We must not allow negative color values be passed to
gdImageFillToBorder(), because that can lead to infinite recursion
since the recursion termination condition will not necessarily be met.
Instead of rolling our own bounds check we use clip_1d() as it's done
in gdImageLine() and in external libgd. We must not pass the image
width and height, respectively, but rather the largest ordinate value
that is allowed to be accessed, i.e. width-1 and height-1,
respectively.
The color components are supposed to be in range 0..255, so we must not
cast them to `signed char`, what can be the default for `char`.
Port of <https://github.com/libgd/libgd/commit/77c8d359>.
The recently introduced test_image_equals_file() doesn't properly work for
palette images, because in this case only the palette indexes are compared,
what can lead to false positives and negatives as shown in the added test.
To fix that we convert palette images to truecolor, what is supposed to be
faster than calling imagecolorsforindex() for each pixel.
We furthermore rely on PHP's refcounting to free unused images; after all,
this is not C.
Firstly, there is no fundamental difference between running the tests on
freetype before and after 2.4.10, so we re-join the respective test cases.
Secondly, there are some minor differences in the results depending on
architecture, freetype versions and config options, so we cater to that
by allowing small deviations from the expectations.
Apparently, different FreeType versions render the string differently, so
we have to cater to these slight differences. Unfortunately, the testing
framework doesn't yet offer a clean API for this, so we work around the
limitations by parsing and checking the output of test_image_equals_file().
The behavior of imagecropauto($im, IMG_CROP_DEFAULT) is rather different
for the bundled and an external libgd, see
<https://github.com/libgd/libgd/issues/298>.
Therefore we split the test for now.
This test case fails with external libgd, because libgd < 2.2.4 converts
palette images to truecolor (<https://github.com/libgd/libgd/issues/297>).
However, even if there happens no truecolor conversion, it is not guaranteed
that the palette indexes of the original and the cropped image are identical
(this is, for instance, not the case with current libgd master where the bug
has been fixed). Therefore we check for the expected RGBA value instead of
the palette index.
The bundled GD test suites makes heavy use of md5() to verify the result
of drawing operations. This leads to fragile tests (even a slight change
in a PNG header would cause failure, and of course there is the
possibility of collisions), and even worse, eventual test failures are
rather unrevealing.
Therefore we replace all md5() verification with a simplistic
test_image_equals_file(), which is basically a simplified port of libgd's
gdTestImageCompareToFile(), adapted to the needs of PHPTs.
In the long run better tests helpers should be introduced (see also
<http://news.php.net/php.internals/94081>), but for now this solution
is preferable over the former.
(cherry picked from commit 24f9e96792)
gdImageTrueColorToPalette() is sometimes wasteful by putting multiple white
color entries into the palette. This is caused by an obvious typo, where
to avoid a division by zero when `total` is zero, `count` is checked instead
of `total`.
We fix this issue to improve the quality of the color quantization.
Cf. <https://github.com/libgd/libgd/commit/24b4550f>
PNG allows identical images to be stored differently what makes nearly all
tests checking the MD5 hash of the PNG representation fail with external
libgd. For now, we use the GD format instead, which doesn't allow for such
differences.
Of course, this md5() checking should be replaced by a image diffing feature
in the long run.
Calling imagesetstyle() with an empty $styles array caused gdImageSetStyle()
to be called with `noOfPixels==0`, what could have lead to OOB reads.
Actually, this issue will be fixed in libgd, but to avoid issues when older
libgd is in use, we simply disallow passing an empty $styles array to
imagesetstyle(), what wouldn't serve a useful purpose anyway.
The test claims that it would be "checking all the values in returned array",
but due to the use of %a it actually skipped elements. We fix that by using
%s instead.
No pixel of a filled arc must ever be drawn multiple times. Otherwise we get
artifacts regarding transparency. That happens with the current
implementation of gdImageFilledArc() unless gdChord or gdNoFill are set.
When gdPie is set, however, the filled arc is drawn in wedges, which are
polygons of three points, and so some overlap is natural.
To resolve the issue, we stick with the current algorithm of calculating the
wedges, but instead of drawing each polygon separately, we put the relevant
points in a large array, and draw a single polygon. That also is supposed to
improve the performance considerably.
Note that this modification will change the results when gdImageSetStyle()
or gdImageSetBrush() are used, but we believe that this modification is also
an improvement in this regard, even though it still might not make much
sense to use these functions with gdImageFilledArc().
The respective fix for libgd is
<https://github.com/libgd/libgd/commit/e7e20d6>.
