deps/lightrec/lightrec.c

   1 // SPDX-License-Identifier: LGPL-2.1-or-later
   2 /*
   3  * Copyright (C) 2014-2021 Paul Cercueil <paul@crapouillou.net>
   4  */
   5
   6 #include "blockcache.h"
   7 #include "debug.h"
   8 #include "disassembler.h"
   9 #include "emitter.h"
  10 #include "interpreter.h"
  11 #include "lightrec-config.h"
  12 #include "lightning-wrapper.h"
  13 #include "lightrec.h"
  14 #include "memmanager.h"
  15 #include "reaper.h"
  16 #include "recompiler.h"
  17 #include "regcache.h"
  18 #include "optimizer.h"
  19 #include "tlsf/tlsf.h"
  20
  21 #include <errno.h>
  22 #include <inttypes.h>
  23 #include <limits.h>
  24 #if ENABLE_THREADED_COMPILER
  25 #include <stdatomic.h>
  26 #endif
  27 #include <stdbool.h>
  28 #include <stddef.h>
  29 #include <string.h>
  30
  31 static struct block * lightrec_precompile_block(struct lightrec_state *state,
  32                                                 u32 pc);
  33 static bool lightrec_block_is_fully_tagged(const struct block *block);
  34
  35 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
  36 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
  37
  38 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
  39                                 void *host, u32 addr, u8 data)
  40 {
  41         *(u8 *)host = data;
  42
  43         if (!state->invalidate_from_dma_only)
  44                 lightrec_invalidate(state, addr, 1);
  45 }
  46
  47 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
  48                                 void *host, u32 addr, u16 data)
  49 {
  50         *(u16 *)host = HTOLE16(data);
  51
  52         if (!state->invalidate_from_dma_only)
  53                 lightrec_invalidate(state, addr, 2);
  54 }
  55
  56 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
  57                                 void *host, u32 addr, u32 data)
  58 {
  59         *(u32 *)host = HTOLE32(data);
  60
  61         if (!state->invalidate_from_dma_only)
  62                 lightrec_invalidate(state, addr, 4);
  63 }
  64
  65 static u8 lightrec_default_lb(struct lightrec_state *state,
  66                               u32 opcode, void *host, u32 addr)
  67 {
  68         return *(u8 *)host;
  69 }
  70
  71 static u16 lightrec_default_lh(struct lightrec_state *state,
  72                                u32 opcode, void *host, u32 addr)
  73 {
  74         return LE16TOH(*(u16 *)host);
  75 }
  76
  77 static u32 lightrec_default_lw(struct lightrec_state *state,
  78                                u32 opcode, void *host, u32 addr)
  79 {
  80         return LE32TOH(*(u32 *)host);
  81 }
  82
  83 static const struct lightrec_mem_map_ops lightrec_default_ops = {
  84         .sb = lightrec_default_sb,
  85         .sh = lightrec_default_sh,
  86         .sw = lightrec_default_sw,
  87         .lb = lightrec_default_lb,
  88         .lh = lightrec_default_lh,
  89         .lw = lightrec_default_lw,
  90 };
  91
  92 static void __segfault_cb(struct lightrec_state *state, u32 addr,
  93                           const struct block *block)
  94 {
  95         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
  96         pr_err("Segmentation fault in recompiled code: invalid "
  97                "load/store at address 0x%08x\n", addr);
  98         if (block)
  99                 pr_err("Was executing block PC 0x%08x\n", block->pc);
 100 }
 101
 102 static void lightrec_swl(struct lightrec_state *state,
 103                          const struct lightrec_mem_map_ops *ops,
 104                          u32 opcode, void *host, u32 addr, u32 data)
 105 {
 106         unsigned int shift = addr & 0x3;
 107         unsigned int mask = shift < 3 ? GENMASK(31, (shift + 1) * 8) : 0;
 108         u32 old_data;
 109
 110         /* Align to 32 bits */
 111         addr &= ~3;
 112         host = (void *)((uintptr_t)host & ~3);
 113
 114         old_data = ops->lw(state, opcode, host, addr);
 115
 116         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
 117
 118         ops->sw(state, opcode, host, addr, data);
 119 }
 120
 121 static void lightrec_swr(struct lightrec_state *state,
 122                          const struct lightrec_mem_map_ops *ops,
 123                          u32 opcode, void *host, u32 addr, u32 data)
 124 {
 125         unsigned int shift = addr & 0x3;
 126         unsigned int mask = (1 << (shift * 8)) - 1;
 127         u32 old_data;
 128
 129         /* Align to 32 bits */
 130         addr &= ~3;
 131         host = (void *)((uintptr_t)host & ~3);
 132
 133         old_data = ops->lw(state, opcode, host, addr);
 134
 135         data = (data << (shift * 8)) | (old_data & mask);
 136
 137         ops->sw(state, opcode, host, addr, data);
 138 }
 139
 140 static void lightrec_swc2(struct lightrec_state *state, union code op,
 141                           const struct lightrec_mem_map_ops *ops,
 142                           void *host, u32 addr)
 143 {
 144         u32 data = lightrec_mfc2(state, op.i.rt);
 145
 146         ops->sw(state, op.opcode, host, addr, data);
 147 }
 148
 149 static u32 lightrec_lwl(struct lightrec_state *state,
 150                         const struct lightrec_mem_map_ops *ops,
 151                         u32 opcode, void *host, u32 addr, u32 data)
 152 {
 153         unsigned int shift = addr & 0x3;
 154         unsigned int mask = (1 << (24 - shift * 8)) - 1;
 155         u32 old_data;
 156
 157         /* Align to 32 bits */
 158         addr &= ~3;
 159         host = (void *)((uintptr_t)host & ~3);
 160
 161         old_data = ops->lw(state, opcode, host, addr);
 162
 163         return (data & mask) | (old_data << (24 - shift * 8));
 164 }
 165
 166 static u32 lightrec_lwr(struct lightrec_state *state,
 167                         const struct lightrec_mem_map_ops *ops,
 168                         u32 opcode, void *host, u32 addr, u32 data)
 169 {
 170         unsigned int shift = addr & 0x3;
 171         unsigned int mask = shift ? GENMASK(31, 32 - shift * 8) : 0;
 172         u32 old_data;
 173
 174         /* Align to 32 bits */
 175         addr &= ~3;
 176         host = (void *)((uintptr_t)host & ~3);
 177
 178         old_data = ops->lw(state, opcode, host, addr);
 179
 180         return (data & mask) | (old_data >> (shift * 8));
 181 }
 182
 183 static void lightrec_lwc2(struct lightrec_state *state, union code op,
 184                           const struct lightrec_mem_map_ops *ops,
 185                           void *host, u32 addr)
 186 {
 187         u32 data = ops->lw(state, op.opcode, host, addr);
 188
 189         lightrec_mtc2(state, op.i.rt, data);
 190 }
 191
 192 static void lightrec_invalidate_map(struct lightrec_state *state,
 193                 const struct lightrec_mem_map *map, u32 addr, u32 len)
 194 {
 195         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
 196                 memset(lut_address(state, lut_offset(addr)), 0,
 197                        ((len + 3) / 4) * lut_elm_size(state));
 198         }
 199 }
 200
 201 static enum psx_map
 202 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
 203 {
 204         const struct lightrec_mem_map *map;
 205         unsigned int i;
 206
 207         for (i = 0; i < state->nb_maps; i++) {
 208                 map = &state->maps[i];
 209
 210                 if (kaddr >= map->pc && kaddr < map->pc + map->length)
 211                         return (enum psx_map) i;
 212         }
 213
 214         return PSX_MAP_UNKNOWN;
 215 }
 216
 217 const struct lightrec_mem_map *
 218 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 219 {
 220         const struct lightrec_mem_map *map;
 221         enum psx_map idx;
 222         u32 addr;
 223
 224         idx = lightrec_get_map_idx(state, kaddr);
 225         if (idx == PSX_MAP_UNKNOWN)
 226                 return NULL;
 227
 228         map = &state->maps[idx];
 229         addr = kaddr - map->pc;
 230
 231         while (map->mirror_of)
 232                 map = map->mirror_of;
 233
 234         if (host)
 235                 *host = map->address + addr;
 236
 237         return map;
 238 }
 239
 240 u32 lightrec_rw(struct lightrec_state *state, union code op,
 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
