Re: Reinstate mips16/nomips16 attribute support

[Sandra Loosemore <sandra at codesourcery dot com>]

Richard Sandiford wrote:

> Changing the implementation of -mflip-mips16 has indeed shown up a few
> hitherto-unknown problems, which I'll detail below.  However, I should
> say that this decision has been a bone of contention between the patch's
> co-authors; Sandra thinks -mflip-mips16 should be implemented using the
> separate mips_set_current_function mechanism described earlier, not least
> because of problem (4) below.

Just to clarify the nature of my concerns about this....  -mflip-mips16 
was originally intended just as a back end testing hook to exercise the 
new target reinitialization and MIPS-specific mode-changing code. 
Turning it into a C only front-end testing hook has made it less useful 
for that purpose, since it looked like the C++ front end was generating 
some different code patterns and hitting some different bugs than I'd 
seen in C.  (E.g., I was starting to poke at a bug still remaining from 
the C++ testing that seemed to be a reload problem, but that I couldn't 
readily duplicate by adding explicit attributes, because the things that 
were failing were implicitly-generated constructors and destructors.) 
So it seems that this change is a net loss in terms of testing coverage.

-Sandra