Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!seismo!sundc!pitstop!sun!oliveb!felix!zemon
From: stephent@tekchips.TEK.COM (Steve Tyler)
Newsgroups: comp.unix.ultrix
Subject: Why Ultrix C library is slow
Message-ID: <9433@felix.UUCP>
Date: Tue, 13-Oct-87 19:38:41 EDT
Article-I.D.: felix.9433
Posted: Tue Oct 13 19:38:41 1987
Date-Received: Fri, 16-Oct-87 04:58:24 EDT
Sender: zemon@felix.UUCP
Organization: Tektronix Inc., Beaverton, Or.
Lines: 21
Approved: zemon@felix.UUCP

Reply-Path:


Some code compiled with the Ultrix C compiler runs 40% slower than
when compiled with the 4.3BSD C compiler.

I discovered this by comparing execution profiles of the same code compiled
with the two compilers.  In the Ultrix version, frexp(), which is a C library
routine that separates a floating point number into mantissa and exponent,
was the single most time-consuming routine.  By disassembling frexp() in the
compiled code with adb, I found that the Ultrix version is 27 instructions and
the 4.3BSD version is only 5 instructions.

For many C library routines there is a C code version and an assembly language
version.  Clearly DEC is using the slower, larger C code version of frexp().
This version computes the exponent by repetitively dividing its argument by 2
until it is less than 1 (or multiplying, if the argument already is less than
one).  The number of divides (or multiplies) is the exponent and the resulting
double is the mantissa.

The assembly language version simply extracts the appropriate bit fields from
the argument to get the exponent and mantissa.  Of course this is not portable
while the C code is, but does Ultrix run on anything besides DEC/VAX hardware
anyway?