|
1 /* |
|
2 * Tiny Code Generator for QEMU |
|
3 * |
|
4 * Copyright (c) 2008 Fabrice Bellard |
|
5 * |
|
6 * Permission is hereby granted, free of charge, to any person obtaining a copy |
|
7 * of this software and associated documentation files (the "Software"), to deal |
|
8 * in the Software without restriction, including without limitation the rights |
|
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
|
10 * copies of the Software, and to permit persons to whom the Software is |
|
11 * furnished to do so, subject to the following conditions: |
|
12 * |
|
13 * The above copyright notice and this permission notice shall be included in |
|
14 * all copies or substantial portions of the Software. |
|
15 * |
|
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
|
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
|
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
|
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
|
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
|
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
|
22 * THE SOFTWARE. |
|
23 */ |
|
24 |
|
25 /* define it to suppress various consistency checks (faster) */ |
|
26 #define NDEBUG |
|
27 |
|
28 /* define it to use liveness analysis (better code) */ |
|
29 #define USE_LIVENESS_ANALYSIS |
|
30 |
|
31 #include <assert.h> |
|
32 #include <stdarg.h> |
|
33 #include <stdlib.h> |
|
34 #include <stdio.h> |
|
35 #include <string.h> |
|
36 #include <inttypes.h> |
|
37 #ifdef _WIN32 |
|
38 #include <malloc.h> |
|
39 #endif |
|
40 #ifdef _AIX |
|
41 #include <alloca.h> |
|
42 #endif |
|
43 |
|
44 #include "config.h" |
|
45 #include "qemu-common.h" |
|
46 #include "cache-utils.h" |
|
47 |
|
48 /* Note: the long term plan is to reduce the dependancies on the QEMU |
|
49 CPU definitions. Currently they are used for qemu_ld/st |
|
50 instructions */ |
|
51 #define NO_CPU_IO_DEFS |
|
52 #include "cpu.h" |
|
53 #include "exec-all.h" |
|
54 |
|
55 #include "tcg-op.h" |
|
56 #include "elf.h" |
|
57 |
|
58 |
|
59 static void patch_reloc(uint8_t *code_ptr, int type, |
|
60 tcg_target_long value, tcg_target_long addend); |
|
61 |
|
62 TCGOpDef tcg_op_defs[] = { |
|
63 #define DEF(s, n, copy_size) { #s, 0, 0, n, n, 0, copy_size }, |
|
64 #define DEF2(s, iargs, oargs, cargs, flags) { #s, iargs, oargs, cargs, iargs + oargs + cargs, flags, 0 }, |
|
65 #include "tcg-opc.h" |
|
66 #undef DEF |
|
67 #undef DEF2 |
|
68 }; |
|
69 |
|
70 static TCGRegSet tcg_target_available_regs[2]; |
|
71 static TCGRegSet tcg_target_call_clobber_regs; |
|
72 |
|
73 /* XXX: move that inside the context */ |
|
74 uint16_t *gen_opc_ptr; |
|
75 TCGArg *gen_opparam_ptr; |
|
76 |
|
77 static inline void tcg_out8(TCGContext *s, uint8_t v) |
|
78 { |
|
79 *s->code_ptr++ = v; |
|
80 } |
|
81 |
|
82 static inline void tcg_out16(TCGContext *s, uint16_t v) |
|
83 { |
|
84 *(uint16_t *)s->code_ptr = v; |
|
85 s->code_ptr += 2; |
|
86 } |
|
87 |
|
88 static inline void tcg_out32(TCGContext *s, uint32_t v) |
|
89 { |
|
90 *(uint32_t *)s->code_ptr = v; |
|
91 s->code_ptr += 4; |
|
92 } |
|
93 |
|
94 /* label relocation processing */ |
|
95 |
|
96 void tcg_out_reloc(TCGContext *s, uint8_t *code_ptr, int type, |
|
97 int label_index, long addend) |
|
98 { |
|
99 TCGLabel *l; |
|
100 TCGRelocation *r; |
|
101 |
|
102 l = &s->labels[label_index]; |
|
103 if (l->has_value) { |
|
104 /* FIXME: This may break relocations on RISC targets that |
|
105 modify instruction fields in place. The caller may not have |
|
106 written the initial value. */ |
|
107 patch_reloc(code_ptr, type, l->u.value, addend); |
|
108 } else { |
|
109 /* add a new relocation entry */ |
|
110 r = tcg_malloc(sizeof(TCGRelocation)); |
|
111 r->type = type; |
|
112 r->ptr = code_ptr; |
|
113 r->addend = addend; |
|
114 r->next = l->u.first_reloc; |
|
115 l->u.first_reloc = r; |
|
116 } |
|
117 } |
|
118 |
|
119 static void tcg_out_label(TCGContext *s, int label_index, |
|
120 tcg_target_long value) |
|
121 { |
|
122 TCGLabel *l; |
|
123 TCGRelocation *r; |
|
124 |
|
125 l = &s->labels[label_index]; |
|
126 if (l->has_value) |
|
127 tcg_abort(); |
|
128 r = l->u.first_reloc; |
|
129 while (r != NULL) { |
|
130 patch_reloc(r->ptr, r->type, value, r->addend); |
|
131 r = r->next; |
|
132 } |
|
133 l->has_value = 1; |
|
134 l->u.value = value; |
|
135 } |
|
136 |
|
137 int gen_new_label(void) |
|
138 { |
|
139 TCGContext *s = &tcg_ctx; |
|
140 int idx; |
|
141 TCGLabel *l; |
|
142 |
|
143 if (s->nb_labels >= TCG_MAX_LABELS) |
|
144 tcg_abort(); |
|
145 idx = s->nb_labels++; |
|
146 l = &s->labels[idx]; |
|
147 l->has_value = 0; |
|
148 l->u.first_reloc = NULL; |
|
149 return idx; |
|
150 } |
|
151 |
|
152 #include "tcg-target.c" |
|
153 |
|
154 /* pool based memory allocation */ |
|
155 void *tcg_malloc_internal(TCGContext *s, int size) |
|
156 { |
|
157 TCGPool *p; |
|
158 int pool_size; |
|
159 |
|
160 if (size > TCG_POOL_CHUNK_SIZE) { |
|
161 /* big malloc: insert a new pool (XXX: could optimize) */ |
|
162 p = qemu_malloc(sizeof(TCGPool) + size); |
|
163 p->size = size; |
|
164 if (s->pool_current) |
|
165 s->pool_current->next = p; |
|
166 else |
|
167 s->pool_first = p; |
|
168 p->next = s->pool_current; |
|
169 } else { |
|
170 p = s->pool_current; |
|
171 if (!p) { |
|
172 p = s->pool_first; |
|
173 if (!p) |
|
174 goto new_pool; |
|
175 } else { |
|
176 if (!p->next) { |
|
177 new_pool: |
|
178 pool_size = TCG_POOL_CHUNK_SIZE; |
|
179 p = qemu_malloc(sizeof(TCGPool) + pool_size); |
|
180 p->size = pool_size; |
|
181 p->next = NULL; |
|
182 if (s->pool_current) |
|
183 s->pool_current->next = p; |
|
184 else |
|
185 s->pool_first = p; |
|
186 } else { |
|
187 p = p->next; |
|
188 } |
|
189 } |
|
190 } |
|
191 s->pool_current = p; |
|
192 s->pool_cur = p->data + size; |
|
193 s->pool_end = p->data + p->size; |
|
194 return p->data; |
|
195 } |
|
196 |
|
197 void tcg_pool_reset(TCGContext *s) |
|
198 { |
|
199 s->pool_cur = s->pool_end = NULL; |
|
200 s->pool_current = NULL; |
|
201 } |
|
202 |
|
203 void tcg_context_init(TCGContext *s) |
|
204 { |
|
205 int op, total_args, n; |
|
206 TCGOpDef *def; |
|
207 TCGArgConstraint *args_ct; |
|
208 int *sorted_args; |
|
209 |
|
210 memset(s, 0, sizeof(*s)); |
|
211 s->temps = s->static_temps; |
|
212 s->nb_globals = 0; |
|
213 |
|
214 /* Count total number of arguments and allocate the corresponding |
|
215 space */ |
|
216 total_args = 0; |
|
217 for(op = 0; op < NB_OPS; op++) { |
|
218 def = &tcg_op_defs[op]; |
|
219 n = def->nb_iargs + def->nb_oargs; |
|
220 total_args += n; |
|
221 } |
|
222 |
|
223 args_ct = qemu_malloc(sizeof(TCGArgConstraint) * total_args); |
|
224 sorted_args = qemu_malloc(sizeof(int) * total_args); |
|
225 |
|
226 for(op = 0; op < NB_OPS; op++) { |
|
227 def = &tcg_op_defs[op]; |
|
228 def->args_ct = args_ct; |
|
229 def->sorted_args = sorted_args; |
|
230 n = def->nb_iargs + def->nb_oargs; |
|
231 sorted_args += n; |
|
232 args_ct += n; |
|
233 } |
|
234 |
|
235 tcg_target_init(s); |
|
236 |
|
237 /* init global prologue and epilogue */ |
|
238 s->code_buf = code_gen_prologue; |
|
239 s->code_ptr = s->code_buf; |
|
240 tcg_target_qemu_prologue(s); |
|
241 flush_icache_range((unsigned long)s->code_buf, |
|
242 (unsigned long)s->code_ptr); |
|
243 } |
|
244 |
|
245 void tcg_set_frame(TCGContext *s, int reg, |
|
246 tcg_target_long start, tcg_target_long size) |
|
247 { |
|
248 s->frame_start = start; |
|
249 s->frame_end = start + size; |
|
250 s->frame_reg = reg; |
|
251 } |
|
252 |
|
253 void tcg_func_start(TCGContext *s) |
|
254 { |
|
255 int i; |
|
256 tcg_pool_reset(s); |
|
257 s->nb_temps = s->nb_globals; |
|
258 for(i = 0; i < (TCG_TYPE_COUNT * 2); i++) |
|
259 s->first_free_temp[i] = -1; |
|
260 s->labels = tcg_malloc(sizeof(TCGLabel) * TCG_MAX_LABELS); |
|
261 s->nb_labels = 0; |
|
262 s->current_frame_offset = s->frame_start; |
|
263 |
|
264 gen_opc_ptr = gen_opc_buf; |
|
265 gen_opparam_ptr = gen_opparam_buf; |
|
266 } |
|
267 |
|
268 static inline void tcg_temp_alloc(TCGContext *s, int n) |
|
269 { |
|
270 if (n > TCG_MAX_TEMPS) |
|
271 tcg_abort(); |
|
272 } |
|
273 |
|
274 static inline int tcg_global_reg_new_internal(TCGType type, int reg, |
|
275 const char *name) |
|
276 { |
|
277 TCGContext *s = &tcg_ctx; |
|
278 TCGTemp *ts; |
