[build] E: Build libffi.so alongside libffi.a

[esaurez]

Summary

Produce a position-independent libffi.so in addition to the existing static libffi.a, mirroring the pattern already established by esaurez/libxml2#1. The .so leaves libposix/libc/libm symbols unresolved so they bind to the host executable's .dynsym at dlopen time — the same model used by all CPython extension modules on Nanvix.

Why

This unblocks the Group B unbundling of _ctypes in CPython (see tracking note). Today _ctypes.cpython-312.so statically bundles a copy of libffi (~700 KB of duplication). With libffi.so available, _ctypes.so can drop to a thin module + DT_NEEDED libffi.so (~80 KB), and any other consumers that want to use libffi at runtime can share the same .so.

Beyond the size win, this also fixes a latent correctness issue: libffi.a as previously built was unusable under ld --whole-archive, which made it impossible to pull libffi into python.elf's .dynsym for the Group A unbundling strategy. The root cause and fix are documented below.

Implementation

Fix 1 — libffi.a nested-archive bundling

Change

Detail

NANVIX_LIBS now uses -L/-l form

Was: explicit /sysroot/lib/libposix.a /toolchain/lib/libc.a /toolchain/lib/libm.a passed via configure-time LDFLAGS. Libtool's static-link path treats explicit .a paths as "convenience archives" and bundles their members into the resulting libffi.a via ar cr. The result before this fix: $ ar t libffi.a libposix.a libc.a libm.a lt1-libposix.a lt2-libc.a lt3-libm.a prep_cif.o ... Consumers that try ld --whole-archive libffi.a then fail because ld cannot expand a nested archive. Switching to -L/-l form lets autoconf's AC_PROG_CC conftest still link a hello-world (so configure passes) while libtool records the libs as system deplibs in libffi.la and does NOT bundle them into libffi.a.

After the fix:

$ ar t libffi.a
prep_cif.o
types.o
raw_api.o
java_raw_api.o
closures.o
tramp.o
ffi.o
sysv.o

Fix 2 — libffi.so build target

Change	Detail
-fPIC in CFLAGS	Same .o files reusable for .a and .so
Keep --disable-shared	libtool's shared-lib detection does not know about i686-nanvix; the .so is linked manually
New $(SHAREDLIB) target	i686-nanvix-gcc -shared -fPIC -nostdlib -Wl,--whole-archive libffi.a -Wl,--no-whole-archive -Wl,-soname,libffi.so -Wl,-z,noexecstack -o libffi.so
test-functional extended	Verifies DT_SONAME=libffi.so, ffi_call / ffi_prep_cif exported in .dynsym, and that libffi.a is free of nested-archive members
package / verify-package	Ship both .a and .so
.nanvix/z.py _OUTPUT_FILES	Copy libffi.so back from the Docker build dir

Validation

$ readelf -d i686-pc-nanvix/.libs/libffi.so
 0x0000000e (SONAME)  Library soname: [libffi.so]

$ nm -D i686-pc-nanvix/.libs/libffi.so | grep -E '^[0-9a-f]+ T ffi_' | head
00001796 T ffi_call
000017c2 T ffi_call_go
00001167 T ffi_closure_alloc
000011a8 T ffi_closure_free
00000c1e T ffi_get_struct_offsets
00000b09 T ffi_prep_cif
00000b3a T ffi_prep_cif_var
00001ab0 T ffi_prep_closure_loc
00001ce0 T ffi_raw_call

$ ar t i686-pc-nanvix/.libs/libffi.a
prep_cif.o
types.o
raw_api.o
java_raw_api.o
closures.o
tramp.o
ffi.o
sysv.o

End-to-end runtime resolution (CPython dlopen → _ctypes.so → DT_NEEDED libffi.so → loader walks chain → resolves libposix/libc/libm UND against python.elf's .dynsym) will be validated in the follow-up CPython integration PR.

Runtime dependencies

• esaurez/nanvix#27 — .init_array invocation and DT_NEEDED chain walking in the user-space loader. (libffi has no global constructors of its own, but DT_RUNPATH is needed by the consumer chain.)

libffi.so does not introduce a diamond, so esaurez/nanvix#28 is not required for this PR specifically; consumers that combine libffi with other shared libs may still rely on nanvix#28 indirectly.

Downstream consumer

• esaurez/cpython — esaurez/cpython#14 will rework _ctypes.cpython-312.so to DT_NEEDED libffi.so so the ~700 KB libffi copy is shared instead of bundled per-extension.

[esaurez]

Superseded by upstream PR nanvix#235, which carries the same .so build addition -- rebased onto current upstream version branch, scope-trimmed (CI workflow downgrade, nanvix.toml version downgrade, .gitignore tweaks, .zutils-version downgrade, z.ps1/z.sh additions all dropped as unrelated env drift), and esaurez/* / python.elf / nanvix-todo references cleaned from commit message and PR body. Closing this fork PR; tracking continues upstream.