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 #define GENMASK(h, l) \
  32         (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
  33
  34 static struct block * lightrec_precompile_block(struct lightrec_state *state,
  35                                                 u32 pc);
  36 static bool lightrec_block_is_fully_tagged(const struct block *block);
  37
  38 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
  39 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
  40
  41 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
  42                                 void *host, u32 addr, u8 data)
  43 {
  44         *(u8 *)host = data;
  45
  46         if (!state->invalidate_from_dma_only)
  47                 lightrec_invalidate(state, addr, 1);
  48 }
  49
  50 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
  51                                 void *host, u32 addr, u16 data)
  52 {
  53         *(u16 *)host = HTOLE16(data);
  54
  55         if (!state->invalidate_from_dma_only)
  56                 lightrec_invalidate(state, addr, 2);
  57 }
  58
  59 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
  60                                 void *host, u32 addr, u32 data)
  61 {
  62         *(u32 *)host = HTOLE32(data);
  63
  64         if (!state->invalidate_from_dma_only)
  65                 lightrec_invalidate(state, addr, 4);
  66 }
  67
  68 static u8 lightrec_default_lb(struct lightrec_state *state,
  69                               u32 opcode, void *host, u32 addr)
  70 {
  71         return *(u8 *)host;
  72 }
  73
  74 static u16 lightrec_default_lh(struct lightrec_state *state,
  75                                u32 opcode, void *host, u32 addr)
  76 {
  77         return LE16TOH(*(u16 *)host);
  78 }
  79
  80 static u32 lightrec_default_lw(struct lightrec_state *state,
  81                                u32 opcode, void *host, u32 addr)
  82 {
  83         return LE32TOH(*(u32 *)host);
  84 }
  85
  86 static const struct lightrec_mem_map_ops lightrec_default_ops = {
  87         .sb = lightrec_default_sb,
  88         .sh = lightrec_default_sh,
  89         .sw = lightrec_default_sw,
  90         .lb = lightrec_default_lb,
  91         .lh = lightrec_default_lh,
  92         .lw = lightrec_default_lw,
  93 };
  94
  95 static void __segfault_cb(struct lightrec_state *state, u32 addr,
  96                           const struct block *block)
  97 {
  98         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
  99         pr_err("Segmentation fault in recompiled code: invalid "
 100                "load/store at address 0x%08x\n", addr);
 101         if (block)
 102                 pr_err("Was executing block PC 0x%08x\n", block->pc);
 103 }
 104
 105 static void lightrec_swl(struct lightrec_state *state,
 106                          const struct lightrec_mem_map_ops *ops,
 107                          u32 opcode, void *host, u32 addr, u32 data)
 108 {
 109         unsigned int shift = addr & 0x3;
 110         unsigned int mask = GENMASK(31, (shift + 1) * 8);
 111         u32 old_data;
 112
 113         /* Align to 32 bits */
 114         addr &= ~3;
 115         host = (void *)((uintptr_t)host & ~3);
 116
 117         old_data = ops->lw(state, opcode, host, addr);
 118
 119         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
 120
 121         ops->sw(state, opcode, host, addr, data);
 122 }
 123
 124 static void lightrec_swr(struct lightrec_state *state,
 125                          const struct lightrec_mem_map_ops *ops,
 126                          u32 opcode, void *host, u32 addr, u32 data)
 127 {
 128         unsigned int shift = addr & 0x3;
 129         unsigned int mask = (1 << (shift * 8)) - 1;
 130         u32 old_data;
 131
 132         /* Align to 32 bits */
 133         addr &= ~3;
 134         host = (void *)((uintptr_t)host & ~3);
 135
 136         old_data = ops->lw(state, opcode, host, addr);
 137
 138         data = (data << (shift * 8)) | (old_data & mask);
 139
 140         ops->sw(state, opcode, host, addr, data);
 141 }
 142
 143 static void lightrec_swc2(struct lightrec_state *state, union code op,
 144                           const struct lightrec_mem_map_ops *ops,
 145                           void *host, u32 addr)
 146 {
 147         u32 data = lightrec_mfc2(state, op.i.rt);
 148
 149         ops->sw(state, op.opcode, host, addr, data);
 150 }
 151
 152 static u32 lightrec_lwl(struct lightrec_state *state,
 153                         const struct lightrec_mem_map_ops *ops,
 154                         u32 opcode, void *host, u32 addr, u32 data)
 155 {
 156         unsigned int shift = addr & 0x3;
 157         unsigned int mask = (1 << (24 - shift * 8)) - 1;
 158         u32 old_data;
 159
 160         /* Align to 32 bits */
 161         addr &= ~3;
 162         host = (void *)((uintptr_t)host & ~3);
 163
 164         old_data = ops->lw(state, opcode, host, addr);
 165
 166         return (data & mask) | (old_data << (24 - shift * 8));
 167 }
 168
 169 static u32 lightrec_lwr(struct lightrec_state *state,
 170                         const struct lightrec_mem_map_ops *ops,
 171                         u32 opcode, void *host, u32 addr, u32 data)
 172 {
 173         unsigned int shift = addr & 0x3;
 174         unsigned int mask = GENMASK(31, 32 - shift * 8);
 175         u32 old_data;
 176
 177         /* Align to 32 bits */
 178         addr &= ~3;
 179         host = (void *)((uintptr_t)host & ~3);
 180
 181         old_data = ops->lw(state, opcode, host, addr);
 182
 183         return (data & mask) | (old_data >> (shift * 8));
 184 }
 185
 186 static void lightrec_lwc2(struct lightrec_state *state, union code op,
 187                           const struct lightrec_mem_map_ops *ops,
 188                           void *host, u32 addr)
 189 {
 190         u32 data = ops->lw(state, op.opcode, host, addr);
 191
 192         lightrec_mtc2(state, op.i.rt, data);
 193 }
 194
 195 static void lightrec_invalidate_map(struct lightrec_state *state,
 196                 const struct lightrec_mem_map *map, u32 addr, u32 len)
 197 {
 198         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
 199                 memset(lut_address(state, lut_offset(addr)), 0,
 200                        ((len + 3) / 4) * lut_elm_size(state));
 201         }
 202 }
 203
 204 enum psx_map
 205 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
 206 {
 207         const struct lightrec_mem_map *map;
 208         unsigned int i;
 209
 210         for (i = 0; i < state->nb_maps; i++) {
 211                 map = &state->maps[i];
 212
 213                 if (kaddr >= map->pc && kaddr < map->pc + map->length)
 214                         return (enum psx_map) i;
 215         }
 216
 217         return PSX_MAP_UNKNOWN;
 218 }
 219
 220 const struct lightrec_mem_map *
 221 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 222 {
 223         const struct lightrec_mem_map *map;
 224         enum psx_map idx;
 225         u32 addr;
 226
 227         idx = lightrec_get_map_idx(state, kaddr);
 228         if (idx == PSX_MAP_UNKNOWN)
 229                 return NULL;
 230
 231         map = &state->maps[idx];
 232         addr = kaddr - map->pc;
 233
 234         while (map->mirror_of)
 235                 map = map->mirror_of;
 236
 237         if (host)
 238                 *host = map->address + addr;
 239
 240         return map;
 241 }
 242
 243 u32 lightrec_rw(struct lightrec_state *state, union code op,
