?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
new beta-0.90.0
[luatex.git] / source / libs / luajit / LuaJIT-src / src / lj_snap.c
blob8638d9edbfd8667872cb3a472134c39d713be18b
1 /*
2 ** Snapshot handling.
3 ** Copyright (C) 2005-2016 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 /* Pass IR on to next optimization in chain (FOLD). */
30 #define emitir(ot, a, b) (lj_ir_set(J, (ot), (a), (b)), lj_opt_fold(J))
31
32 /* Emit raw IR without passing through optimizations. */
33 #define emitir_raw(ot, a, b) (lj_ir_set(J, (ot), (a), (b)), lj_ir_emit(J))
34
35 /* -- Snapshot buffer allocation ------------------------------------------ */
36
37 /* Grow snapshot buffer. */
38 void lj_snap_grow_buf_(jit_State *J, MSize need)
39 {
40 MSize maxsnap = (MSize)J->param[JIT_P_maxsnap];
41 if (need > maxsnap)
42 lj_trace_err(J, LJ_TRERR_SNAPOV);
43 lj_mem_growvec(J->L, J->snapbuf, J->sizesnap, maxsnap, SnapShot);
44 J->cur.snap = J->snapbuf;
45 }
46
47 /* Grow snapshot map buffer. */
48 void lj_snap_grow_map_(jit_State *J, MSize need)
49 {
50 if (need < 2*J->sizesnapmap)
51 need = 2*J->sizesnapmap;
52 else if (need < 64)
53 need = 64;
54 J->snapmapbuf = (SnapEntry *)lj_mem_realloc(J->L, J->snapmapbuf,
55 J->sizesnapmap*sizeof(SnapEntry), need*sizeof(SnapEntry));
56 J->cur.snapmap = J->snapmapbuf;
57 J->sizesnapmap = need;
58 }
59
60 /* -- Snapshot generation ------------------------------------------------- */
61
62 /* Add all modified slots to the snapshot. */
63 static MSize snapshot_slots(jit_State *J, SnapEntry *map, BCReg nslots)
64 {
65 IRRef retf = J->chain[IR_RETF]; /* Limits SLOAD restore elimination. */
66 BCReg s;
67 MSize n = 0;
68 for (s = 0; s < nslots; s++) {
69 TRef tr = J->slot[s];
70 IRRef ref = tref_ref(tr);
71 if (ref) {
72 SnapEntry sn = SNAP_TR(s, tr);
73 IRIns *ir = &J->cur.ir[ref];
74 if (!(sn & (SNAP_CONT|SNAP_FRAME)) &&
75 ir->o == IR_SLOAD && ir->op1 == s && ref > retf) {
76 /* No need to snapshot unmodified non-inherited slots. */
77 if (!(ir->op2 & IRSLOAD_INHERIT))
78 continue;
79 /* No need to restore readonly slots and unmodified non-parent slots. */
80 if (!(LJ_DUALNUM && (ir->op2 & IRSLOAD_CONVERT)) &&
81 (ir->op2 & (IRSLOAD_READONLY|IRSLOAD_PARENT)) != IRSLOAD_PARENT)
82 sn |= SNAP_NORESTORE;
83 }
84 if (LJ_SOFTFP && irt_isnum(ir->t))
85 sn |= SNAP_SOFTFPNUM;
86 map[n++] = sn;
87 }
88 }
89 return n;
90 }
91
92 /* Add frame links at the end of the snapshot. */
93 static BCReg snapshot_framelinks(jit_State *J, SnapEntry *map)
94 {
95 cTValue *frame = J->L->base - 1;
96 cTValue *lim = J->L->base - J->baseslot;
97 GCfunc *fn = frame_func(frame);
98 cTValue *ftop = isluafunc(fn) ? (frame+funcproto(fn)->framesize) : J->L->top;
99 MSize f = 0;
100 lua_assert(!LJ_FR2); /* TODO_FR2: store 64 bit PCs. */
101 map[f++] = SNAP_MKPC(J->pc); /* The current PC is always the first entry. */
102 while (frame > lim) { /* Backwards traversal of all frames above base. */
103 if (frame_islua(frame)) {
