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 static struct block * lightrec_precompile_block(struct lightrec_state *state,
  32                                                 u32 pc);
  33 static bool lightrec_block_is_fully_tagged(const struct block *block);
  34
  35 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
  36 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
  37
  38 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
  39                                 void *host, u32 addr, u32 data)
  40 {
  41         *(u8 *)host = (u8)data;
  42
  43         if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
  44                 lightrec_invalidate(state, addr, 1);
  45 }
  46
  47 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
  48                                 void *host, u32 addr, u32 data)
  49 {
  50         *(u16 *)host = HTOLE16((u16)data);
  51
  52         if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
  53                 lightrec_invalidate(state, addr, 2);
  54 }
  55
  56 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
  57                                 void *host, u32 addr, u32 data)
  58 {
  59         *(u32 *)host = HTOLE32(data);
  60
  61         if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
  62                 lightrec_invalidate(state, addr, 4);
  63 }
  64
  65 static u8 lightrec_default_lb(struct lightrec_state *state,
  66                               u32 opcode, void *host, u32 addr)
  67 {
  68         return *(u8 *)host;
  69 }
  70
  71 static u16 lightrec_default_lh(struct lightrec_state *state,
  72                                u32 opcode, void *host, u32 addr)
  73 {
  74         return LE16TOH(*(u16 *)host);
  75 }
  76
  77 static u32 lightrec_default_lw(struct lightrec_state *state,
  78                                u32 opcode, void *host, u32 addr)
  79 {
  80         return LE32TOH(*(u32 *)host);
  81 }
  82
  83 static u32 lightrec_default_lwu(struct lightrec_state *state,
  84                                 u32 opcode, void *host, u32 addr)
  85 {
  86         u32 val;
  87
  88         memcpy(&val, host, 4);
  89
  90         return LE32TOH(val);
  91 }
  92
  93 static void lightrec_default_swu(struct lightrec_state *state, u32 opcode,
  94                                  void *host, u32 addr, u32 data)
  95 {
  96         data = HTOLE32(data);
  97
  98         memcpy(host, &data, 4);
  99
 100         if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
 101                 lightrec_invalidate(state, addr & ~0x3, 8);
 102 }
 103
 104 static const struct lightrec_mem_map_ops lightrec_default_ops = {
 105         .sb = lightrec_default_sb,
 106         .sh = lightrec_default_sh,
 107         .sw = lightrec_default_sw,
 108         .lb = lightrec_default_lb,
 109         .lh = lightrec_default_lh,
 110         .lw = lightrec_default_lw,
 111         .lwu = lightrec_default_lwu,
 112         .swu = lightrec_default_swu,
 113 };
 114
 115 static void __segfault_cb(struct lightrec_state *state, u32 addr,
 116                           const struct block *block)
 117 {
 118         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 119         pr_err("Segmentation fault in recompiled code: invalid "
 120                "load/store at address "PC_FMT"\n", addr);
 121         if (block)
 122                 pr_err("Was executing block "PC_FMT"\n", block->pc);
 123 }
 124
 125 static void lightrec_swl(struct lightrec_state *state,
 126                          const struct lightrec_mem_map_ops *ops,
 127                          u32 opcode, void *host, u32 addr, u32 data)
 128 {
 129         unsigned int shift = addr & 0x3;
 130         unsigned int mask = shift < 3 ? GENMASK(31, (shift + 1) * 8) : 0;
 131         u32 old_data;
 132
 133         /* Align to 32 bits */
 134         addr &= ~3;
 135         host = (void *)((uintptr_t)host & ~3);
 136
 137         old_data = ops->lw(state, opcode, host, addr);
 138
 139         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
 140
 141         ops->sw(state, opcode, host, addr, data);
 142 }
 143
 144 static void lightrec_swr(struct lightrec_state *state,
 145                          const struct lightrec_mem_map_ops *ops,
 146                          u32 opcode, void *host, u32 addr, u32 data)
 147 {
 148         unsigned int shift = addr & 0x3;
 149         unsigned int mask = (1 << (shift * 8)) - 1;
 150         u32 old_data;
 151
 152         /* Align to 32 bits */
 153         addr &= ~3;
 154         host = (void *)((uintptr_t)host & ~3);
 155
 156         old_data = ops->lw(state, opcode, host, addr);
 157
 158         data = (data << (shift * 8)) | (old_data & mask);
 159
 160         ops->sw(state, opcode, host, addr, data);
 161 }
 162
 163 static void lightrec_swc2(struct lightrec_state *state, union code op,
 164                           const struct lightrec_mem_map_ops *ops,
 165                           void *host, u32 addr)
 166 {
 167         u32 data = lightrec_mfc2(state, op.i.rt);
 168
 169         ops->sw(state, op.opcode, host, addr, data);
 170 }
 171
 172 static u32 lightrec_lwl(struct lightrec_state *state,
 173                         const struct lightrec_mem_map_ops *ops,
 174                         u32 opcode, void *host, u32 addr, u32 data)
 175 {
 176         unsigned int shift = addr & 0x3;
 177         unsigned int mask = (1 << (24 - shift * 8)) - 1;
 178         u32 old_data;
 179
 180         /* Align to 32 bits */
 181         addr &= ~3;
 182         host = (void *)((uintptr_t)host & ~3);
 183
 184         old_data = ops->lw(state, opcode, host, addr);
 185
 186         return (data & mask) | (old_data << (24 - shift * 8));
 187 }
 188
 189 static u32 lightrec_lwr(struct lightrec_state *state,
 190                         const struct lightrec_mem_map_ops *ops,
 191                         u32 opcode, void *host, u32 addr, u32 data)
 192 {
 193         unsigned int shift = addr & 0x3;
 194         unsigned int mask = shift ? GENMASK(31, 32 - shift * 8) : 0;
 195         u32 old_data;
 196
 197         /* Align to 32 bits */
 198         addr &= ~3;
 199         host = (void *)((uintptr_t)host & ~3);
 200
 201         old_data = ops->lw(state, opcode, host, addr);
 202
 203         return (data & mask) | (old_data >> (shift * 8));
 204 }
 205
 206 static void lightrec_lwc2(struct lightrec_state *state, union code op,
 207                           const struct lightrec_mem_map_ops *ops,
 208                           void *host, u32 addr)
 209 {
 210         u32 data = ops->lw(state, op.opcode, host, addr);
 211
 212         lightrec_mtc2(state, op.i.rt, data);
 213 }
 214
 215 static void lightrec_invalidate_map(struct lightrec_state *state,
 216                 const struct lightrec_mem_map *map, u32 addr, u32 len)
 217 {
 218         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
 219                 memset(lut_address(state, lut_offset(addr)), 0,
 220                        ((len + 3) / 4) * lut_elm_size(state));
 221         }
 222 }
 223
 224 static enum psx_map
 225 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
 226 {
 227         const struct lightrec_mem_map *map;
 228         unsigned int i;
 229
 230         for (i = 0; i < state->nb_maps; i++) {
 231                 map = &state->maps[i];
 232
 233                 if (kaddr >= map->pc && kaddr < map->pc + map->length)
 234                         return (enum psx_map) i;
 235         }
 236
 237         return PSX_MAP_UNKNOWN;
 238 }
 239
 240 const struct lightrec_mem_map *
 241 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 242 {
 243         const struct lightrec_mem_map *map;
 244         enum psx_map idx;
 245         u32 addr;
 246
 247         idx = lightrec_get_map_idx(state, kaddr);
 248         if (idx == PSX_MAP_UNKNOWN)
 249                 return NULL;
 250
 251         map = &state->maps[idx];
 252         addr = kaddr - map->pc;
 253
 254         while (map->mirror_of)
 255                 map = map->mirror_of;
 256
 257         if (host)
 258                 *host = map->address + addr;
 259
 260         return map;
 261 }
 262
 263 u32 lightrec_rw(struct lightrec_state *state, union code op, u32 base,
 264                 u32 data, u32 *flags, struct block *block, u16 offset)
 265 {
 266         const struct lightrec_mem_map *map;
 267         const struct lightrec_mem_map_ops *ops;
 268         u32 opcode = op.opcode;
 269         bool was_tagged = true;
 270         u16 old_flags;
 271         u32 addr;
 272         void *host;
 273
 274         addr = kunseg(base + (s16) op.i.imm);
 275
 276         map = lightrec_get_map(state, &host, addr);
 277         if (!map) {
 278                 __segfault_cb(state, addr, block);
 279                 return 0;
 280         }
 281
 282         if (flags)
 283                 was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(*flags);
 284
 285         if (likely(!map->ops)) {
 286                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags)) {
 287                         /* Force parallel port accesses as HW accesses, because
 288                          * the direct-I/O emitters can't differenciate it. */
 289                         if (unlikely(map == &state->maps[PSX_MAP_PARALLEL_PORT]))
 290                                 *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 291                         /* If the base register is 0x0, be extra suspicious.