 432         if (op.i.op == OP_SWC2) {
 433                 val = lightrec_mfc2(state, op.i.rt);
 434         } else if (op.r.rs == OP_CP2_BASIC_MFC2)
 435                 val = lightrec_mfc2(state, op.r.rd);
 436         else {
 437                 val = state->regs.cp2c[op.r.rd];
 438
 439                 switch (op.r.rd) {
 440                 case 4:
 441                 case 12:
 442                 case 20:
 443                 case 26:
 444                 case 27:
 445                 case 29:
 446                 case 30:
 447                         val = (u32)(s16)val;
 448                         fallthrough;
 449                 default:
 450                         break;
 451                 }
 452         }
 453
 454         if (state->ops.cop2_notify)
 455                 (*state->ops.cop2_notify)(state, op.opcode, val);
 456
 457         return val;
 458 }
 459
 460 static void lightrec_mfc_cb(struct lightrec_state *state, union code op)
 461 {
 462         u32 rt = lightrec_mfc(state, op);
 463
 464         if (op.i.op == OP_SWC2)
 465                 state->cp2_temp_reg = rt;
 466         else if (op.r.rt)
 467                 state->regs.gpr[op.r.rt] = rt;
 468 }
 469
 470 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 471 {
 472         u32 status, oldstatus, cause;
 473
 474         switch (reg) {
 475         case 1:
 476         case 4:
 477         case 8:
 478         case 14:
 479         case 15:
 480                 /* Those registers are read-only */
 481                 return;
 482         default:
 483                 break;
 484         }
 485
 486         if (reg == 12) {
 487                 status = state->regs.cp0[12];
 488                 oldstatus = status;
 489
 490                 if (status & ~data & BIT(16)) {
 491                         state->ops.enable_ram(state, true);
 492                         lightrec_invalidate_all(state);
 493                 } else if (~status & data & BIT(16)) {
 494                         state->ops.enable_ram(state, false);
 495                 }
 496         }
 497
 498         if (reg == 13) {
 499                 state->regs.cp0[13] &= ~0x300;
 500                 state->regs.cp0[13] |= data & 0x300;
 501         } else {
 502                 state->regs.cp0[reg] = data;
 503         }
 504
 505         if (reg == 12 || reg == 13) {
 506                 cause = state->regs.cp0[13];
 507                 status = state->regs.cp0[12];
 508
 509                 /* Handle software interrupts */
 510                 if (!!(status & cause & 0x300) & status)
 511                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 512
 513                 /* Handle hardware interrupts */
 514                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
 515                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 516         }
 517 }
 518
 519 static u32 count_leading_bits(s32 data)
 520 {
 521         u32 cnt = 33;
 522
 523 #ifdef __has_builtin
 524 #if __has_builtin(__builtin_clrsb)
 525         return 1 + __builtin_clrsb(data);
 526 #endif
 527 #endif
 528
 529         data = (data ^ (data >> 31)) << 1;
 530
 531         do {
 532                 cnt -= 1;
 533                 data >>= 1;
 534         } while (data);
 535
 536         return cnt;
 537 }
 538
 539 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
 540 {
 541         switch (reg) {
 542         case 15:
 543                 state->regs.cp2d[12] = state->regs.cp2d[13];
 544                 state->regs.cp2d[13] = state->regs.cp2d[14];
 545                 state->regs.cp2d[14] = data;
 546                 break;
 547         case 28:
 548                 state->regs.cp2d[9] = (data << 7) & 0xf80;
 549                 state->regs.cp2d[10] = (data << 2) & 0xf80;
 550                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
 551                 break;
 552         case 31:
 553                 return;
 554         case 30:
 555                 state->regs.cp2d[31] = count_leading_bits((s32) data);
 556                 fallthrough;
 557         default:
 558                 state->regs.cp2d[reg] = data;
 559                 break;
 560         }
 561 }
 562
 563 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 564 {
 565         switch (reg) {
 566         case 4:
 567         case 12:
 568         case 20:
 569         case 26:
 570         case 27:
 571         case 29:
 572         case 30:
 573                 store_u16(&state->regs.cp2c[reg], data);
 574                 break;
 575         case 31:
 576                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
 577                 fallthrough;
 578         default:
 579                 state->regs.cp2c[reg] = data;
 580                 break;
 581         }
 582 }
 583
 584 void lightrec_mtc(struct lightrec_state *state, union code op, u8 reg, u32 data)
 585 {
 586         if (op.i.op == OP_CP0) {
 587                 lightrec_mtc0(state, reg, data);
 588         } else {
 589                 if (op.i.op == OP_LWC2 || op.r.rs != OP_CP2_BASIC_CTC2)
 590                         lightrec_mtc2(state, reg, data);
 591                 else
 592                         lightrec_ctc2(state, reg, data);
 593
 594                 if (state->ops.cop2_notify)
 595                         (*state->ops.cop2_notify)(state, op.opcode, data);
 596         }
 597 }
 598
 599 static void lightrec_mtc_cb(struct lightrec_state *state, u32 arg)
 600 {
 601         union code op = (union code) arg;
 602         u32 data;
 603         u8 reg;
 604
 605         if (op.i.op == OP_LWC2) {
 606                 data = state->cp2_temp_reg;
 607                 reg = op.i.rt;
 608         } else {
 609                 data = state->regs.gpr[op.r.rt];
 610                 reg = op.r.rd;
 611         }
 612
 613         lightrec_mtc(state, op, reg, data);
 614 }
 615
 616 void lightrec_rfe(struct lightrec_state *state)
 617 {
 618         u32 status;
 619
 620         /* Read CP0 Status register (r12) */
 621         status = state->regs.cp0[12];
 622
 623         /* Switch the bits */
 624         status = ((status & 0x3c) >> 2) | (status & ~0xf);
 625
 626         /* Write it back */
 627         lightrec_mtc0(state, 12, status);
 628 }
 629
 630 void lightrec_cp(struct lightrec_state *state, union code op)
 631 {
 632         if (op.i.op == OP_CP0) {
 633                 pr_err("Invalid CP opcode to coprocessor #0\n");
 634                 return;
 635         }
 636
 637         (*state->ops.cop2_op)(state, op.opcode);
 638 }
 639
 640 static void lightrec_cp_cb(struct lightrec_state *state, u32 arg)
 641 {
 642         lightrec_cp(state, (union code) arg);
 643 }
 644
 645 static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 646 {
 647         struct block *block = lightrec_find_block(state->block_cache, pc);
 648         u8 old_flags;
 649
 650         if (block && lightrec_block_is_outdated(state, block)) {
 651                 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
 652
 653                 old_flags = block_set_flags(block, BLOCK_IS_DEAD);
 654                 if (!(old_flags & BLOCK_IS_DEAD)) {
 655                         /* Make sure the recompiler isn't processing the block
 656                          * we'll destroy */
 657                         if (ENABLE_THREADED_COMPILER)
 658                                 lightrec_recompiler_remove(state->rec, block);
 659
 660                         lightrec_unregister_block(state->block_cache, block);
 661                         remove_from_code_lut(state->block_cache, block);
 662                         lightrec_free_block(state, block);
 663                 }
 664
 665                 block = NULL;
 666         }
 667
 668         if (!block) {
 669                 block = lightrec_precompile_block(state, pc);
 670                 if (!block) {
 671                         pr_err("Unable to recompile block at PC 0x%x\n", pc);
 672                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 673                         return NULL;
 674                 }
 675
 676                 lightrec_register_block(state->block_cache, block);
 677         }
 678
 679         return block;
 680 }
 681
