Re: Thinking about release

[Rob Landley <rob@landley.net>]

On 07/10/2013 10:37:55 PM, Rich Felker wrote:
> On Thu, Jul 11, 2013 at 10:44:16AM +1200, Andre Renaud wrote:
> > > This results in 95MB/s on my platform (up from 65MB/s for the  
> existing
> > > memcpy.c, and down from 105MB/s with the asm optimised version).  
> It is
> > > essentially identically readable to the existing memcpy.c. I'm not
> > > really famiilar with any other cpu architectures, so I'm not sure  
> if
> > > this would improve, or hurt, performance on other platforms.
> >
> > Reviewing the assembler that is produced, it appears that GCC will
> > never generate an ldm/stm instruction (load/store multiple) that  
> reads
> > into more than 4 registers, where as the optimised assembler does  
> them
> > that read 8 (ie: 8 * 32bit reads in a single instruction). I've  
> tried
> 
> For the asm, could we make it more than 8? 10 seems easy, 12 seems
> doubtful. I don't see a fundamental reason it needs to be a power of
> two, unless the cache line alignment really helps and isn't just
> cargo-culting. (This is something I'd still like to know about the
> asm: whether it's doing unnecessary stuff that does not help
> performance.)

You're going to hit bus bandwidth at some point, and that's likely to  
be a power of two.

> > various tricks/optimisations with the C code, and can't convince GCC
> > to do more than 4. I assume that this is probably where the  
> remaining
> > 10MB/s is between these two variants.
> 
> Yes, I suspect so. One slightly crazy idea I had was to write the
> function in C with just inline asm for the inner ldm/stm loop. The
> build system does not yet have support for .c files in the arch dirs
> instead of .s files, but it could be added.

Does it have support for a header definining a macro containing the  
assembly bit?

> > Rich - do you have any comments on whether either the C or assembler
> > variants of memcpy might be suitable for inclusion in musl?
> 
> I would say either might be, but it looks like if we want competitive
> performance, some asm will be needed (either inline or full). My
> leaning would be to go for something simpler than the asm you've been
> experimenting with, but with same or better performance, if this is
> possible. I realize the code is not that big as-is, in terms of binary
> size, but it's big from an "understanding it" perspective and I don't
> like big asm blobs that are hard for somebody to look at and say "oh
> yeah, this is clearly right".
> 
> Anyway, the big questions I'd still like to get answered before moving
> forward is whether the cache line alignment has any benefit.

I'd expect so. Fundamentally what the processor is doing is fetching  
and writing cachelines. What it does to the contents of the cachelines  
is just annotating that larger operation.

(Several days behind on email, as usual...)

> Rich

Rob