__STDC_VERSION__ check in C++ code

[0-wiz-0]

Havard Eidnes added a patch to the pkgsrc package for scipy with the following comment:

Add a patch which fixes an obviously bogus preprocessor conditional;
in our case, __STDC_VERSION__ isn't defined when built as C++.
The fix isn't completely right, it insists on <fenv.h> if built as C++.
Not entirely unreasonable, and makes this build on NetBSD/powerpc as well,
which doesn't like the redefinition of fegetround() and fesetround().

I don't know the proper fix.

Here's his patch:

--- scipy/special/_round.h.orig 2018-05-05 17:10:11.000000000 +0000
+++ scipy/special/_round.h
@@ -49,7 +53,7 @@ double add_round_down(double a, double b
 
 
 /* Helper code for testing _round.h. */
-#if __STDC_VERSION__ >= 199901L
+#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || defined(__cplusplus)
 /* We have C99 */
 #include <fenv.h>
 #else

[sturlamolden]

The C++ standard say that definition of __STDC_VERSION__ is implementation dependent. I see that we have cases of C include files that use this macro and are imported in C++, e.g. scipy/special/_round.h and scipy/_lib/_c99compat.h, and probably others.

First, it is not en error to use __STDC_VERSION__ in C++ because the standard allows this macro to be defined. We just have to accept that it will not build on some arcane systems.

Second, we can of course check for __cplusplus and define __STDC_VERSION__ if needed. However, the question is then to decide which value is appropriate, which would also depend on the arcane C++ implementation in question. We would then need preprocessor pragmas for any thinkable implementation, i.e. worst common denominator programming. SciPy is not going to be like OpenSSL if I have a say. Let ill behaving systems break on build-time. That is far better than introducing potential run-time errors that will materialize according to Murphy's law. Say we define __STDC_VERSION__ but accidentally give it the wrong value for some arcane system we don't know about. Then we can just sit back and wait for the next Ariane 5 incident to happen.

[rgommers]

Somehow I managed to miss or ignore this issue for the last 4 years, apologies. The patch looks fine to me, and there's no problem with including a simple one line change. It's indeed a bug to unconditionally use __STDC_VERSION__.

Included in gh-16551.