Which nodes should be included in CFG?

Question

I'm writing a source-code-level optimizer for a language (LSL) whose server-side compiler is non-optimizing. The optimizer is working and performing well in some fields like constant folding, but to make it stronger, I wanted to build a CFG, to enable/simplify jump optimization, dead code removal, and later benefit from SSA.

Now, since it's source-level, I can't afford to convert the code to something like 3AC, otherwise I risk creating a version that is actually less optimal than the original.

So my question is: What would I include in the CFG? The logical answer would be to store AST nodes that represent whole statements. The problem I'm facing is that expressions may call user functions, and these functions may alter global variables, which would affect the SSA form, which is why I wanted to include the functions as part of the same graph, rather than different graphs.