Sturla Molden wrote:
The advantage is that you get C code that don't have to do error checking everywhere.
Really, this is a big disadvantage. In C, you have to check error codes. Anything which gives the illusion you can get away with it is wrong IMHO. Technically, setjmp/longjmp are also problematic for many reasons (often the same as C++ exceptions: it is difficult to get right). At least with goto you can often understand what's wrong - not so much with longjmp/setjmp - if the error handling is too complicated to handle with simple goto, then the code itself is too complicated. If you need to handle deeply nested errors, then using the C API for python exception is good enough. I agree we are not as good as we should in numpy for error handling at the C level, but I think it boils down to not having a unified error handling (that is -3 may mean could not allocate in one function, and could not convert the argument to something in another). But that can be solved with a unified set of error codes and associated strings + a simple log system (that's how sndfile does it for example, and I really like it). cheers, David