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 #if ENABLE_TINYMM
  31 #include <tinymm.h>
  32 #endif
  33
  34 #define GENMASK(h, l) \
  35         (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
  36
  37 static struct block * lightrec_precompile_block(struct lightrec_state *state,
  38                                                 u32 pc);
  39 static bool lightrec_block_is_fully_tagged(const struct block *block);
  40
  41 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
  42 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
  43
  44 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
  45                                 void *host, u32 addr, u8 data)
  46 {
  47         *(u8 *)host = data;
  48
  49         if (!state->invalidate_from_dma_only)
  50                 lightrec_invalidate(state, addr, 1);
  51 }
  52
  53 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
  54                                 void *host, u32 addr, u16 data)
  55 {
  56         *(u16 *)host = HTOLE16(data);
  57
  58         if (!state->invalidate_from_dma_only)
  59                 lightrec_invalidate(state, addr, 2);
  60 }
  61
  62 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
  63                                 void *host, u32 addr, u32 data)
  64 {
  65         *(u32 *)host = HTOLE32(data);
  66
  67         if (!state->invalidate_from_dma_only)
  68                 lightrec_invalidate(state, addr, 4);
  69 }
  70
  71 static u8 lightrec_default_lb(struct lightrec_state *state,
  72                               u32 opcode, void *host, u32 addr)
  73 {
  74         return *(u8 *)host;
  75 }
  76
  77 static u16 lightrec_default_lh(struct lightrec_state *state,
  78                                u32 opcode, void *host, u32 addr)
  79 {
  80         return LE16TOH(*(u16 *)host);
  81 }
  82
  83 static u32 lightrec_default_lw(struct lightrec_state *state,
  84                                u32 opcode, void *host, u32 addr)
  85 {
  86         return LE32TOH(*(u32 *)host);
  87 }
  88
  89 static const struct lightrec_mem_map_ops lightrec_default_ops = {
  90         .sb = lightrec_default_sb,
  91         .sh = lightrec_default_sh,
  92         .sw = lightrec_default_sw,
  93         .lb = lightrec_default_lb,
  94         .lh = lightrec_default_lh,
  95         .lw = lightrec_default_lw,
  96 };
  97
  98 static void __segfault_cb(struct lightrec_state *state, u32 addr,
  99                           const struct block *block)
 100 {
 101         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 102         pr_err("Segmentation fault in recompiled code: invalid "
 103                "load/store at address 0x%08x\n", addr);
 104         if (block)
 105                 pr_err("Was executing block PC 0x%08x\n", block->pc);
 106 }
 107
 108 static void lightrec_swl(struct lightrec_state *state,
 109                          const struct lightrec_mem_map_ops *ops,
 110                          u32 opcode, void *host, u32 addr, u32 data)
 111 {
 112         unsigned int shift = addr & 0x3;
 113         unsigned int mask = GENMASK(31, (shift + 1) * 8);
 114         u32 old_data;
 115
 116         /* Align to 32 bits */
 117         addr &= ~3;
 118         host = (void *)((uintptr_t)host & ~3);
 119
 120         old_data = ops->lw(state, opcode, host, addr);
 121
 122         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
 123
 124         ops->sw(state, opcode, host, addr, data);
 125 }
 126
 127 static void lightrec_swr(struct lightrec_state *state,
 128                          const struct lightrec_mem_map_ops *ops,
 129                          u32 opcode, void *host, u32 addr, u32 data)
 130 {
 131         unsigned int shift = addr & 0x3;
 132         unsigned int mask = (1 << (shift * 8)) - 1;
 133         u32 old_data;
 134
 135         /* Align to 32 bits */
 136         addr &= ~3;
 137         host = (void *)((uintptr_t)host & ~3);
 138
 139         old_data = ops->lw(state, opcode, host, addr);
 140
 141         data = (data << (shift * 8)) | (old_data & mask);
 142
 143         ops->sw(state, opcode, host, addr, data);
 144 }
 145
 146 static void lightrec_swc2(struct lightrec_state *state, union code op,
 147                           const struct lightrec_mem_map_ops *ops,
 148                           void *host, u32 addr)
 149 {
 150         u32 data = lightrec_mfc2(state, op.i.rt);
 151
 152         ops->sw(state, op.opcode, host, addr, data);
 153 }
 154
 155 static u32 lightrec_lwl(struct lightrec_state *state,
 156                         const struct lightrec_mem_map_ops *ops,
 157                         u32 opcode, void *host, u32 addr, u32 data)
 158 {
 159         unsigned int shift = addr & 0x3;
 160         unsigned int mask = (1 << (24 - shift * 8)) - 1;
 161         u32 old_data;
 162
 163         /* Align to 32 bits */
 164         addr &= ~3;
 165         host = (void *)((uintptr_t)host & ~3);
 166
 167         old_data = ops->lw(state, opcode, host, addr);
 168
 169         return (data & mask) | (old_data << (24 - shift * 8));
 170 }
 171
 172 static u32 lightrec_lwr(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 = GENMASK(31, 32 - shift * 8);
 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 >> (shift * 8));
 187 }
 188
 189 static void lightrec_lwc2(struct lightrec_state *state, union code op,
 190                           const struct lightrec_mem_map_ops *ops,
 191                           void *host, u32 addr)
 192 {
 193         u32 data = ops->lw(state, op.opcode, host, addr);
 194
 195         lightrec_mtc2(state, op.i.rt, data);
 196 }
 197
 198 static void lightrec_invalidate_map(struct lightrec_state *state,
 199                 const struct lightrec_mem_map *map, u32 addr, u32 len)
 200 {
 201         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
 202                 memset(lut_address(state, lut_offset(addr)), 0,
 203                        ((len + 3) / 4) * lut_elm_size(state));
 204         }
 205 }
 206
 207 enum psx_map
 208 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
 209 {
 210         const struct lightrec_mem_map *map;
 211         unsigned int i;
 212
 213         for (i = 0; i < state->nb_maps; i++) {
 214                 map = &state->maps[i];
 215
 216                 if (kaddr >= map->pc && kaddr < map->pc + map->length)
 217                         return (enum psx_map) i;
 218         }
 219
 220         return PSX_MAP_UNKNOWN;
 221 }
 222
 223 const struct lightrec_mem_map *
 224 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 225 {
 226         const struct lightrec_mem_map *map;
 227         enum psx_map idx;
 228         u32 addr;
 229
 230         idx = lightrec_get_map_idx(state, kaddr);
 231         if (idx == PSX_MAP_UNKNOWN)
 232                 return NULL;
 233
 234         map = &state->maps[idx];
 235         addr = kaddr - map->pc;
 236
 237         while (map->mirror_of)
 238                 map = map->mirror_of;
 239
 240         if (host)
 241                 *host = map->address + addr;
 242
 243         return map;
 244 }
 245
 246 u32 lightrec_rw(struct lightrec_state *state, union code op,
 247                 u32 addr, u32 data, u16 *flags, struct block *block)
 248 {
 249         const struct lightrec_mem_map *map;
 250         const struct lightrec_mem_map_ops *ops;
 251         u32 opcode = op.opcode;
 252         void *host;
 253
 254         addr += (s16) op.i.imm;
 255
 256         map = lightrec_get_map(state, &host, kunseg(addr));
 257         if (!map) {
 258                 __segfault_cb(state, addr, block);
 259                 return 0;
 260         }
 261
 262         if (unlikely(map->ops)) {
 263                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 264                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 265
 266                 ops = map->ops;
 267         } else {
 268                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 269                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
