Re: [musl] [PATCH 2/2] Add big-endian support to ARM assembler memcpy

[Andre McCurdy <armccurdy@gmail.com>]

On Wed, Jan 15, 2020 at 7:46 AM Rich Felker <dalias@libc.org> wrote:
> On Fri, Sep 13, 2019 at 01:38:34PM -0700, Andre McCurdy wrote:
> > On Fri, Sep 13, 2019 at 11:59 AM Rich Felker <dalias@libc.org> wrote:
> > > On Fri, Sep 13, 2019 at 11:44:32AM -0700, Andre McCurdy wrote:
> > > > Allow the existing ARM assembler memcpy implementation to be used for
> > > > both big and little endian targets.
> > >
> > > Nice. I don't want to merge this just before release, but as long as
> > > it looks ok I should be able to review and merge it afterward.
> > >
> > > Note that I'd really like to replace this giant file with C using
> > > inline asm just for the inner block copies and C for all the flow
> > > control, but I don't mind merging this first as long as it's correct.
> >
> > Sounds good. I'll wait for your feedback after the upcoming release.
>
> Sorry this dropped off my radar. I'd like to merge at least the thumb
> part since it's simple enough to review quickly and users have
> actually complained about memcpy being slow on armv7 with -mthumb as
> default.

Interesting. I wonder what the reference was against which the musl C
code was compared? From my own benchmarking I didn't find the musl
assembler to be much faster than the C code. There are armv6 and maybe
early armv7 CPUs where explicit prefetch instructions make a huge
difference (much more so than C -vs- assembler). Did the users who
complained about musl memcpy() compare against a memcpy() which uses
prefetch? For armv7 using NEON may help, although the latest armv7
cores seem to perform very well with plain old C code too. There are
lots of trade offs so it's impossible for a single implementation to
be universally optimal. The "arm-mem" routines used on Raspberry Pi
seem to be a very fast for many targets, but unfortunately the armv6
memcpy generates mis-aligned accesses so isn't suitable for armv5.

  https://github.com/bavison/arm-mem/