Disable constant folding to not miss checks in constant expressions

[sim642]

When CIL folds constant expressions, Goblint never sees the operations, and thus does not appear to perform all the necessary overflow checks. This used to be optional for our first SV-COMP no-overflow implementation in #149.

This reveals new things which probably should be evaluated speculatively.

TODO

• Fix tests.
• Merge Fold constant conditional expressions, even without lowerConstants cil#215.
• Add missing cases to base eval_rv_base.

[sim642]

I was worried this might hurt performance, but sv-benchmarks level01 evaluation even suggests the opposite: OLS regression shows 5% speedup overall.
I don't know if this is due to disabling lowerConstants on the CIL side or not doing constFold in base, or both. I'll try to do an extra run to determine that.

The reason I was worried about performance is related to programs with many or large constant-foldable expressions. In theory, it might be cheaper to fold them directly, instead of using multiple of our int domains to reach the same constant result.
However, I now also see one way the performance could've improved: most expressions are not foldable, so constFold will just make an unnecessary copy on each evaluation (of each subexpression!).

[sim642]

I don't know if this is due to disabling lowerConstants on the CIL side or not doing constFold in base, or both. I'll try to do an extra run to determine that.

Re-enabling lowerConstants again after this didn't have any impact, so I suppose the difference comes all from not constFold-ing repeatedly.

I think it'd make sense to have lowerConstants enabled to not re-evaluate enumerators on every use because that doesn't really happen at runtime. But it requires some cleanup in CIL to decouple lowerConstants from other unrelated uses (like conditional folding).