adapt to libfirm API changes
[cparser] / driver / firm_opt.c
1 /**
2  *
3  * @file firm_opt.c -- Firm-generating back end optimizations.
4  *
5  * (C) 2005-2010  Michael Beck   beck@ipd.info.uni-karlsruhe.de
6  *
7  * $Id$
8  */
9
10 #include <config.h>
11
12 #include <stdbool.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include <stdbool.h>
16 #include <assert.h>
17 #include <libfirm/firm.h>
18
19 #include "firm_opt.h"
20 #include "firm_codegen.h"
21 #include "firm_cmdline.h"
22 #include "firm_timing.h"
23 #include "ast2firm.h"
24
25 static ir_timer_t *t_vcg_dump;
26 static ir_timer_t *t_verify;
27 static ir_timer_t *t_all_opt;
28 static bool do_irg_opt(ir_graph *irg, const char *name);
29
30 /** dump all the graphs depending on cond */
31
32 static void dump_all(const char *suffix)
33 {
34         if (!firm_dump.ir_graph)
35                 return;
36
37         timer_push(t_vcg_dump);
38         dump_all_ir_graphs(suffix);
39         timer_pop(t_vcg_dump);
40 }
41
42 /* entities of runtime functions */
43 ir_entity_ptr rts_entities[rts_max];
44
45 /**
46  * Map runtime functions.
47  */
48 static void rts_map(void)
49 {
50         static const struct {
51                 ir_entity_ptr *ent; /**< address of the rts entity */
52                 i_mapper_func func; /**< mapper function. */
53         } mapper[] = {
54                 /* integer */
55                 { &rts_entities[rts_abs],     i_mapper_abs },
56                 { &rts_entities[rts_labs],    i_mapper_abs },
57                 { &rts_entities[rts_llabs],   i_mapper_abs },
58                 { &rts_entities[rts_imaxabs], i_mapper_abs },
59
60                 /* double -> double */
61                 { &rts_entities[rts_fabs],    i_mapper_abs },
62                 { &rts_entities[rts_sqrt],    i_mapper_sqrt },
63                 { &rts_entities[rts_cbrt],    i_mapper_cbrt },
64                 { &rts_entities[rts_pow],     i_mapper_pow },
65                 { &rts_entities[rts_exp],     i_mapper_exp },
66                 { &rts_entities[rts_exp2],    i_mapper_exp },
67                 { &rts_entities[rts_exp10],   i_mapper_exp },
68                 { &rts_entities[rts_log],     i_mapper_log },
69                 { &rts_entities[rts_log2],    i_mapper_log2 },
70                 { &rts_entities[rts_log10],   i_mapper_log10 },
71                 { &rts_entities[rts_sin],     i_mapper_sin },
72                 { &rts_entities[rts_cos],     i_mapper_cos },
73                 { &rts_entities[rts_tan],     i_mapper_tan },
74                 { &rts_entities[rts_asin],    i_mapper_asin },
75                 { &rts_entities[rts_acos],    i_mapper_acos },
76                 { &rts_entities[rts_atan],    i_mapper_atan },
77                 { &rts_entities[rts_sinh],    i_mapper_sinh },
78                 { &rts_entities[rts_cosh],    i_mapper_cosh },
79                 { &rts_entities[rts_tanh],    i_mapper_tanh },
80
81                 /* float -> float */
82                 { &rts_entities[rts_fabsf],   i_mapper_abs },
83                 { &rts_entities[rts_sqrtf],   i_mapper_sqrt },
84                 { &rts_entities[rts_cbrtf],   i_mapper_cbrt },
85                 { &rts_entities[rts_powf],    i_mapper_pow },
86                 { &rts_entities[rts_expf],    i_mapper_exp },
87                 { &rts_entities[rts_exp2f],   i_mapper_exp },
88                 { &rts_entities[rts_exp10f],  i_mapper_exp },
89                 { &rts_entities[rts_logf],    i_mapper_log },
90                 { &rts_entities[rts_log2f],   i_mapper_log2 },
91                 { &rts_entities[rts_log10f],  i_mapper_log10 },
92                 { &rts_entities[rts_sinf],    i_mapper_sin },
93                 { &rts_entities[rts_cosf],    i_mapper_cos },
94                 { &rts_entities[rts_tanf],    i_mapper_tan },
95                 { &rts_entities[rts_asinf],   i_mapper_asin },
96                 { &rts_entities[rts_acosf],   i_mapper_acos },
97                 { &rts_entities[rts_atanf],   i_mapper_atan },
98                 { &rts_entities[rts_sinhf],   i_mapper_sinh },
99                 { &rts_entities[rts_coshf],   i_mapper_cosh },
100                 { &rts_entities[rts_tanhf],   i_mapper_tanh },
101
102                 /* long double -> long double */
103                 { &rts_entities[rts_fabsl],   i_mapper_abs },
104                 { &rts_entities[rts_sqrtl],   i_mapper_sqrt },
105                 { &rts_entities[rts_cbrtl],   i_mapper_cbrt },
