root / tcg / tcg.c @ 0834c9ea
History | View | Annotate | Download (75.8 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 |
#define USE_TCG_OPTIMIZATIONS
|
28 |
|
29 |
#include "config.h" |
30 |
|
31 |
/* Define to jump the ELF file used to communicate with GDB. */
|
32 |
#undef DEBUG_JIT
|
33 |
|
34 |
#if !defined(CONFIG_DEBUG_TCG) && !defined(NDEBUG)
|
35 |
/* define it to suppress various consistency checks (faster) */
|
36 |
#define NDEBUG
|
37 |
#endif
|
38 |
|
39 |
#include "qemu-common.h" |
40 |
#include "cache-utils.h" |
41 |
#include "host-utils.h" |
42 |
#include "qemu-timer.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 |
|
50 |
#include "tcg-op.h" |
51 |
|
52 |
#if TCG_TARGET_REG_BITS == 64 |
53 |
# define ELF_CLASS ELFCLASS64
|
54 |
#else
|
55 |
# define ELF_CLASS ELFCLASS32
|
56 |
#endif
|
57 |
#ifdef HOST_WORDS_BIGENDIAN
|
58 |
# define ELF_DATA ELFDATA2MSB
|
59 |
#else
|
60 |
# define ELF_DATA ELFDATA2LSB
|
61 |
#endif
|
62 |
|
63 |
#include "elf.h" |
64 |
|
65 |
#if defined(CONFIG_USE_GUEST_BASE) && !defined(TCG_TARGET_HAS_GUEST_BASE)
|
66 |
#error GUEST_BASE not supported on this host.
|
67 |
#endif
|
68 |
|
69 |
/* Forward declarations for functions declared in tcg-target.c and used here. */
|
70 |
static void tcg_target_init(TCGContext *s); |
71 |
static void tcg_target_qemu_prologue(TCGContext *s); |
72 |
static void patch_reloc(uint8_t *code_ptr, int type, |
73 |
tcg_target_long value, tcg_target_long addend); |
74 |
|
75 |
static void tcg_register_jit_int(void *buf, size_t size, |
76 |
void *debug_frame, size_t debug_frame_size)
|
77 |
__attribute__((unused)); |
78 |
|
79 |
/* Forward declarations for functions declared and used in tcg-target.c. */
|
80 |
static int target_parse_constraint(TCGArgConstraint *ct, const char **pct_str); |
81 |
static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, |
82 |
tcg_target_long arg2); |
83 |
static void tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); |
84 |
static void tcg_out_movi(TCGContext *s, TCGType type, |
85 |
TCGReg ret, tcg_target_long arg); |
86 |
static void tcg_out_op(TCGContext *s, TCGOpcode opc, const TCGArg *args, |
87 |
const int *const_args); |
88 |
static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, |
89 |
tcg_target_long arg2); |
90 |
static int tcg_target_const_match(tcg_target_long val, |
91 |
const TCGArgConstraint *arg_ct);
|
92 |
static int tcg_target_get_call_iarg_regs_count(int flags); |
93 |
|
94 |
TCGOpDef tcg_op_defs[] = { |
95 |
#define DEF(s, oargs, iargs, cargs, flags) { #s, oargs, iargs, cargs, iargs + oargs + cargs, flags }, |
96 |
#include "tcg-opc.h" |
97 |
#undef DEF
|
98 |
}; |
99 |
const size_t tcg_op_defs_max = ARRAY_SIZE(tcg_op_defs);
|
100 |
|
101 |
static TCGRegSet tcg_target_available_regs[2]; |
102 |
static TCGRegSet tcg_target_call_clobber_regs;
|
103 |
|
104 |
/* XXX: move that inside the context */
|
105 |
uint16_t *gen_opc_ptr; |
106 |
TCGArg *gen_opparam_ptr; |
107 |
|
108 |
static inline void tcg_out8(TCGContext *s, uint8_t v) |
109 |
{ |
110 |
*s->code_ptr++ = v; |
111 |
} |
112 |
|
113 |
static inline void tcg_out16(TCGContext *s, uint16_t v) |
114 |
{ |
115 |
*(uint16_t *)s->code_ptr = v; |
116 |
s->code_ptr += 2;
|
117 |
} |
118 |
|
119 |
static inline void tcg_out32(TCGContext *s, uint32_t v) |
120 |
{ |
121 |
*(uint32_t *)s->code_ptr = v; |
122 |
s->code_ptr += 4;
|
123 |
} |
124 |
|
125 |
/* label relocation processing */
|
126 |
|
127 |
static void tcg_out_reloc(TCGContext *s, uint8_t *code_ptr, int type, |
128 |
int label_index, long addend) |
129 |
{ |
130 |
TCGLabel *l; |
131 |
TCGRelocation *r; |
132 |
|
133 |
l = &s->labels[label_index]; |
134 |
if (l->has_value) {
|
135 |
/* FIXME: This may break relocations on RISC targets that
|
136 |
modify instruction fields in place. The caller may not have
|
137 |
written the initial value. */
|
138 |
patch_reloc(code_ptr, type, l->u.value, addend); |
139 |
} else {
|
140 |
/* add a new relocation entry */
|
141 |
r = tcg_malloc(sizeof(TCGRelocation));
|
142 |
r->type = type; |
143 |
r->ptr = code_ptr; |
144 |
r->addend = addend; |
145 |
r->next = l->u.first_reloc; |
146 |
l->u.first_reloc = r; |
147 |
} |
148 |
} |
149 |
|
150 |
static void tcg_out_label(TCGContext *s, int label_index, void *ptr) |
151 |
{ |
152 |
TCGLabel *l; |
153 |
TCGRelocation *r; |
154 |
tcg_target_long value = (tcg_target_long)ptr; |
155 |
|
156 |
l = &s->labels[label_index]; |
157 |
if (l->has_value)
|
158 |
tcg_abort(); |
159 |
r = l->u.first_reloc; |
160 |
while (r != NULL) { |
161 |
patch_reloc(r->ptr, r->type, value, r->addend); |
162 |
r = r->next; |
163 |
} |
164 |
l->has_value = 1;
|
165 |
l->u.value = value; |
166 |
} |
167 |
|
168 |
int gen_new_label(void) |
169 |
{ |
170 |
TCGContext *s = &tcg_ctx; |
171 |
int idx;
|
172 |
TCGLabel *l; |
173 |
|
174 |
if (s->nb_labels >= TCG_MAX_LABELS)
|
175 |
tcg_abort(); |
176 |
idx = s->nb_labels++; |
177 |
l = &s->labels[idx]; |
178 |
l->has_value = 0;
|
179 |
l->u.first_reloc = NULL;
|
180 |
return idx;
|
181 |
} |
182 |
|
183 |
#include "tcg-target.c" |
184 |
|
185 |
/* pool based memory allocation */
|
186 |
void *tcg_malloc_internal(TCGContext *s, int size) |
187 |
{ |
188 |
TCGPool *p; |
189 |
int pool_size;
|
190 |
|
191 |
if (size > TCG_POOL_CHUNK_SIZE) {
|
192 |
/* big malloc: insert a new pool (XXX: could optimize) */
|
193 |
p = g_malloc(sizeof(TCGPool) + size);
|
194 |
p->size = size; |
195 |
p->next = s->pool_first_large; |
196 |
s->pool_first_large = p; |
197 |
return p->data;
|
198 |
} else {
|
199 |
p = s->pool_current; |
200 |
if (!p) {
|
201 |
p = s->pool_first; |
202 |
if (!p)
|
203 |
goto new_pool;
|
204 |
} else {
|
205 |
if (!p->next) {
|
206 |
new_pool:
|
207 |
pool_size = TCG_POOL_CHUNK_SIZE; |
208 |
p = g_malloc(sizeof(TCGPool) + pool_size);
|
209 |
p->size = pool_size; |
210 |
p->next = NULL;
|
211 |
if (s->pool_current)
|
212 |
s->pool_current->next = p; |
213 |
else
|
214 |
s->pool_first = p; |
215 |
} else {
|
216 |
p = p->next; |
217 |
} |
218 |
} |
219 |
} |
220 |
s->pool_current = p; |
221 |
s->pool_cur = p->data + size; |
222 |
s->pool_end = p->data + p->size; |
223 |
return p->data;
|
224 |
} |
225 |
|
226 |
void tcg_pool_reset(TCGContext *s)
|
227 |
{ |
228 |
TCGPool *p, *t; |
229 |
for (p = s->pool_first_large; p; p = t) {
|
230 |
t = p->next; |
231 |
g_free(p); |
232 |
} |
233 |
s->pool_first_large = NULL;
|
234 |
s->pool_cur = s->pool_end = NULL;
|
235 |
s->pool_current = NULL;
|
236 |
} |
237 |
|
238 |
void tcg_context_init(TCGContext *s)
|
239 |
{ |
240 |
int op, total_args, n;
|
241 |
TCGOpDef *def; |
242 |
TCGArgConstraint *args_ct; |
243 |
int *sorted_args;
|
244 |
|
245 |
memset(s, 0, sizeof(*s)); |
246 |
s->temps = s->static_temps; |
247 |
s->nb_globals = 0;
|
248 |
|
249 |
/* Count total number of arguments and allocate the corresponding
|
250 |
space */
|
251 |
total_args = 0;
|
252 |
for(op = 0; op < NB_OPS; op++) { |
253 |
def = &tcg_op_defs[op]; |
254 |
n = def->nb_iargs + def->nb_oargs; |
255 |
total_args += n; |
256 |
} |
257 |
|
258 |
args_ct = g_malloc(sizeof(TCGArgConstraint) * total_args);
|
259 |
sorted_args = g_malloc(sizeof(int) * total_args); |
260 |
|
261 |
for(op = 0; op < NB_OPS; op++) { |
262 |
def = &tcg_op_defs[op]; |
263 |
def->args_ct = args_ct; |
264 |
def->sorted_args = sorted_args; |
265 |
n = def->nb_iargs + def->nb_oargs; |
266 |
sorted_args += n; |
267 |
args_ct += n; |
268 |
} |
269 |
|
270 |
tcg_target_init(s); |
271 |
} |
272 |
|
273 |
void tcg_prologue_init(TCGContext *s)
|
274 |
{ |
275 |
/* init global prologue and epilogue */
|
276 |
s->code_buf = code_gen_prologue; |
277 |
s->code_ptr = s->code_buf; |
278 |
tcg_target_qemu_prologue(s); |
279 |
flush_icache_range((tcg_target_ulong)s->code_buf, |
280 |
(tcg_target_ulong)s->code_ptr); |
281 |
} |
282 |
|
283 |
void tcg_set_frame(TCGContext *s, int reg, |
284 |
tcg_target_long start, tcg_target_long size) |
285 |
{ |
286 |
s->frame_start = start; |
287 |
s->frame_end = start + size; |
288 |
s->frame_reg = reg; |
289 |
} |
290 |
|
291 |
void tcg_func_start(TCGContext *s)
|
292 |
{ |
293 |
int i;
|
294 |
tcg_pool_reset(s); |
295 |
s->nb_temps = s->nb_globals; |
296 |
for(i = 0; i < (TCG_TYPE_COUNT * 2); i++) |
297 |
s->first_free_temp[i] = -1;
|
298 |
s->labels = tcg_malloc(sizeof(TCGLabel) * TCG_MAX_LABELS);
|
299 |
s->nb_labels = 0;
|
300 |
s->current_frame_offset = s->frame_start; |
301 |
|
302 |
gen_opc_ptr = gen_opc_buf; |
303 |
gen_opparam_ptr = gen_opparam_buf; |
304 |
} |
305 |
|
306 |
static inline void tcg_temp_alloc(TCGContext *s, int n) |
307 |
{ |
308 |
if (n > TCG_MAX_TEMPS)
|
309 |
tcg_abort(); |
310 |
} |
311 |
|
312 |
static inline int tcg_global_reg_new_internal(TCGType type, int reg, |
313 |
const char *name) |
314 |
{ |
315 |
TCGContext *s = &tcg_ctx; |
316 |
TCGTemp *ts; |
317 |
int idx;
|
318 |
|
319 |
#if TCG_TARGET_REG_BITS == 32 |
320 |
if (type != TCG_TYPE_I32)
|
321 |
tcg_abort(); |
322 |
#endif
|
323 |
if (tcg_regset_test_reg(s->reserved_regs, reg))
|
324 |
tcg_abort(); |
325 |
idx = s->nb_globals; |
326 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
327 |
ts = &s->temps[s->nb_globals]; |
328 |
ts->base_type = type; |
329 |
ts->type = type; |
330 |
ts->fixed_reg = 1;
|
331 |
ts->reg = reg; |
332 |
ts->name = name; |
333 |
s->nb_globals++; |
334 |
tcg_regset_set_reg(s->reserved_regs, reg); |
335 |
return idx;
|
336 |
} |
337 |
|
338 |
TCGv_i32 tcg_global_reg_new_i32(int reg, const char *name) |
339 |
{ |
340 |
int idx;
|
341 |
|
342 |
idx = tcg_global_reg_new_internal(TCG_TYPE_I32, reg, name); |
343 |
return MAKE_TCGV_I32(idx);
|
344 |
} |
345 |
|
346 |
TCGv_i64 tcg_global_reg_new_i64(int reg, const char *name) |
347 |
{ |
348 |
int idx;
|
349 |
|
350 |
idx = tcg_global_reg_new_internal(TCG_TYPE_I64, reg, name); |
351 |
return MAKE_TCGV_I64(idx);
|
352 |
} |
353 |
|
354 |
static inline int tcg_global_mem_new_internal(TCGType type, int reg, |
355 |
tcg_target_long offset, |
356 |
const char *name) |
357 |
{ |
358 |
TCGContext *s = &tcg_ctx; |
359 |
TCGTemp *ts; |
360 |
int idx;
|
