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
  20 #include <errno.h>
  21 #include <inttypes.h>
  22 #include <limits.h>
  23 #if ENABLE_THREADED_COMPILER
  24 #include <stdatomic.h>
  25 #endif
  26 #include <stdbool.h>
  27 #include <stddef.h>
  28 #include <string.h>
  29 #if ENABLE_TINYMM
  30 #include <tinymm.h>
  31 #endif
  32
  33 #define GENMASK(h, l) \
  34         (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
  35
  36 static struct block * lightrec_precompile_block(struct lightrec_state *state,
  37                                                 u32 pc);
  38 static bool lightrec_block_is_fully_tagged(const struct block *block);
  39
  40 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
  41 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
  42
  43 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
  44                                 void *host, u32 addr, u8 data)
  45 {
  46         *(u8 *)host = data;
  47
  48         if (!state->invalidate_from_dma_only)
  49                 lightrec_invalidate(state, addr, 1);
  50 }
  51
  52 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
  53                                 void *host, u32 addr, u16 data)
  54 {
  55         *(u16 *)host = HTOLE16(data);
  56
  57         if (!state->invalidate_from_dma_only)
  58                 lightrec_invalidate(state, addr, 2);
  59 }
  60
  61 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
  62                                 void *host, u32 addr, u32 data)
  63 {
  64         *(u32 *)host = HTOLE32(data);
  65
  66         if (!state->invalidate_from_dma_only)
  67                 lightrec_invalidate(state, addr, 4);
  68 }
  69
  70 static u8 lightrec_default_lb(struct lightrec_state *state,
  71                               u32 opcode, void *host, u32 addr)
  72 {
  73         return *(u8 *)host;
  74 }
  75
  76 static u16 lightrec_default_lh(struct lightrec_state *state,
  77                                u32 opcode, void *host, u32 addr)
  78 {
  79         return LE16TOH(*(u16 *)host);
  80 }
  81
  82 static u32 lightrec_default_lw(struct lightrec_state *state,
  83                                u32 opcode, void *host, u32 addr)
  84 {
  85         return LE32TOH(*(u32 *)host);
  86 }
  87
  88 static const struct lightrec_mem_map_ops lightrec_default_ops = {
  89         .sb = lightrec_default_sb,
  90         .sh = lightrec_default_sh,
  91         .sw = lightrec_default_sw,
  92         .lb = lightrec_default_lb,
  93         .lh = lightrec_default_lh,
  94         .lw = lightrec_default_lw,
  95 };
  96
  97 static void __segfault_cb(struct lightrec_state *state, u32 addr,
  98                           const struct block *block)
  99 {
 100         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 101         pr_err("Segmentation fault in recompiled code: invalid "
 102                "load/store at address 0x%08x\n", addr);
 103         if (block)
 104                 pr_err("Was executing block PC 0x%08x\n", block->pc);
 105 }
 106
 107 static void lightrec_swl(struct lightrec_state *state,
 108                          const struct lightrec_mem_map_ops *ops,
 109                          u32 opcode, void *host, u32 addr, u32 data)
 110 {
 111         unsigned int shift = addr & 0x3;
 112         unsigned int mask = GENMASK(31, (shift + 1) * 8);
 113         u32 old_data;
 114
 115         /* Align to 32 bits */
 116         addr &= ~3;
 117         host = (void *)((uintptr_t)host & ~3);
 118
 119         old_data = ops->lw(state, opcode, host, addr);
 120
 121         data = (data >> ((3 - shift) * 8)) | (old_data & mask);
 122
 123         ops->sw(state, opcode, host, addr, data);
 124 }
 125
 126 static void lightrec_swr(struct lightrec_state *state,
 127                          const struct lightrec_mem_map_ops *ops,
 128                          u32 opcode, void *host, u32 addr, u32 data)
 129 {
 130         unsigned int shift = addr & 0x3;
 131         unsigned int mask = (1 << (shift * 8)) - 1;
 132         u32 old_data;
 133
 134         /* Align to 32 bits */
 135         addr &= ~3;
 136         host = (void *)((uintptr_t)host & ~3);
 137
 138         old_data = ops->lw(state, opcode, host, addr);
 139
 140         data = (data << (shift * 8)) | (old_data & mask);
 141
 142         ops->sw(state, opcode, host, addr, data);
 143 }
 144
 145 static void lightrec_swc2(struct lightrec_state *state, union code op,
 146                           const struct lightrec_mem_map_ops *ops,
 147                           void *host, u32 addr)
 148 {
 149         u32 data = lightrec_mfc2(state, op.i.rt);
 150
 151         ops->sw(state, op.opcode, host, addr, data);
 152 }
 153
 154 static u32 lightrec_lwl(struct lightrec_state *state,
 155                         const struct lightrec_mem_map_ops *ops,
 156                         u32 opcode, void *host, u32 addr, u32 data)
 157 {
 158         unsigned int shift = addr & 0x3;
 159         unsigned int mask = (1 << (24 - shift * 8)) - 1;
 160         u32 old_data;
 161
 162         /* Align to 32 bits */
 163         addr &= ~3;
 164         host = (void *)((uintptr_t)host & ~3);
 165
 166         old_data = ops->lw(state, opcode, host, addr);
 167
 168         return (data & mask) | (old_data << (24 - shift * 8));
 169 }
 170
 171 static u32 lightrec_lwr(struct lightrec_state *state,
 172                         const struct lightrec_mem_map_ops *ops,
 173                         u32 opcode, void *host, u32 addr, u32 data)
 174 {
 175         unsigned int shift = addr & 0x3;
 176         unsigned int mask = GENMASK(31, 32 - shift * 8);
 177         u32 old_data;
 178
 179         /* Align to 32 bits */
 180         addr &= ~3;
 181         host = (void *)((uintptr_t)host & ~3);
 182
 183         old_data = ops->lw(state, opcode, host, addr);
 184
 185         return (data & mask) | (old_data >> (shift * 8));
 186 }
 187
 188 static void lightrec_lwc2(struct lightrec_state *state, union code op,
 189                           const struct lightrec_mem_map_ops *ops,
 190                           void *host, u32 addr)
 191 {
 192         u32 data = ops->lw(state, op.opcode, host, addr);
 193
 194         lightrec_mtc2(state, op.i.rt, data);
 195 }
 196
 197 static void lightrec_invalidate_map(struct lightrec_state *state,
 198                 const struct lightrec_mem_map *map, u32 addr, u32 len)
 199 {
 200         if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
 201                 memset(&state->code_lut[lut_offset(addr)], 0,
 202                        ((len + 3) / 4) * sizeof(void *));
 203         }
 204 }
 205
 206 const struct lightrec_mem_map *
 207 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 208 {
 209         const struct lightrec_mem_map *map;
 210         unsigned int i;
 211         u32 addr;
 212
 213         for (i = 0; i < state->nb_maps; i++) {
 214                 const struct lightrec_mem_map *mapi = &state->maps[i];
 215
 216                 if (kaddr >= mapi->pc && kaddr < mapi->pc + mapi->length) {
 217                         map = mapi;
 218                         break;
 219                 }
 220         }
 221
 222         if (i == state->nb_maps)
 223                 return NULL;
 224
 225         addr = kaddr - map->pc;
 226
 227         while (map->mirror_of)
 228                 map = map->mirror_of;
 229
 230         if (host)
 231                 *host = map->address + addr;
 232
 233         return map;
 234 }
 235
 236 u32 lightrec_rw(struct lightrec_state *state, union code op,
 237                 u32 addr, u32 data, u16 *flags, struct block *block)
 238 {
