Some notes:

1. Stdlib unit tests runs tests in parallel processes but within each process runs the tests sequentially. We take advantage of this setup to ease the amount of concurrency work that is necessary. If at some point this changes, we will need to be much more careful about creating the global set and in fact will need to do the work to create multiple live object sets and only modify the one associated with the given test thread. That would be unfortunate = (... but luckily is not the case currently!

2. Currently the stdlib unit tests does do some leaks checking, but this checking:
a. IIRC is only for specific tests. It does not work across all tests.
b. Only works for classes (i.e. things with deinits). Any heap memory that is not a class and is allocated by allocObject/deallocObject can not be found by such leak detection.

The code used by the benchmark is in DriverUtils.swift in the sections ifdefed by

SWIFT_RUNTIME_ENABLE_LEAK_CHECKER

libobjc's test machinery has a similar mechanic. Some notes from experience there:

• libobjc's mechanism uses the malloc heap enumeration API to add up all heap allocations in the process. Good: this catches everything. Bad: this can be sensitive to non-leaks. For example, preheating caches can be difficult. Some tests use workarounds such as "run the test 10000 times, then pass if memory growth is not more than 4096 bytes".

• It is useful to be able to set the leak mark once and then check it multiple times.

• You may need to cycle a test many times to catch a leak. For example, if some operation incorrectly adds one entry to a global cache every time, you won't notice the leak until you repeat the operation enough times that the cache reallocates. For Swift's use this probably should be controllable externally: perform a small number of leak checking cycles for most testing, and perform a large number of leak checking cycles on some infrequently-run test bot.

• Diagnosing the leaks is easier if there are environment variables or something to control the detector machinery. For example, you sometimes need to dump the set of tracked allocations at the start and end of tracking, or make the test process hang after leaks are found so other debugging tools can be applied.

Resetting assignee on all Starter Bugs not modified since 2018.