On Wed, Aug 28, 2013 at 4:16 PM, Antony Polukhin
2013/8/28 Andrey Semashev
One question: you're using strcmp to compare type_info names on some compilers to solve problems with cross-library comparisons. If symbol visibility is configured properly, cross-library comparisons should work properly, shouldn't they? Or are you aware of platforms that do not support cross-library comparisons at all? I'm asking because the list of such faulty platforms in your code looks overly long. E.g. on Linux type_info can be compared just fine, as long as the type info symbols have default visibility, and the comparison is faster than strcmp.
GCC's Standard library and some other libraries had an issue. They were assuming that if two dynamic libraries have same strings, then dynamic linker will load those symbols to the same address. That's why they were comparing types by pointers to type names instead of comparing type names. Unfortunately assumption is not right for some platforms (ARM) and linkers.
Can these particular configurations be detected at compile time? E.g. by testing for ARM target? Comparing symbol addresses works on x86 Linux, and I'd like to use it.
Oh yes, I'll need to check GCC once more. As I remember that issue was fixed in GCC 4.5 and it is safe to use default std::type_info comparison.
I have a suggestion about type_info_wrapper. How about I extract it from Boost.Log and then you add no-RTTI and strcmp workarounds to it? This way it will become equivalent to your type_index.