|
279 int idx; |
|
280 |
|
281 #if TCG_TARGET_REG_BITS == 32 |
|
282 if (type != TCG_TYPE_I32) |
|
283 tcg_abort(); |
|
284 #endif |
|
285 if (tcg_regset_test_reg(s->reserved_regs, reg)) |
|
286 tcg_abort(); |
|
287 idx = s->nb_globals; |
|
288 tcg_temp_alloc(s, s->nb_globals + 1); |
|
289 ts = &s->temps[s->nb_globals]; |
|
290 ts->base_type = type; |
|
291 ts->type = type; |
|
292 ts->fixed_reg = 1; |
|
293 ts->reg = reg; |
|
294 ts->name = name; |
|
295 s->nb_globals++; |
|
296 tcg_regset_set_reg(s->reserved_regs, reg); |
|
297 return idx; |
|
298 } |
|
299 |
|
300 TCGv_i32 tcg_global_reg_new_i32(int reg, const char *name) |
|
301 { |
|
302 int idx; |
|
303 |
|
304 idx = tcg_global_reg_new_internal(TCG_TYPE_I32, reg, name); |
|
305 return MAKE_TCGV_I32(idx); |
|
306 } |
|
307 |
|
308 TCGv_i64 tcg_global_reg_new_i64(int reg, const char *name) |
|
309 { |
|
310 int idx; |
|
311 |
|
312 idx = tcg_global_reg_new_internal(TCG_TYPE_I64, reg, name); |
|
313 return MAKE_TCGV_I64(idx); |
|
314 } |
|
315 |
|
316 #if TCG_TARGET_REG_BITS == 32 |
|
317 /* temporary hack to avoid register shortage for tcg_qemu_st64() */ |
|
318 TCGv_i64 tcg_global_reg2_new_hack(TCGType type, int reg1, int reg2, |
|
319 const char *name) |
|
320 { |
|
321 TCGContext *s = &tcg_ctx; |
|
322 TCGTemp *ts; |
|
323 int idx; |
|
324 char buf[64]; |
|
325 |
|
326 if (type != TCG_TYPE_I64) |
|
327 tcg_abort(); |
|
328 idx = s->nb_globals; |
|
329 tcg_temp_alloc(s, s->nb_globals + 2); |
|
330 ts = &s->temps[s->nb_globals]; |
|
331 ts->base_type = type; |
|
332 ts->type = TCG_TYPE_I32; |
|
333 ts->fixed_reg = 1; |
|
334 ts->reg = reg1; |
|
335 pstrcpy(buf, sizeof(buf), name); |
|
336 pstrcat(buf, sizeof(buf), "_0"); |
|
337 ts->name = strdup(buf); |
|
338 |
|
339 ts++; |
|
340 ts->base_type = type; |
|
341 ts->type = TCG_TYPE_I32; |
|
342 ts->fixed_reg = 1; |
|
343 ts->reg = reg2; |
|
344 pstrcpy(buf, sizeof(buf), name); |
|
345 pstrcat(buf, sizeof(buf), "_1"); |
|
346 ts->name = strdup(buf); |
|
347 |
|
348 s->nb_globals += 2; |
|
349 return MAKE_TCGV_I64(idx); |
|
350 } |
|
351 #endif |
|
352 |
|
353 static inline int tcg_global_mem_new_internal(TCGType type, int reg, |
|
354 tcg_target_long offset, |
|
355 const char *name) |
|
356 { |
|
357 TCGContext *s = &tcg_ctx; |
|
358 TCGTemp *ts; |
|
359 int idx; |
|
360 |
|
361 idx = s->nb_globals; |
|
362 #if TCG_TARGET_REG_BITS == 32 |
|
363 if (type == TCG_TYPE_I64) { |
|
364 char buf[64]; |
|
365 tcg_temp_alloc(s, s->nb_globals + 2); |
|
366 ts = &s->temps[s->nb_globals]; |
|
367 ts->base_type = type; |
|
368 ts->type = TCG_TYPE_I32; |
|
369 ts->fixed_reg = 0; |
|
370 ts->mem_allocated = 1; |
|
371 ts->mem_reg = reg; |
|
372 #ifdef TCG_TARGET_WORDS_BIGENDIAN |
|
373 ts->mem_offset = offset + 4; |
|
374 #else |
|
375 ts->mem_offset = offset; |
|
376 #endif |
|
377 pstrcpy(buf, sizeof(buf), name); |
|
378 pstrcat(buf, sizeof(buf), "_0"); |
|
379 ts->name = strdup(buf); |
|
380 ts++; |
|
381 |
|
382 ts->base_type = type; |
|
383 ts->type = TCG_TYPE_I32; |
|
384 ts->fixed_reg = 0; |
|
385 ts->mem_allocated = 1; |
|
386 ts->mem_reg = reg; |
|
387 #ifdef TCG_TARGET_WORDS_BIGENDIAN |
|
388 ts->mem_offset = offset; |
|
389 #else |
|
390 ts->mem_offset = offset + 4; |
|
391 #endif |
|
392 pstrcpy(buf, sizeof(buf), name); |
|
393 pstrcat(buf, sizeof(buf), "_1"); |
|
394 ts->name = strdup(buf); |
|
395 |
|
396 s->nb_globals += 2; |
|
397 } else |
|
398 #endif |
|
399 { |
|
400 tcg_temp_alloc(s, s->nb_globals + 1); |
|
401 ts = &s->temps[s->nb_globals]; |
|
402 ts->base_type = type; |
|
403 ts->type = type; |
|
404 ts->fixed_reg = 0; |
|
405 ts->mem_allocated = 1; |
|
406 ts->mem_reg = reg; |
|
407 ts->mem_offset = offset; |
|
408 ts->name = name; |
|
409 s->nb_globals++; |
|
410 } |
|
411 return idx; |
|
412 } |
|
413 |
|
414 TCGv_i32 tcg_global_mem_new_i32(int reg, tcg_target_long offset, |
|
415 const char *name) |
|
416 { |
|
417 int idx; |
|
418 |
|
419 idx = tcg_global_mem_new_internal(TCG_TYPE_I32, reg, offset, name); |
|
420 return MAKE_TCGV_I32(idx); |
|
421 } |
|
422 |
|
423 TCGv_i64 tcg_global_mem_new_i64(int reg, tcg_target_long offset, |
|
424 const char *name) |
|
425 { |
|
426 int idx; |
|
427 |
|
428 idx = tcg_global_mem_new_internal(TCG_TYPE_I64, reg, offset, name); |
|
429 return MAKE_TCGV_I64(idx); |
|
430 } |
|
431 |
|
432 static inline int tcg_temp_new_internal(TCGType type, int temp_local) |
|
433 { |
|
434 TCGContext *s = &tcg_ctx; |
|
435 TCGTemp *ts; |
|
436 int idx, k; |
|
437 |
|
438 k = type; |
|
439 if (temp_local) |
|
440 k += TCG_TYPE_COUNT; |
|
441 idx = s->first_free_temp[k]; |
|
442 if (idx != -1) { |
|
443 /* There is already an available temp with the |
|
444 right type */ |
|
445 ts = &s->temps[idx]; |
|
446 s->first_free_temp[k] = ts->next_free_temp; |
|
447 ts->temp_allocated = 1; |
|
448 assert(ts->temp_local == temp_local); |
|
449 } else { |
|
450 idx = s->nb_temps; |
|
451 #if TCG_TARGET_REG_BITS == 32 |
|
452 if (type == TCG_TYPE_I64) { |
|
453 tcg_temp_alloc(s, s->nb_temps + 2); |
|
454 ts = &s->temps[s->nb_temps]; |
|
455 ts->base_type = type; |
|
456 ts->type = TCG_TYPE_I32; |
|
457 ts->temp_allocated = 1; |
|
458 ts->temp_local = temp_local; |
|
459 ts->name = NULL; |
|
460 ts++; |
|
461 ts->base_type = TCG_TYPE_I32; |
|
462 ts->type = TCG_TYPE_I32; |
|
463 ts->temp_allocated = 1; |
|
464 ts->temp_local = temp_local; |
|
465 ts->name = NULL; |
|
466 s->nb_temps += 2; |
|
467 } else |
|
468 #endif |
|
469 { |
|
470 tcg_temp_alloc(s, s->nb_temps + 1); |
|
471 ts = &s->temps[s->nb_temps]; |
|
472 ts->base_type = type; |
|
473 ts->type = type; |
|
474 ts->temp_allocated = 1; |
|
475 ts->temp_local = temp_local; |
|
476 ts->name = NULL; |
|
477 s->nb_temps++; |
|
478 } |
|
479 } |
|
480 return idx; |
|
481 } |
|
482 |
|
483 TCGv_i32 tcg_temp_new_internal_i32(int temp_local) |
|
484 { |
|
485 int idx; |
|
486 |
|
487 idx = tcg_temp_new_internal(TCG_TYPE_I32, temp_local); |
|
488 return MAKE_TCGV_I32(idx); |
|
489 } |
|
490 |
|
491 TCGv_i64 tcg_temp_new_internal_i64(int temp_local) |
|
492 { |
|
493 int idx; |
|
494 |
|
495 idx = tcg_temp_new_internal(TCG_TYPE_I64, temp_local); |
|
496 return MAKE_TCGV_I64(idx); |
|
497 } |
|
498 |
|
499 static inline void tcg_temp_free_internal(int idx) |
|
500 { |
|
501 TCGContext *s = &tcg_ctx; |
|
502 TCGTemp *ts; |
|
503 int k; |
|
504 |
|
505 assert(idx >= s->nb_globals && idx < s->nb_temps); |
|
506 ts = &s->temps[idx]; |
|
507 assert(ts->temp_allocated != 0); |
|
508 ts->temp_allocated = 0; |
|
509 k = ts->base_type; |
|
510 if (ts->temp_local) |
|
511 k += TCG_TYPE_COUNT; |
|
512 ts->next_free_temp = s->first_free_temp[k]; |
|
513 s->first_free_temp[k] = idx; |
|
514 } |
|
515 |
|
516 void tcg_temp_free_i32(TCGv_i32 arg) |
|
517 { |
|
518 tcg_temp_free_internal(GET_TCGV_I32(arg)); |
|
519 } |
|
520 |
|
521 void tcg_temp_free_i64(TCGv_i64 arg) |
|
522 { |
|
523 tcg_temp_free_internal(GET_TCGV_I64(arg)); |
|
524 } |
|
525 |
|
526 TCGv_i32 tcg_const_i32(int32_t val) |
|
527 { |
|
528 TCGv_i32 t0; |
|
529 t0 = tcg_temp_new_i32(); |
|
530 tcg_gen_movi_i32(t0, val); |
|
531 return t0; |
|
532 } |
|
533 |
|
534 TCGv_i64 tcg_const_i64(int64_t val) |
|
535 { |
|
536 TCGv_i64 t0; |
|
537 t0 = tcg_temp_new_i64(); |
|
538 tcg_gen_movi_i64(t0, val); |
|
539 return t0; |
|
540 } |
|
541 |
|
542 TCGv_i32 tcg_const_local_i32(int32_t val) |
|
543 { |
|
544 TCGv_i32 t0; |
|
545 t0 = tcg_temp_local_new_i32(); |
|
546 tcg_gen_movi_i32(t0, val); |
|
547 return t0; |
|
548 } |
|
549 |
|
550 TCGv_i64 tcg_const_local_i64(int64_t val) |
|
551 { |
|
552 TCGv_i64 t0; |
|
553 t0 = tcg_temp_local_new_i64(); |
|
554 tcg_gen_movi_i64(t0, val); |
|
555 return t0; |
|
556 } |
|
557 |
|
558 void tcg_register_helper(void *func, const char *name) |
|
559 { |
|
560 TCGContext *s = &tcg_ctx; |
|
561 int n; |
|
562 if ((s->nb_helpers + 1) > s->allocated_helpers) { |
|
563 n = s->allocated_helpers; |
|
564 if (n == 0) { |
|
565 n = 4; |
|
566 } else { |
|
567 n *= 2; |
|
568 } |
|
569 s->helpers = realloc(s->helpers, n * sizeof(TCGHelperInfo)); |
|
570 s->allocated_helpers = n; |
|
571 } |
|
572 s->helpers[s->nb_helpers].func = (tcg_target_ulong)func; |
|
573 s->helpers[s->nb_helpers].name = name; |
|
574 s->nb_helpers++; |
|
575 } |
|
576 |
|