 292                          * Some games (e.g. Sled Storm) actually do segmentation
 293                          * faults by using uninitialized pointers, which are
 294                          * later initialized to point to hardware registers. */
 295                         else if (op.i.rs && base == 0x0)
 296                                 *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 297                         else
 298                                 *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
 299                 }
 300
 301                 ops = &lightrec_default_ops;
 302         } else if (flags &&
 303                    LIGHTREC_FLAGS_GET_IO_MODE(*flags) == LIGHTREC_IO_DIRECT_HW) {
 304                 ops = &lightrec_default_ops;
 305         } else {
 306                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 307                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 308
 309                 ops = map->ops;
 310         }
 311
 312         if (!was_tagged) {
 313                 old_flags = block_set_flags(block, BLOCK_SHOULD_RECOMPILE);
 314
 315                 if (!(old_flags & BLOCK_SHOULD_RECOMPILE)) {
 316                         pr_debug("Opcode of block at "PC_FMT" has been tagged"
 317                                  " - flag for recompilation\n", block->pc);
 318
 319                         lut_write(state, lut_offset(block->pc), NULL);
 320                 }
 321         }
 322
 323         switch (op.i.op) {
 324         case OP_SB:
 325                 ops->sb(state, opcode, host, addr, data);
 326                 return 0;
 327         case OP_SH:
 328                 ops->sh(state, opcode, host, addr, data);
 329                 return 0;
 330         case OP_SWL:
 331                 lightrec_swl(state, ops, opcode, host, addr, data);
 332                 return 0;
 333         case OP_SWR:
 334                 lightrec_swr(state, ops, opcode, host, addr, data);
 335                 return 0;
 336         case OP_SW:
 337                 ops->sw(state, opcode, host, addr, data);
 338                 return 0;
 339         case OP_SWC2:
 340                 lightrec_swc2(state, op, ops, host, addr);
 341                 return 0;
 342         case OP_LB:
 343                 return (s32) (s8) ops->lb(state, opcode, host, addr);
 344         case OP_LBU:
 345                 return ops->lb(state, opcode, host, addr);
 346         case OP_LH:
 347                 return (s32) (s16) ops->lh(state, opcode, host, addr);
 348         case OP_LHU:
 349                 return ops->lh(state, opcode, host, addr);
 350         case OP_LWC2:
 351                 lightrec_lwc2(state, op, ops, host, addr);
 352                 return 0;
 353         case OP_LWL:
 354                 return lightrec_lwl(state, ops, opcode, host, addr, data);
 355         case OP_LWR:
 356                 return lightrec_lwr(state, ops, opcode, host, addr, data);
 357         case OP_META_LWU:
 358                 return ops->lwu(state, opcode, host, addr);
 359         case OP_META_SWU:
 360                 ops->swu(state, opcode, host, addr, data);
 361                 return 0;
 362         case OP_LW:
 363         default:
 364                 return ops->lw(state, opcode, host, addr);
 365         }
 366 }
 367
 368 static void lightrec_rw_helper(struct lightrec_state *state,
 369                                union code op, u32 *flags,
 370                                struct block *block, u16 offset)
 371 {
 372         u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
 373                               state->regs.gpr[op.i.rt], flags, block, offset);
 374
 375         switch (op.i.op) {
 376         case OP_LB:
 377         case OP_LBU:
 378         case OP_LH:
 379         case OP_LHU:
 380         case OP_LWL:
 381         case OP_LWR:
 382         case OP_LW:
 383         case OP_META_LWU:
 384                 if (OPT_HANDLE_LOAD_DELAYS && unlikely(!state->in_delay_slot_n)) {
 385                         state->temp_reg = ret;
 386                         state->in_delay_slot_n = 0xff;
 387                 } else if (op.i.rt) {
 388                         state->regs.gpr[op.i.rt] = ret;
 389                 }
 390                 fallthrough;
 391         default:
 392                 break;
 393         }
 394 }
 395
 396 static void lightrec_rw_cb(struct lightrec_state *state, u32 arg)
 397 {
 398         lightrec_rw_helper(state, (union code) arg, NULL, NULL, 0);
 399 }
 400
 401 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 402 {
 403         struct block *block;
 404         struct opcode *op;
 405         u16 offset = (u16)arg;
 406
 407         block = lightrec_find_block_from_lut(state->block_cache,
 408                                              arg >> 16, state->curr_pc);
 409         if (unlikely(!block)) {
 410                 pr_err("rw_generic: No block found in LUT for "PC_FMT" offset 0x%"PRIx16"\n",
 411                          state->curr_pc, offset);
 412                 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 413                 return;
 414         }
 415
 416         op = &block->opcode_list[offset];
 417         lightrec_rw_helper(state, op->c, &op->flags, block, offset);
 418 }
 419
 420 static u32 clamp_s32(s32 val, s32 min, s32 max)
 421 {
 422         return val < min ? min : val > max ? max : val;
 423 }
 424
 425 static u16 load_u16(u32 *ptr)
 426 {
 427         return ((struct u16x2 *) ptr)->l;
 428 }
 429
 430 static void store_u16(u32 *ptr, u16 value)
 431 {
 432         ((struct u16x2 *) ptr)->l = value;
 433 }
 434
 435 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
 436 {
 437         s16 gteir1, gteir2, gteir3;
 438
 439         switch (reg) {
 440         case 1:
 441         case 3:
 442         case 5:
 443         case 8:
 444         case 9:
 445         case 10:
 446         case 11:
 447                 return (s32)(s16) load_u16(&state->regs.cp2d[reg]);
 448         case 7:
 449         case 16:
 450         case 17:
 451         case 18:
 452         case 19:
 453                 return load_u16(&state->regs.cp2d[reg]);
 454         case 28:
 455         case 29:
 456                 gteir1 = (s16) load_u16(&state->regs.cp2d[9]);
 457                 gteir2 = (s16) load_u16(&state->regs.cp2d[10]);
 458                 gteir3 = (s16) load_u16(&state->regs.cp2d[11]);
 459
 460                 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
 461                         clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
 462                         clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
 463         case 15:
 464                 reg = 14;
 465                 fallthrough;
 466         default:
 467                 return state->regs.cp2d[reg];
 468         }
 469 }
 470
 471 u32 lightrec_mfc(struct lightrec_state *state, union code op)
 472 {
 473         u32 val;
 474
 475         if (op.i.op == OP_CP0)
 476                 return state->regs.cp0[op.r.rd];
 477
 478         if (op.i.op == OP_SWC2) {
 479                 val = lightrec_mfc2(state, op.i.rt);
 480         } else if (op.r.rs == OP_CP2_BASIC_MFC2)
 481                 val = lightrec_mfc2(state, op.r.rd);
 482         else {
 483                 val = state->regs.cp2c[op.r.rd];
 484
 485                 switch (op.r.rd) {
 486                 case 4:
 487                 case 12:
 488                 case 20:
 489                 case 26:
 490                 case 27:
 491                 case 29:
 492                 case 30:
 493                         val = (u32)(s16)val;
 494                         fallthrough;
 495                 default:
 496                         break;
 497                 }
 498         }
 499
 500         if (state->ops.cop2_notify)
 501                 (*state->ops.cop2_notify)(state, op.opcode, val);
 502
 503         return val;
 504 }
 505
 506 static void lightrec_mfc_cb(struct lightrec_state *state, union code op)
 507 {
 508         u32 rt = lightrec_mfc(state, op);
 509
 510         if (op.i.op == OP_SWC2)
 511                 state->temp_reg = rt;
 512         else if (op.r.rt)
 513                 state->regs.gpr[op.r.rt] = rt;
 514 }
 515
 516 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 517 {
 518         u32 status, oldstatus, cause;
 519
 520         switch (reg) {
 521         case 1:
 522         case 4:
 523         case 8:
 524         case 14:
 525         case 15:
 526                 /* Those registers are read-only */
 527                 return;
 528         default:
 529                 break;
 530         }
 531
 532         if (reg == 12) {
 533                 status = state->regs.cp0[12];
 534                 oldstatus = status;
 535
 536                 if (status & ~data & BIT(16)) {
 537                         state->ops.enable_ram(state, true);
 538                         lightrec_invalidate_all(state);
 539                 } else if (~status & data & BIT(16)) {
 540                         state->ops.enable_ram(state, false);
 541                 }
 542         }
 543
 544         if (reg == 13) {
 545                 state->regs.cp0[13] &= ~0x300;
 546                 state->regs.cp0[13] |= data & 0x300;
 547         } else {
 548                 state->regs.cp0[reg] = data;
 549         }
 550
