Check LABEL_SELFREF for visible points only

Looked at the effects of checking visible points only when doing some benchmarks for generalizing strdcmp, which got a mention at TravelMapping#600. Posting here to keep the spam out of the official more public repo.

At first I thought the branch misprediction penalty may outweigh the penalty of the extraneous datacheck, but now it's looking like that's probably not the case.

Existing strdcmp, tailored for use with LABEL_SELFREF specifically:

int strdcmp(const char* a, const char* b, const char d)
{   do  if (!*b) return *a == d ? 0 : *a-*b;
    while   (*a++ == *b++);
    return  *--a - *--b;
}

Generalized strdcmp, for wider applicability:

int strdcmp(const char* a, const char* b, const char d)
{   do  if  (!*b || *b==d) return (!*a || *a==d) ? 0 : *a;
        else if (!*a || *a==d) return -*b;
    while   (*a++ == *b++);
    return  *--a - *--b;
}

The benchmark goes between Sorting waypoints in Quadtree and Finding unprocessed wpt files:

size_t accum = 0;
cout << et.et() << "AccumBegin" << endl;
for (HighwaySystem& h : HighwaySystem::syslist)
  for (Route& r : h.routes)
    for (Waypoint& w : r.points)
      if (!w.is_hidden) // comment/uncomment this line as appropriate for benchmarking
      { const char* l = w.label.data() + (w.label[0] == '*');
    string rte = w.route->banner[0] == '-' ? w.route->route : w.route->name_no_abbrev();
    // first check for number match after a slash, if there is one
    const char* slash = strchr(l, '/');
    if (slash++ && strncmp(l, "Old", 3))
    {   const char* digits = rte.data();
        while (!isdigit(*digits) && *digits) ++digits;
        if (    digits != rte.data()+rte.size()
             && !strdcmp(slash, digits, '_') && !w.coloc_same_number(digits)
             || !strdcmp(slash, rte.data(), '_')
           )    ++accum;
    }
      }
cout << et.et() << "AccumEnd @ " << accum << endl;
return 0;//*/

(For use with older commits, & -> *, points -> point_list, and w. -> w-> as needed.)

The flavors of siteupdate_{P,R}{A,V}{S,G}:

Specialized strdcmp
Generalized strdcmp
All waypoints checked for LABEL_SELFREF
Visible waypoints only
Pointers stored in Route::point_list, to Waypoints allocated individually on the heap Branched from 4c8a0c9, most TMArray conversion done but not Regions or Waypoints.
References to Waypoint objects stored in a TMArray Branched from c7aa426, TMArray conversion essentially done, just not squashed.

The results:		spec
ptr all	.1387	.1397
ptr vis	.1410	.1413
ref all	.0876	.0873
ref vis	.0869	.0875

P vs R: The TMArray conversion is a clear win.
S vs G: Before the TMArray conversion, slowed us down 0.3 - 1.0 ms. That's tolerable. :) After, we can see we're within margin of error: the visible-only benchmark slows down 0.6 ms (right in the middle of that same ballpark), and the all-points benchmark measured 0.3 ms faster. Worth noting that this calls strdcmp more times, for all 182 hidden points with a slash. :P Yes. Let's go ahead and generalize strdcmp.

A vs V: When first benchmarking before TMArray-ifying Waypoints, this looked like a loss (branch misprediction penalty?), taking 1.6 - 2.3 ms longer. Now? Within margin of error (+0.2 ms) or even faster (-0.7 ms). Also worth noting is that the benchmark code above doesn't include the rest of the datacheck after the slash test, which includes a `strncmp` fort every point without a post-slash match. Add this in and limiting to visible points is a more clean win:		spec	gen
ref all	.0964	.0962
ref vis	.0926	.0935

Besides... It's The Right Thing To Do!

yakra / DataProcessing

Check LABEL_SELFREF for visible points only #246