Re: Avoid unnecesary GGC runs during LTO

[Richard Biener <rguenther at suse dot de>]

On Fri, 11 Apr 2014, Jan Hubicka wrote:

> 
> Hi,
> while looking into -ftime-report, I noticed that ggc can take up to 10% of WPA memory
> while it does almost nothing: it is run just after streaming that explicitly
> frees memory that becomes unreachable.  The first GGC run usually saves at
> most 1% of memory and then it is never run again.
> I believe this ought to also help in case we get into swap, since ltranses will
> also ggc less.
> 
> Bootstrapped/regtested x86_64-linux, OK?
> 
> Honza
> 
> 	* lto.c (read_cgraph_and_symbols): Grow ggc memory after streaming.
> 	* ggc.h (ggc_grow): New function.
> 	* ggc-none.c (ggc_grow): New function.
> 	* ggc-page.c (ggc_grow): Likewise.
> Index: ggc.h
> ===================================================================
> --- ggc.h	(revision 209170)
> +++ ggc.h	(working copy)
> @@ -225,6 +225,9 @@ extern const char *ggc_alloc_string_stat
>     function is called, not during allocations.  */
>  extern void ggc_collect	(void);
>  
> +/* Assume that all GGC memory is reachable and grow the limits for next collection. */
> +extern void ggc_grow (void);
> +
>  /* Register an additional root table.  This can be useful for some
>     plugins.  Does nothing if the passed pointer is NULL. */
>  extern void ggc_register_root_tab (const struct ggc_root_tab *);
> Index: lto/lto.c
> ===================================================================
> --- lto/lto.c	(revision 209170)
> +++ lto/lto.c	(working copy)
> @@ -2999,6 +3000,10 @@ read_cgraph_and_symbols (unsigned nfiles
>    gimple_canonical_types = NULL;
>    delete canonical_type_hash_cache;
>    canonical_type_hash_cache = NULL;
> +
> +  /* At this stage we know that majority of GGC memory is reachable.  
> +     Growing the limits prevents unnecesary invocation of GGC.  */
> +  ggc_grow ();
>    ggc_collect ();

Isn't the collect here pointless?  I see not in ENABLE_CHECKING, but
shouldn't this be abstracted away, thus call ggc_collect from ggc_grow?
Or maybe rather even for ENABLE_CHECKING adjust G.allocated_last_gc
and simply drop the ggc_collect above ().

Anyway, this is sth for stage1 at this point.

Thanks,
Richard.

>    /* Set the hooks so that all of the ipa passes can read in their data.  */
> Index: ggc-none.c
> ===================================================================
> --- ggc-none.c	(revision 209170)
> +++ ggc-none.c	(working copy)
> @@ -63,3 +63,8 @@ ggc_free (void *p)
>  {
>    free (p);
>  }
> +
> +void
> +ggc_grow (void)
> +{
> +}
> Index: ggc-page.c
> ===================================================================
> --- ggc-page.c	(revision 209170)
> +++ ggc-page.c	(working copy)
> @@ -2095,6 +2095,19 @@ ggc_collect (void)
>      fprintf (G.debug_file, "END COLLECTING\n");
>  }
>  
> +/* Assume that all GGC memory is reachable and grow the limits for next collection. */
> +
> +void
> +ggc_grow (void)
> +{
> +#ifndef ENABLE_CHECKING
> +  G.allocated_last_gc = MAX (G.allocated_last_gc,
> +			     G.allocated);
> +#endif
> +  if (!quiet_flag)
> +    fprintf (stderr, " {GC start %luk} ", (unsigned long) G.allocated / 1024);
> +}
> +
>  /* Print allocation statistics.  */
>  #define SCALE(x) ((unsigned long) ((x) < 1024*10 \
>  		  ? (x) \
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer