *description; /**< description of this option */
113 /* firm optimization options */
114 { X("no-opt"), NULL, 0, "disable all FIRM optimizations" },
115 { X("cse"), &firm_opt.cse, 1, "enable common subexpression elimination" },
116 { X("no-cse"), &firm_opt.cse, 0, "disable common subexpression elimination" },
117 { X("const-fold"), &firm_opt.const_folding, 1, "enable constant folding" },
118 { X("no-const-fold"), &firm_opt.const_folding, 0, "disable constant folding" },
119 { X("inline-max-size=<size>"), NULL, 0, "set maximum size for function inlining" },
120 { X("inline-threshold=<size>"),NULL, 0, "set benefice threshold for function inlining" },
121 { X("confirm"), &firm_opt.confirm, 1, "enable Confirm optimization" },
122 { X("no-confirm"), &firm_opt.confirm, 0, "disable Confirm optimization" },
123 { X("opt-mul"), &firm_opt.muls, 0, "enable multiplication optimization" },
124 { X("no-opt-mul"), &firm_opt.muls, 0, "disable multiplication optimization" },
125 { X("opt-div"), &firm_opt.divs, 0, "enable division optimization" },
126 { X("no-opt-div"), &firm_opt.divs, 0, "disable division optimization" },
127 { X("opt-mod"), &firm_opt.mods, 0, "enable remainder optimization" },
128 { X("no-opt-mod"), &firm_opt.mods, 0, "disable remainder optimization" },
129 { X("opt-alias"), &firm_opt.alias_analysis, 1, "enable alias analysis" },
130 { X("no-opt-alias"), &firm_opt.alias_analysis, 0, "disable alias analysis" },
131 { X("alias"), &firm_opt.no_alias, 0, "aliasing occurs" },
132 { X("no-alias"), &firm_opt.no_alias, 1, "no aliasing occurs" },
133 { X("strict-aliasing"), &firm_opt.strict_alias, 1, "strict alias rules" },
134 { X("no-strict-aliasing"), &firm_opt.strict_alias, 0, "strict alias rules" },
135 { X("clone-threshold=<value>"),NULL, 0, "set clone threshold to <value>" },
137 /* other firm regarding options */
138 { X("verify-off"), &firm_opt.verify, FIRM_VERIFICATION_OFF, "disable node verification" },
139 { X("verify-on"), &firm_opt.verify, FIRM_VERIFICATION_ON, "enable node verification" },
140 { X("verify-report"), &firm_opt.verify, FIRM_VERIFICATION_REPORT, "node verification, report only" },
141 { X("check-all"), &firm_opt.check_all, 1, "enable checking all Firm phases" },
142 { X("no-check-all"), &firm_opt.check_all, 0, "disable checking all Firm phases" },
143 { X("verify-edges-on"), &firm_opt.verify_edges, 1, "enable out edge verification" },
144 { X("verify-edges-off"), &firm_opt.verify_edges, 0, "disable out edge verification" },
147 { X("dump-ir"), &firm_dump.ir_graph, 1, "dump IR graph" },
148 { X("dump-all-types"), &firm_dump.all_types, 1, "dump graph of all types" },
149 { X("dump-no-blocks"), &firm_dump.no_blocks, 1, "dump non-blocked graph" },
150 { X("dump-extbb"), &firm_dump.extbb, 1, "dump extended basic blocks" },
151 { X("dump-all-phases"), &firm_dump.all_phases, 1, "dump graphs for all optimization phases" },
152 { X("dump-filter=<string>"), NULL, 0, "set dumper filter" },
155 { X("stat-before-opt"), &firm_dump.statistic, STAT_BEFORE_OPT, "Firm statistic output before optimizations" },
156 { X("stat-after-opt"), &firm_dump.statistic, STAT_AFTER_OPT, "Firm statistic output after optimizations" },
157 { X("stat-after-lower"), &firm_dump.statistic, STAT_AFTER_LOWER, "Firm statistic output after lowering" },
158 { X("stat-final-ir"), &firm_dump.statistic, STAT_FINAL_IR, "Firm statistic after final optimization" },
159 { X("stat-final"), &firm_dump.statistic, STAT_FINAL, "Firm statistic after code generation" },
160 { X("stat-pattern"), &firm_dump.stat_pattern, 1, "Firm statistic calculates most used pattern" },
161 { X("stat-dag"), &firm_dump.stat_dag, 1, "Firm calculates DAG statistics" },
166 static ir_timer_t *t_vcg_dump;
167 static ir_timer_t *t_verify;
168 static ir_timer_t *t_all_opt;
169 static bool do_irg_opt(ir_graph *irg, const char *name);
171 /** dump all the graphs depending on cond */
173 static void dump_all(const char *suffix)
175 if (!firm_dump.ir_graph)
178 timer_push(t_vcg_dump);
179 dump_all_ir_graphs(suffix);
180 timer_pop(t_vcg_dump);
183 /* entities of runtime functions */
184 ir_entity_ptr rts_entities[rts_max];
187 * Map runtime functions.
