On Wednesday 21 June 2006 01:36, Valeriy E. Ushakov wrote:
> On Tue, Jun 20, 2006 at 00:38:30 +1000, Simon Burge wrote:
> > MIPS with gcc4 is having some problems generating bad code which we sort
> > of worked around but breaking gcc3 in the process.  Nick Hudson and I
> > have come up with the following which does away with the init_fallthru
> > and fini_fallthru functions and just puts _init and _fini directly in
> > the .init and .fini sections.
>
> I'm confused.  crti.c says:
>
> void    _init(void) __attribute__((section(".init")));
> void    _fini(void) __attribute__((section(".fini")));
>
> so these functions are already in .init and .fini respectively.

They are and they are used correctly as the .init/.fini section entry points. 

What is required is a way to generate a function prologue at the start 
of .init/.fini sections and a function epilogue at the end of the .init/.fini 
section. Any other .init/.fini code is run by being in the middle of these.

The function pro/epilogue are put there by careful placing of objects at link 
time. Any additional .init/.fini section code should in objects passed to the 
linker between start and end objects.

The way the code works now is _init/_fini calls init/fini_fallthru where 
init/fini_fallthru provide the funciton prologue. The function call from 
_init was confusing gcc4 on mips because gcc thinks _init_fallthru is a 
global, but the assembler declares it as a local. Changing _init_fallthru to 
a global in the assembler confuses gcc3.

The diff Simon posted does away with the _init/_fini call to 
init/fall_fallthru and generates a _init/_fini that only have the function 
prologue.

Clear as mud?

> SY, Uwe

Nick