Commit | Line | Data |
---|---|---|
4a71579b | 1 | /* |
79bfeef6 | 2 | * Copyright (C) 2012-2023 Free Software Foundation, Inc. |
4a71579b PC |
3 | * |
4 | * This file is part of GNU lightning. | |
5 | * | |
6 | * GNU lightning is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU Lesser General Public License as published | |
8 | * by the Free Software Foundation; either version 3, or (at your option) | |
9 | * any later version. | |
10 | * | |
11 | * GNU lightning is distributed in the hope that it will be useful, but | |
12 | * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
13 | * or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public | |
14 | * License for more details. | |
15 | * | |
16 | * Authors: | |
17 | * Paulo Cesar Pereira de Andrade | |
18 | */ | |
19 | ||
20 | #include <lightning.h> | |
21 | #include <lightning/jit_private.h> | |
40a44dcb PC |
22 | #if HAVE_MMAP |
23 | # include <sys/mman.h> | |
24 | #endif | |
4a71579b PC |
25 | #if defined(__sgi) |
26 | # include <fcntl.h> | |
27 | #endif | |
28 | ||
29 | #ifndef MAP_ANON | |
30 | # define MAP_ANON MAP_ANONYMOUS | |
31 | # ifndef MAP_ANONYMOUS | |
32 | # define MAP_ANONYMOUS 0 | |
33 | # endif | |
34 | #endif | |
35 | ||
36 | #define jit_regload_reload 0 /* convert to reload */ | |
37 | #define jit_regload_delete 1 /* just remove node */ | |
38 | #define jit_regload_isdead 2 /* delete and unset live bit */ | |
39 | ||
40 | /* | |
41 | * Prototypes | |
42 | */ | |
43 | static jit_word_t hash_data(const void*, jit_word_t); | |
44 | ||
45 | #define new_pool() _new_pool(_jit) | |
46 | static void _new_pool(jit_state_t*); | |
47 | ||
48 | #define new_node(u) _new_node(_jit, u) | |
49 | static jit_node_t *_new_node(jit_state_t*, jit_code_t); | |
50 | ||
51 | #define link_node(u) _link_node(_jit, u) | |
52 | static inline jit_node_t *_link_node(jit_state_t*, jit_node_t*); | |
53 | ||
54 | #define del_node(u, v) _del_node(_jit, u, v) | |
55 | static inline void _del_node(jit_state_t*, jit_node_t*, jit_node_t*); | |
56 | ||
57 | #define free_node(u) _free_node(_jit, u) | |
58 | static inline void _free_node(jit_state_t*, jit_node_t*); | |
59 | ||
60 | #define del_label(u, v) _del_label(_jit, u, v) | |
61 | static void _del_label(jit_state_t*, jit_node_t*, jit_node_t*); | |
62 | ||
63 | #define jit_dataset() _jit_dataset(_jit) | |
64 | static void | |
65 | _jit_dataset(jit_state_t *_jit); | |
66 | ||
c0c16242 PC |
67 | #define block_update_set(block, target) _block_update_set(_jit, block, target) |
68 | static jit_bool_t _block_update_set(jit_state_t*, jit_block_t*, jit_block_t*); | |
69 | ||
24d91c0d PC |
70 | #define propagate_backward(block) _propagate_backward(_jit, block) |
71 | static void _propagate_backward(jit_state_t*, jit_block_t*); | |
72 | ||
c0c16242 PC |
73 | #define check_block_again() _check_block_again(_jit) |
74 | static jit_bool_t _check_block_again(jit_state_t*); | |
75 | ||
76 | #define do_setup() _do_setup(_jit) | |
77 | static void _do_setup(jit_state_t*); | |
78 | ||
4a71579b PC |
79 | #define jit_setup(block) _jit_setup(_jit, block) |
80 | static void | |
81 | _jit_setup(jit_state_t *_jit, jit_block_t *block); | |
82 | ||
c0c16242 PC |
83 | #define do_follow(always) _do_follow(_jit, always) |
84 | static void _do_follow(jit_state_t*, jit_bool_t); | |
85 | ||
86 | #define jit_follow(block) _jit_follow(_jit, block) | |
4a71579b | 87 | static void |
c0c16242 | 88 | _jit_follow(jit_state_t *_jit, jit_block_t *block); |
4a71579b PC |
89 | |
90 | #define jit_update(node, live, mask) _jit_update(_jit, node, live, mask) | |
91 | static void | |
92 | _jit_update(jit_state_t *_jit, jit_node_t *node, | |
93 | jit_regset_t *live, jit_regset_t *mask); | |
94 | ||
95 | #define thread_jumps() _thread_jumps(_jit) | |
96 | static void | |
97 | _thread_jumps(jit_state_t *_jit); | |
98 | ||
99 | #define sequential_labels() _sequential_labels(_jit) | |
100 | static void | |
101 | _sequential_labels(jit_state_t *_jit); | |
102 | ||
103 | #define split_branches() _split_branches(_jit) | |
104 | static void | |
105 | _split_branches(jit_state_t *_jit); | |
106 | ||
107 | #define shortcut_jump(prev, node) _shortcut_jump(_jit, prev, node) | |
108 | static jit_bool_t | |
109 | _shortcut_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node); | |
110 | ||
111 | #define redundant_jump(prev, node) _redundant_jump(_jit, prev, node) | |
112 | static jit_bool_t | |
113 | _redundant_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node); | |
114 | ||
115 | static jit_code_t | |
116 | reverse_jump_code(jit_code_t code); | |
117 | ||
118 | #define reverse_jump(prev, node) _reverse_jump(_jit, prev, node) | |
119 | static jit_bool_t | |
120 | _reverse_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node); | |
121 | ||
122 | #define redundant_store(node, jump) _redundant_store(_jit, node, jump) | |
ba3814c1 | 123 | static jit_bool_t |
4a71579b PC |
124 | _redundant_store(jit_state_t *_jit, jit_node_t *node, jit_bool_t jump); |
125 | ||
126 | #define simplify_movr(p, n, k, s) _simplify_movr(_jit, p, n, k, s) | |
127 | static jit_bool_t | |
128 | _simplify_movr(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node, | |
129 | jit_int32_t kind, jit_int32_t size); | |
130 | ||
131 | #define simplify_movi(p, n, k, s) _simplify_movi(_jit, p, n, k, s) | |
132 | static jit_bool_t | |
133 | _simplify_movi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node, | |
134 | jit_int32_t kind, jit_int32_t size); | |
135 | ||
136 | #define simplify_ldxi(prev, node) _simplify_ldxi(_jit, prev, node) | |
137 | static jit_bool_t | |
138 | _simplify_ldxi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node); | |
139 | ||
140 | #define simplify_stxi(prev, node) _simplify_stxi(_jit, prev, node) | |
141 | static jit_bool_t | |
142 | _simplify_stxi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node); | |
143 | ||
144 | #define simplify_spill(node, regno) _simplify_spill(_jit, node, regno) | |
145 | static void | |
146 | _simplify_spill(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno); | |
147 | ||
148 | #define simplify() _simplify(_jit) | |
ba3814c1 | 149 | static jit_bool_t |
4a71579b PC |
150 | _simplify(jit_state_t *_jit); |
151 | ||
152 | #define jit_reg_undef -1 | |
153 | #define jit_reg_static 0 | |
154 | #define jit_reg_change 1 | |
155 | #define register_change_p(n, l, r) _register_change_p(_jit, n, l, r) | |
156 | static jit_int32_t | |
157 | _register_change_p(jit_state_t *_jit, jit_node_t *node, jit_node_t *link, | |
158 | jit_int32_t regno); | |
159 | ||
160 | #define spill_reglive_p(node, regno) _spill_reglive_p(_jit, node, regno) | |
161 | static jit_bool_t | |
162 | _spill_reglive_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno); | |
163 | ||
164 | #define patch_registers() _patch_registers(_jit) | |
165 | static void | |
166 | _patch_registers(jit_state_t *_jit); | |
167 | ||
168 | #define patch_register(n,l,r,p) _patch_register(_jit,n,l,r,p) | |
169 | static void | |
170 | _patch_register(jit_state_t *jit, jit_node_t *node, jit_node_t *link, | |
171 | jit_int32_t regno, jit_int32_t patch); | |
172 | ||
173 | /* | |
174 | * Initialization | |
175 | */ | |
176 | #if !defined(__sgi) | |
177 | #define mmap_fd -1 | |
178 | #endif | |
179 | ||
180 | /* | |
181 | * Implementation | |
182 | */ | |
183 | void | |
b68d544b | 184 | init_jit_with_debug(const char *progname, FILE *dbg_out) |
4a71579b PC |
185 | { |
186 | jit_get_cpu(); | |
b68d544b | 187 | jit_init_debug(progname, dbg_out); |
4a71579b PC |
188 | jit_init_size(); |
189 | } | |
190 | ||
b68d544b PC |
191 | void |
192 | init_jit(const char *progname) | |
193 | { | |
194 | init_jit_with_debug(progname, stderr); | |
195 | } | |
196 | ||
4a71579b PC |
197 | void |
198 | finish_jit(void) | |
199 | { | |
200 | jit_finish_debug(); | |
201 | jit_finish_size(); | |
202 | } | |
203 | ||
204 | jit_int32_t | |
205 | _jit_get_reg(jit_state_t *_jit, jit_int32_t regspec) | |
206 | { | |
207 | jit_int32_t spec; | |
208 | jit_int32_t regno; | |
209 | ||
210 | spec = regspec & ~(jit_class_chk|jit_class_nospill); | |
211 | if (spec & jit_class_named) { | |
212 | regno = jit_regno(spec); | |
213 | if (jit_regset_tstbit(&_jitc->regsav, regno)) | |
214 | /* fail if register is spilled */ | |
215 | goto fail; | |
216 | if (jit_regset_tstbit(&_jitc->regarg, regno)) | |
217 | /* fail if register is an argument to current instruction */ | |
218 | goto fail; | |
219 | if (jit_regset_tstbit(&_jitc->reglive, regno)) { | |
220 | if (regspec & jit_class_nospill) | |
221 | /* fail if register is live and should not spill/reload */ | |
222 | goto fail; | |
223 | goto spill; | |
224 | } | |
225 | jit_regset_setbit(&_jitc->regarg, regno); | |
226 | return (regno); | |
227 | } | |
228 | else | |
229 | assert(jit_class(spec) != 0); | |
230 | ||
231 | if (_jitc->emit) { | |
232 | /* search for a free register matching spec */ | |
233 | for (regno = 0; regno < _jitc->reglen; regno++) { | |
234 | if ((jit_class(_rvs[regno].spec) & spec) == spec && | |
235 | !jit_regset_tstbit(&_jitc->regarg, regno) && | |
79bfeef6 PC |
236 | !jit_regset_tstbit(&_jitc->reglive, regno)) { |
237 | if (jit_regset_tstbit(&_jitc->regmask, regno)) { | |
238 | /* search further, attempting to find a truly known | |
239 | * free register, not just one in unknown state. */ | |
240 | jit_int32_t regfree; | |
241 | ||
242 | for (regfree = regno + 1; | |
243 | regfree < _jitc->reglen; regfree++) { | |
244 | if ((jit_class(_rvs[regfree].spec) & spec) == spec && | |
245 | !jit_regset_tstbit(&_jitc->regarg, regfree) && | |
246 | !jit_regset_tstbit(&_jitc->reglive, regfree) && | |
247 | !jit_regset_tstbit(&_jitc->regmask, regfree)) { | |
248 | regno = regfree; | |
249 | break; | |
250 | } | |
251 | } | |
252 | } | |
4a71579b | 253 | goto regarg; |
79bfeef6 | 254 | } |
4a71579b PC |
255 | } |
256 | ||
257 | /* search for a register matching spec that is not an argument | |
258 | * for the current instruction */ | |
259 | for (regno = 0; regno < _jitc->reglen; regno++) { | |
260 | if ((jit_class(_rvs[regno].spec) & spec) == spec && | |
261 | !jit_regset_tstbit(&_jitc->regsav, regno) && | |
262 | !jit_regset_tstbit(&_jitc->regarg, regno) && | |
263 | !(regspec & jit_class_nospill)) { | |
264 | spill: | |
265 | assert(_jitc->function != NULL); | |
266 | if (spec & jit_class_gpr) { | |
267 | if (!_jitc->function->regoff[regno]) { | |
268 | _jitc->function->regoff[regno] = | |
269 | jit_allocai(sizeof(jit_word_t)); | |
270 | _jitc->again = 1; | |
271 | } | |
272 | #if DEBUG | |
273 | /* emit_stxi must not need temporary registers */ | |
274 | assert(!_jitc->getreg); | |
275 | _jitc->getreg = 1; | |
276 | #endif | |
277 | emit_stxi(_jitc->function->regoff[regno], JIT_FP, regno); | |
278 | #if DEBUG | |
279 | _jitc->getreg = 0; | |
280 | #endif | |
281 | } | |
282 | else { | |
283 | if (!_jitc->function->regoff[regno]) { | |
284 | _jitc->function->regoff[regno] = | |
285 | jit_allocai(sizeof(jit_float64_t)); | |
286 | _jitc->again = 1; | |
287 | } | |
288 | #if DEBUG | |
289 | /* emit_stxi must not need temporary registers */ | |
290 | assert(!_jitc->getreg); | |
291 | _jitc->getreg = 1; | |
292 | #endif | |
293 | emit_stxi_d(_jitc->function->regoff[regno], JIT_FP, regno); | |
294 | #if DEBUG | |
295 | _jitc->getreg = 0; | |
296 | #endif | |
297 | } | |
298 | jit_regset_setbit(&_jitc->regsav, regno); | |
299 | regarg: | |
300 | jit_regset_setbit(&_jitc->regarg, regno); | |
301 | if (jit_class(_rvs[regno].spec) & jit_class_sav) { | |
302 | /* if will modify callee save registers without a | |
303 | * function prolog, better patch this assertion */ | |
304 | assert(_jitc->function != NULL); | |
305 | if (!jit_regset_tstbit(&_jitc->function->regset, regno)) { | |
306 | jit_regset_setbit(&_jitc->function->regset, regno); | |
307 | _jitc->again = 1; | |
308 | } | |
309 | } | |
310 | return (regno); | |
311 | } | |
312 | } | |
313 | } | |
314 | else { | |
315 | /* nospill hint only valid during emit" */ | |
316 | assert(!(regspec & jit_class_nospill)); | |
317 | for (regno = 0; regno < _jitc->reglen; regno++) { | |
318 | if ((jit_class(_rvs[regno].spec) & spec) == spec && | |
319 | !jit_regset_tstbit(&_jitc->regsav, regno) && | |
320 | !jit_regset_tstbit(&_jitc->regarg, regno)) { | |
321 | jit_regset_setbit(&_jitc->regarg, regno); | |
322 | jit_regset_setbit(&_jitc->regsav, regno); | |
323 | jit_save(regno); | |
324 | return (jit_regno_patch|regno); | |
325 | } | |
326 | } | |
327 | } | |
328 | ||
329 | /* Out of hardware registers */ | |
330 | fail: | |
331 | assert(regspec & jit_class_chk); | |
332 | return (JIT_NOREG); | |
333 | } | |
334 | ||
335 | void | |
336 | _jit_unget_reg(jit_state_t *_jit, jit_int32_t regno) | |
337 | { | |
338 | regno = jit_regno(regno); | |
339 | if (jit_regset_tstbit(&_jitc->regsav, regno)) { | |
340 | if (_jitc->emit) { | |
341 | #if DEBUG | |
342 | /* emit_ldxi must not need a temporary register */ | |
343 | assert(!_jitc->getreg); | |
344 | _jitc->getreg = 1; | |
345 | #endif | |
346 | if (jit_class(_rvs[regno].spec) & jit_class_gpr) | |
347 | emit_ldxi(regno, JIT_FP, _jitc->function->regoff[regno]); | |
348 | else | |
349 | emit_ldxi_d(regno, JIT_FP, _jitc->function->regoff[regno]); | |
350 | #if DEBUG | |
351 | /* emit_ldxi must not need a temporary register */ | |
352 | _jitc->getreg = 0; | |
353 | #endif | |
354 | } | |
355 | else | |
356 | jit_load(regno); | |
357 | jit_regset_clrbit(&_jitc->regsav, regno); | |
358 | } | |
359 | #if defined(jit_carry) | |
360 | assert((regno == jit_carry /*&& _NOREG != jit_carry*/) || | |
361 | jit_regset_tstbit(&_jitc->regarg, regno) != 0); | |
362 | #else | |
363 | assert(jit_regset_tstbit(&_jitc->regarg, regno) != 0); | |
364 | #endif | |
365 | jit_regset_clrbit(&_jitc->regarg, regno); | |
366 | } | |
367 | ||
368 | jit_bool_t | |
369 | _jit_callee_save_p(jit_state_t *_jit, jit_int32_t regno) | |
370 | { | |
371 | assert(regno >= 0 && regno < JIT_NOREG); | |
372 | return (!!(_rvs[regno].spec & jit_class_sav)); | |
373 | } | |
374 | ||
375 | extern jit_bool_t | |
376 | _jit_pointer_p(jit_state_t *_jit, jit_pointer_t address) | |
377 | { | |
378 | return ((jit_uint8_t *)address >= _jit->code.ptr && | |
379 | (jit_word_t)address < _jit->pc.w); | |
380 | } | |
381 | ||
382 | #if __ia64__ | |
383 | void | |
384 | jit_regset_com(jit_regset_t *u, jit_regset_t *v) | |
385 | { | |
386 | u->rl = ~v->rl; u->rh = ~v->rh; | |
387 | u->fl = ~v->fl; u->fh = ~v->fh; | |
388 | } | |
389 | ||
390 | void | |
391 | jit_regset_and(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
392 | { | |
393 | u->rl = v->rl & w->rl; u->rh = v->rh & w->rh; | |
394 | u->fl = v->fl & w->fl; u->fh = v->fh & w->fh; | |
395 | } | |
396 | ||
397 | void | |
398 | jit_regset_ior(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
399 | { | |
400 | u->rl = v->rl | w->rl; u->rh = v->rh | w->rh; | |
401 | u->fl = v->fl | w->fl; u->fh = v->fh | w->fh; | |
402 | } | |
403 | ||
404 | void | |
405 | jit_regset_xor(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
406 | { | |
407 | u->rl = v->rl ^ w->rl; u->rh = v->rh ^ w->rh; | |
408 | u->fl = v->fl ^ w->fl; u->fh = v->fh ^ w->fh; | |
409 | } | |
410 | ||
411 | void | |
412 | jit_regset_set(jit_regset_t *u, jit_regset_t *v) | |
413 | { | |
414 | u->rl = v->rl; u->rh = v->rh; | |
415 | u->fl = v->fl; u->fh = v->fh; | |
416 | } | |
417 | ||
418 | void | |
419 | jit_regset_set_mask(jit_regset_t *u, jit_int32_t v) | |
420 | { | |
421 | jit_bool_t w = !!(v & (v - 1)); | |
422 | ||
423 | assert(v >= 0 && v <= 256); | |
424 | if (v == 0) | |
425 | u->rl = u->rh = u->fl = u->fh = -1LL; | |
426 | else if (v <= 64) { | |
427 | u->rl = w ? (1LL << v) - 1 : -1LL; | |
428 | u->rh = u->fl = u->fh = 0; | |
429 | } | |
430 | else if (v <= 128) { | |
431 | u->rl = -1LL; | |
432 | u->rh = w ? (1LL << (v - 64)) - 1 : -1LL; | |
433 | u->fl = u->fh = 0; | |
434 | } | |
435 | else if (v <= 192) { | |
436 | u->rl = u->rh = -1LL; | |
437 | u->fl = w ? (1LL << (v - 128)) - 1 : -1LL; | |
438 | u->fh = 0; | |
439 | } | |
440 | else { | |
441 | u->rl = u->rh = u->fl = -1LL; | |
442 | u->fh = w ? (1LL << (v - 128)) - 1 : -1LL; | |
443 | } | |
444 | } | |
445 | ||
446 | jit_bool_t | |
447 | jit_regset_cmp_ui(jit_regset_t *u, jit_word_t v) | |
448 | { | |
449 | return !((u->rl == v && u->rh == 0 && u->fl == 0 && u->fh == 0)); | |
450 | } | |
451 | ||
452 | void | |
453 | jit_regset_set_ui(jit_regset_t *u, jit_word_t v) | |
454 | { | |
455 | u->rl = v; | |
456 | u->rh = u->fl = u->fh = 0; | |
457 | } | |
458 | ||
459 | jit_bool_t | |
460 | jit_regset_set_p(jit_regset_t *u) | |
461 | { | |
462 | return (u->rl || u->rh || u->fl || u->fh); | |
463 | } | |
464 | ||
465 | void | |
466 | jit_regset_clrbit(jit_regset_t *set, jit_int32_t bit) | |
467 | { | |
468 | assert(bit >= 0 && bit <= 255); | |
469 | if (bit < 64) | |
470 | set->rl &= ~(1LL << bit); | |
471 | else if (bit < 128) | |
472 | set->rh &= ~(1LL << (bit - 64)); | |
473 | else if (bit < 192) | |
474 | set->fl &= ~(1LL << (bit - 128)); | |
475 | else | |
476 | set->fh &= ~(1LL << (bit - 192)); | |
477 | } | |
478 | ||
479 | void | |
480 | jit_regset_setbit(jit_regset_t *set, jit_int32_t bit) | |
481 | { | |
482 | assert(bit >= 0 && bit <= 255); | |
483 | if (bit < 64) | |
484 | set->rl |= 1LL << bit; | |
485 | else if (bit < 128) | |
486 | set->rh |= 1LL << (bit - 64); | |
487 | else if (bit < 192) | |
488 | set->fl |= 1LL << (bit - 128); | |
489 | else | |
490 | set->fh |= 1LL << (bit - 192); | |
491 | } | |
492 | ||
493 | jit_bool_t | |
494 | jit_regset_tstbit(jit_regset_t *set, jit_int32_t bit) | |
495 | { | |
496 | assert(bit >= 0 && bit <= 255); | |
497 | if (bit < 64) | |
498 | return (!!(set->rl & (1LL << bit))); | |
499 | else if (bit < 128) | |
500 | return (!!(set->rh & (1LL << (bit - 64)))); | |
501 | else if (bit < 192) | |
502 | return (!!(set->fl & (1LL << (bit - 128)))); | |
503 | return (!!(set->fh & (1LL << (bit - 192)))); | |
504 | } | |
505 | ||
506 | unsigned long | |
507 | jit_regset_scan1(jit_regset_t *set, jit_int32_t offset) | |
508 | { | |
509 | assert(offset >= 0 && offset <= 255); | |
510 | for (; offset < 64; offset++) { | |
511 | if (set->rl & (1LL << offset)) | |
512 | return (offset); | |
513 | } | |
514 | for (; offset < 128; offset++) { | |
515 | if (set->rh & (1LL << (offset - 64))) | |
516 | return (offset); | |
517 | } | |
518 | for (; offset < 192; offset++) { | |
519 | if (set->fl & (1LL << (offset - 128))) | |
520 | return (offset); | |
521 | } | |
522 | for (; offset < 256; offset++) { | |
523 | if (set->fh & (1LL << (offset - 192))) | |
524 | return (offset); | |
525 | } | |
526 | return (ULONG_MAX); | |
527 | } | |
528 | ||
529 | #elif __sparc__ && __WORDSIZE == 64 | |
530 | void | |
531 | jit_regset_com(jit_regset_t *u, jit_regset_t *v) | |
532 | { | |
533 | u->rl = ~v->rl; u->rh = ~v->rh; | |
534 | } | |
535 | ||
536 | void | |
537 | jit_regset_and(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
538 | { | |
539 | u->rl = v->rl & w->rl; u->rh = v->rh & w->rh; | |
540 | } | |
541 | ||
542 | void | |
543 | jit_regset_ior(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
544 | { | |
545 | u->rl = v->rl | w->rl; u->rh = v->rh | w->rh; | |
546 | } | |
547 | ||
548 | void | |
549 | jit_regset_xor(jit_regset_t *u, jit_regset_t *v, jit_regset_t *w) | |
550 | { | |
551 | u->rl = v->rl ^ w->rl; u->rh = v->rh ^ w->rh; | |
552 | } | |
553 | ||
554 | void | |
555 | jit_regset_set(jit_regset_t *u, jit_regset_t *v) | |
556 | { | |
557 | u->rl = v->rl; u->rh = v->rh; | |
558 | } | |
559 | ||
560 | void | |
561 | jit_regset_set_mask(jit_regset_t *u, jit_int32_t v) | |
562 | { | |
563 | jit_bool_t w = !!(v & (v - 1)); | |
564 | ||
565 | assert(v >= 0 && v <= 128); | |
566 | if (v == 0) | |
567 | u->rl = u->rh = -1LL; | |
568 | else if (v <= 64) { | |
569 | u->rl = w ? (1LL << v) - 1 : -1LL; | |
570 | u->rh = 0; | |
571 | } | |
572 | else { | |
573 | u->rl = -1LL; | |
574 | u->rh = w ? (1LL << (v - 64)) - 1 : -1LL; | |
575 | } | |
576 | } | |
577 | ||
578 | jit_bool_t | |
579 | jit_regset_cmp_ui(jit_regset_t *u, jit_word_t v) | |
580 | { | |
581 | return !((u->rl == v && u->rh == 0)); | |
582 | } | |
583 | ||
584 | void | |
585 | jit_regset_set_ui(jit_regset_t *u, jit_word_t v) | |
586 | { | |
587 | u->rl = v; | |
588 | u->rh = 0; | |
589 | } | |
590 | ||
591 | jit_bool_t | |
592 | jit_regset_set_p(jit_regset_t *u) | |
593 | { | |
594 | return (u->rl || u->rh); | |
595 | } | |
596 | ||
597 | void | |
598 | jit_regset_clrbit(jit_regset_t *set, jit_int32_t bit) | |
599 | { | |
600 | assert(bit >= 0 && bit <= 128); | |
601 | if (bit < 64) | |
602 | set->rl &= ~(1LL << bit); | |
603 | else | |
604 | set->rh &= ~(1LL << (bit - 64)); | |
605 | } | |
606 | ||
607 | void | |
608 | jit_regset_setbit(jit_regset_t *set, jit_int32_t bit) | |
609 | { | |
610 | assert(bit >= 0 && bit <= 127); | |
611 | if (bit < 64) | |
612 | set->rl |= 1LL << bit; | |
613 | else | |
614 | set->rh |= 1LL << (bit - 64); | |
615 | } | |
616 | ||
617 | jit_bool_t | |
618 | jit_regset_tstbit(jit_regset_t *set, jit_int32_t bit) | |
619 | { | |
620 | assert(bit >= 0 && bit <= 127); | |
621 | if (bit < 64) | |
622 | return (!!(set->rl & (1LL << bit))); | |
623 | else | |
624 | return (!!(set->rh & (1LL << (bit - 64)))); | |
625 | } | |
626 | ||
627 | unsigned long | |
628 | jit_regset_scan1(jit_regset_t *set, jit_int32_t offset) | |
629 | { | |
630 | assert(offset >= 0 && offset <= 127); | |
631 | for (; offset < 64; offset++) { | |
632 | if (set->rl & (1LL << offset)) | |
633 | return (offset); | |
634 | } | |
635 | for (; offset < 128; offset++) { | |
636 | if (set->rh & (1LL << (offset - 64))) | |
637 | return (offset); | |
638 | } | |
639 | return (ULONG_MAX); | |
640 | } | |
641 | ||
642 | #else | |
643 | unsigned long | |
644 | jit_regset_scan1(jit_regset_t *set, jit_int32_t offset) | |
645 | { | |
646 | jit_regset_t mask; | |
647 | assert(offset >= 0 && offset <= 63); | |
648 | if ((mask = *set >> offset)) { | |
649 | for (;;) { | |
650 | if (mask & 1) | |
651 | return (offset); | |
652 | mask >>= 1; | |
653 | ++offset; | |
654 | } | |
655 | } | |
656 | return (ULONG_MAX); | |
657 | } | |
658 | #endif | |
659 | ||
660 | void | |
661 | _jit_save(jit_state_t *_jit, jit_int32_t reg) | |
662 | { | |
663 | reg = jit_regno(reg); | |
664 | assert(!_jitc->realize); | |
665 | _jitc->spill[reg] = jit_new_node_w(jit_code_save, reg); | |
666 | } | |
667 | ||
668 | void | |
669 | _jit_load(jit_state_t *_jit, jit_int32_t reg) | |
670 | { | |
671 | jit_node_t *node; | |
672 | ||
673 | reg = jit_regno(reg); | |
674 | assert(!_jitc->realize); | |
675 | assert(_jitc->spill[reg] != NULL); | |
676 | node = jit_new_node_w(jit_code_load, reg); | |
677 | /* create a path to flag the save/load is not required */ | |
678 | node->link = _jitc->spill[reg]; | |
679 | node->link->link = node; | |
680 | _jitc->spill[reg] = NULL; | |
681 | } | |
682 | ||
683 | static jit_word_t | |
684 | hash_data(const void *data, jit_word_t length) | |
685 | { | |
686 | const jit_uint8_t *ptr; | |
687 | jit_word_t i, key; | |
688 | for (i = key = 0, ptr = data; i < length; i++) | |
689 | key = (key << (key & 1)) ^ ptr[i]; | |
690 | return (key); | |
691 | } | |
692 | ||
693 | jit_pointer_t | |
694 | _jit_address(jit_state_t *_jit, jit_node_t *node) | |
695 | { | |
696 | assert(_jitc->done); | |
697 | assert(node != NULL && | |
698 | /* If a node type that is documented to be a fixed marker */ | |
699 | (node->code == jit_code_note || node->code == jit_code_name || | |
700 | /* If another special fixed marker, returned by jit_indirect() */ | |
701 | (node->code == jit_code_label && (node->flag & jit_flag_use) != 0))); | |
702 | return ((jit_pointer_t)node->u.w); | |
703 | } | |
704 | ||
705 | jit_node_t * | |
706 | _jit_data(jit_state_t *_jit, const void *data, | |
707 | jit_word_t length, jit_int32_t align) | |
708 | { | |
709 | jit_word_t key; | |
710 | jit_node_t *node; | |
711 | ||
712 | assert(!_jitc->realize); | |
713 | ||
714 | /* Ensure there is space even if asking for a duplicate */ | |
715 | if (((_jitc->data.offset + 7) & -8) + length > _jit->data.length) { | |
716 | jit_word_t size; | |
717 | ||
718 | size = (_jit->data.length + length + 4096) & - 4095; | |
719 | assert(size >= _jit->data.length); | |
720 | if (_jitc->data.ptr == NULL) | |
721 | jit_alloc((jit_pointer_t *)&_jitc->data.ptr, size); | |
722 | else | |
723 | jit_realloc((jit_pointer_t *)&_jitc->data.ptr, | |
724 | _jit->data.length, size); | |
725 | _jit->data.length = size; | |
726 | } | |
727 | if (_jitc->data.table == NULL) | |