 244                 u32 addr, u32 data, u32 *flags, struct block *block)
 245 {
 246         const struct lightrec_mem_map *map;
 247         const struct lightrec_mem_map_ops *ops;
 248         u32 opcode = op.opcode;
 249         void *host;
 250
 251         addr += (s16) op.i.imm;
 252
 253         map = lightrec_get_map(state, &host, kunseg(addr));
 254         if (!map) {
 255                 __segfault_cb(state, addr, block);
 256                 return 0;
 257         }
 258
 259         if (unlikely(map->ops)) {
 260                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 261                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 262
 263                 ops = map->ops;
 264         } else {
 265                 if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 266                         *flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
 267
 268                 ops = &lightrec_default_ops;
 269         }
 270
 271         switch (op.i.op) {
 272         case OP_SB:
 273                 ops->sb(state, opcode, host, addr, (u8) data);
 274                 return 0;
 275         case OP_SH:
 276                 ops->sh(state, opcode, host, addr, (u16) data);
 277                 return 0;
 278         case OP_SWL:
 279                 lightrec_swl(state, ops, opcode, host, addr, data);
 280                 return 0;
 281         case OP_SWR:
 282                 lightrec_swr(state, ops, opcode, host, addr, data);
 283                 return 0;
 284         case OP_SW:
 285                 ops->sw(state, opcode, host, addr, data);
 286                 return 0;
 287         case OP_SWC2:
 288                 lightrec_swc2(state, op, ops, host, addr);
 289                 return 0;
 290         case OP_LB:
 291                 return (s32) (s8) ops->lb(state, opcode, host, addr);
 292         case OP_LBU:
 293                 return ops->lb(state, opcode, host, addr);
 294         case OP_LH:
 295                 return (s32) (s16) ops->lh(state, opcode, host, addr);
 296         case OP_LHU:
 297                 return ops->lh(state, opcode, host, addr);
 298         case OP_LWC2:
 299                 lightrec_lwc2(state, op, ops, host, addr);
 300                 return 0;
 301         case OP_LWL:
 302                 return lightrec_lwl(state, ops, opcode, host, addr, data);
 303         case OP_LWR:
 304                 return lightrec_lwr(state, ops, opcode, host, addr, data);
 305         case OP_LW:
 306         default:
 307                 return ops->lw(state, opcode, host, addr);
 308         }
 309 }
 310
 311 static void lightrec_rw_helper(struct lightrec_state *state,
 312                                union code op, u32 *flags,
 313                                struct block *block)
 314 {
 315         u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
 316                               state->regs.gpr[op.i.rt], flags, block);
 317
 318         switch (op.i.op) {
 319         case OP_LB:
 320         case OP_LBU:
 321         case OP_LH:
 322         case OP_LHU:
 323         case OP_LWL:
 324         case OP_LWR:
 325         case OP_LW:
 326                 if (op.i.rt)
 327                         state->regs.gpr[op.i.rt] = ret;
 328                 fallthrough;
 329         default:
 330                 break;
 331         }
 332 }
 333
 334 static void lightrec_rw_cb(struct lightrec_state *state, u32 arg)
 335 {
 336         lightrec_rw_helper(state, (union code) arg, NULL, NULL);
 337 }
 338
 339 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 340 {
 341         struct block *block;
 342         struct opcode *op;
 343         bool was_tagged;
 344         u16 offset = (u16)arg;
 345
 346         block = lightrec_find_block_from_lut(state->block_cache,
 347                                              arg >> 16, state->next_pc);
 348         if (unlikely(!block)) {
 349                 pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
 350                          state->next_pc, offset);
 351                 return;
 352         }
 353
 354         op = &block->opcode_list[offset];
 355         was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(op->flags);
 356
 357         lightrec_rw_helper(state, op->c, &op->flags, block);
 358
 359         if (!was_tagged) {
 360                 pr_debug("Opcode of block at PC 0x%08x has been tagged - flag "
 361                          "for recompilation\n", block->pc);
 362
 363                 block->flags |= BLOCK_SHOULD_RECOMPILE;
 364                 lut_write(state, lut_offset(block->pc), NULL);
 365         }
 366 }
 367
 368 static u32 clamp_s32(s32 val, s32 min, s32 max)
 369 {
 370         return val < min ? min : val > max ? max : val;
 371 }
 372
 373 static u16 load_u16(u32 *ptr)
 374 {
 375         return ((struct u16x2 *) ptr)->l;
 376 }
 377
 378 static void store_u16(u32 *ptr, u16 value)
 379 {
 380         ((struct u16x2 *) ptr)->l = value;
 381 }
 382
 383 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
 384 {
 385         s16 gteir1, gteir2, gteir3;
 386
 387         switch (reg) {
 388         case 1:
 389         case 3:
 390         case 5:
 391         case 8:
 392         case 9:
 393         case 10:
 394         case 11:
 395                 return (s32)(s16) load_u16(&state->regs.cp2d[reg]);
 396         case 7:
 397         case 16:
 398         case 17:
 399         case 18:
 400         case 19:
 401                 return load_u16(&state->regs.cp2d[reg]);
 402         case 28:
 403         case 29:
 404                 gteir1 = (s16) load_u16(&state->regs.cp2d[9]);
 405                 gteir2 = (s16) load_u16(&state->regs.cp2d[10]);
 406                 gteir3 = (s16) load_u16(&state->regs.cp2d[11]);
 407
 408                 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
 409                         clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
 410                         clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
 411         case 15:
 412                 reg = 14;
 413                 fallthrough;
 414         default:
 415                 return state->regs.cp2d[reg];
 416         }
 417 }
 418
 419 u32 lightrec_mfc(struct lightrec_state *state, union code op)
 420 {
 421         if (op.i.op == OP_CP0)
 422                 return state->regs.cp0[op.r.rd];
 423         else if (op.r.rs == OP_CP2_BASIC_MFC2)
 424                 return lightrec_mfc2(state, op.r.rd);
 425         else
 426                 return state->regs.cp2c[op.r.rd];
 427 }
 428
 429 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 430 {
 431         u32 status, oldstatus, cause;
 432
 433         switch (reg) {
 434         case 1:
 435         case 4:
 436         case 8:
 437         case 14:
 438         case 15:
 439                 /* Those registers are read-only */
 440                 return;
 441         default:
 442                 break;
 443         }
 444
 445         if (reg == 12) {
 446                 status = state->regs.cp0[12];
 447                 oldstatus = status;
 448
 449                 if (status & ~data & BIT(16)) {
 450                         state->ops.enable_ram(state, true);
 451                         lightrec_invalidate_all(state);
 452                 } else if (~status & data & BIT(16)) {
 453                         state->ops.enable_ram(state, false);
 454                 }
 455         }
 456
 457         if (reg == 13) {
 458                 state->regs.cp0[13] &= ~0x300;
