From owner-cvs-etc  Tue Mar 21 05:57:59 1995
Return-Path: cvs-etc-owner
Received: (from majordom@localhost) by freefall.cdrom.com (8.6.10/8.6.6) id FAA20565 for cvs-etc-outgoing; Tue, 21 Mar 1995 05:57:59 -0800
Received: from time.cdrom.com (time.cdrom.com [192.216.223.46]) by freefall.cdrom.com (8.6.10/8.6.6) with ESMTP id FAA20558; Tue, 21 Mar 1995 05:57:56 -0800
Received: from godzilla.zeta.org.au (godzilla.zeta.org.au [203.2.228.34]) by time.cdrom.com (8.6.11/8.6.9) with ESMTP id FAA17088; Tue, 21 Mar 1995 05:57:40 -0800
Received: (from bde@localhost) by godzilla.zeta.org.au (8.6.9/8.6.9) id XAA16990; Tue, 21 Mar 1995 23:49:34 +1000
Date: Tue, 21 Mar 1995 23:49:34 +1000
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199503211349.XAA16990@godzilla.zeta.org.au>
To: phk@ref.tfs.com, pst@shockwave.com
Subject: Re: cvs commit: src/etc make.conf
Cc: CVS-commiters@time.cdrom.com, bde@zeta.org.au, cvs-etc@time.cdrom.com,
        jkh@freebsd.org, rgrimes@gndrsh.aac.dev.com
Sender: cvs-etc-owner@freebsd.org
Precedence: bulk

>  > We also need dynamic support for the i387 functions.  -DHAVE_FPU is no
>  > good because it can't be used for the distribution libraries.  Something
>  > like
>  > 
>  > 	if (_have_i387)
>  > 		result = _i387_pow(x, y);
>  > 	else
>  > 		result = __ieee754_pow(x, y);
>  > 
>  > would add less time overhead than shared linkage.

>The extra test on every operation is bad.

Let's replace `pow' by `sin'.  pow() isn't an i387 function and is too
complicated to synthesize from a few i387 functions.

To be precise, it costs 6 cycles on a 486 for the _i387_sin case and 5
cycles for the __ieee754_sin case (plus cache misses...)

>Xonsider the following fragment or high-speed linkages with shared libraries
>instead (I don't know how fast or slow shared linkages are):

Shared linkage costs 4 cycles (1 wasted for a stupidly placed nop and much
more for the first call; plus cache misses...).

>	static vec_pow = pow_init;

>	pow (base, exp)
>	{
>		return (*vec_pow)(base, exp);
>	}

This would only cost 2 cycle (plus cache misses...).

Anyone for self modifying code? :-)  The shared library already uses it
to avoid these 2 cycles and it might not be too hard to get the shared
library to patch in the addresses of the i387-specifice functions
instead of the generic one.  Unfortunately , this won't work for
statically linked programs.

The hardware sin() takes 193-279 cycles on a 486 and the msun wrappers
take many more (especially for shared libraries; position-independent
code costs about 10 cycles just for loading the global register), so
another 5 cycles would be hardly noticeable.

Bruce