root / tcg / tcg.c @ 7a3766f3
History | View | Annotate | Download (61.7 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 use liveness analysis (better code) */
|
26 |
#define USE_LIVENESS_ANALYSIS
|
27 |
|
28 |
#include "config.h" |
29 |
|
30 |
#ifndef CONFIG_DEBUG_TCG
|
31 |
/* define it to suppress various consistency checks (faster) */
|
32 |
#define NDEBUG
|
33 |
#endif
|
34 |
|
35 |
#include <stdarg.h> |
36 |
#include <stdlib.h> |
37 |
#include <stdio.h> |
38 |
#include <string.h> |
39 |
#include <inttypes.h> |
40 |
#ifdef _WIN32
|
41 |
#include <malloc.h> |
42 |
#endif
|
43 |
#ifdef _AIX
|
44 |
#include <alloca.h> |
45 |
#endif
|
46 |
|
47 |
#include "qemu-common.h" |
48 |
#include "cache-utils.h" |
49 |
#include "host-utils.h" |
50 |
|
51 |
/* Note: the long term plan is to reduce the dependancies on the QEMU
|
52 |
CPU definitions. Currently they are used for qemu_ld/st
|
53 |
instructions */
|
54 |
#define NO_CPU_IO_DEFS
|
55 |
#include "cpu.h" |
56 |
#include "exec-all.h" |
57 |
|
58 |
#include "tcg-op.h" |
59 |
#include "elf.h" |
60 |
|
61 |
#if defined(CONFIG_USE_GUEST_BASE) && !defined(TCG_TARGET_HAS_GUEST_BASE)
|
62 |
#error GUEST_BASE not supported on this host.
|
63 |
#endif
|
64 |
|
65 |
static void patch_reloc(uint8_t *code_ptr, int type, |
66 |
tcg_target_long value, tcg_target_long addend); |
67 |
|
68 |
static TCGOpDef tcg_op_defs[] = {
|
69 |
#define DEF(s, n, copy_size) { #s, 0, 0, n, n, 0, copy_size }, |
70 |
#define DEF2(s, oargs, iargs, cargs, flags) { #s, oargs, iargs, cargs, iargs + oargs + cargs, flags, 0 }, |
71 |
#include "tcg-opc.h" |
72 |
#undef DEF
|
73 |
#undef DEF2
|
74 |
}; |
75 |
|
76 |
static TCGRegSet tcg_target_available_regs[2]; |
77 |
static TCGRegSet tcg_target_call_clobber_regs;
|
78 |
|
79 |
/* XXX: move that inside the context */
|
80 |
uint16_t *gen_opc_ptr; |
81 |
TCGArg *gen_opparam_ptr; |
82 |
|
83 |
static inline void tcg_out8(TCGContext *s, uint8_t v) |
84 |
{ |
85 |
*s->code_ptr++ = v; |
86 |
} |
87 |
|
88 |
static inline void tcg_out16(TCGContext *s, uint16_t v) |
89 |
{ |
90 |
*(uint16_t *)s->code_ptr = v; |
91 |
s->code_ptr += 2;
|
92 |
} |
93 |
|
94 |
static inline void tcg_out32(TCGContext *s, uint32_t v) |
95 |
{ |
96 |
*(uint32_t *)s->code_ptr = v; |
97 |
s->code_ptr += 4;
|
98 |
} |
99 |
|
100 |
/* label relocation processing */
|
101 |
|
102 |
void tcg_out_reloc(TCGContext *s, uint8_t *code_ptr, int type, |
103 |
int label_index, long addend) |
104 |
{ |
105 |
TCGLabel *l; |
106 |
TCGRelocation *r; |
107 |
|
108 |
l = &s->labels[label_index]; |
109 |
if (l->has_value) {
|
110 |
/* FIXME: This may break relocations on RISC targets that
|
111 |
modify instruction fields in place. The caller may not have
|
112 |
written the initial value. */
|
113 |
patch_reloc(code_ptr, type, l->u.value, addend); |
114 |
} else {
|
115 |
/* add a new relocation entry */
|
116 |
r = tcg_malloc(sizeof(TCGRelocation));
|
117 |
r->type = type; |
118 |
r->ptr = code_ptr; |
119 |
r->addend = addend; |
120 |
r->next = l->u.first_reloc; |
121 |
l->u.first_reloc = r; |
122 |
} |
123 |
} |
124 |
|
125 |
static void tcg_out_label(TCGContext *s, int label_index, |
126 |
tcg_target_long value) |
127 |
{ |
128 |
TCGLabel *l; |
129 |
TCGRelocation *r; |
130 |
|
131 |
l = &s->labels[label_index]; |
132 |
if (l->has_value)
|
133 |
tcg_abort(); |
134 |
r = l->u.first_reloc; |
135 |
while (r != NULL) { |
136 |
patch_reloc(r->ptr, r->type, value, r->addend); |
137 |
r = r->next; |
138 |
} |
139 |
l->has_value = 1;
|
140 |
l->u.value = value; |
141 |
} |
142 |
|
143 |
int gen_new_label(void) |
144 |
{ |
145 |
TCGContext *s = &tcg_ctx; |
146 |
int idx;
|
147 |
TCGLabel *l; |
148 |
|
149 |
if (s->nb_labels >= TCG_MAX_LABELS)
|
150 |
tcg_abort(); |
151 |
idx = s->nb_labels++; |
152 |
l = &s->labels[idx]; |
153 |
l->has_value = 0;
|
154 |
l->u.first_reloc = NULL;
|
155 |
return idx;
|
156 |
} |
157 |
|
158 |
#include "tcg-target.c" |
159 |
|
160 |
/* pool based memory allocation */
|
161 |
void *tcg_malloc_internal(TCGContext *s, int size) |
162 |
{ |
163 |
TCGPool *p; |
164 |
int pool_size;
|
165 |
|
166 |
if (size > TCG_POOL_CHUNK_SIZE) {
|
167 |
/* big malloc: insert a new pool (XXX: could optimize) */
|
168 |
p = qemu_malloc(sizeof(TCGPool) + size);
|
169 |
p->size = size; |
170 |
if (s->pool_current)
|
171 |
s->pool_current->next = p; |
172 |
else
|
173 |
s->pool_first = p; |
174 |
p->next = s->pool_current; |
175 |
} else {
|
176 |
p = s->pool_current; |
177 |
if (!p) {
|
178 |
p = s->pool_first; |
179 |
if (!p)
|
180 |
goto new_pool;
|
181 |
} else {
|
182 |
if (!p->next) {
|
183 |
new_pool:
|
184 |
pool_size = TCG_POOL_CHUNK_SIZE; |
185 |
p = qemu_malloc(sizeof(TCGPool) + pool_size);
|
186 |
p->size = pool_size; |
187 |
p->next = NULL;
|
188 |
if (s->pool_current)
|
189 |
s->pool_current->next = p; |
190 |
else
|
191 |
s->pool_first = p; |
192 |
} else {
|
193 |
p = p->next; |
194 |
} |
195 |
} |
196 |
} |
197 |
s->pool_current = p; |
198 |
s->pool_cur = p->data + size; |
199 |
s->pool_end = p->data + p->size; |
200 |
return p->data;
|
201 |
} |
202 |
|
203 |
void tcg_pool_reset(TCGContext *s)
|
204 |
{ |
205 |
s->pool_cur = s->pool_end = NULL;
|
206 |
s->pool_current = NULL;
|
207 |
} |
208 |
|
209 |
void tcg_context_init(TCGContext *s)
|
210 |
{ |
211 |
int op, total_args, n;
|
212 |
TCGOpDef *def; |
213 |
TCGArgConstraint *args_ct; |
214 |
int *sorted_args;
|
215 |
|
216 |
memset(s, 0, sizeof(*s)); |
217 |
s->temps = s->static_temps; |
218 |
s->nb_globals = 0;
|
219 |
|
220 |
/* Count total number of arguments and allocate the corresponding
|
221 |
space */
|
222 |
total_args = 0;
|
223 |
for(op = 0; op < NB_OPS; op++) { |
224 |
def = &tcg_op_defs[op]; |
225 |
n = def->nb_iargs + def->nb_oargs; |
226 |
total_args += n; |
227 |
} |
228 |
|
229 |
args_ct = qemu_malloc(sizeof(TCGArgConstraint) * total_args);
|
230 |
sorted_args = qemu_malloc(sizeof(int) * total_args); |
231 |
|
232 |
for(op = 0; op < NB_OPS; op++) { |
233 |
def = &tcg_op_defs[op]; |
234 |
def->args_ct = args_ct; |
235 |
def->sorted_args = sorted_args; |
236 |
n = def->nb_iargs + def->nb_oargs; |
237 |
sorted_args += n; |
238 |
args_ct += n; |
239 |
} |
240 |
|
241 |
tcg_target_init(s); |
242 |
|
243 |
/* init global prologue and epilogue */
|
244 |
s->code_buf = code_gen_prologue; |
245 |
s->code_ptr = s->code_buf; |
246 |
tcg_target_qemu_prologue(s); |
247 |
flush_icache_range((unsigned long)s->code_buf, |
248 |
(unsigned long)s->code_ptr); |
249 |
} |
250 |
|
251 |
void tcg_set_frame(TCGContext *s, int reg, |
252 |
tcg_target_long start, tcg_target_long size) |
253 |
{ |
254 |
s->frame_start = start; |
255 |
s->frame_end = start + size; |
256 |
s->frame_reg = reg; |
257 |
} |
258 |
|
259 |
void tcg_func_start(TCGContext *s)
|
260 |
{ |
261 |
int i;
|
262 |
tcg_pool_reset(s); |
263 |
s->nb_temps = s->nb_globals; |
264 |
for(i = 0; i < (TCG_TYPE_COUNT * 2); i++) |
265 |
s->first_free_temp[i] = -1;
|
266 |
s->labels = tcg_malloc(sizeof(TCGLabel) * TCG_MAX_LABELS);
|
267 |
s->nb_labels = 0;
|
268 |
s->current_frame_offset = s->frame_start; |
269 |
|
270 |
gen_opc_ptr = gen_opc_buf; |
271 |
gen_opparam_ptr = gen_opparam_buf; |
272 |
} |
273 |
|
274 |
static inline void tcg_temp_alloc(TCGContext *s, int n) |
275 |
{ |
276 |
if (n > TCG_MAX_TEMPS)
|
277 |
tcg_abort(); |
278 |
} |
279 |
|
280 |
static inline int tcg_global_reg_new_internal(TCGType type, int reg, |
281 |
const char *name) |
282 |
{ |
283 |
TCGContext *s = &tcg_ctx; |
284 |
TCGTemp *ts; |
285 |
int idx;
|
286 |
|
287 |
#if TCG_TARGET_REG_BITS == 32 |
288 |
if (type != TCG_TYPE_I32)
|
289 |
tcg_abort(); |
290 |
#endif
|
291 |
if (tcg_regset_test_reg(s->reserved_regs, reg))
|
292 |
tcg_abort(); |
293 |
idx = s->nb_globals; |
294 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
295 |
ts = &s->temps[s->nb_globals]; |
296 |
ts->base_type = type; |
297 |
ts->type = type; |
298 |
ts->fixed_reg = 1;
|
299 |
ts->reg = reg; |
300 |
ts->name = name; |
301 |
s->nb_globals++; |
302 |
tcg_regset_set_reg(s->reserved_regs, reg); |
303 |
return idx;
|
