[Patch,libffi] Add MIPS64 soft-float support to libffi.

[Anthony Green]

David Daney wrote:
> libffi is lacking support for the mips64 soft-float ABIs.  In addition 
> be being annoying in and of itself, it also causes libjava failures.
>
> The patch adds soft-float support for both n32 and n64 ABIs.  Quite a 
> bit of code had to be touched as the rules for passing structures 
> containing floating point elements are a slightly complex.
>
> While working on the patch I noticed that we were failing a bunch if 
> pointer passing test cases that where recently added.  It turns out 
> that the original closure support for returning pointers was incorrect 
> and this was unmasked by the new test cases.  I wrapped fixes for this 
> into the patch as well (Yes, I know it is technically a no-no).  The 
> change to ffi_java_raw_to_rvalue() is needed with with the fixed 
> closure pointer return.
>
> I would like to add that I think there is still incorrect handling for 
> values of most types on big-endian systems in java_raw_api.c, but 
> since the java interpreter only uses a couple of types, the 
> problematical cases are never used.
>
> The end result is that all libffi tests except for some long double 
> things now pass at -O0.  There are a couple of floating point failures 
> with optimization turned on, but they look like they may be issues 
> with  libgcc's soft-float support.  The long double FAILures seem to 
> be due to lack of support in my glibc's printf.
>
>
> Tested on mips64-unknown-linux-gnu soft-float with both n32 and n64 
> ABIs with no regressions.  Currently testing mips64-unknown-linux-gnu 
> hard-float n32 and n64.
>
> OK to commit given the hard-float testing is good?
Thanks David.  Please commit once you've confirmed good test results.  
I'll move the patch to the libffi repository once I see your commit.

AG