/*
* The routines @emph{ALIGN_synced} and @emph{ALIGN_ordered} allow to
* simply query the alignment status of the two head columns of two
* BATs.
*/
int
ALIGNsynced(BAT *b1, BAT *b2)
{
BATcheck(b1, "ALIGNsynced: bat 1 required", 0);
BATcheck(b2, "ALIGNsynced: bat 2 required", 0);
/* first try to prove head columns are not in sync */
if (BATcount(b1) != BATcount(b2))
return 0;
if (ATOMtype(BAThtype(b1)) != ATOMtype(BAThtype(b2)))
return 0;
if (BAThvoid(b1) && BAThvoid(b2))
return (b1->hseqbase == b2->hseqbase);
/* then try that they are */
if (b1->batCacheid == b2->batCacheid)
return 1; /* same bat. trivial case */
if (BATcount(b1) == 0)
return 1; /* empty bats of same type. trivial case */
if (b1->halign && b1->halign == b2->halign)
return 1; /* columns marked as equal by algorithmics */
if (VIEWparentcol(b1) && ALIGNsynced(BBPcache(VIEWhparent(b1)), b2))
return 1; /* view on same bat --- left recursive def.. */
if (VIEWparentcol(b2) && ALIGNsynced(b1, BBPcache(VIEWhparent(b2))))
return 1; /* view on same bat --- right recursive def.. */
return 0; /* we simply don't know */
}
/*
* Synced BATs
*/
str
BKCisSynced(bit *ret, const bat *bid1, const bat *bid2)
{
BAT *b1, *b2;
if ((b1 = BATdescriptor(*bid1)) == NULL) {
throw(MAL, "bat.isSynced", RUNTIME_OBJECT_MISSING);
}
if ((b2 = BATdescriptor(*bid2)) == NULL) {
BBPunfix(b1->batCacheid);
throw(MAL, "bat.isSynced", RUNTIME_OBJECT_MISSING);
}
*ret = ALIGNsynced(b1, b2) != 0;
BBPunfix(b1->batCacheid);
BBPunfix(b2->batCacheid);
return MAL_SUCCEED;
}
/*
* The prime routine for the BAT layer is to create a new hash index.
* Its argument is the element type and the maximum number of BUNs be
* stored under the hash function.
*/
BAT *
BAThash(BAT *b, BUN masksize)
{
BAT *o = NULL;
lng t0,t1;
(void) t0;
(void) t1;
if (VIEWhparent(b)) {
bat p = VIEWhparent(b);
o = b;
b = BATdescriptor(p);
if (!ALIGNsynced(o, b) || BUNfirst(o) != BUNfirst(b)) {
BBPunfix(b->batCacheid);
b = o;
o = NULL;
}
}
MT_lock_set(&GDKhashLock(ABS(b->batCacheid)), "BAThash");
if (b->H->hash == NULL) {
unsigned int tpe = ATOMstorage(b->htype);
BUN cnt = BATcount(b);
BUN mask;
BUN p = BUNfirst(b), q = BUNlast(b), r;
Hash *h = NULL;
Heap *hp = NULL;
str nme = BBP_physical(b->batCacheid);
BATiter bi = bat_iterator(b);
ALGODEBUG fprintf(stderr, "#BAThash: create hash(" BUNFMT ");\n", BATcount(b));
/* cnt = 0, hopefully there is a proper capacity from
* which we can derive enough information */
if (!cnt)
cnt = BATcapacity(b);
if (b->htype == TYPE_void) {
if (b->hseqbase == oid_nil) {
MT_lock_unset(&GDKhashLock(ABS(b->batCacheid)), "BAThash");
ALGODEBUG fprintf(stderr, "#BAThash: cannot create hash-table on void-NIL column.\n");
return NULL;
}
ALGODEBUG fprintf(stderr, "#BAThash: creating hash-table on void column..\n");
tpe = TYPE_void;
}
/* determine hash mask size p = first; then no dynamic
* scheme */
if (masksize > 0) {
mask = HASHmask(masksize);
} else if (ATOMsize(ATOMstorage(tpe)) == 1) {
mask = (1 << 8);
} else if (ATOMsize(ATOMstorage(tpe)) == 2) {
mask = (1 << 12);
} else if (b->hkey) {
mask = HASHmask(cnt);
} else {
/* dynamic hash: we start with
* HASHmask(cnt/64); if there are too many
* collisions we try HASHmask(cnt/16), then
* HASHmask(cnt/4), and finally
* HASHmask(cnt). */
mask = HASHmask(cnt >> 6);
p += (cnt >> 2); /* try out on first 25% of b */
if (p > q)
p = q;
}
if (mask < 1024)
mask = 1024;
t0 = GDKusec();
do {
BUN nslots = mask >> 3; /* 1/8 full is too full */
r = BUNfirst(b);
if (hp) {
HEAPfree(hp);
GDKfree(hp);
}
if (h) {
ALGODEBUG fprintf(stderr, "#BAThash: retry hash construction\n");
GDKfree(h);
}
/* create the hash structures */
hp = (Heap *) GDKzalloc(sizeof(Heap));
if (hp &&
(hp->filename = GDKmalloc(strlen(nme) + 12)) != NULL)
sprintf(hp->filename, "%s.%chash", nme, b->batCacheid > 0 ? 'h' : 't');
if (hp == NULL ||
hp->filename == NULL ||
(h = HASHnew(hp, ATOMtype(b->htype), BATcapacity(b), mask)) == NULL) {
MT_lock_unset(&GDKhashLock(ABS(b->batCacheid)), "BAThash");
if (hp != NULL) {
GDKfree(hp->filename);
GDKfree(hp);
}
return NULL;
}
switch (tpe) {
case TYPE_bte:
starthash(bte);
break;
case TYPE_sht:
starthash(sht);
break;
case TYPE_int:
case TYPE_flt:
starthash(int);
break;
case TYPE_dbl:
case TYPE_lng:
starthash(lng);
break;
default:
for (; r < p; r++) {
ptr v = BUNhead(bi, r);
BUN c = (BUN) heap_hash_any(b->H->vheap, h, v);
if ( HASHget(h,c) == HASHnil(h) &&
nslots-- == 0)
break; /* mask too full */
HASHputlink(h,r, HASHget(h,c));
HASHput(h,c, r);
}
break;
}
} while (r < p && mask < cnt && (mask <<= 2));
/* finish the hashtable with the current mask */
p = r;
switch (tpe) {
case TYPE_bte:
finishhash(bte);
break;
case TYPE_sht:
finishhash(sht);
break;
case TYPE_int:
case TYPE_flt:
finishhash(int);
break;
case TYPE_dbl:
case TYPE_lng:
finishhash(lng);
break;
default:
for (; p < q; p++) {
ptr v = BUNhead(bi, p);
BUN c = (BUN) heap_hash_any(b->H->vheap, h, v);
HASHputlink(h,p, HASHget(h,c));
HASHput(h,c,p);
}
break;
}
b->H->hash = h;
t1 = GDKusec();
ALGODEBUG
fprintf(stderr, "#BAThash: hash construction "LLFMT" usec\n", t1-t0);
ALGODEBUG HASHcollisions(b,b->H->hash);
}
