xref: /freebsd/contrib/lua/src/ltable.c (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
1 /*
2 ** $Id: ltable.c $
3 ** Lua tables (hash)
4 ** See Copyright Notice in lua.h
5 */
6 
7 #define ltable_c
8 #define LUA_CORE
9 
10 #include "lprefix.h"
11 
12 
13 /*
14 ** Implementation of tables (aka arrays, objects, or hash tables).
15 ** Tables keep its elements in two parts: an array part and a hash part.
16 ** Non-negative integer keys are all candidates to be kept in the array
17 ** part. The actual size of the array is the largest 'n' such that
18 ** more than half the slots between 1 and n are in use.
19 ** Hash uses a mix of chained scatter table with Brent's variation.
20 ** A main invariant of these tables is that, if an element is not
21 ** in its main position (i.e. the 'original' position that its hash gives
22 ** to it), then the colliding element is in its own main position.
23 ** Hence even when the load factor reaches 100%, performance remains good.
24 */
25 
26 #include <math.h>
27 #include <limits.h>
28 
29 #include "lua.h"
30 
31 #include "ldebug.h"
32 #include "ldo.h"
33 #include "lgc.h"
34 #include "lmem.h"
35 #include "lobject.h"
36 #include "lstate.h"
37 #include "lstring.h"
38 #include "ltable.h"
39 #include "lvm.h"
40 
41 
42 /*
43 ** MAXABITS is the largest integer such that MAXASIZE fits in an
44 ** unsigned int.
45 */
46 #define MAXABITS	cast_int(sizeof(int) * CHAR_BIT - 1)
47 
48 
49 /*
50 ** MAXASIZE is the maximum size of the array part. It is the minimum
51 ** between 2^MAXABITS and the maximum size that, measured in bytes,
52 ** fits in a 'size_t'.
53 */
54 #define MAXASIZE	luaM_limitN(1u << MAXABITS, TValue)
55 
56 /*
57 ** MAXHBITS is the largest integer such that 2^MAXHBITS fits in a
58 ** signed int.
59 */
60 #define MAXHBITS	(MAXABITS - 1)
61 
62 
63 /*
64 ** MAXHSIZE is the maximum size of the hash part. It is the minimum
65 ** between 2^MAXHBITS and the maximum size such that, measured in bytes,
66 ** it fits in a 'size_t'.
67 */
68 #define MAXHSIZE	luaM_limitN(1u << MAXHBITS, Node)
69 
70 
71 #define hashpow2(t,n)		(gnode(t, lmod((n), sizenode(t))))
72 
73 #define hashstr(t,str)		hashpow2(t, (str)->hash)
74 #define hashboolean(t,p)	hashpow2(t, p)
75 #define hashint(t,i)		hashpow2(t, i)
76 
77 
78 /*
79 ** for some types, it is better to avoid modulus by power of 2, as
80 ** they tend to have many 2 factors.
81 */
82 #define hashmod(t,n)	(gnode(t, ((n) % ((sizenode(t)-1)|1))))
83 
84 
85 #define hashpointer(t,p)	hashmod(t, point2uint(p))
86 
87 
88 #define dummynode		(&dummynode_)
89 
90 static const Node dummynode_ = {
91   {{NULL}, LUA_VEMPTY,  /* value's value and type */
92    LUA_VNIL, 0, {NULL}}  /* key type, next, and key value */
93 };
94 
95 
96 static const TValue absentkey = {ABSTKEYCONSTANT};
97 
98 
99 
100 /*
101 ** Hash for floating-point numbers.
102 ** The main computation should be just
103 **     n = frexp(n, &i); return (n * INT_MAX) + i
104 ** but there are some numerical subtleties.
105 ** In a two-complement representation, INT_MAX does not has an exact
106 ** representation as a float, but INT_MIN does; because the absolute
107 ** value of 'frexp' is smaller than 1 (unless 'n' is inf/NaN), the
108 ** absolute value of the product 'frexp * -INT_MIN' is smaller or equal
109 ** to INT_MAX. Next, the use of 'unsigned int' avoids overflows when
110 ** adding 'i'; the use of '~u' (instead of '-u') avoids problems with
111 ** INT_MIN.
112 */
113 #if !defined(l_hashfloat)
114 static int l_hashfloat (lua_Number n) {
115   int i;
116   lua_Integer ni;
117   n = l_mathop(frexp)(n, &i) * -cast_num(INT_MIN);
118   if (!lua_numbertointeger(n, &ni)) {  /* is 'n' inf/-inf/NaN? */
119     lua_assert(luai_numisnan(n) || l_mathop(fabs)(n) == cast_num(HUGE_VAL));
120     return 0;
121   }
122   else {  /* normal case */
123     unsigned int u = cast_uint(i) + cast_uint(ni);
124     return cast_int(u <= cast_uint(INT_MAX) ? u : ~u);
125   }
126 }
127 #endif
128 
129 
130 /*
131 ** returns the 'main' position of an element in a table (that is,
132 ** the index of its hash value). The key comes broken (tag in 'ktt'
133 ** and value in 'vkl') so that we can call it on keys inserted into
134 ** nodes.