577 /* Note: we convert the 64 bit args to 32 bit and do some alignment |
|
578 and endian swap. Maybe it would be better to do the alignment |
|
579 and endian swap in tcg_reg_alloc_call(). */ |
|
580 void tcg_gen_callN(TCGContext *s, TCGv_ptr func, unsigned int flags, |
|
581 int sizemask, TCGArg ret, int nargs, TCGArg *args) |
|
582 { |
|
583 int call_type; |
|
584 int i; |
|
585 int real_args; |
|
586 int nb_rets; |
|
587 TCGArg *nparam; |
|
588 *gen_opc_ptr++ = INDEX_op_call; |
|
589 nparam = gen_opparam_ptr++; |
|
590 call_type = (flags & TCG_CALL_TYPE_MASK); |
|
591 if (ret != TCG_CALL_DUMMY_ARG) { |
|
592 #if TCG_TARGET_REG_BITS < 64 |
|
593 if (sizemask & 1) { |
|
594 #ifdef TCG_TARGET_WORDS_BIGENDIAN |
|
595 *gen_opparam_ptr++ = ret + 1; |
|
596 *gen_opparam_ptr++ = ret; |
|
597 #else |
|
598 *gen_opparam_ptr++ = ret; |
|
599 *gen_opparam_ptr++ = ret + 1; |
|
600 #endif |
|
601 nb_rets = 2; |
|
602 } else |
|
603 #endif |
|
604 { |
|
605 *gen_opparam_ptr++ = ret; |
|
606 nb_rets = 1; |
|
607 } |
|
608 } else { |
|
609 nb_rets = 0; |
|
610 } |
|
611 real_args = 0; |
|
612 for (i = 0; i < nargs; i++) { |
|
613 #if TCG_TARGET_REG_BITS < 64 |
|
614 if (sizemask & (2 << i)) { |
|
615 #ifdef TCG_TARGET_I386 |
|
616 /* REGPARM case: if the third parameter is 64 bit, it is |
|
617 allocated on the stack */ |
|
618 if (i == 2 && call_type == TCG_CALL_TYPE_REGPARM) { |
|
619 call_type = TCG_CALL_TYPE_REGPARM_2; |
|
620 flags = (flags & ~TCG_CALL_TYPE_MASK) | call_type; |
|
621 } |
|
622 #endif |
|
623 #ifdef TCG_TARGET_CALL_ALIGN_ARGS |
|
624 /* some targets want aligned 64 bit args */ |
|
625 if (real_args & 1) { |
|
626 *gen_opparam_ptr++ = TCG_CALL_DUMMY_ARG; |
|
627 real_args++; |
|
628 } |
|
629 #endif |
|
630 #ifdef TCG_TARGET_WORDS_BIGENDIAN |
|
631 *gen_opparam_ptr++ = args[i] + 1; |
|
632 *gen_opparam_ptr++ = args[i]; |
|
633 #else |
|
634 *gen_opparam_ptr++ = args[i]; |
|
635 *gen_opparam_ptr++ = args[i] + 1; |
|
636 #endif |
|
637 real_args += 2; |
|
638 } else |
|
639 #endif |
|
640 { |
|
641 *gen_opparam_ptr++ = args[i]; |
|
642 real_args++; |
|
643 } |
|
644 } |
|
645 *gen_opparam_ptr++ = GET_TCGV_PTR(func); |
|
646 |
|
647 *gen_opparam_ptr++ = flags; |
|
648 |
|
649 *nparam = (nb_rets << 16) | (real_args + 1); |
|
650 |
|
651 /* total parameters, needed to go backward in the instruction stream */ |
|
652 *gen_opparam_ptr++ = 1 + nb_rets + real_args + 3; |
|
653 } |
|
654 |
|
655 #if TCG_TARGET_REG_BITS == 32 |
|
656 void tcg_gen_shifti_i64(TCGv_i64 ret, TCGv_i64 arg1, |
|
657 int c, int right, int arith) |
|
658 { |
|
659 if (c == 0) { |
|
660 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1)); |
|
661 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1)); |
|
662 } else if (c >= 32) { |
|
663 c -= 32; |
|
664 if (right) { |
|
665 if (arith) { |
|
666 tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
|
667 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31); |
|
668 } else { |
|
669 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
|
670 tcg_gen_movi_i32(TCGV_HIGH(ret), 0); |
|
671 } |
|
672 } else { |
|
673 tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_LOW(arg1), c); |
|
674 tcg_gen_movi_i32(TCGV_LOW(ret), 0); |
|
675 } |
|
676 } else { |
|
677 TCGv_i32 t0, t1; |
|
678 |
|
679 t0 = tcg_temp_new_i32(); |
|
680 t1 = tcg_temp_new_i32(); |
|
681 if (right) { |
|
682 tcg_gen_shli_i32(t0, TCGV_HIGH(arg1), 32 - c); |
|
683 if (arith) |
|
684 tcg_gen_sari_i32(t1, TCGV_HIGH(arg1), c); |
|
685 else |
|
686 tcg_gen_shri_i32(t1, TCGV_HIGH(arg1), c); |
|
687 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c); |
|
688 tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(ret), t0); |
|
689 tcg_gen_mov_i32(TCGV_HIGH(ret), t1); |
|
690 } else { |
|
691 tcg_gen_shri_i32(t0, TCGV_LOW(arg1), 32 - c); |
|
692 /* Note: ret can be the same as arg1, so we use t1 */ |
|
693 tcg_gen_shli_i32(t1, TCGV_LOW(arg1), c); |
|
694 tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c); |
|
695 tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t0); |
|
696 tcg_gen_mov_i32(TCGV_LOW(ret), t1); |
|
697 } |
|
698 tcg_temp_free_i32(t0); |
|
699 tcg_temp_free_i32(t1); |
|
700 } |
|
701 } |
|
702 #endif |
|
703 |
|
704 static void tcg_reg_alloc_start(TCGContext *s) |
|
705 { |
|
706 int i; |
|
707 TCGTemp *ts; |
|
708 for(i = 0; i < s->nb_globals; i++) { |
|
709 ts = &s->temps[i]; |
|
710 if (ts->fixed_reg) { |
|
711 ts->val_type = TEMP_VAL_REG; |
|
712 } else { |
|
713 ts->val_type = TEMP_VAL_MEM; |
|
714 } |
|
715 } |
|
716 for(i = s->nb_globals; i < s->nb_temps; i++) { |
|
717 ts = &s->temps[i]; |
|
718 ts->val_type = TEMP_VAL_DEAD; |
|
719 ts->mem_allocated = 0; |
|
720 ts->fixed_reg = 0; |
|
721 } |
|
722 for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
|
723 s->reg_to_temp[i] = -1; |
|
724 } |
|
725 } |
|
726 |
|
727 static char *tcg_get_arg_str_idx(TCGContext *s, char *buf, int buf_size, |
|
728 int idx) |
|
729 { |
|
730 TCGTemp *ts; |
|
731 |
|
732 ts = &s->temps[idx]; |
|
733 if (idx < s->nb_globals) { |
|
734 pstrcpy(buf, buf_size, ts->name); |
|
735 } else { |
|
736 if (ts->temp_local) |
|
737 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); |
|
738 else |
|
739 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); |
|
740 } |
|
741 return buf; |
|
742 } |
|
743 |
|
744 char *tcg_get_arg_str_i32(TCGContext *s, char *buf, int buf_size, TCGv_i32 arg) |
|
745 { |
|
746 return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I32(arg)); |
|
747 } |
|
748 |
|
749 char *tcg_get_arg_str_i64(TCGContext *s, char *buf, int buf_size, TCGv_i64 arg) |
|
750 { |
|
751 return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I64(arg)); |
|
752 } |
|
753 |
|
754 static int helper_cmp(const void *p1, const void *p2) |
|
755 { |
|
756 const TCGHelperInfo *th1 = p1; |
|
757 const TCGHelperInfo *th2 = p2; |
|
758 if (th1->func < th2->func) |
|
759 return -1; |
|
760 else if (th1->func == th2->func) |
|
761 return 0; |
|
762 else |
|
763 return 1; |
|
764 } |
|
765 |
|
766 /* find helper definition (Note: A hash table would be better) */ |
|
767 static TCGHelperInfo *tcg_find_helper(TCGContext *s, tcg_target_ulong val) |
|
768 { |
|
769 int m, m_min, m_max; |
|
770 TCGHelperInfo *th; |
|
771 tcg_target_ulong v; |
|
772 |
|
773 if (unlikely(!s->helpers_sorted)) { |
|
774 qsort(s->helpers, s->nb_helpers, sizeof(TCGHelperInfo), |
|
775 helper_cmp); |
|
776 s->helpers_sorted = 1; |
|
777 } |
|
778 |
|
779 /* binary search */ |
|
780 m_min = 0; |
|
781 m_max = s->nb_helpers - 1; |
|
782 while (m_min <= m_max) { |
|
783 m = (m_min + m_max) >> 1; |
|
784 th = &s->helpers[m]; |
|
785 v = th->func; |
|
786 if (v == val) |
|
787 return th; |
|
788 else if (val < v) { |
|
789 m_max = m - 1; |
|
790 } else { |
|
791 m_min = m + 1; |
|
792 } |
|
793 } |
|
794 return NULL; |
|
795 } |
|
796 |
|
797 static const char * const cond_name[] = |
|
798 { |
|
799 [TCG_COND_EQ] = "eq", |
|
800 [TCG_COND_NE] = "ne", |
|
801 [TCG_COND_LT] = "lt", |
|
802 [TCG_COND_GE] = "ge", |
|
803 [TCG_COND_LE] = "le", |
|
804 [TCG_COND_GT] = "gt", |
|
805 [TCG_COND_LTU] = "ltu", |
|
806 [TCG_COND_GEU] = "geu", |
|
807 [TCG_COND_LEU] = "leu", |
|
808 [TCG_COND_GTU] = "gtu" |
|
809 }; |
|
810 |
|
811 void tcg_dump_ops(TCGContext *s, FILE *outfile) |
|
812 { |
|
813 const uint16_t *opc_ptr; |
|
814 const TCGArg *args; |
|
815 TCGArg arg; |
|
816 int c, i, k, nb_oargs, nb_iargs, nb_cargs, first_insn; |
|
817 const TCGOpDef *def; |
|
818 char buf[128]; |
|
819 |
|
820 first_insn = 1; |
|
821 opc_ptr = gen_opc_buf; |
|
822 args = gen_opparam_buf; |
|
823 while (opc_ptr < gen_opc_ptr) { |
|
824 c = *opc_ptr++; |
|
825 def = &tcg_op_defs[c]; |
|
826 if (c == INDEX_op_debug_insn_start) { |
|
827 uint64_t pc; |
|
828 #if TARGET_LONG_BITS > TCG_TARGET_REG_BITS |
|
829 pc = ((uint64_t)args[1] << 32) | args[0]; |
|
830 #else |
|
831 pc = args[0]; |
|
832 #endif |
|
833 if (!first_insn) |
|
834 fprintf(outfile, "\n"); |
|
835 fprintf(outfile, " ---- 0x%" PRIx64, pc); |
|
836 first_insn = 0; |
|
837 nb_oargs = def->nb_oargs; |
|
838 nb_iargs = def->nb_iargs; |
|
839 nb_cargs = def->nb_cargs; |
|
840 } else if (c == INDEX_op_call) { |
|
841 TCGArg arg; |
|
842 |
|
843 /* variable number of arguments */ |
|
844 arg = *args++; |
|
845 nb_oargs = arg >> 16; |
