src/lj_ccallback.c

   1 /*
   2 ** FFI C callback handling.
   3 ** Copyright (C) 2005-2023 Mike Pall. See Copyright Notice in luajit.h
   4 */
   5
   6 #include "lj_obj.h"
   7
   8 #if LJ_HASFFI
   9
  10 #include "lj_gc.h"
  11 #include "lj_err.h"
  12 #include "lj_tab.h"
  13 #include "lj_state.h"
  14 #include "lj_frame.h"
  15 #include "lj_ctype.h"
  16 #include "lj_cconv.h"
  17 #include "lj_ccall.h"
  18 #include "lj_ccallback.h"
  19 #include "lj_target.h"
  20 #include "lj_mcode.h"
  21 #include "lj_trace.h"
  22 #include "lj_vm.h"
  23
  24 /* -- Target-specific handling of callback slots -------------------------- */
  25
  26 #define CALLBACK_MCODE_SIZE     (LJ_PAGESIZE * LJ_NUM_CBPAGE)
  27
  28 #if LJ_OS_NOJIT
  29
  30 /* Callbacks disabled. */
  31 #define CALLBACK_SLOT2OFS(slot) (0*(slot))
  32 #define CALLBACK_OFS2SLOT(ofs)  (0*(ofs))
  33 #define CALLBACK_MAX_SLOT       0
  34
  35 #elif LJ_TARGET_X86ORX64
  36
  37 #define CALLBACK_MCODE_HEAD     (LJ_64 ? 8 : 0)
  38 #define CALLBACK_MCODE_GROUP    (-2+1+2+(LJ_GC64 ? 10 : 5)+(LJ_64 ? 6 : 5))
  39
  40 #define CALLBACK_SLOT2OFS(slot) \
  41   (CALLBACK_MCODE_HEAD + CALLBACK_MCODE_GROUP*((slot)/32) + 4*(slot))
  42
  43 static MSize CALLBACK_OFS2SLOT(MSize ofs)
  44 {
  45   MSize group;
  46   ofs -= CALLBACK_MCODE_HEAD;
  47   group = ofs / (32*4 + CALLBACK_MCODE_GROUP);
  48   return (ofs % (32*4 + CALLBACK_MCODE_GROUP))/4 + group*32;
  49 }
  50
  51 #define CALLBACK_MAX_SLOT \
  52   (((CALLBACK_MCODE_SIZE-CALLBACK_MCODE_HEAD)/(CALLBACK_MCODE_GROUP+4*32))*32)
  53
  54 #elif LJ_TARGET_ARM
  55
  56 #define CALLBACK_MCODE_HEAD             32
  57
  58 #elif LJ_TARGET_ARM64
  59
  60 #define CALLBACK_MCODE_HEAD             32
  61
  62 #elif LJ_TARGET_PPC
  63
  64 #define CALLBACK_MCODE_HEAD             24
  65
  66 #elif LJ_TARGET_MIPS32
  67
  68 #define CALLBACK_MCODE_HEAD             20
  69
  70 #elif LJ_TARGET_MIPS64
  71
  72 #define CALLBACK_MCODE_HEAD             52
  73
  74 #else
  75
  76 /* Missing support for this architecture. */
  77 #define CALLBACK_SLOT2OFS(slot) (0*(slot))
  78 #define CALLBACK_OFS2SLOT(ofs)  (0*(ofs))
  79 #define CALLBACK_MAX_SLOT       0
  80
  81 #endif
  82
  83 #ifndef CALLBACK_SLOT2OFS
  84 #define CALLBACK_SLOT2OFS(slot)         (CALLBACK_MCODE_HEAD + 8*(slot))
  85 #define CALLBACK_OFS2SLOT(ofs)          (((ofs)-CALLBACK_MCODE_HEAD)/8)
  86 #define CALLBACK_MAX_SLOT               (CALLBACK_OFS2SLOT(CALLBACK_MCODE_SIZE))
  87 #endif
  88
  89 /* Convert callback slot number to callback function pointer. */
  90 static void *callback_slot2ptr(CTState *cts, MSize slot)
  91 {
  92   return (uint8_t *)cts->cb.mcode + CALLBACK_SLOT2OFS(slot);
  93 }
  94
  95 /* Convert callback function pointer to slot number. */
  96 MSize lj_ccallback_ptr2slot(CTState *cts, void *p)
  97 {
  98   uintptr_t ofs = (uintptr_t)((uint8_t *)p -(uint8_t *)cts->cb.mcode);
  99   if (ofs < CALLBACK_MCODE_SIZE) {
 100     MSize slot = CALLBACK_OFS2SLOT((MSize)ofs);
 101     if (CALLBACK_SLOT2OFS(slot) == (MSize)ofs)
 102       return slot;
 103   }
 104   return ~0u;  /* Not a known callback function pointer. */
 105 }
 106
 107 /* Initialize machine code for callback function pointers. */
 108 #if LJ_OS_NOJIT
 109 /* Disabled callback support. */
 110 #define callback_mcode_init(g, p)       (p)
 111 #elif LJ_TARGET_X86ORX64
