Thanks for all the explanations.

In that case I restrict this patch to just freeing the buffer from
within driver::finalize only (I think it should be XDELETEVEC
instead of XDELETE, no?).

On Thu, 7 Dec 2023 at 15:42, Jakub Jelinek <jakub@redhat.com> wrote:

> On Thu, Dec 07, 2023 at 03:16:29PM +0000, Costas Argyris wrote:
> > >  Still reachable memory at exit e.g. from valgrind is not a bug.
> >
> > Indeed, this is coming from a valgrind report here:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93019
> >
> > where it was noted that the driver memory leaks could be
> > problematic for JIT.
>
> In the invoke_embedded_driver JIT case yes, that calls driver::finalize (),
> which is why it should be freed before clearing the pointer in there (as
> then it is a real leak).
>
> > So, since using std::vector did reduce the valgrind records
> > by one (I only targeted a single variable to begin with) I took
> > that as a good sign.
> >
> > Regarding adding a call to XDELETE (mdswitches), yes,
> > that would help in the case where driver::finalize () is actually
> > called, which I think is for JIT.    I was trying to take care of the
> > case where it doesn't get called as well, but from what you say
> > I take it that this case is not of interest.
>
> That is wasted compile time on a non-issue.
>
> If you see a JIT issue with definitely lost records, that is something
> that obviously should be fixed (but even in that area I think we've been a
> little bit lazy in the option handling).
> The most important is that the actual compiler binaries (cc1, cc1plus, ...)
> don't leak memory (in the definitely lost kind) like crazy, we have
> --enable-checking=valgrind
> for that purpose, where the driver runs cc1/cc1plus etc. under valgrind,
> but this is very expensive and slow, so usually it is run once during a
> cycle (if at all), on a fast machine could take even in non-bootstrap mode
> a weekend to go through the whole testsuite, then one can look at the
> leaks.
>
>         Jakub
>
>