189 static void rts_map(void)
191 static const struct {
192 ir_entity_ptr *ent; /**< address of the rts entity */
193 i_mapper_func func; /**< mapper function. */
196 { &rts_entities[rts_abs], i_mapper_abs },
197 { &rts_entities[rts_labs], i_mapper_abs },
198 { &rts_entities[rts_llabs], i_mapper_abs },
199 { &rts_entities[rts_imaxabs], i_mapper_abs },
201 /* double -> double */
202 { &rts_entities[rts_fabs], i_mapper_abs },
203 { &rts_entities[rts_sqrt], i_mapper_sqrt },
204 { &rts_entities[rts_cbrt], i_mapper_cbrt },
205 { &rts_entities[rts_pow], i_mapper_pow },
206 { &rts_entities[rts_exp], i_mapper_exp },
207 { &rts_entities[rts_exp2], i_mapper_exp },
208 { &rts_entities[rts_exp10], i_mapper_exp },
209 { &rts_entities[rts_log], i_mapper_log },
210 { &rts_entities[rts_log2], i_mapper_log2 },
211 { &rts_entities[rts_log10], i_mapper_log10 },
212 { &rts_entities[rts_sin], i_mapper_sin },
213 { &rts_entities[rts_cos], i_mapper_cos },
214 { &rts_entities[rts_tan], i_mapper_tan },
215 { &rts_entities[rts_asin], i_mapper_asin },
216 { &rts_entities[rts_acos], i_mapper_acos },
217 { &rts_entities[rts_atan], i_mapper_atan },
218 { &rts_entities[rts_sinh], i_mapper_sinh },
219 { &rts_entities[rts_cosh], i_mapper_cosh },
220 { &rts_entities[rts_tanh], i_mapper_tanh },
223 { &rts_entities[rts_fabsf], i_mapper_abs },
224 { &rts_entities[rts_sqrtf], i_mapper_sqrt },
225 { &rts_entities[rts_cbrtf], i_mapper_cbrt },
226 { &rts_entities[rts_powf], i_mapper_pow },
227 { &rts_entities[rts_expf], i_mapper_exp },
228 { &rts_entities[rts_exp2f], i_mapper_exp },
229 { &rts_entities[rts_exp10f], i_mapper_exp },
230 { &rts_entities[rts_logf], i_mapper_log },
231 { &rts_entities[rts_log2f], i_mapper_log2 },
232 { &rts_entities[rts_log10f], i_mapper_log10 },
233 { &rts_entities[rts_sinf], i_mapper_sin },
234 { &rts_entities[rts_cosf], i_mapper_cos },
235 { &rts_entities[rts_tanf], i_mapper_tan },
236 { &rts_entities[rts_asinf], i_mapper_asin },
237 { &rts_entities[rts_acosf], i_mapper_acos },
238 { &rts_entities[rts_atanf], i_mapper_atan },
239 { &rts_entities[rts_sinhf], i_mapper_sinh },
240 { &rts_entities[rts_coshf], i_mapper_cosh },
241 { &rts_entities[rts_tanhf], i_mapper_tanh },
243 /* long double -> long double */
244 { &rts_entities[rts_fabsl], i_mapper_abs },
245 { &rts_entities[rts_sqrtl], i_mapper_sqrt },
246 { &rts_entities[rts_cbrtl], i_mapper_cbrt },
247 { &rts_entities[rts_powl], i_mapper_pow },
248 { &rts_entities[rts_expl], i_mapper_exp },
249 { &rts_entities[rts_exp2l], i_mapper_exp },
250 { &rts_entities[rts_exp10l], i_mapper_exp },
251 { &rts_entities[rts_logl], i_mapper_log },
252 { &rts_entities[rts_log2l], i_mapper_log2 },
253 { &rts_entities[rts_log10l], i_mapper_log10 },
254 { &rts_entities[rts_sinl], i_mapper_sin },
255 { &rts_entities[rts_cosl], i_mapper_cos },
256 { &rts_entities[rts_tanl], i_mapper_tan },
257 { &rts_entities[rts_asinl], i_mapper_asin },
258 { &rts_entities[rts_acosl], i_mapper_acos },
259 { &rts_entities[rts_atanl], i_mapper_atan },
260 { &rts_entities[rts_sinhl], i_mapper_sinh },
261 { &rts_entities[rts_coshl], i_mapper_cosh },
262 { &rts_entities[rts_tanhl], i_mapper_tanh },
265 { &rts_entities[rts_strcmp], i_mapper_strcmp },
266 { &rts_entities[rts_strncmp], i_mapper_strncmp },
267 { &rts_entities[rts_strcpy], i_mapper_strcpy },
268 { &rts_entities[rts_strlen], i_mapper_strlen },
269 { &rts_entities[rts_memcpy], i_mapper_memcpy },
270 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