 112 static void *callback_mcode_init(global_State *g, uint8_t *page)
 113 {
 114   uint8_t *p = page;
 115   uint8_t *target = (uint8_t *)(void *)lj_vm_ffi_callback;
 116   MSize slot;
 117 #if LJ_64
 118   *(void **)p = target; p += 8;
 119 #endif
 120   for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
 121     /* mov al, slot; jmp group */
 122     *p++ = XI_MOVrib | RID_EAX; *p++ = (uint8_t)slot;
 123     if ((slot & 31) == 31 || slot == CALLBACK_MAX_SLOT-1) {
 124       /* push ebp/rbp; mov ah, slot>>8; mov ebp, &g. */
 125       *p++ = XI_PUSH + RID_EBP;
 126       *p++ = XI_MOVrib | (RID_EAX+4); *p++ = (uint8_t)(slot >> 8);
 127 #if LJ_GC64
 128       *p++ = 0x48; *p++ = XI_MOVri | RID_EBP;
 129       *(uint64_t *)p = (uint64_t)(g); p += 8;
 130 #else
 131       *p++ = XI_MOVri | RID_EBP;
 132       *(int32_t *)p = i32ptr(g); p += 4;
 133 #endif
 134 #if LJ_64
 135       /* jmp [rip-pageofs] where lj_vm_ffi_callback is stored. */
 136       *p++ = XI_GROUP5; *p++ = XM_OFS0 + (XOg_JMP<<3) + RID_EBP;
 137       *(int32_t *)p = (int32_t)(page-(p+4)); p += 4;
 138 #else
 139       /* jmp lj_vm_ffi_callback. */
 140       *p++ = XI_JMP; *(int32_t *)p = target-(p+4); p += 4;
 141 #endif
 142     } else {
 143       *p++ = XI_JMPs; *p++ = (uint8_t)((2+2)*(31-(slot&31)) - 2);
 144     }
 145   }
 146   return p;
 147 }
 148 #elif LJ_TARGET_ARM
 149 static void *callback_mcode_init(global_State *g, uint32_t *page)
 150 {
 151   uint32_t *p = page;
 152   void *target = (void *)lj_vm_ffi_callback;
 153   MSize slot;
 154   /* This must match with the saveregs macro in buildvm_arm.dasc. */
 155   *p++ = ARMI_SUB|ARMF_D(RID_R12)|ARMF_N(RID_R12)|ARMF_M(RID_PC);
 156   *p++ = ARMI_PUSH|ARMF_N(RID_SP)|RSET_RANGE(RID_R4,RID_R11+1)|RID2RSET(RID_LR);
 157   *p++ = ARMI_SUB|ARMI_K12|ARMF_D(RID_R12)|ARMF_N(RID_R12)|CALLBACK_MCODE_HEAD;
 158   *p++ = ARMI_STR|ARMI_LS_P|ARMI_LS_W|ARMF_D(RID_R12)|ARMF_N(RID_SP)|(CFRAME_SIZE-4*9);
 159   *p++ = ARMI_LDR|ARMI_LS_P|ARMI_LS_U|ARMF_D(RID_R12)|ARMF_N(RID_PC);
 160   *p++ = ARMI_LDR|ARMI_LS_P|ARMI_LS_U|ARMF_D(RID_PC)|ARMF_N(RID_PC);
 161   *p++ = u32ptr(g);
 162   *p++ = u32ptr(target);
 163   for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
 164     *p++ = ARMI_MOV|ARMF_D(RID_R12)|ARMF_M(RID_PC);
 165     *p = ARMI_B | ((page-p-2) & 0x00ffffffu);
 166     p++;
 167   }
 168   return p;
 169 }
 170 #elif LJ_TARGET_ARM64
 171 static void *callback_mcode_init(global_State *g, uint32_t *page)
 172 {
 173   uint32_t *p = page;
 174   ASMFunction target = lj_vm_ffi_callback;
 175   MSize slot;
 176   *p++ = A64I_LE(A64I_LDRLx | A64F_D(RID_X11) | A64F_S19(4));
 177   *p++ = A64I_LE(A64I_LDRLx | A64F_D(RID_X10) | A64F_S19(5));
 178   *p++ = A64I_LE(A64I_BR_AUTH | A64F_N(RID_X11));
 179   *p++ = A64I_LE(A64I_NOP);
 180   ((ASMFunction *)p)[0] = target;
 181   ((void **)p)[1] = g;
 182   p += 4;
 183   for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
 184     *p++ = A64I_LE(A64I_MOVZw | A64F_D(RID_X9) | A64F_U16(slot));
 185     *p = A64I_LE(A64I_B | A64F_S26((page-p) & 0x03ffffffu));
 186     p++;
 187   }
 188   return p;
 189 }
 190 #elif LJ_TARGET_PPC
 191 static void *callback_mcode_init(global_State *g, uint32_t *page)
 192 {
 193   uint32_t *p = page;
 194   void *target = (void *)lj_vm_ffi_callback;
 195   MSize slot;
 196   *p++ = PPCI_LIS | PPCF_T(RID_TMP) | (u32ptr(target) >> 16);