304 |
} |
305 |
|
306 |
TCGv_i32 tcg_global_reg_new_i32(int reg, const char *name) |
307 |
{ |
308 |
int idx;
|
309 |
|
310 |
idx = tcg_global_reg_new_internal(TCG_TYPE_I32, reg, name); |
311 |
return MAKE_TCGV_I32(idx);
|
312 |
} |
313 |
|
314 |
TCGv_i64 tcg_global_reg_new_i64(int reg, const char *name) |
315 |
{ |
316 |
int idx;
|
317 |
|
318 |
idx = tcg_global_reg_new_internal(TCG_TYPE_I64, reg, name); |
319 |
return MAKE_TCGV_I64(idx);
|
320 |
} |
321 |
|
322 |
static inline int tcg_global_mem_new_internal(TCGType type, int reg, |
323 |
tcg_target_long offset, |
324 |
const char *name) |
325 |
{ |
326 |
TCGContext *s = &tcg_ctx; |
327 |
TCGTemp *ts; |
328 |
int idx;
|
329 |
|
330 |
idx = s->nb_globals; |
331 |
#if TCG_TARGET_REG_BITS == 32 |
332 |
if (type == TCG_TYPE_I64) {
|
333 |
char buf[64]; |
334 |
tcg_temp_alloc(s, s->nb_globals + 2);
|
335 |
ts = &s->temps[s->nb_globals]; |
336 |
ts->base_type = type; |
337 |
ts->type = TCG_TYPE_I32; |
338 |
ts->fixed_reg = 0;
|
339 |
ts->mem_allocated = 1;
|
340 |
ts->mem_reg = reg; |
341 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
342 |
ts->mem_offset = offset + 4;
|
343 |
#else
|
344 |
ts->mem_offset = offset; |
345 |
#endif
|
346 |
pstrcpy(buf, sizeof(buf), name);
|
347 |
pstrcat(buf, sizeof(buf), "_0"); |
348 |
ts->name = strdup(buf); |
349 |
ts++; |
350 |
|
351 |
ts->base_type = type; |
352 |
ts->type = TCG_TYPE_I32; |
353 |
ts->fixed_reg = 0;
|
354 |
ts->mem_allocated = 1;
|
355 |
ts->mem_reg = reg; |
356 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
357 |
ts->mem_offset = offset; |
358 |
#else
|
359 |
ts->mem_offset = offset + 4;
|
360 |
#endif
|
361 |
pstrcpy(buf, sizeof(buf), name);
|
362 |
pstrcat(buf, sizeof(buf), "_1"); |
363 |
ts->name = strdup(buf); |
364 |
|
365 |
s->nb_globals += 2;
|
366 |
} else
|
367 |
#endif
|
368 |
{ |
369 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
370 |
ts = &s->temps[s->nb_globals]; |
371 |
ts->base_type = type; |
372 |
ts->type = type; |
373 |
ts->fixed_reg = 0;
|
374 |
ts->mem_allocated = 1;
|
375 |
ts->mem_reg = reg; |
376 |
ts->mem_offset = offset; |
377 |
ts->name = name; |
378 |
s->nb_globals++; |
379 |
} |
380 |
return idx;
|
381 |
} |
382 |
|
383 |
TCGv_i32 tcg_global_mem_new_i32(int reg, tcg_target_long offset,
|
384 |
const char *name) |
385 |
{ |
386 |
int idx;
|
387 |
|
388 |
idx = tcg_global_mem_new_internal(TCG_TYPE_I32, reg, offset, name); |
389 |
return MAKE_TCGV_I32(idx);
|
390 |
} |
391 |
|
392 |
TCGv_i64 tcg_global_mem_new_i64(int reg, tcg_target_long offset,
|
393 |
const char *name) |
394 |
{ |
395 |
int idx;
|
396 |
|
397 |
idx = tcg_global_mem_new_internal(TCG_TYPE_I64, reg, offset, name); |
398 |
return MAKE_TCGV_I64(idx);
|
399 |
} |
400 |
|
401 |
static inline int tcg_temp_new_internal(TCGType type, int temp_local) |
402 |
{ |
403 |
TCGContext *s = &tcg_ctx; |
404 |
TCGTemp *ts; |
405 |
int idx, k;
|
406 |
|
407 |
k = type; |
408 |
if (temp_local)
|
409 |
k += TCG_TYPE_COUNT; |
410 |
idx = s->first_free_temp[k]; |
411 |
if (idx != -1) { |
412 |
/* There is already an available temp with the
|
413 |
right type */
|
414 |
ts = &s->temps[idx]; |
415 |
s->first_free_temp[k] = ts->next_free_temp; |
416 |
ts->temp_allocated = 1;
|
417 |
assert(ts->temp_local == temp_local); |
418 |
} else {
|
419 |
idx = s->nb_temps; |
420 |
#if TCG_TARGET_REG_BITS == 32 |
421 |
if (type == TCG_TYPE_I64) {
|
422 |
tcg_temp_alloc(s, s->nb_temps + 2);
|
423 |
ts = &s->temps[s->nb_temps]; |
424 |
ts->base_type = type; |
425 |
ts->type = TCG_TYPE_I32; |
426 |
ts->temp_allocated = 1;
|
427 |
ts->temp_local = temp_local; |
428 |
ts->name = NULL;
|
429 |
ts++; |
430 |
ts->base_type = TCG_TYPE_I32; |
431 |
ts->type = TCG_TYPE_I32; |
432 |
ts->temp_allocated = 1;
|
433 |
ts->temp_local = temp_local; |
434 |
ts->name = NULL;
|
435 |
s->nb_temps += 2;
|
436 |
} else
|
437 |
#endif
|
438 |
{ |
439 |
tcg_temp_alloc(s, s->nb_temps + 1);
|
440 |
ts = &s->temps[s->nb_temps]; |
441 |
ts->base_type = type; |
442 |
ts->type = type; |
443 |
ts->temp_allocated = 1;
|
444 |
ts->temp_local = temp_local; |
445 |
ts->name = NULL;
|
446 |
s->nb_temps++; |
447 |
} |
448 |
} |
449 |
return idx;
|
450 |
} |
451 |
|
452 |
TCGv_i32 tcg_temp_new_internal_i32(int temp_local)
|
453 |
{ |
454 |
int idx;
|
455 |
|
456 |
idx = tcg_temp_new_internal(TCG_TYPE_I32, temp_local); |
457 |
return MAKE_TCGV_I32(idx);
|
458 |
} |
459 |
|
460 |
TCGv_i64 tcg_temp_new_internal_i64(int temp_local)
|
461 |
{ |
462 |
int idx;
|
463 |
|
464 |
idx = tcg_temp_new_internal(TCG_TYPE_I64, temp_local); |
465 |
return MAKE_TCGV_I64(idx);
|
466 |
} |
467 |
|
468 |
static inline void tcg_temp_free_internal(int idx) |
469 |
{ |
470 |
TCGContext *s = &tcg_ctx; |
471 |
TCGTemp *ts; |
472 |
int k;
|
473 |
|
474 |
assert(idx >= s->nb_globals && idx < s->nb_temps); |
475 |
ts = &s->temps[idx]; |
476 |
assert(ts->temp_allocated != 0);
|
477 |
ts->temp_allocated = 0;
|
478 |
k = ts->base_type; |
479 |
if (ts->temp_local)
|
480 |
k += TCG_TYPE_COUNT; |
481 |
ts->next_free_temp = s->first_free_temp[k]; |
482 |
s->first_free_temp[k] = idx; |
483 |
} |
484 |
|
485 |
void tcg_temp_free_i32(TCGv_i32 arg)
|
486 |
{ |
487 |
tcg_temp_free_internal(GET_TCGV_I32(arg)); |
488 |
} |
489 |
|
490 |
void tcg_temp_free_i64(TCGv_i64 arg)
|
491 |
{ |
492 |
tcg_temp_free_internal(GET_TCGV_I64(arg)); |
493 |
} |
494 |
|
495 |
TCGv_i32 tcg_const_i32(int32_t val) |
496 |
{ |
497 |
TCGv_i32 t0; |
498 |
t0 = tcg_temp_new_i32(); |
499 |
tcg_gen_movi_i32(t0, val); |
500 |
return t0;
|
501 |
} |
502 |
|
503 |
TCGv_i64 tcg_const_i64(int64_t val) |
504 |
{ |
505 |
TCGv_i64 t0; |
506 |
t0 = tcg_temp_new_i64(); |
507 |
tcg_gen_movi_i64(t0, val); |
508 |
return t0;
|
509 |
} |
510 |
|
511 |
TCGv_i32 tcg_const_local_i32(int32_t val) |
512 |
{ |
513 |
TCGv_i32 t0; |
514 |
t0 = tcg_temp_local_new_i32(); |
515 |
tcg_gen_movi_i32(t0, val); |
516 |
return t0;
|
517 |
} |
518 |
|
519 |
TCGv_i64 tcg_const_local_i64(int64_t val) |
520 |
{ |
521 |
TCGv_i64 t0; |
522 |
t0 = tcg_temp_local_new_i64(); |
523 |
tcg_gen_movi_i64(t0, val); |
524 |
return t0;
|
525 |
} |
526 |
|
527 |
void tcg_register_helper(void *func, const char *name) |
528 |
{ |
529 |
TCGContext *s = &tcg_ctx; |
530 |
int n;
|
531 |
if ((s->nb_helpers + 1) > s->allocated_helpers) { |
532 |
n = s->allocated_helpers; |
533 |
if (n == 0) { |
534 |
n = 4;
|
535 |
} else {
|
536 |
n *= 2;
|
537 |
} |
538 |
s->helpers = realloc(s->helpers, n * sizeof(TCGHelperInfo));
|
539 |
s->allocated_helpers = n; |
540 |
} |
541 |
s->helpers[s->nb_helpers].func = (tcg_target_ulong)func; |
542 |
s->helpers[s->nb_helpers].name = name; |
543 |
s->nb_helpers++; |
544 |
} |
545 |
|
546 |
/* Note: we convert the 64 bit args to 32 bit and do some alignment
|
547 |
and endian swap. Maybe it would be better to do the alignment
|
548 |
and endian swap in tcg_reg_alloc_call(). */
|
549 |
void tcg_gen_callN(TCGContext *s, TCGv_ptr func, unsigned int flags, |
550 |
int sizemask, TCGArg ret, int nargs, TCGArg *args) |
551 |
{ |
552 |
int call_type;
|
553 |
int i;
|
554 |
int real_args;
|
555 |
int nb_rets;
|
556 |
TCGArg *nparam; |
557 |
*gen_opc_ptr++ = INDEX_op_call; |
558 |
nparam = gen_opparam_ptr++; |
559 |
call_type = (flags & TCG_CALL_TYPE_MASK); |
560 |
if (ret != TCG_CALL_DUMMY_ARG) {
|
561 |
#if TCG_TARGET_REG_BITS < 64 |
562 |
if (sizemask & 1) { |
563 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
564 |
*gen_opparam_ptr++ = ret + 1;
|
565 |
*gen_opparam_ptr++ = ret; |
566 |
#else
|
567 |
*gen_opparam_ptr++ = ret; |
568 |
*gen_opparam_ptr++ = ret + 1;
|
569 |
#endif
|
570 |
nb_rets = 2;
|
571 |
} else
|
572 |
#endif
|
573 |
{ |
574 |
*gen_opparam_ptr++ = ret; |
575 |
nb_rets = 1;
|
576 |
} |
577 |
} else {
|
578 |
nb_rets = 0;
|
579 |
} |
580 |
real_args = 0;
|
581 |
for (i = 0; i < nargs; i++) { |
582 |
#if TCG_TARGET_REG_BITS < 64 |
583 |
if (sizemask & (2 << i)) { |
584 |
#ifdef TCG_TARGET_I386
|
585 |
/* REGPARM case: if the third parameter is 64 bit, it is
|
586 |
allocated on the stack */
|
587 |
if (i == 2 && call_type == TCG_CALL_TYPE_REGPARM) { |
