NetNews Offline 2/NetNews Offline Volume 2.iso/news/comp/std/c/48

Internet Message Format | 1996-08-06 | 1.6 KB

Path: news.cs.ucla.edu!twinsun!not-for-mail
From: [e-mail] (Paul Eggert)
Newsgroups: comp.std.c
Subject: Re: Undefined result vs. int's holding undefined values.
Date: 7 Jan 1996 18:20:39 -0800
Organization: Twin Sun Inc, El Segundo, CA, USA
Message-ID: <4cpv1n$[e-mail]>
References: <4ck70b$[e-mail]> <4ckms5$[e-mail]>  <4cmg0s$[e-mail]> <[e-mail]>
NNTP-Posting-Host: der.twinsun.com

[e-mail] (Thad Smith) writes:

> it is possible to rewrite the function so that it doesn't
> invoke undefined behavior

Yes.  But it slows things down measurably for an interpreter to use
such a clumsy method for integer overflow detection.  The (x + y < x)
!= (y < 0) method requires a comparison, an exclusive OR of the
comparison result bit with a sign bit, and a conditional branch.
The method you suggested requires a comparison, a conditional branch, a
subtraction, another comparison, and another conditional branch.

Since this is comp.std.c, I'll mention that the (x + y < x) != (y < 0) method
and Daniel Wood's example both port to any C implementation
claiming conformance to LIA-1 (ISO/IEC 10967-1:1994(E),
`Information technology -- Language independent arithmetic -- Part 1:
Integer floating point arithmetic' <ftp://crl.dec.com/pub/misc/lia-1-dis.ps.Z>)
if INT_MODULO is true (which is the only plausible value for INT_MODULO
for real-world C compilers).  I doubt whether Wood's compiler vendor
claims LIA-1 conformance today, but this may become more important in
the future, and perhaps Wood can point his vendor at LIA-1 and say that
it's time to read the handwriting on the wall.