[x264-devel] [PATCH 1/1] arm: optimize neon luma intra deblock

[Janne Grunau]

On 2015-09-02 11:48:25 +0300, Martin Storsjö wrote:
> On Wed, 2 Sep 2015, Janne Grunau wrote:
> 
> >Hi Martin,
> >
> >I forgot to rescheduling the beginning of the macro in the last iteration.
> >The floating point compare is a neat trick to test a 64-bit register
> >against 0 and control the program flow based on it. I tend to forget that
> >this is possible.
> >
> >We might have to add vmrs handling to gaspp. Older assembers might
> >understand fmstat instead.
> 
> Hmm, maybe. armasm handles it fine, but old apple binutils doesn't.
> When testing there, I get the following:
> 
> {standard input}:955:ARM register expected -- `vmrs APSR_nzcv,FPSCR'
> 
> When trying with fmstat, I get this instead:
> 
> {standard input}:955:garbage following instruction -- `fmstat APSR_nzcv,FPSCR'

it's plain 'fmstat' or 'fmrx r15, FPSCR', sorry for not being more 
precise.

> Dunno if this is critical for x264 or if everybody build for iOS
> with clang with their built-in assembler nowadays. (At least for
> libav, I've still got a fate instance with the old apple binutils.)

I hope everybody uses clang by now but if possible we should keep the 
old binutils working.

> I can resend with this squashed, or should we stay off this trick for 
> now?

I suspect the old binutils will accept 'fmstat' otherwise it should have 
complained about an unkown instruction. So please squash.

Janne