Previously a random test ordering was obtained by applying std::shuffle
to the tests in declaration order. This has two problems:
- It depends on the declaration order, so the order in which the tests
will be run will be platform-specific.
- When trying to debug accidental inter-test dependencies, it is helpful
to be able to find a minimal subset of tests which exhibits the issue.
However, any change to the set of tests being run will completely
change the test ordering, making it difficult or impossible to reduce
the set of tests being run in any reasonably efficient manner.
Therefore, change the randomization approach to resolve both these
issues.
Generate a random value based on the user-provided RNG seed. Convert
every test case to an integer by hashing a combination of that value
with the test name. Sort the test cases by this integer.
The test names and RNG are platform-independent, so this should be
consistent across platforms. Also, removing one test does not change
the integer value associated with the remaining tests, so they remain in
the same order.
To hash, use the FNV-1a hash, except with the basis being our randomly
selected value rather than the fixed basis set in the algorithm. Cannot
use std::hash, because it is important that the result be
platform-independent.
In the future, we will also want to introduce our own
`uniform_int_distribution` and `uniform_real_distribution` to get
repeatable test runs across different platforms.
This way it is explicit when there is a `StringRef` -> `std::string`
conversion and makes it easier to look for allocations that could
be avoided.
Doing this has already removed one allocation per registered test
case, as there was a completely pointless `StringRef` -> `std::string`
conversion when parsing tags of a test case.
This PR ultimately does 3 things
* Separately tracks matched tests per each filter part (that is, a set of filters separated by an OR (`,`)), which allows Catch2 to report each of the alternative filters that don't match any tests.
* Fixes `-w NoTests` to return non-zero in the process
* Adds tests for `-w NoTests`.
ReusableStringStream holds a std::ostringstream internally, but only exposes the ostream interface.
It caches a pool of ostringstreams in a vector which is currently global, but will be made thread-local.
Altogether this should enable both runtime and compile-time benefits. although more work is needed to realise the compile time opportunities.
Martin Hořeňovský