 682 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 683 {
 684         struct block *block;
 685         bool should_recompile;
 686         void *func;
 687         int err;
 688
 689         do {
 690                 func = lut_read(state, lut_offset(pc));
 691                 if (func && func != state->get_next_block)
 692                         break;
 693
 694                 block = lightrec_get_block(state, pc);
 695
 696                 if (unlikely(!block))
 697                         break;
 698
 699                 if (OPT_REPLACE_MEMSET &&
 700                     block_has_flag(block, BLOCK_IS_MEMSET)) {
 701                         func = state->memset_func;
 702                         break;
 703                 }
 704
 705                 should_recompile = block_has_flag(block, BLOCK_SHOULD_RECOMPILE) &&
 706                         !block_has_flag(block, BLOCK_IS_DEAD);
 707
 708                 if (unlikely(should_recompile)) {
 709                         pr_debug("Block at PC 0x%08x should recompile\n", pc);
 710
 711                         if (ENABLE_THREADED_COMPILER) {
 712                                 lightrec_recompiler_add(state->rec, block);
 713                         } else {
 714                                 err = lightrec_compile_block(state->cstate, block);
 715                                 if (err) {
 716                                         state->exit_flags = LIGHTREC_EXIT_NOMEM;
 717                                         return NULL;
 718                                 }
 719                         }
 720                 }
 721
 722                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
 723                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
 724                 else
 725                         func = block->function;
 726
 727                 if (likely(func))
 728                         break;
 729
 730                 if (unlikely(block_has_flag(block, BLOCK_NEVER_COMPILE))) {
 731                         pc = lightrec_emulate_block(state, block, pc);
 732
 733                 } else if (!ENABLE_THREADED_COMPILER) {
 734                         /* Block wasn't compiled yet - run the interpreter */
 735                         if (block_has_flag(block, BLOCK_FULLY_TAGGED))
 736                                 pr_debug("Block fully tagged, skipping first pass\n");
 737                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 738                                 pc = lightrec_emulate_block(state, block, pc);
 739
 740                         /* Then compile it using the profiled data */
 741                         err = lightrec_compile_block(state->cstate, block);
 742                         if (err) {
 743                                 state->exit_flags = LIGHTREC_EXIT_NOMEM;
 744                                 return NULL;
 745                         }
 746                 } else if (unlikely(block_has_flag(block, BLOCK_IS_DEAD))) {
 747                         /*
 748                          * If the block is dead but has never been compiled,
 749                          * then its function pointer is NULL and we cannot
 750                          * execute the block. In that case, reap all the dead
 751                          * blocks now, and in the next loop we will create a
 752                          * new block.
 753                          */
 754                         lightrec_reaper_reap(state->reaper);
 755                 } else {
 756                         lightrec_recompiler_add(state->rec, block);
 757                 }
 758         } while (state->exit_flags == LIGHTREC_EXIT_NORMAL
 759                  && state->current_cycle < state->target_cycle);
 760
 761         state->next_pc = pc;
 762         return func;
 763 }
 764
 765 static void * lightrec_alloc_code(struct lightrec_state *state, size_t size)
 766 {
 767         void *code;
 768
 769         if (ENABLE_THREADED_COMPILER)
 770                 lightrec_code_alloc_lock(state);
 771
 772         code = tlsf_malloc(state->tlsf, size);
 773
 774         if (ENABLE_THREADED_COMPILER)
 775                 lightrec_code_alloc_unlock(state);
 776
 777         return code;
 778 }
 779
 780 static void lightrec_realloc_code(struct lightrec_state *state,
 781                                   void *ptr, size_t size)
 782 {
 783         /* NOTE: 'size' MUST be smaller than the size specified during
 784          * the allocation. */
 785
 786         if (ENABLE_THREADED_COMPILER)
 787                 lightrec_code_alloc_lock(state);
 788
 789         tlsf_realloc(state->tlsf, ptr, size);
 790
 791         if (ENABLE_THREADED_COMPILER)
 792                 lightrec_code_alloc_unlock(state);
 793 }
 794
 795 static void lightrec_free_code(struct lightrec_state *state, void *ptr)
 796 {
 797         if (ENABLE_THREADED_COMPILER)
 798                 lightrec_code_alloc_lock(state);
 799
 800         tlsf_free(state->tlsf, ptr);
 801
 802         if (ENABLE_THREADED_COMPILER)
 803                 lightrec_code_alloc_unlock(state);
 804 }
 805
 806 static void * lightrec_emit_code(struct lightrec_state *state,
 807                                  const struct block *block,
 808                                  jit_state_t *_jit, unsigned int *size)
 809 {
 810         bool has_code_buffer = ENABLE_CODE_BUFFER && state->tlsf;
 811         jit_word_t code_size, new_code_size;
 812         void *code;
 813
 814         jit_realize();
 815
 816         if (!ENABLE_DISASSEMBLER)
 817                 jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
 818
 819         if (has_code_buffer) {
 820                 jit_get_code(&code_size);
 821                 code = lightrec_alloc_code(state, (size_t) code_size);
 822
 823                 if (!code) {
 824                         if (ENABLE_THREADED_COMPILER) {
 825                                 /* If we're using the threaded compiler, return
 826                                  * an allocation error here. The threaded
 827                                  * compiler will then empty its job queue and
 828                                  * request a code flush using the reaper. */
 829                                 return NULL;
 830                         }
 831
 832                         /* Remove outdated blocks, and try again */
 833                         lightrec_remove_outdated_blocks(state->block_cache, block);
 834
 835                         pr_debug("Re-try to alloc %zu bytes...\n", code_size);
 836
 837                         code = lightrec_alloc_code(state, code_size);
 838                         if (!code) {
 839                                 pr_err("Could not alloc even after removing old blocks!\n");
 840                                 return NULL;
 841                         }
 842                 }
 843
 844                 jit_set_code(code, code_size);
 845         }
 846
 847         code = jit_emit();
 848
 849         jit_get_code(&new_code_size);
 850         lightrec_register(MEM_FOR_CODE, new_code_size);
 851
 852         if (has_code_buffer) {
 853                 lightrec_realloc_code(state, code, (size_t) new_code_size);
 854
 855                 pr_debug("Creating code block at address 0x%" PRIxPTR ", "
 856                          "code size: %" PRIuPTR " new: %" PRIuPTR "\n",
 857                          (uintptr_t) code, code_size, new_code_size);
 858         }
 859
 860         *size = (unsigned int) new_code_size;
 861
 862         if (state->ops.code_inv)
 863                 state->ops.code_inv(code, new_code_size);
 864
 865         return code;
 866 }
 867
 868 static struct block * generate_wrapper(struct lightrec_state *state)
 869 {
 870         struct block *block;
 871         jit_state_t *_jit;
 872         unsigned int i;
 873         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
 874         jit_node_t *to_end[C_WRAPPERS_COUNT - 1];
 875
 876         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 877         if (!block)
 878                 goto err_no_mem;
 879
 880         _jit = jit_new_state();
 881         if (!_jit)
 882                 goto err_free_block;
 883
 884         jit_name("RW wrapper");
 885         jit_note(__FILE__, __LINE__);
 886
 887         /* Wrapper entry point */
 888         jit_prolog();
 889         jit_tramp(256);
 890
 891         /* Add entry points */
 892         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
 893                 jit_ldxi(JIT_R1, LIGHTREC_REG_STATE,
 894                          offsetof(struct lightrec_state, c_wrappers[i]));
 895                 to_end[i - 1] = jit_b();
 896                 addr[i - 1] = jit_indirect();
 897         }
 898
 899         jit_ldxi(JIT_R1, LIGHTREC_REG_STATE,
 900                  offsetof(struct lightrec_state, c_wrappers[0]));
 901
 902         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 903                 jit_patch(to_end[i]);
 904
 905         jit_epilog();
 906         jit_prolog();
 907
 908         /* Save all temporaries on stack */
 909         for (i = 0; i < NUM_TEMPS; i++) {
 910                 if (i + FIRST_TEMP != 1) {
 911                         jit_stxi(offsetof(struct lightrec_state, wrapper_regs[i]),
 912                                  LIGHTREC_REG_STATE, JIT_R(i + FIRST_TEMP));
 913                 }
 914         }
 915
 916         jit_getarg(JIT_R2, jit_arg());
 917
 918         jit_prepare();
 919         jit_pushargr(LIGHTREC_REG_STATE);
 920         jit_pushargr(JIT_R2);
 921
 922         jit_ldxi_ui(JIT_R2, LIGHTREC_REG_STATE,
 923                     offsetof(struct lightrec_state, target_cycle));
 924
 925         /* state->current_cycle = state->target_cycle - delta; */
 926         jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, LIGHTREC_REG_CYCLE);
 927         jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
 928                    LIGHTREC_REG_STATE, LIGHTREC_REG_CYCLE);
 929
 930         /* Call the wrapper function */
 931         jit_finishr(JIT_R1);
 932
 933         /* delta = state->target_cycle - state->current_cycle */;
 934         jit_ldxi_ui(LIGHTREC_REG_CYCLE, LIGHTREC_REG_STATE,
 935                     offsetof(struct lightrec_state, current_cycle));
 936         jit_ldxi_ui(JIT_R1, LIGHTREC_REG_STATE,
 937                     offsetof(struct lightrec_state, target_cycle));
 938         jit_subr(LIGHTREC_REG_CYCLE, JIT_R1, LIGHTREC_REG_CYCLE);
 939
 940         /* Restore temporaries from stack */
 941         for (i = 0; i < NUM_TEMPS; i++) {
 942                 if (i + FIRST_TEMP != 1) {
 943                         jit_ldxi(JIT_R(i + FIRST_TEMP), LIGHTREC_REG_STATE,
 944                                  offsetof(struct lightrec_state, wrapper_regs[i]));
 945                 }
 946         }
 947
 948         jit_ret();
 949         jit_epilog();
 950
 951         block->_jit = _jit;
 952         block->opcode_list = NULL;
 953         block->flags = BLOCK_NO_OPCODE_LIST;
 954         block->nb_ops = 0;
 955
 956         block->function = lightrec_emit_code(state, block, _jit,
 957                                              &block->code_size);
 958         if (!block->function)
 959                 goto err_free_block;
 960
 961         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
 962
 963         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 964                 state->wrappers_eps[i] = jit_address(addr[i]);
 965
 966         if (ENABLE_DISASSEMBLER) {
 967                 pr_debug("Wrapper block:\n");
 968                 jit_disassemble();
 969         }
 970
 971         jit_clear_state();
 972         return block;
 973
 974 err_free_block:
 975         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
 976 err_no_mem:
 977         pr_err("Unable to compile wrapper: Out of memory\n");
 978         return NULL;
 979 }
 980
 981 static u32 lightrec_memset(struct lightrec_state *state)
 982 {
 983         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
 984         void *host;
 985         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
 986         u32 length = state->regs.gpr[5] * 4;
 987
 988         if (!map) {
 989                 pr_err("Unable to find memory map for memset target address "
 990                        "0x%x\n", kunseg_pc);
 991                 return 0;
 992         }
 993
 994         pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
 995                  kunseg_pc, (uintptr_t)host, length);
 996         memset(host, 0, length);
 997
 998         if (!state->invalidate_from_dma_only)
 999                 lightrec_invalidate_map(state, map, kunseg_pc, length);
1000
1001         /* Rough estimation of the number of cycles consumed */
1002         return 8 + 5 * (length  + 3 / 4);
1003 }
1004
1005 static struct block * generate_dispatcher(struct lightrec_state *state)
1006 {
1007         struct block *block;
1008         jit_state_t *_jit;
1009         jit_node_t *to_end, *loop, *addr, *addr2, *addr3;
1010         unsigned int i;
1011         u32 offset;
1012
1013         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1014         if (!block)
1015                 goto err_no_mem;
1016
1017         _jit = jit_new_state();
1018         if (!_jit)
1019                 goto err_free_block;
1020
1021         jit_name("dispatcher");
1022         jit_note(__FILE__, __LINE__);
1023
1024         jit_prolog();
1025         jit_frame(256);
1026
1027         jit_getarg(LIGHTREC_REG_STATE, jit_arg());
1028         jit_getarg(JIT_V0, jit_arg());
1029         jit_getarg(JIT_V1, jit_arg());
1030         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
1031
1032         /* Force all callee-saved registers to be pushed on the stack */
1033         for (i = 0; i < NUM_REGS; i++)
1034                 jit_movr(JIT_V(i + FIRST_REG), JIT_V(i + FIRST_REG));
1035
1036         loop = jit_label();
1037
1038         /* Call the block's code */
1039         jit_jmpr(JIT_V1);
1040
1041         if (OPT_REPLACE_MEMSET) {
1042                 /* Blocks will jump here when they need to call
1043                  * lightrec_memset() */
1044                 addr3 = jit_indirect();
1045
1046                 jit_movr(JIT_V1, LIGHTREC_REG_CYCLE);
1047
1048                 jit_prepare();
1049                 jit_pushargr(LIGHTREC_REG_STATE);
1050                 jit_finishi(lightrec_memset);
1051
1052                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1053                             offsetof(struct lightrec_state, regs.gpr[31]));
1054
1055                 jit_retval(LIGHTREC_REG_CYCLE);
1056                 jit_subr(LIGHTREC_REG_CYCLE, JIT_V1, LIGHTREC_REG_CYCLE);
1057         }
1058
1059         /* The block will jump here, with the number of cycles remaining in
1060          * LIGHTREC_REG_CYCLE */
1061         addr2 = jit_indirect();
1062
1063         /* Store back the next_pc to the lightrec_state structure */
1064         offset = offsetof(struct lightrec_state, next_pc);
1065         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
1066
1067         /* Jump to end if state->target_cycle < state->current_cycle */
1068         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
1069
1070         /* Convert next PC to KUNSEG and avoid mirrors */
1071         jit_andi(JIT_V1, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
1072         jit_rshi_u(JIT_R1, JIT_V1, 28);
1073         jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
1074         jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
1075         jit_movnr(JIT_V1, JIT_R2, JIT_R1);
1076
1077         /* If possible, use the code LUT */
1078         if (!lut_is_32bit(state))
1079                 jit_lshi(JIT_V1, JIT_V1, 1);
1080         jit_addr(JIT_V1, JIT_V1, LIGHTREC_REG_STATE);
1081
1082         offset = offsetof(struct lightrec_state, code_lut);
1083         if (lut_is_32bit(state))
1084                 jit_ldxi_ui(JIT_V1, JIT_V1, offset);
1085         else
1086                 jit_ldxi(JIT_V1, JIT_V1, offset);
1087
1088         /* If we get non-NULL, loop */
1089         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1090
1091         /* The code LUT will be set to this address when the block at the target
1092          * PC has been preprocessed but not yet compiled by the threaded
1093          * recompiler */
1094         addr = jit_indirect();
1095
1096         /* Slow path: call C function get_next_block_func() */
