src/lj_snap.c

   1 /*
   2 ** Snapshot handling.
   3 ** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
   4 */
   5
   6 #define lj_snap_c
   7 #define LUA_CORE
   8
   9 #include "lj_obj.h"
  10
  11 #if LJ_HASJIT
  12
  13 #include "lj_gc.h"
  14 #include "lj_tab.h"
  15 #include "lj_state.h"
  16 #include "lj_frame.h"
  17 #include "lj_bc.h"
  18 #include "lj_ir.h"
  19 #include "lj_jit.h"
  20 #include "lj_iropt.h"
  21 #include "lj_trace.h"
  22 #include "lj_snap.h"
  23 #include "lj_target.h"
  24 #if LJ_HASFFI
  25 #include "lj_ctype.h"
  26 #include "lj_cdata.h"
  27 #endif
  28
  29 /* Some local macros to save typing. Undef'd at the end. */
  30 #define IR(ref)         (&J->cur.ir[(ref)])
  31
  32 /* Pass IR on to next optimization in chain (FOLD). */
  33 #define emitir(ot, a, b)        (lj_ir_set(J, (ot), (a), (b)), lj_opt_fold(J))
  34
  35 /* Emit raw IR without passing through optimizations. */
  36 #define emitir_raw(ot, a, b)    (lj_ir_set(J, (ot), (a), (b)), lj_ir_emit(J))
  37
  38 /* -- Snapshot buffer allocation ------------------------------------------ */
  39
  40 /* Grow snapshot buffer. */
  41 void lj_snap_grow_buf_(jit_State *J, MSize need)
  42 {
  43   MSize maxsnap = (MSize)J->param[JIT_P_maxsnap];
  44   if (need > maxsnap)
  45     lj_trace_err(J, LJ_TRERR_SNAPOV);
  46   lj_mem_growvec(J->L, J->snapbuf, J->sizesnap, maxsnap, SnapShot);
  47   J->cur.snap = J->snapbuf;
  48 }
  49
  50 /* Grow snapshot map buffer. */
  51 void lj_snap_grow_map_(jit_State *J, MSize need)
  52 {
  53   if (need < 2*J->sizesnapmap)
  54     need = 2*J->sizesnapmap;
  55   else if (need < 64)
  56     need = 64;
  57   J->snapmapbuf = (SnapEntry *)lj_mem_realloc(J->L, J->snapmapbuf,
  58                     J->sizesnapmap*sizeof(SnapEntry), need*sizeof(SnapEntry));
  59   J->cur.snapmap = J->snapmapbuf;
  60   J->sizesnapmap = need;
  61 }
  62
  63 /* -- Snapshot generation ------------------------------------------------- */
  64
  65 /* Add all modified slots to the snapshot. */
  66 static MSize snapshot_slots(jit_State *J, SnapEntry *map, BCReg nslots)
  67 {
  68   IRRef retf = J->chain[IR_RETF];  /* Limits SLOAD restore elimination. */
  69   BCReg s;
  70   MSize n = 0;
  71   for (s = 0; s < nslots; s++) {
  72     TRef tr = J->slot[s];
  73     IRRef ref = tref_ref(tr);
  74     if (ref) {
  75       SnapEntry sn = SNAP_TR(s, tr);
  76       IRIns *ir = IR(ref);
  77       if (!(sn & (SNAP_CONT|SNAP_FRAME)) &&
  78           ir->o == IR_SLOAD && ir->op1 == s && ref > retf) {
  79         /* No need to snapshot unmodified non-inherited slots. */
  80         if (!(ir->op2 & IRSLOAD_INHERIT))
  81           continue;
  82         /* No need to restore readonly slots and unmodified non-parent slots. */
  83         if (!(LJ_DUALNUM && (ir->op2 & IRSLOAD_CONVERT)) &&
  84             (ir->op2 & (IRSLOAD_READONLY|IRSLOAD_PARENT)) != IRSLOAD_PARENT)
  85           sn |= SNAP_NORESTORE;
  86       }
  87       if (LJ_SOFTFP && irt_isnum(ir->t))
  88         sn |= SNAP_SOFTFPNUM;
  89       map[n++] = sn;
  90     }
  91   }
  92   return n;
  93 }
  94
  95 /* Add frame links at the end of the snapshot. */
  96 static BCReg snapshot_framelinks(jit_State *J, SnapEntry *map)
  97 {
  98   cTValue *frame = J->L->base - 1;
  99   cTValue *lim = J->L->base - J->baseslot;
 100   cTValue *ftop = frame + funcproto(frame_func(frame))->framesize;
 101   MSize f = 0;
 102   map[f++] = SNAP_MKPC(J->pc);  /* The current PC is always the first entry. */
 103   while (frame > lim) {  /* Backwards traversal of all frames above base. */
 104     if (frame_islua(frame)) {
 105       map[f++] = SNAP_MKPC(frame_pc(frame));
 106       frame = frame_prevl(frame);
 107       if (frame + funcproto(frame_func(frame))->framesize > ftop)
 108         ftop = frame + funcproto(frame_func(frame))->framesize;
 109     } else if (frame_iscont(frame)) {
 110       map[f++] = SNAP_MKFTSZ(frame_ftsz(frame));
 111       map[f++] = SNAP_MKPC(frame_contpc(frame));
 112       frame = frame_prevd(frame);
 113     } else {
 114       lua_assert(!frame_isc(frame));
 115       map[f++] = SNAP_MKFTSZ(frame_ftsz(frame));
 116       frame = frame_prevd(frame);
 117     }
 118   }
 119   lua_assert(f == (MSize)(1 + J->framedepth));
 120   return (BCReg)(ftop - lim);
 121 }
 122
