Re: Weird behavior of nested functions

[Jeroen Dekkers]

On Mon, Nov 10, 2003 at 06:44:35PM +0100, Marco Gerards wrote:
> Jeroen Dekkers <address@hidden> writes:
> 
> > On Mon, Nov 10, 2003 at 10:10:52AM +0100, Yoshinori K. Okuji wrote:
> > > On Monday 10 November 2003 00:11, Marco Gerards wrote:
> > > > Is it possible that some object files for the utils and pupa mix up?
> > > 
> > > No. The files in utils are compiled without -mrtd or -mregparm, because 
> > > otherwise they are incompatible against your C library. So it is 
> > > necessary to 
> > > compile files in PUPA in per-target basis.
> > > 
> > > BTW, are you good at reading assembly code generated by GCC? For this 
> > > kind of 
> > > problem, it is inevitable to take a look at what GCC produced. I myself 
> > > will 
> > > try your code as soon as possible.
> > 
> > It's a gcc bug triggered by the combination of -mregparm=3 and nested
> > functions. The problem is that it wants to put the address of test in
> > %ecx but that register is also used for the third argument with
> > -mregparm=3. Below is a copy of the assembly code generated by gcc.
> 
> How should this be fixed for PUPA?  I assume not supporting some gcc
> versions is no option. ;)
> 
> I see two options:
> 
> - Not using -mregparm=3 anymore.  This can be made dependent on the
>   gcc version with autoconf I assume.
> 
> - When using nested functions with more than 3 parameters an attribute
>   should be used for this function to disable regparm.  This can also
>   be dependent on the gcc version.
> 
> To me the second option seems the best.  Does anyone see other options
> or does someone have an preference?  If not I'll use option 2.

What about not using nested functions at all? They aren't really
necessary.

Jeroen Dekkers