Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!lll-crg!ames!ucbcad!ucbvax!decvax!mcnc!duke!dukempd!cr
From: cr@dukempd.UUCP (Cyrus Rahman)
Newsgroups: comp.bugs.2bsd
Subject: Problem with multiplication of (long)'s in libc/crt.
Message-ID: <276@dukempd.UUCP>
Date: Fri, 28-Nov-86 04:44:33 EST
Article-I.D.: dukempd.276
Posted: Fri Nov 28 04:44:33 1986
Date-Received: Sat, 29-Nov-86 07:51:10 EST
Organization: Duke University Physics Dept.; Durham, N.C.
Lines: 44
Keywords: lmul.s long


	We have recently brought up 2.9bsd on our 11/70, and although
it has a number of bugs it is a generally good system.  It would be
nice if symbolic links really worked, and the quota feature fails in
a spectacular way.  But there is only one bug I am aware of now that
I need to fix quickly:

Subject: The libc/crt routines for hardware mult./divide don't work properly.

Index:	/usr/src/lib/c/crt 2.9bsd

Description:

	The routines run with the appropriate interrupt bits of the
	fpu disabled, so when overflow occurs during long multiplication the
	result is 0 instead of the truncated product.

Repeat-By:

	Run something that uses rand().  This program demonstrates the
	problem as well:

		#include <stdio.h>
		main() {
		long x;

		for (x = 2; x != 0; x *= 2 + 1)
			printf("x is %ld\n");
		}

	The program will terminate if linked with libc.a, but will behave
	properly if linked with libnonfpc.a.

Fix:
	The problem is not difficult to solve, but fixing it properly
	will require understanding how the fpu is managed by the
	kernel and what state C expects to find it in.

	If someone has already done this I would be grateful for
	their advice.
-- 
	Cyrus Rahman				1-919-684-8226
	Duke University Dept. of Physics	cr@dukempd.uucp
	Durham, N.C.      27706			mcnc!duke!dukempd!cr