Re: MATH_EMULATE

["Perry E. Metzger" <perry%piermont.com@localhost>]

Vit Herman <v.herman%sh.cvut.cz@localhost> writes:
> On Mon, 14 Jan 2008, Perry E. Metzger wrote:
>> Is this an actual 386?
>
> Yes.

FYI, those are not supported in NetBSD 4.0...

>> Hrm. I note that ldd claims that you are linking against the correct
>> libm, so sqrt should in fact be handled by the library and not by the
>> non-existent FPU. This machine is a machine where ping actually
>> crashes on sqrt when you try to run it?
>
> It runs just fine until the final phase (after ^Cing it), then crashes
> with the error previously reported.

I know you reported that happening before, but that happens on this
particular machine?

So the deal is roughly this: there are floating point ops going on all
the time as ping runs, so MATH_EMULATE is clearly working at least to
some extent. Then we hit the end, when you kill ping, and at that
point the standard deviation gets calculated using a sqrt call. That
*should* be calling a pure software implementation that doesn't trap
into the kernel at all, but apparently something is going wrong. The
question is what is going wrong and where.

My guess is that something really stupid is happening here that we
simply didn't notice from lack of testing.

-- 
Perry E. Metzger                perry%piermont.com@localhost