 197   *p++ = PPCI_LIS | PPCF_T(RID_R12) | (u32ptr(g) >> 16);
 198   *p++ = PPCI_ORI | PPCF_A(RID_TMP)|PPCF_T(RID_TMP) | (u32ptr(target) & 0xffff);
 199   *p++ = PPCI_ORI | PPCF_A(RID_R12)|PPCF_T(RID_R12) | (u32ptr(g) & 0xffff);
 200   *p++ = PPCI_MTCTR | PPCF_T(RID_TMP);
 201   *p++ = PPCI_BCTR;
 202   for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
 203     *p++ = PPCI_LI | PPCF_T(RID_R11) | slot;
 204     *p = PPCI_B | (((page-p) & 0x00ffffffu) << 2);
 205     p++;
 206   }
 207   return p;
 208 }
 209 #elif LJ_TARGET_MIPS
 210 static void *callback_mcode_init(global_State *g, uint32_t *page)
 211 {
 212   uint32_t *p = page;
 213   uintptr_t target = (uintptr_t)(void *)lj_vm_ffi_callback;
 214   uintptr_t ug = (uintptr_t)(void *)g;
 215   MSize slot;
 216 #if LJ_TARGET_MIPS32
 217   *p++ = MIPSI_LUI | MIPSF_T(RID_R3) | (target >> 16);
 218   *p++ = MIPSI_LUI | MIPSF_T(RID_R2) | (ug >> 16);
 219 #else
 220   *p++ = MIPSI_LUI  | MIPSF_T(RID_R3) | (target >> 48);
 221   *p++ = MIPSI_LUI  | MIPSF_T(RID_R2) | (ug >> 48);
 222   *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | ((target >> 32) & 0xffff);
 223   *p++ = MIPSI_ORI  | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | ((ug >> 32) & 0xffff);
 224   *p++ = MIPSI_DSLL | MIPSF_D(RID_R3)|MIPSF_T(RID_R3) | MIPSF_A(16);
 225   *p++ = MIPSI_DSLL | MIPSF_D(RID_R2)|MIPSF_T(RID_R2) | MIPSF_A(16);
 226   *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | ((target >> 16) & 0xffff);
 227   *p++ = MIPSI_ORI  | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | ((ug >> 16) & 0xffff);
 228   *p++ = MIPSI_DSLL | MIPSF_D(RID_R3)|MIPSF_T(RID_R3) | MIPSF_A(16);
 229   *p++ = MIPSI_DSLL | MIPSF_D(RID_R2)|MIPSF_T(RID_R2) | MIPSF_A(16);
 230 #endif
 231   *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | (target & 0xffff);
 232   *p++ = MIPSI_JR | MIPSF_S(RID_R3);
 233   *p++ = MIPSI_ORI | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | (ug & 0xffff);
 234   for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
 235     *p = MIPSI_B | ((page-p-1) & 0x0000ffffu);
 236     p++;
 237     *p++ = MIPSI_LI | MIPSF_T(RID_R1) | slot;
 238   }
 239   return p;
 240 }
 241 #else
 242 /* Missing support for this architecture. */
 243 #define callback_mcode_init(g, p)       (p)
 244 #endif
 245
 246 /* -- Machine code management --------------------------------------------- */
 247
 248 #if LJ_TARGET_WINDOWS
 249
 250 #define WIN32_LEAN_AND_MEAN
 251 #include <windows.h>
 252
 253 #elif LJ_TARGET_POSIX
 254
 255 #include <sys/mman.h>
 256 #ifndef MAP_ANONYMOUS
 257 #define MAP_ANONYMOUS   MAP_ANON
 258 #endif
 259 #ifdef PROT_MPROTECT
 260 #define CCPROT_CREATE   (PROT_MPROTECT(PROT_EXEC))
 261 #else
 262 #define CCPROT_CREATE   0
 263 #endif
 264
 265 #endif
 266
 267 /* Allocate and initialize area for callback function pointers. */
 268 static void callback_mcode_new(CTState *cts)
 269 {
 270   size_t sz = (size_t)CALLBACK_MCODE_SIZE;
 271   void *p, *pe;
 272   if (CALLBACK_MAX_SLOT == 0)
 273     lj_err_caller(cts->L, LJ_ERR_FFI_CBACKOV);
 274 #if LJ_TARGET_WINDOWS
 275   p = LJ_WIN_VALLOC(NULL, sz, MEM_RESERVE|MEM_COMMIT, PAGE_READWRITE);
 276   if (!p)
 277     lj_err_caller(cts->L, LJ_ERR_FFI_CBACKOV);
 278 #elif LJ_TARGET_POSIX
 279   p = mmap(NULL, sz, (PROT_READ|PROT_WRITE|CCPROT_CREATE), MAP_PRIVATE|MAP_ANONYMOUS,
 280            -1, 0);