1097
1098         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1099                 /* We may call the interpreter - update state->current_cycle */
1100                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1101                            offsetof(struct lightrec_state, target_cycle));
1102                 jit_subr(JIT_V1, JIT_R2, LIGHTREC_REG_CYCLE);
1103                 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
1104                            LIGHTREC_REG_STATE, JIT_V1);
1105         }
1106
1107         jit_prepare();
1108         jit_pushargr(LIGHTREC_REG_STATE);
1109         jit_pushargr(JIT_V0);
1110
1111         /* Save the cycles register if needed */
1112         if (!(ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES))
1113                 jit_movr(JIT_V0, LIGHTREC_REG_CYCLE);
1114
1115         /* Get the next block */
1116         jit_finishi(&get_next_block_func);
1117         jit_retval(JIT_V1);
1118
1119         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1120                 /* The interpreter may have updated state->current_cycle and
1121                  * state->target_cycle - recalc the delta */
1122                 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
1123                            offsetof(struct lightrec_state, current_cycle));
1124                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1125                            offsetof(struct lightrec_state, target_cycle));
1126                 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
1127         } else {
1128                 jit_movr(LIGHTREC_REG_CYCLE, JIT_V0);
1129         }
1130
1131         /* Reset JIT_V0 to the next PC */
1132         jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1133                     offsetof(struct lightrec_state, next_pc));
1134
1135         /* If we get non-NULL, loop */
1136         jit_patch_at(jit_bnei(JIT_V1, 0), loop);
1137
1138         /* When exiting, the recompiled code will jump to that address */
1139         jit_note(__FILE__, __LINE__);
1140         jit_patch(to_end);
1141
1142         jit_retr(LIGHTREC_REG_CYCLE);
1143         jit_epilog();
1144
1145         block->_jit = _jit;
1146         block->opcode_list = NULL;
1147         block->flags = BLOCK_NO_OPCODE_LIST;
1148         block->nb_ops = 0;
1149
1150         block->function = lightrec_emit_code(state, block, _jit,
1151                                              &block->code_size);
1152         if (!block->function)
1153                 goto err_free_block;
1154
1155         state->eob_wrapper_func = jit_address(addr2);
1156         if (OPT_REPLACE_MEMSET)
1157                 state->memset_func = jit_address(addr3);
1158         state->get_next_block = jit_address(addr);
1159
1160         if (ENABLE_DISASSEMBLER) {
1161                 pr_debug("Dispatcher block:\n");
1162                 jit_disassemble();
1163         }
1164
1165         /* We're done! */
1166         jit_clear_state();
1167         return block;
1168
1169 err_free_block:
1170         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1171 err_no_mem:
1172         pr_err("Unable to compile dispatcher: Out of memory\n");
1173         return NULL;
1174 }
1175
1176 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
1177 {
1178         void *host = NULL;
1179
1180         lightrec_get_map(state, &host, kunseg(pc));
1181
1182         const u32 *code = (u32 *)host;
1183         return (union code) LE32TOH(*code);
1184 }
1185
1186 unsigned int lightrec_cycles_of_opcode(union code code)
1187 {
1188         return 2;
1189 }
1190
1191 void lightrec_free_opcode_list(struct lightrec_state *state, struct opcode *ops)
1192 {
1193         struct opcode_list *list = container_of(ops, struct opcode_list, ops);
1194
1195         lightrec_free(state, MEM_FOR_IR,
1196                       sizeof(*list) + list->nb_ops * sizeof(struct opcode),
1197                       list);
1198 }
1199
1200 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1201 {
1202         unsigned int i;
1203         union code c;
1204
1205         for (i = 1; ; i++) {
1206                 c.opcode = LE32TOH(*src++);
1207
1208                 if (is_syscall(c))
1209                         return i;
1210
1211                 if (is_unconditional_jump(c))
1212                         return i + 1;
1213         }
1214 }
1215
1216 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1217                                             const u32 *src, unsigned int *len)
1218 {
1219         struct opcode_list *list;
1220         unsigned int i, length;
1221
1222         length = lightrec_get_mips_block_len(src);
1223
1224         list = lightrec_malloc(state, MEM_FOR_IR,
1225                                sizeof(*list) + sizeof(struct opcode) * length);
1226         if (!list) {
1227                 pr_err("Unable to allocate memory\n");
1228                 return NULL;
1229         }
1230
1231         list->nb_ops = (u16) length;
1232
1233         for (i = 0; i < length; i++) {
1234                 list->ops[i].opcode = LE32TOH(src[i]);
1235                 list->ops[i].flags = 0;
1236         }
1237
1238         *len = length * sizeof(u32);
1239
1240         return list->ops;
1241 }
1242
1243 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1244                                                 u32 pc)
1245 {
1246         struct opcode *list;
1247         struct block *block;
1248         void *host, *addr;
1249         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1250         const u32 *code = (u32 *) host;
1251         unsigned int length;
1252         bool fully_tagged;
1253         u8 block_flags = 0;
1254
1255         if (!map)
1256                 return NULL;
1257
1258         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1259         if (!block) {
1260                 pr_err("Unable to recompile block: Out of memory\n");
1261                 return NULL;
1262         }
1263
1264         list = lightrec_disassemble(state, code, &length);
1265         if (!list) {
1266                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1267                 return NULL;
1268         }
1269
1270         block->pc = pc;
1271         block->_jit = NULL;
1272         block->function = NULL;
1273         block->opcode_list = list;
1274         block->code = code;
1275         block->next = NULL;
1276         block->flags = 0;
1277         block->code_size = 0;
1278         block->precompile_date = state->current_cycle;
1279         block->nb_ops = length / sizeof(u32);
1280
1281         lightrec_optimize(state, block);
1282
1283         length = block->nb_ops * sizeof(u32);
1284
1285         lightrec_register(MEM_FOR_MIPS_CODE, length);
1286
1287         if (ENABLE_DISASSEMBLER) {
1288                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1289                 lightrec_print_disassembly(block, code);
1290         }
1291
1292         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1293
1294         /* If the first opcode is an 'impossible' branch, never compile the
1295          * block */
1296         if (should_emulate(block->opcode_list))
1297                 block_flags |= BLOCK_NEVER_COMPILE;
1298
1299         fully_tagged = lightrec_block_is_fully_tagged(block);
1300         if (fully_tagged)
1301                 block_flags |= BLOCK_FULLY_TAGGED;
1302
1303         if (block_flags)
1304                 block_set_flags(block, block_flags);
1305
1306         block->hash = lightrec_calculate_block_hash(block);
1307
1308         if (OPT_REPLACE_MEMSET && block_has_flag(block, BLOCK_IS_MEMSET))
1309                 addr = state->memset_func;
1310         else
1311                 addr = state->get_next_block;
1312         lut_write(state, lut_offset(pc), addr);
1313
1314         pr_debug("Recompile count: %u\n", state->nb_precompile++);
1315
1316         return block;
1317 }
1318
1319 static bool lightrec_block_is_fully_tagged(const struct block *block)
1320 {
1321         const struct opcode *op;
1322         unsigned int i;
1323
1324         for (i = 0; i < block->nb_ops; i++) {
1325                 op = &block->opcode_list[i];
1326
1327                 /* Verify that all load/stores of the opcode list
