Re: [EXT] Re: Option processing question

[Jan Hubicka <hubicka at ucw dot cz>]

> On Mon, Jan 13, 2020 at 9:28 PM Gary Oblock <goblock@marvell.com> wrote:
> >
> > If I have an optimization set up to run in LTO (call it my_opt) and
> > the flags -flto-partition=one, -flto options and -fmy-opt are set
> > then the optimization might or might not be run depending of whether
> > it can all fit in one partition.
> >
> > What I'm thinking is as long as it's a fatal error detectable anywhere
> > upstream in the compilation to not specify -fmy-opt without -flto-partition=one
> > then all will be well. So how do I detect it at all and where would I put the checking?
> 
> I guess you could always check the symbols used_from_other_partition
> and/or in_other_partition flags?  That is, it might be perfectly possible
> for some corner-cases to do the transform independently with more than one
> partition if the partitioning happens to isolate things perfectly?

Well, if you do optimization at ltrans you basically see symbols in
ohter partitions as external so there is not that much of practical
difference in the implementation of analysis except that it will catch
fewer transforms.

But the idea for keeping -flto-partition=none around was to make it
possible to write a prototype of IPA pass that does not nees to play
with partitioning and incrementally add WPA bits.
It seems that this strategy may work here - having a prototype that
works only w/o WHOPR but produces valid code & does something useful
seems like good first step (which is hard enough at its own).

So I would suggest to look at -flto-partition=none first and keep in
mind that eventually we will want a solution that plays well with
partitioning.

Honza
> 
> Richard.
> 
> > Gary
> > ________________________________
> > From: Richard Biener <richard.guenther@gmail.com>
> > Sent: Monday, January 13, 2020 2:30 AM
> > To: Gary Oblock <goblock@marvell.com>; Jan Hubicka <hubicka@ucw.cz>
> > Cc: gcc@gcc.gnu.org <gcc@gcc.gnu.org>
> > Subject: [EXT] Re: Option processing question
> >
> > External Email
> >
> > ----------------------------------------------------------------------
> > On Sat, Jan 11, 2020 at 4:47 AM Gary Oblock <goblock@marvell.com> wrote:
> > >
> > > I'm writing an LTO optimization that requires "-flto-partition=one" How can I make
> > > sure that this is the case? I've spent hours greping the code and the Internals Doc is
> > > worth less than nothing for something like this.
> >
> > That's of course because you shouldn't be doing this ;)
> >
> > > If you have an answer or even
> > > I good idea of where to look please let me know. Note, for an normal "-fbalh" there's
> > > a flag_blah that I could look at but for this there seems to be zip, nil, diddly squat, etc.
> >
> > At LTRANS time you indeed don't know.
> >
> > But there's -flto-partition=none (none, not one), that you can detect somehow
> > (I think in_lto_p && !flag_ltrans && !flag_wpa).
> >
> > Richard.
> >
> > > Many thanks,
> > >
> > > Gary