Hi Stefano,
On 2022-11-17 3:37 p.m., Stefano Rivera wrote:
Generally speaking python-cffi test failures point to libffi bugs. So,
I'd start by trying to verify the behaviour of the underlying behaviour
in libffi, from C.
The attached patch to python-cffi fixes the test failures on hppa. Please add
to patches
for python-cffi and push upstream if okay.
A couple of years ago the libffi port for hppa-linux was changed from using a
trampoline
executed on the stack to the function descriptor technique used by ia64. This
doesn't
require an executable stack and is simpler. However, function pointers need to
have the
PLABEL bit set in the pointer. As a result, a simple cast can't be used to
convert closure
pointers to function pointers.
python-cffi uses its own closure allocation so the problem can't be fixed in
libffi.
I added a macro CFFI_FN to do the conversion. It shouldn't affect other
architectures.
There is a similar define in libffi.
Regards,
Dave Anglin
--
John David Anglin dave.ang...@bell.net
--- ./c/_cffi_backend.c.save 2022-12-18 13:56:40.706730405 +0000
+++ ./c/_cffi_backend.c 2022-12-18 13:57:03.306944998 +0000
@@ -60,6 +60,13 @@
# endif
#endif
+/* Convert from closure pointer to function pointer. */
+#if defined(__hppa__) && !defined(__LP64__)
+#define CFFI_FN(f) ((void (*)(void))((unsigned int)(f) | 2))
+#else
+#define CFFI_FN(f) ((void (*)(void))f)
+#endif
+
/* Define the following macro ONLY if you trust libffi's version of
* ffi_closure_alloc() more than the code in malloc_closure.h.
@@ -3191,7 +3198,7 @@
Py_BEGIN_ALLOW_THREADS
restore_errno();
- ffi_call(&cif_descr->cif, (void (*)(void))(cd->c_data),
+ ffi_call(&cif_descr->cif, (void (*)(void)) CFFI_FN(cd->c_data),
resultdata, buffer_array);
save_errno();
Py_END_ALLOW_THREADS
@@ -6394,7 +6401,7 @@
goto error;
Py_INCREF(ct);
cd->head.c_type = ct;
- cd->head.c_data = (char *)closure_exec;
+ cd->head.c_data = (char *)CFFI_FN(closure_exec);
cd->head.c_weakreflist = NULL;
closure->user_data = NULL;
cd->closure = closure;