728 | jit_alloc((jit_pointer_t *)&_jitc->data.table, | |
729 | (_jitc->data.size = 16) * sizeof(jit_node_t*)); | |
730 | ||
731 | key = hash_data(data, length) & (_jitc->data.size - 1); | |
732 | node = _jitc->data.table[key]; | |
733 | for (; node; node = node->next) { | |
734 | if (node->v.w == length && | |
735 | memcmp(_jitc->data.ptr + node->u.w, data, length) == 0) | |
736 | break; | |
737 | } | |
738 | ||
739 | if (!node) { | |
740 | node = jit_new_node_no_link(jit_code_data); | |
741 | if (!align) | |
742 | align = length; | |
743 | switch (align) { | |
744 | case 0: case 1: | |
745 | break; | |
746 | case 2: | |
747 | _jitc->data.offset = (_jitc->data.offset + 1) & -2; | |
748 | break; | |
749 | case 3: case 4: | |
750 | _jitc->data.offset = (_jitc->data.offset + 3) & -4; | |
751 | break; | |
752 | default: | |
753 | _jitc->data.offset = (_jitc->data.offset + 7) & -8; | |
754 | break; | |
755 | } | |
756 | node->u.w = _jitc->data.offset; | |
757 | node->v.w = length; | |
758 | jit_memcpy(_jitc->data.ptr + _jitc->data.offset, data, length); | |
759 | _jitc->data.offset += length; | |
760 | ||
761 | node->next = _jitc->data.table[key]; | |
762 | _jitc->data.table[key] = node; | |
763 | ++_jitc->data.count; | |
764 | ||
765 | /* Rehash if more than 75% used table */ | |
766 | if (_jitc->data.count > | |
767 | (_jitc->data.size >> 1) + (_jitc->data.size >> 2) && | |
768 | (_jitc->data.size << 1) > _jitc->data.size) { | |
769 | jit_word_t i; | |
770 | jit_node_t **hash; | |
771 | jit_node_t *next; | |
772 | jit_node_t *temp; | |
773 | ||
774 | jit_alloc((jit_pointer_t *)&hash, | |
775 | (_jitc->data.size << 1) * sizeof(jit_node_t*)); | |
776 | for (i = 0; i < _jitc->data.size; i++) { | |
777 | temp = _jitc->data.table[i]; | |
778 | for (; temp; temp = next) { | |
779 | next = temp->next; | |
780 | key = hash_data(_jitc->data.ptr + temp->u.w, temp->v.w) & | |
781 | ((_jitc->data.size << 1) - 1); | |
782 | temp->next = hash[key]; | |
783 | hash[key] = temp; | |
784 | } | |
785 | } | |
786 | jit_free((jit_pointer_t *)&_jitc->data.table); | |
787 | _jitc->data.table = hash; | |
788 | _jitc->data.size <<= 1; | |
789 | } | |
790 | } | |
791 | ||
792 | return (node); | |
793 | } | |
794 | ||
795 | static void | |
796 | _new_pool(jit_state_t *_jit) | |
797 | { | |
798 | jit_node_t *list; | |
799 | jit_int32_t offset; | |
800 | ||
801 | if (_jitc->pool.offset >= _jitc->pool.length) { | |
802 | jit_int32_t length; | |
803 | ||
804 | length = _jitc->pool.length + 16; | |
805 | jit_realloc((jit_pointer_t *)&_jitc->pool.ptr, | |
806 | _jitc->pool.length * sizeof(jit_node_t *), | |
807 | length * sizeof(jit_node_t *)); | |
808 | _jitc->pool.length = length; | |
809 | } | |
810 | jit_alloc((jit_pointer_t *)(_jitc->pool.ptr + _jitc->pool.offset), | |
811 | sizeof(jit_node_t) * 1024); | |
812 | list = _jitc->pool.ptr[_jitc->pool.offset]; | |
813 | for (offset = 1; offset < 1024; offset++, list++) | |
814 | list->next = list + 1; | |
815 | list->next = _jitc->list; | |
816 | _jitc->list = _jitc->pool.ptr[_jitc->pool.offset]; | |
817 | ++_jitc->pool.offset; | |
818 | } | |
819 | ||
820 | static jit_node_t * | |
821 | _new_node(jit_state_t *_jit, jit_code_t code) | |
822 | { | |
823 | jit_node_t *node; | |
824 | ||
825 | if (_jitc->list == NULL) | |
826 | new_pool(); | |
827 | node = _jitc->list; | |
828 | _jitc->list = node->next; | |
829 | if (_jitc->synth) | |
830 | node->flag |= jit_flag_synth; | |
831 | node->next = NULL; | |
832 | node->code = code; | |
833 | ||
834 | return (node); | |
835 | } | |
836 | ||
837 | static inline jit_node_t * | |
838 | _link_node(jit_state_t *_jit, jit_node_t *node) | |
839 | { | |
840 | if (_jitc->tail) | |
841 | _jitc->tail->next = node; | |
842 | else | |
843 | _jitc->head = node; | |
844 | return (_jitc->tail = node); | |
845 | } | |
846 | ||
847 | static inline void | |
848 | _del_node(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
849 | { | |
850 | if (prev == node) { | |
851 | assert(prev == _jitc->head); | |
852 | _jitc->head = node->next; | |
853 | } | |
854 | else | |
855 | prev->next = node->next; | |
856 | memset(node, 0, sizeof(jit_node_t)); | |
857 | node->next = _jitc->list; | |
858 | _jitc->list = node; | |
859 | } | |
860 | ||
861 | static inline void | |
862 | _free_node(jit_state_t *_jit, jit_node_t *node) | |
863 | { | |
864 | memset(node, 0, sizeof(jit_node_t)); | |
865 | node->next = _jitc->list; | |
866 | _jitc->list = node; | |
867 | } | |
868 | ||
869 | static void | |
870 | _del_label(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
871 | { | |
872 | jit_block_t *block; | |
873 | ||
874 | /* only allow call to del_label on linked labels */ | |
875 | block = _jitc->blocks.ptr + node->v.w; | |
876 | assert(block->label == node); | |
877 | ||
878 | /* del_label() should only be called when optimizing. | |
879 | * This will leave an empty block index */ | |
880 | jit_regset_del(&block->reglive); | |
881 | jit_regset_del(&block->regmask); | |
882 | block->label = NULL; | |
883 | ||
884 | /* redundant, should be already true */ | |
885 | assert(node->link == NULL); | |
886 | del_node(prev, node); | |
887 | } | |
888 | ||
889 | jit_state_t * | |
890 | jit_new_state(void) | |
891 | { | |
892 | jit_state_t *_jit; | |
893 | ||
894 | jit_alloc((jit_pointer_t *)&_jit, sizeof(jit_state_t)); | |
895 | jit_alloc((jit_pointer_t *)&_jitc, sizeof(jit_compiler_t)); | |
896 | jit_regset_new(&_jitc->regarg); | |
897 | jit_regset_new(&_jitc->regsav); | |
898 | jit_regset_new(&_jitc->reglive); | |
899 | jit_regset_new(&_jitc->regmask); | |
79bfeef6 | 900 | jit_regset_new(&_jitc->explive); |
4a71579b PC |
901 | |
902 | jit_init(); | |
903 | ||
904 | jit_alloc((jit_pointer_t *)&_jitc->spill, | |
905 | _jitc->reglen * sizeof(jit_node_t*)); | |
906 | jit_alloc((jit_pointer_t *)&_jitc->gen, | |
907 | _jitc->reglen * sizeof(jit_int32_t)); | |
908 | jit_alloc((jit_pointer_t *)&_jitc->values, | |
909 | _jitc->reglen * sizeof(jit_value_t)); | |
910 | ||
911 | jit_alloc((jit_pointer_t *)&_jitc->patches.ptr, | |
912 | (_jitc->patches.length = 1024) * sizeof(jit_patch_t)); | |
913 | jit_alloc((jit_pointer_t *)&_jitc->functions.ptr, | |
914 | (_jitc->functions.length = 16) * sizeof(jit_function_t)); | |
915 | jit_alloc((jit_pointer_t *)&_jitc->pool.ptr, | |
916 | (_jitc->pool.length = 16) * sizeof(jit_node_t*)); | |
917 | jit_alloc((jit_pointer_t *)&_jitc->blocks.ptr, | |
918 | (_jitc->blocks.length = 16) * sizeof(jit_block_t)); | |
919 | #if __arm__ && DISASSEMBLER | |
920 | jit_alloc((jit_pointer_t *)&_jitc->data_info.ptr, | |
921 | (_jitc->data_info.length = 1024) * sizeof(jit_data_info_t)); | |
922 | #endif | |
923 | ||
924 | /* allocate at most one extra note in case jit_name() is | |
925 | * never called, or called after adding at least one note */ | |
926 | _jit->note.length = 1; | |
927 | _jitc->note.size = sizeof(jit_note_t); | |
928 | ||
929 | return (_jit); | |
930 | } | |
931 | ||
932 | void | |
933 | _jit_clear_state(jit_state_t *_jit) | |
934 | { | |
935 | #if DEVEL_DISASSEMBLER | |
936 | # define jit_really_clear_state() _jit_really_clear_state(_jit) | |
937 | } | |
938 | ||
939 | void _jit_really_clear_state(jit_state_t *_jit) | |
940 | { | |
941 | #endif | |
942 | jit_word_t offset; | |
943 | jit_function_t *function; | |
944 | ||
945 | /* release memory not required at jit execution time and set | |
946 | * pointers to NULL to explicitly know they are released */ | |
947 | _jitc->head = _jitc->tail = NULL; | |
948 | ||
949 | jit_free((jit_pointer_t *)&_jitc->data.table); | |
950 | _jitc->data.size = _jitc->data.count = 0; | |
951 | ||
952 | jit_free((jit_pointer_t *)&_jitc->spill); | |
953 | jit_free((jit_pointer_t *)&_jitc->gen); | |
954 | jit_free((jit_pointer_t *)&_jitc->values); | |
955 | ||
956 | jit_free((jit_pointer_t *)&_jitc->blocks.ptr); | |
957 | ||
958 | jit_free((jit_pointer_t *)&_jitc->patches.ptr); | |
959 | _jitc->patches.offset = _jitc->patches.length = 0; | |
960 | ||
961 | for (offset = 0; offset < _jitc->functions.offset; offset++) { | |
962 | function = _jitc->functions.ptr + offset; | |
963 | jit_free((jit_pointer_t *)&function->regoff); | |
964 | } | |
965 | jit_free((jit_pointer_t *)&_jitc->functions.ptr); | |
966 | _jitc->functions.offset = _jitc->functions.length = 0; | |
967 | _jitc->function = NULL; | |
968 | ||
969 | for (offset = 0; offset < _jitc->pool.offset; offset++) | |
970 | jit_free((jit_pointer_t *)(_jitc->pool.ptr + offset)); | |
971 | jit_free((jit_pointer_t *)&_jitc->pool.ptr); | |
972 | _jitc->pool.offset = _jitc->pool.length = 0; | |
973 | _jitc->list = NULL; | |
974 | ||
975 | _jitc->note.head = _jitc->note.tail = | |
976 | _jitc->note.name = _jitc->note.note = NULL; | |
977 | _jitc->note.base = NULL; | |
978 | ||
979 | #if __arm__ && DISASSEMBLER | |
980 | jit_free((jit_pointer_t *)&_jitc->data_info.ptr); | |
981 | #endif | |
982 | ||
983 | #if (__powerpc__ && _CALL_AIXDESC) || __ia64__ | |
984 | jit_free((jit_pointer_t *)&_jitc->prolog.ptr); | |
985 | #endif | |
986 | ||
987 | #if __ia64__ | |
988 | jit_regset_del(&_jitc->regs); | |
989 | #endif | |
990 | ||
991 | jit_free((jit_pointer_t *)&_jitc); | |
992 | } | |
993 | ||
994 | void | |
995 | _jit_destroy_state(jit_state_t *_jit) | |
996 | { | |
997 | #if DEVEL_DISASSEMBLER | |
998 | jit_really_clear_state(); | |
999 | #endif | |
40a44dcb | 1000 | #if HAVE_MMAP |
4a71579b PC |
1001 | if (!_jit->user_code) |
1002 | munmap(_jit->code.ptr, _jit->code.length); | |
1003 | if (!_jit->user_data) | |
1004 | munmap(_jit->data.ptr, _jit->data.length); | |
40a44dcb | 1005 | #endif |
4a71579b PC |
1006 | jit_free((jit_pointer_t *)&_jit); |
1007 | } | |
1008 | ||
1009 | void | |
1010 | _jit_synth_inc(jit_state_t *_jit) | |
1011 | { | |
1012 | assert(_jitc->synth < 8); | |
1013 | ++_jitc->synth; | |
1014 | } | |
1015 | ||
1016 | jit_node_t * | |
1017 | _jit_new_node(jit_state_t *_jit, jit_code_t code) | |
1018 | { | |
1019 | assert(!_jitc->realize); | |
1020 | return (link_node(new_node(code))); | |
1021 | } | |
1022 | ||
1023 | jit_node_t * | |
1024 | _jit_new_node_no_link(jit_state_t *_jit, jit_code_t code) | |
1025 | { | |
1026 | assert(!_jitc->realize); | |
1027 | return (new_node(code)); | |
1028 | } | |
1029 | ||
1030 | void | |
1031 | _jit_link_node(jit_state_t *_jit, jit_node_t *node) | |
1032 | { | |
1033 | assert(!_jitc->realize); | |
1034 | link_node(node); | |
1035 | } | |
1036 | ||
1037 | void | |
1038 | _jit_synth_dec(jit_state_t *_jit) | |
1039 | { | |
1040 | assert(_jitc->synth > 0); | |
1041 | --_jitc->synth; | |
1042 | } | |
1043 | ||
1044 | jit_node_t * | |
1045 | _jit_new_node_w(jit_state_t *_jit, jit_code_t code, | |
1046 | jit_word_t u) | |
1047 | { | |
1048 | jit_node_t *node = new_node(code); | |
1049 | assert(!_jitc->realize); | |
1050 | node->u.w = u; | |
1051 | return (link_node(node)); | |
1052 | } | |
1053 | ||
1054 | jit_node_t * | |
1055 | _jit_new_node_f(jit_state_t *_jit, jit_code_t code, | |
1056 | jit_float32_t u) | |
1057 | { | |
1058 | jit_node_t *node = new_node(code); | |
1059 | assert(!_jitc->realize); | |
1060 | node->u.f = u; | |
1061 | return (link_node(node)); | |
1062 | } | |
1063 | ||
1064 | jit_node_t * | |
1065 | _jit_new_node_d(jit_state_t *_jit, jit_code_t code, | |
1066 | jit_float64_t u) | |
1067 | { | |
1068 | jit_node_t *node = new_node(code); | |
1069 | assert(!_jitc->realize); | |
1070 | node->u.d = u; | |
1071 | return (link_node(node)); | |
1072 | } | |
1073 | ||
1074 | jit_node_t * | |
1075 | _jit_new_node_p(jit_state_t *_jit, jit_code_t code, | |
1076 | jit_pointer_t u) | |
1077 | { | |
1078 | jit_node_t *node = new_node(code); | |
1079 | assert(!_jitc->realize); | |
1080 | node->u.p = u; | |
1081 | return (link_node(node)); | |
1082 | } | |
1083 | ||
1084 | jit_node_t * | |
1085 | _jit_new_node_ww(jit_state_t *_jit, jit_code_t code, | |
1086 | jit_word_t u, jit_word_t v) | |
1087 | { | |
1088 | jit_node_t *node = new_node(code); | |
1089 | assert(!_jitc->realize); | |
1090 | node->u.w = u; | |
1091 | node->v.w = v; | |
1092 | return (link_node(node)); | |
1093 | } | |
1094 | ||
1095 | jit_node_t * | |
1096 | _jit_new_node_wp(jit_state_t *_jit, jit_code_t code, | |
1097 | jit_word_t u, jit_pointer_t v) | |
1098 | { | |
1099 | return (jit_new_node_ww(code, u, (jit_word_t)v)); | |
1100 | } | |
1101 | ||
1102 | jit_node_t * | |
1103 | _jit_new_node_fp(jit_state_t *_jit, jit_code_t code, | |
1104 | jit_float32_t u, jit_pointer_t v) | |
1105 | { | |
1106 | jit_node_t *node = new_node(code); | |
1107 | assert(!_jitc->realize); | |
1108 | node->u.f = u; | |
1109 | node->v.w = (jit_word_t)v; | |
1110 | return (link_node(node)); | |
1111 | } | |
1112 | ||
1113 | jit_node_t * | |
1114 | _jit_new_node_dp(jit_state_t *_jit, jit_code_t code, | |
1115 | jit_float64_t u, jit_pointer_t v) | |
1116 | { | |
1117 | jit_node_t *node = new_node(code); | |
1118 | assert(!_jitc->realize); | |
1119 | node->u.d = u; | |
1120 | node->v.w = (jit_word_t)v; | |
1121 | return (link_node(node)); | |
1122 | } | |
1123 | ||
1124 | jit_node_t * | |
1125 | _jit_new_node_pw(jit_state_t *_jit, jit_code_t code, | |
1126 | jit_pointer_t u, jit_word_t v) | |
1127 | { | |
1128 | return (jit_new_node_ww(code, (jit_word_t)u, v)); | |
1129 | } | |
1130 | ||
1131 | jit_node_t * | |
1132 | _jit_new_node_wf(jit_state_t *_jit, jit_code_t code, | |
1133 | jit_word_t u, jit_float32_t v) | |
1134 | { | |
1135 | jit_node_t *node = new_node(code); | |
1136 | assert(!_jitc->realize); | |
1137 | node->u.w = u; | |
1138 | node->v.f = v; | |
1139 | return (link_node(node)); | |
1140 | } | |
1141 | ||
1142 | jit_node_t * | |
1143 | _jit_new_node_wd(jit_state_t *_jit, jit_code_t code, | |
1144 | jit_word_t u, jit_float64_t v) | |
1145 | { | |
1146 | jit_node_t *node = new_node(code); | |
1147 | assert(!_jitc->realize); | |
1148 | node->u.w = u; | |
1149 | node->v.d = v; | |
1150 | return (link_node(node)); | |
1151 | } | |
1152 | ||
1153 | jit_node_t * | |
1154 | _jit_new_node_www(jit_state_t *_jit, jit_code_t code, | |
1155 | jit_word_t u, jit_word_t v, jit_word_t w) | |
1156 | { | |
1157 | jit_node_t *node = new_node(code); | |
1158 | assert(!_jitc->realize); | |
1159 | node->u.w = u; | |
1160 | node->v.w = v; | |
1161 | node->w.w = w; | |
1162 | return (link_node(node)); | |
1163 | } | |
1164 | ||
1165 | jit_node_t * | |
1166 | _jit_new_node_qww(jit_state_t *_jit, jit_code_t code, | |
1167 | jit_int32_t l, jit_int32_t h, | |
1168 | jit_word_t v, jit_word_t w) | |
1169 | { | |
1170 | jit_node_t *node = new_node(code); | |
1171 | assert(!_jitc->realize); | |
1172 | assert(l != h); | |
1173 | node->u.q.l = l; | |
1174 | node->u.q.h = h; | |
1175 | node->v.w = v; | |
1176 | node->w.w = w; | |
1177 | return (link_node(node)); | |
1178 | } | |
1179 | ||
ba86ff93 PC |
1180 | jit_node_t * |
1181 | _jit_new_node_wqw(jit_state_t *_jit, jit_code_t code, | |
1182 | jit_word_t u, jit_int32_t l, | |
1183 | jit_int32_t h, jit_word_t w) | |
1184 | { | |
1185 | jit_node_t *node = new_node(code); | |
1186 | assert(!_jitc->realize); | |
1187 | assert(l != h); | |
1188 | node->u.w = u; | |
1189 | node->v.q.l = l; | |
1190 | node->v.q.h = h; | |
1191 | node->w.w = w; | |
1192 | return (link_node(node)); | |
1193 | } | |
1194 | ||
ba3814c1 PC |
1195 | jit_node_t * |
1196 | _jit_new_node_wwq(jit_state_t *_jit, jit_code_t code, | |
1197 | jit_word_t u, jit_word_t v, | |
1198 | jit_int32_t l, jit_int32_t h) | |
1199 | { | |
1200 | jit_node_t *node = new_node(code); | |
1201 | assert(!_jitc->realize); | |
1202 | node->u.w = u; | |
1203 | node->v.w = v; | |
1204 | node->w.q.l = l; | |
1205 | node->w.q.h = h; | |
1206 | return (link_node(node)); | |
1207 | } | |
1208 | ||
4a71579b PC |
1209 | jit_node_t * |
1210 | _jit_new_node_wwf(jit_state_t *_jit, jit_code_t code, | |
1211 | jit_word_t u, jit_word_t v, jit_float32_t w) | |
1212 | { | |
1213 | jit_node_t *node = new_node(code); | |
1214 | assert(!_jitc->realize); | |
1215 | node->u.w = u; | |
1216 | node->v.w = v; | |
1217 | node->w.f = w; | |
1218 | return (link_node(node)); | |
1219 | } | |
1220 | ||
ba86ff93 PC |
1221 | jit_node_t * |
1222 | _jit_new_node_wqf(jit_state_t *_jit, jit_code_t code, | |
1223 | jit_word_t u, jit_int32_t l, | |
1224 | jit_int32_t h, jit_float32_t w) | |
1225 | { | |
1226 | jit_node_t *node = new_node(code); | |
1227 | assert(!_jitc->realize); | |
1228 | assert(l != h); | |
1229 | node->u.w = u; | |
1230 | node->v.q.l = l; | |
1231 | node->v.q.h = h; | |
1232 | node->w.f = w; | |
1233 | return (link_node(node)); | |
1234 | } | |
1235 | ||
4a71579b PC |
1236 | jit_node_t * |
1237 | _jit_new_node_wwd(jit_state_t *_jit, jit_code_t code, | |
1238 | jit_word_t u, jit_word_t v, jit_float64_t w) | |
1239 | { | |
1240 | jit_node_t *node = new_node(code); | |
1241 | assert(!_jitc->realize); | |
1242 | node->u.w = u; | |
1243 | node->v.w = v; | |
1244 | node->w.d = w; | |
1245 | return (link_node(node)); | |
1246 | } | |
1247 | ||
ba86ff93 PC |
1248 | jit_node_t * |
1249 | _jit_new_node_wqd(jit_state_t *_jit, jit_code_t code, | |
1250 | jit_word_t u, jit_int32_t l, | |
1251 | jit_int32_t h, jit_float64_t w) | |
1252 | { | |
1253 | jit_node_t *node = new_node(code); | |
1254 | assert(!_jitc->realize); | |
1255 | assert(l != h); | |
1256 | node->u.w = u; | |
1257 | node->v.q.l = l; | |
1258 | node->v.q.h = h; | |
1259 | node->w.d = w; | |
1260 | return (link_node(node)); | |
1261 | } | |
1262 | ||
4a71579b PC |
1263 | jit_node_t * |
1264 | _jit_new_node_pww(jit_state_t *_jit, jit_code_t code, | |
1265 | jit_pointer_t u, jit_word_t v, jit_word_t w) | |
1266 | { | |
1267 | jit_node_t *node = new_node(code); | |
1268 | assert(!_jitc->realize); | |
1269 | node->u.p = u; | |
1270 | node->v.w = v; | |
1271 | node->w.w = w; | |
1272 | return (link_node(node)); | |
1273 | } | |
1274 | ||
1275 | jit_node_t * | |
1276 | _jit_new_node_pwf(jit_state_t *_jit, jit_code_t code, | |
1277 | jit_pointer_t u, jit_word_t v, jit_float32_t w) | |
1278 | { | |
1279 | jit_node_t *node = new_node(code); | |
1280 | assert(!_jitc->realize); | |
1281 | node->u.p = u; | |
1282 | node->v.w = v; | |
1283 | node->w.f = w; | |
1284 | return (link_node(node)); | |
1285 | } | |
1286 | ||
1287 | jit_node_t * | |
1288 | _jit_new_node_pwd(jit_state_t *_jit, jit_code_t code, | |
1289 | jit_pointer_t u, jit_word_t v, jit_float64_t w) | |
1290 | { | |
1291 | jit_node_t *node = new_node(code); | |
1292 | assert(!_jitc->realize); | |
1293 | node->u.p = u; | |
1294 | node->v.w = v; | |
1295 | node->w.d = w; | |
1296 | return (link_node(node)); | |
1297 | } | |
1298 | ||
1299 | jit_node_t * | |
1300 | _jit_label(jit_state_t *_jit) | |
1301 | { | |
1302 | jit_node_t *node; | |
1303 | ||
1304 | if (!(node = _jitc->tail) || node->code != jit_code_label) { | |
1305 | node = jit_forward(); | |
1306 | jit_link(node); | |
1307 | } | |
1308 | ||
1309 | return (node); | |
1310 | } | |
1311 | ||
1312 | jit_node_t * | |
1313 | _jit_forward(jit_state_t *_jit) | |
1314 | { | |
1315 | return (jit_new_node_no_link(jit_code_label)); | |
1316 | } | |
1317 | ||
1318 | jit_node_t * | |
1319 | _jit_indirect(jit_state_t *_jit) | |
1320 | { | |
1321 | jit_node_t *node; | |
1322 | ||
1323 | node = jit_label(); | |
1324 | node->flag |= jit_flag_use; | |
1325 | ||
1326 | return (node); | |
1327 | } | |
1328 | ||
1329 | void | |
1330 | _jit_link(jit_state_t *_jit, jit_node_t *node) | |
1331 | { | |
1332 | jit_block_t *block; | |
1333 | ||
1334 | assert((node->code == jit_code_label || | |
1335 | node->code == jit_code_prolog || | |
1336 | node->code == jit_code_epilog) && !node->next); | |
1337 | jit_link_node(node); | |
1338 | if (_jitc->blocks.offset >= _jitc->blocks.length) { | |
1339 | jit_word_t length; | |
1340 | ||
1341 | length = _jitc->blocks.length + 16; | |
1342 | jit_realloc((jit_pointer_t *)&_jitc->blocks.ptr, | |
1343 | _jitc->blocks.length * sizeof(jit_block_t), | |
1344 | length * sizeof(jit_block_t)); | |
1345 | _jitc->blocks.length = length; | |
1346 | } | |
1347 | block = _jitc->blocks.ptr + _jitc->blocks.offset; | |
1348 | block->label = node; | |
1349 | node->v.w = _jitc->blocks.offset; | |
1350 | jit_regset_new(&block->reglive); | |
1351 | jit_regset_new(&block->regmask); | |
1352 | ++_jitc->blocks.offset; | |
1353 | } | |
1354 | ||
1355 | jit_bool_t | |
1356 | _jit_forward_p(jit_state_t *_jit, jit_node_t *node) | |
1357 | { | |
1358 | return (node->code == jit_code_label && !node->next && node != _jitc->tail); | |
1359 | } | |
1360 | ||
1361 | jit_bool_t | |
1362 | _jit_indirect_p(jit_state_t *_jit, jit_node_t *node) | |
1363 | { | |
1364 | return (node->code == jit_code_label && !!(node->flag & jit_flag_use)); | |
1365 | } | |
1366 | ||
1367 | jit_bool_t | |
1368 | _jit_target_p(jit_state_t *_jit, jit_node_t *node) | |
1369 | { | |
1370 | return (node->code == jit_code_label && !!node->link); | |
1371 | } | |
1372 | ||
1373 | void | |
1374 | _jit_prepare(jit_state_t *_jit) | |
1375 | { | |
1376 | assert(_jitc->function != NULL); | |
1377 | _jitc->function->call.call = jit_call_default; | |
1378 | _jitc->function->call.argi = | |
1379 | _jitc->function->call.argf = | |
1380 | _jitc->function->call.size = 0; | |
1381 | _jitc->prepare = jit_new_node(jit_code_prepare); | |
1382 | } | |
1383 | ||
1384 | void | |
1385 | _jit_patch(jit_state_t* _jit, jit_node_t *instr) | |
1386 | { | |
1387 | jit_node_t *label; | |
1388 | ||
1389 | if (!(label = _jitc->tail) || label->code != jit_code_label) | |
1390 | label = jit_label(); | |
1391 | jit_patch_at(instr, label); | |
1392 | } | |
1393 | ||
1394 | jit_int32_t | |
1395 | _jit_classify(jit_state_t *_jit, jit_code_t code) | |
1396 | { | |
1397 | jit_int32_t mask; | |
1398 | ||
1399 | switch (code) { | |
1400 | case jit_code_data: case jit_code_save: case jit_code_load: | |
1401 | case jit_code_name: case jit_code_label: case jit_code_note: | |
1402 | case jit_code_prolog: case jit_code_ellipsis: case jit_code_va_push: | |
1403 | case jit_code_epilog: case jit_code_ret: case jit_code_prepare: | |
1404 | mask = 0; | |
1405 | break; | |
1406 | case jit_code_live: case jit_code_va_end: | |
79bfeef6 PC |
1407 | case jit_code_retr_c: case jit_code_retr_uc: |
1408 | case jit_code_retr_s: case jit_code_retr_us: | |
1409 | case jit_code_retr_i: case jit_code_retr_ui: | |
1410 | case jit_code_retr_l: | |
1411 | case jit_code_retr_f: case jit_code_retr_d: | |
1412 | case jit_code_pushargr_c: | |
1413 | case jit_code_pushargr_uc: | |
1414 | case jit_code_pushargr_s: | |
1415 | case jit_code_pushargr_us: | |
1416 | case jit_code_pushargr_i: | |
1417 | case jit_code_pushargr_ui: | |
1418 | case jit_code_pushargr_l: | |
1419 | case jit_code_pushargr_f: | |
4a71579b PC |
1420 | case jit_code_pushargr_d: |
1421 | case jit_code_finishr: /* synthesized will set jit_cc_a0_jmp */ | |
1422 | mask = jit_cc_a0_reg; | |
1423 | break; | |
79bfeef6 PC |
1424 | case jit_code_align: case jit_code_skip: |
1425 | case jit_code_reti_c: case jit_code_reti_uc: | |
1426 | case jit_code_reti_s: case jit_code_reti_us: | |
1427 | case jit_code_reti_i: case jit_code_reti_ui: | |
1428 | case jit_code_reti_l: | |
1429 | case jit_code_pushargi_c: | |
1430 | case jit_code_pushargi_uc: | |
1431 | case jit_code_pushargi_s: | |
1432 | case jit_code_pushargi_us: | |
1433 | case jit_code_pushargi_i: | |
1434 | case jit_code_pushargi_ui: | |
1435 | case jit_code_pushargi_l: | |
1436 | case jit_code_finishi: /* synthesized will set jit_cc_a0_jmp */ | |
4a71579b PC |
1437 | mask = jit_cc_a0_int; |
1438 | break; | |
1439 | case jit_code_reti_f: case jit_code_pushargi_f: | |
1440 | mask = jit_cc_a0_flt; | |
1441 | break; | |
1442 | case jit_code_reti_d: case jit_code_pushargi_d: | |
1443 | mask = jit_cc_a0_dbl; | |
1444 | break; | |
1445 | case jit_code_allocai: | |
1446 | mask = jit_cc_a0_int|jit_cc_a1_int; | |
1447 | break; | |
79bfeef6 PC |
1448 | case jit_code_arg_c: case jit_code_arg_s: |
1449 | case jit_code_arg_i: case jit_code_arg_l: | |
1450 | case jit_code_arg_f: case jit_code_arg_d: | |
4a71579b PC |
1451 | mask = jit_cc_a0_int|jit_cc_a0_arg; |
1452 | break; | |
1453 | case jit_code_calli: case jit_code_jmpi: | |
1454 | mask = jit_cc_a0_jmp; | |
1455 | break; | |
1456 | case jit_code_callr: case jit_code_jmpr: | |
1457 | mask = jit_cc_a0_reg|jit_cc_a0_jmp; | |
1458 | break; | |
1459 | case jit_code_retval_c: case jit_code_retval_uc: | |
1460 | case jit_code_retval_s: case jit_code_retval_us: | |
1461 | case jit_code_retval_i: case jit_code_retval_ui: | |
1462 | case jit_code_retval_l: | |
1463 | case jit_code_retval_f: case jit_code_retval_d: | |
1464 | case jit_code_va_start: | |
1465 | mask = jit_cc_a0_reg|jit_cc_a0_chg; | |
1466 | break; | |
