notaz.gp2x.de / pcsx_rearmed.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
git subrepo pull --force deps/lightrec
[pcsx_rearmed.git] / deps / lightrec / constprop.c
CommitLineData
9259d748
PC		 1// SPDX-License-Identifier: LGPL-2.1-or-later
2/*
3 * Copyright (C) 2022 Paul Cercueil <paul@crapouillou.net>
4 */
5
6#include "constprop.h"
7#include "disassembler.h"
8#include "lightrec-private.h"
9
10#include <stdbool.h>
11#include <string.h>
12
13static u32 get_min_value(const struct constprop_data *d)
14{
15 /* Min value: all sign bits to 1, all unknown bits but MSB to 0 */
16 return (d->value & d->known) | d->sign | (~d->known & BIT(31));
17}
18
19static u32 get_max_value(const struct constprop_data *d)
20{
21 /* Max value: all sign bits to 0, all unknown bits to 1 */
22 return ((d->value & d->known) | ~d->known) & ~d->sign;
23}
24
25static u32 lightrec_same_sign(const struct constprop_data *d1,
26 const struct constprop_data *d2)
27{
28 u32 min1, min2, max1, max2, a, b, c, d;
29
30 min1 = get_min_value(d1);
31 max1 = get_max_value(d1);
32 min2 = get_min_value(d2);
33 max2 = get_max_value(d2);
34
35 a = min1 + min2;
36 b = min1 + max2;
37 c = max1 + min2;
38 d = max1 + max2;
39
40 return ((a & b & c & d) | (~a & ~b & ~c & ~d)) & BIT(31);
41}
42
43static u32 lightrec_get_sign_mask(const struct constprop_data *d)
44{
45 u32 imm;
46
47 if (d->sign)
48 return d->sign;
49
50 imm = (d->value & BIT(31)) ? d->value : ~d->value;
51 imm = ~(imm & d->known);
52 if (imm)
53 imm = 32 - clz32(imm);
54
55 return imm < 32 ? GENMASK(31, imm) : 0;
56}
57
58static void lightrec_propagate_addi(u32 rs, u32 rd,
59 const struct constprop_data *d,
60 struct constprop_data *v)
61{
62 u32 end, bit, sum, min, mask, imm, value;
63 struct constprop_data result = {
64 .value = v[rd].value,
65 .known = v[rd].known,
66 .sign = v[rd].sign,
67 };
68 bool carry = false;
69
70 /* clear unknown bits to ease processing */
71 v[rs].value &= v[rs].known;
72 value = d->value & d->known;
73
74 mask = ~(lightrec_get_sign_mask(d) & lightrec_get_sign_mask(&v[rs]));
75 end = mask ? 32 - clz32(mask) : 0;
76
77 for (bit = 0; bit < 32; bit++) {
78 if (v[rs].known & d->known & BIT(bit)) {
79 /* the bits are known - compute the resulting bit and
80 * the carry */
81 sum = ((u32)carry << bit) + (v[rs].value & BIT(bit))
82 + (value & BIT(bit));
83
84 if (sum & BIT(bit))
85 result.value |= BIT(bit);
86 else
87 result.value &= ~BIT(bit);
88
89 result.known |= BIT(bit);
90 result.sign &= ~BIT(bit);
91 carry = sum & BIT(bit + 1);
92 continue;
93 }
94
95 if (bit >= end) {
96 /* We're past the last significant bits of the values
97 * (extra sign bits excepted).
98 * The destination register will be sign-extended
99 * starting from here (if no carry) or from the next
100 * bit (if carry).