 239         const struct lightrec_mem_map *map;
 240         const struct lightrec_mem_map_ops *ops;
 241         u32 opcode = op.opcode;
 242         void *host;
 243
 244         addr += (s16) op.i.imm;
 245
 246         map = lightrec_get_map(state, &host, kunseg(addr));
 247         if (!map) {
 248                 __segfault_cb(state, addr, block);
 249                 return 0;
 250         }
 251
 252         if (unlikely(map->ops)) {
 253                 if (flags)
 254                         *flags |= LIGHTREC_HW_IO;
 255
 256                 ops = map->ops;
 257         } else {
 258                 if (flags)
 259                         *flags |= LIGHTREC_DIRECT_IO;
 260
 261                 ops = &lightrec_default_ops;
 262         }
 263
 264         switch (op.i.op) {
 265         case OP_SB:
 266                 ops->sb(state, opcode, host, addr, (u8) data);
 267                 return 0;
 268         case OP_SH:
 269                 ops->sh(state, opcode, host, addr, (u16) data);
 270                 return 0;
 271         case OP_SWL:
 272                 lightrec_swl(state, ops, opcode, host, addr, data);
 273                 return 0;
 274         case OP_SWR:
 275                 lightrec_swr(state, ops, opcode, host, addr, data);
 276                 return 0;
 277         case OP_SW:
 278                 ops->sw(state, opcode, host, addr, data);
 279                 return 0;
 280         case OP_SWC2:
 281                 lightrec_swc2(state, op, ops, host, addr);
 282                 return 0;
 283         case OP_LB:
 284                 return (s32) (s8) ops->lb(state, opcode, host, addr);
 285         case OP_LBU:
 286                 return ops->lb(state, opcode, host, addr);
 287         case OP_LH:
 288                 return (s32) (s16) ops->lh(state, opcode, host, addr);
 289         case OP_LHU:
 290                 return ops->lh(state, opcode, host, addr);
 291         case OP_LWC2:
 292                 lightrec_lwc2(state, op, ops, host, addr);
 293                 return 0;
 294         case OP_LWL:
 295                 return lightrec_lwl(state, ops, opcode, host, addr, data);
 296         case OP_LWR:
 297                 return lightrec_lwr(state, ops, opcode, host, addr, data);
 298         case OP_LW:
 299         default:
 300                 return ops->lw(state, opcode, host, addr);
 301         }
 302 }
 303
 304 static void lightrec_rw_helper(struct lightrec_state *state,
 305                                union code op, u16 *flags,
 306                                struct block *block)
 307 {
 308         u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
 309                               state->regs.gpr[op.i.rt], flags, block);
 310
 311         switch (op.i.op) {
 312         case OP_LB:
 313         case OP_LBU:
 314         case OP_LH:
 315         case OP_LHU:
 316         case OP_LWL:
 317         case OP_LWR:
 318         case OP_LW:
 319                 if (op.i.rt)
 320                         state->regs.gpr[op.i.rt] = ret;
 321         default: /* fall-through */
 322                 break;
 323         }
 324 }
 325
 326 static void lightrec_rw_cb(struct lightrec_state *state, union code op)
 327 {
 328         lightrec_rw_helper(state, op, NULL, NULL);
 329 }
 330
 331 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 332 {
 333         struct block *block;
 334         struct opcode *op;
 335         bool was_tagged;
 336         u16 offset = (u16)arg;
 337
 338         block = lightrec_find_block_from_lut(state->block_cache,
 339                                              arg >> 16, state->next_pc);
 340         if (unlikely(!block)) {
 341                 pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
 342                          state->next_pc, offset);
 343                 return;
 344         }
 345
 346         op = &block->opcode_list[offset];
 347         was_tagged = op->flags & (LIGHTREC_HW_IO | LIGHTREC_DIRECT_IO);
 348
 349         lightrec_rw_helper(state, op->c, &op->flags, block);
 350
 351         if (!was_tagged) {
 352                 pr_debug("Opcode of block at PC 0x%08x has been tagged - flag "
 353                          "for recompilation\n", block->pc);
 354
 355                 block->flags |= BLOCK_SHOULD_RECOMPILE;
 356         }
 357 }
 358
 359 static u32 clamp_s32(s32 val, s32 min, s32 max)
 360 {
 361         return val < min ? min : val > max ? max : val;
 362 }
 363
 364 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
 365 {
 366         s16 gteir1, gteir2, gteir3;
 367
 368         switch (reg) {
 369         case 1:
 370         case 3:
 371         case 5:
 372         case 8:
 373         case 9:
 374         case 10:
 375         case 11:
 376                 return (s32)(s16) state->regs.cp2d[reg];
 377         case 7:
 378         case 16:
 379         case 17:
 380         case 18:
 381         case 19:
 382                 return (u16) state->regs.cp2d[reg];
 383         case 28:
 384         case 29:
 385                 gteir1 = (s16) state->regs.cp2d[9];
 386                 gteir2 = (s16) state->regs.cp2d[10];
 387                 gteir3 = (s16) state->regs.cp2d[11];
 388
 389                 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
 390                         clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
 391                         clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
 392         case 15:
 393                 reg = 14;
 394         default: /* fall-through */
 395                 return state->regs.cp2d[reg];
 396         }
 397 }
 398
 399 u32 lightrec_mfc(struct lightrec_state *state, union code op)
 400 {
 401         if (op.i.op == OP_CP0)
 402                 return state->regs.cp0[op.r.rd];
 403         else if (op.r.rs == OP_CP2_BASIC_MFC2)
 404                 return lightrec_mfc2(state, op.r.rd);
 405         else
 406                 return state->regs.cp2c[op.r.rd];
 407 }
 408
 409 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 410 {
 411         u32 status, oldstatus, cause;
 412
 413         switch (reg) {
 414         case 1:
 415         case 4:
 416         case 8:
 417         case 14:
 418         case 15:
 419                 /* Those registers are read-only */
 420                 return;
 421         default:
 422                 break;
 423         }
 424
 425         if (reg == 12) {
 426                 status = state->regs.cp0[12];
 427                 oldstatus = status;
 428
 429                 if (status & ~data & BIT(16)) {
 430                         state->ops.enable_ram(state, true);
 431                         lightrec_invalidate_all(state);
 432                 } else if (~status & data & BIT(16)) {
 433                         state->ops.enable_ram(state, false);
 434                 }
 435         }
 436
 437         if (reg == 13) {
 438                 state->regs.cp0[13] &= ~0x300;
 439                 state->regs.cp0[13] |= data & 0x300;
 440         } else {
 441                 state->regs.cp0[reg] = data;
 442         }
 443
 444         if (reg == 12 || reg == 13) {
 445                 cause = state->regs.cp0[13];
 446                 status = state->regs.cp0[12];
 447
 448                 /* Handle software interrupts */
 449                 if (!!(status & cause & 0x300) & status)
 450                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 451
 452                 /* Handle hardware interrupts */
 453                 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
 454                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 455         }
 456 }
 457
