Thanks to Amir Livne Bar-on for contributing this. It's about to be
used to build confidence in a rewrite of findloop.c in a followup
commit. I also expect that it will be useful later on for testing
other graph algorithms.
This patch is not Amir's original. I've polished the code in various
small ways. The biggest change is that when it outputs a failing
graph, the graph is printed in the form of an Untangle game id, in the
hope that that will make it possible to visualise easily while
debugging the problem!
The new generator works on the same 'combinatorial coordinates' system
as the more recently written Hats and Spectre generators.
When I came up with that technique for the Hats tiling, I was already
tempted to rewrite the Penrose generator on the same principle,
because it has a lot of advantages over the previous technique of
picking a randomly selected patch out of the subdivision of a huge
starting tile. It generates the exact limiting distribution of
possible tiling patches rather than an approximation based on how big
a tile you subdivided; it doesn't use any overly large integers or
overly precise floating point to suffer overflow or significance loss,
because it never has to even _think_ about the coordinates of any
point not in the actual output area.
But I resisted the temptation to throw out our existing Penrose
generator and move to the shiny new algorithm, for just one reason:
backwards compatibility. There's no sensible way to translate existing
Loopy game IDs into the new notation, so they would stop working,
unless we kept the old generator around as well to interpret the
previous system. And although _random seeds_ aren't guaranteed to
generate the same result from one build of Puzzles to the next, I do
try to keep existing descriptive game IDs working.
So if we had to keep the old generator around anyway, it didn't seem
worth writing a new one...
... at least, until we discovered that the old generator has a latent
bug. The function that decides whether each tile is within the target
area, and hence whether to make it part of the output grid, is based
on floating-point calculation of the tile's vertices. So a change in
FP rounding behaviour between two platforms could cause them to
interpret the same grid description differently, invalidating a Loopy
game on that grid. This is _unlikely_, as long as everyone uses IEEE
754 double, but the C standard doesn't actually guarantee that.
We found this out when I started investigating a slight distortion in
large instances of Penrose Loopy. For example, the Loopy random seed
"40x40t11#12345", as of just before this commit, generates a game
description beginning with the Penrose grid string "G-4944,5110,108",
in which you can see a star shape of five darts a few tiles down the
left edge, where two of the radii of the star don't properly line up
with edges in the surrounding shell of kites that they should be
collinear with. This turns out to be due to FP error: not because
_double precision_ ran out, but because the definitions of COS54,
SIN54, COS18 and SIN18 in penrose.c were specified to only 7 digits of
precision. And when you expand a patch of tiling that large, you end
up with integer combinations of those numbers with coefficients about
7 digits long, mostly cancelling out to leave a much smaller output
value, and the inaccuracies in those constants start to be noticeable.
But having noticed that, it then became clear that these badly
computed values were also critical to _correctness_ of the grid. So
they can't be fixed without invalidating old game IDs. _And_ then we
realised that this also means existing platforms might disagree on a
game ID's validity.
So if we have to break compatibility anyway, we should go all the way,
and instead of fixing the old system, introduce the shiny new system
that gets all of this right. Therefore, the new penrose.c uses the
more reliable combinatorial-coordinates system which doesn't deal in
integers that large in the first place. Also, there's no longer any
floating point at all in the calculation of tile vertex coordinates:
the combinations of 1 and sqrt(5) are computed exactly by the
n_times_root_k function. So now a large Penrose grid should never have
obvious failures of alignment like that.
The old system is kept for backwards compatibility. It's not fully
reliable, because that bug hasn't been fixed - but any Penrose Loopy
game ID that was working before on _some_ platform should still work
on that one. However, new Penrose Loopy game IDs are based on
combinatorial coordinates, computed in exact arithmetic, and should be
properly stable.
The new code looks suspiciously like the Spectre code (though
considerably simpler - the Penrose coordinate maps are easy enough
that I just hand-typed them rather than having an elaborate auxiliary
data-generation tool). That's because they were similar enough in
concept to make it possible to clone and hack. But there are enough
divergences that it didn't seem like a good idea to try to merge them
again afterwards - in particular, the fact that output Penrose tiles
are generated by merging two triangular metatiles while Spectres are
subdivisions of their metatiles.
This uses a tile shape very similar to the hat, but the tiling
_structure_ is totally changed so that there aren't any reflected
copies of the tile.
I'm not sure how much difference this makes to gameplay: the two
tilings are very similar for Loopy purposes. But the code was fun to
write, and I think the Spectre shape is noticeably prettier, so I'm
adding this to the collection anyway.
The test programs also generate a pile of SVG images used in the
companion article on my website.
The 'core' library contains almost all the same objects as 'common',
but leaves out hat.c. And the auxiliary program 'hatgen' now links
against that slightly reduced core library instead of 'common'.
This avoids a dependency loop: one of hatgen's jobs is to generate
hat-tables.h, but hat-tables.h is a dependency of it.
Of course, the generated hat-tables.h is already committed, so this
doesn't present a bootstrapping problem in a normal build. But if
someone modifies hatgen.c in order to regenerate hat-tables.h, and
does so in a way that makes it uncompilable, they can't rebuild hatgen
and try again! Of course you can always revert changes with git, but
it's annoying to have to. Better to keep the dependencies non-cyclic
in the first place.
Having stated the principle in the previous commit, I should apply it
consistently. A source file linked into the Puzzles library of common
support code should not also define a main() under ifdef.
This commit only goes as far as the _library_ support modules. It
would be a much bigger job to do the same for all the actual _puzzles_
that have test main()s or standalone-solver main()s. And it's not
necessary, because modifying one of those source files only triggers a
rebuild of _one_ puzzle, not absolutely everything. (Not to mention
that it's quite likely the puzzle and the test main() will need to be
modified in conjunction anyway.)
As in the previous commit, this has required exposing a few internal
API functions as global, and maybe editing them a bit. In particular,
the one-shot internal function that divvy_rectangle() loops on until
it succeeds is now exposed as divvy_rectangle_attempt(), which means
the test program doesn't have to condition a failure counter into the
real function.
I've thrown away penrose-vector-test completely, because that didn't
look like a test program with any ongoing use at all - it was surely
vestigial, while James was getting the vector representation up and
running in the first place.
I noticed while hacking on hat-test recently that it's quite awkward
to be compiling a test main() program that lives in a source file also
built into the Puzzles support library, because every modification to
main() also triggers a rebuild of the library, and thence of all the
actual puzzles. So it's better if such a test main() has its own
source file.
In order to make hat-test work standalone, I've had to move a lot of
hat.c's internal declarations out into a second header file. This also
means making a bunch of internal functions global, which means they're
also in the namespace of programs other than hat-test, which means in
turn that they should have names with less implicit context.
