Re: [PATCH] tree-optimization/109304 - properly handle instrumented aliases

[Jan Hubicka <hubicka@ucw.cz>]

> On Tue, 4 Apr 2023, Jan Hubicka wrote:
> 
> > > On Tue, 28 Mar 2023, Richard Biener wrote:
> > > 
> > > > When adjusting calls to reflect instrumentation we failed to handle
> > > > calls to aliases since they appear to have no body.  Instead resort
> > > > to symtab node availability.  The patch also avoids touching
> > > > internal function calls in a more obvious way (builtins might
> > > > have a body available).
> > > > 
> > > > profiledbootstrap & regtest running on x86_64-unknown-linux-gnu.
> > > > 
> > > > Honza - does this look OK?
> > > > 	PR tree-optimization/109304
> > > > 	* tree-profile.cc (tree_profiling): Use symtab node
> > > > 	availability to decide whether to skip adjusting calls.
> > > > 	Do not adjust calls to internal functions.
> > > > @@ -842,12 +842,15 @@ tree_profiling (void)
> > > >  	    for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
> > > >  	      {
> > > >  		gcall *call = dyn_cast <gcall *> (gsi_stmt (gsi));
> > > > -		if (!call)
> > > > +		if (!call || gimple_call_internal_p (call))
> > > >  		  continue;
> > > >  
> > > >  		/* We do not clear pure/const on decls without body.  */
> > > >  		tree fndecl = gimple_call_fndecl (call);
> > > > -		if (fndecl && !gimple_has_body_p (fndecl))
> > > > +		cgraph_node *callee;
> > > > +		if (fndecl
> > > > +		    && (callee = cgraph_node::get (fndecl))
> > > > +		    && callee->get_availability (node) == AVAIL_NOT_AVAILABLE)
> > 
> > As discussed earlier, the testcase I posted can be adjusted to put the
> > const declared wrapper into another translation unit, so I think we will
> > need to drop the visibility check completely.  But as discussed, it is
> > wrong code issue, but not a regression, so we may go with the
> > availability check as you suggest. So the patch is OK. 
> > 
> > 
> > I wonder if we do not want to drop it everywhere (as we plan for next
> > stage1 anyway).  I think similar ICE as in the PR can be produced with
> > LTO. In normal situation declaration merging will do the right thing:
> > If you have unit A calling const foo externally, it won't get processed
> > by the code above.  However unit B declaring foo will get it downgraded
> > to non-const.
> > 
> > Now at WPA time we will read both A and B and in declaration merging B's
> > definition will prevail.  This won't happen if lto_symtab_merge_p
> > returns false which can probably be triggered by adding warning/error
> > attribute to B's declaration but not to A's.
> > 
> > It is however really side case and I am worried about dropping
> > pure/const from builtin declarations...
> 
> Yeah, that's what I'm worried about as well.  I guess we'd need to
> do the demotion to non-const/pure at WPA time and have a flag
> in the cgraph node indicating instrument_add_{read,write}?  But
> then how should we deal with C++ comdats instrumented in one TU
> but not in another?  Does this mean we should do instrumentation
> at IPA time instead of as small IPA pass before IPA?

I do not think LTO is of any help here.  You can allways call non-LTO
const function from outer-world and that function can will end up
calling back to instrumented const function in your unit which
effectively makes the extenral const function non-const.
> 
> That said, when there's a definition of say strlen in a TU and
> that's instrumented we do want to drop pure from calls but if
> not then we shouldn't worry.
> 
> Without LTO we'd still run into coverage issues but at least
> with LTO we shouldn't ICE?

I am not sure I see your point here...
We could avoid demoting builtins to avoid ICEs and have coverage
mismathces, but how LTO makes difference?

Honza
> 
> Richard.