106                 { &rts_entities[rts_powl],    i_mapper_pow },
107                 { &rts_entities[rts_expl],    i_mapper_exp },
108                 { &rts_entities[rts_exp2l],   i_mapper_exp },
109                 { &rts_entities[rts_exp10l],  i_mapper_exp },
110                 { &rts_entities[rts_logl],    i_mapper_log },
111                 { &rts_entities[rts_log2l],   i_mapper_log2 },
112                 { &rts_entities[rts_log10l],  i_mapper_log10 },
113                 { &rts_entities[rts_sinl],    i_mapper_sin },
114                 { &rts_entities[rts_cosl],    i_mapper_cos },
115                 { &rts_entities[rts_tanl],    i_mapper_tan },
116                 { &rts_entities[rts_asinl],   i_mapper_asin },
117                 { &rts_entities[rts_acosl],   i_mapper_acos },
118                 { &rts_entities[rts_atanl],   i_mapper_atan },
119                 { &rts_entities[rts_sinhl],   i_mapper_sinh },
120                 { &rts_entities[rts_coshl],   i_mapper_cosh },
121                 { &rts_entities[rts_tanhl],   i_mapper_tanh },
122
123                 /* string */
124                 { &rts_entities[rts_strcmp],  i_mapper_strcmp },
125                 { &rts_entities[rts_strncmp], i_mapper_strncmp },
126                 { &rts_entities[rts_strcpy],  i_mapper_strcpy },
127                 { &rts_entities[rts_strlen],  i_mapper_strlen },
128                 { &rts_entities[rts_memcpy],  i_mapper_memcpy },
129                 { &rts_entities[rts_mempcpy], i_mapper_mempcpy },
130                 { &rts_entities[rts_memmove], i_mapper_memmove },
131                 { &rts_entities[rts_memset],  i_mapper_memset },
132                 { &rts_entities[rts_memcmp],  i_mapper_memcmp }
133         };
134         i_record rec[sizeof(mapper)/sizeof(mapper[0])];
135         unsigned i, n_map;
136
137         for (i = n_map = 0; i < sizeof(mapper)/sizeof(mapper[0]); ++i) {
138                 if (*mapper[i].ent != NULL) {
139                         rec[n_map].i_call.kind     = INTRINSIC_CALL;
140                         rec[n_map].i_call.i_ent    = *mapper[i].ent;
141                         rec[n_map].i_call.i_mapper = mapper[i].func;
142                         rec[n_map].i_call.ctx      = NULL;
143                         rec[n_map].i_call.link     = NULL;
144                         ++n_map;
145                 }
146         }
147
148         if (n_map > 0)
149                 lower_intrinsics(rec, n_map, /* part_block_used=*/0);
150 }
151
152 static int *irg_dump_no;
153 static int firm_const_exists;
154
155 static void do_optimize_funccalls(void)
156 {
157         optimize_funccalls(firm_const_exists, NULL);
158 }
159
160 static void do_gcse(ir_graph *irg)
161 {
162         set_opt_global_cse(1);
163         optimize_graph_df(irg);
164         place_code(irg);
165         set_opt_global_cse(0);
166 }
167
168 static void do_lower_highlevel(ir_graph *irg)
169 {
170         lower_highlevel_graph(irg, firm_opt.lower_bitfields);
171 }
172
173 static void do_stred(ir_graph *irg)
174 {
175         opt_osr(irg, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
176 }
177
178 static void after_inline_opt(ir_graph *irg)
179 {
180         do_irg_opt(irg, "scalar-replace");
181         do_irg_opt(irg, "local");
182         do_irg_opt(irg, "control-flow");
183         do_irg_opt(irg, "combo");
184 }
185
186 static void do_inline(void)
187 {
188         inline_functions(firm_opt.inline_maxsize, firm_opt.inline_threshold,
189                          after_inline_opt);
190 }
191
192 static void do_cloning(void)
193 {
194         proc_cloning((float) firm_opt.clone_threshold);
195 }
196
197 static void do_lower_switch(ir_graph *irg)
198 {
199         lower_switch(irg, firm_opt.spare_size);
200 }
201
202 static void do_lower_mux(ir_graph *irg)
203 {
204         lower_mux(irg, NULL);
205 }
206
207 static void do_lower_for_target(void)
208 {
209         const backend_params *be_params = be_get_backend_param();
210         be_params->lower_for_target();
211 }
212
213 static void do_vrp(ir_graph *irg)
214 {
215         set_vrp_data(irg);
216 }
217
218 typedef enum opt_target {
219         OPT_TARGET_IRG, /**< optimization function works on a single graph */
220         OPT_TARGET_IRP  /**< optimization function works on the complete program */
221 } opt_target_t;
222
223 typedef enum opt_flags {
224         OPT_FLAG_NONE         = 0,