588 |
call_type = TCG_CALL_TYPE_REGPARM_2; |
589 |
flags = (flags & ~TCG_CALL_TYPE_MASK) | call_type; |
590 |
} |
591 |
#endif
|
592 |
#ifdef TCG_TARGET_CALL_ALIGN_ARGS
|
593 |
/* some targets want aligned 64 bit args */
|
594 |
if (real_args & 1) { |
595 |
*gen_opparam_ptr++ = TCG_CALL_DUMMY_ARG; |
596 |
real_args++; |
597 |
} |
598 |
#endif
|
599 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
600 |
*gen_opparam_ptr++ = args[i] + 1;
|
601 |
*gen_opparam_ptr++ = args[i]; |
602 |
#else
|
603 |
*gen_opparam_ptr++ = args[i]; |
604 |
*gen_opparam_ptr++ = args[i] + 1;
|
605 |
#endif
|
606 |
real_args += 2;
|
607 |
} else
|
608 |
#endif
|
609 |
{ |
610 |
*gen_opparam_ptr++ = args[i]; |
611 |
real_args++; |
612 |
} |
613 |
} |
614 |
*gen_opparam_ptr++ = GET_TCGV_PTR(func); |
615 |
|
616 |
*gen_opparam_ptr++ = flags; |
617 |
|
618 |
*nparam = (nb_rets << 16) | (real_args + 1); |
619 |
|
620 |
/* total parameters, needed to go backward in the instruction stream */
|
621 |
*gen_opparam_ptr++ = 1 + nb_rets + real_args + 3; |
622 |
} |
623 |
|
624 |
#if TCG_TARGET_REG_BITS == 32 |
625 |
void tcg_gen_shifti_i64(TCGv_i64 ret, TCGv_i64 arg1,
|
626 |
int c, int right, int arith) |
627 |
{ |
628 |
if (c == 0) { |
629 |
tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1)); |
630 |
tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1)); |
631 |
} else if (c >= 32) { |
632 |
c -= 32;
|
633 |
if (right) {
|
634 |
if (arith) {
|
635 |
tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
636 |
tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31);
|
637 |
} else {
|
638 |
tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
639 |
tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
|
640 |
} |
641 |
} else {
|
642 |
tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_LOW(arg1), c); |
643 |
tcg_gen_movi_i32(TCGV_LOW(ret), 0);
|
644 |
} |
645 |
} else {
|
646 |
TCGv_i32 t0, t1; |
647 |
|
648 |
t0 = tcg_temp_new_i32(); |
649 |
t1 = tcg_temp_new_i32(); |
650 |
if (right) {
|
651 |
tcg_gen_shli_i32(t0, TCGV_HIGH(arg1), 32 - c);
|
652 |
if (arith)
|
653 |
tcg_gen_sari_i32(t1, TCGV_HIGH(arg1), c); |
654 |
else
|
655 |
tcg_gen_shri_i32(t1, TCGV_HIGH(arg1), c); |
656 |
tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c); |
657 |
tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(ret), t0); |
658 |
tcg_gen_mov_i32(TCGV_HIGH(ret), t1); |
659 |
} else {
|
660 |
tcg_gen_shri_i32(t0, TCGV_LOW(arg1), 32 - c);
|
661 |
/* Note: ret can be the same as arg1, so we use t1 */
|
662 |
tcg_gen_shli_i32(t1, TCGV_LOW(arg1), c); |
663 |
tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c); |
664 |
tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t0); |
665 |
tcg_gen_mov_i32(TCGV_LOW(ret), t1); |
666 |
} |
667 |
tcg_temp_free_i32(t0); |
668 |
tcg_temp_free_i32(t1); |
669 |
} |
670 |
} |
671 |
#endif
|
672 |
|
673 |
static void tcg_reg_alloc_start(TCGContext *s) |
674 |
{ |
675 |
int i;
|
676 |
TCGTemp *ts; |
677 |
for(i = 0; i < s->nb_globals; i++) { |
678 |
ts = &s->temps[i]; |
679 |
if (ts->fixed_reg) {
|
680 |
ts->val_type = TEMP_VAL_REG; |
681 |
} else {
|
682 |
ts->val_type = TEMP_VAL_MEM; |
683 |
} |
684 |
} |
685 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
686 |
ts = &s->temps[i]; |
687 |
ts->val_type = TEMP_VAL_DEAD; |
688 |
ts->mem_allocated = 0;
|
689 |
ts->fixed_reg = 0;
|
690 |
} |
691 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
692 |
s->reg_to_temp[i] = -1;
|
693 |
} |
694 |
} |
695 |
|
696 |
static char *tcg_get_arg_str_idx(TCGContext *s, char *buf, int buf_size, |
697 |
int idx)
|
698 |
{ |
699 |
TCGTemp *ts; |
700 |
|
701 |
ts = &s->temps[idx]; |
702 |
if (idx < s->nb_globals) {
|
703 |
pstrcpy(buf, buf_size, ts->name); |
704 |
} else {
|
705 |
if (ts->temp_local)
|
706 |
snprintf(buf, buf_size, "loc%d", idx - s->nb_globals);
|
707 |
else
|
708 |
snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals);
|
709 |
} |
710 |
return buf;
|
711 |
} |
712 |
|
713 |
char *tcg_get_arg_str_i32(TCGContext *s, char *buf, int buf_size, TCGv_i32 arg) |
714 |
{ |
715 |
return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I32(arg));
|
716 |
} |
717 |
|
718 |
char *tcg_get_arg_str_i64(TCGContext *s, char *buf, int buf_size, TCGv_i64 arg) |
719 |
{ |
720 |
return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I64(arg));
|
721 |
} |
722 |
|
723 |
static int helper_cmp(const void *p1, const void *p2) |
724 |
{ |
725 |
const TCGHelperInfo *th1 = p1;
|
726 |
const TCGHelperInfo *th2 = p2;
|
727 |
if (th1->func < th2->func)
|
728 |
return -1; |
729 |
else if (th1->func == th2->func) |
730 |
return 0; |
731 |
else
|
732 |
return 1; |
733 |
} |
734 |
|
735 |
/* find helper definition (Note: A hash table would be better) */
|
736 |
static TCGHelperInfo *tcg_find_helper(TCGContext *s, tcg_target_ulong val)
|
737 |
{ |
738 |
int m, m_min, m_max;
|
739 |
TCGHelperInfo *th; |
740 |
tcg_target_ulong v; |
741 |
|
742 |
if (unlikely(!s->helpers_sorted)) {
|
743 |
qsort(s->helpers, s->nb_helpers, sizeof(TCGHelperInfo),
|
744 |
helper_cmp); |
745 |
s->helpers_sorted = 1;
|
746 |
} |
747 |
|
748 |
/* binary search */
|
749 |
m_min = 0;
|
750 |
m_max = s->nb_helpers - 1;
|
751 |
while (m_min <= m_max) {
|
752 |
m = (m_min + m_max) >> 1;
|
753 |
th = &s->helpers[m]; |
754 |
v = th->func; |
755 |
if (v == val)
|
756 |
return th;
|
757 |
else if (val < v) { |
758 |
m_max = m - 1;
|
759 |
} else {
|
760 |
m_min = m + 1;
|
761 |
} |
762 |
} |
763 |
return NULL; |
764 |
} |
765 |
|
766 |
static const char * const cond_name[] = |
767 |
{ |
768 |
[TCG_COND_EQ] = "eq",
|
769 |
[TCG_COND_NE] = "ne",
|
770 |
[TCG_COND_LT] = "lt",
|
771 |
[TCG_COND_GE] = "ge",
|
772 |
[TCG_COND_LE] = "le",
|
773 |
[TCG_COND_GT] = "gt",
|
774 |
[TCG_COND_LTU] = "ltu",
|
775 |
[TCG_COND_GEU] = "geu",
|
776 |
[TCG_COND_LEU] = "leu",
|
777 |
[TCG_COND_GTU] = "gtu"
|
778 |
}; |
779 |
|
780 |
void tcg_dump_ops(TCGContext *s, FILE *outfile)
|
781 |
{ |
782 |
const uint16_t *opc_ptr;
|
783 |
const TCGArg *args;
|
784 |
TCGArg arg; |
785 |
int c, i, k, nb_oargs, nb_iargs, nb_cargs, first_insn;
|
786 |
const TCGOpDef *def;
|
787 |
char buf[128]; |
788 |
|
789 |
first_insn = 1;
|
790 |
opc_ptr = gen_opc_buf; |
791 |
args = gen_opparam_buf; |
792 |
while (opc_ptr < gen_opc_ptr) {
|
793 |
c = *opc_ptr++; |
794 |
def = &tcg_op_defs[c]; |
795 |
if (c == INDEX_op_debug_insn_start) {
|
796 |
uint64_t pc; |
797 |
#if TARGET_LONG_BITS > TCG_TARGET_REG_BITS
|
798 |
pc = ((uint64_t)args[1] << 32) | args[0]; |
799 |
#else
|
800 |
pc = args[0];
|
801 |
#endif
|
802 |
if (!first_insn)
|
803 |
fprintf(outfile, "\n");
|
804 |
fprintf(outfile, " ---- 0x%" PRIx64, pc);
|
805 |
first_insn = 0;
|
806 |
nb_oargs = def->nb_oargs; |
807 |
nb_iargs = def->nb_iargs; |
808 |
nb_cargs = def->nb_cargs; |
809 |
} else if (c == INDEX_op_call) { |
810 |
TCGArg arg; |
811 |
|
812 |
/* variable number of arguments */
|
813 |
arg = *args++; |
814 |
nb_oargs = arg >> 16;
|
815 |
nb_iargs = arg & 0xffff;
|
816 |
nb_cargs = def->nb_cargs; |
817 |
|
818 |
fprintf(outfile, " %s ", def->name);
|
819 |
|
820 |
/* function name */
|
821 |
fprintf(outfile, "%s",
|
822 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + nb_iargs - 1])); |
823 |
/* flags */
|
824 |
fprintf(outfile, ",$0x%" TCG_PRIlx,
|
825 |
args[nb_oargs + nb_iargs]); |
826 |
/* nb out args */
|
827 |
fprintf(outfile, ",$%d", nb_oargs);
|
828 |
for(i = 0; i < nb_oargs; i++) { |
829 |
fprintf(outfile, ",");
|
830 |
fprintf(outfile, "%s",
|
831 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[i]));
|
832 |
} |
833 |
for(i = 0; i < (nb_iargs - 1); i++) { |
834 |
fprintf(outfile, ",");
|
835 |
if (args[nb_oargs + i] == TCG_CALL_DUMMY_ARG) {
|
836 |
fprintf(outfile, "<dummy>");
|
837 |
} else {
|
838 |
fprintf(outfile, "%s",
|
839 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[nb_oargs + i]));
|
840 |
} |
841 |
} |
842 |
} else if (c == INDEX_op_movi_i32 |
843 |
#if TCG_TARGET_REG_BITS == 64 |
844 |
|| c == INDEX_op_movi_i64 |
845 |
#endif
|
846 |
) { |
847 |
tcg_target_ulong val; |