101 * If the source registers are not sign-extended and we
102 * have no carry, the algorithm is done here. */
103
104 if ((v[rs].sign | d->sign) & BIT(bit)) {
105 mask = GENMASK(31, bit);
106
107 if (lightrec_same_sign(&v[rs], d)) {
108 /* Theorical minimum and maximum values
109 * have the same sign; therefore the
110 * sign bits are known. */
111 min = get_min_value(&v[rs])
112 + get_min_value(d);
113 result.value = (min & mask)
114 | (result.value & ~mask);
115 result.known |= mask << carry;
116 result.sign = 0;
117 } else {
118 /* min/max have different signs. */
119 result.sign = mask << 1;
120 result.known &= ~mask;
121 }
122 break;
123 } else if (!carry) {
124 /* Past end bit, no carry; we're done here. */
125 break;
126 }
127 }
128
129 result.known &= ~BIT(bit);
130 result.sign &= ~BIT(bit);
131
132 /* Found an unknown bit in one of the registers.
133 * If the carry and the bit in the other register are both zero,
134 * we can continue the algorithm. */
135 if (!carry && (((d->known & ~value)
136 | (v[rs].known & ~v[rs].value)) & BIT(bit)))
137 continue;
138
139 /* We have an unknown bit in one of the source registers, and we
140 * may generate a carry: there's nothing to do. Everything from
141 * this bit till the next known 0 bit or sign bit will be marked
142 * as unknown. The algorithm can then restart at the following
143 * bit. */
144
145 imm = (v[rs].known & d->known & ~v[rs].value & ~value)
146 | v[rs].sign | d->sign;
147
148 imm &= GENMASK(31, bit);
149 imm = imm ? ctz32(imm) : 31;
150 mask = GENMASK(imm, bit);
151 result.known &= ~mask;
152 result.sign &= ~mask;
153
154 bit = imm;
155 carry = false;
156 }
157
158 v[rd] = result;
159}
160
161static void lightrec_propagate_sub(u32 rs, u32 rt, u32 rd,
162 struct constprop_data *v)
163{
164 struct constprop_data d = {
165 .value = ~v[rt].value,
166 .known = v[rt].known,
167 .sign = v[rt].sign,
168 };
169 u32 imm, mask, bit;
170
171 /* Negate the known Rt value, then propagate as a regular ADD. */
172
173 for (bit = 0; bit < 32; bit++) {
174 if (!(d.known & BIT(bit))) {
175 /* Unknown bit - mark bits unknown up to the next known 0 */
176
177 imm = (d.known & ~d.value) | d.sign;
178 imm &= GENMASK(31, bit);
179 imm = imm ? ctz32(imm) : 31;
180 mask = GENMASK(imm, bit);
181 d.known &= ~mask;
182 d.sign &= ~mask;
183 break;
184 }
185
186 if (!(d.value & BIT(bit))) {
187 /* Bit is 0: we can set our carry, and the algorithm is done. */
188 d.value |= BIT(bit);
189 break;
190 }
191
192 /* Bit is 1 - set to 0 and continue algorithm */
193 d.value &= ~BIT(bit);
194 }
195
196 lightrec_propagate_addi(rs, rd, &d, v);
197}
198
199static void lightrec_propagate_slt(u32 rs, u32 rd, bool is_signed,
200 const struct constprop_data *d,
201 struct constprop_data *v)
