Describe the issue:

The function npy_memchr (in multiarray/common.h) has an optimisation for the case !NPY_ALIGNMENT_REQUIRED && needle == 0 && stride == 1 where it start by find the first nonzero 4-byte block before iterating within the block. However, the coarse-grained loop begins at haystack itself, and since the stride is 1 there's no expectation that this pointer would have any particular alignment.

While this is benign on x86_64, the misaligned access is nevertheless picked up by UBsan as it is undefined behaviour according to the C standard. npy_memchr should first scan each byte until it reaches the first 4-byte aligned block before entering the coarse-grained loop.

How to reproduce:

Build NumPy with UBsan and run a test suite containing a sufficiently large number of np.testing.assert_ calls. Since func_assert_same_pos constructs a boolean mask that is used to index into the arrays under test, with enough calls there's a decent chance that it will end up creating an unaligned boolean array at least once.

Error message:

third_party/py/numpy/core/src/multiarray/common.h:277:35: runtime error: load of misaligned address 0x... for type 'unsigned int', which requires 4 byte alignment
0x615000673205: note: pointer points here
 01 01 01 01 01 01 01  01 01 01 01 01 01 01 01  
             ^ 
    #0 0x... in npy_memchr third_party/py/numpy/core/src/multiarray/common.h:277:35
    #1 0x... in array_boolean_subscript third_party/py/numpy/core/src/multiarray/mapping.c:1113:30
    #2 0x... in array_subscript third_party/py/numpy/core/src/multiarray/mapping.c:1572:30
    #3 0x... in PyObject_GetItem third_party/python_runtime/v3_7/Objects/abstract.c:182:26
...
SUMMARY: UndefinedBehaviorSanitizer: misaligned-pointer-use third_party/py/numpy/core/src/multiarray/common.h:277:35

NumPy/Python version information:

Bug exists at HEAD.