Thu, 21 Oct 2004 02:47:29 +0000
Newsgroups: comp.lang.lisp Subject: Re: CMU CL or SBCL on (FreeBSD/)AMD64 any time soon?
Gareth McCaughan <email@example.com> writes:
> But I don't want to find myself without a decent zippy > Lisp implementation. Several different people seem to have > done work on making SBCL run on AMD64 in the last N months > (generally for Linux rather than FreeBSD); what's the > current state of the world?
Since acquiring an amd64 machine of my own last weekend I've dusted off the SBCL work I was doing earlier in the year (many thanks to Kevin Rosenberg for ssh access to his machines, but it's a lot more comfortable developing locally than over the internet, not to mention that the appeal is so much greater now I own a machine that it'll run on ...)
You asked about "current state" rather than "progress", which I like, becuase otherwise I'd have only have been able to say "it'll be done when it's done". So:
If you're familiar with the SBCL build process I can tell you that when I left off last it was generating plausible-looking cross-compiled core files that got maybe two thirds of the way through the cold-init process (at the end of which is a repl). After that all that remains is building PCL, which is allegedly portable, and the contrib modules, which are mostly actually portable if we limit ourselves to considering sbcl-on-different-architectures.
After picking it up again on Sunday, I broke it utterly by deciding to rewrite the floating point support to use sse2/sse3/xmm (faster, more predictable, less weird) instead of the legacy x87 instructions. That work is mostly complete at least in that it's now back to the point of generating plausible cross-compiled cores, but (whether due to new FP or to having simultaneously forward-ported the branch by several months) it dies at about the second function call. This could be considered a regression.
- 64-bit lisp word (yes, big cores)
- 61 bits of fixnum.
- uses all 16 registers
- sse for floating point (16 xmm registers)
Big known issue between here and the goal: bignum support, which is still based on 32-bit words. Other than that it's mostly a question of seeing what crops up.
I don't know much about the FreeBSD amd64 port, but unless they've don something really weird I can't imagine it being much work to get SBCL up on once we have a working compiler and runtime.
As others have mentioned, the x86 port of SBCL (at least on Linux) runs fine[*] under a 64 bit kernel if furnished with appropriate shared libraries. On my machine it takes around 9-10 minutes to build itself including all contribs: I don't know objectively how zippy this is, but it's about a sevenfold improvement on the box it replaced, so I at least am content for the moment.
[*] Actually, fine except for threads. But if you're on freebsd you don't get them anyway, so I mention that mostly for completeness' sake