[1441] in Athena Bugs
[scs@ADAM.PIKA.MIT.EDU: Re: bugs in shorts (sic) in vax pcc]
daemon@ATHENA.MIT.EDU (Joe Harrington)
Tue Nov 29 16:07:24 1988
Date: Tue, 29 Nov 88 16:06:17 EST
From: Joe Harrington <jh@SOL.MIT.EDU>
To: eichin@ATHENA.MIT.EDU, bugs@ATHENA.MIT.EDU
Cc: jh@ATHENA.MIT.EDU
Reply-To: jh@ATHENA.MIT.EDU
This is a followup from a bug with the vax C compiler discovered in
our lab. The bug was that shorts cast to doubles sometimes produced
junk. The bug exists in all athena vaxen running unix (5.5T, 6.0R) I
could get my hands on. It did not occur on the rt (6.0B using "cc")
or under gcc.
Below is a response from Chris Torek about the problem. Thanks go to
Steve Summit for helping me figure out the vax assembly.
% cat short_bug.c
#include <stdio.h>
/* This is a program to demonstrate in a very few lines how casting a double as
short is not done properly on this computer, with this compiler (sol, cc) */
main()
{
double d = 1.234567;
printf("(short) %f = %d\n", d, (short) d);
}
% short_bug
(short) 1.234567 = 2147418113
Date: Fri, 25 Nov 88 20:59:48 EST
From: scs@ADAM.PIKA.MIT.EDU (Steve Summit)
To: jh@ATHENA.MIT.EDU
From chris@mimsy.umd.edu Wed Nov 23 03:11:09 1988
Received: by ADAM.PIKA.MIT.EDU (5.45/4.7) id AA16339; Wed, 23 Nov 88 03:11:03 EST
Received: by mimsy.umd.edu (5.58/4.7)
id AA18855; Wed, 23 Nov 88 03:10:58 EST
Date: Wed, 23 Nov 88 03:10:58 EST
From: Chris Torek <chris@mimsy.umd.edu>
Message-Id: <8811230810.AA18855@mimsy.umd.edu>
To: scs@ADAM.PIKA.MIT.EDU
Subject: Re: bugs in shorts (sic) in vax pcc
This is fixed in 4.3tahoe: the compiler emits
cvtdw -8(fp),r0
cvtwl r0,-(sp)
<printf>
and
cvtdw -8(fp),r0
cvtwl r0,r0
cvtwl -10(fp),r1
cmpl r0,r1
jeql Lxxx
calls $0,_oops
The latter is far from optimal---
cvtdw -8(fp),r0
cmpw r0,-10(fp)
would do the trick---but is at least correct.
Incidentally, the `cvtwl' instruction on the Tahoe does not actually
convert anything. Apparently there were a number of related bugs in
pcc.tahoe.
Going over to okeeffe, it seems that the tahoe compiler gets it wrong
still, using the sequence
ldd -60(fp)
cvtdl -(sp)
<printf>
and
ldd -60(fp)
cvtdl r0
cvtwl -62(fp),r1
cmpl r0,r1
jeql Lxxx
callf $4,_oops
which compares the extended version of `s' against the truncated-to-
int-but-not-to-short version of `d'.