361 |
|
362 |
idx = s->nb_globals; |
363 |
#if TCG_TARGET_REG_BITS == 32 |
364 |
if (type == TCG_TYPE_I64) {
|
365 |
char buf[64]; |
366 |
tcg_temp_alloc(s, s->nb_globals + 2);
|
367 |
ts = &s->temps[s->nb_globals]; |
368 |
ts->base_type = type; |
369 |
ts->type = TCG_TYPE_I32; |
370 |
ts->fixed_reg = 0;
|
371 |
ts->mem_allocated = 1;
|
372 |
ts->mem_reg = reg; |
373 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
374 |
ts->mem_offset = offset + 4;
|
375 |
#else
|
376 |
ts->mem_offset = offset; |
377 |
#endif
|
378 |
pstrcpy(buf, sizeof(buf), name);
|
379 |
pstrcat(buf, sizeof(buf), "_0"); |
380 |
ts->name = strdup(buf); |
381 |
ts++; |
382 |
|
383 |
ts->base_type = type; |
384 |
ts->type = TCG_TYPE_I32; |
385 |
ts->fixed_reg = 0;
|
386 |
ts->mem_allocated = 1;
|
387 |
ts->mem_reg = reg; |
388 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
389 |
ts->mem_offset = offset; |
390 |
#else
|
391 |
ts->mem_offset = offset + 4;
|
392 |
#endif
|
393 |
pstrcpy(buf, sizeof(buf), name);
|
394 |
pstrcat(buf, sizeof(buf), "_1"); |
395 |
ts->name = strdup(buf); |
396 |
|
397 |
s->nb_globals += 2;
|
398 |
} else
|
399 |
#endif
|
400 |
{ |
401 |
tcg_temp_alloc(s, s->nb_globals + 1);
|
402 |
ts = &s->temps[s->nb_globals]; |
403 |
ts->base_type = type; |
404 |
ts->type = type; |
405 |
ts->fixed_reg = 0;
|
406 |
ts->mem_allocated = 1;
|
407 |
ts->mem_reg = reg; |
408 |
ts->mem_offset = offset; |
409 |
ts->name = name; |
410 |
s->nb_globals++; |
411 |
} |
412 |
return idx;
|
413 |
} |
414 |
|
415 |
TCGv_i32 tcg_global_mem_new_i32(int reg, tcg_target_long offset,
|
416 |
const char *name) |
417 |
{ |
418 |
int idx;
|
419 |
|
420 |
idx = tcg_global_mem_new_internal(TCG_TYPE_I32, reg, offset, name); |
421 |
return MAKE_TCGV_I32(idx);
|
422 |
} |
423 |
|
424 |
TCGv_i64 tcg_global_mem_new_i64(int reg, tcg_target_long offset,
|
425 |
const char *name) |
426 |
{ |
427 |
int idx;
|
428 |
|
429 |
idx = tcg_global_mem_new_internal(TCG_TYPE_I64, reg, offset, name); |
430 |
return MAKE_TCGV_I64(idx);
|
431 |
} |
432 |
|
433 |
static inline int tcg_temp_new_internal(TCGType type, int temp_local) |
434 |
{ |
435 |
TCGContext *s = &tcg_ctx; |
436 |
TCGTemp *ts; |
437 |
int idx, k;
|
438 |
|
439 |
k = type; |
440 |
if (temp_local)
|
441 |
k += TCG_TYPE_COUNT; |
442 |
idx = s->first_free_temp[k]; |
443 |
if (idx != -1) { |
444 |
/* There is already an available temp with the
|
445 |
right type */
|
446 |
ts = &s->temps[idx]; |
447 |
s->first_free_temp[k] = ts->next_free_temp; |
448 |
ts->temp_allocated = 1;
|
449 |
assert(ts->temp_local == temp_local); |
450 |
} else {
|
451 |
idx = s->nb_temps; |
452 |
#if TCG_TARGET_REG_BITS == 32 |
453 |
if (type == TCG_TYPE_I64) {
|
454 |
tcg_temp_alloc(s, s->nb_temps + 2);
|
455 |
ts = &s->temps[s->nb_temps]; |
456 |
ts->base_type = type; |
457 |
ts->type = TCG_TYPE_I32; |
458 |
ts->temp_allocated = 1;
|
459 |
ts->temp_local = temp_local; |
460 |
ts->name = NULL;
|
461 |
ts++; |
462 |
ts->base_type = TCG_TYPE_I32; |
463 |
ts->type = TCG_TYPE_I32; |
464 |
ts->temp_allocated = 1;
|
465 |
ts->temp_local = temp_local; |
466 |
ts->name = NULL;
|
467 |
s->nb_temps += 2;
|
468 |
} else
|
469 |
#endif
|
470 |
{ |
471 |
tcg_temp_alloc(s, s->nb_temps + 1);
|
472 |
ts = &s->temps[s->nb_temps]; |
473 |
ts->base_type = type; |
474 |
ts->type = type; |
475 |
ts->temp_allocated = 1;
|
476 |
ts->temp_local = temp_local; |
477 |
ts->name = NULL;
|
478 |
s->nb_temps++; |
479 |
} |
480 |
} |
481 |
|
482 |
#if defined(CONFIG_DEBUG_TCG)
|
483 |
s->temps_in_use++; |
484 |
#endif
|
485 |
return idx;
|
486 |
} |
487 |
|
488 |
TCGv_i32 tcg_temp_new_internal_i32(int temp_local)
|
489 |
{ |
490 |
int idx;
|
491 |
|
492 |
idx = tcg_temp_new_internal(TCG_TYPE_I32, temp_local); |
493 |
return MAKE_TCGV_I32(idx);
|
494 |
} |
495 |
|
496 |
TCGv_i64 tcg_temp_new_internal_i64(int temp_local)
|
497 |
{ |
498 |
int idx;
|
499 |
|
500 |
idx = tcg_temp_new_internal(TCG_TYPE_I64, temp_local); |
501 |
return MAKE_TCGV_I64(idx);
|
502 |
} |
503 |
|
504 |
static inline void tcg_temp_free_internal(int idx) |
505 |
{ |
506 |
TCGContext *s = &tcg_ctx; |
507 |
TCGTemp *ts; |
508 |
int k;
|
509 |
|
510 |
#if defined(CONFIG_DEBUG_TCG)
|
511 |
s->temps_in_use--; |
512 |
if (s->temps_in_use < 0) { |
513 |
fprintf(stderr, "More temporaries freed than allocated!\n");
|
514 |
} |
515 |
#endif
|
516 |
|
517 |
assert(idx >= s->nb_globals && idx < s->nb_temps); |
518 |
ts = &s->temps[idx]; |
519 |
assert(ts->temp_allocated != 0);
|
520 |
ts->temp_allocated = 0;
|
521 |
k = ts->base_type; |
522 |
if (ts->temp_local)
|
523 |
k += TCG_TYPE_COUNT; |
524 |
ts->next_free_temp = s->first_free_temp[k]; |
525 |
s->first_free_temp[k] = idx; |
526 |
} |
527 |
|
528 |
void tcg_temp_free_i32(TCGv_i32 arg)
|
529 |
{ |
530 |
tcg_temp_free_internal(GET_TCGV_I32(arg)); |
531 |
} |
532 |
|
533 |
void tcg_temp_free_i64(TCGv_i64 arg)
|
534 |
{ |
535 |
tcg_temp_free_internal(GET_TCGV_I64(arg)); |
536 |
} |
537 |
|
538 |
TCGv_i32 tcg_const_i32(int32_t val) |
539 |
{ |
540 |
TCGv_i32 t0; |
541 |
t0 = tcg_temp_new_i32(); |
542 |
tcg_gen_movi_i32(t0, val); |
543 |
return t0;
|
544 |
} |
545 |
|
546 |
TCGv_i64 tcg_const_i64(int64_t val) |
547 |
{ |
548 |
TCGv_i64 t0; |
549 |
t0 = tcg_temp_new_i64(); |
550 |
tcg_gen_movi_i64(t0, val); |
551 |
return t0;
|
552 |
} |
553 |
|
554 |
TCGv_i32 tcg_const_local_i32(int32_t val) |
555 |
{ |
556 |
TCGv_i32 t0; |
557 |
t0 = tcg_temp_local_new_i32(); |
558 |
tcg_gen_movi_i32(t0, val); |
559 |
return t0;
|
560 |
} |
561 |
|
562 |
TCGv_i64 tcg_const_local_i64(int64_t val) |
563 |
{ |
564 |
TCGv_i64 t0; |
565 |
t0 = tcg_temp_local_new_i64(); |
566 |
tcg_gen_movi_i64(t0, val); |
567 |
return t0;
|
568 |
} |
569 |
|
570 |
#if defined(CONFIG_DEBUG_TCG)
|
571 |
void tcg_clear_temp_count(void) |
572 |
{ |
573 |
TCGContext *s = &tcg_ctx; |
574 |
s->temps_in_use = 0;
|
575 |
} |
576 |
|
577 |
int tcg_check_temp_count(void) |
578 |
{ |
579 |
TCGContext *s = &tcg_ctx; |
580 |
if (s->temps_in_use) {
|
581 |
/* Clear the count so that we don't give another
|
582 |
* warning immediately next time around.
|
583 |
*/
|
584 |
s->temps_in_use = 0;
|
585 |
return 1; |
586 |
} |
587 |
return 0; |
588 |
} |
589 |
#endif
|
590 |
|
591 |
void tcg_register_helper(void *func, const char *name) |
592 |
{ |
593 |
TCGContext *s = &tcg_ctx; |
594 |
int n;
|
595 |
if ((s->nb_helpers + 1) > s->allocated_helpers) { |
596 |
n = s->allocated_helpers; |
597 |
if (n == 0) { |
598 |
n = 4;
|
599 |
} else {
|
600 |
n *= 2;
|
601 |
} |
602 |
s->helpers = realloc(s->helpers, n * sizeof(TCGHelperInfo));
|
603 |
s->allocated_helpers = n; |
604 |
} |
605 |
s->helpers[s->nb_helpers].func = (tcg_target_ulong)func; |
606 |
s->helpers[s->nb_helpers].name = name; |
607 |
s->nb_helpers++; |
608 |
} |
609 |
|
610 |
/* Note: we convert the 64 bit args to 32 bit and do some alignment
|
611 |
and endian swap. Maybe it would be better to do the alignment
|
612 |
and endian swap in tcg_reg_alloc_call(). */
|
613 |
void tcg_gen_callN(TCGContext *s, TCGv_ptr func, unsigned int flags, |
614 |
int sizemask, TCGArg ret, int nargs, TCGArg *args) |
615 |
{ |
616 |
int i;
|
617 |
int real_args;
|
618 |
int nb_rets;
|
619 |
TCGArg *nparam; |
620 |
|
621 |
#if defined(TCG_TARGET_EXTEND_ARGS) && TCG_TARGET_REG_BITS == 64 |
622 |
for (i = 0; i < nargs; ++i) { |
623 |
int is_64bit = sizemask & (1 << (i+1)*2); |
624 |
int is_signed = sizemask & (2 << (i+1)*2); |
625 |
if (!is_64bit) {
|
626 |
TCGv_i64 temp = tcg_temp_new_i64(); |
627 |
TCGv_i64 orig = MAKE_TCGV_I64(args[i]); |
628 |
if (is_signed) {
|
629 |
tcg_gen_ext32s_i64(temp, orig); |
630 |
} else {
|
631 |
tcg_gen_ext32u_i64(temp, orig); |
632 |
} |
633 |
args[i] = GET_TCGV_I64(temp); |
634 |
} |
635 |
} |
636 |
#endif /* TCG_TARGET_EXTEND_ARGS */ |
637 |
|
638 |
*gen_opc_ptr++ = INDEX_op_call; |
639 |
nparam = gen_opparam_ptr++; |
640 |
if (ret != TCG_CALL_DUMMY_ARG) {
|
641 |
#if TCG_TARGET_REG_BITS < 64 |
642 |
if (sizemask & 1) { |
643 |
#ifdef TCG_TARGET_WORDS_BIGENDIAN
|
644 |
*gen_opparam_ptr++ = ret + 1;
|
645 |
*gen_opparam_ptr++ = ret; |
646 |
#else
|
647 |
*gen_opparam_ptr++ = ret; |
648 |
*gen_opparam_ptr++ = ret + 1;
|
649 |
#endif
|
650 |
nb_rets = 2;
|
651 |
} else
|
652 |
#endif
|
653 |
{ |
654 |
*gen_opparam_ptr++ = ret; |
655 |
nb_rets = 1;
|
656 |
} |
657 |
} else {
|
658 |
nb_rets = 0;
|
659 |
} |
660 |
real_args = 0;
|
661 |
for (i = 0; i < nargs; i++) { |
662 |
#if TCG_TARGET_REG_BITS < 64 |
663 |
int is_64bit = sizemask & (1 << (i+1)*2); |
664 |
if (is_64bit) {
|
665 |
#ifdef TCG_TARGET_CALL_ALIGN_ARGS
|
666 |
/* some targets want aligned 64 bit args */
|
667 |
if (real_args & 1) { |
668 |
*gen_opparam_ptr++ = TCG_CALL_DUMMY_ARG; |
669 |
real_args++; |
670 |
} |
671 |
#endif
|
672 |
/* If stack grows up, then we will be placing successive
|
673 |
arguments at lower addresses, which means we need to
|
674 |
reverse the order compared to how we would normally
|
675 |
treat either big or little-endian. For those arguments
|
676 |
that will wind up in registers, this still works for
|
677 |
HPPA (the only current STACK_GROWSUP target) since the
|
678 |
argument registers are *also* allocated in decreasing
|
679 |
order. If another such target is added, this logic may
|
680 |
have to get more complicated to differentiate between
|
681 |
stack arguments and register arguments. */
|
682 |
#if defined(TCG_TARGET_WORDS_BIGENDIAN) != defined(TCG_TARGET_STACK_GROWSUP)
|
683 |
*gen_opparam_ptr++ = args[i] + 1;
|
684 |
*gen_opparam_ptr++ = args[i]; |
685 |
#else
|
686 |
*gen_opparam_ptr++ = args[i]; |
687 |
*gen_opparam_ptr++ = args[i] + 1;
|
688 |
#endif
|
689 |
real_args += 2;
|
690 |
continue;
|
691 |
} |
692 |
#endif /* TCG_TARGET_REG_BITS < 64 */ |
693 |
|
694 |
*gen_opparam_ptr++ = args[i]; |
695 |
real_args++; |
696 |
} |
697 |
*gen_opparam_ptr++ = GET_TCGV_PTR(func); |
698 |
|
699 |
*gen_opparam_ptr++ = flags; |
700 |
|
701 |
*nparam = (nb_rets << 16) | (real_args + 1); |
702 |
|
703 |
/* total parameters, needed to go backward in the instruction stream */
|
704 |
