Re: [PATCH] Allocate all target globals using GC for SWITCHABLE_TARGETs

[Jakub Jelinek <jakub at redhat dot com>]

On Thu, Jan 09, 2014 at 08:25:31AM -0800, Richard Henderson wrote:
> > +  p = (struct target_globals_extra *)
> > +      ggc_internal_cleared_alloc_stat (sizeof (struct target_globals_extra)
> > +				       PASS_MEM_STAT);
> > +  g = (struct target_globals *) p;
> > +  g->flag_state = &p->flag_state;
> > +  g->regs = &p->regs;
> >    g->rtl = ggc_alloc_cleared_target_rtl ();
> 
> So, we're relying on something pointing to G, thus keeping the whole P alive?
> I suppose that works but it's fairly ugly that's for sure.

The separate structures aren't really installed individually, they are
always installed together through restore_target_globals.  As long as the
any FUNCTION_DECL with such TARGET_OPTION_NODE exists, it will be reachable.

The reason why it needs to be GC is:
1) in two of these target_* structures there are embedded rtxes etc. the GC
needs to see
2) if all FUNCTION_DECL with such combination of target attributes are GCed,
we'd leak memory

> As for the extra ~500k wasted on x86_64, we can either fix our gc allocator to
> do something sensible with these high-order allocations, or we can do nearly
> this same trick only with libc.  I.e.
> 
>   struct target_globals_extra {
>     struct target_flag_state flag_state;
>     struct target_regs regs;
>     struct target_hard_regs hard_regs;
>     struct target_reload reload;
>     struct target_expmed expmed;
>     struct target_optabs optabs;
>     struct target_cfgloop cfgloop;
>     struct target_ira ira;
>     struct target_ira_int ira_int;
>     struct target_lra_int lra_int;
>     struct target_builtins builtins;
>     struct target_gcse gcse;
>     struct target_bb_reorder bb_reorder;
>     struct target_lower_subreg lower_subreg;
>   } *p;
> 
>   g = ggc_alloc_target_globals ();
>   p = XCNEW (target_globals_extra);
>   ...

That would be fine for 1), but would mean 2).  It is also fine to GC
allocate each structure individually, but some (like bb_reorder) are say
just 4 bytes long, so it might be overkill.

As noted by Richard S., IRA/LRA still puts pointers to heap allocated
objects into some of the structures, so there is some leak anyway, but not
as big.

	Jakub