]> git.proxmox.com Git - qemu.git/blob - tcg/optimize.c
tcg/optimize: simplify or/xor r, a, 0 cases
[qemu.git] / tcg / optimize.c
1 /*
2 * Optimizations for Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2010 Samsung Electronics.
5 * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23 * THE SOFTWARE.
24 */
25
26 #include "config.h"
27
28 #include <stdlib.h>
29 #include <stdio.h>
30
31 #include "qemu-common.h"
32 #include "tcg-op.h"
33
34 #define CASE_OP_32_64(x) \
35 glue(glue(case INDEX_op_, x), _i32): \
36 glue(glue(case INDEX_op_, x), _i64)
37
38 typedef enum {
39 TCG_TEMP_UNDEF = 0,
40 TCG_TEMP_CONST,
41 TCG_TEMP_COPY,
42 TCG_TEMP_HAS_COPY,
43 TCG_TEMP_ANY
44 } tcg_temp_state;
45
46 struct tcg_temp_info {
47 tcg_temp_state state;
48 uint16_t prev_copy;
49 uint16_t next_copy;
50 tcg_target_ulong val;
51 };
52
53 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
54
55 /* Reset TEMP's state to TCG_TEMP_ANY. If TEMP was a representative of some
56 class of equivalent temp's, a new representative should be chosen in this
57 class. */
58 static void reset_temp(TCGArg temp, int nb_temps, int nb_globals)
59 {
60 int i;
61 TCGArg new_base = (TCGArg)-1;
62 if (temps[temp].state == TCG_TEMP_HAS_COPY) {
63 for (i = temps[temp].next_copy; i != temp; i = temps[i].next_copy) {
64 if (i >= nb_globals) {
65 temps[i].state = TCG_TEMP_HAS_COPY;
66 new_base = i;
67 break;
68 }
69 }
70 for (i = temps[temp].next_copy; i != temp; i = temps[i].next_copy) {
71 if (new_base == (TCGArg)-1) {
72 temps[i].state = TCG_TEMP_ANY;
73 } else {
74 temps[i].val = new_base;
75 }
76 }
77 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
78 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
79 } else if (temps[temp].state == TCG_TEMP_COPY) {
80 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
81 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
82 new_base = temps[temp].val;
83 }
84 temps[temp].state = TCG_TEMP_ANY;
85 if (new_base != (TCGArg)-1 && temps[new_base].next_copy == new_base) {
86 temps[new_base].state = TCG_TEMP_ANY;
87 }
88 }
89
90 static int op_bits(TCGOpcode op)
91 {
92 const TCGOpDef *def = &tcg_op_defs[op];
93 return def->flags & TCG_OPF_64BIT ? 64 : 32;
94 }
95
96 static TCGOpcode op_to_movi(TCGOpcode op)
97 {
98 switch (op_bits(op)) {
99 case 32:
100 return INDEX_op_movi_i32;
101 case 64:
102 return INDEX_op_movi_i64;
103 default:
104 fprintf(stderr, "op_to_movi: unexpected return value of "
105 "function op_bits.\n");
106 tcg_abort();
107 }
108 }
109
110 static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args, TCGArg dst,
111 TCGArg src, int nb_temps, int nb_globals)
112 {
113 reset_temp(dst, nb_temps, nb_globals);
114 assert(temps[src].state != TCG_TEMP_COPY);
115 /* Don't try to copy if one of temps is a global or either one
116 is local and another is register */
117 if (src >= nb_globals && dst >= nb_globals &&
118 tcg_arg_is_local(s, src) == tcg_arg_is_local(s, dst)) {
119 assert(temps[src].state != TCG_TEMP_CONST);
120 if (temps[src].state != TCG_TEMP_HAS_COPY) {
121 temps[src].state = TCG_TEMP_HAS_COPY;
122 temps[src].next_copy = src;
123 temps[src].prev_copy = src;
124 }
125 temps[dst].state = TCG_TEMP_COPY;
126 temps[dst].val = src;
127 temps[dst].next_copy = temps[src].next_copy;
128 temps[dst].prev_copy = src;
129 temps[temps[dst].next_copy].prev_copy = dst;
130 temps[src].next_copy = dst;
