PyObject_GetBuffer fails on arrays

Hello, the regex package fails tests with PyPy.

Apparently PyObject_GetBuffer(string, &str_info->view, PyBUF_SIMPLE) does not work on array.array types. The relevant code is here

Test programme:

import regex
import array

a = array.array("c")
regex.compile("bla").match(a)

Result:

pypy test.py
PyObject_GetBuffer failed!
Traceback (most recent call last):
  File "test.py", line 5, in <module>
    regex.compile("bla").match(a)
TypeError: expected string or buffer

Version info:

$ pypy --version
Python 2.7.13 (ab0b9caf307db6592905a80b8faffd69b39005b8, Oct 13 2018, 23:33:37)
[PyPy 6.0.0 with GCC 8.2.0]

I tested, and I get the same error on top of CPython 2.7:

PyObject_GetBuffer(some_array_object, &view, PyBUF_SIMPLE)
TypeError: 'array.array' does not have the buffer interface

so it seems that PyPy is correctly implementing the same behaviour as CPython. The difference is likely inside regex itself: the line just before is #if defined(PYPY_VERSION). Likely, nowadays we should just completely remove the PyPy special case. You can check if that works; if it doesn't, then there is some real PyPy-CPython difference for us to fix; if it does work, then I suggest opening a pull request on the regex package to remove this code.

Hello @Armin Rigo, I've tried to do that but it fails with another error:

Traceback (most recent call last):
  File "test.py", line 5, in <module>
    regex.compile("bla").match(a)
SystemError: An exception was set, but function returned a value

Here's what I did to the regex package:

diff -r 92a74b206f0d regex_2/_regex.c
--- a/regex_2/_regex.c  Thu Nov 22 02:09:43 2018 +0000
+++ b/regex_2/_regex.c  Tue Nov 27 12:27:59 2018 +0100
@@ -17539,27 +17539,6 @@
     }

 #endif
-#if defined(PYPY_VERSION)
-    /* Get pointer to string buffer. */
-    if (PyObject_GetBuffer(string, &str_info->view, PyBUF_SIMPLE) != 0) {
-        printf("PyObject_GetBuffer failed!\n");
-        PyErr_SetString(PyExc_TypeError, "expected string or buffer");
-        return FALSE;
-    }
-
-    if (!str_info->view.buf) {
-        PyBuffer_Release(&str_info->view);
-        PyErr_SetString(PyExc_ValueError, "buffer is NULL");
-        return FALSE;
-    }
-
-    str_info->should_release = TRUE;
-
-    str_info->characters = str_info->view.buf;
-    str_info->length = str_info->view.len;
-    str_info->charsize = 1;
-    str_info->is_unicode = FALSE;
-#else
     /* Get pointer to string buffer. */
     buffer = Py_TYPE(string)->tp_as_buffer;
     str_info->view.len = -1;
@@ -17620,7 +17599,6 @@

     str_info->length = size;
     str_info->is_unicode = FALSE;
-#endif

     return TRUE;
 }

The C code contains a bug: it calls (*buffer->bf_getbuffer)(string, &str_info->view, PyBUF_SIMPLE), but if that returns -1, then it never clears the exception. I think that's why it ends up with the error SystemError: An exception was set, but function returned a value. I guess that the difference is that some other things that it calls happen to swallow the exception on CPython but not on PyPy. If this is correct then it's really a bug of _regex.c that should be fixed in general.