jamesjb and I have been looking at the problems with backtrace on Win32. He
figured out what was going on, and I started figuring out how to fix it.
Anyway, the attached patch fixes the lisp-side notion of where the control
stack is, which makes backtraces work properly (the case which was breaking
earlier was trying to backtrace through an internal-error).
Unfortunately, this patch has a side effect. There are now constant
notifications from VirtualProtect that it has failed with code 0x1e7. This
turns out to be ERROR_INVALID_ADDRESS, and is because SBCL is attempting to
install a stack guard page on uncommitted memory. The stack guard page stuff
doesn't work on win32 -anyway-, so the easiest thing would be to disable it.