135 */
136 static Node *mainposition (const Table *t, int ktt, const Value *kvl) {
137   switch (withvariant(ktt)) {
138     case LUA_VNUMINT:
139       return hashint(t, ivalueraw(*kvl));
140     case LUA_VNUMFLT:
141       return hashmod(t, l_hashfloat(fltvalueraw(*kvl)));
142     case LUA_VSHRSTR:
143       return hashstr(t, tsvalueraw(*kvl));
144     case LUA_VLNGSTR:
145       return hashpow2(t, luaS_hashlongstr(tsvalueraw(*kvl)));
146     case LUA_VFALSE:
147       return hashboolean(t, 0);
148     case LUA_VTRUE:
149       return hashboolean(t, 1);
150     case LUA_VLIGHTUSERDATA:
151       return hashpointer(t, pvalueraw(*kvl));
152     case LUA_VLCF:
153       return hashpointer(t, fvalueraw(*kvl));
154     default:
155       return hashpointer(t, gcvalueraw(*kvl));
156   }
157 }
158 
159 
160 /*
161 ** Returns the main position of an element given as a 'TValue'
162 */
163 static Node *mainpositionTV (const Table *t, const TValue *key) {
164   return mainposition(t, rawtt(key), valraw(key));
165 }
166 
167 
168 /*
169 ** Check whether key 'k1' is equal to the key in node 'n2'. This
170 ** equality is raw, so there are no metamethods. Floats with integer
171 ** values have been normalized, so integers cannot be equal to
172 ** floats. It is assumed that 'eqshrstr' is simply pointer equality, so
173 ** that short strings are handled in the default case.
174 ** A true 'deadok' means to accept dead keys as equal to their original
175 ** values. All dead keys are compared in the default case, by pointer
176 ** identity. (Only collectable objects can produce dead keys.) Note that
177 ** dead long strings are also compared by identity.
178 ** Once a key is dead, its corresponding value may be collected, and
179 ** then another value can be created with the same address. If this
180 ** other value is given to 'next', 'equalkey' will signal a false
181 ** positive. In a regular traversal, this situation should never happen,
182 ** as all keys given to 'next' came from the table itself, and therefore
183 ** could not have been collected. Outside a regular traversal, we
184 ** have garbage in, garbage out. What is relevant is that this false
185 ** positive does not break anything.  (In particular, 'next' will return
186 ** some other valid item on the table or nil.)
187 */
188 static int equalkey (const TValue *k1, const Node *n2, int deadok) {
189   if ((rawtt(k1) != keytt(n2)) &&  /* not the same variants? */
190        !(deadok && keyisdead(n2) && iscollectable(k1)))
191    return 0;  /* cannot be same key */
192   switch (keytt(n2)) {
193     case LUA_VNIL: case LUA_VFALSE: case LUA_VTRUE:
194       return 1;
195     case LUA_VNUMINT:
196       return (ivalue(k1) == keyival(n2));
197     case LUA_VNUMFLT:
198       return luai_numeq(fltvalue(k1), fltvalueraw(keyval(n2)));
199     case LUA_VLIGHTUSERDATA:
200       return pvalue(k1) == pvalueraw(keyval(n2));
201     case LUA_VLCF:
202       return fvalue(k1) == fvalueraw(keyval(n2));
203     case ctb(LUA_VLNGSTR):
204       return luaS_eqlngstr(tsvalue(k1), keystrval(n2));
205     default:
206       return gcvalue(k1) == gcvalueraw(keyval(n2));
207   }
208 }
209 
210 
211 /*
212 ** True if value of 'alimit' is equal to the real size of the array
213 ** part of table 't'. (Otherwise, the array part must be larger than
214 ** 'alimit'.)
215 */
216 #define limitequalsasize(t)	(isrealasize(t) || ispow2((t)->alimit))
217 
218 
219 /*
220 ** Returns the real size of the 'array' array
221 */
222 LUAI_FUNC unsigned int luaH_realasize (const Table *t) {
223   if (limitequalsasize(t))
224     return t->alimit;  /* this is the size */
225   else {
226     unsigned int size = t->alimit;
227     /* compute the smallest power of 2 not smaller than 'n' */
228     size |= (size >> 1);
229     size |= (size >> 2);
230     size |= (size >> 4);
231     size |= (size >> 8);
232     size |= (size >> 16);
233 #if (UINT_MAX >> 30) > 3
234     size |= (size >> 32);  /* unsigned int has more than 32 bits */
235 #endif
236     size++;
237     lua_assert(ispow2(size) && size/2 < t->alimit && t->alimit < size);
238     return size;
239   }
240 }
241 
242 
243 /*
244 ** Check whether real size of the array is a power of 2.
