[InstCombine] We need to recursively enqueue users of instructions whose use-count changed

[LebedevRI]

Bugzilla Link	47238
Version	trunk
OS	Linux
Attachments	anoter testcase
CC	@fhahn,@nikic,@rotateright

Extended Description

And i think we've finally found a fundamental issue :S
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20200817/821034.html

While so far we've managed to get out of all the extra iterations by consistently
enqueueing all the operands to be revisited, i'm not sure it's really that simple.

Here, we have basically the following:

%i0 = cmp with non-canonical pred
%i1 = not %i0
%i2 = some-op %i0

First time around, we don't manage to fold the %i1 (by flipping cmp predicate),
because %i0 has other uses, but manage to fold away %i2, reducing use count of %i0.
We do enqueue %i0, but that does us no good, because we need to revisit %i1,
which we don't during this iteration.

So basically, not only do we need to enqueue the values whose use-count changed,
we also need to recursively enqueue users of these instructions.

[nikic]

We might want to distinguish instructions that have been added to the deferred worklist "for DCE" and adjust the code to not just perform DCE if use count is zero, but also re-add the single user if there is only one use. One-use folds seem like the main thing that would be problematic in this context. We might actually not want to reschedule the instruction itself if it was added for DCE (though this might not work as some transforms do inspect the users of an instruction).