225         OPT_FLAG_ENABLED      = 1 << 0, /**< enable the optimization */
226         OPT_FLAG_NO_DUMP      = 1 << 1, /**< don't dump after transformation */
227         OPT_FLAG_NO_VERIFY    = 1 << 2, /**< don't verify after transformation */
228         OPT_FLAG_HIDE_OPTIONS = 1 << 3, /**< do not automatically process
229                                              -foptions for this transformation */
230         OPT_FLAG_ESSENTIAL    = 1 << 4, /**< output won't work without this pass
231                                              so we need it even with -O0 */
232 } opt_flags_t;
233
234 typedef void (*transform_irg_func)(ir_graph *irg);
235 typedef void (*transform_irp_func)(void);
236
237 typedef struct {
238         opt_target_t  target;
239         const char   *name;
240         union {
241                 transform_irg_func transform_irg;
242                 transform_irp_func transform_irp;
243         } u;
244         const char   *description;
245         opt_flags_t   flags;
246 } opt_config_t;
247
248 static opt_config_t opts[] = {
249 #define IRG(a, b, c, d) { OPT_TARGET_IRG, a, .u.transform_irg = (transform_irg_func)b, c, d }
250 #define IRP(a, b, c, d) { OPT_TARGET_IRP, a, .u.transform_irp = b,                     c, d }
251         IRG("bool",              opt_bool,                 "bool simplification",                                   OPT_FLAG_NONE),
252         IRG("combo",             combo,                    "combined CCE, UCE and GVN",                             OPT_FLAG_NONE),
253         IRG("confirm",           construct_confirms,       "confirm optimisation",                                  OPT_FLAG_HIDE_OPTIONS),
254         IRG("control-flow",      optimize_cf,              "optimization of control-flow",                          OPT_FLAG_HIDE_OPTIONS),
255         IRG("dead",              dead_node_elimination,    "dead node elimination",                                 OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
256         IRG("deconv",            conv_opt,                 "conv node elimination",                                 OPT_FLAG_NONE),
257         IRG("fp-vrp",            fixpoint_vrp,             "fixpoint value range propagation",                      OPT_FLAG_NONE),
258         IRG("frame",             opt_frame_irg,            "remove unused frame entities",                          OPT_FLAG_NONE),
259         IRG("gcse",              do_gcse,                  "global common subexpression elimination",               OPT_FLAG_NONE),
260         IRG("gvn-pre",           do_gvn_pre,               "global value numbering partial redundancy elimination", OPT_FLAG_NONE),
261         IRG("if-conversion",     opt_if_conv,              "if-conversion",                                         OPT_FLAG_NONE),
262         IRG("invert-loops",      do_loop_inversion,        "loop inversion",                                        OPT_FLAG_NONE),
263         IRG("ivopts",            do_stred,                 "induction variable strength reduction",                 OPT_FLAG_NONE),
264         IRG("local",             optimize_graph_df,        "local graph optimizations",                             OPT_FLAG_HIDE_OPTIONS),
265         IRG("lower",             do_lower_highlevel,       "lowering",                                              OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
266         IRG("lower-mux",         do_lower_mux,             "mux lowering",                                          OPT_FLAG_NONE),
267         IRG("lower-switch",      do_lower_switch,          "switch lowering",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
268         IRG("one-return",        normalize_one_return,     "normalisation to 1 return",                             OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
269         IRG("opt-load-store",    optimize_load_store,      "load store optimization",                               OPT_FLAG_NONE),
270         IRG("opt-tail-rec",      opt_tail_rec_irg,         "tail-recursion eliminiation",                           OPT_FLAG_NONE),
271         IRG("parallelize-mem",   opt_parallelize_mem,      "parallelize memory",                                    OPT_FLAG_NONE),
272         IRG("place",             place_code,               "code placement",                                        OPT_FLAG_NONE),