 281   if (p == MAP_FAILED)
 282     lj_err_caller(cts->L, LJ_ERR_FFI_CBACKOV);
 283 #else
 284   /* Fallback allocator. Fails if memory is not executable by default. */
 285   p = lj_mem_new(cts->L, sz);
 286 #endif
 287   cts->cb.mcode = p;
 288   pe = callback_mcode_init(cts->g, p);
 289   UNUSED(pe);
 290   lj_assertCTS((size_t)((char *)pe - (char *)p) <= sz,
 291                "miscalculated CALLBACK_MAX_SLOT");
 292   lj_mcode_sync(p, (char *)p + sz);
 293 #if LJ_TARGET_WINDOWS
 294   {
 295     DWORD oprot;
 296     LJ_WIN_VPROTECT(p, sz, PAGE_EXECUTE_READ, &oprot);
 297   }
 298 #elif LJ_TARGET_POSIX
 299   mprotect(p, sz, (PROT_READ|PROT_EXEC));
 300 #endif
 301 }
 302
 303 /* Free area for callback function pointers. */
 304 void lj_ccallback_mcode_free(CTState *cts)
 305 {
 306   size_t sz = (size_t)CALLBACK_MCODE_SIZE;
 307   void *p = cts->cb.mcode;
 308   if (p == NULL) return;
 309 #if LJ_TARGET_WINDOWS
 310   VirtualFree(p, 0, MEM_RELEASE);
 311   UNUSED(sz);
 312 #elif LJ_TARGET_POSIX
 313   munmap(p, sz);
 314 #else
 315   lj_mem_free(cts->g, p, sz);
 316 #endif
 317 }
 318
 319 /* -- C callback entry ---------------------------------------------------- */
 320
 321 /* Target-specific handling of register arguments. Similar to lj_ccall.c. */
 322 #if LJ_TARGET_X86
 323
 324 #define CALLBACK_HANDLE_REGARG \
 325   if (!isfp) {  /* Only non-FP values may be passed in registers. */ \
 326     if (n > 1) {  /* Anything > 32 bit is passed on the stack. */ \
 327       if (!LJ_ABI_WIN) ngpr = maxgpr;  /* Prevent reordering. */ \
 328     } else if (ngpr + 1 <= maxgpr) { \
 329       sp = &cts->cb.gpr[ngpr]; \
 330       ngpr += n; \
 331       goto done; \
 332     } \
 333   }
 334
 335 #elif LJ_TARGET_X64 && LJ_ABI_WIN
 336
 337 /* Windows/x64 argument registers are strictly positional (use ngpr). */
 338 #define CALLBACK_HANDLE_REGARG \
 339   if (isfp) { \
 340     if (ngpr < maxgpr) { sp = &cts->cb.fpr[ngpr++]; UNUSED(nfpr); goto done; } \
 341   } else { \
 342     if (ngpr < maxgpr) { sp = &cts->cb.gpr[ngpr++]; goto done; } \
 343   }
 344
 345 #elif LJ_TARGET_X64
 346
 347 #define CALLBACK_HANDLE_REGARG \
 348   if (isfp) { \
 349     if (nfpr + n <= CCALL_NARG_FPR) { \
 350       sp = &cts->cb.fpr[nfpr]; \
 351       nfpr += n; \
 352       goto done; \
 353     } \
 354   } else { \
 355     if (ngpr + n <= maxgpr) { \
 356       sp = &cts->cb.gpr[ngpr]; \
 357       ngpr += n; \
 358       goto done; \
 359     } \
 360   }
 361
 362 #elif LJ_TARGET_ARM
 363
 364 #if LJ_ABI_SOFTFP
 365
 366 #define CALLBACK_HANDLE_REGARG_FP1      UNUSED(isfp);
 367 #define CALLBACK_HANDLE_REGARG_FP2
 368
 369 #else
 370
 371 #define CALLBACK_HANDLE_REGARG_FP1 \
 372   if (isfp) { \
 373     if (n == 1) { \
 374       if (fprodd) { \
 375         sp = &cts->cb.fpr[fprodd-1]; \
 376         fprodd = 0; \
 377         goto done; \
 378       } else if (nfpr + 1 <= CCALL_NARG_FPR) { \
 379         sp = &cts->cb.fpr[nfpr++]; \
 380         fprodd = nfpr; \
 381         goto done; \
 382       } \
 383     } else { \
 384       if (nfpr + 1 <= CCALL_NARG_FPR) { \
 385         sp = &cts->cb.fpr[nfpr++]; \
 386         goto done; \
 387       } \
 388     } \
 389     fprodd = 0;  /* No reordering after the first FP value is on stack. */ \
 390   } else {
 391
 392 #define CALLBACK_HANDLE_REGARG_FP2      }
 393
 394 #endif