104 map[f++] = SNAP_MKPC(frame_pc(frame));
105 frame = frame_prevl(frame);
106 } else if (frame_iscont(frame)) {
107 map[f++] = SNAP_MKFTSZ(frame_ftsz(frame));
108 map[f++] = SNAP_MKPC(frame_contpc(frame));
109 frame = frame_prevd(frame);
110 } else {
111 lua_assert(!frame_isc(frame));
112 map[f++] = SNAP_MKFTSZ(frame_ftsz(frame));
113 frame = frame_prevd(frame);
114 continue;
115 }
116 if (frame + funcproto(frame_func(frame))->framesize > ftop)
117 ftop = frame + funcproto(frame_func(frame))->framesize;
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 ((nsnap > 0 && J->cur.snap[nsnap-1].ref == J->cur.nins) ||
149 (J->mergesnap && !irt_isguard(J->guardemit))) {
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)+LJ_FR2);
243 if (LJ_FR2) DEF_SLOT(bc_a(ins)+1);
244 s = bc_a(ins) - ((op == BC_ITERC || op == BC_ITERN) ? 3 : 0);
245 for (; s < top; s++) USE_SLOT(s);
246 for (; s < maxslot; s++) DEF_SLOT(s);
247 if (op == BC_CALLT || op == BC_CALLMT) {
248 for (s = 0; s < bc_a(ins); s++) DEF_SLOT(s);
249 return 0;
250 }
251 } else if (op == BC_KNIL) {
252 for (s = bc_a(ins); s <= bc_d(ins); s++) DEF_SLOT(s);
253 } else if (op == BC_TSETM) {
254 for (s = bc_a(ins)-1; s < maxslot; s++) USE_SLOT(s);
255 }
256 break;
257 default: break;
258 }
259 lua_assert(pc >= proto_bc(J->pt) && pc < proto_bc(J->pt) + J->pt->sizebc);
260 }
261
262 #undef USE_SLOT
263 #undef DEF_SLOT
264
265 return 0; /* unreachable */
266 }
267
268 /* Purge dead slots before the next snapshot. */
269 void lj_snap_purge(jit_State *J)
270 {
271 uint8_t udf[SNAP_USEDEF_SLOTS];
272 BCReg maxslot = J->maxslot;
273 BCReg s = snap_usedef(J, udf, J->pc, maxslot);
274 for (; s < maxslot; s++)
275 if (udf[s] != 0)
276 J->base[s] = 0; /* Purge dead slots. */
277 }
278
279 /* Shrink last snapshot. */
280 void lj_snap_shrink(jit_State *J)
281 {
282 SnapShot *snap = &J->cur.snap[J->cur.nsnap-1];
283 SnapEntry *map = &J->cur.snapmap[snap->mapofs];
284 MSize n, m, nlim, nent = snap->nent;
285 uint8_t udf[SNAP_USEDEF_SLOTS];
286 BCReg maxslot = J->maxslot;
287 BCReg minslot = snap_usedef(J, udf, snap_pc(map[nent]), maxslot);
288 BCReg baseslot = J->baseslot;
289 maxslot += baseslot;
290 minslot += baseslot;
291 snap->nslots = (uint8_t)maxslot;
292 for (n = m = 0; n < nent; n++) { /* Remove unused slots from snapshot. */
293 BCReg s = snap_slot(map[n]);
294 if (s < minslot || (s < maxslot && udf[s-baseslot] == 0))
295 map[m++] = map[n]; /* Only copy used slots. */
296 }
297 snap->nent = (uint8_t)m;
298 nlim = J->cur.nsnapmap - snap->mapofs - 1;
299 while (n <= nlim) map[m++] = map[n++]; /* Move PC + frame links down. */
300 J->cur.nsnapmap = (uint16_t)(snap->mapofs + m); /* Free up space in map. */
301 }
302
303 /* -- Snapshot access ----------------------------------------------------- */
304
305 /* Initialize a Bloom Filter with all renamed refs.
306 ** There are very few renames (often none), so the filter has
307 ** very few bits set. This makes it suitable for negative filtering.
