JezK
Edit File: greenlet_compiler_compat.hpp
/* -*- indent-tabs-mode: nil; tab-width: 4; -*- */ #ifndef GREENLET_COMPILER_COMPAT_HPP #define GREENLET_COMPILER_COMPAT_HPP /** * Definitions to aid with compatibility with different compilers. * * .. caution:: Use extreme care with G_NOEXCEPT. * Some compilers and runtimes, specifically gcc/libgcc/libstdc++ on * Linux, implement stack unwinding by throwing an uncatchable * exception, one that specifically does not appear to be an active * exception to the rest of the runtime. If this happens while we're in a G_NOEXCEPT function, * we have violated our dynamic exception contract, and so the runtime * will call std::terminate(), which kills the process with the * unhelpful message "terminate called without an active exception". * * This has happened in this scenario: A background thread is running * a greenlet that has made a native call and released the GIL. * Meanwhile, the main thread finishes and starts shutting down the * interpreter. When the background thread is scheduled again and * attempts to obtain the GIL, it notices that the interpreter is * exiting and calls ``pthread_exit()``. This in turn starts to unwind * the stack by throwing that exception. But we had the ``PyCall`` * functions annotated as G_NOEXCEPT, so the runtime terminated us. * * #2 0x00007fab26fec2b7 in std::terminate() () from /lib/x86_64-linux-gnu/libstdc++.so.6 * #3 0x00007fab26febb3c in __gxx_personality_v0 () from /lib/x86_64-linux-gnu/libstdc++.so.6 * #4 0x00007fab26f34de6 in ?? () from /lib/x86_64-linux-gnu/libgcc_s.so.1 * #6 0x00007fab276a34c6 in __GI___pthread_unwind at ./nptl/unwind.c:130 * #7 0x00007fab2769bd3a in __do_cancel () at ../sysdeps/nptl/pthreadP.h:280 * #8 __GI___pthread_exit (value=value@entry=0x0) at ./nptl/pthread_exit.c:36 * #9 0x000000000052e567 in PyThread_exit_thread () at ../Python/thread_pthread.h:370 * #10 0x00000000004d60b5 in take_gil at ../Python/ceval_gil.h:224 * #11 0x00000000004d65f9 in PyEval_RestoreThread at ../Python/ceval.c:467 * #12 0x000000000060cce3 in setipaddr at ../Modules/socketmodule.c:1203 * #13 0x00000000006101cd in socket_gethostbyname */ /* The compiler used for Python 2.7 on Windows doesn't include either stdint.h or cstdint.h. Nor does it understand nullptr or have std::shared_ptr. = delete, etc Sigh. */ #if defined(_MSC_VER) && _MSC_VER <= 1500 typedef unsigned long long uint64_t; typedef signed long long int64_t; typedef unsigned int uint32_t; // C++ defines NULL to be 0, which is ambiguous // with an integer in certain cases, and won't autoconvert to a // pointer in other cases. #define nullptr NULL #define G_HAS_METHOD_DELETE 0 // Use G_EXPLICIT_OP as the prefix for operator methods // that should be explicit. Old MSVC doesn't support explicit operator // methods. #define G_EXPLICIT_OP #define G_NOEXCEPT throw() // This version doesn't support "objects with internal linkage" // in non-type template arguments. Translation: function pointer // template arguments cannot be for static functions. #define G_FP_TMPL_STATIC #else // Newer, reasonable compilers implementing C++11 or so. #include <cstdint> #define G_HAS_METHOD_DELETE 1 #define G_EXPLICIT_OP explicit #define G_NOEXCEPT noexcept # if defined(__clang__) # define G_FP_TMPL_STATIC static # else // GCC has no problem allowing static function pointers, but emits // tons of warnings about "whose type uses the anonymous namespace [-Wsubobject-linkage]" # define G_FP_TMPL_STATIC # endif #endif #if G_HAS_METHOD_DELETE == 1 # define G_NO_COPIES_OF_CLS(Cls) private: \ Cls(const Cls& other) = delete; \ Cls& operator=(const Cls& other) = delete # define G_NO_ASSIGNMENT_OF_CLS(Cls) private: \ Cls& operator=(const Cls& other) = delete # define G_NO_COPY_CONSTRUCTOR_OF_CLS(Cls) private: \ Cls(const Cls& other) = delete; #else # define G_NO_COPIES_OF_CLS(Cls) private: \ Cls(const Cls& other); \ Cls& operator=(const Cls& other) # define G_NO_ASSIGNMENT_OF_CLS(Cls) private: \ Cls& operator=(const Cls& other) # define G_NO_COPY_CONSTRUCTOR_OF_CLS(Cls) private: \ Cls(const Cls& other); #endif // CAUTION: MSVC is stupidly picky: // // "The compiler ignores, without warning, any __declspec keywords // placed after * or & and in front of the variable identifier in a // declaration." // (https://docs.microsoft.com/en-us/cpp/cpp/declspec?view=msvc-160) // // So pointer return types must be handled differently (because of the // trailing *), or you get inscrutable compiler warnings like "error // C2059: syntax error: ''" #if defined(__GNUC__) || defined(__clang__) /* We used to check for GCC 4+ or 3.4+, but those compilers are laughably out of date. Just assume they support it. */ # define GREENLET_NOINLINE_SUPPORTED # define GREENLET_NOINLINE(name) __attribute__((noinline)) name # define GREENLET_NOINLINE_P(rtype, name) rtype __attribute__((noinline)) name # define UNUSED(x) UNUSED_ ## x __attribute__((__unused__)) #elif defined(_MSC_VER) /* We used to check for && (_MSC_VER >= 1300) but that's also out of date. */ # define GREENLET_NOINLINE_SUPPORTED # define GREENLET_NOINLINE(name) __declspec(noinline) name # define GREENLET_NOINLINE_P(rtype, name) __declspec(noinline) rtype name # define UNUSED(x) UNUSED_ ## x #endif #if defined(_MSC_VER) # define G_NOEXCEPT_WIN32 G_NOEXCEPT #else # define G_NOEXCEPT_WIN32 #endif #endif