*gen_opparam_ptr++ = 1 + nb_rets + real_args + 3; |
705 |
|
706 |
#if defined(TCG_TARGET_EXTEND_ARGS) && TCG_TARGET_REG_BITS == 64 |
707 |
for (i = 0; i < nargs; ++i) { |
708 |
int is_64bit = sizemask & (1 << (i+1)*2); |
709 |
if (!is_64bit) {
|
710 |
TCGv_i64 temp = MAKE_TCGV_I64(args[i]); |
711 |
tcg_temp_free_i64(temp); |
712 |
} |
713 |
} |
714 |
#endif /* TCG_TARGET_EXTEND_ARGS */ |
715 |
} |
716 |
|
717 |
#if TCG_TARGET_REG_BITS == 32 |
718 |
void tcg_gen_shifti_i64(TCGv_i64 ret, TCGv_i64 arg1,
|
719 |
int c, int right, int arith) |
720 |
{ |
721 |
if (c == 0) { |
722 |
tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1)); |
723 |
tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1)); |
724 |
} else if (c >= 32) { |
725 |
c -= 32;
|
726 |
if (right) {
|
727 |
if (arith) {
|
728 |
tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
729 |
tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31);
|
730 |
} else {
|
731 |
tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c); |
732 |
tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
|
733 |
} |
734 |
} else {
|
735 |
tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_LOW(arg1), c); |
736 |
tcg_gen_movi_i32(TCGV_LOW(ret), 0);
|
737 |
} |
738 |
} else {
|
739 |
TCGv_i32 t0, t1; |
740 |
|
741 |
t0 = tcg_temp_new_i32(); |
742 |
t1 = tcg_temp_new_i32(); |
743 |
if (right) {
|
744 |
tcg_gen_shli_i32(t0, TCGV_HIGH(arg1), 32 - c);
|
745 |
if (arith)
|
746 |
tcg_gen_sari_i32(t1, TCGV_HIGH(arg1), c); |
747 |
else
|
748 |
tcg_gen_shri_i32(t1, TCGV_HIGH(arg1), c); |
749 |
tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c); |
750 |
tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(ret), t0); |
751 |
tcg_gen_mov_i32(TCGV_HIGH(ret), t1); |
752 |
} else {
|
753 |
tcg_gen_shri_i32(t0, TCGV_LOW(arg1), 32 - c);
|
754 |
/* Note: ret can be the same as arg1, so we use t1 */
|
755 |
tcg_gen_shli_i32(t1, TCGV_LOW(arg1), c); |
756 |
tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c); |
757 |
tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t0); |
758 |
tcg_gen_mov_i32(TCGV_LOW(ret), t1); |
759 |
} |
760 |
tcg_temp_free_i32(t0); |
761 |
tcg_temp_free_i32(t1); |
762 |
} |
763 |
} |
764 |
#endif
|
765 |
|
766 |
|
767 |
static void tcg_reg_alloc_start(TCGContext *s) |
768 |
{ |
769 |
int i;
|
770 |
TCGTemp *ts; |
771 |
for(i = 0; i < s->nb_globals; i++) { |
772 |
ts = &s->temps[i]; |
773 |
if (ts->fixed_reg) {
|
774 |
ts->val_type = TEMP_VAL_REG; |
775 |
} else {
|
776 |
ts->val_type = TEMP_VAL_MEM; |
777 |
} |
778 |
} |
779 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
780 |
ts = &s->temps[i]; |
781 |
ts->val_type = TEMP_VAL_DEAD; |
782 |
ts->mem_allocated = 0;
|
783 |
ts->fixed_reg = 0;
|
784 |
} |
785 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
786 |
s->reg_to_temp[i] = -1;
|
787 |
} |
788 |
} |
789 |
|
790 |
static char *tcg_get_arg_str_idx(TCGContext *s, char *buf, int buf_size, |
791 |
int idx)
|
792 |
{ |
793 |
TCGTemp *ts; |
794 |
|
795 |
assert(idx >= 0 && idx < s->nb_temps);
|
796 |
ts = &s->temps[idx]; |
797 |
assert(ts); |
798 |
if (idx < s->nb_globals) {
|
799 |
pstrcpy(buf, buf_size, ts->name); |
800 |
} else {
|
801 |
if (ts->temp_local)
|
802 |
snprintf(buf, buf_size, "loc%d", idx - s->nb_globals);
|
803 |
else
|
804 |
snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals);
|
805 |
} |
806 |
return buf;
|
807 |
} |
808 |
|
809 |
char *tcg_get_arg_str_i32(TCGContext *s, char *buf, int buf_size, TCGv_i32 arg) |
810 |
{ |
811 |
return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I32(arg));
|
812 |
} |
813 |
|
814 |
char *tcg_get_arg_str_i64(TCGContext *s, char *buf, int buf_size, TCGv_i64 arg) |
815 |
{ |
816 |
return tcg_get_arg_str_idx(s, buf, buf_size, GET_TCGV_I64(arg));
|
817 |
} |
818 |
|
819 |
static int helper_cmp(const void *p1, const void *p2) |
820 |
{ |
821 |
const TCGHelperInfo *th1 = p1;
|
822 |
const TCGHelperInfo *th2 = p2;
|
823 |
if (th1->func < th2->func)
|
824 |
return -1; |
825 |
else if (th1->func == th2->func) |
826 |
return 0; |
827 |
else
|
828 |
return 1; |
829 |
} |
830 |
|
831 |
/* find helper definition (Note: A hash table would be better) */
|
832 |
static TCGHelperInfo *tcg_find_helper(TCGContext *s, tcg_target_ulong val)
|
833 |
{ |
834 |
int m, m_min, m_max;
|
835 |
TCGHelperInfo *th; |
836 |
tcg_target_ulong v; |
837 |
|
838 |
if (unlikely(!s->helpers_sorted)) {
|
839 |
qsort(s->helpers, s->nb_helpers, sizeof(TCGHelperInfo),
|
840 |
helper_cmp); |
841 |
s->helpers_sorted = 1;
|
842 |
} |
843 |
|
844 |
/* binary search */
|
845 |
m_min = 0;
|
846 |
m_max = s->nb_helpers - 1;
|
847 |
while (m_min <= m_max) {
|
848 |
m = (m_min + m_max) >> 1;
|
849 |
th = &s->helpers[m]; |
850 |
v = th->func; |
851 |
if (v == val)
|
852 |
return th;
|
853 |
else if (val < v) { |
854 |
m_max = m - 1;
|
855 |
} else {
|
856 |
m_min = m + 1;
|
857 |
} |
858 |
} |
859 |
return NULL; |
860 |
} |
861 |
|
862 |
static const char * const cond_name[] = |
863 |
{ |
864 |
[TCG_COND_EQ] = "eq",
|
865 |
[TCG_COND_NE] = "ne",
|
866 |
[TCG_COND_LT] = "lt",
|
867 |
[TCG_COND_GE] = "ge",
|
868 |
[TCG_COND_LE] = "le",
|
869 |
[TCG_COND_GT] = "gt",
|
870 |
[TCG_COND_LTU] = "ltu",
|
871 |
[TCG_COND_GEU] = "geu",
|
872 |
[TCG_COND_LEU] = "leu",
|
873 |
[TCG_COND_GTU] = "gtu"
|
874 |
}; |
875 |
|
876 |
void tcg_dump_ops(TCGContext *s)
|
877 |
{ |
878 |
const uint16_t *opc_ptr;
|
879 |
const TCGArg *args;
|
880 |
TCGArg arg; |
881 |
TCGOpcode c; |
882 |
int i, k, nb_oargs, nb_iargs, nb_cargs, first_insn;
|
883 |
const TCGOpDef *def;
|
884 |
char buf[128]; |
885 |
|
886 |
first_insn = 1;
|
887 |
opc_ptr = gen_opc_buf; |
888 |
args = gen_opparam_buf; |
889 |
while (opc_ptr < gen_opc_ptr) {
|
890 |
c = *opc_ptr++; |
891 |
def = &tcg_op_defs[c]; |
892 |
if (c == INDEX_op_debug_insn_start) {
|
893 |
uint64_t pc; |
894 |
#if TARGET_LONG_BITS > TCG_TARGET_REG_BITS
|
895 |
pc = ((uint64_t)args[1] << 32) | args[0]; |
896 |
#else
|
897 |
pc = args[0];
|
898 |
#endif
|
899 |
if (!first_insn) {
|
900 |
qemu_log("\n");
|
901 |
} |
902 |
qemu_log(" ---- 0x%" PRIx64, pc);
|
903 |
first_insn = 0;
|
904 |
nb_oargs = def->nb_oargs; |
905 |
nb_iargs = def->nb_iargs; |
906 |
nb_cargs = def->nb_cargs; |
907 |
} else if (c == INDEX_op_call) { |
908 |
TCGArg arg; |
909 |
|
910 |
/* variable number of arguments */
|
911 |
arg = *args++; |
912 |
nb_oargs = arg >> 16;
|
913 |
nb_iargs = arg & 0xffff;
|
914 |
nb_cargs = def->nb_cargs; |
915 |
|
916 |
qemu_log(" %s ", def->name);
|
917 |
|
918 |
/* function name */
|
919 |
qemu_log("%s",
|
920 |
tcg_get_arg_str_idx(s, buf, sizeof(buf),
|
921 |
args[nb_oargs + nb_iargs - 1]));
|
922 |
/* flags */
|
923 |
qemu_log(",$0x%" TCG_PRIlx, args[nb_oargs + nb_iargs]);
|
924 |
/* nb out args */
|
925 |
qemu_log(",$%d", nb_oargs);
|
926 |
for(i = 0; i < nb_oargs; i++) { |
927 |
qemu_log(",");
|
928 |
qemu_log("%s", tcg_get_arg_str_idx(s, buf, sizeof(buf), |
929 |
args[i])); |
930 |
} |
931 |
for(i = 0; i < (nb_iargs - 1); i++) { |
932 |
qemu_log(",");
|
933 |
if (args[nb_oargs + i] == TCG_CALL_DUMMY_ARG) {
|
934 |
qemu_log("<dummy>");
|
935 |
} else {
|
936 |
qemu_log("%s", tcg_get_arg_str_idx(s, buf, sizeof(buf), |
937 |
args[nb_oargs + i])); |
938 |
} |
939 |
} |
940 |
} else if (c == INDEX_op_movi_i32 |
941 |
#if TCG_TARGET_REG_BITS == 64 |
942 |
|| c == INDEX_op_movi_i64 |
943 |
#endif
|
944 |
) { |
945 |
tcg_target_ulong val; |
946 |
TCGHelperInfo *th; |
947 |
|
948 |
nb_oargs = def->nb_oargs; |
949 |
nb_iargs = def->nb_iargs; |
950 |
nb_cargs = def->nb_cargs; |
951 |
qemu_log(" %s %s,$", def->name,
|
952 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), args[0])); |
953 |
val = args[1];
|
954 |
th = tcg_find_helper(s, val); |
955 |
if (th) {
|
956 |
qemu_log("%s", th->name);
|
957 |
} else {
|
958 |
if (c == INDEX_op_movi_i32) {
|
959 |
qemu_log("0x%x", (uint32_t)val);
|
960 |
} else {
|
961 |
qemu_log("0x%" PRIx64 , (uint64_t)val);
|
962 |
} |
963 |
} |
964 |
} else {
|
965 |
qemu_log(" %s ", def->name);
|
966 |
if (c == INDEX_op_nopn) {
|
967 |
/* variable number of arguments */
|
968 |
nb_cargs = *args; |
969 |
nb_oargs = 0;
|
970 |
nb_iargs = 0;
|
971 |
} else {
|
972 |
nb_oargs = def->nb_oargs; |
973 |
nb_iargs = def->nb_iargs; |
974 |
nb_cargs = def->nb_cargs; |
975 |
} |
976 |
|
977 |
k = 0;
|
978 |
for(i = 0; i < nb_oargs; i++) { |
979 |
if (k != 0) { |
980 |
qemu_log(",");
|
981 |
} |
982 |
qemu_log("%s", tcg_get_arg_str_idx(s, buf, sizeof(buf), |
983 |
args[k++])); |
984 |
} |
985 |
for(i = 0; i < nb_iargs; i++) { |
986 |
if (k != 0) { |
987 |
qemu_log(",");
|
988 |
} |
989 |
qemu_log("%s", tcg_get_arg_str_idx(s, buf, sizeof(buf), |
990 |
args[k++])); |
991 |
} |
992 |
switch (c) {
|
993 |
case INDEX_op_brcond_i32:
|
994 |
case INDEX_op_setcond_i32:
|
995 |
case INDEX_op_movcond_i32:
|
996 |
#if TCG_TARGET_REG_BITS == 32 |
997 |
case INDEX_op_brcond2_i32:
|
998 |
case INDEX_op_setcond2_i32:
|
999 |
#else
|
1000 |
case INDEX_op_brcond_i64:
|
1001 |
case INDEX_op_setcond_i64:
|
1002 |
case INDEX_op_movcond_i64:
|
1003 |
#endif
|
1004 |
if (args[k] < ARRAY_SIZE(cond_name) && cond_name[args[k]]) {
|
1005 |
qemu_log(",%s", cond_name[args[k++]]);
|
1006 |
} else {
|
1007 |
qemu_log(",$0x%" TCG_PRIlx, args[k++]);
|
1008 |
} |
1009 |
i = 1;
|
1010 |
break;
|
1011 |
default:
|
1012 |
i = 0;
|
1013 |
break;
|
1014 |
} |
1015 |
for(; i < nb_cargs; i++) {
|
1016 |
if (k != 0) { |
1017 |
qemu_log(",");
|
1018 |
} |
1019 |
arg = args[k++]; |
1020 |
qemu_log("$0x%" TCG_PRIlx, arg);
|
1021 |
} |
1022 |
} |
1023 |
qemu_log("\n");
|
1024 |
args += nb_iargs + nb_oargs + nb_cargs; |
1025 |
} |
1026 |
} |
1027 |
|
1028 |
/* we give more priority to constraints with less registers */
|
1029 |
static int get_constraint_priority(const TCGOpDef *def, int k) |
1030 |
{ |
1031 |
const TCGArgConstraint *arg_ct;
|
1032 |
|
1033 |
int i, n;
|
1034 |
arg_ct = &def->args_ct[k]; |
1035 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
1036 |
/* an alias is equivalent to a single register */
|
1037 |
n = 1;
|
1038 |
} else {
|
1039 |
if (!(arg_ct->ct & TCG_CT_REG))
|
1040 |
