This regression might have been around for a couple emacs versions.  I
don't recall when it started.  Probably about the same time as the
window management overhaul.

Assuming that you are in the same directory as the included foo.c
file...

Compile the file with debugging support:
    gcc -ggdb -o foo foo.c

Then start emacs:
    emacs -Q

Split the frame into two windows:
    C-x 4 f foo.c RET

Debugging sequence 1

    M-x gdb RET
    RET         (accepting default "gdb -i=me foo")
    start RET

At this point, notice that *gud-foo* buffer is gone, as foo.c has
replaced it.  At some point in time in the past, foo.c would end up
replacing the other window (the *scratch* buffer).  That would be the
desired behavior, as I would be able to continue using the *gud* buffer.

    C-x 4 b RET   (accepting default "*gud-foo*")
    n RET RET

Now *input/output of foo* replaced the *gud-foo* buffer.  I can't say
much about this, as I normally use gdb a different way.  But it seems
sub-optimal.  For my usual debugging habits, see Sequence 2.

Debugging sequence 2

Same set-up as for sequence 1, then as follows:

    M-x gud-gdb RET
    RET               (accepting default "gdb --fullname foo")
    start RET

At this point, notice that *gud-foo* buffer is gone, as foo.c has
replaced it.  At some point in time in the past, foo.c would end up
replacing the other window (the *scratch* buffer).  That would be the
desired behavior, as I would be able to continue using the *gud* buffer.

If need be, I can create a slightly more elaborate example containing
more than one C file, and we can see *gud-foo* getting buried again as
other C files pop up.