 459                 state->regs.cp0[13] |= data & 0x300;
 460         } else {
 461                 state->regs.cp0[reg] = data;
 462         }
 463
 464         if (reg == 12 || reg == 13) {
 465                 cause = state->regs.cp0[13];
 466                 status = state->regs.cp0[12];
 467
 468                 /* Handle software interrupts */
 469                 if (!!(status & cause & 0x300) & status)
 470                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 471
 472                 /* Handle hardware interrupts */
 473                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
 474                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 475         }
 476 }
 477
 478 static u32 count_leading_bits(s32 data)
 479 {
 480         u32 cnt = 33;
 481
 482 #ifdef __has_builtin
 483 #if __has_builtin(__builtin_clrsb)
 484         return 1 + __builtin_clrsb(data);
 485 #endif
 486 #endif
 487
 488         data = (data ^ (data >> 31)) << 1;
 489
 490         do {
 491                 cnt -= 1;
 492                 data >>= 1;
 493         } while (data);
 494
 495         return cnt;
 496 }
 497
 498 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
 499 {
 500         switch (reg) {
 501         case 15:
 502                 state->regs.cp2d[12] = state->regs.cp2d[13];
 503                 state->regs.cp2d[13] = state->regs.cp2d[14];
 504                 state->regs.cp2d[14] = data;
 505                 break;
 506         case 28:
 507                 state->regs.cp2d[9] = (data << 7) & 0xf80;
 508                 state->regs.cp2d[10] = (data << 2) & 0xf80;
 509                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
 510                 break;
 511         case 31:
 512                 return;
 513         case 30:
 514                 state->regs.cp2d[31] = count_leading_bits((s32) data);
 515                 fallthrough;
 516         default:
 517                 state->regs.cp2d[reg] = data;
 518                 break;
 519         }
 520 }
 521
 522 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 523 {
 524         switch (reg) {
 525         case 4:
 526         case 12:
 527         case 20:
 528         case 26:
 529         case 27:
 530         case 29:
 531         case 30:
 532                 store_u16(&state->regs.cp2c[reg], data);
 533                 break;
 534         case 31:
 535                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
 536                 fallthrough;
 537         default:
 538                 state->regs.cp2c[reg] = data;
 539                 break;
 540         }
 541 }
 542
 543 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
 544 {
 545         if (op.i.op == OP_CP0)
 546                 lightrec_mtc0(state, op.r.rd, data);
 547         else if (op.r.rs == OP_CP2_BASIC_CTC2)
 548                 lightrec_ctc2(state, op.r.rd, data);
 549         else
 550                 lightrec_mtc2(state, op.r.rd, data);
 551 }
 552
 553 static void lightrec_mtc_cb(struct lightrec_state *state, u32 arg)
 554 {
 555         union code op = (union code) arg;
 556
 557         lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
 558 }
 559
 560 void lightrec_rfe(struct lightrec_state *state)
 561 {
 562         u32 status;
 563
 564         /* Read CP0 Status register (r12) */
 565         status = state->regs.cp0[12];
 566
 567         /* Switch the bits */
 568         status = ((status & 0x3c) >> 2) | (status & ~0xf);
 569
 570         /* Write it back */
 571         lightrec_mtc0(state, 12, status);
 572 }
 573
 574 void lightrec_cp(struct lightrec_state *state, union code op)
 575 {
 576         if (op.i.op == OP_CP0) {
 577                 pr_err("Invalid CP opcode to coprocessor #0\n");
 578                 return;
 579         }
 580
 581         (*state->ops.cop2_op)(state, op.opcode);
 582 }
 583
 584 static void lightrec_cp_cb(struct lightrec_state *state, u32 arg)
 585 {
 586         lightrec_cp(state, (union code) arg);
 587 }
 588
 589 static void lightrec_syscall_cb(struct lightrec_state *state)
 590 {
 591         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
 592 }
 593
 594 static void lightrec_break_cb(struct lightrec_state *state)
 595 {
 596         lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
 597 }
 598
 599 static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 600 {
 601         struct block *block = lightrec_find_block(state->block_cache, pc);
 602
 603         if (block && lightrec_block_is_outdated(state, block)) {
 604                 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
 605
 606                 /* Make sure the recompiler isn't processing the block we'll
 607                  * destroy */
 608                 if (ENABLE_THREADED_COMPILER)
 609                         lightrec_recompiler_remove(state->rec, block);
 610
 611                 lightrec_unregister_block(state->block_cache, block);
 612                 remove_from_code_lut(state->block_cache, block);
 613                 lightrec_free_block(state, block);
 614                 block = NULL;
 615         }
 616
 617         if (!block) {
 618                 block = lightrec_precompile_block(state, pc);
 619                 if (!block) {
 620                         pr_err("Unable to recompile block at PC 0x%x\n", pc);
 621                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 622                         return NULL;
 623                 }
 624
 625                 lightrec_register_block(state->block_cache, block);
 626         }
 627
 628         return block;
 629 }
 630
 631 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 632 {
 633         struct block *block;
 634         bool should_recompile;
 635         void *func;
 636         int err;
 637
 638         for (;;) {
 639                 func = lut_read(state, lut_offset(pc));
 640                 if (func && func != state->get_next_block)
 641                         break;
 642
 643                 block = lightrec_get_block(state, pc);
 644
 645                 if (unlikely(!block))
 646                         break;
 647
 648                 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
 649                         func = state->memset_func;
 650                         break;
 651                 }
 652
 653                 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
 654                         !(block->flags & BLOCK_IS_DEAD);
 655
 656                 if (unlikely(should_recompile)) {
 657                         pr_debug("Block at PC 0x%08x should recompile\n", pc);
 658
 659                         lightrec_unregister(MEM_FOR_CODE, block->code_size);
 660
 661                         if (ENABLE_THREADED_COMPILER) {
 662                                 lightrec_recompiler_add(state->rec, block);
 663                         } else {
 664                                 err = lightrec_compile_block(state->cstate, block);
 665                                 if (err) {
 666                                         state->exit_flags = LIGHTREC_EXIT_NOMEM;
 667                                         return NULL;
 668                                 }
 669                         }
 670                 }
 671
 672                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
 673                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
 674                 else
 675                         func = block->function;
 676
 677                 if (likely(func))
