Likely GMP bug

[Marco Bodrato]

Ciao,

Il Mer, 30 Maggio 2018 10:20 am, Niels Möller ha scritto:
> "Marco Bodrato" <bodrato at mail.dm.unipi.it> writes:

> Thinking about micro optimizations... Consider

>>       count_trailing_zeros (c, ulimb);
>>       ulimb = (ulimb >> 1) >> c;

> vs

>>       count_trailing_zeros (c, ulimb);
>>       ulimb >>= (c + 1);

> I wonder if maybe it's preferable to always use the first variant? Both

You are saying that my patch https://gmplib.org/repo/gmp/rev/e4849ae7c974
was not an emergency workaround, but a cleverly subtle optimization?
Well, I have to admit that I was not aware... but thanks :-)

> first variant, the first shift can be issued in parallel with ctz,

If we could hint to the compiler...
  HINT (c >= 0 && c + 1 < ....)
then they should be equivalent. The optimiser of the compiler can choose
the best one for the target environment :-)

If we define c as unsigned, we do not need other hints, the compiler has
enough information to understand that ulimb>>(c+1) can be implemented as
(ulimb>>1)>>c for all non-undefined cases. The vice-versa is not as easy.

Maybe, of course, that current compilers does not take care of this
possible optimisation...

Ĝis,
m

-- 
http://bodrato.it/