Re: [Sbcl-devel] Replacing SIGTRAP by SIGILL

[<wil...@ai...>]

On Fri, Sep 14, 2007 at 03:39:40PM +0200, Andreas Bogk wrote:
> I recently spent several days trying to debug a problem in a FFI
> library, using gdb.  As you might know, this is less than fun, due to
> gdb's inability to process its own SIGTRAPs differently from SIGTRAPs
> raised by the debuggee.  And sbcl uses SIGTRAP quite a lot.
> 
> So I looked into ways to get rid of SIGTRAP usage in sbcl, to make my
> debugging life easier.  I found code for Darwin, which uses SIGILL
> instead of SIGTRAP, and an illegal instruction instead of "int 3" to
> trigger it.  Enabling that code on my platform (x86/Linux) was
> straightforward, worked pretty well, and provided me a much smoother
> debugging experience.
> 
> I suggest making this behaviour the default.  The cost is a two byte
> opcode instead of a one byte opcode for the trap instruction, which in
> total should be lost in the noise.  The win is much better debuggability
> with gdb.

I think it would be useful to have this in the sources, but I'm at
least mildly opposed to enabling it in the default build. I can see
its practical importance, because gdb is so ubiquitous. But in some
abstract sense it seems clearly the wrong thing --- something we would
never consider doing if gdb were less ill-behaved. And debugging FFI
code is a sufficiently hard-core niche that I'd expect almost anyone
doing that would be capable of rebuilding if necessary.

Even without esthetic wrong-thing arguments, if the 80kbyte figure
ISTR from IRC is correct, it would be about 0.3% of the core on my
machine. It don't think it's unreasonable to say that 0.3% is "lost in
the noise" --- but I also think that once you do, you should address
the possibility that the proportion of cores where the benefit will be
used might be lost in the noise.:-|

-- 
William Harold Newman <wil...@ai...>
PGP key fingerprint 85 CE 1C BA 79 8D 51 8C  B9 25 FB EE E0 C3 E5 7C
"gdb is the enemy of your enemy" -- Dan Barlow