Re: [PATCH] Fix LTO option streaming

[Richard Biener <rguenther at suse dot de>]

On Thu, 23 Apr 2015, Jan Hubicka wrote:

> > 
> > It looks like when transitioning to using target and optimization
> > option nodes for compile-time to link-time option streaming you
> > didn't adjust lto-opts.c nor lto-wrapper.c.  The following fixes
> 
> Yep, I assumed that lto-wrapper's merging is now redundant for optimization
> options, while it is still needed for global functions (-fPIC) that was
> not covered by my changes and since this was all bit late, it seemed I can
> just keep it for next stage1.
> 
> > the target option side (for SWITCHABLE_TARGET).  Do not record
> > any target options in the lto_opts section.
> > 
> > Honza - I suppose we don't have any testcase that this works, I'll
> > try to come up with sth.  This also looks like a correctness issue
> > to me.
> 
> Thanks, that would be a good idea. So with SWITCHABLE_TARGET
> we have problem that after switching the options are misinterpretted?
> I think Jakub made patch to simply skip target-option-node streaming here,
> how this is supposed to work.

Not all targets support target attributes and I think those that
support them are those that are SWITCHABLE_TARGET - but maybe I am
mistaken?

> > 
> > We can do similar changes for optimize options, now, for all targets, no?
> 
> I am not quite sure about this; not all options are per-function nature.
> I need to audit all the target options for LTO safety, I did some work
> on this for optimization options, but the target options are still very wild.
> For example I am not quite convinced mixing -malign-double -mno-align-double units
> will work as expected.

Well, but once types are laid out specifying sth else at link time
won't help.  But yes, if we don't stream the option we have no chance
to complain in lto-wrapper.  It's just that we seem to not enter
the merging machinery for single-file testcases so I chose to "fix"
lto option streaming rather than lto-wrapper ... I'll have a closer
look to lto-wrapper.  But to complain about target option mismatches
we'd have to call a target hook from lto-wrapper (or finally re-structure
all this and merge options in lto1 itself).

Richard.

> Honza
> > 
> > Thanks,
> > Richard.
> > 
> > 2015-04-23  Richard Biener  <rguenther@suse.de>
> > 
> > 	* lto-opts.c (lto_write_options): Do not record target options
> > 	if we are targeting a SWITCHABLE_TARGET target.  We use target
> > 	options on functions to transfer target flags from compile to
> > 	link time.
> > 
> > Index: gcc/lto-opts.c
> > ===================================================================
> > *** gcc/lto-opts.c	(revision 222360)
> > --- gcc/lto-opts.c	(working copy)
> > *************** lto_write_options (void)
> > *** 219,224 ****
> > --- 219,230 ----
> >   	  && lto_stream_offload_p)
> >          continue;
> >   
> > +       /* Do not store target-specific options if we target a
> > +          SWITCHABLE_TARGET target.  */
> > +       if ((cl_options[option->opt_index].flags & CL_TARGET)
> > + 	  && SWITCHABLE_TARGET)
> > + 	continue;
> > + 
> >         /* Drop options created from the gcc driver that will be rejected
> >   	 when passed on to the driver again.  */
> >         if (cl_options[option->opt_index].cl_reject_driver)
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)