Re: [MIT-Scheme-devel] Ubuntu 11.04 and FE_DFL_ENV.

[Matt Birkholz]

> From: Taylor R Campbell <address@hidden>
> Date: Sat, 11 Jun 2011 16:48:10 +0000
> 
> [...]
>    The second patch implements BORKED_FENV.
> 
> That this is necessary is a bit surprising.

Let me sing it for you, Brother.  I "fixed" this easily enough and
verified, via debugging traces and fegetexcept() in gdb, that ALL
exceptions were masked.  Yet gdb said I was trapping in Pango, looking
for a font or sump'n.

I worried about uninitialized struct members, that other thread, etc,
etc.  I compared my gtk-event-viewer to its straight-C cousin gtkev,
which (oh, the terror!) could mask and unmask exceptions, set AND
clear them... reliably!!  I eventually got down on my hands and knees
and looked at the bloody native bloody floating-point bloody
instructions.  I am flabbergasted that it started working with (just!)
those two patches.

> Can you set a breakpoint in gdb on alienate_float_environment, both
> with libc's fe* and with Scheme's fe*, and step through the machine
> instructions to compare what they do differently?

Of course!  Can't we all?  :-)

> I looked at the glibc source code, and I see no substantial
> difference beyond fnstcw vs fstcw in fedisableexcept (which shouldnt
> make a difference here).

There is also the sse_p issue, x87_clear_exceptions (fnclex) versus
x87_read/write_environment (fnstenv&fldcw/fldenv), and all that stuff
I don't know anything about, like T7500 (Core2Duo) fpu bugs.

I pounded Chaos into a corner.  I rest now.

How about someone else verifying that I am not hallucinating?  Perhaps
I am just the unfortunate victim of a heat issue in my old notebook.
One must never underestimate The Great Adversary!

> [...]  I guess the only way to test this would be to allocate a
> stack large enough that it doesn't fit in the low 4 GB of the
> virtual address space, which would fall outside the domain of our
> automated testing facilities...

I don't remember which f instruction trapped.  I just assumed they all
would have a problem with esp. :-)

The (patched) emulation worked well on Ubuntu 11.04 amd64 too, BTW,
yet (surprise!) is no longer needed there. ?