1328                  * Check all loads/stores of the opcode list and mark the
1329                  * block as fully compiled if they all have been tagged. */
1330                 switch (op->c.i.op) {
1331                 case OP_LB:
1332                 case OP_LH:
1333                 case OP_LWL:
1334                 case OP_LW:
1335                 case OP_LBU:
1336                 case OP_LHU:
1337                 case OP_LWR:
1338                 case OP_SB:
1339                 case OP_SH:
1340                 case OP_SWL:
1341                 case OP_SW:
1342                 case OP_SWR:
1343                 case OP_LWC2:
1344                 case OP_SWC2:
1345                         if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
1346                                 return false;
1347                         fallthrough;
1348                 default:
1349                         continue;
1350                 }
1351         }
1352
1353         return true;
1354 }
1355
1356 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1357 {
1358         struct block *block = data;
1359
1360         pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1361         lightrec_unregister_block(state->block_cache, block);
1362         lightrec_free_block(state, block);
1363 }
1364
1365 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1366 {
1367         _jit_destroy_state(data);
1368 }
1369
1370 static void lightrec_free_function(struct lightrec_state *state, void *fn)
1371 {
1372         if (ENABLE_CODE_BUFFER && state->tlsf) {
1373                 pr_debug("Freeing code block at 0x%" PRIxPTR "\n", (uintptr_t) fn);
1374                 lightrec_free_code(state, fn);
1375         }
1376 }
1377
1378 static void lightrec_reap_function(struct lightrec_state *state, void *data)
1379 {
1380         lightrec_free_function(state, data);
1381 }
1382
1383 static void lightrec_reap_opcode_list(struct lightrec_state *state, void *data)
1384 {
1385         lightrec_free_opcode_list(state, data);
1386 }
1387
1388 int lightrec_compile_block(struct lightrec_cstate *cstate,
1389                            struct block *block)
1390 {
1391         struct lightrec_state *state = cstate->state;
1392         struct lightrec_branch_target *target;
1393         bool fully_tagged = false;
1394         struct block *block2;
1395         struct opcode *elm;
1396         jit_state_t *_jit, *oldjit;
1397         jit_node_t *start_of_block;
1398         bool skip_next = false;
1399         void *old_fn, *new_fn;
1400         size_t old_code_size;
1401         unsigned int i, j;
1402         u8 old_flags;
1403         u32 offset;
1404
1405         fully_tagged = lightrec_block_is_fully_tagged(block);
1406         if (fully_tagged)
1407                 block_set_flags(block, BLOCK_FULLY_TAGGED);
1408
1409         _jit = jit_new_state();
1410         if (!_jit)
1411                 return -ENOMEM;
1412
1413         oldjit = block->_jit;
1414         old_fn = block->function;
1415         old_code_size = block->code_size;
1416         block->_jit = _jit;
1417
1418         lightrec_regcache_reset(cstate->reg_cache);
1419         lightrec_preload_pc(cstate->reg_cache);
1420
1421         cstate->cycles = 0;
1422         cstate->nb_local_branches = 0;
1423         cstate->nb_targets = 0;
1424
1425         jit_prolog();
1426         jit_tramp(256);
1427
1428         start_of_block = jit_label();
1429
1430         for (i = 0; i < block->nb_ops; i++) {
1431                 elm = &block->opcode_list[i];
1432
1433                 if (skip_next) {
1434                         skip_next = false;
1435                         continue;
1436                 }
1437
1438                 if (should_emulate(elm)) {
1439                         pr_debug("Branch at offset 0x%x will be emulated\n",
1440                                  i << 2);
1441
1442                         lightrec_emit_eob(cstate, block, i);
1443                         skip_next = !op_flag_no_ds(elm->flags);
1444                 } else {
1445                         lightrec_rec_opcode(cstate, block, i);
1446                         skip_next = !op_flag_no_ds(elm->flags) && has_delay_slot(elm->c);
1447 #if _WIN32
1448                         /* FIXME: GNU Lightning on Windows seems to use our
1449                          * mapped registers as temporaries. Until the actual bug
1450                          * is found and fixed, unconditionally mark our
1451                          * registers as live here. */
1452                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1453 #endif
1454                 }
1455
1456                 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1457         }
1458
1459         for (i = 0; i < cstate->nb_local_branches; i++) {
1460                 struct lightrec_branch *branch = &cstate->local_branches[i];
1461
1462                 pr_debug("Patch local branch to offset 0x%x\n",
1463                          branch->target << 2);
1464
1465                 if (branch->target == 0) {
1466                         jit_patch_at(branch->branch, start_of_block);
1467                         continue;
1468                 }
1469
1470                 for (j = 0; j < cstate->nb_targets; j++) {
1471                         if (cstate->targets[j].offset == branch->target) {
1472                                 jit_patch_at(branch->branch,
1473                                              cstate->targets[j].label);
1474                                 break;
1475                         }
1476                 }
1477
1478                 if (j == cstate->nb_targets)
1479                         pr_err("Unable to find branch target\n");
1480         }
1481
1482         jit_ret();
1483         jit_epilog();
1484
1485         new_fn = lightrec_emit_code(state, block, _jit, &block->code_size);
1486         if (!new_fn) {
1487                 if (!ENABLE_THREADED_COMPILER)
1488                         pr_err("Unable to compile block!\n");
1489                 block->_jit = oldjit;
1490                 jit_clear_state();
1491                 _jit_destroy_state(_jit);
1492                 return -ENOMEM;
1493         }
1494
1495         /* Pause the reaper, because lightrec_reset_lut_offset() may try to set
1496          * the old block->function pointer to the code LUT. */
1497         if (ENABLE_THREADED_COMPILER)
1498                 lightrec_reaper_pause(state->reaper);
1499
1500         block->function = new_fn;
1501         block_clear_flags(block, BLOCK_SHOULD_RECOMPILE);
1502
1503         /* Add compiled function to the LUT */
1504         lut_write(state, lut_offset(block->pc), block->function);
1505
1506         if (ENABLE_THREADED_COMPILER)
1507                 lightrec_reaper_continue(state->reaper);
1508
1509         /* Detect old blocks that have been covered by the new one */
1510         for (i = 0; i < cstate->nb_targets; i++) {
1511                 target = &cstate->targets[i];
1512
1513                 if (!target->offset)
1514                         continue;
1515
1516                 offset = block->pc + target->offset * sizeof(u32);
1517
1518                 /* Pause the reaper while we search for the block until we set
1519                  * the BLOCK_IS_DEAD flag, otherwise the block may be removed
1520                  * under our feet. */
1521                 if (ENABLE_THREADED_COMPILER)
1522                         lightrec_reaper_pause(state->reaper);
1523
1524                 block2 = lightrec_find_block(state->block_cache, offset);
1525                 if (block2) {
1526                         /* No need to check if block2 is compilable - it must
1527                          * be, otherwise block wouldn't be compilable either */
1528
1529                         /* Set the "block dead" flag to prevent the dynarec from
1530                          * recompiling this block */
1531                         old_flags = block_set_flags(block2, BLOCK_IS_DEAD);
1532                 }
1533
1534                 if (ENABLE_THREADED_COMPILER) {
1535                         lightrec_reaper_continue(state->reaper);
1536
1537                         /* If block2 was pending for compilation, cancel it.
