Baslinux mailing list

["Anthony J. Albert" <albert AT polaris.umpi.maine.edu>]

On 15 Jan 2004 at 12:24, James Miller wrote:
>On Thu, 15 Jan 2004, Anthony J. Albert wrote:
>> I have tried to run compiles on a 4MB 386SX, with BL1.x installed on
>> the hard drive, and I was only able to get the simplest programs to
>> compile. gcc kept bombing out on anything much more complex than "Hello
>> World", until I added additional memory.
>
>Thanks for that information.  I do have the BL1 compiler to install, if I
>get adventurous.  Maybe with a math coprocessor compiling would work
>better? (just a guess: I don't really have a very clear idea of what a
>coprocessor actually does)  I've asked a friend if he might be able to dig
>one up from among his scrap heap.

Okay, here's what the math coprocessor is & does:

A math co-processor is a separate processing unit from the CPU which 
can do math much faster than the CPU - especially multiplies, divides, 
and floating-point math.  So, if the speed-up is enough, the CPU can 
send math instructions to the co-processor, and then receive the answer 
more quickly than the CPU can do the operation itself.

In the x86 line of processors, the math co-processor was initially a 
separate chip, then came to be combined on the die with the CPU. 
Usually, the designation of the co-processor is similar to the CPU, 
with a "7" instead of the 6.

So the 8086 has an 8087 math co-processor, the 80286 an 80287, and the 
80386 an 80387.  The 80486 had a math co-processor incorporated inside 
the package... but the 80486SX designation indicated a cheaper 
processor, without the math co-processor.  Some motherboards I've seen 
would allow you to use an 80387 with an 80486SX, to provide the co-
processor support.  From the Pentium on up, though, the math 
instructions are done directly on-chip, so no co-processor is needed... 
at least until the highly math-intensive graphics came around, so many 
video cards these days have a "DSP" chip, or Digital Signal Processor -
which could be considered a form of math co-processor.

Now, getting back to the original question: Compiling a program with a 
math co-processor, as compared to compiling it without, which is 
faster?  Well, the answer is going to be "it depends".

If the compilation process uses a lot of multiplies, divides, or 
floating point operations, then it can benefit from having the math co-
processor.  However, if it doesn't use a lot of those, then odds are 
that there will be no benefit.

It would be my guess, without testing, that something like a kernel 
compile would have a noticable speed-up with a math co-processor 
present, but that for compiling something like a single driver (usually 
only a few hundred KB) might not benefit noticably.

Hope this was informative,
Anthony Albert

===========================================================
Anthony J. Albert                     albert AT umpi.maine.edu
Systems and Software Support Specialist          Postmaster
Computer Services - University of Maine, Presque Isle
        "This is only temporary, unless it works."
                        --- Red Green