[llvm-commits] AddressSanitizer, first patch

[Chandler Carruth]

FYI,

I've looked again at this patch recently, and talked to several others about
it...

I think there are some big issues with the implementation, bit I think they
could probably be addressed. The one that I see the most is the x86 assembly
just being dumped as a blob. I think that really needs to be resolved before
this can move forward.

It seems like this should be do-able in some way with the llvm.trap
intrinsic. If anything, we might need another intrinsic or an extension of
llvm.trap that allows feeding the SIGILL signal handler the
data arguments it needs; I'd need to read up on exactly what is required for
that to figure out how best to represent that in IR, but it shouldn't be
that hard.

That said, maybe to improve the discussion of the patch and make reviewing
it easier, you could just switch to emitting the call to the runtime library
function in all cases? That should be a strictly simpler patch, and then how
to improve performance by avoiding the runtime call can be a followup.

Maybe others could comment on other implementation issues? I'm sadly not an
expert at LLVM...
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20110908/ff04cfb6/attachment.html>