271 { &rts_entities[rts_memmove], i_mapper_memmove },
272 { &rts_entities[rts_memset], i_mapper_memset },
273 { &rts_entities[rts_memcmp], i_mapper_memcmp }
275 i_record rec[sizeof(mapper)/sizeof(mapper[0])];
278 for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
279 if (*mapper[i].ent != NULL) {
280 rec[n_map].i_call.kind = INTRINSIC_CALL;
281 rec[n_map].i_call.i_ent = *mapper[i].ent;
282 rec[n_map].i_call.i_mapper = mapper[i].func;
283 rec[n_map].i_call.ctx = NULL;
284 rec[n_map].i_call.link = NULL;
290 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
293 static int *irg_dump_no;
295 static void do_stred(ir_graph *irg)
297 opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
300 static void after_inline_opt(ir_graph *irg)
302 do_irg_opt(irg, "scalar-replace");
303 do_irg_opt(irg, "local");
304 do_irg_opt(irg, "control-flow");
305 do_irg_opt(irg, "combo");
308 static void do_inline(void)
310 inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
314 static void do_cloning(void)
316 proc_cloning((float) firm_opt.clone_threshold);
319 static void do_lower_mux(ir_graph *irg)
321 lower_mux(irg, NULL);
324 static void do_gcse(ir_graph *irg)
326 set_opt_global_cse(1);
327 optimize_graph_df(irg);
328 set_opt_global_cse(0);
331 static void lower_blockcopy(ir_graph *irg)
333 lower_CopyB(irg, 128, 4);
336 typedef enum opt_target {
337 OPT_TARGET_IRG, /**< optimization function works on a single graph */
338 OPT_TARGET_IRP /**< optimization function works on the complete program */
341 typedef enum opt_flags {
343 OPT_FLAG_ENABLED = 1 << 0, /**< enable the optimization */
344 OPT_FLAG_NO_DUMP = 1 << 1, /**< don't dump after transformation */
345 OPT_FLAG_NO_VERIFY = 1 << 2, /**< don't verify after transformation */
346 OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
347 -foptions for this transformation */
348 OPT_FLAG_ESSENTIAL = 1 << 4, /**< output won't work without this pass
349 so we need it even with -O0 */
352 typedef void (*transform_irg_func)(ir_graph *irg);
353 typedef void (*transform_irp_func)(void);
359 transform_irg_func transform_irg;
360 transform_irp_func transform_irp;
362 const char *description;
366 static opt_config_t opts[] = {
367 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
368 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b, c, d }
369 IRG("bool", opt_bool, "bool simplification", OPT_FLAG_NONE),
370 IRG("combo", combo, "combined CCE, UCE and GVN", OPT_FLAG_NONE),
371 IRG("confirm", construct_confirms, "confirm optimization", OPT_FLAG_HIDE_OPTIONS),
372 IRG("control-flow", optimize_cf, "optimization of control-flow", OPT_FLAG_HIDE_OPTIONS),
373 IRG("dead", dead_node_elimination, "dead node elimination", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
374 IRG("deconv", conv_opt, "conv node elimination", OPT_FLAG_NONE),
375 IRG("fp-vrp", fixpoint_vrp, "fixpoint value range propagation", OPT_FLAG_NONE),
376 IRG("frame", opt_frame_irg, "remove unused frame entities", OPT_FLAG_NONE),
377 IRG("gvn-pre", do_gvn_pre, "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
378 IRG("if-conversion", opt_if_conv, "if-conversion", OPT_FLAG_NONE),
379 IRG("invert-loops", do_loop_inversion, "loop inversion", OPT_FLAG_NONE),
380 IRG("ivopts", do_stred, "induction variable strength reduction", OPT_FLAG_NONE),
381 IRG("local", optimize_graph_df, "local graph optimizations", OPT_FLAG_HIDE_OPTIONS),