308 */
309 static BloomFilter snap_renamefilter(GCtrace *T, SnapNo lim)
310 {
311 BloomFilter rfilt = 0;
312 IRIns *ir;
313 for (ir = &T->ir[T->nins-1]; ir->o == IR_RENAME; ir--)
314 if (ir->op2 <= lim)
315 bloomset(rfilt, ir->op1);
316 return rfilt;
317 }
318
319 /* Process matching renames to find the original RegSP. */
320 static RegSP snap_renameref(GCtrace *T, SnapNo lim, IRRef ref, RegSP rs)
321 {
322 IRIns *ir;
323 for (ir = &T->ir[T->nins-1]; ir->o == IR_RENAME; ir--)
324 if (ir->op1 == ref && ir->op2 <= lim)
325 rs = ir->prev;
326 return rs;
327 }
328
329 /* Copy RegSP from parent snapshot to the parent links of the IR. */
330 IRIns *lj_snap_regspmap(GCtrace *T, SnapNo snapno, IRIns *ir)
331 {
332 SnapShot *snap = &T->snap[snapno];
333 SnapEntry *map = &T->snapmap[snap->mapofs];
334 BloomFilter rfilt = snap_renamefilter(T, snapno);
335 MSize n = 0;
336 IRRef ref = 0;
337 for ( ; ; ir++) {
338 uint32_t rs;
339 if (ir->o == IR_SLOAD) {
340 if (!(ir->op2 & IRSLOAD_PARENT)) break;
341 for ( ; ; n++) {
342 lua_assert(n < snap->nent);
343 if (snap_slot(map[n]) == ir->op1) {
344 ref = snap_ref(map[n++]);
345 break;
346 }
347 }
348 } else if (LJ_SOFTFP && ir->o == IR_HIOP) {
349 ref++;
350 } else if (ir->o == IR_PVAL) {
351 ref = ir->op1 + REF_BIAS;
352 } else {
353 break;
354 }
355 rs = T->ir[ref].prev;
356 if (bloomtest(rfilt, ref))
357 rs = snap_renameref(T, snapno, ref, rs);
358 ir->prev = (uint16_t)rs;
359 lua_assert(regsp_used(rs));
360 }
361 return ir;
362 }
363
364 /* -- Snapshot replay ----------------------------------------------------- */
365
366 /* Replay constant from parent trace. */
367 static TRef snap_replay_const(jit_State *J, IRIns *ir)
368 {
369 /* Only have to deal with constants that can occur in stack slots. */
370 switch ((IROp)ir->o) {
371 case IR_KPRI: return TREF_PRI(irt_type(ir->t));
372 case IR_KINT: return lj_ir_kint(J, ir->i);
373 case IR_KGC: return lj_ir_kgc(J, ir_kgc(ir), irt_t(ir->t));
374 case IR_KNUM: return lj_ir_k64(J, IR_KNUM, ir_knum(ir));
375 case IR_KINT64: return lj_ir_k64(J, IR_KINT64, ir_kint64(ir));
376 case IR_KPTR: return lj_ir_kptr(J, ir_kptr(ir)); /* Continuation. */
377 default: lua_assert(0); return TREF_NIL; break;
378 }
379 }
380
381 /* De-duplicate parent reference. */
382 static TRef snap_dedup(jit_State *J, SnapEntry *map, MSize nmax, IRRef ref)
383 {
384 MSize j;
385 for (j = 0; j < nmax; j++)
386 if (snap_ref(map[j]) == ref)
387 return J->slot[snap_slot(map[j])] & ~(SNAP_CONT|SNAP_FRAME);
388 return 0;
389 }
390
391 /* Emit parent reference with de-duplication. */
392 static TRef snap_pref(jit_State *J, GCtrace *T, SnapEntry *map, MSize nmax,
393 BloomFilter seen, IRRef ref)
394 {
395 IRIns *ir = &T->ir[ref];
396 TRef tr;
397 if (irref_isk(ref))
398 tr = snap_replay_const(J, ir);
399 else if (!regsp_used(ir->prev))
400 tr = 0;
401 else if (!bloomtest(seen, ref) || (tr = snap_dedup(J, map, nmax, ref)) == 0)
402 tr = emitir(IRT(IR_PVAL, irt_type(ir->t)), ref - REF_BIAS, 0);
403 return tr;
404 }
405
406 /* Check whether a sunk store corresponds to an allocation. Slow path. */
407 static int snap_sunk_store2(GCtrace *T, IRIns *ira, IRIns *irs)
408 {