 395
 396 #define CALLBACK_HANDLE_REGARG \
 397   CALLBACK_HANDLE_REGARG_FP1 \
 398   if (n > 1) ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
 399   if (ngpr + n <= maxgpr) { \
 400     sp = &cts->cb.gpr[ngpr]; \
 401     ngpr += n; \
 402     goto done; \
 403   } CALLBACK_HANDLE_REGARG_FP2
 404
 405 #elif LJ_TARGET_ARM64
 406
 407 #define CALLBACK_HANDLE_REGARG \
 408   if (isfp) { \
 409     if (nfpr + n <= CCALL_NARG_FPR) { \
 410       sp = &cts->cb.fpr[nfpr]; \
 411       nfpr += n; \
 412       goto done; \
 413     } else { \
 414       nfpr = CCALL_NARG_FPR;  /* Prevent reordering. */ \
 415     } \
 416   } else { \
 417     if (!LJ_TARGET_OSX && n > 1) \
 418       ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
 419     if (ngpr + n <= maxgpr) { \
 420       sp = &cts->cb.gpr[ngpr]; \
 421       ngpr += n; \
 422       goto done; \
 423     } else { \
 424       ngpr = CCALL_NARG_GPR;  /* Prevent reordering. */ \
 425     } \
 426   }
 427
 428 #elif LJ_TARGET_PPC
 429
 430 #define CALLBACK_HANDLE_GPR \
 431   if (n > 1) { \
 432     lj_assertCTS(((LJ_ABI_SOFTFP && ctype_isnum(cta->info)) ||  /* double. */ \
 433                  ctype_isinteger(cta->info)) && n == 2,  /* int64_t. */ \
 434                  "bad GPR type"); \
 435     ngpr = (ngpr + 1u) & ~1u;  /* Align int64_t to regpair. */ \
 436   } \
 437   if (ngpr + n <= maxgpr) { \
 438     sp = &cts->cb.gpr[ngpr]; \
 439     ngpr += n; \
 440     goto done; \
 441   }
 442
 443 #if LJ_ABI_SOFTFP
 444 #define CALLBACK_HANDLE_REGARG \
 445   CALLBACK_HANDLE_GPR \
 446   UNUSED(isfp);
 447 #else
 448 #define CALLBACK_HANDLE_REGARG \
 449   if (isfp) { \
 450     if (nfpr + 1 <= CCALL_NARG_FPR) { \
 451       sp = &cts->cb.fpr[nfpr++]; \
 452       cta = ctype_get(cts, CTID_DOUBLE);  /* FPRs always hold doubles. */ \
 453       goto done; \
 454     } \
 455   } else {  /* Try to pass argument in GPRs. */ \
 456     CALLBACK_HANDLE_GPR \
 457   }
 458 #endif
 459
 460 #if !LJ_ABI_SOFTFP
 461 #define CALLBACK_HANDLE_RET \
 462   if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
 463     *(double *)dp = *(float *)dp;  /* FPRs always hold doubles. */
 464 #endif
 465
 466 #elif LJ_TARGET_MIPS32
 467
 468 #define CALLBACK_HANDLE_GPR \
 469   if (n > 1) ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
 470   if (ngpr + n <= maxgpr) { \
 471     sp = &cts->cb.gpr[ngpr]; \
 472     ngpr += n; \
 473     goto done; \
 474   }
 475
 476 #if !LJ_ABI_SOFTFP      /* MIPS32 hard-float */
 477 #define CALLBACK_HANDLE_REGARG \
 478   if (isfp && nfpr < CCALL_NARG_FPR) {  /* Try to pass argument in FPRs. */ \
 479     sp = (void *)((uint8_t *)&cts->cb.fpr[nfpr] + ((LJ_BE && n==1) ? 4 : 0)); \
 480     nfpr++; ngpr += n; \
 481     goto done; \
 482   } else {  /* Try to pass argument in GPRs. */ \
 483     nfpr = CCALL_NARG_FPR; \
 484     CALLBACK_HANDLE_GPR \
 485   }
 486 #else                   /* MIPS32 soft-float */
 487 #define CALLBACK_HANDLE_REGARG \
 488   CALLBACK_HANDLE_GPR \
 489   UNUSED(isfp);
 490 #endif
 491
 492 #define CALLBACK_HANDLE_RET \
 493   if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
 494     ((float *)dp)[1] = *(float *)dp;
 495
 496 #elif LJ_TARGET_MIPS64
 497
 498 #if !LJ_ABI_SOFTFP      /* MIPS64 hard-float */
 499 #define CALLBACK_HANDLE_REGARG \
 500   if (ngpr + n <= maxgpr) { \
