Re: [PATCH 1/2] symtab: also change RTL decl name

[Jan Hubicka <hubicka@ucw.cz>]

> On Mon, 21 Nov 2022 20:02:49 +0100
> Jan Hubicka <hubicka@ucw.cz> wrote:
> 
> > > Hi Honza, Ping.
> > > Regtests cleanly for c,fortran,c++,ada,d,go,lto,objc,obj-c++
> > > Ok?
> > > I'd need this for attribute target_clones for the Fortran FE.  
> > Sorry for delay here.
> > > >  void
> > > > @@ -303,6 +301,10 @@ symbol_table::change_decl_assembler_name (tree decl, tree name)
> > > >  	warning (0, "%qD renamed after being referenced in assembly", decl);
> > > >  
> > > >        SET_DECL_ASSEMBLER_NAME (decl, name);
> > > > +      /* Set the new name in rtl.  */
> > > > +      if (DECL_RTL_SET_P (decl))
> > > > +	XSTR (XEXP (DECL_RTL (decl), 0), 0) = IDENTIFIER_POINTER (name);  
> > 
> > I am not quite sure how safe this is.  We generally produce DECL_RTL
> > when we produce assembly file.  So if DECL_RTL is set then we probably
> > already output the original function name and it is too late to change
> > it.
> 
> AFAICS we make_decl_rtl in the fortran FE in trans_function_start.

I see, it may be a relic of something that is no longer necessary.  Can
you see why one needs DECL_RTL so early?
> 
> > 
> > Also RTL is shared so changing it in-place is going to rewrite all the
> > existing RTL expressions using it.
> > 
> > Why the DECL_RTL is produced for function you want to rename?
> 
> I think the fortran FE sets it quite early when lowering a function.
> Later, when the ME creates the target_clones, it wants to rename the
> initial function to initial_fun.default for the default target.
> That's where the change_decl_assembler_name is called (only on the
> decl).
> But nobody changes the RTL name, so the ifunc (which should be the
> initial, unchanged name) is properly emitted but
> assemble_start_function uses the same, unchanged, initial fnname it
> later obtains by get_fnname_from_decl which fetches the (wrong) initial
> name where it should use the .default target name.
> See
> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/605081.html
> 
> I'm open to other suggestions to make this work in a different way, of
> course. Maybe we're missing some magic somewhere that might share the
> name between the fndecl and the RTL XSTR so the RTL is magically
> updated by that single SET_ECL_ASSEMBLER_NAME in
> change_decl_assembler_name? But i didn't quite see where that'd be?

I think we should start by understanding why Fortran FE produces
DECL_RTL early.  It was written before symbol table code emerged, it may
be simply an oversight I made while converting FE to symbol table.

Honza
> 
> thanks,
> 
> > Honza
> > > > +
> > > >        if (alias)
> > > >  	{
> > > >  	  IDENTIFIER_TRANSPARENT_ALIAS (name) = 1;  
> > >   
>