409 if (irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
410 irs->o == IR_FSTORE || irs->o == IR_XSTORE) {
411 IRIns *irk = &T->ir[irs->op1];
412 if (irk->o == IR_AREF || irk->o == IR_HREFK)
413 irk = &T->ir[irk->op1];
414 return (&T->ir[irk->op1] == ira);
415 }
416 return 0;
417 }
418
419 /* Check whether a sunk store corresponds to an allocation. Fast path. */
420 static LJ_AINLINE int snap_sunk_store(GCtrace *T, IRIns *ira, IRIns *irs)
421 {
422 if (irs->s != 255)
423 return (ira + irs->s == irs); /* Fast check. */
424 return snap_sunk_store2(T, ira, irs);
425 }
426
427 /* Replay snapshot state to setup side trace. */
428 void lj_snap_replay(jit_State *J, GCtrace *T)
429 {
430 SnapShot *snap = &T->snap[J->exitno];
431 SnapEntry *map = &T->snapmap[snap->mapofs];
432 MSize n, nent = snap->nent;
433 BloomFilter seen = 0;
434 int pass23 = 0;
435 J->framedepth = 0;
436 /* Emit IR for slots inherited from parent snapshot. */
437 for (n = 0; n < nent; n++) {
438 SnapEntry sn = map[n];
439 BCReg s = snap_slot(sn);
440 IRRef ref = snap_ref(sn);
441 IRIns *ir = &T->ir[ref];
442 TRef tr;
443 /* The bloom filter avoids O(nent^2) overhead for de-duping slots. */
444 if (bloomtest(seen, ref) && (tr = snap_dedup(J, map, n, ref)) != 0)
445 goto setslot;
446 bloomset(seen, ref);
447 if (irref_isk(ref)) {
448 tr = snap_replay_const(J, ir);
449 } else if (!regsp_used(ir->prev)) {
450 pass23 = 1;
451 lua_assert(s != 0);
452 tr = s;
453 } else {
454 IRType t = irt_type(ir->t);
455 uint32_t mode = IRSLOAD_INHERIT|IRSLOAD_PARENT;
456 if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM)) t = IRT_NUM;
457 if (ir->o == IR_SLOAD) mode |= (ir->op2 & IRSLOAD_READONLY);
458 tr = emitir_raw(IRT(IR_SLOAD, t), s, mode);
459 }
460 setslot:
461 J->slot[s] = tr | (sn&(SNAP_CONT|SNAP_FRAME)); /* Same as TREF_* flags. */
462 J->framedepth += ((sn & (SNAP_CONT|SNAP_FRAME)) && s);
463 if ((sn & SNAP_FRAME))
464 J->baseslot = s+1;
465 }
466 if (pass23) {
467 IRIns *irlast = &T->ir[snap->ref];
468 pass23 = 0;
469 /* Emit dependent PVALs. */
470 for (n = 0; n < nent; n++) {
471 SnapEntry sn = map[n];
472 IRRef refp = snap_ref(sn);
473 IRIns *ir = &T->ir[refp];
474 if (regsp_reg(ir->r) == RID_SUNK) {
475 if (J->slot[snap_slot(sn)] != snap_slot(sn)) continue;
476 pass23 = 1;
477 lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP ||
478 ir->o == IR_CNEW || ir->o == IR_CNEWI);
479 if (ir->op1 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op1);
480 if (ir->op2 >= T->nk) snap_pref(J, T, map, nent, seen, ir->op2);
481 if (LJ_HASFFI && ir->o == IR_CNEWI) {
482 if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP)
483 snap_pref(J, T, map, nent, seen, (ir+1)->op2);
484 } else {
485 IRIns *irs;
486 for (irs = ir+1; irs < irlast; irs++)
487 if (irs->r == RID_SINK && snap_sunk_store(T, ir, irs)) {
488 if (snap_pref(J, T, map, nent, seen, irs->op2) == 0)
489 snap_pref(J, T, map, nent, seen, T->ir[irs->op2].op1);
490 else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
491 irs+1 < irlast && (irs+1)->o == IR_HIOP)
492 snap_pref(J, T, map, nent, seen, (irs+1)->op2);
493 }
494 }
495 } else if (!irref_isk(refp) && !regsp_used(ir->prev)) {
