root / tcg / tcg.c @ bdffd4a9
History | View | Annotate | Download (62.7 kB)
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 |
|
41 |
#include "config.h" |
42 |
#include "qemu-common.h" |
43 |
|
44 |
/* Note: the long term plan is to reduce the dependancies on the QEMU
|
45 |
CPU definitions. Currently they are used for qemu_ld/st
|
46 |
instructions */
|
47 |
#define NO_CPU_IO_DEFS
|
48 |
#include "cpu.h" |
49 |
#include "exec-all.h" |
50 |
|
51 |
#include "tcg-op.h" |
52 |
#include "elf.h" |
53 |
|
54 |
|
55 |
static void patch_reloc(uint8_t *code_ptr, int type, |
56 |
tcg_target_long value, tcg_target_long addend); |
57 |
|
58 |
TCGOpDef tcg_op_defs[] = { |
59 |
#define DEF(s, n, copy_size) { #s, 0, 0, n, n, 0, copy_size }, |
60 |
#define DEF2(s, iargs, oargs, cargs, flags) { #s, iargs, oargs, cargs, iargs + oargs + cargs, flags, 0 }, |
61 |
#include "tcg-opc.h" |
62 |
#undef DEF
|
63 |
#undef DEF2
|
64 |
}; |
65 |
|
66 |
TCGRegSet tcg_target_available_regs[2];
|
67 |
TCGRegSet tcg_target_call_clobber_regs; |
68 |
|
69 |
/* XXX: move that inside the context */
|
70 |
uint16_t *gen_opc_ptr; |
71 |
TCGArg *gen_opparam_ptr; |
72 |
|
73 |
static inline void tcg_out8(TCGContext *s, uint8_t v) |
74 |
{ |
75 |
*s->code_ptr++ = v; |
76 |
} |
77 |
|
78 |
static inline void tcg_out16(TCGContext *s, uint16_t v) |
79 |
{ |
80 |
*(uint16_t *)s->code_ptr = v; |
81 |
s->code_ptr += 2;
|
82 |
} |
83 |
|
84 |
static inline void tcg_out32(TCGContext *s, uint32_t v) |
85 |
{ |
86 |
*(uint32_t *)s->code_ptr = v; |
87 |
s->code_ptr += 4;
|
88 |
} |
89 |
|
90 |
/* label relocation processing */
|
91 |
|
92 |
void tcg_out_reloc(TCGContext *s, uint8_t *code_ptr, int type, |
93 |
int label_index, long addend) |
94 |
{ |
95 |
TCGLabel *l; |
96 |
TCGRelocation *r; |
97 |
|
98 |
l = &s->labels[label_index]; |
99 |
if (l->has_value) {
|
100 |
/* FIXME: This may break relocations on RISC targets that
|
101 |
modify instruction fields in place. The caller may not have
|
102 |
written the initial value. */
|
103 |
patch_reloc(code_ptr, type, l->u.value, addend); |
104 |
} else {
|
105 |
/* add a new relocation entry */
|
106 |
r = tcg_malloc(sizeof(TCGRelocation));
|
107 |
r->type = type; |
108 |
r->ptr = code_ptr; |
109 |
r->addend = addend; |
110 |
r->next = l->u.first_reloc; |
111 |
l->u.first_reloc = r; |
112 |
} |
113 |
} |
114 |
|
115 |
static void tcg_out_label(TCGContext *s, int label_index, |
116 |
tcg_target_long value) |
117 |
{ |
118 |
TCGLabel *l; |
119 |
TCGRelocation *r; |
120 |
|
121 |
l = &s->labels[label_index]; |
122 |
if (l->has_value)
|
123 |
tcg_abort(); |
124 |
r = l->u.first_reloc; |
125 |
while (r != NULL) { |
126 |
patch_reloc(r->ptr, r->type, value, r->addend); |
127 |
r = r->next; |
128 |
} |
129 |
l->has_value = 1;
|
130 |
l->u.value = value; |
131 |
} |
132 |
|
133 |
int gen_new_label(void) |
134 |
{ |
135 |
TCGContext *s = &tcg_ctx; |
136 |
int idx;
|
137 |
TCGLabel *l; |
138 |
|
139 |
if (s->nb_labels >= TCG_MAX_LABELS)
|
140 |
tcg_abort(); |
141 |
idx = s->nb_labels++; |
142 |
l = &s->labels[idx]; |
143 |
l->has_value = 0;
|
144 |
l->u.first_reloc = NULL;
|
145 |
return idx;
|
146 |
} |
147 |
|
148 |
#include "tcg-target.c" |
149 |
|
150 |
/* pool based memory allocation */
|
151 |
void *tcg_malloc_internal(TCGContext *s, int size) |
152 |
{ |
153 |
TCGPool *p; |
154 |
int pool_size;
|
155 |
|
156 |
if (size > TCG_POOL_CHUNK_SIZE) {
|
157 |
/* big malloc: insert a new pool (XXX: could optimize) */
|
158 |
p = qemu_malloc(sizeof(TCGPool) + size);
|
159 |
p->size = size; |
160 |
if (s->pool_current)
|
161 |
s->pool_current->next = p; |
162 |
else
|
163 |
s->pool_first = p; |
164 |
p->next = s->pool_current; |
165 |
} else {
|
166 |
p = s->pool_current; |
167 |
if (!p) {
|
168 |
p = s->pool_first; |
169 |
if (!p)
|
170 |
goto new_pool;
|
171 |
} else {
|
172 |
if (!p->next) {
|
173 |
new_pool:
|
174 |
pool_size = TCG_POOL_CHUNK_SIZE; |
175 |
p = qemu_malloc(sizeof(TCGPool) + pool_size);
|
176 |
p->size = pool_size; |
177 |
p->next = NULL;
|
178 |
if (s->pool_current)
|
179 |
s->pool_current->next = p; |
180 |
else
|
181 |
s->pool_first = p; |
182 |
} else {
|
183 |
p = p->next; |
184 |
} |
185 |
} |
186 |
} |
187 |
s->pool_current = p; |
188 |
s->pool_cur = p->data + size; |
189 |
s->pool_end = p->data + p->size; |
190 |
return p->data;
|
191 |
} |
192 |
|
193 |
void tcg_pool_reset(TCGContext *s)
|
194 |
{ |
195 |
s->pool_cur = s->pool_end = NULL;
|
196 |
s->pool_current = NULL;
|
197 |
} |
198 |
|
199 |
void tcg_context_init(TCGContext *s)
|
200 |
{ |
201 |
int op, total_args, n;
|
202 |
TCGOpDef *def; |
203 |
TCGArgConstraint *args_ct; |
204 |
int *sorted_args;
|
205 |
|
206 |
memset(s, 0, sizeof(*s)); |
207 |
s->temps = s->static_temps; |
208 |
s->nb_globals = 0;
|
209 |
|
210 |
/* Count total number of arguments and allocate the corresponding
|
211 |
space */
|
212 |
total_args = 0;
|
213 |
for(op = 0; op < NB_OPS; op++) { |
214 |
def = &tcg_op_defs[op]; |
215 |
n = def->nb_iargs + def->nb_oargs; |
216 |
total_args += n; |
217 |
} |
218 |
|
219 |
args_ct = qemu_malloc(sizeof(TCGArgConstraint) * total_args);
|
220 |
sorted_args = qemu_malloc(sizeof(int) * total_args); |
221 |
|
222 |
for(op = 0; op < NB_OPS; op++) { |
223 |
def = &tcg_op_defs[op]; |
224 |
def->args_ct = args_ct; |
225 |
def->sorted_args = sorted_args; |
226 |
n = def->nb_iargs + def->nb_oargs; |
227 |
sorted_args += n; |
228 |
args_ct += n; |
229 |
} |
230 |
|
231 |
tcg_target_init(s); |
232 |
|
233 |
/* init global prologue and epilogue */
|
234 |
s->code_buf = code_gen_prologue; |
235 |
s->code_ptr = s->code_buf; |
236 |
tcg_target_qemu_prologue(s); |
237 |
flush_icache_range((unsigned long)s->code_buf, |
238 |
(unsigned long)s->code_ptr); |
239 |
} |
240 |
|
241 |
void tcg_set_frame(TCGContext *s, int reg, |
242 |
tcg_target_long start, tcg_target_long size) |
243 |
{ |
244 |
s->frame_start = start; |
245 |
s->frame_end = start + size; |
246 |
s->frame_reg = reg; |
247 |
} |
248 |
|
249 |
void tcg_func_start(TCGContext *s)
|
250 |
{ |
251 |
int i;
|
252 |
tcg_pool_reset(s); |
253 |
s->nb_temps = s->nb_globals; |
254 |
for(i = 0; i < (TCG_TYPE_COUNT * 2); i++) |
255 |
s->first_free_temp[i] = -1;
|
256 |
s->labels = tcg_malloc(sizeof(TCGLabel) * TCG_MAX_LABELS);
|
257 |
s->nb_labels = 0;
|
258 |
s->current_frame_offset = s->frame_start; |
259 |
|
260 |
gen_opc_ptr = gen_opc_buf; |
261 |
gen_opparam_ptr = gen_opparam_buf; |
262 |
} |
263 |
|
264 |
static inline void tcg_temp_alloc(TCGContext *s, int n) |
265 |
{ |
266 |
if (n > TCG_MAX_TEMPS)
|
267 |
tcg_abort(); |
268 |
} |
269 |
|
270 |
TCGv tcg_global_reg_new(TCGType type, int reg, const char *name) |
271 |
{ |
272 |
TCGContext *s = &tcg_ctx; |
273 |
TCGTemp *ts; |
274 |
int idx;
|
275 |
|
276 |
#if TCG_TARGET_REG_BITS == 32 |
277 |
if (type != TCG_TYPE_I32)
|
278 |
tcg_abort(); |
279 |
#endif
|
280 |
if (tcg_regset_test_reg(s->reserved_regs, reg))
|
281 |
tcg_abort(); |
282 |
idx = s->nb_globals; |
283 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
284 |
ts = &s->temps[s->nb_globals]; |
285 |
ts->base_type = type; |
286 |
ts->type = type; |
287 |
ts->fixed_reg = 1;
|
288 |
ts->reg = reg; |
289 |
ts->name = name; |
290 |
s->nb_globals++; |
291 |
tcg_regset_set_reg(s->reserved_regs, reg); |
292 |
return MAKE_TCGV(idx);
|
293 |
} |
294 |
|
295 |
#if TCG_TARGET_REG_BITS == 32 |
296 |
/* temporary hack to avoid register shortage for tcg_qemu_st64() */
|
297 |
TCGv tcg_global_reg2_new_hack(TCGType type, int reg1, int reg2, |
298 |
const char *name) |
299 |
{ |
300 |
TCGContext *s = &tcg_ctx; |
301 |
TCGTemp *ts; |
302 |
int idx;
|
303 |
char buf[64]; |
304 |
|
305 |
if (type != TCG_TYPE_I64)
|
306 |
tcg_abort(); |
307 |
idx = s->nb_globals; |
308 |
tcg_temp_alloc(s, s->nb_globals + 2);
|
309 |
ts = &s->temps[s->nb_globals]; |
310 |
ts->base_type = type; |
311 |
ts->type = TCG_TYPE_I32; |
312 |
ts->fixed_reg = 1;
|
313 |
ts->reg = reg1; |
314 |
pstrcpy(buf, sizeof(buf), name);
|
315 |
pstrcat(buf, sizeof(buf), "_0"); |
316 |
ts->name = strdup(buf); |
317 |
|
318 |
ts++; |
319 |
ts->base_type = type; |
320 |
ts->type = TCG_TYPE_I32; |
321 |
ts->fixed_reg = 1;
|
322 |
ts->reg = reg2; |
323 |
pstrcpy(buf, sizeof(buf), name);
|
324 |
pstrcat(buf, sizeof(buf), "_1"); |
325 |
ts->name = strdup(buf); |
326 |
|
327 |
s->nb_globals += 2;
|
328 |
return MAKE_TCGV(idx);
|
329 |
} |
330 |
#endif
|
331 |
|
332 |
TCGv tcg_global_mem_new(TCGType type, int reg, tcg_target_long offset,
|
333 |
const char *name) |
334 |
{ |
335 |
TCGContext *s = &tcg_ctx; |
336 |
TCGTemp *ts; |
337 |
int idx;
|
338 |
|
339 |
idx = s->nb_globals; |
340 |
#if TCG_TARGET_REG_BITS == 32 |
341 |
if (type == TCG_TYPE_I64) {
|
342 |
char buf[64]; |
343 |
tcg_temp_alloc(s, s->nb_globals + 2);
|
344 |
ts = &s->temps[s->nb_globals]; |
345 |
ts->base_type = type; |
346 |
ts->type = TCG_TYPE_I32; |
347 |
ts->fixed_reg = 0;
|
348 |
ts->mem_allocated = 1;
|
349 |
ts->mem_reg = reg; |
350 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
351 |
ts->mem_offset = offset + 4;
|
352 |
#else
|
353 |
ts->mem_offset = offset; |
354 |
#endif
|
355 |
pstrcpy(buf, sizeof(buf), name);
|
356 |
pstrcat(buf, sizeof(buf), "_0"); |
357 |
ts->name = strdup(buf); |
358 |
ts++; |
359 |
|
360 |
ts->base_type = type; |
361 |
ts->type = TCG_TYPE_I32; |
362 |
ts->fixed_reg = 0;
|
363 |
ts->mem_allocated = 1;
|
364 |
ts->mem_reg = reg; |
365 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
366 |
ts->mem_offset = offset; |
367 |
#else
|
368 |
ts->mem_offset = offset + 4;
|
369 |
#endif
|
370 |
pstrcpy(buf, sizeof(buf), name);
|
371 |
pstrcat(buf, sizeof(buf), "_1"); |
372 |
ts->name = strdup(buf); |
373 |
|
374 |
s->nb_globals += 2;
|
375 |
} else
|
376 |
#endif
|
377 |
{ |
378 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
379 |
ts = &s->temps[s->nb_globals]; |
380 |
ts->base_type = type; |
381 |
ts->type = type; |
382 |
ts->fixed_reg = 0;
|
383 |
ts->mem_allocated = 1;
|
384 |
ts->mem_reg = reg; |
385 |
ts->mem_offset = offset; |
386 |
ts->name = name; |
387 |
s->nb_globals++; |
388 |
} |
389 |
return MAKE_TCGV(idx);
|
390 |
} |
391 |
|
392 |
TCGv tcg_temp_new_internal(TCGType type, int temp_local)
|
393 |
{ |
394 |
TCGContext *s = &tcg_ctx; |
395 |
TCGTemp *ts; |
396 |
int idx, k;
|
397 |
|
398 |
k = type; |
399 |
if (temp_local)
|
400 |
k += TCG_TYPE_COUNT; |
401 |
idx = s->first_free_temp[k]; |
402 |
if (idx != -1) { |
403 |
/* There is already an available temp with the
|
404 |
right type */
|
405 |
ts = &s->temps[idx]; |
406 |
s->first_free_temp[k] = ts->next_free_temp; |
407 |
ts->temp_allocated = 1;
|
408 |
assert(ts->temp_local == temp_local); |
409 |
} else {
|
410 |
idx = s->nb_temps; |
411 |
#if TCG_TARGET_REG_BITS == 32 |
412 |
if (type == TCG_TYPE_I64) {
|
413 |
tcg_temp_alloc(s, s->nb_temps + 2);
|
414 |
ts = &s->temps[s->nb_temps]; |
415 |
ts->base_type = type; |
416 |
ts->type = TCG_TYPE_I32; |
417 |
ts->temp_allocated = 1;
|
418 |
ts->temp_local = temp_local; |
419 |
ts->name = NULL;
|
420 |
ts++; |
421 |
ts->base_type = TCG_TYPE_I32; |
422 |
ts->type = TCG_TYPE_I32; |
423 |
ts->temp_allocated = 1;
|
424 |
ts->temp_local = temp_local; |
425 |
ts->name = NULL;
|
426 |
s->nb_temps += 2;
|
427 |
} else
|
428 |
#endif
|
429 |
{ |
430 |
tcg_temp_alloc(s, s->nb_temps + 1);
|
431 |
ts = &s->temps[s->nb_temps]; |
432 |
ts->base_type = type; |
433 |
ts->type = type; |
434 |
ts->temp_allocated = 1;
|
435 |
ts->temp_local = temp_local; |
436 |
ts->name = NULL;
|
437 |
s->nb_temps++; |
438 |
} |
439 |
} |
440 |
return MAKE_TCGV(idx);
|
441 |
} |
442 |
|
443 |
void tcg_temp_free(TCGv arg)
|
444 |
{ |
445 |
TCGContext *s = &tcg_ctx; |
446 |
TCGTemp *ts; |
447 |
int idx = GET_TCGV(arg);
|
448 |
int k;
|
449 |
|
450 |
assert(idx >= s->nb_globals && idx < s->nb_temps); |
451 |
ts = &s->temps[idx]; |
452 |
assert(ts->temp_allocated != 0);
|
453 |
ts->temp_allocated = 0;
|
454 |
k = ts->base_type; |
455 |
if (ts->temp_local)
|
456 |
k += TCG_TYPE_COUNT; |
457 |
ts->next_free_temp = s->first_free_temp[k]; |
458 |
s->first_free_temp[k] = idx; |
459 |
} |
460 |
|
461 |
|
462 |
TCGv tcg_const_i32(int32_t val) |
463 |
{ |
464 |
TCGv t0; |
465 |
t0 = tcg_temp_new(TCG_TYPE_I32); |
466 |
tcg_gen_movi_i32(t0, val); |
467 |
return t0;
|
468 |
} |
469 |
|
470 |
TCGv tcg_const_i64(int64_t val) |
471 |
{ |
472 |
TCGv t0; |
473 |
t0 = tcg_temp_new(TCG_TYPE_I64); |
474 |
tcg_gen_movi_i64(t0, val); |
475 |
return t0;
|
476 |
} |
477 |
|
478 |
TCGv tcg_const_local_i32(int32_t val) |
479 |
{ |
480 |
TCGv t0; |
481 |
t0 = tcg_temp_local_new(TCG_TYPE_I32); |
482 |
tcg_gen_movi_i32(t0, val); |
483 |
return t0;
|
484 |
} |
485 |
|
486 |
TCGv tcg_const_local_i64(int64_t val) |
487 |
{ |
488 |
TCGv t0; |
489 |
t0 = tcg_temp_local_new(TCG_TYPE_I64); |
490 |
tcg_gen_movi_i64(t0, val); |
491 |
return t0;
|
492 |
} |
493 |
|
494 |
void tcg_register_helper(void *func, const char *name) |
495 |
{ |
496 |
TCGContext *s = &tcg_ctx; |
497 |
int n;
|
498 |
if ((s->nb_helpers + 1) > s->allocated_helpers) { |
499 |
n = s->allocated_helpers; |
500 |
if (n == 0) { |
501 |
n = 4;
|
502 |
} else {
|
503 |
n *= 2;
|
504 |
} |
505 |
s->helpers = realloc(s->helpers, n * sizeof(TCGHelperInfo));
|
506 |
s->allocated_helpers = n; |
507 |
} |
508 |
s->helpers[s->nb_helpers].func = (tcg_target_ulong)func; |
509 |
s->helpers[s->nb_helpers].name = name; |
510 |
s->nb_helpers++; |
511 |
} |
512 |
|
513 |
static inline TCGType tcg_get_base_type(TCGContext *s, TCGv arg) |
514 |
{ |
515 |
return s->temps[GET_TCGV(arg)].base_type;
|
516 |
} |
517 |
|
518 |
static void tcg_gen_call_internal(TCGContext *s, TCGv func, |
519 |
unsigned int flags, |
520 |
unsigned int nb_rets, const TCGv *rets, |
521 |
unsigned int nb_params, const TCGv *params) |
522 |
{ |
523 |
int i;
|
524 |
*gen_opc_ptr++ = INDEX_op_call; |
525 |
*gen_opparam_ptr++ = (nb_rets << 16) | (nb_params + 1); |
526 |
for(i = 0; i < nb_rets; i++) { |
527 |
*gen_opparam_ptr++ = GET_TCGV(rets[i]); |
528 |
} |
529 |
for(i = 0; i < nb_params; i++) { |
530 |
*gen_opparam_ptr++ = GET_TCGV(params[i]); |
531 |
} |
532 |
*gen_opparam_ptr++ = GET_TCGV(func); |
533 |
|
534 |
*gen_opparam_ptr++ = flags; |
535 |
/* total parameters, needed to go backward in the instruction stream */
|
536 |
*gen_opparam_ptr++ = 1 + nb_rets + nb_params + 3; |
537 |
} |
538 |
|
539 |
|
540 |
#if TCG_TARGET_REG_BITS < 64 |
541 |
/* Note: we convert the 64 bit args to 32 bit and do some alignment
|
542 |
and endian swap. Maybe it would be better to do the alignment
|
543 |
and endian swap in tcg_reg_alloc_call(). */
|
544 |
void tcg_gen_call(TCGContext *s, TCGv func, unsigned int flags, |
545 |
unsigned int nb_rets, const TCGv *rets, |
546 |
unsigned int nb_params, const TCGv *args1) |
547 |
{ |
548 |
TCGv ret, *args2, rets_2[2], arg;
|
549 |
int j, i, call_type;
|
550 |
|
551 |
if (nb_rets == 1) { |
552 |
ret = rets[0];
|
553 |
if (tcg_get_base_type(s, ret) == TCG_TYPE_I64) {
|
554 |
nb_rets = 2;
|
555 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
556 |
rets_2[0] = TCGV_HIGH(ret);
|
557 |
rets_2[1] = ret;
|
558 |
#else
|
559 |
rets_2[0] = ret;
|
560 |
rets_2[1] = TCGV_HIGH(ret);
|
561 |
#endif
|
562 |
rets = rets_2; |
563 |
} |
564 |
} |
565 |
args2 = alloca((nb_params * 3) * sizeof(TCGv)); |
566 |
j = 0;
|
567 |
call_type = (flags & TCG_CALL_TYPE_MASK); |
568 |
for(i = 0; i < nb_params; i++) { |
569 |
arg = args1[i]; |
570 |
if (tcg_get_base_type(s, arg) == TCG_TYPE_I64) {
|
571 |
#ifdef TCG_TARGET_I386
|
572 |
/* REGPARM case: if the third parameter is 64 bit, it is
|
573 |
allocated on the stack */
|
574 |
if (j == 2 && call_type == TCG_CALL_TYPE_REGPARM) { |
575 |
call_type = TCG_CALL_TYPE_REGPARM_2; |
576 |
flags = (flags & ~TCG_CALL_TYPE_MASK) | call_type; |
577 |
} |
578 |
args2[j++] = arg; |
579 |
args2[j++] = TCGV_HIGH(arg); |
580 |
#else
|
581 |
#ifdef TCG_TARGET_CALL_ALIGN_ARGS
|
582 |
/* some targets want aligned 64 bit args */
|
583 |
if (j & 1) { |
584 |
args2[j++] = TCG_CALL_DUMMY_ARG; |
585 |
} |
586 |
#endif
|
587 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
588 |
args2[j++] = TCGV_HIGH(arg); |
589 |
args2[j++] = arg; |
590 |
#else
|
591 |
args2[j++] = arg; |
592 |
args2[j++] = TCGV_HIGH(arg); |
593 |
#endif
|
594 |
#endif
|
595 |
} else {
|
596 |
args2[j++] = arg; |
597 |
} |
598 |
} |
599 |
tcg_gen_call_internal(s, func, flags, |
600 |
nb_rets, rets, j, args2); |
601 |
} |
602 |
#else
|
603 |
void tcg_gen_call(TCGContext *s, TCGv func, unsigned int flags, |
604 |
unsigned int nb_rets, const TCGv *rets, |
605 |
unsigned int nb_params, const TCGv *args1) |
606 |
{ |
607 |
tcg_gen_call_internal(s, func, flags, |
608 |
nb_rets, rets, nb_params, args1); |
609 |
} |
610 |
#endif
|
611 |
|
612 |
#if TCG_TARGET_REG_BITS == 32 |
613 |
void tcg_gen_shifti_i64(TCGv ret, TCGv arg1,
|
614 |
int c, int right, int arith) |
615 |
{ |
616 |
if (c == 0) { |
617 |
tcg_gen_mov_i32(ret, arg1); |
618 |
tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1)); |
619 |
} else if (c >= 32) { |
620 |
c -= 32;
|
621 |
if (right) {
|
622 |
if (arith) {
|
623 |
tcg_gen_sari_i32(ret, TCGV_HIGH(arg1), c); |
624 |
tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31);
|
625 |
} else {
|
626 |
tcg_gen_shri_i32(ret, TCGV_HIGH(arg1), c); |
627 |
tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
|
628 |
} |
629 |
} else {
|
630 |
tcg_gen_shli_i32(TCGV_HIGH(ret), arg1, c); |
631 |
tcg_gen_movi_i32(ret, 0);
|
632 |
} |
633 |
} else {
|
634 |
TCGv t0, t1; |
635 |
|
636 |
t0 = tcg_temp_new(TCG_TYPE_I32); |
637 |
t1 = tcg_temp_new(TCG_TYPE_I32); |
638 |
if (right) {
|
639 |
tcg_gen_shli_i32(t0, TCGV_HIGH(arg1), 32 - c);
|
640 |
if (arith)
|
641 |
tcg_gen_sari_i32(t1, TCGV_HIGH(arg1), c); |
642 |
else
|
643 |
tcg_gen_shri_i32(t1, TCGV_HIGH(arg1), c); |
644 |
tcg_gen_shri_i32(ret, arg1, c); |
645 |
tcg_gen_or_i32(ret, ret, t0); |
646 |
tcg_gen_mov_i32(TCGV_HIGH(ret), t1); |
647 |
} else {
|
648 |
tcg_gen_shri_i32(t0, arg1, 32 - c);
|
649 |
/* Note: ret can be the same as arg1, so we use t1 */
|
650 |
tcg_gen_shli_i32(t1, arg1, c); |
651 |
tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c); |
652 |
tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t0); |
653 |
tcg_gen_mov_i32(ret, t1); |
654 |
} |
655 |
tcg_temp_free(t0); |
656 |
tcg_temp_free(t1); |
657 |
} |
658 |
} |
659 |
#endif
|
660 |
|
661 |
static void tcg_reg_alloc_start(TCGContext *s) |
662 |
{ |
663 |
int i;
|
664 |
TCGTemp *ts; |
665 |
for(i = 0; i < s->nb_globals; i++) { |
666 |
ts = &s->temps[i]; |
667 |
if (ts->fixed_reg) {
|
668 |
ts->val_type = TEMP_VAL_REG; |
669 |
} else {
|
670 |
ts->val_type = TEMP_VAL_MEM; |
671 |
} |
672 |
} |
673 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
674 |
ts = &s->temps[i]; |
675 |
ts->val_type = TEMP_VAL_DEAD; |
676 |
ts->mem_allocated = 0;
|
677 |
ts->fixed_reg = 0;
|
678 |
} |
679 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
680 |
s->reg_to_temp[i] = -1;
|
681 |
} |
682 |
} |
683 |
|
684 |
static char *tcg_get_arg_str_idx(TCGContext *s, char *buf, int buf_size, |
685 |
int idx)
|
686 |
{ |
687 |
TCGTemp *ts; |
688 |
|
689 |
ts = &s->temps[idx]; |
690 |
if (idx < s->nb_globals) {
|
691 |
pstrcpy(buf, buf_size, ts->name); |
692 |
} else {
|
693 |
if (ts->temp_local)
|
694 |
snprintf(buf, buf_size, "loc%d", idx - s->nb_globals);
|
695 |
else
|
696 |
snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals);
|
697 |
} |
698 |
return buf;
|
699 |
} |
700 |
|
701 |
char *tcg_get_arg_str(TCGContext *s, char *buf, int buf_size, TCGv arg) |
702 |
{ |
703 |
return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV(arg));
|
704 |
} |
705 |
|
706 |
static int helper_cmp(const void *p1, const void *p2) |
707 |
{ |
708 |
const TCGHelperInfo *th1 = p1;
|
709 |
const TCGHelperInfo *th2 = p2;
|
710 |
if (th1->func < th2->func)
|
711 |
return -1; |
712 |
else if (th1->func == th2->func) |
713 |
return 0; |
714 |
else
|
715 |
return 1; |
716 |
} |
717 |
|
718 |
/* find helper definition (Note: A hash table would be better) */
|
719 |
static TCGHelperInfo *tcg_find_helper(TCGContext *s, tcg_target_ulong val)
|
720 |
{ |
721 |
int m, m_min, m_max;
|
722 |
TCGHelperInfo *th; |
723 |
tcg_target_ulong v; |
724 |
|
725 |
if (unlikely(!s->helpers_sorted)) {
|
726 |
qsort(s->helpers, s->nb_helpers, sizeof(TCGHelperInfo),
|
727 |
helper_cmp); |
728 |
s->helpers_sorted = 1;
|
729 |
} |
730 |
|
731 |
/* binary search */
|
732 |
m_min = 0;
|
733 |
m_max = s->nb_helpers - 1;
|
734 |
while (m_min <= m_max) {
|
735 |
m = (m_min + m_max) >> 1;
|
736 |
th = &s->helpers[m]; |
737 |
v = th->func; |
738 |
if (v == val)
|
739 |
return th;
|
740 |
else if (val < v) { |
741 |
m_max = m - 1;
|
742 |
} else {
|
743 |
m_min = m + 1;
|
744 |
} |
745 |
} |
746 |
return NULL; |
747 |
} |
748 |
|
749 |
static const char * const cond_name[] = |
750 |
{ |
751 |
[TCG_COND_EQ] = "eq",
|
752 |
[TCG_COND_NE] = "ne",
|
753 |
[TCG_COND_LT] = "lt",
|
754 |
[TCG_COND_GE] = "ge",
|
755 |
[TCG_COND_LE] = "le",
|
756 |
[TCG_COND_GT] = "gt",
|
757 |
[TCG_COND_LTU] = "ltu",
|
758 |
[TCG_COND_GEU] = "geu",
|
759 |
[TCG_COND_LEU] = "leu",
|
760 |
[TCG_COND_GTU] = "gtu"
|
761 |
}; |
762 |
|
763 |
void tcg_dump_ops(TCGContext *s, FILE *outfile)
|
764 |
{ |
765 |
const uint16_t *opc_ptr;
|
766 |
const TCGArg *args;
|
767 |
TCGArg arg; |
768 |
int c, i, k, nb_oargs, nb_iargs, nb_cargs, first_insn;
|
769 |
const TCGOpDef *def;
|
770 |
char buf[128]; |
771 |
|
772 |
first_insn = 1;
|
773 |
opc_ptr = gen_opc_buf; |
774 |
args = gen_opparam_buf; |
775 |
while (opc_ptr < gen_opc_ptr) {
|
776 |
c = *opc_ptr++; |
777 |
def = &tcg_op_defs[c]; |
778 |
if (c == INDEX_op_debug_insn_start) {
|
779 |
uint64_t pc; |
780 |
#if TARGET_LONG_BITS > TCG_TARGET_REG_BITS
|
781 |
pc = ((uint64_t)args[1] << 32) | args[0]; |
782 |
#else
|
783 |
pc = args[0];
|
784 |
#endif
|
785 |
if (!first_insn)
|
786 |
fprintf(outfile, "\n");
|
787 |
fprintf(outfile, " ---- 0x%" PRIx64, pc);
|
788 |
first_insn = 0;
|
789 |
nb_oargs = def->nb_oargs; |
790 |
nb_iargs = def->nb_iargs; |
791 |
nb_cargs = def->nb_cargs; |
792 |
} else if (c == INDEX_op_call) { |
793 |
TCGArg arg; |
794 |
|
795 |
/* variable number of arguments */
|
796 |
arg = *args++; |
797 |
nb_oargs = arg >> 16;
|
798 |
nb_iargs = arg & 0xffff;
|
799 |
nb_cargs = def->nb_cargs; |
800 |
|
801 |
fprintf(outfile, " %s ", def->name);
|
802 |
|
803 |
/* function name */
|
804 |
fprintf(outfile, "%s",
|
805 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + nb_iargs - 1])); |
806 |
/* flags */
|
807 |
fprintf(outfile, ",$0x%" TCG_PRIlx,
|
808 |
args[nb_oargs + nb_iargs]); |
809 |
/* nb out args */
|
810 |
fprintf(outfile, ",$%d", nb_oargs);
|
811 |
for(i = 0; i < nb_oargs; i++) { |
812 |
fprintf(outfile, ",");
|
813 |
fprintf(outfile, "%s",
|
814 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[i]));
|
815 |
} |
816 |
for(i = 0; i < (nb_iargs - 1); i++) { |
817 |
fprintf(outfile, ",");
|
818 |
if (args[nb_oargs + i] == TCG_CALL_DUMMY_ARG) {
|
819 |
fprintf(outfile, "<dummy>");
|
820 |
} else {
|
821 |
fprintf(outfile, "%s",
|
822 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + i]));
|
823 |
} |
824 |
} |
825 |
} else if (c == INDEX_op_movi_i32 |
826 |
#if TCG_TARGET_REG_BITS == 64 |
827 |
|| c == INDEX_op_movi_i64 |
828 |
#endif
|
829 |
) { |
830 |
tcg_target_ulong val; |
831 |
TCGHelperInfo *th; |
832 |
|
833 |
nb_oargs = def->nb_oargs; |
834 |
nb_iargs = def->nb_iargs; |
835 |
nb_cargs = def->nb_cargs; |
836 |
fprintf(outfile, " %s %s,$", def->name,
|
837 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[0])); |
838 |
val = args[1];
|
839 |
th = tcg_find_helper(s, val); |
840 |
if (th) {
|
841 |
fprintf(outfile, th->name); |
842 |
} else {
|
843 |
if (c == INDEX_op_movi_i32)
|
844 |
fprintf(outfile, "0x%x", (uint32_t)val);
|
845 |
else
|
846 |
fprintf(outfile, "0x%" PRIx64 , (uint64_t)val);
|
847 |
} |
848 |
} else {
|
849 |
fprintf(outfile, " %s ", def->name);
|
850 |
if (c == INDEX_op_nopn) {
|
851 |
/* variable number of arguments */
|
852 |
nb_cargs = *args; |
853 |
nb_oargs = 0;
|
854 |
nb_iargs = 0;
|
855 |
} else {
|
856 |
nb_oargs = def->nb_oargs; |
857 |
nb_iargs = def->nb_iargs; |
858 |
nb_cargs = def->nb_cargs; |
859 |
} |
860 |
|
861 |
k = 0;
|
862 |