 270
 271                 ops = &lightrec_default_ops;
 272         }
 273
 274         switch (op.i.op) {
 275         case OP_SB:
 276                 ops->sb(state, opcode, host, addr, (u8) data);
 277                 return 0;
 278         case OP_SH:
 279                 ops->sh(state, opcode, host, addr, (u16) data);
 280                 return 0;
 281         case OP_SWL:
 282                 lightrec_swl(state, ops, opcode, host, addr, data);
 283                 return 0;
 284         case OP_SWR:
 285                 lightrec_swr(state, ops, opcode, host, addr, data);
 286                 return 0;
 287         case OP_SW:
 288                 ops->sw(state, opcode, host, addr, data);
 289                 return 0;
 290         case OP_SWC2:
 291                 lightrec_swc2(state, op, ops, host, addr);
 292                 return 0;
 293         case OP_LB:
 294                 return (s32) (s8) ops->lb(state, opcode, host, addr);
 295         case OP_LBU:
 296                 return ops->lb(state, opcode, host, addr);
 297         case OP_LH:
 298                 return (s32) (s16) ops->lh(state, opcode, host, addr);
 299         case OP_LHU:
 300                 return ops->lh(state, opcode, host, addr);
 301         case OP_LWC2:
 302                 lightrec_lwc2(state, op, ops, host, addr);
 303                 return 0;
 304         case OP_LWL:
 305                 return lightrec_lwl(state, ops, opcode, host, addr, data);
 306         case OP_LWR:
 307                 return lightrec_lwr(state, ops, opcode, host, addr, data);
 308         case OP_LW:
 309         default:
 310                 return ops->lw(state, opcode, host, addr);
 311         }
 312 }
 313
 314 static void lightrec_rw_helper(struct lightrec_state *state,
 315                                union code op, u16 *flags,
 316                                struct block *block)
 317 {
 318         u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
 319                               state->regs.gpr[op.i.rt], flags, block);
 320
 321         switch (op.i.op) {
 322         case OP_LB:
 323         case OP_LBU:
 324         case OP_LH:
 325         case OP_LHU:
 326         case OP_LWL:
 327         case OP_LWR:
 328         case OP_LW:
 329                 if (op.i.rt)
 330                         state->regs.gpr[op.i.rt] = ret;
 331         default: /* fall-through */
 332                 break;
 333         }
 334 }
 335
 336 static void lightrec_rw_cb(struct lightrec_state *state)
 337 {
 338         lightrec_rw_helper(state, (union code)state->c_wrapper_arg, NULL, NULL);
 339 }
 340
 341 static void lightrec_rw_generic_cb(struct lightrec_state *state)
 342 {
 343         struct block *block;
 344         struct opcode *op;
 345         bool was_tagged;
 346         u32 arg = state->c_wrapper_arg;
 347         u16 offset = (u16)arg;
 348
 349         block = lightrec_find_block_from_lut(state->block_cache,
 350                                              arg >> 16, state->next_pc);
 351         if (unlikely(!block)) {
 352                 pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
 353                          state->next_pc, offset);
 354                 return;
 355         }
 356
 357         op = &block->opcode_list[offset];
 358         was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(op->flags);
 359
 360         lightrec_rw_helper(state, op->c, &op->flags, block);
 361
 362         if (!was_tagged) {
 363                 pr_debug("Opcode of block at PC 0x%08x has been tagged - flag "
 364                          "for recompilation\n", block->pc);
 365
 366                 block->flags |= BLOCK_SHOULD_RECOMPILE;
 367         }
 368 }
 369
 370 static u32 clamp_s32(s32 val, s32 min, s32 max)
 371 {
 372         return val < min ? min : val > max ? max : val;
 373 }
 374
 375 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
 376 {
 377         s16 gteir1, gteir2, gteir3;
 378
 379         switch (reg) {
 380         case 1:
 381         case 3:
 382         case 5:
 383         case 8:
 384         case 9:
 385         case 10:
 386         case 11:
 387                 return (s32)(s16) state->regs.cp2d[reg];
 388         case 7:
 389         case 16:
 390         case 17:
 391         case 18:
 392         case 19:
 393                 return (u16) state->regs.cp2d[reg];
 394         case 28:
 395         case 29:
 396                 gteir1 = (s16) state->regs.cp2d[9];
 397                 gteir2 = (s16) state->regs.cp2d[10];
 398                 gteir3 = (s16) state->regs.cp2d[11];
 399
 400                 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
 401                         clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
 402                         clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
 403         case 15:
 404                 reg = 14;
 405         default: /* fall-through */
 406                 return state->regs.cp2d[reg];
 407         }
 408 }
 409
 410 u32 lightrec_mfc(struct lightrec_state *state, union code op)
 411 {
 412         if (op.i.op == OP_CP0)
 413                 return state->regs.cp0[op.r.rd];
 414         else if (op.r.rs == OP_CP2_BASIC_MFC2)
 415                 return lightrec_mfc2(state, op.r.rd);
 416         else
 417                 return state->regs.cp2c[op.r.rd];
 418 }
 419
 420 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 421 {
 422         u32 status, oldstatus, cause;
 423
 424         switch (reg) {
 425         case 1:
 426         case 4:
 427         case 8:
 428         case 14:
 429         case 15:
 430                 /* Those registers are read-only */
 431                 return;
 432         default:
 433                 break;
 434         }
 435
 436         if (reg == 12) {
 437                 status = state->regs.cp0[12];
 438                 oldstatus = status;
 439
 440                 if (status & ~data & BIT(16)) {
 441                         state->ops.enable_ram(state, true);
 442                         lightrec_invalidate_all(state);
 443                 } else if (~status & data & BIT(16)) {
 444                         state->ops.enable_ram(state, false);
 445                 }
 446         }
 447
 448         if (reg == 13) {
 449                 state->regs.cp0[13] &= ~0x300;
 450                 state->regs.cp0[13] |= data & 0x300;
 451         } else {
 452                 state->regs.cp0[reg] = data;
 453         }
 454
 455         if (reg == 12 || reg == 13) {
 456                 cause = state->regs.cp0[13];
 457                 status = state->regs.cp0[12];
 458
 459                 /* Handle software interrupts */
 460                 if (!!(status & cause & 0x300) & status)
 461                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 462
 463                 /* Handle hardware interrupts */
 464                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
 465                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 466         }
 467 }
 468
 469 static u32 count_leading_bits(s32 data)
 470 {
 471         u32 cnt = 33;
 472
 473 #ifdef __has_builtin
 474 #if __has_builtin(__builtin_clrsb)
 475         return 1 + __builtin_clrsb(data);
 476 #endif
 477 #endif
 478
 479         data = (data ^ (data >> 31)) << 1;
 480
 481         do {
 482                 cnt -= 1;
 483                 data >>= 1;
 484         } while (data);
 485
 486         return cnt;
 487 }
 488
 489 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
 490 {
 491         switch (reg) {
 492         case 15:
 493                 state->regs.cp2d[12] = state->regs.cp2d[13];
 494                 state->regs.cp2d[13] = state->regs.cp2d[14];
 495                 state->regs.cp2d[14] = data;
 496                 break;
 497         case 28:
 498                 state->regs.cp2d[9] = (data << 7) & 0xf80;
 499                 state->regs.cp2d[10] = (data << 2) & 0xf80;