 501     sp = isfp ? (void*) &cts->cb.fpr[ngpr] : (void*) &cts->cb.gpr[ngpr]; \
 502     ngpr += n; \
 503     goto done; \
 504   }
 505 #else                   /* MIPS64 soft-float */
 506 #define CALLBACK_HANDLE_REGARG \
 507   if (ngpr + n <= maxgpr) { \
 508     UNUSED(isfp); \
 509     sp = (void*) &cts->cb.gpr[ngpr]; \
 510     ngpr += n; \
 511     goto done; \
 512   }
 513 #endif
 514
 515 #define CALLBACK_HANDLE_RET \
 516   if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
 517     ((float *)dp)[1] = *(float *)dp;
 518
 519 #else
 520 #error "Missing calling convention definitions for this architecture"
 521 #endif
 522
 523 /* Convert and push callback arguments to Lua stack. */
 524 static void callback_conv_args(CTState *cts, lua_State *L)
 525 {
 526   TValue *o = L->top;
 527   intptr_t *stack = cts->cb.stack;
 528   MSize slot = cts->cb.slot;
 529   CTypeID id = 0, rid, fid;
 530   int gcsteps = 0;
 531   CType *ct;
 532   GCfunc *fn;
 533   int fntp;
 534   MSize ngpr = 0, nsp = 0, maxgpr = CCALL_NARG_GPR;
 535 #if CCALL_NARG_FPR
 536   MSize nfpr = 0;
 537 #if LJ_TARGET_ARM
 538   MSize fprodd = 0;
 539 #endif
 540 #endif
 541
 542   if (slot < cts->cb.sizeid && (id = cts->cb.cbid[slot]) != 0) {
 543     ct = ctype_get(cts, id);
 544     rid = ctype_cid(ct->info);  /* Return type. x86: +(spadj<<16). */
 545     fn = funcV(lj_tab_getint(cts->miscmap, (int32_t)slot));
 546     fntp = LJ_TFUNC;
 547   } else {  /* Must set up frame first, before throwing the error. */
 548     ct = NULL;
 549     rid = 0;
 550     fn = (GCfunc *)L;
 551     fntp = LJ_TTHREAD;
 552   }
 553   /* Continuation returns from callback. */
 554   if (LJ_FR2) {
 555     (o++)->u64 = LJ_CONT_FFI_CALLBACK;
 556     (o++)->u64 = rid;
 557   } else {
 558     o->u32.lo = LJ_CONT_FFI_CALLBACK;
 559     o->u32.hi = rid;
 560     o++;
 561   }
 562   setframe_gc(o, obj2gco(fn), fntp);
 563   if (LJ_FR2) o++;
 564   setframe_ftsz(o, ((char *)(o+1) - (char *)L->base) + FRAME_CONT);
 565   L->top = L->base = ++o;
 566   if (!ct)
 567     lj_err_caller(cts->L, LJ_ERR_FFI_BADCBACK);
 568   if (isluafunc(fn))
 569     setcframe_pc(L->cframe, proto_bc(funcproto(fn))+1);
 570   lj_state_checkstack(L, LUA_MINSTACK);  /* May throw. */
 571   o = L->base;  /* Might have been reallocated. */
 572
 573 #if LJ_TARGET_X86
 574   /* x86 has several different calling conventions. */
 575   switch (ctype_cconv(ct->info)) {
 576   case CTCC_FASTCALL: maxgpr = 2; break;
 577   case CTCC_THISCALL: maxgpr = 1; break;
 578   default: maxgpr = 0; break;
 579   }
 580 #endif
 581
 582   fid = ct->sib;
 583   while (fid) {
 584     CType *ctf = ctype_get(cts, fid);
 585     if (!ctype_isattrib(ctf->info)) {
 586       CType *cta;
 587       void *sp;
 588       CTSize sz;
 589       int isfp;
 590       MSize n;
 591       lj_assertCTS(ctype_isfield(ctf->info), "field expected");
 592       cta = ctype_rawchild(cts, ctf);
 593       isfp = ctype_isfp(cta->info);
 594       sz = (cta->size + CTSIZE_PTR-1) & ~(CTSIZE_PTR-1);
 595       n = sz / CTSIZE_PTR;  /* Number of GPRs or stack slots needed. */
 596
 597       CALLBACK_HANDLE_REGARG  /* Handle register arguments. */
 598
 599       /* Otherwise pass argument on stack. */
 600       if (CCALL_ALIGN_STACKARG && LJ_32 && sz == 8)
 601         nsp = (nsp + 1) & ~1u;  /* Align 64 bit argument on stack. */
 602       sp = &stack[nsp];
