[FFmpeg-devel] [PATCH] vp9: 16bpp tm/dc/h/v intra pred simd (mostly sse2) functions.

[Ronald S. Bultje]

Hi,

On Mon, Oct 5, 2015 at 7:39 AM, Ronald S. Bultje <rsbultje at gmail.com> wrote:

> Hi,
>
> On Mon, Oct 5, 2015 at 5:54 AM, Henrik Gramner <henrik at gramner.com> wrote:
>
>> On Mon, Oct 5, 2015 at 10:55 AM, Matt Oliver <protogonoi at gmail.com>
>> wrote:
>> > This patch has broken the 32bit msvc builds:
>> >
>> http://fate.ffmpeg.org/report.cgi?time=20151005065109&slot=x86_32-msvc12-windows-native
>> >
>> > I had a look through the code but couldnt find the cause. The error
>> message
>> > just points to the code line for a macro instantiation and I'm not
>> familiar
>> > enough with the code to be able to work out where within the macro the
>> > error is actually occurring and why. So ill leave this one to someone
>> more
>> > familiar with the code.
>>
>> I see where the problem is.
>>
>> Using stack space in cglobal requires an extra register if the stack
>> alignment is less than mmsize (32-bit msvc only has 4-byte stack
>> alignment), and if the function already utilizes all available
>> registers that will result in failure.
>>
>> The fix is either to reduce the number of registers used by the
>> function or to only enable the function on x86-32 if the stack is at
>> least 16-byte aligned. x86inc has the variable STACK_ALIGNMENT which
>> can be tested for this, in C there's the HAVE_ALIGNED_STACK define
>> which is set if the stack is 16-byte aligned.
>
>
> Yeah, it should use 6 regs, I'll send a patch.
>

Should be fixed now, apologies for the breakage.

Ronald