builtin_return_addr vs frame_pointer_needed vs -O3

[DJ Delorie]

OK, I'm looking at a 20010122-1 failure (ix86-linux), which seems to
be caused by test4a() not having a valid frame through which to find
the right return address.  So, I use SETUP_FRAME_ADDRESSES to set
frame_pointer_needed, but it doesn't work.  A few printfs later, I see
that gcc is parsing a bunch of functions all at once, but deferring
the asm output until later, when frame_pointer_needed is no longer
valid.

So, what's the right way for a target to tell
expand_builtin_return_addr that the function calling it (for count>0
at least) needs a valid frame?