273         IRG("reassociation",     optimize_reassociation,   "reassociation",                                         OPT_FLAG_NONE),
274         IRG("remove-confirms",   remove_confirms,          "confirm removal",                                       OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
275         IRG("remove-phi-cycles", remove_phi_cycles,        "removal of phi cycles",                                 OPT_FLAG_HIDE_OPTIONS),
276         IRG("scalar-replace",    scalar_replacement_opt,   "scalar replacement",                                    OPT_FLAG_NONE),
277         IRG("shape-blocks",      shape_blocks,             "block shaping",                                         OPT_FLAG_NONE),
278         IRG("thread-jumps",      opt_jumpthreading,        "path-sensitive jumpthreading",                          OPT_FLAG_NONE),
279         IRG("unroll-loops",      do_loop_unrolling,        "loop unrolling",                                        OPT_FLAG_NONE),
280         IRG("vrp",               do_vrp,                   "value range propagation",                               OPT_FLAG_NONE),
281         IRP("inline",            do_inline,                "inlining",                                              OPT_FLAG_NONE),
282         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),
283         IRP("target-lowering",   do_lower_for_target,      "lowering necessary for target architecture",            OPT_FLAG_HIDE_OPTIONS | OPT_FLAG_ESSENTIAL),
284         IRP("opt-func-call",     do_optimize_funccalls,    "function call optimization",                            OPT_FLAG_NONE),
285         IRP("opt-proc-clone",    do_cloning,               "procedure cloning",                                     OPT_FLAG_NONE),
286         IRP("remove-unused",     garbage_collect_entities, "removal of unused functions/variables",                 OPT_FLAG_NO_DUMP | OPT_FLAG_NO_VERIFY),
287         IRP("rts",               rts_map,                  "optimization of known library functions",               OPT_FLAG_HIDE_OPTIONS),
288 #undef IRP
289 #undef IRG
290 };
291 static const int n_opts = sizeof(opts) / sizeof(opts[0]);
292 ir_timer_t *timers[sizeof(opts)/sizeof(opts[0])];
293
294 static opt_config_t *get_opt(const char *name)
295 {
296         int i;
297         for (i = 0; i < n_opts; ++i) {
298                 opt_config_t *config = &opts[i];
299                 if (strcmp(config->name, name) == 0)
300                         return config;
301         }
302
303         return NULL;
304 }
305
306 static void set_opt_enabled(const char *name, bool enabled)
307 {
308         opt_config_t *config = get_opt(name);
309         config->flags = (config->flags & ~OPT_FLAG_ENABLED)
310                 | (enabled ? OPT_FLAG_ENABLED : 0);
311 }
312
313 static bool get_opt_enabled(const char *name)
314 {
315         opt_config_t *config = get_opt(name);
316         return (config->flags & OPT_FLAG_ENABLED) != 0;
317 }
318
319 /**
320  * perform an optimisation on a single graph
321  *
322  * @return  true if something changed, false otherwise
323  */
324 static bool do_irg_opt(ir_graph *irg, const char *name)
325 {
326         ir_graph     *old_irg;
327         opt_config_t *config = get_opt(name);
328         size_t        n      = config - opts;
329         assert(config != NULL);
330         assert(config->target == OPT_TARGET_IRG);
331         if (! (config->flags & OPT_FLAG_ENABLED))
332                 return false;
333
334         old_irg          = current_ir_graph;
335         current_ir_graph = irg;
336
337         timer_push(timers[n]);
338         config->u.transform_irg(irg);
339         timer_pop(timers[n]);
340
341         if (firm_dump.all_phases && firm_dump.ir_graph) {
342                 dump_ir_graph(irg, name);
343         }
344
345         if (firm_opt.check_all) {
346                 timer_push(t_verify);
347                 irg_verify(irg, VERIFY_ENFORCE_SSA);
348                 timer_pop(t_verify);
349         }
350
351         current_ir_graph = old_irg;
352         return true;
353 }
354
355 static void do_irp_opt(const char *name)
356 {
357         opt_config_t *config = get_opt(name);
358         size_t        n      = config - opts;
359         assert(config->target == OPT_TARGET_IRP);
360         if (! (config->flags & OPT_FLAG_ENABLED))
361                 return;
362
363         timer_push(timers[n]);
364         config->u.transform_irp();
365         timer_pop(timers[n]);