 123 /* Take a snapshot of the current stack. */
 124 static void snapshot_stack(jit_State *J, SnapShot *snap, MSize nsnapmap)
 125 {
 126   BCReg nslots = J->baseslot + J->maxslot;
 127   MSize nent;
 128   SnapEntry *p;
 129   /* Conservative estimate. */
 130   lj_snap_grow_map(J, nsnapmap + nslots + (MSize)J->framedepth+1);
 131   p = &J->cur.snapmap[nsnapmap];
 132   nent = snapshot_slots(J, p, nslots);
 133   snap->topslot = (uint8_t)snapshot_framelinks(J, p + nent);
 134   snap->mapofs = (uint16_t)nsnapmap;
 135   snap->ref = (IRRef1)J->cur.nins;
 136   snap->nent = (uint8_t)nent;
 137   snap->nslots = (uint8_t)nslots;
 138   snap->count = 0;
 139   J->cur.nsnapmap = (uint16_t)(nsnapmap + nent + 1 + J->framedepth);
 140 }
 141
 142 /* Add or merge a snapshot. */
 143 void lj_snap_add(jit_State *J)
 144 {
 145   MSize nsnap = J->cur.nsnap;
 146   MSize nsnapmap = J->cur.nsnapmap;
 147   /* Merge if no ins. inbetween or if requested and no guard inbetween. */
 148   if (J->mergesnap ? !irt_isguard(J->guardemit) :
 149       (nsnap > 0 && J->cur.snap[nsnap-1].ref == J->cur.nins)) {
 150     if (nsnap == 1) {  /* But preserve snap #0 PC. */
 151       emitir_raw(IRT(IR_NOP, IRT_NIL), 0, 0);
 152       goto nomerge;
 153     }
 154     nsnapmap = J->cur.snap[--nsnap].mapofs;
 155   } else {
 156   nomerge:
 157     lj_snap_grow_buf(J, nsnap+1);
 158     J->cur.nsnap = (uint16_t)(nsnap+1);
 159   }
 160   J->mergesnap = 0;
 161   J->guardemit.irt = 0;
 162   snapshot_stack(J, &J->cur.snap[nsnap], nsnapmap);
 163 }
 164
 165 /* -- Snapshot modification ----------------------------------------------- */
 166
 167 #define SNAP_USEDEF_SLOTS       (LJ_MAX_JSLOTS+LJ_STACK_EXTRA)
 168
 169 /* Find unused slots with reaching-definitions bytecode data-flow analysis. */
 170 static BCReg snap_usedef(jit_State *J, uint8_t *udf,
 171                          const BCIns *pc, BCReg maxslot)
 172 {
 173   BCReg s;
 174   GCobj *o;
 175
 176   if (maxslot == 0) return 0;
 177 #ifdef LUAJIT_USE_VALGRIND
 178   /* Avoid errors for harmless reads beyond maxslot. */
 179   memset(udf, 1, SNAP_USEDEF_SLOTS);
 180 #else
 181   memset(udf, 1, maxslot);
 182 #endif
 183
 184   /* Treat open upvalues as used. */
 185   o = gcref(J->L->openupval);
 186   while (o) {
 187     if (uvval(gco2uv(o)) < J->L->base) break;
 188     udf[uvval(gco2uv(o)) - J->L->base] = 0;
 189     o = gcref(o->gch.nextgc);
 190   }
 191
 192 #define USE_SLOT(s)             udf[(s)] &= ~1
 193 #define DEF_SLOT(s)             udf[(s)] *= 3
 194
 195   /* Scan through following bytecode and check for uses/defs. */
 196   lua_assert(pc >= proto_bc(J->pt) && pc < proto_bc(J->pt) + J->pt->sizebc);
 197   for (;;) {
 198     BCIns ins = *pc++;
 199     BCOp op = bc_op(ins);
 200     switch (bcmode_b(op)) {
 201     case BCMvar: USE_SLOT(bc_b(ins)); break;
 202     default: break;
 203     }
 204     switch (bcmode_c(op)) {
 205     case BCMvar: USE_SLOT(bc_c(ins)); break;
 206     case BCMrbase:
 207       lua_assert(op == BC_CAT);
 208       for (s = bc_b(ins); s <= bc_c(ins); s++) USE_SLOT(s);
 209       for (; s < maxslot; s++) DEF_SLOT(s);
 210       break;
 211     case BCMjump:
 212     handle_jump: {
 213       BCReg minslot = bc_a(ins);
 214       if (op >= BC_FORI && op <= BC_JFORL) minslot += FORL_EXT;
 215       else if (op >= BC_ITERL && op <= BC_JITERL) minslot += bc_b(pc[-2])-1;
 216       else if (op == BC_UCLO) { pc += bc_j(ins); break; }
 217       for (s = minslot; s < maxslot; s++) DEF_SLOT(s);
 218       return minslot < maxslot ? minslot : maxslot;
 219       }
 220     case BCMlit:
 221       if (op == BC_JFORL || op == BC_JITERL || op == BC_JLOOP) {
 222         goto handle_jump;
 223       } else if (bc_isret(op)) {
 224         BCReg top = op == BC_RETM ? maxslot : (bc_a(ins) + bc_d(ins)-1);
 225         for (s = 0; s < bc_a(ins); s++) DEF_SLOT(s);
 226         for (; s < top; s++) USE_SLOT(s);
 227         for (; s < maxslot; s++) DEF_SLOT(s);
 228         return 0;
 229       }
 230       break;
 231     case BCMfunc: return maxslot;  /* NYI: will abort, anyway. */
 232     default: break;
 233     }
 234     switch (bcmode_a(op)) {
 235     case BCMvar: USE_SLOT(bc_a(ins)); break;
 236     case BCMdst:
 237        if (!(op == BC_ISTC || op == BC_ISFC)) DEF_SLOT(bc_a(ins));
 238        break;
 239     case BCMbase:
 240       if (op >= BC_CALLM && op <= BC_VARG) {
 241         BCReg top = (op == BC_CALLM || op == BC_CALLMT || bc_c(ins) == 0) ?
 242                     maxslot : (bc_a(ins) + bc_c(ins));
 243         s = bc_a(ins) - ((op == BC_ITERC || op == BC_ITERN) ? 3 : 0);
 244         for (; s < top; s++) USE_SLOT(s);
 245         for (; s < maxslot; s++) DEF_SLOT(s);
 246         if (op == BC_CALLT || op == BC_CALLMT) {
 247           for (s = 0; s < bc_a(ins); s++) DEF_SLOT(s);
 248           return 0;
 249         }
 250       } else if (op == BC_KNIL) {
 251         for (s = bc_a(ins); s <= bc_d(ins); s++) DEF_SLOT(s);
 252       } else if (op == BC_TSETM) {
 253         for (s = bc_a(ins)-1; s < maxslot; s++) USE_SLOT(s);
 254       }
 255       break;
 256     default: break;
 257     }
 258     lua_assert(pc >= proto_bc(J->pt) && pc < proto_bc(J->pt) + J->pt->sizebc);
 259   }
 260
 261 #undef USE_SLOT
 262 #undef DEF_SLOT
 263
 264   return 0;  /* unreachable */
 265 }
 266
 267 /* Purge dead slots before the next snapshot. */
 268 void lj_snap_purge(jit_State *J)
 269 {
 270   uint8_t udf[SNAP_USEDEF_SLOTS];
 271   BCReg maxslot = J->maxslot;
 272   BCReg s = snap_usedef(J, udf, J->pc, maxslot);
 273   for (; s < maxslot; s++)
 274     if (udf[s] != 0)
 275       J->base[s] = 0;  /* Purge dead slots. */
 276 }
 277
 278 /* Shrink last snapshot. */
 279 void lj_snap_shrink(jit_State *J)
 280 {
 281   SnapShot *snap = &J->cur.snap[J->cur.nsnap-1];
 282   SnapEntry *map = &J->cur.snapmap[snap->mapofs];
 283   MSize n, m, nlim, nent = snap->nent;
 284   uint8_t udf[SNAP_USEDEF_SLOTS];
 285   BCReg maxslot = J->maxslot;
 286   BCReg minslot = snap_usedef(J, udf, snap_pc(map[nent]), maxslot);
 287   BCReg baseslot = J->baseslot;
 288   maxslot += baseslot;
 289   minslot += baseslot;
 290   snap->nslots = (uint8_t)maxslot;
 291   for (n = m = 0; n < nent; n++) {  /* Remove unused slots from snapshot. */
 292     BCReg s = snap_slot(map[n]);
 293     if (s < minslot || (s < maxslot && udf[s-baseslot] == 0))
 294       map[m++] = map[n];  /* Only copy used slots. */
 295   }
 296   snap->nent = (uint8_t)m;
 297   nlim = J->cur.nsnapmap - snap->mapofs - 1;
 298   while (n <= nlim) map[m++] = map[n++];  /* Move PC + frame links down. */
 299   J->cur.nsnapmap = (uint16_t)(snap->mapofs + m);  /* Free up space in map. */
 300 }
 301
 302 /* -- Snapshot access ----------------------------------------------------- */
 303
 304 /* Initialize a Bloom Filter with all renamed refs.
 305 ** There are very few renames (often none), so the filter has
 306 ** very few bits set. This makes it suitable for negative filtering.
