Add at-sigatomic macro (closes #12309)

[kmsquire]

cc: @stevengj @JeffBezanson @vtjnash

[JeffBezanson]

Should be $(esc(code)).

[kmsquire]

Fixed, thanks.

[vtjnash]

i think this should encourage use of the lambda form, rather than adding additional syntax:

help?> disable_sigint
search: disable_sigint

Base.disable_sigint(f::Function)

   Disable Ctrl-C handler during execution of a function, for calling
   external code that is not interrupt safe. Intended to be called
   using "do" block syntax as follows:

      disable_sigint() do
          # interrupt-unsafe code
          ...
      end

[kmsquire]

i think this should encourage use of the lambda form, rather than adding additional syntax

Any suggestions on how it should do that?

I find this slightly cleaner for one line ccalls, and I'm generally not a fan of the lambda forms because of an preconceived notion that they might be inefficient (because, well, they use lambdas).

Any reason why the lambda forms aren't used in GTK.jl?

(See also the discussion in #12309)

[vtjnash]

Gtk.jl has complicated needs due to COPY_STACKS interfering with event loop integration (prior to that it was using the lambda form of disable_sigint).

the syntax does look pretty sensible for ccall, although there's another experiment in making it zero-cost that I want to try first before adding this

inlining can (in the future) handle the lambda transformation to make it equivalent to this, so i'm not too convinced by that argument

[kmsquire]

inlining can (in the future) handle the lambda transformation to make it equivalent to this, so i'm not too convinced by that argument

Okay. There's been talk about various kinds of function inlining for a few years now. I'm sure we're a lot closer than we were in 2012, but you should be able to understand why I would make that statement. ;-)

Looking forward to your zero-cost solution.

[yuyichao]

Now that we have jb/functions, is this still needed?

[kmsquire]

I argued above that this is cleaner for ccalls. @vtjnash said he had a zero-cost experiment he wanted to try first, and neither of us followed up.

[vtjnash]

no, you can close this with your branch yyc/threads/safepoint-signal

[yuyichao]

Yep, the zero-cost approach is #16174 and I was doing my pre-PR opening search for related issues/prs.

FWIW, if there's still good reason to prefer a macro with #16174 and jb/functions, the try-catch will not be necessary anymore.

[kmsquire]

#16174 looks nice! I look forward to trying that with VideoIO sometime (although that package isn't in a very good state right now).

You can definitely close this and #12309 with the merge of #16174.