366
367         if (firm_dump.ir_graph && firm_dump.all_phases) {
368                 int i;
369                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
370                         ir_graph *irg = get_irp_irg(i);
371                         dump_ir_graph(irg, name);
372                 }
373         }
374
375         if (firm_opt.check_all) {
376                 int i;
377                 timer_push(t_verify);
378                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
379                         irg_verify(get_irp_irg(i), VERIFY_ENFORCE_SSA);
380                 }
381                 timer_pop(t_verify);
382         }
383 }
384
385 /**
386  * Enable transformations which should be always safe (and cheap) to perform
387  */
388 static void enable_safe_defaults(void)
389 {
390         set_opt_enabled("remove-unused", true);
391         set_opt_enabled("opt-tail-rec", true);
392         set_opt_enabled("opt-func-call", true);
393         set_opt_enabled("reassociation", true);
394         set_opt_enabled("control-flow", true);
395         set_opt_enabled("local", true);
396         set_opt_enabled("lower-const", true);
397         set_opt_enabled("scalar-replace", true);
398         set_opt_enabled("place", true);
399         set_opt_enabled("confirm", true);
400         set_opt_enabled("opt-load-store", true);
401         set_opt_enabled("lower", true);
402         set_opt_enabled("deconv", true);
403         set_opt_enabled("remove-confirms", true);
404         set_opt_enabled("ivopts", true);
405         set_opt_enabled("dead", true);
406         set_opt_enabled("lower-switch", true);
407         set_opt_enabled("remove-phi-cycles", true);
408         set_opt_enabled("frame", true);
409         set_opt_enabled("combo", true);
410         set_opt_enabled("invert-loops", true);
411         set_opt_enabled("target-lowering", true);
412 }
413
414 /**
415  * run all the Firm optimizations
416  *
417  * @param input_filename     the name of the (main) source file
418  */
419 static void do_firm_optimizations(const char *input_filename)
420 {
421         int      i;
422         unsigned aa_opt;
423
424         set_opt_alias_analysis(firm_opt.alias_analysis);
425
426         aa_opt = aa_opt_no_opt;
427         if (firm_opt.strict_alias)
428                 aa_opt |= aa_opt_type_based | aa_opt_byte_type_may_alias;
429         if (firm_opt.no_alias)
430                 aa_opt = aa_opt_no_alias;
431
432         set_irp_memory_disambiguator_options(aa_opt);
433
434         /* parameter passing code should set them directly sometime... */
435         set_opt_enabled("rts", !firm_opt.freestanding);
436         set_opt_enabled("gcse", firm_opt.gcse);
437         set_opt_enabled("place", !firm_opt.gcse);
438         set_opt_enabled("confirm", firm_opt.confirm);
439         set_opt_enabled("remove-confirms", firm_opt.confirm);
440
441         /* osr supersedes remove_phi_cycles */
442         if (get_opt_enabled("ivopts"))
443                 set_opt_enabled("remove-phi-cycles", false);
444
445         timer_start(t_all_opt);
446
447         do_irp_opt("rts");
448
449         /* first step: kill dead code */
450         for (i = 0; i < get_irp_n_irgs(); i++) {
451                 ir_graph *irg = get_irp_irg(i);
452                 do_irg_opt(irg, "combo");
453                 do_irg_opt(irg, "local");
454                 do_irg_opt(irg, "control-flow");
455         }
456
457         do_irp_opt("remove-unused");
458         for (i = 0; i < get_irp_n_irgs(); ++i) {
459                 ir_graph *irg = get_irp_irg(i);
460                 do_irg_opt(irg, "opt-tail-rec");
461         }
462         do_irp_opt("opt-func-call");
463         do_irp_opt("lower-const");
464
465         for (i = 0; i < get_irp_n_irgs(); i++) {
466                 ir_graph *irg = get_irp_irg(i);
467
468                 do_irg_opt(irg, "scalar-replace");
469                 do_irg_opt(irg, "invert-loops");
470                 do_irg_opt(irg, "unroll-loops");
471                 do_irg_opt(irg, "local");
472                 do_irg_opt(irg, "reassociation");
473                 do_irg_opt(irg, "local");
474                 do_irg_opt(irg, "gcse");
475
476                 if (firm_opt.confirm) {
477                         /* Confirm construction currently can only handle blocks with only
478                            one control flow predecessor. Calling optimize_cf here removes
