this isn't required by the spec and PHPStan chokes on it. I was previously having it ignore these errors, but it turns out that PHPStan is not making use of extended typeinfo provided if it can't parse the tag, which is problematic on level 6 and also a problem for array-of-type. Therefore, we are going to have to take the hit.
this was making the location table point to an offset that did not yet exist, which caused the region header consistency check to discard the region as corrupted the next time it was loaded.
This is better for performance because these then don't need to be reevaluated every time they are called.
When encountering an unqualified function or constant reference, PHP will first try to locate a symbol in the current namespace by that name, and then fall back to the global namespace.
This short-circuits the check, which has substantial performance effects in some cases - in particular, ord(), chr() and strlen() show ~1500x faster calls when they are fully qualified.
However, this doesn't mean that PM is getting a massive amount faster. In real world terms, this translates to about 10-15% performance improvement.
But before anyone gets excited, you should know that the CodeOptimizer in the PreProcessor repo has been applying fully-qualified symbol optimizations to Jenkins builds for years, which is one of the reasons why Jenkins builds have better performance than home-built or source installations.
We're choosing to do this for the sake of future SafePHP integration and also to be able to get rid of the buggy CodeOptimizer, so that phar and source are more consistent.
I don't know why this would be missing, but in some cases it is, as seen in the crash archive. Whatever the case, we shouldn't be shitting the bed because of this.
This takes advantage of two key behaviours of PHP:
1. Assigning a string does not copy the string
2. Changing an offset in a string causes the string to be copied.
These two factors combined, along with the fact that blocklight and skylight arrays are usually all-zeros, allow us to produce a significant memory usage reduction of loaded chunks.
A freshly generated PM world with 3,332 chunks loaded drops from 310MB to 200MB memory usage with these changes applied.
this process of fast-serialization, fast-deserialize, network-serialize is an order of magnitude slower than just doing the network encode directly on the main thread, and also copies more useless data.
For the main thread, the figures were something like 3x more expensive, and then an extra 7x for deserialization on the worker thread. This is a ridiculously large overhead.
