Messages in this thread | | | Date | Sun, 2 Nov 1997 13:58:17 -0800 | From | Richard Henderson <> | Subject | Re: Linux on Merced |
| |
On Sat, Nov 01, 1997 at 07:19:04AM +0100, Andi Kleen wrote: > (Solaris is still completely 32bit - Linux is the only OS which > uses the 64bit features of the chip)
Well no, there is the HAL port of Solaris 2.5 (by/for Fujitsu?) -- but that's not what you're going to run on anything you get from Sun.
> I agree that it would be harder to port gcc to EPIC. It seems that it is > a much more radical design than even v9 sparc (very big register file, > VLIW, ...) AFAIK gcc hasn't been ported to a VLIW machine yet. It's > probably doable though.
I know of one VLIW port, though it will probably never be released, but that is sort of irrelevant. The contraints for ia64 are less strict than a true VLIW implementation, since ia64 _will_ be fully interlocked.
There are a number of simplifications that can be made to get GCC up on an EPIC implementation rather quickly.
For instance, treat the 64 predicate bits as 32 condition code pairs. Ignore predication until just before FINAL, when we convert jumps around small basic blocks to a sequence of predicated instructions. It wouldn't be as good as teaching the GCSE, FLOW and JUMP about predicates, but it wouldn't completely suck either.
The Haifa scheduler in EGCS already knows how to do speculative loads, so we can already take advantage of that. DaveM has experimented a bit with that already, since Sparc v9 provides non-faulting loads as well.
r~
| |