382 IRG("lower", lower_highlevel_graph, "lowering", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
383 IRG("lower-mux", do_lower_mux, "mux lowering", OPT_FLAG_NONE),
384 IRG("opt-load-store", optimize_load_store, "load store optimization", OPT_FLAG_NONE),
385 IRG("opt-tail-rec", opt_tail_rec_irg, "tail-recursion eliminiation", OPT_FLAG_NONE),
386 IRG("parallelize-mem", opt_parallelize_mem, "parallelize memory", OPT_FLAG_NONE),
387 IRG("gcse", do_gcse, "global common subexpression eliminiation", OPT_FLAG_NONE),
388 IRG("place", place_code, "code placement", OPT_FLAG_NONE),
389 IRG("reassociation", optimize_reassociation, "reassociation", OPT_FLAG_NONE),
390 IRG("remove-confirms", remove_confirms, "confirm removal", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
391 IRG("remove-phi-cycles", remove_phi_cycles, "removal of phi cycles", OPT_FLAG_HIDE_OPTIONS),
392 IRG("scalar-replace", scalar_replacement_opt, "scalar replacement", OPT_FLAG_NONE),
393 IRG("shape-blocks", shape_blocks, "block shaping", OPT_FLAG_NONE),
394 IRG("thread-jumps", opt_jumpthreading, "path-sensitive jumpthreading", OPT_FLAG_NONE),
395 IRG("unroll-loops", do_loop_unrolling, "loop unrolling", OPT_FLAG_NONE),
396 IRG("vrp", set_vrp_data, "value range propagation", OPT_FLAG_NONE),
397 IRP("inline", do_inline, "inlining", OPT_FLAG_NONE),
398 IRP("lower-const", lower_const_code, "lowering of constant code", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY | OPT_FLAG_ESSENTIAL),
399 IRG("lower-blockcopy", lower_blockcopy, "replace small block copies with load/store sequences", OPT_FLAG_NO_DUMP),
400 IRP("target-lowering", be_lower_for_target, "lowering necessary for target architecture", OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
401 IRP("opt-func-call", optimize_funccalls, "function call optimization", OPT_FLAG_NONE),
402 IRP("opt-proc-clone", do_cloning, "procedure cloning", OPT_FLAG_NONE),
403 IRP("remove-unused", garbage_collect_entities, "removal of unused functions/variables", OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
404 IRP("rts", rts_map, "optimization of known library functions", OPT_FLAG_NONE),
405 IRP("opt-cc", mark_private_methods, "calling conventions optimization", OPT_FLAG_NONE),
409 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
410 static ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
412 static opt_config_t *get_opt(const char *name)
415 for (i = 0; i < n_opts; ++i) {
416 opt_config_t *config = &opts[i];
417 if (strcmp(config->name, name) == 0)
424 static void set_opt_enabled(const char *name, bool enabled)
426 opt_config_t *config = get_opt(name);
427 config->flags = (config->flags & ~OPT_FLAG_ENABLED)
428 | (enabled ? OPT_FLAG_ENABLED : 0);
431 static bool get_opt_enabled(const char *name)
433 opt_config_t *config = get_opt(name);
434 return (config->flags & OPT_FLAG_ENABLED) != 0;
438 * perform an optimization on a single graph
440 * @return true if something changed, false otherwise
442 static bool do_irg_opt(ir_graph *irg, const char *name)
445 opt_config_t *config = get_opt(name);
446 size_t n = config - opts;
447 assert(config != NULL);
448 assert(config->target == OPT_TARGET_IRG);
449 if (! (config->flags & OPT_FLAG_ENABLED))
452 old_irg = current_ir_graph;
453 current_ir_graph = irg;
455 timer_push(timers[n]);
456 config->u.transform_irg(irg);
457 timer_pop(timers[n]);
459 if (firm_dump.all_phases && firm_dump.ir_graph) {
460 dump_ir_graph(irg, name);
463 if (firm_opt.check_all) {
464 timer_push(t_verify);
