`a += b` may silently generates incorrect code

[Shizhi Tang]

We distinguish a = b and a[...] = b, where the former is an assignment in stage-1, and the latter is in stage-2. This is well defined because a = b goes to a Python reference assignment, and a[...] = b goes to __setitem__.

However, there is not a clear distinction between a += b and a[...] += b: they both go to __iadd__, which is in stage-2. The real problem is that when an exception raise from __iadd__, Python will automatically fall back to a = a + b for a += b, which is in stage-1, and this leads to an incorrect code. An example of such exceptions is modifying a read-only tensor.

[Shizhi Tang]

This issue is not true.