Fix variadic arguments on arm64 darwin ABI

[glandium]

Normal arguments that spill on the stack are packed, but not variadic
arguments. This is handled correctly for their placement already, but
code generated on the callee side with va_list expects word-size
sign-extension, so we need to fill the entire word.

[lawrence-danna-apple]

Can you provide a test demonstrating the incorrect behavior that this fixes?

[glandium]

It turns out the fix was only necessary because the callee had va_arg() calls with undefined behavior.

[glandium]

Come to think of it, this may still be necessary.

The problem happens when the callee does e.g. va_arg(list, int), but the caller passed a smaller type, e.g. short, which is promoted to int per the C standard. The caller may cast back to the original type (short), but it should get a valid int from the promotion, but doesn't because libffi doesn't fill all the bits.

A small test that shows the problem looks like the following:

#include <stdio.h>
#include <ffi.h>

int main() {
	short val = -42;
	char *s = "Hello %x\n";
	ffi_cif cif;
	ffi_type* argtypes[] = { &ffi_type_pointer, &ffi_type_sshort };
	void* values[] = { &s, &val };
	ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 2, &ffi_type_void, argtypes);
	ffi_call(&cif, printf, NULL, values);

	printf(s, val);

	return 0;
}

It could be argued that the caller maybe should do the promotion itself, but it could be argued back that ffi should then return an error when it's passed varargs arguments that are supposed to be promoted.

(Note that technically, to printf a short it should be %hx, in which case there is no difference between the libffi call and the direct call, but the point is exactly for when the callee takes shortcuts with the argument type ; also note that clang doesn't warn about the format string for printf("%x", (short)-42) but does for e.g. "%lx").

[fkistner]

I am also seeing a problem with the ushort return value in libffi.bhaible/test-call.c -DDGTEST=25 -O2 macOS 11 arm64. Could this be related?

[glandium]

That would be something else, since it affects a return value, not a variadic argument.

[fkistner]

Thought this might've also been related to integer promotion, but seems it's a problem in the test setup. Filed as #598.

[atgreen]

libffi was recently changed to reject variable argument types that require promotion. This task is left to the caller.

[dkocher]

What does this mean to JNA which has previously cherry picked this PR?

[atgreen]

I don't think the fact that they are carrying this patch will break anything, but the right thing for them to do is to promote these types themselves, just like promote float to double for varargs: java-native-access/jna#463

The Ruby FFI layer also does its own promotion above libffi: https://github.com/ffi/ffi/blob/6d14c0a9107c0d5febb3bf92a60d2581e768fa2f/ext/ffi_c/Variadic.c