 307 */
 308 static BloomFilter snap_renamefilter(GCtrace *T, SnapNo lim)
 309 {
 310   BloomFilter rfilt = 0;
 311   IRIns *ir;
 312   for (ir = &T->ir[T->nins-1]; ir->o == IR_RENAME; ir--)
 313     if (ir->op2 <= lim)
 314       bloomset(rfilt, ir->op1);
 315   return rfilt;
 316 }
 317
 318 /* Process matching renames to find the original RegSP. */
 319 static RegSP snap_renameref(GCtrace *T, SnapNo lim, IRRef ref, RegSP rs)
 320 {
 321   IRIns *ir;
 322   for (ir = &T->ir[T->nins-1]; ir->o == IR_RENAME; ir--)
 323     if (ir->op1 == ref && ir->op2 <= lim)
 324       rs = ir->prev;
 325   return rs;
 326 }
 327
 328 /* Copy RegSP from parent snapshot to the parent links of the IR. */
 329 IRIns *lj_snap_regspmap(GCtrace *T, SnapNo snapno, IRIns *ir)
 330 {
 331   SnapShot *snap = &T->snap[snapno];
 332   SnapEntry *map = &T->snapmap[snap->mapofs];
 333   BloomFilter rfilt = snap_renamefilter(T, snapno);
 334   MSize n = 0;
 335   IRRef ref = 0;
 336   for ( ; ; ir++) {
 337     uint32_t rs;
 338     if (ir->o == IR_SLOAD) {
 339       if (!(ir->op2 & IRSLOAD_PARENT)) break;
 340       for ( ; ; n++) {
 341         lua_assert(n < snap->nent);
 342         if (snap_slot(map[n]) == ir->op1) {
 343           ref = snap_ref(map[n++]);
 344           break;
 345         }
 346       }
 347     } else if (LJ_SOFTFP && ir->o == IR_HIOP) {
 348       ref++;
 349     } else if (ir->o == IR_PVAL) {
 350       ref = ir->op1 + REF_BIAS;
 351     } else {
 352       break;
 353     }
 354     rs = T->ir[ref].prev;
 355     if (bloomtest(rfilt, ref))
 356       rs = snap_renameref(T, snapno, ref, rs);
 357     ir->prev = (uint16_t)rs;
 358     lua_assert(regsp_used(rs));
 359   }
 360   return ir;
 361 }
 362
 363 /* -- Snapshot replay ----------------------------------------------------- */
 364
 365 /* Replay constant from parent trace. */
 366 static TRef snap_replay_const(jit_State *J, IRIns *ir)
 367 {
 368   /* Only have to deal with constants that can occur in stack slots. */
 369   switch ((IROp)ir->o) {
 370   case IR_KPRI: return TREF_PRI(irt_type(ir->t));
 371   case IR_KINT: return lj_ir_kint(J, ir->i);
 372   case IR_KGC: return lj_ir_kgc(J, ir_kgc(ir), irt_t(ir->t));
 373   case IR_KNUM: return lj_ir_k64(J, IR_KNUM, ir_knum(ir));
 374   case IR_KINT64: return lj_ir_k64(J, IR_KINT64, ir_kint64(ir));
 375   case IR_KPTR: return lj_ir_kptr(J, ir_kptr(ir));  /* Continuation. */
 376   default: lua_assert(0); return TREF_NIL; break;
 377   }
 378 }
 379
 380 /* De-duplicate parent reference. */
 381 static TRef snap_dedup(jit_State *J, SnapEntry *map, MSize nmax, IRRef ref)
 382 {
 383   MSize j;
 384   for (j = 0; j < nmax; j++)
 385     if (snap_ref(map[j]) == ref)
 386       return J->slot[snap_slot(map[j])] & ~(SNAP_CONT|SNAP_FRAME);
 387   return 0;
 388 }
 389
 390 /* Emit parent reference with de-duplication. */
 391 static TRef snap_pref(jit_State *J, GCtrace *T, SnapEntry *map, MSize nmax,
 392                       BloomFilter seen, IRRef ref)
 393 {
 394   IRIns *ir = &T->ir[ref];
 395   TRef tr;
 396   if (irref_isk(ref))
 397     tr = snap_replay_const(J, ir);
 398   else if (!regsp_used(ir->prev))
 399     tr = 0;
 400   else if (!bloomtest(seen, ref) || (tr = snap_dedup(J, map, nmax, ref)) == 0)
 401     tr = emitir(IRT(IR_PVAL, irt_type(ir->t)), ref - REF_BIAS, 0);
 402   return tr;
 403 }
 404
 405 /* Check whether a sunk store corresponds to an allocation. Slow path. */
 406 static int snap_sunk_store2(jit_State *J, IRIns *ira, IRIns *irs)
 407 {
 408   if (irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
 409       irs->o == IR_FSTORE || irs->o == IR_XSTORE) {
 410     IRIns *irk = IR(irs->op1);
 411     if (irk->o == IR_AREF || irk->o == IR_HREFK)
 412       irk = IR(irk->op1);
 413     return (IR(irk->op1) == ira);
 414   }
 415   return 0;
 416 }
 417
 418 /* Check whether a sunk store corresponds to an allocation. Fast path. */
 419 static LJ_AINLINE int snap_sunk_store(jit_State *J, IRIns *ira, IRIns *irs)
 420 {
 421   if (irs->s != 255)
 422     return (ira + irs->s == irs);  /* Fast check. */
 423   return snap_sunk_store2(J, ira, irs);
 424 }
 425
 426 /* Replay snapshot state to setup side trace. */
 427 void lj_snap_replay(jit_State *J, GCtrace *T)
 428 {
 429   SnapShot *snap = &T->snap[J->exitno];
 430   SnapEntry *map = &T->snapmap[snap->mapofs];
 431   MSize n, nent = snap->nent;
 432   BloomFilter seen = 0;
 433   int pass23 = 0;
 434   J->framedepth = 0;
 435   /* Emit IR for slots inherited from parent snapshot. */
