Re: [musl] Release prep for 1.2.1, and afterwards

[Rich Felker <dalias@libc.org>]

On Fri, Jun 26, 2020 at 10:40:49AM +0200, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2020-06-25 21:20:06 -0400]:
> > On Thu, Jun 25, 2020 at 05:15:42PM -0400, Rich Felker wrote:
> > > On Thu, Jun 25, 2020 at 04:50:24PM -0400, Rich Felker wrote:
> > > > > > > but it would be nice if we could get the aarch64
> > > > > > > memcpy patch in (the c implementation is really
> > > > > > > slow and i've seen ppl compare aarch64 vs x86
> > > > > > > server performance with some benchmark on alpine..)
> > > > > > 
> > > > > > OK, I'll look again.
> > > > > 
> > > > > thanks.
> > > > > 
> > > > > (there are more aarch64 string functions in the
> > > > > optimized-routines github repo but i think they
> > > > > are not as important as memcpy/memmove/memset)
> > > > 
> > > > I found the code. Can you commend on performance and whether memset is
> > > > needed? (The C memset should be rather good already, moreso than
> > > > memcpy.)
> 
> the asm seems faster in all measurements but there is
> a lot of variance with different size/alignment cases.
> 
> the avg improvement on typical workload and the possible
> improvements across various cases and cores i'd expect:
> 
> memcpy typical: 1.6x-1.7x
> memcpy possible: 1.2x-3.1x
> 
> memset typical: 1.1x-1.4x
> memset possible: 1.0x-2.6x
> 
> > > Are the assumptions (v8-a, unaligned access) documented in memcpy.S
> > > valid for all presently supportable aarch64?
> 
> yes, unaligned access on normal memory in userspace
> is valid (part of the base abi on linux).
> 
> iirc a core can be configured to trap unaligned access
> and it is not valid on device memory so e.g. such
> memcpy would not work in the kernel. but avoiding
> unaligned access in memcpy is not enough to fix that,
> the compiler will generate unaligned load for
> 
> int f(char *p)
> {
>     int i;
>     __builtin_memcpy(&i,p,sizeof i);
>     return i;
> }
> 
> > > 
> > > A couple comments for merging if we do, that aren't hard requirements
> > > but preferences:
> > > 
> > > - I'd like to expand out the macros from ../asmdefs.h since that won't
> > >   be available and they just hide things (I guess they're attractive
> > >   for Apple/macho users or something but not relevant to musl) and
> > >   since the symbol name lines need to be changed anyway to public
> > >   name. "Local var name" macros are ok to leave; changing them would
> > >   be too error-prone and they make the code more readable anyway.
> 
> the weird macros are there so the code is similar to glibc
> asm code (which adds cfi annotation and optionally adds
> profile hooks to entry etc)
> 
> > > 
> > > - I'd prefer not to have memmove logic in memcpy since it makes it
> > >   larger and implies that misuse of memcpy when you mean memmove is
> > >   supported usage. I'd be happy with an approach like x86 though,
> > >   defining an __memcpy_fwd alias and having memmove tail call to that
> > >   unless len>128 and reverse is needed, or just leaving memmove.c.
> 
> in principle the code should be called memmove, not memcpy,
> since it satisfies the memmove contract, which of course
> works for memcpy too. so tail calling memmove from memcpy
> makes more sense but memcpy is more performance critical
> than memmove, so we probably should not add extra branches
> there..
> 
> > 
> > Something like the attached.
> 
> looks good to me.

I think you saw already, but just to make it clear on the list too,
it's upstream now. I'm open to further improvements like doing
memmove (either as a separate copy of the full implementation or some
minimal branch-to-__memcpy_fwd approach) but I think what's already
there is sufficient to solve the main practical performance issues
users were hitting that made aarch64 look bad in relation to x86_64.

I'd still like to revisit the topic of minimizing the per-arch code
needed for this so that all archs can benefit from the basic logic,
too.

Rich