1467 | case jit_code_getarg_c: case jit_code_getarg_uc: | |
1468 | case jit_code_getarg_s: case jit_code_getarg_us: | |
1469 | case jit_code_getarg_i: case jit_code_getarg_ui: | |
1470 | case jit_code_getarg_l: | |
1471 | case jit_code_getarg_f: case jit_code_getarg_d: | |
1472 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_arg; | |
1473 | break; | |
79bfeef6 PC |
1474 | case jit_code_putargr_c:case jit_code_putargr_uc: |
1475 | case jit_code_putargr_s:case jit_code_putargr_us: | |
1476 | case jit_code_putargr_i:case jit_code_putargr_ui: | |
1477 | case jit_code_putargr_l: | |
1478 | case jit_code_putargr_f:case jit_code_putargr_d: | |
4a71579b PC |
1479 | mask = jit_cc_a0_reg|jit_cc_a1_arg; |
1480 | break; | |
79bfeef6 PC |
1481 | case jit_code_putargi_c:case jit_code_putargi_uc: |
1482 | case jit_code_putargi_s:case jit_code_putargi_us: | |
1483 | case jit_code_putargi_i:case jit_code_putargi_ui: | |
1484 | case jit_code_putargi_l: | |
4a71579b PC |
1485 | mask = jit_cc_a0_int|jit_cc_a1_arg; |
1486 | break; | |
1487 | case jit_code_putargi_f: | |
1488 | mask = jit_cc_a0_flt|jit_cc_a1_arg; | |
1489 | break; | |
1490 | case jit_code_putargi_d: | |
1491 | mask = jit_cc_a0_dbl|jit_cc_a1_arg; | |
1492 | break; | |
1493 | case jit_code_movi: case jit_code_ldi_c: case jit_code_ldi_uc: | |
1494 | case jit_code_ldi_s: case jit_code_ldi_us: case jit_code_ldi_i: | |
1495 | case jit_code_ldi_ui: case jit_code_ldi_l: case jit_code_ldi_f: | |
1496 | case jit_code_ldi_d: | |
ba86ff93 PC |
1497 | case jit_code_negi: case jit_code_comi: |
1498 | case jit_code_exti_c: case jit_code_exti_uc: | |
1499 | case jit_code_exti_s: case jit_code_exti_us: | |
1500 | case jit_code_exti_i: case jit_code_exti_ui: | |
1501 | case jit_code_bswapi_us:case jit_code_bswapi_ui: | |
1502 | case jit_code_bswapi_ul: | |
1503 | case jit_code_htoni_us: case jit_code_htoni_ui: | |
1504 | case jit_code_htoni_ul: | |
1505 | case jit_code_cloi: case jit_code_clzi: | |
1506 | case jit_code_ctoi: case jit_code_ctzi: | |
1507 | case jit_code_rbiti: case jit_code_popcnti: | |
1508 | case jit_code_movi_w_f: case jit_code_movi_w_d: | |
4a71579b PC |
1509 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_int; |
1510 | break; | |
ba86ff93 PC |
1511 | case jit_code_unldi: case jit_code_unldi_u: case jit_code_unldi_x: |
1512 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_int|jit_cc_a2_int; | |
1513 | break; | |
1514 | case jit_code_movi_ww_d: | |
1515 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_int|jit_cc_a2_int; | |
1516 | break; | |
4a71579b | 1517 | case jit_code_movi_f: case jit_code_movi_f_w: |
ba86ff93 PC |
1518 | case jit_code_negi_f: case jit_code_absi_f: |
1519 | case jit_code_sqrti_f: | |
4a71579b PC |
1520 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_flt; |
1521 | break; | |
1522 | case jit_code_movi_d: case jit_code_movi_d_w: | |
ba86ff93 PC |
1523 | case jit_code_negi_d: case jit_code_absi_d: |
1524 | case jit_code_sqrti_d: | |
4a71579b PC |
1525 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_dbl; |
1526 | break; | |
1527 | case jit_code_movi_d_ww: | |
1528 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a1_chg| | |
1529 | jit_cc_a2_dbl; | |
1530 | break; | |
1531 | case jit_code_negr: case jit_code_comr: case jit_code_movr: | |
1532 | case jit_code_extr_c: case jit_code_extr_uc: case jit_code_extr_s: | |
1533 | case jit_code_extr_us: case jit_code_extr_i: case jit_code_extr_ui: | |
1534 | case jit_code_truncr_f_i: case jit_code_truncr_f_l: | |
1535 | case jit_code_truncr_d_i: case jit_code_truncr_d_l: | |
1536 | case jit_code_htonr_us: case jit_code_htonr_ui: case jit_code_htonr_ul: | |
40a44dcb | 1537 | case jit_code_bswapr_us: case jit_code_bswapr_ui: case jit_code_bswapr_ul: |
4a71579b PC |
1538 | case jit_code_ldr_c: case jit_code_ldr_uc: |
1539 | case jit_code_ldr_s: case jit_code_ldr_us: case jit_code_ldr_i: | |
1540 | case jit_code_ldr_ui: case jit_code_ldr_l: case jit_code_negr_f: | |
1541 | case jit_code_absr_f: case jit_code_sqrtr_f: case jit_code_movr_f: | |
1542 | case jit_code_extr_f: case jit_code_extr_d_f: case jit_code_ldr_f: | |
1543 | case jit_code_negr_d: case jit_code_absr_d: case jit_code_sqrtr_d: | |
1544 | case jit_code_movr_d: case jit_code_extr_d: case jit_code_extr_f_d: | |
ba86ff93 | 1545 | case jit_code_ldr_d: case jit_code_rbitr: case jit_code_popcntr: |
79bfeef6 PC |
1546 | case jit_code_clor: case jit_code_clzr: |
1547 | case jit_code_ctor: case jit_code_ctzr: | |
4a71579b PC |
1548 | case jit_code_movr_w_f: case jit_code_movr_f_w: |
1549 | case jit_code_movr_w_d: case jit_code_movr_d_w: | |
1550 | case jit_code_va_arg: case jit_code_va_arg_d: | |
1551 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg; | |
1552 | break; | |
1553 | case jit_code_movr_d_ww: | |
1554 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a1_chg| | |
1555 | jit_cc_a2_reg; | |
1556 | break; | |
1557 | case jit_code_addi: case jit_code_addxi: case jit_code_addci: | |
1558 | case jit_code_subi: case jit_code_subxi: case jit_code_subci: | |
1559 | case jit_code_rsbi: | |
ba86ff93 PC |
1560 | case jit_code_muli: case jit_code_hmuli: case jit_code_hmuli_u: |
1561 | case jit_code_divi: case jit_code_divi_u: | |
4a71579b PC |
1562 | case jit_code_remi: case jit_code_remi_u: case jit_code_andi: |
1563 | case jit_code_ori: case jit_code_xori: case jit_code_lshi: | |
ba86ff93 PC |
1564 | case jit_code_rshi: case jit_code_rshi_u: case jit_code_lroti: |
1565 | case jit_code_rroti: case jit_code_lti: | |
4a71579b PC |
1566 | case jit_code_lti_u: case jit_code_lei: case jit_code_lei_u: |
1567 | case jit_code_eqi: case jit_code_gei: case jit_code_gei_u: | |
1568 | case jit_code_gti: case jit_code_gti_u: case jit_code_nei: | |
1569 | case jit_code_ldxi_c: case jit_code_ldxi_uc: case jit_code_ldxi_s: | |
1570 | case jit_code_ldxi_us: case jit_code_ldxi_i: case jit_code_ldxi_ui: | |
1571 | case jit_code_ldxi_l: case jit_code_ldxi_f: case jit_code_ldxi_d: | |
ba86ff93 | 1572 | case jit_code_unldr: case jit_code_unldr_u: case jit_code_unldr_x: |
4a71579b PC |
1573 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_int; |
1574 | break; | |
1575 | case jit_code_qmuli: case jit_code_qmuli_u: | |
1576 | case jit_code_qdivi: case jit_code_qdivi_u: | |
ba86ff93 PC |
1577 | case jit_code_qlshi: case jit_code_qlshi_u: |
1578 | case jit_code_qrshi: case jit_code_qrshi_u: | |
4a71579b PC |
1579 | mask = jit_cc_a0_reg|jit_cc_a0_rlh|jit_cc_a0_chg| |
1580 | jit_cc_a1_reg|jit_cc_a2_int; | |
1581 | break; | |
1582 | case jit_code_addi_f: case jit_code_subi_f: case jit_code_rsbi_f: | |
1583 | case jit_code_muli_f: case jit_code_divi_f: case jit_code_lti_f: | |
1584 | case jit_code_lei_f: case jit_code_eqi_f: case jit_code_gei_f: | |
1585 | case jit_code_gti_f: case jit_code_nei_f: case jit_code_unlti_f: | |
1586 | case jit_code_unlei_f: case jit_code_uneqi_f: case jit_code_ungei_f: | |
1587 | case jit_code_ungti_f: case jit_code_ltgti_f: case jit_code_ordi_f: | |
1588 | case jit_code_unordi_f: | |
1589 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_flt; | |
1590 | break; | |
1591 | case jit_code_addi_d: case jit_code_subi_d: case jit_code_rsbi_d: | |
1592 | case jit_code_muli_d: case jit_code_divi_d: case jit_code_lti_d: | |
1593 | case jit_code_lei_d: case jit_code_eqi_d: case jit_code_gei_d: | |
1594 | case jit_code_gti_d: case jit_code_nei_d: case jit_code_unlti_d: | |
1595 | case jit_code_unlei_d: case jit_code_uneqi_d: case jit_code_ungei_d: | |
1596 | case jit_code_ungti_d: case jit_code_ltgti_d: case jit_code_ordi_d: | |
1597 | case jit_code_unordi_d: | |
1598 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_dbl; | |
1599 | break; | |
1600 | case jit_code_addr: case jit_code_addxr: case jit_code_addcr: | |
1601 | case jit_code_subr: case jit_code_subxr: case jit_code_subcr: | |
ba86ff93 PC |
1602 | case jit_code_mulr: case jit_code_hmulr: case jit_code_hmulr_u: |
1603 | case jit_code_divr: case jit_code_divr_u: | |
4a71579b PC |
1604 | case jit_code_remr: case jit_code_remr_u: case jit_code_andr: |
1605 | case jit_code_orr: case jit_code_xorr: case jit_code_lshr: | |
ba86ff93 PC |
1606 | case jit_code_rshr: case jit_code_rshr_u: case jit_code_lrotr: |
1607 | case jit_code_rrotr: case jit_code_ltr: | |
4a71579b PC |
1608 | case jit_code_ltr_u: case jit_code_ler: case jit_code_ler_u: |
1609 | case jit_code_eqr: case jit_code_ger: case jit_code_ger_u: | |
1610 | case jit_code_gtr: case jit_code_gtr_u: case jit_code_ner: | |
1611 | case jit_code_ldxr_c: case jit_code_ldxr_uc: case jit_code_ldxr_s: | |
1612 | case jit_code_ldxr_us: case jit_code_ldxr_i: case jit_code_ldxr_ui: | |
1613 | case jit_code_ldxr_l: case jit_code_addr_f: case jit_code_subr_f: | |
1614 | case jit_code_mulr_f: case jit_code_divr_f: case jit_code_ltr_f: | |
1615 | case jit_code_ler_f: case jit_code_eqr_f: case jit_code_ger_f: | |
1616 | case jit_code_gtr_f: case jit_code_ner_f: case jit_code_unltr_f: | |
1617 | case jit_code_unler_f: case jit_code_uneqr_f: case jit_code_unger_f: | |
1618 | case jit_code_ungtr_f: case jit_code_ltgtr_f: case jit_code_ordr_f: | |
1619 | case jit_code_unordr_f: case jit_code_ldxr_f: case jit_code_addr_d: | |
1620 | case jit_code_subr_d: case jit_code_mulr_d: case jit_code_divr_d: | |
1621 | case jit_code_ltr_d: case jit_code_ler_d: case jit_code_eqr_d: | |
1622 | case jit_code_ger_d: case jit_code_gtr_d: case jit_code_ner_d: | |
1623 | case jit_code_unltr_d: case jit_code_unler_d: case jit_code_uneqr_d: | |
1624 | case jit_code_unger_d: case jit_code_ungtr_d: case jit_code_ltgtr_d: | |
1625 | case jit_code_ordr_d: case jit_code_unordr_d: case jit_code_ldxr_d: | |
1626 | case jit_code_movr_ww_d: | |
1627 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_reg; | |
1628 | break; | |
1629 | case jit_code_qmulr: case jit_code_qmulr_u: | |
1630 | case jit_code_qdivr: case jit_code_qdivr_u: | |
ba86ff93 PC |
1631 | case jit_code_qlshr: case jit_code_qlshr_u: |
1632 | case jit_code_qrshr: case jit_code_qrshr_u: | |
4a71579b PC |
1633 | mask = jit_cc_a0_reg|jit_cc_a0_rlh|jit_cc_a0_chg| |
1634 | jit_cc_a1_reg|jit_cc_a2_reg; | |
1635 | break; | |
1636 | case jit_code_sti_c: case jit_code_sti_s: case jit_code_sti_i: | |
1637 | case jit_code_sti_l: case jit_code_sti_f: case jit_code_sti_d: | |
1638 | mask = jit_cc_a0_int|jit_cc_a1_reg; | |
1639 | break; | |
ba86ff93 PC |
1640 | case jit_code_unsti: case jit_code_unsti_x: |
1641 | mask = jit_cc_a0_int|jit_cc_a1_reg|jit_cc_a2_int; | |
1642 | break; | |
4a71579b PC |
1643 | case jit_code_blti: case jit_code_blti_u: case jit_code_blei: |
1644 | case jit_code_blei_u: case jit_code_beqi: case jit_code_bgei: | |
1645 | case jit_code_bgei_u: case jit_code_bgti: case jit_code_bgti_u: | |
1646 | case jit_code_bnei: case jit_code_bmsi: case jit_code_bmci: | |
1647 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a2_int; | |
1648 | break; | |
1649 | case jit_code_blti_f: case jit_code_blei_f: case jit_code_beqi_f: | |
1650 | case jit_code_bgei_f: case jit_code_bgti_f: case jit_code_bnei_f: | |
1651 | case jit_code_bunlti_f: case jit_code_bunlei_f: case jit_code_buneqi_f: | |
1652 | case jit_code_bungei_f: case jit_code_bungti_f: case jit_code_bltgti_f: | |
1653 | case jit_code_bordi_f: case jit_code_bunordi_f: | |
1654 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a2_flt; | |
1655 | break; | |
1656 | case jit_code_blti_d: case jit_code_blei_d: case jit_code_beqi_d: | |
1657 | case jit_code_bgei_d: case jit_code_bgti_d: case jit_code_bnei_d: | |
1658 | case jit_code_bunlti_d: case jit_code_bunlei_d: case jit_code_buneqi_d: | |
1659 | case jit_code_bungei_d: case jit_code_bungti_d: case jit_code_bltgti_d: | |
1660 | case jit_code_bordi_d: case jit_code_bunordi_d: | |
1661 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a2_dbl; | |
1662 | break; | |
1663 | case jit_code_allocar: /* synthesized instructions make it | |
1664 | * equivalent to jit_cc_a0_chg */ | |
1665 | case jit_code_str_c: case jit_code_str_s: case jit_code_str_i: | |
1666 | case jit_code_str_l: case jit_code_str_f: case jit_code_str_d: | |
1667 | mask = jit_cc_a0_reg|jit_cc_a1_reg; | |
1668 | break; | |
ba86ff93 PC |
1669 | case jit_code_unstr: case jit_code_unstr_x: |
1670 | mask = jit_cc_a0_reg|jit_cc_a1_reg|jit_cc_a2_int; | |
1671 | break; | |
4a71579b PC |
1672 | case jit_code_stxi_c: case jit_code_stxi_s: case jit_code_stxi_i: |
1673 | case jit_code_stxi_l: case jit_code_stxi_f: case jit_code_stxi_d: | |
1674 | mask = jit_cc_a0_int|jit_cc_a1_reg|jit_cc_a2_reg; | |
1675 | break; | |
1676 | case jit_code_bltr: case jit_code_bltr_u: case jit_code_bler: | |
1677 | case jit_code_bler_u: case jit_code_beqr: case jit_code_bger: | |
1678 | case jit_code_bger_u: case jit_code_bgtr: case jit_code_bgtr_u: | |
1679 | case jit_code_bner: case jit_code_bmsr: case jit_code_bmcr: | |
1680 | case jit_code_bltr_f: case jit_code_bler_f: case jit_code_beqr_f: | |
1681 | case jit_code_bger_f: case jit_code_bgtr_f: case jit_code_bner_f: | |
1682 | case jit_code_bunltr_f: case jit_code_bunler_f: case jit_code_buneqr_f: | |
1683 | case jit_code_bunger_f: case jit_code_bungtr_f: case jit_code_bltgtr_f: | |
1684 | case jit_code_bordr_f: case jit_code_bunordr_f:case jit_code_bltr_d: | |
1685 | case jit_code_bler_d: case jit_code_beqr_d: case jit_code_bger_d: | |
1686 | case jit_code_bgtr_d: case jit_code_bner_d: case jit_code_bunltr_d: | |
1687 | case jit_code_bunler_d: case jit_code_buneqr_d: case jit_code_bunger_d: | |
1688 | case jit_code_bungtr_d: case jit_code_bltgtr_d: case jit_code_bordr_d: | |
1689 | case jit_code_bunordr_d: | |
1690 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a2_reg; | |
1691 | break; | |
1692 | case jit_code_boaddi: case jit_code_boaddi_u: case jit_code_bxaddi: | |
1693 | case jit_code_bxaddi_u: case jit_code_bosubi: case jit_code_bosubi_u: | |
1694 | case jit_code_bxsubi: case jit_code_bxsubi_u: | |
1695 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a1_chg|jit_cc_a2_int; | |
1696 | break; | |
1697 | case jit_code_stxr_c: case jit_code_stxr_s: case jit_code_stxr_i: | |
1698 | case jit_code_stxr_l: case jit_code_stxr_f: case jit_code_stxr_d: | |
1699 | mask = jit_cc_a0_reg|jit_cc_a1_reg|jit_cc_a2_reg; | |
1700 | break; | |
1701 | case jit_code_boaddr: case jit_code_boaddr_u: case jit_code_bxaddr: | |
1702 | case jit_code_bxaddr_u: case jit_code_bosubr: case jit_code_bosubr_u: | |
1703 | case jit_code_bxsubr: case jit_code_bxsubr_u: | |
1704 | mask = jit_cc_a0_jmp|jit_cc_a1_reg|jit_cc_a1_chg|jit_cc_a2_reg; | |
1705 | break; | |
40a44dcb PC |
1706 | case jit_code_movnr: case jit_code_movzr: |
1707 | mask = jit_cc_a0_reg|jit_cc_a0_cnd|jit_cc_a1_reg|jit_cc_a2_reg; | |
1708 | break; | |
ba3814c1 PC |
1709 | case jit_code_casr: |
1710 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg| | |
1711 | jit_cc_a2_reg|jit_cc_a2_rlh; | |
1712 | break; | |
1713 | case jit_code_casi: | |
1714 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_int| | |
1715 | jit_cc_a2_reg|jit_cc_a2_rlh; | |
1716 | break; | |
ba86ff93 PC |
1717 | case jit_code_extr: case jit_code_extr_u: case jit_code_depr: |
1718 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_rlh; | |
1719 | break; | |
1720 | case jit_code_exti: case jit_code_exti_u: case jit_code_depi: | |
1721 | mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_int|jit_cc_a2_rlh; | |
1722 | break; | |
1723 | case jit_code_fmar_f: case jit_code_fmar_d: | |
1724 | case jit_code_fmsr_f: case jit_code_fmsr_d: | |
1725 | case jit_code_fnmar_f: case jit_code_fnmar_d: | |
1726 | case jit_code_fnmsr_f: case jit_code_fnmsr_d: | |
1727 | mask = jit_cc_a0_reg|jit_cc_a0_chg| | |
1728 | jit_cc_a1_reg|jit_cc_a1_rlh|jit_cc_a2_reg; | |
1729 | break; | |
1730 | case jit_code_fmai_f: case jit_code_fmsi_f: | |
1731 | case jit_code_fnmai_f: case jit_code_fnmsi_f: | |
1732 | mask = jit_cc_a0_reg|jit_cc_a0_chg| | |
1733 | jit_cc_a1_reg|jit_cc_a1_rlh|jit_cc_a2_flt; | |
1734 | break; | |
1735 | case jit_code_fmai_d: case jit_code_fmsi_d: | |
1736 | case jit_code_fnmai_d: case jit_code_fnmsi_d: | |
1737 | mask = jit_cc_a0_reg|jit_cc_a0_chg| | |
1738 | jit_cc_a1_reg|jit_cc_a1_rlh|jit_cc_a2_dbl; | |
1739 | break; | |
4a71579b PC |
1740 | default: |
1741 | abort(); | |
1742 | } | |
1743 | ||
1744 | return (mask); | |
1745 | } | |
1746 | ||
1747 | void | |
1748 | _jit_patch_abs(jit_state_t *_jit, jit_node_t *instr, jit_pointer_t address) | |
1749 | { | |
d481fb64 | 1750 | #ifndef NDEBUG |
4a71579b | 1751 | jit_int32_t mask; |
d481fb64 | 1752 | #endif |
4a71579b PC |
1753 | switch (instr->code) { |
1754 | case jit_code_movi: case jit_code_ldi_c: case jit_code_ldi_uc: | |
1755 | case jit_code_ldi_s: case jit_code_ldi_us: case jit_code_ldi_i: | |
1756 | case jit_code_ldi_ui: case jit_code_ldi_l: case jit_code_ldi_f: | |
1757 | case jit_code_ldi_d: | |
1758 | instr->v.p = address; | |
1759 | break; | |
1760 | case jit_code_sti_c: case jit_code_sti_s: case jit_code_sti_i: | |
1761 | case jit_code_sti_l: case jit_code_sti_f: case jit_code_sti_d: | |
1762 | instr->u.p = address; | |
1763 | break; | |
1764 | default: | |
d481fb64 | 1765 | #ifndef NDEBUG |
4a71579b | 1766 | mask = jit_classify(instr->code); |
d481fb64 | 1767 | #endif |
4a71579b PC |
1768 | assert((mask & (jit_cc_a0_reg|jit_cc_a0_jmp)) == jit_cc_a0_jmp); |
1769 | instr->u.p = address; | |
1770 | } | |
1771 | } | |
1772 | ||
1773 | void | |
1774 | _jit_patch_at(jit_state_t *_jit, jit_node_t *instr, jit_node_t *label) | |
1775 | { | |
d481fb64 | 1776 | #ifndef NDEBUG |
4a71579b | 1777 | jit_int32_t mask; |
d481fb64 | 1778 | #endif |
4a71579b PC |
1779 | assert(!(instr->flag & jit_flag_node)); |
1780 | instr->flag |= jit_flag_node; | |
1781 | switch (instr->code) { | |
1782 | case jit_code_movi: | |
1783 | assert(label->code == jit_code_label || | |
1784 | label->code == jit_code_data); | |
1785 | instr->v.n = label; | |
1786 | if (label->code == jit_code_data) | |
1787 | instr->flag |= jit_flag_data; | |
1788 | break; | |
1789 | case jit_code_jmpi: | |
1790 | assert(label->code == jit_code_label || | |
1791 | label->code == jit_code_epilog); | |
1792 | instr->u.n = label; | |
1793 | break; | |
1794 | default: | |
d481fb64 | 1795 | #ifndef NDEBUG |
4a71579b | 1796 | mask = jit_classify(instr->code); |
d481fb64 | 1797 | #endif |
4a71579b PC |
1798 | assert((mask & (jit_cc_a0_reg|jit_cc_a0_jmp)) == jit_cc_a0_jmp); |
1799 | assert(label->code == jit_code_label); | |
1800 | instr->u.n = label; | |
1801 | break; | |
1802 | } | |
1803 | /* link field is used as list of nodes associated with a given label */ | |
1804 | instr->link = label->link; | |
1805 | label->link = instr; | |
1806 | } | |
1807 | ||
c0c16242 PC |
1808 | static void |
1809 | _do_setup(jit_state_t *_jit) | |
4a71579b | 1810 | { |
4a71579b PC |
1811 | jit_block_t *block; |
1812 | jit_word_t offset; | |
1813 | ||
4a71579b PC |
1814 | /* create initial mapping of live register values |
1815 | * at the start of a basic block */ | |
1816 | for (offset = 0; offset < _jitc->blocks.offset; offset++) { | |
1817 | block = _jitc->blocks.ptr + offset; | |
79bfeef6 | 1818 | if (!block->label) |
4a71579b | 1819 | continue; |
79bfeef6 PC |
1820 | if (block->label->code == jit_code_epilog) { |
1821 | jit_regset_setbit(&block->reglive, JIT_RET); | |
1822 | jit_regset_setbit(&block->reglive, JIT_FRET); | |
1823 | jit_regset_com(&block->regmask, &block->reglive); | |
1824 | continue; | |
1825 | } | |
c0c16242 | 1826 | jit_setup(block); |
4a71579b | 1827 | } |
c0c16242 | 1828 | } |
4a71579b | 1829 | |
c0c16242 PC |
1830 | static jit_bool_t |
1831 | _block_update_set(jit_state_t *_jit, | |
1832 | jit_block_t *block, jit_block_t *target) | |
1833 | { | |
1834 | jit_regset_t regmask; | |
1835 | ||
1836 | jit_regset_ior(®mask, &block->reglive, &target->reglive); | |
1837 | jit_regset_and(®mask, ®mask, &block->regmask); | |
1838 | if (jit_regset_set_p(®mask)) { | |
1839 | jit_regset_ior(&block->reglive, &block->reglive, ®mask); | |
1840 | jit_regset_and(®mask, &block->reglive, &block->regmask); | |
1841 | jit_regset_com(®mask, ®mask); | |
1842 | jit_regset_and(&block->regmask, &block->regmask, ®mask); | |
1843 | block->again = 1; | |
1844 | return (1); | |
1845 | } | |
1846 | return (0); | |
1847 | } | |
4a71579b | 1848 | |
24d91c0d PC |
1849 | static void |
1850 | _propagate_backward(jit_state_t *_jit, jit_block_t *block) | |
1851 | { | |
1852 | jit_block_t *prev; | |
1853 | jit_word_t offset; | |
1854 | ||
1855 | for (offset = block->label->v.w - 1; | |
1856 | offset >= 0; --offset) { | |
1857 | prev = _jitc->blocks.ptr + offset; | |
1858 | if (!block_update_set(prev, block) || | |
1859 | !(prev->label->flag & jit_flag_head)) | |
1860 | break; | |
1861 | } | |
1862 | } | |
1863 | ||
c0c16242 PC |
1864 | static jit_bool_t |
1865 | _check_block_again(jit_state_t *_jit) | |
1866 | { | |
1867 | jit_int32_t todo; | |
1868 | jit_word_t offset; | |
1869 | jit_node_t *node, *label; | |
1870 | jit_block_t *block, *target; | |
4a71579b | 1871 | |
c0c16242 PC |
1872 | todo = 0; |
1873 | for (offset = 0; offset < _jitc->blocks.offset; offset++) { | |
1874 | block = _jitc->blocks.ptr + offset; | |
1875 | if (block->again) { | |
ba3814c1 | 1876 | todo = 1; |
c0c16242 | 1877 | break; |
ba3814c1 | 1878 | } |
c0c16242 PC |
1879 | } |
1880 | /* If no block changed state */ | |
1881 | if (!todo) | |
1882 | return (0); | |
ba3814c1 | 1883 | |
c0c16242 PC |
1884 | do { |
1885 | todo = 0; | |
1886 | block = NULL; | |
ba3814c1 | 1887 | for (node = _jitc->head; node; node = node->next) { |
c0c16242 PC |
1888 | /* Special jumps that match jit_cc_a0_jmp */ |
1889 | if (node->code == jit_code_calli || node->code == jit_code_callr) | |
1890 | continue; | |
1891 | ||
1892 | /* Remember current label */ | |
1893 | if (node->code == jit_code_label || | |
1894 | node->code == jit_code_prolog || | |
1895 | node->code == jit_code_epilog) { | |
1896 | ||
1897 | /* If previous block does not pass through */ | |
1898 | if (!(node->flag & jit_flag_head)) | |
1899 | block = NULL; | |
1900 | ||
1901 | target = _jitc->blocks.ptr + node->v.w; | |
24d91c0d PC |
1902 | if (block && target->again && block_update_set(block, target)) { |
1903 | propagate_backward(block); | |
ba3814c1 | 1904 | todo = 1; |
24d91c0d | 1905 | } |
c0c16242 | 1906 | block = target; |
ba3814c1 | 1907 | } |
c0c16242 PC |
1908 | /* If not the first jmpi */ |
1909 | else if (block) { | |
1910 | /* If a jump to dynamic address or if a jump to raw address */ | |
1911 | if (!(jit_classify(node->code) & jit_cc_a0_jmp) || | |
1912 | !(node->flag & jit_flag_node)) | |
1913 | continue; | |
1914 | label = node->u.n; | |
1915 | /* Mark predecessor needs updating due to target change */ | |
1916 | target = _jitc->blocks.ptr + label->v.w; | |
24d91c0d PC |
1917 | if (target->again && block_update_set(block, target)) { |
1918 | propagate_backward(block); | |
ba3814c1 | 1919 | todo = 1; |
24d91c0d | 1920 | } |
ba3814c1 PC |
1921 | } |
1922 | } | |
c0c16242 PC |
1923 | } |
1924 | while (todo); | |
1925 | ||
79bfeef6 | 1926 | return (todo); |
c0c16242 | 1927 | } |
ba3814c1 | 1928 | |
c0c16242 PC |
1929 | static void |
1930 | _do_follow(jit_state_t *_jit, jit_bool_t always) | |
1931 | { | |
1932 | jit_block_t *block; | |
1933 | jit_word_t offset; | |
1934 | ||
1935 | /* set live state of registers not referenced in a block, but | |
1936 | * referenced in a jump target or normal flow */ | |
1937 | for (offset = 0; offset < _jitc->blocks.offset; offset++) { | |
1938 | block = _jitc->blocks.ptr + offset; | |
1939 | if (!block->label || block->label->code == jit_code_epilog) | |
1940 | continue; | |
1941 | if (always || block->again) { | |
1942 | block->again = 0; | |
1943 | jit_follow(block); | |
1944 | } | |
1945 | } | |
1946 | } | |
1947 | ||
1948 | void | |
1949 | _jit_optimize(jit_state_t *_jit) | |
1950 | { | |
1951 | jit_bool_t jump; | |
1952 | jit_bool_t todo; | |
1953 | jit_int32_t mask; | |
1954 | jit_node_t *node; | |
1955 | jit_block_t *block; | |
1956 | jit_word_t offset; | |
79bfeef6 | 1957 | jit_regset_t regmask; |
c0c16242 PC |
1958 | |
1959 | todo = 0; | |
1960 | _jitc->function = NULL; | |
1961 | ||
1962 | thread_jumps(); | |
1963 | sequential_labels(); | |
1964 | split_branches(); | |
1965 | do_setup(); | |
1966 | do_follow(1); | |
1967 | ||
79bfeef6 PC |
1968 | jit_regset_set_ui(®mask, 0); |
1969 | for (offset = 0; offset < _jitc->reglen; offset++) { | |
1970 | if ((jit_class(_rvs[offset].spec) & (jit_class_gpr|jit_class_fpr)) && | |
1971 | (jit_class(_rvs[offset].spec) & jit_class_sav) == jit_class_sav) | |
1972 | jit_regset_setbit(®mask, offset); | |
1973 | } | |
1974 | ||