 551         if (reg == 12 || reg == 13) {
 552                 cause = state->regs.cp0[13];
 553                 status = state->regs.cp0[12];
 554
 555                 /* Handle software interrupts */
 556                 if ((!!(status & cause & 0x300)) & status)
 557                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 558
 559                 /* Handle hardware interrupts */
 560                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
 561                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 562         }
 563 }
 564
 565 static u32 count_leading_bits(s32 data)
 566 {
 567         u32 cnt = 33;
 568
 569 #ifdef __has_builtin
 570 #if __has_builtin(__builtin_clrsb)
 571         return 1 + __builtin_clrsb(data);
 572 #endif
 573 #endif
 574
 575         data = (data ^ (data >> 31)) << 1;
 576
 577         do {
 578                 cnt -= 1;
 579                 data >>= 1;
 580         } while (data);
 581
 582         return cnt;
 583 }
 584
 585 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
 586 {
 587         switch (reg) {
 588         case 15:
 589                 state->regs.cp2d[12] = state->regs.cp2d[13];
 590                 state->regs.cp2d[13] = state->regs.cp2d[14];
 591                 state->regs.cp2d[14] = data;
 592                 break;
 593         case 28:
 594                 state->regs.cp2d[9] = (data << 7) & 0xf80;
 595                 state->regs.cp2d[10] = (data << 2) & 0xf80;
 596                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
 597                 break;
 598         case 31:
 599                 return;
 600         case 30:
 601                 state->regs.cp2d[31] = count_leading_bits((s32) data);
 602                 fallthrough;
 603         default:
 604                 state->regs.cp2d[reg] = data;
 605                 break;
 606         }
 607 }
 608
 609 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 610 {
 611         switch (reg) {
 612         case 4:
 613         case 12:
 614         case 20:
 615         case 26:
 616         case 27:
 617         case 29:
 618         case 30:
 619                 store_u16(&state->regs.cp2c[reg], data);
 620                 break;
 621         case 31:
 622                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
 623                 fallthrough;
 624         default:
 625                 state->regs.cp2c[reg] = data;
 626                 break;
 627         }
 628 }
 629
 630 void lightrec_mtc(struct lightrec_state *state, union code op, u8 reg, u32 data)
 631 {
 632         if (op.i.op == OP_CP0) {
 633                 lightrec_mtc0(state, reg, data);
 634         } else {
 635                 if (op.i.op == OP_LWC2 || op.r.rs != OP_CP2_BASIC_CTC2)
 636                         lightrec_mtc2(state, reg, data);
 637                 else
 638                         lightrec_ctc2(state, reg, data);
 639
 640                 if (state->ops.cop2_notify)
 641                         (*state->ops.cop2_notify)(state, op.opcode, data);
 642         }
 643 }
 644
 645 static void lightrec_mtc_cb(struct lightrec_state *state, u32 arg)
 646 {
 647         union code op = (union code) arg;
 648         u32 data;
 649         u8 reg;
 650
 651         if (op.i.op == OP_LWC2) {
 652                 data = state->temp_reg;
 653                 reg = op.i.rt;
 654         } else {
 655                 data = state->regs.gpr[op.r.rt];
 656                 reg = op.r.rd;
 657         }
 658
 659         lightrec_mtc(state, op, reg, data);
 660 }
 661
 662 void lightrec_rfe(struct lightrec_state *state)
 663 {
 664         u32 status;
 665
 666         /* Read CP0 Status register (r12) */
 667         status = state->regs.cp0[12];
 668
 669         /* Switch the bits */
 670         status = ((status & 0x3c) >> 2) | (status & ~0xf);
 671
 672         /* Write it back */
 673         lightrec_mtc0(state, 12, status);
 674 }
 675
 676 void lightrec_cp(struct lightrec_state *state, union code op)
 677 {
 678         if (op.i.op == OP_CP0) {
 679                 pr_err("Invalid CP opcode to coprocessor #0\n");
 680                 return;
 681         }
 682
 683         (*state->ops.cop2_op)(state, op.opcode);
 684 }
 685
 686 static void lightrec_cp_cb(struct lightrec_state *state, u32 arg)
 687 {
 688         lightrec_cp(state, (union code) arg);
 689 }
 690
 691 static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 692 {
 693         struct block *block = lightrec_find_block(state->block_cache, pc);
 694         u8 old_flags;
 695
 696         if (block && lightrec_block_is_outdated(state, block)) {
 697                 pr_debug("Block at "PC_FMT" is outdated!\n", block->pc);
 698
 699                 old_flags = block_set_flags(block, BLOCK_IS_DEAD);
 700                 if (!(old_flags & BLOCK_IS_DEAD)) {
 701                         /* Make sure the recompiler isn't processing the block
 702                          * we'll destroy */
 703                         if (ENABLE_THREADED_COMPILER)
 704                                 lightrec_recompiler_remove(state->rec, block);
 705
 706                         lightrec_unregister_block(state->block_cache, block);
 707                         remove_from_code_lut(state->block_cache, block);
 708                         lightrec_free_block(state, block);
 709                 }
 710
 711                 block = NULL;
 712         }
 713
 714         if (!block) {
 715                 block = lightrec_precompile_block(state, pc);
 716                 if (!block) {
 717                         pr_err("Unable to recompile block at "PC_FMT"\n", pc);
 718                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 719                         return NULL;
 720                 }
 721
 722                 lightrec_register_block(state->block_cache, block);
 723         }
 724
 725         return block;
 726 }
 727
 728 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 729 {
 730         struct block *block;
 731         bool should_recompile;
 732         void *func;
 733         int err;
 734
 735         do {
 736                 func = lut_read(state, lut_offset(pc));
 737                 if (func && func != state->get_next_block)
 738                         break;
 739
 740                 block = lightrec_get_block(state, pc);
 741
 742                 if (unlikely(!block))
 743                         break;
 744
 745                 if (OPT_REPLACE_MEMSET &&
 746                     block_has_flag(block, BLOCK_IS_MEMSET)) {
 747                         func = state->memset_func;
 748                         break;
 749                 }
 750
 751                 should_recompile = block_has_flag(block, BLOCK_SHOULD_RECOMPILE) &&
 752                         !block_has_flag(block, BLOCK_NEVER_COMPILE) &&
 753                         !block_has_flag(block, BLOCK_IS_DEAD);
 754
 755                 if (unlikely(should_recompile)) {
 756                         pr_debug("Block at "PC_FMT" should recompile\n", pc);
 757
 758                         if (ENABLE_THREADED_COMPILER) {
 759                                 lightrec_recompiler_add(state->rec, block);
 760                         } else {
 761                                 err = lightrec_compile_block(state->cstate, block);
 762                                 if (err) {
 763                                         state->exit_flags = LIGHTREC_EXIT_NOMEM;
 764                                         return NULL;
 765                                 }
 766                         }
 767                 }
 768
 769                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
 770                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
 771                 else
 772                         func = block->function;
 773
 774                 if (likely(func))
 775                         break;
 776
 777                 if (unlikely(block_has_flag(block, BLOCK_NEVER_COMPILE))) {
 778                         pc = lightrec_emulate_block(state, block, pc);
 779
 780                 } else if (!ENABLE_THREADED_COMPILER) {
 781                         /* Block wasn't compiled yet - run the interpreter */
 782                         if (block_has_flag(block, BLOCK_FULLY_TAGGED))
 783                                 pr_debug("Block fully tagged, skipping first pass\n");
 784                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 785                                 pc = lightrec_emulate_block(state, block, pc);
 786
 787                         /* Then compile it using the profiled data */
 788                         err = lightrec_compile_block(state->cstate, block);
 789                         if (err) {
 790                                 state->exit_flags = LIGHTREC_EXIT_NOMEM;
 791                                 return NULL;
 792                         }
 793                 } else if (unlikely(block_has_flag(block, BLOCK_IS_DEAD))) {
 794                         /*
 795                          * If the block is dead but has never been compiled,
 796                          * then its function pointer is NULL and we cannot
 797                          * execute the block. In that case, reap all the dead
 798                          * blocks now, and in the next loop we will create a
 799                          * new block.
