Update lightrec 20220911 (#687)
[pcsx_rearmed.git] / deps / lightrec / lightrec.c
1 // SPDX-License-Identifier: LGPL-2.1-or-later
2 /*
3  * Copyright (C) 2014-2021 Paul Cercueil <paul@crapouillou.net>
4  */
5
6 #include "blockcache.h"
7 #include "debug.h"
8 #include "disassembler.h"
9 #include "emitter.h"
10 #include "interpreter.h"
11 #include "lightrec-config.h"
12 #include "lightning-wrapper.h"
13 #include "lightrec.h"
14 #include "memmanager.h"
15 #include "reaper.h"
16 #include "recompiler.h"
17 #include "regcache.h"
18 #include "optimizer.h"
19 #include "tlsf/tlsf.h"
20
21 #include <errno.h>
22 #include <inttypes.h>
23 #include <limits.h>
24 #if ENABLE_THREADED_COMPILER
25 #include <stdatomic.h>
26 #endif
27 #include <stdbool.h>
28 #include <stddef.h>
29 #include <string.h>
30
31 #define GENMASK(h, l) \
32         (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
33
34 static struct block * lightrec_precompile_block(struct lightrec_state *state,
35                                                 u32 pc);
36 static bool lightrec_block_is_fully_tagged(const struct block *block);
37
38 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
39 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
40
41 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
42                                 void *host, u32 addr, u8 data)
43 {
44         *(u8 *)host = data;
45
46         if (!state->invalidate_from_dma_only)
47                 lightrec_invalidate(state, addr, 1);
48 }
49
50 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
51                                 void *host, u32 addr, u16 data)
52 {
53         *(u16 *)host = HTOLE16(data);
54
55         if (!state->invalidate_from_dma_only)
56                 lightrec_invalidate(state, addr, 2);
57 }
58
59 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
60                                 void *host, u32 addr, u32 data)
61 {
62         *(u32 *)host = HTOLE32(data);
63
64         if (!state->invalidate_from_dma_only)
65                 lightrec_invalidate(state, addr, 4);
66 }
67
68 static u8 lightrec_default_lb(struct lightrec_state *state,
69                               u32 opcode, void *host, u32 addr)
70 {
71         return *(u8 *)host;
72 }
73
74 static u16 lightrec_default_lh(struct lightrec_state *state,
75                                u32 opcode, void *host, u32 addr)
76 {
77         return LE16TOH(*(u16 *)host);
78 }
79
80 static u32 lightrec_default_lw(struct lightrec_state *state,
81                                u32 opcode, void *host, u32 addr)
82 {
83         return LE32TOH(*(u32 *)host);
84 }
85
86 static const struct lightrec_mem_map_ops lightrec_default_ops = {
87         .sb = lightrec_default_sb,
88         .sh = lightrec_default_sh,
89         .sw = lightrec_default_sw,
90         .lb = lightrec_default_lb,
91         .lh = lightrec_default_lh,
92         .lw = lightrec_default_lw,
93 };
94
95 static void __segfault_cb(struct lightrec_state *state, u32 addr,
96                           const struct block *block)
97 {
98         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
99         pr_err("Segmentation fault in recompiled code: invalid "
100                "load/store at address 0x%08x\n", addr);
101         if (block)
102                 pr_err("Was executing block PC 0x%08x\n", block->pc);
103 }
104
105 static void lightrec_swl(struct lightrec_state *state,
106                          const struct lightrec_mem_map_ops *ops,
107                          u32 opcode, void *host, u32 addr, u32 data)
108 {
109         unsigned int shift = addr & 0x3;
110         unsigned int mask = GENMASK(31, (shift + 1) * 8);
111         u32 old_data;
112
113         /* Align to 32 bits */
114         addr &= ~3;
115         host = (void *)((uintptr_t)host & ~3);
116
117         old_data = ops->lw(state, opcode, host, addr);
118
119         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
120
121         ops->sw(state, opcode, host, addr, data);
122 }
123
124 static void lightrec_swr(struct lightrec_state *state,
125                          const struct lightrec_mem_map_ops *ops,
126                          u32 opcode, void *host, u32 addr, u32 data)
127 {
128         unsigned int shift = addr & 0x3;
129         unsigned int mask = (1 << (shift * 8)) - 1;
130         u32 old_data;
131
132         /* Align to 32 bits */
133         addr &= ~3;
134         host = (void *)((uintptr_t)host & ~3);
135
136         old_data = ops->lw(state, opcode, host, addr);
137
138         data = (data << (shift * 8)) | (old_data & mask);
139
140         ops->sw(state, opcode, host, addr, data);
141 }
142
143 static void lightrec_swc2(struct lightrec_state *state, union code op,
144                           const struct lightrec_mem_map_ops *ops,
145                           void *host, u32 addr)
146 {
147         u32 data = lightrec_mfc2(state, op.i.rt);
148
149         ops->sw(state, op.opcode, host, addr, data);
150 }
151
152 static u32 lightrec_lwl(struct lightrec_state *state,
153                         const struct lightrec_mem_map_ops *ops,
154                         u32 opcode, void *host, u32 addr, u32 data)
155 {
156         unsigned int shift = addr & 0x3;
157         unsigned int mask = (1 << (24 - shift * 8)) - 1;
158         u32 old_data;
159
160         /* Align to 32 bits */
161         addr &= ~3;
162         host = (void *)((uintptr_t)host & ~3);
163
164         old_data = ops->lw(state, opcode, host, addr);
165
166         return (data & mask) | (old_data << (24 - shift * 8));
167 }
168
169 static u32 lightrec_lwr(struct lightrec_state *state,
170                         const struct lightrec_mem_map_ops *ops,
171                         u32 opcode, void *host, u32 addr, u32 data)
172 {
173         unsigned int shift = addr & 0x3;
174         unsigned int mask = GENMASK(31, 32 - shift * 8);
175         u32 old_data;
176
177         /* Align to 32 bits */
178         addr &= ~3;
179         host = (void *)((uintptr_t)host & ~3);
180
181         old_data = ops->lw(state, opcode, host, addr);
182
183         return (data & mask) | (old_data >> (shift * 8));
184 }
185
186 static void lightrec_lwc2(struct lightrec_state *state, union code op,
187                           const struct lightrec_mem_map_ops *ops,
188                           void *host, u32 addr)
189 {
190         u32 data = ops->lw(state, op.opcode, host, addr);
191
192         lightrec_mtc2(state, op.i.rt, data);
193 }
194
195 static void lightrec_invalidate_map(struct lightrec_state *state,
196                 const struct lightrec_mem_map *map, u32 addr, u32 len)
197 {
198         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
199                 memset(lut_address(state, lut_offset(addr)), 0,
200                        ((len + 3) / 4) * lut_elm_size(state));
201         }
202 }
203
204 enum psx_map
205 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
206 {
207         const struct lightrec_mem_map *map;
208         unsigned int i;
209
210         for (i = 0; i < state->nb_maps; i++) {
211                 map = &state->maps[i];
212
213                 if (kaddr >= map->pc && kaddr < map->pc + map->length)
214                         return (enum psx_map) i;
215         }
216
217         return PSX_MAP_UNKNOWN;
218 }
219
220 const struct lightrec_mem_map *
221 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
222 {
223         const struct lightrec_mem_map *map;
224         enum psx_map idx;
225         u32 addr;
226
227         idx = lightrec_get_map_idx(state, kaddr);
228         if (idx == PSX_MAP_UNKNOWN)
229                 return NULL;
230
231         map = &state->maps[idx];
232         addr = kaddr - map->pc;
233
234         while (map->mirror_of)
235                 map = map->mirror_of;
236
237         if (host)
238                 *host = map->address + addr;
239
240         return map;
241 }
242
243 u32 lightrec_rw(struct lightrec_state *state, union code op,
244                 u32 addr, u32 data, u32 *flags, struct block *block)
245 {
246         const struct lightrec_mem_map *map;
247         const struct lightrec_mem_map_ops *ops;
248         u32 opcode = op.opcode;
249         void *host;
250