465 irg_verify(irg, VERIFY_ENFORCE_SSA);
469 current_ir_graph = old_irg;
473 static void do_irp_opt(const char *name)
475 opt_config_t *config = get_opt(name);
476 size_t n = config - opts;
477 assert(config->target == OPT_TARGET_IRP);
478 if (! (config->flags & OPT_FLAG_ENABLED))
481 timer_push(timers[n]);
482 config->u.transform_irp();
483 timer_pop(timers[n]);
485 if (firm_dump.ir_graph && firm_dump.all_phases) {
487 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
488 ir_graph *irg = get_irp_irg(i);
489 dump_ir_graph(irg, name);
493 if (firm_opt.check_all) {
495 timer_push(t_verify);
496 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
497 irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
504 * Enable transformations which should be always safe (and cheap) to perform
506 static void enable_safe_defaults(void)
508 set_opt_enabled("remove-unused", true);
509 set_opt_enabled("opt-tail-rec", true);
510 set_opt_enabled("opt-func-call", true);
511 set_opt_enabled("reassociation", true);
512 set_opt_enabled("control-flow", true);
513 set_opt_enabled("local", true);
514 set_opt_enabled("lower-const", true);
515 set_opt_enabled("scalar-replace", true);
516 set_opt_enabled("place", true);
517 set_opt_enabled("gcse", true);
518 set_opt_enabled("confirm", true);
519 set_opt_enabled("opt-load-store", true);
520 set_opt_enabled("lower", true);
521 set_opt_enabled("lower-blockcopy", true);
522 set_opt_enabled("deconv", true);
523 set_opt_enabled("remove-confirms", true);
524 set_opt_enabled("ivopts", true);
525 set_opt_enabled("dead", true);
526 set_opt_enabled("remove-phi-cycles", true);
527 set_opt_enabled("frame", true);
528 set_opt_enabled("combo", true);
529 set_opt_enabled("invert-loops", true);
530 set_opt_enabled("target-lowering", true);
531 set_opt_enabled("rts", true);
532 set_opt_enabled("parallelize-mem", true);
533 set_opt_enabled("opt-cc", true);
537 * run all the Firm optimizations
539 * @param input_filename the name of the (main) source file
541 static void do_firm_optimizations(const char *input_filename)
546 set_opt_alias_analysis(firm_opt.alias_analysis);
548 aa_opt = aa_opt_no_opt;
549 if (firm_opt.strict_alias)
550 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
551 if (firm_opt.no_alias)
552 aa_opt = aa_opt_no_alias;
554 set_irp_memory_disambiguator_options(aa_opt);
556 /* parameter passing code should set them directly sometime... */
557 set_opt_enabled("confirm", firm_opt.confirm);
558 set_opt_enabled("remove-confirms", firm_opt.confirm);
560 /* osr supersedes remove_phi_cycles */
561 if (get_opt_enabled("ivopts"))
562 set_opt_enabled("remove-phi-cycles", false);
564 timer_start(t_all_opt);
568 /* first step: kill dead code */
569 for (i = 0; i < get_irp_n_irgs(); i++) {
570 ir_graph *irg = get_irp_irg(i);
571 do_irg_opt(irg, "combo");
572 do_irg_opt(irg, "local");
573 do_irg_opt(irg, "control-flow");
576 do_irp_opt("remove-unused");
577 for (i = 0; i < get_irp_n_irgs(); ++i) {
578 ir_graph *irg = get_irp_irg(i);
579 do_irg_opt(irg, "opt-tail-rec");
581 do_irp_opt("opt-func-call");
582 do_irp_opt("lower-const");
584 for (i = 0; i < get_irp_n_irgs(); i++) {
585 ir_graph *irg = get_irp_irg(i);
587 do_irg_opt(irg, "scalar-replace");
588 do_irg_opt(irg, "invert-loops");
589 do_irg_opt(irg, "unroll-loops");
590 do_irg_opt(irg, "local");
591 do_irg_opt(irg, "reassociation");
592 do_irg_opt(irg, "local");
593 do_irg_opt(irg, "gcse");
594 do_irg_opt(irg, "place");
