> > Instead of disabling the optimization, can't you make KMSAN aware of the > "has_zero()" call (which I believe is the part that will trip up KMSAN) and > just not report it? > > -- Steve I don't think this is the right thing to do. If we read a short string - e.g. only 4 bytes long, using read_word_at_a_time(), we'll also read uninitialized bytes after the trailing zero, which will lead to false positives we're trying to avoid. But if there's no trailing zero, or if the first 4 bytes are uninitialized as well, we want to catch that, so we must not ignore functions that operate on those bytes. We could theoretically make read_word_at_a_time() mask out the bytes past the trailing zero in KMSAN mode, but IIUC that function may read binary data as well.