251         addr += (s16) op.i.imm;
252
253         map = lightrec_get_map(state, &host, kunseg(addr));
254         if (!map) {
255                 __segfault_cb(state, addr, block);
256                 return 0;
257         }
258
259
260         if (likely(!map->ops)) {
261                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
262                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
263
264                 ops = &lightrec_default_ops;
265         } else if (flags &&
266                    LIGHTREC_FLAGS_GET_IO_MODE(*flags) == LIGHTREC_IO_DIRECT_HW) {
267                 ops = &lightrec_default_ops;
268         } else {
269                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
270                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
271
272                 ops = map->ops;
273         }
274
275         switch (op.i.op) {
276         case OP_SB:
277                 ops->sb(state, opcode, host, addr, (u8) data);
278                 return 0;
279         case OP_SH:
280                 ops->sh(state, opcode, host, addr, (u16) data);
281                 return 0;
282         case OP_SWL:
283                 lightrec_swl(state, ops, opcode, host, addr, data);
284                 return 0;
285         case OP_SWR:
286                 lightrec_swr(state, ops, opcode, host, addr, data);
287                 return 0;
288         case OP_SW:
289                 ops->sw(state, opcode, host, addr, data);
290                 return 0;
291         case OP_SWC2:
292                 lightrec_swc2(state, op, ops, host, addr);
293                 return 0;
294         case OP_LB:
295                 return (s32) (s8) ops->lb(state, opcode, host, addr);
296         case OP_LBU:
297                 return ops->lb(state, opcode, host, addr);
298         case OP_LH:
299                 return (s32) (s16) ops->lh(state, opcode, host, addr);
300         case OP_LHU:
301                 return ops->lh(state, opcode, host, addr);
302         case OP_LWC2:
303                 lightrec_lwc2(state, op, ops, host, addr);
304                 return 0;
305         case OP_LWL:
306                 return lightrec_lwl(state, ops, opcode, host, addr, data);
307         case OP_LWR:
308                 return lightrec_lwr(state, ops, opcode, host, addr, data);
309         case OP_LW:
310         default:
311                 return ops->lw(state, opcode, host, addr);
312         }
313 }
314
315 static void lightrec_rw_helper(struct lightrec_state *state,
316                                union code op, u32 *flags,
317                                struct block *block)
318 {
319         u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
320                               state->regs.gpr[op.i.rt], flags, block);
321
322         switch (op.i.op) {
323         case OP_LB:
324         case OP_LBU:
325         case OP_LH:
326         case OP_LHU:
327         case OP_LWL:
328         case OP_LWR:
329         case OP_LW:
330                 if (op.i.rt)
331                         state->regs.gpr[op.i.rt] = ret;
332                 fallthrough;
333         default:
334                 break;
335         }
336 }
337
338 static void lightrec_rw_cb(struct lightrec_state *state, u32 arg)
339 {
340         lightrec_rw_helper(state, (union code) arg, NULL, NULL);
341 }
342
343 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
344 {
345         struct block *block;
346         struct opcode *op;
347         bool was_tagged;
348         u16 offset = (u16)arg;
349         u16 old_flags;
350
351         block = lightrec_find_block_from_lut(state->block_cache,
352                                              arg >> 16, state->next_pc);
353         if (unlikely(!block)) {
354                 pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
355                          state->next_pc, offset);
356                 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
357                 return;
358         }
359
360         op = &block->opcode_list[offset];
361         was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(op->flags);
362
363         lightrec_rw_helper(state, op->c, &op->flags, block);
364
365         if (!was_tagged) {
366                 old_flags = block_set_flags(block, BLOCK_SHOULD_RECOMPILE);
367
368                 if (!(old_flags & BLOCK_SHOULD_RECOMPILE)) {
369                         pr_debug("Opcode of block at PC 0x%08x has been tagged"
370                                  " - flag for recompilation\n", block->pc);
371
372                         lut_write(state, lut_offset(block->pc), NULL);
373                 }
374         }
375 }
376
377 static u32 clamp_s32(s32 val, s32 min, s32 max)
378 {
379         return val < min ? min : val > max ? max : val;
380 }
381
382 static u16 load_u16(u32 *ptr)
383 {
384         return ((struct u16x2 *) ptr)->l;
385 }
386
387 static void store_u16(u32 *ptr, u16 value)
388 {
389         ((struct u16x2 *) ptr)->l = value;
390 }
391
392 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
393 {
394         s16 gteir1, gteir2, gteir3;
395
396         switch (reg) {
397         case 1:
398         case 3:
399         case 5:
400         case 8:
401         case 9:
402         case 10:
403         case 11:
404                 return (s32)(s16) load_u16(&state->regs.cp2d[reg]);
405         case 7:
406         case 16:
407         case 17:
408         case 18:
409         case 19:
410                 return load_u16(&state->regs.cp2d[reg]);
411         case 28:
412         case 29:
413                 gteir1 = (s16) load_u16(&state->regs.cp2d[9]);
414                 gteir2 = (s16) load_u16(&state->regs.cp2d[10]);
415                 gteir3 = (s16) load_u16(&state->regs.cp2d[11]);
416
417                 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
418                         clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
419                         clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
420         case 15:
421                 reg = 14;
422                 fallthrough;
423         default:
424                 return state->regs.cp2d[reg];
425         }
426 }
427
428 u32 lightrec_mfc(struct lightrec_state *state, union code op)
429 {
430         u32 val;
431
432         if (op.i.op == OP_CP0)
433                 return state->regs.cp0[op.r.rd];
434         else if (op.r.rs == OP_CP2_BASIC_MFC2)
435                 return lightrec_mfc2(state, op.r.rd);
436
437         val = state->regs.cp2c[op.r.rd];
438
439         switch (op.r.rd) {
440         case 4:
441         case 12:
442         case 20:
443         case 26:
444         case 27:
445         case 29:
446         case 30:
447                 return (u32)(s16)val;
448         default:
449                 return val;
450         }
451 }
452
453 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
454 {
455         u32 status, oldstatus, cause;
456
457         switch (reg) {
458         case 1:
459         case 4:
460         case 8:
461         case 14:
462         case 15:
463                 /* Those registers are read-only */
464                 return;
465         default:
466                 break;
467         }
468
469         if (reg == 12) {
470                 status = state->regs.cp0[12];
471                 oldstatus = status;
472
473                 if (status & ~data & BIT(16)) {
474                         state->ops.enable_ram(state, true);
475                         lightrec_invalidate_all(state);
476                 } else if (~status & data & BIT(16)) {
477                         state->ops.enable_ram(state, false);
478                 }
479         }
480
481         if (reg == 13) {
482                 state->regs.cp0[13] &= ~0x300;
483                 state->regs.cp0[13] |= data & 0x300;
484         } else {
485                 state->regs.cp0[reg] = data;
486         }
487
488         if (reg == 12 || reg == 13) {
489                 cause = state->regs.cp0[13];
490                 status = state->regs.cp0[12];
491
492                 /* Handle software interrupts */
493                 if (!!(status & cause & 0x300) & status)
494                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
495
