forwarded message from Lal George
Stephen Weeks
MLton@sourcelight.com
Fri, 8 Dec 2000 13:05:40 -0800 (PST)
--DpCXMC9NIj
Content-Type: text/plain; charset=us-ascii
Content-Description: message body text
Content-Transfer-Encoding: 7bit
Just in case y'all aren't on the SML/NJ mailing list. Here was a pointer they
gave to their new x86 floating point generator.
http://cm.bell-labs.com/cm/cs/what/smlnj/compiler-notes/x86-fp.ps
--DpCXMC9NIj
Content-Type: message/rfc822
Content-Description: forwarded message
Content-Transfer-Encoding: 7bit
Received: from maguro.epr.com ([198.3.162.27]) by exchange.epr.com with SMTP (Microsoft Exchange Internet Mail Service Version 5.5.2650.21)
id WNKA3TZD; Fri, 8 Dec 2000 12:43:03 -0800
Received: from magrathea.epr.com (firewall-user@magrathea.epr.com [198.3.160.1])
by maguro.epr.com (8.9.3/8.9.3) with ESMTP id MAA15780
for <sweeks@intertrust.com>; Fri, 8 Dec 2000 12:47:01 -0800 (PST)
Received: (from uucp@localhost) by magrathea.epr.com (8.9.3/8.7.3) id MAA11902 for <sweeks@intertrust.com>; Fri, 8 Dec 2000 12:46:48 -0800 (PST)
Received: from nodnsquery(199.222.69.4) by magrathea.epr.com via smap (V5.5)
id xma011399; Fri, 8 Dec 00 12:46:05 -0800
Received: from dirty.research.bell-labs.com (dirty.research.bell-labs.com [204.178.16.6])
by mail.acm.org (8.9.3/8.9.3) with SMTP id PAA04826
for <sweeks@acm.org>; Fri, 8 Dec 2000 15:46:02 -0500
Received: from nslocum.cs.bell-labs.com ([135.104.8.38]) by dirty; Fri Dec 8 15:44:46 EST 2000
Received: (from george@localhost)
by nslocum.cs.bell-labs.com (8.9.3/8.9.3) id PAA29578744
for sweeks@acm.org; Fri, 8 Dec 2000 15:44:46 -0500 (EST)
Message-Id: <200012082044.PAA29578744@nslocum.cs.bell-labs.com>
From: Lal George <george@research.bell-labs.com>
To: sweeks@acm.org
Subject: SML/NJ: 110.31 NEWS
Date: Fri, 8 Dec 2000 15:44:46 -0500 (EST)
S M L / N J
1 1 0 . 3 1 N E W S
December 8, 2000
WARNING
This version is intended for compiler hackers. The
version ought to be stable, however we have not run
our full regression testing.
http://cm.bell-labs.com/cm/cs/what/smlnj/index.html
Summary:
o Socket related bug fixes.
o Improvements to CM autoloading.
o General cleanup in the use of CM libraries in the compiler.
o A new x86 fp compilation strategy.
o Removal of regmaps from MLRISC.
--------------------------------
Bug Fixes:
1514. sockets c-library broken
1582. SysErr exception connecting to socket
1585. getpeername in sockets
--------------------------------
CM:
Drastically improved link traversal code, resulting is faster load
times for CM and CMB.
Changed CM tool-plugin mechanism. See new manual
Made pickle-lib.cm and eliminated use of comp-lib.cm
--------------------------------
SML/NJ Library:
Fixed "where" clause to GraphSCCFn.
--------------------------------
MLRISC:
A. Intel X86 floating point:
As of 110.31, there is an alternative floating point code generator
and register allocator for the x86. Since this is still experimental,
by default this is turned off. To turn this on, do:
CM.autoload "$smlnj/compiler.cm";
Compiler.Control.MLRISC.getFlag "x86-fast-fp" := true;
The new floating point code generator treats the x86 fp stack as
7 registers, plus one temporary, and directly allocates floating point
values into these registers. Currently, fp parameter passing is still
done through memory, so the new code generator only benefits floating
point heavy loops. However, code compiled under the old and new
code generator can coexist.
