Varargs support via FFI functions #121
Conversation
In order to call varargs functions using ffi_prep_cif_var we need to know the fixed argument count for the target method. This change adds such a field to CallContext allowing it to select the appropriate function.
|
As expected, the pre-built linux binaries failed. Will mark this as non-draft once I know they work with the changes, and we can incrementally get the non-Darwin binaries rebuilt. |
In order to avoid changing the JNI stub or requiring that currently-working platforms rebuild that stub, this patch moves the fixed param count into the high word of the existing `flags` parameter passed into newCallContext. Older builds will ignore this value — since they only mask off low word bits — and continue to use the non-variadic FFI logic that already works on those platforms. Once rebuilt, they will start to see the fixed count and use it for the variadic FFI logic.
|
Most recent commit moves the fixed param count into the high word of the I fully acknowledge the slight hackiness of passing the fixed count in the flags field but any other solution would require changing the C stub and introducing fallbacks on platforms where it has not been rebuilt. That feels hackier to me. We will need to address the versioning of the native stub better (see #93). |
JRuby FFI duplicates some logic, notably logic from VariadicInvoker, due to private or package-private utilities in jnr-ffi. In this case, it needs to be able to set up the CallContext for the function using the fixed param count, but there's no constructor to do this and no utility function to create the CallContext manually. This duplication will need to be cleaned up but that is out of scope for this change. I add this constructor so that the duped code in JRuby can make the proper call.
See the following PRs: * jnr/jffi#121 * jnr/jnr-ffi#292 * jnr/jnr-posix#174
This PR modifies variadic function call logic to use the
ffi_prep_cif_varfunction provided by libffi for passing varargs. This primarily improves support on Apple Silicon, which (at least under Darwin) structures the call stack for varargs differently than other platforms (a mix of registers and stack).This is the first part of fixes to better support Apple Silicon. Additional fixes are required in the upstream libraries like jnr-ffi and jnr-posix.
Note that at the time I submit this initial PR it still requires the native library to be rebuilt for each platform. Given the challenges of doing so right now, I hope to find a way to make the Java changes work with the older builds of the C library. We will address build and versioning issues in the future.