496                 /* Handle hardware interrupts */
497                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
498                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
499         }
500 }
501
502 static u32 count_leading_bits(s32 data)
503 {
504         u32 cnt = 33;
505
506 #ifdef __has_builtin
507 #if __has_builtin(__builtin_clrsb)
508         return 1 + __builtin_clrsb(data);
509 #endif
510 #endif
511
512         data = (data ^ (data >> 31)) << 1;
513
514         do {
515                 cnt -= 1;
516                 data >>= 1;
517         } while (data);
518
519         return cnt;
520 }
521
522 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
523 {
524         switch (reg) {
525         case 15:
526                 state->regs.cp2d[12] = state->regs.cp2d[13];
527                 state->regs.cp2d[13] = state->regs.cp2d[14];
528                 state->regs.cp2d[14] = data;
529                 break;
530         case 28:
531                 state->regs.cp2d[9] = (data << 7) & 0xf80;
532                 state->regs.cp2d[10] = (data << 2) & 0xf80;
533                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
534                 break;
535         case 31:
536                 return;
537         case 30:
538                 state->regs.cp2d[31] = count_leading_bits((s32) data);
539                 fallthrough;
540         default:
541                 state->regs.cp2d[reg] = data;
542                 break;
543         }
544 }
545
546 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
547 {
548         switch (reg) {
549         case 4:
550         case 12:
551         case 20:
552         case 26:
553         case 27:
554         case 29:
555         case 30:
556                 store_u16(&state->regs.cp2c[reg], data);
557                 break;
558         case 31:
559                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
560                 fallthrough;
561         default:
562                 state->regs.cp2c[reg] = data;
563                 break;
564         }
565 }
566
567 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
568 {
569         if (op.i.op == OP_CP0)
570                 lightrec_mtc0(state, op.r.rd, data);
571         else if (op.r.rs == OP_CP2_BASIC_CTC2)
572                 lightrec_ctc2(state, op.r.rd, data);
573         else
574                 lightrec_mtc2(state, op.r.rd, data);
575 }
576
577 static void lightrec_mtc_cb(struct lightrec_state *state, u32 arg)
578 {
579         union code op = (union code) arg;
580
581         lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
582 }
583
584 void lightrec_rfe(struct lightrec_state *state)
585 {
586         u32 status;
587
588         /* Read CP0 Status register (r12) */
589         status = state->regs.cp0[12];
590
591         /* Switch the bits */
592         status = ((status & 0x3c) >> 2) | (status & ~0xf);
593
594         /* Write it back */
595         lightrec_mtc0(state, 12, status);
596 }
597
598 void lightrec_cp(struct lightrec_state *state, union code op)
599 {
600         if (op.i.op == OP_CP0) {
601                 pr_err("Invalid CP opcode to coprocessor #0\n");
602                 return;
603         }
604
605         (*state->ops.cop2_op)(state, op.opcode);
606 }
607
608 static void lightrec_cp_cb(struct lightrec_state *state, u32 arg)
609 {
610         lightrec_cp(state, (union code) arg);
611 }
612
613 static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
614 {
615         struct block *block = lightrec_find_block(state->block_cache, pc);
616         u8 old_flags;
617
618         if (block && lightrec_block_is_outdated(state, block)) {
619                 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
620
621                 old_flags = block_set_flags(block, BLOCK_IS_DEAD);
622                 if (!(old_flags & BLOCK_IS_DEAD)) {
623                         /* Make sure the recompiler isn't processing the block
624                          * we'll destroy */
625                         if (ENABLE_THREADED_COMPILER)
626                                 lightrec_recompiler_remove(state->rec, block);
627
628                         lightrec_unregister_block(state->block_cache, block);
629                         remove_from_code_lut(state->block_cache, block);
630                         lightrec_free_block(state, block);
631                 }
632
633                 block = NULL;
634         }
635
636         if (!block) {
637                 block = lightrec_precompile_block(state, pc);
638                 if (!block) {
639                         pr_err("Unable to recompile block at PC 0x%x\n", pc);
640                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
641                         return NULL;
642                 }
643
644                 lightrec_register_block(state->block_cache, block);
645         }
646
647         return block;
648 }
649
650 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
651 {
652         struct block *block;
653         bool should_recompile;
654         void *func;
655         int err;
656
657         for (;;) {
658                 func = lut_read(state, lut_offset(pc));
659                 if (func && func != state->get_next_block)
660                         break;
661
662                 block = lightrec_get_block(state, pc);
663
664                 if (unlikely(!block))
665                         break;
666
667                 if (OPT_REPLACE_MEMSET &&
668                     block_has_flag(block, BLOCK_IS_MEMSET)) {
669                         func = state->memset_func;
670                         break;
671                 }
672
673                 should_recompile = block_has_flag(block, BLOCK_SHOULD_RECOMPILE) &&
674                         !block_has_flag(block, BLOCK_IS_DEAD);
675
676                 if (unlikely(should_recompile)) {
677                         pr_debug("Block at PC 0x%08x should recompile\n", pc);
678
679                         if (ENABLE_THREADED_COMPILER) {
680                                 lightrec_recompiler_add(state->rec, block);
681                         } else {
682                                 err = lightrec_compile_block(state->cstate, block);
683                                 if (err) {
684                                         state->exit_flags = LIGHTREC_EXIT_NOMEM;
685                                         return NULL;
686                                 }
687                         }
688                 }
689
690                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
691                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
692                 else
693                         func = block->function;
694
695                 if (likely(func))
696                         break;
697
698                 if (unlikely(block_has_flag(block, BLOCK_NEVER_COMPILE))) {
699                         pc = lightrec_emulate_block(state, block, pc);
700
701                 } else if (!ENABLE_THREADED_COMPILER) {
702                         /* Block wasn't compiled yet - run the interpreter */
703                         if (block_has_flag(block, BLOCK_FULLY_TAGGED))
704                                 pr_debug("Block fully tagged, skipping first pass\n");
705                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
706                                 pc = lightrec_emulate_block(state, block, pc);
707
708                         /* Then compile it using the profiled data */
709                         err = lightrec_compile_block(state->cstate, block);
710                         if (err) {
711                                 state->exit_flags = LIGHTREC_EXIT_NOMEM;
712                                 return NULL;
713                         }
714                 } else if (unlikely(block_has_flag(block, BLOCK_IS_DEAD))) {
715                         /*
716                          * If the block is dead but has never been compiled,
717                          * then its function pointer is NULL and we cannot
718                          * execute the block. In that case, reap all the dead
719                          * blocks now, and in the next loop we will create a
720                          * new block.