496 lua_assert(ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
497 J->slot[snap_slot(sn)] = snap_pref(J, T, map, nent, seen, ir->op1);
498 }
499 }
500 /* Replay sunk instructions. */
501 for (n = 0; pass23 && n < nent; n++) {
502 SnapEntry sn = map[n];
503 IRRef refp = snap_ref(sn);
504 IRIns *ir = &T->ir[refp];
505 if (regsp_reg(ir->r) == RID_SUNK) {
506 TRef op1, op2;
507 if (J->slot[snap_slot(sn)] != snap_slot(sn)) { /* De-dup allocs. */
508 J->slot[snap_slot(sn)] = J->slot[J->slot[snap_slot(sn)]];
509 continue;
510 }
511 op1 = ir->op1;
512 if (op1 >= T->nk) op1 = snap_pref(J, T, map, nent, seen, op1);
513 op2 = ir->op2;
514 if (op2 >= T->nk) op2 = snap_pref(J, T, map, nent, seen, op2);
515 if (LJ_HASFFI && ir->o == IR_CNEWI) {
516 if (LJ_32 && refp+1 < T->nins && (ir+1)->o == IR_HIOP) {
517 lj_needsplit(J); /* Emit joining HIOP. */
518 op2 = emitir_raw(IRT(IR_HIOP, IRT_I64), op2,
519 snap_pref(J, T, map, nent, seen, (ir+1)->op2));
520 }
521 J->slot[snap_slot(sn)] = emitir(ir->ot & ~(IRT_MARK|IRT_ISPHI), op1, op2);
522 } else {
523 IRIns *irs;
524 TRef tr = emitir(ir->ot, op1, op2);
525 J->slot[snap_slot(sn)] = tr;
526 for (irs = ir+1; irs < irlast; irs++)
527 if (irs->r == RID_SINK && snap_sunk_store(T, ir, irs)) {
528 IRIns *irr = &T->ir[irs->op1];
529 TRef val, key = irr->op2, tmp = tr;
530 if (irr->o != IR_FREF) {
531 IRIns *irk = &T->ir[key];
532 if (irr->o == IR_HREFK)
533 key = lj_ir_kslot(J, snap_replay_const(J, &T->ir[irk->op1]),
534 irk->op2);
535 else
536 key = snap_replay_const(J, irk);
537 if (irr->o == IR_HREFK || irr->o == IR_AREF) {
538 IRIns *irf = &T->ir[irr->op1];
539 tmp = emitir(irf->ot, tmp, irf->op2);
540 }
541 }
542 tmp = emitir(irr->ot, tmp, key);
543 val = snap_pref(J, T, map, nent, seen, irs->op2);
544 if (val == 0) {
545 IRIns *irc = &T->ir[irs->op2];
546 lua_assert(irc->o == IR_CONV && irc->op2 == IRCONV_NUM_INT);
547 val = snap_pref(J, T, map, nent, seen, irc->op1);
548 val = emitir(IRTN(IR_CONV), val, IRCONV_NUM_INT);
549 } else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
550 irs+1 < irlast && (irs+1)->o == IR_HIOP) {
551 IRType t = IRT_I64;
552 if (LJ_SOFTFP && irt_type((irs+1)->t) == IRT_SOFTFP)
553 t = IRT_NUM;
554 lj_needsplit(J);
555 if (irref_isk(irs->op2) && irref_isk((irs+1)->op2)) {
556 uint64_t k = (uint32_t)T->ir[irs->op2].i +
557 ((uint64_t)T->ir[(irs+1)->op2].i << 32);
558 val = lj_ir_k64(J, t == IRT_I64 ? IR_KINT64 : IR_KNUM,
559 lj_ir_k64_find(J, k));
560 } else {
561 val = emitir_raw(IRT(IR_HIOP, t), val,
562 snap_pref(J, T, map, nent, seen, (irs+1)->op2));
563 }
564 tmp = emitir(IRT(irs->o, t), tmp, val);
565 continue;
566 }
567 tmp = emitir(irs->ot, tmp, val);
568 } else if (LJ_HASFFI && irs->o == IR_XBAR && ir->o == IR_CNEW) {
569 emitir(IRT(IR_XBAR, IRT_NIL), 0, 0);
570 }
571 }
572 }
573 }
574 }
575 J->base = J->slot + J->baseslot;
576 J->maxslot = snap->nslots - J->baseslot;
577 lj_snap_add(J);
578 if (pass23) /* Need explicit GC step _after_ initial snapshot. */
579 emitir_raw(IRTG(IR_GCSTEP, IRT_NIL), 0, 0);
580 }
581
582 /* -- Snapshot restore ---------------------------------------------------- */
583
584 static void snap_unsink(jit_State *J, GCtrace *T, ExitState *ex,
