Re: [PATCH 0/8] ARC updates to uClibc

On February 16, 2015 6:15:15 AM GMT+01:00, Vineet Gupta <[email protected]> wrote:
>On Saturday 14 February 2015 07:41 PM, Bernhard Reutner-Fischer wrote:
>
>>Please consider merging the cumulative ARC updates to uClibc.
>>The important one being support for variable page size and
>>optimization of sigaction code path.
>
>
>Can you please elaborate on 6 and 8?
>LGTM otherwise.
>Thanks,
>
>
>When we ran LMBench under perf, lat_sig install showed hot spots in
>userspace memcpy and memset.
>Turns out that those were for small sized buffers, leading to greater
>micro-architectural penalty for ARC and more importantly the user of
>those functions in this case seemed extraneous anyways:
>
>* lat_sig -> sigemtpyset -> memset for 8 bytes was pointless, given
>_SIGSET_NWORDS is 2 for most arches
>* lat_sig -> arc-specific-sigaction -> memcpy for 20 bytes was also not
>fully justified given ARC code clobbers a few fields right after the
>memcpy.
>
>so 8/8 removed the out-of-line memset from sigemptyset, making it
>fallback to inline variant in bits/sigset.h - similar to what
>sigfillset already does.

Fair enough.

>
>While it at I also did some arch specific adjustment in sigaction path
>- inlining the rt_sigaction syscall stub detour to reduce branch return
>stack mispredicts etc - which is what 6/8 does !

This sounds suspicious.
IIRC we already had that argument, last time around _dl_do_reloc and _dl_do_lazy_reloc.
Could it be that your port has a bug here ( missed optimisation ) around ifunc handling? Sounds like back then on ARM https://gcc.gnu.org/PR40887#c6

What am I missing?

Thanks,