Re: [avr-libc-dev] PROGMEM section variables throw away by --gc-sections?

[Dean]

----- Original Message ----- 
From: "Rolf Ebert" <address@hidden>
To: "Dean" <address@hidden>
Cc: <address@hidden>
Sent: Wednesday, November 14, 2007 5:13 PM
Subject: Re: [avr-libc-dev] PROGMEM section variables throw away 
by --gc-sections?


> Dean schrieb:
> > Hi guys,
> >
> > I've been noticing some strange behaviour with my makefile arguments.
> >
> > Firstly, LD.exe crashes when I use -mshort-calls with the --relax linker 
> > option, but *only* if I *don't* use -ffunction-sections 
> > and --gc-sections. With the latter two options enabled both -mshort-calls 
> > and --relax work just fine.
> >
> > Secondly, -ffunction-sections and --gc-sections seems to throw away some 
> > of my PROGMEM variables. I rely on a few PROGMEM variables being embedded 
> > into the binary so I can later distinguish which version it was compiled 
> > from, similar to:
> >
> > char PROGMEM buildtime[] = __TIME__;
> >
> > --gc-sections seems to remove this, although I thought it was only 
> > designed to remove unused functions and not PROGMEM data. Is there a 
> > workaround to this?
>
> --gc_sections removes all sections, that the linker thinks are unused. Did 
> you compile also with the -fdata-sections compiler switch?

No, didn't try that. I was under the hazy impression that -fdata-sections 
didn't work for the AVR target? I'll give it a try.

> The assembler symbol of buildtime is not modified by -ffunction-sections, 
> only by -fdata-sections.
>
> I can also imagine a missing KEEP instruction in the linker script
>
> Rolf

This is all getting wierder. It's only triggered by --relax, --gc-sections 
and -mshort-calls on my large projects - test cases don't work. 
Incidentally, removing -ffunction-sections causes my USB demo apps to 
misbehave, even when --gc-sections isn't enabled. I though 
that -ffunction-sections essentially does nothing without --gc-sections? I 
tried to compare the listings and found some very minor differences, but 
nothing to indicate why the firmware screwed up when -ffunction-sections was 
turned off. Removing --relax fixes the problem and allows me to freely 
turn -ffunction-sections on and off without faulty firmware.

I've no idea what's going on, but there's obviously some sort of confict (if 
not multiple ones) between -mshort-calls, -ffunction-sections, --relax 
and --gc-sections. It only seems to affect my large complex project, not 
simple ones I've tested it on. Wierd.