|
846 nb_iargs = arg & 0xffff; |
|
847 nb_cargs = def->nb_cargs; |
|
848 |
|
849 fprintf(outfile, " %s ", def->name); |
|
850 |
|
851 /* function name */ |
|
852 fprintf(outfile, "%s", |
|
853 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + nb_iargs - 1])); |
|
854 /* flags */ |
|
855 fprintf(outfile, ",$0x%" TCG_PRIlx, |
|
856 args[nb_oargs + nb_iargs]); |
|
857 /* nb out args */ |
|
858 fprintf(outfile, ",$%d", nb_oargs); |
|
859 for(i = 0; i < nb_oargs; i++) { |
|
860 fprintf(outfile, ","); |
|
861 fprintf(outfile, "%s", |
|
862 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[i])); |
|
863 } |
|
864 for(i = 0; i < (nb_iargs - 1); i++) { |
|
865 fprintf(outfile, ","); |
|
866 if (args[nb_oargs + i] == TCG_CALL_DUMMY_ARG) { |
|
867 fprintf(outfile, "<dummy>"); |
|
868 } else { |
|
869 fprintf(outfile, "%s", |
|
870 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + i])); |
|
871 } |
|
872 } |
|
873 } else if (c == INDEX_op_movi_i32 |
|
874 #if TCG_TARGET_REG_BITS == 64 |
|
875 || c == INDEX_op_movi_i64 |
|
876 #endif |
|
877 ) { |
|
878 tcg_target_ulong val; |
|
879 TCGHelperInfo *th; |
|
880 |
|
881 nb_oargs = def->nb_oargs; |
|
882 nb_iargs = def->nb_iargs; |
|
883 nb_cargs = def->nb_cargs; |
|
884 fprintf(outfile, " %s %s,$", def->name, |
|
885 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[0])); |
|
886 val = args[1]; |
|
887 th = tcg_find_helper(s, val); |
|
888 if (th) { |
|
889 fprintf(outfile, "%s", th->name); |
|
890 } else { |
|
891 if (c == INDEX_op_movi_i32) |
|
892 fprintf(outfile, "0x%x", (uint32_t)val); |
|
893 else |
|
894 fprintf(outfile, "0x%" PRIx64 , (uint64_t)val); |
|
895 } |
|
896 } else { |
|
897 fprintf(outfile, " %s ", def->name); |
|
898 if (c == INDEX_op_nopn) { |
|
899 /* variable number of arguments */ |
|
900 nb_cargs = *args; |
|
901 nb_oargs = 0; |
|
902 nb_iargs = 0; |
|
903 } else { |
|
904 nb_oargs = def->nb_oargs; |
|
905 nb_iargs = def->nb_iargs; |
|
906 nb_cargs = def->nb_cargs; |
|
907 } |
|
908 |
|
909 k = 0; |
|
910 for(i = 0; i < nb_oargs; i++) { |
|
911 if (k != 0) |
|
912 fprintf(outfile, ","); |
|
913 fprintf(outfile, "%s", |
|
914 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++])); |
|
915 } |
|
916 for(i = 0; i < nb_iargs; i++) { |
|
917 if (k != 0) |
|
918 fprintf(outfile, ","); |
|
919 fprintf(outfile, "%s", |
|
920 tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++])); |
|
921 } |
|
922 if (c == INDEX_op_brcond_i32 |
|
923 #if TCG_TARGET_REG_BITS == 32 |
|
924 || c == INDEX_op_brcond2_i32 |
|
925 #elif TCG_TARGET_REG_BITS == 64 |
|
926 || c == INDEX_op_brcond_i64 |
|
927 #endif |
|
928 ) { |
|
929 if (args[k] < ARRAY_SIZE(cond_name) && cond_name[args[k]]) |
|
930 fprintf(outfile, ",%s", cond_name[args[k++]]); |
|
931 else |
|
932 fprintf(outfile, ",$0x%" TCG_PRIlx, args[k++]); |
|
933 i = 1; |
|
934 } |
|
935 else |
|
936 i = 0; |
|
937 for(; i < nb_cargs; i++) { |
|
938 if (k != 0) |
|
939 fprintf(outfile, ","); |
|
940 arg = args[k++]; |
|
941 fprintf(outfile, "$0x%" TCG_PRIlx, arg); |
|
942 } |
|
943 } |
|
944 fprintf(outfile, "\n"); |
|
945 args += nb_iargs + nb_oargs + nb_cargs; |
|
946 } |
|
947 } |
|
948 |
|
949 /* we give more priority to constraints with less registers */ |
|
950 static int get_constraint_priority(const TCGOpDef *def, int k) |
|
951 { |
|
952 const TCGArgConstraint *arg_ct; |
|
953 |
|
954 int i, n; |
|
955 arg_ct = &def->args_ct[k]; |
|
956 if (arg_ct->ct & TCG_CT_ALIAS) { |
|
957 /* an alias is equivalent to a single register */ |
|
958 n = 1; |
|
959 } else { |
|
960 if (!(arg_ct->ct & TCG_CT_REG)) |
|
961 return 0; |
|
962 n = 0; |
|
963 for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
|
964 if (tcg_regset_test_reg(arg_ct->u.regs, i)) |
|
965 n++; |
|
966 } |
|
967 } |
|
968 return TCG_TARGET_NB_REGS - n + 1; |
|
969 } |
|
970 |
|
971 /* sort from highest priority to lowest */ |
|
972 static void sort_constraints(TCGOpDef *def, int start, int n) |
|
973 { |
|
974 int i, j, p1, p2, tmp; |
|
975 |
|
976 for(i = 0; i < n; i++) |
|
977 def->sorted_args[start + i] = start + i; |
|
978 if (n <= 1) |
|
979 return; |
|
980 for(i = 0; i < n - 1; i++) { |
|
981 for(j = i + 1; j < n; j++) { |
|
982 p1 = get_constraint_priority(def, def->sorted_args[start + i]); |
|
983 p2 = get_constraint_priority(def, def->sorted_args[start + j]); |
|
984 if (p1 < p2) { |
|
985 tmp = def->sorted_args[start + i]; |
|
986 def->sorted_args[start + i] = def->sorted_args[start + j]; |
|
987 def->sorted_args[start + j] = tmp; |
|
988 } |
|
989 } |
|
990 } |
|
991 } |
|
992 |
|
993 void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs) |
|
994 { |
|
995 int op; |
|
996 TCGOpDef *def; |
|
997 const char *ct_str; |
|
998 int i, nb_args; |
|
999 |
|
1000 for(;;) { |
|
1001 if (tdefs->op < 0) |
|
1002 break; |
|
1003 op = tdefs->op; |
|
1004 assert(op >= 0 && op < NB_OPS); |
|
1005 def = &tcg_op_defs[op]; |
|
1006 nb_args = def->nb_iargs + def->nb_oargs; |
|
1007 for(i = 0; i < nb_args; i++) { |
|
1008 ct_str = tdefs->args_ct_str[i]; |
|
1009 tcg_regset_clear(def->args_ct[i].u.regs); |
|
1010 def->args_ct[i].ct = 0; |
|
1011 if (ct_str[0] >= '0' && ct_str[0] <= '9') { |
|
1012 int oarg; |
|
1013 oarg = ct_str[0] - '0'; |
|
1014 assert(oarg < def->nb_oargs); |
|
1015 assert(def->args_ct[oarg].ct & TCG_CT_REG); |
|
1016 /* TCG_CT_ALIAS is for the output arguments. The input |
|
1017 argument is tagged with TCG_CT_IALIAS. */ |
|
1018 def->args_ct[i] = def->args_ct[oarg]; |
|
1019 def->args_ct[oarg].ct = TCG_CT_ALIAS; |
|
1020 def->args_ct[oarg].alias_index = i; |
|
1021 def->args_ct[i].ct |= TCG_CT_IALIAS; |
|
1022 def->args_ct[i].alias_index = oarg; |
|
1023 } else { |
|
1024 for(;;) { |
|
1025 if (*ct_str == '\0') |
|
1026 break; |
|
1027 switch(*ct_str) { |
|
1028 case 'i': |
|
1029 def->args_ct[i].ct |= TCG_CT_CONST; |
|
1030 ct_str++; |
|
1031 break; |
|
1032 default: |
|
1033 if (target_parse_constraint(&def->args_ct[i], &ct_str) < 0) { |
|
1034 fprintf(stderr, "Invalid constraint '%s' for arg %d of operation '%s'\n", |
|
1035 ct_str, i, def->name); |
|
1036 exit(1); |
|
1037 } |
|
1038 } |
|
1039 } |
|
1040 } |
|
1041 } |
|
1042 |
|
1043 /* sort the constraints (XXX: this is just an heuristic) */ |
|
1044 sort_constraints(def, 0, def->nb_oargs); |
|
1045 sort_constraints(def, def->nb_oargs, def->nb_iargs); |
|
1046 |
|
1047 #if 0 |
|
1048 { |
|
1049 int i; |
|
1050 |
|
1051 printf("%s: sorted=", def->name); |
|
1052 for(i = 0; i < def->nb_oargs + def->nb_iargs; i++) |
|
1053 printf(" %d", def->sorted_args[i]); |
|
1054 printf("\n"); |
|
1055 } |
|
1056 #endif |
|
1057 tdefs++; |
|
1058 } |
|
1059 |
|
1060 } |
|
1061 |
|
1062 #ifdef USE_LIVENESS_ANALYSIS |
|
1063 |
|
1064 /* set a nop for an operation using 'nb_args' */ |
|
1065 static inline void tcg_set_nop(TCGContext *s, uint16_t *opc_ptr, |
|
1066 TCGArg *args, int nb_args) |
|
1067 { |
|
1068 if (nb_args == 0) { |
|
1069 *opc_ptr = INDEX_op_nop; |
|
1070 } else { |
|
1071 *opc_ptr = INDEX_op_nopn; |
|
1072 args[0] = nb_args; |
|
1073 args[nb_args - 1] = nb_args; |
|
1074 } |
|
1075 } |
|
1076 |
|
1077 /* liveness analysis: end of function: globals are live, temps are |
|
1078 dead. */ |
|
1079 /* XXX: at this stage, not used as there would be little gains because |
|
1080 most TBs end with a conditional jump. */ |
|
1081 static inline void tcg_la_func_end(TCGContext *s, uint8_t *dead_temps) |
|
1082 { |
|
1083 memset(dead_temps, 0, s->nb_globals); |
|
1084 memset(dead_temps + s->nb_globals, 1, s->nb_temps - s->nb_globals); |
|
1085 } |
|
1086 |
|
1087 /* liveness analysis: end of basic block: globals are live, temps are |
|
1088 dead, local temps are live. */ |
|
1089 static inline void tcg_la_bb_end(TCGContext *s, uint8_t *dead_temps) |
|
1090 { |
|
1091 int i; |
|
1092 TCGTemp *ts; |
|
1093 |
|
1094 memset(dead_temps, 0, s->nb_globals); |
|
1095 ts = &s->temps[s->nb_globals]; |
|
1096 for(i = s->nb_globals; i < s->nb_temps; i++) { |
|
1097 if (ts->temp_local) |
|
1098 dead_temps[i] = 0; |
|
1099 else |
|
1100 dead_temps[i] = 1; |
|
1101 ts++; |
|
1102 } |
|
1103 } |
|
1104 |
|
1105 /* Liveness analysis : update the opc_dead_iargs array to tell if a |
|
1106 given input arguments is dead. Instructions updating dead |
|
1107 temporaries are removed. */ |
