[FFmpeg-devel] [PATCH] avcodec/proresdec: align dequantization matrix buffers

[James Almer]

On 11/15/2017 11:01 PM, Carl Eugen Hoyos wrote:
> 2017-11-16 2:52 GMT+01:00 James Almer <jamrial at gmail.com>:
>> On 11/15/2017 10:35 PM, Carl Eugen Hoyos wrote:
>>> 2017-11-16 2:29 GMT+01:00 James Almer <jamrial at gmail.com>:
>>>
>>>> The OP configure line is a massive dump of pointless
>>>> "--disable" options typical from Gentoo builds,
>>>
>>> Good to know we agree on something.
>>>
>>>> so i didn't even bother looking at it for specific things.
>>>
>>>> And now that i look at yours they are completely different
>>>> as well
>>>
>>> And I thought I spent several hours today only to allow you
>>> to reproduce a crash to ease testing by providing the
>>> necessary configure switches.
>>> They are of course identical to what the op provided, do
>>> you really suggest I added those stupidities for fun?
>>
>> Yours was "--enable-small --toolchain=hardened --disable-avx",
>> and only the latter is in the OP configure line. They didn't use
>> neither small or hardened.
> 
> So you are still saying that I added the options for fun?

No, i didn't say or even imply that. What i said is what's written in
that email: Since the two command lines seemed pretty different, then
perhaps said options were not related to the crash and it's a target
dependent issue.

Again, i didn't intend to offend you in any way whatsoever.

> (Thank you, as said it did take significantly more than
> an hour to find them.)
> 
> The op used an equivalent for both small and hardened in
> his configure line that imo is not supported by us, I found
> a supported configure line that has the same effect on
> config.mak and the resulting binary.
> 
> I knew that you were able to solve the riddle without
> the configure line but I still believe it makes sense
> for you to also being able to test a fix.
> 
> I do not understand why you wrote you cannot reproduce
> when you did not even try to.

Because at no point i assumed the configure line you or the OP used
could be needed to reproduce a crash like this. A wrong assumption in
retrospect, as STRIDE and av_malloc alignment can be affected by
configure options, and stack alignment (as it was the issue here) by
compiler options.
So since i was testing with a different OS than you two, i figured it
was pretty much target dependent and decided to send the patch stating
it was untested.

> Since the issue is not easily reproducible with clang, I
> still wonder if it is reproducible on Windows.

gcc -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2 -fno-strict-overflow
-fstack-protector-all -fPIE -c -o /tmp/ffconf.OTilhXct/test.o
/tmp/ffconf.OTilhXct/test.c
gcc -Wl,-z,relro -Wl,-z,now -fPIE -pie -o /tmp/ffconf.OTilhXct/test.exe
/tmp/ffconf.OTilhXct/test.o
F:/msys/mingw64/bin/../lib/gcc/x86_64-w64-mingw32/7.2.0/../../../../x86_64-w64-mingw32/bin/ld.exe:
unrecognized option '-z'
F:/msys/mingw64/bin/../lib/gcc/x86_64-w64-mingw32/7.2.0/../../../../x86_64-w64-mingw32/bin/ld.exe:
use the --help option for usage information
collect2.exe: error: ld returned 1 exit status
C compiler test failed.

configure can't even succeed with --toolchain=hardened on mingw-w64, so
looks like i wouldn't have been able to reproduce it to begin with.