1538                          * If it's being compiled right now, wait until it
1539                          * finishes. */
1540                         if (block2)
1541                                 lightrec_recompiler_remove(state->rec, block2);
1542                 }
1543
1544                 /* We know from now on that block2 (if present) isn't going to
1545                  * be compiled. We can override the LUT entry with our new
1546                  * block's entry point. */
1547                 offset = lut_offset(block->pc) + target->offset;
1548                 lut_write(state, offset, jit_address(target->label));
1549
1550                 if (block2) {
1551                         pr_debug("Reap block 0x%08x as it's covered by block "
1552                                  "0x%08x\n", block2->pc, block->pc);
1553
1554                         /* Finally, reap the block. */
1555                         if (!ENABLE_THREADED_COMPILER) {
1556                                 lightrec_unregister_block(state->block_cache, block2);
1557                                 lightrec_free_block(state, block2);
1558                         } else if (!(old_flags & BLOCK_IS_DEAD)) {
1559                                 lightrec_reaper_add(state->reaper,
1560                                                     lightrec_reap_block,
1561                                                     block2);
1562                         }
1563                 }
1564         }
1565
1566         if (ENABLE_DISASSEMBLER) {
1567                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1568                 jit_disassemble();
1569         }
1570
1571         jit_clear_state();
1572
1573         if (fully_tagged)
1574                 old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1575
1576         if (fully_tagged && !(old_flags & BLOCK_NO_OPCODE_LIST)) {
1577                 pr_debug("Block PC 0x%08x is fully tagged"
1578                          " - free opcode list\n", block->pc);
1579
1580                 if (ENABLE_THREADED_COMPILER) {
1581                         lightrec_reaper_add(state->reaper,
1582                                             lightrec_reap_opcode_list,
1583                                             block->opcode_list);
1584                 } else {
1585                         lightrec_free_opcode_list(state, block->opcode_list);
1586                 }
1587         }
1588
1589         if (oldjit) {
1590                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1591                          block->pc);
1592
1593                 if (ENABLE_THREADED_COMPILER) {
1594                         lightrec_reaper_add(state->reaper,
1595                                             lightrec_reap_jit, oldjit);
1596                         lightrec_reaper_add(state->reaper,
1597                                             lightrec_reap_function, old_fn);
1598                 } else {
1599                         _jit_destroy_state(oldjit);
1600                         lightrec_free_function(state, old_fn);
1601                 }
1602
1603                 lightrec_unregister(MEM_FOR_CODE, old_code_size);
1604         }
1605
1606         return 0;
1607 }
1608
1609 static void lightrec_print_info(struct lightrec_state *state)
1610 {
1611         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1612                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1613                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1614                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1615                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1616                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1617                         lightrec_get_total_mem_usage() / 1024,
1618                        lightrec_get_average_ipi());
1619                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1620         }
1621 }
1622
1623 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1624 {
1625         s32 (*func)(struct lightrec_state *, u32, void *, s32) = (void *)state->dispatcher->function;
1626         void *block_trace;
1627         s32 cycles_delta;
1628
1629         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1630
1631         /* Handle the cycle counter overflowing */
1632         if (unlikely(target_cycle < state->current_cycle))
1633                 target_cycle = UINT_MAX;
1634
1635         state->target_cycle = target_cycle;
1636         state->next_pc = pc;
1637
1638         block_trace = get_next_block_func(state, pc);
1639         if (block_trace) {
1640                 cycles_delta = state->target_cycle - state->current_cycle;
1641
1642                 cycles_delta = (*func)(state, state->next_pc,
1643                                        block_trace, cycles_delta);
1644
1645                 state->current_cycle = state->target_cycle - cycles_delta;
1646         }
1647
1648         if (ENABLE_THREADED_COMPILER)
1649                 lightrec_reaper_reap(state->reaper);
1650
1651         if (LOG_LEVEL >= INFO_L)
1652                 lightrec_print_info(state);
1653
1654         return state->next_pc;
1655 }
1656
1657 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc,
1658                              u32 target_cycle)
1659 {
1660         struct block *block;
1661
1662         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1663         state->target_cycle = target_cycle;
1664
1665         do {
1666                 block = lightrec_get_block(state, pc);
1667                 if (!block)
1668                         break;
1669
1670                 pc = lightrec_emulate_block(state, block, pc);
1671
1672                 if (ENABLE_THREADED_COMPILER)
1673                         lightrec_reaper_reap(state->reaper);
1674         } while (state->current_cycle < state->target_cycle);
1675
1676         if (LOG_LEVEL >= INFO_L)
1677                 lightrec_print_info(state);
1678
1679         return pc;
1680 }
1681
1682 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1683 {
1684         u8 old_flags;
1685
1686         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1687         old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
1688
1689         if (!(old_flags & BLOCK_NO_OPCODE_LIST))
1690                 lightrec_free_opcode_list(state, block->opcode_list);
1691         if (block->_jit)
1692                 _jit_destroy_state(block->_jit);
1693         if (block->function) {
1694                 lightrec_free_function(state, block->function);
1695                 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1696         }
1697         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1698 }
1699
1700 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1701 {
1702         struct lightrec_cstate *cstate;
1703
1704         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1705         if (!cstate)
1706                 return NULL;
1707
1708         cstate->reg_cache = lightrec_regcache_init(state);
1709         if (!cstate->reg_cache) {
1710                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1711                 return NULL;
1712         }
1713
1714         cstate->state = state;
1715
1716         return cstate;
1717 }
1718
1719 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1720 {
1721         lightrec_free_regcache(cstate->reg_cache);
1722         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1723 }
1724
1725 struct lightrec_state * lightrec_init(char *argv0,
1726                                       const struct lightrec_mem_map *map,
1727                                       size_t nb,
1728                                       const struct lightrec_ops *ops)
1729 {
1730         const struct lightrec_mem_map *codebuf_map = &map[PSX_MAP_CODE_BUFFER];
1731         struct lightrec_state *state;
1732         uintptr_t addr;
1733         void *tlsf = NULL;
1734         bool with_32bit_lut = false;
1735         size_t lut_size;
1736
1737         /* Sanity-check ops */
1738         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1739                 pr_err("Missing callbacks in lightrec_ops structure\n");
1740                 return NULL;
1741         }
1742
1743         if (ops->cop2_notify)