245 ** (If it is not, 'alimit' cannot be changed to any other value
246 ** without changing the real size.)
247 */
248 static int ispow2realasize (const Table *t) {
249   return (!isrealasize(t) || ispow2(t->alimit));
250 }
251 
252 
253 static unsigned int setlimittosize (Table *t) {
254   t->alimit = luaH_realasize(t);
255   setrealasize(t);
256   return t->alimit;
257 }
258 
259 
260 #define limitasasize(t)	check_exp(isrealasize(t), t->alimit)
261 
262 
263 
264 /*
265 ** "Generic" get version. (Not that generic: not valid for integers,
266 ** which may be in array part, nor for floats with integral values.)
267 ** See explanation about 'deadok' in function 'equalkey'.
268 */
269 static const TValue *getgeneric (Table *t, const TValue *key, int deadok) {
270   Node *n = mainpositionTV(t, key);
271   for (;;) {  /* check whether 'key' is somewhere in the chain */
272     if (equalkey(key, n, deadok))
273       return gval(n);  /* that's it */
274     else {
275       int nx = gnext(n);
276       if (nx == 0)
277         return &absentkey;  /* not found */
278       n += nx;
279     }
280   }
281 }
282 
283 
284 /*
285 ** returns the index for 'k' if 'k' is an appropriate key to live in
286 ** the array part of a table, 0 otherwise.
287 */
288 static unsigned int arrayindex (lua_Integer k) {
289   if (l_castS2U(k) - 1u < MAXASIZE)  /* 'k' in [1, MAXASIZE]? */
290     return cast_uint(k);  /* 'key' is an appropriate array index */
291   else
292     return 0;
293 }
294 
295 
296 /*
297 ** returns the index of a 'key' for table traversals. First goes all
298 ** elements in the array part, then elements in the hash part. The
299 ** beginning of a traversal is signaled by 0.
300 */
301 static unsigned int findindex (lua_State *L, Table *t, TValue *key,
302                                unsigned int asize) {
303   unsigned int i;
304   if (ttisnil(key)) return 0;  /* first iteration */
305   i = ttisinteger(key) ? arrayindex(ivalue(key)) : 0;
306   if (i - 1u < asize)  /* is 'key' inside array part? */
307     return i;  /* yes; that's the index */
308   else {
309     const TValue *n = getgeneric(t, key, 1);
310     if (unlikely(isabstkey(n)))
311       luaG_runerror(L, "invalid key to 'next'");  /* key not found */
312     i = cast_int(nodefromval(n) - gnode(t, 0));  /* key index in hash table */
313     /* hash elements are numbered after array ones */
314     return (i + 1) + asize;
315   }
316 }
317 
318 
319 int luaH_next (lua_State *L, Table *t, StkId key) {
320   unsigned int asize = luaH_realasize(t);
321   unsigned int i = findindex(L, t, s2v(key), asize);  /* find original key */
322   for (; i < asize; i++) {  /* try first array part */
323     if (!isempty(&t->array[i])) {  /* a non-empty entry? */
324       setivalue(s2v(key), i + 1);
325       setobj2s(L, key + 1, &t->array[i]);
326       return 1;
327     }
328   }
329   for (i -= asize; cast_int(i) < sizenode(t); i++) {  /* hash part */
330     if (!isempty(gval(gnode(t, i)))) {  /* a non-empty entry? */
331       Node *n = gnode(t, i);
332       getnodekey(L, s2v(key), n);
333       setobj2s(L, key + 1, gval(n));
334       return 1;
335     }
336   }
337   return 0;  /* no more elements */
338 }
339 
340 
341 static void freehash (lua_State *L, Table *t) {
342   if (!isdummy(t))
343     luaM_freearray(L, t->node, cast_sizet(sizenode(t)));
344 }
345 
346 
347 /*
348 ** {=============================================================
349 ** Rehash
350 ** ==============================================================
351 */
352 
353 /*
354 ** Compute the optimal size for the array part of table 't'. 'nums' is a
355 ** "count array" where 'nums[i]' is the number of integers in the table
356 ** between 2^(i - 1) + 1 and 2^i. 'pna' enters with the total number of
357 ** integer keys in the table and leaves with the number of keys that
358 ** will go to the array part; return the optimal size.  (The condition
359 ** 'twotoi > 0' in the for loop stops the loop if 'twotoi' overflows.)