1975 | /* Figure out labels that are only reached with a jump */ | |
c0c16242 PC |
1976 | jump = 1; |
1977 | for (node = _jitc->head; node; node = node->next) { | |
1978 | switch (node->code) { | |
1979 | case jit_code_label: | |
79bfeef6 | 1980 | if (!jump) { |
c0c16242 | 1981 | node->flag |= jit_flag_head; |
79bfeef6 PC |
1982 | if (!node->link) { |
1983 | /* Block is dead code or only reachable with an | |
1984 | * indirect jumps. In such condition, must assume | |
1985 | * all callee save registers are live. */ | |
1986 | block = _jitc->blocks.ptr + node->v.w; | |
1987 | jit_regset_ior(&block->reglive, | |
1988 | &block->reglive, ®mask); | |
1989 | /* Cleanup regmask */ | |
1990 | block_update_set(block, block); | |
1991 | } | |
1992 | } | |
c0c16242 PC |
1993 | break; |
1994 | case jit_code_jmpi: case jit_code_jmpr: | |
1995 | case jit_code_epilog: | |
1996 | jump = 1; | |
1997 | break; | |
1998 | case jit_code_data: case jit_code_note: | |
1999 | break; | |
2000 | default: | |
2001 | jump = 0; | |
2002 | break; | |
2003 | } | |
2004 | } | |
2005 | ||
ba86ff93 PC |
2006 | patch_registers(); |
2007 | if (simplify()) | |
2008 | todo = 1; | |
2009 | ||
c0c16242 PC |
2010 | for (node = _jitc->head; node; node = node->next) { |
2011 | mask = jit_classify(node->code); | |
2012 | if (mask & jit_cc_a0_reg) | |
2013 | node->u.w &= ~jit_regno_patch; | |
2014 | if (mask & jit_cc_a1_reg) | |
2015 | node->v.w &= ~jit_regno_patch; | |
2016 | if (mask & jit_cc_a2_reg) | |
2017 | node->w.w &= ~jit_regno_patch; | |
2018 | if (node->code == jit_code_beqi) { | |
2019 | if (redundant_store(node, 1)) { | |
2020 | block = _jitc->blocks.ptr + ((jit_node_t *)node->u.n)->v.w; | |
2021 | block->again = 1; | |
2022 | todo = 1; | |
2023 | } | |
2024 | } | |
2025 | else if (node->code == jit_code_bnei) { | |
2026 | if (redundant_store(node, 0)) { | |
2027 | block = _jitc->blocks.ptr + ((jit_node_t *)node->u.n)->v.w; | |
2028 | block->again = 1; | |
2029 | todo = 1; | |
2030 | } | |
4a71579b PC |
2031 | } |
2032 | } | |
2033 | ||
c0c16242 PC |
2034 | if (!todo) |
2035 | todo = check_block_again(); | |
2036 | ||
2037 | /* If instructions were removed or first pass did modify the entry | |
2038 | * state of any block */ | |
2039 | if (todo) { | |
2040 | do_setup(); | |
2041 | todo = 0; | |
2042 | do { | |
2043 | do_follow(0); | |
2044 | /* If any block again has the entry state modified. */ | |
2045 | todo = check_block_again(); | |
2046 | } while (todo); | |
2047 | } | |
2048 | ||
4a71579b PC |
2049 | for (node = _jitc->head; node; node = node->next) { |
2050 | mask = jit_classify(node->code); | |
2051 | if (mask & jit_cc_a0_reg) | |
2052 | node->u.w &= ~jit_regno_patch; | |
2053 | if (mask & jit_cc_a1_reg) | |
2054 | node->v.w &= ~jit_regno_patch; | |
2055 | if (mask & jit_cc_a2_reg) | |
2056 | node->w.w &= ~jit_regno_patch; | |
ba3814c1 PC |
2057 | if (node->code == jit_code_prolog) |
2058 | _jitc->function = _jitc->functions.ptr + node->w.w; | |
2059 | else if(node->code == jit_code_epilog) | |
2060 | _jitc->function = NULL; | |
2061 | else { | |
4a71579b | 2062 | #if JIT_HASH_CONSTS |
ba3814c1 PC |
2063 | if (mask & jit_cc_a0_flt) { |
2064 | node->u.p = jit_data(&node->u.f, sizeof(jit_float32_t), 4); | |
2065 | node->flag |= jit_flag_node | jit_flag_data; | |
2066 | } | |
2067 | else if (mask & jit_cc_a0_dbl) { | |
2068 | node->u.p = jit_data(&node->u.d, sizeof(jit_float64_t), 8); | |
2069 | node->flag |= jit_flag_node | jit_flag_data; | |
2070 | } | |
2071 | else if (mask & jit_cc_a1_flt) { | |
2072 | node->v.p = jit_data(&node->v.f, sizeof(jit_float32_t), 4); | |
2073 | node->flag |= jit_flag_node | jit_flag_data; | |
2074 | } | |
2075 | else if (mask & jit_cc_a1_dbl) { | |
2076 | node->v.p = jit_data(&node->v.d, sizeof(jit_float64_t), 8); | |
2077 | node->flag |= jit_flag_node | jit_flag_data; | |
2078 | } | |
2079 | else if (mask & jit_cc_a2_flt) { | |
2080 | node->w.p = jit_data(&node->w.f, sizeof(jit_float32_t), 4); | |
2081 | node->flag |= jit_flag_node | jit_flag_data; | |
2082 | } | |
2083 | else if (mask & jit_cc_a2_dbl) { | |
2084 | node->w.p = jit_data(&node->w.d, sizeof(jit_float64_t), 8); | |
2085 | node->flag |= jit_flag_node | jit_flag_data; | |
2086 | } | |
4a71579b | 2087 | #endif |
ba3814c1 PC |
2088 | if (_jitc->function) { |
2089 | if ((mask & (jit_cc_a0_reg|jit_cc_a0_chg)) == | |
2090 | (jit_cc_a0_reg|jit_cc_a0_chg)) { | |
2091 | if (mask & jit_cc_a0_rlh) { | |
2092 | jit_regset_setbit(&_jitc->function->regset, | |
2093 | jit_regno(node->u.q.l)); | |
4a71579b | 2094 | jit_regset_setbit(&_jitc->function->regset, |
ba3814c1 PC |
2095 | jit_regno(node->u.q.h)); |
2096 | } | |
2097 | else | |
4a71579b | 2098 | jit_regset_setbit(&_jitc->function->regset, |
ba3814c1 | 2099 | jit_regno(node->u.w)); |
4a71579b | 2100 | } |
ba3814c1 PC |
2101 | if ((mask & (jit_cc_a1_reg|jit_cc_a1_chg)) == |
2102 | (jit_cc_a1_reg|jit_cc_a1_chg)) | |
2103 | jit_regset_setbit(&_jitc->function->regset, | |
2104 | jit_regno(node->v.w)); | |
2105 | if ((mask & (jit_cc_a2_reg|jit_cc_a2_chg)) == | |
2106 | (jit_cc_a2_reg|jit_cc_a2_chg)) | |
2107 | jit_regset_setbit(&_jitc->function->regset, | |
2108 | jit_regno(node->w.w)); | |
2109 | } | |
4a71579b PC |
2110 | } |
2111 | } | |
2112 | } | |
2113 | ||
2114 | void | |
2115 | _jit_reglive(jit_state_t *_jit, jit_node_t *node) | |
2116 | { | |
2117 | jit_int32_t spec; | |
2118 | jit_int32_t value; | |
2119 | jit_block_t *block; | |
2120 | ||
2121 | switch (node->code) { | |
2122 | case jit_code_label: case jit_code_prolog: case jit_code_epilog: | |
2123 | block = _jitc->blocks.ptr + node->v.w; | |
2124 | jit_regset_set(&_jitc->reglive, &block->reglive); | |
79bfeef6 PC |
2125 | jit_regset_set_ui(&_jitc->explive, 0); |
2126 | break; | |
2127 | case jit_code_live: | |
2128 | jit_regset_setbit(&_jitc->explive, node->u.w); | |
4a71579b PC |
2129 | break; |
2130 | case jit_code_callr: | |
2131 | value = jit_regno(node->u.w); | |
2132 | if (!(node->u.w & jit_regno_patch)) { | |
2133 | jit_regset_setbit(&_jitc->reglive, value); | |
2134 | } | |
2135 | case jit_code_calli: | |
2136 | for (value = 0; value < _jitc->reglen; value++) { | |
2137 | spec = jit_class(_rvs[value].spec); | |
2138 | if ((spec & jit_class_arg) && jit_regarg_p(node, value)) | |
2139 | jit_regset_setbit(&_jitc->reglive, value); | |
2140 | else if (!(spec & jit_class_sav)) | |
2141 | jit_regset_clrbit(&_jitc->reglive, value); | |
2142 | } | |
2143 | break; | |
2144 | default: | |
2145 | value = jit_classify(node->code); | |
2146 | if (value & jit_cc_a0_reg) { | |
2147 | if (value & jit_cc_a0_rlh) { | |
2148 | if (!(node->u.q.l & jit_regno_patch)) { | |
2149 | if (value & jit_cc_a0_chg) { | |
2150 | jit_regset_clrbit(&_jitc->reglive, node->u.q.l); | |
2151 | jit_regset_setbit(&_jitc->regmask, node->u.q.l); | |
2152 | } | |
2153 | else | |
2154 | jit_regset_setbit(&_jitc->reglive, node->u.q.l); | |
2155 | } | |
2156 | if (!(node->u.q.h & jit_regno_patch)) { | |
2157 | if (value & jit_cc_a0_chg) { | |
2158 | jit_regset_clrbit(&_jitc->reglive, node->u.q.h); | |
2159 | jit_regset_setbit(&_jitc->regmask, node->u.q.h); | |
2160 | } | |
2161 | else | |
2162 | jit_regset_setbit(&_jitc->reglive, node->u.q.h); | |
2163 | } | |
2164 | } | |
2165 | else { | |
2166 | if (!(node->u.w & jit_regno_patch)) { | |
2167 | if (value & jit_cc_a0_chg) { | |
2168 | jit_regset_clrbit(&_jitc->reglive, node->u.w); | |
2169 | jit_regset_setbit(&_jitc->regmask, node->u.w); | |
2170 | } | |
2171 | else | |
2172 | jit_regset_setbit(&_jitc->reglive, node->u.w); | |
2173 | } | |
2174 | } | |
2175 | } | |
ba86ff93 PC |
2176 | if (value & jit_cc_a1_reg) { |
2177 | if (value & jit_cc_a1_rlh) { | |
2178 | /* Assume registers are not changed */ | |
2179 | if (!(node->v.q.l & jit_regno_patch)) | |
2180 | jit_regset_setbit(&_jitc->reglive, node->v.q.l); | |
2181 | if (!(node->v.q.h & jit_regno_patch)) | |
2182 | jit_regset_setbit(&_jitc->reglive, node->v.q.h); | |
2183 | } | |
2184 | else if (!(node->v.w & jit_regno_patch)) { | |
2185 | if (value & jit_cc_a1_chg) { | |
2186 | jit_regset_clrbit(&_jitc->reglive, node->v.w); | |
2187 | jit_regset_setbit(&_jitc->regmask, node->v.w); | |
2188 | } | |
2189 | else | |
2190 | jit_regset_setbit(&_jitc->reglive, node->v.w); | |
4a71579b | 2191 | } |
4a71579b | 2192 | } |
ba3814c1 PC |
2193 | if (value & jit_cc_a2_reg) { |
2194 | if (value & jit_cc_a2_rlh) { | |
2195 | /* Assume registers are not changed */ | |
2196 | if (!(node->w.q.l & jit_regno_patch)) | |
2197 | jit_regset_setbit(&_jitc->reglive, node->w.q.l); | |
2198 | if (!(node->w.q.h & jit_regno_patch)) | |
2199 | jit_regset_setbit(&_jitc->reglive, node->w.q.h); | |
2200 | } | |
2201 | else { | |
2202 | if (!(node->w.w & jit_regno_patch)) { | |
2203 | if (value & jit_cc_a2_chg) { | |
2204 | jit_regset_clrbit(&_jitc->reglive, node->w.w); | |
2205 | jit_regset_setbit(&_jitc->regmask, node->w.w); | |
2206 | } | |
2207 | else | |
2208 | jit_regset_setbit(&_jitc->reglive, node->w.w); | |
2209 | } | |
4a71579b | 2210 | } |
4a71579b PC |
2211 | } |
2212 | if (jit_regset_set_p(&_jitc->regmask)) { | |
2213 | jit_update(node->next, &_jitc->reglive, &_jitc->regmask); | |
2214 | if (jit_regset_set_p(&_jitc->regmask)) { | |
2215 | /* any unresolved live state is considered as live */ | |
2216 | jit_regset_ior(&_jitc->reglive, | |
2217 | &_jitc->reglive, &_jitc->regmask); | |
2218 | jit_regset_set_ui(&_jitc->regmask, 0); | |
2219 | } | |
2220 | } | |
2221 | break; | |
2222 | } | |
2223 | } | |
2224 | ||
2225 | void | |
2226 | _jit_regarg_set(jit_state_t *_jit, jit_node_t *node, jit_int32_t value) | |
2227 | { | |
2228 | #if GET_JIT_SIZE | |
2229 | jit_size_prepare(); | |
2230 | #endif | |
2231 | if (value & jit_cc_a0_reg) { | |
2232 | if (value & jit_cc_a0_rlh) { | |
2233 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->u.q.l)); | |
2234 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->u.q.h)); | |
2235 | } | |
2236 | else | |
2237 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->u.w)); | |
2238 | } | |
ba86ff93 PC |
2239 | if (value & jit_cc_a1_reg) { |
2240 | if (value & jit_cc_a1_rlh) { | |
2241 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->v.q.l)); | |
2242 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->v.q.h)); | |
2243 | } | |
2244 | else | |
2245 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->v.w)); | |
2246 | } | |
ba3814c1 PC |
2247 | if (value & jit_cc_a2_reg) { |
2248 | if (value & jit_cc_a2_rlh) { | |
2249 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->w.q.l)); | |
2250 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->w.q.h)); | |
2251 | } | |
2252 | else | |
2253 | jit_regset_setbit(&_jitc->regarg, jit_regno(node->w.w)); | |
2254 | } | |
79bfeef6 PC |
2255 | /* Prevent incorrect detection of running out of registers |
2256 | * if will need to patch jump, and all registers have been | |
2257 | * used in the current block. */ | |
2258 | if (node->code == jit_code_jmpi && (node->flag & jit_flag_node)) { | |
2259 | jit_node_t *label = node->u.n; | |
2260 | jit_block_t *block = _jitc->blocks.ptr + label->v.w; | |
2261 | jit_regset_set(&_jitc->reglive, &block->reglive); | |
2262 | jit_regset_set(&_jitc->regmask, &block->regmask); | |
2263 | if (jit_regset_set_p(&_jitc->explive)) { | |
2264 | jit_regset_ior(&_jitc->reglive, &block->reglive, &_jitc->explive); | |
2265 | jit_regset_xor(&_jitc->regmask, &_jitc->regmask, &_jitc->explive); | |
2266 | } | |
2267 | } | |
4a71579b PC |
2268 | } |
2269 | ||
2270 | void | |
2271 | _jit_regarg_clr(jit_state_t *_jit, jit_node_t *node, jit_int32_t value) | |
2272 | { | |
2273 | #if GET_JIT_SIZE | |
2274 | jit_size_collect(node); | |
2275 | #endif | |
2276 | if (value & jit_cc_a0_reg) { | |
2277 | if (value & jit_cc_a0_rlh) { | |
2278 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->u.q.l)); | |
2279 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->u.q.h)); | |
2280 | } | |
2281 | else | |
2282 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->u.w)); | |
2283 | } | |
ba86ff93 PC |
2284 | if (value & jit_cc_a1_reg) { |
2285 | if (value & jit_cc_a1_rlh) { | |
2286 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->v.q.l)); | |
2287 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->v.q.h)); | |
2288 | } | |
2289 | else | |
2290 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->v.w)); | |
2291 | } | |
ba3814c1 PC |
2292 | if (value & jit_cc_a2_reg) { |
2293 | if (value & jit_cc_a2_rlh) { | |
2294 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->w.q.l)); | |
2295 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->w.q.h)); | |
2296 | } | |
2297 | else | |
2298 | jit_regset_clrbit(&_jitc->regarg, jit_regno(node->w.w)); | |
2299 | } | |
4a71579b PC |
2300 | } |
2301 | ||
2302 | void | |
2303 | _jit_realize(jit_state_t *_jit) | |
2304 | { | |
2305 | assert(!_jitc->realize); | |
2306 | if (_jitc->function) | |
2307 | jit_epilog(); | |
2308 | jit_optimize(); | |
2309 | _jitc->realize = 1; | |
2310 | ||
2311 | /* ensure it is aligned */ | |
2312 | _jitc->data.offset = (_jitc->data.offset + 7) & -8; | |
2313 | ||
2314 | #if GET_JIT_SIZE | |
2315 | /* Heuristic to guess code buffer size */ | |
2316 | _jitc->mult = 4; | |
2317 | _jit->code.length = _jitc->pool.length * 1024 * _jitc->mult; | |
2318 | #else | |
2319 | _jit->code.length = jit_get_size(); | |
2320 | #endif | |
2321 | } | |
2322 | ||
2323 | void | |
2324 | _jit_dataset(jit_state_t *_jit) | |
2325 | { | |
2326 | jit_uint8_t *ptr; | |
2327 | jit_node_t *node; | |
2328 | jit_word_t offset; | |
2329 | #if defined(__sgi) | |
2330 | int mmap_fd; | |
2331 | #endif | |
2332 | ||
2333 | assert(!_jitc->dataset); | |
40a44dcb PC |
2334 | #if !HAVE_MMAP |
2335 | assert(_jit->user_data); | |
2336 | #else | |
4a71579b PC |
2337 | if (!_jit->user_data) { |
2338 | ||
2339 | /* create read only data buffer */ | |
2340 | _jit->data.length = (_jitc->data.offset + | |
2341 | /* reserve space for annotations */ | |
2342 | _jitc->note.size + 4095) & -4096; | |
2343 | #if defined(__sgi) | |
2344 | mmap_fd = open("/dev/zero", O_RDWR); | |
2345 | #endif | |
2346 | _jit->data.ptr = mmap(NULL, _jit->data.length, | |
2347 | PROT_READ | PROT_WRITE, | |
2348 | MAP_PRIVATE | MAP_ANON, mmap_fd, 0); | |
2349 | assert(_jit->data.ptr != MAP_FAILED); | |
2350 | #if defined(__sgi) | |
2351 | close(mmap_fd); | |
2352 | #endif | |
2353 | } | |
40a44dcb | 2354 | #endif /* !HAVE_MMAP */ |
4a71579b PC |
2355 | |
2356 | if (!_jitc->no_data) | |
2357 | jit_memcpy(_jit->data.ptr, _jitc->data.ptr, _jitc->data.offset); | |
2358 | ||
2359 | if (_jitc->no_note) { | |
2360 | /* Space for one note is always allocated, so revert it here | |
2361 | * if after jit_new_state was called, it is also requested to | |
2362 | * not generate annotation information */ | |
2363 | _jit->note.length = 0; | |
2364 | _jitc->note.size = 0; | |
2365 | } | |
2366 | else { | |
2367 | _jitc->note.base = _jit->data.ptr; | |
2368 | if (!_jitc->no_data) | |
2369 | _jitc->note.base += _jitc->data.offset; | |
2370 | memset(_jitc->note.base, 0, _jitc->note.size); | |
2371 | } | |
2372 | ||
2373 | if (_jit->user_data) | |
2374 | /* Need the temporary hashed data until jit_emit is finished */ | |
2375 | ptr = _jitc->no_data ? _jitc->data.ptr : _jit->data.ptr; | |
2376 | else { | |
2377 | ptr = _jit->data.ptr; | |
2378 | /* Temporary hashed data no longer required */ | |
2379 | jit_free((jit_pointer_t *)&_jitc->data.ptr); | |
2380 | } | |
2381 | ||
2382 | for (offset = 0; offset < _jitc->data.size; offset++) { | |
2383 | for (node = _jitc->data.table[offset]; node; node = node->next) { | |
2384 | node->flag |= jit_flag_patch; | |
2385 | node->u.w = (jit_word_t)(ptr + node->u.w); | |
2386 | } | |
2387 | } | |
2388 | ||
2389 | _jitc->dataset = 1; | |
2390 | } | |
2391 | ||
2392 | jit_pointer_t | |
2393 | _jit_get_code(jit_state_t *_jit, jit_word_t *length) | |
2394 | { | |
2395 | assert(_jitc->realize); | |
2396 | if (length) { | |
2397 | if (_jitc->done) | |
2398 | /* If code already generated, return exact size of code */ | |
2399 | *length = _jit->pc.uc - _jit->code.ptr; | |
2400 | else | |
2401 | /* Else return current size of the code buffer */ | |
2402 | *length = _jit->code.length; | |
2403 | } | |
2404 | ||
2405 | return (_jit->code.ptr); | |
2406 | } | |
2407 | ||
2408 | void | |
2409 | _jit_set_code(jit_state_t *_jit, jit_pointer_t ptr, jit_word_t length) | |
2410 | { | |
2411 | assert(_jitc->realize); | |
2412 | _jit->code.ptr = ptr; | |
2413 | _jit->code.length = length; | |
2414 | _jit->user_code = 1; | |
2415 | } | |
2416 | ||
2417 | jit_pointer_t | |
2418 | _jit_get_data(jit_state_t *_jit, jit_word_t *data_size, jit_word_t *note_size) | |
2419 | { | |
2420 | assert(_jitc->realize); | |
2421 | if (data_size) | |
2422 | *data_size = _jitc->data.offset; | |
2423 | if (note_size) | |
2424 | *note_size = _jitc->note.size; | |
2425 | return (_jit->data.ptr); | |
2426 | } | |
2427 | ||
2428 | void | |
2429 | _jit_set_data(jit_state_t *_jit, jit_pointer_t ptr, | |
2430 | jit_word_t length, jit_word_t flags) | |
2431 | { | |
2432 | assert(_jitc->realize); | |
2433 | if (flags & JIT_DISABLE_DATA) | |
2434 | _jitc->no_data = 1; | |
2435 | else | |
2436 | assert(length >= _jitc->data.offset); | |
2437 | if (flags & JIT_DISABLE_NOTE) | |
2438 | _jitc->no_note = 1; | |
2439 | else { | |
2440 | if (flags & JIT_DISABLE_DATA) | |
2441 | assert(length >= _jitc->note.size); | |
2442 | else | |
2443 | assert(length >= _jitc->data.offset + _jitc->note.size); | |
2444 | } | |
2445 | _jit->data.ptr = ptr; | |
2446 | _jit->data.length = length; | |
2447 | _jit->user_data = 1; | |
2448 | } | |
2449 | ||
2450 | jit_pointer_t | |
2451 | _jit_emit(jit_state_t *_jit) | |
2452 | { | |
2453 | jit_pointer_t code; | |
2454 | jit_node_t *node; | |
2455 | size_t length; | |
d481fb64 | 2456 | #ifndef NDEBUG |
4a71579b | 2457 | int result; |
d481fb64 | 2458 | #endif |
4a71579b PC |
2459 | #if defined(__sgi) |
2460 | int mmap_fd; | |
2461 | #endif | |
c0c16242 | 2462 | int mmap_prot, mmap_flags; |
4a71579b PC |
2463 | |
2464 | if (!_jitc->realize) | |
2465 | jit_realize(); | |
2466 | ||
2467 | if (!_jitc->dataset) | |
2468 | jit_dataset(); | |
2469 | ||
2470 | _jitc->emit = 1; | |
2471 | ||
40a44dcb PC |
2472 | #if !HAVE_MMAP |
2473 | assert(_jit->user_code); | |
2474 | #else | |
4a71579b | 2475 | if (!_jit->user_code) { |
c0c16242 | 2476 | mmap_prot = PROT_READ | PROT_WRITE; |
79bfeef6 | 2477 | #if !(__OpenBSD__ || __APPLE__) |
c0c16242 PC |
2478 | mmap_prot |= PROT_EXEC; |
2479 | #endif | |
2480 | #if __NetBSD__ | |
2481 | mmap_prot = PROT_MPROTECT(mmap_prot); | |
2482 | mmap_flags = 0; | |
2483 | #else | |
2484 | mmap_flags = MAP_PRIVATE; | |
2485 | #endif | |
2486 | mmap_flags |= MAP_ANON; | |
4a71579b PC |
2487 | #if defined(__sgi) |
2488 | mmap_fd = open("/dev/zero", O_RDWR); | |
2489 | #endif | |
2490 | _jit->code.ptr = mmap(NULL, _jit->code.length, | |
c0c16242 | 2491 | mmap_prot, mmap_flags, mmap_fd, 0); |
4a71579b PC |
2492 | assert(_jit->code.ptr != MAP_FAILED); |
2493 | } | |
40a44dcb | 2494 | #endif /* !HAVE_MMAP */ |
4a71579b PC |
2495 | _jitc->code.end = _jit->code.ptr + _jit->code.length - |
2496 | jit_get_max_instr(); | |
2497 | _jit->pc.uc = _jit->code.ptr; | |
2498 | ||
2499 | for (;;) { | |
c0c16242 | 2500 | #if __NetBSD__ |
d481fb64 PC |
2501 | #ifndef NDEBUG |
2502 | result = | |
2503 | #endif | |
2504 | mprotect(_jit->code.ptr, _jit->code.length, PROT_READ | PROT_WRITE); | |
c0c16242 PC |
2505 | assert(result == 0); |
2506 | #endif | |
4a71579b PC |
2507 | if ((code = emit_code()) == NULL) { |
2508 | _jitc->patches.offset = 0; | |
2509 | for (node = _jitc->head; node; node = node->next) { | |
2510 | if (node->link && | |
2511 | (node->code == jit_code_label || | |
2512 | node->code == jit_code_epilog)) | |
2513 | node->flag &= ~jit_flag_patch; | |
2514 | } | |
40a44dcb PC |
2515 | #if !HAVE_MMAP |
2516 | assert(_jit->user_code); | |
2517 | #else | |
4a71579b PC |
2518 | if (_jit->user_code) |
2519 | goto fail; | |
2520 | #if GET_JIT_SIZE | |
2521 | ++_jitc->mult; | |
2522 | length = _jitc->pool.length * 1024 * _jitc->mult; | |
2523 | #else | |
2524 | /* Should only happen on very special cases */ | |
2525 | length = _jit->code.length + 4096; | |
2526 | #endif | |
2527 | ||
2528 | #if !HAVE_MREMAP | |
2529 | munmap(_jit->code.ptr, _jit->code.length); | |
2530 | #endif | |
2531 | ||
2532 | #if HAVE_MREMAP | |
2533 | # if __NetBSD__ | |
2534 | _jit->code.ptr = mremap(_jit->code.ptr, _jit->code.length, | |
2535 | _jit->code.ptr, length, 0); | |
2536 | # else | |
2537 | _jit->code.ptr = mremap(_jit->code.ptr, _jit->code.length, | |
2538 | length, MREMAP_MAYMOVE, NULL); | |
2539 | # endif | |
2540 | #else | |
2541 | _jit->code.ptr = mmap(NULL, length, | |
79bfeef6 | 2542 | mmap_prot, mmap_flags, mmap_fd, 0); |
4a71579b PC |
2543 | #endif |
2544 | ||
2545 | assert(_jit->code.ptr != MAP_FAILED); | |
2546 | _jit->code.length = length; | |
2547 | _jitc->code.end = _jit->code.ptr + _jit->code.length - | |
2548 | jit_get_max_instr(); | |
2549 | _jit->pc.uc = _jit->code.ptr; | |
40a44dcb | 2550 | #endif /* !HAVE_MMAP */ |
4a71579b PC |
2551 | } |
2552 | else | |
2553 | break; | |
2554 | } | |
2555 | ||
2556 | #if defined(__sgi) | |
2557 | if (!_jit->user_code) | |
2558 | close(mmap_fd); | |
2559 | #endif | |
2560 | ||
2561 | _jitc->done = 1; | |
2562 | if (!_jitc->no_note) | |
2563 | jit_annotate(); | |
2564 | ||
2565 | if (_jit->user_data) | |
2566 | jit_free((jit_pointer_t *)&_jitc->data.ptr); | |
40a44dcb | 2567 | #if HAVE_MMAP |
4a71579b | 2568 | else { |
d481fb64 PC |
2569 | # ifndef NDEBUG |
2570 | result = | |
2571 | # endif | |
2572 | mprotect(_jit->data.ptr, _jit->data.length, PROT_READ); | |
4a71579b PC |
2573 | assert(result == 0); |
2574 | } | |
2575 | if (!_jit->user_code) { | |
ba86ff93 | 2576 | _jit->code.protect = _jit->pc.uc - _jit->code.ptr; |
c0c16242 PC |
2577 | # if __riscv && __WORDSIZE == 64 |
2578 | /* FIXME should start adding consts at a page boundary */ | |
ba86ff93 | 2579 | _jit->code.protect -= _jitc->consts.hash.count * sizeof(jit_word_t); |
c0c16242 | 2580 | # endif |
d481fb64 PC |
2581 | # ifndef NDEBUG |
2582 | result = | |
2583 | mprotect(_jit->code.ptr, _jit->code.protect, PROT_READ | PROT_EXEC); | |
2584 | # endif | |
4a71579b PC |
2585 | assert(result == 0); |
2586 | } | |
40a44dcb | 2587 | #endif /* HAVE_MMAP */ |
4a71579b PC |
2588 | |
2589 | return (_jit->code.ptr); | |
2590 | fail: | |
2591 | return (NULL); | |
2592 | } | |
2593 | ||
79bfeef6 PC |
2594 | void |
2595 | _jit_protect(jit_state_t *_jit) | |
2596 | { | |
2597 | #if !HAVE_MMAP | |
2598 | assert (_jit->user_code); | |
2599 | #else | |
d481fb64 | 2600 | # ifndef NDEBUG |
79bfeef6 | 2601 | int result; |
d481fb64 | 2602 | # endif |
79bfeef6 | 2603 | if (_jit->user_code) return; |
d481fb64 PC |
2604 | # ifndef NDEBUG |
2605 | result = | |
2606 | #endif | |
2607 | mprotect (_jit->code.ptr, _jit->code.protect, PROT_READ | PROT_EXEC); | |
79bfeef6 PC |
2608 | assert (result == 0); |
2609 | #endif | |
2610 | } | |
2611 | ||
2612 | void | |
2613 | _jit_unprotect(jit_state_t *_jit) | |
2614 | { | |
2615 | #if !HAVE_MMAP | |
2616 | assert (_jit->user_code); | |
2617 | #else | |
d481fb64 | 2618 | # ifndef NDEBUG |
79bfeef6 | 2619 | int result; |
d481fb64 | 2620 | # endif |
79bfeef6 | 2621 | if (_jit->user_code) return; |
d481fb64 PC |
2622 | # ifndef NDEBUG |
2623 | result = | |
2624 | # endif | |
2625 | mprotect (_jit->code.ptr, _jit->code.protect, PROT_READ | PROT_WRITE); | |
79bfeef6 PC |
2626 | assert (result == 0); |
2627 | #endif | |
2628 | } | |
2629 | ||
4a71579b PC |
2630 | void |
2631 | _jit_frame(jit_state_t *_jit, jit_int32_t frame) | |
2632 | { | |
2633 | jit_trampoline(frame, 1); | |
2634 | } | |
2635 | ||
2636 | void | |
2637 | _jit_tramp(jit_state_t *_jit, jit_int32_t frame) | |
2638 | { | |
2639 | jit_trampoline(frame, 0); | |
2640 | } | |
2641 | ||
2642 | void | |
2643 | _jit_trampoline(jit_state_t *_jit, jit_int32_t frame, jit_bool_t prolog) | |
2644 | { | |
2645 | jit_int32_t regno; | |
2646 | ||
2647 | /* Must be called after prolog, actually, just to simplify | |
2648 | * tests and know there is a current function and that | |
2649 | * _jitc->function->self.aoff is at the before any alloca value */ | |
2650 | assert(_jitc->tail && _jitc->tail->code == jit_code_prolog); | |
2651 | ||
2652 | /* + 24 for 3 possible spilled temporaries (that could be a double) */ | |
2653 | frame += 24; | |
2654 | #if defined(__hppa__) | |
2655 | frame += _jitc->function->self.aoff; | |
2656 | #else | |
2657 | frame -= _jitc->function->self.aoff; | |
2658 | #endif | |