 800                          */
 801                         lightrec_reaper_reap(state->reaper);
 802                 } else {
 803                         lightrec_recompiler_add(state->rec, block);
 804                 }
 805         } while (state->exit_flags == LIGHTREC_EXIT_NORMAL
 806                  && state->current_cycle < state->target_cycle);
 807
 808         state->curr_pc = pc;
 809         return func;
 810 }
 811
 812 static void * lightrec_alloc_code(struct lightrec_state *state, size_t size)
 813 {
 814         void *code;
 815
 816         if (ENABLE_THREADED_COMPILER)
 817                 lightrec_code_alloc_lock(state);
 818
 819         code = tlsf_malloc(state->tlsf, size);
 820
 821         if (ENABLE_THREADED_COMPILER)
 822                 lightrec_code_alloc_unlock(state);
 823
 824         return code;
 825 }
 826
 827 static void lightrec_realloc_code(struct lightrec_state *state,
 828                                   void *ptr, size_t size)
 829 {
 830         /* NOTE: 'size' MUST be smaller than the size specified during
 831          * the allocation. */
 832
 833         if (ENABLE_THREADED_COMPILER)
 834                 lightrec_code_alloc_lock(state);
 835
 836         tlsf_realloc(state->tlsf, ptr, size);
 837
 838         if (ENABLE_THREADED_COMPILER)
 839                 lightrec_code_alloc_unlock(state);
 840 }
 841
 842 static void lightrec_free_code(struct lightrec_state *state, void *ptr)
 843 {
 844         if (ENABLE_THREADED_COMPILER)
 845                 lightrec_code_alloc_lock(state);
 846
 847         tlsf_free(state->tlsf, ptr);
 848
 849         if (ENABLE_THREADED_COMPILER)
 850                 lightrec_code_alloc_unlock(state);
 851 }
 852
 853 static char lightning_code_data[0x80000];
 854
 855 static void * lightrec_emit_code(struct lightrec_state *state,
 856                                  const struct block *block,
 857                                  jit_state_t *_jit, unsigned int *size)
 858 {
 859         bool has_code_buffer = ENABLE_CODE_BUFFER && state->tlsf;
 860         jit_word_t code_size, new_code_size;
 861         void *code;
 862
 863         jit_realize();
 864
 865         if (ENABLE_DISASSEMBLER)
 866                 jit_set_data(lightning_code_data, sizeof(lightning_code_data), 0);
 867         else
 868                 jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
 869
 870         if (has_code_buffer) {
 871                 jit_get_code(&code_size);
 872                 code = lightrec_alloc_code(state, (size_t) code_size);
 873
 874                 if (!code) {
 875                         if (ENABLE_THREADED_COMPILER) {
 876                                 /* If we're using the threaded compiler, return
 877                                  * an allocation error here. The threaded
 878                                  * compiler will then empty its job queue and
 879                                  * request a code flush using the reaper. */
 880                                 return NULL;
 881                         }
 882
 883                         /* Remove outdated blocks, and try again */
 884                         lightrec_remove_outdated_blocks(state->block_cache, block);
 885
 886                         pr_debug("Re-try to alloc %zu bytes...\n", code_size);
 887
 888                         code = lightrec_alloc_code(state, code_size);
 889                         if (!code) {
 890                                 pr_err("Could not alloc even after removing old blocks!\n");
 891                                 return NULL;
 892                         }
 893                 }
 894
 895                 jit_set_code(code, code_size);
 896         }
 897
 898         code = jit_emit();
 899
 900         jit_get_code(&new_code_size);
 901         lightrec_register(MEM_FOR_CODE, new_code_size);
 902
 903         if (has_code_buffer) {
 904                 lightrec_realloc_code(state, code, (size_t) new_code_size);
 905
 906                 pr_debug("Creating code block at address 0x%" PRIxPTR ", "
 907                          "code size: %" PRIuPTR " new: %" PRIuPTR "\n",
 908                          (uintptr_t) code, code_size, new_code_size);
 909         }
 910
 911         *size = (unsigned int) new_code_size;
 912
 913         if (state->ops.code_inv)
 914                 state->ops.code_inv(code, new_code_size);
 915
 916         return code;
 917 }
 918
 919 static struct block * generate_wrapper(struct lightrec_state *state)
 920 {
 921         struct block *block;
 922         jit_state_t *_jit;
 923         unsigned int i;
 924         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
 925         jit_node_t *to_end[C_WRAPPERS_COUNT - 1];
 926         u8 tmp = JIT_R1;
 927
 928 #ifdef __sh__
 929         /* On SH, GBR-relative loads target the r0 register.
 930          * Use it as the temporary register to factorize the move to
 931          * JIT_R1. */
 932         if (LIGHTREC_REG_STATE == _GBR)
 933                 tmp = _R0;
 934 #endif
 935
 936         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 937         if (!block)
 938                 goto err_no_mem;
 939
 940         _jit = jit_new_state();
 941         if (!_jit)
 942                 goto err_free_block;
 943
 944         jit_name("RW wrapper");
 945         jit_note(__FILE__, __LINE__);
 946
 947         /* Wrapper entry point */
 948         jit_prolog();
 949         jit_tramp(256);
 950
 951         /* Add entry points */
 952         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
 953                 jit_ldxi(tmp, LIGHTREC_REG_STATE,
 954                          offsetof(struct lightrec_state, c_wrappers[i]));
 955                 to_end[i - 1] = jit_b();
 956                 addr[i - 1] = jit_indirect();
 957         }
 958
 959         jit_ldxi(tmp, LIGHTREC_REG_STATE,
 960                  offsetof(struct lightrec_state, c_wrappers[0]));
 961
 962         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 963                 jit_patch(to_end[i]);
 964         jit_movr(JIT_R1, tmp);
 965
 966         jit_epilog();
 967         jit_prolog();
 968
 969         /* Save all temporaries on stack */
 970         for (i = 0; i < NUM_TEMPS; i++) {
 971                 if (i + FIRST_TEMP != 1) {
 972                         jit_stxi(offsetof(struct lightrec_state, wrapper_regs[i]),
 973                                  LIGHTREC_REG_STATE, JIT_R(i + FIRST_TEMP));
 974                 }
 975         }
 976
 977         jit_getarg(JIT_R2, jit_arg());
 978
 979         jit_prepare();
 980         jit_pushargr(LIGHTREC_REG_STATE);
 981         jit_pushargr(JIT_R2);
 982
 983         jit_ldxi_ui(JIT_R2, LIGHTREC_REG_STATE,
 984                     offsetof(struct lightrec_state, target_cycle));
 985
 986         /* state->current_cycle = state->target_cycle - delta; */
 987         jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, LIGHTREC_REG_CYCLE);
 988         jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
 989                    LIGHTREC_REG_STATE, LIGHTREC_REG_CYCLE);
 990
 991         /* Call the wrapper function */
 992         jit_finishr(JIT_R1);
 993
 994         /* delta = state->target_cycle - state->current_cycle */;
 995         jit_ldxi_ui(LIGHTREC_REG_CYCLE, LIGHTREC_REG_STATE,
 996                     offsetof(struct lightrec_state, current_cycle));
 997         jit_ldxi_ui(JIT_R1, LIGHTREC_REG_STATE,
 998                     offsetof(struct lightrec_state, target_cycle));
 999         jit_subr(LIGHTREC_REG_CYCLE, JIT_R1, LIGHTREC_REG_CYCLE);
1000
1001         /* Restore temporaries from stack */
1002         for (i = 0; i < NUM_TEMPS; i++) {
1003                 if (i + FIRST_TEMP != 1) {
1004                         jit_ldxi(JIT_R(i + FIRST_TEMP), LIGHTREC_REG_STATE,
1005                                  offsetof(struct lightrec_state, wrapper_regs[i]));
1006                 }
1007         }
1008
1009         jit_ret();
1010         jit_epilog();
1011
1012         block->_jit = _jit;
1013         block->opcode_list = NULL;
1014         block->flags = BLOCK_NO_OPCODE_LIST;
1015         block->nb_ops = 0;
1016
1017         block->function = lightrec_emit_code(state, block, _jit,
1018                                              &block->code_size);
1019         if (!block->function)
1020                 goto err_free_block;
1021
1022         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
1023
1024         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
1025                 state->wrappers_eps[i] = jit_address(addr[i]);
1026
1027         if (ENABLE_DISASSEMBLER) {
1028                 pr_debug("Wrapper block:\n");
1029                 jit_disassemble();
1030         }
1031
1032         jit_clear_state();
1033         return block;
1034
1035 err_free_block:
1036         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1037 err_no_mem:
1038         pr_err("Unable to compile wrapper: Out of memory\n");
1039         return NULL;
1040 }
1041
1042 static u32 lightrec_memset(struct lightrec_state *state)
1043 {
1044         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
1045         void *host;
1046         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
1047         u32 length = state->regs.gpr[5] * 4;
1048
1049         if (!map) {
1050                 pr_err("Unable to find memory map for memset target address "PC_FMT"\n",
1051                        kunseg_pc);
1052                 return 0;
1053         }
1054
1055         pr_debug("Calling host memset, "PC_FMT" (host address 0x%"PRIxPTR") for %u bytes\n",
1056                  kunseg_pc, (uintptr_t)host, length);
1057         memset(host, 0, length);
1058
1059         if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
1060                 lightrec_invalidate_map(state, map, kunseg_pc, length);
1061
1062         /* Rough estimation of the number of cycles consumed */
1063         return 8 + 5 * (length  + 3 / 4);
1064 }
1065
1066 static u32 lightrec_check_load_delay(struct lightrec_state *state, u32 pc, u8 reg)
1067 {
1068         struct block *block;
1069         union code first_op;
1070
1071         first_op = lightrec_read_opcode(state, pc);
1072
1073         if (likely(!opcode_reads_register(first_op, reg))) {
1074                 state->regs.gpr[reg] = state->temp_reg;
