What does #error do? - gcc

Question

Elsewhere (here on SO, for example) and even my own memory tells me that #error will cause compilation to terminate, but if I compile this:
#ifndef FOO
# error "FOO not defined."
#endif
int main() {
return "exit now!!";
}
I get:
g++ -Wall -o /dev/null main.cpp
main.cpp:2:4: error: #error "FOO not defined."
# error "FOO not defined."
^~~~~
main.cpp: In function 'int main()':
main.cpp:8:10: error: invalid conversion from 'const char*' to 'int' [-fpermissive]
return "exit!!";
^~~~~~~~
This is for g++ versions 5.4.0 (Ubuntu 5.4.0-6ubuntu1~16.04.4 20160609) and 7.0.0 (built myself 20161128). (There is a similar question asked here about MS Visual Studio 2015.)
According to the GNU CPP manual:
The directive ‘#error’ causes the preprocessor to report a fatal error. The tokens forming the rest of the line following ‘#error’ are used as the error message.
So, perhaps I just don't know what a "fatal error" is. The gcc option -Wfatal-errors would imply that a fatal error is one that would cause compilation to terminate, but nothing in the manuals ever defines it (unless I missed it). An answer to this question, however, is a bit more guarded and says that:
[#error] renders the translation unit ill-formed (i.e., it causes compilation to fail)
which I don't read as saying the same thing as saying compilation terminates at that point.
So, is this behaviour of gcc correct?

Answer 1

"Error" and "fatal error" were not separate categories in gcc until fairly recently. People would say "fatal error" just to make sure you understood that the compilation would not produce the expected object file. "warning" and "error" were always separate, but that wasn't obvious to everyone. You might think a warning is a type of error if you weren't told otherwise.
I don't know whether there was ever a version of gcc that actually terminated immediately after the #error, but I doubt it. It's possible that in the era when cpp was actually a separate program, it would have terminated at the end of the preprocessing phase without calling the compiler proper; that would have resulted in less error messages. (But not when -pipe was used to run them simultaneously!)
You could report a bug against the cpp manual. It's using "fatal error" in a way that is no longer consistent with the use of the term in the rest of the gcc manual.