|
1108 static void tcg_liveness_analysis(TCGContext *s) |
|
1109 { |
|
1110 int i, op_index, op, nb_args, nb_iargs, nb_oargs, arg, nb_ops; |
|
1111 TCGArg *args; |
|
1112 const TCGOpDef *def; |
|
1113 uint8_t *dead_temps; |
|
1114 unsigned int dead_iargs; |
|
1115 |
|
1116 gen_opc_ptr++; /* skip end */ |
|
1117 |
|
1118 nb_ops = gen_opc_ptr - gen_opc_buf; |
|
1119 |
|
1120 /* XXX: make it really dynamic */ |
|
1121 s->op_dead_iargs = tcg_malloc(OPC_BUF_SIZE * sizeof(uint16_t)); |
|
1122 |
|
1123 dead_temps = tcg_malloc(s->nb_temps); |
|
1124 memset(dead_temps, 1, s->nb_temps); |
|
1125 |
|
1126 args = gen_opparam_ptr; |
|
1127 op_index = nb_ops - 1; |
|
1128 while (op_index >= 0) { |
|
1129 op = gen_opc_buf[op_index]; |
|
1130 def = &tcg_op_defs[op]; |
|
1131 switch(op) { |
|
1132 case INDEX_op_call: |
|
1133 { |
|
1134 int call_flags; |
|
1135 |
|
1136 nb_args = args[-1]; |
|
1137 args -= nb_args; |
|
1138 nb_iargs = args[0] & 0xffff; |
|
1139 nb_oargs = args[0] >> 16; |
|
1140 args++; |
|
1141 call_flags = args[nb_oargs + nb_iargs]; |
|
1142 |
|
1143 /* pure functions can be removed if their result is not |
|
1144 used */ |
|
1145 if (call_flags & TCG_CALL_PURE) { |
|
1146 for(i = 0; i < nb_oargs; i++) { |
|
1147 arg = args[i]; |
|
1148 if (!dead_temps[arg]) |
|
1149 goto do_not_remove_call; |
|
1150 } |
|
1151 tcg_set_nop(s, gen_opc_buf + op_index, |
|
1152 args - 1, nb_args); |
|
1153 } else { |
|
1154 do_not_remove_call: |
|
1155 |
|
1156 /* output args are dead */ |
|
1157 for(i = 0; i < nb_oargs; i++) { |
|
1158 arg = args[i]; |
|
1159 dead_temps[arg] = 1; |
|
1160 } |
|
1161 |
|
1162 /* globals are live (they may be used by the call) */ |
|
1163 memset(dead_temps, 0, s->nb_globals); |
|
1164 |
|
1165 /* input args are live */ |
|
1166 dead_iargs = 0; |
|
1167 for(i = 0; i < nb_iargs; i++) { |
|
1168 arg = args[i + nb_oargs]; |
|
1169 if (arg != TCG_CALL_DUMMY_ARG) { |
|
1170 if (dead_temps[arg]) { |
|
1171 dead_iargs |= (1 << i); |
|
1172 } |
|
1173 dead_temps[arg] = 0; |
|
1174 } |
|
1175 } |
|
1176 s->op_dead_iargs[op_index] = dead_iargs; |
|
1177 } |
|
1178 args--; |
|
1179 } |
|
1180 break; |
|
1181 case INDEX_op_set_label: |
|
1182 args--; |
|
1183 /* mark end of basic block */ |
|
1184 tcg_la_bb_end(s, dead_temps); |
|
1185 break; |
|
1186 case INDEX_op_debug_insn_start: |
|
1187 args -= def->nb_args; |
|
1188 break; |
|
1189 case INDEX_op_nopn: |
|
1190 nb_args = args[-1]; |
|
1191 args -= nb_args; |
|
1192 break; |
|
1193 case INDEX_op_discard: |
|
1194 args--; |
|
1195 /* mark the temporary as dead */ |
|
1196 dead_temps[args[0]] = 1; |
|
1197 break; |
|
1198 case INDEX_op_end: |
|
1199 break; |
|
1200 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ |
|
1201 default: |
|
1202 args -= def->nb_args; |
|
1203 nb_iargs = def->nb_iargs; |
|
1204 nb_oargs = def->nb_oargs; |
|
1205 |
|
1206 /* Test if the operation can be removed because all |
|
1207 its outputs are dead. We assume that nb_oargs == 0 |
|
1208 implies side effects */ |
|
1209 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { |
|
1210 for(i = 0; i < nb_oargs; i++) { |
|
1211 arg = args[i]; |
|
1212 if (!dead_temps[arg]) |
|
1213 goto do_not_remove; |
|
1214 } |
|
1215 tcg_set_nop(s, gen_opc_buf + op_index, args, def->nb_args); |
|
1216 #ifdef CONFIG_PROFILER |
|
1217 s->del_op_count++; |
|
1218 #endif |
|
1219 } else { |
|
1220 do_not_remove: |
|
1221 |
|
1222 /* output args are dead */ |
|
1223 for(i = 0; i < nb_oargs; i++) { |
|
1224 arg = args[i]; |
|
1225 dead_temps[arg] = 1; |
|
1226 } |
|
1227 |
|
1228 /* if end of basic block, update */ |
|
1229 if (def->flags & TCG_OPF_BB_END) { |
|
1230 tcg_la_bb_end(s, dead_temps); |
|
1231 } else if (def->flags & TCG_OPF_CALL_CLOBBER) { |
|
1232 /* globals are live */ |
|
1233 memset(dead_temps, 0, s->nb_globals); |
|
1234 } |
|
1235 |
|
1236 /* input args are live */ |
|
1237 dead_iargs = 0; |
|
1238 for(i = 0; i < nb_iargs; i++) { |
|
1239 arg = args[i + nb_oargs]; |
|
1240 if (dead_temps[arg]) { |
|
1241 dead_iargs |= (1 << i); |
|
1242 } |
|
1243 dead_temps[arg] = 0; |
|
1244 } |
|
1245 s->op_dead_iargs[op_index] = dead_iargs; |
|
1246 } |
|
1247 break; |
|
1248 } |
|
1249 op_index--; |
|
1250 } |
|
1251 |
|
1252 if (args != gen_opparam_buf) |
|
1253 tcg_abort(); |
|
1254 } |
|
1255 #else |
|
1256 /* dummy liveness analysis */ |
|
1257 void tcg_liveness_analysis(TCGContext *s) |
|
1258 { |
|
1259 int nb_ops; |
|
1260 nb_ops = gen_opc_ptr - gen_opc_buf; |
|
1261 |
|
1262 s->op_dead_iargs = tcg_malloc(nb_ops * sizeof(uint16_t)); |
|
1263 memset(s->op_dead_iargs, 0, nb_ops * sizeof(uint16_t)); |
|
1264 } |
|
1265 #endif |
|
1266 |
|
1267 #ifndef NDEBUG |
|
1268 static void dump_regs(TCGContext *s) |
|
1269 { |
|
1270 TCGTemp *ts; |
|
1271 int i; |
|
1272 char buf[64]; |
|
1273 |
|
1274 for(i = 0; i < s->nb_temps; i++) { |
|
1275 ts = &s->temps[i]; |
|
1276 printf(" %10s: ", tcg_get_arg_str_idx(s, buf, sizeof(buf), i)); |
|
1277 switch(ts->val_type) { |
|
1278 case TEMP_VAL_REG: |
|
1279 printf("%s", tcg_target_reg_names[ts->reg]); |
|
1280 break; |
|
1281 case TEMP_VAL_MEM: |
|
1282 printf("%d(%s)", (int)ts->mem_offset, tcg_target_reg_names[ts->mem_reg]); |
|
1283 break; |
|
1284 case TEMP_VAL_CONST: |
|
1285 printf("$0x%" TCG_PRIlx, ts->val); |
|
1286 break; |
|
1287 case TEMP_VAL_DEAD: |
|
1288 printf("D"); |
|
1289 break; |
|
1290 default: |
|
1291 printf("???"); |
|
1292 break; |
|
1293 } |
|
1294 printf("\n"); |
|
1295 } |
|
1296 |
|
1297 for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
|
1298 if (s->reg_to_temp[i] >= 0) { |
|
1299 printf("%s: %s\n", |
|
1300 tcg_target_reg_names[i], |
|
1301 tcg_get_arg_str_idx(s, buf, sizeof(buf), s->reg_to_temp[i])); |
|
1302 } |
|
1303 } |
|
1304 } |
|
1305 |
|
1306 static void check_regs(TCGContext *s) |
|
1307 { |
|
1308 int reg, k; |
|
1309 TCGTemp *ts; |
|
1310 char buf[64]; |
|
1311 |
|
1312 for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
|
1313 k = s->reg_to_temp[reg]; |
|
1314 if (k >= 0) { |
|
1315 ts = &s->temps[k]; |
|
1316 if (ts->val_type != TEMP_VAL_REG || |
|
1317 ts->reg != reg) { |
|
1318 printf("Inconsistency for register %s:\n", |
|
1319 tcg_target_reg_names[reg]); |
|
1320 goto fail; |
|
1321 } |
|
1322 } |
|
1323 } |
|
1324 for(k = 0; k < s->nb_temps; k++) { |
|
1325 ts = &s->temps[k]; |
|
1326 if (ts->val_type == TEMP_VAL_REG && |
|
1327 !ts->fixed_reg && |
|
1328 s->reg_to_temp[ts->reg] != k) { |
|
1329 printf("Inconsistency for temp %s:\n", |
|
1330 tcg_get_arg_str_idx(s, buf, sizeof(buf), k)); |
|
1331 fail: |
|
1332 printf("reg state:\n"); |
|
1333 dump_regs(s); |
|
1334 tcg_abort(); |
|
1335 } |
|
1336 } |
|
1337 } |
|
1338 #endif |
|
1339 |
|
1340 static void temp_allocate_frame(TCGContext *s, int temp) |
|
1341 { |
|
1342 TCGTemp *ts; |
|
1343 ts = &s->temps[temp]; |
|
1344 s->current_frame_offset = (s->current_frame_offset + sizeof(tcg_target_long) - 1) & ~(sizeof(tcg_target_long) - 1); |
|
1345 if (s->current_frame_offset + sizeof(tcg_target_long) > s->frame_end) |
|
1346 tcg_abort(); |
|
1347 ts->mem_offset = s->current_frame_offset; |
|
1348 ts->mem_reg = s->frame_reg; |
|
1349 ts->mem_allocated = 1; |
|
1350 s->current_frame_offset += sizeof(tcg_target_long); |
|
1351 } |
|
1352 |
|
1353 /* free register 'reg' by spilling the corresponding temporary if necessary */ |
|
1354 static void tcg_reg_free(TCGContext *s, int reg) |
|
1355 { |
|
1356 TCGTemp *ts; |
|
1357 int temp; |
|
1358 |
|
1359 temp = s->reg_to_temp[reg]; |
|
1360 if (temp != -1) { |
|
1361 ts = &s->temps[temp]; |
|
1362 assert(ts->val_type == TEMP_VAL_REG); |
|
1363 if (!ts->mem_coherent) { |
|
1364 if (!ts->mem_allocated) |
|
1365 temp_allocate_frame(s, temp); |
|
1366 tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1367 } |
|
1368 ts->val_type = TEMP_VAL_MEM; |
|
1369 s->reg_to_temp[reg] = -1; |
|
1370 } |
|
1371 } |
|
1372 |
|
1373 /* Allocate a register belonging to reg1 & ~reg2 */ |
|
1374 static int tcg_reg_alloc(TCGContext *s, TCGRegSet reg1, TCGRegSet reg2) |
|
1375 { |
|
1376 int i, reg; |
|
1377 TCGRegSet reg_ct; |
|
1378 |
|
1379 tcg_regset_andnot(reg_ct, reg1, reg2); |
|
1380 |
|
1381 /* first try free registers */ |
|
1382 for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
|
