The one interesting piece that needed to be fixed up was the type of an `Indirect[Read|Write]SideEffect` operand/result. If the parameter type is a pointer or reference to an incomplete type, we need to set the type of the side effect memory access to `Unknown`, because we don't model incomplete types in the IR type system.
I also added minimal support for `__assume` (generated as a `NoOp`), because lack of `__assume` support got in the way of debugging the other issue above.
This change gives a slight performance improvement and makes the QL code
shorter. It introduces some magic numbers in the code, but those are
confined to the `Pos` and `Spec` classes.
We get a speed-up because the evaluator has built-in support for integer
literals in the `OUTPUT` of `JOIN` operations, whereas `newtype`s have
to be explicitly joined on. As a result, a predicate like
`CFG::straightLineSparse#ffff` drops from 262 pipeline nodes to 242.
I measured performance on https://github.com/jluttine/suitesparse, which
is one of the projects that had the biggest slowdown when enabling the
QL CFG on lgtm.com. I took two measurements before this change and two
after. The `CFG.qll` stage took 117s and 112s before, and it took 106s
and 107s after.
This test used `getAQlClass`, which caused it to break when new classes
were added anywhere in the libraries. That's now avoided by switching to
`getCanonicalQLClass`. It turns out that `getCanonicalQLClass` didn't
support arithmetic expressions on complex numbers, so that support had
to be added.
This fixes the test `cpp/ql/test/library-tests/constexpr_if/cfg.ql`,
which broke when the QL CFG was enabled.
The new cases are just copy-pastes of the `IfStmt` cases (they don't
share a useful common superclass) with added checks for whether their
constant value equals 0.
This change is needed when enabling the QL CFG on certain snapshots such
as notaz/picodrive. It removes the `bbNotInLoop` predicate, which was
always a liability because it's inherently quadratic. The real slowdown
came in `skipLoop`, where all true-upon-entry loops were crossed with
all definitions of variables that should take their definition from the
loop body.