for(i = 0; i < nb_oargs; i++) { |
863 |
if (k != 0) |
864 |
fprintf(outfile, ",");
|
865 |
fprintf(outfile, "%s",
|
866 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++]));
|
867 |
} |
868 |
for(i = 0; i < nb_iargs; i++) { |
869 |
if (k != 0) |
870 |
fprintf(outfile, ",");
|
871 |
fprintf(outfile, "%s",
|
872 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++]));
|
873 |
} |
874 |
if (c == INDEX_op_brcond_i32
|
875 |
#if TCG_TARGET_REG_BITS == 32 |
876 |
|| c == INDEX_op_brcond2_i32 |
877 |
#elif TCG_TARGET_REG_BITS == 64 |
878 |
|| c == INDEX_op_brcond_i64 |
879 |
#endif
|
880 |
) { |
881 |
if (args[k] < ARRAY_SIZE(cond_name) && cond_name[args[k]])
|
882 |
fprintf(outfile, ",%s", cond_name[args[k++]]);
|
883 |
else
|
884 |
fprintf(outfile, ",$0x%" TCG_PRIlx, args[k++]);
|
885 |
i = 1;
|
886 |
} |
887 |
else
|
888 |
i = 0;
|
889 |
for(; i < nb_cargs; i++) {
|
890 |
if (k != 0) |
891 |
fprintf(outfile, ",");
|
892 |
arg = args[k++]; |
893 |
fprintf(outfile, "$0x%" TCG_PRIlx, arg);
|
894 |
} |
895 |
} |
896 |
fprintf(outfile, "\n");
|
897 |
args += nb_iargs + nb_oargs + nb_cargs; |
898 |
} |
899 |
} |
900 |
|
901 |
/* we give more priority to constraints with less registers */
|
902 |
static int get_constraint_priority(const TCGOpDef *def, int k) |
903 |
{ |
904 |
const TCGArgConstraint *arg_ct;
|
905 |
|
906 |
int i, n;
|
907 |
arg_ct = &def->args_ct[k]; |
908 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
909 |
/* an alias is equivalent to a single register */
|
910 |
n = 1;
|
911 |
} else {
|
912 |
if (!(arg_ct->ct & TCG_CT_REG))
|
913 |
return 0; |
914 |
n = 0;
|
915 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
916 |
if (tcg_regset_test_reg(arg_ct->u.regs, i))
|
917 |
n++; |
918 |
} |
919 |
} |
920 |
return TCG_TARGET_NB_REGS - n + 1; |
921 |
} |
922 |
|
923 |
/* sort from highest priority to lowest */
|
924 |
static void sort_constraints(TCGOpDef *def, int start, int n) |
925 |
{ |
926 |
int i, j, p1, p2, tmp;
|
927 |
|
928 |
for(i = 0; i < n; i++) |
929 |
def->sorted_args[start + i] = start + i; |
930 |
if (n <= 1) |
931 |
return;
|
932 |
for(i = 0; i < n - 1; i++) { |
933 |
for(j = i + 1; j < n; j++) { |
934 |
p1 = get_constraint_priority(def, def->sorted_args[start + i]); |
935 |
p2 = get_constraint_priority(def, def->sorted_args[start + j]); |
936 |
if (p1 < p2) {
|
937 |
tmp = def->sorted_args[start + i]; |
938 |
def->sorted_args[start + i] = def->sorted_args[start + j]; |
939 |
def->sorted_args[start + j] = tmp; |
940 |
} |
941 |
} |
942 |
} |
943 |
} |
944 |
|
945 |
void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs) |
946 |
{ |
947 |
int op;
|
948 |
TCGOpDef *def; |
949 |
const char *ct_str; |
950 |
int i, nb_args;
|
951 |
|
952 |
for(;;) {
|
953 |
if (tdefs->op < 0) |
954 |
break;
|
955 |
op = tdefs->op; |
956 |
assert(op >= 0 && op < NB_OPS);
|
957 |
def = &tcg_op_defs[op]; |
958 |
nb_args = def->nb_iargs + def->nb_oargs; |
959 |
for(i = 0; i < nb_args; i++) { |
960 |
ct_str = tdefs->args_ct_str[i]; |
961 |
tcg_regset_clear(def->args_ct[i].u.regs); |
962 |
def->args_ct[i].ct = 0;
|
963 |
if (ct_str[0] >= '0' && ct_str[0] <= '9') { |
964 |
int oarg;
|
965 |
oarg = ct_str[0] - '0'; |
966 |
assert(oarg < def->nb_oargs); |
967 |
assert(def->args_ct[oarg].ct & TCG_CT_REG); |
968 |
/* TCG_CT_ALIAS is for the output arguments. The input
|
969 |
argument is tagged with TCG_CT_IALIAS. */
|
970 |
def->args_ct[i] = def->args_ct[oarg]; |
971 |
def->args_ct[oarg].ct = TCG_CT_ALIAS; |
972 |
def->args_ct[oarg].alias_index = i; |
973 |
def->args_ct[i].ct |= TCG_CT_IALIAS; |
974 |
def->args_ct[i].alias_index = oarg; |
975 |
} else {
|
976 |
for(;;) {
|
977 |
if (*ct_str == '\0') |
978 |
break;
|
979 |
switch(*ct_str) {
|
980 |
case 'i': |
981 |
def->args_ct[i].ct |= TCG_CT_CONST; |
982 |
ct_str++; |
983 |
break;
|
984 |
default:
|
985 |
if (target_parse_constraint(&def->args_ct[i], &ct_str) < 0) { |
986 |
fprintf(stderr, "Invalid constraint '%s' for arg %d of operation '%s'\n",
|
987 |
ct_str, i, def->name); |
988 |
exit(1);
|
989 |
} |
990 |
} |
991 |
} |
992 |
} |
993 |
} |
994 |
|
995 |
/* sort the constraints (XXX: this is just an heuristic) */
|
996 |
sort_constraints(def, 0, def->nb_oargs);
|
997 |
sort_constraints(def, def->nb_oargs, def->nb_iargs); |
998 |
|
999 |
#if 0
|
1000 |
{
|
1001 |
int i;
|
1002 |
|
1003 |
printf("%s: sorted=", def->name);
|
1004 |
for(i = 0; i < def->nb_oargs + def->nb_iargs; i++)
|
1005 |
printf(" %d", def->sorted_args[i]);
|
1006 |
printf("\n");
|
1007 |
}
|
1008 |
#endif
|
1009 |
tdefs++; |
1010 |
} |
1011 |
|
1012 |
} |
1013 |
|
1014 |
#ifdef USE_LIVENESS_ANALYSIS
|
1015 |
|
1016 |
/* set a nop for an operation using 'nb_args' */
|
1017 |
static inline void tcg_set_nop(TCGContext *s, uint16_t *opc_ptr, |
1018 |
TCGArg *args, int nb_args)
|
1019 |
{ |
1020 |
if (nb_args == 0) { |
1021 |
*opc_ptr = INDEX_op_nop; |
1022 |
} else {
|
1023 |
*opc_ptr = INDEX_op_nopn; |
1024 |
args[0] = nb_args;
|
1025 |
args[nb_args - 1] = nb_args;
|
1026 |
} |
1027 |
} |
1028 |
|
1029 |
/* liveness analysis: end of function: globals are live, temps are
|
1030 |
dead. */
|
1031 |
/* XXX: at this stage, not used as there would be little gains because
|
1032 |
most TBs end with a conditional jump. */
|
1033 |
static inline void tcg_la_func_end(TCGContext *s, uint8_t *dead_temps) |
1034 |
{ |
1035 |
memset(dead_temps, 0, s->nb_globals);
|
1036 |
memset(dead_temps + s->nb_globals, 1, s->nb_temps - s->nb_globals);
|
1037 |
} |
1038 |
|
1039 |
/* liveness analysis: end of basic block: globals are live, temps are
|
1040 |
dead, local temps are live. */
|
1041 |
static inline void tcg_la_bb_end(TCGContext *s, uint8_t *dead_temps) |
1042 |
{ |
1043 |
int i;
|
1044 |
TCGTemp *ts; |
1045 |
|
1046 |
memset(dead_temps, 0, s->nb_globals);
|
1047 |
ts = &s->temps[s->nb_globals]; |
1048 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1049 |
if (ts->temp_local)
|
1050 |
dead_temps[i] = 0;
|
1051 |
else
|
1052 |
dead_temps[i] = 1;
|
1053 |
ts++; |
1054 |
} |
1055 |
} |
1056 |
|
1057 |
/* Liveness analysis : update the opc_dead_iargs array to tell if a
|
1058 |
given input arguments is dead. Instructions updating dead
|
1059 |
temporaries are removed. */
|
1060 |
static void tcg_liveness_analysis(TCGContext *s) |
1061 |
{ |
1062 |
int i, op_index, op, nb_args, nb_iargs, nb_oargs, arg, nb_ops;
|
1063 |
TCGArg *args; |
1064 |
const TCGOpDef *def;
|
1065 |
uint8_t *dead_temps; |
1066 |
unsigned int dead_iargs; |
1067 |
|
1068 |
gen_opc_ptr++; /* skip end */
|
1069 |
|
1070 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1071 |
|
1072 |
/* XXX: make it really dynamic */
|
1073 |
s->op_dead_iargs = tcg_malloc(OPC_BUF_SIZE * sizeof(uint16_t));
|
1074 |
|
1075 |
dead_temps = tcg_malloc(s->nb_temps); |
1076 |
memset(dead_temps, 1, s->nb_temps);
|
1077 |
|
1078 |
args = gen_opparam_ptr; |
1079 |
op_index = nb_ops - 1;
|
1080 |
while (op_index >= 0) { |
1081 |
op = gen_opc_buf[op_index]; |
1082 |
def = &tcg_op_defs[op]; |
1083 |
switch(op) {
|
1084 |
case INDEX_op_call:
|
1085 |
{ |
1086 |
int call_flags;
|
1087 |
|
1088 |
nb_args = args[-1];
|
1089 |
args -= nb_args; |
1090 |
nb_iargs = args[0] & 0xffff; |
1091 |
nb_oargs = args[0] >> 16; |
1092 |
args++; |
1093 |
call_flags = args[nb_oargs + nb_iargs]; |
1094 |
|
1095 |
/* pure functions can be removed if their result is not
|
1096 |
used */
|
1097 |
if (call_flags & TCG_CALL_PURE) {
|
1098 |
for(i = 0; i < nb_oargs; i++) { |
1099 |
arg = args[i]; |
1100 |
if (!dead_temps[arg])
|
1101 |
goto do_not_remove_call;
|
1102 |
} |
1103 |
tcg_set_nop(s, gen_opc_buf + op_index, |
1104 |
args - 1, nb_args);
|
1105 |
} else {
|
1106 |
do_not_remove_call:
|
1107 |
|
1108 |
/* output args are dead */
|
1109 |
for(i = 0; i < nb_oargs; i++) { |
1110 |
arg = args[i]; |
1111 |
dead_temps[arg] = 1;
|
1112 |
} |
1113 |
|
1114 |
/* globals are live (they may be used by the call) */
|
1115 |
memset(dead_temps, 0, s->nb_globals);
|
1116 |
|
1117 |
/* input args are live */
|
1118 |
dead_iargs = 0;
|
1119 |
for(i = 0; i < nb_iargs; i++) { |
1120 |
