Spell of Mastery Progress

Ok. The new voxel rendering algorithm performs 1.4 times faster. Not much, but voxel models can also be split into 16-bit chunks, which locally use only 16-bit addressing. That should speedup the algorithm further, if the bottleneck is with the memory bandwidth.

I've also discovered that GCC's __builtin_popcount is slower than a lookup table, since it branches into subroutine __popcountdi2. Why GCC doesnt use the x86 popcnt opcode? No idea. No idea. But it probably would be easier to answer why Stallman molests little children.

Implemented auto keyword. Which make declaration type be the type of its initializer.

Here is the code handling auto vars:
static void pg_decl(cnode_t *n) {
cnode_t *type = n->head;
cnode_t *vars = n->tail;
cnode_t *s = is_struct(type);
if (s) {
pg_struct(s, type, vars);
} else {
indent(); log("vars:\n");
if (type->id==N_TYPE_STOR && type->head->id == (AUTO|N_TERM)) {
cnode_t *expr = vars->head->tail;
sym_t *s = pg_declof(expr);
if (!s) goto normal;
int ptrc = ptr_level(s->decl)+pg_declof_ptr_lv;
if (ptrc < 0) ptrc = 0;
char *tname = pg_typename(s->type);
char *ps = malloc(ptrc+1);
for (int i = 0; i < ptrc; i++) ps[i] = '*';
ps[ptrc] = 0;
char *ts = sjoin(tname,ps);
free(ps);
patch_t *p = new_patch(P_AUTO, n->sofs);
p->as[0].s = ts;
p->as[1].i = type->eofs;
type = s->type;
return;
}
normal:
pg_vars(type, vars);
}
}

It doesn't allow more than one auto var inside single statement, because that is a rewriter, not a true compiler. So all other vars get the type of the first one.