721                          */
722                         lightrec_reaper_reap(state->reaper);
723                 } else {
724                         lightrec_recompiler_add(state->rec, block);
725                 }
726
727                 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
728                     state->current_cycle >= state->target_cycle)
729                         break;
730         }
731
732         state->next_pc = pc;
733         return func;
734 }
735
736 static void * lightrec_alloc_code(struct lightrec_state *state, size_t size)
737 {
738         void *code;
739
740         if (ENABLE_THREADED_COMPILER)
741                 lightrec_code_alloc_lock(state);
742
743         code = tlsf_malloc(state->tlsf, size);
744
745         if (ENABLE_THREADED_COMPILER)
746                 lightrec_code_alloc_unlock(state);
747
748         return code;
749 }
750
751 static void lightrec_realloc_code(struct lightrec_state *state,
752                                   void *ptr, size_t size)
753 {
754         /* NOTE: 'size' MUST be smaller than the size specified during
755          * the allocation. */
756
757         if (ENABLE_THREADED_COMPILER)
758                 lightrec_code_alloc_lock(state);
759
760         tlsf_realloc(state->tlsf, ptr, size);
761
762         if (ENABLE_THREADED_COMPILER)
763                 lightrec_code_alloc_unlock(state);
764 }
765
766 static void lightrec_free_code(struct lightrec_state *state, void *ptr)
767 {
768         if (ENABLE_THREADED_COMPILER)
769                 lightrec_code_alloc_lock(state);
770
771         tlsf_free(state->tlsf, ptr);
772
773         if (ENABLE_THREADED_COMPILER)
774                 lightrec_code_alloc_unlock(state);
775 }
776
777 static void * lightrec_emit_code(struct lightrec_state *state,
778                                  const struct block *block,
779                                  jit_state_t *_jit, unsigned int *size)
780 {
781         bool has_code_buffer = ENABLE_CODE_BUFFER && state->tlsf;
782         jit_word_t code_size, new_code_size;
783         void *code;
784
785         jit_realize();
786
787         if (!ENABLE_DISASSEMBLER)
788                 jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
789
790         if (has_code_buffer) {
791                 jit_get_code(&code_size);
792                 code = lightrec_alloc_code(state, (size_t) code_size);
793
794                 if (!code) {
795                         if (ENABLE_THREADED_COMPILER) {
796                                 /* If we're using the threaded compiler, return
797                                  * an allocation error here. The threaded
798                                  * compiler will then empty its job queue and
799                                  * request a code flush using the reaper. */
800                                 return NULL;
801                         }
802
803                         /* Remove outdated blocks, and try again */
804                         lightrec_remove_outdated_blocks(state->block_cache, block);
805
806                         pr_debug("Re-try to alloc %zu bytes...\n", code_size);
807
808                         code = lightrec_alloc_code(state, code_size);
809                         if (!code) {
810                                 pr_err("Could not alloc even after removing old blocks!\n");
811                                 return NULL;
812                         }
813                 }
814
815                 jit_set_code(code, code_size);
816         }
817
818         code = jit_emit();
819
820         jit_get_code(&new_code_size);
821         lightrec_register(MEM_FOR_CODE, new_code_size);
822
823         if (has_code_buffer) {
824                 lightrec_realloc_code(state, code, (size_t) new_code_size);
825
826                 pr_debug("Creating code block at address 0x%" PRIxPTR ", "
827                          "code size: %" PRIuPTR " new: %" PRIuPTR "\n",
828                          (uintptr_t) code, code_size, new_code_size);
829         }
830
831         *size = (unsigned int) new_code_size;
832
833         return code;
834 }
835
836 static struct block * generate_wrapper(struct lightrec_state *state)
837 {
838         struct block *block;
839         jit_state_t *_jit;
840         unsigned int i;
841         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
842         jit_node_t *to_end[C_WRAPPERS_COUNT - 1];
843
844         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
845         if (!block)
846                 goto err_no_mem;
847
848         _jit = jit_new_state();
849         if (!_jit)
850                 goto err_free_block;
851
852         jit_name("RW wrapper");
853         jit_note(__FILE__, __LINE__);
854
855         /* Wrapper entry point */
856         jit_prolog();
857         jit_tramp(256);
858
859         /* Add entry points */
860         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
861                 jit_ldxi(JIT_R1, LIGHTREC_REG_STATE,
862                          offsetof(struct lightrec_state, c_wrappers[i]));
863                 to_end[i - 1] = jit_b();
864                 addr[i - 1] = jit_indirect();
865         }
866
867         jit_ldxi(JIT_R1, LIGHTREC_REG_STATE,
868                  offsetof(struct lightrec_state, c_wrappers[0]));
869
870         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
871                 jit_patch(to_end[i]);
872
873         jit_epilog();
874         jit_prolog();
875
876         /* Save all temporaries on stack */
877         for (i = 0; i < NUM_TEMPS; i++) {
878                 if (i + FIRST_TEMP != 1) {
879                         jit_stxi(offsetof(struct lightrec_state, wrapper_regs[i]),
880                                  LIGHTREC_REG_STATE, JIT_R(i + FIRST_TEMP));
881                 }
882         }
883
884         jit_getarg(JIT_R2, jit_arg());
885
886         jit_prepare();
887         jit_pushargr(LIGHTREC_REG_STATE);
888         jit_pushargr(JIT_R2);
889
890         jit_ldxi_ui(JIT_R2, LIGHTREC_REG_STATE,
891                     offsetof(struct lightrec_state, target_cycle));
892
893         /* state->current_cycle = state->target_cycle - delta; */
894         jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, LIGHTREC_REG_CYCLE);
895         jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
896                    LIGHTREC_REG_STATE, LIGHTREC_REG_CYCLE);
897
898         /* Call the wrapper function */
899         jit_finishr(JIT_R1);
900
901         /* delta = state->target_cycle - state->current_cycle */;
902         jit_ldxi_ui(LIGHTREC_REG_CYCLE, LIGHTREC_REG_STATE,
903                     offsetof(struct lightrec_state, current_cycle));
904         jit_ldxi_ui(JIT_R1, LIGHTREC_REG_STATE,
905                     offsetof(struct lightrec_state, target_cycle));
906         jit_subr(LIGHTREC_REG_CYCLE, JIT_R1, LIGHTREC_REG_CYCLE);
907
908         /* Restore temporaries from stack */
909         for (i = 0; i < NUM_TEMPS; i++) {
910                 if (i + FIRST_TEMP != 1) {
911                         jit_ldxi(JIT_R(i + FIRST_TEMP), LIGHTREC_REG_STATE,
912                                  offsetof(struct lightrec_state, wrapper_regs[i]));
913                 }
914         }
915
916         jit_ret();
917         jit_epilog();
918
919         block->_jit = _jit;
920         block->opcode_list = NULL;
921         block->flags = BLOCK_NO_OPCODE_LIST;
922         block->nb_ops = 0;
923
924         block->function = lightrec_emit_code(state, block, _jit,
925                                              &block->code_size);
926         if (!block->function)
927                 goto err_free_block;
928
929         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
930
931         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
932                 state->wrappers_eps[i] = jit_address(addr[i]);
933
934         if (ENABLE_DISASSEMBLER) {
935                 pr_debug("Wrapper block:\n");
936                 jit_disassemble();
937         }
938
939         jit_clear_state();
940         return block;
941
942 err_free_block:
943         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
944 err_no_mem:
945         pr_err("Unable to compile wrapper: Out of memory\n");
946         return NULL;
947 }
948
949 static u32 lightrec_memset(struct lightrec_state *state)
950 {
951         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
952         void *host;
953         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
954         u32 length = state->regs.gpr[5] * 4;
955
956         if (!map) {
957                 pr_err("Unable to find memory map for memset target address "
958                        "0x%x\n", kunseg_pc);
959                 return 0;
960         }
961
962         pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
963                  kunseg_pc, (uintptr_t)host, length);
964         memset(host, 0, length);
965
966         if (!state->invalidate_from_dma_only)
967                 lightrec_invalidate_map(state, map, kunseg_pc, length);
968
969         /* Rough estimation of the number of cycles consumed */
970         return 8 + 5 * (length  + 3 / 4);
971 }
972
973 static struct block * generate_dispatcher(struct lightrec_state *state)
974 {
975         struct block *block;
976         jit_state_t *_jit;
977         jit_node_t *to_end, *loop, *addr, *addr2, *addr3;
978         unsigned int i;
979         u32 offset;
980
981         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
982         if (!block)
983                 goto err_no_mem;
984
985         _jit = jit_new_state();
986         if (!_jit)
987                 goto err_free_block;
988
989         jit_name("dispatcher");
990         jit_note(__FILE__, __LINE__);