202{
203 unsigned int bit;
204
205 if (is_signed && (v[rs].known & d->known
206 & (v[rs].value ^ d->value) & BIT(31))) {
207 /* If doing a signed comparison and the two bits 31 are known
208 * to be opposite, we can deduce the value. */
209 v[rd].value = v[rs].value >> 31;
210 v[rd].known = 0xffffffff;
211 v[rd].sign = 0;
212 return;
213 }
214
215 for (bit = 32; bit > 0; bit--) {
216 if (!(v[rs].known & d->known & BIT(bit - 1))) {
217 /* One bit is unknown and we cannot figure out which
218 * value is smaller. We still know that the upper 31
219 * bits are zero. */
220 v[rd].value = 0;
221 v[rd].known = 0xfffffffe;
222 v[rd].sign = 0;
223 break;
224 }
225
226 /* The two bits are equal - continue to the next bit. */
227 if (~(v[rs].value ^ d->value) & BIT(bit - 1))
228 continue;
229
230 /* The two bits aren't equal; we can therefore deduce which
231 * value is smaller. */
232 v[rd].value = !(v[rs].value & BIT(bit - 1));
233 v[rd].known = 0xffffffff;
234 v[rd].sign = 0;
235 break;
236 }
237
238 if (bit == 0) {
239 /* rs == rt and all bits are known */
240 v[rd].value = 0;
241 v[rd].known = 0xffffffff;
242 v[rd].sign = 0;
243 }
244}
245
cb72ea13 246void lightrec_consts_propagate(const struct block *block,
9259d748
PC		 247 unsigned int idx,
248 struct constprop_data *v)
249{
cb72ea13 250 const struct opcode *list = block->opcode_list;
9259d748 251 union code c;
cb72ea13 252 u32 imm, flags;
9259d748
PC		 253
254 if (idx == 0)
255 return;
256
257 /* Register $zero is always, well, zero */
258 v[0].value = 0;
259 v[0].sign = 0;
260 v[0].known = 0xffffffff;
261
262 if (op_flag_sync(list[idx].flags)) {
263 memset(&v[1], 0, sizeof(*v) * 31);
264 return;
265 }
266
cb72ea13
PC		 267 flags = list[idx - 1].flags;
268
269 if (idx > 1 && !op_flag_sync(flags)) {
270 if (op_flag_no_ds(flags))
271 c = list[idx - 1].c;
272 else
273 c = list[idx - 2].c;
9259d748
PC		 274
275 switch (c.i.op) {
276 case OP_BNE:
277 /* After a BNE $zero + delay slot, we know that the
278 * branch wasn't taken, and therefore the other register
279 * is zero. */
280 if (c.i.rs == 0) {
281 v[c.i.rt].value = 0;
282 v[c.i.rt].sign = 0;
283 v[c.i.rt].known = 0xffffffff;
284 } else if (c.i.rt == 0) {
285 v[c.i.rs].value = 0;
286 v[c.i.rs].sign = 0;
287 v[c.i.rs].known = 0xffffffff;
288 }
289 break;
290 case OP_BLEZ:
291 v[c.i.rs].value &= ~BIT(31);
292 v[c.i.rs].known |= BIT(31);
293 fallthrough;
294 case OP_BEQ:
295 /* TODO: handle non-zero? */
296 break;
297 case OP_REGIMM:
298 switch (c.r.rt) {
299 case OP_REGIMM_BLTZ:
300 case OP_REGIMM_BLTZAL:
301 v[c.i.rs].value &= ~BIT(31);
302 v[c.i.rs].known |= BIT(31);
303 break;
304 case OP_REGIMM_BGEZ:
305 case OP_REGIMM_BGEZAL:
306 v[c.i.rs].value |= BIT(31);
307 v[c.i.rs].known |= BIT(31);
