On Thu, Nov 06, 2014 at 11:10:37AM +0100, Jakub Jelinek wrote:
> Sorry for being slow on this, but in the current algorithm you are using
> the latest vector element from each bb only anyway, so the
> && tree_int_cst_compare (cur_align, align) == 0
> is unnecessary, even if the condition isn't true
On Thu, Nov 06, 2014 at 10:56:17AM +0100, Marek Polacek wrote:
> --- gcc/sanopt.c
> +++ gcc/sanopt.c
> @@ -130,7 +130,30 @@ sanopt_optimize_walker (basic_block bb, struct
> sanopt_ctx *ctx)
> /* At this point we shouldn't have any statements
>tha
On Wed, Nov 05, 2014 at 11:08:27PM +0100, Jakub Jelinek wrote:
> I wonder if for the case where you don't remove, gimple_bb (g) == gimple_bb
> (stmt) and tree_int_cst_compare (cur_align, align) == 0 it wouldn't be
> worthwhile to v.pop (); so that the vector doesn't grow up unnecessarily too
Good
On Wed, Nov 05, 2014 at 10:37:46PM +0100, Marek Polacek wrote:
> 2014-11-05 Marek Polacek
>
> * sanopt.c (sanopt_optimize_walker): Limit removal of the checks.
> testsuite/
> * c-c++-common/ubsan/align-2.c: Add dg-output.
> * c-c++-common/ubsan/align-4.c: Likewise.
> * c
This patch limits optimizing UBSAN_NULL checks a bit, so that we
don't lose diagnostics.
If we know we're on a path where segv would have already occured,
we can drop further checks (e.g. dereferencing a NULL pointer - we
can't recover from that). Also if we're not recovering or using
-fsanitize-u
