Very Computer

Hi,

I'm running on a

SunOS enigma 5.8 Generic_108528-12 sun4u sparc SUNW,Ultra-5_10

Using cc: Sun WorkShop 6 update 2 C 5.3 2001/05/15,
I build a 32 bit version (no special compile flags except for -O)
and a 64 bit version ( -xtarget=ultra -xarch=v9 -dalign ).

I then build a 32 bit executable and a 64 bit executable and run them.

Running on the same machine, my 64 bit executable is approximately 20%
SLOWER and I'm trying to get a handle of why this should be the case.
The computations being done are fairly floating point intensive.

How is that the emulated 32 bit program runs faster than the 64 bit
version?

Thanks,

Ted

That's possible.

Quote:> How is that the emulated 32 bit program runs faster than the 64 bit
> version?

What do you mean by "emulated" there?  Why would anything be emulated in
your examples?

It is possible to construct a program that uses loops that will fit into
the on-CPU chip cache with 32-bit pointers, but will not fit into the
cache when it uses 64-bit pointers (simply because the size of the code
comprising the loop is now larger).  When this happens, it may cause a
severe performance reduction.

--

Senior Technical Consultant         TAOS            http://www.taos.com/
Got some Dr Pepper?                           San Francisco, CA bay area
         < This line left intentionally blank to confuse you. >

First things first: use the same flags, where possible.  Use -O
for both, and add -xarch=v9 for the 64-bit build.  If you use
xtarget and/or dalign, use it for both builds.

WHat results do you get when you recompile them like this?

Quote:> How is that the emulated 32 bit program runs faster than the 64 bit
> version?

It's not an emulated program.  The UltraSPARC processor runs 32-bit
AND 64-bit code natively.

In the 64-bit programming model (called LP64), longs and pointers
are 64 bits wide, rather than 32.  This means that less of them
will fit into the cache, having a performance impact.

HTH,

--
Rich Teer, SCNA, SCSA

President,
Rite Online Inc.

Voice: +1 (250) 979-1638
URL: http://www.rite-online.net

Remember that with 64 bit opcodes your instruction word is twice as
large.  A 64 bit chip simply has to do more work to get things done.
So what's the advantage?  Bandwidth.  When dealing with large
datasets, 64 bit apps really shine.  A (possibly bad) analogy:

A corvette can travel at 150mph and carry 100 pounds of cement at a
time
A Mack truck can travel 80mph and carry 1000 pounds of cement at a
time

If you're moving small amounts of data, 32 bit will be faster.  If you
need to move large amounts of data, a 64 bit app can do it faster.

I'm sure I've messed up a few points and I'm sure someone will correct
them.

Check this thread:
http://groups.google.com/groups?hl=en&lr=&ie=UTF-8&oe=UTF-8&threadm=3...

SPARCv9 instructions are still only 32 bits wide.

--
Wishing you good fortune,
--Robin Kay-- (komadori)

..

Quote:

> I'm sure I've messed up a few points and I'm sure someone will correct
> them.

The point about opcode lengths is wrong, for one. SPARC opcodes are 4
bytes long, period.

--
Tony