2659 | _jitc->function->frame = frame; | |
2660 | if (prolog) | |
2661 | _jitc->function->define_frame = 1; | |
2662 | else | |
2663 | _jitc->function->assume_frame = 1; | |
2664 | for (regno = 0; regno < _jitc->reglen; regno++) | |
2665 | if (jit_class(_rvs[regno].spec) & jit_class_sav) | |
2666 | jit_regset_setbit(&_jitc->function->regset, regno); | |
2667 | } | |
2668 | ||
2669 | /* Compute initial reglive and regmask set values of a basic block. | |
2670 | * reglive is the set of known live registers | |
2671 | * regmask is the set of registers not referenced in the block | |
2672 | * Registers in regmask might be live. | |
2673 | */ | |
2674 | static void | |
2675 | _jit_setup(jit_state_t *_jit, jit_block_t *block) | |
2676 | { | |
2677 | jit_node_t *node; | |
2678 | jit_bool_t live; | |
2679 | unsigned long value; | |
2680 | ||
2681 | jit_regset_set_mask(&block->regmask, _jitc->reglen); | |
2682 | for (value = 0; value < _jitc->reglen; ++value) | |
2683 | if (!(jit_class(_rvs[value].spec) & (jit_class_gpr|jit_class_fpr))) | |
2684 | jit_regset_clrbit(&block->regmask, value); | |
2685 | ||
2686 | for (node = block->label->next; node; node = node->next) { | |
2687 | switch (node->code) { | |
2688 | case jit_code_label: case jit_code_prolog: | |
2689 | case jit_code_epilog: | |
2690 | return; | |
2691 | default: | |
2692 | /* Check argument registers in reverse order to properly | |
2693 | * handle registers that are both, argument and result */ | |
2694 | value = jit_classify(node->code); | |
ba86ff93 | 2695 | if (value & jit_cc_a2_reg) { |
4a71579b | 2696 | live = !(value & jit_cc_a2_chg); |
ba86ff93 PC |
2697 | if (value & jit_cc_a2_rlh) { |
2698 | /* Assume will not modify a pair in second argument */ | |
2699 | assert(live); | |
2700 | if (!(node->w.q.l & jit_regno_patch) && | |
2701 | jit_regset_tstbit(&block->regmask, node->w.q.l)) { | |
2702 | jit_regset_clrbit(&block->regmask, node->w.q.l); | |
2703 | } | |
2704 | if (!(node->w.q.h & jit_regno_patch) && | |
2705 | jit_regset_tstbit(&block->regmask, node->w.q.h)) { | |
2706 | jit_regset_clrbit(&block->regmask, node->w.q.h); | |
2707 | } | |
2708 | } | |
2709 | else { | |
2710 | if (!(node->w.w & jit_regno_patch) && | |
2711 | jit_regset_tstbit(&block->regmask, node->w.w)) { | |
2712 | jit_regset_clrbit(&block->regmask, node->w.w); | |
2713 | if (live) | |
2714 | jit_regset_setbit(&block->reglive, node->w.w); | |
2715 | } | |
2716 | } | |
4a71579b | 2717 | } |
ba86ff93 | 2718 | if (value & jit_cc_a1_reg) { |
4a71579b | 2719 | live = !(value & jit_cc_a1_chg); |
ba86ff93 PC |
2720 | if (value & jit_cc_a1_rlh) { |
2721 | /* Assume will not modify a pair in second argument */ | |
2722 | assert(live); | |
2723 | if (!(node->v.q.l & jit_regno_patch) && | |
2724 | jit_regset_tstbit(&block->regmask, node->v.q.l)) { | |
2725 | jit_regset_clrbit(&block->regmask, node->v.q.l); | |
2726 | } | |
2727 | if (!(node->v.q.h & jit_regno_patch) && | |
2728 | jit_regset_tstbit(&block->regmask, node->v.q.h)) { | |
2729 | jit_regset_clrbit(&block->regmask, node->v.q.h); | |
2730 | } | |
2731 | } | |
2732 | else { | |
2733 | if (!(node->v.w & jit_regno_patch) && | |
2734 | jit_regset_tstbit(&block->regmask, node->v.w)) { | |
2735 | jit_regset_clrbit(&block->regmask, node->v.w); | |
2736 | if (live) | |
2737 | jit_regset_setbit(&block->reglive, node->v.w); | |
2738 | } | |
2739 | } | |
4a71579b PC |
2740 | } |
2741 | if (value & jit_cc_a0_reg) { | |
2742 | live = !(value & jit_cc_a0_chg); | |
2743 | if (value & jit_cc_a0_rlh) { | |
2744 | if (!(node->u.q.l & jit_regno_patch) && | |
2745 | jit_regset_tstbit(&block->regmask, node->u.q.l)) { | |
2746 | jit_regset_clrbit(&block->regmask, node->u.q.l); | |
2747 | if (live) | |
2748 | jit_regset_setbit(&block->reglive, node->u.q.l); | |
2749 | } | |
2750 | if (!(node->u.q.h & jit_regno_patch) && | |
2751 | jit_regset_tstbit(&block->regmask, node->u.q.h)) { | |
2752 | jit_regset_clrbit(&block->regmask, node->u.q.h); | |
2753 | if (live) | |
2754 | jit_regset_setbit(&block->reglive, node->u.q.h); | |
2755 | } | |
2756 | } | |
2757 | else { | |
2758 | if (!(node->u.w & jit_regno_patch) && | |
2759 | jit_regset_tstbit(&block->regmask, node->u.w)) { | |
2760 | jit_regset_clrbit(&block->regmask, node->u.w); | |
2761 | if (live) | |
2762 | jit_regset_setbit(&block->reglive, node->u.w); | |
2763 | } | |
2764 | } | |
2765 | } | |
2766 | break; | |
2767 | } | |
2768 | } | |
2769 | } | |
2770 | ||
2771 | /* Update regmask and reglive of blocks at entry point of branch targets | |
2772 | * or normal flow that have a live register not used in this block. | |
2773 | */ | |
2774 | static void | |
c0c16242 | 2775 | _jit_follow(jit_state_t *_jit, jit_block_t *block) |
4a71579b PC |
2776 | { |
2777 | jit_node_t *node; | |
2778 | jit_block_t *next; | |
2779 | jit_int32_t spec; | |
2780 | jit_int32_t regno; | |
2781 | unsigned long value; | |
2782 | jit_node_t *label; | |
2783 | jit_regset_t reglive; | |
2784 | jit_regset_t regmask; | |
2785 | jit_regset_t regtemp; | |
2786 | ||
2787 | jit_regset_set(®live, &block->reglive); | |
2788 | jit_regset_set(®mask, &block->regmask); | |
2789 | for (node = block->label->next; node; node = node->next) { | |
2790 | switch (node->code) { | |
2791 | case jit_code_label: | |
2792 | /* Do not consider jmpi and jmpr cannot jump to the | |
2793 | * next instruction. */ | |
2794 | next = _jitc->blocks.ptr + node->v.w; | |
2795 | /* Set of live registers in next block that are at unknown | |
2796 | * state in this block. */ | |
2797 | jit_regset_and(®temp, ®mask, &next->reglive); | |
2798 | if (jit_regset_set_p(®temp)) { | |
2799 | /* Add live state of next block to current block. */ | |
2800 | jit_regset_ior(&block->reglive, &block->reglive, ®temp); | |
2801 | /* Remove from unknown state bitmask. */ | |
2802 | jit_regset_com(®temp, ®temp); | |
2803 | jit_regset_and(&block->regmask, &block->regmask, ®temp); | |
c0c16242 | 2804 | block->again = 1; |
4a71579b PC |
2805 | } |
2806 | case jit_code_prolog: | |
2807 | case jit_code_epilog: | |
2808 | return; | |
2809 | case jit_code_callr: | |
2810 | value = jit_regno(node->u.w); | |
2811 | if (!(node->u.w & jit_regno_patch)) { | |
2812 | if (jit_regset_tstbit(®mask, value)) { | |
2813 | jit_regset_clrbit(®mask, value); | |
2814 | jit_regset_setbit(®live, value); | |
2815 | } | |
2816 | } | |
2817 | case jit_code_calli: | |
2818 | for (value = 0; value < _jitc->reglen; ++value) { | |
2819 | value = jit_regset_scan1(®mask, value); | |
2820 | if (value >= _jitc->reglen) | |
2821 | break; | |
2822 | spec = jit_class(_rvs[value].spec); | |
2823 | if (!(spec & jit_class_sav)) | |
2824 | jit_regset_clrbit(®mask, value); | |
2825 | if ((spec & jit_class_arg) && jit_regarg_p(node, value)) | |
2826 | jit_regset_setbit(®live, value); | |
2827 | } | |
2828 | break; | |
2829 | default: | |
2830 | value = jit_classify(node->code); | |
2831 | if (value & jit_cc_a2_reg) { | |
ba3814c1 PC |
2832 | if (value & jit_cc_a2_rlh) { |
2833 | if (!(node->w.q.l & jit_regno_patch)) { | |
2834 | /* Assume register is not changed */ | |
2835 | if (jit_regset_tstbit(®mask, node->w.q.l)) | |
2836 | jit_regset_clrbit(®mask, node->w.q.l); | |
2837 | } | |
2838 | if (!(node->w.q.h & jit_regno_patch)) { | |
2839 | if (jit_regset_tstbit(®mask, node->w.q.h)) | |
2840 | jit_regset_clrbit(®mask, node->w.q.h); | |
2841 | } | |
2842 | } | |
2843 | else { | |
ba86ff93 PC |
2844 | if (!(node->w.w & jit_regno_patch)) { |
2845 | if (jit_regset_tstbit(®mask, node->w.w)) { | |
2846 | jit_regset_clrbit(®mask, node->w.w); | |
2847 | if (!(value & jit_cc_a2_chg)) | |
2848 | jit_regset_setbit(®live, node->w.w); | |
ba3814c1 | 2849 | } |
4a71579b PC |
2850 | } |
2851 | } | |
2852 | } | |
2853 | if (value & jit_cc_a1_reg) { | |
ba86ff93 PC |
2854 | if (value & jit_cc_a1_rlh) { |
2855 | if (!(node->v.q.l & jit_regno_patch)) { | |
2856 | /* Assume register is not changed */ | |
2857 | if (jit_regset_tstbit(®mask, node->v.q.l)) | |
2858 | jit_regset_clrbit(®mask, node->v.q.l); | |
2859 | } | |
2860 | if (!(node->v.q.h & jit_regno_patch)) { | |
2861 | if (jit_regset_tstbit(®mask, node->v.q.h)) | |
2862 | jit_regset_clrbit(®mask, node->v.q.h); | |
2863 | } | |
2864 | } | |
2865 | else { | |
2866 | if (!(node->v.w & jit_regno_patch)) { | |
2867 | if (jit_regset_tstbit(®mask, node->v.w)) { | |
2868 | jit_regset_clrbit(®mask, node->v.w); | |
2869 | if (!(value & jit_cc_a1_chg)) | |
2870 | jit_regset_setbit(®live, node->v.w); | |
2871 | } | |
4a71579b PC |
2872 | } |
2873 | } | |
2874 | } | |
2875 | if (value & jit_cc_a0_reg) { | |
2876 | if (value & jit_cc_a0_rlh) { | |
2877 | if (!(node->u.q.l & jit_regno_patch)) { | |
2878 | if (jit_regset_tstbit(®mask, node->u.q.l)) { | |
2879 | jit_regset_clrbit(®mask, node->u.q.l); | |
2880 | if (!(value & jit_cc_a0_chg)) | |
2881 | jit_regset_setbit(®live, node->u.q.l); | |
2882 | } | |
2883 | } | |
2884 | if (!(node->u.q.h & jit_regno_patch)) { | |
2885 | if (jit_regset_tstbit(®mask, node->u.q.h)) { | |
2886 | jit_regset_clrbit(®mask, node->u.q.h); | |
2887 | if (!(value & jit_cc_a0_chg)) | |
2888 | jit_regset_setbit(®live, node->u.q.h); | |
2889 | } | |
2890 | } | |
2891 | } | |
2892 | else { | |
2893 | if (!(node->u.w & jit_regno_patch)) { | |
2894 | if (jit_regset_tstbit(®mask, node->u.w)) { | |
2895 | jit_regset_clrbit(®mask, node->u.w); | |
2896 | if (!(value & jit_cc_a0_chg)) | |
2897 | jit_regset_setbit(®live, node->u.w); | |
2898 | } | |
2899 | } | |
2900 | } | |
2901 | } | |
2902 | if (value & jit_cc_a0_jmp) { | |
2903 | if (node->flag & jit_flag_node) { | |
2904 | label = node->u.n; | |
2905 | /* Do not consider jmpi and jmpr cannot jump to the | |
2906 | * next instruction. */ | |
2907 | next = _jitc->blocks.ptr + label->v.w; | |
2908 | jit_regset_and(®temp, ®mask, &next->reglive); | |
2909 | if (jit_regset_set_p(®temp)) { | |
2910 | /* Add live state. */ | |
2911 | jit_regset_ior(&block->reglive, | |
2912 | &block->reglive, ®temp); | |
2913 | /* Remove from unknown state bitmask. */ | |
2914 | jit_regset_com(®temp, ®temp); | |
2915 | jit_regset_and(&block->regmask, | |
2916 | &block->regmask, ®temp); | |
c0c16242 | 2917 | block->again = 1; |
4a71579b PC |
2918 | } |
2919 | } | |
2920 | else { | |
2921 | /* Jump to unknown location. | |
2922 | * This is a pitfall of the implementation. | |
2923 | * Only jmpi to not a jit code should reach here, | |
2924 | * or a jmpr of a computed address. | |
2925 | * Because the implementation needs jit_class_nospill | |
2926 | * registers, must treat jmpr as a function call. This | |
2927 | * means that only JIT_Vn registers can be trusted on | |
2928 | * arrival of jmpr. | |
2929 | */ | |
ba3814c1 | 2930 | jit_regset_set_ui(®mask, 0); |
4a71579b PC |
2931 | for (regno = 0; regno < _jitc->reglen; regno++) { |
2932 | spec = jit_class(_rvs[regno].spec); | |
ba3814c1 PC |
2933 | if ((spec & (jit_class_gpr|jit_class_fpr)) && |
2934 | (spec & jit_class_sav)) | |
2935 | jit_regset_setbit(®mask, regno); | |
4a71579b PC |
2936 | } |
2937 | /* Assume non callee save registers are live due | |
2938 | * to jump to unknown location. */ | |
2939 | /* Treat all callee save as live. */ | |
ba3814c1 | 2940 | jit_regset_ior(&block->reglive, ®live, ®mask); |
4a71579b | 2941 | /* Treat anything else as dead. */ |
ba3814c1 | 2942 | return; |
4a71579b PC |
2943 | } |
2944 | } | |
2945 | break; | |
2946 | } | |
2947 | } | |
2948 | } | |
2949 | ||
2950 | /* Follow code generation up to finding a label or end of code. | |
2951 | * When finding a label, update the set of live registers. | |
2952 | * On branches, update based on taken branch or normal flow. | |
2953 | */ | |
2954 | static void | |
2955 | _jit_update(jit_state_t *_jit, jit_node_t *node, | |
2956 | jit_regset_t *live, jit_regset_t *mask) | |
2957 | { | |
2958 | jit_int32_t spec; | |
2959 | jit_int32_t regno; | |
2960 | unsigned long value; | |
2961 | jit_block_t *block; | |
2962 | jit_node_t *label; | |
2963 | jit_regset_t regtemp; | |
2964 | ||
2965 | for (; node; node = node->next) { | |
2966 | if (jit_regset_set_p(mask) == 0) | |
2967 | break; | |
2968 | switch (node->code) { | |
2969 | case jit_code_label: | |
2970 | block = _jitc->blocks.ptr + node->v.w; | |
2971 | jit_regset_and(®temp, mask, &block->reglive); | |
2972 | if (jit_regset_set_p(®temp)) { | |
2973 | /* Add live state. */ | |
2974 | jit_regset_ior(live, live, ®temp); | |
2975 | /* Remove from unknown state bitmask. */ | |
2976 | jit_regset_com(®temp, ®temp); | |
2977 | jit_regset_and(mask, mask, ®temp); | |
2978 | } | |
2979 | return; | |
2980 | case jit_code_prolog: | |
2981 | jit_regset_set_ui(mask, 0); | |
2982 | return; | |
2983 | case jit_code_epilog: | |
2984 | jit_regset_set_ui(mask, 0); | |
2985 | return; | |
2986 | case jit_code_callr: | |
2987 | value = jit_regno(node->u.w); | |
2988 | if (!(node->u.w & jit_regno_patch)) { | |
2989 | if (jit_regset_tstbit(mask, value)) { | |
2990 | jit_regset_clrbit(mask, value); | |
2991 | jit_regset_setbit(live, value); | |
2992 | } | |
2993 | } | |
2994 | case jit_code_calli: | |
2995 | for (value = 0; value < _jitc->reglen; ++value) { | |
2996 | value = jit_regset_scan1(mask, value); | |
2997 | if (value >= _jitc->reglen) | |
2998 | break; | |
2999 | spec = jit_class(_rvs[value].spec); | |
3000 | if (!(spec & jit_class_sav)) | |
3001 | jit_regset_clrbit(mask, value); | |
3002 | if ((spec & jit_class_arg) && jit_regarg_p(node, value)) | |
3003 | jit_regset_setbit(live, value); | |
3004 | } | |
3005 | break; | |
3006 | default: | |
3007 | value = jit_classify(node->code); | |
3008 | if (value & jit_cc_a2_reg) { | |
ba3814c1 PC |
3009 | if (value & jit_cc_a2_rlh) { |
3010 | if (!(node->w.q.l & jit_regno_patch)) { | |
3011 | /* Assume register is not changed */ | |
3012 | if (jit_regset_tstbit(mask, node->w.q.l)) | |
3013 | jit_regset_clrbit(mask, node->w.q.l); | |
3014 | } | |
3015 | if (!(node->w.q.h & jit_regno_patch)) { | |
3016 | if (jit_regset_tstbit(mask, node->w.q.h)) | |
3017 | jit_regset_clrbit(mask, node->w.q.h); | |
3018 | } | |
3019 | } | |
3020 | else { | |
3021 | if (!(node->w.w & jit_regno_patch)) { | |
3022 | if (jit_regset_tstbit(mask, node->w.w)) { | |
3023 | jit_regset_clrbit(mask, node->w.w); | |
3024 | if (!(value & jit_cc_a2_chg)) | |
3025 | jit_regset_setbit(live, node->w.w); | |
3026 | } | |
4a71579b PC |
3027 | } |
3028 | } | |
3029 | } | |
3030 | if (value & jit_cc_a1_reg) { | |
ba86ff93 PC |
3031 | if (value & jit_cc_a1_rlh) { |
3032 | if (!(node->v.q.l & jit_regno_patch)) { | |
3033 | /* Assume register is not changed */ | |
3034 | if (jit_regset_tstbit(mask, node->v.q.l)) | |
3035 | jit_regset_clrbit(mask, node->v.q.l); | |
3036 | } | |
3037 | if (!(node->v.q.h & jit_regno_patch)) { | |
3038 | if (jit_regset_tstbit(mask, node->v.q.h)) | |
3039 | jit_regset_clrbit(mask, node->v.q.h); | |
3040 | } | |
3041 | } | |
3042 | else { | |
3043 | if (!(node->v.w & jit_regno_patch)) { | |
3044 | if (jit_regset_tstbit(mask, node->v.w)) { | |
3045 | jit_regset_clrbit(mask, node->v.w); | |
3046 | if (!(value & jit_cc_a1_chg)) | |
3047 | jit_regset_setbit(live, node->v.w); | |
3048 | } | |
4a71579b PC |
3049 | } |
3050 | } | |
3051 | } | |
3052 | if (value & jit_cc_a0_reg) { | |
3053 | if (value & jit_cc_a0_rlh) { | |
3054 | if (!(node->u.q.l & jit_regno_patch)) { | |
3055 | if (jit_regset_tstbit(mask, node->u.q.l)) { | |
3056 | jit_regset_clrbit(mask, node->u.q.l); | |
3057 | if (!(value & jit_cc_a0_chg)) | |
3058 | jit_regset_setbit(live, node->u.q.l); | |
3059 | } | |
3060 | } | |
3061 | if (!(node->u.q.h & jit_regno_patch)) { | |
3062 | if (jit_regset_tstbit(mask, node->u.q.h)) { | |
3063 | jit_regset_clrbit(mask, node->u.q.h); | |
3064 | if (!(value & jit_cc_a0_chg)) | |
3065 | jit_regset_setbit(live, node->u.q.h); | |
3066 | } | |
3067 | } | |
3068 | } | |
3069 | else { | |
3070 | if (!(node->u.w & jit_regno_patch)) { | |
3071 | if (jit_regset_tstbit(mask, node->u.w)) { | |
3072 | jit_regset_clrbit(mask, node->u.w); | |
3073 | if (!(value & jit_cc_a0_chg)) | |
3074 | jit_regset_setbit(live, node->u.w); | |
3075 | } | |
3076 | } | |
3077 | } | |
3078 | } | |
3079 | if (value & jit_cc_a0_jmp) { | |
3080 | if (node->flag & jit_flag_node) { | |
3081 | label = node->u.n; | |
3082 | /* Do not consider jmpi and jmpr cannot jump to the | |
3083 | * next instruction. */ | |
3084 | block = _jitc->blocks.ptr + label->v.w; | |
3085 | jit_regset_and(®temp, mask, &block->reglive); | |
3086 | if (jit_regset_set_p(®temp)) { | |
3087 | /* Add live state. */ | |
3088 | jit_regset_ior(live, live, ®temp); | |
3089 | /* Remove from unknown state bitmask. */ | |
3090 | jit_regset_com(®temp, ®temp); | |
3091 | jit_regset_and(mask, mask, ®temp); | |
3092 | } | |
3093 | } | |
3094 | else { | |
3095 | /* Jump to unknown location. | |
3096 | * This is a pitfall of the implementation. | |
3097 | * Only jmpi to not a jit code should reach here, | |
3098 | * or a jmpr of a computed address. | |
3099 | * Because the implementation needs jit_class_nospill | |
3100 | * registers, must treat jmpr as a function call. This | |
3101 | * means that only JIT_Vn registers can be trusted on | |
3102 | * arrival of jmpr. | |
3103 | */ | |
ba3814c1 | 3104 | jit_regset_set_ui(mask, 0); |
4a71579b PC |
3105 | for (regno = 0; regno < _jitc->reglen; regno++) { |
3106 | spec = jit_class(_rvs[regno].spec); | |
ba3814c1 PC |
3107 | if ((spec & (jit_class_gpr|jit_class_fpr)) && |
3108 | (spec & jit_class_sav)) | |
3109 | jit_regset_setbit(mask, regno); | |
4a71579b PC |
3110 | } |
3111 | /* Assume non callee save registers are live due | |
3112 | * to jump to unknown location. */ | |
3113 | /* Treat all callee save as live. */ | |
3114 | jit_regset_ior(live, live, mask); | |
79bfeef6 PC |
3115 | /* Prevent explicitly set as live registers to |
3116 | * be used as a temporary for the jmpi. */ | |
3117 | jit_regset_ior(live, live, &_jitc->explive); | |
4a71579b | 3118 | /* Treat anything else as dead. */ |
ba3814c1 | 3119 | return; |
4a71579b PC |
3120 | } |
3121 | } | |
3122 | break; | |
3123 | } | |
3124 | } | |
3125 | } | |
3126 | ||
3127 | static void | |
3128 | _thread_jumps(jit_state_t *_jit) | |
3129 | { | |
3130 | jit_node_t *prev; | |
3131 | jit_node_t *node; | |
3132 | jit_node_t *next; | |
3133 | jit_int32_t mask; | |
3134 | ||
3135 | for (prev = node = _jitc->head; node;) { | |
3136 | next = node->next; | |
3137 | switch (node->code) { | |
3138 | case jit_code_jmpi: | |
3139 | if (redundant_jump(prev, node)) { | |
3140 | node = prev; | |
3141 | continue; | |
3142 | } | |
3143 | if (shortcut_jump(prev, node)) | |
3144 | continue; | |
3145 | break; | |
3146 | case jit_code_jmpr: | |
3147 | case jit_code_callr: case jit_code_calli: | |
3148 | /* non optimizable jump like code */ | |
3149 | break; | |
3150 | default: | |
3151 | mask = jit_classify(node->code); | |
3152 | if (mask & jit_cc_a0_jmp) { | |
3153 | if (reverse_jump(prev, node) || | |
3154 | shortcut_jump(prev, node)) | |
3155 | continue; | |
3156 | } | |
3157 | break; | |
3158 | } | |
3159 | prev = node; | |
3160 | node = next; | |
3161 | } | |
3162 | } | |
3163 | ||
3164 | static void | |
3165 | _sequential_labels(jit_state_t *_jit) | |
3166 | { | |
3167 | jit_node_t *jump; | |
3168 | jit_node_t *link; | |
3169 | jit_node_t *prev; | |
3170 | jit_node_t *next; | |
3171 | jit_node_t *node; | |
3172 | ||
3173 | for (prev = node = _jitc->head; node; node = next) { | |
3174 | next = node->next; | |
3175 | if (node->code == jit_code_label) { | |
3176 | if (!node->flag) { | |
3177 | if (!node->link) { | |
3178 | del_label(prev, node); | |
3179 | continue; | |
3180 | } | |
3181 | if (prev != node && prev->code == jit_code_label) { | |
3182 | if ((jump = node->link)) { | |
3183 | for (; jump; jump = link) { | |
3184 | link = jump->link; | |
79bfeef6 PC |
3185 | if (jump->code == jit_code_movi) |
3186 | jump->v.n = prev; | |
3187 | else | |
3188 | jump->u.n = prev; | |
4a71579b PC |
3189 | jump->link = prev->link; |
3190 | prev->link = jump; | |
3191 | } | |
3192 | node->link = NULL; | |
3193 | } | |
3194 | del_label(prev, node); | |
3195 | continue; | |
3196 | } | |
3197 | } | |
3198 | if (next && next->code == jit_code_label && !next->flag) { | |
3199 | if ((jump = next->link)) { | |
3200 | for (; jump; jump = link) { | |
3201 | link = jump->link; | |
79bfeef6 PC |
3202 | if (jump->code == jit_code_movi) |
3203 | jump->v.n = node; | |
3204 | else | |
3205 | jump->u.n = node; | |
4a71579b PC |
3206 | jump->link = node->link; |
3207 | node->link = jump; | |
3208 | } | |
3209 | next->link = NULL; | |
3210 | } | |
3211 | del_label(node, next); | |
3212 | next = node->next; | |
3213 | continue; | |
3214 | } | |
3215 | } | |
3216 | prev = node; | |
3217 | } | |
3218 | } | |
3219 | ||
3220 | static void | |
3221 | _split_branches(jit_state_t *_jit) | |
3222 | { | |
3223 | jit_node_t *node; | |
3224 | jit_node_t *next; | |
3225 | jit_node_t *label; | |
3226 | jit_block_t *block; | |
24d91c0d PC |
3227 | jit_block_t *blocks; |
3228 | jit_word_t offset; | |
3229 | jit_word_t length; | |
4a71579b | 3230 | |
24d91c0d PC |
3231 | length = _jitc->blocks.length; |
3232 | jit_alloc((jit_pointer_t *)&blocks, length * sizeof(jit_block_t)); | |
3233 | if ((node = _jitc->head) && | |
3234 | (node->code == jit_code_label || node->code == jit_code_prolog)) { | |
3235 | block = _jitc->blocks.ptr + node->v.w; | |
3236 | memcpy(blocks, block, sizeof(jit_block_t)); | |
3237 | node->v.w = 0; | |
3238 | offset = 1; | |
3239 | } | |
3240 | else | |
3241 | offset = 0; | |
4a71579b PC |
3242 | for (node = _jitc->head; node; node = next) { |
3243 | if ((next = node->next)) { | |
3244 | if (next->code == jit_code_label || | |
3245 | next->code == jit_code_prolog || | |
24d91c0d PC |
3246 | next->code == jit_code_epilog) { |
3247 | if (offset >= length) { | |
3248 | jit_realloc((jit_pointer_t *)&blocks, | |
3249 | length * sizeof(jit_block_t), | |
3250 | (length + 16) * sizeof(jit_block_t)); | |
3251 | length += 16; | |
3252 | } | |
3253 | block = _jitc->blocks.ptr + next->v.w; | |
3254 | memcpy(blocks + offset, block, sizeof(jit_block_t)); | |
3255 | next->v.w = offset++; | |
3256 | } | |
4a71579b | 3257 | /* split block on branches */ |
24d91c0d | 3258 | else if (jit_classify(node->code) & jit_cc_a0_jmp) { |
4a71579b PC |
3259 | label = new_node(jit_code_label); |
3260 | label->next = next; | |
3261 | node->next = label; | |
24d91c0d PC |
3262 | if (offset >= length) { |
3263 | jit_realloc((jit_pointer_t *)&blocks, | |
3264 | length * sizeof(jit_block_t), | |
3265 | (length + 16) * sizeof(jit_block_t)); | |
3266 | length += 16; | |
4a71579b | 3267 | } |
24d91c0d | 3268 | block = blocks + offset; |
4a71579b | 3269 | block->label = label; |
24d91c0d | 3270 | label->v.w = offset++; |
4a71579b PC |
3271 | jit_regset_new(&block->reglive); |
3272 | jit_regset_new(&block->regmask); | |
4a71579b PC |
3273 | } |
3274 | } | |
3275 | } | |
24d91c0d PC |
3276 | jit_free((jit_pointer_t *)&_jitc->blocks.ptr); |
3277 | _jitc->blocks.ptr = blocks; | |
3278 | _jitc->blocks.offset = offset; | |
3279 | _jitc->blocks.length = length; | |
4a71579b PC |
3280 | } |
3281 | ||
3282 | static jit_bool_t | |
3283 | _shortcut_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
3284 | { | |
3285 | jit_bool_t cond; | |
3286 | jit_node_t *jump; | |
3287 | jit_node_t *next; | |
3288 | jit_node_t *temp; | |
3289 | ||
3290 | if (!(node->flag & jit_flag_node)) | |
3291 | return (0); | |
3292 | assert(node->code != jit_code_jmpr); | |
3293 | cond = node->code != jit_code_jmpi; | |
3294 | jump = node->u.n; | |
3295 | for (next = jump->next; next; next = next->next) { | |
3296 | switch (next->code) { | |
3297 | case jit_code_jmpi: | |
3298 | if (!(next->flag & jit_flag_node)) | |
3299 | return (0); | |
3300 | if (jump->link == node) | |
3301 | jump->link = node->link; | |
3302 | else { | |
3303 | for (temp = jump->link; | |
3304 | temp->link != node; | |
3305 | temp = temp->link) | |
3306 | assert(temp != NULL); | |
3307 | temp->link = node->link; | |
3308 | } | |
3309 | jump = next->u.n; | |
3310 | node->u.n = jump; | |
3311 | node->link = jump->link; | |
3312 | jump->link = node; | |
3313 | return (1); | |
3314 | case jit_code_jmpr: | |
3315 | if (cond) | |
3316 | return (0); | |
3317 | node->code = jit_code_jmpr; | |
3318 | node->u.w = next->u.w; | |
3319 | node->link = NULL; | |
3320 | node->flag &= ~jit_flag_node; | |
3321 | return (1); | |
3322 | case jit_code_note: case jit_code_label: | |
3323 | break; | |
3324 | default: | |
3325 | return (0); | |
3326 | } | |
3327 | } | |
3328 | return (0); | |
3329 | } | |
3330 | ||
3331 | static jit_bool_t | |