 678                         break;
 679
 680                 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
 681                         pc = lightrec_emulate_block(state, block, pc);
 682
 683                 } else if (!ENABLE_THREADED_COMPILER) {
 684                         /* Block wasn't compiled yet - run the interpreter */
 685                         if (block->flags & BLOCK_FULLY_TAGGED)
 686                                 pr_debug("Block fully tagged, skipping first pass\n");
 687                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 688                                 pc = lightrec_emulate_block(state, block, pc);
 689
 690                         /* Then compile it using the profiled data */
 691                         err = lightrec_compile_block(state->cstate, block);
 692                         if (err) {
 693                                 state->exit_flags = LIGHTREC_EXIT_NOMEM;
 694                                 return NULL;
 695                         }
 696                 } else {
 697                         lightrec_recompiler_add(state->rec, block);
 698                 }
 699
 700                 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
 701                     state->current_cycle >= state->target_cycle)
 702                         break;
 703         }
 704
 705         state->next_pc = pc;
 706         return func;
 707 }
 708
 709 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
 710                               void (*f)(struct lightrec_state *, u32), u32 arg)
 711 {
 712         state->current_cycle = state->target_cycle - cycles_delta;
 713
 714         (*f)(state, arg);
 715
 716         return state->target_cycle - state->current_cycle;
 717 }
 718
 719 static void * lightrec_alloc_code(struct lightrec_state *state, size_t size)
 720 {
 721         void *code;
 722
 723         if (ENABLE_THREADED_COMPILER)
 724                 lightrec_code_alloc_lock(state);
 725
 726         code = tlsf_malloc(state->tlsf, size);
 727
 728         if (ENABLE_THREADED_COMPILER)
 729                 lightrec_code_alloc_unlock(state);
 730
 731         return code;
 732 }
 733
 734 static void lightrec_realloc_code(struct lightrec_state *state,
 735                                   void *ptr, size_t size)
 736 {
 737         /* NOTE: 'size' MUST be smaller than the size specified during
 738          * the allocation. */
 739
 740         if (ENABLE_THREADED_COMPILER)
 741                 lightrec_code_alloc_lock(state);
 742
 743         tlsf_realloc(state->tlsf, ptr, size);
 744
 745         if (ENABLE_THREADED_COMPILER)
 746                 lightrec_code_alloc_unlock(state);
 747 }
 748
 749 static void lightrec_free_code(struct lightrec_state *state, void *ptr)
 750 {
 751         if (ENABLE_THREADED_COMPILER)
 752                 lightrec_code_alloc_lock(state);
 753
 754         tlsf_free(state->tlsf, ptr);
 755
 756         if (ENABLE_THREADED_COMPILER)
 757                 lightrec_code_alloc_unlock(state);
 758 }
 759
 760 static void * lightrec_emit_code(struct lightrec_state *state,
 761                                  const struct block *block,
 762                                  jit_state_t *_jit, unsigned int *size)
 763 {
 764         bool has_code_buffer = ENABLE_CODE_BUFFER && state->tlsf;
 765         jit_word_t code_size, new_code_size;
 766         void *code;
 767
 768         jit_realize();
 769
 770         if (!ENABLE_DISASSEMBLER)
 771                 jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
 772
 773         if (has_code_buffer) {
 774                 jit_get_code(&code_size);
 775                 code = lightrec_alloc_code(state, (size_t) code_size);
 776
 777                 if (!code) {
 778                         if (ENABLE_THREADED_COMPILER) {
 779                                 /* If we're using the threaded compiler, return
 780                                  * an allocation error here. The threaded
 781                                  * compiler will then empty its job queue and
 782                                  * request a code flush using the reaper. */
 783                                 return NULL;
 784                         }
 785
 786                         /* Remove outdated blocks, and try again */
 787                         lightrec_remove_outdated_blocks(state->block_cache, block);
 788
 789                         pr_debug("Re-try to alloc %zu bytes...\n", code_size);
 790
 791                         code = lightrec_alloc_code(state, code_size);
 792                         if (!code) {
 793                                 pr_err("Could not alloc even after removing old blocks!\n");
 794                                 return NULL;
 795                         }
 796                 }
 797
 798                 jit_set_code(code, code_size);
 799         }
 800
 801         code = jit_emit();
 802
 803         jit_get_code(&new_code_size);
 804         lightrec_register(MEM_FOR_CODE, new_code_size);
 805
 806         if (has_code_buffer) {
 807                 lightrec_realloc_code(state, code, (size_t) new_code_size);
 808
 809                 pr_debug("Creating code block at address 0x%" PRIxPTR ", "
 810                          "code size: %" PRIuPTR " new: %" PRIuPTR "\n",
 811                          (uintptr_t) code, code_size, new_code_size);
 812         }
 813
 814         *size = (unsigned int) new_code_size;
 815
 816         return code;
 817 }
 818
 819 static struct block * generate_wrapper(struct lightrec_state *state)
 820 {
 821         struct block *block;
 822         jit_state_t *_jit;
 823         unsigned int i;
 824         int stack_ptr;
 825         jit_node_t *to_tramp, *to_fn_epilog;
 826         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
 827
 828         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 829         if (!block)
 830                 goto err_no_mem;
 831
 832         _jit = jit_new_state();
 833         if (!_jit)
 834                 goto err_free_block;
 835
 836         jit_name("RW wrapper");
 837         jit_note(__FILE__, __LINE__);
 838
 839         /* Wrapper entry point */
 840         jit_prolog();
 841         jit_tramp(256);
 842
 843         /* Add entry points; separate them by opcodes that increment
 844          * LIGHTREC_REG_STATE (since we cannot touch other registers).
 845          * The difference will then tell us which C function to call. */
 846         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
 847                 jit_addi(LIGHTREC_REG_STATE, LIGHTREC_REG_STATE, __WORDSIZE / 8);
 848                 addr[i - 1] = jit_indirect();
 849         }
 850
 851         jit_epilog();
 852         jit_prolog();
 853
 854         stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
 855
 856         /* Save all temporaries on stack */
 857         for (i = 0; i < NUM_TEMPS; i++)
 858                 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
 859
 860         jit_getarg(JIT_R1, jit_arg());
 861
 862         /* Jump to the trampoline */
 863         to_tramp = jit_jmpi();
 864
 865         /* The trampoline will jump back here */
 866         to_fn_epilog = jit_label();
 867
 868         /* Restore temporaries from stack */
 869         for (i = 0; i < NUM_TEMPS; i++)
 870                 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
 871
 872         jit_ret();
 873         jit_epilog();
 874
 875         /* Trampoline entry point.
