Re: [cfg-branch] unroll-new cleanups

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello.

> > But this is not the point!!! The point is, that this loop iterates (2^bits of
> > i - 1) times; unless you know exactly what "bits of i" is, you cannot
> > calculate the number of iterations exactly. If you think you can count this
> > safely, then do it and everything will work fine (I just was not quite sure
> > about it, so I've chosen more conservative way); but you cannot just ignore it.
> 
> I see what you are shooting for.  It is already taken care for in count_loop_iterations:
> 
>   /* Given that iteration_var is going to iterate over its own mode,
>      not HOST_WIDE_INT, disregard higher bits that might have come
>      into the picture due to sign extension of initial and final
>      values.  */
>   abs_diff &= ((unsigned HOST_WIDE_INT) 1
>                << (GET_MODE_BITSIZE (GET_MODE (desc->var)) - 1)
>                << 1) - 1;
> 
> so no need to worry about it anymore.

OK, I have overseen this. Then probably the best way to take care of loop

for (i=10;i<0;i++)

is really in count_loop_iterations (set niter to 0 in this case).

Zdenek