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