1075         } else {
1076                 block = lightrec_get_block(state, pc);
1077                 if (unlikely(!block)) {
1078                         pr_err("Unable to get block at "PC_FMT"\n", pc);
1079                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
1080                         pc = 0;
1081                 } else {
1082                         pc = lightrec_handle_load_delay(state, block, pc, reg);
1083                 }
1084         }
1085
1086         return pc;
1087 }
1088
1089 static void update_cycle_counter_before_c(jit_state_t *_jit)
1090 {
1091         /* update state->current_cycle */
1092         jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1093                    offsetof(struct lightrec_state, target_cycle));
1094         jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
1095         jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
1096                    LIGHTREC_REG_STATE, JIT_R1);
1097 }
1098
1099 static void update_cycle_counter_after_c(jit_state_t *_jit)
1100 {
1101         /* Recalc the delta */
1102         jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
1103                    offsetof(struct lightrec_state, current_cycle));
1104         jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1105                    offsetof(struct lightrec_state, target_cycle));
1106         jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
1107 }
1108
1109 static void sync_next_pc(jit_state_t *_jit)
1110 {
1111         if (lightrec_store_next_pc()) {
1112                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1113                             offsetof(struct lightrec_state, next_pc));
1114         }
1115 }
1116
1117 static struct block * generate_dispatcher(struct lightrec_state *state)
1118 {
1119         struct block *block;
1120         jit_state_t *_jit;
1121         jit_node_t *to_end, *loop, *addr, *addr2, *addr3, *addr4, *addr5, *jmp, *jmp2;
1122         unsigned int i;
1123         u32 offset;
1124
1125         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1126         if (!block)
1127                 goto err_no_mem;
1128
1129         _jit = jit_new_state();
1130         if (!_jit)
1131                 goto err_free_block;
1132
1133         jit_name("dispatcher");
1134         jit_note(__FILE__, __LINE__);
1135
1136         jit_prolog();
1137         jit_frame(256);
1138
1139         jit_getarg(LIGHTREC_REG_STATE, jit_arg());
1140         jit_getarg(JIT_V0, jit_arg());
1141         jit_getarg(JIT_V1, jit_arg());
1142         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
1143
1144         /* Force all callee-saved registers to be pushed on the stack */
1145         for (i = 0; i < NUM_REGS; i++)
1146                 jit_movr(JIT_V(i + FIRST_REG), JIT_V(i + FIRST_REG));
1147
1148         loop = jit_label();
1149
1150         /* Call the block's code */
1151         jit_jmpr(JIT_V1);
1152
1153         if (OPT_REPLACE_MEMSET) {
1154                 /* Blocks will jump here when they need to call
1155                  * lightrec_memset() */
1156                 addr3 = jit_indirect();
1157
1158                 jit_movr(JIT_V1, LIGHTREC_REG_CYCLE);
1159
1160                 jit_prepare();
1161                 jit_pushargr(LIGHTREC_REG_STATE);
1162
1163                 jit_finishi(lightrec_memset);
1164                 jit_retval(LIGHTREC_REG_CYCLE);
1165
1166                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1167                             offsetof(struct lightrec_state, regs.gpr[31]));
1168                 jit_subr(LIGHTREC_REG_CYCLE, JIT_V1, LIGHTREC_REG_CYCLE);
1169
1170                 if (OPT_DETECT_IMPOSSIBLE_BRANCHES || OPT_HANDLE_LOAD_DELAYS)
1171                         jmp = jit_b();
1172         }
1173
1174         if (OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1175                 /* Blocks will jump here when they reach a branch that should
1176                  * be executed with the interpreter, passing the branch's PC
1177                  * in JIT_V0 and the address of the block in JIT_V1. */
1178                 addr4 = jit_indirect();
1179
1180                 sync_next_pc(_jit);
1181                 update_cycle_counter_before_c(_jit);
1182
1183                 jit_prepare();
1184                 jit_pushargr(LIGHTREC_REG_STATE);
1185                 jit_pushargr(JIT_V1);
1186                 jit_pushargr(JIT_V0);
1187                 jit_finishi(lightrec_emulate_block);
1188
1189                 jit_retval(JIT_V0);
1190
1191                 update_cycle_counter_after_c(_jit);
1192
1193                 if (OPT_HANDLE_LOAD_DELAYS)
1194                         jmp2 = jit_b();
1195
1196         }
1197
1198         if (OPT_HANDLE_LOAD_DELAYS) {
1199                 /* Blocks will jump here when they reach a branch with a load
1200                  * opcode in its delay slot. The delay slot has already been
1201                  * executed; the load value is in (state->temp_reg), and the
1202                  * register number is in JIT_V1.
1203                  * Jump to a C function which will evaluate the branch target's
1204                  * first opcode, to make sure that it does not read the register
1205                  * in question; and if it does, handle it accordingly. */
1206                 addr5 = jit_indirect();
1207
1208                 sync_next_pc(_jit);
1209                 update_cycle_counter_before_c(_jit);
1210
1211                 jit_prepare();
1212                 jit_pushargr(LIGHTREC_REG_STATE);
1213                 jit_pushargr(JIT_V0);
1214                 jit_pushargr(JIT_V1);
1215                 jit_finishi(lightrec_check_load_delay);
1216
1217                 jit_retval(JIT_V0);
1218
1219                 update_cycle_counter_after_c(_jit);
1220         }
1221
1222         /* The block will jump here, with the number of cycles remaining in
1223          * LIGHTREC_REG_CYCLE */
1224         addr2 = jit_indirect();
1225
1226         sync_next_pc(_jit);
1227
1228         if (OPT_HANDLE_LOAD_DELAYS && OPT_DETECT_IMPOSSIBLE_BRANCHES)
1229               jit_patch(jmp2);
1230
1231         if (OPT_REPLACE_MEMSET
1232             && (OPT_DETECT_IMPOSSIBLE_BRANCHES || OPT_HANDLE_LOAD_DELAYS)) {
1233                 jit_patch(jmp);
1234         }
1235
1236         /* Store back the next PC to the lightrec_state structure */
1237         offset = offsetof(struct lightrec_state, curr_pc);
1238         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
1239
1240         /* Jump to end if state->target_cycle < state->current_cycle */
1241         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
1242
1243         /* Convert next PC to KUNSEG and avoid mirrors */
1244         jit_andi(JIT_V1, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
1245         jit_rshi_u(JIT_R1, JIT_V1, 28);
1246         jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
1247         jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
1248         jit_movnr(JIT_V1, JIT_R2, JIT_R1);
1249
1250         /* If possible, use the code LUT */
1251         if (!lut_is_32bit(state))
1252                 jit_lshi(JIT_V1, JIT_V1, 1);
1253         jit_add_state(JIT_V1, JIT_V1);
1254
1255         offset = offsetof(struct lightrec_state, code_lut);
1256         if (lut_is_32bit(state))
1257                 jit_ldxi_ui(JIT_V1, JIT_V1, offset);
1258         else
1259                 jit_ldxi(JIT_V1, JIT_V1, offset);
1260
1261         /* If we get non-NULL, loop */
1262         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1263
1264         /* The code LUT will be set to this address when the block at the target
1265          * PC has been preprocessed but not yet compiled by the threaded
1266          * recompiler */
1267         addr = jit_indirect();
1268
1269         /* Slow path: call C function get_next_block_func() */
1270
1271         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1272                 /* We may call the interpreter - update state->current_cycle */
1273                 update_cycle_counter_before_c(_jit);
1274         }
1275
1276         jit_prepare();
1277         jit_pushargr(LIGHTREC_REG_STATE);
1278         jit_pushargr(JIT_V0);
1279
1280         /* Save the cycles register if needed */
1281         if (!(ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES))
1282                 jit_movr(JIT_V0, LIGHTREC_REG_CYCLE);
1283
1284         /* Get the next block */
1285         jit_finishi(&get_next_block_func);
1286         jit_retval(JIT_V1);
1287
1288         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1289                 /* The interpreter may have updated state->current_cycle and
1290                  * state->target_cycle - recalc the delta */
1291                 update_cycle_counter_after_c(_jit);
1292         } else {
1293                 jit_movr(LIGHTREC_REG_CYCLE, JIT_V0);
1294         }
1295
1296         /* Reset JIT_V0 to the next PC */
1297         jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1298                     offsetof(struct lightrec_state, curr_pc));
1299
1300         /* If we get non-NULL, loop */
1301         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1302
1303         /* When exiting, the recompiled code will jump to that address */
1304         jit_note(__FILE__, __LINE__);
1305         jit_patch(to_end);
1306
1307         jit_retr(LIGHTREC_REG_CYCLE);
1308         jit_epilog();
1309
1310         block->_jit = _jit;
1311         block->opcode_list = NULL;
1312         block->flags = BLOCK_NO_OPCODE_LIST;
1313         block->nb_ops = 0;
1314
1315         block->function = lightrec_emit_code(state, block, _jit,
1316                                              &block->code_size);
1317         if (!block->function)
1318                 goto err_free_block;
1319
1320         state->eob_wrapper_func = jit_address(addr2);
1321         if (OPT_DETECT_IMPOSSIBLE_BRANCHES)
1322                 state->interpreter_func = jit_address(addr4);
1323         if (OPT_HANDLE_LOAD_DELAYS)
1324                 state->ds_check_func = jit_address(addr5);
1325         if (OPT_REPLACE_MEMSET)
1326                 state->memset_func = jit_address(addr3);
1327         state->get_next_block = jit_address(addr);
1328
1329         if (ENABLE_DISASSEMBLER) {