1383 reg = tcg_target_reg_alloc_order[i]; |
|
1384 if (tcg_regset_test_reg(reg_ct, reg) && s->reg_to_temp[reg] == -1) |
|
1385 return reg; |
|
1386 } |
|
1387 |
|
1388 /* XXX: do better spill choice */ |
|
1389 for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
|
1390 reg = tcg_target_reg_alloc_order[i]; |
|
1391 if (tcg_regset_test_reg(reg_ct, reg)) { |
|
1392 tcg_reg_free(s, reg); |
|
1393 return reg; |
|
1394 } |
|
1395 } |
|
1396 |
|
1397 tcg_abort(); |
|
1398 } |
|
1399 |
|
1400 /* save a temporary to memory. 'allocated_regs' is used in case a |
|
1401 temporary registers needs to be allocated to store a constant. */ |
|
1402 static void temp_save(TCGContext *s, int temp, TCGRegSet allocated_regs) |
|
1403 { |
|
1404 TCGTemp *ts; |
|
1405 int reg; |
|
1406 |
|
1407 ts = &s->temps[temp]; |
|
1408 if (!ts->fixed_reg) { |
|
1409 switch(ts->val_type) { |
|
1410 case TEMP_VAL_REG: |
|
1411 tcg_reg_free(s, ts->reg); |
|
1412 break; |
|
1413 case TEMP_VAL_DEAD: |
|
1414 ts->val_type = TEMP_VAL_MEM; |
|
1415 break; |
|
1416 case TEMP_VAL_CONST: |
|
1417 reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
|
1418 allocated_regs); |
|
1419 if (!ts->mem_allocated) |
|
1420 temp_allocate_frame(s, temp); |
|
1421 tcg_out_movi(s, ts->type, reg, ts->val); |
|
1422 tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1423 ts->val_type = TEMP_VAL_MEM; |
|
1424 break; |
|
1425 case TEMP_VAL_MEM: |
|
1426 break; |
|
1427 default: |
|
1428 tcg_abort(); |
|
1429 } |
|
1430 } |
|
1431 } |
|
1432 |
|
1433 /* save globals to their cannonical location and assume they can be |
|
1434 modified be the following code. 'allocated_regs' is used in case a |
|
1435 temporary registers needs to be allocated to store a constant. */ |
|
1436 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) |
|
1437 { |
|
1438 int i; |
|
1439 |
|
1440 for(i = 0; i < s->nb_globals; i++) { |
|
1441 temp_save(s, i, allocated_regs); |
|
1442 } |
|
1443 } |
|
1444 |
|
1445 /* at the end of a basic block, we assume all temporaries are dead and |
|
1446 all globals are stored at their canonical location. */ |
|
1447 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) |
|
1448 { |
|
1449 TCGTemp *ts; |
|
1450 int i; |
|
1451 |
|
1452 for(i = s->nb_globals; i < s->nb_temps; i++) { |
|
1453 ts = &s->temps[i]; |
|
1454 if (ts->temp_local) { |
|
1455 temp_save(s, i, allocated_regs); |
|
1456 } else { |
|
1457 if (ts->val_type == TEMP_VAL_REG) { |
|
1458 s->reg_to_temp[ts->reg] = -1; |
|
1459 } |
|
1460 ts->val_type = TEMP_VAL_DEAD; |
|
1461 } |
|
1462 } |
|
1463 |
|
1464 save_globals(s, allocated_regs); |
|
1465 } |
|
1466 |
|
1467 #define IS_DEAD_IARG(n) ((dead_iargs >> (n)) & 1) |
|
1468 |
|
1469 static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args) |
|
1470 { |
|
1471 TCGTemp *ots; |
|
1472 tcg_target_ulong val; |
|
1473 |
|
1474 ots = &s->temps[args[0]]; |
|
1475 val = args[1]; |
|
1476 |
|
1477 if (ots->fixed_reg) { |
|
1478 /* for fixed registers, we do not do any constant |
|
1479 propagation */ |
|
1480 tcg_out_movi(s, ots->type, ots->reg, val); |
|
1481 } else { |
|
1482 /* The movi is not explicitly generated here */ |
|
1483 if (ots->val_type == TEMP_VAL_REG) |
|
1484 s->reg_to_temp[ots->reg] = -1; |
|
1485 ots->val_type = TEMP_VAL_CONST; |
|
1486 ots->val = val; |
|
1487 } |
|
1488 } |
|
1489 |
|
1490 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, |
|
1491 const TCGArg *args, |
|
1492 unsigned int dead_iargs) |
|
1493 { |
|
1494 TCGTemp *ts, *ots; |
|
1495 int reg; |
|
1496 const TCGArgConstraint *arg_ct; |
|
1497 |
|
1498 ots = &s->temps[args[0]]; |
|
1499 ts = &s->temps[args[1]]; |
|
1500 arg_ct = &def->args_ct[0]; |
|
1501 |
|
1502 /* XXX: always mark arg dead if IS_DEAD_IARG(0) */ |
|
1503 if (ts->val_type == TEMP_VAL_REG) { |
|
1504 if (IS_DEAD_IARG(0) && !ts->fixed_reg && !ots->fixed_reg) { |
|
1505 /* the mov can be suppressed */ |
|
1506 if (ots->val_type == TEMP_VAL_REG) |
|
1507 s->reg_to_temp[ots->reg] = -1; |
|
1508 reg = ts->reg; |
|
1509 s->reg_to_temp[reg] = -1; |
|
1510 ts->val_type = TEMP_VAL_DEAD; |
|
1511 } else { |
|
1512 if (ots->val_type == TEMP_VAL_REG) { |
|
1513 reg = ots->reg; |
|
1514 } else { |
|
1515 reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
|
1516 } |
|
1517 if (ts->reg != reg) { |
|
1518 tcg_out_mov(s, reg, ts->reg); |
|
1519 } |
|
1520 } |
|
1521 } else if (ts->val_type == TEMP_VAL_MEM) { |
|
1522 if (ots->val_type == TEMP_VAL_REG) { |
|
1523 reg = ots->reg; |
|
1524 } else { |
|
1525 reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
|
1526 } |
|
1527 tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1528 } else if (ts->val_type == TEMP_VAL_CONST) { |
|
1529 if (ots->fixed_reg) { |
|
1530 reg = ots->reg; |
|
1531 tcg_out_movi(s, ots->type, reg, ts->val); |
|
1532 } else { |
|
1533 /* propagate constant */ |
|
1534 if (ots->val_type == TEMP_VAL_REG) |
|
1535 s->reg_to_temp[ots->reg] = -1; |
|
1536 ots->val_type = TEMP_VAL_CONST; |
|
1537 ots->val = ts->val; |
|
1538 return; |
|
1539 } |
|
1540 } else { |
|
1541 tcg_abort(); |
|
1542 } |
|
1543 s->reg_to_temp[reg] = args[0]; |
|
1544 ots->reg = reg; |
|
1545 ots->val_type = TEMP_VAL_REG; |
|
1546 ots->mem_coherent = 0; |
|
1547 } |
|
1548 |
|
1549 static void tcg_reg_alloc_op(TCGContext *s, |
|
1550 const TCGOpDef *def, int opc, |
|
1551 const TCGArg *args, |
|
1552 unsigned int dead_iargs) |
|
1553 { |
|
1554 TCGRegSet allocated_regs; |
|
1555 int i, k, nb_iargs, nb_oargs, reg; |
|
1556 TCGArg arg; |
|
1557 const TCGArgConstraint *arg_ct; |
|
1558 TCGTemp *ts; |
|
1559 TCGArg new_args[TCG_MAX_OP_ARGS]; |
|
1560 int const_args[TCG_MAX_OP_ARGS]; |
|
1561 |
|
1562 nb_oargs = def->nb_oargs; |
|
1563 nb_iargs = def->nb_iargs; |
|
1564 |
|
1565 /* copy constants */ |
|
1566 memcpy(new_args + nb_oargs + nb_iargs, |
|
1567 args + nb_oargs + nb_iargs, |
|
1568 sizeof(TCGArg) * def->nb_cargs); |
|
1569 |
|
1570 /* satisfy input constraints */ |
|
1571 tcg_regset_set(allocated_regs, s->reserved_regs); |
|
1572 for(k = 0; k < nb_iargs; k++) { |
|
1573 i = def->sorted_args[nb_oargs + k]; |
|
1574 arg = args[i]; |
|
1575 arg_ct = &def->args_ct[i]; |
|
1576 ts = &s->temps[arg]; |
|
1577 if (ts->val_type == TEMP_VAL_MEM) { |
|
1578 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1579 tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1580 ts->val_type = TEMP_VAL_REG; |
|
1581 ts->reg = reg; |
|
1582 ts->mem_coherent = 1; |
|
1583 s->reg_to_temp[reg] = arg; |
|
1584 } else if (ts->val_type == TEMP_VAL_CONST) { |
|
1585 if (tcg_target_const_match(ts->val, arg_ct)) { |
|
1586 /* constant is OK for instruction */ |
|
1587 const_args[i] = 1; |
|
1588 new_args[i] = ts->val; |
|
1589 goto iarg_end; |
|
1590 } else { |
|
1591 /* need to move to a register */ |
|
1592 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1593 tcg_out_movi(s, ts->type, reg, ts->val); |
|
1594 ts->val_type = TEMP_VAL_REG; |
|
1595 ts->reg = reg; |
|
1596 ts->mem_coherent = 0; |
|
1597 s->reg_to_temp[reg] = arg; |
|
1598 } |
|
1599 } |
|
1600 assert(ts->val_type == TEMP_VAL_REG); |
|
1601 if (arg_ct->ct & TCG_CT_IALIAS) { |
|
1602 if (ts->fixed_reg) { |
|
1603 /* if fixed register, we must allocate a new register |
|
1604 if the alias is not the same register */ |
|
1605 if (arg != args[arg_ct->alias_index]) |
|
1606 goto allocate_in_reg; |
|
1607 } else { |
|
1608 /* if the input is aliased to an output and if it is |
|
1609 not dead after the instruction, we must allocate |
|
1610 a new register and move it */ |
|
1611 if (!IS_DEAD_IARG(i - nb_oargs)) |
|
1612 goto allocate_in_reg; |
|
1613 } |
|
1614 } |
|
1615 reg = ts->reg; |
|
1616 if (tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
|
1617 /* nothing to do : the constraint is satisfied */ |
|
1618 } else { |
|
1619 allocate_in_reg: |
|
1620 /* allocate a new register matching the constraint |
|
1621 and move the temporary register into it */ |
|
1622 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1623 tcg_out_mov(s, reg, ts->reg); |
|
1624 } |
|
1625 new_args[i] = reg; |
|
1626 const_args[i] = 0; |
|
1627 tcg_regset_set_reg(allocated_regs, reg); |
|
1628 iarg_end: ; |
|
1629 } |
|
1630 |
|
1631 if (def->flags & TCG_OPF_BB_END) { |
|
1632 tcg_reg_alloc_bb_end(s, allocated_regs); |
|