 876          * The sole purpose of the trampoline is to cheese Lightning not to
 877          * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
 878          * do want to return to the caller with this register modified. */
 879         jit_prolog();
 880         jit_tramp(256);
 881         jit_patch(to_tramp);
 882
 883         /* Retrieve the wrapper function */
 884         jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
 885                  offsetof(struct lightrec_state, c_wrappers));
 886
 887         /* Restore LIGHTREC_REG_STATE to its correct value */
 888         jit_movi(LIGHTREC_REG_STATE, (uintptr_t) state);
 889
 890         jit_prepare();
 891         jit_pushargr(LIGHTREC_REG_STATE);
 892         jit_pushargr(LIGHTREC_REG_CYCLE);
 893         jit_pushargr(JIT_R0);
 894         jit_pushargr(JIT_R1);
 895         jit_finishi(c_function_wrapper);
 896         jit_retval_i(LIGHTREC_REG_CYCLE);
 897
 898         jit_patch_at(jit_jmpi(), to_fn_epilog);
 899         jit_epilog();
 900
 901         block->_jit = _jit;
 902         block->opcode_list = NULL;
 903         block->flags = 0;
 904         block->nb_ops = 0;
 905
 906         block->function = lightrec_emit_code(state, block, _jit,
 907                                              &block->code_size);
 908         if (!block->function)
 909                 goto err_free_block;
 910
 911         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
 912
 913         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 914                 state->wrappers_eps[i] = jit_address(addr[i]);
 915
 916         if (ENABLE_DISASSEMBLER) {
 917                 pr_debug("Wrapper block:\n");
 918                 jit_disassemble();
 919         }
 920
 921         jit_clear_state();
 922         return block;
 923
 924 err_free_block:
 925         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
 926 err_no_mem:
 927         pr_err("Unable to compile wrapper: Out of memory\n");
 928         return NULL;
 929 }
 930
 931 static u32 lightrec_memset(struct lightrec_state *state)
 932 {
 933         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
 934         void *host;
 935         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
 936         u32 length = state->regs.gpr[5] * 4;
 937
 938         if (!map) {
 939                 pr_err("Unable to find memory map for memset target address "
 940                        "0x%x\n", kunseg_pc);
 941                 return 0;
 942         }
 943
 944         pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
 945                  kunseg_pc, (uintptr_t)host, length);
 946         memset(host, 0, length);
 947
 948         if (!state->invalidate_from_dma_only)
 949                 lightrec_invalidate_map(state, map, kunseg_pc, length);
 950
 951         /* Rough estimation of the number of cycles consumed */
 952         return 8 + 5 * (length  + 3 / 4);
 953 }
 954
 955 static struct block * generate_dispatcher(struct lightrec_state *state)
 956 {
 957         struct block *block;
 958         jit_state_t *_jit;
 959         jit_node_t *to_end, *loop, *addr, *addr2, *addr3;
 960         unsigned int i;
 961         u32 offset;
 962
 963         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 964         if (!block)
 965                 goto err_no_mem;
 966
 967         _jit = jit_new_state();
 968         if (!_jit)
 969                 goto err_free_block;
 970
 971         jit_name("dispatcher");
 972         jit_note(__FILE__, __LINE__);
 973
 974         jit_prolog();
 975         jit_frame(256);
 976
 977         jit_getarg(JIT_R0, jit_arg());
 978         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
 979
 980         /* Force all callee-saved registers to be pushed on the stack */
 981         for (i = 0; i < NUM_REGS; i++)
 982                 jit_movr(JIT_V(i), JIT_V(i));
 983
 984         /* Pass lightrec_state structure to blocks, using the last callee-saved
 985          * register that Lightning provides */
 986         jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
 987
 988         loop = jit_label();
 989
 990         /* Call the block's code */
 991         jit_jmpr(JIT_R0);
 992
 993         if (OPT_REPLACE_MEMSET) {
 994                 /* Blocks will jump here when they need to call
 995                  * lightrec_memset() */
 996                 addr3 = jit_indirect();
 997
 998                 jit_prepare();
 999                 jit_pushargr(LIGHTREC_REG_STATE);
1000                 jit_finishi(lightrec_memset);
1001
1002                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
1003                             offsetof(struct lightrec_state, regs.gpr[31]));
1004
1005                 jit_retval(JIT_R0);
1006                 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
1007         }
1008
1009         /* The block will jump here, with the number of cycles remaining in
1010          * LIGHTREC_REG_CYCLE */
1011         addr2 = jit_indirect();
1012
1013         /* Store back the next_pc to the lightrec_state structure */
1014         offset = offsetof(struct lightrec_state, next_pc);
1015         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
1016
1017         /* Jump to end if state->target_cycle < state->current_cycle */
1018         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
1019
1020         /* Convert next PC to KUNSEG and avoid mirrors */
1021         jit_andi(JIT_R0, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
1022         jit_rshi_u(JIT_R1, JIT_R0, 28);
1023         jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
1024         jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
1025         jit_movnr(JIT_R0, JIT_R2, JIT_R1);
1026
1027         /* If possible, use the code LUT */
1028         if (!lut_is_32bit(state))
1029                 jit_lshi(JIT_R0, JIT_R0, 1);
1030         jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
1031
1032         offset = offsetof(struct lightrec_state, code_lut);
1033         if (lut_is_32bit(state))
1034                 jit_ldxi_ui(JIT_R0, JIT_R0, offset);
1035         else
1036                 jit_ldxi(JIT_R0, JIT_R0, offset);
1037
1038         /* If we get non-NULL, loop */
1039         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
1040
1041         /* Slow path: call C function get_next_block_func() */
1042
1043         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1044                 /* We may call the interpreter - update state->current_cycle */
1045                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1046                            offsetof(struct lightrec_state, target_cycle));
1047                 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
1048                 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
1049                            LIGHTREC_REG_STATE, JIT_R1);
1050         }
1051
1052         /* The code LUT will be set to this address when the block at the target
1053          * PC has been preprocessed but not yet compiled by the threaded
1054          * recompiler */
1055         addr = jit_indirect();
1056
1057         /* Get the next block */
1058         jit_prepare();
1059         jit_pushargr(LIGHTREC_REG_STATE);
1060         jit_pushargr(JIT_V0);
1061         jit_finishi(&get_next_block_func);
1062         jit_retval(JIT_R0);
1063
1064         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
1065                 /* The interpreter may have updated state->current_cycle and
1066                  * state->target_cycle - recalc the delta */
1067                 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
1068                            offsetof(struct lightrec_state, current_cycle));
1069                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
1070                            offsetof(struct lightrec_state, target_cycle));
1071                 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
1072         }
1073
1074         /* If we get non-NULL, loop */
1075         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
1076
1077         /* When exiting, the recompiled code will jump to that address */
1078         jit_note(__FILE__, __LINE__);
1079         jit_patch(to_end);
1080
1081         jit_retr(LIGHTREC_REG_CYCLE);
1082         jit_epilog();
1083
1084         block->_jit = _jit;
1085         block->opcode_list = NULL;
1086         block->flags = 0;
1087         block->nb_ops = 0;
1088
1089         block->function = lightrec_emit_code(state, block, _jit,
1090                                              &block->code_size);
1091         if (!block->function)
1092                 goto err_free_block;
1093
1094         state->eob_wrapper_func = jit_address(addr2);
1095         if (OPT_REPLACE_MEMSET)
1096                 state->memset_func = jit_address(addr3);
1097         state->get_next_block = jit_address(addr);
1098
1099         if (ENABLE_DISASSEMBLER) {
1100                 pr_debug("Dispatcher block:\n");
1101                 jit_disassemble();
1102         }
1103
1104         /* We're done! */
1105         jit_clear_state();
1106         return block;
1107
1108 err_free_block:
1109         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1110 err_no_mem:
1111         pr_err("Unable to compile dispatcher: Out of memory\n");
1112         return NULL;
1113 }
1114
1115 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
1116 {
1117         void *host = NULL;
1118
1119         lightrec_get_map(state, &host, kunseg(pc));
1120
1121         const u32 *code = (u32 *)host;
1122         return (union code) LE32TOH(*code);
1123 }
1124
1125 unsigned int lightrec_cycles_of_opcode(union code code)
1126 {
1127         return 2;
1128 }
1129
1130 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
1131 {
1132         lightrec_free(state, MEM_FOR_IR,
1133                       sizeof(*block->opcode_list) * block->nb_ops,
1134                       block->opcode_list);
1135 }
1136
1137 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1138 {
1139         unsigned int i;
1140         union code c;
1141
1142         for (i = 1; ; i++) {
1143                 c.opcode = LE32TOH(*src++);
1144
1145                 if (is_syscall(c))
1146                         return i;
1147
1148                 if (is_unconditional_jump(c))
1149                         return i + 1;
1150         }
1151 }
1152
1153 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1154                                             const u32 *src, unsigned int *len)
1155 {
1156         struct opcode *list;
1157         unsigned int i, length;
1158
1159         length = lightrec_get_mips_block_len(src);
1160
1161         list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
1162         if (!list) {