 458 static u32 count_leading_bits(s32 data)
 459 {
 460         u32 cnt = 33;
 461
 462 #ifdef __has_builtin
 463 #if __has_builtin(__builtin_clrsb)
 464         return 1 + __builtin_clrsb(data);
 465 #endif
 466 #endif
 467
 468         data = (data ^ (data >> 31)) << 1;
 469
 470         do {
 471                 cnt -= 1;
 472                 data >>= 1;
 473         } while (data);
 474
 475         return cnt;
 476 }
 477
 478 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
 479 {
 480         switch (reg) {
 481         case 15:
 482                 state->regs.cp2d[12] = state->regs.cp2d[13];
 483                 state->regs.cp2d[13] = state->regs.cp2d[14];
 484                 state->regs.cp2d[14] = data;
 485                 break;
 486         case 28:
 487                 state->regs.cp2d[9] = (data << 7) & 0xf80;
 488                 state->regs.cp2d[10] = (data << 2) & 0xf80;
 489                 state->regs.cp2d[11] = (data >> 3) & 0xf80;
 490                 break;
 491         case 31:
 492                 return;
 493         case 30:
 494                 state->regs.cp2d[31] = count_leading_bits((s32) data);
 495         default: /* fall-through */
 496                 state->regs.cp2d[reg] = data;
 497                 break;
 498         }
 499 }
 500
 501 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 502 {
 503         switch (reg) {
 504         case 4:
 505         case 12:
 506         case 20:
 507         case 26:
 508         case 27:
 509         case 29:
 510         case 30:
 511                 data = (s32)(s16) data;
 512                 break;
 513         case 31:
 514                 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
 515         default: /* fall-through */
 516                 break;
 517         }
 518
 519         state->regs.cp2c[reg] = data;
 520 }
 521
 522 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
 523 {
 524         if (op.i.op == OP_CP0)
 525                 lightrec_mtc0(state, op.r.rd, data);
 526         else if (op.r.rs == OP_CP2_BASIC_CTC2)
 527                 lightrec_ctc2(state, op.r.rd, data);
 528         else
 529                 lightrec_mtc2(state, op.r.rd, data);
 530 }
 531
 532 static void lightrec_mtc_cb(struct lightrec_state *state, union code op)
 533 {
 534         lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
 535 }
 536
 537 void lightrec_rfe(struct lightrec_state *state)
 538 {
 539         u32 status;
 540
 541         /* Read CP0 Status register (r12) */
 542         status = state->regs.cp0[12];
 543
 544         /* Switch the bits */
 545         status = ((status & 0x3c) >> 2) | (status & ~0xf);
 546
 547         /* Write it back */
 548         lightrec_mtc0(state, 12, status);
 549 }
 550
 551 void lightrec_cp(struct lightrec_state *state, union code op)
 552 {
 553         if (op.i.op == OP_CP0) {
 554                 pr_err("Invalid CP opcode to coprocessor #0\n");
 555                 return;
 556         }
 557
 558         (*state->ops.cop2_op)(state, op.opcode);
 559 }
 560
 561 static void lightrec_syscall_cb(struct lightrec_state *state, union code op)
 562 {
 563         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
 564 }
 565
 566 static void lightrec_break_cb(struct lightrec_state *state, union code op)
 567 {
 568         lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
 569 }
 570
 571 struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 572 {
 573         struct block *block = lightrec_find_block(state->block_cache, pc);
 574
 575         if (block && lightrec_block_is_outdated(state, block)) {
 576                 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
 577
 578                 /* Make sure the recompiler isn't processing the block we'll
 579                  * destroy */
 580                 if (ENABLE_THREADED_COMPILER)
 581                         lightrec_recompiler_remove(state->rec, block);
 582
 583                 lightrec_unregister_block(state->block_cache, block);
 584                 remove_from_code_lut(state->block_cache, block);
 585                 lightrec_free_block(state, block);
 586                 block = NULL;
 587         }
 588
 589         if (!block) {
 590                 block = lightrec_precompile_block(state, pc);
 591                 if (!block) {
 592                         pr_err("Unable to recompile block at PC 0x%x\n", pc);
 593                         lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 594                         return NULL;
 595                 }
 596
 597                 lightrec_register_block(state->block_cache, block);
 598         }
 599
 600         return block;
 601 }
 602
 603 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 604 {
 605         struct block *block;
 606         bool should_recompile;
 607         void *func;
 608
 609         for (;;) {
 610                 func = state->code_lut[lut_offset(pc)];
 611                 if (func && func != state->get_next_block)
 612                         break;
 613
 614                 block = lightrec_get_block(state, pc);
 615
 616                 if (unlikely(!block))
 617                         break;
 618
 619                 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
 620                         func = state->memset_func;
 621                         break;
 622                 }
 623
 624                 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
 625                         !(block->flags & BLOCK_IS_DEAD);
 626
 627                 if (unlikely(should_recompile)) {
 628                         pr_debug("Block at PC 0x%08x should recompile\n", pc);
 629
 630                         lightrec_unregister(MEM_FOR_CODE, block->code_size);
 631
 632                         if (ENABLE_THREADED_COMPILER)
 633                                 lightrec_recompiler_add(state->rec, block);
 634                         else
 635                                 lightrec_compile_block(state->cstate, block);
 636                 }
 637
 638                 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
 639                         func = lightrec_recompiler_run_first_pass(state, block, &pc);
 640                 else
 641                         func = block->function;
 642
 643                 if (likely(func))
 644                         break;
 645
 646                 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
 647                         pc = lightrec_emulate_block(state, block, pc);
 648
 649                 } else if (!ENABLE_THREADED_COMPILER) {
 650                         /* Block wasn't compiled yet - run the interpreter */
 651                         if (block->flags & BLOCK_FULLY_TAGGED)
 652                                 pr_debug("Block fully tagged, skipping first pass\n");
 653                         else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 654                                 pc = lightrec_emulate_block(state, block, pc);
 655
 656                         /* Then compile it using the profiled data */
 657                         lightrec_compile_block(state->cstate, block);
 658                 } else {
 659                         lightrec_recompiler_add(state->rec, block);
 660                 }
 661
 662                 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
 663                     state->current_cycle >= state->target_cycle)
 664                         break;
 665         }
 666
 667         state->next_pc = pc;
 668         return func;
 669 }
 670
 671 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