1633 } else { |
|
1634 /* mark dead temporaries and free the associated registers */ |
|
1635 for(i = 0; i < nb_iargs; i++) { |
|
1636 arg = args[nb_oargs + i]; |
|
1637 if (IS_DEAD_IARG(i)) { |
|
1638 ts = &s->temps[arg]; |
|
1639 if (!ts->fixed_reg) { |
|
1640 if (ts->val_type == TEMP_VAL_REG) |
|
1641 s->reg_to_temp[ts->reg] = -1; |
|
1642 ts->val_type = TEMP_VAL_DEAD; |
|
1643 } |
|
1644 } |
|
1645 } |
|
1646 |
|
1647 if (def->flags & TCG_OPF_CALL_CLOBBER) { |
|
1648 /* XXX: permit generic clobber register list ? */ |
|
1649 for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
|
1650 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) { |
|
1651 tcg_reg_free(s, reg); |
|
1652 } |
|
1653 } |
|
1654 /* XXX: for load/store we could do that only for the slow path |
|
1655 (i.e. when a memory callback is called) */ |
|
1656 |
|
1657 /* store globals and free associated registers (we assume the insn |
|
1658 can modify any global. */ |
|
1659 save_globals(s, allocated_regs); |
|
1660 } |
|
1661 |
|
1662 /* satisfy the output constraints */ |
|
1663 tcg_regset_set(allocated_regs, s->reserved_regs); |
|
1664 for(k = 0; k < nb_oargs; k++) { |
|
1665 i = def->sorted_args[k]; |
|
1666 arg = args[i]; |
|
1667 arg_ct = &def->args_ct[i]; |
|
1668 ts = &s->temps[arg]; |
|
1669 if (arg_ct->ct & TCG_CT_ALIAS) { |
|
1670 reg = new_args[arg_ct->alias_index]; |
|
1671 } else { |
|
1672 /* if fixed register, we try to use it */ |
|
1673 reg = ts->reg; |
|
1674 if (ts->fixed_reg && |
|
1675 tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
|
1676 goto oarg_end; |
|
1677 } |
|
1678 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1679 } |
|
1680 tcg_regset_set_reg(allocated_regs, reg); |
|
1681 /* if a fixed register is used, then a move will be done afterwards */ |
|
1682 if (!ts->fixed_reg) { |
|
1683 if (ts->val_type == TEMP_VAL_REG) |
|
1684 s->reg_to_temp[ts->reg] = -1; |
|
1685 ts->val_type = TEMP_VAL_REG; |
|
1686 ts->reg = reg; |
|
1687 /* temp value is modified, so the value kept in memory is |
|
1688 potentially not the same */ |
|
1689 ts->mem_coherent = 0; |
|
1690 s->reg_to_temp[reg] = arg; |
|
1691 } |
|
1692 oarg_end: |
|
1693 new_args[i] = reg; |
|
1694 } |
|
1695 } |
|
1696 |
|
1697 /* emit instruction */ |
|
1698 tcg_out_op(s, opc, new_args, const_args); |
|
1699 |
|
1700 /* move the outputs in the correct register if needed */ |
|
1701 for(i = 0; i < nb_oargs; i++) { |
|
1702 ts = &s->temps[args[i]]; |
|
1703 reg = new_args[i]; |
|
1704 if (ts->fixed_reg && ts->reg != reg) { |
|
1705 tcg_out_mov(s, ts->reg, reg); |
|
1706 } |
|
1707 } |
|
1708 } |
|
1709 |
|
1710 #ifdef TCG_TARGET_STACK_GROWSUP |
|
1711 #define STACK_DIR(x) (-(x)) |
|
1712 #else |
|
1713 #define STACK_DIR(x) (x) |
|
1714 #endif |
|
1715 |
|
1716 static int tcg_reg_alloc_call(TCGContext *s, const TCGOpDef *def, |
|
1717 int opc, const TCGArg *args, |
|
1718 unsigned int dead_iargs) |
|
1719 { |
|
1720 int nb_iargs, nb_oargs, flags, nb_regs, i, reg, nb_params; |
|
1721 TCGArg arg, func_arg; |
|
1722 TCGTemp *ts; |
|
1723 tcg_target_long stack_offset, call_stack_size, func_addr; |
|
1724 int const_func_arg, allocate_args; |
|
1725 TCGRegSet allocated_regs; |
|
1726 const TCGArgConstraint *arg_ct; |
|
1727 |
|
1728 arg = *args++; |
|
1729 |
|
1730 nb_oargs = arg >> 16; |
|
1731 nb_iargs = arg & 0xffff; |
|
1732 nb_params = nb_iargs - 1; |
|
1733 |
|
1734 flags = args[nb_oargs + nb_iargs]; |
|
1735 |
|
1736 nb_regs = tcg_target_get_call_iarg_regs_count(flags); |
|
1737 if (nb_regs > nb_params) |
|
1738 nb_regs = nb_params; |
|
1739 |
|
1740 /* assign stack slots first */ |
|
1741 /* XXX: preallocate call stack */ |
|
1742 call_stack_size = (nb_params - nb_regs) * sizeof(tcg_target_long); |
|
1743 call_stack_size = (call_stack_size + TCG_TARGET_STACK_ALIGN - 1) & |
|
1744 ~(TCG_TARGET_STACK_ALIGN - 1); |
|
1745 allocate_args = (call_stack_size > TCG_STATIC_CALL_ARGS_SIZE); |
|
1746 if (allocate_args) { |
|
1747 tcg_out_addi(s, TCG_REG_CALL_STACK, -STACK_DIR(call_stack_size)); |
|
1748 } |
|
1749 |
|
1750 stack_offset = TCG_TARGET_CALL_STACK_OFFSET; |
|
1751 for(i = nb_regs; i < nb_params; i++) { |
|
1752 arg = args[nb_oargs + i]; |
|
1753 #ifdef TCG_TARGET_STACK_GROWSUP |
|
1754 stack_offset -= sizeof(tcg_target_long); |
|
1755 #endif |
|
1756 if (arg != TCG_CALL_DUMMY_ARG) { |
|
1757 ts = &s->temps[arg]; |
|
1758 if (ts->val_type == TEMP_VAL_REG) { |
|
1759 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, stack_offset); |
|
1760 } else if (ts->val_type == TEMP_VAL_MEM) { |
|
1761 reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
|
1762 s->reserved_regs); |
|
1763 /* XXX: not correct if reading values from the stack */ |
|
1764 tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1765 tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
|
1766 } else if (ts->val_type == TEMP_VAL_CONST) { |
|
1767 reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
|
1768 s->reserved_regs); |
|
1769 /* XXX: sign extend may be needed on some targets */ |
|
1770 tcg_out_movi(s, ts->type, reg, ts->val); |
|
1771 tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
|
1772 } else { |
|
1773 tcg_abort(); |
|
1774 } |
|
1775 } |
|
1776 #ifndef TCG_TARGET_STACK_GROWSUP |
|
1777 stack_offset += sizeof(tcg_target_long); |
|
1778 #endif |
|
1779 } |
|
1780 |
|
1781 /* assign input registers */ |
|
1782 tcg_regset_set(allocated_regs, s->reserved_regs); |
|
1783 for(i = 0; i < nb_regs; i++) { |
|
1784 arg = args[nb_oargs + i]; |
|
1785 if (arg != TCG_CALL_DUMMY_ARG) { |
|
1786 ts = &s->temps[arg]; |
|
1787 reg = tcg_target_call_iarg_regs[i]; |
|
1788 tcg_reg_free(s, reg); |
|
1789 if (ts->val_type == TEMP_VAL_REG) { |
|
1790 if (ts->reg != reg) { |
|
1791 tcg_out_mov(s, reg, ts->reg); |
|
1792 } |
|
1793 } else if (ts->val_type == TEMP_VAL_MEM) { |
|
1794 tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1795 } else if (ts->val_type == TEMP_VAL_CONST) { |
|
1796 /* XXX: sign extend ? */ |
|
1797 tcg_out_movi(s, ts->type, reg, ts->val); |
|
1798 } else { |
|
1799 tcg_abort(); |
|
1800 } |
|
1801 tcg_regset_set_reg(allocated_regs, reg); |
|
1802 } |
|
1803 } |
|
1804 |
|
1805 /* assign function address */ |
|
1806 func_arg = args[nb_oargs + nb_iargs - 1]; |
|
1807 arg_ct = &def->args_ct[0]; |
|
1808 ts = &s->temps[func_arg]; |
|
1809 func_addr = ts->val; |
|
1810 const_func_arg = 0; |
|
1811 if (ts->val_type == TEMP_VAL_MEM) { |
|
1812 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1813 tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
|
1814 func_arg = reg; |
|
1815 tcg_regset_set_reg(allocated_regs, reg); |
|
1816 } else if (ts->val_type == TEMP_VAL_REG) { |
|
1817 reg = ts->reg; |
|
1818 if (!tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
|
1819 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1820 tcg_out_mov(s, reg, ts->reg); |
|
1821 } |
|
1822 func_arg = reg; |
|
1823 tcg_regset_set_reg(allocated_regs, reg); |
|
1824 } else if (ts->val_type == TEMP_VAL_CONST) { |
|
1825 if (tcg_target_const_match(func_addr, arg_ct)) { |
|
1826 const_func_arg = 1; |
|
1827 func_arg = func_addr; |
|
1828 } else { |
|
1829 reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
|
1830 tcg_out_movi(s, ts->type, reg, func_addr); |
|
1831 func_arg = reg; |
|
1832 tcg_regset_set_reg(allocated_regs, reg); |
|
1833 } |
|
1834 } else { |
|
1835 tcg_abort(); |
|
1836 } |
|
1837 |
|
1838 |
|
1839 /* mark dead temporaries and free the associated registers */ |
|
1840 for(i = 0; i < nb_iargs; i++) { |
|
1841 arg = args[nb_oargs + i]; |
|
1842 if (IS_DEAD_IARG(i)) { |
|
1843 ts = &s->temps[arg]; |
|
1844 if (!ts->fixed_reg) { |
|
1845 if (ts->val_type == TEMP_VAL_REG) |
|
1846 s->reg_to_temp[ts->reg] = -1; |
|
1847 ts->val_type = TEMP_VAL_DEAD; |
|
1848 } |
|
1849 } |
|
1850 } |
|
1851 |
|
1852 /* clobber call registers */ |
|
1853 for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
|
1854 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) { |
|
1855 tcg_reg_free(s, reg); |
|
1856 } |
|
1857 } |
|
1858 |
|
1859 /* store globals and free associated registers (we assume the call |
|
1860 can modify any global. */ |
|
1861 save_globals(s, allocated_regs); |
|
1862 |
|
1863 tcg_out_op(s, opc, &func_arg, &const_func_arg); |
|
1864 |
|
