Re: [PATCH v2] Add __pure2 to __locale_ctype_ptr(_l)

[Corinna Vinschen <vinschen at redhat dot com>]

On Nov  6 18:40, Wilco Dijkstra wrote:
> Corinna Vinschen wrote: 
> > On Oct 31 13:34, Wilco Dijkstra wrote:
> > > The newlib ctype functons are extremely inefficient due to repeatedly
> > > calling __locale_ctype_ptr for every single use of a ctype macro, even
> > > in a tight loop.  Improve this by adding the missing __pure2 attribute so
> > > the pointer can be cached just like in GLIBC, resulting in > 2x speedup
> > > in loops.
> >
> > Apart from Craig's comment, how did you test this?  I checked this on
> > x86_64 Cygwin with input of 300 Megs, and I didn't see any difference
> > between using the current code and an additional __pure2.  A tight loop
> > around these 300 Megs of input always took 2.3 - 2.4 secs.
> >
> > My testcase is attached.  Input file as parameter with no punctuation
> > chars (so the printf is deliberately never called, but GCC's optimization
> > doesn't discard the result of ispunct.
> 
> I tried the inner loop from your benchmark on an array and I get a speedup
> of 3.1x on an AArch64 server with -O3. There seems to be an optimization bug
> in GCC with -O2 since it won't lift the pure call, apparently due to the
> side-effect in c = *p++. If I change to a for loop with the p++ done separately,
> I get the expected code with -O2.

I tried again on an x86_64 system running gcc 6.4.0, with -O2, -O3, with
the original while as well as a for loop, with and without __pure2, and
I don't see any difference whatsoever.

So it seems this is very much architecture dependent.  We can surely add
__pure2 if it makes a difference on some platforms.

> Btw is there a simple recipe to link newlib on a Linux system? I wasn't able
> to get it to link with the libc.a, so I had to hack a __locale_ctype_ptr function
> by copying bits of code from various newlib files...

I don't know.  I didn't look into the linux branch for ages.


Corinna

-- 
Corinna Vinschen
Cygwin Maintainer
Red Hat

Attachment: signature.asc
Description: PGP signature