Tobias Meggendorfer

Huh! I didn't expect this to come out of a minor version change. Well, that at least explains what happens in that particular case ... Thanks for trying with an...

Aha! The 17.0.9 changelog includes `hotspot/compiler C2 Blackholes should allow load optimizations` Since your performance was similar when the blackhole was moved outside the loop, I am guessing this is...

> for 3 accesses or more the rule is to cache I'm fine with that. Its just important to note that a) this primarily is a style decision, not performance...

> Well, the rule tries to avoid the bad case. It is physically impossible to try with all type combinations and all instances whether there will be a negative impact...

> fastutil has been written > 20y ago, and at that time a series of compilation optimizations made reserve loop faster. I think things change and it's possible current JVM...

if I recall correctly, there are none indeed! It would certainly be nice to have it, not only from code style but also performance (`EntrySet#forEach` allocates objects)

A docker-compatible approach would be to `touch` / write a file. Scripts could easily check the modification time regularly or use filesystem events (inotify) for instant response. EDIT: Actually, you...

Related comment: When an expected_verdict can be anything, does it make sense to write `expected_verdict: "False(xy)"` instead of adding `subproperty: xy` to the task-definition? (reasoning: `subproperty` is domain-specific) Also: Right...

> But if we put something like the precision into a separate key, then we should make it consistently always in separate keys. But that would be part of the...

I will double check the following: * Brief usage instructions for both tools * Separate python API doc file * Restructure the benchexec description