 500                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
 501                 break;
 502         case 31:
 503                 return;
 504         case 30:
 505                 state->regs.cp2d[31] = count_leading_bits((s32) data);
 506         default: /* fall-through */
 507                 state->regs.cp2d[reg] = data;
 508                 break;
 509         }
 510 }
 511
 512 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 513 {
 514         switch (reg) {
 515         case 4:
 516         case 12:
 517         case 20:
 518         case 26:
 519         case 27:
 520         case 29:
 521         case 30:
 522                 data = (s32)(s16) data;
 523                 break;
 524         case 31:
 525                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
 526         default: /* fall-through */
 527                 break;
 528         }
 529
 530         state->regs.cp2c[reg] = data;
 531 }
 532
 533 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
 534 {
 535         if (op.i.op == OP_CP0)
 536                 lightrec_mtc0(state, op.r.rd, data);
 537         else if (op.r.rs == OP_CP2_BASIC_CTC2)
 538                 lightrec_ctc2(state, op.r.rd, data);
 539         else
 540                 lightrec_mtc2(state, op.r.rd, data);
 541 }
 542
 543 static void lightrec_mtc_cb(struct lightrec_state *state)
 544 {
 545         union code op = (union code) state->c_wrapper_arg;
 546
 547         lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
 548 }
 549
 550 void lightrec_rfe(struct lightrec_state *state)
 551 {
 552         u32 status;
 553
 554         /* Read CP0 Status register (r12) */
 555         status = state->regs.cp0[12];
 556
 557         /* Switch the bits */
 558         status = ((status & 0x3c) >> 2) | (status & ~0xf);
 559
 560         /* Write it back */
 561         lightrec_mtc0(state, 12, status);
 562 }
 563
 564 void lightrec_cp(struct lightrec_state *state, union code op)
 565 {
 566         if (op.i.op == OP_CP0) {
 567                 pr_err("Invalid CP opcode to coprocessor #0\n");
 568                 return;
 569         }
 570
 571         (*state->ops.cop2_op)(state, op.opcode);
 572 }
 573
 574 static void lightrec_cp_cb(struct lightrec_state *state)
 575 {
 576         lightrec_cp(state, (union code) state->c_wrapper_arg);
 577 }
 578
 579 static void lightrec_syscall_cb(struct lightrec_state *state)
 580 {
 581         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
 582 }
 583
 584 static void lightrec_break_cb(struct lightrec_state *state)
 585 {
 586         lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
 587 }
 588
 589 struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 590 {
 591         struct block *block = lightrec_find_block(state->block_cache, pc);
 592
 593         if (block && lightrec_block_is_outdated(state, block)) {
 594                 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
 595
 596                 /* Make sure the recompiler isn't processing the block we'll
 597                  * destroy */
 598                 if (ENABLE_THREADED_COMPILER)
 599                         lightrec_recompiler_remove(state->rec, block);
 600
 601                 lightrec_unregister_block(state->block_cache, block);
 602                 remove_from_code_lut(state->block_cache, block);
 603                 lightrec_free_block(state, block);
 604                 block = NULL;
 605         }
 606
 607         if (!block) {
 608                 block = lightrec_precompile_block(state, pc);
 609                 if (!block) {
 610                         pr_err("Unable to recompile block at PC 0x%x\n", pc);
 611                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 612                         return NULL;
 613                 }
 614
 615                 lightrec_register_block(state->block_cache, block);
 616         }
 617
 618         return block;
 619 }
 620
 621 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 622 {
 623         struct block *block;
 624         bool should_recompile;
 625         void *func;
 626
 627         for (;;) {
 628                 func = lut_read(state, pc);
 629                 if (func && func != state->get_next_block)
 630                         break;
 631
 632                 block = lightrec_get_block(state, pc);
 633
 634                 if (unlikely(!block))
 635                         break;
 636
 637                 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
 638                         func = state->memset_func;
 639                         break;
 640                 }
 641
 642                 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
 643                         !(block->flags & BLOCK_IS_DEAD);
 644
 645                 if (unlikely(should_recompile)) {
 646                         pr_debug("Block at PC 0x%08x should recompile\n", pc);
 647
 648                         lightrec_unregister(MEM_FOR_CODE, block->code_size);
 649
 650                         if (ENABLE_THREADED_COMPILER)
 651                                 lightrec_recompiler_add(state->rec, block);
 652                         else
 653                                 lightrec_compile_block(state->cstate, block);
 654                 }
 655
 656                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
 657                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
 658                 else
 659                         func = block->function;
 660
 661                 if (likely(func))
 662                         break;
 663
 664                 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
 665                         pc = lightrec_emulate_block(state, block, pc);
 666
 667                 } else if (!ENABLE_THREADED_COMPILER) {
 668                         /* Block wasn't compiled yet - run the interpreter */
 669                         if (block->flags & BLOCK_FULLY_TAGGED)
 670                                 pr_debug("Block fully tagged, skipping first pass\n");
 671                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 672                                 pc = lightrec_emulate_block(state, block, pc);
 673
 674                         /* Then compile it using the profiled data */
 675                         lightrec_compile_block(state->cstate, block);
 676                 } else {
 677                         lightrec_recompiler_add(state->rec, block);
 678                 }
 679
 680                 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
 681                     state->current_cycle >= state->target_cycle)
 682                         break;
 683         }
 684
 685         state->next_pc = pc;
 686         return func;
 687 }
 688
 689 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
 690                               void (*f)(struct lightrec_state *))
 691 {
 692         state->current_cycle = state->target_cycle - cycles_delta;
 693
 694         (*f)(state);
 695
 696         return state->target_cycle - state->current_cycle;
 697 }
 698
 699 static void * lightrec_emit_code(struct lightrec_state *state,
 700                                  jit_state_t *_jit, unsigned int *size)
 701 {
 702         bool has_code_buffer = ENABLE_CODE_BUFFER && state->tlsf;
 703         jit_word_t code_size, new_code_size;
 704         void *code;
 705
 706         jit_realize();
 707
 708         if (!ENABLE_DISASSEMBLER)
 709                 jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
 710
 711         if (has_code_buffer) {
 712                 jit_get_code(&code_size);
 713                 code = tlsf_malloc(state->tlsf, (size_t) code_size);
 714                 if (!code)
 715                         return NULL;
 716
 717                 jit_set_code(code, code_size);
 718         }
 719
 720         code = jit_emit();
 721
 722         jit_get_code(&new_code_size);
