Re: GCC

[Dave Hudson <dave at cyclicode dot net>]

Hi,

Martin wrote:
> It strikes me this is part of a larger problem.  It seems there is (in
> many companies / institutions there is a large divide between the people
> who design processors and the people who write compilers.

Well it's going a little off-topic, but this doesn't always happen.  The 
gcc port I've been optimizing for the Ubicom IP3000 series has been part 
of a joint development effort that has been optimizing the instruction 
set, the compiler and the operating system and protocol stacks that will 
be its primary function.  It's not perfect but we've learned from a few 
mistakes in the first ISA definition and the before we finish the second 
version we'll make sure we have a working compiler and simulator first :-)

I would note though that good as our gcc port is we have still ended up 
adopting several machine-specific optimizations where we can make 
opportunistic gains within the machine-dependent reorg pass.  Attempting 
to target these earlier pessimizes the code quite badly, but we have to 
use customized RTL-level optimizations because we may be replacing say 2 
or 3 instructions spread over perhaps 4 or 5 instructions in total. 
While this is at the more acceptable end of processor-specific 
optimization I do have some sympathy with folks who find that gcc can't 
quite do what they want at times.

There will always be situations where for very good reasons specific 
instructions will be added to an ISA but that may be hard to utilize 
well within gcc.  There may well be very good reasons why these 
instructions make sense though and arguing that a more pure-RISC 
strategy (one that makes it easy to target with compilers) will solve 
all of the world's problems isn't the right answer either ;-)

FWIW though I'd note that I've generally found that the most bizarre 
instructions do not originate with the processor designers but from 
operating system and application developers who demand magic 
instructions to save cycles in critical places.  These are usually also 
the same people who complain  when the compiler can't use their magic 
instructions very well :-)


Regards,
Dave