 603       nsp += n;
 604
 605     done:
 606       if (LJ_BE && cta->size < CTSIZE_PTR
 607 #if LJ_TARGET_MIPS64
 608           && !(isfp && nsp)
 609 #endif
 610          )
 611         sp = (void *)((uint8_t *)sp + CTSIZE_PTR-cta->size);
 612       gcsteps += lj_cconv_tv_ct(cts, cta, 0, o++, sp);
 613     }
 614     fid = ctf->sib;
 615   }
 616   L->top = o;
 617 #if LJ_TARGET_X86
 618   /* Store stack adjustment for returns from non-cdecl callbacks. */
 619   if (ctype_cconv(ct->info) != CTCC_CDECL) {
 620 #if LJ_FR2
 621     (L->base-3)->u64 |= (nsp << (16+2));
 622 #else
 623     (L->base-2)->u32.hi |= (nsp << (16+2));
 624 #endif
 625   }
 626 #endif
 627   while (gcsteps-- > 0)
 628     lj_gc_check(L);
 629 }
 630
 631 /* Convert Lua object to callback result. */
 632 static void callback_conv_result(CTState *cts, lua_State *L, TValue *o)
 633 {
 634 #if LJ_FR2
 635   CType *ctr = ctype_raw(cts, (uint16_t)(L->base-3)->u64);
 636 #else
 637   CType *ctr = ctype_raw(cts, (uint16_t)(L->base-2)->u32.hi);
 638 #endif
 639 #if LJ_TARGET_X86
 640   cts->cb.gpr[2] = 0;
 641 #endif
 642   if (!ctype_isvoid(ctr->info)) {
 643     uint8_t *dp = (uint8_t *)&cts->cb.gpr[0];
 644 #if CCALL_NUM_FPR
 645     if (ctype_isfp(ctr->info))
 646       dp = (uint8_t *)&cts->cb.fpr[0];
 647 #endif
 648 #if LJ_TARGET_ARM64 && LJ_BE
 649     if (ctype_isfp(ctr->info) && ctr->size == sizeof(float))
 650       dp = (uint8_t *)&cts->cb.fpr[0].f[1];
 651 #endif
 652     lj_cconv_ct_tv(cts, ctr, dp, o, 0);
 653 #ifdef CALLBACK_HANDLE_RET
 654     CALLBACK_HANDLE_RET
 655 #endif
 656     /* Extend returned integers to (at least) 32 bits. */
 657     if (ctype_isinteger_or_bool(ctr->info) && ctr->size < 4) {
 658       if (ctr->info & CTF_UNSIGNED)
 659         *(uint32_t *)dp = ctr->size == 1 ? (uint32_t)*(uint8_t *)dp :
 660                                            (uint32_t)*(uint16_t *)dp;
 661       else
 662         *(int32_t *)dp = ctr->size == 1 ? (int32_t)*(int8_t *)dp :
 663                                           (int32_t)*(int16_t *)dp;
 664     }
 665 #if LJ_TARGET_MIPS64 || (LJ_TARGET_ARM64 && LJ_BE)
 666     /* Always sign-extend results to 64 bits. Even a soft-fp 'float'. */
 667     if (ctr->size <= 4 &&
 668         (LJ_ABI_SOFTFP || ctype_isinteger_or_bool(ctr->info)))
 669       *(int64_t *)dp = (int64_t)*(int32_t *)dp;
 670 #endif
 671 #if LJ_TARGET_X86
 672     if (ctype_isfp(ctr->info))
 673       cts->cb.gpr[2] = ctr->size == sizeof(float) ? 1 : 2;
 674 #endif
 675   }
 676 }
 677
 678 /* Enter callback. */
 679 lua_State * LJ_FASTCALL lj_ccallback_enter(CTState *cts, void *cf)
 680 {
 681   lua_State *L = cts->L;
 682   global_State *g = cts->g;
 683   lj_assertG(L != NULL, "uninitialized cts->L in callback");
 684   if (tvref(g->jit_base)) {
 685     setstrV(L, L->top++, lj_err_str(L, LJ_ERR_FFI_BADCBACK));
 686     if (g->panic) g->panic(L);
 687     exit(EXIT_FAILURE);
 688   }
 689   lj_trace_abort(g);  /* Never record across callback. */
 690   /* Setup C frame. */
 691   cframe_prev(cf) = L->cframe;
 692   setcframe_L(cf, L);
 693   cframe_errfunc(cf) = -1;
 694   cframe_nres(cf) = 0;
 695   L->cframe = cf;
 696   callback_conv_args(cts, L);
 697   return L;  /* Now call the function on this stack. */
 698 }
 699
 700 /* Leave callback. */
 701 void LJ_FASTCALL lj_ccallback_leave(CTState *cts, TValue *o)
 702 {