585 SnapNo snapno, BloomFilter rfilt,
586 IRIns *ir, TValue *o);
587
588 /* Restore a value from the trace exit state. */
589 static void snap_restoreval(jit_State *J, GCtrace *T, ExitState *ex,
590 SnapNo snapno, BloomFilter rfilt,
591 IRRef ref, TValue *o)
592 {
593 IRIns *ir = &T->ir[ref];
594 IRType1 t = ir->t;
595 RegSP rs = ir->prev;
596 if (irref_isk(ref)) { /* Restore constant slot. */
597 lj_ir_kvalue(J->L, o, ir);
598 return;
599 }
600 if (LJ_UNLIKELY(bloomtest(rfilt, ref)))
601 rs = snap_renameref(T, snapno, ref, rs);
602 lua_assert(!LJ_GC64); /* TODO_GC64: handle 64 bit references. */
603 if (ra_hasspill(regsp_spill(rs))) { /* Restore from spill slot. */
604 int32_t *sps = &ex->spill[regsp_spill(rs)];
605 if (irt_isinteger(t)) {
606 setintV(o, *sps);
607 #if !LJ_SOFTFP
608 } else if (irt_isnum(t)) {
609 o->u64 = *(uint64_t *)sps;
610 #endif
611 } else if (LJ_64 && irt_islightud(t)) {
612 /* 64 bit lightuserdata which may escape already has the tag bits. */
613 o->u64 = *(uint64_t *)sps;
614 } else {
615 lua_assert(!irt_ispri(t)); /* PRI refs never have a spill slot. */
616 setgcV(J->L, o, (GCobj *)(uintptr_t)*(GCSize *)sps, irt_toitype(t));
617 }
618 } else { /* Restore from register. */
619 Reg r = regsp_reg(rs);
620 if (ra_noreg(r)) {
621 lua_assert(ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
622 snap_restoreval(J, T, ex, snapno, rfilt, ir->op1, o);
623 if (LJ_DUALNUM) setnumV(o, (lua_Number)intV(o));
624 return;
625 } else if (irt_isinteger(t)) {
626 setintV(o, (int32_t)ex->gpr[r-RID_MIN_GPR]);
627 #if !LJ_SOFTFP
628 } else if (irt_isnum(t)) {
629 setnumV(o, ex->fpr[r-RID_MIN_FPR]);
630 #endif
631 } else if (LJ_64 && irt_is64(t)) {
632 /* 64 bit values that already have the tag bits. */
633 o->u64 = ex->gpr[r-RID_MIN_GPR];
634 } else if (irt_ispri(t)) {
635 setpriV(o, irt_toitype(t));
636 } else {
637 setgcV(J->L, o, (GCobj *)ex->gpr[r-RID_MIN_GPR], irt_toitype(t));
638 }
639 }
640 }
641
642 #if LJ_HASFFI
643 /* Restore raw data from the trace exit state. */
644 static void snap_restoredata(GCtrace *T, ExitState *ex,
645 SnapNo snapno, BloomFilter rfilt,
646 IRRef ref, void *dst, CTSize sz)
647 {
648 IRIns *ir = &T->ir[ref];
649 RegSP rs = ir->prev;
650 int32_t *src;
651 uint64_t tmp;
652 if (irref_isk(ref)) {
653 if (ir->o == IR_KNUM || ir->o == IR_KINT64) {
654 src = mref(ir->ptr, int32_t);
655 } else if (sz == 8) {
656 tmp = (uint64_t)(uint32_t)ir->i;
657 src = (int32_t *)&tmp;
658 } else {
659 src = &ir->i;
660 }
661 } else {
662 if (LJ_UNLIKELY(bloomtest(rfilt, ref)))
663 rs = snap_renameref(T, snapno, ref, rs);
664 if (ra_hasspill(regsp_spill(rs))) {
665 src = &ex->spill[regsp_spill(rs)];
666 if (sz == 8 && !irt_is64(ir->t)) {
667 tmp = (uint64_t)(uint32_t)*src;
668 src = (int32_t *)&tmp;
669 }
670 } else {
671 Reg r = regsp_reg(rs);
672 if (ra_noreg(r)) {
673 /* Note: this assumes CNEWI is never used for SOFTFP split numbers. */
674 lua_assert(sz == 8 && ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT);
675 snap_restoredata(T, ex, snapno, rfilt, ir->op1, dst, 4);
676 *(lua_Number *)dst = (lua_Number)*(int32_t *)dst;
677 return;
678 }
679 src = (int32_t *)&ex->gpr[r-RID_MIN_GPR];
680 #if !LJ_SOFTFP