 436   for (n = 0; n < nent; n++) {
 437     SnapEntry sn = map[n];
 438     BCReg s = snap_slot(sn);
 439     IRRef ref = snap_ref(sn);
 440     IRIns *ir = &T->ir[ref];
 441     TRef tr;
 442     /* The bloom filter avoids O(nent^2) overhead for de-duping slots. */
 443     if (bloomtest(seen, ref) && (tr = snap_dedup(J, map, n, ref)) != 0)
 444       goto setslot;
 445     bloomset(seen, ref);
 446     if (irref_isk(ref)) {
 447       tr = snap_replay_const(J, ir);
 448     } else if (!regsp_used(ir->prev)) {
 449       pass23 = 1;
 450       lua_assert(s != 0);
 451       tr = s;
 452     } else {
 453       IRType t = irt_type(ir->t);
 454       uint32_t mode = IRSLOAD_INHERIT|IRSLOAD_PARENT;
 455       if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM)) t = IRT_NUM;
 456       if (ir->o == IR_SLOAD) mode |= (ir->op2 & IRSLOAD_READONLY);
 457       tr = emitir_raw(IRT(IR_SLOAD, t), s, mode);
 458     }
 459   setslot:
 460     J->slot[s] = tr | (sn&(SNAP_CONT|SNAP_FRAME));  /* Same as TREF_* flags. */
 461     J->framedepth += ((sn & (SNAP_CONT|SNAP_FRAME)) && s);
 462     if ((sn & SNAP_FRAME))
 463       J->baseslot = s+1;
 464   }
 465   if (pass23) {
 466     IRIns *irlast = &T->ir[snap->ref];
 467     pass23 = 0;
 468     /* Emit dependent PVALs. */
 469     for (n = 0; n < nent; n++) {
 470       SnapEntry sn = map[n];
 471       IRRef refp = snap_ref(sn);
 472       IRIns *ir = &T->ir[refp];
 473       if (regsp_reg(ir->r) == RID_SUNK) {
 474         if (J->slot[snap_slot(sn)] != snap_slot(sn)) continue;
 475         pass23 = 1;
 476         lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP ||
 477                    ir->o == IR_CNEW || ir->o == IR_CNEWI);
 478         if (ir->op1 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op1);
 479         if (ir->op2 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op2);
 480         if (LJ_HASFFI && ir->o == IR_CNEWI) {
 481           if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP)
 482             snap_pref(J, T, map, nent, seen, (ir+1)->op2);
 483         } else {
 484           IRIns *irs;
 485           for (irs = ir+1; irs < irlast; irs++)
 486             if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
 487               if (snap_pref(J, T, map, nent, seen, irs->op2) == 0)
 488                 snap_pref(J, T, map, nent, seen, T->ir[irs->op2].op1);
 489               else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
 490                        irs+1 < irlast && (irs+1)->o == IR_HIOP)
 491                 snap_pref(J, T, map, nent, seen, (irs+1)->op2);
 492             }
 493         }
 494       } else if (!irref_isk(refp) && !regsp_used(ir->prev)) {
 495         lua_assert(ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
 496         J->slot[snap_slot(sn)] = snap_pref(J, T, map, nent, seen, ir->op1);
 497       }
 498     }
 499     /* Replay sunk instructions. */
 500     for (n = 0; pass23 && n < nent; n++) {
 501       SnapEntry sn = map[n];
 502       IRRef refp = snap_ref(sn);
 503       IRIns *ir = &T->ir[refp];
 504       if (regsp_reg(ir->r) == RID_SUNK) {
 505         TRef op1, op2;
 506         if (J->slot[snap_slot(sn)] != snap_slot(sn)) {  /* De-dup allocs. */
 507           J->slot[snap_slot(sn)] = J->slot[J->slot[snap_slot(sn)]];
 508           continue;
 509         }
 510         op1 = ir->op1;
 511         if (op1 >= T->nk) op1 = snap_pref(J, T, map, nent, seen, op1);
 512         op2 = ir->op2;
 513         if (op2 >= T->nk) op2 = snap_pref(J, T, map, nent, seen, op2);
 514         if (LJ_HASFFI && ir->o == IR_CNEWI) {
 515           if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP) {
 516             lj_needsplit(J);  /* Emit joining HIOP. */
 517             op2 = emitir_raw(IRT(IR_HIOP, IRT_I64), op2,
 518                              snap_pref(J, T, map, nent, seen, (ir+1)->op2));
 519           }
 520           J->slot[snap_slot(sn)] = emitir(ir->ot, op1, op2);
 521         } else {
 522           IRIns *irs;
 523           TRef tr = emitir(ir->ot, op1, op2);
 524           J->slot[snap_slot(sn)] = tr;
 525           for (irs = ir+1; irs < irlast; irs++)
 526             if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
 527               IRIns *irr = &T->ir[irs->op1];
 528               TRef val, key = irr->op2, tmp = tr;
 529               if (irr->o != IR_FREF) {
 530                 IRIns *irk = &T->ir[key];
 531                 if (irr->o == IR_HREFK)
