C++: Split new range analysis into constant and relative stages

[rdmarsh2]

This splits the new range analysis into a constant stage and a relative stage, so that the constant stage can be used for overflow detection and potentially include some nonlinear recursion, while limiting the total performance impact by keeping relative bounds out of that recursion.

[MathiasVP]

I see there are some result changes caused by this PR. That's not expected, is it?

[MathiasVP]

A couple of comments. Mostly about keeping the files language-neutral

[MathiasVP]

This is also the only language-specific parts of this file, right? I think we should move it out. As far as I can see, the only place where IR is used is in

this.(SemanticBound::SemSsaBound).getExpr(0) instanceof IR::PhiInstruction

which I hope we can refactor 🤞.

[MathiasVP]

It would be good to check the performance impact of instantiating RangeStage twice. You can do that in one of two ways, I'd say:

• Run the buffer-overflows query suite we have in DCA
• Create a branch from DO NOT MERGE: C++: Replace simple range analysis uses by semantic range analysis uses #12505 and merge this PR into that branch, and run DCA on that.

[MathiasVP]

I took the liberty of starting option two here.

[MathiasVP]

DCA looks fine 🎉

[MathiasVP]

Also seeing compilation failures on CI:

FAILED(COMPILATION) /home/runner/work/semmle-code/semmle-code/ql/cpp/ql/test/library-tests/ir/modulus-analysis/ModulusAnalysis.ql

[MathiasVP]

One small comment. Otherwise, this LGTM once DCA is happy

Edit: DCA finished and shows no new performance concerns. So my only two requests are:

• Remove the unnecessary TODO comment, and
• Get rid of the one IR-specific construct that's now used directly in RangeAnalysisImpl.qll.