991
992         jit_prolog();
993         jit_frame(256);
994
995         jit_getarg(JIT_V1, jit_arg());
996         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
997
998         /* Force all callee-saved registers to be pushed on the stack */
999         for (i = 0; i < NUM_REGS; i++)
1000                 jit_movr(JIT_V(i + FIRST_REG), JIT_V(i + FIRST_REG));
1001
1002         /* Pass lightrec_state structure to blocks, using the last callee-saved
1003          * register that Lightning provides */
1004         jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
1005
1006         loop = jit_label();
1007
1008         /* Call the block's code */
1009         jit_jmpr(JIT_V1);
1010
1011         if (OPT_REPLACE_MEMSET) {
1012                 /* Blocks will jump here when they need to call
1013                  * lightrec_memset() */
1014                 addr3 = jit_indirect();
1015
1016                 jit_movr(JIT_V1, LIGHTREC_REG_CYCLE);
1017
1018                 jit_prepare();
1019                 jit_pushargr(LIGHTREC_REG_STATE);
1020                 jit_finishi(lightrec_memset);
1021
1022                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1023                             offsetof(struct lightrec_state, regs.gpr[31]));
1024
1025                 jit_retval(LIGHTREC_REG_CYCLE);
1026                 jit_subr(LIGHTREC_REG_CYCLE, JIT_V1, LIGHTREC_REG_CYCLE);
1027         }
1028
1029         /* The block will jump here, with the number of cycles remaining in
1030          * LIGHTREC_REG_CYCLE */
1031         addr2 = jit_indirect();
1032
1033         /* Store back the next_pc to the lightrec_state structure */
1034         offset = offsetof(struct lightrec_state, next_pc);
1035         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
1036
1037         /* Jump to end if state->target_cycle < state->current_cycle */
1038         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
1039
1040         /* Convert next PC to KUNSEG and avoid mirrors */
1041         jit_andi(JIT_V1, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
1042         jit_rshi_u(JIT_R1, JIT_V1, 28);
1043         jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
1044         jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
1045         jit_movnr(JIT_V1, JIT_R2, JIT_R1);
1046
1047         /* If possible, use the code LUT */
1048         if (!lut_is_32bit(state))
1049                 jit_lshi(JIT_V1, JIT_V1, 1);
1050         jit_addr(JIT_V1, JIT_V1, LIGHTREC_REG_STATE);
1051
1052         offset = offsetof(struct lightrec_state, code_lut);
1053         if (lut_is_32bit(state))
1054                 jit_ldxi_ui(JIT_V1, JIT_V1, offset);
1055         else
1056                 jit_ldxi(JIT_V1, JIT_V1, offset);
1057
1058         /* If we get non-NULL, loop */
1059         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1060
1061         /* The code LUT will be set to this address when the block at the target
1062          * PC has been preprocessed but not yet compiled by the threaded
1063          * recompiler */
1064         addr = jit_indirect();
1065
1066         /* Slow path: call C function get_next_block_func() */
1067
1068         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1069                 /* We may call the interpreter - update state->current_cycle */
1070                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1071                            offsetof(struct lightrec_state, target_cycle));
1072                 jit_subr(JIT_V1, JIT_R2, LIGHTREC_REG_CYCLE);
1073                 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
1074                            LIGHTREC_REG_STATE, JIT_V1);
1075         }
1076
1077         jit_prepare();
1078         jit_pushargr(LIGHTREC_REG_STATE);
1079         jit_pushargr(JIT_V0);
1080
1081         /* Save the cycles register if needed */
1082         if (!(ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES))
1083                 jit_movr(JIT_V0, LIGHTREC_REG_CYCLE);
1084
1085         /* Get the next block */
1086         jit_finishi(&get_next_block_func);
1087         jit_retval(JIT_V1);
1088
1089         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1090                 /* The interpreter may have updated state->current_cycle and
1091                  * state->target_cycle - recalc the delta */
1092                 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
1093                            offsetof(struct lightrec_state, current_cycle));
1094                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1095                            offsetof(struct lightrec_state, target_cycle));
1096                 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
1097         } else {
1098                 jit_movr(LIGHTREC_REG_CYCLE, JIT_V0);
1099         }
1100
1101         /* If we get non-NULL, loop */
1102         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1103
1104         /* When exiting, the recompiled code will jump to that address */
1105         jit_note(__FILE__, __LINE__);
1106         jit_patch(to_end);
1107
1108         jit_retr(LIGHTREC_REG_CYCLE);
1109         jit_epilog();
1110
1111         block->_jit = _jit;
1112         block->opcode_list = NULL;
1113         block->flags = BLOCK_NO_OPCODE_LIST;
1114         block->nb_ops = 0;
1115
1116         block->function = lightrec_emit_code(state, block, _jit,
1117                                              &block->code_size);
1118         if (!block->function)
1119                 goto err_free_block;
1120
1121         state->eob_wrapper_func = jit_address(addr2);
1122         if (OPT_REPLACE_MEMSET)
1123                 state->memset_func = jit_address(addr3);
1124         state->get_next_block = jit_address(addr);
1125
1126         if (ENABLE_DISASSEMBLER) {
1127                 pr_debug("Dispatcher block:\n");
1128                 jit_disassemble();
1129         }
1130
1131         /* We're done! */
1132         jit_clear_state();
1133         return block;
1134
1135 err_free_block:
1136         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1137 err_no_mem:
1138         pr_err("Unable to compile dispatcher: Out of memory\n");
1139         return NULL;
1140 }
1141
1142 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
1143 {
1144         void *host = NULL;
1145
1146         lightrec_get_map(state, &host, kunseg(pc));
1147
1148         const u32 *code = (u32 *)host;
1149         return (union code) LE32TOH(*code);
1150 }
1151
1152 unsigned int lightrec_cycles_of_opcode(union code code)
1153 {
1154         return 2;
1155 }
1156
1157 void lightrec_free_opcode_list(struct lightrec_state *state, struct opcode *ops)
1158 {
1159         struct opcode_list *list = container_of(ops, struct opcode_list, ops);
1160
1161         lightrec_free(state, MEM_FOR_IR,
1162                       sizeof(*list) + list->nb_ops * sizeof(struct opcode),
1163                       list);
1164 }
1165
1166 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1167 {
1168         unsigned int i;
1169         union code c;
1170
1171         for (i = 1; ; i++) {
1172                 c.opcode = LE32TOH(*src++);
1173
1174                 if (is_syscall(c))
1175                         return i;
1176
1177                 if (is_unconditional_jump(c))
1178                         return i + 1;
1179         }
1180 }
1181
1182 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1183                                             const u32 *src, unsigned int *len)
1184 {
1185         struct opcode_list *list;
1186         unsigned int i, length;
1187
1188         length = lightrec_get_mips_block_len(src);
1189
1190         list = lightrec_malloc(state, MEM_FOR_IR,
1191                                sizeof(*list) + sizeof(struct opcode) * length);
1192         if (!list) {
1193                 pr_err("Unable to allocate memory\n");
1194                 return NULL;
1195         }
1196
1197         list->nb_ops = (u16) length;
1198
1199         for (i = 0; i < length; i++) {
1200                 list->ops[i].opcode = LE32TOH(src[i]);
1201                 list->ops[i].flags = 0;
1202         }
1203
1204         *len = length * sizeof(u32);
1205
1206         return list->ops;
1207 }
1208
1209 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1210                                                 u32 pc)
1211 {
1212         struct opcode *list;
1213         struct block *block;
1214         void *host, *addr;
1215         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1216         const u32 *code = (u32 *) host;
1217         unsigned int length;
1218         bool fully_tagged;
1219         u8 block_flags = 0;
1220
1221         if (!map)
1222                 return NULL;
1223
1224         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1225         if (!block) {
1226                 pr_err("Unable to recompile block: Out of memory\n");
1227                 return NULL;
1228         }
1229
1230         list = lightrec_disassemble(state, code, &length);
1231         if (!list) {
1232                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1233                 return NULL;
1234         }
1235
1236         block->pc = pc;
1237         block->_jit = NULL;
1238         block->function = NULL;
1239         block->opcode_list = list;
1240         block->code = code;
1241         block->next = NULL;
1242         block->flags = 0;
1243         block->code_size = 0;
1244         block->precompile_date = state->current_cycle;
1245         block->nb_ops = length / sizeof(u32);
1246
1247         lightrec_optimize(state, block);
1248
1249         length = block->nb_ops * sizeof(u32);
1250
