Invalid writes prevent valid writes

Hi,

On the Xeon Phi, the invalid write to a[-1] prevents both the prior and subsequent writes to a[0] from occurring:

__kernel void A(__global int* a) {
  a[0] = 1;
  a[-1] = 2;
  a[0] = 3;
}

Though the programs is technically invalid, I found this inconsistency interesting, since in all other OpenCL drivers I have tested (including 4 other Intel drivers), the write to a[-1] will not affect the final result of a[0] = 3. On the Xeon Phi, neither a[0] = 1 or a[0] = 3 are executed. Full test case is: http://paste.ubuntu.com/25216421/

The device is a "Intel(R) Many Integrated Core Acceleration Card", the host OS is Ubuntu 16.04, and the driver version is:

$ cat /etc/OpenCL/vendors/intel64.icd
/opt/intel/opencl-1.2-4.4.0.117/lib64/libintelocl.so

Cheers,
Chris