596 if (firm_opt.confirm) {
597 /* Confirm construction currently can only handle blocks with only
598 one control flow predecessor. Calling optimize_cf here removes
599 Bad predecessors and help the optimization of switch constructs.
601 do_irg_opt(irg, "control-flow");
602 do_irg_opt(irg, "confirm");
603 do_irg_opt(irg, "vrp");
604 do_irg_opt(irg, "local");
607 do_irg_opt(irg, "control-flow");
608 do_irg_opt(irg, "opt-load-store");
609 do_irg_opt(irg, "fp-vrp");
610 do_irg_opt(irg, "lower");
611 do_irg_opt(irg, "deconv");
612 do_irg_opt(irg, "thread-jumps");
613 do_irg_opt(irg, "remove-confirms");
614 do_irg_opt(irg, "gvn-pre");
615 do_irg_opt(irg, "gcse");
616 do_irg_opt(irg, "place");
617 do_irg_opt(irg, "control-flow");
619 if (do_irg_opt(irg, "if-conversion")) {
620 do_irg_opt(irg, "local");
621 do_irg_opt(irg, "control-flow");
623 /* this doesn't make too much sense but tests the mux destruction... */
624 do_irg_opt(irg, "lower-mux");
626 do_irg_opt(irg, "bool");
627 do_irg_opt(irg, "shape-blocks");
628 do_irg_opt(irg, "ivopts");
629 do_irg_opt(irg, "local");
630 do_irg_opt(irg, "dead");
633 do_irp_opt("inline");
634 do_irp_opt("opt-proc-clone");
636 for (i = 0; i < get_irp_n_irgs(); i++) {
637 ir_graph *irg = get_irp_irg(i);
638 do_irg_opt(irg, "local");
639 do_irg_opt(irg, "control-flow");
640 do_irg_opt(irg, "thread-jumps");
641 do_irg_opt(irg, "local");
642 do_irg_opt(irg, "control-flow");
644 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
645 do_irg_opt(irg, "local");
646 do_irg_opt(irg, "vrp");
647 do_irg_opt(irg, "local");
648 do_irg_opt(irg, "vrp");
652 if (firm_dump.ir_graph) {
653 /* recompute backedges for nicer dumps */
654 for (i = 0; i < get_irp_n_irgs(); i++)
655 construct_cf_backedges(get_irp_irg(i));
660 if (firm_dump.statistic & STAT_AFTER_OPT)
661 stat_dump_snapshot(input_filename, "opt");
663 timer_stop(t_all_opt);
669 * @param input_filename the name of the (main) source file
671 static void do_firm_lowering(const char *input_filename)
675 /* enable architecture dependent optimizations */
676 arch_dep_set_opts((arch_dep_opts_t)
677 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
678 (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
679 (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
680 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
681 ir_graph *irg = get_irp_irg(i);
682 do_irg_opt(irg, "reassociation");
683 do_irg_opt(irg, "local");
686 do_irp_opt("target-lowering");
688 if (firm_dump.statistic & STAT_AFTER_LOWER)
689 stat_dump_snapshot(input_filename, "low");
691 timer_start(t_all_opt);
693 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
694 ir_graph *irg = get_irp_irg(i);
696 do_irg_opt(irg, "lower-blockcopy");
697 do_irg_opt(irg, "local");
698 do_irg_opt(irg, "deconv");
699 do_irg_opt(irg, "control-flow");
700 do_irg_opt(irg, "opt-load-store");
701 do_irg_opt(irg, "gcse");
702 do_irg_opt(irg, "place");
703 do_irg_opt(irg, "control-flow");
705 if (do_irg_opt(irg, "vrp")) {
706 do_irg_opt(irg, "local");
707 do_irg_opt(irg, "control-flow");
708 do_irg_opt(irg, "vrp");
709 do_irg_opt(irg, "local");
710 do_irg_opt(irg, "control-flow");
713 if (do_irg_opt(irg, "if-conversion")) {
714 do_irg_opt(irg, "local");
715 do_irg_opt(irg, "control-flow");
718 set_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2);
719 do_irg_opt(irg, "local");
721 do_irg_opt(irg, "parallelize-mem");
722 do_irg_opt(irg, "frame");
724 do_irp_opt("remove-unused");
725 do_irp_opt("opt-cc");
726 timer_stop(t_all_opt);
729 if (firm_dump.statistic & STAT_FINAL) {
730 stat_dump_snapshot(input_filename, "final");
735 * Initialize for the Firm-generating back end.
