C++: Don't count every conversion as a use

[MathiasVP]

This PR reduces the number of SSA reads by not considering conversions as reads. This partially fixes the looping behavior observed in #11963 (but it doesn't fix the one observed in #11975).

It should also provide a small speedup boost since we're reducing the amount of reads considered in the SSA phase.

Commit-by-commit review recommended. The meat of the PR is in 9de8d5c (which includes conversions from the set of uses). However, I saw a number of missing results from just that commit, and it took some time to debug why. The reason is as follows:

We mark a number of ad-hoc things as conversions for iterator flow (to handle temporary variables created by rvalue-to-lvalue conversions). When we mark these as conversions we'd also excluding those from the set of uses, and this breaks a bunch of things. So 7ecc346 adds an additional column to the conversionFlow predicate that we can use to decide whether to consider those ad-hoc conversions. This column should be _ for most things, but must be false when we're excluding the conversions in isUse.

Finally, the last two commits just accepts test changes. These all appear to be fine 🎉.

[MathiasVP]

Converting this to a draft while investigating the result changes (on both CI and DCA). Done. The tests should be okay now. Will start another DCA run.

[geoffw0]

LGTM. I'm definitely seeking that additional reassurance from DCA though. In the tests I see that we lose quite a lot of taint steps (as we'd expect) and some duplicate results (which is a good change).

[MathiasVP]

LGTM. I'm definitely seeking that additional reassurance from DCA though. In the tests I see that we lose quite a lot of taint steps (as we'd expect) and some duplicate results (which is a good change).

Indeed. I've started a DCA run as well. I expected quite a lot of "lost results" in DCA. Mostly because we're getting less duplication, but also because we're fixing a class of dataflow looping that's demonstrated by the sink_then_source_1 test.

[MathiasVP]

I've looked through all the results on DCA. They all appear to be either:

• Result deduplications (as we've seen in our query tests on this PR), or
• FPs removed because of the false flows that's removed (similar to the ones in the dataflow-library tests).