 672                               void (*f)(struct lightrec_state *, u32 d),
 673                               u32 d)
 674 {
 675         state->current_cycle = state->target_cycle - cycles_delta;
 676
 677         (*f)(state, d);
 678
 679         return state->target_cycle - state->current_cycle;
 680 }
 681
 682 static struct block * generate_wrapper(struct lightrec_state *state)
 683 {
 684         struct block *block;
 685         jit_state_t *_jit;
 686         unsigned int i;
 687         int stack_ptr;
 688         jit_word_t code_size;
 689         jit_node_t *to_tramp, *to_fn_epilog;
 690         jit_node_t *addr[C_WRAPPERS_COUNT - 1];
 691
 692         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 693         if (!block)
 694                 goto err_no_mem;
 695
 696         _jit = jit_new_state();
 697         if (!_jit)
 698                 goto err_free_block;
 699
 700         jit_name("RW wrapper");
 701         jit_note(__FILE__, __LINE__);
 702
 703         /* Wrapper entry point */
 704         jit_prolog();
 705         jit_tramp(256);
 706
 707         /* Add entry points; separate them by opcodes that increment
 708          * LIGHTREC_REG_STATE (since we cannot touch other registers).
 709          * The difference will then tell us which C function to call. */
 710         for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
 711                 jit_addi(LIGHTREC_REG_STATE, LIGHTREC_REG_STATE, __WORDSIZE / 8);
 712                 addr[i - 1] = jit_indirect();
 713         }
 714
 715         jit_epilog();
 716         jit_prolog();
 717
 718         stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
 719
 720         /* Save all temporaries on stack */
 721         for (i = 0; i < NUM_TEMPS; i++)
 722                 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
 723
 724         /* Jump to the trampoline */
 725         to_tramp = jit_jmpi();
 726
 727         /* The trampoline will jump back here */
 728         to_fn_epilog = jit_label();
 729
 730         /* Restore temporaries from stack */
 731         for (i = 0; i < NUM_TEMPS; i++)
 732                 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
 733
 734         jit_ret();
 735         jit_epilog();
 736
 737         /* Trampoline entry point.
 738          * The sole purpose of the trampoline is to cheese Lightning not to
 739          * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
 740          * do want to return to the caller with this register modified. */
 741         jit_prolog();
 742         jit_tramp(256);
 743         jit_patch(to_tramp);
 744
 745         /* Retrieve the wrapper function */
 746         jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
 747                  offsetof(struct lightrec_state, c_wrappers));
 748
 749         /* Restore LIGHTREC_REG_STATE to its correct value */
 750         jit_movi(LIGHTREC_REG_STATE, (uintptr_t) state);
 751
 752         jit_prepare();
 753         jit_pushargr(LIGHTREC_REG_STATE);
 754         jit_pushargr(LIGHTREC_REG_CYCLE);
 755         jit_pushargr(JIT_R0);
 756         jit_pushargr(JIT_R1);
 757         jit_finishi(c_function_wrapper);
 758         jit_retval_i(LIGHTREC_REG_CYCLE);
 759
 760         jit_patch_at(jit_jmpi(), to_fn_epilog);
 761         jit_epilog();
 762
 763         block->_jit = _jit;
 764         block->function = jit_emit();
 765         block->opcode_list = NULL;
 766         block->flags = 0;
 767         block->nb_ops = 0;
 768
 769         state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
 770
 771         for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
 772                 state->wrappers_eps[i] = jit_address(addr[i]);
 773
 774         jit_get_code(&code_size);
 775         lightrec_register(MEM_FOR_CODE, code_size);
 776
 777         block->code_size = code_size;
 778
 779         if (ENABLE_DISASSEMBLER) {
 780                 pr_debug("Wrapper block:\n");
 781                 jit_disassemble();
 782         }
 783
 784         jit_clear_state();
 785         return block;
 786
 787 err_free_block:
 788         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
 789 err_no_mem:
 790         pr_err("Unable to compile wrapper: Out of memory\n");
 791         return NULL;
 792 }
 793
 794 static u32 lightrec_memset(struct lightrec_state *state)
 795 {
 796         u32 kunseg_pc = kunseg(state->regs.gpr[4]);
 797         void *host;
 798         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
 799         u32 length = state->regs.gpr[5] * 4;
 800
 801         if (!map) {
 802                 pr_err("Unable to find memory map for memset target address "
 803                        "0x%x\n", kunseg_pc);
 804                 return 0;
 805         }
 806
 807         pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
 808                  kunseg_pc, (uintptr_t)host, length);
 809         memset(host, 0, length);
 810
 811         if (!state->invalidate_from_dma_only)
 812                 lightrec_invalidate_map(state, map, kunseg_pc, length);
 813
 814         /* Rough estimation of the number of cycles consumed */
 815         return 8 + 5 * (length  + 3 / 4);
 816 }
 817
 818 static struct block * generate_dispatcher(struct lightrec_state *state)
 819 {
 820         struct block *block;
 821         jit_state_t *_jit;
 822         jit_node_t *to_end, *to_c, *loop, *addr, *addr2, *addr3;
 823         unsigned int i;
 824         u32 offset, ram_len;
 825         jit_word_t code_size;
 826
 827         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 828         if (!block)
 829                 goto err_no_mem;
 830
 831         _jit = jit_new_state();
 832         if (!_jit)
 833                 goto err_free_block;
 834
 835         jit_name("dispatcher");
 836         jit_note(__FILE__, __LINE__);
 837
 838         jit_prolog();
 839         jit_frame(256);
 840
 841         jit_getarg(JIT_R0, jit_arg());
 842         jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
 843
 844         /* Force all callee-saved registers to be pushed on the stack */
 845         for (i = 0; i < NUM_REGS; i++)
 846                 jit_movr(JIT_V(i), JIT_V(i));
 847
 848         /* Pass lightrec_state structure to blocks, using the last callee-saved
 849          * register that Lightning provides */
 850         jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
 851
 852         loop = jit_label();
 853
 854         /* Call the block's code */
 855         jit_jmpr(JIT_R0);
 856
 857         if (OPT_REPLACE_MEMSET) {
 858                 /* Blocks will jump here when they need to call
 859                  * lightrec_memset() */
 860                 addr3 = jit_indirect();
 861
 862                 jit_prepare();
 863                 jit_pushargr(LIGHTREC_REG_STATE);
 864                 jit_finishi(lightrec_memset);
 865
 866                 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
 867                             offsetof(struct lightrec_state, regs.gpr[31]));
 868
 869                 jit_retval(JIT_R0);
 870                 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
 871         }
 872
 873         /* The block will jump here, with the number of cycles remaining in
 874          * LIGHTREC_REG_CYCLE */
 875         addr2 = jit_indirect();
 876
 877         /* Store back the next_pc to the lightrec_state structure */
 878         offset = offsetof(struct lightrec_state, next_pc);
 879         jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
 880
 881         /* Jump to end if state->target_cycle < state->current_cycle */
 882         to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