681 if (r >= RID_MAX_GPR) {
682 src = (int32_t *)&ex->fpr[r-RID_MIN_FPR];
683 #if LJ_TARGET_PPC
684 if (sz == 4) { /* PPC FPRs are always doubles. */
685 *(float *)dst = (float)*(double *)src;
686 return;
687 }
688 #else
689 if (LJ_BE && sz == 4) src++;
690 #endif
691 }
692 #endif
693 }
694 }
695 lua_assert(sz == 1 || sz == 2 || sz == 4 || sz == 8);
696 if (sz == 4) *(int32_t *)dst = *src;
697 else if (sz == 8) *(int64_t *)dst = *(int64_t *)src;
698 else if (sz == 1) *(int8_t *)dst = (int8_t)*src;
699 else *(int16_t *)dst = (int16_t)*src;
700 }
701 #endif
702
703 /* Unsink allocation from the trace exit state. Unsink sunk stores. */
704 static void snap_unsink(jit_State *J, GCtrace *T, ExitState *ex,
705 SnapNo snapno, BloomFilter rfilt,
706 IRIns *ir, TValue *o)
707 {
708 lua_assert(ir->o == IR_TNEW || ir->o == IR_TDUP ||
709 ir->o == IR_CNEW || ir->o == IR_CNEWI);
710 #if LJ_HASFFI
711 if (ir->o == IR_CNEW || ir->o == IR_CNEWI) {
712 CTState *cts = ctype_cts(J->L);
713 CTypeID id = (CTypeID)T->ir[ir->op1].i;
714 CTSize sz;
715 CTInfo info = lj_ctype_info(cts, id, &sz);
716 GCcdata *cd = lj_cdata_newx(cts, id, sz, info);
717 setcdataV(J->L, o, cd);
718 if (ir->o == IR_CNEWI) {
719 uint8_t *p = (uint8_t *)cdataptr(cd);
720 lua_assert(sz == 4 || sz == 8);
721 if (LJ_32 && sz == 8 && ir+1 < T->ir + T->nins && (ir+1)->o == IR_HIOP) {
722 snap_restoredata(T, ex, snapno, rfilt, (ir+1)->op2, LJ_LE?p+4:p, 4);
723 if (LJ_BE) p += 4;
724 sz = 4;
725 }
726 snap_restoredata(T, ex, snapno, rfilt, ir->op2, p, sz);
727 } else {
728 IRIns *irs, *irlast = &T->ir[T->snap[snapno].ref];
729 for (irs = ir+1; irs < irlast; irs++)
730 if (irs->r == RID_SINK && snap_sunk_store(T, ir, irs)) {
731 IRIns *iro = &T->ir[T->ir[irs->op1].op2];
732 uint8_t *p = (uint8_t *)cd;
733 CTSize szs;
734 lua_assert(irs->o == IR_XSTORE && T->ir[irs->op1].o == IR_ADD);
735 lua_assert(iro->o == IR_KINT || iro->o == IR_KINT64);
736 if (irt_is64(irs->t)) szs = 8;
737 else if (irt_isi8(irs->t) || irt_isu8(irs->t)) szs = 1;
738 else if (irt_isi16(irs->t) || irt_isu16(irs->t)) szs = 2;
739 else szs = 4;
740 if (LJ_64 && iro->o == IR_KINT64)
741 p += (int64_t)ir_k64(iro)->u64;
742 else
743 p += iro->i;
744 lua_assert(p >= (uint8_t *)cdataptr(cd) &&
745 p + szs <= (uint8_t *)cdataptr(cd) + sz);
746 if (LJ_32 && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
747 lua_assert(szs == 4);
748 snap_restoredata(T, ex, snapno, rfilt, (irs+1)->op2, LJ_LE?p+4:p,4);
749 if (LJ_BE) p += 4;
750 }
751 snap_restoredata(T, ex, snapno, rfilt, irs->op2, p, szs);
752 }
753 }
754 } else
755 #endif
756 {
757 IRIns *irs, *irlast;
758 GCtab *t = ir->o == IR_TNEW ? lj_tab_new(J->L, ir->op1, ir->op2) :
759 lj_tab_dup(J->L, ir_ktab(&T->ir[ir->op1]));
760 settabV(J->L, o, t);
761 irlast = &T->ir[T->snap[snapno].ref];
762 for (irs = ir+1; irs < irlast; irs++)
763 if (irs->r == RID_SINK && snap_sunk_store(T, ir, irs)) {
764 IRIns *irk = &T->ir[irs->op1];
765 TValue tmp, *val;
766 lua_assert(irs->o == IR_ASTORE || irs->o == IR_HSTORE ||
767 irs->o == IR_FSTORE);
768 if (irk->o == IR_FREF) {
769 lua_assert(irk->op2 == IRFL_TAB_META);
770 snap_restoreval(J, T, ex, snapno, rfilt, irs->op2, &tmp);