360 */
361 static unsigned int computesizes (unsigned int nums[], unsigned int *pna) {
362   int i;
363   unsigned int twotoi;  /* 2^i (candidate for optimal size) */
364   unsigned int a = 0;  /* number of elements smaller than 2^i */
365   unsigned int na = 0;  /* number of elements to go to array part */
366   unsigned int optimal = 0;  /* optimal size for array part */
367   /* loop while keys can fill more than half of total size */
368   for (i = 0, twotoi = 1;
369        twotoi > 0 && *pna > twotoi / 2;
370        i++, twotoi *= 2) {
371     a += nums[i];
372     if (a > twotoi/2) {  /* more than half elements present? */
373       optimal = twotoi;  /* optimal size (till now) */
374       na = a;  /* all elements up to 'optimal' will go to array part */
375     }
376   }
377   lua_assert((optimal == 0 || optimal / 2 < na) && na <= optimal);
378   *pna = na;
379   return optimal;
380 }
381 
382 
383 static int countint (lua_Integer key, unsigned int *nums) {
384   unsigned int k = arrayindex(key);
385   if (k != 0) {  /* is 'key' an appropriate array index? */
386     nums[luaO_ceillog2(k)]++;  /* count as such */
387     return 1;
388   }
389   else
390     return 0;
391 }
392 
393 
394 /*
395 ** Count keys in array part of table 't': Fill 'nums[i]' with
396 ** number of keys that will go into corresponding slice and return
397 ** total number of non-nil keys.
398 */
399 static unsigned int numusearray (const Table *t, unsigned int *nums) {
400   int lg;
401   unsigned int ttlg;  /* 2^lg */
402   unsigned int ause = 0;  /* summation of 'nums' */
403   unsigned int i = 1;  /* count to traverse all array keys */
404   unsigned int asize = limitasasize(t);  /* real array size */
405   /* traverse each slice */
406   for (lg = 0, ttlg = 1; lg <= MAXABITS; lg++, ttlg *= 2) {
407     unsigned int lc = 0;  /* counter */
408     unsigned int lim = ttlg;
409     if (lim > asize) {
410       lim = asize;  /* adjust upper limit */
411       if (i > lim)
412         break;  /* no more elements to count */
413     }
414     /* count elements in range (2^(lg - 1), 2^lg] */
415     for (; i <= lim; i++) {
416       if (!isempty(&t->array[i-1]))
417         lc++;
418     }
419     nums[lg] += lc;
420     ause += lc;
421   }
422   return ause;
423 }
424 
425 
426 static int numusehash (const Table *t, unsigned int *nums, unsigned int *pna) {
427   int totaluse = 0;  /* total number of elements */
428   int ause = 0;  /* elements added to 'nums' (can go to array part) */
429   int i = sizenode(t);
430   while (i--) {
431     Node *n = &t->node[i];
432     if (!isempty(gval(n))) {
433       if (keyisinteger(n))
434         ause += countint(keyival(n), nums);
435       totaluse++;
436     }
437   }
438   *pna += ause;
439   return totaluse;
440 }
441 
442 
443 /*
444 ** Creates an array for the hash part of a table with the given
445 ** size, or reuses the dummy node if size is zero.
446 ** The computation for size overflow is in two steps: the first
447 ** comparison ensures that the shift in the second one does not
448 ** overflow.
449 */
450 static void setnodevector (lua_State *L, Table *t, unsigned int size) {
451   if (size == 0) {  /* no elements to hash part? */
452     t->node = cast(Node *, dummynode);  /* use common 'dummynode' */
453     t->lsizenode = 0;
454     t->lastfree = NULL;  /* signal that it is using dummy node */
455   }
456   else {
457     int i;
458     int lsize = luaO_ceillog2(size);
459     if (lsize > MAXHBITS || (1u << lsize) > MAXHSIZE)
460       luaG_runerror(L, "table overflow");
461     size = twoto(lsize);
462     t->node = luaM_newvector(L, size, Node);
463     for (i = 0; i < (int)size; i++) {
464       Node *n = gnode(t, i);
465       gnext(n) = 0;
466       setnilkey(n);
467       setempty(gval(n));
468     }
469     t->lsizenode = cast_byte(lsize);
470     t->lastfree = gnode(t, size);  /* all positions are free */
471   }
472 }
473 
474 
475 /*
476 ** (Re)insert all elements from the hash part of 'ot' into table 't'.
