The object::cache::threadmania test consistently crashes on Win32 x64. This is because the pthread API against which the test is written specifies a thread's return value as being pointer-sized. On Win32, the return value of a thread is 32 bits on all platforms, and so a pointer doesn't fit platforms with larger pointer sizes. When the return value is sized-up back to a pointer and dereferenced, the test case crashes.

This change fixes the problem by turning git_thread into a structure on Win32 with room for the thread handle, entry procedure pointer, and a void * parameter / return value member. To accomodate the programming model, the git_thread_join macro has been modified to require the address of a git_thread instead of passing it by value. This avoids passing a structure by value. The other git_thread function macros in use already accepted a git_thread * and did not require modification.

A randomly generated magic number incorporated into win32\pthread.c is used to check for clean thread exit. In the case of unclean exit, the value member of the git_win32_thread structure still contains the opaque value passed into the thread entry procedure, and should not be returned to the caller of git_thread_join as a proper return value. Furthermore, we should ensure that all threads started by the library terminate gracefully, so the implementation also raises an assert in the case of unclean thread termination.

While here I also fixed up a couple of warnings in global.c that MinGW complains about, and simplified the initialization logic for the Vista+ reader-writer lock function pointers.