Re: -funsafe-loop-optimizations

[Richard Guenther <richard dot guenther at gmail dot com>]

On Sat, 1 Jan 2005 22:10:56 +0100, Zdenek Dvorak
<rakdver@atrey.karlin.mff.cuni.cz> wrote:
> Hello,
> 
> > > several times some people sounded opinion that in addition to
> > > improving
> > > analyses in loop optimizer, we might also add a flag that would enable
> > > the compiler to assume that the compiled code does not contain loops
> > > behaving in "weird" way (loops whose control induction variables
> > > overflows, infinite loops with nontrivial exit conditions).
> > >
> > > I remember several people were against this possibility.  Therefore
> > > I send the proposal for the patch implementing this flag first here
> > > to the gcc list, to give the chance to discuss about it before I try
> > > to test and submit it properly (or not).
> >
> > Another idea: why don't you generate both the safe and the unsafe code testing
> > at runtime for the condition (e.g. n==0)? This is also what we do in autovect
> > to test for aligned/unaligned access. We could use this new optimization flag
> > to disable generate of the safe version of the loop, but we would still be
> > generating decent loop with the default options.
> 
> I have thought about this.  But the runtime checks are not for free, and
> increase in code size also would not be insignificant.

You could do the transformation based on value-profiling and putting
the unlikely case in a cold section.  How many times do loop transformations
trigger with -funsafe-loop-optimizations in gcc sources anyway compared
to current behavior?  Are they always profitable?  Which are the common
cases?

Richard