[Bug tree-optimization/70731] With -Ofast, reorder floating-point and integer operands to minimize conversions

[josh at joshtriplett dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70731

--- Comment #4 from Josh Triplett <josh at joshtriplett dot org> ---
(In reply to Marc Glisse from comment #3)
> (In reply to Josh Triplett from comment #2)
> > That's a fair point.  Perhaps it should go into a separate optimization
> > option, then, though it still seems in the spirit of -Ofast.  (If overflow
> > is a concern, the application would hopefully be checking for that
> > separately; GCC also already has various optimizations that assume overflow
> > cannot occur.)
> 
> The code might look like:
> return (double)lsym + lbits + dsym + (double)dbits;
> (we don't see the difference)
> that is, the application is very explicitly using double, precisely to avoid
> integer overflow issues.
> 
> Assuming that no overflow happens in test2 is normal, it is guaranteed by
> the standard. Assuming that with test1 the integers are small enough that if
> we reorder things no overflow will happen... Well maybe someone else will be
> more optimistic than I am. Maybe if you could restrict it to the case where
> VRP information guarantees that there is no overflow? That might be rare
> enough that it won't matter to you though :-(

In cases where GCC can *guarantee* no overflow, it can optimize those by
default.

However, I think it's worth having an optimization option that allows GCC to do
this even when not guaranteed.  Or, at a minimum, perhaps GCC could notice
cases where such reordering would reduce conversions, and have a warning option
to flag them?