Re: Was: Re: Loop optimizer misses simple optimisation?

[John Vickers <jvickers@acorn.com>]

On Wed 06 May, Joern Rennecke wrote:
> > int a, b, c, d;
> > ...
> >   c = b / c;
> >   d = b % c;
> >   
> > becomes:
> > 
> >   c = a / b;
> >   d = a - c * b;  

Ho hum.  Perhaps the transformation above is not necessarily
to the advantage of the user. 

> This would have to depend on the cost of the modulo operation - and if b
> is constant or can be made constant by constant-propagation, this also
> ight make a difference to the cost.

Yes indeed.  We have a lot of ways of skinning the cat,
and we should check the costs, assuming sensible costs
are available ;-).

I should say that I haven't investigated what egcs is doing
in more than a rather cursory way.

I think I was over-generalising without checking my facts -
the targets I'm interested in don't have either SImode divide or divmod.
If the div needs a libcall, it looks like expmed.c::expand_divmod
gives up on trying to optimise the mod.

This is what I was really getting at.


By the way, it seems like there are a few odd tricks which expand_divmod
misses e.g. divisor so big that repeated subtraction is efficient,
or mod when the divisor is (2^n +/-1)*2^m.

John.