Date: Thu, 17 Jan 2008 16:39:07 +0000 (UTC) From: Bruce Evans <bde@FreeBSD.org> To: src-committers@FreeBSD.org, cvs-src@FreeBSD.org, cvs-all@FreeBSD.org Subject: cvs commit: src/lib/libc/amd64 _fpmath.h src/lib/libc/i386 _fpmath.h src/lib/libc/ia64 _fpmath.h Message-ID: <200801171639.m0HGd7pn089315@repoman.freebsd.org>
next in thread | raw e-mail | index | archive | help
bde 2008-01-17 16:39:07 UTC
FreeBSD src repository
Modified files:
lib/libc/amd64 _fpmath.h
lib/libc/i386 _fpmath.h
lib/libc/ia64 _fpmath.h
Log:
Add an alternative view of the bits in an 80-bit long double (64+16
instead of 32+32+15+1) on all arches that have such long doubles (amd64,
ia64 and i386). Large objects should be be accessed in large units,
and the 32+32+15+1[+padding] decomposition asks for almost the opposite
of that, sometimes resulting in very slow accesses depending on how
well the compiler ignores what we ask for and converts to the best
units for the given machine. E.g., on Athlons, there is a 10-20 cycle
penalty for accessing the middle 32-bit word immediately after an
80-bit store.
Whether actually using the alternative view is better is very machine-
dependent. A 32+32+16 view is probably best with old 32-bit systems
and gcc through 4.2.1. The compiler should mostly avoid the view and
generate best accesses, but gcc-4.2.1 is far from doing that. I think
64+16 is best for now. Similarly for doubles -- they should be using
64+0 especially on 64-bit machines, but fdlibm uses 32+32 extensively
for them. Fortunately, in 64-bit mode for doubles, gcc already ignores
the 32+32-bit view and generates best accesses in many cases.
Revision Changes Path
1.7 +6 -1 src/lib/libc/amd64/_fpmath.h
1.6 +6 -1 src/lib/libc/i386/_fpmath.h
1.7 +12 -1 src/lib/libc/ia64/_fpmath.h
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200801171639.m0HGd7pn089315>