848 |
TCGHelperInfo *th; |
849 |
|
850 |
nb_oargs = def->nb_oargs; |
851 |
nb_iargs = def->nb_iargs; |
852 |
nb_cargs = def->nb_cargs; |
853 |
fprintf(outfile, " %s %s,$", def->name,
|
854 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[0])); |
855 |
val = args[1];
|
856 |
th = tcg_find_helper(s, val); |
857 |
if (th) {
|
858 |
fprintf(outfile, "%s", th->name);
|
859 |
} else {
|
860 |
if (c == INDEX_op_movi_i32)
|
861 |
fprintf(outfile, "0x%x", (uint32_t)val);
|
862 |
else
|
863 |
fprintf(outfile, "0x%" PRIx64 , (uint64_t)val);
|
864 |
} |
865 |
} else {
|
866 |
fprintf(outfile, " %s ", def->name);
|
867 |
if (c == INDEX_op_nopn) {
|
868 |
/* variable number of arguments */
|
869 |
nb_cargs = *args; |
870 |
nb_oargs = 0;
|
871 |
nb_iargs = 0;
|
872 |
} else {
|
873 |
nb_oargs = def->nb_oargs; |
874 |
nb_iargs = def->nb_iargs; |
875 |
nb_cargs = def->nb_cargs; |
876 |
} |
877 |
|
878 |
k = 0;
|
879 |
for(i = 0; i < nb_oargs; i++) { |
880 |
if (k != 0) |
881 |
fprintf(outfile, ",");
|
882 |
fprintf(outfile, "%s",
|
883 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++]));
|
884 |
} |
885 |
for(i = 0; i < nb_iargs; i++) { |
886 |
if (k != 0) |
887 |
fprintf(outfile, ",");
|
888 |
fprintf(outfile, "%s",
|
889 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[k++]));
|
890 |
} |
891 |
if (c == INDEX_op_brcond_i32
|
892 |
#if TCG_TARGET_REG_BITS == 32 |
893 |
|| c == INDEX_op_brcond2_i32 |
894 |
#elif TCG_TARGET_REG_BITS == 64 |
895 |
|| c == INDEX_op_brcond_i64 |
896 |
#endif
|
897 |
) { |
898 |
if (args[k] < ARRAY_SIZE(cond_name) && cond_name[args[k]])
|
899 |
fprintf(outfile, ",%s", cond_name[args[k++]]);
|
900 |
else
|
901 |
fprintf(outfile, ",$0x%" TCG_PRIlx, args[k++]);
|
902 |
i = 1;
|
903 |
} |
904 |
else
|
905 |
i = 0;
|
906 |
for(; i < nb_cargs; i++) {
|
907 |
if (k != 0) |
908 |
fprintf(outfile, ",");
|
909 |
arg = args[k++]; |
910 |
fprintf(outfile, "$0x%" TCG_PRIlx, arg);
|
911 |
} |
912 |
} |
913 |
fprintf(outfile, "\n");
|
914 |
args += nb_iargs + nb_oargs + nb_cargs; |
915 |
} |
916 |
} |
917 |
|
918 |
/* we give more priority to constraints with less registers */
|
919 |
static int get_constraint_priority(const TCGOpDef *def, int k) |
920 |
{ |
921 |
const TCGArgConstraint *arg_ct;
|
922 |
|
923 |
int i, n;
|
924 |
arg_ct = &def->args_ct[k]; |
925 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
926 |
/* an alias is equivalent to a single register */
|
927 |
n = 1;
|
928 |
} else {
|
929 |
if (!(arg_ct->ct & TCG_CT_REG))
|
930 |
return 0; |
931 |
n = 0;
|
932 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
933 |
if (tcg_regset_test_reg(arg_ct->u.regs, i))
|
934 |
n++; |
935 |
} |
936 |
} |
937 |
return TCG_TARGET_NB_REGS - n + 1; |
938 |
} |
939 |
|
940 |
/* sort from highest priority to lowest */
|
941 |
static void sort_constraints(TCGOpDef *def, int start, int n) |
942 |
{ |
943 |
int i, j, p1, p2, tmp;
|
944 |
|
945 |
for(i = 0; i < n; i++) |
946 |
def->sorted_args[start + i] = start + i; |
947 |
if (n <= 1) |
948 |
return;
|
949 |
for(i = 0; i < n - 1; i++) { |
950 |
for(j = i + 1; j < n; j++) { |
951 |
p1 = get_constraint_priority(def, def->sorted_args[start + i]); |
952 |
p2 = get_constraint_priority(def, def->sorted_args[start + j]); |
953 |
if (p1 < p2) {
|
954 |
tmp = def->sorted_args[start + i]; |
955 |
def->sorted_args[start + i] = def->sorted_args[start + j]; |
956 |
def->sorted_args[start + j] = tmp; |
957 |
} |
958 |
} |
959 |
} |
960 |
} |
961 |
|
962 |
void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs) |
963 |
{ |
964 |
int op;
|
965 |
TCGOpDef *def; |
966 |
const char *ct_str; |
967 |
int i, nb_args;
|
968 |
|
969 |
for(;;) {
|
970 |
if (tdefs->op < 0) |
971 |
break;
|
972 |
op = tdefs->op; |
973 |
assert(op >= 0 && op < NB_OPS);
|
974 |
def = &tcg_op_defs[op]; |
975 |
nb_args = def->nb_iargs + def->nb_oargs; |
976 |
for(i = 0; i < nb_args; i++) { |
977 |
ct_str = tdefs->args_ct_str[i]; |
978 |
tcg_regset_clear(def->args_ct[i].u.regs); |
979 |
def->args_ct[i].ct = 0;
|
980 |
if (ct_str[0] >= '0' && ct_str[0] <= '9') { |
981 |
int oarg;
|
982 |
oarg = ct_str[0] - '0'; |
983 |
assert(oarg < def->nb_oargs); |
984 |
assert(def->args_ct[oarg].ct & TCG_CT_REG); |
985 |
/* TCG_CT_ALIAS is for the output arguments. The input
|
986 |
argument is tagged with TCG_CT_IALIAS. */
|
987 |
def->args_ct[i] = def->args_ct[oarg]; |
988 |
def->args_ct[oarg].ct = TCG_CT_ALIAS; |
989 |
def->args_ct[oarg].alias_index = i; |
990 |
def->args_ct[i].ct |= TCG_CT_IALIAS; |
991 |
def->args_ct[i].alias_index = oarg; |
992 |
} else {
|
993 |
for(;;) {
|
994 |
if (*ct_str == '\0') |
995 |
break;
|
996 |
switch(*ct_str) {
|
997 |
case 'i': |
998 |
def->args_ct[i].ct |= TCG_CT_CONST; |
999 |
ct_str++; |
1000 |
break;
|
1001 |
default:
|
1002 |
if (target_parse_constraint(&def->args_ct[i], &ct_str) < 0) { |
1003 |
fprintf(stderr, "Invalid constraint '%s' for arg %d of operation '%s'\n",
|
1004 |
ct_str, i, def->name); |
1005 |
exit(1);
|
1006 |
} |
1007 |
} |
1008 |
} |
1009 |
} |
1010 |
} |
1011 |
|
1012 |
/* sort the constraints (XXX: this is just an heuristic) */
|
1013 |
sort_constraints(def, 0, def->nb_oargs);
|
1014 |
sort_constraints(def, def->nb_oargs, def->nb_iargs); |
1015 |
|
1016 |
#if 0
|
1017 |
{
|
1018 |
int i;
|
1019 |
|
1020 |
printf("%s: sorted=", def->name);
|
1021 |
for(i = 0; i < def->nb_oargs + def->nb_iargs; i++)
|
1022 |
printf(" %d", def->sorted_args[i]);
|
1023 |
printf("\n");
|
1024 |
}
|
1025 |
#endif
|
1026 |
tdefs++; |
1027 |
} |
1028 |
|
1029 |
} |
1030 |
|
1031 |
#ifdef USE_LIVENESS_ANALYSIS
|
1032 |
|
1033 |
/* set a nop for an operation using 'nb_args' */
|
1034 |
static inline void tcg_set_nop(TCGContext *s, uint16_t *opc_ptr, |
1035 |
TCGArg *args, int nb_args)
|
1036 |
{ |
1037 |
if (nb_args == 0) { |
1038 |
*opc_ptr = INDEX_op_nop; |
1039 |
} else {
|
1040 |
*opc_ptr = INDEX_op_nopn; |
1041 |
args[0] = nb_args;
|
1042 |
args[nb_args - 1] = nb_args;
|
1043 |
} |
1044 |
} |
1045 |
|
1046 |
/* liveness analysis: end of function: globals are live, temps are
|
1047 |
dead. */
|
1048 |
/* XXX: at this stage, not used as there would be little gains because
|
1049 |
most TBs end with a conditional jump. */
|
1050 |
static inline void tcg_la_func_end(TCGContext *s, uint8_t *dead_temps) |
1051 |
{ |
1052 |
memset(dead_temps, 0, s->nb_globals);
|
1053 |
memset(dead_temps + s->nb_globals, 1, s->nb_temps - s->nb_globals);
|
1054 |
} |
1055 |
|
1056 |
/* liveness analysis: end of basic block: globals are live, temps are
|
1057 |
dead, local temps are live. */
|
1058 |
static inline void tcg_la_bb_end(TCGContext *s, uint8_t *dead_temps) |
1059 |
{ |
1060 |
int i;
|
1061 |
TCGTemp *ts; |
1062 |
|
1063 |
memset(dead_temps, 0, s->nb_globals);
|
1064 |
ts = &s->temps[s->nb_globals]; |
1065 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1066 |
if (ts->temp_local)
|
1067 |
dead_temps[i] = 0;
|
1068 |
else
|
1069 |
dead_temps[i] = 1;
|
1070 |
ts++; |
1071 |
} |
1072 |
} |
1073 |
|
1074 |
/* Liveness analysis : update the opc_dead_iargs array to tell if a
|
1075 |
given input arguments is dead. Instructions updating dead
|
1076 |
temporaries are removed. */
|
1077 |
static void tcg_liveness_analysis(TCGContext *s) |
1078 |
{ |
1079 |
int i, op_index, op, nb_args, nb_iargs, nb_oargs, arg, nb_ops;
|
1080 |
TCGArg *args; |
1081 |
const TCGOpDef *def;
|
1082 |
uint8_t *dead_temps; |
1083 |
unsigned int dead_iargs; |
1084 |
|
1085 |
gen_opc_ptr++; /* skip end */
|
1086 |
|
1087 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1088 |
|
1089 |
s->op_dead_iargs = tcg_malloc(nb_ops * sizeof(uint16_t));
|
1090 |
|
1091 |
dead_temps = tcg_malloc(s->nb_temps); |
1092 |
memset(dead_temps, 1, s->nb_temps);
|
1093 |
|
1094 |
args = gen_opparam_ptr; |
1095 |
op_index = nb_ops - 1;
|
1096 |
while (op_index >= 0) { |
1097 |
op = gen_opc_buf[op_index]; |
1098 |
def = &tcg_op_defs[op]; |
1099 |
switch(op) {
|
1100 |
case INDEX_op_call:
|
1101 |
{ |
1102 |
int call_flags;
|
1103 |
|
1104 |
nb_args = args[-1];
|
1105 |
args -= nb_args; |
1106 |
nb_iargs = args[0] & 0xffff; |
