Re: Put scope blocks on a diet

[Jan Hubicka <jh at suse dot cz>]

> On Aug  6, 2007, Jan Hubicka <jh@suse.cz> wrote:
> 
> > I would like to test the attached patch.  It prevents the removal before
> > inlining so UIDs should remain the same.  However I do get massive
> > regressions on bootstrap-debug with or without the block removal code.
> > I will now rebuild and see if the set of failing files changes by 
> > disabling block removal, but if you do have some testcases handy, it
> > would help.
> 
> > I wonder if we need to keep those at all - dwarf2 already uses abstract
> > origin to avoid need to saveinfo on variables in each copy.  We can
> > probably just keep around the dead variables on separate list in their
> > ABSTRACT_ORIGIN -g or not?
> 
> This patch of yours (except for the ChangeLog entry) fixes the
> -g-changes-generated-code in the trunk, so this should go in.

Thanks for confirmation.  I was still planning to re-test it but didn't
get around it yet.
> 
> However, it breaks the vta branch, because we still keep debug
> annotations for variables in blocks that your patch removes.  I
> suppose I could just drop such annotations in the floor for now, but,
> in the long run, should I?


The pass is basically walking the function body looking for what scope
blocks are used (ie having some statements in or some variables declared
live).  If you have another annotations (I didn't looked into your code
yet), I guess you just need to update the marking stage to prevent them
from being removed if they still might end up being output to debug
info.

Honza
> 

> for gcc/ChangeLog.vta
> from  Jan Hubicka  <jh@suse.cz>
> 
> 	* tree-ssa-live.c (remove_unused_scope_block_p): Drop
> 	declarations and blocks only after inlining.
> 
> Index: gcc/tree-ssa-live.c
> ===================================================================
> --- gcc/tree-ssa-live.c	(revision 127183)
> +++ gcc/tree-ssa-live.c	(working copy)
> @@ -494,7 +495,11 @@ remove_unused_scope_block_p (tree scope)
>  	 only the used variables for cfgexpand's memory packing saving quite
>  	 a lot of memory.  */
>        else if (debug_info_level != DINFO_LEVEL_NORMAL
> -	       && debug_info_level != DINFO_LEVEL_VERBOSE)
> +	       && debug_info_level != DINFO_LEVEL_VERBOSE
> +	       /* Removing declarations before inlining is going to affect
> +		  DECL_UID that in turn is going to affect hashtables and
> +		  code generation.  */
> +	       && cfun->after_inlining)
>  	{
>  	  *t = TREE_CHAIN (*t);
>  	  next = t;

> 
> -- 
> Alexandre Oliva         http://www.lsd.ic.unicamp.br/~oliva/
> FSF Latin America Board Member         http://www.fsfla.org/
> Red Hat Compiler Engineer   aoliva@{redhat.com, gcc.gnu.org}
> Free Software Evangelist  oliva@{lsd.ic.unicamp.br, gnu.org}