1251         lightrec_register(MEM_FOR_MIPS_CODE, length);
1252
1253         if (ENABLE_DISASSEMBLER) {
1254                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1255                 lightrec_print_disassembly(block, code);
1256         }
1257
1258         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1259
1260         /* If the first opcode is an 'impossible' branch, never compile the
1261          * block */
1262         if (should_emulate(block->opcode_list))
1263                 block_flags |= BLOCK_NEVER_COMPILE;
1264
1265         fully_tagged = lightrec_block_is_fully_tagged(block);
1266         if (fully_tagged)
1267                 block_flags |= BLOCK_FULLY_TAGGED;
1268
1269         if (block_flags)
1270                 block_set_flags(block, block_flags);
1271
1272         block->hash = lightrec_calculate_block_hash(block);
1273
1274         if (OPT_REPLACE_MEMSET && block_has_flag(block, BLOCK_IS_MEMSET))
1275                 addr = state->memset_func;
1276         else
1277                 addr = state->get_next_block;
1278         lut_write(state, lut_offset(pc), addr);
1279
1280         pr_debug("Recompile count: %u\n", state->nb_precompile++);
1281
1282         return block;
1283 }
1284
1285 static bool lightrec_block_is_fully_tagged(const struct block *block)
1286 {
1287         const struct opcode *op;
1288         unsigned int i;
1289
1290         for (i = 0; i < block->nb_ops; i++) {
1291                 op = &block->opcode_list[i];
1292
1293                 /* Verify that all load/stores of the opcode list
1294                  * Check all loads/stores of the opcode list and mark the
1295                  * block as fully compiled if they all have been tagged. */
1296                 switch (op->c.i.op) {
1297                 case OP_LB:
1298                 case OP_LH:
1299                 case OP_LWL:
1300                 case OP_LW:
1301                 case OP_LBU:
1302                 case OP_LHU:
1303                 case OP_LWR:
1304                 case OP_SB:
1305                 case OP_SH:
1306                 case OP_SWL:
1307                 case OP_SW:
1308                 case OP_SWR:
1309                 case OP_LWC2:
1310                 case OP_SWC2:
1311                         if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
1312                                 return false;
1313                         fallthrough;
1314                 default:
1315                         continue;
1316                 }
1317         }
1318
1319         return true;
1320 }
1321
1322 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1323 {
1324         struct block *block = data;
1325
1326         pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1327         lightrec_unregister_block(state->block_cache, block);
1328         lightrec_free_block(state, block);
1329 }
1330
1331 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1332 {
1333         _jit_destroy_state(data);
1334 }
1335
1336 static void lightrec_free_function(struct lightrec_state *state, void *fn)
1337 {
1338         if (ENABLE_CODE_BUFFER && state->tlsf) {
1339                 pr_debug("Freeing code block at 0x%" PRIxPTR "\n", (uintptr_t) fn);
1340                 lightrec_free_code(state, fn);
1341         }
1342 }
1343
1344 static void lightrec_reap_function(struct lightrec_state *state, void *data)
1345 {
1346         lightrec_free_function(state, data);
1347 }
1348
1349 static void lightrec_reap_opcode_list(struct lightrec_state *state, void *data)
1350 {
1351         lightrec_free_opcode_list(state, data);
1352 }
1353
1354 int lightrec_compile_block(struct lightrec_cstate *cstate,
1355                            struct block *block)
1356 {
1357         struct lightrec_state *state = cstate->state;
1358         struct lightrec_branch_target *target;
1359         bool fully_tagged = false;
1360         struct block *block2;
1361         struct opcode *elm;
1362         jit_state_t *_jit, *oldjit;
1363         jit_node_t *start_of_block;
1364         bool skip_next = false;
1365         void *old_fn, *new_fn;
1366         size_t old_code_size;
1367         unsigned int i, j;
1368         u8 old_flags;
1369         u32 offset;
1370
1371         fully_tagged = lightrec_block_is_fully_tagged(block);
1372         if (fully_tagged)
1373                 block_set_flags(block, BLOCK_FULLY_TAGGED);
1374
1375         _jit = jit_new_state();
1376         if (!_jit)
1377                 return -ENOMEM;
1378
1379         oldjit = block->_jit;
1380         old_fn = block->function;
1381         old_code_size = block->code_size;
1382         block->_jit = _jit;
1383
1384         lightrec_regcache_reset(cstate->reg_cache);
1385         cstate->cycles = 0;
1386         cstate->nb_local_branches = 0;
1387         cstate->nb_targets = 0;
1388
1389         jit_prolog();
1390         jit_tramp(256);
1391
1392         start_of_block = jit_label();
1393
1394         for (i = 0; i < block->nb_ops; i++) {
1395                 elm = &block->opcode_list[i];
1396
1397                 if (skip_next) {
1398                         skip_next = false;
1399                         continue;
1400                 }
1401
1402                 if (should_emulate(elm)) {
1403                         pr_debug("Branch at offset 0x%x will be emulated\n",
1404                                  i << 2);
1405
1406                         lightrec_emit_eob(cstate, block, i, false);
1407                         skip_next = !op_flag_no_ds(elm->flags);
1408                 } else {
1409                         lightrec_rec_opcode(cstate, block, i);
1410                         skip_next = !op_flag_no_ds(elm->flags) && has_delay_slot(elm->c);
1411 #if _WIN32
1412                         /* FIXME: GNU Lightning on Windows seems to use our
1413                          * mapped registers as temporaries. Until the actual bug
1414                          * is found and fixed, unconditionally mark our
1415                          * registers as live here. */
1416                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1417 #endif
1418                 }
1419
1420                 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1421         }
1422
1423         for (i = 0; i < cstate->nb_local_branches; i++) {
1424                 struct lightrec_branch *branch = &cstate->local_branches[i];
1425
1426                 pr_debug("Patch local branch to offset 0x%x\n",
1427                          branch->target << 2);
1428
1429                 if (branch->target == 0) {
1430                         jit_patch_at(branch->branch, start_of_block);
1431                         continue;
1432                 }
1433
1434                 for (j = 0; j < cstate->nb_targets; j++) {
1435                         if (cstate->targets[j].offset == branch->target) {
1436                                 jit_patch_at(branch->branch,
1437                                              cstate->targets[j].label);
1438                                 break;
1439                         }
1440                 }
1441
1442                 if (j == cstate->nb_targets)
1443                         pr_err("Unable to find branch target\n");
1444         }
1445
1446         jit_ret();
1447         jit_epilog();
1448
1449         new_fn = lightrec_emit_code(state, block, _jit, &block->code_size);
1450         if (!new_fn) {
1451                 if (!ENABLE_THREADED_COMPILER)
1452                         pr_err("Unable to compile block!\n");
1453                 block->_jit = oldjit;
1454                 jit_clear_state();
1455                 _jit_destroy_state(_jit);
1456                 return -ENOMEM;
1457         }
1458
1459         /* Pause the reaper, because lightrec_reset_lut_offset() may try to set
1460          * the old block->function pointer to the code LUT. */
1461         if (ENABLE_THREADED_COMPILER)
1462                 lightrec_reaper_pause(state->reaper);
1463
1464         block->function = new_fn;
1465         block_clear_flags(block, BLOCK_SHOULD_RECOMPILE);
1466
1467         /* Add compiled function to the LUT */
1468         lut_write(state, lut_offset(block->pc), block->function);
1469
1470         if (ENABLE_THREADED_COMPILER)
1471                 lightrec_reaper_continue(state->reaper);
1472
1473         /* Detect old blocks that have been covered by the new one */
1474         for (i = 0; i < cstate->nb_targets; i++) {
1475                 target = &cstate->targets[i];
1476
1477                 if (!target->offset)
1478                         continue;
1479
1480                 offset = block->pc + target->offset * sizeof(u32);
1481
1482                 /* Pause the reaper while we search for the block until we set
1483                  * the BLOCK_IS_DEAD flag, otherwise the block may be removed
1484                  * under our feet. */
1485                 if (ENABLE_THREADED_COMPILER)
1486                         lightrec_reaper_pause(state->reaper);
1487
1488                 block2 = lightrec_find_block(state->block_cache, offset);
1489                 if (block2) {
1490                         /* No need to check if block2 is compilable - it must
1491                          * be, otherwise block wouldn't be compilable either */
1492
1493                         /* Set the "block dead" flag to prevent the dynarec from
1494                          * recompiling this block */
1495                         old_flags = block_set_flags(block2, BLOCK_IS_DEAD);
1496                 }
1497
1498                 if (ENABLE_THREADED_COMPILER) {
1499                         lightrec_reaper_continue(state->reaper);
1500
1501                         /* If block2 was pending for compilation, cancel it.
