Re: memchr() performance

[Rich Felker <dalias@libc.org>]

On Mon, Sep 19, 2016 at 03:29:53PM +0200, Georg Sauthoff wrote:
> On Sun, Sep 18, 2016 at 10:40:36PM +0200, Szabolcs Nagy wrote:
> 
> Hello,
> 
> > * Georg Sauthoff <mail@georg.so> [2016-09-18 20:54:22 +0200]:
> 
> [..]
> 
> > > On recent Intel CPUs it is even slower than a naive implementation:
> > > 
> > > https://gms.tf/stdfind-and-memchr-optimizations.html#measurements
> > > https://gms.tf/sparc-and-ppc-find-benchmark-results.html
> > > 
> > > Of course, on x86, other implementations that use SIMD instructions
> > > perform even better.
> 
> > yes simd is expected to be faster.
>  
> > but that needs asm which is expensive to maintain (there is no
> > portable simd language extension for c and there is the aliasing
> > issue: the reinterpret_cast in your code is formally ub).
> 
> you mean because the vector-word pointer returned by reinterpret_cast is
> used to access vector-words in the memory passed via a char pointer and
> this is not covered by the ISO C++ strict aliasing rules?
> 
> Yes. Sure, ub means that anything can happen, but this case should be ok
> with GCC - if the function is compiled in isolation in its own
> translation unit.
> 
> I mean, there isn't much possibiltiy for reordering due to the
> application of strict-aliasing-rules that would yield a different
> result.
> 
> There are no aliased write accesses.
> 
> Btw, the current musl memchr() implementation has similar aliased
> accesses - there, unsigned characters are aliased via a size_t pointer. 

Yes, that's a known bug with a pending patch - see:

http://www.openwall.com/lists/musl/2016/04/27/8

Rich