 723         lightrec_register(MEM_FOR_CODE, new_code_size);
 724
 725         if (has_code_buffer) {
 726                 tlsf_realloc(state->tlsf, code, new_code_size);
 727
 728                 pr_debug("Creating code block at address 0x%" PRIxPTR ", "
 729                          "code size: %" PRIuPTR " new: %" PRIuPTR "\n",
 730                          (uintptr_t) code, code_size, new_code_size);
 731         }
 732
 733         *size = (unsigned int) new_code_size;
 734
 735         return code;
 736 }
 737
 738 static struct block * generate_wrapper(struct lightrec_state *state)
 739 {
 740         struct block *block;
 741         jit_state_t *_jit;
 742         unsigned int i;
 743         int stack_ptr;
 744         jit_node_t *to_tramp, *to_fn_epilog;
 745         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
 746
 747         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 748         if (!block)
 749                 goto err_no_mem;
 750
 751         _jit = jit_new_state();
 752         if (!_jit)
 753                 goto err_free_block;
 754
 755         jit_name("RW wrapper");
 756         jit_note(__FILE__, __LINE__);
 757
 758         /* Wrapper entry point */
 759         jit_prolog();
 760         jit_tramp(256);
 761
 762         /* Add entry points; separate them by opcodes that increment
 763          * LIGHTREC_REG_STATE (since we cannot touch other registers).
 764          * The difference will then tell us which C function to call. */
 765         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
 766                 jit_addi(LIGHTREC_REG_STATE, LIGHTREC_REG_STATE, __WORDSIZE / 8);
 767                 addr[i - 1] = jit_indirect();
 768         }
 769
 770         jit_epilog();
 771         jit_prolog();
 772
 773         stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
 774
 775         /* Save all temporaries on stack */
 776         for (i = 0; i < NUM_TEMPS; i++)
 777                 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
 778
 779         /* Jump to the trampoline */
 780         to_tramp = jit_jmpi();
 781
 782         /* The trampoline will jump back here */
 783         to_fn_epilog = jit_label();
 784
 785         /* Restore temporaries from stack */
 786         for (i = 0; i < NUM_TEMPS; i++)
 787                 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
 788
 789         jit_ret();
 790         jit_epilog();
 791
 792         /* Trampoline entry point.
 793          * The sole purpose of the trampoline is to cheese Lightning not to
 794          * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
 795          * do want to return to the caller with this register modified. */
 796         jit_prolog();
 797         jit_tramp(256);
 798         jit_patch(to_tramp);
 799
 800         /* Retrieve the wrapper function */
 801         jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
 802                  offsetof(struct lightrec_state, c_wrappers));
 803
 804         /* Restore LIGHTREC_REG_STATE to its correct value */
 805         jit_movi(LIGHTREC_REG_STATE, (uintptr_t) state);
 806
 807         jit_prepare();
 808         jit_pushargr(LIGHTREC_REG_STATE);
 809         jit_pushargr(LIGHTREC_REG_CYCLE);
 810         jit_pushargr(JIT_R0);
 811         jit_finishi(c_function_wrapper);
 812         jit_retval_i(LIGHTREC_REG_CYCLE);
 813
 814         jit_patch_at(jit_jmpi(), to_fn_epilog);
 815         jit_epilog();
 816
 817         block->_jit = _jit;
 818         block->opcode_list = NULL;
 819         block->flags = 0;
 820         block->nb_ops = 0;
 821
 822         block->function = lightrec_emit_code(state, _jit,
 823                                              &block->code_size);
 824         if (!block->function)
 825                 goto err_free_block;
 826
 827         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
 828
 829         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 830                 state->wrappers_eps[i] = jit_address(addr[i]);
 831
 832         if (ENABLE_DISASSEMBLER) {
 833                 pr_debug("Wrapper block:\n");
 834                 jit_disassemble();
 835         }
 836
 837         jit_clear_state();
 838         return block;
 839
 840 err_free_block:
 841         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
 842 err_no_mem:
 843         pr_err("Unable to compile wrapper: Out of memory\n");
 844         return NULL;
 845 }
 846
 847 static u32 lightrec_memset(struct lightrec_state *state)
 848 {
 849         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
 850         void *host;
 851         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
 852         u32 length = state->regs.gpr[5] * 4;
 853
 854         if (!map) {
 855                 pr_err("Unable to find memory map for memset target address "
 856                        "0x%x\n", kunseg_pc);
 857                 return 0;
 858         }
 859
 860         pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
 861                  kunseg_pc, (uintptr_t)host, length);
 862         memset(host, 0, length);
 863
 864         if (!state->invalidate_from_dma_only)
 865                 lightrec_invalidate_map(state, map, kunseg_pc, length);
 866
 867         /* Rough estimation of the number of cycles consumed */
 868         return 8 + 5 * (length  + 3 / 4);
 869 }
 870
 871 static struct block * generate_dispatcher(struct lightrec_state *state)
 872 {
 873         struct block *block;
 874         jit_state_t *_jit;
 875         jit_node_t *to_end, *loop, *addr, *addr2, *addr3;
 876         unsigned int i;
 877         u32 offset;
 878
 879         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 880         if (!block)
 881                 goto err_no_mem;
 882
 883         _jit = jit_new_state();
 884         if (!_jit)
 885                 goto err_free_block;
 886
 887         jit_name("dispatcher");
 888         jit_note(__FILE__, __LINE__);
 889
 890         jit_prolog();
 891         jit_frame(256);
 892
 893         jit_getarg(JIT_R0, jit_arg());
 894         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
 895
 896         /* Force all callee-saved registers to be pushed on the stack */
 897         for (i = 0; i < NUM_REGS; i++)
 898                 jit_movr(JIT_V(i), JIT_V(i));
 899
 900         /* Pass lightrec_state structure to blocks, using the last callee-saved
 901          * register that Lightning provides */
 902         jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
 903
 904         loop = jit_label();
 905
 906         /* Call the block's code */
 907         jit_jmpr(JIT_R0);
 908
 909         if (OPT_REPLACE_MEMSET) {
 910                 /* Blocks will jump here when they need to call
 911                  * lightrec_memset() */
 912                 addr3 = jit_indirect();
 913
 914                 jit_prepare();
 915                 jit_pushargr(LIGHTREC_REG_STATE);
 916                 jit_finishi(lightrec_memset);
 917
 918                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
 919                             offsetof(struct lightrec_state, regs.gpr[31]));
 920
 921                 jit_retval(JIT_R0);
 922                 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
 923         }
 924
 925         /* The block will jump here, with the number of cycles remaining in
 926          * LIGHTREC_REG_CYCLE */
 927         addr2 = jit_indirect();
 928
 929         /* Store back the next_pc to the lightrec_state structure */
 930         offset = offsetof(struct lightrec_state, next_pc);
 931         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
 932
 933         /* Jump to end if state->target_cycle < state->current_cycle */
 934         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
 935
 936         /* Convert next PC to KUNSEG and avoid mirrors */
 937         jit_andi(JIT_R0, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
 938         jit_rshi_u(JIT_R1, JIT_R0, 28);
 939         jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
 940         jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
 941         jit_movnr(JIT_R0, JIT_R2, JIT_R1);
 942
 943         /* If possible, use the code LUT */
 944         if (!lut_is_32bit(state))
 945                 jit_lshi(JIT_R0, JIT_R0, 1);
 946         jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
 947
 948         offset = offsetof(struct lightrec_state, code_lut);
 949         if (lut_is_32bit(state))
 950                 jit_ldxi_ui(JIT_R0, JIT_R0, offset);
 951         else
 952                 jit_ldxi(JIT_R0, JIT_R0, offset);
 953
 954         /* If we get non-NULL, loop */
 955         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
 956
 957         /* Slow path: call C function get_next_block_func() */
 958
 959         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 960                 /* We may call the interpreter - update state->current_cycle */
 961                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
 962                            offsetof(struct lightrec_state, target_cycle));
 963                 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
 964                 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
 965                            LIGHTREC_REG_STATE, JIT_R1);
 966         }
 967
 968         /* The code LUT will be set to this address when the block at the target
 969          * PC has been preprocessed but not yet compiled by the threaded
 970          * recompiler */
 971         addr = jit_indirect();
 972
 973         /* Get the next block */
 974         jit_prepare();
 975         jit_pushargr(LIGHTREC_REG_STATE);
 976         jit_pushargr(JIT_V0);
 977         jit_finishi(&get_next_block_func);
 978         jit_retval(JIT_R0);
 979
 980         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 981                 /* The interpreter may have updated state->current_cycle and
 982                  * state->target_cycle - recalc the delta */
 983                 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
 984                            offsetof(struct lightrec_state, current_cycle));
 985                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
 986                            offsetof(struct lightrec_state, target_cycle));
 987                 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
 988         }
 989
 990         /* If we get non-NULL, loop */
 991         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
 992
 993         /* When exiting, the recompiled code will jump to that address */
 994         jit_note(__FILE__, __LINE__);
 995         jit_patch(to_end);
 996
 997         jit_retr(LIGHTREC_REG_CYCLE);
 998         jit_epilog();
 999
1000         block->_jit = _jit;
1001         block->opcode_list = NULL;
1002         block->flags = 0;
1003         block->nb_ops = 0;
1004
1005         block->function = lightrec_emit_code(state, _jit,
1006                                              &block->code_size);
1007         if (!block->function)
1008                 goto err_free_block;
1009
1010         state->eob_wrapper_func = jit_address(addr2);
1011         if (OPT_REPLACE_MEMSET)
1012                 state->memset_func = jit_address(addr3);
1013         state->get_next_block = jit_address(addr);
1014
1015         if (ENABLE_DISASSEMBLER) {
1016                 pr_debug("Dispatcher block:\n");
1017                 jit_disassemble();
1018         }
1019
1020         /* We're done! */
1021         jit_clear_state();
1022         return block;
1023
1024 err_free_block:
1025         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1026 err_no_mem:
1027         pr_err("Unable to compile dispatcher: Out of memory\n");
1028         return NULL;
1029 }
1030
1031 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
1032 {
1033         void *host = NULL;
1034
1035         lightrec_get_map(state, &host, kunseg(pc));
1036
1037         const u32 *code = (u32 *)host;
1038         return (union code) LE32TOH(*code);
1039 }
1040
1041 unsigned int lightrec_cycles_of_opcode(union code code)
1042 {
1043         return 2;
1044 }
1045
1046 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
1047 {
1048         lightrec_free(state, MEM_FOR_IR,
1049                       sizeof(*block->opcode_list) * block->nb_ops,
1050                       block->opcode_list);
1051 }
1052
1053 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1054 {
1055         unsigned int i;
1056         union code c;
1057
1058         for (i = 1; ; i++) {
1059                 c.opcode = LE32TOH(*src++);
1060
1061                 if (is_syscall(c))
1062                         return i;
1063
1064                 if (is_unconditional_jump(c))
1065                         return i + 1;
1066         }
1067 }
1068
1069 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1070                                             const u32 *src, unsigned int *len)
1071 {
1072         struct opcode *list;
1073         unsigned int i, length;
1074
1075         length = lightrec_get_mips_block_len(src);
1076
1077         list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
1078         if (!list) {
1079                 pr_err("Unable to allocate memory\n");
1080                 return NULL;
1081         }
1082
1083         for (i = 0; i < length; i++) {
1084                 list[i].opcode = LE32TOH(src[i]);
1085                 list[i].flags = 0;
1086         }
1087
1088         *len = length * sizeof(u32);
1089
1090         return list;
1091 }
1092
1093 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1094                                                 u32 pc)
1095 {
1096         struct opcode *list;
1097         struct block *block;
1098         void *host;
1099         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1100         const u32 *code = (u32 *) host;
1101         unsigned int length;
1102         bool fully_tagged;
1103
1104         if (!map)
1105                 return NULL;
1106
1107         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1108         if (!block) {
1109                 pr_err("Unable to recompile block: Out of memory\n");
1110                 return NULL;
1111         }
1112
1113         list = lightrec_disassemble(state, code, &length);
1114         if (!list) {
1115                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1116                 return NULL;
1117         }
1118
1119         block->pc = pc;
1120         block->_jit = NULL;
1121         block->function = NULL;
1122         block->opcode_list = list;
1123         block->code = code;
1124         block->next = NULL;
1125         block->flags = 0;
1126         block->code_size = 0;
1127 #if ENABLE_THREADED_COMPILER
1128         block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1129 #endif
1130         block->nb_ops = length / sizeof(u32);
1131
1132         lightrec_optimize(state, block);
1133
1134         length = block->nb_ops * sizeof(u32);
1135
1136         lightrec_register(MEM_FOR_MIPS_CODE, length);
1137
1138         if (ENABLE_DISASSEMBLER) {
1139                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1140                 lightrec_print_disassembly(block, code);
1141         }
1142
1143         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1144
1145         /* If the first opcode is an 'impossible' branch, never compile the
1146          * block */
1147         if (should_emulate(block->opcode_list))
1148                 block->flags |= BLOCK_NEVER_COMPILE;
1149
1150         fully_tagged = lightrec_block_is_fully_tagged(block);
1151         if (fully_tagged)
1152                 block->flags |= BLOCK_FULLY_TAGGED;
1153
1154         if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1155                 lut_write(state, lut_offset(pc), state->memset_func);
1156
1157         block->hash = lightrec_calculate_block_hash(block);
1158