479                            Bad predecessors and help the optimization of switch constructs.
480                          */
481                         do_irg_opt(irg, "control-flow");
482                         do_irg_opt(irg, "confirm");
483                         do_irg_opt(irg, "vrp");
484                         do_irg_opt(irg, "local");
485                 }
486
487                 do_irg_opt(irg, "control-flow");
488                 do_irg_opt(irg, "opt-load-store");
489                 do_irg_opt(irg, "fp-vrp");
490                 do_irg_opt(irg, "lower");
491                 do_irg_opt(irg, "deconv");
492                 do_irg_opt(irg, "thread-jumps");
493                 do_irg_opt(irg, "remove-confirms");
494                 do_irg_opt(irg, "gvn-pre");
495                 do_irg_opt(irg, "place");
496                 do_irg_opt(irg, "control-flow");
497
498                 if (do_irg_opt(irg, "if-conversion")) {
499                         do_irg_opt(irg, "local");
500                         do_irg_opt(irg, "control-flow");
501                 }
502                 /* this doesn't make too much sense but tests the mux destruction... */
503                 do_irg_opt(irg, "lower-mux");
504
505                 do_irg_opt(irg, "bool");
506                 do_irg_opt(irg, "shape-blocks");
507                 do_irg_opt(irg, "lower-switch");
508                 do_irg_opt(irg, "ivopts");
509                 do_irg_opt(irg, "local");
510                 do_irg_opt(irg, "dead");
511                 do_irg_opt(irg, "frame");
512         }
513
514         do_irp_opt("inline");
515         do_irp_opt("opt-proc-clone");
516
517         for (i = 0; i < get_irp_n_irgs(); i++) {
518                 ir_graph *irg = get_irp_irg(i);
519                 do_irg_opt(irg, "local");
520                 do_irg_opt(irg, "control-flow");
521                 do_irg_opt(irg, "thread-jumps");
522                 do_irg_opt(irg, "local");
523                 do_irg_opt(irg, "control-flow");
524
525                 if( do_irg_opt(irg, "vrp") ) { // if vrp is enabled
526                         do_irg_opt(irg, "local");
527                         do_irg_opt(irg, "vrp");
528                         do_irg_opt(irg, "local");
529                         do_irg_opt(irg, "vrp");
530                 }
531         }
532
533         if (firm_dump.ir_graph) {
534                 /* recompute backedges for nicer dumps */
535                 for (i = 0; i < get_irp_n_irgs(); i++)
536                         construct_cf_backedges(get_irp_irg(i));
537         }
538
539         dump_all("opt");
540
541         if (firm_dump.statistic & STAT_AFTER_OPT)
542                 stat_dump_snapshot(input_filename, "opt");
543
544         timer_stop(t_all_opt);
545 }
546
547 /**
548  * do Firm lowering
549  *
550  * @param input_filename  the name of the (main) source file
551  */
552 static void do_firm_lowering(const char *input_filename)
553 {
554         int i;
555
556         do_irp_opt("target-lowering");
557
558         if (firm_dump.statistic & STAT_AFTER_LOWER)
559                 stat_dump_snapshot(input_filename, "low");
560
561         dump_all("low");
562
563         if (firm_opt.enabled) {
564                 timer_start(t_all_opt);
565
566                 /* run reassociation first on all graphs BEFORE the architecture
567                    dependent optimizations are enabled */
568                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
569                         ir_graph *irg = get_irp_irg(i);
570                         do_irg_opt(irg, "reassociation");
571                 }
572
573                 /* enable architecture dependent optimizations */
574                 arch_dep_set_opts((arch_dep_opts_t)
575                                 ((firm_opt.muls ? arch_dep_mul_to_shift : arch_dep_none) |
576                                  (firm_opt.divs ? arch_dep_div_by_const : arch_dep_none) |
577                                  (firm_opt.mods ? arch_dep_mod_by_const : arch_dep_none) ));
578
579                 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
580                         ir_graph *irg = get_irp_irg(i);
581
582                         do_irg_opt(irg, "local");
583                         do_irg_opt(irg, "gcse");
584                         do_irg_opt(irg, "control-flow");
585                         do_irg_opt(irg, "opt-load-store");
586                         do_irg_opt(irg, "local");
587                         do_irg_opt(irg, "control-flow");
588
589                         if (do_irg_opt(irg, "vrp")) {
590                                 do_irg_opt(irg, "local");
591                                 do_irg_opt(irg, "control-flow");
592                                 do_irg_opt(irg, "vrp");
593                                 do_irg_opt(irg, "local");
594                                 do_irg_opt(irg, "control-flow");
595                         }
596
597                         if (do_irg_opt(irg, "if-conversion")) {
598                                 do_irg_opt(irg, "local");
599                                 do_irg_opt(irg, "control-flow");
600                         }
601
602                         do_irg_opt(irg, "parallelize-mem");
603                 }
604                 timer_stop(t_all_opt);
605
606                 do_irp_opt("remove-unused");
607
608                 dump_all("low-opt");
609         }
610
611         if (firm_opt.cc_opt)
612                 mark_private_methods();
613
614         /* set the phase to low */
615         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
616                 ir_graph *irg = get_irp_irg(i);
617                 set_irg_phase_state(irg, phase_low);
618         }
619         set_irp_phase_state(phase_low);
620
621         if (firm_dump.statistic & STAT_FINAL) {
622                 stat_dump_snapshot(input_filename, "final");
623         }
624 }
625
626 /**
627  * Initialize for the Firm-generating back end.
