Re: [Sbcl-devel] cleaning up the runtime

[William H. N. <wil...@ai...>]

On Tue, Jan 29, 2002 at 04:13:49PM +0000, Daniel Barlow wrote:
> Here are some of the things I'm currently speculating about.  Would
> patches for these be accepted?
>
> 1) Make greater use of $arch-$os-os.c: any function in $arch-arch.c whose
> contents are decided entirely or substantially by a #ifdef some_os 
> (e.g. arch_set_pseudo_atomic_interrupted in alpha-arch.c), or vice
> versa anything in $os-os.c bracketed entirely by #ifdef some_arch
> (like sigsegv_handler in linux-os.c) is a candidate for being moved
> into $arch-$os-os.c.  {arch,os}_install_interrupt_handlers, too
> 
> 2) Add some machine-independent way to set the fpu status, ideally using
> the same definitions as the Lisp side does (so, they need to be
> exported in sbcl.h)
> 
> (I did consider making the FPU settings avilable only to C and having
> the Lisp code do FFI calls to get at it, but that seems incredibly
> wasteful for something which typically only really wants to set a register)
> 
> 3) Linux 2.0 has been available for as long as anyone can remember: we
> may as well retire the (broken, anyway) mmap-in-16M-chunks and ask for
> our dynamic spaces all in one chunk.
> 
> 4) (more ambitious) have some lisp->c translation code generate the
> data structures that monitor.c and all the related ldb stuff, so it
> doesn't need to be maintained in parallel by hand

All of these sound good to me.

In the case of 1, I'd also remark that when the natural minimal #ifdef
doesn't surround the entire construct (where by construct I mean the
thing you'd move to another file, a function definition or structure
definition or whatever) then I'd also encourage patches which separate
out the smaller architecture-dependent piece, then provide
per-architecture definitions of the architecture-dependent piece.

I was going to use "struct call_info" as an example, but on reflection
I think my first impulse there would be to try to make all architectures
use what the Alpha currently uses, machine pointers like
  struct call_frame *frame;
instead of pointers punned into fixed-size words like
  u32 frame
So maybe a better example would be memory_fault_handler() in 
bsd-os.c, where instead of writing FreeBSD and OpenBSD variants
of memory_fault_handler(), I'd suggest rewriting it as
  static void
  memory_fault_handler(int signal, siginfo_t *siginfo, void *void_context)
  {
      /* This is the part that is #ifdef'ed in the current sources. */
      void fault_addr = os_siginfo_fault_address(siginfo);

      /* This part is system-independent. */
      if (!gencgc_handle_wp_violation(fault_addr)) {
  	interrupt_handle_now(signal, siginfo, void_context);
      }
  }
and then implementing FreeBSD and OpenBSD variants of
os_siginfo_fault_address().

> And feel free to add more ideas as you think of them ...

The lack of Lisp->C translation of layouts and other information, as
in point 4, is a special case of general failure of the code to
propagate layout information cleanly. Any cleanups which help move
toward the ideal where a change in a layout in objdef.lisp would just
work everywhere would be great. (As per the KLUDGE notes at the head
of objdef.lisp, the failure to propagate changes isn't just a failure
to propagate them into ldb, but a failure to propagate them to vital
mainstream non-debugger parts of the system too.)

-- 
William Harold Newman <wil...@ai...>
"Look on my works, ye Mighty, and despair!" -- Ozymandias, King of Kings 
PGP key fingerprint 85 CE 1C BA 79 8D 51 8C  B9 25 FB EE E0 C3 E5 7C