|
|
`absl::Uniform(tag, rng, a, b)` has some restrictions on the values it can produce in that it will always be in the range specified by `a` and `b`, but these restrictions can be violated by `absl::MockingBitGen`. This makes it easier than necessary to introduce a bug in tests using a mock RNG.
We can fix this by making `MockingBitGen` emit a runtime error if the value produced is out of bounds.
Immediately fixing all the internal buggy uses of `MockingBitGen` is currently infeasible, so the plan is this:
1. Add turned-off validation to `MockingBitGen` to avoid the costs of maintaining unsubmitted code.
2. Temporarily migrate the internal buggy use cases to keep the current behavior, to be fixed later.
3. Turn on validation for `MockingBitGen`.
4. Fix the internal buggy use cases over time.
---
A few of the different categories of errors I found:
- `Call(tag, rng, a, b) -> a or b`, for open/half-open intervals (i.e. incorrect boundary condition). This case happens quite a lot, e.g. by specifying `absl::Uniform<double>(rng, 0, 1)` to return `1.0`.
- `Call(tag, rng, 0, 1) -> 42` (i.e. return an arbitrary value). These may be straightforward to fix by just returning an in-range value, or sometimes they are difficult to fix because other data structures depend on those values.
PiperOrigin-RevId: 635503223
Change-Id: I9293ab78e79450e2b7b682dcb05149f238ecc550
|