131 }
132 gen_args[0] = dst;
133 gen_args[1] = src;
134 }
135
136 static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val,
137 int nb_temps, int nb_globals)
138 {
139 reset_temp(dst, nb_temps, nb_globals);
140 temps[dst].state = TCG_TEMP_CONST;
141 temps[dst].val = val;
142 gen_args[0] = dst;
143 gen_args[1] = val;
144 }
145
146 static TCGOpcode op_to_mov(TCGOpcode op)
147 {
148 switch (op_bits(op)) {
149 case 32:
150 return INDEX_op_mov_i32;
151 case 64:
152 return INDEX_op_mov_i64;
153 default:
154 fprintf(stderr, "op_to_mov: unexpected return value of "
155 "function op_bits.\n");
156 tcg_abort();
157 }
158 }
159
160 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
161 {
162 switch (op) {
163 CASE_OP_32_64(add):
164 return x + y;
165
166 CASE_OP_32_64(sub):
167 return x - y;
168
169 CASE_OP_32_64(mul):
170 return x * y;
171
172 CASE_OP_32_64(and):
173 return x & y;
174
175 CASE_OP_32_64(or):
176 return x | y;
177
178 CASE_OP_32_64(xor):
179 return x ^ y;
180
181 case INDEX_op_shl_i32:
182 return (uint32_t)x << (uint32_t)y;
183
184 case INDEX_op_shl_i64:
185 return (uint64_t)x << (uint64_t)y;
186
187 case INDEX_op_shr_i32:
188 return (uint32_t)x >> (uint32_t)y;
189
190 case INDEX_op_shr_i64:
191 return (uint64_t)x >> (uint64_t)y;
192
193 case INDEX_op_sar_i32:
194 return (int32_t)x >> (int32_t)y;
195
196 case INDEX_op_sar_i64:
197 return (int64_t)x >> (int64_t)y;
198
199 case INDEX_op_rotr_i32:
200 x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
201 return x;
202
203 case INDEX_op_rotr_i64:
204 x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
205 return x;
206
207 case INDEX_op_rotl_i32:
208 x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
209 return x;
210
211 case INDEX_op_rotl_i64:
212 x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
213 return x;
214
215 CASE_OP_32_64(not):
216 return ~x;
217
218 CASE_OP_32_64(neg):
219 return -x;
220
221 CASE_OP_32_64(andc):
222 return x & ~y;
223
224 CASE_OP_32_64(orc):
225 return x | ~y;
226
227 CASE_OP_32_64(eqv):
228 return ~(x ^ y);
229
230 CASE_OP_32_64(nand):
231 return ~(x & y);
232
233 CASE_OP_32_64(nor):
234 return ~(x | y);
235
236 CASE_OP_32_64(ext8s):
237 return (int8_t)x;
238
239 CASE_OP_32_64(ext16s):
240 return (int16_t)x;
241
242 CASE_OP_32_64(ext8u):
243 return (uint8_t)x;
244
245 CASE_OP_32_64(ext16u):
246 return (uint16_t)x;
247
248 case INDEX_op_ext32s_i64:
249 return (int32_t)x;
250
251 case INDEX_op_ext32u_i64:
252 return (uint32_t)x;
253
254 default:
255 fprintf(stderr,
256 "Unrecognized operation %d in do_constant_folding.\n", op);
257 tcg_abort();
258 }
259 }
260
261 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
262 {
263 TCGArg res = do_constant_folding_2(op, x, y);
264 if (op_bits(op) == 32) {
265 res &= 0xffffffff;
266 }
267 return res;
268 }
269
270 /* Propagate constants and copies, fold constant expressions. */
271 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
272 TCGArg *args, TCGOpDef *tcg_op_defs)
273 {
274 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
275 TCGOpcode op;
276 const TCGOpDef *def;
277 TCGArg *gen_args;
278 TCGArg tmp;
279 /* Array VALS has an element for each temp.
280 If this temp holds a constant then its value is kept in VALS' element.
281 If this temp is a copy of other ones then this equivalence class'
282 representative is kept in VALS' element.
