Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!shadooby!accuvax.nwu.edu!tank!ncar!unmvax!pprg.unm.edu!hc!lll-winken!uunet!portal!cup.portal.com!Don_A_Corbitt
From: Don_A_Corbitt@cup.portal.com
Newsgroups: comp.arch
Subject: Re: Query about miserable M68882 performance [really 80-bit not
Message-ID: <16733@cup.portal.com>
Date: 6 Apr 89 01:16:00 GMT
References: <2583@tank.uchicago.edu> <7829@pyr.gatech.EDU> <16543@winchester.mips.COM>
Organization: The Portal System (TM)
Lines: 18

[request for info on intermediate results kept in FP chip]
Trivial answer for Turbo C using 80x87.  It keeps intermediate results of a 
single expression in the '87 as 80 bit values.  Also, all float/double/long 
double functions actually return an 80 bit value in the FP top of stack.

This was a hassle for my Numerical Analysis class, where they wanted us to
show the difference between 32 and 64 bit arithmetic (had to store each 
intermediate result to a memory variable, and reload through an aliased
pointer :->)

Seems to work fine.  Code is simpler (stack-based instructions), results are 
'better' (although numerical analysis people have complained - see above).
Of course, the 387 supposedly runs faster when one of the operands is in
_memory_ as opposed to being in a register!  (Is this because it knows a
memory op is only 32 bits long, so it doesn't need to do the entire 80 bit
operation?  Confusion reigns)
	Don Corbitt, occasional grad student
	Don_A_Corbitt@cup.portal.com