On Wed, Nov 05, 2014 at 12:19:22PM +0300, Yury Gribov wrote:
> On 11/03/2014 05:27 PM, Marek Polacek wrote:
> >Another shot at optimizing redundant UBSAN_NULL statements.
> >
> >This time we walk the dominator tree - that should result in
> >more effective optimization - and keep a list of UBSAN_NULL
> >statements that dominate the current block, see the comment
> >before sanopt_optimize_walker.
>
> Marek,
>
> A general question - have you considered coding this as a dataflow loop
> instead of dominator walk? That would allow to also remove checks for
> variables defined via PHI nodes provided that all arguments of PHI have
> already been checked.
I'd be afraid that we'd turn sanopt into another var-tracking that way,
with possibly huge hash tables being copied on write, merging of the tables
etc., with big memory and time requirements, having to add --param limits
to give up if the sum size of the tables go over certain limit. Or can you
explain how this problem is different from the var-tracking problem?
The way Marek has coded it up is pretty cheap optimization.
BTW, as discussed privately with Marek last time, we probably want to
optimize UBSAN_NULL (etc.) only if -fno-sanitize-recover=null (etc.)
or if location_t is the same, otherwise such optimizations lead to only
one problem being reported instead of all of them.
Jakub