308 /* TODO: handle non-zero? */
309 break;
310 }
311 break;
312 default:
313 break;
314 }
315 }
316
317 c = list[idx - 1].c;
318
319 switch (c.i.op) {
320 case OP_SPECIAL:
321 switch (c.r.op) {
322 case OP_SPECIAL_SLL:
323 v[c.r.rd].value = v[c.r.rt].value << c.r.imm;
324 v[c.r.rd].known = (v[c.r.rt].known << c.r.imm)
325 | (BIT(c.r.imm) - 1);
326 v[c.r.rd].sign = v[c.r.rt].sign << c.r.imm;
327 break;
328
329 case OP_SPECIAL_SRL:
330 v[c.r.rd].value = v[c.r.rt].value >> c.r.imm;
331 v[c.r.rd].known = (v[c.r.rt].known >> c.r.imm)
332 | (BIT(c.r.imm) - 1 << 32 - c.r.imm);
333 v[c.r.rd].sign = c.r.imm ? 0 : v[c.r.rt].sign;
334 break;
335
336 case OP_SPECIAL_SRA:
337 v[c.r.rd].value = (s32)v[c.r.rt].value >> c.r.imm;
338 v[c.r.rd].known = (s32)v[c.r.rt].known >> c.r.imm;
339 v[c.r.rd].sign = (s32)v[c.r.rt].sign >> c.r.imm;
340 break;
341
342 case OP_SPECIAL_SLLV:
343 if ((v[c.r.rs].known & 0x1f) == 0x1f) {
344 imm = v[c.r.rs].value & 0x1f;
345 v[c.r.rd].value = v[c.r.rt].value << imm;
346 v[c.r.rd].known = (v[c.r.rt].known << imm)
347 | (BIT(imm) - 1);
348 v[c.r.rd].sign = v[c.r.rt].sign << imm;
349 } else {
350 v[c.r.rd].known = 0;
351 v[c.r.rd].sign = 0;
352 }
353 break;
354
355 case OP_SPECIAL_SRLV:
356 if ((v[c.r.rs].known & 0x1f) == 0x1f) {
357 imm = v[c.r.rs].value & 0x1f;
358 v[c.r.rd].value = v[c.r.rt].value >> imm;
359 v[c.r.rd].known = (v[c.r.rt].known >> imm)
360 | (BIT(imm) - 1 << 32 - imm);
361 if (imm)
362 v[c.r.rd].sign = 0;
363 } else {
364 v[c.r.rd].known = 0;
365 v[c.r.rd].sign = 0;
366 }
367 break;
368
369 case OP_SPECIAL_SRAV:
370 if ((v[c.r.rs].known & 0x1f) == 0x1f) {
371 imm = v[c.r.rs].value & 0x1f;
372 v[c.r.rd].value = (s32)v[c.r.rt].value >> imm;
373 v[c.r.rd].known = (s32)v[c.r.rt].known >> imm;
374 v[c.r.rd].sign = (s32)v[c.r.rt].sign >> imm;
375 } else {
376 v[c.r.rd].known = 0;
377 v[c.r.rd].sign = 0;
378 }
379 break;
380
381 case OP_SPECIAL_ADD:
382 case OP_SPECIAL_ADDU:
383 if (is_known_zero(v, c.r.rs))
384 v[c.r.rd] = v[c.r.rt];
385 else if (is_known_zero(v, c.r.rt))
386 v[c.r.rd] = v[c.r.rs];
387 else
388 lightrec_propagate_addi(c.r.rs, c.r.rd, &v[c.r.rt], v);
389 break;
390
391 case OP_SPECIAL_SUB:
392 case OP_SPECIAL_SUBU:
393 if (c.r.rs == c.r.rt) {
394 v[c.r.rd].value = 0;
395 v[c.r.rd].known = 0xffffffff;
396 v[c.r.rd].sign = 0;
397 } else {
398 lightrec_propagate_sub(c.r.rs, c.r.rt, c.r.rd, v);
399 }
400 break;
401
402 case OP_SPECIAL_AND:
403 v[c.r.rd].known = (v[c.r.rt].known & v[c.r.rs].known)
404 | (~v[c.r.rt].value & v[c.r.rt].known)
405 | (~v[c.r.rs].value & v[c.r.rs].known);
406 v[c.r.rd].value = v[c.r.rt].value & v[c.r.rs].value & v[c.r.rd].known;
407 v[c.r.rd].sign = v[c.r.rt].sign & v[c.r.rs].sign;
408 break;
409
410 case OP_SPECIAL_OR:
411 v[c.r.rd].known = (v[c.r.rt].known & v[c.r.rs].known)
412 | (v[c.r.rt].value & v[c.r.rt].known)
413 | (v[c.r.rs].value & v[c.r.rs].known);
414 v[c.r.rd].value = (v[c.r.rt].value | v[c.r.rs].value) & v[c.r.rd].known;
415 v[c.r.rd].sign = v[c.r.rt].sign & v[c.r.rs].sign;
416 break;
417
418 case OP_SPECIAL_XOR:
419 v[c.r.rd].value = v[c.r.rt].value ^ v[c.r.rs].value;
420 v[c.r.rd].known = v[c.r.rt].known & v[c.r.rs].known;
421 v[c.r.rd].sign = v[c.r.rt].sign & v[c.r.rs].sign;
422 break;
423
424 case OP_SPECIAL_NOR:
425 v[c.r.rd].known = (v[c.r.rt].known & v[c.r.rs].known)
426 | (v[c.r.rt].value & v[c.r.rt].known)
427 | (v[c.r.rs].value & v[c.r.rs].known);
428 v[c.r.rd].value = ~(v[c.r.rt].value | v[c.r.rs].value) & v[c.r.rd].known;
429 v[c.r.rd].sign = v[c.r.rt].sign & v[c.r.rs].sign;
430 break;
431
432 case OP_SPECIAL_SLT:
433 case OP_SPECIAL_SLTU:
434 lightrec_propagate_slt(c.r.rs, c.r.rd,
435 c.r.op == OP_SPECIAL_SLT,
436 &v[c.r.rt], v);
437 break;
438
439 case OP_SPECIAL_MULT:
440 case OP_SPECIAL_MULTU:
441 case OP_SPECIAL_DIV:
442 case OP_SPECIAL_DIVU:
443 if (OPT_FLAG_MULT_DIV && c.r.rd) {
444 v[c.r.rd].known = 0;
445 v[c.r.rd].sign = 0;
446 }
447 if (OPT_FLAG_MULT_DIV && c.r.imm) {
448 v[c.r.imm].known = 0;
449 v[c.r.imm].sign = 0;
450 }
451 break;
452
453 case OP_SPECIAL_MFLO:
454 case OP_SPECIAL_MFHI:
455 v[c.r.rd].known = 0;
456 v[c.r.rd].sign = 0;
457 break;
cb72ea13
PC		 458
459 case OP_SPECIAL_JALR:
460 v[c.r.rd].known = 0xffffffff;
461 v[c.r.rd].sign = 0;
462 v[c.r.rd].value = block->pc + (idx + 2 << 2);
463 break;
464
9259d748
PC		 465 default:
466 break;
467 }
468 break;
469
470 case OP_META_MULT2:
471 case OP_META_MULTU2:
472 if (OPT_FLAG_MULT_DIV && c.r.rd) {
473 if (c.r.op < 32) {
474 v[c.r.rd].value = v[c.r.rs].value << c.r.op;
475 v[c.r.rd].known = (v[c.r.rs].known << c.r.op)
476 | (BIT(c.r.op) - 1);
477 v[c.r.rd].sign = v[c.r.rs].sign << c.r.op;
478 } else {
479 v[c.r.rd].value = 0;
480 v[c.r.rd].known = 0xffffffff;
481 v[c.r.rd].sign = 0;
482 }
483 }
484
485 if (OPT_FLAG_MULT_DIV && c.r.imm) {
486 if (c.r.op >= 32) {
487 v[c.r.imm].value = v[c.r.rs].value << c.r.op - 32;
488 v[c.r.imm].known = (v[c.r.rs].known << c.r.op - 32)
489 | (BIT(c.r.op - 32) - 1);
490 v[c.r.imm].sign = v[c.r.rs].sign << c.r.op - 32;
491 } else if (c.i.op == OP_META_MULT2) {
492 v[c.r.imm].value = (s32)v[c.r.rs].value >> 32 - c.r.op;
493 v[c.r.imm].known = (s32)v[c.r.rs].known >> 32 - c.r.op;
494 v[c.r.imm].sign = (s32)v[c.r.rs].sign >> 32 - c.r.op;
495 } else {
496 v[c.r.imm].value = v[c.r.rs].value >> 32 - c.r.op;
497 v[c.r.imm].known = v[c.r.rs].known >> 32 - c.r.op;
498 v[c.r.imm].sign = v[c.r.rs].sign >> 32 - c.r.op;
499 }
500 }
501 break;
502
503 case OP_REGIMM:
504 break;
505
506 case OP_ADDI:
507 case OP_ADDIU:
508 if (c.i.imm) {
509 struct constprop_data d = {
510 .value = (s32)(s16)c.i.imm,
511 .known = 0xffffffff,
512 .sign = 0,
513 };
514
515 lightrec_propagate_addi(c.i.rs, c.i.rt, &d, v);
516 } else {
517 /* immediate is zero - that's just a register copy. */
518 v[c.i.rt] = v[c.i.rs];
519 }
520 break;
521
522 case OP_SLTI:
523 case OP_SLTIU:
524 {
525 struct constprop_data d = {
526 .value = (s32)(s16)c.i.imm,
527 .known = 0xffffffff,
528 .sign = 0,
529 };
530
531 lightrec_propagate_slt(c.i.rs, c.i.rt,
532 c.i.op == OP_SLTI, &d, v);
533 }
534 break;
535
536 case OP_ANDI:
537 v[c.i.rt].value = v[c.i.rs].value & c.i.imm;
538 v[c.i.rt].known = v[c.i.rs].known | ~c.i.imm;
539 v[c.i.rt].sign = 0;
540 break;
541
542 case OP_ORI:
543 v[c.i.rt].value = v[c.i.rs].value | c.i.imm;
544 v[c.i.rt].known = v[c.i.rs].known | c.i.imm;
545 v[c.i.rt].sign = (v[c.i.rs].sign & 0xffff) ? 0xffff0000 : v[c.i.rs].sign;
546 break;
547
548 case OP_XORI:
549 v[c.i.rt].value = v[c.i.rs].value ^ c.i.imm;
550 v[c.i.rt].known = v[c.i.rs].known;
551 v[c.i.rt].sign = (v[c.i.rs].sign & 0xffff) ? 0xffff0000 : v[c.i.rs].sign;
552 break;
553
554 case OP_LUI:
555 v[c.i.rt].value = c.i.imm << 16;
556 v[c.i.rt].known = 0xffffffff;
557 v[c.i.rt].sign = 0;
558 break;
559
560 case OP_CP0:
561 switch (c.r.rs) {
562 case OP_CP0_MFC0:
563 case OP_CP0_CFC0:
564 v[c.r.rt].known = 0;
565 v[c.r.rt].sign = 0;
566 break;
567 default:
568 break;
569 }
570 break;
571
572 case OP_CP2:
573 if (c.r.op == OP_CP2_BASIC) {
574 switch (c.r.rs) {
575 case OP_CP2_BASIC_MFC2:
576 switch (c.r.rd) {
577 case 1:
578 case 3:
579 case 5:
580 case 8:
581 case 9:
582 case 10:
583 case 11:
584 /* Signed 16-bit */
585 v[c.r.rt].known = 0;
586 v[c.r.rt].sign = 0xffff8000;
587 break;
588 case 7:
589 case 16:
590 case 17:
591 case 18:
592 case 19:
593 /* Unsigned 16-bit */
594 v[c.r.rt].value = 0;
595 v[c.r.rt].known = 0xffff0000;
596 v[c.r.rt].sign = 0;
597 break;
598 default:
599 /* 32-bit */
600 v[c.r.rt].known = 0;
601 v[c.r.rt].sign = 0;
602 break;
603 }
604 break;
605 case OP_CP2_BASIC_CFC2:
606 switch (c.r.rd) {
607 case 4:
608 case 12:
609 case 20:
610 case 26:
611 case 27:
612 case 29:
613 case 30:
614 /* Signed 16-bit */
615 v[c.r.rt].known = 0;
616 v[c.r.rt].sign = 0xffff8000;
617 break;
618 default:
619 /* 32-bit */
620 v[c.r.rt].known = 0;
621 v[c.r.rt].sign = 0;
622 break;
623 }
624 break;
625 }
626 }
627 break;
628 case OP_LB:
629 v[c.i.rt].known = 0;
630 v[c.i.rt].sign = 0xffffff80;
631 break;
632 case OP_LH:
633 v[c.i.rt].known = 0;
634 v[c.i.rt].sign = 0xffff8000;
635 break;
636 case OP_LBU:
637 v[c.i.rt].value = 0;
638 v[c.i.rt].known = 0xffffff00;
639 v[c.i.rt].sign = 0;
640 break;
641 case OP_LHU:
642 v[c.i.rt].value = 0;
643 v[c.i.rt].known = 0xffff0000;