1502                          * If it's being compiled right now, wait until it
1503                          * finishes. */
1504                         if (block2)
1505                                 lightrec_recompiler_remove(state->rec, block2);
1506                 }
1507
1508                 /* We know from now on that block2 (if present) isn't going to
1509                  * be compiled. We can override the LUT entry with our new
1510                  * block's entry point. */
1511                 offset = lut_offset(block->pc) + target->offset;
1512                 lut_write(state, offset, jit_address(target->label));
1513
1514                 if (block2) {
1515                         pr_debug("Reap block 0x%08x as it's covered by block "
1516                                  "0x%08x\n", block2->pc, block->pc);
1517
1518                         /* Finally, reap the block. */
1519                         if (!ENABLE_THREADED_COMPILER) {
1520                                 lightrec_unregister_block(state->block_cache, block2);
1521                                 lightrec_free_block(state, block2);
1522                         } else if (!(old_flags & BLOCK_IS_DEAD)) {
1523                                 lightrec_reaper_add(state->reaper,
1524                                                     lightrec_reap_block,
1525                                                     block2);
1526                         }
1527                 }
1528         }
1529
1530         if (ENABLE_DISASSEMBLER) {
1531                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1532                 jit_disassemble();
1533         }
1534
1535         jit_clear_state();
1536
1537         if (fully_tagged)
1538                 old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1539
1540         if (fully_tagged && !(old_flags & BLOCK_NO_OPCODE_LIST)) {
1541                 pr_debug("Block PC 0x%08x is fully tagged"
1542                          " - free opcode list\n", block->pc);
1543
1544                 if (ENABLE_THREADED_COMPILER) {
1545                         lightrec_reaper_add(state->reaper,
1546                                             lightrec_reap_opcode_list,
1547                                             block->opcode_list);
1548                 } else {
1549                         lightrec_free_opcode_list(state, block->opcode_list);
1550                 }
1551         }
1552
1553         if (oldjit) {
1554                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1555                          block->pc);
1556
1557                 if (ENABLE_THREADED_COMPILER) {
1558                         lightrec_reaper_add(state->reaper,
1559                                             lightrec_reap_jit, oldjit);
1560                         lightrec_reaper_add(state->reaper,
1561                                             lightrec_reap_function, old_fn);
1562                 } else {
1563                         _jit_destroy_state(oldjit);
1564                         lightrec_free_function(state, old_fn);
1565                 }
1566
1567                 lightrec_unregister(MEM_FOR_CODE, old_code_size);
1568         }
1569
1570         return 0;
1571 }
1572
1573 static void lightrec_print_info(struct lightrec_state *state)
1574 {
1575         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1576                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1577                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1578                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1579                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1580                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1581                         lightrec_get_total_mem_usage() / 1024,
1582                        lightrec_get_average_ipi());
1583                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1584         }
1585 }
1586
1587 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1588 {
1589         s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1590         void *block_trace;
1591         s32 cycles_delta;
1592
1593         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1594
1595         /* Handle the cycle counter overflowing */
1596         if (unlikely(target_cycle < state->current_cycle))
1597                 target_cycle = UINT_MAX;
1598
1599         state->target_cycle = target_cycle;
1600         state->next_pc = pc;
1601
1602         block_trace = get_next_block_func(state, pc);
1603         if (block_trace) {
1604                 cycles_delta = state->target_cycle - state->current_cycle;
1605
1606                 cycles_delta = (*func)(block_trace, cycles_delta);
1607
1608                 state->current_cycle = state->target_cycle - cycles_delta;
1609         }
1610
1611         if (ENABLE_THREADED_COMPILER)
1612                 lightrec_reaper_reap(state->reaper);
1613
1614         if (LOG_LEVEL >= INFO_L)
1615                 lightrec_print_info(state);
1616
1617         return state->next_pc;
1618 }
1619
1620 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc,
1621                              u32 target_cycle)
1622 {
1623         struct block *block;
1624
1625         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1626         state->target_cycle = target_cycle;
1627
1628         do {
1629                 block = lightrec_get_block(state, pc);
1630                 if (!block)
1631                         break;
1632
1633                 pc = lightrec_emulate_block(state, block, pc);
1634
1635                 if (ENABLE_THREADED_COMPILER)
1636                         lightrec_reaper_reap(state->reaper);
1637         } while (state->current_cycle < state->target_cycle);
1638
1639         if (LOG_LEVEL >= INFO_L)
1640                 lightrec_print_info(state);
1641
1642         return pc;
1643 }
1644
1645 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1646 {
1647         u8 old_flags;
1648
1649         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1650         old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1651
1652         if (!(old_flags & BLOCK_NO_OPCODE_LIST))
1653                 lightrec_free_opcode_list(state, block->opcode_list);
1654         if (block->_jit)
1655                 _jit_destroy_state(block->_jit);
1656         if (block->function) {
1657                 lightrec_free_function(state, block->function);
1658                 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1659         }
1660         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1661 }
1662
1663 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1664 {
1665         struct lightrec_cstate *cstate;
1666
1667         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1668         if (!cstate)
1669                 return NULL;
1670
1671         cstate->reg_cache = lightrec_regcache_init(state);
1672         if (!cstate->reg_cache) {
1673                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1674                 return NULL;
1675         }
1676
1677         cstate->state = state;
1678
1679         return cstate;
1680 }
1681
1682 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1683 {
1684         lightrec_free_regcache(cstate->reg_cache);
1685         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1686 }
1687
1688 struct lightrec_state * lightrec_init(char *argv0,
1689                                       const struct lightrec_mem_map *map,
1690                                       size_t nb,
1691                                       const struct lightrec_ops *ops)
1692 {
1693         const struct lightrec_mem_map *codebuf_map = &map[PSX_MAP_CODE_BUFFER];
1694         struct lightrec_state *state;
1695         uintptr_t addr;
1696         void *tlsf = NULL;
1697         bool with_32bit_lut = false;
1698         size_t lut_size;
1699
1700         /* Sanity-check ops */
1701         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1702                 pr_err("Missing callbacks in lightrec_ops structure\n");
