Re: -mregparm=3 (was Re: [PATCH] i386 do_machine_check() is redundant.

[Daniel Jacobowitz <dan@debian.org>]

On Mon, Sep 29, 2003 at 10:20:45PM +0200, Mikulas Patocka wrote:
> > > > Use machine_check_vector in the entry code instead.
> > >
> > > This is wrong. You just lost the "asmlinkage" thing, which means that it
> > > breaks when asmlinkage matters.
> > >
> > > And yes, asmlinkage _can_ matter, even on x86. It disasbles regparm, for
> > > one thing, so it makes a huge difference if the kernel is compiled with
> > > -mregparm=3 (which used to work, and which I'd love to do, but gcc has
> > > often been a tad fragile).
> >
> > gcc 3.2 and later are supposed to be ok (eg during 3.2 development a
> > long standing bug with regparm was fixed and now is believed to work)...
> > since our makefiles check gcc version already... this can be made gcc
> > version dependent as well for sure..
> 
> They are still buggy. gcc 3.3.1 miscompiles itself with -mregparm=3
> (without -O or -O2 it works). (I am too lazy to spend several days trying
> to find exactly which function in gcc was miscompiled, maybe I do it one
> day). gcc 2.95.3 compiles gcc 3.3.1 with -mregparm=3 -O2 correctly.
> gcc 3.4 doesn't seem to be better.
> 
> gcc 2.7.2.3 has totally broken -mregparm=3, even quite simple programs
> fail.

You can't build GCC with -mregparm=3.  It changes the interface to
system functions.  So unless your libc happened to be built with
-mregparm=3, and extensively hacked to expect arguments in registers to
the assembly stubs, it can't work.

It's interesting for kernel code, whole distributions, or things which
are careful to have a glue layer.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer