Re: [PATCH] More early LTO dwarf2out bits

[Richard Biener <rguenther at suse dot de>]

On Tue, 19 Sep 2017, Jakub Jelinek wrote:

> On Thu, Oct 13, 2016 at 09:16:07AM +0200, Richard Biener wrote:
> > 
> > This merges a few more bits guarding stuff with ! early_dwarf, mostly
> > to avoid creating locations that involve addresses of decls early
> > but also to avoid wasting work for BLOCK_NONLOCALIZED_VARs.
> > 
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, gdb testsuite
> > tested on the same arch (from the gdb 7.12 branch), applied to trunk.
> > 
> > Richard.
> > 
> > 2016-10-13  Richard Biener  <rguenther@suse.de>
> > 
> > 	* dwarf2out.c (tree_add_const_value_attribute): Do not try
> > 	rtl_for_decl_init during early phase.
> > 	(gen_variable_die): Do not create locations during early phase.
> > 	(gen_label_die): Likewise.
> > 	(decls_for_scope): Do not waste time handling BLOCK_NONLOCALIZED_VARs
> > 	twice.
> 
> As Martin reported, this has very undesirable effect on the size of DWARF
> debug info.
> 
> The problem is that add_const_value_attribute has lots of smarts to handle
> various kinds of constants, which the
>   if (CHAR_BIT == 8 && BITS_PER_UNIT == 8
>       && initializer_constant_valid_p (init, type))
> block doesn't implement, it relies on all the smarts to be done earlier.
> One option is to duplicate that all on trees, another option is
> to use the rtl_for_decl_init + add_const_value_attribute way
> even for early_dwarf for constants, something like:
> 
> -  if (! early_dwarf)
> +  if (! early_dwarf || CONSTANT_CLASS_P (init))
>      {
>        rtl = rtl_for_decl_init (init, type);
>        if (rtl)
>  	return add_const_value_attribute (die, rtl);
>      }
>    if (CHAR_BIT == 8 && BITS_PER_UNIT == 8
>        && initializer_constant_valid_p (init, type))

If we are sure that rtl_for_decl_init will never create symbol references
for CONSTANT_CLASS_P (init) that's fine, but yes, in the end it would be
nice to not have overlapping rtl/tree handling and have all constants
handled in the tree path...

Richard.

> For the duplicate on trees, some cases might not be that hard, like
> handling of INTEGER_CSTs, just see if they are negative and fit into shwi,
> (then add_AT_int), or fit into uhwi (then add_AT_unsigned); maybe that
> would be enough for now.
> 
> > --- gcc/dwarf2out.c	(revision 241022)
> > +++ gcc/dwarf2out.c	(working copy)
> > @@ -17958,12 +17958,15 @@ tree_add_const_value_attribute (dw_die_r
> >    init = t;
> >    gcc_assert (!DECL_P (init));
> >  
> > -  rtl = rtl_for_decl_init (init, type);
> > -  if (rtl)
> > -    return add_const_value_attribute (die, rtl);
> > +  if (! early_dwarf)
> > +    {
> > +      rtl = rtl_for_decl_init (init, type);
> > +      if (rtl)
> > +	return add_const_value_attribute (die, rtl);
> > +    }
> >    /* If the host and target are sane, try harder.  */
> > -  else if (CHAR_BIT == 8 && BITS_PER_UNIT == 8
> > -	   && initializer_constant_valid_p (init, type))
> > +  if (CHAR_BIT == 8 && BITS_PER_UNIT == 8
> > +      && initializer_constant_valid_p (init, type))
> >      {
> >        HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (init));
> >        if (size > 0 && (int) size == size)
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)