 883
 884         /* Convert next PC to KUNSEG and avoid mirrors */
 885         ram_len = state->maps[PSX_MAP_KERNEL_USER_RAM].length;
 886         jit_andi(JIT_R0, JIT_V0, 0x10000000 | (ram_len - 1));
 887         to_c = jit_bgei(JIT_R0, ram_len);
 888
 889         /* Fast path: code is running from RAM, use the code LUT */
 890         if (__WORDSIZE == 64)
 891                 jit_lshi(JIT_R0, JIT_R0, 1);
 892         jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
 893         jit_ldxi(JIT_R0, JIT_R0, offsetof(struct lightrec_state, code_lut));
 894
 895         /* If we get non-NULL, loop */
 896         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
 897
 898         /* Slow path: call C function get_next_block_func() */
 899         jit_patch(to_c);
 900
 901         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 902                 /* We may call the interpreter - update state->current_cycle */
 903                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
 904                            offsetof(struct lightrec_state, target_cycle));
 905                 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
 906                 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
 907                            LIGHTREC_REG_STATE, JIT_R1);
 908         }
 909
 910         /* The code LUT will be set to this address when the block at the target
 911          * PC has been preprocessed but not yet compiled by the threaded
 912          * recompiler */
 913         addr = jit_indirect();
 914
 915         /* Get the next block */
 916         jit_prepare();
 917         jit_pushargr(LIGHTREC_REG_STATE);
 918         jit_pushargr(JIT_V0);
 919         jit_finishi(&get_next_block_func);
 920         jit_retval(JIT_R0);
 921
 922         if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 923                 /* The interpreter may have updated state->current_cycle and
 924                  * state->target_cycle - recalc the delta */
 925                 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
 926                            offsetof(struct lightrec_state, current_cycle));
 927                 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
 928                            offsetof(struct lightrec_state, target_cycle));
 929                 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
 930         }
 931
 932         /* If we get non-NULL, loop */
 933         jit_patch_at(jit_bnei(JIT_R0, 0), loop);
 934
 935         /* When exiting, the recompiled code will jump to that address */
 936         jit_note(__FILE__, __LINE__);
 937         jit_patch(to_end);
 938
 939         jit_retr(LIGHTREC_REG_CYCLE);
 940         jit_epilog();
 941
 942         block->_jit = _jit;
 943         block->function = jit_emit();
 944         block->opcode_list = NULL;
 945         block->flags = 0;
 946         block->nb_ops = 0;
 947
 948         jit_get_code(&code_size);
 949         lightrec_register(MEM_FOR_CODE, code_size);
 950
 951         block->code_size = code_size;
 952
 953         state->eob_wrapper_func = jit_address(addr2);
 954         if (OPT_REPLACE_MEMSET)
 955                 state->memset_func = jit_address(addr3);
 956         state->get_next_block = jit_address(addr);
 957
 958         if (ENABLE_DISASSEMBLER) {
 959                 pr_debug("Dispatcher block:\n");
 960                 jit_disassemble();
 961         }
 962
 963         /* We're done! */
 964         jit_clear_state();
 965         return block;
 966
 967 err_free_block:
 968         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
 969 err_no_mem:
 970         pr_err("Unable to compile dispatcher: Out of memory\n");
 971         return NULL;
 972 }
 973
 974 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
 975 {
 976         void *host = NULL;
 977
 978         lightrec_get_map(state, &host, kunseg(pc));
 979
 980         const u32 *code = (u32 *)host;
 981         return (union code) *code;
 982 }
 983
 984 unsigned int lightrec_cycles_of_opcode(union code code)
 985 {
 986         return 2;
 987 }
 988
 989 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
 990 {
 991         lightrec_free(state, MEM_FOR_IR,
 992                       sizeof(*block->opcode_list) * block->nb_ops,
 993                       block->opcode_list);
 994 }
 995
 996 static unsigned int lightrec_get_mips_block_len(const u32 *src)
 997 {
 998         unsigned int i;
 999         union code c;
1000
1001         for (i = 1; ; i++) {
1002                 c.opcode = LE32TOH(*src++);
1003
1004                 if (is_syscall(c))
1005                         return i;
1006
1007                 if (is_unconditional_jump(c))
1008                         return i + 1;
1009         }
1010 }
1011
1012 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1013                                             const u32 *src, unsigned int *len)
1014 {
1015         struct opcode *list;
1016         unsigned int i, length;
1017
1018         length = lightrec_get_mips_block_len(src);
1019
1020         list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
1021         if (!list) {
1022                 pr_err("Unable to allocate memory\n");
1023                 return NULL;
1024         }
1025
1026         for (i = 0; i < length; i++) {
1027                 list[i].opcode = LE32TOH(src[i]);
1028                 list[i].flags = 0;
1029         }
1030
1031         *len = length * sizeof(u32);
1032
1033         return list;
1034 }
1035
1036 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1037                                                 u32 pc)
1038 {
1039         struct opcode *list;
1040         struct block *block;
1041         void *host;
1042         const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1043         const u32 *code = (u32 *) host;
1044         unsigned int length;
1045         bool fully_tagged;
1046
1047         if (!map)
1048                 return NULL;
1049
1050         block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1051         if (!block) {
1052                 pr_err("Unable to recompile block: Out of memory\n");
1053                 return NULL;
1054         }
1055
1056         list = lightrec_disassemble(state, code, &length);
1057         if (!list) {
1058                 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1059                 return NULL;
1060         }
1061
1062         block->pc = pc;
1063         block->_jit = NULL;
1064         block->function = NULL;
1065         block->opcode_list = list;
1066         block->code = code;
1067         block->next = NULL;
1068         block->flags = 0;
1069         block->code_size = 0;
1070 #if ENABLE_THREADED_COMPILER
1071         block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1072 #endif
1073         block->nb_ops = length / sizeof(u32);
1074
1075         lightrec_optimize(state, block);
1076
1077         length = block->nb_ops * sizeof(u32);
1078
1079         lightrec_register(MEM_FOR_MIPS_CODE, length);
1080
1081         if (ENABLE_DISASSEMBLER) {
1082                 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1083                 lightrec_print_disassembly(block, code);
1084         }
1085
1086         pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1087
1088         /* If the first opcode is an 'impossible' branch, never compile the
1089          * block */
1090         if (should_emulate(block->opcode_list))
1091                 block->flags |= BLOCK_NEVER_COMPILE;
1092
1093         fully_tagged = lightrec_block_is_fully_tagged(block);
1094         if (fully_tagged)
1095                 block->flags |= BLOCK_FULLY_TAGGED;
1096
1097         if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1098                 state->code_lut[lut_offset(pc)] = state->memset_func;
1099
1100         block->hash = lightrec_calculate_block_hash(block);
1101
1102         pr_debug("Recompile count: %u\n", state->nb_precompile++);
1103
1104         return block;
1105 }
1106
1107 static bool lightrec_block_is_fully_tagged(const struct block *block)
1108 {
1109         const struct opcode *op;
1110         unsigned int i;
1111
1112         for (i = 0; i < block->nb_ops; i++) {
1113                 op = &block->opcode_list[i];
1114
1115                 /* Verify that all load/stores of the opcode list
1116                  * Check all loads/stores of the opcode list and mark the
1117                  * block as fully compiled if they all have been tagged. */
1118                 switch (op->c.i.op) {
1119                 case OP_LB:
1120                 case OP_LH:
1121                 case OP_LWL:
1122                 case OP_LW:
1123                 case OP_LBU:
1124                 case OP_LHU:
1125                 case OP_LWR:
1126                 case OP_SB:
1127                 case OP_SH:
1128                 case OP_SWL:
1129                 case OP_SW:
1130                 case OP_SWR:
1131                 case OP_LWC2:
1132                 case OP_SWC2:
1133                         if (!(op->flags & (LIGHTREC_DIRECT_IO |
1134                                            LIGHTREC_HW_IO)))
1135                                 return false;
1136                 default: /* fall-through */
1137                         continue;
1138                 }
1139         }
1140
1141         return true;
1142 }
1143
1144 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1145 {
1146         struct block *block = data;
1147
1148         pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1149         lightrec_unregister_block(state->block_cache, block);
1150         lightrec_free_block(state, block);
1151 }
1152
1153 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1154 {
1155         _jit_destroy_state(data);
1156 }
1157
1158 int lightrec_compile_block(struct lightrec_cstate *cstate,
1159                            struct block *block)
1160 {
1161         struct lightrec_state *state = cstate->state;
1162         struct lightrec_branch_target *target;
1163         bool op_list_freed = false, fully_tagged = false;
1164         struct block *block2;
1165         struct opcode *elm;
1166         jit_state_t *_jit, *oldjit;
1167         jit_node_t *start_of_block;
1168         bool skip_next = false;
1169         jit_word_t code_size;
1170         unsigned int i, j;
1171         u32 offset;
1172
1173         fully_tagged = lightrec_block_is_fully_tagged(block);
1174         if (fully_tagged)
1175                 block->flags |= BLOCK_FULLY_TAGGED;
1176
1177         _jit = jit_new_state();
1178         if (!_jit)
1179                 return -ENOMEM;
1180
1181         oldjit = block->_jit;
1182         block->_jit = _jit;
1183
1184         lightrec_regcache_reset(cstate->reg_cache);
1185         cstate->cycles = 0;
1186         cstate->nb_branches = 0;
1187         cstate->nb_local_branches = 0;
1188         cstate->nb_targets = 0;
1189
1190         jit_prolog();
1191         jit_tramp(256);
1192
1193         start_of_block = jit_label();
1194
1195         for (i = 0; i < block->nb_ops; i++) {
1196                 elm = &block->opcode_list[i];
1197
1198                 if (skip_next) {
1199                         skip_next = false;
1200                         continue;
1201                 }
1202
1203                 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1204
1205                 if (should_emulate(elm)) {
1206                         pr_debug("Branch at offset 0x%x will be emulated\n",
1207                                  i << 2);
1208
1209                         lightrec_emit_eob(cstate, block, i, false);
1210                         skip_next = !(elm->flags & LIGHTREC_NO_DS);
1211                 } else {
1212                         lightrec_rec_opcode(cstate, block, i);
1213                         skip_next = has_delay_slot(elm->c) &&
1214                                 !(elm->flags & LIGHTREC_NO_DS);
1215 #if _WIN32
1216                         /* FIXME: GNU Lightning on Windows seems to use our
1217                          * mapped registers as temporaries. Until the actual bug
1218                          * is found and fixed, unconditionally mark our
1219                          * registers as live here. */
1220                         lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1221 #endif
1222                 }
1223         }
1224
1225         for (i = 0; i < cstate->nb_branches; i++)
1226                 jit_patch(cstate->branches[i]);
1227
1228         for (i = 0; i < cstate->nb_local_branches; i++) {
1229                 struct lightrec_branch *branch = &cstate->local_branches[i];
1230
1231                 pr_debug("Patch local branch to offset 0x%x\n",
1232                          branch->target << 2);
1233
1234                 if (branch->target == 0) {
1235                         jit_patch_at(branch->branch, start_of_block);
1236                         continue;
1237                 }
1238
1239                 for (j = 0; j < cstate->nb_targets; j++) {
1240                         if (cstate->targets[j].offset == branch->target) {
1241                                 jit_patch_at(branch->branch,
1242                                              cstate->targets[j].label);
1243                                 break;
1244                         }
1245                 }
1246
1247                 if (j == cstate->nb_targets)
1248                         pr_err("Unable to find branch target\n");
1249         }
1250
1251         jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
1252                  offsetof(struct lightrec_state, eob_wrapper_func));
1253
1254         jit_jmpr(JIT_R0);
1255
1256         jit_ret();
1257         jit_epilog();
1258
1259         block->function = jit_emit();
1260         block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1261
1262         /* Add compiled function to the LUT */
1263         state->code_lut[lut_offset(block->pc)] = block->function;
1264
1265         if (ENABLE_THREADED_COMPILER) {
1266                 /* Since we might try to reap the same block multiple times,
1267                  * we need the reaper to wait until everything has been
1268                  * submitted, so that the duplicate entries can be dropped. */
1269                 lightrec_reaper_pause(state->reaper);
1270         }
1271
1272         /* Detect old blocks that have been covered by the new one */
1273         for (i = 0; i < cstate->nb_targets; i++) {
1274                 target = &cstate->targets[i];
1275
1276                 if (!target->offset)
1277                         continue;
1278
1279                 offset = block->pc + target->offset * sizeof(u32);
1280                 block2 = lightrec_find_block(state->block_cache, offset);
1281                 if (block2) {
1282                         /* No need to check if block2 is compilable - it must
1283                          * be, otherwise block wouldn't be compilable either */
1284
1285                         /* Set the "block dead" flag to prevent the dynarec from
1286                          * recompiling this block */
1287                         block2->flags |= BLOCK_IS_DEAD;
1288
1289                         /* If block2 was pending for compilation, cancel it.