1107 |
nb_oargs = args[0] >> 16; |
1108 |
args++; |
1109 |
call_flags = args[nb_oargs + nb_iargs]; |
1110 |
|
1111 |
/* pure functions can be removed if their result is not
|
1112 |
used */
|
1113 |
if (call_flags & TCG_CALL_PURE) {
|
1114 |
for(i = 0; i < nb_oargs; i++) { |
1115 |
arg = args[i]; |
1116 |
if (!dead_temps[arg])
|
1117 |
goto do_not_remove_call;
|
1118 |
} |
1119 |
tcg_set_nop(s, gen_opc_buf + op_index, |
1120 |
args - 1, nb_args);
|
1121 |
} else {
|
1122 |
do_not_remove_call:
|
1123 |
|
1124 |
/* output args are dead */
|
1125 |
for(i = 0; i < nb_oargs; i++) { |
1126 |
arg = args[i]; |
1127 |
dead_temps[arg] = 1;
|
1128 |
} |
1129 |
|
1130 |
if (!(call_flags & TCG_CALL_CONST)) {
|
1131 |
/* globals are live (they may be used by the call) */
|
1132 |
memset(dead_temps, 0, s->nb_globals);
|
1133 |
} |
1134 |
|
1135 |
/* input args are live */
|
1136 |
dead_iargs = 0;
|
1137 |
for(i = 0; i < nb_iargs; i++) { |
1138 |
arg = args[i + nb_oargs]; |
1139 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1140 |
if (dead_temps[arg]) {
|
1141 |
dead_iargs |= (1 << i);
|
1142 |
} |
1143 |
dead_temps[arg] = 0;
|
1144 |
} |
1145 |
} |
1146 |
s->op_dead_iargs[op_index] = dead_iargs; |
1147 |
} |
1148 |
args--; |
1149 |
} |
1150 |
break;
|
1151 |
case INDEX_op_set_label:
|
1152 |
args--; |
1153 |
/* mark end of basic block */
|
1154 |
tcg_la_bb_end(s, dead_temps); |
1155 |
break;
|
1156 |
case INDEX_op_debug_insn_start:
|
1157 |
args -= def->nb_args; |
1158 |
break;
|
1159 |
case INDEX_op_nopn:
|
1160 |
nb_args = args[-1];
|
1161 |
args -= nb_args; |
1162 |
break;
|
1163 |
case INDEX_op_discard:
|
1164 |
args--; |
1165 |
/* mark the temporary as dead */
|
1166 |
dead_temps[args[0]] = 1; |
1167 |
break;
|
1168 |
case INDEX_op_end:
|
1169 |
break;
|
1170 |
/* XXX: optimize by hardcoding common cases (e.g. triadic ops) */
|
1171 |
default:
|
1172 |
args -= def->nb_args; |
1173 |
nb_iargs = def->nb_iargs; |
1174 |
nb_oargs = def->nb_oargs; |
1175 |
|
1176 |
/* Test if the operation can be removed because all
|
1177 |
its outputs are dead. We assume that nb_oargs == 0
|
1178 |
implies side effects */
|
1179 |
if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { |
1180 |
for(i = 0; i < nb_oargs; i++) { |
1181 |
arg = args[i]; |
1182 |
if (!dead_temps[arg])
|
1183 |
goto do_not_remove;
|
1184 |
} |
1185 |
tcg_set_nop(s, gen_opc_buf + op_index, args, def->nb_args); |
1186 |
#ifdef CONFIG_PROFILER
|
1187 |
s->del_op_count++; |
1188 |
#endif
|
1189 |
} else {
|
1190 |
do_not_remove:
|
1191 |
|
1192 |
/* output args are dead */
|
1193 |
for(i = 0; i < nb_oargs; i++) { |
1194 |
arg = args[i]; |
1195 |
dead_temps[arg] = 1;
|
1196 |
} |
1197 |
|
1198 |
/* if end of basic block, update */
|
1199 |
if (def->flags & TCG_OPF_BB_END) {
|
1200 |
tcg_la_bb_end(s, dead_temps); |
1201 |
} else if (def->flags & TCG_OPF_CALL_CLOBBER) { |
1202 |
/* globals are live */
|
1203 |
memset(dead_temps, 0, s->nb_globals);
|
1204 |
} |
1205 |
|
1206 |
/* input args are live */
|
1207 |
dead_iargs = 0;
|
1208 |
for(i = 0; i < nb_iargs; i++) { |
1209 |
arg = args[i + nb_oargs]; |
1210 |
if (dead_temps[arg]) {
|
1211 |
dead_iargs |= (1 << i);
|
1212 |
} |
1213 |
dead_temps[arg] = 0;
|
1214 |
} |
1215 |
s->op_dead_iargs[op_index] = dead_iargs; |
1216 |
} |
1217 |
break;
|
1218 |
} |
1219 |
op_index--; |
1220 |
} |
1221 |
|
1222 |
if (args != gen_opparam_buf)
|
1223 |
tcg_abort(); |
1224 |
} |
1225 |
#else
|
1226 |
/* dummy liveness analysis */
|
1227 |
void tcg_liveness_analysis(TCGContext *s)
|
1228 |
{ |
1229 |
int nb_ops;
|
1230 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1231 |
|
1232 |
s->op_dead_iargs = tcg_malloc(nb_ops * sizeof(uint16_t));
|
1233 |
memset(s->op_dead_iargs, 0, nb_ops * sizeof(uint16_t)); |
1234 |
} |
1235 |
#endif
|
1236 |
|
1237 |
#ifndef NDEBUG
|
1238 |
static void dump_regs(TCGContext *s) |
1239 |
{ |
1240 |
TCGTemp *ts; |
1241 |
int i;
|
1242 |
char buf[64]; |
1243 |
|
1244 |
for(i = 0; i < s->nb_temps; i++) { |
1245 |
ts = &s->temps[i]; |
1246 |
printf(" %10s: ", tcg_get_arg_str_idx(s, buf, sizeof(buf), i)); |
1247 |
switch(ts->val_type) {
|
1248 |
case TEMP_VAL_REG:
|
1249 |
printf("%s", tcg_target_reg_names[ts->reg]);
|
1250 |
break;
|
1251 |
case TEMP_VAL_MEM:
|
1252 |
printf("%d(%s)", (int)ts->mem_offset, tcg_target_reg_names[ts->mem_reg]); |
1253 |
break;
|
1254 |
case TEMP_VAL_CONST:
|
1255 |
printf("$0x%" TCG_PRIlx, ts->val);
|
1256 |
break;
|
1257 |
case TEMP_VAL_DEAD:
|
1258 |
printf("D");
|
1259 |
break;
|
1260 |
default:
|
1261 |
printf("???");
|
1262 |
break;
|
1263 |
} |
1264 |
printf("\n");
|
1265 |
} |
1266 |
|
1267 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
1268 |
if (s->reg_to_temp[i] >= 0) { |
1269 |
printf("%s: %s\n",
|
1270 |
tcg_target_reg_names[i], |
1271 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), s->reg_to_temp[i]));
|
1272 |
} |
1273 |
} |
1274 |
} |
1275 |
|
1276 |
static void check_regs(TCGContext *s) |
1277 |
{ |
1278 |
int reg, k;
|
1279 |
TCGTemp *ts; |
1280 |
char buf[64]; |
1281 |
|
1282 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1283 |
k = s->reg_to_temp[reg]; |
1284 |
if (k >= 0) { |
1285 |
ts = &s->temps[k]; |
1286 |
if (ts->val_type != TEMP_VAL_REG ||
|
1287 |
ts->reg != reg) { |
1288 |
printf("Inconsistency for register %s:\n",
|
1289 |
tcg_target_reg_names[reg]); |
1290 |
goto fail;
|
1291 |
} |
1292 |
} |
1293 |
} |
1294 |
for(k = 0; k < s->nb_temps; k++) { |
1295 |
ts = &s->temps[k]; |
1296 |
if (ts->val_type == TEMP_VAL_REG &&
|
1297 |
!ts->fixed_reg && |
1298 |
s->reg_to_temp[ts->reg] != k) { |
1299 |
printf("Inconsistency for temp %s:\n",
|
1300 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), k));
|
1301 |
fail:
|
1302 |
printf("reg state:\n");
|
1303 |
dump_regs(s); |
1304 |
tcg_abort(); |
1305 |
} |
1306 |
} |
1307 |
} |
1308 |
#endif
|
1309 |
|
1310 |
static void temp_allocate_frame(TCGContext *s, int temp) |
1311 |
{ |
1312 |
TCGTemp *ts; |
1313 |
ts = &s->temps[temp]; |
1314 |
s->current_frame_offset = (s->current_frame_offset + sizeof(tcg_target_long) - 1) & ~(sizeof(tcg_target_long) - 1); |
1315 |
if (s->current_frame_offset + sizeof(tcg_target_long) > s->frame_end) |
1316 |
tcg_abort(); |
1317 |
ts->mem_offset = s->current_frame_offset; |
1318 |
ts->mem_reg = s->frame_reg; |
1319 |
ts->mem_allocated = 1;
|
1320 |
s->current_frame_offset += sizeof(tcg_target_long);
|
1321 |
} |
1322 |
|
1323 |
/* free register 'reg' by spilling the corresponding temporary if necessary */
|
1324 |
static void tcg_reg_free(TCGContext *s, int reg) |
1325 |
{ |
1326 |
TCGTemp *ts; |
1327 |
int temp;
|
1328 |
|
1329 |
temp = s->reg_to_temp[reg]; |
1330 |
if (temp != -1) { |
1331 |
ts = &s->temps[temp]; |
1332 |
assert(ts->val_type == TEMP_VAL_REG); |
1333 |
if (!ts->mem_coherent) {
|
1334 |
if (!ts->mem_allocated)
|
1335 |
temp_allocate_frame(s, temp); |
1336 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1337 |
} |
1338 |
ts->val_type = TEMP_VAL_MEM; |
1339 |
s->reg_to_temp[reg] = -1;
|
1340 |
} |
1341 |
} |
1342 |
|
1343 |
/* Allocate a register belonging to reg1 & ~reg2 */
|
1344 |
static int tcg_reg_alloc(TCGContext *s, TCGRegSet reg1, TCGRegSet reg2) |
1345 |
{ |
1346 |
int i, reg;
|
1347 |
TCGRegSet reg_ct; |
1348 |
|
1349 |
tcg_regset_andnot(reg_ct, reg1, reg2); |
1350 |
|
1351 |
/* first try free registers */
|
1352 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1353 |
reg = tcg_target_reg_alloc_order[i]; |
1354 |
if (tcg_regset_test_reg(reg_ct, reg) && s->reg_to_temp[reg] == -1) |
1355 |
return reg;
|
1356 |
} |
1357 |
|
1358 |
/* XXX: do better spill choice */
|
1359 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1360 |
reg = tcg_target_reg_alloc_order[i]; |
1361 |
if (tcg_regset_test_reg(reg_ct, reg)) {
|
1362 |
tcg_reg_free(s, reg); |
1363 |
return reg;
|
1364 |
} |
1365 |
} |
1366 |
|
1367 |
tcg_abort(); |
1368 |
} |
1369 |
|
1370 |
/* save a temporary to memory. 'allocated_regs' is used in case a
|
1371 |
temporary registers needs to be allocated to store a constant. */
|
1372 |