737 void gen_firm_init(void)
739 unsigned pattern = 0;
742 for (i = 0; i < n_opts; ++i) {
743 timers[i] = ir_timer_new();
744 timer_register(timers[i], opts[i].description);
746 t_verify = ir_timer_new();
747 timer_register(t_verify, "Firm: verify pass");
748 t_vcg_dump = ir_timer_new();
749 timer_register(t_vcg_dump, "Firm: vcg dumping");
750 t_all_opt = ir_timer_new();
751 timer_register(t_all_opt, "Firm: all optimizations");
753 if (firm_dump.stat_pattern)
754 pattern |= FIRMSTAT_PATTERN_ENABLED;
756 if (firm_dump.stat_dag)
757 pattern |= FIRMSTAT_COUNT_DAG;
760 firm_init_stat(firm_dump.statistic == STAT_NONE ?
761 0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
762 | FIRMSTAT_COUNT_CONSTS | pattern);
764 edges_init_dbg(firm_opt.verify_edges);
766 /* Sel node cannot produce NULL pointers */
767 set_opt_sel_based_null_check_elim(1);
769 /* dynamic dispatch works currently only if whole world scenarios */
770 set_opt_dyn_meth_dispatch(0);
772 /* do not run architecture dependent optimizations in building phase */
773 arch_dep_set_opts(arch_dep_none);
775 do_node_verification((firm_verification_t) firm_opt.verify);
777 ir_add_dump_flags(ir_dump_flag_group_extbb);
778 if (firm_dump.no_blocks)
779 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
782 set_opt_constant_folding(firm_opt.const_folding);
783 set_opt_algebraic_simplification(firm_opt.const_folding);
784 set_opt_cse(firm_opt.cse);
785 set_opt_global_cse(0);
786 set_opt_unreachable_code(1);
790 * Called, after the Firm generation is completed,
791 * do all optimizations and backend call here.
793 * @param out a file handle for the output, may be NULL
794 * @param input_filename the name of the (main) source file
795 * @param c_mode non-zero if "C" was compiled
797 void gen_firm_finish(FILE *out, const char *input_filename)
801 /* the general for dumping option must be set, or the others will not work*/
803 = (bool) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
805 ir_add_dump_flags(ir_dump_flag_keepalive_edges
806 | ir_dump_flag_consts_local | ir_dump_flag_dominance);
807 ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
809 /* FIXME: cloning might ADD new graphs. */
810 irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
812 if (firm_dump.all_types) {
813 dump_ir_prog_ext(dump_typegraph, "types.vcg");
816 /* finalize all graphs */
817 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
818 ir_graph *irg = get_irp_irg(i);
819 irg_finalize_cons(irg);
823 timer_push(t_verify);
827 /* all graphs are finalized, set the irp phase to high */
828 set_irp_phase_state(phase_high);
830 /* BEWARE: kill unreachable code before doing compound lowering */
831 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
832 ir_graph *irg = get_irp_irg(i);
833 do_irg_opt(irg, "control-flow");
834 do_irg_opt(irg, "lower-blockcopy");
837 if (firm_dump.statistic & STAT_BEFORE_OPT) {
838 stat_dump_snapshot(input_filename, "noopt");
841 do_firm_optimizations(input_filename);
842 do_firm_lowering(input_filename);
844 /* set the phase to low */
845 for (i = get_irp_n_irgs() - 1; i >= 0; --i)
846 set_irg_phase_state(get_irp_irg(i), phase_low);
848 if (firm_dump.statistic & STAT_FINAL_IR)
849 stat_dump_snapshot(input_filename, "final-ir");
851 /* run the code generator */
852 ir_timer_t *timer = ir_timer_new();
853 timer_register(timer, "Firm: backend");
855 be_main(out, input_filename);