1163                 pr_err("Unable to allocate memory\n");
1164                 return NULL;
1165         }
1166
1167         for (i = 0; i < length; i++) {
1168                 list[i].opcode = LE32TOH(src[i]);
1169                 list[i].flags = 0;
1170         }
1171
1172         *len = length * sizeof(u32);
1173
1174         return list;
1175 }
1176
1177 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1178                                                 u32 pc)
1179 {
1180         struct opcode *list;
1181         struct block *block;
1182         void *host;
1183         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1184         const u32 *code = (u32 *) host;
1185         unsigned int length;
1186         bool fully_tagged;
1187
1188         if (!map)
1189                 return NULL;
1190
1191         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1192         if (!block) {
1193                 pr_err("Unable to recompile block: Out of memory\n");
1194                 return NULL;
1195         }
1196
1197         list = lightrec_disassemble(state, code, &length);
1198         if (!list) {
1199                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1200                 return NULL;
1201         }
1202
1203         block->pc = pc;
1204         block->_jit = NULL;
1205         block->function = NULL;
1206         block->opcode_list = list;
1207         block->code = code;
1208         block->next = NULL;
1209         block->flags = 0;
1210         block->code_size = 0;
1211         block->precompile_date = state->current_cycle;
1212 #if ENABLE_THREADED_COMPILER
1213         block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1214 #endif
1215         block->nb_ops = length / sizeof(u32);
1216
1217         lightrec_optimize(state, block);
1218
1219         length = block->nb_ops * sizeof(u32);
1220
1221         lightrec_register(MEM_FOR_MIPS_CODE, length);
1222
1223         if (ENABLE_DISASSEMBLER) {
1224                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1225                 lightrec_print_disassembly(block, code);
1226         }
1227
1228         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1229
1230         /* If the first opcode is an 'impossible' branch, never compile the
1231          * block */
1232         if (should_emulate(block->opcode_list))
1233                 block->flags |= BLOCK_NEVER_COMPILE;
1234
1235         fully_tagged = lightrec_block_is_fully_tagged(block);
1236         if (fully_tagged)
1237                 block->flags |= BLOCK_FULLY_TAGGED;
1238
1239         if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1240                 lut_write(state, lut_offset(pc), state->memset_func);
1241
1242         block->hash = lightrec_calculate_block_hash(block);
1243
1244         pr_debug("Recompile count: %u\n", state->nb_precompile++);
1245
1246         return block;
1247 }
1248
1249 static bool lightrec_block_is_fully_tagged(const struct block *block)
1250 {
1251         const struct opcode *op;
1252         unsigned int i;
1253
1254         for (i = 0; i < block->nb_ops; i++) {
1255                 op = &block->opcode_list[i];
1256
1257                 /* Verify that all load/stores of the opcode list
1258                  * Check all loads/stores of the opcode list and mark the
1259                  * block as fully compiled if they all have been tagged. */
1260                 switch (op->c.i.op) {
1261                 case OP_LB:
1262                 case OP_LH:
1263                 case OP_LWL:
1264                 case OP_LW:
1265                 case OP_LBU:
1266                 case OP_LHU:
1267                 case OP_LWR:
1268                 case OP_SB:
1269                 case OP_SH:
1270                 case OP_SWL:
1271                 case OP_SW:
1272                 case OP_SWR:
1273                 case OP_LWC2:
1274                 case OP_SWC2:
1275                         if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
1276                                 return false;
1277                         fallthrough;
1278                 default:
1279                         continue;
1280                 }
1281         }
1282
1283         return true;
1284 }
1285
1286 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1287 {
1288         struct block *block = data;
1289
1290         pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1291         lightrec_unregister_block(state->block_cache, block);
1292         lightrec_free_block(state, block);
1293 }
1294
1295 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1296 {
1297         _jit_destroy_state(data);
1298 }
1299
1300 static void lightrec_free_function(struct lightrec_state *state, void *fn)
1301 {
1302         if (ENABLE_CODE_BUFFER && state->tlsf) {
1303                 pr_debug("Freeing code block at 0x%" PRIxPTR "\n", (uintptr_t) fn);
1304                 lightrec_free_code(state, fn);
1305         }
1306 }
1307
1308 static void lightrec_reap_function(struct lightrec_state *state, void *data)
1309 {
1310         lightrec_free_function(state, data);
1311 }
1312
1313 int lightrec_compile_block(struct lightrec_cstate *cstate,
1314                            struct block *block)
1315 {
1316         struct lightrec_state *state = cstate->state;
1317         struct lightrec_branch_target *target;
1318         bool op_list_freed = false, fully_tagged = false;
1319         struct block *block2;
1320         struct opcode *elm;
1321         jit_state_t *_jit, *oldjit;
1322         jit_node_t *start_of_block;
1323         bool skip_next = false;
1324         void *old_fn, *new_fn;
1325         unsigned int i, j;
1326         u32 offset;
1327
1328         fully_tagged = lightrec_block_is_fully_tagged(block);
1329         if (fully_tagged)
1330                 block->flags |= BLOCK_FULLY_TAGGED;
1331
1332         _jit = jit_new_state();
1333         if (!_jit)
1334                 return -ENOMEM;
1335
1336         oldjit = block->_jit;
1337         old_fn = block->function;
1338         block->_jit = _jit;
1339
1340         lightrec_regcache_reset(cstate->reg_cache);
1341         cstate->cycles = 0;
1342         cstate->nb_branches = 0;
1343         cstate->nb_local_branches = 0;
1344         cstate->nb_targets = 0;
1345
1346         jit_prolog();
1347         jit_tramp(256);
1348
1349         start_of_block = jit_label();
1350
1351         for (i = 0; i < block->nb_ops; i++) {
1352                 elm = &block->opcode_list[i];
1353
1354                 if (skip_next) {
1355                         skip_next = false;
1356                         continue;
1357                 }
1358
1359                 if (should_emulate(elm)) {
1360                         pr_debug("Branch at offset 0x%x will be emulated\n",
1361                                  i << 2);
1362
1363                         lightrec_emit_eob(cstate, block, i, false);
1364                         skip_next = !op_flag_no_ds(elm->flags);
1365                 } else {
1366                         lightrec_rec_opcode(cstate, block, i);
1367                         skip_next = !op_flag_no_ds(elm->flags) && has_delay_slot(elm->c);
1368 #if _WIN32
1369                         /* FIXME: GNU Lightning on Windows seems to use our
1370                          * mapped registers as temporaries. Until the actual bug
1371                          * is found and fixed, unconditionally mark our
1372                          * registers as live here. */
1373                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1374 #endif
1375                 }
1376
1377                 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1378         }
1379
1380         for (i = 0; i < cstate->nb_branches; i++)
1381                 jit_patch(cstate->branches[i]);
1382
1383         for (i = 0; i < cstate->nb_local_branches; i++) {
1384                 struct lightrec_branch *branch = &cstate->local_branches[i];
1385
1386                 pr_debug("Patch local branch to offset 0x%x\n",
1387                          branch->target << 2);
1388
1389                 if (branch->target == 0) {
1390                         jit_patch_at(branch->branch, start_of_block);
1391                         continue;
1392                 }
1393
1394                 for (j = 0; j < cstate->nb_targets; j++) {
1395                         if (cstate->targets[j].offset == branch->target) {
1396                                 jit_patch_at(branch->branch,
1397                                              cstate->targets[j].label);
1398                                 break;
1399                         }
1400                 }
1401
1402                 if (j == cstate->nb_targets)
1403                         pr_err("Unable to find branch target\n");
1404         }
1405
1406         jit_patch_abs(jit_jmpi(), state->eob_wrapper_func);
1407         jit_ret();
1408         jit_epilog();
1409
1410         new_fn = lightrec_emit_code(state, block, _jit, &block->code_size);
1411         if (!new_fn) {
1412                 if (!ENABLE_THREADED_COMPILER)
1413                         pr_err("Unable to compile block!\n");
1414                 block->_jit = oldjit;
1415                 _jit_destroy_state(_jit);
1416                 return -ENOMEM;
1417         }
1418
1419         block->function = new_fn;
1420         block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1421
1422         /* Add compiled function to the LUT */
1423         lut_write(state, lut_offset(block->pc), block->function);
1424
1425         if (ENABLE_THREADED_COMPILER) {
1426                 /* Since we might try to reap the same block multiple times,
1427                  * we need the reaper to wait until everything has been
1428                  * submitted, so that the duplicate entries can be dropped. */
1429                 lightrec_reaper_pause(state->reaper);
1430         }
1431
1432         /* Detect old blocks that have been covered by the new one */
1433         for (i = 0; i < cstate->nb_targets; i++) {
1434                 target = &cstate->targets[i];
1435
1436                 if (!target->offset)
1437                         continue;
1438
1439                 offset = block->pc + target->offset * sizeof(u32);
1440                 block2 = lightrec_find_block(state->block_cache, offset);
1441                 if (block2) {
1442                         /* No need to check if block2 is compilable - it must
1443                          * be, otherwise block wouldn't be compilable either */
1444
1445                         /* Set the "block dead" flag to prevent the dynarec from
1446                          * recompiling this block */
1447                         block2->flags |= BLOCK_IS_DEAD;
1448
1449                         /* If block2 was pending for compilation, cancel it.
