Reverting PR 452

[lemire]

PR #452 by @CharlesChen888 introduced an optimization that could be 'argued' to be safe. @feng-y reported problems and they blamed this PR.

I am not able to fault PR #452 when using mainstream compilers, but with the Intel compiler, we fail the tests (see my observation #506).

So I am reverting it.

It might be possible to make this safe under all compilers, but it will require more care.

[CharlesChen888]

Is someone doing further inspections? I have little knowledge about Intel's compilers, and I am curious why PR #452 is causing problems.

[lemire]

@CharlesChen888 I am not sure it was ever safe. It may simply be a case where one compiler happens to expose a bug through our tests that is not triggered by other compilers.

Please see...
#476
where @feng-y report problems. They never provided a reproducible example, unfortunately.

If you take the latest version of the Intel compiler (ICX), and try to run our tests, you should be able to reproduce an issue. I am unable to demonstrate because the Intel compiler is not available in our CI, but I can assure you that the problem is real and reproducible. The inplace union gave the wrong answer, just like @feng-y reported.

If you could check it out with the Intel compiler (the supported on, ICX, not ICC), then that would be great. I am not going to work on this.