Cast `row_width` to `size_t` before multiplication when computing
`row_info->rowbytes`, to avoid 32-bit truncation on LP64 systems.
Fix the cast order in `png_do_expand_palette` where the widening
cast was applied after the multiply rather than before it.
Reviewed-by: Cosmin Truta <ctruta@gmail.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Replace hand-strength-reduced idioms to make the source code easier to
read and to audit:
- Replace `i << 2` with `i * 4`;
- Replace `(i << 1) + i` with `i * 3`;
- Replace `sizeof(png_color)` with 3;
- Replace `sizeof(png_uint_32)` with 4;
- Use `index * CONSTANT` consistently in all stride expressions.
The optimizing compiler will emit identical machine code.
In particular, the expression `sizeof(png_uint_32)` used in this context
was a semantic misnomer: its value should represent bytes per pixel in
the RGBA format, not the size of the underlying integer type.
Both `png_do_expand_palette_rgba8_neon` and
`png_do_expand_palette_rgb8_neon` advanced in fixed-size chunks without
guarding the final iteration, allowing out-of-bounds reads and writes
when the row width is not a multiple of the chunk size.
Restrict the NEON loop to full chunks only, remove the now-unnecessary
post-loop adjustment, and undo the `*ddp` pre-adjustment before the
pointer handoff to the scalar fallback.
Reported-by: Amemoyoi <Amemoyoi@users.noreply.github.com>
Co-authored-by: Amemoyoi <Amemoyoi@users.noreply.github.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Following up on commit 7c67f70da1:
Replace 22 deprecated pointer typedefs with their underlying types.
The typedefs replaced are:
1. Core structure types:
- png_structp, png_structrp, png_structpp and const variants
- png_infop, png_inforp, png_infopp and const variants
2. Utility structure types:
- png_row_infop, png_row_infopp
- png_imagep, png_imagepp
- png_controlp
3. Color structure types:
- png_colorp, png_colorpp and const variants
- png_color_8p, png_color_8pp and const variants
- png_color_16p, png_color_16pp and const variants
4. Text and time structure types:
- png_textp, png_textpp and const variants
- png_timep, png_timepp and const variants
5. Chunk structure types:
- png_unknown_chunkp, png_unknown_chunkpp and const variants
- png_sPLT_tp, png_sPLT_tpp and const variants
- png_sPLT_entryp, png_sPLT_entrypp and const variants
6. Internal types:
- png_compression_bufferp
- png_dsortp, png_dsortpp
Remove all remaining uses of spurious PNG_RESTRICT qualifiers.
Non-aliasing has already been guaranteed by the type-based aliasing
rules combined with the libpng usage pattern of a single pointer per
type.
Update the user manual accordingly.
This completes the systematic replacement of deprecated pointer
typedefs across the entire libpng codebase. The typedef declarations
remain in png.h for backward compatibility, marked as deprecated.
Reviewed-by: Chris Blume <ProgramMax@gmail.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Following up on commit ddce3ff853:
1. Replace 28 deprecated pointer typedefs with their underlying types.
The typedefs replaced are:
- Basic: png_voidp, png_charp and const variants
- Byte: png_bytep, png_const_bytep, png_bytepp
- Numeric:
- Integer: 16/32-bit integer pointer variants
- Fixed-point: png_fixed_point_p and variants
- Floating-point: png_doublep and variants
2. Remove a spurious PNG_RESTRICT qualifier. Non-aliasing has already
been guaranteed by the type-based aliasing rules.
3. Update the user manual accordingly.
Reviewed-by: Chris Blume <ProgramMax@gmail.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Make the error directives uniformly consistent with one another.
Remove surrounding quotes and trailing punctuation, and rephrase
some of the error messages and some of their associated comments
for either brevity or clarity.
The leading blank lines are apparently an artefact of an older source
control system. They are not required and they look like accidents,
because starting a source file with a blank line is not a regular habit
of software developers nowadays.
This is a cherry-pick of commit 37cc20add8
from branch 'libpng18'.
Signed-off-by: John Bowler <jbowler@acm.org>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
The leading blank lines are apparently an artefact of an older source
control system. They are not required and they look like accidents,
because starting a source file with a blank line is not a regular habit
of software developers nowadays.
Signed-off-by: John Bowler <jbowler@acm.org>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
The change removes the need for build configuration of target specific
code such as SIMD enhancements.
The extensible framework is based on checks in the new file named
pngtarget.h and allows per target description of requirements and
capabilities using macros with the prefix PNG_TARGET_. These macros are
interpreted by the core libpng code in a target independent way.
The PNG_TARGET_ prefixed macros describe the target specific
implementation and document how to include it. The actual inclusion of
the code is done, at present, by a single target independent core file
named pngsimd.c.
