Modifying a locally-defined non-volatile variable below the setjmp()
return point results in undefined behavior whereby the variable may not
have the expected value after setjmp() returns.
Fixes#379
The primary impetus for this is to eliminate build warnings, such as
(32-bit only)
section "__textcoal_nt" is deprecated
object file (XXXXXX.o) was built for newer OSX version (10.XX) than
being linked (10.5)
Upgrading to GCC 6 results in neutral performance for compression,
a measured average overall decompression speedup of 2.5% for 64-bit
code, and a measured average overall decompression speedup of -4.3% for
32-bit code on a 3 GHz Core i7. The 4.3% slow-down for 32-bit code is
deemed acceptable, given that 32-bit macOS apps are deprecated.
Splitting the pointer arithmetic in GET_SYM() into a separate add and
sub instruction was an attempt to work around an error ("invalid operand
type") that occurred when assembling the file with NASM. However, this
created a link error on macOS ("ld: illegal text-relocation to
'_jconst_huff_encode_one_block' in
simd/CMakeFiles/simd.dir/i386/jchuff-sse2.asm.o from
'_jsimd_huff_encode_one_block_sse2' in
simd/CMakeFiles/simd.dir/i386/jchuff-sse2.asm.o for architecture i386")
and also changed the alignment of the code in ways that might have
affected the previous benchmark results (which took a great deal of time
to obtain.) Ultimately, the path of least resistance is just to
require NASM 2.13 or later.
This commit improves the C and SSE2 Huffman encoding implementations in
the following ways:
- Avoid using xmm8-xmm15 in the x86-64 SSE2 implementation. There is no
actual need to use those registers, and avoiding them produces a
cleaner WIN64 function entry/exit-- as well as shorter code, since REX
prefixes can be avoided (this is helpful on certain CPUs, such as
Intel Atom, for which instruction fetch and decoding can be a
bottleneck.)
- Optimize register usage so that fewer REX prefixes and
register-register moves are needed.
- Use the bit counter to store the number of free bits in the bit buffer
rather than the number of bits in the bit buffer. This changes the
method for inserting a code into the bit buffer to:
(put_buffer |= code << (free_bits -= code_size));
As a result:
* Only one bit counter needs to stay in a register (we just keep it in
cl.)
* The bit buffer contents are already properly aligned to be written
out (after a byte swap.)
* Adjusting the free bits counter and checking if the bit buffer is
full can be combined into a single operation.
* We can wait to flush the bit buffer until the buffer is actually
full and not just in danger of becoming full. Thus, eight bytes can
be flushed at a time.
- Speed is quite sensitive to the alignment of branch target labels, so
insert some padding and remove branches from the flush code.
(Flushing this way isn't actually faster when compared to using
branches, but the branchless code doesn't need extra alignment and is
thus smaller.)
- Speculatively write out the bit buffer as a single 8-byte write,
falling back to a byte-by-byte write only if there are any 0xFF bytes
in the bit buffer that need to be encoded as 0xFF 0x00.
- Use MMX registers for the 32-bit implementation (so the bit buffer can
be 64 bits wide.)
- Slightly reduce overall function code size.
- Eliminate or combine a few SSE instructions.
- Make some minor improvements to instruction scheduling.
- Adjust flush_bits() in jchuff.c to handle cases in which the bit
buffer has less than 7 free bits (apparently that couldn't happen
before.)
Based on:
947a09defa262ebb6b816e9a091221
See change log for performance claims.
Closes#292
This macro is a relic of libjpeg's historic need to support a wide
variety of C compilers with varying degrees of compatibility. Such was
necessary during the open systems era, because C compilers were often
supplied by the system vendor. Prior to 1989, there was no C standard
per se, and even after ANSI C became a thing, there were still compilers
in use that didn't conform to it (libjpeg was first released in 1991.)
Realistically, only a handful of C compilers are in widespread use these
days, and all modern C compilers should support structure assignment.
... to avoid backward compatibility issues with GCC 4-6 MinGW
toolchains. Apparently GCC 7+ MinGW toolchains introduce a link-time
dependency with internal MinGW CRT functions that are meant to provide
compatibility with Microsoft's Universal CRT (ucrt) library, but those
internal functions are not available in GCC 4-6 MinGW toolchains. This
made it impossible to use the official builds of libjpeg.a and
libturbojpeg.a with GCC 4-6 MinGW toolchains (a fatal link error--
"undefined reference to '__imp___acrt_iob_func'"-- occurred.)
This problem was not immediately apparent after switching to the MSYS2
implementation of MinGW (d6d7b53968)
because, for a while, MSYS2 was still using GCC 5 and 6.
Refer to libjpeg-turbo/libjpeg-turbo#382
byte, word, dword, qword, oword, and yword are all assembler keywords,
so it makes sense to use lowercase for these so as not to mistake them
for macros or constants.
(AKA "Java for OS X systems.") This implementation of Java 1.6 is long
obsolete and not supported on any version of macOS past High Sierra.
Oracle no longer provides a 32-bit JVM on macOS, so it is no longer
necessary to provide a 32-bit version of the TurboJPEG Java wrapper on
macOS.
If the TurboJPEG instance passed to tjDecodeYUV[Planes]() was previously
used to decompress a progressive JPEG image, then we need to disable the
progressive decompression parameters in the underlying libjpeg instance
before calling jinit_master_decompress().
This commit also modifies the build system so that the "tjtest" target
will test for this issue, and it corrects a previous oversight in the
build system whereby tjbenchtest did not test progressive
compression/decompression unless WITH_JAVA was true.
(regression introduced by 5b177b3cab)
The SSE2 implementation of progressive Huffman encoding performed
extraneous iterations when the scan length was a multiple of 16.
Based on:
bb7f1ef983Fixes#335Closes#367
- Re-purpose the non-SIMD test to test with MSan as well.
- Re-purpose the ASan test to test with UBSan as well.
- Use the default Travis build environment rather than specifying Ubuntu
14.04. I think I added 'dist: trusty' back when 14.04 was newer than
the default, but now it's older than the default.
- Enable verbose output for any unit tests that fail (so we can see the
sanitizer output.)
Prevent several integer overflow issues and subsequent segfaults that
occurred when attempting to compress or decompress gigapixel images with
the TurboJPEG API:
- Modify tjBufSize(), tjBufSizeYUV2(), and tjPlaneSizeYUV() to avoid
integer overflow when computing the return values and to return an
error if such an overflow is unavoidable.
- Modify tjunittest to validate the above.
- Modify tjCompress2(), tjEncodeYUVPlanes(), tjDecompress2(), and
tjDecodeYUVPlanes() to avoid integer overflow when computing the row
pointers in the 64-bit TurboJPEG C API.
- Modify TJBench (both C and Java versions) to avoid overflowing the
size argument to malloc()/new and to fail gracefully if such an
overflow is unavoidable.
In general, this allows gigapixel images to be accommodated by the
64-bit TurboJPEG C API when using automatic JPEG buffer (re)allocation.
Such images cannot currently be accommodated without automatic JPEG
buffer (re)allocation, due to the fact that tjAlloc() accepts a 32-bit
integer argument (oops.) Such images cannot be accommodated in the
TurboJPEG Java API due to the fact that Java always uses a signed 32-bit
integer as an array index.
Fixes#361
This commit modifies h1v2_fancy_upsample() so that it uses an ordered
dither pattern, similar to that of h2v1_fancy_upsample(), rounding up or
down the result for alternate pixels rather than always rounding down.
This ensures that the decompression error pattern for a 4:4:0 JPEG image
will be similar to the rotated decompression error pattern for a 4:2:2
JPEG image. Thus, the final result will be similar regardless of
whether a 4:2:2 JPEG image is rotated or transposed before or after
decompression.
Closes#356
xcode7.3 is based on El Capitan, which is EOL, and Homebrew no longer
provides El Cap bottles (pre-compiled binaries.) Thus, Homebrew was
trying to build GCC 5, YASM, and the other packages we need from source,
which caused the Mac CI builds to time out. I tried goading Homebrew
into installing GCC 5.5.0_2 and YASM 1.3.0_1, which still have El Cap
bottles available, by using the URLs of those specific versions of the
formulae (from the Homebrew GitHub repository) as package names. This
failed, however, because 'brew bundle' converted the URLs to all
lowercase. I then tried explicitly installing the old formulae by using
'brew install' with the aforementioned URLs (bypassing the Travis
Homebrew addon), but Homebrew still tried to build all of the
dependencies from source.
Upgrading to Xcode 8.3.x necessitated regression testing the performance
on iOS, but that proved less of a pain than figuring out how to install
all of the old Homebrew bottles we needed. Also, this future-proofs the
CI builds against the inevitable discontinuation of the xcode7.3 image.
Note that Xcode 8.3.x improves iOS 64-bit decompression performance
significantly relative to Xcode 7.2.x or 7.3.x. iOS 32-bit performance
unfortunately regresses by as much as 5%, but it can't be helped (32-bit
iOS apps are no longer supported on iOS 11+ anyhow, and the next major
release of libjpeg-turbo will remove support for them as well.)
... including, but not limited to:
- unused macros
- private functions not marked as static
- unprototyped global functions
- variable shadowing
(detected by various non-default GCC 8 warning options)
According to Intel's manual [1], "If a value entered for CPUID.EAX is
higher than the maximum input value for basic or extended function for
that processor then the data for the highest basic information leaf is
returned."
Right now, libjpeg-turbo doesn't first check that leaf 07H is supported
before attempting to use it, so the ostensible AVX2 bit (Bit 05) of the
CPUID result might actually be Bit 05 from a lower leaf. That bit might
be set, even if the CPU doesn't support AVX2.
This commit modifies the x86 and x86-64 SIMD feature detection code so
that it first checks whether CPUID leaf 07H is supported before
attempting to use it to check for AVX2 instruction support.
DRC:
This commit should fix
https://bugzilla.mozilla.org/show_bug.cgi?id=1520760
However, I have not personally been able to reproduce that issue,
despite using a Nehalem (pre-AVX2) CPU on which the maximum CPUID leaf
has been limited via a BIOS setting.
Closes#348
[1]
"Intel® 64 and IA-32 Architectures Software Developer's Manual, Volume 2 (2A, 2B, 2C & 2D): Instruction Set Reference, A-Z", https://software.intel.com/sites/default/files/managed/a4/60/325383-sdm-vol-2abcd.pdf, page 3-192.
Same as d3a3a73f64 but in the fast decode
path. It was necessary to use a different-sized test image in order to
trigger the error in this location.
Refer to #347
