Ignore less cached data when more is available

[jrieken]

This PR helps with #101685. While drilling into a potential startup regression I have realised that calling createCachedData can yield less data than returned in previous rounds. It's not entirely clear to me why but having less cached data will negatively impact startup performance.

[jrieken]

@hashseed can you help me understand how this goes? For VS Code we use NodeJS's createCachedData-API (Node.js: 12.8.1, V8: 7.8.279.23-electron.0). We call createCachedData repeatedly with the idea that as more code is being parsed more cached data is created. However, it seems that it gets less and less, e.g for VS Code's main bundle we have the following cached data sizes: 9.4MB, 5MB, 3MB (3 calls, immediate after startup, after ~30sec, after ~3min).

[hashseed]

Maybe code got garbage-collected?

[jrieken]

Ah, that would explain it. So, at startup all the code is parsed (to some level of depth) and then GC'ed when unused. That makes sense to me. We should take this has a hint that our main bundle is too large... Thanks

[hashseed]

Not quite. V8 usually does not compile functions before it's executed for the first time. So this indicates to me that some code was used at startup, but is no longer used afterwards. V8 also has a code aging mechanism to get rid of stale code that has not executed for a longer time period.

[jrieken]

hm... any flags that I can set to trace this?

[alexdima]

👍