1330                 pr_debug("Dispatcher block:\n");
1331                 jit_disassemble();
1332         }
1333
1334         /* We're done! */
1335         jit_clear_state();
1336         return block;
1337
1338 err_free_block:
1339         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1340 err_no_mem:
1341         pr_err("Unable to compile dispatcher: Out of memory\n");
1342         return NULL;
1343 }
1344
1345 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
1346 {
1347         void *host = NULL;
1348
1349         lightrec_get_map(state, &host, kunseg(pc));
1350
1351         const u32 *code = (u32 *)host;
1352         return (union code) LE32TOH(*code);
1353 }
1354
1355 unsigned int lightrec_cycles_of_opcode(const struct lightrec_state *state,
1356                                        union code code)
1357 {
1358         return state->cycles_per_op;
1359 }
1360
1361 void lightrec_free_opcode_list(struct lightrec_state *state, struct opcode *ops)
1362 {
1363         struct opcode_list *list = container_of(ops, struct opcode_list, ops);
1364
1365         lightrec_free(state, MEM_FOR_IR,
1366                       sizeof(*list) + list->nb_ops * sizeof(struct opcode),
1367                       list);
1368 }
1369
1370 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1371 {
1372         unsigned int i;
1373         union code c;
1374
1375         for (i = 1; ; i++) {
1376                 c.opcode = LE32TOH(*src++);
1377
1378                 if (is_syscall(c))
1379                         return i;
1380
1381                 if (is_unconditional_jump(c))
1382                         return i + 1;
1383         }
1384 }
1385
1386 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1387                                             const u32 *src, unsigned int *len)
1388 {
1389         struct opcode_list *list;
1390         unsigned int i, length;
1391
1392         length = lightrec_get_mips_block_len(src);
1393
1394         list = lightrec_malloc(state, MEM_FOR_IR,
1395                                sizeof(*list) + sizeof(struct opcode) * length);
1396         if (!list) {
1397                 pr_err("Unable to allocate memory\n");
1398                 return NULL;
1399         }
1400
1401         list->nb_ops = (u16) length;
1402
1403         for (i = 0; i < length; i++) {
1404                 list->ops[i].opcode = LE32TOH(src[i]);
1405                 list->ops[i].flags = 0;
1406         }
1407
1408         *len = length * sizeof(u32);
1409
1410         return list->ops;
1411 }
1412
1413 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1414                                                 u32 pc)
1415 {
1416         struct opcode *list;
1417         struct block *block;
1418         void *host, *addr;
1419         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1420         const u32 *code = (u32 *) host;
1421         unsigned int length;
1422         bool fully_tagged;
1423         u8 block_flags = 0;
1424
1425         if (!map)
1426                 return NULL;
1427
1428         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1429         if (!block) {
1430                 pr_err("Unable to recompile block: Out of memory\n");
1431                 return NULL;
1432         }
1433
1434         list = lightrec_disassemble(state, code, &length);
1435         if (!list) {
1436                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1437                 return NULL;
1438         }
1439
1440         block->pc = pc;
1441         block->_jit = NULL;
1442         block->function = NULL;
1443         block->opcode_list = list;
1444         block->code = code;
1445         block->next = NULL;
1446         block->flags = 0;
1447         block->code_size = 0;
1448         block->precompile_date = state->current_cycle;
1449         block->nb_ops = length / sizeof(u32);
1450
1451         lightrec_optimize(state, block);
1452
1453         length = block->nb_ops * sizeof(u32);
1454
1455         lightrec_register(MEM_FOR_MIPS_CODE, length);
1456
1457         if (ENABLE_DISASSEMBLER) {
1458                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1459                 lightrec_print_disassembly(block, code);
1460         }
1461
1462         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1463
1464         fully_tagged = lightrec_block_is_fully_tagged(block);
1465         if (fully_tagged)
1466                 block_flags |= BLOCK_FULLY_TAGGED;
1467
1468         if (block_flags)
1469                 block_set_flags(block, block_flags);
1470
1471         block->hash = lightrec_calculate_block_hash(block);
1472
1473         if (OPT_REPLACE_MEMSET && block_has_flag(block, BLOCK_IS_MEMSET))
1474                 addr = state->memset_func;
1475         else
1476                 addr = state->get_next_block;
1477         lut_write(state, lut_offset(pc), addr);
1478
1479         pr_debug("Blocks created: %u\n", ++state->nb_precompile);
1480
1481         return block;
1482 }
1483
1484 static bool lightrec_block_is_fully_tagged(const struct block *block)
1485 {
1486         const struct opcode *op;
1487         unsigned int i;
1488
1489         for (i = 0; i < block->nb_ops; i++) {
1490                 op = &block->opcode_list[i];
1491
1492                 /* If we have one branch that must be emulated, we cannot trash
1493                  * the opcode list. */
1494                 if (should_emulate(op))
1495                         return false;
1496
1497                 /* Check all loads/stores of the opcode list and mark the
1498                  * block as fully compiled if they all have been tagged. */
1499                 switch (op->c.i.op) {
1500                 case OP_LB:
1501                 case OP_LH:
1502                 case OP_LWL:
1503                 case OP_LW:
1504                 case OP_LBU:
1505                 case OP_LHU:
1506                 case OP_LWR:
1507                 case OP_SB:
1508                 case OP_SH:
1509                 case OP_SWL:
1510                 case OP_SW:
1511                 case OP_SWR:
1512                 case OP_LWC2:
1513                 case OP_SWC2:
1514                 case OP_META_LWU:
1515                 case OP_META_SWU:
1516                         if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
1517                                 return false;
1518                         fallthrough;
1519                 default:
1520                         continue;
1521                 }
1522         }
1523
1524         return true;
1525 }
1526
1527 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1528 {
1529         struct block *block = data;
1530
1531         pr_debug("Reap dead block at "PC_FMT"\n", block->pc);
1532         lightrec_unregister_block(state->block_cache, block);
1533         lightrec_free_block(state, block);
1534 }
1535
1536 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1537 {
1538         _jit_destroy_state(data);
1539 }
1540
1541 static void lightrec_free_function(struct lightrec_state *state, void *fn)
1542 {
1543         if (ENABLE_CODE_BUFFER && state->tlsf) {
1544                 pr_debug("Freeing code block at 0x%" PRIxPTR "\n", (uintptr_t) fn);
1545                 lightrec_free_code(state, fn);
1546         }
1547 }
1548
1549 static void lightrec_reap_function(struct lightrec_state *state, void *data)
1550 {
1551         lightrec_free_function(state, data);
1552 }
1553
1554 static void lightrec_reap_opcode_list(struct lightrec_state *state, void *data)
1555 {
1556         lightrec_free_opcode_list(state, data);
1557 }
1558
1559 int lightrec_compile_block(struct lightrec_cstate *cstate,
1560                            struct block *block)
1561 {
1562         u32 was_dead[ARRAY_SIZE(cstate->targets) / 8];
1563         struct lightrec_state *state = cstate->state;
1564         struct lightrec_branch_target *target;
1565         bool fully_tagged = false;
1566         struct block *block2;
1567         struct opcode *elm;
1568         jit_state_t *_jit, *oldjit;
1569         jit_node_t *start_of_block;
1570         bool skip_next = false;
1571         void *old_fn, *new_fn;
1572         size_t old_code_size;
1573         unsigned int i, j;
1574         u8 old_flags;
1575         u32 offset;
1576
1577         fully_tagged = lightrec_block_is_fully_tagged(block);
1578         if (fully_tagged)
1579                 block_set_flags(block, BLOCK_FULLY_TAGGED);
1580
1581         _jit = jit_new_state();
1582         if (!_jit)
1583                 return -ENOMEM;
1584
1585         oldjit = block->_jit;
1586         old_fn = block->function;
1587         old_code_size = block->code_size;
1588         block->_jit = _jit;
1589
1590         lightrec_regcache_reset(cstate->reg_cache);
1591
1592         if (OPT_PRELOAD_PC && (block->flags & BLOCK_PRELOAD_PC))
1593                 lightrec_preload_pc(cstate->reg_cache, _jit);
1594
1595         cstate->cycles = 0;
1596         cstate->nb_local_branches = 0;
1597         cstate->nb_targets = 0;
1598         cstate->no_load_delay = false;
1599
1600         jit_prolog();
1601         jit_tramp(256);
1602
1603         start_of_block = jit_label();
1604
1605         for (i = 0; i < block->nb_ops; i++) {
1606                 elm = &block->opcode_list[i];
1607
1608                 if (skip_next) {
1609                         skip_next = false;
1610                         continue;
1611                 }
1612
1613                 if (should_emulate(elm)) {
1614                         pr_debug("Branch at offset 0x%x will be emulated\n",
1615                                  i << 2);
1616
1617                         lightrec_emit_jump_to_interpreter(cstate, block, i);
1618                         skip_next = !op_flag_no_ds(elm->flags);
1619                 } else {
1620                         lightrec_rec_opcode(cstate, block, i);
1621                         skip_next = !op_flag_no_ds(elm->flags) && has_delay_slot(elm->c);
1622 #if _WIN32
1623                         /* FIXME: GNU Lightning on Windows seems to use our
1624                          * mapped registers as temporaries. Until the actual bug
1625                          * is found and fixed, unconditionally mark our
1626                          * registers as live here. */
1627                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1628 #endif