1703                 return NULL;
1704         }
1705
1706         if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER
1707             && codebuf_map->address) {
1708                 tlsf = tlsf_create_with_pool(codebuf_map->address,
1709                                              codebuf_map->length);
1710                 if (!tlsf) {
1711                         pr_err("Unable to initialize code buffer\n");
1712                         return NULL;
1713                 }
1714
1715                 if (__WORDSIZE == 64) {
1716                         addr = (uintptr_t) codebuf_map->address + codebuf_map->length - 1;
1717                         with_32bit_lut = addr == (u32) addr;
1718                 }
1719         }
1720
1721         if (with_32bit_lut)
1722                 lut_size = CODE_LUT_SIZE * 4;
1723         else
1724                 lut_size = CODE_LUT_SIZE * sizeof(void *);
1725
1726         init_jit(argv0);
1727
1728         state = calloc(1, sizeof(*state) + lut_size);
1729         if (!state)
1730                 goto err_finish_jit;
1731
1732         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) + lut_size);
1733
1734         state->tlsf = tlsf;
1735         state->with_32bit_lut = with_32bit_lut;
1736
1737         state->block_cache = lightrec_blockcache_init(state);
1738         if (!state->block_cache)
1739                 goto err_free_state;
1740
1741         if (ENABLE_THREADED_COMPILER) {
1742                 state->rec = lightrec_recompiler_init(state);
1743                 if (!state->rec)
1744                         goto err_free_block_cache;
1745
1746                 state->reaper = lightrec_reaper_init(state);
1747                 if (!state->reaper)
1748                         goto err_free_recompiler;
1749         } else {
1750                 state->cstate = lightrec_create_cstate(state);
1751                 if (!state->cstate)
1752                         goto err_free_block_cache;
1753         }
1754
1755         state->nb_maps = nb;
1756         state->maps = map;
1757
1758         memcpy(&state->ops, ops, sizeof(*ops));
1759
1760         state->dispatcher = generate_dispatcher(state);
1761         if (!state->dispatcher)
1762                 goto err_free_reaper;
1763
1764         state->c_wrapper_block = generate_wrapper(state);
1765         if (!state->c_wrapper_block)
1766                 goto err_free_dispatcher;
1767
1768         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1769         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1770         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1771         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
1772
1773         map = &state->maps[PSX_MAP_BIOS];
1774         state->offset_bios = (uintptr_t)map->address - map->pc;
1775
1776         map = &state->maps[PSX_MAP_SCRATCH_PAD];
1777         state->offset_scratch = (uintptr_t)map->address - map->pc;
1778
1779         map = &state->maps[PSX_MAP_HW_REGISTERS];
1780         state->offset_io = (uintptr_t)map->address - map->pc;
1781
1782         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1783         state->offset_ram = (uintptr_t)map->address - map->pc;
1784
1785         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1786             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1787             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1788                 state->mirrors_mapped = true;
1789
1790         if (state->offset_bios == 0 &&
1791             state->offset_scratch == 0 &&
1792             state->offset_ram == 0 &&
1793             state->offset_io == 0 &&
1794             state->mirrors_mapped) {
1795                 pr_info("Memory map is perfect. Emitted code will be best.\n");
1796         } else {
1797                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1798         }
1799
1800         if (state->with_32bit_lut)
1801                 pr_info("Using 32-bit LUT\n");
1802
1803         return state;
1804
1805 err_free_dispatcher:
1806         lightrec_free_block(state, state->dispatcher);
1807 err_free_reaper:
1808         if (ENABLE_THREADED_COMPILER)
1809                 lightrec_reaper_destroy(state->reaper);
1810 err_free_recompiler:
1811         if (ENABLE_THREADED_COMPILER)
1812                 lightrec_free_recompiler(state->rec);
1813         else
1814                 lightrec_free_cstate(state->cstate);
1815 err_free_block_cache:
1816         lightrec_free_block_cache(state->block_cache);
1817 err_free_state:
1818         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1819                             lut_elm_size(state) * CODE_LUT_SIZE);
1820         free(state);
1821 err_finish_jit:
1822         finish_jit();
1823         if (ENABLE_CODE_BUFFER && tlsf)
1824                 tlsf_destroy(tlsf);
1825         return NULL;
1826 }
1827
1828 void lightrec_destroy(struct lightrec_state *state)
1829 {
1830         /* Force a print info on destroy*/
1831         state->current_cycle = ~state->current_cycle;
1832         lightrec_print_info(state);
1833
1834         lightrec_free_block_cache(state->block_cache);
1835         lightrec_free_block(state, state->dispatcher);
1836         lightrec_free_block(state, state->c_wrapper_block);
1837
1838         if (ENABLE_THREADED_COMPILER) {
1839                 lightrec_free_recompiler(state->rec);
1840                 lightrec_reaper_destroy(state->reaper);
1841         } else {
1842                 lightrec_free_cstate(state->cstate);
1843         }
1844
1845         finish_jit();
1846         if (ENABLE_CODE_BUFFER && state->tlsf)
1847                 tlsf_destroy(state->tlsf);
1848
1849         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1850                             lut_elm_size(state) * CODE_LUT_SIZE);
1851         free(state);
1852 }
1853
1854 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1855 {
1856         u32 kaddr = kunseg(addr & ~0x3);
1857         enum psx_map idx = lightrec_get_map_idx(state, kaddr);
1858
1859         switch (idx) {
1860         case PSX_MAP_MIRROR1:
1861         case PSX_MAP_MIRROR2:
1862         case PSX_MAP_MIRROR3:
1863                 /* Handle mirrors */
1864                 kaddr &= RAM_SIZE - 1;
1865                 fallthrough;
1866         case PSX_MAP_KERNEL_USER_RAM:
1867                 break;
1868         default:
1869                 return;
1870         }
1871
1872         memset(lut_address(state, lut_offset(kaddr)), 0,
1873                ((len + 3) / 4) * lut_elm_size(state));
1874 }
1875
1876 void lightrec_invalidate_all(struct lightrec_state *state)
1877 {
1878         memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
1879 }
1880
1881 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1882 {
1883         if (state->invalidate_from_dma_only != dma_only)
1884                 lightrec_invalidate_all(state);
1885
1886         state->invalidate_from_dma_only = dma_only;
1887 }
1888
1889 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1890 {
1891         if (flags != LIGHTREC_EXIT_NORMAL) {
1892                 state->exit_flags |= flags;
1893                 state->target_cycle = state->current_cycle;
1894         }
1895 }
1896
1897 u32 lightrec_exit_flags(struct lightrec_state *state)
1898 {
1899         return state->exit_flags;
1900 }
1901
1902 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1903 {
1904         return state->current_cycle;
1905 }
1906
1907 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1908 {
1909         state->current_cycle = cycles;
1910
1911         if (state->target_cycle < cycles)
1912                 state->target_cycle = cycles;
1913 }
1914
1915 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1916 {
1917         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1918                 if (cycles < state->current_cycle)
1919                         cycles = state->current_cycle;
1920
1921                 state->target_cycle = cycles;
1922         }
1923 }
1924
1925 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1926 {
1927         return &state->regs;
1928 }