Re: -funsafe-loop-optimizations

[Dale Johannesen <dalej at apple dot com>]

On Jan 1, 2005, at 2:05 PM, Zdenek Dvorak wrote:
> Hello,
>
> > Is it maybe possible in some cases to look at how the induction
> > variable is used (e.g. as an array index and we know that it is
> > undefined to have an index outside the array domain), and decide
> > on whether the loop must be finite or not based on that?
>
> we do this on tree level.

The standard also limits how big an object can be (more exactly, does 
not
require a compiler to handle bigger objects).   So if the overflow case 
would
result in addressing a bigger object, we can assume that doesn't happen,
even if we don't know how big the allocated size was.  (There are cases 
in
Spec where the allocated size is read in from a file, so even 
whole-program
opt can't find it.)  This is true even when the object access isn't 
connected
to the index variable, for example:

   for (p=0, i=1; i<=n; i++)
     foo (*p++) ;

The limit in the standard is 64K-1 bytes, but as a practical matter it 
would
certainly be wise to pick a larger limit, possibly target-dependent.