644 v[c.i.rt].sign = 0;
645 break;
646 case OP_LWL:
647 case OP_LWR:
648 /* LWL/LWR don't write the full register if the address is
649 * unaligned, so we only need to know the low 2 bits */
650 if (v[c.i.rs].known & 0x3) {
651 imm = (v[c.i.rs].value & 0x3) * 8;
652
653 if (c.i.op == OP_LWL) {
654 imm = BIT(24 - imm) - 1;
655 v[c.i.rt].sign &= ~imm;
656 } else {
657 imm = imm ? GENMASK(31, 32 - imm) : 0;
658 v[c.i.rt].sign = 0;
659 }
cb72ea13 660 v[c.i.rt].known &= imm;
9259d748
PC		 661 break;
662 }
663 fallthrough;
664 case OP_LW:
665 v[c.i.rt].known = 0;
666 v[c.i.rt].sign = 0;
667 break;
cb72ea13
PC		 668 case OP_META:
669 switch (c.m.op) {
670 case OP_META_MOV:
671 v[c.m.rd] = v[c.m.rs];
672 break;
9259d748 673
cb72ea13
PC		 674 case OP_META_EXTC:
675 v[c.m.rd].value = (s32)(s8)v[c.m.rs].value;
676 if (v[c.m.rs].known & BIT(7)) {
677 v[c.m.rd].known = v[c.m.rs].known | 0xffffff00;
678 v[c.m.rd].sign = 0;
679 } else {
680 v[c.m.rd].known = v[c.m.rs].known & 0x7f;
681 v[c.m.rd].sign = 0xffffff80;
682 }
683 break;
684
685 case OP_META_EXTS:
686 v[c.m.rd].value = (s32)(s16)v[c.m.rs].value;
687 if (v[c.m.rs].known & BIT(15)) {
688 v[c.m.rd].known = v[c.m.rs].known | 0xffff0000;
689 v[c.m.rd].sign = 0;
690 } else {
691 v[c.m.rd].known = v[c.m.rs].known & 0x7fff;
692 v[c.m.rd].sign = 0xffff8000;
693 }
694 break;
695
696 case OP_META_COM:
697 v[c.m.rd].known = v[c.m.rs].known;
698 v[c.m.rd].value = ~v[c.m.rs].value;
699 v[c.m.rd].sign = v[c.m.rs].sign;
700 break;
701 default:
702 break;
9259d748
PC		 703 }
704 break;
cb72ea13
PC		 705 case OP_JAL:
706 v[31].known = 0xffffffff;
707 v[31].sign = 0;
708 v[31].value = block->pc + (idx + 2 << 2);
709 break;
9259d748
PC		 710
711 default:
712 break;
713 }
714
715 /* Reset register 0 which may have been used as a target */
716 v[0].value = 0;
717 v[0].sign = 0;
718 v[0].known = 0xffffffff;
719}
720
721enum psx_map
722lightrec_get_constprop_map(const struct lightrec_state *state,
723 const struct constprop_data *v, u8 reg, s16 imm)
724{
725 const struct lightrec_mem_map *map;
726 unsigned int i;
727 u32 min, max;
728
729 min = get_min_value(&v[reg]) + imm;
730 max = get_max_value(&v[reg]) + imm;
731
732 /* Handle the case where max + imm overflows */
733 if ((min & 0xe0000000) != (max & 0xe0000000))
734 return PSX_MAP_UNKNOWN;
735
736 pr_debug("Min: 0x%08x max: 0x%08x Known: 0x%08x Sign: 0x%08x\n",
737 min, max, v[reg].known, v[reg].sign);
738
739 min = kunseg(min);
740 max = kunseg(max);
741
742 for (i = 0; i < state->nb_maps; i++) {
743 map = &state->maps[i];
744
745 if (min >= map->pc && min < map->pc + map->length
746 && max >= map->pc && max < map->pc + map->length)
747 return (enum psx_map) i;
748 }
749
750 return PSX_MAP_UNKNOWN;
751}
ARM optimized PCSX fork