This approach ensures that no special configuration in the build system
is required on any system. It is not required on a system which is
compiling for the build host where we can reasonably assume that the
compiler is correctly set up. Neither is it required for cross builds
so long as the toolchain is correctly set up. In this case the
toolchain means the cross build system and any ancilliary programs it
requires.
More details of the implementation are included in the source files
named pngtarget.h and pngsimd.c. These files document in C comments how
additional target specific code might be added.
The changes consist primarily of movement of target specific code from
the core libpng source files to the existing target specific
subdirectories such as the ones named intel and arm. This isolates the
code into the corresponding subdirectory and eliminates the requirement
for changes across the libpng structure.
As part of this the ARM processor NEON instruction set specific palette
to RGB code has been generalised to allow implementation in other
architectures. The resultant updated version of the original (ARM Ltd)
implementation has been moved to the subdirectory named arm.
In addition some selected changes have been included to allow testing
with higher warning levels than the typical compiler defaults.
Reviewed-by: Cosmin Truta <ctruta@gmail.com>
Signed-off-by: John Bowler <jbowler@acm.org>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
This changes the arm/palette_neon_intrinsics.c code from using a bare
cast to using the portable macro `png_aligncastconst`. This removes
the `-Wcast-align` warning.
The cast is safe because the pointer was allocated by `malloc(1024)`,
and `malloc` is required to return a pointer with sufficient alignment
for anything the allocation can accomodate.
Nevertheless, this is a quick and dirty fix; the use of `png_bytep` in
the `png_struct` declaration of the pointer is machine-specific. Other
architectures may need different types and, consequently, the only type
which is portable to other companies' CPUs is `(void*)`.
This is a cherry-pick of commit a604b12c11
from branch 'libpng18'.
Reviewed-by: Cosmin Truta <ctruta@gmail.com>
Signed-off-by: John Bowler <jbowler@acm.org>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
This changes the arm/palette_neon_intrinsics.c code from using a bare
cast to using the portable macro `png_aligncastconst`. This removes
the `-Wcast-align` warning.
The cast is safe because the pointer was allocated by `malloc(1024)`,
and `malloc` is required to return a pointer with sufficient alignment
for anything the allocation can accomodate.
Nevertheless, this is a quick and dirty fix; the use of `png_bytep` in
the `png_struct` declaration of the pointer is machine-specific. Other
architectures may need different types and, consequently, the only type
which is portable to other companies' CPUs is `(void*)`.
Reviewed-by: Cosmin Truta <ctruta@gmail.com>
Signed-off-by: John Bowler <jbowler@acm.org>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
In the previous commit 9e538750d9
we removed the obsolete assembler implementation `filter_neon.S`.
In this commit we add a stand-in for the original file, restoring
the original source tree structure, for the benefit of continuing
hassle-free libpng source upgrades in the 1.6.x line.
This file contains hand-coded assembler implementations of the filter
functions for 32-bit Arm platforms. These are only used when the
compiler doesn't support neon intrinsics (added to GCC 4.3 in 2008) or
is exactly GCC 4.5.4 (released 2012), both of which are sufficiently
unlikely to be true that it's fair to say the assembler is no longer
used.
This commit deletes filter_neon.S and removes the now obsolete
preprocessor logic in pngpriv.h.
Signed-off-by: Bill Roberts <bill.roberts@arm.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Following up on the previous commit by Martin Storsjö:
If the availability of ARM Neon is not only known at compile time,
but also expected by the compiler's target architecture, there should
be no further need for a run-time check. In fact, such a check would
be counter-productive.
ARM Neon is expected to be unconditionally available on ARM64.
Issue an #error if PNG_ARM_NEON_CHECK_SUPPORTED is defined on an
ARM64-based platform.
For the plain ARM platforms where ARM Neon checking is not yet
implemented, issue a slightly more descriptive #error, including
a suggested course of action.
MSVC's <arm64_neon.h> header contains cl.exe intrinsics that are not
recognized by clang-cl.
Contributed-by: Ryan VanderMeulen <rvandermeulen@mozilla.com>
Contributed-by: Mike Klein <mtklein@google.com>
Signed-off-by: Cosmin Truta <ctruta@gmail.com>
Move deallocation of riffled_palette from png_write_destroy to
png_read_destroy. The reader (not the writer) is the owner of
riffled_palette.
Move allocation and initialization of riffled_palette from
png_do_read_transformations to png_init_palette_transformations.
Allow riffled_palette inside png_struct only if the ARM Neon
optimizations are enabled.
Rename png_riffle_palette_rgba to png_riffle_palette_rgba8, etc.,
to better indicate the strict applicability of these routines.
Fix an unused parameter warning in the build configurations where
riffled palette optimization is not enabled.
Fix indentation.
