[tex-k] mf segfaults if compiled with -g3 (bug in mfmpi386.asm)

[Olaf Weber]

Karl Berry writes:

> Hi Ed,
>     Hope I'm sending this to the right address 

> Yes.

>     and that this makes sense.

> It sounds right to me.  (Olaf?)

> I'm not sure we really need that assembly code at all, these days.  I
> wonder if compiler output would be better ...

It is clear we don't exactly _need_ the code, but it _is_ subtantially
better than the compiler-generated code, because the default
implementation is complicated by being restricted to 32 bit arithmetic
and the other alternative uses floating point.

A proper alternative is to check for 64-bit 'long' or 'long long' and
do the calculation in that type if available, falling back on the
floating-point in other cases.  But some benchmarking is necessary to
see what difference this makes, and in which direction: the code with
64bit integer arithmetic might be more compact yet lose on those
platforms where fp is substantially faster than integer arithmetic.

-- 
Olaf Weber

               (This space left blank for technical reasons.)