283 If this temp is neither copy nor constant then corresponding VALS'
284 element is unused. */
285
286 nb_temps = s->nb_temps;
287 nb_globals = s->nb_globals;
288 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
289
290 nb_ops = tcg_opc_ptr - gen_opc_buf;
291 gen_args = args;
292 for (op_index = 0; op_index < nb_ops; op_index++) {
293 op = gen_opc_buf[op_index];
294 def = &tcg_op_defs[op];
295 /* Do copy propagation */
296 if (!(def->flags & (TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS))) {
297 assert(op != INDEX_op_call);
298 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
299 if (temps[args[i]].state == TCG_TEMP_COPY) {
300 args[i] = temps[args[i]].val;
301 }
302 }
303 }
304
305 /* For commutative operations make constant second argument */
306 switch (op) {
307 CASE_OP_32_64(add):
308 CASE_OP_32_64(mul):
309 CASE_OP_32_64(and):
310 CASE_OP_32_64(or):
311 CASE_OP_32_64(xor):
312 CASE_OP_32_64(eqv):
313 CASE_OP_32_64(nand):
314 CASE_OP_32_64(nor):
315 if (temps[args[1]].state == TCG_TEMP_CONST) {
316 tmp = args[1];
317 args[1] = args[2];
318 args[2] = tmp;
319 }
320 break;
321 default:
322 break;
323 }
324
325 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
326 switch (op) {
327 CASE_OP_32_64(add):
328 CASE_OP_32_64(sub):
329 CASE_OP_32_64(shl):
330 CASE_OP_32_64(shr):
331 CASE_OP_32_64(sar):
332 CASE_OP_32_64(rotl):
333 CASE_OP_32_64(rotr):
334 CASE_OP_32_64(or):
335 CASE_OP_32_64(xor):
336 if (temps[args[1]].state == TCG_TEMP_CONST) {
337 /* Proceed with possible constant folding. */
338 break;
339 }
340 if (temps[args[2]].state == TCG_TEMP_CONST
341 && temps[args[2]].val == 0) {
342 if ((temps[args[0]].state == TCG_TEMP_COPY
343 && temps[args[0]].val == args[1])
344 || args[0] == args[1]) {
345 args += 3;
346 gen_opc_buf[op_index] = INDEX_op_nop;
347 } else {
348 gen_opc_buf[op_index] = op_to_mov(op);
349 tcg_opt_gen_mov(s, gen_args, args[0], args[1],
350 nb_temps, nb_globals);
351 gen_args += 2;
352 args += 3;
353 }
354 continue;
355 }
356 break;
357 default:
358 break;
359 }
360
361 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
362 switch (op) {
363 CASE_OP_32_64(mul):
364 if ((temps[args[2]].state == TCG_TEMP_CONST
365 && temps[args[2]].val == 0)) {
366 gen_opc_buf[op_index] = op_to_movi(op);
367 tcg_opt_gen_movi(gen_args, args[0], 0, nb_temps, nb_globals);
368 args += 3;
369 gen_args += 2;
370 continue;
371 }
372 break;
373 default:
374 break;
375 }
376
377 /* Simplify expression for "op r, a, a => mov r, a" cases */
378 switch (op) {
379 CASE_OP_32_64(or):
380 CASE_OP_32_64(and):
381 if (args[1] == args[2]) {
382 if (args[1] == args[0]) {
383 args += 3;
384 gen_opc_buf[op_index] = INDEX_op_nop;
385 } else {
386 gen_opc_buf[op_index] = op_to_mov(op);
387 tcg_opt_gen_mov(s, gen_args, args[0], args[1], nb_temps,
388 nb_globals);
389 gen_args += 2;
390 args += 3;
391 }
392 continue;
393 }
394 break;
395 default:
396 break;
397 }
398
399 /* Propagate constants through copy operations and do constant
400 folding. Constants will be substituted to arguments by register
401 allocator where needed and possible. Also detect copies. */
402 switch (op) {
403 CASE_OP_32_64(mov):
404 if ((temps[args[1]].state == TCG_TEMP_COPY
405 && temps[args[1]].val == args[0])
406 || args[0] == args[1]) {
407 args += 2;
408 gen_opc_buf[op_index] = INDEX_op_nop;
409 break;
410 }
411 if (temps[args[1]].state != TCG_TEMP_CONST) {
412 tcg_opt_gen_mov(s, gen_args, args[0], args[1],
413 nb_temps, nb_globals);
414 gen_args += 2;
415 args += 2;
416 break;
417 }
418 /* Source argument is constant. Rewrite the operation and
419 let movi case handle it. */
420 op = op_to_movi(op);
421 gen_opc_buf[op_index] = op;
422 args[1] = temps[args[1]].val;
423 /* fallthrough */
424 CASE_OP_32_64(movi):
425 tcg_opt_gen_movi(gen_args, args[0], args[1], nb_temps, nb_globals);
426 gen_args += 2;
427 args += 2;
428 break;
429 CASE_OP_32_64(not):
430 CASE_OP_32_64(neg):
431 CASE_OP_32_64(ext8s):
432 CASE_OP_32_64(ext8u):
433 CASE_OP_32_64(ext16s):
434 CASE_OP_32_64(ext16u):
435 case INDEX_op_ext32s_i64:
436 case INDEX_op_ext32u_i64:
437 if (temps[args[1]].state == TCG_TEMP_CONST) {
438 gen_opc_buf[op_index] = op_to_movi(op);
439 tmp = do_constant_folding(op, temps[args[1]].val, 0);
440 tcg_opt_gen_movi(gen_args, args[0], tmp, nb_temps, nb_globals);
441 gen_args += 2;
442 args += 2;
443 break;
444 } else {
445 reset_temp(args[0], nb_temps, nb_globals);
446 gen_args[0] = args[0];
447 gen_args[1] = args[1];
448 gen_args += 2;
449 args += 2;
450 break;
451 }
452 CASE_OP_32_64(add):
453 CASE_OP_32_64(sub):
454 CASE_OP_32_64(mul):
455 CASE_OP_32_64(or):
456 CASE_OP_32_64(and):
457 CASE_OP_32_64(xor):
458 CASE_OP_32_64(shl):
459 CASE_OP_32_64(shr):
460 CASE_OP_32_64(sar):
461 CASE_OP_32_64(rotl):
462 CASE_OP_32_64(rotr):
463 CASE_OP_32_64(andc):
464 CASE_OP_32_64(orc):
465 CASE_OP_32_64(eqv):
466 CASE_OP_32_64(nand):
467 CASE_OP_32_64(nor):
468 if (temps[args[1]].state == TCG_TEMP_CONST
469 && temps[args[2]].state == TCG_TEMP_CONST) {
470 gen_opc_buf[op_index] = op_to_movi(op);
471 tmp = do_constant_folding(op, temps[args[1]].val,
472 temps[args[2]].val);
473 tcg_opt_gen_movi(gen_args, args[0], tmp, nb_temps, nb_globals);
474 gen_args += 2;
475 args += 3;
476 break;
477 } else {
478 reset_temp(args[0], nb_temps, nb_globals);
479 gen_args[0] = args[0];
480 gen_args[1] = args[1];
481 gen_args[2] = args[2];
482 gen_args += 3;
483 args += 3;
484 break;
485 }
486 case INDEX_op_call:
487 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
488 if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) {
489 for (i = 0; i < nb_globals; i++) {
490 reset_temp(i, nb_temps, nb_globals);
491 }
492 }
493 for (i = 0; i < (args[0] >> 16); i++) {
494 reset_temp(args[i + 1], nb_temps, nb_globals);
495 }
496 i = nb_call_args + 3;
497 while (i) {
498 *gen_args = *args;
499 args++;
500 gen_args++;
501 i--;
502 }
503 break;
504 case INDEX_op_set_label:
505 case INDEX_op_jmp:
506 case INDEX_op_br:
507 CASE_OP_32_64(brcond):
508 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
509 for (i = 0; i < def->nb_args; i++) {
510 *gen_args = *args;
511 args++;
512 gen_args++;
513 }
514 break;
515 default:
516 /* Default case: we do know nothing about operation so no
517 propagation is done. We only trash output args. */
518 for (i = 0; i < def->nb_oargs; i++) {
519 reset_temp(args[i], nb_temps, nb_globals);
520 }
521 for (i = 0; i < def->nb_args; i++) {
522 gen_args[i] = args[i];
523 }
524 args += def->nb_args;
525 gen_args += def->nb_args;
526 break;
527 }
528 }
529
530 return gen_args;
531 }
532
533 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
534 TCGArg *args, TCGOpDef *tcg_op_defs)
535 {
536 TCGArg *res;
537 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
538 return res;
539 }