628  */
629 void gen_firm_init(void)
630 {
631         unsigned pattern = 0;
632         int      i;
633
634         for (i = 0; i < n_opts; ++i) {
635                 timers[i] = ir_timer_new();
636                 timer_register(timers[i], opts[i].description);
637         }
638         t_verify = ir_timer_new();
639         timer_register(t_verify, "Firm: verify pass");
640         t_vcg_dump = ir_timer_new();
641         timer_register(t_vcg_dump, "Firm: vcg dumping");
642         t_all_opt = ir_timer_new();
643         timer_register(t_all_opt, "Firm: all optimizations");
644
645         if (firm_dump.stat_pattern)
646                 pattern |= FIRMSTAT_PATTERN_ENABLED;
647
648         if (firm_dump.stat_dag)
649                 pattern |= FIRMSTAT_COUNT_DAG;
650
651         ir_init(NULL);
652         firm_init_stat(firm_dump.statistic == STAT_NONE ?
653                         0 : FIRMSTAT_ENABLED | FIRMSTAT_COUNT_STRONG_OP
654                         | FIRMSTAT_COUNT_CONSTS | pattern);
655
656         edges_init_dbg(firm_opt.verify_edges);
657
658         /* Sel node cannot produce NULL pointers */
659         set_opt_sel_based_null_check_elim(1);
660
661         /* dynamic dispatch works currently only if whole world scenarios */
662         set_opt_dyn_meth_dispatch(0);
663
664         /* do not run architecture dependent optimizations in building phase */
665         arch_dep_set_opts(arch_dep_none);
666
667         do_node_verification((firm_verification_t) firm_opt.verify);
668         if (firm_dump.filter != NULL)
669                 ir_set_dump_filter(firm_dump.filter);
670         if (firm_dump.extbb)
671                 ir_add_dump_flags(ir_dump_flag_group_extbb);
672         if (firm_dump.no_blocks)
673                 ir_remove_dump_flags(ir_dump_flag_blocks_as_subgraphs);
674
675         if (firm_opt.enabled) {
676                 set_optimize(1);
677                 set_opt_constant_folding(firm_opt.const_folding);
678                 set_opt_algebraic_simplification(firm_opt.const_folding);
679                 set_opt_cse(firm_opt.cse);
680                 set_opt_global_cse(0);
681                 set_opt_unreachable_code(1);
682                 set_opt_control_flow(firm_opt.control_flow);
683                 set_opt_control_flow_weak_simplification(1);
684                 set_opt_control_flow_strong_simplification(1);
685         } else {
686                 set_optimize(0);
687         }
688 }
689
690 /**
691  * Called, after the Firm generation is completed,
692  * do all optimizations and backend call here.