1629                 }
1630
1631                 cstate->cycles += lightrec_cycles_of_opcode(state, elm->c);
1632         }
1633
1634         for (i = 0; i < cstate->nb_local_branches; i++) {
1635                 struct lightrec_branch *branch = &cstate->local_branches[i];
1636
1637                 pr_debug("Patch local branch to offset 0x%x\n",
1638                          branch->target << 2);
1639
1640                 if (branch->target == 0) {
1641                         jit_patch_at(branch->branch, start_of_block);
1642                         continue;
1643                 }
1644
1645                 for (j = 0; j < cstate->nb_targets; j++) {
1646                         if (cstate->targets[j].offset == branch->target) {
1647                                 jit_patch_at(branch->branch,
1648                                              cstate->targets[j].label);
1649                                 break;
1650                         }
1651                 }
1652
1653                 if (j == cstate->nb_targets)
1654                         pr_err("Unable to find branch target\n");
1655         }
1656
1657         jit_ret();
1658         jit_epilog();
1659
1660         new_fn = lightrec_emit_code(state, block, _jit, &block->code_size);
1661         if (!new_fn) {
1662                 if (!ENABLE_THREADED_COMPILER)
1663                         pr_err("Unable to compile block!\n");
1664                 block->_jit = oldjit;
1665                 jit_clear_state();
1666                 _jit_destroy_state(_jit);
1667                 return -ENOMEM;
1668         }
1669
1670         /* Pause the reaper, because lightrec_reset_lut_offset() may try to set
1671          * the old block->function pointer to the code LUT. */
1672         if (ENABLE_THREADED_COMPILER)
1673                 lightrec_reaper_pause(state->reaper);
1674
1675         block->function = new_fn;
1676         block_clear_flags(block, BLOCK_SHOULD_RECOMPILE);
1677
1678         /* Add compiled function to the LUT */
1679         lut_write(state, lut_offset(block->pc), block->function);
1680
1681         /* Detect old blocks that have been covered by the new one */
1682         for (i = 0; ENABLE_THREADED_COMPILER && i < cstate->nb_targets; i++) {
1683                 target = &cstate->targets[i];
1684
1685                 if (!target->offset)
1686                         continue;
1687
1688                 offset = block->pc + target->offset * sizeof(u32);
1689
1690                 block2 = lightrec_find_block(state->block_cache, offset);
1691                 if (block2) {
1692                         /* No need to check if block2 is compilable - it must
1693                          * be, otherwise block wouldn't be compilable either */
1694
1695                         /* Set the "block dead" flag to prevent the dynarec from
1696                          * recompiling this block */
1697                         old_flags = block_set_flags(block2, BLOCK_IS_DEAD);
1698
1699                         if (old_flags & BLOCK_IS_DEAD)
1700                                 was_dead[i / 32] |= BIT(i % 32);
1701                         else
1702                                 was_dead[i / 32] &= ~BIT(i % 32);
1703                 }
1704
1705                 /* If block2 was pending for compilation, cancel it.
1706                  * If it's being compiled right now, wait until it finishes. */
1707                 if (block2)
1708                         lightrec_recompiler_remove(state->rec, block2);
1709         }
1710
1711         for (i = 0; i < cstate->nb_targets; i++) {
1712                 target = &cstate->targets[i];
1713
1714                 if (!target->offset)
1715                         continue;
1716
1717                 /* We know from now on that block2 (if present) isn't going to
1718                  * be compiled. We can override the LUT entry with our new
1719                  * block's entry point. */
1720                 offset = lut_offset(block->pc) + target->offset;
1721                 lut_write(state, offset, jit_address(target->label));
1722
1723                 offset = block->pc + target->offset * sizeof(u32);
1724                 block2 = lightrec_find_block(state->block_cache, offset);
1725                 if (block2) {
1726                         pr_debug("Reap block 0x%08x as it's covered by block "
1727                                  "0x%08x\n", block2->pc, block->pc);
1728
1729                         /* Finally, reap the block. */
1730                         if (!ENABLE_THREADED_COMPILER) {
1731                                 lightrec_unregister_block(state->block_cache, block2);
1732                                 lightrec_free_block(state, block2);
1733                         } else if (!(was_dead[i / 32] & BIT(i % 32))) {
1734                                 lightrec_reaper_add(state->reaper,
1735                                                     lightrec_reap_block,
1736                                                     block2);
1737                         }
1738                 }
1739         }
1740
1741         if (ENABLE_THREADED_COMPILER)
1742                 lightrec_reaper_continue(state->reaper);
1743
1744         if (ENABLE_DISASSEMBLER) {
1745                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1746                 jit_disassemble();
1747         }
1748
1749         jit_clear_state();
1750
1751         if (fully_tagged)
1752                 old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1753
1754         if (fully_tagged && !(old_flags & BLOCK_NO_OPCODE_LIST)) {
1755                 pr_debug("Block "PC_FMT" is fully tagged"
1756                          " - free opcode list\n", block->pc);
1757
1758                 if (ENABLE_THREADED_COMPILER) {
1759                         lightrec_reaper_add(state->reaper,
1760                                             lightrec_reap_opcode_list,
1761                                             block->opcode_list);
1762                 } else {
1763                         lightrec_free_opcode_list(state, block->opcode_list);
1764                 }
1765         }
1766
1767         if (oldjit) {
1768                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1769                          block->pc);
1770
1771                 if (ENABLE_THREADED_COMPILER) {
1772                         lightrec_reaper_add(state->reaper,
1773                                             lightrec_reap_jit, oldjit);
1774                         lightrec_reaper_add(state->reaper,
1775                                             lightrec_reap_function, old_fn);
1776                 } else {
1777                         _jit_destroy_state(oldjit);
1778                         lightrec_free_function(state, old_fn);
1779                 }
1780
1781                 lightrec_unregister(MEM_FOR_CODE, old_code_size);
1782         }
1783
1784         pr_debug("Blocks compiled: %u\n", ++state->nb_compile);
1785
1786         return 0;
1787 }
1788
1789 static void lightrec_print_info(struct lightrec_state *state)
1790 {
1791         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1792                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1793                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1794                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1795                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1796                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1797                         lightrec_get_total_mem_usage() / 1024,
1798                        lightrec_get_average_ipi());
1799                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1800         }
1801 }
1802
1803 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1804 {
1805         s32 (*func)(struct lightrec_state *, u32, void *, s32) = (void *)state->dispatcher->function;
1806         void *block_trace;
1807         s32 cycles_delta;
1808
1809         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1810
1811         /* Handle the cycle counter overflowing */
1812         if (unlikely(target_cycle < state->current_cycle))
1813                 target_cycle = UINT_MAX;
1814
1815         state->target_cycle = target_cycle;
1816         state->curr_pc = pc;
1817
1818         block_trace = get_next_block_func(state, pc);
1819         if (block_trace) {
1820                 cycles_delta = state->target_cycle - state->current_cycle;
1821
1822                 cycles_delta = (*func)(state, state->curr_pc,
1823                                        block_trace, cycles_delta);
1824
1825                 state->current_cycle = state->target_cycle - cycles_delta;
1826         }
1827
1828         if (ENABLE_THREADED_COMPILER)
1829                 lightrec_reaper_reap(state->reaper);
1830
1831         if (LOG_LEVEL >= INFO_L)
1832                 lightrec_print_info(state);
1833
1834         return state->curr_pc;
1835 }
1836
1837 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc,
1838                              u32 target_cycle)
1839 {
1840         struct block *block;
1841
1842         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1843         state->target_cycle = target_cycle;
1844
1845         do {
1846                 block = lightrec_get_block(state, pc);
1847                 if (!block)
1848                         break;
1849
1850                 pc = lightrec_emulate_block(state, block, pc);
1851
1852                 if (ENABLE_THREADED_COMPILER)
1853                         lightrec_reaper_reap(state->reaper);
1854         } while (state->current_cycle < state->target_cycle);
1855
1856         if (LOG_LEVEL >= INFO_L)
1857                 lightrec_print_info(state);
1858
1859         return pc;
1860 }
1861
1862 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1863 {
1864         u8 old_flags;
1865
1866         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1867         old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1868
1869         if (!(old_flags & BLOCK_NO_OPCODE_LIST))
1870                 lightrec_free_opcode_list(state, block->opcode_list);
1871         if (block->_jit)
1872                 _jit_destroy_state(block->_jit);
1873         if (block->function) {
1874                 lightrec_free_function(state, block->function);
