ffi: adapt to return value braindamage in libffi.

This change gets TXR FFI working on Power PC 64,
big endian. Many thanks to the GCC Compile Farm.

Turns out, the libffi API has changed some years ago,
with outdated documentation lingering.  Return values
for the basic integral types must all be handled with
a buffer of type ffi_arg and casting. This is true in
both ffi_call and closures: in both directions.

For instance, if a foreign function returns char, we
must retrieve a ffi_arg value (which might be as
large as 64 bits).  Then we must cast this value to
char. In actual fact, on a big endian system with an
8 byte ffi_arg, the char value may be located in byte
7 of the 8 byte buffer, not in byte 0.

FFI's own test suite got this wrong and had to
be patched back in 2013: it was doing things the way
we are doing them:

https://sourceware.org/ml/libffi-discuss/2013/msg00199.html

The doc was updated at the same time to tell the truth:

https://sourceware.org/ml/libffi-discuss/2013/msg00200.html

Luckily, we can fix TXR's ffi module without damaging its
framework. The object model is flexible enough to absorb the
change. Basically, we add two new methods to the txr_ffi_type
objects: virtual functions rget and rput.  We only add these
when targetting big endian.  These have the same type
signature as get and put, but are specialized for handling
return values. For all integer types smaller than 64 bits,
these functions are separately implemented with distinct
semanics which handle the FFI API requirements with ffi_arg
and casting and all. For all other types, rget and rput
are identical to get and put.

* ffi.c (pad_retval, ifbe, ifbe2): New macros, conditionally
defined for big and little endian.
(struct txr_ffi_type): New members rput and rget.
(ffi_i8_rput, ffi_i8_rget, ffi_u8_rput, ffi_u8_rget,
ffi_i16_rput, ffi_i16_rget, ffi_u16_rput, ffi_u16_rget,
ffi_i32_rput, ffi_i32_rget, ffi_u32_rput, ffi_u32_rget,
ffi_char_rput, ffi_char_rget, ffi_uchar_rput, ffi_uchar_rget,
ffi_bchar_rget, ffi_short_rput, ffi_short_rget,
ffi_ushort_rput, ffi_ushort_rget, ffi_int_rput, ffi_int_rget,
ffi_uint_rput, ffi_uint_rget, ffi_long_rput, ffi_long_rget,
ffi_ulong_rput, ffi_ulong_rget, ffi_wchar_rput): New
functions.
(make_ffi_type_builtin): Take rput and rget arguments,
regardless of platform. On big endian, store these in the
corresponding members. If they are null, duplicate get and put
instead.
(ffi_type_compile, ffi_init_types): Specify the rput and rget
functions for the basic types, using the ifbe macro which
nullifies the references to functions when they don't exist on
little endian. For all other types, pass the new arguments as
null.
(ffi_call_wrap): Pad the return vaue buffer size to a minimum
size on big endian; it must be at least sizeof (ffi_alloc).
Use the rget method for the return value on big endian.
(ffi_closure_dispatch, ffi_closure_dispatch_safe): Use rput on
big endian for storing return value.

0 files changed, 0 insertions, 0 deletions