arg = args[i + nb_oargs]; |
1121 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1122 |
if (dead_temps[arg]) {
|
1123 |
dead_iargs |= (1 << i);
|
1124 |
} |
1125 |
dead_temps[arg] = 0;
|
1126 |
} |
1127 |
} |
1128 |
s->op_dead_iargs[op_index] = dead_iargs; |
1129 |
} |
1130 |
args--; |
1131 |
} |
1132 |
break;
|
1133 |
case INDEX_op_set_label:
|
1134 |
args--; |
1135 |
/* mark end of basic block */
|
1136 |
tcg_la_bb_end(s, dead_temps); |
1137 |
break;
|
1138 |
case INDEX_op_debug_insn_start:
|
1139 |
args -= def->nb_args; |
1140 |
break;
|
1141 |
case INDEX_op_nopn:
|
1142 |
nb_args = args[-1];
|
1143 |
args -= nb_args; |
1144 |
break;
|
1145 |
case INDEX_op_discard:
|
1146 |
args--; |
1147 |
/* mark the temporary as dead */
|
1148 |
dead_temps[args[0]] = 1; |
1149 |
break;
|
1150 |
case INDEX_op_end:
|
1151 |
break;
|
1152 |
/* XXX: optimize by hardcoding common cases (e.g. triadic ops) */
|
1153 |
default:
|
1154 |
if (op > INDEX_op_end) {
|
1155 |
args -= def->nb_args; |
1156 |
nb_iargs = def->nb_iargs; |
1157 |
nb_oargs = def->nb_oargs; |
1158 |
|
1159 |
/* Test if the operation can be removed because all
|
1160 |
its outputs are dead. We assume that nb_oargs == 0
|
1161 |
implies side effects */
|
1162 |
if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { |
1163 |
for(i = 0; i < nb_oargs; i++) { |
1164 |
arg = args[i]; |
1165 |
if (!dead_temps[arg])
|
1166 |
goto do_not_remove;
|
1167 |
} |
1168 |
tcg_set_nop(s, gen_opc_buf + op_index, args, def->nb_args); |
1169 |
#ifdef CONFIG_PROFILER
|
1170 |
s->del_op_count++; |
1171 |
#endif
|
1172 |
} else {
|
1173 |
do_not_remove:
|
1174 |
|
1175 |
/* output args are dead */
|
1176 |
for(i = 0; i < nb_oargs; i++) { |
1177 |
arg = args[i]; |
1178 |
dead_temps[arg] = 1;
|
1179 |
} |
1180 |
|
1181 |
/* if end of basic block, update */
|
1182 |
if (def->flags & TCG_OPF_BB_END) {
|
1183 |
tcg_la_bb_end(s, dead_temps); |
1184 |
} else if (def->flags & TCG_OPF_CALL_CLOBBER) { |
1185 |
/* globals are live */
|
1186 |
memset(dead_temps, 0, s->nb_globals);
|
1187 |
} |
1188 |
|
1189 |
/* input args are live */
|
1190 |
dead_iargs = 0;
|
1191 |
for(i = 0; i < nb_iargs; i++) { |
1192 |
arg = args[i + nb_oargs]; |
1193 |
if (dead_temps[arg]) {
|
1194 |
dead_iargs |= (1 << i);
|
1195 |
} |
1196 |
dead_temps[arg] = 0;
|
1197 |
} |
1198 |
s->op_dead_iargs[op_index] = dead_iargs; |
1199 |
} |
1200 |
} else {
|
1201 |
/* legacy dyngen operations */
|
1202 |
args -= def->nb_args; |
1203 |
/* mark end of basic block */
|
1204 |
tcg_la_bb_end(s, dead_temps); |
1205 |
} |
1206 |
break;
|
1207 |
} |
1208 |
op_index--; |
1209 |
} |
1210 |
|
1211 |
if (args != gen_opparam_buf)
|
1212 |
tcg_abort(); |
1213 |
} |
1214 |
#else
|
1215 |
/* dummy liveness analysis */
|
1216 |
void tcg_liveness_analysis(TCGContext *s)
|
1217 |
{ |
1218 |
int nb_ops;
|
1219 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1220 |
|
1221 |
s->op_dead_iargs = tcg_malloc(nb_ops * sizeof(uint16_t));
|
1222 |
memset(s->op_dead_iargs, 0, nb_ops * sizeof(uint16_t)); |
1223 |
} |
1224 |
#endif
|
1225 |
|
1226 |
#ifndef NDEBUG
|
1227 |
static void dump_regs(TCGContext *s) |
1228 |
{ |
1229 |
TCGTemp *ts; |
1230 |
int i;
|
1231 |
char buf[64]; |
1232 |
|
1233 |
for(i = 0; i < s->nb_temps; i++) { |
1234 |
ts = &s->temps[i]; |
1235 |
printf(" %10s: ", tcg_get_arg_str_idx(s, buf, sizeof(buf), i)); |
1236 |
switch(ts->val_type) {
|
1237 |
case TEMP_VAL_REG:
|
1238 |
printf("%s", tcg_target_reg_names[ts->reg]);
|
1239 |
break;
|
1240 |
case TEMP_VAL_MEM:
|
1241 |
printf("%d(%s)", (int)ts->mem_offset, tcg_target_reg_names[ts->mem_reg]); |
1242 |
break;
|
1243 |
case TEMP_VAL_CONST:
|
1244 |
printf("$0x%" TCG_PRIlx, ts->val);
|
1245 |
break;
|
1246 |
case TEMP_VAL_DEAD:
|
1247 |
printf("D");
|
1248 |
break;
|
1249 |
default:
|
1250 |
printf("???");
|
1251 |
break;
|
1252 |
} |
1253 |
printf("\n");
|
1254 |
} |
1255 |
|
1256 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
1257 |
if (s->reg_to_temp[i] >= 0) { |
1258 |
printf("%s: %s\n",
|
1259 |
tcg_target_reg_names[i], |
1260 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), s->reg_to_temp[i]));
|
1261 |
} |
1262 |
} |
1263 |
} |
1264 |
|
1265 |
static void check_regs(TCGContext *s) |
1266 |
{ |
1267 |
int reg, k;
|
1268 |
TCGTemp *ts; |
1269 |
char buf[64]; |
1270 |
|
1271 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1272 |
k = s->reg_to_temp[reg]; |
1273 |
if (k >= 0) { |
1274 |
ts = &s->temps[k]; |
1275 |
if (ts->val_type != TEMP_VAL_REG ||
|
1276 |
ts->reg != reg) { |
1277 |
printf("Inconsistency for register %s:\n",
|
1278 |
tcg_target_reg_names[reg]); |
1279 |
goto fail;
|
1280 |
} |
1281 |
} |
1282 |
} |
1283 |
for(k = 0; k < s->nb_temps; k++) { |
1284 |
ts = &s->temps[k]; |
1285 |
if (ts->val_type == TEMP_VAL_REG &&
|
1286 |
!ts->fixed_reg && |
1287 |
s->reg_to_temp[ts->reg] != k) { |
1288 |
printf("Inconsistency for temp %s:\n",
|
1289 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), k));
|
1290 |
fail:
|
1291 |
printf("reg state:\n");
|
1292 |
dump_regs(s); |
1293 |
tcg_abort(); |
1294 |
} |
1295 |
} |
1296 |
} |
1297 |
#endif
|
1298 |
|
1299 |
static void temp_allocate_frame(TCGContext *s, int temp) |
1300 |
{ |
1301 |
TCGTemp *ts; |
1302 |
ts = &s->temps[temp]; |
1303 |
s->current_frame_offset = (s->current_frame_offset + sizeof(tcg_target_long) - 1) & ~(sizeof(tcg_target_long) - 1); |
1304 |
if (s->current_frame_offset + sizeof(tcg_target_long) > s->frame_end) |
1305 |
tcg_abort(); |
1306 |
ts->mem_offset = s->current_frame_offset; |
1307 |
ts->mem_reg = s->frame_reg; |
1308 |
ts->mem_allocated = 1;
|
1309 |
s->current_frame_offset += sizeof(tcg_target_long);
|
1310 |
} |
1311 |
|
1312 |
/* free register 'reg' by spilling the corresponding temporary if necessary */
|
1313 |
static void tcg_reg_free(TCGContext *s, int reg) |
1314 |
{ |
1315 |
TCGTemp *ts; |
1316 |
int temp;
|
1317 |
|
1318 |
temp = s->reg_to_temp[reg]; |
1319 |
if (temp != -1) { |
1320 |
ts = &s->temps[temp]; |
1321 |
assert(ts->val_type == TEMP_VAL_REG); |
1322 |
if (!ts->mem_coherent) {
|
1323 |
if (!ts->mem_allocated)
|
1324 |
temp_allocate_frame(s, temp); |
1325 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1326 |
} |
1327 |
ts->val_type = TEMP_VAL_MEM; |
1328 |
s->reg_to_temp[reg] = -1;
|
1329 |
} |
1330 |
} |
1331 |
|
1332 |
/* Allocate a register belonging to reg1 & ~reg2 */
|
1333 |
static int tcg_reg_alloc(TCGContext *s, TCGRegSet reg1, TCGRegSet reg2) |
1334 |
{ |
1335 |
int i, reg;
|
1336 |
TCGRegSet reg_ct; |
1337 |
|
1338 |
tcg_regset_andnot(reg_ct, reg1, reg2); |
1339 |
|
1340 |
/* first try free registers */
|
1341 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1342 |
reg = tcg_target_reg_alloc_order[i]; |
1343 |
if (tcg_regset_test_reg(reg_ct, reg) && s->reg_to_temp[reg] == -1) |
1344 |
return reg;
|
1345 |
} |
1346 |
|
1347 |
/* XXX: do better spill choice */
|
1348 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1349 |
reg = tcg_target_reg_alloc_order[i]; |
1350 |
if (tcg_regset_test_reg(reg_ct, reg)) {
|
1351 |
tcg_reg_free(s, reg); |
1352 |
return reg;
|
1353 |
} |
1354 |
} |
1355 |
|
1356 |
tcg_abort(); |
1357 |
} |
1358 |
|
1359 |
/* save a temporary to memory. 'allocated_regs' is used in case a
|
1360 |
temporary registers needs to be allocated to store a constant. */
|
1361 |
static void temp_save(TCGContext *s, int temp, TCGRegSet allocated_regs) |
1362 |
{ |
1363 |
TCGTemp *ts; |
1364 |
int reg;
|
1365 |
|
1366 |
ts = &s->temps[temp]; |
1367 |
if (!ts->fixed_reg) {
|
1368 |
switch(ts->val_type) {
|
1369 |
case TEMP_VAL_REG:
|
1370 |
tcg_reg_free(s, ts->reg); |
1371 |
break;
|
1372 |
case TEMP_VAL_DEAD:
|
1373 |
ts->val_type = TEMP_VAL_MEM; |
1374 |
break;
|
1375 |
case TEMP_VAL_CONST:
|
1376 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1377 |
allocated_regs); |
1378 |
if (!ts->mem_allocated)
|
1379 |
temp_allocate_frame(s, temp); |
1380 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1381 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1382 |
ts->val_type = TEMP_VAL_MEM; |
1383 |
break;
|
1384 |
