From: gsw AT aloft DOT att DOT com
Date: Fri, 16 Jul 93 14:18:09 EDT
Original-From: aloft!gsw (gerald.s.williams)
To: djgpp AT sun DOT soe DOT clarkson DOT edu
Subject: Re: How do I use emu387 only when necessary?

Jeremy Mathers <pynq AT midway DOT uchicago DOT edu> writes:

> I remember asking this question a long time ago - in the context of
> "What will happen if I run a DJGPP compiled program that does floating
> point on a 486 DX, but happen to have the emu field set in the GO32 var?"
> 
> At that time, I was told (by who I cannot remember) that it would use
> the chip (i.e., would "do the right thing").  Was I misled?

I don't know what GO32 v1.10 does, but v1.09 definitely uses the EMU387
emulator code in that case.

In fact, we are using this as a work-around for the case where our GCC
cross-compiler does constant folding causing unhandled floating point
exceptions, which causes our cross-compiler to go belly-up (we are
currently using DJGPP to compile the compiler).

Actually, this is an indication of a problem with GO32.  Code running
under the emulator does not work exactly the same way as code using
floating point hardware, since unhandled coprocessor exceptions (we've
seen it with divide by zero) cause GO32 to fail.  In our case, it is
unacceptible for our cross-compiler to abort in response to any input
file, so this is a critical problem.  I would expect that divide by
zero exceptions should cause an infinity to be generated, since I've
seen that method used, at least when IEEE floating point arithmetic is
being used.

,-----------------.
|Gerald S Williams|
|gsw AT aloft DOT att DOT com|
|  (215)439-7237  |
`-----------------'

      _  |     ____/    _  |
     /   /    /        /   /      /                         /
    /   /  ____ |     ____/   _  _/     __ |     __ |      /       ____/
   /   /        /    /          /      /   /    /   /     /     ____
______/  ______/  __/         ___/   _____/   _____/    ___/   ______/