The algorithm is described in:
http://cm.bell-labs.com/cm/cs/what/smlnj/compiler-notes/x86-fp.ps
Benchmarks:
We compared Version 110.30 compiling the PCLubIN entry in the
ICFP'00 programming context.
110.30 new fp Speedup
(ICFP00, PCLubIN)
chess.gml 22.16 20.98 5.63%
cone-fractal.gml 5.70 5.45 4.51%
cylinder.gml 1.61 1.58 2.28%
dice.gml 7.33 6.88 6.57%
ellipsoid.gml 1.35 1.30 4.16%
fov.gml 2.63 2.51 4.70%
fractal.gml 42.08 41.03 2.56%
golf.gml 3.09 2.95 4.75%
holes.gml 3.72 3.50 6.40%
house.gml 1.41 1.33 5.71%
intercyl.gml 3.02 2.78 8.41%
large.gml 8.01 7.81 2.64%
pipe.gml 6.35 5.78 10.01%
snowgoon.gml 4.70 4.31 8.95%
spheres.gml 1.26 1.17 6.98%
spotlight.gml 0.71 0.68 4.69%
By inlining Array2 in the same benchmark we get the following results:
110.30 new fp Speedup
chess.gml 21.85s 21.46s 1.83%
cone-fractal.gml 5.82s 5.47s 6.28%
cylinder.gml 1.57s 1.61s -2.85%
dice.gml 7.57s 6.85s 10.50%
ellipsoid.gml 1.33s 1.25s 6.74%
fov.gml 2.75s 2.57s 7.01%
fractal.gml 22.64s 21.52s 5.20%
golf.gml 3.04s 2.92s 4.25%
holes.gml 3.66s 3.48s 5.11%
house.gml 1.39s 1.29s 7.74%
intercyl.gml 3.00s 2.78s 7.91%
large.gml 7.91s 7.82s 1.13%
pipe.gml 6.44s 5.65s 13.98%
snowgoon.gml 4.75s 4.29s 10.53%
spheres.gml 1.22s 1.12s 8.36%
spotlight.gml 0.71s 0.68s 5.62%
Results from other benchmarks:
barnes-hut 1.714 1.696 1.0%
fft 0.954 0.906 5.2%
mandelbrot 19.91 14.99 32.8%
matrix-multiply(a) 47.77 45.81 4.3%
matrix-multiply(b) 17.04 15.42 10.5%
simple 3.02 2.69 12.3%
tsp 1.75 1.656 5.6%
NOTE: Matrix multiply(b) has all bounds checking removed.
Each test is run 10 times and I take the average.
Overall, the numbers do not improve as much as I was hoping, except for
mandelbrot. The following benchmarks compare smlnj with mlton and C:
mandelbrot fft barnes-hut
sml/nj 110.30 19.91 0.96 1.71
sml/nj new fp 14.99 0.90 1.71
gcc -O 14.83
gcc -O6 14.01 0.68
mlton -O6 17.46 1.04 1.62
(version 200000906)
B. Internal 'regmap' Changes:
1. Changed interface to CELLS and the type of cell, cellkind, cellset etc.
2. No more regmaps!! The attributes of cells, including its current color,
are accessible from CELLS interface. Cells can now take arbitrary
annotations. [They will also have a width attribute in the next
go around.]
3. Interface of STREAM etc have changed (again, no more regmap).
4. Some MLTREE constructors, like, IF, BCC, JMP, CALL, etc have
been simplified. CVTI2I has been renamed into SX and ZX respectively,
following the lambda rtl convention.
5. The old RA interface was getting too complicated. There are now
two functors, RISC_RA (in ra/risc_ra.sml) and X86RA (in x86/ra/x86RA.sml)
which abstract out from all the ugly business. The first is for
RISC machines, and the second is for x86. Please let us know if you
use these functors.
6. The cell change broke the peephole phases, because they used to
pattern match on specific cell number. I (Allen) hacked up a simple
tool to translate fake SML with where clauses into real ML. This makes
it much easier to write the rules. Seems to work. (See Tools/WhereGen)
--DpCXMC9NIj--