case TEMP_VAL_MEM:
|
1385 |
break;
|
1386 |
default:
|
1387 |
tcg_abort(); |
1388 |
} |
1389 |
} |
1390 |
} |
1391 |
|
1392 |
/* save globals to their cannonical location and assume they can be
|
1393 |
modified be the following code. 'allocated_regs' is used in case a
|
1394 |
temporary registers needs to be allocated to store a constant. */
|
1395 |
static void save_globals(TCGContext *s, TCGRegSet allocated_regs) |
1396 |
{ |
1397 |
int i;
|
1398 |
|
1399 |
for(i = 0; i < s->nb_globals; i++) { |
1400 |
temp_save(s, i, allocated_regs); |
1401 |
} |
1402 |
} |
1403 |
|
1404 |
/* at the end of a basic block, we assume all temporaries are dead and
|
1405 |
all globals are stored at their canonical location. */
|
1406 |
static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) |
1407 |
{ |
1408 |
TCGTemp *ts; |
1409 |
int i;
|
1410 |
|
1411 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1412 |
ts = &s->temps[i]; |
1413 |
if (ts->temp_local) {
|
1414 |
temp_save(s, i, allocated_regs); |
1415 |
} else {
|
1416 |
if (ts->val_type == TEMP_VAL_REG) {
|
1417 |
s->reg_to_temp[ts->reg] = -1;
|
1418 |
} |
1419 |
ts->val_type = TEMP_VAL_DEAD; |
1420 |
} |
1421 |
} |
1422 |
|
1423 |
save_globals(s, allocated_regs); |
1424 |
} |
1425 |
|
1426 |
#define IS_DEAD_IARG(n) ((dead_iargs >> (n)) & 1) |
1427 |
|
1428 |
static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args) |
1429 |
{ |
1430 |
TCGTemp *ots; |
1431 |
tcg_target_ulong val; |
1432 |
|
1433 |
ots = &s->temps[args[0]];
|
1434 |
val = args[1];
|
1435 |
|
1436 |
if (ots->fixed_reg) {
|
1437 |
/* for fixed registers, we do not do any constant
|
1438 |
propagation */
|
1439 |
tcg_out_movi(s, ots->type, ots->reg, val); |
1440 |
} else {
|
1441 |
/* The movi is not explicitly generated here */
|
1442 |
if (ots->val_type == TEMP_VAL_REG)
|
1443 |
s->reg_to_temp[ots->reg] = -1;
|
1444 |
ots->val_type = TEMP_VAL_CONST; |
1445 |
ots->val = val; |
1446 |
} |
1447 |
} |
1448 |
|
1449 |
static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, |
1450 |
const TCGArg *args,
|
1451 |
unsigned int dead_iargs) |
1452 |
{ |
1453 |
TCGTemp *ts, *ots; |
1454 |
int reg;
|
1455 |
const TCGArgConstraint *arg_ct;
|
1456 |
|
1457 |
ots = &s->temps[args[0]];
|
1458 |
ts = &s->temps[args[1]];
|
1459 |
arg_ct = &def->args_ct[0];
|
1460 |
|
1461 |
/* XXX: always mark arg dead if IS_DEAD_IARG(0) */
|
1462 |
if (ts->val_type == TEMP_VAL_REG) {
|
1463 |
if (IS_DEAD_IARG(0) && !ts->fixed_reg && !ots->fixed_reg) { |
1464 |
/* the mov can be suppressed */
|
1465 |
if (ots->val_type == TEMP_VAL_REG)
|
1466 |
s->reg_to_temp[ots->reg] = -1;
|
1467 |
reg = ts->reg; |
1468 |
s->reg_to_temp[reg] = -1;
|
1469 |
ts->val_type = TEMP_VAL_DEAD; |
1470 |
} else {
|
1471 |
if (ots->val_type == TEMP_VAL_REG) {
|
1472 |
reg = ots->reg; |
1473 |
} else {
|
1474 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1475 |
} |
1476 |
if (ts->reg != reg) {
|
1477 |
tcg_out_mov(s, reg, ts->reg); |
1478 |
} |
1479 |
} |
1480 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1481 |
if (ots->val_type == TEMP_VAL_REG) {
|
1482 |
reg = ots->reg; |
1483 |
} else {
|
1484 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1485 |
} |
1486 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1487 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1488 |
if (ots->fixed_reg) {
|
1489 |
reg = ots->reg; |
1490 |
tcg_out_movi(s, ots->type, reg, ts->val); |
1491 |
} else {
|
1492 |
/* propagate constant */
|
1493 |
if (ots->val_type == TEMP_VAL_REG)
|
1494 |
s->reg_to_temp[ots->reg] = -1;
|
1495 |
ots->val_type = TEMP_VAL_CONST; |
1496 |
ots->val = ts->val; |
1497 |
return;
|
1498 |
} |
1499 |
} else {
|
1500 |
tcg_abort(); |
1501 |
} |
1502 |
s->reg_to_temp[reg] = args[0];
|
1503 |
ots->reg = reg; |
1504 |
ots->val_type = TEMP_VAL_REG; |
1505 |
ots->mem_coherent = 0;
|
1506 |
} |
1507 |
|
1508 |
static void tcg_reg_alloc_op(TCGContext *s, |
1509 |
const TCGOpDef *def, int opc, |
1510 |
const TCGArg *args,
|
1511 |
unsigned int dead_iargs) |
1512 |
{ |
1513 |
TCGRegSet allocated_regs; |
1514 |
int i, k, nb_iargs, nb_oargs, reg;
|
1515 |
TCGArg arg; |
1516 |
const TCGArgConstraint *arg_ct;
|
1517 |
TCGTemp *ts; |
1518 |
TCGArg new_args[TCG_MAX_OP_ARGS]; |
1519 |
int const_args[TCG_MAX_OP_ARGS];
|
1520 |
|
1521 |
nb_oargs = def->nb_oargs; |
1522 |
nb_iargs = def->nb_iargs; |
1523 |
|
1524 |
/* copy constants */
|
1525 |
memcpy(new_args + nb_oargs + nb_iargs, |
1526 |
args + nb_oargs + nb_iargs, |
1527 |
sizeof(TCGArg) * def->nb_cargs);
|
1528 |
|
1529 |
/* satisfy input constraints */
|
1530 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1531 |
for(k = 0; k < nb_iargs; k++) { |
1532 |
i = def->sorted_args[nb_oargs + k]; |
1533 |
arg = args[i]; |
1534 |
arg_ct = &def->args_ct[i]; |
1535 |
ts = &s->temps[arg]; |
1536 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1537 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1538 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1539 |
ts->val_type = TEMP_VAL_REG; |
1540 |
ts->reg = reg; |
1541 |
ts->mem_coherent = 1;
|
1542 |
s->reg_to_temp[reg] = arg; |
1543 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1544 |
if (tcg_target_const_match(ts->val, arg_ct)) {
|
1545 |
/* constant is OK for instruction */
|
1546 |
const_args[i] = 1;
|
1547 |
new_args[i] = ts->val; |
1548 |
goto iarg_end;
|
1549 |
} else {
|
1550 |
/* need to move to a register */
|
1551 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1552 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1553 |
ts->val_type = TEMP_VAL_REG; |
1554 |
ts->reg = reg; |
1555 |
ts->mem_coherent = 0;
|
1556 |
s->reg_to_temp[reg] = arg; |
1557 |
} |
1558 |
} |
1559 |
assert(ts->val_type == TEMP_VAL_REG); |
1560 |
if (arg_ct->ct & TCG_CT_IALIAS) {
|
1561 |
if (ts->fixed_reg) {
|
1562 |
/* if fixed register, we must allocate a new register
|
1563 |
if the alias is not the same register */
|
1564 |
if (arg != args[arg_ct->alias_index])
|
1565 |
goto allocate_in_reg;
|
1566 |
} else {
|
1567 |
/* if the input is aliased to an output and if it is
|
1568 |
not dead after the instruction, we must allocate
|
1569 |
a new register and move it */
|
1570 |
if (!IS_DEAD_IARG(i - nb_oargs))
|
1571 |
goto allocate_in_reg;
|
1572 |
} |
1573 |
} |
1574 |
reg = ts->reg; |
1575 |
if (tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1576 |
/* nothing to do : the constraint is satisfied */
|
1577 |
} else {
|
1578 |
allocate_in_reg:
|
1579 |
/* allocate a new register matching the constraint
|
1580 |
and move the temporary register into it */
|
1581 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1582 |
tcg_out_mov(s, reg, ts->reg); |
1583 |
} |
1584 |
new_args[i] = reg; |
1585 |
const_args[i] = 0;
|
1586 |
tcg_regset_set_reg(allocated_regs, reg); |
1587 |
iarg_end: ;
|
1588 |
} |
1589 |
|
1590 |
if (def->flags & TCG_OPF_BB_END) {
|
1591 |
tcg_reg_alloc_bb_end(s, allocated_regs); |
1592 |
} else {
|
1593 |
/* mark dead temporaries and free the associated registers */
|
1594 |
for(i = 0; i < nb_iargs; i++) { |
1595 |
arg = args[nb_oargs + i]; |
1596 |
if (IS_DEAD_IARG(i)) {
|
1597 |
ts = &s->temps[arg]; |
1598 |
if (!ts->fixed_reg) {
|
1599 |
if (ts->val_type == TEMP_VAL_REG)
|
1600 |
s->reg_to_temp[ts->reg] = -1;
|
1601 |
ts->val_type = TEMP_VAL_DEAD; |
1602 |
} |
1603 |
} |
1604 |
} |
1605 |
|
1606 |
if (def->flags & TCG_OPF_CALL_CLOBBER) {
|
1607 |
/* XXX: permit generic clobber register list ? */
|
1608 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1609 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1610 |
tcg_reg_free(s, reg); |
1611 |
} |
1612 |
} |
1613 |
/* XXX: for load/store we could do that only for the slow path
|
1614 |
(i.e. when a memory callback is called) */
|
1615 |
|
1616 |
/* store globals and free associated registers (we assume the insn
|
1617 |
can modify any global. */
|
1618 |
save_globals(s, allocated_regs); |
1619 |
} |
1620 |
|
1621 |
/* satisfy the output constraints */
|
1622 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1623 |
for(k = 0; k < nb_oargs; k++) { |
