------- Comment #20 from dj at redhat dot com 2010-08-12 16:57 ------- Just for fun, I compiled this test case with various levels of optimization. It works fine without optimization or with -O1, but segfaults at -O2 or -O3.
That indicates that the program only works by coincidence, not by design - you've made assumptions about how GCC will interpret your sources, and those assumptions are wrong. In this case, your assumption is that "bug_example_2" will always be a separate function, and will always be called as a separate function, and thus that you can assume some knowledge of the internals of the stack layout. The C language does *not* require that a function which is called, be called as a separate function, only that the semantics of the call be the same as far as the C language requires. The C language allows GCC to implement that function call in any way it chooses - and GCC chooses to implement it without actually doing a function call, but by copying the function body to the callee. At least, it does when optimizing. Without optimization, it *happens* to do what you expect. It will also do what you expect if bug_example_2 and bug_example are in separate source files - *then* the "cdecl" standard you refer to applies, because cross-object calls are limited by the compatibility standards. However - if you use gcc to link as well, gcc has the option of optimizing those calls *also*. So, GCC is "cdecl" compliant because *if* there's a function call, *then* the *stack* is laid out the same. However, the "cdecl" standard does *not* require that your program work, because C allows the optimizer to avoid the actual function call completely when the callee and caller are in the same scope. Note: you can tell gcc to not inline a function with __attribute__((noinline)) in which case a call to it is always an actual call to it, but it would be easier to just use the standard methods for accessing parameters so that it *always* works. Also, with full optimization enabled, your code crashes with MSVC also. Please file a bug report with Microsoft. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45265