477 */
478 static void reinsert (lua_State *L, Table *ot, Table *t) {
479   int j;
480   int size = sizenode(ot);
481   for (j = 0; j < size; j++) {
482     Node *old = gnode(ot, j);
483     if (!isempty(gval(old))) {
484       /* doesn't need barrier/invalidate cache, as entry was
485          already present in the table */
486       TValue k;
487       getnodekey(L, &k, old);
488       setobjt2t(L, luaH_set(L, t, &k), gval(old));
489     }
490   }
491 }
492 
493 
494 /*
495 ** Exchange the hash part of 't1' and 't2'.
496 */
497 static void exchangehashpart (Table *t1, Table *t2) {
498   lu_byte lsizenode = t1->lsizenode;
499   Node *node = t1->node;
500   Node *lastfree = t1->lastfree;
501   t1->lsizenode = t2->lsizenode;
502   t1->node = t2->node;
503   t1->lastfree = t2->lastfree;
504   t2->lsizenode = lsizenode;
505   t2->node = node;
506   t2->lastfree = lastfree;
507 }
508 
509 
510 /*
511 ** Resize table 't' for the new given sizes. Both allocations (for
512 ** the hash part and for the array part) can fail, which creates some
513 ** subtleties. If the first allocation, for the hash part, fails, an
514 ** error is raised and that is it. Otherwise, it copies the elements from
515 ** the shrinking part of the array (if it is shrinking) into the new
516 ** hash. Then it reallocates the array part.  If that fails, the table
517 ** is in its original state; the function frees the new hash part and then
518 ** raises the allocation error. Otherwise, it sets the new hash part
519 ** into the table, initializes the new part of the array (if any) with
520 ** nils and reinserts the elements of the old hash back into the new
521 ** parts of the table.
522 */
523 void luaH_resize (lua_State *L, Table *t, unsigned int newasize,
524                                           unsigned int nhsize) {
525   unsigned int i;
526   Table newt;  /* to keep the new hash part */
527   unsigned int oldasize = setlimittosize(t);
528   TValue *newarray;
529   /* create new hash part with appropriate size into 'newt' */
530   setnodevector(L, &newt, nhsize);
531   if (newasize < oldasize) {  /* will array shrink? */
532     t->alimit = newasize;  /* pretend array has new size... */
533     exchangehashpart(t, &newt);  /* and new hash */
534     /* re-insert into the new hash the elements from vanishing slice */
535     for (i = newasize; i < oldasize; i++) {
536       if (!isempty(&t->array[i]))
537         luaH_setint(L, t, i + 1, &t->array[i]);
538     }
539     t->alimit = oldasize;  /* restore current size... */
540     exchangehashpart(t, &newt);  /* and hash (in case of errors) */
541   }
542   /* allocate new array */
543   newarray = luaM_reallocvector(L, t->array, oldasize, newasize, TValue);
544   if (unlikely(newarray == NULL && newasize > 0)) {  /* allocation failed? */
545     freehash(L, &newt);  /* release new hash part */
546     luaM_error(L);  /* raise error (with array unchanged) */
547   }
548   /* allocation ok; initialize new part of the array */
549   exchangehashpart(t, &newt);  /* 't' has the new hash ('newt' has the old) */
550   t->array = newarray;  /* set new array part */
551   t->alimit = newasize;
552   for (i = oldasize; i < newasize; i++)  /* clear new slice of the array */
553      setempty(&t->array[i]);
554   /* re-insert elements from old hash part into new parts */
555   reinsert(L, &newt, t);  /* 'newt' now has the old hash */
556   freehash(L, &newt);  /* free old hash part */
557 }
558 
559 
560 void luaH_resizearray (lua_State *L, Table *t, unsigned int nasize) {
561   int nsize = allocsizenode(t);
562   luaH_resize(L, t, nasize, nsize);
563 }
564 
565 /*
566 ** nums[i] = number of keys 'k' where 2^(i - 1) < k <= 2^i
567 */
568 static void rehash (lua_State *L, Table *t, const TValue *ek) {
569   unsigned int asize;  /* optimal size for array part */
570   unsigned int na;  /* number of keys in the array part */
571   unsigned int nums[MAXABITS + 1];
572   int i;
573   int totaluse;
574   for (i = 0; i <= MAXABITS; i++) nums[i] = 0;  /* reset counts */
575   setlimittosize(t);
576   na = numusearray(t, nums);  /* count keys in array part */
577   totaluse = na;  /* all those keys are integer keys */
578   totaluse += numusehash(t, nums, &na);  /* count keys in hash part */
579   /* count extra key */
580   if (ttisinteger(ek))
581     na += countint(ivalue(ek), nums);
582   totaluse++;
583   /* compute new size for array part */
584   asize = computesizes(nums, &na);
585   /* resize the table to new computed sizes */
586   luaH_resize(L, t, asize, totaluse - na);
587 }
588 
589 
590 
591 /*
592 ** }=============================================================
593 */
594 
595 
596 Table *luaH_new (lua_State *L) {
597   GCObject *o = luaC_newobj(L, LUA_VTABLE, sizeof(Table));
598   Table *t = gco2t(o);
599   t->metatable = NULL;
600   t->flags = cast_byte(maskflags);  /* table has no metamethod fields */
601   t->array = NULL;
602   t->alimit = 0;
603   setnodevector(L, t, 0);
604   return t;
605 }
606 
607 
608 void luaH_free (lua_State *L, Table *t) {
609   freehash(L, t);
610   luaM_freearray(L, t->array, luaH_realasize(t));
611   luaM_free(L, t);
612 }
613 
614 
615 static Node *getfreepos (Table *t) {
616   if (!isdummy(t)) {
617     while (t->lastfree > t->node) {
618       t->lastfree--;
619       if (keyisnil(t->lastfree))
620         return t->lastfree;
621     }
622   }
623   return NULL;  /* could not find a free place */
624 }
625 
626 
627 
628 /*
629 ** inserts a new key into a hash table; first, check whether key's main
630 ** position is free. If not, check whether colliding node is in its main
631 ** position or not: if it is not, move colliding node to an empty place and
632 ** put new key in its main position; otherwise (colliding node is in its main
633 ** position), new key goes to an empty position.