1624 |
i = def->sorted_args[k]; |
1625 |
arg = args[i]; |
1626 |
arg_ct = &def->args_ct[i]; |
1627 |
ts = &s->temps[arg]; |
1628 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
1629 |
reg = new_args[arg_ct->alias_index]; |
1630 |
} else {
|
1631 |
/* if fixed register, we try to use it */
|
1632 |
reg = ts->reg; |
1633 |
if (ts->fixed_reg &&
|
1634 |
tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
1635 |
goto oarg_end;
|
1636 |
} |
1637 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1638 |
} |
1639 |
tcg_regset_set_reg(allocated_regs, reg); |
1640 |
/* if a fixed register is used, then a move will be done afterwards */
|
1641 |
if (!ts->fixed_reg) {
|
1642 |
if (ts->val_type == TEMP_VAL_REG)
|
1643 |
s->reg_to_temp[ts->reg] = -1;
|
1644 |
ts->val_type = TEMP_VAL_REG; |
1645 |
ts->reg = reg; |
1646 |
/* temp value is modified, so the value kept in memory is
|
1647 |
potentially not the same */
|
1648 |
ts->mem_coherent = 0;
|
1649 |
s->reg_to_temp[reg] = arg; |
1650 |
} |
1651 |
oarg_end:
|
1652 |
new_args[i] = reg; |
1653 |
} |
1654 |
} |
1655 |
|
1656 |
/* emit instruction */
|
1657 |
tcg_out_op(s, opc, new_args, const_args); |
1658 |
|
1659 |
/* move the outputs in the correct register if needed */
|
1660 |
for(i = 0; i < nb_oargs; i++) { |
1661 |
ts = &s->temps[args[i]]; |
1662 |
reg = new_args[i]; |
1663 |
if (ts->fixed_reg && ts->reg != reg) {
|
1664 |
tcg_out_mov(s, ts->reg, reg); |
1665 |
} |
1666 |
} |
1667 |
} |
1668 |
|
1669 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1670 |
#define STACK_DIR(x) (-(x))
|
1671 |
#else
|
1672 |
#define STACK_DIR(x) (x)
|
1673 |
#endif
|
1674 |
|
1675 |
static int tcg_reg_alloc_call(TCGContext *s, const TCGOpDef *def, |
1676 |
int opc, const TCGArg *args, |
1677 |
unsigned int dead_iargs) |
1678 |
{ |
1679 |
int nb_iargs, nb_oargs, flags, nb_regs, i, reg, nb_params;
|
1680 |
TCGArg arg, func_arg; |
1681 |
TCGTemp *ts; |
1682 |
tcg_target_long stack_offset, call_stack_size, func_addr; |
1683 |
int const_func_arg, allocate_args;
|
1684 |
TCGRegSet allocated_regs; |
1685 |
const TCGArgConstraint *arg_ct;
|
1686 |
|
1687 |
arg = *args++; |
1688 |
|
1689 |
nb_oargs = arg >> 16;
|
1690 |
nb_iargs = arg & 0xffff;
|
1691 |
nb_params = nb_iargs - 1;
|
1692 |
|
1693 |
flags = args[nb_oargs + nb_iargs]; |
1694 |
|
1695 |
nb_regs = tcg_target_get_call_iarg_regs_count(flags); |
1696 |
if (nb_regs > nb_params)
|
1697 |
nb_regs = nb_params; |
1698 |
|
1699 |
/* assign stack slots first */
|
1700 |
/* XXX: preallocate call stack */
|
1701 |
call_stack_size = (nb_params - nb_regs) * sizeof(tcg_target_long);
|
1702 |
call_stack_size = (call_stack_size + TCG_TARGET_STACK_ALIGN - 1) &
|
1703 |
~(TCG_TARGET_STACK_ALIGN - 1);
|
1704 |
allocate_args = (call_stack_size > TCG_STATIC_CALL_ARGS_SIZE); |
1705 |
if (allocate_args) {
|
1706 |
tcg_out_addi(s, TCG_REG_CALL_STACK, -STACK_DIR(call_stack_size)); |
1707 |
} |
1708 |
|
1709 |
stack_offset = TCG_TARGET_CALL_STACK_OFFSET; |
1710 |
for(i = nb_regs; i < nb_params; i++) {
|
1711 |
arg = args[nb_oargs + i]; |
1712 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1713 |
stack_offset -= sizeof(tcg_target_long);
|
1714 |
#endif
|
1715 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1716 |
ts = &s->temps[arg]; |
1717 |
if (ts->val_type == TEMP_VAL_REG) {
|
1718 |
tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, stack_offset); |
1719 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1720 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1721 |
s->reserved_regs); |
1722 |
/* XXX: not correct if reading values from the stack */
|
1723 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1724 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1725 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1726 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1727 |
s->reserved_regs); |
1728 |
/* XXX: sign extend may be needed on some targets */
|
1729 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1730 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1731 |
} else {
|
1732 |
tcg_abort(); |
1733 |
} |
1734 |
} |
1735 |
#ifndef TCG_TARGET_STACK_GROWSUP
|
1736 |
stack_offset += sizeof(tcg_target_long);
|
1737 |
#endif
|
1738 |
} |
1739 |
|
1740 |
/* assign input registers */
|
1741 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1742 |
for(i = 0; i < nb_regs; i++) { |
1743 |
arg = args[nb_oargs + i]; |
1744 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1745 |
ts = &s->temps[arg]; |
1746 |
reg = tcg_target_call_iarg_regs[i]; |
1747 |
tcg_reg_free(s, reg); |
1748 |
if (ts->val_type == TEMP_VAL_REG) {
|
1749 |
if (ts->reg != reg) {
|
1750 |
tcg_out_mov(s, reg, ts->reg); |
1751 |
} |
1752 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1753 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1754 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1755 |
/* XXX: sign extend ? */
|
1756 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1757 |
} else {
|
1758 |
tcg_abort(); |
1759 |
} |
1760 |
tcg_regset_set_reg(allocated_regs, reg); |
1761 |
} |
1762 |
} |
1763 |
|
1764 |
/* assign function address */
|
1765 |
func_arg = args[nb_oargs + nb_iargs - 1];
|
1766 |
arg_ct = &def->args_ct[0];
|
1767 |
ts = &s->temps[func_arg]; |
1768 |
func_addr = ts->val; |
1769 |
const_func_arg = 0;
|
1770 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1771 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1772 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1773 |
func_arg = reg; |
1774 |
tcg_regset_set_reg(allocated_regs, reg); |
1775 |
} else if (ts->val_type == TEMP_VAL_REG) { |
1776 |
reg = ts->reg; |
1777 |
if (!tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1778 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1779 |
tcg_out_mov(s, reg, ts->reg); |
1780 |
} |
1781 |
func_arg = reg; |
1782 |
tcg_regset_set_reg(allocated_regs, reg); |
1783 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1784 |
if (tcg_target_const_match(func_addr, arg_ct)) {
|
1785 |
const_func_arg = 1;
|
1786 |
func_arg = func_addr; |
1787 |
} else {
|
1788 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1789 |
tcg_out_movi(s, ts->type, reg, func_addr); |
1790 |
func_arg = reg; |
1791 |
tcg_regset_set_reg(allocated_regs, reg); |
1792 |
} |
1793 |
} else {
|
1794 |
tcg_abort(); |
1795 |
} |
1796 |
|
1797 |
|
1798 |
/* mark dead temporaries and free the associated registers */
|
1799 |
for(i = 0; i < nb_iargs; i++) { |
1800 |
arg = args[nb_oargs + i]; |
1801 |
if (IS_DEAD_IARG(i)) {
|
1802 |
ts = &s->temps[arg]; |
1803 |
if (!ts->fixed_reg) {
|
1804 |
if (ts->val_type == TEMP_VAL_REG)
|
1805 |
s->reg_to_temp[ts->reg] = -1;
|
1806 |
ts->val_type = TEMP_VAL_DEAD; |
1807 |
} |
1808 |
} |
1809 |
} |
1810 |
|
1811 |
/* clobber call registers */
|
1812 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1813 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1814 |
tcg_reg_free(s, reg); |
1815 |
} |
1816 |
} |
1817 |
|
1818 |
/* store globals and free associated registers (we assume the call
|
1819 |
can modify any global. */
|
1820 |
save_globals(s, allocated_regs); |
1821 |
|
1822 |
tcg_out_op(s, opc, &func_arg, &const_func_arg); |
1823 |
|
1824 |
if (allocate_args) {
|
1825 |
tcg_out_addi(s, TCG_REG_CALL_STACK, STACK_DIR(call_stack_size)); |
1826 |
} |
1827 |
|
1828 |
/* assign output registers and emit moves if needed */
|
1829 |
for(i = 0; i < nb_oargs; i++) { |
1830 |
arg = args[i]; |
1831 |
ts = &s->temps[arg]; |
1832 |
reg = tcg_target_call_oarg_regs[i]; |
1833 |
assert(s->reg_to_temp[reg] == -1);
|
1834 |
if (ts->fixed_reg) {
|
1835 |
if (ts->reg != reg) {
|
1836 |
tcg_out_mov(s, ts->reg, reg); |
1837 |
} |
1838 |
} else {
|
1839 |
if (ts->val_type == TEMP_VAL_REG)
|
1840 |
s->reg_to_temp[ts->reg] = -1;
|
1841 |
ts->val_type = TEMP_VAL_REG; |
1842 |
ts->reg = reg; |
1843 |
ts->mem_coherent = 0;
|
1844 |
s->reg_to_temp[reg] = arg; |
1845 |
} |
1846 |
} |
1847 |
|
1848 |
return nb_iargs + nb_oargs + def->nb_cargs + 1; |
1849 |
} |
1850 |
|
1851 |
#ifdef CONFIG_PROFILER
|
1852 |
|
1853 |