771 /* NOBARRIER: The table is new (marked white). */
772 setgcref(t->metatable, obj2gco(tabV(&tmp)));
773 } else {
774 irk = &T->ir[irk->op2];
775 if (irk->o == IR_KSLOT) irk = &T->ir[irk->op1];
776 lj_ir_kvalue(J->L, &tmp, irk);
777 val = lj_tab_set(J->L, t, &tmp);
778 /* NOBARRIER: The table is new (marked white). */
779 snap_restoreval(J, T, ex, snapno, rfilt, irs->op2, val);
780 if (LJ_SOFTFP && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
781 snap_restoreval(J, T, ex, snapno, rfilt, (irs+1)->op2, &tmp);
782 val->u32.hi = tmp.u32.lo;
783 }
784 }
785 }
786 }
787 }
788
789 /* Restore interpreter state from exit state with the help of a snapshot. */
790 const BCIns *lj_snap_restore(jit_State *J, void *exptr)
791 {
792 ExitState *ex = (ExitState *)exptr;
793 SnapNo snapno = J->exitno; /* For now, snapno == exitno. */
794 GCtrace *T = traceref(J, J->parent);
795 SnapShot *snap = &T->snap[snapno];
796 MSize n, nent = snap->nent;
797 SnapEntry *map = &T->snapmap[snap->mapofs];
798 SnapEntry *flinks = &T->snapmap[snap_nextofs(T, snap)-1];
799 ptrdiff_t ftsz0;
800 TValue *frame;
801 BloomFilter rfilt = snap_renamefilter(T, snapno);
802 const BCIns *pc = snap_pc(map[nent]);
803 lua_State *L = J->L;
804
805 /* Set interpreter PC to the next PC to get correct error messages. */
806 setcframe_pc(cframe_raw(L->cframe), pc+1);
807
808 /* Make sure the stack is big enough for the slots from the snapshot. */
809 if (LJ_UNLIKELY(L->base + snap->topslot >= tvref(L->maxstack))) {
810 L->top = curr_topL(L);
811 lj_state_growstack(L, snap->topslot - curr_proto(L)->framesize);
812 }
813
814 /* Fill stack slots with data from the registers and spill slots. */
815 frame = L->base-1;
816 ftsz0 = frame_ftsz(frame); /* Preserve link to previous frame in slot #0. */
817 for (n = 0; n < nent; n++) {
818 SnapEntry sn = map[n];
819 if (!(sn & SNAP_NORESTORE)) {
820 TValue *o = &frame[snap_slot(sn)];
821 IRRef ref = snap_ref(sn);
822 IRIns *ir = &T->ir[ref];
823 if (ir->r == RID_SUNK) {
824 MSize j;
825 for (j = 0; j < n; j++)
826 if (snap_ref(map[j]) == ref) { /* De-duplicate sunk allocations. */
827 copyTV(L, o, &frame[snap_slot(map[j])]);
828 goto dupslot;
829 }
830 snap_unsink(J, T, ex, snapno, rfilt, ir, o);
831 dupslot:
832 continue;
833 }
834 snap_restoreval(J, T, ex, snapno, rfilt, ref, o);
835 if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM) && tvisint(o)) {
836 TValue tmp;
837 snap_restoreval(J, T, ex, snapno, rfilt, ref+1, &tmp);
838 o->u32.hi = tmp.u32.lo;
839 } else if ((sn & (SNAP_CONT|SNAP_FRAME))) {
840 lua_assert(!LJ_FR2); /* TODO_FR2: store 64 bit PCs. */
841 /* Overwrite tag with frame link. */
842 setframe_ftsz(o, snap_slot(sn) != 0 ? (int32_t)*flinks-- : ftsz0);
843 L->base = o+1;
844 }
845 }
846 }
847 lua_assert(map + nent == flinks);
848
849 /* Compute current stack top. */
850 switch (bc_op(*pc)) {
851 default:
852 if (bc_op(*pc) < BC_FUNCF) {
853 L->top = curr_topL(L);
854 break;
855 }
856 /* fallthrough */
857 case BC_CALLM: case BC_CALLMT: case BC_RETM: case BC_TSETM:
858 L->top = frame + snap->nslots;
859 break;
860 }
861 return pc;
862 }
863
864 #undef emitir_raw
865 #undef emitir
866
867 #endif
LuaTeX git mirror