gcc 3.1 bug: flag_no_inline incorrectly set

[Jason Merrill]

The existing code wants to disable the RTL inliner and enable the tree
inliner unless flag_instrument_function_entry_exit is set.  I believe this
is correct.

The behavior was changed somewhat by Aldy's patch to always enable the tree
inliner, in order to support __attribute ((always_inline)), but I don't
think the impact on real code is significant; the same patch created
flag_really_no_inline so that only functions with that attribute will be
inlined when inlining is supposed to be disabled.

What exactly is the bug?

Jason