3332 | _redundant_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
3333 | { | |
3334 | jit_node_t *local_prev; | |
3335 | jit_node_t *local_next; | |
3336 | ||
3337 | if (!(node->flag & jit_flag_node)) | |
3338 | return (0); | |
3339 | for (local_prev = node, local_next = node->next; | |
3340 | local_next; | |
3341 | local_prev = local_next, local_next = local_next->next) { | |
3342 | ||
3343 | switch (local_next->code) { | |
3344 | case jit_code_label: case jit_code_epilog: | |
3345 | if (node->u.n == local_next) { | |
3346 | if (local_next->link == node) | |
3347 | local_next->link = node->link; | |
3348 | else { | |
3349 | for (local_prev = local_next->link; | |
3350 | local_prev->link != node; | |
3351 | local_prev = local_prev->link) | |
3352 | assert(local_prev != NULL); | |
3353 | local_prev->link = node->link; | |
3354 | } | |
3355 | del_node(prev, node); | |
3356 | return (1); | |
3357 | } | |
3358 | break; | |
3359 | case jit_code_name: case jit_code_note: | |
4a71579b PC |
3360 | break; |
3361 | default: | |
3362 | return (0); | |
3363 | } | |
3364 | } | |
3365 | return (0); | |
3366 | } | |
3367 | ||
3368 | static jit_code_t | |
3369 | reverse_jump_code(jit_code_t code) | |
3370 | { | |
3371 | switch (code) { | |
3372 | case jit_code_bltr: return (jit_code_bger); | |
3373 | case jit_code_blti: return (jit_code_bgei); | |
3374 | case jit_code_bltr_u: return (jit_code_bger_u); | |
3375 | case jit_code_blti_u: return (jit_code_bgei_u); | |
3376 | case jit_code_bler: return (jit_code_bgtr); | |
3377 | case jit_code_blei: return (jit_code_bgti); | |
3378 | case jit_code_bler_u: return (jit_code_bgtr_u); | |
3379 | case jit_code_blei_u: return (jit_code_bgti_u); | |
3380 | case jit_code_beqr: return (jit_code_bner); | |
3381 | case jit_code_beqi: return (jit_code_bnei); | |
3382 | case jit_code_bger: return (jit_code_bltr); | |
3383 | case jit_code_bgei: return (jit_code_blti); | |
3384 | case jit_code_bger_u: return (jit_code_bltr_u); | |
3385 | case jit_code_bgei_u: return (jit_code_blti_u); | |
3386 | case jit_code_bgtr: return (jit_code_bler); | |
3387 | case jit_code_bgti: return (jit_code_blei); | |
3388 | case jit_code_bgtr_u: return (jit_code_bler_u); | |
3389 | case jit_code_bgti_u: return (jit_code_blei_u); | |
3390 | case jit_code_bner: return (jit_code_beqr); | |
3391 | case jit_code_bnei: return (jit_code_beqi); | |
3392 | case jit_code_bmsr: return (jit_code_bmcr); | |
3393 | case jit_code_bmsi: return (jit_code_bmci); | |
3394 | case jit_code_bmcr: return (jit_code_bmsr); | |
3395 | case jit_code_bmci: return (jit_code_bmsi); | |
3396 | case jit_code_bltr_f: return (jit_code_bunger_f); | |
3397 | case jit_code_blti_f: return (jit_code_bungei_f); | |
3398 | case jit_code_bler_f: return (jit_code_bungtr_f); | |
3399 | case jit_code_blei_f: return (jit_code_bungti_f); | |
3400 | ||
3401 | case jit_code_beqr_f: return (jit_code_bner_f); | |
3402 | case jit_code_beqi_f: return (jit_code_bnei_f); | |
3403 | ||
3404 | case jit_code_bger_f: return (jit_code_bunltr_f); | |
3405 | case jit_code_bgei_f: return (jit_code_bunlti_f); | |
3406 | case jit_code_bgtr_f: return (jit_code_bunler_f); | |
3407 | case jit_code_bgti_f: return (jit_code_bunlei_f); | |
3408 | ||
3409 | case jit_code_bner_f: return (jit_code_beqr_f); | |
79bfeef6 | 3410 | case jit_code_bnei_f: return (jit_code_beqi_f); |
4a71579b PC |
3411 | |
3412 | case jit_code_bunltr_f: return (jit_code_bger_f); | |
3413 | case jit_code_bunlti_f: return (jit_code_bgei_f); | |
3414 | case jit_code_bunler_f: return (jit_code_bgtr_f); | |
3415 | case jit_code_bunlei_f: return (jit_code_bgti_f); | |
3416 | ||
3417 | case jit_code_buneqr_f: return (jit_code_bltgtr_f); | |
3418 | case jit_code_buneqi_f: return (jit_code_bltgti_f); | |
3419 | ||
3420 | case jit_code_bunger_f: return (jit_code_bltr_f); | |
3421 | case jit_code_bungei_f: return (jit_code_blti_f); | |
3422 | case jit_code_bungtr_f: return (jit_code_bler_f); | |
3423 | case jit_code_bungti_f: return (jit_code_blei_f); | |
3424 | ||
3425 | case jit_code_bltgtr_f: return (jit_code_buneqr_f); | |
3426 | case jit_code_bltgti_f: return (jit_code_buneqi_f); | |
3427 | ||
3428 | case jit_code_bordr_f: return (jit_code_bunordr_f); | |
3429 | case jit_code_bordi_f: return (jit_code_bunordi_f); | |
3430 | case jit_code_bunordr_f:return (jit_code_bordr_f); | |
3431 | case jit_code_bunordi_f:return (jit_code_bordi_f); | |
3432 | case jit_code_bltr_d: return (jit_code_bunger_d); | |
3433 | case jit_code_blti_d: return (jit_code_bungei_d); | |
3434 | case jit_code_bler_d: return (jit_code_bungtr_d); | |
3435 | case jit_code_blei_d: return (jit_code_bungti_d); | |
3436 | ||
3437 | case jit_code_beqr_d: return (jit_code_bner_d); | |
3438 | case jit_code_beqi_d: return (jit_code_bnei_d); | |
3439 | ||
3440 | case jit_code_bger_d: return (jit_code_bunltr_d); | |
3441 | case jit_code_bgei_d: return (jit_code_bunlti_d); | |
3442 | case jit_code_bgtr_d: return (jit_code_bunler_d); | |
3443 | case jit_code_bgti_d: return (jit_code_bunlei_d); | |
3444 | ||
3445 | case jit_code_bner_d: return (jit_code_beqr_d); | |
3446 | case jit_code_bnei_d: return (jit_code_beqi_d); | |
3447 | ||
3448 | case jit_code_bunltr_d: return (jit_code_bger_d); | |
3449 | case jit_code_bunlti_d: return (jit_code_bgei_d); | |
3450 | case jit_code_bunler_d: return (jit_code_bgtr_d); | |
3451 | case jit_code_bunlei_d: return (jit_code_bgti_d); | |
3452 | ||
3453 | case jit_code_buneqr_d: return (jit_code_bltgtr_d); | |
3454 | case jit_code_buneqi_d: return (jit_code_bltgti_d); | |
3455 | ||
3456 | case jit_code_bunger_d: return (jit_code_bltr_d); | |
3457 | case jit_code_bungei_d: return (jit_code_blti_d); | |
3458 | case jit_code_bungtr_d: return (jit_code_bler_d); | |
3459 | case jit_code_bungti_d: return (jit_code_blei_d); | |
3460 | ||
3461 | case jit_code_bltgtr_d: return (jit_code_buneqr_d); | |
3462 | case jit_code_bltgti_d: return (jit_code_buneqi_d); | |
3463 | ||
3464 | case jit_code_bordr_d: return (jit_code_bunordr_d); | |
3465 | case jit_code_bordi_d: return (jit_code_bunordi_d); | |
3466 | case jit_code_bunordr_d:return (jit_code_bordr_d); | |
3467 | case jit_code_bunordi_d:return (jit_code_bordi_d); | |
3468 | case jit_code_boaddr: return (jit_code_bxaddr); | |
3469 | case jit_code_boaddi: return (jit_code_bxaddi); | |
3470 | case jit_code_boaddr_u: return (jit_code_bxaddr_u); | |
3471 | case jit_code_boaddi_u: return (jit_code_bxaddi_u); | |
3472 | case jit_code_bxaddr: return (jit_code_boaddr); | |
3473 | case jit_code_bxaddi: return (jit_code_boaddi); | |
3474 | case jit_code_bxaddr_u: return (jit_code_boaddr_u); | |
3475 | case jit_code_bxaddi_u: return (jit_code_boaddi_u); | |
3476 | case jit_code_bosubr: return (jit_code_bxsubr); | |
3477 | case jit_code_bosubi: return (jit_code_bxsubi); | |
3478 | case jit_code_bosubr_u: return (jit_code_bxsubr_u); | |
3479 | case jit_code_bosubi_u: return (jit_code_bxsubi_u); | |
3480 | case jit_code_bxsubr: return (jit_code_bosubr); | |
3481 | case jit_code_bxsubi: return (jit_code_bosubi); | |
3482 | case jit_code_bxsubr_u: return (jit_code_bosubr_u); | |
3483 | case jit_code_bxsubi_u: return (jit_code_bosubi_u); | |
3484 | default: abort(); /* invalid jump code */ | |
3485 | } | |
3486 | } | |
3487 | ||
3488 | /* | |
3489 | * change common pattern: | |
3490 | * <cond_jump L0> <jump L1> <label L0> | |
3491 | * into | |
3492 | * <reverse_cond_jump L1> | |
3493 | */ | |
3494 | static jit_bool_t | |
3495 | _reverse_jump(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
3496 | { | |
3497 | jit_node_t *local_prev; | |
3498 | jit_node_t *local_next; | |
3499 | jit_node_t *local_jump; | |
3500 | ||
3501 | if (!(node->flag & jit_flag_node)) | |
3502 | return (0); | |
3503 | /* =><cond_jump L0> <jump L1> <label L0> */ | |
3504 | local_next = node->next; | |
3505 | if (local_next->code != jit_code_jmpi || | |
3506 | !(local_next->flag & jit_flag_node)) | |
3507 | return (0); | |
3508 | /* <cond_jump L0> =><jump L1> <label L0> */ | |
3509 | ||
3510 | local_jump = local_next->u.n; | |
3511 | for (local_prev = local_next, local_next = local_next->next; | |
3512 | local_next; | |
3513 | local_prev = local_next, local_next = local_next->next) { | |
3514 | switch (local_next->code) { | |
3515 | case jit_code_label: case jit_code_epilog: | |
3516 | if (node->u.n == local_next) { | |
3517 | if (local_next->link == node) | |
3518 | local_next->link = node->link; | |
3519 | else { | |
3520 | for (local_prev = local_next->link; | |
3521 | local_prev->link != node; | |
3522 | local_prev = local_prev->link) | |
3523 | assert(local_prev != NULL); | |
3524 | local_prev->link = node->link; | |
3525 | } | |
3526 | del_node(node, node->next); | |
3527 | node->code = reverse_jump_code(node->code); | |
3528 | node->u.n = local_jump; | |
3529 | node->link = local_jump->link; | |
3530 | local_jump->link = node; | |
3531 | return (1); | |
3532 | } | |
3533 | break; | |
3534 | case jit_code_note: | |
3535 | break; | |
3536 | default: | |
3537 | return (0); | |
3538 | } | |
3539 | } | |
3540 | return (0); | |
3541 | } | |
3542 | ||
ba3814c1 | 3543 | static jit_bool_t |
4a71579b PC |
3544 | _redundant_store(jit_state_t *_jit, jit_node_t *node, jit_bool_t jump) |
3545 | { | |
3546 | jit_node_t *iter; | |
3547 | jit_node_t *prev; | |
3548 | jit_word_t word; | |
3549 | jit_int32_t spec; | |
3550 | jit_int32_t regno; | |
ba3814c1 | 3551 | jit_bool_t result; |
4a71579b PC |
3552 | |
3553 | if (jump) { | |
3554 | prev = node->u.n; | |
3555 | if (prev->code == jit_code_epilog) | |
ba3814c1 | 3556 | return (0); |
4a71579b PC |
3557 | assert(prev->code == jit_code_label); |
3558 | if ((prev->flag & jit_flag_head) || node->link || prev->link != node) | |
3559 | /* multiple sources */ | |
ba3814c1 | 3560 | return (0); |
4a71579b PC |
3561 | /* if there are sequential labels it will return below */ |
3562 | } | |
3563 | else | |
3564 | prev = node; | |
ba3814c1 | 3565 | result = 0; |
4a71579b PC |
3566 | word = node->w.w; |
3567 | regno = jit_regno(node->v.w); | |
3568 | for (iter = prev->next; iter; prev = iter, iter = iter->next) { | |
3569 | switch (iter->code) { | |
3570 | case jit_code_label: case jit_code_prolog: | |
3571 | case jit_code_epilog: | |
ba3814c1 | 3572 | return (result); |
4a71579b PC |
3573 | case jit_code_movi: |
3574 | if (regno == jit_regno(iter->u.w)) { | |
3575 | if (iter->flag || iter->v.w != word) | |
ba3814c1 PC |
3576 | return (result); |
3577 | result = 1; | |
4a71579b PC |
3578 | del_node(prev, iter); |
3579 | iter = prev; | |
3580 | } | |
3581 | break; | |
3582 | default: | |
3583 | spec = jit_classify(iter->code); | |
3584 | if (spec & jit_cc_a0_jmp) | |
ba3814c1 | 3585 | return (result); |
4a71579b PC |
3586 | if ((spec & (jit_cc_a0_reg|jit_cc_a0_chg)) == |
3587 | (jit_cc_a0_reg|jit_cc_a0_chg)) { | |
3588 | if (spec & jit_cc_a0_rlh) { | |
3589 | if (regno == jit_regno(iter->u.q.l) || | |
3590 | regno == jit_regno(iter->u.q.h)) | |
ba3814c1 | 3591 | return (result); |
4a71579b PC |
3592 | } |
3593 | else { | |
3594 | if (regno == jit_regno(iter->u.w)) | |
ba3814c1 | 3595 | return (result); |
4a71579b PC |
3596 | } |
3597 | } | |
3598 | if ((spec & (jit_cc_a1_reg|jit_cc_a1_chg)) == | |
3599 | (jit_cc_a1_reg|jit_cc_a1_chg)) { | |
3600 | if (regno == jit_regno(iter->v.w)) | |
ba3814c1 | 3601 | return (result); |
4a71579b PC |
3602 | } |
3603 | if ((spec & (jit_cc_a2_reg|jit_cc_a2_chg)) == | |
3604 | (jit_cc_a2_reg|jit_cc_a2_chg)) { | |
3605 | if (regno == jit_regno(iter->w.w)) | |
ba3814c1 | 3606 | return (result); |
4a71579b PC |
3607 | } |
3608 | break; | |
3609 | } | |
3610 | } | |
c0c16242 PC |
3611 | |
3612 | return (result); | |
4a71579b PC |
3613 | } |
3614 | ||
3615 | static jit_bool_t | |
3616 | _simplify_movr(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node, | |
3617 | jit_int32_t kind, jit_int32_t size) | |
3618 | { | |
3619 | jit_int32_t regno; | |
3620 | jit_int32_t right; | |
ba86ff93 | 3621 | jit_int32_t spec; |
4a71579b PC |
3622 | jit_value_t *value; |
3623 | ||
3624 | regno = jit_regno(node->u.w); | |
3625 | right = jit_regno(node->v.w); | |
3626 | value = _jitc->values + regno; | |
ba86ff93 PC |
3627 | |
3628 | spec = jit_class(_rvs[regno].spec); | |
3629 | if (!(spec & (jit_class_gpr | jit_class_fpr))) { | |
3630 | /* reserved register */ | |
3631 | value->kind = 0; | |
3632 | ++_jitc->gen[regno]; | |
3633 | return (0); | |
3634 | } | |
3635 | ||
4a71579b PC |
3636 | if ((value->kind == jit_kind_register && |
3637 | jit_regno(value->base.q.l) == right && | |
3638 | value->base.q.h == _jitc->gen[right]) || | |
3639 | (value->kind == kind && _jitc->values[right].kind == kind && | |
3640 | memcmp(&value->base.w, &_jitc->values[right].base.w, size) == 0)) { | |
3641 | del_node(prev, node); | |
3642 | return (1); | |
3643 | } | |
3644 | if (_jitc->values[right].kind == jit_kind_word) | |
3645 | jit_memcpy(value, _jitc->values + right, sizeof(jit_value_t)); | |
3646 | else { | |
3647 | value->kind = jit_kind_register; | |
3648 | value->base.q.l = right; | |
3649 | value->base.q.h = _jitc->gen[right]; | |
3650 | } | |
3651 | ++_jitc->gen[regno]; | |
3652 | ||
3653 | return (0); | |
3654 | } | |
3655 | ||
3656 | static jit_bool_t | |
3657 | _simplify_movi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node, | |
3658 | jit_int32_t kind, jit_int32_t size) | |
3659 | { | |
3660 | jit_value_t *value; | |
3661 | jit_int32_t spec; | |
3662 | jit_int32_t regno; | |
3663 | jit_int32_t offset; | |
3664 | ||
3665 | regno = jit_regno(node->u.w); | |
3666 | value = _jitc->values + regno; | |
3667 | if (node->flag & jit_flag_node) { | |
3668 | /* set to undefined if value will be patched */ | |
3669 | value->kind = 0; | |
3670 | ++_jitc->gen[regno]; | |
3671 | return (0); | |
3672 | } | |
ba86ff93 PC |
3673 | spec = jit_class(_rvs[regno].spec); |
3674 | if (!(spec & (jit_class_gpr | jit_class_fpr))) { | |
3675 | /* reserved register */ | |
3676 | value->kind = 0; | |
3677 | ++_jitc->gen[regno]; | |
3678 | return (0); | |
3679 | } | |
4a71579b PC |
3680 | if (value->kind == kind) { |
3681 | if (memcmp(&node->v.w, &value->base.w, size) == 0) { | |
3682 | del_node(prev, node); | |
3683 | return (1); | |
3684 | } | |
4a71579b PC |
3685 | if (kind == jit_kind_word) |
3686 | spec &= jit_class_gpr; | |
3687 | else | |
3688 | spec &= (jit_class_xpr | jit_class_fpr); | |
3689 | for (offset = 0; offset < _jitc->reglen; offset++) { | |
3690 | if (_jitc->values[offset].kind == kind && | |
3691 | memcmp(&node->v.w, &_jitc->values[offset].base.w, size) == 0 && | |
3692 | (jit_class(_rvs[offset].spec) & spec) == spec) { | |
3693 | if (kind == jit_kind_word) | |
3694 | node->code = jit_code_movr; | |
3695 | else if (kind == jit_kind_float32) | |
3696 | node->code = jit_code_movr_f; | |
3697 | else | |
3698 | node->code = jit_code_movr_d; | |
3699 | node->v.w = offset; | |
3700 | jit_memcpy(value, _jitc->values + offset, sizeof(jit_value_t)); | |
3701 | ++_jitc->gen[regno]; | |
3702 | return (0); | |
3703 | } | |
3704 | } | |
3705 | } | |
3706 | value->kind = kind; | |
3707 | jit_memcpy(&value->base.w, &node->v.w, size); | |
3708 | ++_jitc->gen[regno]; | |
3709 | ||
3710 | return (0); | |
3711 | } | |
3712 | ||
3713 | /* simple/safe redundandy test not checking if another register | |
3714 | * holds the same value | |
3715 | */ | |
3716 | static jit_bool_t | |
3717 | _simplify_ldxi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
3718 | { | |
3719 | jit_value_t *value; | |
3720 | jit_int32_t regno; | |
3721 | jit_int32_t right; | |
3722 | ||
3723 | regno = jit_regno(node->u.w); | |
3724 | right = jit_regno(node->v.w); | |
3725 | value = _jitc->values + regno; | |
3726 | if (regno != right && | |
3727 | value->kind == jit_kind_code && value->code == node->code && | |
3728 | value->base.q.l == right && value->base.q.h == _jitc->gen[right] && | |
3729 | node->w.w == value->disp.w) { | |
3730 | del_node(prev, node); | |
3731 | return (1); | |
3732 | } | |
3733 | value->kind = jit_kind_code; | |
3734 | value->code = node->code; | |
3735 | value->base.q.l = right; | |
3736 | value->base.q.h = _jitc->gen[right]; | |
3737 | value->disp.w = node->w.w; | |
3738 | ++_jitc->gen[regno]; | |
3739 | ||
3740 | return (0); | |
3741 | } | |
3742 | ||
3743 | static jit_bool_t | |
3744 | _simplify_stxi(jit_state_t *_jit, jit_node_t *prev, jit_node_t *node) | |
3745 | { | |
3746 | jit_value_t *value; | |
3747 | jit_int32_t regno; | |
3748 | jit_int32_t right; | |
3749 | jit_int32_t offset; | |
3750 | ||
3751 | regno = jit_regno(node->w.w); | |
3752 | right = jit_regno(node->v.w); | |
3753 | value = _jitc->values + regno; | |
3754 | ||
3755 | /* check for redundant store after load */ | |
3756 | if (regno != right && | |
3757 | value->kind == jit_kind_code && value->code == node->code && | |
3758 | value->base.q.l == right && value->base.q.h == _jitc->gen[right] && | |
3759 | node->u.w == value->disp.w) { | |
3760 | del_node(prev, node); | |
3761 | return (1); | |
3762 | } | |
3763 | ||
3764 | /* assume anything can alias, and invalidate tracked values */ | |
3765 | for (offset = 0; offset < _jitc->reglen; offset++) { | |
3766 | if (_jitc->values[offset].kind == jit_kind_code) { | |
3767 | _jitc->values[offset].kind = 0; | |
3768 | ++_jitc->gen[offset]; | |
3769 | } | |
3770 | } | |
3771 | ||
3772 | /* no multiple information, so, if set to a constant, | |
3773 | * prefer to keep that information */ | |
3774 | if (value->kind == 0) { | |
4a71579b PC |
3775 | switch (node->code) { |
3776 | /* no information about signed/unsigned either */ | |
3777 | case jit_code_stxi_c: value->code = jit_code_ldxi_c; break; | |
3778 | case jit_code_stxi_s: value->code = jit_code_ldxi_s; break; | |
3779 | case jit_code_stxi_i: value->code = jit_code_ldxi_i; break; | |
3780 | case jit_code_stxi_l: value->code = jit_code_ldxi_l; break; | |
3781 | case jit_code_stxi_f: value->code = jit_code_ldxi_f; break; | |
3782 | case jit_code_stxi_d: value->code = jit_code_ldxi_d; break; | |
3783 | default: abort(); | |
3784 | } | |
3785 | value->kind = jit_kind_code; | |
3786 | value->base.q.l = right; | |
3787 | value->base.q.h = _jitc->gen[right]; | |
3788 | value->disp.w = node->u.w; | |
3789 | } | |
3790 | ||
3791 | return (0); | |
3792 | } | |
3793 | ||
3794 | /* usually there should be only one store in the | |
3795 | * jit_get_reg/jit_unget_reg, but properly handle | |
3796 | * multiple ones by moving the save node */ | |
3797 | static void | |
3798 | _simplify_spill(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno) | |
3799 | { | |
3800 | jit_node_t *save; | |
3801 | jit_node_t *temp; | |
3802 | ||
3803 | if ((temp = _jitc->spill[regno]) && (save = temp->next) != node) { | |
3804 | temp->next = save->next; | |
3805 | save->next = node->next; | |
3806 | node->next = save; | |
3807 | _jitc->spill[regno] = node; | |
3808 | } | |
3809 | } | |
3810 | ||
3811 | /* checks for simple cases where a register is set more than | |
3812 | * once to the same value, and is a common pattern of calls | |
3813 | * to jit_pushargi and jit_pushargr | |
3814 | */ | |
ba3814c1 | 3815 | static jit_bool_t |
4a71579b PC |
3816 | _simplify(jit_state_t *_jit) |
3817 | { | |
3818 | jit_node_t *prev; | |
3819 | jit_node_t *node; | |
3820 | jit_node_t *next; | |
3821 | jit_int32_t info; | |
3822 | jit_int32_t regno; | |
ba3814c1 | 3823 | jit_bool_t result; |
4a71579b | 3824 | |
ba3814c1 | 3825 | result = 0; |
4a71579b PC |
3826 | for (prev = NULL, node = _jitc->head; node; prev = node, node = next) { |
3827 | next = node->next; | |
3828 | switch (node->code) { | |
3829 | case jit_code_label: case jit_code_prolog: | |
3830 | case jit_code_callr: case jit_code_calli: | |
3831 | reset: | |
3832 | memset(_jitc->gen, 0, sizeof(jit_int32_t) * _jitc->reglen); | |
3833 | memset(_jitc->values, 0, sizeof(jit_value_t) * _jitc->reglen); | |
3834 | break; | |
3835 | case jit_code_save: | |
3836 | _jitc->spill[jit_regno(node->u.w)] = prev; | |
3837 | break; | |
3838 | case jit_code_load: | |
3839 | regno = jit_regno(node->u.w); | |
3840 | if (register_change_p(node->link->next, node, regno) != | |
3841 | jit_reg_change) { | |
3842 | /* spill not required due to optimizing common | |
3843 | * redundancy case of calling jit_get_reg/jit_unget_reg | |
3844 | * and then setting the register to the value it is | |
3845 | * already holding */ | |
3846 | patch_register(node->link->next, node, | |
3847 | jit_regno_patch|regno, regno); | |
ba3814c1 | 3848 | result = 1; |
4a71579b PC |
3849 | del_node(_jitc->spill[regno], node->link); |
3850 | del_node(prev, node); | |
3851 | node = prev; | |
3852 | } | |
3853 | _jitc->spill[regno] = NULL; | |
3854 | break; | |
3855 | case jit_code_movr: | |
3856 | regno = jit_regno(node->u.w); | |
3857 | if (simplify_movr(prev, node, | |
ba3814c1 PC |
3858 | jit_kind_word, sizeof(jit_word_t))) { |
3859 | result = 1; | |
4a71579b | 3860 | simplify_spill(node = prev, regno); |
ba3814c1 | 3861 | } |
4a71579b PC |
3862 | break; |
3863 | case jit_code_movi: | |
3864 | regno = jit_regno(node->u.w); | |
3865 | if (simplify_movi(prev, node, | |
ba3814c1 PC |
3866 | jit_kind_word, sizeof(jit_word_t))) { |
3867 | result = 1; | |
4a71579b | 3868 | simplify_spill(node = prev, regno); |
ba3814c1 | 3869 | } |
4a71579b PC |
3870 | break; |
3871 | case jit_code_movr_f: | |
3872 | regno = jit_regno(node->u.w); | |
3873 | if (simplify_movr(prev, node, | |
ba3814c1 PC |
3874 | jit_kind_float32, sizeof(jit_float32_t))) { |
3875 | result = 1; | |
4a71579b | 3876 | simplify_spill(node = prev, regno); |
ba3814c1 | 3877 | } |
4a71579b PC |
3878 | break; |
3879 | case jit_code_movi_f: | |
3880 | regno = jit_regno(node->u.w); | |
3881 | if (simplify_movi(prev, node, | |
ba3814c1 PC |
3882 | jit_kind_float32, sizeof(jit_float32_t))) { |
3883 | result = 1; | |
4a71579b | 3884 | simplify_spill(node = prev, regno); |
ba3814c1 | 3885 | } |
4a71579b PC |
3886 | break; |
3887 | case jit_code_movr_d: | |
3888 | regno = jit_regno(node->u.w); | |
3889 | if (simplify_movr(prev, node, | |
ba3814c1 PC |
3890 | jit_kind_float64, sizeof(jit_float64_t))) { |
3891 | result = 1; | |
4a71579b | 3892 | simplify_spill(node = prev, regno); |
ba3814c1 | 3893 | } |
4a71579b PC |
3894 | break; |
3895 | case jit_code_movi_d: | |
3896 | regno = jit_regno(node->u.w); | |
3897 | if (simplify_movi(prev, node, | |
ba3814c1 PC |
3898 | jit_kind_float64, sizeof(jit_float64_t))) { |
3899 | result = 1; | |
4a71579b | 3900 | simplify_spill(node = prev, regno); |
ba3814c1 | 3901 | } |
4a71579b PC |
3902 | break; |
3903 | case jit_code_ldxi_c: case jit_code_ldxi_uc: | |
3904 | case jit_code_ldxi_s: case jit_code_ldxi_us: | |
3905 | case jit_code_ldxi_i: case jit_code_ldxi_ui: | |
3906 | case jit_code_ldxi_l: | |
3907 | case jit_code_ldxi_f: case jit_code_ldxi_d: | |
3908 | regno = jit_regno(node->u.w); | |
ba3814c1 PC |
3909 | if (simplify_ldxi(prev, node)) { |
3910 | result = 1; | |
4a71579b | 3911 | simplify_spill(node = prev, regno); |
ba3814c1 | 3912 | } |
4a71579b PC |
3913 | break; |
3914 | case jit_code_stxi_c: case jit_code_stxi_s: | |
3915 | case jit_code_stxi_i: case jit_code_stxi_l: | |
3916 | case jit_code_stxi_f: case jit_code_stxi_d: | |
3917 | regno = jit_regno(node->u.w); | |
ba3814c1 PC |
3918 | if (simplify_stxi(prev, node)) { |
3919 | result = 1; | |
4a71579b | 3920 | simplify_spill(node = prev, regno); |
ba3814c1 | 3921 | } |
4a71579b PC |
3922 | break; |
3923 | default: | |
3924 | info = jit_classify(node->code); | |
3925 | if (info & jit_cc_a0_jmp) | |
3926 | /* labels are not implicitly added when not taking | |
3927 | * a conditional branch */ | |
3928 | goto reset; | |
3929 | if (info & jit_cc_a0_chg) { | |
3930 | if (info & jit_cc_a0_rlh) { | |
3931 | regno = jit_regno(node->u.q.l); | |
3932 | _jitc->values[regno].kind = 0; | |
3933 | ++_jitc->gen[regno]; | |
3934 | regno = jit_regno(node->u.q.h); | |
3935 | _jitc->values[regno].kind = 0; | |
3936 | ++_jitc->gen[regno]; | |
3937 | } | |
3938 | else { | |
3939 | regno = jit_regno(node->u.w); | |
3940 | _jitc->values[regno].kind = 0; | |
3941 | ++_jitc->gen[regno]; | |
3942 | } | |
3943 | } | |
3944 | if (info & jit_cc_a1_chg) { | |
ba86ff93 PC |
3945 | #if 0 |
3946 | /* Assume registers are not changed */ | |
3947 | if (info & jit_cc_a1_rlh) { | |
3948 | regno = jit_regno(node->v.q.l); | |
3949 | _jitc->values[regno].kind = 0; | |
3950 | ++_jitc->gen[regno]; | |
3951 | regno = jit_regno(node->v.q.h); | |
3952 | _jitc->values[regno].kind = 0; | |
3953 | ++_jitc->gen[regno]; | |
3954 | } | |
3955 | else { | |
3956 | #endif | |
3957 | regno = jit_regno(node->v.w); | |