634 */
635 TValue *luaH_newkey (lua_State *L, Table *t, const TValue *key) {
636   Node *mp;
637   TValue aux;
638   if (unlikely(ttisnil(key)))
639     luaG_runerror(L, "table index is nil");
640   else if (ttisfloat(key)) {
641     lua_Number f = fltvalue(key);
642     lua_Integer k;
643     if (luaV_flttointeger(f, &k, F2Ieq)) {  /* does key fit in an integer? */
644       setivalue(&aux, k);
645       key = &aux;  /* insert it as an integer */
646     }
647     else if (unlikely(luai_numisnan(f)))
648       luaG_runerror(L, "table index is NaN");
649   }
650   mp = mainpositionTV(t, key);
651   if (!isempty(gval(mp)) || isdummy(t)) {  /* main position is taken? */
652     Node *othern;
653     Node *f = getfreepos(t);  /* get a free place */
654     if (f == NULL) {  /* cannot find a free place? */
655       rehash(L, t, key);  /* grow table */
656       /* whatever called 'newkey' takes care of TM cache */
657       return luaH_set(L, t, key);  /* insert key into grown table */
658     }
659     lua_assert(!isdummy(t));
660     othern = mainposition(t, keytt(mp), &keyval(mp));
661     if (othern != mp) {  /* is colliding node out of its main position? */
662       /* yes; move colliding node into free position */
663       while (othern + gnext(othern) != mp)  /* find previous */
664         othern += gnext(othern);
665       gnext(othern) = cast_int(f - othern);  /* rechain to point to 'f' */
666       *f = *mp;  /* copy colliding node into free pos. (mp->next also goes) */
667       if (gnext(mp) != 0) {
668         gnext(f) += cast_int(mp - f);  /* correct 'next' */
669         gnext(mp) = 0;  /* now 'mp' is free */
670       }
671       setempty(gval(mp));
672     }
673     else {  /* colliding node is in its own main position */
674       /* new node will go into free position */
675       if (gnext(mp) != 0)
676         gnext(f) = cast_int((mp + gnext(mp)) - f);  /* chain new position */
677       else lua_assert(gnext(f) == 0);
678       gnext(mp) = cast_int(f - mp);
679       mp = f;
680     }
681   }
682   setnodekey(L, mp, key);
683   luaC_barrierback(L, obj2gco(t), key);
684   lua_assert(isempty(gval(mp)));
685   return gval(mp);
686 }
687 
688 
689 /*
690 ** Search function for integers. If integer is inside 'alimit', get it
691 ** directly from the array part. Otherwise, if 'alimit' is not equal to
692 ** the real size of the array, key still can be in the array part. In
693 ** this case, try to avoid a call to 'luaH_realasize' when key is just
694 ** one more than the limit (so that it can be incremented without
695 ** changing the real size of the array).
