On Fri, 13 Feb 2004, Jun Sun wrote:
> > > 2. It changes inline-assembly function prologues to be embedded within the
> > > functions, which makes them a bit safer as they can now explicitly refer
> > > to the "regs" struct and assures the code won't be removed or reordered.
> >
> > It is possible that gcc changes one of the registers before save_static
I think it is possible, too, but it doesn't happen now as gcc tries not
to change static callee-saved registers if possible as that's expensive.
Only changes to "s8" seem inevitable if the frame pointer is used, but for
MIPS the kernel is always built with "-fomit-frame-pointer", so the
restriction doesn't apply.
> > and I can't imagine there's a reliable way to fix this in the inline
> > version.
>
> Yes. I still remember this bug vividly. It took me quite a few days
> to track it down.
The patch makes the code safer than what we have now, but it would still
need to be verified periodically.
> I really wish there is a more reliable and systematic way to do this,
> even at some expense of a few more instructions ...
If we want to tolerate performance loss, then it's easily doable. That
can be done with the current setup, with a jump instruction to the
referred function added at the end and "__attribute__((used))" or perhaps
"asm("foo")" added to the function declaration.
I can choose this path if we agree on it.
Maciej
--
+ Maciej W. Rozycki, Technical University of Gdansk, Poland +
+--------------------------------------------------------------+
+ e-mail: macro@ds2.pg.gda.pl, PGP key available +
|