[FFmpeg-devel] Patch: Inline asm fixes for Intel compiler on Windows
protogonoi at gmail.com
Mon Mar 17 05:32:59 CET 2014
> If this patchset is applied it will need someone to maintain it
> otherwise it will probably be more broken than working most of the
Figured as much. The reason ive been updating these patches for the last
few months is that I have my own repo that I use for compiling ffmpeg for
Windows that I use in a project im working on. So for a while atleast I
will be compiling ffmpeg using icl (and the proposed changes) so its easier
for me to locate bugs and I will be fixing them in my own repo so its not
to much effort to provide them upstream.
The rebasing is kind of a pain, but thats mainly because these patches were
initially written up last year and have been working in my repo for ages.
The rebasing is just so I can get the patches relative to upstream head so
that you guys can apply it easily. And all the rebasing is pretty much
always because configure is modified pretty much every day. Once the
configure changes are in from the initial patches future updates will
mostly be just re-tweaking some asm and shouldn't have so many issues and
wouldnt need constant rebasing. Either way consider me able and willing to
support the icl asm for awhile.
> > +
> > # system capabilities
> > log2_deps="!libc_msvcrt"
> is this actually needed ?
Possibly not, the nonlocal labels should not be used if inline asm is
disabled as the code will never get called. But as far as the config.h
header is concerned I figured for completeness there is no point having the
option enabled in the header when it clearly cant be because all inline asm
is disabled. I assumed it was logical to include it but if you want it
removed anyway let me know.
this looks a bit more complex than needed
> you should be able to just unconditionally add the
> NAMED_CONSTRAINTS_ADD() to the existing macro
I though of doing that but there are several locations ( line 392, 759, 781
to name a few) where there are no additional named constraints added. There
are also several locations where only 1 named constraint is required. So by
adding all of them to the existing macro then it would be adding named
constraints that are only needed in some but not all instances. This raises
compilation warnings about unused named references so I opted to keep
things as clean and efficient as possible. So again Ill leave this up to
you. If youd prefer the possible safety of just adding all named
constraints to the existing macro but have some of them be unused at
various times then let me know.
This also goes for the RGB_PACK24_B_OPERANDS as combining them will result
in additional unneeded constraints being added to asm blocks that dont use
More information about the ffmpeg-devel