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