Re: [musl] [PATCH] aarch64: add optimized memcpy, memmove and memset

[Rich Felker <dalias@libc.org>]

On Wed, Mar 25, 2020 at 10:45:45PM +0100, Szabolcs Nagy wrote:
> minimal edits to upstream version for easier updates
> and because this code was benchmarked across many cores.
> 
> gcc generates slow code for the current c implementations.
> 
> the integer memcpy was chosen instead of the simd one,
> this performs better on little cores, i think this is
> the more conservative choice for now.

I think this was discussed before on IRC, and I'm not particularly
opposed to these especially since aarch64 is one of the most important
archs these days. However I would really like to avoid adding more asm
source files with the function flow written in asm when the only thing
that really needs to benefit from asm is the inner loop body. I know
nothing has happened on this front since we last talked about it, so
it's very possible that the answer is just "we need something with
decent performance in the short term and nobody has cycles to spend on
doing it better right now and so we we should just use the asm
files"...

> note: there are upcoming security architectures which
> may mean updates to these functions (BTI - landing pads,
> PAUTH - return address signing, MTE - 16byte tag granule
> may affect optimized strcmp etc, not relevant yet), but
> runtime support for these will need other libc changes.

If these mattered they'd be another reason to prefer having the
function in C with minimal inline asm or just extensions for unaligned
loads/stores, but MTE is the only one of these that's interesting and
it doesn't conflict with any current code in musl at all (nothing does
unaligned overreads; they have to be assumed to be able to fault
anyway).

Rich