Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!uunet!tut.cis.ohio-state.edu!elephant.cis.ohio-state.edu!weide
From: weide@elephant.cis.ohio-state.edu (Bruce Weide)
Newsgroups: comp.software-eng
Subject: Re: Reusability considered harmful??(!!)
Message-ID: <88431@tut.cis.ohio-state.edu>
Date: 14 Feb 91 21:41:02 GMT
References: <6108@stpstn.UUCP> <87829@tut.cis.ohio-state.edu> <6248@stpstn.UUCP> <1991Feb11.103100.112@skyler.mavd.honeywell.com>
Sender: news@tut.cis.ohio-state.edu
Reply-To: Bruce Weide <weide@cis.ohio-state.edu>
Organization: Ohio State University Computer and Information Science
Lines: 41

In article <1991Feb11.103100.112@skyler.mavd.honeywell.com>
djbailey@skyler.mavd.honeywell.com writes:
>
>A sine routine ought to be a very good candidate for a reusable
>component, but we got bit by a subtle problem.  This is one case where
>the algorithm selected was very important and I am interested in how a
>"specification and object code" system would handle it.
>

Handling this problem (a sine routine that isn't too accurate for some
particular range of values) in a "specification and object code" reuse
scenario is based on the idea that a single abstract concept can have
multiple plug-compatible implementations.  Bill Ogden's recent note
about the specification of such an abstraction as a sine routine is
important.  The specification is relational, i.e., it doesn't say that
the routine must return the mathematical sine of its argument; it says
that the value returned is within some bound of the true sine.  (The
allowable error may even be a parameter to the reusable component that
exports the sine routine.  Call it a tolerance if you want :-).)

Now there are many possible implementations of a sine function that
satisfy this behavioral specification.  Any of them should work.
However, some may be more or less efficient, in one sense or another.
Perhaps some are faster for arguments near zero, others for arguments
away from zero, or whatever.  It is not necessarily true that one
implementation dominates the others in all relevant performance
dimensions.

If I sell you the behavioral specification along with object code (and
preformance information for each implementation) -- but NOT source
code -- you still have all the information you need in your example.
Of course, I should also guarantee that the object code works, but
that will be needed in any case.  Will you trust me if I say that,
even if I don't provide source code?  (Possibly not... do you trust,
say, the Macintosh toolbox?  Seems pretty robust to me, and I've never
seen the source code for it.)  But certainly in principle there is no
reason to believe that your own methods of certifying correctness of
implementations should be any better than those of, say, an
independent laboratory that offers its official seal of approval.

	-Bruce