 532                   key = lj_ir_kslot(J, snap_replay_const(J, &T->ir[irk->op1]),
 533                                     irk->op2);
 534                 else
 535                   key = snap_replay_const(J, irk);
 536                 if (irr->o == IR_HREFK || irr->o == IR_AREF) {
 537                   IRIns *irf = &T->ir[irr->op1];
 538                   tmp = emitir(irf->ot, tmp, irf->op2);
 539                 }
 540               }
 541               tmp = emitir(irr->ot, tmp, key);
 542               val = snap_pref(J, T, map, nent, seen, irs->op2);
 543               if (val == 0) {
 544                 IRIns *irc = &T->ir[irs->op2];
 545                 lua_assert(irc->o == IR_CONV && irc->op2 == IRCONV_NUM_INT);
 546                 val = snap_pref(J, T, map, nent, seen, irc->op1);
 547                 val = emitir(IRTN(IR_CONV), val, IRCONV_NUM_INT);
 548               } else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
 549                          irs+1 < irlast && (irs+1)->o == IR_HIOP) {
 550                 IRType t = IRT_I64;
 551                 if (LJ_SOFTFP && irt_type((irs+1)->t) == IRT_SOFTFP)
 552                   t = IRT_NUM;
 553                 lj_needsplit(J);
 554                 if (irref_isk(irs->op2) && irref_isk((irs+1)->op2)) {
 555                   uint64_t k = (uint32_t)T->ir[irs->op2].i +
 556                                ((uint64_t)T->ir[(irs+1)->op2].i << 32);
 557                   val = lj_ir_k64(J, t == IRT_I64 ? IR_KINT64 : IR_KNUM,
 558                                   lj_ir_k64_find(J, k));
 559                 } else {
 560                   val = emitir_raw(IRT(IR_HIOP, t), val,
 561                           snap_pref(J, T, map, nent, seen, (irs+1)->op2));
 562                 }
 563                 tmp = emitir(IRT(irs->o, t), tmp, val);
 564                 continue;
 565               }
 566               tmp = emitir(irs->ot, tmp, val);
 567             }
 568         }
 569       }
 570     }
 571   }
 572   J->base = J->slot + J->baseslot;
 573   J->maxslot = snap->nslots - J->baseslot;
 574   lj_snap_add(J);
 575   if (pass23)  /* Need explicit GC step _after_ initial snapshot. */
 576     emitir_raw(IRTG(IR_GCSTEP, IRT_NIL), 0, 0);
 577 }
 578
 579 /* -- Snapshot restore ---------------------------------------------------- */
 580
 581 static void snap_unsink(jit_State *J, GCtrace *T, ExitState *ex,
 582                         SnapNo snapno, BloomFilter rfilt,
 583                         IRIns *ir, TValue *o);
 584
 585 /* Restore a value from the trace exit state. */
 586 static void snap_restoreval(jit_State *J, GCtrace *T, ExitState *ex,
 587                             SnapNo snapno, BloomFilter rfilt,
 588                             IRRef ref, TValue *o)
 589 {
 590   IRIns *ir = &T->ir[ref];
 591   IRType1 t = ir->t;
 592   RegSP rs = ir->prev;
 593   if (irref_isk(ref)) {  /* Restore constant slot. */
 594     lj_ir_kvalue(J->L, o, ir);
 595     return;
 596   }
 597   if (LJ_UNLIKELY(bloomtest(rfilt, ref)))
 598     rs = snap_renameref(T, snapno, ref, rs);
 599   if (ra_hasspill(regsp_spill(rs))) {  /* Restore from spill slot. */
 600     int32_t *sps = &ex->spill[regsp_spill(rs)];
 601     if (irt_isinteger(t)) {
 602       setintV(o, *sps);
 603 #if !LJ_SOFTFP
 604     } else if (irt_isnum(t)) {
 605       o->u64 = *(uint64_t *)sps;
 606 #endif
 607     } else if (LJ_64 && irt_islightud(t)) {
 608       /* 64 bit lightuserdata which may escape already has the tag bits. */
 609       o->u64 = *(uint64_t *)sps;
 610     } else {
 611       lua_assert(!irt_ispri(t));  /* PRI refs never have a spill slot. */
 612       setgcrefi(o->gcr, *sps);
 613       setitype(o, irt_toitype(t));
 614     }
 615   } else {  /* Restore from register. */
 616     Reg r = regsp_reg(rs);
 617     if (ra_noreg(r)) {
 618       lua_assert(ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
 619       snap_restoreval(J, T, ex, snapno, rfilt, ir->op1, o);
 620       if (LJ_DUALNUM) setnumV(o, (lua_Number)intV(o));
 621       return;
 622     } else if (irt_isinteger(t)) {
 623       setintV(o, (int32_t)ex->gpr[r-RID_MIN_GPR]);
 624 #if !LJ_SOFTFP
 625     } else if (irt_isnum(t)) {
 626       setnumV(o, ex->fpr[r-RID_MIN_FPR]);
 627 #endif
 628     } else if (LJ_64 && irt_islightud(t)) {
 629       /* 64 bit lightuserdata which may escape already has the tag bits. */
 630       o->u64 = ex->gpr[r-RID_MIN_GPR];
 631     } else {
 632       if (!irt_ispri(t))
 633         setgcrefi(o->gcr, ex->gpr[r-RID_MIN_GPR]);