static int64_t dyngen_table_op_count[NB_OPS];
|
1854 |
|
1855 |
void dump_op_count(void) |
1856 |
{ |
1857 |
int i;
|
1858 |
FILE *f; |
1859 |
f = fopen("/tmp/op1.log", "w"); |
1860 |
for(i = 0; i < INDEX_op_end; i++) { |
1861 |
fprintf(f, "%s %" PRId64 "\n", tcg_op_defs[i].name, dyngen_table_op_count[i]); |
1862 |
} |
1863 |
fclose(f); |
1864 |
f = fopen("/tmp/op2.log", "w"); |
1865 |
for(i = INDEX_op_end; i < NB_OPS; i++) {
|
1866 |
fprintf(f, "%s %" PRId64 "\n", tcg_op_defs[i].name, dyngen_table_op_count[i]); |
1867 |
} |
1868 |
fclose(f); |
1869 |
} |
1870 |
#endif
|
1871 |
|
1872 |
|
1873 |
static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf, |
1874 |
long search_pc)
|
1875 |
{ |
1876 |
int opc, op_index;
|
1877 |
const TCGOpDef *def;
|
1878 |
unsigned int dead_iargs; |
1879 |
const TCGArg *args;
|
1880 |
|
1881 |
#ifdef DEBUG_DISAS
|
1882 |
if (unlikely(loglevel & CPU_LOG_TB_OP)) {
|
1883 |
fprintf(logfile, "OP:\n");
|
1884 |
tcg_dump_ops(s, logfile); |
1885 |
fprintf(logfile, "\n");
|
1886 |
} |
1887 |
#endif
|
1888 |
|
1889 |
#ifdef CONFIG_PROFILER
|
1890 |
s->la_time -= profile_getclock(); |
1891 |
#endif
|
1892 |
tcg_liveness_analysis(s); |
1893 |
#ifdef CONFIG_PROFILER
|
1894 |
s->la_time += profile_getclock(); |
1895 |
#endif
|
1896 |
|
1897 |
#ifdef DEBUG_DISAS
|
1898 |
if (unlikely(loglevel & CPU_LOG_TB_OP_OPT)) {
|
1899 |
fprintf(logfile, "OP after la:\n");
|
1900 |
tcg_dump_ops(s, logfile); |
1901 |
fprintf(logfile, "\n");
|
1902 |
} |
1903 |
#endif
|
1904 |
|
1905 |
tcg_reg_alloc_start(s); |
1906 |
|
1907 |
s->code_buf = gen_code_buf; |
1908 |
s->code_ptr = gen_code_buf; |
1909 |
|
1910 |
args = gen_opparam_buf; |
1911 |
op_index = 0;
|
1912 |
|
1913 |
for(;;) {
|
1914 |
opc = gen_opc_buf[op_index]; |
1915 |
#ifdef CONFIG_PROFILER
|
1916 |
dyngen_table_op_count[opc]++; |
1917 |
#endif
|
1918 |
def = &tcg_op_defs[opc]; |
1919 |
#if 0
|
1920 |
printf("%s: %d %d %d\n", def->name,
|
1921 |
def->nb_oargs, def->nb_iargs, def->nb_cargs);
|
1922 |
// dump_regs(s);
|
1923 |
#endif
|
1924 |
switch(opc) {
|
1925 |
case INDEX_op_mov_i32:
|
1926 |
#if TCG_TARGET_REG_BITS == 64 |
1927 |
case INDEX_op_mov_i64:
|
1928 |
#endif
|
1929 |
dead_iargs = s->op_dead_iargs[op_index]; |
1930 |
tcg_reg_alloc_mov(s, def, args, dead_iargs); |
1931 |
break;
|
1932 |
case INDEX_op_movi_i32:
|
1933 |
#if TCG_TARGET_REG_BITS == 64 |
1934 |
case INDEX_op_movi_i64:
|
1935 |
#endif
|
1936 |
tcg_reg_alloc_movi(s, args); |
1937 |
break;
|
1938 |
case INDEX_op_debug_insn_start:
|
1939 |
/* debug instruction */
|
1940 |
break;
|
1941 |
case INDEX_op_nop:
|
1942 |
case INDEX_op_nop1:
|
1943 |
case INDEX_op_nop2:
|
1944 |
case INDEX_op_nop3:
|
1945 |
break;
|
1946 |
case INDEX_op_nopn:
|
1947 |
args += args[0];
|
1948 |
goto next;
|
1949 |
case INDEX_op_discard:
|
1950 |
{ |
1951 |
TCGTemp *ts; |
1952 |
ts = &s->temps[args[0]];
|
1953 |
/* mark the temporary as dead */
|
1954 |
if (!ts->fixed_reg) {
|
1955 |
if (ts->val_type == TEMP_VAL_REG)
|
1956 |
s->reg_to_temp[ts->reg] = -1;
|
1957 |
ts->val_type = TEMP_VAL_DEAD; |
1958 |
} |
1959 |
} |
1960 |
break;
|
1961 |
case INDEX_op_set_label:
|
1962 |
tcg_reg_alloc_bb_end(s, s->reserved_regs); |
1963 |
tcg_out_label(s, args[0], (long)s->code_ptr); |
1964 |
break;
|
1965 |
case INDEX_op_call:
|
1966 |
dead_iargs = s->op_dead_iargs[op_index]; |
1967 |
args += tcg_reg_alloc_call(s, def, opc, args, dead_iargs); |
1968 |
goto next;
|
1969 |
case INDEX_op_end:
|
1970 |
goto the_end;
|
1971 |
|
1972 |
#ifdef CONFIG_DYNGEN_OP
|
1973 |
case 0 ... INDEX_op_end - 1: |
1974 |
/* legacy dyngen ops */
|
1975 |
#ifdef CONFIG_PROFILER
|
1976 |
s->old_op_count++; |
1977 |
#endif
|
1978 |
tcg_reg_alloc_bb_end(s, s->reserved_regs); |
1979 |
if (search_pc >= 0) { |
1980 |
s->code_ptr += def->copy_size; |
1981 |
args += def->nb_args; |
1982 |
} else {
|
1983 |
args = dyngen_op(s, opc, args); |
1984 |
} |
1985 |
goto next;
|
1986 |
#endif
|
1987 |
default:
|
1988 |
/* Note: in order to speed up the code, it would be much
|
1989 |
faster to have specialized register allocator functions for
|
1990 |
some common argument patterns */
|
1991 |
dead_iargs = s->op_dead_iargs[op_index]; |
1992 |
tcg_reg_alloc_op(s, def, opc, args, dead_iargs); |
1993 |
break;
|
1994 |
} |
1995 |
args += def->nb_args; |
1996 |
next:
|
1997 |
if (search_pc >= 0 && search_pc < s->code_ptr - gen_code_buf) { |
1998 |
return op_index;
|
1999 |
} |
2000 |
op_index++; |
2001 |
#ifndef NDEBUG
|
2002 |
check_regs(s); |
2003 |
#endif
|
2004 |
} |
2005 |
the_end:
|
2006 |
return -1; |
2007 |
} |
2008 |
|
2009 |
int dyngen_code(TCGContext *s, uint8_t *gen_code_buf)
|
2010 |
{ |
2011 |
#ifdef CONFIG_PROFILER
|
2012 |
{ |
2013 |
int n;
|
2014 |
n = (gen_opc_ptr - gen_opc_buf); |
2015 |
s->op_count += n; |
2016 |
if (n > s->op_count_max)
|
2017 |
s->op_count_max = n; |
2018 |
|
2019 |
s->temp_count += s->nb_temps; |
2020 |
if (s->nb_temps > s->temp_count_max)
|
2021 |
s->temp_count_max = s->nb_temps; |
2022 |
} |
2023 |
#endif
|
2024 |
|
2025 |
tcg_gen_code_common(s, gen_code_buf, -1);
|
2026 |
|
2027 |
/* flush instruction cache */
|
2028 |
flush_icache_range((unsigned long)gen_code_buf, |
2029 |
(unsigned long)s->code_ptr); |
2030 |
return s->code_ptr - gen_code_buf;
|
2031 |
} |
2032 |
|
2033 |
/* Return the index of the micro operation such as the pc after is <
|
2034 |
offset bytes from the start of the TB. The contents of gen_code_buf must
|
2035 |
not be changed, though writing the same values is ok.
|
2036 |
Return -1 if not found. */
|
2037 |
int dyngen_code_search_pc(TCGContext *s, uint8_t *gen_code_buf, long offset) |
2038 |
{ |
2039 |
return tcg_gen_code_common(s, gen_code_buf, offset);
|
2040 |
} |
2041 |
|
2042 |
#ifdef CONFIG_PROFILER
|
2043 |
void tcg_dump_info(FILE *f,
|
2044 |
int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
2045 |
{ |
2046 |
TCGContext *s = &tcg_ctx; |
2047 |
int64_t tot; |
2048 |
|
2049 |
tot = s->interm_time + s->code_time; |
2050 |
cpu_fprintf(f, "JIT cycles %" PRId64 " (%0.3f s at 2.4 GHz)\n", |
2051 |
tot, tot / 2.4e9); |
2052 |
cpu_fprintf(f, "translated TBs %" PRId64 " (aborted=%" PRId64 " %0.1f%%)\n", |
2053 |
s->tb_count, |
2054 |
s->tb_count1 - s->tb_count, |
2055 |
s->tb_count1 ? (double)(s->tb_count1 - s->tb_count) / s->tb_count1 * 100.0 : 0); |
2056 |
cpu_fprintf(f, "avg ops/TB %0.1f max=%d\n",
|
2057 |
s->tb_count ? (double)s->op_count / s->tb_count : 0, s->op_count_max); |
2058 |
cpu_fprintf(f, "old ops/total ops %0.1f%%\n",
|
2059 |
s->op_count ? (double)s->old_op_count / s->op_count * 100.0 : 0); |
2060 |
cpu_fprintf(f, "deleted ops/TB %0.2f\n",
|
2061 |
s->tb_count ? |
2062 |
(double)s->del_op_count / s->tb_count : 0); |
2063 |
cpu_fprintf(f, "avg temps/TB %0.2f max=%d\n",
|
2064 |
s->tb_count ? |
2065 |
(double)s->temp_count / s->tb_count : 0, |
2066 |
s->temp_count_max); |
2067 |
|
2068 |
cpu_fprintf(f, "cycles/op %0.1f\n",
|
2069 |
s->op_count ? (double)tot / s->op_count : 0); |
2070 |
cpu_fprintf(f, "cycles/in byte %0.1f\n",
|
2071 |
s->code_in_len ? (double)tot / s->code_in_len : 0); |
2072 |
cpu_fprintf(f, "cycles/out byte %0.1f\n",
|
2073 |
s->code_out_len ? (double)tot / s->code_out_len : 0); |
2074 |
if (tot == 0) |
2075 |
tot = 1;
|
2076 |
cpu_fprintf(f, " gen_interm time %0.1f%%\n",
|
2077 |
(double)s->interm_time / tot * 100.0); |
2078 |
cpu_fprintf(f, " gen_code time %0.1f%%\n",
|
2079 |
(double)s->code_time / tot * 100.0); |
2080 |
cpu_fprintf(f, "liveness/code time %0.1f%%\n",
|
2081 |
(double)s->la_time / (s->code_time ? s->code_time : 1) * 100.0); |
2082 |
cpu_fprintf(f, "cpu_restore count %" PRId64 "\n", |
2083 |
s->restore_count); |
2084 |
cpu_fprintf(f, " avg cycles %0.1f\n",
|
2085 |
s->restore_count ? (double)s->restore_time / s->restore_count : 0); |
2086 |
{ |
2087 |
extern void dump_op_count(void); |
2088 |
dump_op_count(); |
2089 |
} |
2090 |
} |
2091 |
#else
|
2092 |
void tcg_dump_info(FILE *f,
|
2093 |
int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
2094 |
{ |
2095 |
cpu_fprintf(f, "[TCG profiler not compiled]\n");
|
2096 |
} |
2097 |
#endif
|