1290                          * If it's being compiled right now, wait until it
1291                          * finishes. */
1292                         if (ENABLE_THREADED_COMPILER)
1293                                 lightrec_recompiler_remove(state->rec, block2);
1294                 }
1295
1296                 /* We know from now on that block2 (if present) isn't going to
1297                  * be compiled. We can override the LUT entry with our new
1298                  * block's entry point. */
1299                 offset = lut_offset(block->pc) + target->offset;
1300                 state->code_lut[offset] = jit_address(target->label);
1301
1302                 if (block2) {
1303                         pr_debug("Reap block 0x%08x as it's covered by block "
1304                                  "0x%08x\n", block2->pc, block->pc);
1305
1306                         /* Finally, reap the block. */
1307                         if (ENABLE_THREADED_COMPILER) {
1308                                 lightrec_reaper_add(state->reaper,
1309                                                     lightrec_reap_block,
1310                                                     block2);
1311                         } else {
1312                                 lightrec_unregister_block(state->block_cache, block2);
1313                                 lightrec_free_block(state, block2);
1314                         }
1315                 }
1316         }
1317
1318         if (ENABLE_THREADED_COMPILER)
1319                 lightrec_reaper_continue(state->reaper);
1320
1321         jit_get_code(&code_size);
1322         lightrec_register(MEM_FOR_CODE, code_size);
1323
1324         block->code_size = code_size;
1325
1326         if (ENABLE_DISASSEMBLER) {
1327                 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1328                 jit_disassemble();
1329         }
1330
1331         jit_clear_state();
1332
1333 #if ENABLE_THREADED_COMPILER
1334         if (fully_tagged)
1335                 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1336 #endif
1337         if (fully_tagged && !op_list_freed) {
1338                 pr_debug("Block PC 0x%08x is fully tagged"
1339                          " - free opcode list\n", block->pc);
1340                 lightrec_free_opcode_list(state, block);
1341                 block->opcode_list = NULL;
1342         }
1343
1344         if (oldjit) {
1345                 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1346                          block->pc);
1347
1348                 if (ENABLE_THREADED_COMPILER)
1349                         lightrec_reaper_add(state->reaper,
1350                                             lightrec_reap_jit, oldjit);
1351                 else
1352                         _jit_destroy_state(oldjit);
1353         }
1354
1355         return 0;
1356 }
1357
1358 static void lightrec_print_info(struct lightrec_state *state)
1359 {
1360         if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1361                 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1362                         "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1363                         lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1364                         lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1365                         lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1366                         lightrec_get_total_mem_usage() / 1024,
1367                        lightrec_get_average_ipi());
1368                 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1369         }
1370 }
1371
1372 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1373 {
1374         s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1375         void *block_trace;
1376         s32 cycles_delta;
1377
1378         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1379
1380         /* Handle the cycle counter overflowing */
1381         if (unlikely(target_cycle < state->current_cycle))
1382                 target_cycle = UINT_MAX;
1383
1384         state->target_cycle = target_cycle;
1385         state->next_pc = pc;
1386
1387         block_trace = get_next_block_func(state, pc);
1388         if (block_trace) {
1389                 cycles_delta = state->target_cycle - state->current_cycle;
1390
1391                 cycles_delta = (*func)(block_trace, cycles_delta);
1392
1393                 state->current_cycle = state->target_cycle - cycles_delta;
1394         }
1395
1396         if (ENABLE_THREADED_COMPILER)
1397                 lightrec_reaper_reap(state->reaper);
1398
1399         if (LOG_LEVEL >= INFO_L)
1400                 lightrec_print_info(state);
1401
1402         return state->next_pc;
1403 }
1404
1405 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1406 {
1407         return lightrec_execute(state, pc, state->current_cycle);
1408 }
1409
1410 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1411 {
1412         struct block *block = lightrec_get_block(state, pc);
1413         if (!block)
1414                 return 0;
1415
1416         state->exit_flags = LIGHTREC_EXIT_NORMAL;
1417
1418         pc = lightrec_emulate_block(state, block, pc);
1419
1420         if (LOG_LEVEL >= INFO_L)
1421                 lightrec_print_info(state);
1422
1423         return pc;
1424 }
1425
1426 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1427 {
1428         lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1429         if (block->opcode_list)
1430                 lightrec_free_opcode_list(state, block);
1431         if (block->_jit)
1432                 _jit_destroy_state(block->_jit);
1433         lightrec_unregister(MEM_FOR_CODE, block->code_size);
1434         lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1435 }
1436
1437 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1438 {
1439         struct lightrec_cstate *cstate;
1440
1441         cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1442         if (!cstate)
1443                 return NULL;
1444
1445         cstate->reg_cache = lightrec_regcache_init(state);
1446         if (!cstate->reg_cache) {
1447                 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1448                 return NULL;
1449         }
1450
1451         cstate->state = state;
1452
1453         return cstate;
1454 }
1455
1456 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1457 {
1458         lightrec_free_regcache(cstate->reg_cache);
1459         lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1460 }
1461
1462 struct lightrec_state * lightrec_init(char *argv0,
1463                                       const struct lightrec_mem_map *map,
1464                                       size_t nb,
1465                                       const struct lightrec_ops *ops)
1466 {
1467         struct lightrec_state *state;
1468
1469         /* Sanity-check ops */
1470         if (!ops || !ops->cop2_op || !ops->enable_ram) {
1471                 pr_err("Missing callbacks in lightrec_ops structure\n");
1472                 return NULL;
1473         }
1474
1475         init_jit(argv0);