 703   lua_State *L = cts->L;
 704   GCfunc *fn;
 705   TValue *obase = L->base;
 706   L->base = L->top;  /* Keep continuation frame for throwing errors. */
 707   if (o >= L->base) {
 708     /* PC of RET* is lost. Point to last line for result conv. errors. */
 709     fn = curr_func(L);
 710     if (isluafunc(fn)) {
 711       GCproto *pt = funcproto(fn);
 712       setcframe_pc(L->cframe, proto_bc(pt)+pt->sizebc+1);
 713     }
 714   }
 715   callback_conv_result(cts, L, o);
 716   /* Finally drop C frame and continuation frame. */
 717   L->top -= 2+2*LJ_FR2;
 718   L->base = obase;
 719   L->cframe = cframe_prev(L->cframe);
 720   cts->cb.slot = 0;  /* Blacklist C function that called the callback. */
 721 }
 722
 723 /* -- C callback management ----------------------------------------------- */
 724
 725 /* Get an unused slot in the callback slot table. */
 726 static MSize callback_slot_new(CTState *cts, CType *ct)
 727 {
 728   CTypeID id = ctype_typeid(cts, ct);
 729   CTypeID1 *cbid = cts->cb.cbid;
 730   MSize top;
 731   for (top = cts->cb.topid; top < cts->cb.sizeid; top++)
 732     if (LJ_LIKELY(cbid[top] == 0))
 733       goto found;
 734 #if CALLBACK_MAX_SLOT
 735   if (top >= CALLBACK_MAX_SLOT)
 736 #endif
 737     lj_err_caller(cts->L, LJ_ERR_FFI_CBACKOV);
 738   if (!cts->cb.mcode)
 739     callback_mcode_new(cts);
 740   lj_mem_growvec(cts->L, cbid, cts->cb.sizeid, CALLBACK_MAX_SLOT, CTypeID1);
 741   cts->cb.cbid = cbid;
 742   memset(cbid+top, 0, (cts->cb.sizeid-top)*sizeof(CTypeID1));
 743 found:
 744   cbid[top] = id;
 745   cts->cb.topid = top+1;
 746   return top;
 747 }
 748
 749 /* Check for function pointer and supported argument/result types. */
 750 static CType *callback_checkfunc(CTState *cts, CType *ct)
 751 {
 752   int narg = 0;
 753   if (!ctype_isptr(ct->info) || (LJ_64 && ct->size != CTSIZE_PTR))
 754     return NULL;
 755   ct = ctype_rawchild(cts, ct);
 756   if (ctype_isfunc(ct->info)) {
 757     CType *ctr = ctype_rawchild(cts, ct);
 758     CTypeID fid = ct->sib;
 759     if (!(ctype_isvoid(ctr->info) || ctype_isenum(ctr->info) ||
 760           ctype_isptr(ctr->info) || (ctype_isnum(ctr->info) && ctr->size <= 8)))
 761       return NULL;
 762     if ((ct->info & CTF_VARARG))
 763       return NULL;
 764     while (fid) {
 765       CType *ctf = ctype_get(cts, fid);
 766       if (!ctype_isattrib(ctf->info)) {
 767         CType *cta;
 768         lj_assertCTS(ctype_isfield(ctf->info), "field expected");
 769         cta = ctype_rawchild(cts, ctf);
 770         if (!(ctype_isenum(cta->info) || ctype_isptr(cta->info) ||
 771               (ctype_isnum(cta->info) && cta->size <= 8)) ||
 772             ++narg >= LUA_MINSTACK-3)
 773           return NULL;
 774       }
 775       fid = ctf->sib;
 776     }
 777     return ct;
 778   }
 779   return NULL;
 780 }
 781
 782 /* Create a new callback and return the callback function pointer. */
 783 void *lj_ccallback_new(CTState *cts, CType *ct, GCfunc *fn)
 784 {
 785   ct = callback_checkfunc(cts, ct);
 786   if (ct) {
 787     MSize slot = callback_slot_new(cts, ct);
 788     GCtab *t = cts->miscmap;
 789     setfuncV(cts->L, lj_tab_setint(cts->L, t, (int32_t)slot), fn);
 790     lj_gc_anybarriert(cts->L, t);
 791     return callback_slot2ptr(cts, slot);
 792   }
 793   return NULL;  /* Bad conversion. */
 794 }
 795
 796 #endif