693  *
694  * @param out                a file handle for the output, may be NULL
695  * @param input_filename     the name of the (main) source file
696  * @param c_mode             non-zero if "C" was compiled
697  * @param new_firm_const_exists  non-zero, if the const attribute was used on functions
698  */
699 void gen_firm_finish(FILE *out, const char *input_filename,
700                      int new_firm_const_exists)
701 {
702         int i;
703
704 #if 0
705         if (firm_opt.enable_statev) {
706                 char buf[1024];
707                 snprintf(buf, sizeof(buf), "%s.ev", input_filename);
708                 ir_stat_ev_begin(input_filename, firm_opt.statev_filter);
709                 ir_stat_ev_compilation_unit(input_filename);
710         }
711 #endif
712
713         firm_const_exists = new_firm_const_exists;
714
715         /* the general for dumping option must be set, or the others will not work*/
716         firm_dump.ir_graph
717                 = (a_byte) (firm_dump.ir_graph | firm_dump.all_phases | firm_dump.extbb);
718
719         ir_add_dump_flags(ir_dump_flag_keepalive_edges
720                         | ir_dump_flag_consts_local | ir_dump_flag_dominance);
721         ir_remove_dump_flags(ir_dump_flag_loops | ir_dump_flag_ld_names);
722
723         /* FIXME: cloning might ADD new graphs. */
724         irg_dump_no = calloc(get_irp_last_idx(), sizeof(*irg_dump_no));
725
726         if (firm_dump.all_types) {
727                 dump_ir_prog_ext(dump_typegraph, "types.vcg");
728         }
729
730         /* finalize all graphs */
731         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
732                 ir_graph *irg = get_irp_irg(i);
733                 irg_finalize_cons(irg);
734         }
735         dump_all("");
736
737         timer_push(t_verify);
738         tr_verify();
739         timer_pop(t_verify);
740
741         /* all graphs are finalized, set the irp phase to high */
742         set_irp_phase_state(phase_high);
743
744         /* BEWARE: kill unreachable code before doing compound lowering */
745         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
746                 ir_graph *irg = get_irp_irg(i);
747                 do_irg_opt(irg, "control-flow");
748         }
749
750         /* lower all compound call return values */
751         lower_compound_params();
752
753         /* lower copyb nodes */
754         for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
755                 ir_graph *irg = get_irp_irg(i);
756                 lower_CopyB(irg, 128, 4);
757         }
758
759         if (firm_dump.statistic & STAT_BEFORE_OPT) {
760                 stat_dump_snapshot(input_filename, "noopt");
761         }
762
763         if (firm_opt.enabled)
764                 do_firm_optimizations(input_filename);
765
766         if (firm_opt.lower)
767                 do_firm_lowering(input_filename);
768
769         /* set the phase to low */
770         for (i = get_irp_n_irgs() - 1; i >= 0; --i)
771                 set_irg_phase_state(get_irp_irg(i), phase_low);
772
773         if (firm_dump.statistic & STAT_FINAL_IR)
774                 stat_dump_snapshot(input_filename, "final-ir");
775
776         /* run the code generator */
777         if (firm_be_opt.selection != BE_NONE)
778                 do_codegen(out, input_filename);
779
780         if (firm_dump.statistic & STAT_FINAL)
781                 stat_dump_snapshot(input_filename, "final");
782 }
783
784 void disable_all_opts(void)
785 {
786         for (int i = 0; i < n_opts; ++i) {
787                 opt_config_t *config = &opts[i];
788                 if (config->flags & OPT_FLAG_ESSENTIAL) {
789                         config->flags |= OPT_FLAG_ENABLED;
790                 } else {
791                         config->flags &= ~OPT_FLAG_ENABLED;
792                 }
793         }
794 }
795
796 int firm_opt_option(const char *opt)
797 {
798         bool enable = true;
799         if (strncmp(opt, "no-", 3) == 0) {
800                 enable = false;
801                 opt = opt + 3;
802         }
803
804         opt_config_t *config = get_opt(opt);
805         if (config == NULL || (config->flags & OPT_FLAG_HIDE_OPTIONS))
806                 return 0;
807
808         config->flags &= ~OPT_FLAG_ENABLED;
809         config->flags |= enable ? OPT_FLAG_ENABLED : 0;
810         return 1;
811 }
812
813 void firm_opt_option_help(void)
814 {
815         int i;
816
817         for (i = 0; i < n_opts; ++i) {
818                 char buf[1024];
819                 char buf2[1024];
820
821                 const opt_config_t *config = &opts[i];
822                 if (config->flags & OPT_FLAG_HIDE_OPTIONS)
823                         continue;
824
825                 snprintf(buf2, sizeof(buf2), "firm: enable %s", config->description);
826                 print_option_help(config->name, buf2);
827                 snprintf(buf, sizeof(buf), "no-%s", config->name);
828                 snprintf(buf2, sizeof(buf2), "firm: disable %s", config->description);
829                 print_option_help(buf, buf2);
830         }
831 }
832
833 /**
834  * Do very early initializations
835  */
836 void firm_early_init(void)
837 {
838         /* arg: need this here for command line options */
839         be_opt_register();
840
841         enable_safe_defaults();
842 }