1159         pr_debug("Recompile count: %u\n", state->nb_precompile++);
1160
1161         return block;
1162 }
1163
1164 static bool lightrec_block_is_fully_tagged(const struct block *block)
1165 {
1166         const struct opcode *op;
1167         unsigned int i;
1168
1169         for (i = 0; i < block->nb_ops; i++) {
1170                 op = &block->opcode_list[i];
1171
1172                 /* Verify that all load/stores of the opcode list
1173                  * Check all loads/stores of the opcode list and mark the
1174                  * block as fully compiled if they all have been tagged. */
1175                 switch (op->c.i.op) {
1176                 case OP_LB:
1177                 case OP_LH:
1178                 case OP_LWL:
1179                 case OP_LW:
1180                 case OP_LBU:
1181                 case OP_LHU:
1182                 case OP_LWR:
1183                 case OP_SB:
1184                 case OP_SH:
1185                 case OP_SWL:
1186                 case OP_SW:
1187                 case OP_SWR:
1188                 case OP_LWC2:
1189                 case OP_SWC2:
1190                         if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
1191                                 return false;
1192                 default: /* fall-through */
1193                         continue;
1194                 }
1195         }
1196
1197         return true;
1198 }
1199
1200 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1201 {
1202         struct block *block = data;
1203
1204         pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1205         lightrec_unregister_block(state->block_cache, block);
1206         lightrec_free_block(state, block);
1207 }
1208
1209 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1210 {
1211         _jit_destroy_state(data);
1212 }
1213
1214 static void lightrec_free_function(struct lightrec_state *state, void *fn)
1215 {
1216         if (ENABLE_CODE_BUFFER && state->tlsf) {
1217                 pr_debug("Freeing code block at 0x%" PRIxPTR "\n", (uintptr_t) fn);
1218                 tlsf_free(state->tlsf, fn);
1219         }
1220 }
1221
1222 static void lightrec_reap_function(struct lightrec_state *state, void *data)
1223 {
1224         lightrec_free_function(state, data);
1225 }
1226
1227 int lightrec_compile_block(struct lightrec_cstate *cstate,
1228                            struct block *block)
1229 {
1230         struct lightrec_state *state = cstate->state;
1231         struct lightrec_branch_target *target;
1232         bool op_list_freed = false, fully_tagged = false;
1233         struct block *block2;
1234         struct opcode *elm;
1235         jit_state_t *_jit, *oldjit;
1236         jit_node_t *start_of_block;
1237         bool skip_next = false;
1238         void *old_fn;
1239         unsigned int i, j;
1240         u32 offset;
1241
1242         fully_tagged = lightrec_block_is_fully_tagged(block);
1243         if (fully_tagged)
1244                 block->flags |= BLOCK_FULLY_TAGGED;
1245
1246         _jit = jit_new_state();
1247         if (!_jit)
1248                 return -ENOMEM;
1249
1250         oldjit = block->_jit;
1251         old_fn = block->function;
1252         block->_jit = _jit;
1253
1254         lightrec_regcache_reset(cstate->reg_cache);
1255         cstate->cycles = 0;
1256         cstate->nb_branches = 0;
1257         cstate->nb_local_branches = 0;
1258         cstate->nb_targets = 0;
1259
1260         jit_prolog();
1261         jit_tramp(256);
1262
1263         start_of_block = jit_label();
1264
1265         for (i = 0; i < block->nb_ops; i++) {
1266                 elm = &block->opcode_list[i];
1267
1268                 if (skip_next) {
1269                         skip_next = false;
1270                         continue;
1271                 }
1272
1273                 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1274
1275                 if (should_emulate(elm)) {
1276                         pr_debug("Branch at offset 0x%x will be emulated\n",
1277                                  i << 2);
1278
1279                         lightrec_emit_eob(cstate, block, i, false);
1280                         skip_next = !(elm->flags & LIGHTREC_NO_DS);
1281                 } else {
1282                         lightrec_rec_opcode(cstate, block, i);
1283                         skip_next = has_delay_slot(elm->c) &&
1284                                 !(elm->flags & LIGHTREC_NO_DS);
1285 #if _WIN32
1286                         /* FIXME: GNU Lightning on Windows seems to use our
1287                          * mapped registers as temporaries. Until the actual bug
1288                          * is found and fixed, unconditionally mark our
1289                          * registers as live here. */
1290                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1291 #endif
1292                 }
1293         }
1294
1295         for (i = 0; i < cstate->nb_branches; i++)
1296                 jit_patch(cstate->branches[i]);
1297
1298         for (i = 0; i < cstate->nb_local_branches; i++) {
1299                 struct lightrec_branch *branch = &cstate->local_branches[i];
1300
1301                 pr_debug("Patch local branch to offset 0x%x\n",
1302                          branch->target << 2);
1303
1304                 if (branch->target == 0) {
1305                         jit_patch_at(branch->branch, start_of_block);
1306                         continue;
1307                 }
1308
1309                 for (j = 0; j < cstate->nb_targets; j++) {
1310                         if (cstate->targets[j].offset == branch->target) {
1311                                 jit_patch_at(branch->branch,
1312                                              cstate->targets[j].label);
1313                                 break;
1314                         }
1315                 }
1316
1317                 if (j == cstate->nb_targets)
1318                         pr_err("Unable to find branch target\n");
1319         }
1320
1321         jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
1322                  offsetof(struct lightrec_state, eob_wrapper_func));
1323
1324         jit_jmpr(JIT_R0);
1325
1326         jit_ret();
1327         jit_epilog();
1328
1329         block->function = lightrec_emit_code(state, _jit,
1330                                              &block->code_size);
1331         if (!block->function) {
1332                 pr_err("Unable to compile block!\n");
1333         }
1334
1335         block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1336
1337         /* Add compiled function to the LUT */
1338         lut_write(state, lut_offset(block->pc), block->function);
1339
1340         if (ENABLE_THREADED_COMPILER) {
1341                 /* Since we might try to reap the same block multiple times,
1342                  * we need the reaper to wait until everything has been
1343                  * submitted, so that the duplicate entries can be dropped. */
1344                 lightrec_reaper_pause(state->reaper);
1345         }
1346
1347         /* Detect old blocks that have been covered by the new one */
1348         for (i = 0; i < cstate->nb_targets; i++) {
1349                 target = &cstate->targets[i];
1350
1351                 if (!target->offset)
1352                         continue;
1353
1354                 offset = block->pc + target->offset * sizeof(u32);
1355                 block2 = lightrec_find_block(state->block_cache, offset);
1356                 if (block2) {
1357                         /* No need to check if block2 is compilable - it must
1358                          * be, otherwise block wouldn't be compilable either */
1359
1360                         /* Set the "block dead" flag to prevent the dynarec from
1361                          * recompiling this block */
1362                         block2->flags |= BLOCK_IS_DEAD;
1363
1364                         /* If block2 was pending for compilation, cancel it.