 634       setitype(o, irt_toitype(t));
 635     }
 636   }
 637 }
 638
 639 #if LJ_HASFFI
 640 /* Restore raw data from the trace exit state. */
 641 static void snap_restoredata(GCtrace *T, ExitState *ex,
 642                              SnapNo snapno, BloomFilter rfilt,
 643                              IRRef ref, void *dst, CTSize sz)
 644 {
 645   IRIns *ir = &T->ir[ref];
 646   RegSP rs = ir->prev;
 647   int32_t *src;
 648   uint64_t tmp;
 649   if (irref_isk(ref)) {
 650     if (ir->o == IR_KNUM || ir->o == IR_KINT64) {
 651       src = mref(ir->ptr, int32_t);
 652     } else if (sz == 8) {
 653       tmp = (uint64_t)(uint32_t)ir->i;
 654       src = (int32_t *)&tmp;
 655     } else {
 656       src = &ir->i;
 657     }
 658   } else {
 659     if (LJ_UNLIKELY(bloomtest(rfilt, ref)))
 660       rs = snap_renameref(T, snapno, ref, rs);
 661     if (ra_hasspill(regsp_spill(rs))) {
 662       src = &ex->spill[regsp_spill(rs)];
 663       if (sz == 8 && !irt_is64(ir->t)) {
 664         tmp = (uint64_t)(uint32_t)*src;
 665         src = (int32_t *)&tmp;
 666       }
 667     } else {
 668       Reg r = regsp_reg(rs);
 669       if (ra_noreg(r)) {
 670         /* Note: this assumes CNEWI is never used for SOFTFP split numbers. */
 671         lua_assert(sz == 8 && ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
 672         snap_restoredata(T, ex, snapno, rfilt, ir->op1, dst, 4);
 673         *(lua_Number *)dst = (lua_Number)*(int32_t *)dst;
 674         return;
 675       }
 676       src = (int32_t *)&ex->gpr[r-RID_MIN_GPR];
 677 #if !LJ_SOFTFP
 678       if (r >= RID_MAX_GPR) {
 679         src = (int32_t *)&ex->fpr[r-RID_MIN_FPR];
 680 #if LJ_TARGET_PPC
 681         if (sz == 4) {  /* PPC FPRs are always doubles. */
 682           *(float *)dst = (float)*(double *)src;
 683           return;
 684         }
 685 #else
 686         if (LJ_BE && sz == 4) src++;
 687 #endif
 688       }
 689 #endif
 690     }
 691   }
 692   lua_assert(sz == 1 || sz == 2 || sz == 4 || sz == 8);
 693   if (sz == 4) *(int32_t *)dst = *src;
 694   else if (sz == 8) *(int64_t *)dst = *(int64_t *)src;
 695   else if (sz == 1) *(int8_t *)dst = (int8_t)*src;
 696   else *(int16_t *)dst = (int16_t)*src;
 697 }
 698 #endif
 699
 700 /* Unsink allocation from the trace exit state. Unsink sunk stores. */
 701 static void snap_unsink(jit_State *J, GCtrace *T, ExitState *ex,
 702                         SnapNo snapno, BloomFilter rfilt,
 703                         IRIns *ir, TValue *o)
 704 {
 705   lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP ||
 706              ir->o == IR_CNEW || ir->o == IR_CNEWI);
 707 #if LJ_HASFFI
 708   if (ir->o == IR_CNEW || ir->o == IR_CNEWI) {
 709     CTState *cts = ctype_ctsG(J2G(J));
 710     CTypeID id = (CTypeID)T->ir[ir->op1].i;
 711     CTSize sz = lj_ctype_size(cts, id);
 712     GCcdata *cd = lj_cdata_new(cts, id, sz);
 713     setcdataV(J->L, o, cd);
 714     if (ir->o == IR_CNEWI) {
 715       uint8_t *p = (uint8_t *)cdataptr(cd);
 716       lua_assert(sz == 4 || sz == 8);
 717       if (LJ_32 && sz == 8 && ir+1 < T->ir + T->nins && (ir+1)->o == IR_HIOP) {
 718         snap_restoredata(T, ex, snapno, rfilt, (ir+1)->op2, LJ_LE?p+4:p, 4);
 719         if (LJ_BE) p += 4;
 720         sz = 4;
 721       }
 722       snap_restoredata(T, ex, snapno, rfilt, ir->op2, p, sz);
 723     } else {
 724       IRIns *irs, *irlast = &T->ir[T->snap[snapno].ref];
 725       for (irs = ir+1; irs < irlast; irs++)
 726         if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
 727           IRIns *iro = &T->ir[T->ir[irs->op1].op2];
 728           uint8_t *p = (uint8_t *)cd;
 729           CTSize szs;
 730           lua_assert(irs->o == IR_XSTORE && T->ir[irs->op1].o == IR_ADD);
 731           lua_assert(iro->o == IR_KINT || iro->o == IR_KINT64);
 732           if (irt_is64(irs->t)) szs = 8;
 733           else if (irt_isi8(irs->t) || irt_isu8(irs->t)) szs = 1;
 734           else if (irt_isi16(irs->t) || irt_isu16(irs->t)) szs = 2;
 735           else szs = 4;
 736           if (LJ_64 && iro->o == IR_KINT64)
 737             p += (int64_t)ir_k64(iro)->u64;
 738           else
 739             p += iro->i;
 740           lua_assert(p >= (uint8_t *)cdataptr(cd) &&
 741                      p + szs <= (uint8_t *)cdataptr(cd) + sz);
 742           if (LJ_32 && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
 743             lua_assert(szs == 4);
 744             snap_restoredata(T, ex, snapno, rfilt, (irs+1)->op2, LJ_LE?p+4:p,4);
 745             if (LJ_BE) p += 4;
 746           }
 747           snap_restoredata(T, ex, snapno, rfilt, irs->op2, p, szs);
 748         }
 749     }
 750   } else
 751 #endif
 752   {
 753     IRIns *irs, *irlast;
 754     GCtab *t = ir->o == IR_TNEW ? lj_tab_new(J->L, ir->op1, ir->op2) :
 755                                   lj_tab_dup(J->L, ir_ktab(&T->ir[ir->op1]));
 756     settabV(J->L, o, t);
 757     irlast = &T->ir[T->snap[snapno].ref];
 758     for (irs = ir+1; irs < irlast; irs++)
 759       if (irs->r == RID_SINK && snap_sunk_store(J, ir, irs)) {
 760         IRIns *irk = &T->ir[irs->op1];
 761         TValue tmp, *val;
 762         lua_assert(irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
 763                    irs->o == IR_FSTORE);
 764         if (irk->o == IR_FREF) {
 765           lua_assert(irk->op2 == IRFL_TAB_META);
 766           snap_restoreval(J, T, ex, snapno, rfilt, irs->op2, &tmp);
 767           /* NOBARRIER: The table is new (marked white). */
 768           setgcref(t->metatable, obj2gco(tabV(&tmp)));
 769         } else {
 770           irk = &T->ir[irk->op2];
 771           if (irk->o == IR_KSLOT) irk = &T->ir[irk->op1];
 772           lj_ir_kvalue(J->L, &tmp, irk);
 773           val = lj_tab_set(J->L, t, &tmp);
 774           /* NOBARRIER: The table is new (marked white). */
 775           snap_restoreval(J, T, ex, snapno, rfilt, irs->op2, val);
 776           if (LJ_SOFTFP && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
 777             snap_restoreval(J, T, ex, snapno, rfilt, (irs+1)->op2, &tmp);
 778             val->u32.hi = tmp.u32.lo;
 779           }
 780         }
 781       }
 782   }
 783 }
 784
 785 /* Restore interpreter state from exit state with the help of a snapshot. */
 786 const BCIns *lj_snap_restore(jit_State *J, void *exptr)
 787 {
 788   ExitState *ex = (ExitState *)exptr;
 789   SnapNo snapno = J->exitno;  /* For now, snapno == exitno. */
 790   GCtrace *T = traceref(J, J->parent);
 791   SnapShot *snap = &T->snap[snapno];
 792   MSize n, nent = snap->nent;
 793   SnapEntry *map = &T->snapmap[snap->mapofs];
 794   SnapEntry *flinks = &T->snapmap[snap_nextofs(T, snap)-1];
 795   int32_t ftsz0;
 796   TValue *frame;
 797   BloomFilter rfilt = snap_renamefilter(T, snapno);
 798   const BCIns *pc = snap_pc(map[nent]);
 799   lua_State *L = J->L;
 800
 801   /* Set interpreter PC to the next PC to get correct error messages. */
 802   setcframe_pc(cframe_raw(L->cframe), pc+1);
 803
 804   /* Make sure the stack is big enough for the slots from the snapshot. */
 805   if (LJ_UNLIKELY(L->base + snap->topslot >= tvref(L->maxstack))) {
 806     L->top = curr_topL(L);
 807     lj_state_growstack(L, snap->topslot - curr_proto(L)->framesize);
 808   }
 809
 810   /* Fill stack slots with data from the registers and spill slots. */
 811   frame = L->base-1;
 812   ftsz0 = frame_ftsz(frame);  /* Preserve link to previous frame in slot #0. */
 813   for (n = 0; n < nent; n++) {
 814     SnapEntry sn = map[n];
 815     if (!(sn & SNAP_NORESTORE)) {
 816       TValue *o = &frame[snap_slot(sn)];
 817       IRRef ref = snap_ref(sn);
 818       IRIns *ir = &T->ir[ref];
 819       if (ir->r == RID_SUNK) {
 820         MSize j;
 821         for (j = 0; j < n; j++)
 822           if (snap_ref(map[j]) == ref) {  /* De-duplicate sunk allocations. */
 823             copyTV(L, o, &frame[snap_slot(map[j])]);
 824             goto dupslot;
 825           }
 826         snap_unsink(J, T, ex, snapno, rfilt, ir, o);
 827       dupslot:
 828         continue;
 829       }
 830       snap_restoreval(J, T, ex, snapno, rfilt, ref, o);
 831       if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM) && tvisint(o)) {
 832         TValue tmp;
 833         snap_restoreval(J, T, ex, snapno, rfilt, ref+1, &tmp);
 834         o->u32.hi = tmp.u32.lo;
 835       } else if ((sn & (SNAP_CONT|SNAP_FRAME))) {
 836         /* Overwrite tag with frame link. */
 837         o->fr.tp.ftsz = snap_slot(sn) != 0 ? (int32_t)*flinks-- : ftsz0;
 838         L->base = o+1;
 839       }
 840     }
 841   }
 842   lua_assert(map + nent == flinks);
 843
 844   /* Compute current stack top. */
 845   switch (bc_op(*pc)) {
 846   case BC_CALLM: case BC_CALLMT: case BC_RETM: case BC_TSETM:
 847     L->top = frame + snap->nslots;
 848     break;
 849   default:
 850     L->top = curr_topL(L);
 851     break;
 852   }
 853   return pc;
 854 }
 855
 856 #undef IR
 857 #undef emitir_raw
 858 #undef emitir
 859
 860 #endif