develooper Front page | perl.perl5.porters | Postings from November 2014

Re: Do pad name string buffers need to be pointer-aligned?

Thread Previous | Thread Next
From:
Aaron Crane
Date:
November 30, 2014 15:24
Subject:
Re: Do pad name string buffers need to be pointer-aligned?
Message ID:
CACmk_tvkTETxz2eFc-bJfXEhWRjAexuvv4zw4Z036OJHpWfXVA@mail.gmail.com
Father Chrysostomos <sprout@cpan.org> wrote:
> I could fiddle things a bit to get the string buffer to begin at
> xpadn_flags+1.  That would get us down to 6 pointers for four-char
> names like $self
>
> A couple of years ago someone pointed out that memcmp is faster at
> pointer-aligned addresses.  But how much faster?  Is it noticeable?

My apologies for not replying to this sooner.

I see that your current smoke-me/pn branch contains code for both
options, selectable at compile time, and finds the low-memory version
slightly faster than the version that pointer-aligns the buffers. My
understanding is that, although memcmp() is faster for pointer-aligned
arguments on at least some platforms, it's hard for that advantage to
kick in for short buffers, since most of the cost is in loading the
relevant cache lines. Since many (most?) pad names are less than a
dozen or so bytes long, I don't find it terribly surprising that
saving memory is a slight win over pointer-aligning the buffers.

That said, it would be interesting to use DaveM's new cachegrind-based
Porting/bench.pl to compare Perls compiled with PERL_PADNAME_MINIMAL
and PERL_PADNAME_ALIGNED, on code using lexical variables with names
of a variety of lengths.

One other note on smoke-me/pn: there seems to be a missing line in the
commit message for 549200d "Minimise the size of padname + string
buffer" — perhaps the line began "#if" and was therefore swallowed by
"git commit"?

-- 
Aaron Crane ** http://aaroncrane.co.uk/

Thread Previous | Thread Next


nntp.perl.org: Perl Programming lists via nntp and http.
Comments to Ask Bjørn Hansen at ask@perl.org | Group listing | About