1365                          * If it's being compiled right now, wait until it
1366                          * finishes. */
1367                         if (ENABLE_THREADED_COMPILER)
1368                                 lightrec_recompiler_remove(state->rec, block2);
1369                 }
1370
1371                 /* We know from now on that block2 (if present) isn't going to
1372                  * be compiled. We can override the LUT entry with our new
1373                  * block's entry point. */
1374                 offset = lut_offset(block->pc) + target->offset;
1375                 lut_write(state, offset, jit_address(target->label));
1376
1377                 if (block2) {
1378                         pr_debug("Reap block 0x%08x as it's covered by block "
1379                                  "0x%08x\n", block2->pc, block->pc);
1380
1381                         /* Finally, reap the block. */
1382                         if (ENABLE_THREADED_COMPILER) {
1383                                 lightrec_reaper_add(state->reaper,
1384                                                     lightrec_reap_block,
1385                                                     block2);
1386                         } else {
1387                                 lightrec_unregister_block(state->block_cache, block2);
1388                                 lightrec_free_block(state, block2);
1389                         }
1390                 }
1391         }
1392
1393         if (ENABLE_THREADED_COMPILER)
1394                 lightrec_reaper_continue(state->reaper);
1395
1396         if (ENABLE_DISASSEMBLER) {
1397                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1398                 jit_disassemble();
1399         }
1400
1401         jit_clear_state();
1402
1403 #if ENABLE_THREADED_COMPILER
1404         if (fully_tagged)
1405                 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1406 #endif
1407         if (fully_tagged && !op_list_freed) {
1408                 pr_debug("Block PC 0x%08x is fully tagged"
1409                          " - free opcode list\n", block->pc);
1410                 lightrec_free_opcode_list(state, block);
1411                 block->opcode_list = NULL;
1412         }
1413
1414         if (oldjit) {
1415                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1416                          block->pc);
1417
1418                 if (ENABLE_THREADED_COMPILER) {
1419                         lightrec_reaper_add(state->reaper,
1420                                             lightrec_reap_jit, oldjit);
1421                         lightrec_reaper_add(state->reaper,
1422                                             lightrec_reap_function, old_fn);
1423                 } else {
1424                         _jit_destroy_state(oldjit);
1425                         lightrec_free_function(state, old_fn);
1426                 }
1427         }
1428
1429         return 0;
1430 }
1431
1432 static void lightrec_print_info(struct lightrec_state *state)
1433 {
1434         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1435                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1436                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1437                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1438                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1439                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1440                         lightrec_get_total_mem_usage() / 1024,
1441                        lightrec_get_average_ipi());
1442                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1443         }
1444 }
1445
1446 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1447 {
1448         s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1449         void *block_trace;
1450         s32 cycles_delta;
1451
1452         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1453
1454         /* Handle the cycle counter overflowing */
1455         if (unlikely(target_cycle < state->current_cycle))
1456                 target_cycle = UINT_MAX;
1457
1458         state->target_cycle = target_cycle;
1459         state->next_pc = pc;
1460
1461         block_trace = get_next_block_func(state, pc);
1462         if (block_trace) {
1463                 cycles_delta = state->target_cycle - state->current_cycle;
1464
1465                 cycles_delta = (*func)(block_trace, cycles_delta);
1466
1467                 state->current_cycle = state->target_cycle - cycles_delta;
1468         }
1469
1470         if (ENABLE_THREADED_COMPILER)
1471                 lightrec_reaper_reap(state->reaper);
1472
1473         if (LOG_LEVEL >= INFO_L)
1474                 lightrec_print_info(state);
1475
1476         return state->next_pc;
1477 }
1478
1479 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1480 {
1481         return lightrec_execute(state, pc, state->current_cycle);
1482 }
1483
1484 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1485 {
1486         struct block *block = lightrec_get_block(state, pc);
1487         if (!block)
1488                 return 0;
1489
1490         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1491
1492         pc = lightrec_emulate_block(state, block, pc);
1493
1494         if (LOG_LEVEL >= INFO_L)
1495                 lightrec_print_info(state);
1496
1497         return pc;
1498 }
1499
1500 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1501 {
1502         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1503         if (block->opcode_list)
1504                 lightrec_free_opcode_list(state, block);
1505         if (block->_jit)
1506                 _jit_destroy_state(block->_jit);
1507         lightrec_free_function(state, block->function);
1508         lightrec_unregister(MEM_FOR_CODE, block->code_size);
1509         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1510 }
1511
1512 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1513 {
1514         struct lightrec_cstate *cstate;
1515
1516         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1517         if (!cstate)
1518                 return NULL;
1519
1520         cstate->reg_cache = lightrec_regcache_init(state);
1521         if (!cstate->reg_cache) {
1522                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1523                 return NULL;
1524         }
1525
1526         cstate->state = state;
1527
1528         return cstate;
1529 }
1530
1531 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1532 {
1533         lightrec_free_regcache(cstate->reg_cache);
1534         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1535 }
1536
1537 struct lightrec_state * lightrec_init(char *argv0,
1538                                       const struct lightrec_mem_map *map,
1539                                       size_t nb,
1540                                       const struct lightrec_ops *ops)
1541 {
1542         const struct lightrec_mem_map *codebuf_map;
1543         struct lightrec_state *state;
1544         uintptr_t addr;
1545         void *tlsf = NULL;
1546         bool with_32bit_lut = false;
1547         size_t lut_size;
1548
1549         /* Sanity-check ops */
1550         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1551                 pr_err("Missing callbacks in lightrec_ops structure\n");
1552                 return NULL;
1553         }
1554
1555         if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER) {
1556                 codebuf_map = &map[PSX_MAP_CODE_BUFFER];
1557
1558                 tlsf = tlsf_create_with_pool(codebuf_map->address,
1559                                              codebuf_map->length);
1560                 if (!tlsf) {
1561                         pr_err("Unable to initialize code buffer\n");