1744                 pr_debug("Optional cop2_notify callback in lightrec_ops\n");
1745         else
1746                 pr_debug("No optional cop2_notify callback in lightrec_ops\n");
1747
1748         if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER
1749             && codebuf_map->address) {
1750                 tlsf = tlsf_create_with_pool(codebuf_map->address,
1751                                              codebuf_map->length);
1752                 if (!tlsf) {
1753                         pr_err("Unable to initialize code buffer\n");
1754                         return NULL;
1755                 }
1756
1757                 if (__WORDSIZE == 64) {
1758                         addr = (uintptr_t) codebuf_map->address + codebuf_map->length - 1;
1759                         with_32bit_lut = addr == (u32) addr;
1760                 }
1761         }
1762
1763         if (with_32bit_lut)
1764                 lut_size = CODE_LUT_SIZE * 4;
1765         else
1766                 lut_size = CODE_LUT_SIZE * sizeof(void *);
1767
1768         init_jit(argv0);
1769
1770         state = calloc(1, sizeof(*state) + lut_size);
1771         if (!state)
1772                 goto err_finish_jit;
1773
1774         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) + lut_size);
1775
1776         state->tlsf = tlsf;
1777         state->with_32bit_lut = with_32bit_lut;
1778
1779         state->block_cache = lightrec_blockcache_init(state);
1780         if (!state->block_cache)
1781                 goto err_free_state;
1782
1783         if (ENABLE_THREADED_COMPILER) {
1784                 state->rec = lightrec_recompiler_init(state);
1785                 if (!state->rec)
1786                         goto err_free_block_cache;
1787
1788                 state->reaper = lightrec_reaper_init(state);
1789                 if (!state->reaper)
1790                         goto err_free_recompiler;
1791         } else {
1792                 state->cstate = lightrec_create_cstate(state);
1793                 if (!state->cstate)
1794                         goto err_free_block_cache;
1795         }
1796
1797         state->nb_maps = nb;
1798         state->maps = map;
1799
1800         memcpy(&state->ops, ops, sizeof(*ops));
1801
1802         state->dispatcher = generate_dispatcher(state);
1803         if (!state->dispatcher)
1804                 goto err_free_reaper;
1805
1806         state->c_wrapper_block = generate_wrapper(state);
1807         if (!state->c_wrapper_block)
1808                 goto err_free_dispatcher;
1809
1810         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1811         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1812         state->c_wrappers[C_WRAPPER_MFC] = lightrec_mfc_cb;
1813         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1814         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
1815
1816         map = &state->maps[PSX_MAP_BIOS];
1817         state->offset_bios = (uintptr_t)map->address - map->pc;
1818
1819         map = &state->maps[PSX_MAP_SCRATCH_PAD];
1820         state->offset_scratch = (uintptr_t)map->address - map->pc;
1821
1822         map = &state->maps[PSX_MAP_HW_REGISTERS];
1823         state->offset_io = (uintptr_t)map->address - map->pc;
1824
1825         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1826         state->offset_ram = (uintptr_t)map->address - map->pc;
1827
1828         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1829             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1830             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1831                 state->mirrors_mapped = true;
1832
1833         if (state->offset_bios == 0 &&
1834             state->offset_scratch == 0 &&
1835             state->offset_ram == 0 &&
1836             state->offset_io == 0 &&
1837             state->mirrors_mapped) {
1838                 pr_info("Memory map is perfect. Emitted code will be best.\n");
1839         } else {
1840                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1841         }
1842
1843         if (state->with_32bit_lut)
1844                 pr_info("Using 32-bit LUT\n");
1845
1846         return state;
1847
1848 err_free_dispatcher:
1849         lightrec_free_block(state, state->dispatcher);
1850 err_free_reaper:
1851         if (ENABLE_THREADED_COMPILER)
1852                 lightrec_reaper_destroy(state->reaper);
1853 err_free_recompiler:
1854         if (ENABLE_THREADED_COMPILER)
1855                 lightrec_free_recompiler(state->rec);
1856         else
1857                 lightrec_free_cstate(state->cstate);
1858 err_free_block_cache:
1859         lightrec_free_block_cache(state->block_cache);
1860 err_free_state:
1861         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1862                             lut_elm_size(state) * CODE_LUT_SIZE);
1863         free(state);
1864 err_finish_jit:
1865         finish_jit();
1866         if (ENABLE_CODE_BUFFER && tlsf)
1867                 tlsf_destroy(tlsf);
1868         return NULL;
1869 }
1870
1871 void lightrec_destroy(struct lightrec_state *state)
1872 {
1873         /* Force a print info on destroy*/
1874         state->current_cycle = ~state->current_cycle;
1875         lightrec_print_info(state);
1876
1877         lightrec_free_block_cache(state->block_cache);
1878         lightrec_free_block(state, state->dispatcher);
1879         lightrec_free_block(state, state->c_wrapper_block);
1880
1881         if (ENABLE_THREADED_COMPILER) {
1882                 lightrec_free_recompiler(state->rec);
1883                 lightrec_reaper_destroy(state->reaper);
1884         } else {
1885                 lightrec_free_cstate(state->cstate);
1886         }
1887
1888         finish_jit();
1889         if (ENABLE_CODE_BUFFER && state->tlsf)
1890                 tlsf_destroy(state->tlsf);
1891
1892         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1893                             lut_elm_size(state) * CODE_LUT_SIZE);
1894         free(state);
1895 }
1896
1897 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1898 {
1899         u32 kaddr = kunseg(addr & ~0x3);
1900         enum psx_map idx = lightrec_get_map_idx(state, kaddr);
1901
1902         switch (idx) {
1903         case PSX_MAP_MIRROR1:
1904         case PSX_MAP_MIRROR2:
1905         case PSX_MAP_MIRROR3:
1906                 /* Handle mirrors */
1907                 kaddr &= RAM_SIZE - 1;
1908                 fallthrough;
1909         case PSX_MAP_KERNEL_USER_RAM:
1910                 break;
1911         default:
1912                 return;
1913         }
1914
1915         memset(lut_address(state, lut_offset(kaddr)), 0,
1916                ((len + 3) / 4) * lut_elm_size(state));
1917 }
1918
1919 void lightrec_invalidate_all(struct lightrec_state *state)
1920 {
1921         memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
1922 }
1923
1924 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1925 {
1926         if (state->invalidate_from_dma_only != dma_only)
1927                 lightrec_invalidate_all(state);
1928
1929         state->invalidate_from_dma_only = dma_only;
1930 }
1931
1932 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1933 {
1934         if (flags != LIGHTREC_EXIT_NORMAL) {
1935                 state->exit_flags |= flags;
1936                 state->target_cycle = state->current_cycle;
1937         }
1938 }
1939
1940 u32 lightrec_exit_flags(struct lightrec_state *state)
1941 {
1942         return state->exit_flags;
1943 }
1944
1945 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1946 {
1947         return state->current_cycle;
1948 }
1949
1950 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1951 {
1952         state->current_cycle = cycles;
1953
1954         if (state->target_cycle < cycles)
1955                 state->target_cycle = cycles;
1956 }
1957
1958 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1959 {
1960         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1961                 if (cycles < state->current_cycle)
1962                         cycles = state->current_cycle;
1963
1964                 state->target_cycle = cycles;
1965         }
1966 }
1967
1968 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1969 {
1970         return &state->regs;
1971 }