1450                          * If it's being compiled right now, wait until it
1451                          * finishes. */
1452                         if (ENABLE_THREADED_COMPILER)
1453                                 lightrec_recompiler_remove(state->rec, block2);
1454                 }
1455
1456                 /* We know from now on that block2 (if present) isn't going to
1457                  * be compiled. We can override the LUT entry with our new
1458                  * block's entry point. */
1459                 offset = lut_offset(block->pc) + target->offset;
1460                 lut_write(state, offset, jit_address(target->label));
1461
1462                 if (block2) {
1463                         pr_debug("Reap block 0x%08x as it's covered by block "
1464                                  "0x%08x\n", block2->pc, block->pc);
1465
1466                         /* Finally, reap the block. */
1467                         if (ENABLE_THREADED_COMPILER) {
1468                                 lightrec_reaper_add(state->reaper,
1469                                                     lightrec_reap_block,
1470                                                     block2);
1471                         } else {
1472                                 lightrec_unregister_block(state->block_cache, block2);
1473                                 lightrec_free_block(state, block2);
1474                         }
1475                 }
1476         }
1477
1478         if (ENABLE_THREADED_COMPILER)
1479                 lightrec_reaper_continue(state->reaper);
1480
1481         if (ENABLE_DISASSEMBLER) {
1482                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1483                 jit_disassemble();
1484         }
1485
1486         jit_clear_state();
1487
1488 #if ENABLE_THREADED_COMPILER
1489         if (fully_tagged)
1490                 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1491 #endif
1492         if (fully_tagged && !op_list_freed) {
1493                 pr_debug("Block PC 0x%08x is fully tagged"
1494                          " - free opcode list\n", block->pc);
1495                 lightrec_free_opcode_list(state, block);
1496                 block->opcode_list = NULL;
1497         }
1498
1499         if (oldjit) {
1500                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1501                          block->pc);
1502
1503                 if (ENABLE_THREADED_COMPILER) {
1504                         lightrec_reaper_add(state->reaper,
1505                                             lightrec_reap_jit, oldjit);
1506                         lightrec_reaper_add(state->reaper,
1507                                             lightrec_reap_function, old_fn);
1508                 } else {
1509                         _jit_destroy_state(oldjit);
1510                         lightrec_free_function(state, old_fn);
1511                 }
1512         }
1513
1514         return 0;
1515 }
1516
1517 static void lightrec_print_info(struct lightrec_state *state)
1518 {
1519         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1520                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1521                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1522                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1523                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1524                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1525                         lightrec_get_total_mem_usage() / 1024,
1526                        lightrec_get_average_ipi());
1527                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1528         }
1529 }
1530
1531 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1532 {
1533         s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1534         void *block_trace;
1535         s32 cycles_delta;
1536
1537         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1538
1539         /* Handle the cycle counter overflowing */
1540         if (unlikely(target_cycle < state->current_cycle))
1541                 target_cycle = UINT_MAX;
1542
1543         state->target_cycle = target_cycle;
1544         state->next_pc = pc;
1545
1546         block_trace = get_next_block_func(state, pc);
1547         if (block_trace) {
1548                 cycles_delta = state->target_cycle - state->current_cycle;
1549
1550                 cycles_delta = (*func)(block_trace, cycles_delta);
1551
1552                 state->current_cycle = state->target_cycle - cycles_delta;
1553         }
1554
1555         if (ENABLE_THREADED_COMPILER)
1556                 lightrec_reaper_reap(state->reaper);
1557
1558         if (LOG_LEVEL >= INFO_L)
1559                 lightrec_print_info(state);
1560
1561         return state->next_pc;
1562 }
1563
1564 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1565 {
1566         return lightrec_execute(state, pc, state->current_cycle);
1567 }
1568
1569 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1570 {
1571         struct block *block = lightrec_get_block(state, pc);
1572         if (!block)
1573                 return 0;
1574
1575         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1576
1577         pc = lightrec_emulate_block(state, block, pc);
1578
1579         if (LOG_LEVEL >= INFO_L)
1580                 lightrec_print_info(state);
1581
1582         return pc;
1583 }
1584
1585 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1586 {
1587         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1588         if (block->opcode_list)
1589                 lightrec_free_opcode_list(state, block);
1590         if (block->_jit)
1591                 _jit_destroy_state(block->_jit);
1592         if (block->function) {
1593                 lightrec_free_function(state, block->function);
1594                 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1595         }
1596         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1597 }
1598
1599 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1600 {
1601         struct lightrec_cstate *cstate;
1602
1603         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1604         if (!cstate)
1605                 return NULL;
1606
1607         cstate->reg_cache = lightrec_regcache_init(state);
1608         if (!cstate->reg_cache) {
1609                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1610                 return NULL;
1611         }
1612
1613         cstate->state = state;
1614
1615         return cstate;
1616 }
1617
1618 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1619 {
1620         lightrec_free_regcache(cstate->reg_cache);
1621         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1622 }
1623
1624 struct lightrec_state * lightrec_init(char *argv0,
1625                                       const struct lightrec_mem_map *map,
1626                                       size_t nb,
1627                                       const struct lightrec_ops *ops)
1628 {
1629         const struct lightrec_mem_map *codebuf_map = &map[PSX_MAP_CODE_BUFFER];
1630         struct lightrec_state *state;
1631         uintptr_t addr;
1632         void *tlsf = NULL;
1633         bool with_32bit_lut = false;
1634         size_t lut_size;
1635
1636         /* Sanity-check ops */
1637         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1638                 pr_err("Missing callbacks in lightrec_ops structure\n");
1639                 return NULL;
1640         }
1641
1642         if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER
1643             && codebuf_map->address) {
1644                 tlsf = tlsf_create_with_pool(codebuf_map->address,
1645                                              codebuf_map->length);
