These changes cause 16-bit arithmetic to be used for 8-bit data in the gamma
corrected compose and grayscale operations. The arithmetic errors have
three sources all of which are fixed in this commit:
1) 8-bit linear calculations produce massive errors for lower intensity
values.
2) The old 16-bit "16 to 8" gamma table code erroneously wrote the lowest
output value into a table entry which corresponded to multiple output
values (so where the value written should have been the closest to the
transformed input value.)
3) In a number of cases the code to access the 16-bit table did not round;
it did a simple shift, which was wrong and made the side effects of (2)
even worse.
The new gamma code does not have the 16-to-8 problem at the cost of slighly
more calculations and the algorithm used to minimize the number of
calculations has been extended to all the 16-bit tables; it has advantages
for any significant gamma correction.
checks. The code now does a png_error if an attempt is made to do the
row initialization twice; this is an application error and it has
serious consequences because the transform data in png_struct is
changed by each call.
This turns on PNG_FLAG_DETECT_UNINITIALIZED to detect app code that
fails to call one of the 'start' routines (not enabled in libpng-1.5
because it is technically an API change, since it did normally work
before.) It also makes duplicate calls to png_read_start_row (an
internal function called at the start of the image read) benign, as
they were before changes to use png_inflate_claim. Somehow webkit is
causing this to happen; this is probably a mis-feature in the zlib
changes so this commit is only a work-round.
compiler issues slightly different warnings from those issued by the
current vesions of GCC. This eliminates those warnings by
adding/removing casts and small code rewrites.
those uses of png_memcpy that were doing a simple assignment to assignments
(all those cases where the thing being copied is a non-array C L-value.)
Added some error checking to png_set_*() routines and removed the
reference to the non-exported function png_memcpy() from example.c. Fixed
the Visual C 64-bit build - it requires jmp_buf to be aligned, but it had
become misaligned.
Added new "png_structrp" typedef. Because of the
way libpng works both png_info and png_struct are always accessed via a
single pointer. This means adding C99 'restrict' to the pointer gives
the compiler some opportunity to optimize the code. This change allows that.
The logical shift fix for Microsoft Visual C is required by other compilers,
so this enables that fix for all compilers when using compile-time constants.
Under MSYS 'byte' is a name declared in a system header file, so we
changed the name of a local variable to avoid the warnings that result.
These fixes attend to most of the errors revealed in pngvalid, however doing
the gamma work twice results in inaccuracies that can't be easily fixed.
There is now a warning in the code if this is going to happen.