static void temp_save(TCGContext *s, int temp, TCGRegSet allocated_regs) |
1373 |
{ |
1374 |
TCGTemp *ts; |
1375 |
int reg;
|
1376 |
|
1377 |
ts = &s->temps[temp]; |
1378 |
if (!ts->fixed_reg) {
|
1379 |
switch(ts->val_type) {
|
1380 |
case TEMP_VAL_REG:
|
1381 |
tcg_reg_free(s, ts->reg); |
1382 |
break;
|
1383 |
case TEMP_VAL_DEAD:
|
1384 |
ts->val_type = TEMP_VAL_MEM; |
1385 |
break;
|
1386 |
case TEMP_VAL_CONST:
|
1387 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1388 |
allocated_regs); |
1389 |
if (!ts->mem_allocated)
|
1390 |
temp_allocate_frame(s, temp); |
1391 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1392 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1393 |
ts->val_type = TEMP_VAL_MEM; |
1394 |
break;
|
1395 |
case TEMP_VAL_MEM:
|
1396 |
break;
|
1397 |
default:
|
1398 |
tcg_abort(); |
1399 |
} |
1400 |
} |
1401 |
} |
1402 |
|
1403 |
/* save globals to their cannonical location and assume they can be
|
1404 |
modified be the following code. 'allocated_regs' is used in case a
|
1405 |
temporary registers needs to be allocated to store a constant. */
|
1406 |
static void save_globals(TCGContext *s, TCGRegSet allocated_regs) |
1407 |
{ |
1408 |
int i;
|
1409 |
|
1410 |
for(i = 0; i < s->nb_globals; i++) { |
1411 |
temp_save(s, i, allocated_regs); |
1412 |
} |
1413 |
} |
1414 |
|
1415 |
/* at the end of a basic block, we assume all temporaries are dead and
|
1416 |
all globals are stored at their canonical location. */
|
1417 |
static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) |
1418 |
{ |
1419 |
TCGTemp *ts; |
1420 |
int i;
|
1421 |
|
1422 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1423 |
ts = &s->temps[i]; |
1424 |
if (ts->temp_local) {
|
1425 |
temp_save(s, i, allocated_regs); |
1426 |
} else {
|
1427 |
if (ts->val_type == TEMP_VAL_REG) {
|
1428 |
s->reg_to_temp[ts->reg] = -1;
|
1429 |
} |
1430 |
ts->val_type = TEMP_VAL_DEAD; |
1431 |
} |
1432 |
} |
1433 |
|
1434 |
save_globals(s, allocated_regs); |
1435 |
} |
1436 |
|
1437 |
#define IS_DEAD_IARG(n) ((dead_iargs >> (n)) & 1) |
1438 |
|
1439 |
static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args) |
1440 |
{ |
1441 |
TCGTemp *ots; |
1442 |
tcg_target_ulong val; |
1443 |
|
1444 |
ots = &s->temps[args[0]];
|
1445 |
val = args[1];
|
1446 |
|
1447 |
if (ots->fixed_reg) {
|
1448 |
/* for fixed registers, we do not do any constant
|
1449 |
propagation */
|
1450 |
tcg_out_movi(s, ots->type, ots->reg, val); |
1451 |
} else {
|
1452 |
/* The movi is not explicitly generated here */
|
1453 |
if (ots->val_type == TEMP_VAL_REG)
|
1454 |
s->reg_to_temp[ots->reg] = -1;
|
1455 |
ots->val_type = TEMP_VAL_CONST; |
1456 |
ots->val = val; |
1457 |
} |
1458 |
} |
1459 |
|
1460 |
static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, |
1461 |
const TCGArg *args,
|
1462 |
unsigned int dead_iargs) |
1463 |
{ |
1464 |
TCGTemp *ts, *ots; |
1465 |
int reg;
|
1466 |
const TCGArgConstraint *arg_ct;
|
1467 |
|
1468 |
ots = &s->temps[args[0]];
|
1469 |
ts = &s->temps[args[1]];
|
1470 |
arg_ct = &def->args_ct[0];
|
1471 |
|
1472 |
/* XXX: always mark arg dead if IS_DEAD_IARG(0) */
|
1473 |
if (ts->val_type == TEMP_VAL_REG) {
|
1474 |
if (IS_DEAD_IARG(0) && !ts->fixed_reg && !ots->fixed_reg) { |
1475 |
/* the mov can be suppressed */
|
1476 |
if (ots->val_type == TEMP_VAL_REG)
|
1477 |
s->reg_to_temp[ots->reg] = -1;
|
1478 |
reg = ts->reg; |
1479 |
s->reg_to_temp[reg] = -1;
|
1480 |
ts->val_type = TEMP_VAL_DEAD; |
1481 |
} else {
|
1482 |
if (ots->val_type == TEMP_VAL_REG) {
|
1483 |
reg = ots->reg; |
1484 |
} else {
|
1485 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1486 |
} |
1487 |
if (ts->reg != reg) {
|
1488 |
tcg_out_mov(s, reg, ts->reg); |
1489 |
} |
1490 |
} |
1491 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1492 |
if (ots->val_type == TEMP_VAL_REG) {
|
1493 |
reg = ots->reg; |
1494 |
} else {
|
1495 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1496 |
} |
1497 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1498 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1499 |
if (ots->fixed_reg) {
|
1500 |
reg = ots->reg; |
1501 |
tcg_out_movi(s, ots->type, reg, ts->val); |
1502 |
} else {
|
1503 |
/* propagate constant */
|
1504 |
if (ots->val_type == TEMP_VAL_REG)
|
1505 |
s->reg_to_temp[ots->reg] = -1;
|
1506 |
ots->val_type = TEMP_VAL_CONST; |
1507 |
ots->val = ts->val; |
1508 |
return;
|
1509 |
} |
1510 |
} else {
|
1511 |
tcg_abort(); |
1512 |
} |
1513 |
s->reg_to_temp[reg] = args[0];
|
1514 |
ots->reg = reg; |
1515 |
ots->val_type = TEMP_VAL_REG; |
1516 |
ots->mem_coherent = 0;
|
1517 |
} |
1518 |
|
1519 |
static void tcg_reg_alloc_op(TCGContext *s, |
1520 |
const TCGOpDef *def, int opc, |
1521 |
const TCGArg *args,
|
1522 |
unsigned int dead_iargs) |
1523 |
{ |
1524 |
TCGRegSet allocated_regs; |
1525 |
int i, k, nb_iargs, nb_oargs, reg;
|
1526 |
TCGArg arg; |
1527 |
const TCGArgConstraint *arg_ct;
|
1528 |
TCGTemp *ts; |
1529 |
TCGArg new_args[TCG_MAX_OP_ARGS]; |
1530 |
int const_args[TCG_MAX_OP_ARGS];
|
1531 |
|
1532 |
nb_oargs = def->nb_oargs; |
1533 |
nb_iargs = def->nb_iargs; |
1534 |
|
1535 |
/* copy constants */
|
1536 |
memcpy(new_args + nb_oargs + nb_iargs, |
1537 |
args + nb_oargs + nb_iargs, |
1538 |
sizeof(TCGArg) * def->nb_cargs);
|
1539 |
|
1540 |
/* satisfy input constraints */
|
1541 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1542 |
for(k = 0; k < nb_iargs; k++) { |
1543 |
i = def->sorted_args[nb_oargs + k]; |
1544 |
arg = args[i]; |
1545 |
arg_ct = &def->args_ct[i]; |
1546 |
ts = &s->temps[arg]; |
1547 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1548 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1549 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1550 |
ts->val_type = TEMP_VAL_REG; |
1551 |
ts->reg = reg; |
1552 |
ts->mem_coherent = 1;
|
1553 |
s->reg_to_temp[reg] = arg; |
1554 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1555 |
if (tcg_target_const_match(ts->val, arg_ct)) {
|
1556 |
/* constant is OK for instruction */
|
1557 |
const_args[i] = 1;
|
1558 |
new_args[i] = ts->val; |
1559 |
goto iarg_end;
|
1560 |
} else {
|
1561 |
/* need to move to a register */
|
1562 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1563 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1564 |
ts->val_type = TEMP_VAL_REG; |
1565 |
ts->reg = reg; |
1566 |
ts->mem_coherent = 0;
|
1567 |
s->reg_to_temp[reg] = arg; |
1568 |
} |
1569 |
} |
1570 |
assert(ts->val_type == TEMP_VAL_REG); |
1571 |
if (arg_ct->ct & TCG_CT_IALIAS) {
|
1572 |
if (ts->fixed_reg) {
|
1573 |
/* if fixed register, we must allocate a new register
|
1574 |
if the alias is not the same register */
|
1575 |
if (arg != args[arg_ct->alias_index])
|
1576 |
goto allocate_in_reg;
|
1577 |
} else {
|
1578 |
/* if the input is aliased to an output and if it is
|
1579 |
not dead after the instruction, we must allocate
|
1580 |
a new register and move it */
|
1581 |
if (!IS_DEAD_IARG(i - nb_oargs))
|
1582 |
goto allocate_in_reg;
|
1583 |
} |
1584 |
} |
1585 |
reg = ts->reg; |
1586 |
if (tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1587 |
/* nothing to do : the constraint is satisfied */
|
1588 |
} else {
|
1589 |
allocate_in_reg:
|
1590 |
/* allocate a new register matching the constraint
|
1591 |
and move the temporary register into it */
|
1592 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1593 |
tcg_out_mov(s, reg, ts->reg); |
1594 |
} |
1595 |
new_args[i] = reg; |
1596 |
const_args[i] = 0;
|
1597 |
tcg_regset_set_reg(allocated_regs, reg); |
1598 |
iarg_end: ;
|
1599 |
} |
1600 |
|
1601 |
if (def->flags & TCG_OPF_BB_END) {
|
1602 |
tcg_reg_alloc_bb_end(s, allocated_regs); |
1603 |
} else {
|
1604 |
/* mark dead temporaries and free the associated registers */
|
1605 |
for(i = 0; i < nb_iargs; i++) { |
1606 |
arg = args[nb_oargs + i]; |
1607 |
if (IS_DEAD_IARG(i)) {
|
1608 |
ts = &s->temps[arg]; |
1609 |
if (!ts->fixed_reg) {
|
1610 |
if (ts->val_type == TEMP_VAL_REG)
|
1611 |
s->reg_to_temp[ts->reg] = -1;
|
1612 |
ts->val_type = TEMP_VAL_DEAD; |
1613 |
} |
1614 |
