The original code, removed in the "rewrite" patch, was incorrect for
large structures, and required dynamic allocation of a trampoline on
every ffi_call.
Instead, allocate a 4k entry table of all possible structure returns.
The table is 80k, but is read-only and dynamically paged, which ought
to be better than allocating the trampoline.
This is difficult to test with gcc. One can only use -O0 at present.
See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63668.
6eec410f15