696 */
697 const TValue *luaH_getint (Table *t, lua_Integer key) {
698   if (l_castS2U(key) - 1u < t->alimit)  /* 'key' in [1, t->alimit]? */
699     return &t->array[key - 1];
700   else if (!limitequalsasize(t) &&  /* key still may be in the array part? */
701            (l_castS2U(key) == t->alimit + 1 ||
702             l_castS2U(key) - 1u < luaH_realasize(t))) {
703     t->alimit = cast_uint(key);  /* probably '#t' is here now */
704     return &t->array[key - 1];
705   }
706   else {
707     Node *n = hashint(t, key);
708     for (;;) {  /* check whether 'key' is somewhere in the chain */
709       if (keyisinteger(n) && keyival(n) == key)
710         return gval(n);  /* that's it */
711       else {
712         int nx = gnext(n);
713         if (nx == 0) break;
714         n += nx;
715       }
716     }
717     return &absentkey;
718   }
719 }
720 
721 
722 /*
723 ** search function for short strings
724 */
725 const TValue *luaH_getshortstr (Table *t, TString *key) {
726   Node *n = hashstr(t, key);
727   lua_assert(key->tt == LUA_VSHRSTR);
728   for (;;) {  /* check whether 'key' is somewhere in the chain */
729     if (keyisshrstr(n) && eqshrstr(keystrval(n), key))
730       return gval(n);  /* that's it */
731     else {
732       int nx = gnext(n);
733       if (nx == 0)
734         return &absentkey;  /* not found */
735       n += nx;
736     }
737   }
738 }
739 
740 
741 const TValue *luaH_getstr (Table *t, TString *key) {
742   if (key->tt == LUA_VSHRSTR)
743     return luaH_getshortstr(t, key);
744   else {  /* for long strings, use generic case */
745     TValue ko;
746     setsvalue(cast(lua_State *, NULL), &ko, key);
747     return getgeneric(t, &ko, 0);
748   }
749 }
750 
751 
752 /*
753 ** main search function
754 */
755 const TValue *luaH_get (Table *t, const TValue *key) {
756   switch (ttypetag(key)) {
757     case LUA_VSHRSTR: return luaH_getshortstr(t, tsvalue(key));
758     case LUA_VNUMINT: return luaH_getint(t, ivalue(key));
759     case LUA_VNIL: return &absentkey;
760     case LUA_VNUMFLT: {
761       lua_Integer k;
762       if (luaV_flttointeger(fltvalue(key), &k, F2Ieq)) /* integral index? */
763         return luaH_getint(t, k);  /* use specialized version */
764       /* else... */
765     }  /* FALLTHROUGH */
766     default:
767       return getgeneric(t, key, 0);
768   }
769 }
770 
771 
772 /*
773 ** beware: when using this function you probably need to check a GC
774 ** barrier and invalidate the TM cache.
775 */
776 TValue *luaH_set (lua_State *L, Table *t, const TValue *key) {
777   const TValue *p = luaH_get(t, key);
778   if (!isabstkey(p))
779     return cast(TValue *, p);
780   else return luaH_newkey(L, t, key);
781 }
782 
783 
784 void luaH_setint (lua_State *L, Table *t, lua_Integer key, TValue *value) {
785   const TValue *p = luaH_getint(t, key);
786   TValue *cell;
787   if (!isabstkey(p))
788     cell = cast(TValue *, p);
789   else {
790     TValue k;
791     setivalue(&k, key);
792     cell = luaH_newkey(L, t, &k);
793   }
794   setobj2t(L, cell, value);
795 }
796 
797 
798 /*
799 ** Try to find a boundary in the hash part of table 't'. From the
800 ** caller, we know that 'j' is zero or present and that 'j + 1' is
801 ** present. We want to find a larger key that is absent from the
802 ** table, so that we can do a binary search between the two keys to
803 ** find a boundary. We keep doubling 'j' until we get an absent index.
804 ** If the doubling would overflow, we try LUA_MAXINTEGER. If it is
805 ** absent, we are ready for the binary search. ('j', being max integer,
806 ** is larger or equal to 'i', but it cannot be equal because it is
807 ** absent while 'i' is present; so 'j > i'.) Otherwise, 'j' is a
808 ** boundary. ('j + 1' cannot be a present integer key because it is
809 ** not a valid integer in Lua.)
810 */
811 static lua_Unsigned hash_search (Table *t, lua_Unsigned j) {
812   lua_Unsigned i;
813   if (j == 0) j++;  /* the caller ensures 'j + 1' is present */
814   do {
815     i = j;  /* 'i' is a present index */
816     if (j <= l_castS2U(LUA_MAXINTEGER) / 2)
817       j *= 2;
818     else {
819       j = LUA_MAXINTEGER;
820       if (isempty(luaH_getint(t, j)))  /* t[j] not present? */
821         break;  /* 'j' now is an absent index */
822       else  /* weird case */
823         return j;  /* well, max integer is a boundary... */
824     }
825   } while (!isempty(luaH_getint(t, j)));  /* repeat until an absent t[j] */
826   /* i < j  &&  t[i] present  &&  t[j] absent */
827   while (j - i > 1u) {  /* do a binary search between them */
828     lua_Unsigned m = (i + j) / 2;
829     if (isempty(luaH_getint(t, m))) j = m;
830     else i = m;
831   }
832   return i;
833 }
834 
835 
836 static unsigned int binsearch (const TValue *array, unsigned int i,
837                                                     unsigned int j) {
838   while (j - i > 1u) {  /* binary search */
839     unsigned int m = (i + j) / 2;
840     if (isempty(&array[m - 1])) j = m;
841     else i = m;
842   }
843   return i;
844 }
845 
846 
847 /*
848 ** Try to find a boundary in table 't'. (A 'boundary' is an integer index
849 ** such that t[i] is present and t[i+1] is absent, or 0 if t[1] is absent
850 ** and 'maxinteger' if t[maxinteger] is present.)
