Re: [Patch,AVR]: PR18145: do_copy_data & do_clear_bss only if needed

[Georg-Johann Lay <avr at gjlay dot de>]

Weddington, Eric schrieb:
> > -----Original Message-----
> > From: Richard Henderson [mailto:rth@redhat.com]
> > Sent: Tuesday, April 19, 2011 1:31 PM
> > To: Georg-Johann Lay
> > Cc: gcc-patches@gcc.gnu.org; Weddington, Eric; Denis Chertykov; Anatoly
> > Sokolov
> > Subject: Re: [Patch,AVR]: PR18145: do_copy_data & do_clear_bss only if
> > needed
> >
> > On 04/18/2011 10:20 AM, Georg-Johann Lay wrote:
> >
> > > +avr_asm_named_section (const char *name, unsigned int flags, tree decl)
> > > +{
> > > +  if (!avr_need_copy_data_p)
> > > +    avr_need_copy_data_p = ((0 == strncmp (name, ".data", 5)
> > > +                             || 0 == strncmp (name, ".rodata", 7)
> > > +                             || 0 == strncmp (name, ".gnu.linkonce.",
> >
> > 14)));
> >
> > > +
> > > +  if (!avr_need_clear_bss_p)
> > > +    avr_need_clear_bss_p = (0 == strncmp (name, ".bss", 4));
> >
> > Have a look at FLAGS.  I expect that you can reference those to
> > categorize the data rather than hard-coding the section names.
> > In particular I believe your ".gnu.linkonce" test is wrong.
> >
> > It's not clear to me what you're looking for wrt "data".

> We're looking for the existence of global initialized data variables.
>
> We have 2 small subroutines in our libgcc, one to set everything in
> .bss to zero, and another to copy the initializations from .text (in
> flash) to the RAM based variables in .data. These subroutines have
> always been included in the startup code whether they were needed or
> not. The purpose of this patch is to check whether these subroutines
> are actually needed or not (i.e. if anything actually exists in .bss,
> or in .data). If either of these sections are actually empty, then we
> don't want to link in the corresponding subroutine, which helps us
> save some code space on small devices.
>
> AFAICT, I've never seen the avr target generate anything in .rodata
> and .gnu.linkonce sections. I think these might be holdovers from
> some other target code and/or there for some completeness reason, I'm
> not really sure.

When I wrote this patch I looked at the default linker script to see 
what goes into .data resp .bss; the hard-coded section maned reflect 
that. For the linkonce stuff I found no explanation (grepping the web 
yields bunch of questions from people having trouble with it) and in the 
gcc wiki you find nothing (except you have already a link to what you seek).

The .gnu.linkonce is an overestimation, .gnu.linkonce.d. is perhaps 
better, but exceptions are not supported on avr, anyway. So an 
overestimation won't hurt.

Then I wonder where .noinit, .progmem.data, .progmem.gcc_sw_table
would go?

Moreover, with checking section names it might be easier for users that 
want some private sections treated as cleared by __do_clear_bss or 
initialized by __do_copy_data.

AFAIK the section attribute does not allow to set section flags?

Eric, you don't see .rodata because of avr.c:avr_asm_init_sections
 readonly_data_section = data_section;
To see .rodata give -fdata-sections for const not in progmem.

The only thing I am a bit uncomfortable with is that there is no command 
line option to unconditionally reference the __do's.

Johann