3958 | _jitc->values[regno].kind = 0; | |
3959 | ++_jitc->gen[regno]; | |
3960 | #if 0 | |
3961 | } | |
3962 | #endif | |
4a71579b PC |
3963 | } |
3964 | if (info & jit_cc_a2_chg) { | |
ba3814c1 PC |
3965 | #if 0 |
3966 | /* Assume registers are not changed */ | |
3967 | if (info & jit_cc_a2_rlh) { | |
3968 | regno = jit_regno(node->w.q.l); | |
3969 | _jitc->values[regno].kind = 0; | |
3970 | ++_jitc->gen[regno]; | |
3971 | regno = jit_regno(node->w.q.h); | |
3972 | _jitc->values[regno].kind = 0; | |
3973 | ++_jitc->gen[regno]; | |
3974 | } | |
3975 | else { | |
3976 | #endif | |
3977 | regno = jit_regno(node->w.w); | |
3978 | _jitc->values[regno].kind = 0; | |
3979 | ++_jitc->gen[regno]; | |
3980 | #if 0 | |
3981 | } | |
3982 | #endif | |
4a71579b PC |
3983 | } |
3984 | break; | |
3985 | } | |
3986 | } | |
ba3814c1 | 3987 | return (result); |
4a71579b PC |
3988 | } |
3989 | ||
3990 | static jit_int32_t | |
3991 | _register_change_p(jit_state_t *_jit, jit_node_t *node, jit_node_t *link, | |
3992 | jit_int32_t regno) | |
3993 | { | |
3994 | jit_int32_t value; | |
ba86ff93 | 3995 | jit_bool_t use, change; |
4a71579b PC |
3996 | |
3997 | for (; node != link; node = node->next) { | |
3998 | switch (node->code) { | |
3999 | case jit_code_label: case jit_code_prolog: | |
ba86ff93 PC |
4000 | /* lack of extra information so assume it is live */ |
4001 | return (jit_reg_static); | |
4a71579b PC |
4002 | case jit_code_callr: case jit_code_calli: |
4003 | if (!(jit_class(_rvs[regno].spec) & jit_class_sav)) | |
4004 | return (jit_reg_undef); | |
4005 | break; | |
4006 | default: | |
4007 | value = jit_classify(node->code); | |
ba86ff93 PC |
4008 | use = change = 0; |
4009 | if (value & jit_cc_a0_rlh) { | |
4010 | if (node->u.q.l == regno || node->u.q.h == regno) { | |
4011 | if (value & jit_cc_a0_chg) | |
4012 | change = 1; | |
4013 | use = !change; | |
4014 | } | |
4015 | } | |
4016 | else if ((value & jit_cc_a0_reg) && node->u.w == regno) { | |
4017 | if (value & jit_cc_a0_chg) | |
4018 | change = 1; | |
4019 | use = !change; | |
4020 | } | |
4021 | if (!use && | |
4022 | (value & jit_cc_a1_reg) && node->v.w == regno) { | |
4023 | if (value & jit_cc_a1_chg) | |
4024 | change = 1; | |
4025 | use = !change; | |
4026 | } | |
4027 | if (!use && | |
4028 | (value & jit_cc_a2_reg) && node->w.w == regno) { | |
4029 | if (value & jit_cc_a2_chg) | |
4030 | change = 1; | |
4031 | use = !change; | |
4032 | } | |
4a71579b | 4033 | /* lack of extra information */ |
ba86ff93 PC |
4034 | if (!use && (value & (jit_cc_a0_jmp|jit_cc_a0_cnd)) && |
4035 | /* In case of indirect branches, always consider | |
4036 | * callee save registers as live. */ | |
4037 | !(jit_class(_rvs[regno].spec) & jit_class_sav)) | |
4038 | change = 1; | |
4039 | if (!use && change) | |
4a71579b PC |
4040 | return (jit_reg_change); |
4041 | } | |
4042 | } | |
4043 | ||
4044 | return (jit_reg_static); | |
4045 | } | |
4046 | ||
4047 | /* most of this could be done at the same time as generating jit, but | |
4048 | * avoid complications on different cpu backends and patch spill/loads | |
4049 | * here, by simulating jit generation */ | |
4050 | static jit_bool_t | |
4051 | _spill_reglive_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno) | |
4052 | { | |
4053 | if (!jit_regset_tstbit(&_jitc->reglive, regno)) { | |
4054 | jit_regset_setbit(&_jitc->regmask, regno); | |
4055 | jit_update(node->next, &_jitc->reglive, &_jitc->regmask); | |
4056 | if (!jit_regset_tstbit(&_jitc->reglive, regno) && | |
4057 | register_change_p(node->next, node->link, regno) != jit_reg_change) | |
4058 | return (0); | |
4059 | } | |
4060 | ||
4061 | return (1); | |
4062 | } | |
4063 | ||
4064 | static void | |
4065 | _patch_registers(jit_state_t *_jit) | |
4066 | { | |
4067 | jit_node_t *prev; | |
4068 | jit_node_t *node; | |
4069 | jit_node_t *next; | |
4070 | jit_int32_t info; | |
4071 | jit_int32_t spec; | |
4072 | jit_int32_t regno; | |
4073 | jit_int32_t value; | |
4074 | ||
4075 | _jitc->function = NULL; | |
4076 | ||
4077 | jit_reglive_setup(); | |
4078 | for (prev = NULL, node = _jitc->head; node; node = next) { | |
4079 | next = node->next; | |
4080 | ||
4081 | info = jit_classify(node->code); | |
4082 | jit_regarg_set(node, info); | |
4083 | ||
4084 | switch (node->code) { | |
4085 | case jit_code_save: | |
4086 | regno = jit_regno(node->u.w); | |
4087 | if (!spill_reglive_p(node, regno)) { | |
4088 | /* register is not live, just remove spill/reload */ | |
4089 | jit_regarg_clr(node, info); | |
4090 | node->link->v.w = jit_regload_delete; | |
4091 | del_node(prev, node); | |
4092 | continue; | |
4093 | } | |
4094 | else { | |
4095 | /* try to find a free register of the same class */ | |
4096 | spec = jit_class(_rvs[regno].spec) & ~jit_class_arg; | |
4097 | for (value = 0; value < _jitc->reglen; value++) { | |
4098 | if (value != regno && | |
4099 | ((jit_class(_rvs[value].spec) & spec) & | |
4100 | ~jit_class_arg) == spec && | |
4101 | !jit_regset_tstbit(&_jitc->regarg, value) && | |
4102 | !spill_reglive_p(node, value)) | |
4103 | break; | |
4104 | } | |
4105 | if (value < _jitc->reglen) { | |
4106 | jit_regarg_clr(node, info); | |
4107 | patch_register(node->next, node->link, | |
4108 | jit_regno_patch|node->u.w, | |
4109 | jit_regno_patch|value); | |
4110 | /* mark as live just in case there are nested | |
4111 | * register patches, so that next patch will | |
4112 | * not want to use the same register */ | |
4113 | jit_regset_setbit(&_jitc->reglive, value); | |
4114 | /* register is not live, just remove spill/reload */ | |
4115 | node->link->v.w = jit_regload_isdead; | |
4116 | del_node(prev, node); | |
4117 | continue; | |
4118 | } | |
4119 | else { | |
4120 | /* failed to find a free register */ | |
4121 | if (spec & jit_class_gpr) { | |
4122 | if (!_jitc->function->regoff[regno]) | |
4123 | _jitc->function->regoff[regno] = | |
4124 | jit_allocai(sizeof(jit_word_t)); | |
4125 | #if __WORDSIZE == 32 | |
4126 | node->code = jit_code_stxi_i; | |
4127 | #else | |
4128 | node->code = jit_code_stxi_l; | |
4129 | #endif | |
4130 | } | |
4131 | else { | |
4132 | node->code = jit_code_stxi_d; | |
4133 | if (!_jitc->function->regoff[regno]) | |
4134 | _jitc->function->regoff[regno] = | |
4135 | jit_allocai(sizeof(jit_float64_t)); | |
4136 | } | |
4137 | node->u.w = _jitc->function->regoff[regno]; | |
4138 | node->v.w = JIT_FP; | |
4139 | node->w.w = regno; | |
4140 | node->link = NULL; | |
4141 | } | |
4142 | } | |
4143 | break; | |
4144 | case jit_code_load: | |
4145 | regno = jit_regno(node->u.w); | |
4146 | if (node->v.w) { | |
4147 | if (node->v.w == jit_regload_isdead) | |
4148 | jit_regset_clrbit(&_jitc->reglive, regno); | |
4149 | del_node(prev, node); | |
4150 | continue; | |
4151 | } | |
4152 | spec = jit_class(_rvs[regno].spec); | |
4153 | if (spec & jit_class_gpr) { | |
4154 | #if __WORDSIZE == 32 | |
4155 | node->code = jit_code_ldxi_i; | |
4156 | #else | |
4157 | node->code = jit_code_ldxi_l; | |
4158 | #endif | |
4159 | } | |
4160 | else | |
4161 | node->code = jit_code_ldxi_d; | |
4162 | node->v.w = regno; | |
4163 | node->v.w = JIT_FP; | |
4164 | node->w.w = _jitc->function->regoff[regno]; | |
4165 | node->link = NULL; | |
4166 | break; | |
4167 | case jit_code_prolog: | |
4168 | _jitc->function = _jitc->functions.ptr + node->w.w; | |
4169 | break; | |
4170 | case jit_code_epilog: | |
4171 | _jitc->function = NULL; | |
4172 | break; | |
4173 | default: | |
4174 | break; | |
4175 | } | |
4176 | ||
4177 | jit_regarg_clr(node, info); | |
4178 | /* update register live state */ | |
4179 | jit_reglive(node); | |
4180 | prev = node; | |
4181 | } | |
4182 | } | |
4183 | ||
4184 | static void | |
4185 | _patch_register(jit_state_t *_jit, jit_node_t *node, jit_node_t *link, | |
4186 | jit_int32_t regno, jit_int32_t patch) | |
4187 | { | |
4188 | jit_int32_t value; | |
4189 | ||
4190 | for (; node != link; node = node->next) { | |
4191 | value = jit_classify(node->code); | |
4192 | if (value & jit_cc_a0_reg) { | |
4193 | if (value & jit_cc_a0_rlh) { | |
4194 | if (node->u.q.l == regno) | |
4195 | node->u.q.l = patch; | |
4196 | if (node->u.q.h == regno) | |
4197 | node->u.q.h = patch; | |
4198 | } | |
4199 | else { | |
4200 | if (node->u.w == regno) | |
4201 | node->u.w = patch; | |
4202 | } | |
4203 | } | |
ba86ff93 PC |
4204 | if (value & jit_cc_a1_reg) { |
4205 | if (value & jit_cc_a1_rlh) { | |
4206 | if (node->v.q.l == regno) | |
4207 | node->v.q.l = patch; | |
4208 | if (node->v.q.h == regno) | |
4209 | node->v.q.h = patch; | |
4210 | } | |
4211 | else { | |
4212 | if (node->v.w == regno) | |
4213 | node->v.w = patch; | |
4214 | } | |
4215 | } | |
ba3814c1 PC |
4216 | if (value & jit_cc_a2_reg) { |
4217 | if (value & jit_cc_a2_rlh) { | |
4218 | if (node->w.q.l == regno) | |
4219 | node->w.q.l = patch; | |
4220 | if (node->w.q.h == regno) | |
4221 | node->w.q.h = patch; | |
4222 | } | |
4223 | else { | |
4224 | if (node->w.w == regno) | |
4225 | node->w.w = patch; | |
4226 | } | |
4227 | } | |
4a71579b PC |
4228 | } |
4229 | } | |
4230 | ||
40a44dcb PC |
4231 | #if __BYTE_ORDER == __LITTLE_ENDIAN |
4232 | # define htonr_us(r0,r1) bswapr_us(r0,r1) | |
4233 | # define htonr_ui(r0,r1) bswapr_ui(r0,r1) | |
4234 | # if __WORDSIZE == 64 | |
4235 | # define htonr_ul(r0,r1) bswapr_ul(r0,r1) | |
4236 | # endif | |
4237 | #else | |
4238 | # define htonr_us(r0,r1) extr_us(r0,r1) | |
4239 | # if __WORDSIZE == 32 | |
4240 | # define htonr_ui(r0,r1) movr(r0,r1) | |
4241 | # else | |
4242 | # define htonr_ui(r0,r1) extr_ui(r0,r1) | |
4243 | # define htonr_ul(r0,r1) movr(r0,r1) | |
4244 | # endif | |
4245 | #endif | |
4246 | ||
4247 | static maybe_unused void | |
4248 | generic_bswapr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1); | |
4249 | static maybe_unused void | |
4250 | generic_bswapr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1); | |
4251 | #if __WORDSIZE == 64 | |
4252 | static maybe_unused void | |
4253 | generic_bswapr_ul(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1); | |
4254 | #endif | |
4255 | ||
ba86ff93 PC |
4256 | #define depi(r0, i0, i1, i2) _depi(_jit, r0, i0, i1, i2) |
4257 | static void _depi(jit_state_t*,jit_int32_t, jit_word_t, jit_word_t, jit_word_t); | |
4258 | #define negi(r0, i0) _negi(_jit, r0, i0) | |
4259 | static void _negi(jit_state_t*, jit_int32_t, jit_word_t); | |
4260 | #define comi(r0, i0) _comi(_jit, r0, i0) | |
4261 | static void _comi(jit_state_t*, jit_int32_t, jit_word_t); | |
4262 | #define exti_c(r0, i0) _exti_c(_jit, r0, i0) | |
4263 | static void _exti_c(jit_state_t*, jit_int32_t, jit_word_t); | |
4264 | #define exti_uc(r0, i0) _exti_uc(_jit, r0, i0) | |
4265 | static void _exti_uc(jit_state_t*, jit_int32_t, jit_word_t); | |
4266 | #define exti_s(r0, i0) _exti_s(_jit, r0, i0) | |
4267 | static void _exti_s(jit_state_t*, jit_int32_t, jit_word_t); | |
4268 | #define exti_us(r0, i0) _exti_us(_jit, r0, i0) | |
4269 | static void _exti_us(jit_state_t*, jit_int32_t, jit_word_t); | |
4270 | #if __WORDSIZE == 64 | |
4271 | #define exti_i(r0, i0) _exti_i(_jit, r0, i0) | |
4272 | static void _exti_i(jit_state_t*, jit_int32_t, jit_word_t); | |
4273 | #define exti_ui(r0, i0) _exti_ui(_jit, r0, i0) | |
4274 | static void _exti_ui(jit_state_t*, jit_int32_t, jit_word_t); | |
4275 | #endif | |
4276 | #define bswapi_us(r0, i0) _bswapi_us(_jit, r0, i0) | |
4277 | static void _bswapi_us(jit_state_t*, jit_int32_t, jit_word_t); | |
4278 | #define bswapi_ui(r0, i0) _bswapi_ui(_jit, r0, i0) | |
4279 | static void _bswapi_ui(jit_state_t*, jit_int32_t, jit_word_t); | |
4280 | #if __WORDSIZE == 64 | |
4281 | # define bswapi_ul(r0, i0) _bswapi_ul(_jit, r0, i0) | |
4282 | static void _bswapi_ul(jit_state_t*, jit_int32_t, jit_word_t); | |
4283 | #endif | |
4284 | #define htoni_us(r0, i0) _htoni_us(_jit, r0, i0) | |
4285 | static void _htoni_us(jit_state_t*, jit_int32_t, jit_word_t); | |
4286 | #define htoni_ui(r0, i0) _htoni_ui(_jit, r0, i0) | |
4287 | static void _htoni_ui(jit_state_t*, jit_int32_t, jit_word_t); | |
4288 | #if __WORDSIZE == 64 | |
4289 | # define htoni_ul(r0, i0) _htoni_ul(_jit, r0, i0) | |
4290 | static void _htoni_ul(jit_state_t*, jit_int32_t, jit_word_t); | |
4291 | #endif | |
4292 | # define movi_f_w(r0, i0) _movi_f_w(_jit, r0, i0) | |
4293 | static void _movi_f_w(jit_state_t*, jit_int32_t, jit_float32_t); | |
d481fb64 | 4294 | #if __WORDSIZE == 32 && !(defined(__mips__) && NEW_ABI) |
ba86ff93 PC |
4295 | # define movi_d_ww(r0, r1, i0) _movi_d_ww(_jit, r0, r1, i0) |
4296 | static void _movi_d_ww(jit_state_t*, jit_int32_t, jit_int32_t, jit_float64_t); | |
4297 | #else | |
4298 | # define movi_d_w(r0, i0) _movi_d_w(_jit, r0, i0) | |
4299 | static void _movi_d_w(jit_state_t*, jit_int32_t, jit_float64_t); | |
4300 | #endif | |
4301 | #define cloi(r0, i0) _cloi(_jit, r0, i0) | |
4302 | static void _cloi(jit_state_t*, jit_int32_t, jit_word_t); | |
4303 | #define clzi(r0, i0) _clzi(_jit, r0, i0) | |
4304 | static void _clzi(jit_state_t*, jit_int32_t, jit_word_t); | |
4305 | #define ctoi(r0, i0) _ctoi(_jit, r0, i0) | |
4306 | static void _ctoi(jit_state_t*, jit_int32_t, jit_word_t); | |
4307 | #define ctzi(r0, i0) _ctzi(_jit, r0, i0) | |
4308 | static void _ctzi(jit_state_t*, jit_int32_t, jit_word_t); | |
4309 | #define rbiti(r0, i0) _rbiti(_jit, r0, i0) | |
4310 | static void _rbiti(jit_state_t*, jit_int32_t, jit_word_t); | |
4311 | #define popcnti(r0, i0) _popcnti(_jit, r0, i0) | |
4312 | static void _popcnti(jit_state_t*, jit_int32_t, jit_word_t); | |
4313 | #define exti(r0, i0, i1, i2) _exti(_jit, r0, i0, i1, i2) | |
4314 | static void _exti(jit_state_t*,jit_int32_t, jit_word_t, jit_word_t, jit_word_t); | |
4315 | #define exti_u(r0, i0, i1, i2) _exti_u(_jit, r0, i0, i1, i2) | |
4316 | static void _exti_u(jit_state_t*, | |
4317 | jit_int32_t, jit_word_t, jit_word_t, jit_word_t); | |
4318 | #define generic_unldr(r0, r1, i0) _generic_unldr(_jit, r0, r1, i0) | |
4319 | static void _generic_unldr(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t); | |
4320 | #define generic_unldi(r0, i0, i1) _generic_unldi(_jit, r0, i0, i1) | |
4321 | static void _generic_unldi(jit_state_t*, jit_int32_t, jit_word_t, jit_word_t); | |
4322 | #define generic_unldr_u(r0, r1, i0) _generic_unldr_u(_jit, r0, r1, i0) | |
4323 | static void _generic_unldr_u(jit_state_t*, | |
4324 | jit_int32_t, jit_int32_t, jit_word_t); | |
4325 | #define generic_unldi_u(r0, i0, i1) _generic_unldi_u(_jit, r0, i0, i1) | |
4326 | static void _generic_unldi_u(jit_state_t*, jit_int32_t, jit_word_t, jit_word_t); | |
4327 | #define generic_unstr(r0, r1, i0) _generic_unstr(_jit, r0, r1, i0) | |
4328 | static void _generic_unstr(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t); | |
4329 | #define generic_unsti(i0, r0, i1) _generic_unsti(_jit, i0, r0, i1) | |
4330 | static void _generic_unsti(jit_state_t*, jit_word_t, jit_int32_t, jit_word_t); | |
4331 | #if !defined(__i386__) && !defined(__x86_64__) | |
4332 | # define generic_unldr_x(r0, r1, i0) _generic_unldr_x(_jit, r0, r1, i0) | |
4333 | static void _generic_unldr_x(jit_state_t*, | |
4334 | jit_int32_t, jit_int32_t, jit_word_t); | |
4335 | # define generic_unldi_x(r0, i0, i1) _generic_unldi_x(_jit, r0, i0, i1) | |
4336 | static void _generic_unldi_x(jit_state_t*, jit_int32_t, jit_word_t, jit_word_t); | |
4337 | # define generic_unstr_x(r0, r1, i0) _generic_unstr_x(_jit, r0, r1, i0) | |
4338 | static void _generic_unstr_x(jit_state_t*, | |
4339 | jit_int32_t, jit_int32_t, jit_word_t); | |
4340 | # define generic_unsti_x(i0, r0, i1) _generic_unsti_x(_jit, i0, r0, i1) | |
4341 | static void _generic_unsti_x(jit_state_t*, jit_word_t, jit_int32_t, jit_word_t); | |
4342 | #endif | |
79bfeef6 PC |
4343 | #define patch_alist(revert) _patch_alist(_jit, revert) |
4344 | static maybe_unused void _patch_alist(jit_state_t *_jit, jit_bool_t revert); | |
4345 | ||
4a71579b PC |
4346 | #if defined(__i386__) || defined(__x86_64__) |
4347 | # include "jit_x86.c" | |
4348 | #elif defined(__mips__) | |
4349 | # include "jit_mips.c" | |
4350 | #elif defined(__arm__) | |
4351 | # include "jit_arm.c" | |
4352 | #elif defined(__powerpc__) | |
4353 | # include "jit_ppc.c" | |
4354 | #elif defined(__sparc__) | |
4355 | # include "jit_sparc.c" | |
4356 | #elif defined(__ia64__) | |
4357 | # include "jit_ia64.c" | |
4358 | #elif defined(__hppa__) | |
4359 | # include "jit_hppa.c" | |
4360 | #elif defined(__aarch64__) | |
4361 | # include "jit_aarch64.c" | |
4362 | #elif defined(__s390__) || defined(__s390x__) | |
4363 | # include "jit_s390.c" | |
4364 | #elif defined(__alpha__) | |
4365 | # include "jit_alpha.c" | |
4366 | #elif defined(__riscv) | |
4367 | # include "jit_riscv.c" | |
24d91c0d PC |
4368 | #elif defined(__loongarch__) |
4369 | # include "jit_loongarch.c" | |
4a71579b | 4370 | #endif |
40a44dcb PC |
4371 | |
4372 | static maybe_unused void | |
4373 | generic_bswapr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1) | |
4374 | { | |
4375 | jit_int32_t reg = jit_get_reg(jit_class_gpr); | |
4376 | ||
4377 | rshi(rn(reg), r1, 8); | |
4378 | andi(r0, r1, 0xff); | |
4379 | andi(rn(reg), rn(reg), 0xff); | |
4380 | lshi(r0, r0, 8); | |
4381 | orr(r0, r0, rn(reg)); | |
4382 | ||
4383 | jit_unget_reg(reg); | |
4384 | } | |
4385 | ||
4386 | static maybe_unused void | |
4387 | generic_bswapr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1) | |
4388 | { | |
4389 | jit_int32_t reg = jit_get_reg(jit_class_gpr); | |
4390 | ||
4391 | rshi(rn(reg), r1, 16); | |
4392 | bswapr_us(r0, r1); | |
4393 | bswapr_us(rn(reg), rn(reg)); | |
4394 | lshi(r0, r0, 16); | |
4395 | orr(r0, r0, rn(reg)); | |
4396 | ||
4397 | jit_unget_reg(reg); | |
4398 | } | |
4399 | ||
4400 | #if __WORDSIZE == 64 | |
4401 | static maybe_unused void | |
4402 | generic_bswapr_ul(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1) | |
4403 | { | |
4404 | jit_int32_t reg = jit_get_reg(jit_class_gpr); | |
4405 | ||
4406 | rshi_u(rn(reg), r1, 32); | |
4407 | bswapr_ui(r0, r1); | |
4408 | bswapr_ui(rn(reg), rn(reg)); | |
4409 | lshi(r0, r0, 32); | |
4410 | orr(r0, r0, rn(reg)); | |
4411 | ||
4412 | jit_unget_reg(reg); | |
4413 | } | |
4414 | #endif | |
79bfeef6 | 4415 | |
ba86ff93 PC |
4416 | static void |
4417 | _depi(jit_state_t *_jit, | |
4418 | jit_int32_t r0, jit_word_t i0, jit_word_t i1, jit_word_t i2) | |
4419 | { | |
4420 | jit_int32_t reg; | |
4421 | reg = jit_get_reg(jit_class_gpr); | |
4422 | movi(rn(reg), i0); | |
4423 | depr(r0, rn(reg), i1, i2); | |
4424 | jit_unget_reg(reg); | |
4425 | } | |
4426 | ||
4427 | static void | |
4428 | _negi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4429 | { | |
4430 | movi(r0, -i0); | |
4431 | } | |
4432 | ||
4433 | static void | |
4434 | _comi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4435 | { | |
4436 | movi(r0, ~i0); | |
4437 | } | |
4438 | ||
4439 | static void | |
4440 | _exti_c(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4441 | { | |
4442 | movi(r0, (jit_int8_t)i0); | |
4443 | } | |
4444 | ||
4445 | static void | |
4446 | _exti_uc(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4447 | { | |
4448 | movi(r0, (jit_uint8_t)i0); | |
4449 | } | |
4450 | ||
4451 | static void | |
4452 | _exti_s(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4453 | { | |
4454 | movi(r0, (jit_int16_t)i0); | |
4455 | } | |
4456 | ||
4457 | static void | |
4458 | _exti_us(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4459 | { | |
4460 | movi(r0, (jit_uint16_t)i0); | |
4461 | } | |
4462 | ||
4463 | #if __WORDSIZE == 64 | |
4464 | static void | |
4465 | _exti_i(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4466 | { | |
4467 | movi(r0, (jit_int32_t)i0); | |
4468 | } | |
4469 | ||
4470 | static void | |
4471 | _exti_ui(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4472 | { | |
4473 | movi(r0, (jit_uint32_t)i0); | |
4474 | } | |
4475 | #endif | |
4476 | ||
4477 | static void | |
4478 | _bswapi_us(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4479 | { | |
4480 | union { | |
4481 | jit_uint16_t us; | |
4482 | jit_uint8_t v[2]; | |
4483 | } l, h; | |
4484 | l.us = i0; | |
4485 | h.v[0] = l.v[1]; | |
4486 | h.v[1] = l.v[0]; | |
4487 | movi(r0, h.us); | |
4488 | } | |
4489 | ||
4490 | static void | |
4491 | _bswapi_ui(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4492 | { | |
4493 | union { | |
4494 | jit_uint32_t ui; | |
4495 | jit_uint8_t v[4]; | |
4496 | } l, h; | |
4497 | l.ui = i0; | |
4498 | h.v[0] = l.v[3]; | |
4499 | h.v[1] = l.v[2]; | |
4500 | h.v[2] = l.v[1]; | |
4501 | h.v[3] = l.v[0]; | |
4502 | movi(r0, h.ui); | |
4503 | } | |
4504 | ||
4505 | #if __WORDSIZE == 64 | |
4506 | static void | |
4507 | _bswapi_ul(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4508 | { | |
4509 | union { | |
4510 | jit_uint64_t ul; | |
4511 | jit_uint8_t v[8]; | |
4512 | } l, h; | |
4513 | l.ul = i0; | |
4514 | h.v[0] = l.v[7]; | |
4515 | h.v[1] = l.v[6]; | |
4516 | h.v[2] = l.v[5]; | |
4517 | h.v[3] = l.v[4]; | |
4518 | h.v[4] = l.v[3]; | |
4519 | h.v[5] = l.v[2]; | |
4520 | h.v[6] = l.v[1]; | |
4521 | h.v[7] = l.v[0]; | |
4522 | movi(r0, h.ul); | |
4523 | } | |
4524 | #endif | |
4525 | ||
4526 | static void | |
4527 | _htoni_us(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4528 | { | |
4529 | #if __BYTE_ORDER == __LITTLE_ENDIAN | |
4530 | bswapi_us(r0, i0); | |
4531 | #else | |
4532 | exti_us(r0, i0); | |
4533 | #endif | |
4534 | } | |
4535 | ||
4536 | static void | |
4537 | _htoni_ui(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4538 | { | |
4539 | #if __BYTE_ORDER == __LITTLE_ENDIAN | |
4540 | bswapi_ui(r0, i0); | |
4541 | #else | |
4542 | # if __WORDSIZE == 32 | |
4543 | movi(r0, i0); | |
4544 | # else | |
4545 | exti_ui(r0, i0); | |
4546 | # endif | |
4547 | #endif | |
4548 | } | |
4549 | ||
4550 | #if __WORDSIZE == 64 | |
4551 | static void | |
4552 | _htoni_ul(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4553 | { | |
4554 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
4555 | bswapi_ul(r0, i0); | |
4556 | # else | |
4557 | movi(r0, i0); | |
4558 | # endif | |
4559 | } | |
4560 | #endif | |
4561 | ||
4562 | static void | |
4563 | _movi_f_w(jit_state_t *_jit, jit_int32_t r0, jit_float32_t i0) | |
4564 | { | |
4565 | union { | |
4566 | jit_int32_t i; | |
4567 | jit_float32_t f; | |
4568 | } data; | |
4569 | data.f = i0; | |
4570 | #if defined(__ia64__) | |
4571 | /* Should be used only in this case (with out0 == 120) */ | |
4572 | if (r0 >= 120) | |
4573 | r0 = _jitc->rout + (r0 - 120); | |
4574 | #endif | |
4575 | movi(r0, data.i); | |
4576 | } | |
4577 | ||
d481fb64 | 4578 | #if __WORDSIZE == 32 && !(defined(__mips__) && NEW_ABI) |
ba86ff93 PC |
4579 | static void |
4580 | _movi_d_ww(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_float64_t i0) | |
4581 | { | |
4582 | union { | |
4583 | jit_int32_t i[2]; | |
4584 | jit_float64_t d; | |
4585 | } data; | |
4586 | data.d = i0; | |
4587 | /* Mips does not change byte order of double values */ | |
4588 | # if __BYTE_ORDER == __LITTLE_ENDIAN || defined(__mips__) | |
4589 | movi(r0, data.i[0]); | |
4590 | movi(r1, data.i[1]); | |
4591 | # else | |
4592 | movi(r1, data.i[0]); | |
4593 | movi(r0, data.i[1]); | |
4594 | # endif | |
4595 | } | |
4596 | ||
4597 | #else | |
4598 | static void | |
4599 | _movi_d_w(jit_state_t *_jit, jit_int32_t r0, jit_float64_t i0) | |
4600 | { | |
4601 | union { | |
4602 | jit_int64_t l; | |
4603 | jit_float64_t d; | |
4604 | } data; | |
4605 | data.d = i0; | |
4606 | # if defined(__ia64__) | |
4607 | /* Should be used only in this case (with out0 == 120) */ | |
4608 | if (r0 >= 120) | |
4609 | r0 = _jitc->rout + (r0 - 120); | |
4610 | # endif | |
4611 | movi(r0, data.l); | |
4612 | } | |
4613 | #endif | |
4614 | ||
4615 | void | |
4616 | _jit_negi_f(jit_state_t *_jit, jit_fpr_t u, jit_float32_t v) | |
4617 | { | |
4618 | jit_inc_synth_wf(negi_f, u, v); | |
4619 | jit_movi_f(u, v); | |
4620 | jit_negr_f(u, u); | |
4621 | jit_dec_synth(); | |
4622 | } | |
4623 | ||
4624 | void | |
4625 | _jit_absi_f(jit_state_t *_jit, jit_fpr_t u, jit_float32_t v) | |
4626 | { | |
4627 | jit_inc_synth_wf(absi_f, u, v); | |
4628 | jit_movi_f(u, v); | |
4629 | jit_absr_f(u, u); | |
4630 | jit_dec_synth(); | |
4631 | } | |
4632 | ||
4633 | void | |
4634 | _jit_sqrti_f(jit_state_t *_jit, jit_fpr_t u, jit_float32_t v) | |
4635 | { | |
4636 | jit_inc_synth_wf(sqrti_f, u, v); | |
4637 | jit_movi_f(u, v); | |
4638 | jit_sqrtr_f(u, u); | |
4639 | jit_dec_synth(); | |
4640 | } | |
4641 | ||
4642 | void | |
4643 | _jit_fmai_f(jit_state_t *_jit, | |
4644 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float32_t x) | |
4645 | { | |
4646 | jit_int32_t y; | |