1875                 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1876         }
1877         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1878 }
1879
1880 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1881 {
1882         struct lightrec_cstate *cstate;
1883
1884         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1885         if (!cstate)
1886                 return NULL;
1887
1888         cstate->reg_cache = lightrec_regcache_init(state);
1889         if (!cstate->reg_cache) {
1890                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1891                 return NULL;
1892         }
1893
1894         cstate->state = state;
1895
1896         return cstate;
1897 }
1898
1899 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1900 {
1901         lightrec_free_regcache(cstate->reg_cache);
1902         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1903 }
1904
1905 struct lightrec_state * lightrec_init(char *argv0,
1906                                       const struct lightrec_mem_map *map,
1907                                       size_t nb,
1908                                       const struct lightrec_ops *ops)
1909 {
1910         const struct lightrec_mem_map *codebuf_map = &map[PSX_MAP_CODE_BUFFER];
1911         struct lightrec_state *state;
1912         uintptr_t addr;
1913         void *tlsf = NULL;
1914         bool with_32bit_lut = false;
1915         size_t lut_size;
1916
1917         /* Sanity-check ops */
1918         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1919                 pr_err("Missing callbacks in lightrec_ops structure\n");
1920                 return NULL;
1921         }
1922
1923         if (ops->cop2_notify)
1924                 pr_debug("Optional cop2_notify callback in lightrec_ops\n");
1925         else
1926                 pr_debug("No optional cop2_notify callback in lightrec_ops\n");
1927
1928         if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER
1929             && codebuf_map->address) {
1930                 tlsf = tlsf_create_with_pool(codebuf_map->address,
1931                                              codebuf_map->length);
1932                 if (!tlsf) {
1933                         pr_err("Unable to initialize code buffer\n");
1934                         return NULL;
1935                 }
1936
1937                 if (__WORDSIZE == 64) {
1938                         addr = (uintptr_t) codebuf_map->address + codebuf_map->length - 1;
1939                         with_32bit_lut = addr == (u32) addr;
1940                 }
1941         }
1942
1943         if (with_32bit_lut)
1944                 lut_size = CODE_LUT_SIZE * 4;
1945         else
1946                 lut_size = CODE_LUT_SIZE * sizeof(void *);
1947
1948         init_jit_with_debug(argv0, stdout);
1949
1950         state = calloc(1, sizeof(*state) + lut_size);
1951         if (!state)
1952                 goto err_finish_jit;
1953
1954         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) + lut_size);
1955
1956         state->tlsf = tlsf;
1957         state->with_32bit_lut = with_32bit_lut;
1958         state->in_delay_slot_n = 0xff;
1959         state->cycles_per_op = 2;
1960
1961         state->block_cache = lightrec_blockcache_init(state);
1962         if (!state->block_cache)
1963                 goto err_free_state;
1964
1965         if (ENABLE_THREADED_COMPILER) {
1966                 state->rec = lightrec_recompiler_init(state);
1967                 if (!state->rec)
1968                         goto err_free_block_cache;
1969
1970                 state->reaper = lightrec_reaper_init(state);
1971                 if (!state->reaper)
1972                         goto err_free_recompiler;
1973         } else {
1974                 state->cstate = lightrec_create_cstate(state);
1975                 if (!state->cstate)
1976                         goto err_free_block_cache;
1977         }
1978
1979         state->nb_maps = nb;
1980         state->maps = map;
1981
1982         memcpy(&state->ops, ops, sizeof(*ops));
1983
1984         state->dispatcher = generate_dispatcher(state);
1985         if (!state->dispatcher)
1986                 goto err_free_reaper;
1987
1988         state->c_wrapper_block = generate_wrapper(state);
1989         if (!state->c_wrapper_block)
1990                 goto err_free_dispatcher;
1991
1992         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1993         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1994         state->c_wrappers[C_WRAPPER_MFC] = lightrec_mfc_cb;
1995         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1996         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
1997
1998         map = &state->maps[PSX_MAP_BIOS];
1999         state->offset_bios = (uintptr_t)map->address - map->pc;
2000
2001         map = &state->maps[PSX_MAP_SCRATCH_PAD];
2002         state->offset_scratch = (uintptr_t)map->address - map->pc;
2003
2004         map = &state->maps[PSX_MAP_HW_REGISTERS];
2005         state->offset_io = (uintptr_t)map->address - map->pc;
2006
2007         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
2008         state->offset_ram = (uintptr_t)map->address - map->pc;
2009
2010         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
2011             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
2012             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
2013                 state->mirrors_mapped = true;
2014
2015         if (state->offset_bios == 0 &&
2016             state->offset_scratch == 0 &&
2017             state->offset_ram == 0 &&
2018             state->offset_io == 0 &&
2019             state->mirrors_mapped) {
2020                 pr_info("Memory map is perfect. Emitted code will be best.\n");
2021         } else {
2022                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
2023         }
2024
2025         if (state->with_32bit_lut)
2026                 pr_info("Using 32-bit LUT\n");
2027
2028         return state;
2029
2030 err_free_dispatcher:
2031         lightrec_free_block(state, state->dispatcher);
2032 err_free_reaper:
2033         if (ENABLE_THREADED_COMPILER)
2034                 lightrec_reaper_destroy(state->reaper);
2035 err_free_recompiler:
2036         if (ENABLE_THREADED_COMPILER)
2037                 lightrec_free_recompiler(state->rec);
2038         else
2039                 lightrec_free_cstate(state->cstate);
2040 err_free_block_cache:
2041         lightrec_free_block_cache(state->block_cache);
2042 err_free_state:
2043         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
2044                             lut_elm_size(state) * CODE_LUT_SIZE);
2045         free(state);
2046 err_finish_jit:
2047         finish_jit();
2048         if (ENABLE_CODE_BUFFER && tlsf)
2049                 tlsf_destroy(tlsf);
2050         return NULL;
2051 }
2052
2053 void lightrec_destroy(struct lightrec_state *state)
2054 {
2055         /* Force a print info on destroy*/
2056         state->current_cycle = ~state->current_cycle;
2057         lightrec_print_info(state);
2058
2059         lightrec_free_block_cache(state->block_cache);
2060         lightrec_free_block(state, state->dispatcher);
2061         lightrec_free_block(state, state->c_wrapper_block);
2062
2063         if (ENABLE_THREADED_COMPILER) {
2064                 lightrec_free_recompiler(state->rec);
2065                 lightrec_reaper_destroy(state->reaper);
2066         } else {
2067                 lightrec_free_cstate(state->cstate);
2068         }
2069
2070         finish_jit();
2071         if (ENABLE_CODE_BUFFER && state->tlsf)
2072                 tlsf_destroy(state->tlsf);
2073
2074         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
2075                             lut_elm_size(state) * CODE_LUT_SIZE);
2076         free(state);
2077 }
2078
2079 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
2080 {
2081         u32 kaddr = kunseg(addr & ~0x3);
2082         enum psx_map idx = lightrec_get_map_idx(state, kaddr);
2083
2084         switch (idx) {
2085         case PSX_MAP_MIRROR1:
2086         case PSX_MAP_MIRROR2:
2087         case PSX_MAP_MIRROR3:
2088                 /* Handle mirrors */
2089                 kaddr &= RAM_SIZE - 1;
2090                 fallthrough;
2091         case PSX_MAP_KERNEL_USER_RAM:
2092                 break;
2093         default:
2094                 return;
2095         }
2096
2097         memset(lut_address(state, lut_offset(kaddr)), 0,
2098                ((len + 3) / 4) * lut_elm_size(state));
2099 }
2100
2101 void lightrec_invalidate_all(struct lightrec_state *state)
2102 {
2103         memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
2104 }
2105
2106 void lightrec_set_unsafe_opt_flags(struct lightrec_state *state, u32 flags)
2107 {
2108         if ((flags ^ state->opt_flags) & LIGHTREC_OPT_INV_DMA_ONLY)
2109                 lightrec_invalidate_all(state);
2110
2111         state->opt_flags = flags;
2112 }
2113
2114 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
2115 {
2116         if (flags != LIGHTREC_EXIT_NORMAL) {
2117                 state->exit_flags |= flags;
2118                 state->target_cycle = state->current_cycle;
2119         }
2120 }
2121
2122 u32 lightrec_exit_flags(struct lightrec_state *state)
2123 {
2124         return state->exit_flags;
2125 }
2126
2127 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
2128 {
2129         return state->current_cycle;
2130 }
2131
2132 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
2133 {
2134         state->current_cycle = cycles;
2135
2136         if (state->target_cycle < cycles)
2137                 state->target_cycle = cycles;
2138 }
2139
2140 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
2141 {
2142         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
2143                 if (cycles < state->current_cycle)
2144                         cycles = state->current_cycle;
2145
2146                 state->target_cycle = cycles;
2147         }
2148 }
2149
2150 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
2151 {
2152         return &state->regs;
2153 }
2154
2155 void lightrec_set_cycles_per_opcode(struct lightrec_state *state, u32 cycles)
2156 {
2157         state->cycles_per_op = cycles;
2158 }