Set return code at beginning of termination function

[nmlgc]

DOSBox also sets it before any other checks or assignments. Required for Turbo C++ 4.0J's TCC.EXE and TLINK.EXE to correctly return their exit code to the host OS, and to build systems that rely on non-zero exit codes to indicate failure.

Once again, the actual cause behind the bug is more complicated:

• Both TCC and TLINK spawn RTM.EXE as a TSR.
• This TSR offers various helper functions that run using the PSP of the RTM.EXE process. This is done by overriding the current PSP using INT 21h, AH=50h and resetting it to its original value before returning.
• RTM.EXE also properly terminates itself via INT 21h, AH=4C shortly before the parent process does.
• This means that msdos_process_terminate() gets called three times:
  • Call 1) TSR's RTM,
  • call 2) exits RTM with exit code 0, and
  • call 3) terminates the parent process with the intended exit code.
• msdos_process_terminate() always frees the process entry, even in the function's TSR mode.
• Three terminations for two processes mean that call 3) will always fail with a throw(0x1f); and not run to the bottom of the function where retval is set. Therefore, retval stays at the 0 set during call 2), ignoring the actually intended exit code.
• Just leaking the process entry won't fix this issue due to the way the PSP is continuously overridden:
  • MS-DOS Player's implementation of INT 21h, AH=50h permanently changes the PSP of the process entry.
  • This causes both the parent's and RTM's entry to both point to RTM's PSP during RTM function calls.
  • Since RTM's termination must be run with RTM's PSP as the active one, RTM doesn't get to revert this final temporary PSP change.
  • This leaves the parent's PSP permanently set to RTM's PSP, …
  • …causing call 3) to never actually target the parent process.
• It would work if we also simplified msdos_int_21h_50h() to a single current_psp = CPU_BX; assignment, matching regular DOS. But that's two workarounds instead of one.

In the long term, we probably want to remove the process_t structure and only emulate its fields in terms of the single SDA and multiple PSP structures. For example, MS-DOS exclusively keeps a process's maximum number of files in the PSP, not in any separate process structure. Just like the new_process allocation hacks, process_t also dates back to the very first build from 2010, and looks like an equally temporary hack.