} |
1615 |
} |
1616 |
|
1617 |
if (def->flags & TCG_OPF_CALL_CLOBBER) {
|
1618 |
/* XXX: permit generic clobber register list ? */
|
1619 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1620 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1621 |
tcg_reg_free(s, reg); |
1622 |
} |
1623 |
} |
1624 |
/* XXX: for load/store we could do that only for the slow path
|
1625 |
(i.e. when a memory callback is called) */
|
1626 |
|
1627 |
/* store globals and free associated registers (we assume the insn
|
1628 |
can modify any global. */
|
1629 |
save_globals(s, allocated_regs); |
1630 |
} |
1631 |
|
1632 |
/* satisfy the output constraints */
|
1633 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1634 |
for(k = 0; k < nb_oargs; k++) { |
1635 |
i = def->sorted_args[k]; |
1636 |
arg = args[i]; |
1637 |
arg_ct = &def->args_ct[i]; |
1638 |
ts = &s->temps[arg]; |
1639 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
1640 |
reg = new_args[arg_ct->alias_index]; |
1641 |
} else {
|
1642 |
/* if fixed register, we try to use it */
|
1643 |
reg = ts->reg; |
1644 |
if (ts->fixed_reg &&
|
1645 |
tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
1646 |
goto oarg_end;
|
1647 |
} |
1648 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1649 |
} |
1650 |
tcg_regset_set_reg(allocated_regs, reg); |
1651 |
/* if a fixed register is used, then a move will be done afterwards */
|
1652 |
if (!ts->fixed_reg) {
|
1653 |
if (ts->val_type == TEMP_VAL_REG)
|
1654 |
s->reg_to_temp[ts->reg] = -1;
|
1655 |
ts->val_type = TEMP_VAL_REG; |
1656 |
ts->reg = reg; |
1657 |
/* temp value is modified, so the value kept in memory is
|
1658 |
potentially not the same */
|
1659 |
ts->mem_coherent = 0;
|
1660 |
s->reg_to_temp[reg] = arg; |
1661 |
} |
1662 |
oarg_end:
|
1663 |
new_args[i] = reg; |
1664 |
} |
1665 |
} |
1666 |
|
1667 |
/* emit instruction */
|
1668 |
tcg_out_op(s, opc, new_args, const_args); |
1669 |
|
1670 |
/* move the outputs in the correct register if needed */
|
1671 |
for(i = 0; i < nb_oargs; i++) { |
1672 |
ts = &s->temps[args[i]]; |
1673 |
reg = new_args[i]; |
1674 |
if (ts->fixed_reg && ts->reg != reg) {
|
1675 |
tcg_out_mov(s, ts->reg, reg); |
1676 |
} |
1677 |
} |
1678 |
} |
1679 |
|
1680 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1681 |
#define STACK_DIR(x) (-(x))
|
1682 |
#else
|
1683 |
#define STACK_DIR(x) (x)
|
1684 |
#endif
|
1685 |
|
1686 |
static int tcg_reg_alloc_call(TCGContext *s, const TCGOpDef *def, |
1687 |
int opc, const TCGArg *args, |
1688 |
unsigned int dead_iargs) |
1689 |
{ |
1690 |
int nb_iargs, nb_oargs, flags, nb_regs, i, reg, nb_params;
|
1691 |
TCGArg arg, func_arg; |
1692 |
TCGTemp *ts; |
1693 |
tcg_target_long stack_offset, call_stack_size, func_addr; |
1694 |
int const_func_arg, allocate_args;
|
1695 |
TCGRegSet allocated_regs; |
1696 |
const TCGArgConstraint *arg_ct;
|
1697 |
|
1698 |
arg = *args++; |
1699 |
|
1700 |
nb_oargs = arg >> 16;
|
1701 |
nb_iargs = arg & 0xffff;
|
1702 |
nb_params = nb_iargs - 1;
|
1703 |
|
1704 |
flags = args[nb_oargs + nb_iargs]; |
1705 |
|
1706 |
nb_regs = tcg_target_get_call_iarg_regs_count(flags); |
1707 |
if (nb_regs > nb_params)
|
1708 |
nb_regs = nb_params; |
1709 |
|
1710 |
/* assign stack slots first */
|
1711 |
/* XXX: preallocate call stack */
|
1712 |
call_stack_size = (nb_params - nb_regs) * sizeof(tcg_target_long);
|
1713 |
call_stack_size = (call_stack_size + TCG_TARGET_STACK_ALIGN - 1) &
|
1714 |
~(TCG_TARGET_STACK_ALIGN - 1);
|
1715 |
allocate_args = (call_stack_size > TCG_STATIC_CALL_ARGS_SIZE); |
1716 |
if (allocate_args) {
|
1717 |
tcg_out_addi(s, TCG_REG_CALL_STACK, -STACK_DIR(call_stack_size)); |
1718 |
} |
1719 |
|
1720 |
stack_offset = TCG_TARGET_CALL_STACK_OFFSET; |
1721 |
for(i = nb_regs; i < nb_params; i++) {
|
1722 |
arg = args[nb_oargs + i]; |
1723 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1724 |
stack_offset -= sizeof(tcg_target_long);
|
1725 |
#endif
|
1726 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1727 |
ts = &s->temps[arg]; |
1728 |
if (ts->val_type == TEMP_VAL_REG) {
|
1729 |
tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, stack_offset); |
1730 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1731 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1732 |
s->reserved_regs); |
1733 |
/* XXX: not correct if reading values from the stack */
|
1734 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1735 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1736 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1737 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1738 |
s->reserved_regs); |
1739 |
/* XXX: sign extend may be needed on some targets */
|
1740 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1741 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1742 |
} else {
|
1743 |
tcg_abort(); |
1744 |
} |
1745 |
} |
1746 |
#ifndef TCG_TARGET_STACK_GROWSUP
|
1747 |
stack_offset += sizeof(tcg_target_long);
|
1748 |
#endif
|
1749 |
} |
1750 |
|
1751 |
/* assign input registers */
|
1752 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1753 |
for(i = 0; i < nb_regs; i++) { |
1754 |
arg = args[nb_oargs + i]; |
1755 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1756 |
ts = &s->temps[arg]; |
1757 |
reg = tcg_target_call_iarg_regs[i]; |
1758 |
tcg_reg_free(s, reg); |
1759 |
if (ts->val_type == TEMP_VAL_REG) {
|
1760 |
if (ts->reg != reg) {
|
1761 |
tcg_out_mov(s, reg, ts->reg); |
1762 |
} |
1763 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1764 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1765 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1766 |
/* XXX: sign extend ? */
|
1767 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1768 |
} else {
|
1769 |
tcg_abort(); |
1770 |
} |
1771 |
tcg_regset_set_reg(allocated_regs, reg); |
1772 |
} |
1773 |
} |
1774 |
|
1775 |
/* assign function address */
|
1776 |
func_arg = args[nb_oargs + nb_iargs - 1];
|
1777 |
arg_ct = &def->args_ct[0];
|
1778 |
ts = &s->temps[func_arg]; |
1779 |
func_addr = ts->val; |
1780 |
const_func_arg = 0;
|
1781 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1782 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1783 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1784 |
func_arg = reg; |
1785 |
tcg_regset_set_reg(allocated_regs, reg); |
1786 |
} else if (ts->val_type == TEMP_VAL_REG) { |
1787 |
reg = ts->reg; |
1788 |
if (!tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1789 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1790 |
tcg_out_mov(s, reg, ts->reg); |
1791 |
} |
1792 |
func_arg = reg; |
1793 |
tcg_regset_set_reg(allocated_regs, reg); |
1794 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1795 |
if (tcg_target_const_match(func_addr, arg_ct)) {
|
1796 |
const_func_arg = 1;
|
1797 |
func_arg = func_addr; |
1798 |
} else {
|
1799 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1800 |
tcg_out_movi(s, ts->type, reg, func_addr); |
1801 |
func_arg = reg; |
1802 |
tcg_regset_set_reg(allocated_regs, reg); |
1803 |
} |
1804 |
} else {
|
1805 |
tcg_abort(); |
1806 |
} |
1807 |
|
1808 |
|
1809 |
/* mark dead temporaries and free the associated registers */
|
1810 |
for(i = 0; i < nb_iargs; i++) { |
1811 |
arg = args[nb_oargs + i]; |
1812 |
if (IS_DEAD_IARG(i)) {
|
1813 |
ts = &s->temps[arg]; |
1814 |
if (!ts->fixed_reg) {
|
1815 |
if (ts->val_type == TEMP_VAL_REG)
|
1816 |
s->reg_to_temp[ts->reg] = -1;
|
1817 |
ts->val_type = TEMP_VAL_DEAD; |
1818 |
} |
1819 |
} |
1820 |
} |
1821 |
|
1822 |
/* clobber call registers */
|
1823 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1824 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1825 |
tcg_reg_free(s, reg); |
1826 |
} |
1827 |
} |
1828 |
|
1829 |
/* store globals and free associated registers (we assume the call
|
1830 |
can modify any global. */
|
1831 |
if (!(flags & TCG_CALL_CONST)) {
|
1832 |
save_globals(s, allocated_regs); |
1833 |
} |
1834 |
|
1835 |
tcg_out_op(s, opc, &func_arg, &const_func_arg); |
1836 |
|
1837 |
if (allocate_args) {
|
1838 |