return 0; |
1041 |
n = 0;
|
1042 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
1043 |
if (tcg_regset_test_reg(arg_ct->u.regs, i))
|
1044 |
n++; |
1045 |
} |
1046 |
} |
1047 |
return TCG_TARGET_NB_REGS - n + 1; |
1048 |
} |
1049 |
|
1050 |
/* sort from highest priority to lowest */
|
1051 |
static void sort_constraints(TCGOpDef *def, int start, int n) |
1052 |
{ |
1053 |
int i, j, p1, p2, tmp;
|
1054 |
|
1055 |
for(i = 0; i < n; i++) |
1056 |
def->sorted_args[start + i] = start + i; |
1057 |
if (n <= 1) |
1058 |
return;
|
1059 |
for(i = 0; i < n - 1; i++) { |
1060 |
for(j = i + 1; j < n; j++) { |
1061 |
p1 = get_constraint_priority(def, def->sorted_args[start + i]); |
1062 |
p2 = get_constraint_priority(def, def->sorted_args[start + j]); |
1063 |
if (p1 < p2) {
|
1064 |
tmp = def->sorted_args[start + i]; |
1065 |
def->sorted_args[start + i] = def->sorted_args[start + j]; |
1066 |
def->sorted_args[start + j] = tmp; |
1067 |
} |
1068 |
} |
1069 |
} |
1070 |
} |
1071 |
|
1072 |
void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs) |
1073 |
{ |
1074 |
TCGOpcode op; |
1075 |
TCGOpDef *def; |
1076 |
const char *ct_str; |
1077 |
int i, nb_args;
|
1078 |
|
1079 |
for(;;) {
|
1080 |
if (tdefs->op == (TCGOpcode)-1) |
1081 |
break;
|
1082 |
op = tdefs->op; |
1083 |
assert((unsigned)op < NB_OPS);
|
1084 |
def = &tcg_op_defs[op]; |
1085 |
#if defined(CONFIG_DEBUG_TCG)
|
1086 |
/* Duplicate entry in op definitions? */
|
1087 |
assert(!def->used); |
1088 |
def->used = 1;
|
1089 |
#endif
|
1090 |
nb_args = def->nb_iargs + def->nb_oargs; |
1091 |
for(i = 0; i < nb_args; i++) { |
1092 |
ct_str = tdefs->args_ct_str[i]; |
1093 |
/* Incomplete TCGTargetOpDef entry? */
|
1094 |
assert(ct_str != NULL);
|
1095 |
tcg_regset_clear(def->args_ct[i].u.regs); |
1096 |
def->args_ct[i].ct = 0;
|
1097 |
if (ct_str[0] >= '0' && ct_str[0] <= '9') { |
1098 |
int oarg;
|
1099 |
oarg = ct_str[0] - '0'; |
1100 |
assert(oarg < def->nb_oargs); |
1101 |
assert(def->args_ct[oarg].ct & TCG_CT_REG); |
1102 |
/* TCG_CT_ALIAS is for the output arguments. The input
|
1103 |
argument is tagged with TCG_CT_IALIAS. */
|
1104 |
def->args_ct[i] = def->args_ct[oarg]; |
1105 |
def->args_ct[oarg].ct = TCG_CT_ALIAS; |
1106 |
def->args_ct[oarg].alias_index = i; |
1107 |
def->args_ct[i].ct |= TCG_CT_IALIAS; |
1108 |
def->args_ct[i].alias_index = oarg; |
1109 |
} else {
|
1110 |
for(;;) {
|
1111 |
if (*ct_str == '\0') |
1112 |
break;
|
1113 |
switch(*ct_str) {
|
1114 |
case 'i': |
1115 |
def->args_ct[i].ct |= TCG_CT_CONST; |
1116 |
ct_str++; |
1117 |
break;
|
1118 |
default:
|
1119 |
if (target_parse_constraint(&def->args_ct[i], &ct_str) < 0) { |
1120 |
fprintf(stderr, "Invalid constraint '%s' for arg %d of operation '%s'\n",
|
1121 |
ct_str, i, def->name); |
1122 |
exit(1);
|
1123 |
} |
1124 |
} |
1125 |
} |
1126 |
} |
1127 |
} |
1128 |
|
1129 |
/* TCGTargetOpDef entry with too much information? */
|
1130 |
assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL);
|
1131 |
|
1132 |
/* sort the constraints (XXX: this is just an heuristic) */
|
1133 |
sort_constraints(def, 0, def->nb_oargs);
|
1134 |
sort_constraints(def, def->nb_oargs, def->nb_iargs); |
1135 |
|
1136 |
#if 0
|
1137 |
{
|
1138 |
int i;
|
1139 |
|
1140 |
printf("%s: sorted=", def->name);
|
1141 |
for(i = 0; i < def->nb_oargs + def->nb_iargs; i++)
|
1142 |
printf(" %d", def->sorted_args[i]);
|
1143 |
printf("\n");
|
1144 |
}
|
1145 |
#endif
|
1146 |
tdefs++; |
1147 |
} |
1148 |
|
1149 |
#if defined(CONFIG_DEBUG_TCG)
|
1150 |
i = 0;
|
1151 |
for (op = 0; op < ARRAY_SIZE(tcg_op_defs); op++) { |
1152 |
const TCGOpDef *def = &tcg_op_defs[op];
|
1153 |
if (op < INDEX_op_call
|
1154 |
|| op == INDEX_op_debug_insn_start |
1155 |
|| (def->flags & TCG_OPF_NOT_PRESENT)) { |
1156 |
/* Wrong entry in op definitions? */
|
1157 |
if (def->used) {
|
1158 |
fprintf(stderr, "Invalid op definition for %s\n", def->name);
|
1159 |
i = 1;
|
1160 |
} |
1161 |
} else {
|
1162 |
/* Missing entry in op definitions? */
|
1163 |
if (!def->used) {
|
1164 |
fprintf(stderr, "Missing op definition for %s\n", def->name);
|
1165 |
i = 1;
|
1166 |
} |
1167 |
} |
1168 |
} |
1169 |
if (i == 1) { |
1170 |
tcg_abort(); |
1171 |
} |
1172 |
#endif
|
1173 |
} |
1174 |
|
1175 |
#ifdef USE_LIVENESS_ANALYSIS
|
1176 |
|
1177 |
/* set a nop for an operation using 'nb_args' */
|
1178 |
static inline void tcg_set_nop(TCGContext *s, uint16_t *opc_ptr, |
1179 |
TCGArg *args, int nb_args)
|
1180 |
{ |
1181 |
if (nb_args == 0) { |
1182 |
*opc_ptr = INDEX_op_nop; |
1183 |
} else {
|
1184 |
*opc_ptr = INDEX_op_nopn; |
1185 |
args[0] = nb_args;
|
1186 |
args[nb_args - 1] = nb_args;
|
1187 |
} |
1188 |
} |
1189 |
|
1190 |
/* liveness analysis: end of function: globals are live, temps are
|
1191 |
dead. */
|
1192 |
/* XXX: at this stage, not used as there would be little gains because
|
1193 |
most TBs end with a conditional jump. */
|
1194 |
static inline void tcg_la_func_end(TCGContext *s, uint8_t *dead_temps) |
1195 |
{ |
1196 |
memset(dead_temps, 0, s->nb_globals);
|
1197 |
memset(dead_temps + s->nb_globals, 1, s->nb_temps - s->nb_globals);
|
1198 |
} |
1199 |
|
1200 |
/* liveness analysis: end of basic block: globals are live, temps are
|
1201 |
dead, local temps are live. */
|
1202 |
static inline void tcg_la_bb_end(TCGContext *s, uint8_t *dead_temps) |
1203 |
{ |
1204 |
int i;
|
1205 |
TCGTemp *ts; |
1206 |
|
1207 |
memset(dead_temps, 0, s->nb_globals);
|
1208 |
ts = &s->temps[s->nb_globals]; |
1209 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1210 |
if (ts->temp_local)
|
1211 |
dead_temps[i] = 0;
|
1212 |
else
|
1213 |
dead_temps[i] = 1;
|
1214 |
ts++; |
1215 |
} |
1216 |
} |
1217 |
|
1218 |
/* Liveness analysis : update the opc_dead_args array to tell if a
|
1219 |
given input arguments is dead. Instructions updating dead
|
1220 |
temporaries are removed. */
|
1221 |
static void tcg_liveness_analysis(TCGContext *s) |
1222 |
{ |
1223 |
int i, op_index, nb_args, nb_iargs, nb_oargs, arg, nb_ops;
|
1224 |
TCGOpcode op; |
1225 |
TCGArg *args; |
1226 |
const TCGOpDef *def;
|
1227 |
uint8_t *dead_temps; |
1228 |
unsigned int dead_args; |
1229 |
|
1230 |
gen_opc_ptr++; /* skip end */
|
1231 |
|
1232 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1233 |
|
1234 |
s->op_dead_args = tcg_malloc(nb_ops * sizeof(uint16_t));
|
1235 |
|
1236 |
dead_temps = tcg_malloc(s->nb_temps); |
1237 |
memset(dead_temps, 1, s->nb_temps);
|
1238 |
|
1239 |
args = gen_opparam_ptr; |
1240 |
op_index = nb_ops - 1;
|
1241 |
while (op_index >= 0) { |
1242 |
op = gen_opc_buf[op_index]; |
1243 |
def = &tcg_op_defs[op]; |
1244 |
switch(op) {
|
1245 |
case INDEX_op_call:
|
1246 |
{ |
1247 |
int call_flags;
|
1248 |
|
1249 |
nb_args = args[-1];
|
1250 |
args -= nb_args; |
1251 |
nb_iargs = args[0] & 0xffff; |
1252 |
nb_oargs = args[0] >> 16; |
1253 |
args++; |
1254 |
call_flags = args[nb_oargs + nb_iargs]; |
1255 |
|
1256 |
/* pure functions can be removed if their result is not
|
1257 |
used */
|
1258 |
if (call_flags & TCG_CALL_PURE) {
|
1259 |
for(i = 0; i < nb_oargs; i++) { |
1260 |
arg = args[i]; |
1261 |
if (!dead_temps[arg])
|
1262 |
goto do_not_remove_call;
|
1263 |
} |
1264 |
tcg_set_nop(s, gen_opc_buf + op_index, |
1265 |
args - 1, nb_args);
|
1266 |
} else {
|
1267 |
do_not_remove_call:
|
1268 |
|
1269 |
/* output args are dead */
|
1270 |
dead_args = 0;
|
1271 |
for(i = 0; i < nb_oargs; i++) { |
1272 |
arg = args[i]; |
1273 |
if (dead_temps[arg]) {
|
1274 |
dead_args |= (1 << i);
|
1275 |
} |
1276 |
dead_temps[arg] = 1;
|
1277 |
} |
1278 |
|
1279 |
if (!(call_flags & TCG_CALL_CONST)) {
|
1280 |
/* globals are live (they may be used by the call) */
|
1281 |
memset(dead_temps, 0, s->nb_globals);
|
1282 |
} |
1283 |
|
1284 |
/* input args are live */
|
1285 |
for(i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
|
1286 |
arg = args[i]; |
1287 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1288 |
if (dead_temps[arg]) {
|
1289 |
dead_args |= (1 << i);
|
1290 |
} |
1291 |
dead_temps[arg] = 0;
|
1292 |
} |
1293 |
} |
1294 |
s->op_dead_args[op_index] = dead_args; |
1295 |
} |
1296 |
args--; |
1297 |
} |
1298 |
break;
|
1299 |
case INDEX_op_debug_insn_start:
|
1300 |
args -= def->nb_args; |
1301 |
break;
|
1302 |
case INDEX_op_nopn:
|
1303 |
nb_args = args[-1];
|
1304 |
args -= nb_args; |
1305 |
break;
|
1306 |
case INDEX_op_discard:
|
1307 |
args--; |
1308 |
/* mark the temporary as dead */
|
1309 |
dead_temps[args[0]] = 1; |
1310 |
break;
|
1311 |
case INDEX_op_end:
|
1312 |
break;
|
1313 |
/* XXX: optimize by hardcoding common cases (e.g. triadic ops) */
|
1314 |
default:
|
1315 |
args -= def->nb_args; |
1316 |
nb_iargs = def->nb_iargs; |
1317 |
nb_oargs = def->nb_oargs; |
1318 |
|
1319 |
/* Test if the operation can be removed because all
|
1320 |
its outputs are dead. We assume that nb_oargs == 0
|
1321 |
implies side effects */
|
1322 |
if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { |
1323 |
for(i = 0; i < nb_oargs; i++) { |
1324 |
arg = args[i]; |
1325 |
if (!dead_temps[arg])
|
1326 |
goto do_not_remove;
|
1327 |
} |
1328 |
tcg_set_nop(s, gen_opc_buf + op_index, args, def->nb_args); |
1329 |
#ifdef CONFIG_PROFILER
|
1330 |
s->del_op_count++; |
1331 |
#endif
|
1332 |
} else {
|
1333 |
do_not_remove:
|
1334 |
|
1335 |
/* output args are dead */
|
1336 |
dead_args = 0;
|
1337 |
for(i = 0; i < nb_oargs; i++) { |
1338 |
arg = args[i]; |
1339 |
if (dead_temps[arg]) {
|
1340 |
dead_args |= (1 << i);
|
1341 |
} |
1342 |
dead_temps[arg] = 1;
|
1343 |
} |
1344 |
|
1345 |
/* if end of basic block, update */
|
1346 |
if (def->flags & TCG_OPF_BB_END) {
|
1347 |
tcg_la_bb_end(s, dead_temps); |
1348 |
} else if (def->flags & TCG_OPF_CALL_CLOBBER) { |
1349 |
/* globals are live */
|
1350 |
memset(dead_temps, 0, s->nb_globals);
|
1351 |
} |
1352 |
|
1353 |
/* input args are live */
|
1354 |
for(i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
|
1355 |
arg = args[i]; |
1356 |
if (dead_temps[arg]) {
|
1357 |
dead_args |= (1 << i);
|
1358 |
} |
1359 |
dead_temps[arg] = 0;
|
1360 |
} |
1361 |