4647 | jit_inc_synth_wqf(fmai_f, u, v, w, x); | |
4648 | if (u != v && u != w) { | |
4649 | jit_movi_f(u, x); | |
4650 | jit_fmar_f(u, v, w, u); | |
4651 | } | |
4652 | else { | |
4653 | y = jit_get_reg(jit_class_fpr); | |
4654 | jit_movi_f(y, x); | |
4655 | jit_fmar_f(u, v, w, y); | |
4656 | jit_unget_reg(y); | |
4657 | } | |
4658 | jit_dec_synth(); | |
4659 | } | |
4660 | ||
4661 | void | |
4662 | _jit_fmsi_f(jit_state_t *_jit, | |
4663 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float32_t x) | |
4664 | { | |
4665 | jit_int32_t y; | |
4666 | jit_inc_synth_wqf(fmai_f, u, v, w, x); | |
4667 | if (u != v && u != w) { | |
4668 | jit_movi_f(u, x); | |
4669 | jit_fmsr_f(u, v, w, u); | |
4670 | } | |
4671 | else { | |
4672 | y = jit_get_reg(jit_class_fpr); | |
4673 | jit_movi_f(y, x); | |
4674 | jit_fmsr_f(u, v, w, y); | |
4675 | jit_unget_reg(y); | |
4676 | } | |
4677 | jit_dec_synth(); | |
4678 | } | |
4679 | ||
4680 | void | |
4681 | _jit_fnmai_f(jit_state_t *_jit, | |
4682 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float32_t x) | |
4683 | { | |
4684 | jit_int32_t y; | |
4685 | jit_inc_synth_wqf(fmai_f, u, v, w, x); | |
4686 | if (u != v && u != w) { | |
4687 | jit_movi_f(u, x); | |
4688 | jit_fnmar_f(u, v, w, u); | |
4689 | } | |
4690 | else { | |
4691 | y = jit_get_reg(jit_class_fpr); | |
4692 | jit_movi_f(y, x); | |
4693 | jit_fnmar_f(u, v, w, y); | |
4694 | jit_unget_reg(y); | |
4695 | } | |
4696 | jit_dec_synth(); | |
4697 | } | |
4698 | ||
4699 | void | |
4700 | _jit_fnmsi_f(jit_state_t *_jit, | |
4701 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float32_t x) | |
4702 | { | |
4703 | jit_int32_t y; | |
4704 | jit_inc_synth_wqf(fmai_f, u, v, w, x); | |
4705 | if (u != v && u != w) { | |
4706 | jit_movi_f(u, x); | |
4707 | jit_fnmsr_f(u, v, w, u); | |
4708 | } | |
4709 | else { | |
4710 | y = jit_get_reg(jit_class_fpr); | |
4711 | jit_movi_f(y, x); | |
4712 | jit_fnmsr_f(u, v, w, y); | |
4713 | jit_unget_reg(y); | |
4714 | } | |
4715 | jit_dec_synth(); | |
4716 | } | |
4717 | ||
4718 | void | |
4719 | _jit_negi_d(jit_state_t *_jit, jit_fpr_t u, jit_float64_t v) | |
4720 | { | |
4721 | jit_inc_synth_wd(negi_d, u, v); | |
4722 | jit_movi_d(u, v); | |
4723 | jit_negr_d(u, u); | |
4724 | jit_dec_synth(); | |
4725 | } | |
4726 | ||
4727 | void | |
4728 | _jit_absi_d(jit_state_t *_jit, jit_fpr_t u, jit_float64_t v) | |
4729 | { | |
4730 | jit_inc_synth_wd(absi_d, u, v); | |
4731 | jit_movi_d(u, v); | |
4732 | jit_absr_d(u, u); | |
4733 | jit_dec_synth(); | |
4734 | } | |
4735 | ||
4736 | void | |
4737 | _jit_sqrti_d(jit_state_t *_jit, jit_fpr_t u, jit_float64_t v) | |
4738 | { | |
4739 | jit_inc_synth_wd(sqrti_d, u, v); | |
4740 | jit_movi_d(u, v); | |
4741 | jit_sqrtr_d(u, u); | |
4742 | jit_dec_synth(); | |
4743 | } | |
4744 | ||
4745 | void | |
4746 | _jit_fmai_d(jit_state_t *_jit, | |
4747 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float64_t x) | |
4748 | { | |
4749 | jit_int32_t y; | |
4750 | jit_inc_synth_wqd(fmai_d, u, v, w, x); | |
4751 | if (u != v && u != w) { | |
4752 | jit_movi_d(u, x); | |
4753 | jit_fmar_d(u, v, w, u); | |
4754 | } | |
4755 | else { | |
4756 | y = jit_get_reg(jit_class_fpr); | |
4757 | jit_movi_d(y, x); | |
4758 | jit_fmar_d(u, v, w, y); | |
4759 | jit_unget_reg(y); | |
4760 | } | |
4761 | jit_dec_synth(); | |
4762 | } | |
4763 | ||
4764 | void | |
4765 | _jit_fmsi_d(jit_state_t *_jit, | |
4766 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float64_t x) | |
4767 | { | |
4768 | jit_int32_t y; | |
4769 | jit_inc_synth_wqd(fmai_d, u, v, w, x); | |
4770 | if (u != v && u != w) { | |
4771 | jit_movi_d(u, x); | |
4772 | jit_fmsr_d(u, v, w, u); | |
4773 | } | |
4774 | else { | |
4775 | y = jit_get_reg(jit_class_fpr); | |
4776 | jit_movi_d(y, x); | |
4777 | jit_fmsr_d(u, v, w, y); | |
4778 | jit_unget_reg(y); | |
4779 | } | |
4780 | jit_dec_synth(); | |
4781 | } | |
4782 | ||
4783 | void | |
4784 | _jit_fnmai_d(jit_state_t *_jit, | |
4785 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float64_t x) | |
4786 | { | |
4787 | jit_int32_t y; | |
4788 | jit_inc_synth_wqd(fmai_d, u, v, w, x); | |
4789 | if (u != v && u != w) { | |
4790 | jit_movi_d(u, x); | |
4791 | jit_fnmar_d(u, v, w, u); | |
4792 | } | |
4793 | else { | |
4794 | y = jit_get_reg(jit_class_fpr); | |
4795 | jit_movi_d(y, x); | |
4796 | jit_fnmar_d(u, v, w, y); | |
4797 | jit_unget_reg(y); | |
4798 | } | |
4799 | jit_dec_synth(); | |
4800 | } | |
4801 | ||
4802 | void | |
4803 | _jit_fnmsi_d(jit_state_t *_jit, | |
4804 | jit_fpr_t u, jit_fpr_t v, jit_fpr_t w, jit_float64_t x) | |
4805 | { | |
4806 | jit_int32_t y; | |
4807 | jit_inc_synth_wqd(fmai_d, u, v, w, x); | |
4808 | if (u != v && u != w) { | |
4809 | jit_movi_d(u, x); | |
4810 | jit_fnmsr_d(u, v, w, u); | |
4811 | } | |
4812 | else { | |
4813 | y = jit_get_reg(jit_class_fpr); | |
4814 | jit_movi_d(y, x); | |
4815 | jit_fnmsr_d(u, v, w, y); | |
4816 | jit_unget_reg(y); | |
4817 | } | |
4818 | jit_dec_synth(); | |
4819 | } | |
4820 | ||
4821 | static void | |
4822 | _cloi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4823 | { | |
4824 | clzi(r0, ~i0); | |
4825 | } | |
4826 | ||
4827 | static void | |
4828 | _clzi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4829 | { | |
4830 | #if __WORDSIZE == 64 && _WIN32 | |
4831 | movi(r0, (i0) ? __builtin_clzll(i0) : __WORDSIZE); | |
4832 | #else | |
4833 | movi(r0, (i0) ? __builtin_clzl(i0) : __WORDSIZE); | |
4834 | #endif | |
4835 | } | |
4836 | ||
4837 | static void | |
4838 | _ctoi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4839 | { | |
4840 | ctzi(r0, ~i0); | |
4841 | } | |
4842 | ||
4843 | static void | |
4844 | _ctzi(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4845 | { | |
4846 | #if __WORDSIZE == 64 && _WIN32 | |
4847 | movi(r0, (i0) ? __builtin_ctzll(i0) : __WORDSIZE); | |
4848 | #else | |
4849 | movi(r0, (i0) ? __builtin_ctzl(i0) : __WORDSIZE); | |
4850 | #endif | |
4851 | } | |
4852 | ||
4853 | static void | |
4854 | _rbiti(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4855 | { | |
4856 | jit_int32_t i; | |
4857 | union { | |
4858 | jit_uword_t w; | |
4859 | jit_uint8_t v[__WORDSIZE >> 3]; | |
4860 | } u, v; | |
4861 | static const unsigned char swap_tab[256] = { | |
4862 | 0, 128, 64, 192, 32, 160, 96, 224, | |
4863 | 16, 144, 80, 208, 48, 176, 112, 240, | |
4864 | 8, 136, 72, 200, 40, 168, 104, 232, | |
4865 | 24, 152, 88, 216 ,56, 184, 120, 248, | |
4866 | 4, 132, 68, 196, 36, 164, 100, 228, | |
4867 | 20, 148, 84, 212, 52, 180, 116, 244, | |
4868 | 12, 140, 76, 204, 44, 172, 108, 236, | |
4869 | 28, 156, 92, 220, 60, 188, 124, 252, | |
4870 | 2, 130, 66, 194, 34, 162, 98, 226, | |
4871 | 18, 146, 82, 210, 50, 178, 114, 242, | |
4872 | 10, 138, 74, 202, 42, 170, 106, 234, | |
4873 | 26, 154, 90, 218, 58, 186, 122, 250, | |
4874 | 6, 134, 70, 198, 38, 166, 102, 230, | |
4875 | 22, 150, 86, 214, 54, 182, 118, 246, | |
4876 | 14, 142, 78, 206, 46, 174, 110, 238, | |
4877 | 30, 158, 94, 222, 62, 190, 126, 254, | |
4878 | 1, 129, 65, 193, 33, 161, 97, 225, | |
4879 | 17, 145, 81, 209, 49, 177, 113, 241, | |
4880 | 9, 137, 73, 201, 41, 169, 105, 233, | |
4881 | 25, 153, 89, 217, 57, 185, 121, 249, | |
4882 | 5, 133, 69, 197, 37, 165, 101, 229, | |
4883 | 21, 149, 85, 213, 53, 181, 117, 245, | |
4884 | 13, 141, 77, 205, 45, 173, 109, 237, | |
4885 | 29, 157, 93, 221, 61, 189, 125, 253, | |
4886 | 3, 131, 67, 195, 35, 163, 99, 227, | |
4887 | 19, 147, 83, 211, 51, 179, 115, 243, | |
4888 | 11, 139, 75, 203, 43, 171, 107, 235, | |
4889 | 27, 155, 91, 219, 59, 187, 123, 251, | |
4890 | 7, 135, 71, 199, 39, 167, 103, 231, | |
4891 | 23, 151, 87, 215, 55, 183, 119, 247, | |
4892 | 15, 143, 79, 207, 47, 175, 111, 239, | |
4893 | 31, 159, 95, 223, 63, 191, 127, 255 | |
4894 | }; | |
4895 | u.w = i0; | |
4896 | for (i = 0; i < sizeof(jit_word_t); ++i) | |
4897 | v.v[i] = swap_tab[u.v[sizeof(jit_word_t) - i - 1]]; | |
4898 | movi(r0, v.w); | |
4899 | } | |
4900 | ||
4901 | static void | |
4902 | _popcnti(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0) | |
4903 | { | |
4904 | #if __WORDSIZE == 64 && _WIN32 | |
4905 | movi(r0, (i0) ? __builtin_popcountll(i0) : __WORDSIZE); | |
4906 | #else | |
4907 | movi(r0, (i0) ? __builtin_popcountl(i0) : __WORDSIZE); | |
4908 | #endif | |
4909 | } | |
4910 | ||
4911 | static void _exti(jit_state_t *_jit, | |
4912 | jit_int32_t r0, jit_word_t i0, jit_word_t i1, jit_word_t i2) | |
4913 | { | |
4914 | #if __BYTE_ORDER == __BIG_ENDIAN | |
4915 | i1 = __WORDSIZE - (i1 + i2); | |
4916 | #endif | |
4917 | i0 <<= __WORDSIZE - (i1 + i2); | |
4918 | i0 >>= __WORDSIZE - i2; | |
4919 | movi(r0, i0); | |
4920 | } | |
4921 | ||
4922 | static void _exti_u(jit_state_t *_jit, | |
4923 | jit_int32_t r0, jit_word_t i0, jit_word_t i1, jit_word_t i2) | |
4924 | { | |
4925 | jit_word_t t; | |
4926 | #if __BYTE_ORDER == __BIG_ENDIAN | |
4927 | i1 = __WORDSIZE - (i1 + i2); | |
4928 | #endif | |
4929 | if (i1) | |
4930 | i0 >>= __WORDSIZE - i2; | |
4931 | #if __WORDSIZE == 64 && !_WIN32 | |
4932 | i0 &= (1L << i2) - 1; | |
4933 | #else | |
4934 | i0 &= (1LL << i2) - 1; | |
4935 | #endif | |
4936 | movi(r0, i0); | |
4937 | } | |
4938 | ||
4939 | static void | |
4940 | _generic_unldr(jit_state_t *_jit, | |
4941 | jit_int32_t r0, jit_int32_t r1, jit_word_t i0) | |
4942 | { | |
4943 | jit_int32_t t0, r2; | |
4944 | assert(i0 >= 1 && i0 <= sizeof(jit_word_t)); | |
4945 | if (i0 & (i0 - 1)) { | |
4946 | t0 = jit_get_reg(jit_class_gpr); | |
4947 | r2 = rn(t0); | |
4948 | movr(r2, r1); | |
4949 | } | |
4950 | switch (i0) { | |
4951 | case 1: | |
4952 | ldr_c(r0, r1); | |
4953 | break; | |
4954 | case 2: | |
4955 | ldr_s(r0, r1); | |
4956 | break; | |
4957 | case 3: | |
4958 | #if __BYTE_ORDER == __LITTLE_ENDIAN | |
4959 | ldr_us(r0, r2); | |
4960 | ldxi_c(r2, r2, 2); | |
4961 | lshi(r2, r2, 16); | |
4962 | #else | |
4963 | ldr_c(r0, r2); | |
4964 | lshi(r0, r0, 16); | |
4965 | ldxi_us(r2, r2, 1); | |
4966 | #endif | |
4967 | break; | |
4968 | #if __WORDSIZE == 32 | |
4969 | default: | |
4970 | ldr_i(r0, r1); | |
4971 | break; | |
4972 | #else | |
4973 | case 4: | |
4974 | ldr_i(r0, r1); | |
4975 | break; | |
4976 | case 5: | |
4977 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
4978 | ldr_ui(r0, r2); | |
4979 | ldxi_c(r2, r2, 4); | |
4980 | lshi(r2, r2, 32); | |
4981 | # else | |
4982 | ldr_i(r0, r2); | |
4983 | lshi(r0, r0, 8); | |
4984 | ldxi_uc(r2, r2, 4); | |
4985 | # endif | |
4986 | break; | |
4987 | case 6: | |
4988 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
4989 | ldr_ui(r0, r2); | |
4990 | ldxi_s(r2, r2, 4); | |
4991 | lshi(r2, r2, 32); | |
4992 | # else | |
4993 | ldr_i(r0, r2); | |
4994 | lshi(r0, r0, 16); | |
4995 | ldxi_us(r2, r2, 4); | |
4996 | # endif | |
4997 | break; | |
4998 | case 7: | |
4999 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5000 | ldr_ui(r0, r2); | |
5001 | ldxi_i(r2, r2, 4); | |
5002 | lshi(r2, r2, 40); | |
5003 | rshi(r2, r2, 8); | |
5004 | # else | |
5005 | ldr_i(r0, r2); | |
5006 | lshi(r0, r0, 24); | |
5007 | ldxi_ui(r2, r2, 4); | |
5008 | rshi(r2, r2, 8); | |
5009 | # endif | |
5010 | break; | |
5011 | default: | |
5012 | ldr_l(r0, r1); | |
5013 | break; | |
5014 | #endif | |
5015 | } | |
5016 | if (i0 & (i0 - 1)) { | |
5017 | orr(r0, r0, r2); | |
5018 | jit_unget_reg(t0); | |
5019 | } | |
5020 | } | |
5021 | ||
5022 | static void | |
5023 | _generic_unldi(jit_state_t *_jit, | |
5024 | jit_int32_t r0, jit_word_t i0, jit_word_t i1) | |
5025 | { | |
5026 | jit_int32_t t0, r2; | |
5027 | assert(i1 >= 1 && i1 <= sizeof(jit_word_t)); | |
5028 | if (i1 & (i1 - 1)) { | |
5029 | t0 = jit_get_reg(jit_class_gpr); | |
5030 | r2 = rn(t0); | |
5031 | } | |
5032 | switch (i1) { | |
5033 | case 1: | |
5034 | ldi_c(r0, i0); | |
5035 | break; | |
5036 | case 2: | |
5037 | ldi_s(r0, i0); | |
5038 | break; | |
5039 | case 3: | |
5040 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5041 | ldi_us(r0, i0); | |
5042 | ldi_c(r2, i0 + 2); | |
5043 | lshi(r2, r2, 16); | |
5044 | # else | |
5045 | ldi_c(r0, i0); | |
5046 | lshi(r0, r0, 16); | |
5047 | ldi_us(r2, i0 + 1); | |
5048 | # endif | |
5049 | break; | |
5050 | # if __WORDSIZE == 32 | |
5051 | default: | |
5052 | ldi_i(r0, i0); | |
5053 | break; | |
5054 | # else | |
5055 | case 4: | |
5056 | ldi_i(r0, i0); | |
5057 | break; | |
5058 | case 5: | |
5059 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5060 | ldi_ui(r0, i0); | |
5061 | ldi_c(r2, i0 + 4); | |
5062 | lshi(r2, r2, 32); | |
5063 | # else | |
5064 | ldi_i(r0, i0); | |
5065 | lshi(r0, r0, 8); | |
5066 | ldi_uc(r2, i0 + 4); | |
5067 | # endif | |
5068 | break; | |
5069 | case 6: | |
5070 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5071 | ldi_ui(r0, i0); | |
5072 | ldi_s(r2, i0 + 4); | |
5073 | lshi(r2, r2, 32); | |
5074 | # else | |
5075 | ldi_i(r0, i0); | |
5076 | lshi(r0, r0, 16); | |
5077 | ldi_us(r2, i0 + 4); | |
5078 | # endif | |
5079 | break; | |
5080 | case 7: | |
5081 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5082 | ldi_ui(r0, i0); | |
5083 | ldi_i(r2, i0 + 4); | |
5084 | lshi(r2, r2, 40); | |
5085 | rshi(r2, r2, 8); | |
5086 | # else | |
5087 | ldi_i(r0, i0); | |
5088 | lshi(r0, r0, 24); | |
5089 | ldi_ui(r2, i0 + 4); | |
5090 | rshi(r2, r2, 8); | |
5091 | # endif | |
5092 | break; | |
5093 | default: | |
5094 | ldi_l(r0, i0); | |
5095 | break; | |
5096 | # endif | |
5097 | } | |
5098 | if (i1 & (i1 - 1)) { | |
5099 | orr(r0, r0, r2); | |
5100 | jit_unget_reg(t0); | |
5101 | } | |
5102 | } | |
5103 | ||
5104 | static void | |
5105 | _generic_unldr_u(jit_state_t *_jit, | |
5106 | jit_int32_t r0, jit_int32_t r1, jit_word_t i0) | |
5107 | { | |
5108 | jit_int32_t t0, r2; | |
5109 | assert(i0 >= 1 && i0 <= sizeof(jit_word_t)); | |
5110 | if (i0 & (i0 - 1)) { | |
5111 | t0 = jit_get_reg(jit_class_gpr); | |
5112 | r2 = rn(t0); | |
5113 | movr(r2, r1); | |
5114 | } | |
5115 | switch (i0) { | |
5116 | case 1: | |
5117 | ldr_uc(r0, r1); | |
5118 | break; | |
5119 | case 2: | |
5120 | ldr_us(r0, r1); | |
5121 | break; | |
5122 | case 3: | |
5123 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5124 | ldr_us(r0, r2); | |
5125 | ldxi_uc(r2, r2, 2); | |
5126 | lshi(r2, r2, 16); | |
5127 | # else | |
5128 | ldr_uc(r0, r2); | |
5129 | lshi(r0, r0, 16); | |
5130 | ldxi_us(r2, r2, 1); | |
5131 | # endif | |
5132 | break; | |
5133 | # if __WORDSIZE == 32 | |
5134 | default: | |
5135 | ldr_i(r0, r1); | |
5136 | break; | |
5137 | # else | |
5138 | case 4: | |
5139 | ldr_ui(r0, r1); | |
5140 | break; | |
5141 | case 5: | |
5142 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5143 | ldr_ui(r0, r2); | |
5144 | ldxi_uc(r2, r2, 4); | |
5145 | lshi(r2, r2, 32); | |
5146 | # else | |
5147 | ldr_ui(r0, r2); | |
5148 | lshi(r0, r0, 8); | |
5149 | ldxi_uc(r2, r2, 4); | |
5150 | # endif | |
5151 | break; | |
5152 | case 6: | |
5153 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5154 | ldr_ui(r0, r2); | |
5155 | ldxi_us(r2, r2, 4); | |
5156 | lshi(r2, r2, 32); | |
5157 | # else | |
5158 | ldr_ui(r0, r2); | |
5159 | lshi(r0, r0, 16); | |
5160 | ldxi_us(r2, r2, 4); | |
5161 | # endif | |
5162 | break; | |
5163 | case 7: | |
5164 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5165 | ldr_ui(r0, r2); | |
5166 | ldxi_ui(r2, r2, 4); | |
5167 | lshi(r2, r2, 40); | |
5168 | rshi_u(r2, r2, 8); | |
5169 | # else | |
5170 | ldr_ui(r0, r2); | |
5171 | lshi(r0, r0, 24); | |
5172 | ldxi_ui(r2, r2, 4); | |
5173 | rshi(r2, r2, 8); | |
5174 | # endif | |
5175 | break; | |
5176 | default: | |
5177 | ldr_l(r0, r1); | |
5178 | break; | |
5179 | # endif | |
5180 | } | |
5181 | if (i0 & (i0 - 1)) { | |
5182 | orr(r0, r0, r2); | |
5183 | jit_unget_reg(t0); | |
5184 | } | |
5185 | } | |
5186 | ||
5187 | static void | |
5188 | _generic_unldi_u(jit_state_t *_jit, | |
5189 | jit_int32_t r0, jit_word_t i0, jit_word_t i1) | |
5190 | { | |
5191 | jit_int32_t t0, r2; | |
5192 | assert(i1 >= 1 && i1 <= sizeof(jit_word_t)); | |
5193 | if (i1 & (i1 - 1)) { | |
5194 | t0 = jit_get_reg(jit_class_gpr); | |
5195 | r2 = rn(t0); | |
5196 | } | |
5197 | switch (i1) { | |
5198 | case 1: | |
5199 | ldi_uc(r0, i0); | |
5200 | break; | |
5201 | case 2: | |
5202 | ldi_us(r0, i0); | |
5203 | break; | |
5204 | case 3: | |
5205 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5206 | ldi_us(r0, i0); | |
5207 | ldi_uc(r2, i0 + 2); | |
5208 | lshi(r2, r2, 16); | |
5209 | # else | |
5210 | ldi_uc(r0, i0); | |
5211 | lshi(r0, r0, 16); | |
5212 | ldi_us(r2, i0 + 1); | |
5213 | # endif | |
5214 | break; | |
5215 | # if __WORDSIZE == 32 | |
5216 | default: | |
5217 | ldi_i(r0, i0); | |
5218 | break; | |
5219 | # else | |
5220 | case 4: | |
5221 | ldi_ui(r0, i0); | |
5222 | break; | |
5223 | case 5: | |
5224 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5225 | ldi_ui(r0, i0); | |
5226 | ldi_uc(r2, i0 + 4); | |
5227 | lshi(r2, r2, 32); | |
5228 | # else | |
5229 | ldi_ui(r0, i0); | |
5230 | lshi(r0, r0, 8); | |
5231 | ldi_uc(r2, i0 + 4); | |
5232 | # endif | |
5233 | break; | |
5234 | case 6: | |
5235 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5236 | ldi_ui(r0, i0); | |
5237 | ldi_us(r2, i0 + 4); | |
5238 | lshi(r2, r2, 32); | |
5239 | # else | |
5240 | ldi_ui(r0, i0); | |
5241 | lshi(r0, r0, 16); | |
5242 | ldi_us(r2, i0 + 4); | |
5243 | # endif | |
5244 | break; | |
5245 | case 7: | |
5246 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5247 | ldi_ui(r0, i0); | |
5248 | ldi_ui(r2, i0 + 4); | |
5249 | lshi(r2, r2, 40); | |
5250 | rshi_u(r2, r2, 8); | |
5251 | # else | |
5252 | ldi_ui(r0, i0); | |
5253 | lshi(r0, r0, 24); | |
5254 | ldi_ui(r2, i0 + 4); | |
5255 | rshi(r2, r2, 8); | |
5256 | # endif | |
5257 | break; | |
5258 | default: | |
5259 | ldi_l(r0, i0); | |
5260 | break; | |
5261 | # endif | |
5262 | } | |
5263 | if (i1 & (i1 - 1)) { | |
5264 | orr(r0, r0, r2); | |
5265 | jit_unget_reg(t0); | |
5266 | } | |
5267 | } | |
5268 | ||
5269 | static void | |
5270 | _generic_unstr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0) | |
5271 | { | |
5272 | jit_word_t t0, r2; | |
5273 | assert(i0 > 0 && i0 <= sizeof(jit_word_t)); | |
5274 | if (i0 & (i0 - 1)) { | |
5275 | t0 = jit_get_reg(jit_class_gpr); | |
5276 | r2 = rn(t0); | |
5277 | } | |
5278 | switch (i0) { | |
5279 | case 1: | |
5280 | str_c(r0, r1); | |
5281 | break; | |
5282 | case 2: | |
5283 | str_s(r0, r1); | |
5284 | break; | |
5285 | case 3: | |
5286 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5287 | str_s(r0, r1); | |
5288 | rshi(r2, r1, 16); | |
5289 | stxi_c(2, r0, r2); | |
5290 | # else | |
5291 | stxi_c(2, r0, r1); | |
5292 | rshi(r2, r1, 8); | |
5293 | str_s(r0, r2); | |
5294 | # endif | |
5295 | break; | |
5296 | # if __WORDSIZE == 32 | |
5297 | default: | |
5298 | str_i(r0, r1); | |
5299 | break; | |
5300 | # else | |
5301 | case 4: | |
5302 | str_i(r0, r1); | |
5303 | break; | |
5304 | case 5: | |
5305 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5306 | str_i(r0, r1); | |
5307 | rshi(r2, r1, 32); | |
5308 | stxi_c(4, r0, r2); | |
5309 | # else | |
5310 | stxi_c(4, r0, r1); | |
5311 | rshi(r2, r1, 8); | |
5312 | str_i(r0, r2); | |
5313 | # endif | |
5314 | break; | |
5315 | case 6: | |
5316 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5317 | str_i(r0, r1); | |
5318 | rshi(r2, r1, 32); | |
5319 | stxi_s(4, r0, r2); | |
5320 | # else | |
5321 | stxi_s(4, r0, r1); | |
5322 | rshi(r2, r1, 16); | |
5323 | str_i(r0, r2); | |
5324 | # endif | |
5325 | break; | |
5326 | case 7: | |
5327 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5328 | str_i(r0, r1); | |
5329 | rshi(r2, r1, 32); | |
5330 | stxi_s(4, r0, r2); | |
5331 | rshi(r2, r2, 16); | |
5332 | stxi_c(6, r0, r2); | |
5333 | # else | |
5334 | stxi_c(6, r0, r1); | |
5335 | rshi(r2, r1, 8); | |
5336 | stxi_s(4, r0, r2); | |
5337 | rshi(r2, r2, 16); | |
5338 | str_i(r0, r2); | |
5339 | # endif | |
5340 | break; | |
5341 | default: | |
5342 | str_l(r0, r1); | |
5343 | break; | |
5344 | # endif | |
5345 | } | |
5346 | if (i0 & (i0 - 1)) | |
5347 | jit_unget_reg(t0); | |
5348 | } | |
5349 | ||
5350 | static void | |
5351 | _generic_unsti(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1) | |
5352 | { | |
5353 | jit_word_t t0, r2; | |
5354 | assert(i1 > 0 && i1 <= sizeof(jit_word_t)); | |
5355 | if (i1 & (i1 - 1)) { | |
5356 | t0 = jit_get_reg(jit_class_gpr); | |
5357 | r2 = rn(t0); | |
5358 | } | |
5359 | switch (i1) { | |
5360 | case 1: | |
5361 | sti_c(i0, r0); | |
5362 | break; | |
5363 | case 2: | |
5364 | sti_s(i0, r0); | |
5365 | break; | |
5366 | case 3: | |
5367 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5368 | sti_s(i0, r0); | |
5369 | rshi(r2, r0, 16); | |
5370 | sti_c(2 + i0, r2); | |
5371 | # else | |
5372 | sti_c(2 + i0, r0); | |
5373 | rshi(r2, r0, 8); | |
5374 | sti_s(i0, r2); | |
5375 | # endif | |
5376 | break; | |
5377 | # if __WORDSIZE == 32 | |
5378 | default: | |
5379 | sti_i(i0, r0); | |
5380 | break; | |
5381 | # else | |
5382 | case 4: | |
5383 | sti_i(i0, r0); | |
5384 | break; | |
5385 | case 5: | |
5386 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5387 | sti_i(i0, r0); | |
5388 | rshi(r2, r0, 32); | |
5389 | sti_c(4 + i0, r2); | |
5390 | # else | |
5391 | stxi_c(4, i0, r0); | |
5392 | rshi(r2, r0, 8); | |
5393 | sti_i(i0, r2); | |
5394 | # endif | |
5395 | break; | |
5396 | case 6: | |
5397 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5398 | sti_i(i0, r0); | |
5399 | rshi(r2, r0, 32); | |
5400 | sti_s(4 + i0, r2); | |
5401 | # else | |
5402 | sti_s(4 + i0, r0); | |
5403 | rshi(r2, r0, 16); | |
5404 | sti_i(i0, r2); | |
5405 | # endif | |
5406 | break; | |
5407 | case 7: | |
5408 | # if __BYTE_ORDER == __LITTLE_ENDIAN | |
5409 | sti_i(i0, r0); | |
5410 | rshi(r2, r0, 32); | |
5411 | sti_s(4 + i0, r2); | |
5412 | rshi(r2, r2, 16); | |
5413 | sti_c(6 + i0, r2); | |
5414 | # else | |
5415 | sti_c(6 + i0, r0); | |
5416 | rshi(r2, r0, 8); | |
5417 | sti_s(4 + i0, r2); | |
5418 | rshi(r2, r2, 16); | |
5419 | sti_i(i0, r2); | |
5420 | # endif | |
5421 | break; | |
5422 | default: | |
5423 | sti_l(i0, r0); | |
5424 | break; | |
5425 | # endif | |
5426 | } | |
5427 | if (i1 & (i1 - 1)) | |
5428 | jit_unget_reg(t0); | |
5429 | } | |
5430 | ||
5431 | #if !defined(__i386__) && !defined(__x86_64__) && !defined(__arm__) | |
5432 | static void | |
5433 | _generic_unldr_x(jit_state_t *_jit, | |
5434 | jit_int32_t r0, jit_int32_t r1, jit_word_t i0) | |
5435 | { | |
5436 | assert(i0 == 4 || i0 == 8); | |
5437 | if (i0 == 4) | |
5438 | ldr_f(r0, r1); | |
5439 | else | |
5440 | ldr_d(r0, r1); | |
5441 | } | |
5442 | ||
5443 | static void | |
5444 | _generic_unldi_x(jit_state_t *_jit, | |
5445 | jit_int32_t r0, jit_word_t i0, jit_word_t i1) | |
5446 | { | |
5447 | assert(i1 == 4 || i1 == 8); | |
5448 | if (i1 == 4) | |
5449 | ldi_f(r0, i0); | |
5450 | else | |
5451 | ldi_d(r0, i0); | |
5452 | } | |
5453 | ||
5454 | static void | |
5455 | _generic_unstr_x(jit_state_t *_jit, | |
5456 | jit_int32_t r0, jit_int32_t r1, jit_word_t i0) | |
5457 | { | |
5458 | assert(i0 == 4 || i0 == 8); | |
5459 | if (i0 == 4) | |
5460 | str_f(r0, r1); | |
5461 | else | |
5462 | str_d(r0, r1); | |
5463 | } | |
5464 | ||
5465 | static void | |
5466 | _generic_unsti_x(jit_state_t *_jit, | |
5467 | jit_word_t i0, jit_int32_t r0, jit_word_t i1) | |
5468 | { | |
5469 | assert(i1 == 4 || i1 == 8); | |
5470 | if (i1 == 4) | |
5471 | sti_f(i0, r0); | |
5472 | else | |
5473 | sti_d(i0, r0); | |
5474 | } | |
5475 | #endif | |
5476 | ||
79bfeef6 PC |
5477 | #if defined(stack_framesize) |
5478 | static maybe_unused void | |
5479 | _patch_alist(jit_state_t *_jit, jit_bool_t revert) | |
5480 | { | |
5481 | jit_int32_t diff; | |
5482 | jit_node_t *node; | |
5483 | diff = jit_diffsize(); | |
5484 | if (diff) { | |
5485 | if (revert) | |
5486 | diff = -diff; | |
5487 | for (node = _jitc->function->alist; node; node = node->link) { | |
5488 | switch (node->code) { | |
5489 | case jit_code_ldxi_c: case jit_code_ldxi_uc: | |
5490 | case jit_code_ldxi_s: case jit_code_ldxi_us: | |
5491 | case jit_code_ldxi_i: | |
5492 | #if __WORDSIZE == 64 | |
5493 | case jit_code_ldxi_ui: case jit_code_ldxi_l: | |
5494 | #endif | |
5495 | case jit_code_ldxi_f: case jit_code_ldxi_d: | |
5496 | node->w.w -= diff; | |
5497 | break; | |
5498 | case jit_code_stxi_c: case jit_code_stxi_s: | |
5499 | case jit_code_stxi_i: | |
5500 | #if __WORDSIZE == 64 | |
5501 | case jit_code_stxi_l: | |
5502 | #endif | |
5503 | case jit_code_stxi_f: case jit_code_stxi_d: | |
5504 | node->u.w -= diff; | |
5505 | break; | |
5506 | default: | |
5507 | abort(); | |
5508 | } | |
5509 | } | |
5510 | } | |
5511 | } | |
5512 | #endif |