Comparing the CI build time between master and this PR, the unit test now costs 30% longer time. My gut is it comes from the cost of creating every single context and evaluating @babel/polyfill and babel-helpers-in-memory.

Instead of running common modules for every single exec test, can we leverage the cacheData option? We can generate code cache after running common modules, and supply it to every single exec test context.

Makes every exec.js test run in a new VM context (necessary for the new tests)

Can we separate this into a new PR? After the code cache, running yarn jest object-rest-spread now costs 8 secs vs. 4 sec on master.

Note that we can circumvent this issue by adding some suffix.

Object.defineProperty(Object.prototype, `NOSET${__filename}`, {
  get(value) {
    // noop
  },
});

I am good on the loose mode changes, but we should see if we can do better on the exec context.