MSVC: Compiler bug on seemingly trivial C code?

Here's a simple, 45 line console application that reproduces the bug on my Win XP 32-bit system.

The relevant parts:

input_watcher_wndproc belongs to a window that was set-up via RegisterRawInputDevices to watch raw keyboard input events.

  LRESULT APIENTRY
  input_watcher_wndproc(HWND hwnd, UINT umsg, WPARAM wparam, LPARAM lparam) {
    switch(umsg) {
    case WM_INPUT:
      analyze_raw_input((HRAWINPUT)lparam);
      DefWindowProc(hwnd, umsg, wparam, lparam);
      return 0;
    }
    return DefWindowProc(hwnd, umsg, wparam, lparam);
  }

  void开发者_如何学JAVA analyze_raw_input(HRAWINPUT raw_input) {
    RAWINPUTHEADER header;
    UINT size;
    GetRawInputData(raw_input, RID_HEADER, &header, &size, sizeof(RAWINPUTHEADER));
    printf("raw input device handle: %d\n", header.hDevice);
  }

When compiled with MSVC 9 or 10, the program always incorrectly reports the device handle to be "1".

However, if I apply some seemingly non-functional changes, like transforming the switch-statement into an if-statement...

    if (umsg == WM_INPUT) {
      analyze_raw_input((HRAWINPUT)lparam);
      DefWindowProc(hwnd, umsg, wparam, lparam);
      return 0;
    }

... or when adding an arbitrary instruction before the call to analyze_raw_input ...

  switch(umsg) {
   case WM_INPUT:
     puts("foo");
     analyze_raw_input((HRAWINPUT)lparam);
     DefWindowProc(hwnd, umsg, wparam, lparam);
     return 0;
   }

...the code runs fine and outputs the corrent device handle.

Here's the modified, working source code, with just the print statement added.

Any ideas what might cause this behaviour?

Make sure to initialize 'size' to be sizeof(RAWINPUTHEADER) before calling GetRawInputData since this is an __inout parameter.