1562                         return NULL;
1563                 }
1564
1565                 if (__WORDSIZE == 64) {
1566                         addr = (uintptr_t) codebuf_map->address + codebuf_map->length - 1;
1567                         with_32bit_lut = addr == (u32) addr;
1568                 }
1569         }
1570
1571         if (with_32bit_lut)
1572                 lut_size = CODE_LUT_SIZE * 4;
1573         else
1574                 lut_size = CODE_LUT_SIZE * sizeof(void *);
1575
1576         init_jit(argv0);
1577
1578         state = calloc(1, sizeof(*state) + lut_size);
1579         if (!state)
1580                 goto err_finish_jit;
1581
1582         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) + lut_size);
1583
1584         state->tlsf = tlsf;
1585         state->with_32bit_lut = with_32bit_lut;
1586
1587 #if ENABLE_TINYMM
1588         state->tinymm = tinymm_init(malloc, free, 4096);
1589         if (!state->tinymm)
1590                 goto err_free_state;
1591 #endif
1592
1593         state->block_cache = lightrec_blockcache_init(state);
1594         if (!state->block_cache)
1595                 goto err_free_tinymm;
1596
1597         if (ENABLE_THREADED_COMPILER) {
1598                 state->rec = lightrec_recompiler_init(state);
1599                 if (!state->rec)
1600                         goto err_free_block_cache;
1601
1602                 state->reaper = lightrec_reaper_init(state);
1603                 if (!state->reaper)
1604                         goto err_free_recompiler;
1605         } else {
1606                 state->cstate = lightrec_create_cstate(state);
1607                 if (!state->cstate)
1608                         goto err_free_block_cache;
1609         }
1610
1611         state->nb_maps = nb;
1612         state->maps = map;
1613
1614         memcpy(&state->ops, ops, sizeof(*ops));
1615
1616         state->dispatcher = generate_dispatcher(state);
1617         if (!state->dispatcher)
1618                 goto err_free_reaper;
1619
1620         state->c_wrapper_block = generate_wrapper(state);
1621         if (!state->c_wrapper_block)
1622                 goto err_free_dispatcher;
1623
1624         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1625         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1626         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1627         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
1628         state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1629         state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1630
1631         map = &state->maps[PSX_MAP_BIOS];
1632         state->offset_bios = (uintptr_t)map->address - map->pc;
1633
1634         map = &state->maps[PSX_MAP_SCRATCH_PAD];
1635         state->offset_scratch = (uintptr_t)map->address - map->pc;
1636
1637         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1638         state->offset_ram = (uintptr_t)map->address - map->pc;
1639
1640         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1641             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1642             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1643                 state->mirrors_mapped = true;
1644
1645         if (state->offset_bios == 0 &&
1646             state->offset_scratch == 0 &&
1647             state->offset_ram == 0 &&
1648             state->mirrors_mapped) {
1649                 pr_info("Memory map is perfect. Emitted code will be best.\n");
1650         } else {
1651                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1652         }
1653
1654         if (state->with_32bit_lut)
1655                 pr_info("Using 32-bit LUT\n");
1656
1657         return state;
1658
1659 err_free_dispatcher:
1660         lightrec_free_block(state, state->dispatcher);
1661 err_free_reaper:
1662         if (ENABLE_THREADED_COMPILER)
1663                 lightrec_reaper_destroy(state->reaper);
1664 err_free_recompiler:
1665         if (ENABLE_THREADED_COMPILER)
1666                 lightrec_free_recompiler(state->rec);
1667         else
1668                 lightrec_free_cstate(state->cstate);
1669 err_free_block_cache:
1670         lightrec_free_block_cache(state->block_cache);
1671 err_free_tinymm:
1672 #if ENABLE_TINYMM
1673         tinymm_shutdown(state->tinymm);
1674 err_free_state:
1675 #endif
1676         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1677                             lut_elm_size(state) * CODE_LUT_SIZE);
1678         free(state);
1679 err_finish_jit:
1680         finish_jit();
1681         if (ENABLE_CODE_BUFFER && tlsf)
1682                 tlsf_destroy(tlsf);
1683         return NULL;
1684 }
1685
1686 void lightrec_destroy(struct lightrec_state *state)
1687 {
1688         /* Force a print info on destroy*/
1689         state->current_cycle = ~state->current_cycle;
1690         lightrec_print_info(state);
1691
1692         if (ENABLE_THREADED_COMPILER) {
1693                 lightrec_free_recompiler(state->rec);
1694                 lightrec_reaper_destroy(state->reaper);
1695         } else {
1696                 lightrec_free_cstate(state->cstate);
1697         }
1698
1699         lightrec_free_block_cache(state->block_cache);
1700         lightrec_free_block(state, state->dispatcher);
1701         lightrec_free_block(state, state->c_wrapper_block);
1702         finish_jit();
1703         if (ENABLE_CODE_BUFFER && state->tlsf)
1704                 tlsf_destroy(state->tlsf);
1705
1706 #if ENABLE_TINYMM
1707         tinymm_shutdown(state->tinymm);
1708 #endif
1709         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1710                             lut_elm_size(state) * CODE_LUT_SIZE);
1711         free(state);
1712 }
1713
1714 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1715 {
1716         u32 kaddr = kunseg(addr & ~0x3);
1717         const struct lightrec_mem_map *map = lightrec_get_map(state, NULL, kaddr);
1718
1719         if (map) {
1720                 if (map != &state->maps[PSX_MAP_KERNEL_USER_RAM])
1721                         return;
1722
1723                 /* Handle mirrors */
1724                 kaddr &= (state->maps[PSX_MAP_KERNEL_USER_RAM].length - 1);
1725
1726                 lightrec_invalidate_map(state, map, kaddr, len);
1727         }
1728 }
1729
1730 void lightrec_invalidate_all(struct lightrec_state *state)
1731 {
1732         memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
1733 }
1734
1735 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1736 {
1737         if (state->invalidate_from_dma_only != dma_only)
1738                 lightrec_invalidate_all(state);
1739
1740         state->invalidate_from_dma_only = dma_only;
1741 }
1742
1743 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1744 {
1745         if (flags != LIGHTREC_EXIT_NORMAL) {
1746                 state->exit_flags |= flags;
1747                 state->target_cycle = state->current_cycle;
1748         }
1749 }
1750
1751 u32 lightrec_exit_flags(struct lightrec_state *state)
1752 {
1753         return state->exit_flags;
1754 }
1755
1756 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1757 {
1758         return state->current_cycle;
1759 }
1760
1761 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1762 {
1763         state->current_cycle = cycles;
1764
1765         if (state->target_cycle < cycles)
1766                 state->target_cycle = cycles;
1767 }
1768
1769 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1770 {
1771         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1772                 if (cycles < state->current_cycle)
1773                         cycles = state->current_cycle;
1774
1775                 state->target_cycle = cycles;
1776         }
1777 }
1778
1779 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1780 {
1781         return &state->regs;
1782 }