s->op_dead_args[op_index] = dead_args; |
1362 |
} |
1363 |
break;
|
1364 |
} |
1365 |
op_index--; |
1366 |
} |
1367 |
|
1368 |
if (args != gen_opparam_buf)
|
1369 |
tcg_abort(); |
1370 |
} |
1371 |
#else
|
1372 |
/* dummy liveness analysis */
|
1373 |
static void tcg_liveness_analysis(TCGContext *s) |
1374 |
{ |
1375 |
int nb_ops;
|
1376 |
nb_ops = gen_opc_ptr - gen_opc_buf; |
1377 |
|
1378 |
s->op_dead_args = tcg_malloc(nb_ops * sizeof(uint16_t));
|
1379 |
memset(s->op_dead_args, 0, nb_ops * sizeof(uint16_t)); |
1380 |
} |
1381 |
#endif
|
1382 |
|
1383 |
#ifndef NDEBUG
|
1384 |
static void dump_regs(TCGContext *s) |
1385 |
{ |
1386 |
TCGTemp *ts; |
1387 |
int i;
|
1388 |
char buf[64]; |
1389 |
|
1390 |
for(i = 0; i < s->nb_temps; i++) { |
1391 |
ts = &s->temps[i]; |
1392 |
printf(" %10s: ", tcg_get_arg_str_idx(s, buf, sizeof(buf), i)); |
1393 |
switch(ts->val_type) {
|
1394 |
case TEMP_VAL_REG:
|
1395 |
printf("%s", tcg_target_reg_names[ts->reg]);
|
1396 |
break;
|
1397 |
case TEMP_VAL_MEM:
|
1398 |
printf("%d(%s)", (int)ts->mem_offset, tcg_target_reg_names[ts->mem_reg]); |
1399 |
break;
|
1400 |
case TEMP_VAL_CONST:
|
1401 |
printf("$0x%" TCG_PRIlx, ts->val);
|
1402 |
break;
|
1403 |
case TEMP_VAL_DEAD:
|
1404 |
printf("D");
|
1405 |
break;
|
1406 |
default:
|
1407 |
printf("???");
|
1408 |
break;
|
1409 |
} |
1410 |
printf("\n");
|
1411 |
} |
1412 |
|
1413 |
for(i = 0; i < TCG_TARGET_NB_REGS; i++) { |
1414 |
if (s->reg_to_temp[i] >= 0) { |
1415 |
printf("%s: %s\n",
|
1416 |
tcg_target_reg_names[i], |
1417 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), s->reg_to_temp[i]));
|
1418 |
} |
1419 |
} |
1420 |
} |
1421 |
|
1422 |
static void check_regs(TCGContext *s) |
1423 |
{ |
1424 |
int reg, k;
|
1425 |
TCGTemp *ts; |
1426 |
char buf[64]; |
1427 |
|
1428 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1429 |
k = s->reg_to_temp[reg]; |
1430 |
if (k >= 0) { |
1431 |
ts = &s->temps[k]; |
1432 |
if (ts->val_type != TEMP_VAL_REG ||
|
1433 |
ts->reg != reg) { |
1434 |
printf("Inconsistency for register %s:\n",
|
1435 |
tcg_target_reg_names[reg]); |
1436 |
goto fail;
|
1437 |
} |
1438 |
} |
1439 |
} |
1440 |
for(k = 0; k < s->nb_temps; k++) { |
1441 |
ts = &s->temps[k]; |
1442 |
if (ts->val_type == TEMP_VAL_REG &&
|
1443 |
!ts->fixed_reg && |
1444 |
s->reg_to_temp[ts->reg] != k) { |
1445 |
printf("Inconsistency for temp %s:\n",
|
1446 |
tcg_get_arg_str_idx(s, buf, sizeof(buf), k));
|
1447 |
fail:
|
1448 |
printf("reg state:\n");
|
1449 |
dump_regs(s); |
1450 |
tcg_abort(); |
1451 |
} |
1452 |
} |
1453 |
} |
1454 |
#endif
|
1455 |
|
1456 |
static void temp_allocate_frame(TCGContext *s, int temp) |
1457 |
{ |
1458 |
TCGTemp *ts; |
1459 |
ts = &s->temps[temp]; |
1460 |
#ifndef __sparc_v9__ /* Sparc64 stack is accessed with offset of 2047 */ |
1461 |
s->current_frame_offset = (s->current_frame_offset + |
1462 |
(tcg_target_long)sizeof(tcg_target_long) - 1) & |
1463 |
~(sizeof(tcg_target_long) - 1); |
1464 |
#endif
|
1465 |
if (s->current_frame_offset + (tcg_target_long)sizeof(tcg_target_long) > |
1466 |
s->frame_end) { |
1467 |
tcg_abort(); |
1468 |
} |
1469 |
ts->mem_offset = s->current_frame_offset; |
1470 |
ts->mem_reg = s->frame_reg; |
1471 |
ts->mem_allocated = 1;
|
1472 |
s->current_frame_offset += (tcg_target_long)sizeof(tcg_target_long);
|
1473 |
} |
1474 |
|
1475 |
/* free register 'reg' by spilling the corresponding temporary if necessary */
|
1476 |
static void tcg_reg_free(TCGContext *s, int reg) |
1477 |
{ |
1478 |
TCGTemp *ts; |
1479 |
int temp;
|
1480 |
|
1481 |
temp = s->reg_to_temp[reg]; |
1482 |
if (temp != -1) { |
1483 |
ts = &s->temps[temp]; |
1484 |
assert(ts->val_type == TEMP_VAL_REG); |
1485 |
if (!ts->mem_coherent) {
|
1486 |
if (!ts->mem_allocated)
|
1487 |
temp_allocate_frame(s, temp); |
1488 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1489 |
} |
1490 |
ts->val_type = TEMP_VAL_MEM; |
1491 |
s->reg_to_temp[reg] = -1;
|
1492 |
} |
1493 |
} |
1494 |
|
1495 |
/* Allocate a register belonging to reg1 & ~reg2 */
|
1496 |
static int tcg_reg_alloc(TCGContext *s, TCGRegSet reg1, TCGRegSet reg2) |
1497 |
{ |
1498 |
int i, reg;
|
1499 |
TCGRegSet reg_ct; |
1500 |
|
1501 |
tcg_regset_andnot(reg_ct, reg1, reg2); |
1502 |
|
1503 |
/* first try free registers */
|
1504 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1505 |
reg = tcg_target_reg_alloc_order[i]; |
1506 |
if (tcg_regset_test_reg(reg_ct, reg) && s->reg_to_temp[reg] == -1) |
1507 |
return reg;
|
1508 |
} |
1509 |
|
1510 |
/* XXX: do better spill choice */
|
1511 |
for(i = 0; i < ARRAY_SIZE(tcg_target_reg_alloc_order); i++) { |
1512 |
reg = tcg_target_reg_alloc_order[i]; |
1513 |
if (tcg_regset_test_reg(reg_ct, reg)) {
|
1514 |
tcg_reg_free(s, reg); |
1515 |
return reg;
|
1516 |
} |
1517 |
} |
1518 |
|
1519 |
tcg_abort(); |
1520 |
} |
1521 |
|
1522 |
/* save a temporary to memory. 'allocated_regs' is used in case a
|
1523 |
temporary registers needs to be allocated to store a constant. */
|
1524 |
static void temp_save(TCGContext *s, int temp, TCGRegSet allocated_regs) |
1525 |
{ |
1526 |
TCGTemp *ts; |
1527 |
int reg;
|
1528 |
|
1529 |
ts = &s->temps[temp]; |
1530 |
if (!ts->fixed_reg) {
|
1531 |
switch(ts->val_type) {
|
1532 |
case TEMP_VAL_REG:
|
1533 |
tcg_reg_free(s, ts->reg); |
1534 |
break;
|
1535 |
case TEMP_VAL_DEAD:
|
1536 |
ts->val_type = TEMP_VAL_MEM; |
1537 |
break;
|
1538 |
case TEMP_VAL_CONST:
|
1539 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1540 |
allocated_regs); |
1541 |
if (!ts->mem_allocated)
|
1542 |
temp_allocate_frame(s, temp); |
1543 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1544 |
tcg_out_st(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1545 |
ts->val_type = TEMP_VAL_MEM; |
1546 |
break;
|
1547 |
case TEMP_VAL_MEM:
|
1548 |
break;
|
1549 |
default:
|
1550 |
tcg_abort(); |
1551 |
} |
1552 |
} |
1553 |
} |
1554 |
|
1555 |
/* save globals to their canonical location and assume they can be
|
1556 |
modified be the following code. 'allocated_regs' is used in case a
|
1557 |
temporary registers needs to be allocated to store a constant. */
|
1558 |
static void save_globals(TCGContext *s, TCGRegSet allocated_regs) |
1559 |
{ |
1560 |
int i;
|
1561 |
|
1562 |
for(i = 0; i < s->nb_globals; i++) { |
1563 |
temp_save(s, i, allocated_regs); |
1564 |
} |
1565 |
} |
1566 |
|
1567 |
/* at the end of a basic block, we assume all temporaries are dead and
|
1568 |
all globals are stored at their canonical location. */
|
1569 |
static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) |
1570 |
{ |
1571 |
TCGTemp *ts; |
1572 |
int i;
|
1573 |
|
1574 |
for(i = s->nb_globals; i < s->nb_temps; i++) {
|
1575 |
ts = &s->temps[i]; |
1576 |
if (ts->temp_local) {
|
1577 |
temp_save(s, i, allocated_regs); |
1578 |
} else {
|
1579 |
if (ts->val_type == TEMP_VAL_REG) {
|
1580 |
s->reg_to_temp[ts->reg] = -1;
|
1581 |
} |
1582 |
ts->val_type = TEMP_VAL_DEAD; |
1583 |
} |
1584 |
} |
1585 |
|
1586 |
save_globals(s, allocated_regs); |
1587 |
} |
1588 |
|
1589 |
#define IS_DEAD_ARG(n) ((dead_args >> (n)) & 1) |
1590 |
|
1591 |
static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args) |
1592 |
{ |
1593 |
TCGTemp *ots; |
1594 |
tcg_target_ulong val; |
1595 |
|
1596 |
ots = &s->temps[args[0]];
|
1597 |
val = args[1];
|
1598 |
|
1599 |
if (ots->fixed_reg) {
|
1600 |
/* for fixed registers, we do not do any constant
|
1601 |
propagation */
|
1602 |
tcg_out_movi(s, ots->type, ots->reg, val); |
1603 |
} else {
|
1604 |
/* The movi is not explicitly generated here */
|
1605 |
if (ots->val_type == TEMP_VAL_REG)
|
1606 |
s->reg_to_temp[ots->reg] = -1;
|
1607 |
ots->val_type = TEMP_VAL_CONST; |
1608 |
ots->val = val; |
1609 |
} |
1610 |
} |
1611 |
|
1612 |
static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, |
1613 |
const TCGArg *args,
|
1614 |
unsigned int dead_args) |
1615 |
{ |
1616 |
TCGTemp *ts, *ots; |
1617 |
int reg;
|
1618 |
const TCGArgConstraint *arg_ct;
|
1619 |
|
1620 |
ots = &s->temps[args[0]];
|
1621 |
ts = &s->temps[args[1]];
|
1622 |
arg_ct = &def->args_ct[0];
|
1623 |
|
1624 |
/* XXX: always mark arg dead if IS_DEAD_ARG(1) */
|
1625 |
if (ts->val_type == TEMP_VAL_REG) {
|
1626 |
if (IS_DEAD_ARG(1) && !ts->fixed_reg && !ots->fixed_reg) { |
1627 |
/* the mov can be suppressed */
|
1628 |
if (ots->val_type == TEMP_VAL_REG)
|
1629 |
s->reg_to_temp[ots->reg] = -1;
|
1630 |
reg = ts->reg; |
1631 |
s->reg_to_temp[reg] = -1;
|
1632 |
ts->val_type = TEMP_VAL_DEAD; |
1633 |
} else {
|
1634 |
if (ots->val_type == TEMP_VAL_REG) {
|
1635 |
reg = ots->reg; |
1636 |
} else {
|
1637 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1638 |
} |
1639 |
if (ts->reg != reg) {
|
1640 |
tcg_out_mov(s, ots->type, reg, ts->reg); |
1641 |
} |
1642 |
} |
1643 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1644 |
if (ots->val_type == TEMP_VAL_REG) {
|
1645 |
reg = ots->reg; |
1646 |
} else {
|
1647 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, s->reserved_regs); |
1648 |
} |
1649 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1650 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1651 |
if (ots->fixed_reg) {
|
1652 |
reg = ots->reg; |
1653 |
tcg_out_movi(s, ots->type, reg, ts->val); |
1654 |
} else {
|
1655 |
/* propagate constant */
|
1656 |
if (ots->val_type == TEMP_VAL_REG)
|
1657 |
s->reg_to_temp[ots->reg] = -1;
|
1658 |
ots->val_type = TEMP_VAL_CONST; |
1659 |
ots->val = ts->val; |
1660 |
return;
|
1661 |
} |
1662 |
} else {
|
1663 |
tcg_abort(); |
1664 |
} |
1665 |
s->reg_to_temp[reg] = args[0];
|
1666 |
ots->reg = reg; |
1667 |
ots->val_type = TEMP_VAL_REG; |
1668 |
ots->mem_coherent = 0;
|
1669 |
} |
1670 |
|
1671 |
static void tcg_reg_alloc_op(TCGContext *s, |
1672 |
const TCGOpDef *def, TCGOpcode opc,
|
1673 |
const TCGArg *args,
|
1674 |
unsigned int dead_args) |
1675 |
{ |
1676 |
TCGRegSet allocated_regs; |
1677 |
int i, k, nb_iargs, nb_oargs, reg;
|
1678 |
TCGArg arg; |
1679 |
const TCGArgConstraint *arg_ct;
|
1680 |
TCGTemp *ts; |
1681 |
TCGArg new_args[TCG_MAX_OP_ARGS]; |
1682 |
int const_args[TCG_MAX_OP_ARGS];
|
1683 |
|
1684 |
