Test failure is the same as #15789

I think this also needs to disallow changing the strides when ndarray.size == 0 in PyArray_CheckStrides.

Could maybe add the test from 912aedd with a @pytest.mark.skipif(not NPY_RELAXED_STRIDES_CHECKING ...)

Regarding PyArray_CheckStrides, I think you're suggesting

--- numpy\core\src\common\mem_overlap.c	Tue Aug 20 01:12:25 2019
+++ numpy\core\src\common\mem_overlap.c	Sat Mar 21 18:46:06 2020
@@ -666,10 +666,7 @@
 
     for (i = 0; i < nd; i++) {
         if (dims[i] == 0) {
-            /* If the array size is zero, return an empty range */
-            *lower_offset = 0;
-            *upper_offset = 0;
-            return;
+            continue;
         }
         /* Expand either upwards or downwards depending on stride */
         max_axis_offset = strides[i] * (dims[i] - 1);

I was thinking more of disallowing assignment of anything but 0 to strides. Currently this succeeds:

a = np.zeros((10, 9, 0))
assert a.size == 0
a.strides = (13, 17, 19)

@eric-wieser any thoughts on moving this forward? The "uhoh" messages are annoying.

Why did we close this instead of merging it? Was something important missing or was it a matter of this fix exposing yet another already-problematic API?

I think it was just closed as part of trying to close a few stalled PRs, just reopen if you like. We removed NPY_RELAXED_STRIDES_CHECKING=0 now, so this should be even simpler.

Maybe this can now be simplified to just memsetting strides to 0? I suppose there might be some (weird) ways to abuse run into the UB still, OTOH, I still doubt that the UB matters in practice usually.

Yeah, I think @charris made the reasonable decision to close out all my stalled PRs, as I haven't found much time in the last year or so to pick them back up again