1865 if (allocate_args) { |
|
1866 tcg_out_addi(s, TCG_REG_CALL_STACK, STACK_DIR(call_stack_size)); |
|
1867 } |
|
1868 |
|
1869 /* assign output registers and emit moves if needed */ |
|
1870 for(i = 0; i < nb_oargs; i++) { |
|
1871 arg = args[i]; |
|
1872 ts = &s->temps[arg]; |
|
1873 reg = tcg_target_call_oarg_regs[i]; |
|
1874 assert(s->reg_to_temp[reg] == -1); |
|
1875 if (ts->fixed_reg) { |
|
1876 if (ts->reg != reg) { |
|
1877 tcg_out_mov(s, ts->reg, reg); |
|
1878 } |
|
1879 } else { |
|
1880 if (ts->val_type == TEMP_VAL_REG) |
|
1881 s->reg_to_temp[ts->reg] = -1; |
|
1882 ts->val_type = TEMP_VAL_REG; |
|
1883 ts->reg = reg; |
|
1884 ts->mem_coherent = 0; |
|
1885 s->reg_to_temp[reg] = arg; |
|
1886 } |
|
1887 } |
|
1888 |
|
1889 return nb_iargs + nb_oargs + def->nb_cargs + 1; |
|
1890 } |
|
1891 |
|
1892 #ifdef CONFIG_PROFILER |
|
1893 |
|
1894 static int64_t tcg_table_op_count[NB_OPS]; |
|
1895 |
|
1896 void dump_op_count(void) |
|
1897 { |
|
1898 int i; |
|
1899 FILE *f; |
|
1900 f = fopen("/tmp/op.log", "w"); |
|
1901 for(i = INDEX_op_end; i < NB_OPS; i++) { |
|
1902 fprintf(f, "%s %" PRId64 "\n", tcg_op_defs[i].name, tcg_table_op_count[i]); |
|
1903 } |
|
1904 fclose(f); |
|
1905 } |
|
1906 #endif |
|
1907 |
|
1908 |
|
1909 static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf, |
|
1910 long search_pc) |
|
1911 { |
|
1912 int opc, op_index; |
|
1913 const TCGOpDef *def; |
|
1914 unsigned int dead_iargs; |
|
1915 const TCGArg *args; |
|
1916 |
|
1917 #ifdef DEBUG_DISAS |
|
1918 if (unlikely(loglevel & CPU_LOG_TB_OP)) { |
|
1919 fprintf(logfile, "OP:\n"); |
|
1920 tcg_dump_ops(s, logfile); |
|
1921 fprintf(logfile, "\n"); |
|
1922 } |
|
1923 #endif |
|
1924 |
|
1925 #ifdef CONFIG_PROFILER |
|
1926 s->la_time -= profile_getclock(); |
|
1927 #endif |
|
1928 tcg_liveness_analysis(s); |
|
1929 #ifdef CONFIG_PROFILER |
|
1930 s->la_time += profile_getclock(); |
|
1931 #endif |
|
1932 |
|
1933 #ifdef DEBUG_DISAS |
|
1934 if (unlikely(loglevel & CPU_LOG_TB_OP_OPT)) { |
|
1935 fprintf(logfile, "OP after la:\n"); |
|
1936 tcg_dump_ops(s, logfile); |
|
1937 fprintf(logfile, "\n"); |
|
1938 } |
|
1939 #endif |
|
1940 |
|
1941 tcg_reg_alloc_start(s); |
|
1942 |
|
1943 s->code_buf = gen_code_buf; |
|
1944 s->code_ptr = gen_code_buf; |
|
1945 |
|
1946 args = gen_opparam_buf; |
|
1947 op_index = 0; |
|
1948 |
|
1949 for(;;) { |
|
1950 opc = gen_opc_buf[op_index]; |
|
1951 #ifdef CONFIG_PROFILER |
|
1952 tcg_table_op_count[opc]++; |
|
1953 #endif |
|
1954 def = &tcg_op_defs[opc]; |
|
1955 #if 0 |
|
1956 printf("%s: %d %d %d\n", def->name, |
|
1957 def->nb_oargs, def->nb_iargs, def->nb_cargs); |
|
1958 // dump_regs(s); |
|
1959 #endif |
|
1960 switch(opc) { |
|
1961 case INDEX_op_mov_i32: |
|
1962 #if TCG_TARGET_REG_BITS == 64 |
|
1963 case INDEX_op_mov_i64: |
|
1964 #endif |
|
1965 dead_iargs = s->op_dead_iargs[op_index]; |
|
1966 tcg_reg_alloc_mov(s, def, args, dead_iargs); |
|
1967 break; |
|
1968 case INDEX_op_movi_i32: |
|
1969 #if TCG_TARGET_REG_BITS == 64 |
|
1970 case INDEX_op_movi_i64: |
|
1971 #endif |
|
1972 tcg_reg_alloc_movi(s, args); |
|
1973 break; |
|
1974 case INDEX_op_debug_insn_start: |
|
1975 /* debug instruction */ |
|
1976 break; |
|
1977 case INDEX_op_nop: |
|
1978 case INDEX_op_nop1: |
|
1979 case INDEX_op_nop2: |
|
1980 case INDEX_op_nop3: |
|
1981 break; |
|
1982 case INDEX_op_nopn: |
|
1983 args += args[0]; |
|
1984 goto next; |
|
1985 case INDEX_op_discard: |
|
1986 { |
|
1987 TCGTemp *ts; |
|
1988 ts = &s->temps[args[0]]; |
|
1989 /* mark the temporary as dead */ |
|
1990 if (!ts->fixed_reg) { |
|
1991 if (ts->val_type == TEMP_VAL_REG) |
|
1992 s->reg_to_temp[ts->reg] = -1; |
|
1993 ts->val_type = TEMP_VAL_DEAD; |
|
1994 } |
|
1995 } |
|
1996 break; |
|
1997 case INDEX_op_set_label: |
|
1998 tcg_reg_alloc_bb_end(s, s->reserved_regs); |
|
1999 tcg_out_label(s, args[0], (long)s->code_ptr); |
|
2000 break; |
|
2001 case INDEX_op_call: |
|
2002 dead_iargs = s->op_dead_iargs[op_index]; |
|
2003 args += tcg_reg_alloc_call(s, def, opc, args, dead_iargs); |
|
2004 goto next; |
|
2005 case INDEX_op_end: |
|
2006 goto the_end; |
|
2007 default: |
|
2008 /* Note: in order to speed up the code, it would be much |
|
2009 faster to have specialized register allocator functions for |
|
2010 some common argument patterns */ |
|
2011 dead_iargs = s->op_dead_iargs[op_index]; |
|
2012 tcg_reg_alloc_op(s, def, opc, args, dead_iargs); |
|
2013 break; |
|
2014 } |
|
2015 args += def->nb_args; |
|
2016 next: |
|
2017 if (search_pc >= 0 && search_pc < s->code_ptr - gen_code_buf) { |
|
2018 return op_index; |
|
2019 } |
|
2020 op_index++; |
|
2021 #ifndef NDEBUG |
|
2022 check_regs(s); |
|
2023 #endif |
|
2024 } |
|
2025 the_end: |
|
2026 return -1; |
|
2027 } |
|
2028 |
|
2029 int tcg_gen_code(TCGContext *s, uint8_t *gen_code_buf) |
|
2030 { |
|
2031 #ifdef CONFIG_PROFILER |
|
2032 { |
|
2033 int n; |
|
2034 n = (gen_opc_ptr - gen_opc_buf); |
|
2035 s->op_count += n; |
|
2036 if (n > s->op_count_max) |
|
2037 s->op_count_max = n; |
|
2038 |
|
2039 s->temp_count += s->nb_temps; |
|
2040 if (s->nb_temps > s->temp_count_max) |
|
2041 s->temp_count_max = s->nb_temps; |
|
2042 } |
|
2043 #endif |
|
2044 |
|
2045 tcg_gen_code_common(s, gen_code_buf, -1); |
|
2046 |
|
2047 /* flush instruction cache */ |
|
2048 flush_icache_range((unsigned long)gen_code_buf, |
|
2049 (unsigned long)s->code_ptr); |
|
2050 return s->code_ptr - gen_code_buf; |
|
2051 } |
|
2052 |
|
2053 /* Return the index of the micro operation such as the pc after is < |
|
2054 offset bytes from the start of the TB. The contents of gen_code_buf must |
|
2055 not be changed, though writing the same values is ok. |
|
2056 Return -1 if not found. */ |
|
2057 int tcg_gen_code_search_pc(TCGContext *s, uint8_t *gen_code_buf, long offset) |
|
2058 { |
|
2059 return tcg_gen_code_common(s, gen_code_buf, offset); |
|
2060 } |
|
2061 |
|
2062 #ifdef CONFIG_PROFILER |
|
2063 void tcg_dump_info(FILE *f, |
|
2064 int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
|
2065 { |
|
2066 TCGContext *s = &tcg_ctx; |
|
2067 int64_t tot; |
|
2068 |
|
2069 tot = s->interm_time + s->code_time; |
|
2070 cpu_fprintf(f, "JIT cycles %" PRId64 " (%0.3f s at 2.4 GHz)\n", |
|
2071 tot, tot / 2.4e9); |
|
2072 cpu_fprintf(f, "translated TBs %" PRId64 " (aborted=%" PRId64 " %0.1f%%)\n", |
|
2073 s->tb_count, |
|
2074 s->tb_count1 - s->tb_count, |
|
2075 s->tb_count1 ? (double)(s->tb_count1 - s->tb_count) / s->tb_count1 * 100.0 : 0); |
|
2076 cpu_fprintf(f, "avg ops/TB %0.1f max=%d\n", |
|
2077 s->tb_count ? (double)s->op_count / s->tb_count : 0, s->op_count_max); |
|
2078 cpu_fprintf(f, "old ops/total ops %0.1f%%\n", |
|
2079 s->op_count ? (double)s->old_op_count / s->op_count * 100.0 : 0); |
|
2080 cpu_fprintf(f, "deleted ops/TB %0.2f\n", |
|
2081 s->tb_count ? |
|
2082 (double)s->del_op_count / s->tb_count : 0); |
|
2083 cpu_fprintf(f, "avg temps/TB %0.2f max=%d\n", |
|
2084 s->tb_count ? |
|
2085 (double)s->temp_count / s->tb_count : 0, |
|
2086 s->temp_count_max); |
|
2087 |
|
2088 cpu_fprintf(f, "cycles/op %0.1f\n", |
|
2089 s->op_count ? (double)tot / s->op_count : 0); |
|
2090 cpu_fprintf(f, "cycles/in byte %0.1f\n", |
|
2091 s->code_in_len ? (double)tot / s->code_in_len : 0); |
|
2092 cpu_fprintf(f, "cycles/out byte %0.1f\n", |
|
2093 s->code_out_len ? (double)tot / s->code_out_len : 0); |
|
2094 if (tot == 0) |
|
2095 tot = 1; |
|
2096 cpu_fprintf(f, " gen_interm time %0.1f%%\n", |
|
2097 (double)s->interm_time / tot * 100.0); |
|
2098 cpu_fprintf(f, " gen_code time %0.1f%%\n", |
|
2099 (double)s->code_time / tot * 100.0); |
|
2100 cpu_fprintf(f, "liveness/code time %0.1f%%\n", |
|
2101 (double)s->la_time / (s->code_time ? s->code_time : 1) * 100.0); |
|
2102 cpu_fprintf(f, "cpu_restore count %" PRId64 "\n", |
|
2103 s->restore_count); |
|
2104 cpu_fprintf(f, " avg cycles %0.1f\n", |
|
2105 s->restore_count ? (double)s->restore_time / s->restore_count : 0); |
|
2106 { |
|
2107 extern void dump_op_count(void); |
|
2108 dump_op_count(); |
|
2109 } |
|
2110 } |
|
2111 #else |
|
2112 void tcg_dump_info(FILE *f, |
|
2113 int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
|
2114 { |
|
2115 cpu_fprintf(f, "[TCG profiler not compiled]\n"); |
|
2116 } |
|
2117 #endif |