nb_oargs = def->nb_oargs; |
1685 |
nb_iargs = def->nb_iargs; |
1686 |
|
1687 |
/* copy constants */
|
1688 |
memcpy(new_args + nb_oargs + nb_iargs, |
1689 |
args + nb_oargs + nb_iargs, |
1690 |
sizeof(TCGArg) * def->nb_cargs);
|
1691 |
|
1692 |
/* satisfy input constraints */
|
1693 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1694 |
for(k = 0; k < nb_iargs; k++) { |
1695 |
i = def->sorted_args[nb_oargs + k]; |
1696 |
arg = args[i]; |
1697 |
arg_ct = &def->args_ct[i]; |
1698 |
ts = &s->temps[arg]; |
1699 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1700 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1701 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1702 |
ts->val_type = TEMP_VAL_REG; |
1703 |
ts->reg = reg; |
1704 |
ts->mem_coherent = 1;
|
1705 |
s->reg_to_temp[reg] = arg; |
1706 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1707 |
if (tcg_target_const_match(ts->val, arg_ct)) {
|
1708 |
/* constant is OK for instruction */
|
1709 |
const_args[i] = 1;
|
1710 |
new_args[i] = ts->val; |
1711 |
goto iarg_end;
|
1712 |
} else {
|
1713 |
/* need to move to a register */
|
1714 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1715 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1716 |
ts->val_type = TEMP_VAL_REG; |
1717 |
ts->reg = reg; |
1718 |
ts->mem_coherent = 0;
|
1719 |
s->reg_to_temp[reg] = arg; |
1720 |
} |
1721 |
} |
1722 |
assert(ts->val_type == TEMP_VAL_REG); |
1723 |
if (arg_ct->ct & TCG_CT_IALIAS) {
|
1724 |
if (ts->fixed_reg) {
|
1725 |
/* if fixed register, we must allocate a new register
|
1726 |
if the alias is not the same register */
|
1727 |
if (arg != args[arg_ct->alias_index])
|
1728 |
goto allocate_in_reg;
|
1729 |
} else {
|
1730 |
/* if the input is aliased to an output and if it is
|
1731 |
not dead after the instruction, we must allocate
|
1732 |
a new register and move it */
|
1733 |
if (!IS_DEAD_ARG(i)) {
|
1734 |
goto allocate_in_reg;
|
1735 |
} |
1736 |
} |
1737 |
} |
1738 |
reg = ts->reg; |
1739 |
if (tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1740 |
/* nothing to do : the constraint is satisfied */
|
1741 |
} else {
|
1742 |
allocate_in_reg:
|
1743 |
/* allocate a new register matching the constraint
|
1744 |
and move the temporary register into it */
|
1745 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1746 |
tcg_out_mov(s, ts->type, reg, ts->reg); |
1747 |
} |
1748 |
new_args[i] = reg; |
1749 |
const_args[i] = 0;
|
1750 |
tcg_regset_set_reg(allocated_regs, reg); |
1751 |
iarg_end: ;
|
1752 |
} |
1753 |
|
1754 |
if (def->flags & TCG_OPF_BB_END) {
|
1755 |
tcg_reg_alloc_bb_end(s, allocated_regs); |
1756 |
} else {
|
1757 |
/* mark dead temporaries and free the associated registers */
|
1758 |
for(i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
|
1759 |
arg = args[i]; |
1760 |
if (IS_DEAD_ARG(i)) {
|
1761 |
ts = &s->temps[arg]; |
1762 |
if (!ts->fixed_reg) {
|
1763 |
if (ts->val_type == TEMP_VAL_REG)
|
1764 |
s->reg_to_temp[ts->reg] = -1;
|
1765 |
ts->val_type = TEMP_VAL_DEAD; |
1766 |
} |
1767 |
} |
1768 |
} |
1769 |
|
1770 |
if (def->flags & TCG_OPF_CALL_CLOBBER) {
|
1771 |
/* XXX: permit generic clobber register list ? */
|
1772 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1773 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1774 |
tcg_reg_free(s, reg); |
1775 |
} |
1776 |
} |
1777 |
/* XXX: for load/store we could do that only for the slow path
|
1778 |
(i.e. when a memory callback is called) */
|
1779 |
|
1780 |
/* store globals and free associated registers (we assume the insn
|
1781 |
can modify any global. */
|
1782 |
save_globals(s, allocated_regs); |
1783 |
} |
1784 |
|
1785 |
/* satisfy the output constraints */
|
1786 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1787 |
for(k = 0; k < nb_oargs; k++) { |
1788 |
i = def->sorted_args[k]; |
1789 |
arg = args[i]; |
1790 |
arg_ct = &def->args_ct[i]; |
1791 |
ts = &s->temps[arg]; |
1792 |
if (arg_ct->ct & TCG_CT_ALIAS) {
|
1793 |
reg = new_args[arg_ct->alias_index]; |
1794 |
} else {
|
1795 |
/* if fixed register, we try to use it */
|
1796 |
reg = ts->reg; |
1797 |
if (ts->fixed_reg &&
|
1798 |
tcg_regset_test_reg(arg_ct->u.regs, reg)) { |
1799 |
goto oarg_end;
|
1800 |
} |
1801 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1802 |
} |
1803 |
tcg_regset_set_reg(allocated_regs, reg); |
1804 |
/* if a fixed register is used, then a move will be done afterwards */
|
1805 |
if (!ts->fixed_reg) {
|
1806 |
if (ts->val_type == TEMP_VAL_REG)
|
1807 |
s->reg_to_temp[ts->reg] = -1;
|
1808 |
if (IS_DEAD_ARG(i)) {
|
1809 |
ts->val_type = TEMP_VAL_DEAD; |
1810 |
} else {
|
1811 |
ts->val_type = TEMP_VAL_REG; |
1812 |
ts->reg = reg; |
1813 |
/* temp value is modified, so the value kept in memory is
|
1814 |
potentially not the same */
|
1815 |
ts->mem_coherent = 0;
|
1816 |
s->reg_to_temp[reg] = arg; |
1817 |
} |
1818 |
} |
1819 |
oarg_end:
|
1820 |
new_args[i] = reg; |
1821 |
} |
1822 |
} |
1823 |
|
1824 |
/* emit instruction */
|
1825 |
tcg_out_op(s, opc, new_args, const_args); |
1826 |
|
1827 |
/* move the outputs in the correct register if needed */
|
1828 |
for(i = 0; i < nb_oargs; i++) { |
1829 |
ts = &s->temps[args[i]]; |
1830 |
reg = new_args[i]; |
1831 |
if (ts->fixed_reg && ts->reg != reg) {
|
1832 |
tcg_out_mov(s, ts->type, ts->reg, reg); |
1833 |
} |
1834 |
} |
1835 |
} |
1836 |
|
1837 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1838 |
#define STACK_DIR(x) (-(x))
|
1839 |
#else
|
1840 |
#define STACK_DIR(x) (x)
|
1841 |
#endif
|
1842 |
|
1843 |
static int tcg_reg_alloc_call(TCGContext *s, const TCGOpDef *def, |
1844 |
TCGOpcode opc, const TCGArg *args,
|
1845 |
unsigned int dead_args) |
1846 |
{ |
1847 |
int nb_iargs, nb_oargs, flags, nb_regs, i, reg, nb_params;
|
1848 |
TCGArg arg, func_arg; |
1849 |
TCGTemp *ts; |
1850 |
tcg_target_long stack_offset, call_stack_size, func_addr; |
1851 |
int const_func_arg, allocate_args;
|
1852 |
TCGRegSet allocated_regs; |
1853 |
const TCGArgConstraint *arg_ct;
|
1854 |
|
1855 |
arg = *args++; |
1856 |
|
1857 |
nb_oargs = arg >> 16;
|
1858 |
nb_iargs = arg & 0xffff;
|
1859 |
nb_params = nb_iargs - 1;
|
1860 |
|
1861 |
flags = args[nb_oargs + nb_iargs]; |
1862 |
|
1863 |
nb_regs = tcg_target_get_call_iarg_regs_count(flags); |
1864 |
if (nb_regs > nb_params)
|
1865 |
nb_regs = nb_params; |
1866 |
|
1867 |
/* assign stack slots first */
|
1868 |
call_stack_size = (nb_params - nb_regs) * sizeof(tcg_target_long);
|
1869 |
call_stack_size = (call_stack_size + TCG_TARGET_STACK_ALIGN - 1) &
|
1870 |
~(TCG_TARGET_STACK_ALIGN - 1);
|
1871 |
allocate_args = (call_stack_size > TCG_STATIC_CALL_ARGS_SIZE); |
1872 |
if (allocate_args) {
|
1873 |
/* XXX: if more than TCG_STATIC_CALL_ARGS_SIZE is needed,
|
1874 |
preallocate call stack */
|
1875 |
tcg_abort(); |
1876 |
} |
1877 |
|
1878 |
stack_offset = TCG_TARGET_CALL_STACK_OFFSET; |
1879 |
for(i = nb_regs; i < nb_params; i++) {
|
1880 |
arg = args[nb_oargs + i]; |
1881 |
#ifdef TCG_TARGET_STACK_GROWSUP
|
1882 |
stack_offset -= sizeof(tcg_target_long);
|
1883 |
#endif
|
1884 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1885 |
ts = &s->temps[arg]; |
1886 |
if (ts->val_type == TEMP_VAL_REG) {
|
1887 |
tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, stack_offset); |
1888 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1889 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1890 |
s->reserved_regs); |
1891 |
/* XXX: not correct if reading values from the stack */
|
1892 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1893 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1894 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1895 |
reg = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], |
1896 |
s->reserved_regs); |
1897 |
/* XXX: sign extend may be needed on some targets */
|
1898 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1899 |
tcg_out_st(s, ts->type, reg, TCG_REG_CALL_STACK, stack_offset); |
1900 |
} else {
|
1901 |
tcg_abort(); |
1902 |
} |
1903 |
} |
1904 |
#ifndef TCG_TARGET_STACK_GROWSUP
|
1905 |
stack_offset += sizeof(tcg_target_long);
|
1906 |
#endif
|
1907 |
} |
1908 |
|
1909 |
/* assign input registers */
|
1910 |
tcg_regset_set(allocated_regs, s->reserved_regs); |
1911 |
for(i = 0; i < nb_regs; i++) { |
1912 |
arg = args[nb_oargs + i]; |
1913 |
if (arg != TCG_CALL_DUMMY_ARG) {
|
1914 |
ts = &s->temps[arg]; |
1915 |
reg = tcg_target_call_iarg_regs[i]; |
1916 |
tcg_reg_free(s, reg); |
1917 |
if (ts->val_type == TEMP_VAL_REG) {
|
1918 |
if (ts->reg != reg) {
|
1919 |
tcg_out_mov(s, ts->type, reg, ts->reg); |
1920 |
} |
1921 |
} else if (ts->val_type == TEMP_VAL_MEM) { |
1922 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1923 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1924 |
/* XXX: sign extend ? */
|
1925 |
tcg_out_movi(s, ts->type, reg, ts->val); |
1926 |
} else {
|
1927 |
tcg_abort(); |
1928 |
} |
1929 |
tcg_regset_set_reg(allocated_regs, reg); |
1930 |
} |
1931 |
} |
1932 |
|
1933 |
/* assign function address */
|
1934 |
func_arg = args[nb_oargs + nb_iargs - 1];
|
1935 |
arg_ct = &def->args_ct[0];
|
1936 |
ts = &s->temps[func_arg]; |
1937 |
func_addr = ts->val; |
1938 |
const_func_arg = 0;
|
1939 |
if (ts->val_type == TEMP_VAL_MEM) {
|
1940 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1941 |
tcg_out_ld(s, ts->type, reg, ts->mem_reg, ts->mem_offset); |
1942 |
func_arg = reg; |
1943 |
tcg_regset_set_reg(allocated_regs, reg); |
1944 |
} else if (ts->val_type == TEMP_VAL_REG) { |
1945 |
reg = ts->reg; |
1946 |
if (!tcg_regset_test_reg(arg_ct->u.regs, reg)) {
|
1947 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1948 |
tcg_out_mov(s, ts->type, reg, ts->reg); |
1949 |
} |
1950 |
func_arg = reg; |
1951 |
tcg_regset_set_reg(allocated_regs, reg); |
1952 |
} else if (ts->val_type == TEMP_VAL_CONST) { |
1953 |
if (tcg_target_const_match(func_addr, arg_ct)) {
|
1954 |
const_func_arg = 1;
|
1955 |
func_arg = func_addr; |
1956 |
} else {
|
1957 |
reg = tcg_reg_alloc(s, arg_ct->u.regs, allocated_regs); |
1958 |
