[9fans] 8c vs. gcc shootout

[9fans] 8c vs. gcc shootout

[anothy]

i know little of benchmarking, but i'm not clear on how we
could build this test and have the results really be very
meaningfull. that is, how does one account for things like
the differing header files? would using, say, Inferno's lib9.h
and friends in both cases be reasonable? how 'bout the
underlying OS's differences, like syscall time or differences
in the scheduler? are they considered negligable (if we're
dealing with times on the order of a minute, i'd certainly
hope so)? should we be building something to run on the
raw hardware?

i've got Plan 9 and FreeBSD w/ GCC running on identical
hardware. if someone could provide me with (or point me
at) suitable code to run the compile and execute tests on,
i'd be happy to do so.

how significant are the penalties imposed by using APE
likely to be? in one respect, it'd be nice to use identical
code, but then GCC has the advantage of running it its
native environment, while 8c has to emulate.
ア

Re: [9fans] 8c vs. gcc shootout

[Thomas Bushnell, BSG]

anothy@cosym.net writes:

> i've got Plan 9 and FreeBSD w/ GCC running on identical
> hardware. if someone could provide me with (or point me
> at) suitable code to run the compile and execute tests on,
> i'd be happy to do so.

The problem is that you want to control not just for header files but
for the general time the system takes.

I think it's patently obvious that the Plan 9 kernel itself is much
more tightly coded than either BSD or Linux.

It would be *even better* if the Plan 9 kernel were being compiled
with an optimizing compiler.

So the real test of compile speed is:

Compile one system with 8c, and then use the 8c compiler.

Compile the *same* system with GCC -O3, and then use GCC with no -O.

Thomas

Re: [9fans] 8c vs. gcc shootout

[forsyth]

>>It would be *even better* if the Plan 9 kernel were being compiled
>>with an optimizing compiler.

it didn't make any appreciable difference when that was tried years ago (not by us)
with Inferno, compiled with gcc on various platforms.
it didn't make any appreciable difference with my own O/S on the sun3 and i386,
and i suspect that's because in all those cases there was little code on
which heavy-duty optimisation makes much difference.  there is a fair amount
of function calling but relatively little crunching.  the main crunching that's
done is data movement or graphics and both are optimised separately using
non-compiler-specific techniques.