1646                 if (!tlsf) {
1647                         pr_err("Unable to initialize code buffer\n");
1648                         return NULL;
1649                 }
1650
1651                 if (__WORDSIZE == 64) {
1652                         addr = (uintptr_t) codebuf_map->address + codebuf_map->length - 1;
1653                         with_32bit_lut = addr == (u32) addr;
1654                 }
1655         }
1656
1657         if (with_32bit_lut)
1658                 lut_size = CODE_LUT_SIZE * 4;
1659         else
1660                 lut_size = CODE_LUT_SIZE * sizeof(void *);
1661
1662         init_jit(argv0);
1663
1664         state = calloc(1, sizeof(*state) + lut_size);
1665         if (!state)
1666                 goto err_finish_jit;
1667
1668         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) + lut_size);
1669
1670         state->tlsf = tlsf;
1671         state->with_32bit_lut = with_32bit_lut;
1672
1673         state->block_cache = lightrec_blockcache_init(state);
1674         if (!state->block_cache)
1675                 goto err_free_state;
1676
1677         if (ENABLE_THREADED_COMPILER) {
1678                 state->rec = lightrec_recompiler_init(state);
1679                 if (!state->rec)
1680                         goto err_free_block_cache;
1681
1682                 state->reaper = lightrec_reaper_init(state);
1683                 if (!state->reaper)
1684                         goto err_free_recompiler;
1685         } else {
1686                 state->cstate = lightrec_create_cstate(state);
1687                 if (!state->cstate)
1688                         goto err_free_block_cache;
1689         }
1690
1691         state->nb_maps = nb;
1692         state->maps = map;
1693
1694         memcpy(&state->ops, ops, sizeof(*ops));
1695
1696         state->dispatcher = generate_dispatcher(state);
1697         if (!state->dispatcher)
1698                 goto err_free_reaper;
1699
1700         state->c_wrapper_block = generate_wrapper(state);
1701         if (!state->c_wrapper_block)
1702                 goto err_free_dispatcher;
1703
1704         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1705         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1706         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1707         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
1708         state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1709         state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1710
1711         map = &state->maps[PSX_MAP_BIOS];
1712         state->offset_bios = (uintptr_t)map->address - map->pc;
1713
1714         map = &state->maps[PSX_MAP_SCRATCH_PAD];
1715         state->offset_scratch = (uintptr_t)map->address - map->pc;
1716
1717         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1718         state->offset_ram = (uintptr_t)map->address - map->pc;
1719
1720         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1721             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1722             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1723                 state->mirrors_mapped = true;
1724
1725         if (state->offset_bios == 0 &&
1726             state->offset_scratch == 0 &&
1727             state->offset_ram == 0 &&
1728             state->mirrors_mapped) {
1729                 pr_info("Memory map is perfect. Emitted code will be best.\n");
1730         } else {
1731                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1732         }
1733
1734         if (state->with_32bit_lut)
1735                 pr_info("Using 32-bit LUT\n");
1736
1737         return state;
1738
1739 err_free_dispatcher:
1740         lightrec_free_block(state, state->dispatcher);
1741 err_free_reaper:
1742         if (ENABLE_THREADED_COMPILER)
1743                 lightrec_reaper_destroy(state->reaper);
1744 err_free_recompiler:
1745         if (ENABLE_THREADED_COMPILER)
1746                 lightrec_free_recompiler(state->rec);
1747         else
1748                 lightrec_free_cstate(state->cstate);
1749 err_free_block_cache:
1750         lightrec_free_block_cache(state->block_cache);
1751 err_free_state:
1752         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1753                             lut_elm_size(state) * CODE_LUT_SIZE);
1754         free(state);
1755 err_finish_jit:
1756         finish_jit();
1757         if (ENABLE_CODE_BUFFER && tlsf)
1758                 tlsf_destroy(tlsf);
1759         return NULL;
1760 }
1761
1762 void lightrec_destroy(struct lightrec_state *state)
1763 {
1764         /* Force a print info on destroy*/
1765         state->current_cycle = ~state->current_cycle;
1766         lightrec_print_info(state);
1767
1768         lightrec_free_block_cache(state->block_cache);
1769         lightrec_free_block(state, state->dispatcher);
1770         lightrec_free_block(state, state->c_wrapper_block);
1771
1772         if (ENABLE_THREADED_COMPILER) {
1773                 lightrec_free_recompiler(state->rec);
1774                 lightrec_reaper_destroy(state->reaper);
1775         } else {
1776                 lightrec_free_cstate(state->cstate);
1777         }
1778
1779         finish_jit();
1780         if (ENABLE_CODE_BUFFER && state->tlsf)
1781                 tlsf_destroy(state->tlsf);
1782
1783         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1784                             lut_elm_size(state) * CODE_LUT_SIZE);
1785         free(state);
1786 }
1787
1788 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1789 {
1790         u32 kaddr = kunseg(addr & ~0x3);
1791         enum psx_map idx = lightrec_get_map_idx(state, kaddr);
1792
1793         switch (idx) {
1794         case PSX_MAP_MIRROR1:
1795         case PSX_MAP_MIRROR2:
1796         case PSX_MAP_MIRROR3:
1797                 /* Handle mirrors */
1798                 kaddr &= RAM_SIZE - 1;
1799                 fallthrough;
1800         case PSX_MAP_KERNEL_USER_RAM:
1801                 break;
1802         default:
1803                 return;
1804         }
1805
1806         memset(lut_address(state, lut_offset(kaddr)), 0,
1807                ((len + 3) / 4) * lut_elm_size(state));
1808 }
1809
1810 void lightrec_invalidate_all(struct lightrec_state *state)
1811 {
1812         memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
1813 }
1814
1815 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1816 {
1817         if (state->invalidate_from_dma_only != dma_only)
1818                 lightrec_invalidate_all(state);
1819
1820         state->invalidate_from_dma_only = dma_only;
1821 }
1822
1823 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1824 {
1825         if (flags != LIGHTREC_EXIT_NORMAL) {
1826                 state->exit_flags |= flags;
1827                 state->target_cycle = state->current_cycle;
1828         }
1829 }
1830
1831 u32 lightrec_exit_flags(struct lightrec_state *state)
1832 {
1833         return state->exit_flags;
1834 }
1835
1836 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1837 {
1838         return state->current_cycle;
1839 }
1840
1841 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1842 {
1843         state->current_cycle = cycles;
1844
1845         if (state->target_cycle < cycles)
1846                 state->target_cycle = cycles;
1847 }
1848
1849 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1850 {
1851         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1852                 if (cycles < state->current_cycle)
1853                         cycles = state->current_cycle;
1854
1855                 state->target_cycle = cycles;
1856         }
1857 }
1858
1859 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1860 {
1861         return &state->regs;
1862 }