tcg_out_movi(s, ts->type, reg, func_addr); |
1959 |
func_arg = reg; |
1960 |
tcg_regset_set_reg(allocated_regs, reg); |
1961 |
} |
1962 |
} else {
|
1963 |
tcg_abort(); |
1964 |
} |
1965 |
|
1966 |
|
1967 |
/* mark dead temporaries and free the associated registers */
|
1968 |
for(i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
|
1969 |
arg = args[i]; |
1970 |
if (IS_DEAD_ARG(i)) {
|
1971 |
ts = &s->temps[arg]; |
1972 |
if (!ts->fixed_reg) {
|
1973 |
if (ts->val_type == TEMP_VAL_REG)
|
1974 |
s->reg_to_temp[ts->reg] = -1;
|
1975 |
ts->val_type = TEMP_VAL_DEAD; |
1976 |
} |
1977 |
} |
1978 |
} |
1979 |
|
1980 |
/* clobber call registers */
|
1981 |
for(reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { |
1982 |
if (tcg_regset_test_reg(tcg_target_call_clobber_regs, reg)) {
|
1983 |
tcg_reg_free(s, reg); |
1984 |
} |
1985 |
} |
1986 |
|
1987 |
/* store globals and free associated registers (we assume the call
|
1988 |
can modify any global. */
|
1989 |
if (!(flags & TCG_CALL_CONST)) {
|
1990 |
save_globals(s, allocated_regs); |
1991 |
} |
1992 |
|
1993 |
tcg_out_op(s, opc, &func_arg, &const_func_arg); |
1994 |
|
1995 |
/* assign output registers and emit moves if needed */
|
1996 |
for(i = 0; i < nb_oargs; i++) { |
1997 |
arg = args[i]; |
1998 |
ts = &s->temps[arg]; |
1999 |
reg = tcg_target_call_oarg_regs[i]; |
2000 |
assert(s->reg_to_temp[reg] == -1);
|
2001 |
if (ts->fixed_reg) {
|
2002 |
if (ts->reg != reg) {
|
2003 |
tcg_out_mov(s, ts->type, ts->reg, reg); |
2004 |
} |
2005 |
} else {
|
2006 |
if (ts->val_type == TEMP_VAL_REG)
|
2007 |
s->reg_to_temp[ts->reg] = -1;
|
2008 |
if (IS_DEAD_ARG(i)) {
|
2009 |
ts->val_type = TEMP_VAL_DEAD; |
2010 |
} else {
|
2011 |
ts->val_type = TEMP_VAL_REG; |
2012 |
ts->reg = reg; |
2013 |
ts->mem_coherent = 0;
|
2014 |
s->reg_to_temp[reg] = arg; |
2015 |
} |
2016 |
} |
2017 |
} |
2018 |
|
2019 |
return nb_iargs + nb_oargs + def->nb_cargs + 1; |
2020 |
} |
2021 |
|
2022 |
#ifdef CONFIG_PROFILER
|
2023 |
|
2024 |
static int64_t tcg_table_op_count[NB_OPS];
|
2025 |
|
2026 |
static void dump_op_count(void) |
2027 |
{ |
2028 |
int i;
|
2029 |
FILE *f; |
2030 |
f = fopen("/tmp/op.log", "w"); |
2031 |
for(i = INDEX_op_end; i < NB_OPS; i++) {
|
2032 |
fprintf(f, "%s %" PRId64 "\n", tcg_op_defs[i].name, tcg_table_op_count[i]); |
2033 |
} |
2034 |
fclose(f); |
2035 |
} |
2036 |
#endif
|
2037 |
|
2038 |
|
2039 |
static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf, |
2040 |
long search_pc)
|
2041 |
{ |
2042 |
TCGOpcode opc; |
2043 |
int op_index;
|
2044 |
const TCGOpDef *def;
|
2045 |
unsigned int dead_args; |
2046 |
const TCGArg *args;
|
2047 |
|
2048 |
#ifdef DEBUG_DISAS
|
2049 |
if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP))) {
|
2050 |
qemu_log("OP:\n");
|
2051 |
tcg_dump_ops(s); |
2052 |
qemu_log("\n");
|
2053 |
} |
2054 |
#endif
|
2055 |
|
2056 |
#ifdef CONFIG_PROFILER
|
2057 |
s->opt_time -= profile_getclock(); |
2058 |
#endif
|
2059 |
|
2060 |
#ifdef USE_TCG_OPTIMIZATIONS
|
2061 |
gen_opparam_ptr = |
2062 |
tcg_optimize(s, gen_opc_ptr, gen_opparam_buf, tcg_op_defs); |
2063 |
#endif
|
2064 |
|
2065 |
#ifdef CONFIG_PROFILER
|
2066 |
s->opt_time += profile_getclock(); |
2067 |
s->la_time -= profile_getclock(); |
2068 |
#endif
|
2069 |
|
2070 |
tcg_liveness_analysis(s); |
2071 |
|
2072 |
#ifdef CONFIG_PROFILER
|
2073 |
s->la_time += profile_getclock(); |
2074 |
#endif
|
2075 |
|
2076 |
#ifdef DEBUG_DISAS
|
2077 |
if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT))) {
|
2078 |
qemu_log("OP after optimization and liveness analysis:\n");
|
2079 |
tcg_dump_ops(s); |
2080 |
qemu_log("\n");
|
2081 |
} |
2082 |
#endif
|
2083 |
|
2084 |
tcg_reg_alloc_start(s); |
2085 |
|
2086 |
s->code_buf = gen_code_buf; |
2087 |
s->code_ptr = gen_code_buf; |
2088 |
|
2089 |
args = gen_opparam_buf; |
2090 |
op_index = 0;
|
2091 |
|
2092 |
for(;;) {
|
2093 |
opc = gen_opc_buf[op_index]; |
2094 |
#ifdef CONFIG_PROFILER
|
2095 |
tcg_table_op_count[opc]++; |
2096 |
#endif
|
2097 |
def = &tcg_op_defs[opc]; |
2098 |
#if 0
|
2099 |
printf("%s: %d %d %d\n", def->name,
|
2100 |
def->nb_oargs, def->nb_iargs, def->nb_cargs);
|
2101 |
// dump_regs(s);
|
2102 |
#endif
|
2103 |
switch(opc) {
|
2104 |
case INDEX_op_mov_i32:
|
2105 |
#if TCG_TARGET_REG_BITS == 64 |
2106 |
case INDEX_op_mov_i64:
|
2107 |
#endif
|
2108 |
dead_args = s->op_dead_args[op_index]; |
2109 |
tcg_reg_alloc_mov(s, def, args, dead_args); |
2110 |
break;
|
2111 |
case INDEX_op_movi_i32:
|
2112 |
#if TCG_TARGET_REG_BITS == 64 |
2113 |
case INDEX_op_movi_i64:
|
2114 |
#endif
|
2115 |
tcg_reg_alloc_movi(s, args); |
2116 |
break;
|
2117 |
case INDEX_op_debug_insn_start:
|
2118 |
/* debug instruction */
|
2119 |
break;
|
2120 |
case INDEX_op_nop:
|
2121 |
case INDEX_op_nop1:
|
2122 |
case INDEX_op_nop2:
|
2123 |
case INDEX_op_nop3:
|
2124 |
break;
|
2125 |
case INDEX_op_nopn:
|
2126 |
args += args[0];
|
2127 |
goto next;
|
2128 |
case INDEX_op_discard:
|
2129 |
{ |
2130 |
TCGTemp *ts; |
2131 |
ts = &s->temps[args[0]];
|
2132 |
/* mark the temporary as dead */
|
2133 |
if (!ts->fixed_reg) {
|
2134 |
if (ts->val_type == TEMP_VAL_REG)
|
2135 |
s->reg_to_temp[ts->reg] = -1;
|
2136 |
ts->val_type = TEMP_VAL_DEAD; |
2137 |
} |
2138 |
} |
2139 |
break;
|
2140 |
case INDEX_op_set_label:
|
2141 |
tcg_reg_alloc_bb_end(s, s->reserved_regs); |
2142 |
tcg_out_label(s, args[0], s->code_ptr);
|
2143 |
break;
|
2144 |
case INDEX_op_call:
|
2145 |
dead_args = s->op_dead_args[op_index]; |
2146 |
args += tcg_reg_alloc_call(s, def, opc, args, dead_args); |
2147 |
goto next;
|
2148 |
case INDEX_op_end:
|
2149 |
goto the_end;
|
2150 |
default:
|
2151 |
/* Sanity check that we've not introduced any unhandled opcodes. */
|
2152 |
if (def->flags & TCG_OPF_NOT_PRESENT) {
|
2153 |
tcg_abort(); |
2154 |
} |
2155 |
/* Note: in order to speed up the code, it would be much
|
2156 |
faster to have specialized register allocator functions for
|
2157 |
some common argument patterns */
|
2158 |
dead_args = s->op_dead_args[op_index]; |
2159 |
tcg_reg_alloc_op(s, def, opc, args, dead_args); |
2160 |
break;
|
2161 |
} |
2162 |
args += def->nb_args; |
2163 |
next:
|
2164 |
if (search_pc >= 0 && search_pc < s->code_ptr - gen_code_buf) { |
2165 |
return op_index;
|
2166 |
} |
2167 |
op_index++; |
2168 |
#ifndef NDEBUG
|
2169 |
check_regs(s); |
2170 |
#endif
|
2171 |
} |
2172 |
the_end:
|
2173 |
return -1; |
2174 |
} |
2175 |
|
2176 |
int tcg_gen_code(TCGContext *s, uint8_t *gen_code_buf)
|
2177 |
{ |
2178 |
#ifdef CONFIG_PROFILER
|
2179 |
{ |
2180 |
int n;
|
2181 |
n = (gen_opc_ptr - gen_opc_buf); |
2182 |
s->op_count += n; |
2183 |
if (n > s->op_count_max)
|
2184 |
s->op_count_max = n; |
2185 |
|
2186 |
s->temp_count += s->nb_temps; |
2187 |
if (s->nb_temps > s->temp_count_max)
|
2188 |
s->temp_count_max = s->nb_temps; |
2189 |
} |
2190 |
#endif
|
2191 |
|
2192 |
tcg_gen_code_common(s, gen_code_buf, -1);
|
2193 |
|
2194 |
/* flush instruction cache */
|
2195 |
flush_icache_range((tcg_target_ulong)gen_code_buf, |
2196 |
(tcg_target_ulong)s->code_ptr); |
2197 |
|
2198 |
return s->code_ptr - gen_code_buf;
|
2199 |
} |
2200 |
|
2201 |
/* Return the index of the micro operation such as the pc after is <
|
2202 |
offset bytes from the start of the TB. The contents of gen_code_buf must
|
2203 |
not be changed, though writing the same values is ok.
|
2204 |
Return -1 if not found. */
|
2205 |
int tcg_gen_code_search_pc(TCGContext *s, uint8_t *gen_code_buf, long offset) |
2206 |
{ |
2207 |
return tcg_gen_code_common(s, gen_code_buf, offset);
|
2208 |
} |
2209 |
|
2210 |
#ifdef CONFIG_PROFILER
|
2211 |
void tcg_dump_info(FILE *f, fprintf_function cpu_fprintf)
|
2212 |
{ |
2213 |
TCGContext *s = &tcg_ctx; |
2214 |
int64_t tot; |
2215 |
|
2216 |
tot = s->interm_time + s->code_time; |
2217 |
cpu_fprintf(f, "JIT cycles %" PRId64 " (%0.3f s at 2.4 GHz)\n", |
2218 |
tot, tot / 2.4e9); |
2219 |
cpu_fprintf(f, "translated TBs %" PRId64 " (aborted=%" PRId64 " %0.1f%%)\n", |
2220 |
s->tb_count, |
2221 |
s->tb_count1 - s->tb_count, |
2222 |
s->tb_count1 ? (double)(s->tb_count1 - s->tb_count) / s->tb_count1 * 100.0 : 0); |
2223 |
cpu_fprintf(f, "avg ops/TB %0.1f max=%d\n",
|
2224 |
s->tb_count ? (double)s->op_count / s->tb_count : 0, s->op_count_max); |
2225 |
cpu_fprintf(f, "deleted ops/TB %0.2f\n",
|
2226 |
s->tb_count ? |
2227 |
(double)s->del_op_count / s->tb_count : 0); |
2228 |
cpu_fprintf(f, "avg temps/TB %0.2f max=%d\n",
|
2229 |
s->tb_count ? |
2230 |
(double)s->temp_count / s->tb_count : 0, |
2231 |
s->temp_count_max); |
2232 |
|
2233 |
cpu_fprintf(f, "cycles/op %0.1f\n",
|
2234 |
s->op_count ? (double)tot / s->op_count : 0); |
2235 |
cpu_fprintf(f, "cycles/in byte %0.1f\n",
|
2236 |
s->code_in_len ? (double)tot / s->code_in_len : 0); |
2237 |
cpu_fprintf(f, "cycles/out byte %0.1f\n",
|
2238 |
s->code_out_len ? (double)tot / s->code_out_len : 0); |
2239 |
if (tot == 0) |
2240 |
tot = 1;
|
2241 |
cpu_fprintf(f, " gen_interm time %0.1f%%\n",
|
2242 |
(double)s->interm_time / tot * 100.0); |
2243 |
cpu_fprintf(f, " gen_code time %0.1f%%\n",
|
2244 |
(double)s->code_time / tot * 100.0); |
2245 |
cpu_fprintf(f, "optim./code time %0.1f%%\n",
|
2246 |
(double)s->opt_time / (s->code_time ? s->code_time : 1) |
2247 |
* 100.0); |
2248 |
cpu_fprintf(f, "liveness/code time %0.1f%%\n",
|
2249 |
(double)s->la_time / (s->code_time ? s->code_time : 1) * 100.0); |
2250 |
cpu_fprintf(f, "cpu_restore count %" PRId64 "\n", |
2251 |
s->restore_count); |
2252 |
cpu_fprintf(f, " avg cycles %0.1f\n",
|
2253 |
s->restore_count ? (double)s->restore_time / s->restore_count : 0); |
2254 |
|
2255 |
dump_op_count(); |
2256 |
} |
2257 |
#else
|
2258 |
void tcg_dump_info(FILE *f, fprintf_function cpu_fprintf)
|
2259 |
{ |
2260 |
cpu_fprintf(f, "[TCG profiler not compiled]\n");
|
2261 |
} |
2262 |
#endif
|
2263 |
|
2264 |
#ifdef ELF_HOST_MACHINE
|
2265 |
/* In order to use this feature, the backend needs to do three things:
|
2266 |
|
2267 |
(1) Define ELF_HOST_MACHINE to indicate both what value to
|
2268 |
put into the ELF image and to indicate support for the feature.