1476
1477         state = calloc(1, sizeof(*state) +
1478                        sizeof(*state->code_lut) * CODE_LUT_SIZE);
1479         if (!state)
1480                 goto err_finish_jit;
1481
1482         lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) +
1483                           sizeof(*state->code_lut) * CODE_LUT_SIZE);
1484
1485 #if ENABLE_TINYMM
1486         state->tinymm = tinymm_init(malloc, free, 4096);
1487         if (!state->tinymm)
1488                 goto err_free_state;
1489 #endif
1490
1491         state->block_cache = lightrec_blockcache_init(state);
1492         if (!state->block_cache)
1493                 goto err_free_tinymm;
1494
1495         if (ENABLE_THREADED_COMPILER) {
1496                 state->rec = lightrec_recompiler_init(state);
1497                 if (!state->rec)
1498                         goto err_free_block_cache;
1499
1500                 state->reaper = lightrec_reaper_init(state);
1501                 if (!state->reaper)
1502                         goto err_free_recompiler;
1503         } else {
1504                 state->cstate = lightrec_create_cstate(state);
1505                 if (!state->cstate)
1506                         goto err_free_block_cache;
1507         }
1508
1509         state->nb_maps = nb;
1510         state->maps = map;
1511
1512         memcpy(&state->ops, ops, sizeof(*ops));
1513
1514         state->dispatcher = generate_dispatcher(state);
1515         if (!state->dispatcher)
1516                 goto err_free_reaper;
1517
1518         state->c_wrapper_block = generate_wrapper(state);
1519         if (!state->c_wrapper_block)
1520                 goto err_free_dispatcher;
1521
1522         state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1523         state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1524         state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1525         state->c_wrappers[C_WRAPPER_CP] = lightrec_cp;
1526         state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1527         state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1528
1529         map = &state->maps[PSX_MAP_BIOS];
1530         state->offset_bios = (uintptr_t)map->address - map->pc;
1531
1532         map = &state->maps[PSX_MAP_SCRATCH_PAD];
1533         state->offset_scratch = (uintptr_t)map->address - map->pc;
1534
1535         map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1536         state->offset_ram = (uintptr_t)map->address - map->pc;
1537
1538         if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1539             state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1540             state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1541                 state->mirrors_mapped = true;
1542
1543         if (state->offset_bios == 0 &&
1544             state->offset_scratch == 0 &&
1545             state->offset_ram == 0 &&
1546             state->mirrors_mapped) {
1547                 pr_info("Memory map is perfect. Emitted code will be best.\n");
1548         } else {
1549                 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1550         }
1551
1552         return state;
1553
1554 err_free_dispatcher:
1555         lightrec_free_block(state, state->dispatcher);
1556 err_free_reaper:
1557         if (ENABLE_THREADED_COMPILER)
1558                 lightrec_reaper_destroy(state->reaper);
1559 err_free_recompiler:
1560         if (ENABLE_THREADED_COMPILER)
1561                 lightrec_free_recompiler(state->rec);
1562         else
1563                 lightrec_free_cstate(state->cstate);
1564 err_free_block_cache:
1565         lightrec_free_block_cache(state->block_cache);
1566 err_free_tinymm:
1567 #if ENABLE_TINYMM
1568         tinymm_shutdown(state->tinymm);
1569 err_free_state:
1570 #endif
1571         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1572                             sizeof(*state->code_lut) * CODE_LUT_SIZE);
1573         free(state);
1574 err_finish_jit:
1575         finish_jit();
1576         return NULL;
1577 }
1578
1579 void lightrec_destroy(struct lightrec_state *state)
1580 {
1581         /* Force a print info on destroy*/
1582         state->current_cycle = ~state->current_cycle;
1583         lightrec_print_info(state);
1584
1585         if (ENABLE_THREADED_COMPILER) {
1586                 lightrec_free_recompiler(state->rec);
1587                 lightrec_reaper_destroy(state->reaper);
1588         } else {
1589                 lightrec_free_cstate(state->cstate);
1590         }
1591
1592         lightrec_free_block_cache(state->block_cache);
1593         lightrec_free_block(state, state->dispatcher);
1594         lightrec_free_block(state, state->c_wrapper_block);
1595         finish_jit();
1596
1597 #if ENABLE_TINYMM
1598         tinymm_shutdown(state->tinymm);
1599 #endif
1600         lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1601                             sizeof(*state->code_lut) * CODE_LUT_SIZE);
1602         free(state);
1603 }
1604
1605 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1606 {
1607         u32 kaddr = kunseg(addr & ~0x3);
1608         const struct lightrec_mem_map *map = lightrec_get_map(state, NULL, kaddr);
1609
1610         if (map) {
1611                 if (map != &state->maps[PSX_MAP_KERNEL_USER_RAM])
1612                         return;
1613
1614                 /* Handle mirrors */
1615                 kaddr &= (state->maps[PSX_MAP_KERNEL_USER_RAM].length - 1);
1616
1617                 lightrec_invalidate_map(state, map, kaddr, len);
1618         }
1619 }
1620
1621 void lightrec_invalidate_all(struct lightrec_state *state)
1622 {
1623         memset(state->code_lut, 0, sizeof(*state->code_lut) * CODE_LUT_SIZE);
1624 }
1625
1626 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1627 {
1628         if (state->invalidate_from_dma_only != dma_only)
1629                 lightrec_invalidate_all(state);
1630
1631         state->invalidate_from_dma_only = dma_only;
1632 }
1633
1634 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1635 {
1636         if (flags != LIGHTREC_EXIT_NORMAL) {
1637                 state->exit_flags |= flags;
1638                 state->target_cycle = state->current_cycle;
1639         }
1640 }
1641
1642 u32 lightrec_exit_flags(struct lightrec_state *state)
1643 {
1644         return state->exit_flags;
1645 }
1646
1647 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1648 {
1649         return state->current_cycle;
1650 }
1651
1652 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1653 {
1654         state->current_cycle = cycles;
1655
1656         if (state->target_cycle < cycles)
1657                 state->target_cycle = cycles;
1658 }
1659
1660 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1661 {
1662         if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1663                 if (cycles < state->current_cycle)
1664                         cycles = state->current_cycle;
1665
1666                 state->target_cycle = cycles;
1667         }
1668 }
1669
1670 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1671 {
1672         return &state->regs;
1673 }