tcg_out_addi(s, TCG_REG_CALL_STACK, STACK_DIR(call_stack_size)); |
1839 |
} |
1840 |
|
1841 |
/* assign output registers and emit moves if needed */
|
1842 |
for(i = 0; i < nb_oargs; i++) { |
1843 |
arg = args[i]; |
1844 |
ts = &s->temps[arg]; |
1845 |
reg = tcg_target_call_oarg_regs[i]; |
1846 |
assert(s->reg_to_temp[reg] == -1);
|
1847 |
if (ts->fixed_reg) {
|
1848 |
if (ts->reg != reg) {
|
1849 |
tcg_out_mov(s, ts->reg, reg); |
1850 |
} |
1851 |
} else {
|
1852 |
if (ts->val_type == TEMP_VAL_REG)
|
1853 |
s->reg_to_temp[ts->reg] = -1;
|
1854 |
ts->val_type = TEMP_VAL_REG; |
1855 |
ts->reg = reg; |
1856 |
ts->mem_coherent = 0;
|
1857 |
s->reg_to_temp[reg] = arg; |
1858 |
} |
1859 |
} |
1860 |
|
1861 |
return nb_iargs + nb_oargs + def->nb_cargs + 1; |
1862 |
} |
1863 |
|
1864 |
#ifdef CONFIG_PROFILER
|
1865 |
|
1866 |
static int64_t tcg_table_op_count[NB_OPS];
|
1867 |
|
1868 |
static void dump_op_count(void) |
1869 |
{ |
1870 |
int i;
|
1871 |
FILE *f; |
1872 |
f = fopen("/tmp/op.log", "w"); |
1873 |
for(i = INDEX_op_end; i < NB_OPS; i++) {
|
1874 |
fprintf(f, "%s %" PRId64 "\n", tcg_op_defs[i].name, tcg_table_op_count[i]); |
1875 |
} |
1876 |
fclose(f); |
1877 |
} |
1878 |
#endif
|
1879 |
|
1880 |
|
1881 |
static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf, |
1882 |
long search_pc)
|
1883 |
{ |
1884 |
int opc, op_index;
|
1885 |
const TCGOpDef *def;
|
1886 |
unsigned int dead_iargs; |
1887 |
const TCGArg *args;
|
1888 |
|
1889 |
#ifdef DEBUG_DISAS
|
1890 |
if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP))) {
|
1891 |
qemu_log("OP:\n");
|
1892 |
tcg_dump_ops(s, logfile); |
1893 |
qemu_log("\n");
|
1894 |
} |
1895 |
#endif
|
1896 |
|
1897 |
#ifdef CONFIG_PROFILER
|
1898 |
s->la_time -= profile_getclock(); |
1899 |
#endif
|
1900 |
tcg_liveness_analysis(s); |
1901 |
#ifdef CONFIG_PROFILER
|
1902 |
s->la_time += profile_getclock(); |
1903 |
#endif
|
1904 |
|
1905 |
#ifdef DEBUG_DISAS
|
1906 |
if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT))) {
|
1907 |
qemu_log("OP after liveness analysis:\n");
|
1908 |
tcg_dump_ops(s, logfile); |
1909 |
qemu_log("\n");
|
1910 |
} |
1911 |
#endif
|
1912 |
|
1913 |
tcg_reg_alloc_start(s); |
1914 |
|
1915 |
s->code_buf = gen_code_buf; |
1916 |
s->code_ptr = gen_code_buf; |
1917 |
|
1918 |
args = gen_opparam_buf; |
1919 |
op_index = 0;
|
1920 |
|
1921 |
for(;;) {
|
1922 |
opc = gen_opc_buf[op_index]; |
1923 |
#ifdef CONFIG_PROFILER
|
1924 |
tcg_table_op_count[opc]++; |
1925 |
#endif
|
1926 |
def = &tcg_op_defs[opc]; |
1927 |
#if 0
|
1928 |
printf("%s: %d %d %d\n", def->name,
|
1929 |
def->nb_oargs, def->nb_iargs, def->nb_cargs);
|
1930 |
// dump_regs(s);
|
1931 |
#endif
|
1932 |
switch(opc) {
|
1933 |
case INDEX_op_mov_i32:
|
1934 |
#if TCG_TARGET_REG_BITS == 64 |
1935 |
case INDEX_op_mov_i64:
|
1936 |
#endif
|
1937 |
dead_iargs = s->op_dead_iargs[op_index]; |
1938 |
tcg_reg_alloc_mov(s, def, args, dead_iargs); |
1939 |
break;
|
1940 |
case INDEX_op_movi_i32:
|
1941 |
#if TCG_TARGET_REG_BITS == 64 |
1942 |
case INDEX_op_movi_i64:
|
1943 |
#endif
|
1944 |
tcg_reg_alloc_movi(s, args); |
1945 |
break;
|
1946 |
case INDEX_op_debug_insn_start:
|
1947 |
/* debug instruction */
|
1948 |
break;
|
1949 |
case INDEX_op_nop:
|
1950 |
case INDEX_op_nop1:
|
1951 |
case INDEX_op_nop2:
|
1952 |
case INDEX_op_nop3:
|
1953 |
break;
|
1954 |
case INDEX_op_nopn:
|
1955 |
args += args[0];
|
1956 |
goto next;
|
1957 |
case INDEX_op_discard:
|
1958 |
{ |
1959 |
TCGTemp *ts; |
1960 |
ts = &s->temps[args[0]];
|
1961 |
/* mark the temporary as dead */
|
1962 |
if (!ts->fixed_reg) {
|
1963 |
if (ts->val_type == TEMP_VAL_REG)
|
1964 |
s->reg_to_temp[ts->reg] = -1;
|
1965 |
ts->val_type = TEMP_VAL_DEAD; |
1966 |
} |
1967 |
} |
1968 |
break;
|
1969 |
case INDEX_op_set_label:
|
1970 |
tcg_reg_alloc_bb_end(s, s->reserved_regs); |
1971 |
tcg_out_label(s, args[0], (long)s->code_ptr); |
1972 |
break;
|
1973 |
case INDEX_op_call:
|
1974 |
dead_iargs = s->op_dead_iargs[op_index]; |
1975 |
args += tcg_reg_alloc_call(s, def, opc, args, dead_iargs); |
1976 |
goto next;
|
1977 |
case INDEX_op_end:
|
1978 |
goto the_end;
|
1979 |
default:
|
1980 |
/* Note: in order to speed up the code, it would be much
|
1981 |
faster to have specialized register allocator functions for
|
1982 |
some common argument patterns */
|
1983 |
dead_iargs = s->op_dead_iargs[op_index]; |
1984 |
tcg_reg_alloc_op(s, def, opc, args, dead_iargs); |
1985 |
break;
|
1986 |
} |
1987 |
args += def->nb_args; |
1988 |
next:
|
1989 |
if (search_pc >= 0 && search_pc < s->code_ptr - gen_code_buf) { |
1990 |
return op_index;
|
1991 |
} |
1992 |
op_index++; |
1993 |
#ifndef NDEBUG
|
1994 |
check_regs(s); |
1995 |
#endif
|
1996 |
} |
1997 |
the_end:
|
1998 |
return -1; |
1999 |
} |
2000 |
|
2001 |
int tcg_gen_code(TCGContext *s, uint8_t *gen_code_buf)
|
2002 |
{ |
2003 |
#ifdef CONFIG_PROFILER
|
2004 |
{ |
2005 |
int n;
|
2006 |
n = (gen_opc_ptr - gen_opc_buf); |
2007 |
s->op_count += n; |
2008 |
if (n > s->op_count_max)
|
2009 |
s->op_count_max = n; |
2010 |
|
2011 |
s->temp_count += s->nb_temps; |
2012 |
if (s->nb_temps > s->temp_count_max)
|
2013 |
s->temp_count_max = s->nb_temps; |
2014 |
} |
2015 |
#endif
|
2016 |
|
2017 |
tcg_gen_code_common(s, gen_code_buf, -1);
|
2018 |
|
2019 |
/* flush instruction cache */
|
2020 |
flush_icache_range((unsigned long)gen_code_buf, |
2021 |
(unsigned long)s->code_ptr); |
2022 |
return s->code_ptr - gen_code_buf;
|
2023 |
} |
2024 |
|
2025 |
/* Return the index of the micro operation such as the pc after is <
|
2026 |
offset bytes from the start of the TB. The contents of gen_code_buf must
|
2027 |
not be changed, though writing the same values is ok.
|
2028 |
Return -1 if not found. */
|
2029 |
int tcg_gen_code_search_pc(TCGContext *s, uint8_t *gen_code_buf, long offset) |
2030 |
{ |
2031 |
return tcg_gen_code_common(s, gen_code_buf, offset);
|
2032 |
} |
2033 |
|
2034 |
#ifdef CONFIG_PROFILER
|
2035 |
void tcg_dump_info(FILE *f,
|
2036 |
int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
2037 |
{ |
2038 |
TCGContext *s = &tcg_ctx; |
2039 |
int64_t tot; |
2040 |
|
2041 |
tot = s->interm_time + s->code_time; |
2042 |
cpu_fprintf(f, "JIT cycles %" PRId64 " (%0.3f s at 2.4 GHz)\n", |
2043 |
tot, tot / 2.4e9); |
2044 |
cpu_fprintf(f, "translated TBs %" PRId64 " (aborted=%" PRId64 " %0.1f%%)\n", |
2045 |
s->tb_count, |
2046 |
s->tb_count1 - s->tb_count, |
2047 |
s->tb_count1 ? (double)(s->tb_count1 - s->tb_count) / s->tb_count1 * 100.0 : 0); |
2048 |
cpu_fprintf(f, "avg ops/TB %0.1f max=%d\n",
|
2049 |
s->tb_count ? (double)s->op_count / s->tb_count : 0, s->op_count_max); |
2050 |
cpu_fprintf(f, "deleted ops/TB %0.2f\n",
|
2051 |
s->tb_count ? |
2052 |
(double)s->del_op_count / s->tb_count : 0); |
2053 |
cpu_fprintf(f, "avg temps/TB %0.2f max=%d\n",
|
2054 |
s->tb_count ? |
2055 |
(double)s->temp_count / s->tb_count : 0, |
2056 |
s->temp_count_max); |
2057 |
|
2058 |
cpu_fprintf(f, "cycles/op %0.1f\n",
|
2059 |
s->op_count ? (double)tot / s->op_count : 0); |
2060 |
cpu_fprintf(f, "cycles/in byte %0.1f\n",
|
2061 |
s->code_in_len ? (double)tot / s->code_in_len : 0); |
2062 |
cpu_fprintf(f, "cycles/out byte %0.1f\n",
|
2063 |
s->code_out_len ? (double)tot / s->code_out_len : 0); |
2064 |
if (tot == 0) |
2065 |
tot = 1;
|
2066 |
cpu_fprintf(f, " gen_interm time %0.1f%%\n",
|
2067 |
(double)s->interm_time / tot * 100.0); |
2068 |
cpu_fprintf(f, " gen_code time %0.1f%%\n",
|
2069 |
(double)s->code_time / tot * 100.0); |
2070 |
cpu_fprintf(f, "liveness/code time %0.1f%%\n",
|
2071 |
(double)s->la_time / (s->code_time ? s->code_time : 1) * 100.0); |
2072 |
cpu_fprintf(f, "cpu_restore count %" PRId64 "\n", |
2073 |
s->restore_count); |
2074 |
cpu_fprintf(f, " avg cycles %0.1f\n",
|
2075 |
s->restore_count ? (double)s->restore_time / s->restore_count : 0); |
2076 |
|
2077 |
dump_op_count(); |
2078 |
} |
2079 |
#else
|
2080 |
void tcg_dump_info(FILE *f,
|
2081 |
int (*cpu_fprintf)(FILE *f, const char *fmt, ...)) |
2082 |
{ |
2083 |
cpu_fprintf(f, "[TCG profiler not compiled]\n");
|
2084 |
} |
2085 |
#endif
|