|
2269 |
|
2270 |
(2) Define tcg_register_jit. This should create a buffer containing
|
2271 |
the contents of a .debug_frame section that describes the post-
|
2272 |
prologue unwind info for the tcg machine.
|
2273 |
|
2274 |
(3) Call tcg_register_jit_int, with the constructed .debug_frame.
|
2275 |
*/
|
2276 |
|
2277 |
/* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */
|
2278 |
typedef enum { |
2279 |
JIT_NOACTION = 0,
|
2280 |
JIT_REGISTER_FN, |
2281 |
JIT_UNREGISTER_FN |
2282 |
} jit_actions_t; |
2283 |
|
2284 |
struct jit_code_entry {
|
2285 |
struct jit_code_entry *next_entry;
|
2286 |
struct jit_code_entry *prev_entry;
|
2287 |
const void *symfile_addr; |
2288 |
uint64_t symfile_size; |
2289 |
}; |
2290 |
|
2291 |
struct jit_descriptor {
|
2292 |
uint32_t version; |
2293 |
uint32_t action_flag; |
2294 |
struct jit_code_entry *relevant_entry;
|
2295 |
struct jit_code_entry *first_entry;
|
2296 |
}; |
2297 |
|
2298 |
void __jit_debug_register_code(void) __attribute__((noinline)); |
2299 |
void __jit_debug_register_code(void) |
2300 |
{ |
2301 |
asm(""); |
2302 |
} |
2303 |
|
2304 |
/* Must statically initialize the version, because GDB may check
|
2305 |
the version before we can set it. */
|
2306 |
struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; |
2307 |
|
2308 |
/* End GDB interface. */
|
2309 |
|
2310 |
static int find_string(const char *strtab, const char *str) |
2311 |
{ |
2312 |
const char *p = strtab + 1; |
2313 |
|
2314 |
while (1) { |
2315 |
if (strcmp(p, str) == 0) { |
2316 |
return p - strtab;
|
2317 |
} |
2318 |
p += strlen(p) + 1;
|
2319 |
} |
2320 |
} |
2321 |
|
2322 |
static void tcg_register_jit_int(void *buf_ptr, size_t buf_size, |
2323 |
void *debug_frame, size_t debug_frame_size)
|
2324 |
{ |
2325 |
struct __attribute__((packed)) DebugInfo {
|
2326 |
uint32_t len; |
2327 |
uint16_t version; |
2328 |
uint32_t abbrev; |
2329 |
uint8_t ptr_size; |
2330 |
uint8_t cu_die; |
2331 |
uint16_t cu_lang; |
2332 |
uintptr_t cu_low_pc; |
2333 |
uintptr_t cu_high_pc; |
2334 |
uint8_t fn_die; |
2335 |
char fn_name[16]; |
2336 |
uintptr_t fn_low_pc; |
2337 |
uintptr_t fn_high_pc; |
2338 |
uint8_t cu_eoc; |
2339 |
}; |
2340 |
|
2341 |
struct ElfImage {
|
2342 |
ElfW(Ehdr) ehdr; |
2343 |
ElfW(Phdr) phdr; |
2344 |
ElfW(Shdr) shdr[7];
|
2345 |
ElfW(Sym) sym[2];
|
2346 |
struct DebugInfo di;
|
2347 |
uint8_t da[24];
|
2348 |
char str[80]; |
2349 |
}; |
2350 |
|
2351 |
struct ElfImage *img;
|
2352 |
|
2353 |
static const struct ElfImage img_template = { |
2354 |
.ehdr = { |
2355 |
.e_ident[EI_MAG0] = ELFMAG0, |
2356 |
.e_ident[EI_MAG1] = ELFMAG1, |
2357 |
.e_ident[EI_MAG2] = ELFMAG2, |
2358 |
.e_ident[EI_MAG3] = ELFMAG3, |
2359 |
.e_ident[EI_CLASS] = ELF_CLASS, |
2360 |
.e_ident[EI_DATA] = ELF_DATA, |
2361 |
.e_ident[EI_VERSION] = EV_CURRENT, |
2362 |
.e_type = ET_EXEC, |
2363 |
.e_machine = ELF_HOST_MACHINE, |
2364 |
.e_version = EV_CURRENT, |
2365 |
.e_phoff = offsetof(struct ElfImage, phdr),
|
2366 |
.e_shoff = offsetof(struct ElfImage, shdr),
|
2367 |
.e_ehsize = sizeof(ElfW(Shdr)),
|
2368 |
.e_phentsize = sizeof(ElfW(Phdr)),
|
2369 |
.e_phnum = 1,
|
2370 |
.e_shentsize = sizeof(ElfW(Shdr)),
|
2371 |
.e_shnum = ARRAY_SIZE(img->shdr), |
2372 |
.e_shstrndx = ARRAY_SIZE(img->shdr) - 1,
|
2373 |
#ifdef ELF_HOST_FLAGS
|
2374 |
.e_flags = ELF_HOST_FLAGS, |
2375 |
#endif
|
2376 |
#ifdef ELF_OSABI
|
2377 |
.e_ident[EI_OSABI] = ELF_OSABI, |
2378 |
#endif
|
2379 |
}, |
2380 |
.phdr = { |
2381 |
.p_type = PT_LOAD, |
2382 |
.p_flags = PF_X, |
2383 |
}, |
2384 |
.shdr = { |
2385 |
[0] = { .sh_type = SHT_NULL },
|
2386 |
/* Trick: The contents of code_gen_buffer are not present in
|
2387 |
this fake ELF file; that got allocated elsewhere. Therefore
|
2388 |
we mark .text as SHT_NOBITS (similar to .bss) so that readers
|
2389 |
will not look for contents. We can record any address. */
|
2390 |
[1] = { /* .text */ |
2391 |
.sh_type = SHT_NOBITS, |
2392 |
.sh_flags = SHF_EXECINSTR | SHF_ALLOC, |
2393 |
}, |
2394 |
[2] = { /* .debug_info */ |
2395 |
.sh_type = SHT_PROGBITS, |
2396 |
.sh_offset = offsetof(struct ElfImage, di),
|
2397 |
.sh_size = sizeof(struct DebugInfo), |
2398 |
}, |
2399 |
[3] = { /* .debug_abbrev */ |
2400 |
.sh_type = SHT_PROGBITS, |
2401 |
.sh_offset = offsetof(struct ElfImage, da),
|
2402 |
.sh_size = sizeof(img->da),
|
2403 |
}, |
2404 |
[4] = { /* .debug_frame */ |
2405 |
.sh_type = SHT_PROGBITS, |
2406 |
.sh_offset = sizeof(struct ElfImage), |
2407 |
}, |
2408 |
[5] = { /* .symtab */ |
2409 |
.sh_type = SHT_SYMTAB, |
2410 |
.sh_offset = offsetof(struct ElfImage, sym),
|
2411 |
.sh_size = sizeof(img->sym),
|
2412 |
.sh_info = 1,
|
2413 |
.sh_link = ARRAY_SIZE(img->shdr) - 1,
|
2414 |
.sh_entsize = sizeof(ElfW(Sym)),
|
2415 |
}, |
2416 |
[6] = { /* .strtab */ |
2417 |
.sh_type = SHT_STRTAB, |
2418 |
.sh_offset = offsetof(struct ElfImage, str),
|
2419 |
.sh_size = sizeof(img->str),
|
2420 |
} |
2421 |
}, |
2422 |
.sym = { |
2423 |
[1] = { /* code_gen_buffer */ |
2424 |
.st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), |
2425 |
.st_shndx = 1,
|
2426 |
} |
2427 |
}, |
2428 |
.di = { |
2429 |
.len = sizeof(struct DebugInfo) - 4, |
2430 |
.version = 2,
|
2431 |
.ptr_size = sizeof(void *), |
2432 |
.cu_die = 1,
|
2433 |
.cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ |
2434 |
.fn_die = 2,
|
2435 |
.fn_name = "code_gen_buffer"
|
2436 |
}, |
2437 |
.da = { |
2438 |
1, /* abbrev number (the cu) */ |
2439 |
0x11, 1, /* DW_TAG_compile_unit, has children */ |
2440 |
0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ |
2441 |
0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ |
2442 |
0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ |
2443 |
0, 0, /* end of abbrev */ |
2444 |
2, /* abbrev number (the fn) */ |
2445 |
0x2e, 0, /* DW_TAG_subprogram, no children */ |
2446 |
0x3, 0x8, /* DW_AT_name, DW_FORM_string */ |
2447 |
0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ |
2448 |
0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ |
2449 |
0, 0, /* end of abbrev */ |
2450 |
0 /* no more abbrev */ |
2451 |
}, |
2452 |
.str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" |
2453 |
".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", |
2454 |
}; |
2455 |
|
2456 |
/* We only need a single jit entry; statically allocate it. */
|
2457 |
static struct jit_code_entry one_entry; |
2458 |
|
2459 |
uintptr_t buf = (uintptr_t)buf_ptr; |
2460 |
size_t img_size = sizeof(struct ElfImage) + debug_frame_size; |
2461 |
|
2462 |
img = g_malloc(img_size); |
2463 |
*img = img_template; |
2464 |
memcpy(img + 1, debug_frame, debug_frame_size);
|
2465 |
|
2466 |
img->phdr.p_vaddr = buf; |
2467 |
img->phdr.p_paddr = buf; |
2468 |
img->phdr.p_memsz = buf_size; |
2469 |
|
2470 |
img->shdr[1].sh_name = find_string(img->str, ".text"); |
2471 |
img->shdr[1].sh_addr = buf;
|
2472 |
img->shdr[1].sh_size = buf_size;
|
2473 |
|
2474 |
img->shdr[2].sh_name = find_string(img->str, ".debug_info"); |
2475 |
img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); |
2476 |
|
2477 |
img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); |
2478 |
img->shdr[4].sh_size = debug_frame_size;
|
2479 |
|
2480 |
img->shdr[5].sh_name = find_string(img->str, ".symtab"); |
2481 |
img->shdr[6].sh_name = find_string(img->str, ".strtab"); |
2482 |
|
2483 |
img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); |
2484 |
img->sym[1].st_value = buf;
|
2485 |
img->sym[1].st_size = buf_size;
|
2486 |
|
2487 |
img->di.cu_low_pc = buf; |
2488 |
img->di.cu_high_pc = buf_size; |
2489 |
img->di.fn_low_pc = buf; |
2490 |
img->di.fn_high_pc = buf_size; |
2491 |
|
2492 |
#ifdef DEBUG_JIT
|
2493 |
/* Enable this block to be able to debug the ELF image file creation.
|
2494 |
One can use readelf, objdump, or other inspection utilities. */
|
2495 |
{ |
2496 |
FILE *f = fopen("/tmp/qemu.jit", "w+b"); |
2497 |
if (f) {
|
2498 |
if (fwrite(img, img_size, 1, f) != img_size) { |
2499 |
/* Avoid stupid unused return value warning for fwrite. */
|
2500 |
} |
2501 |
fclose(f); |
2502 |
} |
2503 |
} |
2504 |
#endif
|
2505 |
|
2506 |
one_entry.symfile_addr = img; |
2507 |
one_entry.symfile_size = img_size; |
2508 |
|
2509 |
__jit_debug_descriptor.action_flag = JIT_REGISTER_FN; |
2510 |
__jit_debug_descriptor.relevant_entry = &one_entry; |
2511 |
__jit_debug_descriptor.first_entry = &one_entry; |
2512 |
__jit_debug_register_code(); |
2513 |
} |
2514 |
#else
|
2515 |
/* No support for the feature. Provide the entry point expected by exec.c,
|
2516 |
and implement the internal function we declared earlier. */
|
2517 |
|
2518 |
static void tcg_register_jit_int(void *buf, size_t size, |
2519 |
void *debug_frame, size_t debug_frame_size)
|
2520 |
{ |
2521 |
} |
2522 |
|
2523 |
void tcg_register_jit(void *buf, size_t buf_size) |
2524 |
{ |
2525 |
} |
2526 |
#endif /* ELF_HOST_MACHINE */ |