Re: [Sbcl-devel] (scale-float 1.0 most-positive-fixnum) doesn't fail on freebsd

[Andreas F. <as...@bo...>]

On 2004-07-19, Christophe Rhodes <cs...@ca...> wrote:
> Andreas Fuchs <as...@bo...> writes:
>> debugger invoked on a FLOATING-POINT-OVERFLOW in thread 87635:
>>   arithmetic error FLOATING-POINT-OVERFLOW signalled
>> Operation was SCALE-FLOAT, operands (1.0 536870911).
>
> I'm betting the floating point modes are incorrectly being reset
> after taking the ABORT or TOPLEVEL restarts from the debugger.  What
> happens on the second overflowing call to scale-float?

Further calls to scale-float do the same as the first one; You were
close, though. Both forms work correctly in a fresh SBCL image. It
only fails to load when float.pure.lisp is loaded; I found these three
forms to cause the problems I described above (of which the last two
are the more interesting):

(defmacro raises-error? (form &optional (error-subtype-spec 'error))
  `(typep (nth-value 1 (ignore-errors ,form)) ',error-subtype-spec))

(ignore-errors (funcall (fdefinition 'float-radix) "notfloat"))

(assert (raises-error? (scale-float 1.0 most-positive-fixnum)
		       floating-point-overflow))

;;;;

another interesting thing is that (float-radix "notfloat") will not be
sufficient to trigger the bug; you have to funcall the fdefinition or
the symbol 'float-radix, otherwise it appears to work correctly (and
scale-float raises the error).

Hm,
-- 
Andreas Fuchs, <as...@bo...>, as...@ja..., antifuchs