This works by adding data-flow edges to skip over array expressions when
reading from arrays. On the post-update side, there was already code to
skip over array expressions when storing to arrays. That happens in
`valueToUpdate` in `AddressFlow.qll`, which needed just a small tweak to
support assignments with non-field expressions at the top-level LHS,
like `*a = ...` or `a[0] = ...`.
The new code in `AddressFlow.qll` is copy-pasted from `EscapesTree.qll`,
and there is already a note in these files saying that they share a lot
of code and must be maintained in sync.
Flow from a definition by reference of a field into its object was
working inconsistently and in a very syntax-dependent way. For a
function `f` receiving a reference, `f(a->x)` could propagate data back
to `a` via the _reverse read_ mechanism in the shared data-flow library,
but for a function `g` receiving a pointer, `g(&a->x)` would not work.
And `f((*a).x)` would not work either.
In all cases, the issue was that the shared data-flow library propagates
data backwards between `PostUpdateNode`s only, but there is no
`PostUpdateNode` for `a->x` in `g(&a->x)`. This pull request inserts
such post-update nodes where appropriate and links them to their
neighbors. In this exapmle, flow back from the output parameter of `g`
passes first to the `PostUpdateNode` of `&`, then to the (new)
`PostUpdateNode` of `a->x`, and finally, as a _reverse read_ with the
appropriate field projection, to `a`.
This case was added in dccc0f4db. The surrounding code has changed a lot
since then, and the case no longer seems to have an effect except to
create some dead ends and possibly cycles in the local flow graph.
The data flow library conflates pointers and their objects in some
places but not others. For example, a member function call `x.f()` will
cause flow from `x` of type `T` to `this` of type `T*` inside `f`. It
might be ideal to avoid that conflation, but that's not realistic
without using the IR.
We've had good experience in the taint tracking library with conflating
pointers and objects, and it improves results for field flow, so perhaps
it's time to try it out for all data flow.
g++ doesn't support this code:
sorry, unimplemented: non-trivial designated initializers not supported
twoIntFields sSwapped = { .m2 = source(), .m1 = 0 };
so we need to build it in clang mode.
This removes a lot of flow steps, but it all seems to be flow that was
present twice: both exiting a `PartialDefNode` and a
`DefinitionByReferenceNode`. All `DefinitionByReferenceNode`s are now
`PartialDefNode`s.
This commit changes how data flow works in the following code.
MyType x = source();
defineByReference(&x);
sink(x);
The question here is whether there should be flow from `source` to
`sink`. Such flow is desirable if `defineByReference` doesn't write to
all of `x`, but it's undesirable if `defineByReference` is a typical
init function in `C` that writes to every field or if
`defineByReference` is `memcpy` or `memset` on the full range.
Before 1.20.0, there would be flow from `source` to `sink` in case `x`
happened to be modeled with `BlockVar` but not in case `x` happened to
be modelled with SSA. The choice of modelling depends on an analysis of
how `x` is used elsewhere in the function, and it's supposed to be an
internal implementation detail that there are two ways to model
variables. In 1.20.0, I changed the `BlockVar` behavior so it worked the
same as SSA, never allowing that flow. It turns out that this change
broke a customer's query.
This commit reverts `BlockVar` to its old behavior of letting flow
propagate past the `defineByReference` call and then regains consistency
by changing all variables that are ever defined by reference to be
modelled with `BlockVar` instead of SSA. This means we now get too much
flow in certain cases, but that appears to be better overall than
getting too little flow. See also the discussion in CPP-336.