851 ** (In the next explanation, we use Lua indices, that is, with base 1.
852 ** The code itself uses base 0 when indexing the array part of the table.)
853 ** The code starts with 'limit = t->alimit', a position in the array
854 ** part that may be a boundary.
855 **
856 ** (1) If 't[limit]' is empty, there must be a boundary before it.
857 ** As a common case (e.g., after 't[#t]=nil'), check whether 'limit-1'
858 ** is present. If so, it is a boundary. Otherwise, do a binary search
859 ** between 0 and limit to find a boundary. In both cases, try to
860 ** use this boundary as the new 'alimit', as a hint for the next call.
861 **
862 ** (2) If 't[limit]' is not empty and the array has more elements
863 ** after 'limit', try to find a boundary there. Again, try first
864 ** the special case (which should be quite frequent) where 'limit+1'
865 ** is empty, so that 'limit' is a boundary. Otherwise, check the
866 ** last element of the array part. If it is empty, there must be a
867 ** boundary between the old limit (present) and the last element
868 ** (absent), which is found with a binary search. (This boundary always
869 ** can be a new limit.)
870 **
871 ** (3) The last case is when there are no elements in the array part
872 ** (limit == 0) or its last element (the new limit) is present.
873 ** In this case, must check the hash part. If there is no hash part
874 ** or 'limit+1' is absent, 'limit' is a boundary.  Otherwise, call
875 ** 'hash_search' to find a boundary in the hash part of the table.
876 ** (In those cases, the boundary is not inside the array part, and
877 ** therefore cannot be used as a new limit.)
878 */
879 lua_Unsigned luaH_getn (Table *t) {
880   unsigned int limit = t->alimit;
881   if (limit > 0 && isempty(&t->array[limit - 1])) {  /* (1)? */
882     /* there must be a boundary before 'limit' */
883     if (limit >= 2 && !isempty(&t->array[limit - 2])) {
884       /* 'limit - 1' is a boundary; can it be a new limit? */
885       if (ispow2realasize(t) && !ispow2(limit - 1)) {
886         t->alimit = limit - 1;
887         setnorealasize(t);  /* now 'alimit' is not the real size */
888       }
889       return limit - 1;
890     }
891     else {  /* must search for a boundary in [0, limit] */
892       unsigned int boundary = binsearch(t->array, 0, limit);
893       /* can this boundary represent the real size of the array? */
894       if (ispow2realasize(t) && boundary > luaH_realasize(t) / 2) {
895         t->alimit = boundary;  /* use it as the new limit */
896         setnorealasize(t);
897       }
898       return boundary;
899     }
900   }
901   /* 'limit' is zero or present in table */
902   if (!limitequalsasize(t)) {  /* (2)? */
903     /* 'limit' > 0 and array has more elements after 'limit' */
904     if (isempty(&t->array[limit]))  /* 'limit + 1' is empty? */
905       return limit;  /* this is the boundary */
906     /* else, try last element in the array */
907     limit = luaH_realasize(t);
908     if (isempty(&t->array[limit - 1])) {  /* empty? */
909       /* there must be a boundary in the array after old limit,
910          and it must be a valid new limit */
911       unsigned int boundary = binsearch(t->array, t->alimit, limit);
912       t->alimit = boundary;
913       return boundary;
914     }
915     /* else, new limit is present in the table; check the hash part */
916   }
917   /* (3) 'limit' is the last element and either is zero or present in table */
918   lua_assert(limit == luaH_realasize(t) &&
919              (limit == 0 || !isempty(&t->array[limit - 1])));
920   if (isdummy(t) || isempty(luaH_getint(t, cast(lua_Integer, limit + 1))))
921     return limit;  /* 'limit + 1' is absent */
922   else  /* 'limit + 1' is also present */
923     return hash_search(t, limit);
924 }
925 
926 
927 
928 #if defined(LUA_DEBUG)
929 
930 /* export these functions for the test library */
931 
932 Node *luaH_mainposition (const Table *t, const TValue *key) {
933   return mainpositionTV(t, key);
934 }
935 
936 int luaH_isdummy (const Table *t) { return isdummy(t); }
937 
938 #endif
939