Subject: the recently-posted floating-point performance program
To: None <>
From: Tim Chase <>
List: current-users
Date: 12/07/1993 16:48:30
Regarding the "bad floating-point performance" program that was
recently posted to this list:  Assuming both machines had
floating-point hardware, it's important to note that the normal "-lm"
library on NetBSD is reaaaaaaal slow (compared to the x87 intrinsics).
If you want better performance, you need to build the one using the
sources (recently renamed to .S) in /usr/src/lib/libm/i387.  Also, you
can hack the compiler's md file to make the compiler use intrinsics
(and can also use the -ffast-math compiler flag) to speed it up the

						- Tim Chase

PS:  Has anyone thought of making a target in the stock
/usr/src/lib/libm/Makefile to build a 387-using version of the library
which could be installed as libm387.a? As it sits, you've gotta change
the SRCS & stuff and fiddle with the common stuff to get a working x387
library.  One other thing... as it sits now, the .S.o rule doesn't work
for the i387 .S files.  They also need preprocessing.