858 if (firm_dump.statistic & STAT_FINAL)
859 stat_dump_snapshot(input_filename, "final");
862 static void disable_all_opts(void)
864 firm_opt.cse = false;
865 firm_opt.confirm = false;
866 firm_opt.muls = false;
867 firm_opt.divs = false;
868 firm_opt.mods = false;
869 firm_opt.alias_analysis = false;
870 firm_opt.strict_alias = false;
871 firm_opt.no_alias = false;
873 for (int i = 0; i < n_opts; ++i) {
874 opt_config_t *config = &opts[i];
875 if (config->flags & OPT_FLAG_ESSENTIAL) {
876 config->flags |= OPT_FLAG_ENABLED;
878 config->flags &= ~OPT_FLAG_ENABLED;
883 static bool firm_opt_option(const char *opt)
886 if (strncmp(opt, "no-", 3) == 0) {
891 opt_config_t *config = get_opt(opt);
892 if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
895 config->flags &= ~OPT_FLAG_ENABLED;
896 config->flags |= enable ? OPT_FLAG_ENABLED : 0;
900 void firm_option_help(print_option_help_func print_option_help)
902 print_option_help(firm_options[0].option, firm_options[0].description);
904 for (int i = 0; i < n_opts; ++i) {
908 const opt_config_t *config = &opts[i];
909 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
912 snprintf(buf, sizeof(buf), "-f%s", config->name);
913 snprintf(buf2, sizeof(buf2), "enable %s", config->description);
914 print_option_help(buf, buf2);
915 snprintf(buf, sizeof(buf), "-fno-%s", config->name);
916 snprintf(buf2, sizeof(buf2), "disable %s", config->description);
917 print_option_help(buf, buf2);
920 size_t const n_options = sizeof(firm_options)/sizeof(firm_options[0]);
921 for (size_t k = 0; k < n_options; ++k) {
924 snprintf(buf, sizeof(buf), "-f%s", firm_options[k].option);
925 snprintf(buf2, sizeof(buf2), "%s", firm_options[k].description);
926 print_option_help(buf, buf2);
930 int firm_option(const char *const opt)
933 if ((val = strstart(opt, "dump-filter="))) {
934 ir_set_dump_filter(val);
936 } else if ((val = strstart(opt, "clone-threshold="))) {
937 sscanf(val, "%d", &firm_opt.clone_threshold);
939 } else if ((val = strstart(opt, "inline-max-size="))) {
940 sscanf(val, "%u", &firm_opt.inline_maxsize);
942 } else if ((val = strstart(opt, "inline-threshold="))) {
943 sscanf(val, "%u", &firm_opt.inline_threshold);
945 } else if (streq(opt, "no-opt")) {
950 size_t const len = strlen(opt);
951 size_t const n_options = sizeof(firm_options)/sizeof(firm_options[0]);
952 for (size_t i = n_options; i != 0;) {
953 struct params const* const o = &firm_options[--i];
954 if (len == o->opt_len && strncmp(opt, o->option, len) == 0) {
955 /* statistic options do accumulate */
956 if (o->flag == &firm_dump.statistic)
957 *o->flag = (bool) (*o->flag | o->set);
965 /* maybe this enables/disables optimizations */
966 if (firm_opt_option(opt))
972 static void set_be_option(const char *arg)
974 int res = be_parse_arg(arg);
979 static void set_option(const char *arg)
981 int res = firm_option(arg);
986 void choose_optimization_pack(int level)
988 /* apply optimization level */
991 set_option("no-opt");
994 set_option("no-inline");
998 /* use_builtins = true; */
1001 set_option("thread-jumps");
1002 set_option("if-conversion");
1005 set_option("strict-aliasing");
1006 set_option("inline");
1007 set_option("fp-vrp");
1008 set_option("deconv");
1009 set_be_option("omitfp");
1015 * Do very early initializations
1017 void firm_early_init(void)
1019 /* arg: need this here for command line options */
1022 enable_safe_defaults();