Add @gcsafe ccall

[kpamnany]

From GPUToolbox.jl.

[kpamnany]

@martinholters: looks like an upstream change has broken a test you added? Can you take a look please?

[martinholters]

Huh? AFAICT, the problem originates in

Compat.jl/test/runtests.jl

Lines 158 to 174 in 8819abe

	function _retrieve_exception_stack(;with_backtraces::Bool)
	exception_stack = try
	try
	# Generate the first exception:
	__not_a_binding__
	catch
	# Catch the first exception, and generate a second exception
	# during what would be handling of the first exception:
	1 ÷ 0
	end
	catch
	# Retrieve an ExceptionStack with both exceptions,
	# and bind `exception_stack` (at the top of this block) thereto:
	current_exceptions(;backtrace=with_backtraces)
	end
	return exception_stack
	end

which was added in #746 by @c42f and then modified by @Sacha0 in #758. But I can take a look non-the-less, of course. Presumable the error caused by accessing __not_a_binding__ has changed by the binding partitioning work? Maybe explicitly throwing an exception would be less brittle?

[kpamnany]

So sorry for the mistaken tag @martinholters! I didn't look too closely... maybe @Sacha0 can take a look at it?

[martinholters]

No worries. And it's just the "Suggestion: check for spelling errors or missing imports." that has been added to the exception printing at some point that causes the test failure. Fix at #836.

[martinholters]

Any chance to make this work on v1.6 and v1.7?

[kpamnany]

Any chance to make this work on v1.6 and v1.7?

@vchuravy would know... when he gets back and reviews this.

[vchuravy]

I don't see why this code should be lower bounded to 1.8?

[kpamnany]

On Julia 1.6:

gcsafe_ccall: Error During Test at /Users/kpamnany/Work/Compat.jl/test/runtests.jl:1086
  Test threw exception
  Expression: gc_safe_ccall() isa UInt64
  could not load symbol "jl_rand":
  dlsym(RTLD_DEFAULT, jl_rand): symbol not found
<snip>

[kpamnany]

Calling any Julia C function marked JL_NOTSAFEPOINT will serve, yeah?

[kpamnany]

Switched to calling jl_errno.

[KristofferC]

I got asked to review here so I give my personal opinion. I don't think Compat with a bunch of internals that tries to backport features from later Julia versions is a good idea in general. I don't think adding a Compat dependency (which transitively gives you dependencies on very large stdlibs) is very nice, especially when it is done on small packages that themselves have many other packages depending on them. I personally rather see some copy-pasting if something like this is really needed.

[kpamnany]

@KristofferC's points make sense to me. I can copy-paste this macro instead; that was also an alternative @vchuravy had suggested.