Use LRU cache for ESMReaders. When cache capacity is reached close least
recently used ESMReader. Remember the file name if a reader was open. Once the
reader requested again open the file if there is stored name for it. Put
released ESMReader to the back of the free items list. Close ESMReader's from
the front of the free items list.
Cached item can be used only by one client at the same time. If the same item is
requested twice exception is thrown. This should never happen in practice. If
this happens need to fix the client logic.
It's allowed to go over the capacity limit when requesting different readers.
Ideally this should never happen but there will be system error anyway
signalizing about too many open files. Need to fix client logic in this case.
All places that were using a vector of ESMReaders now using the cache. Cache is
local for each use case and there is no need for a thread safety.
This PR aims to start addressing `ESM` design issues that have silenced errors we incorporated into groundcover `ESM` loading approaches.
- We move the resolution of `parentFileIndices` from `ESMStore` to `ESMReader` as suggested in a `TODO` comment.
- We improve a highly misleading comment which downplayed the significance of `parentFileIndices`.
- We document important preconditions.
- We move a user facing error message to the highest level and improve its context.
- We remove an inappropriate `setGlobalReaderList` method. We now pass this reader list into the method that requires it.
- We remove a thoroughly pointless optimisation of `Store<ESM::LandTexture>`'s construction that has unnecessarily depended on `getGlobalReaderList`.
There should be no functional changes for `master`, but this PR should remove an issue blocking PR #3208.
With this PR we restore @elsid 's optimisations of countRecords we have unintentionally discarded in PR #3197. In addition, we give it a more appropriate name and add comments concerning its peculiar background.
The idea is to avoid std::map lookup for each CellRef. Instead generate a
sequence of added and removed RefNums into a vector then order them by RefNum
using a stable sort that preserves relative order of elements with the same
RefNum. RefIDs are stored in a different vector to avoid std::string move ctor
calls when swapping elements while sorting. Reversed iteration over added and
removed RefNums for each unique RefNum is an equivalent to what map-based
algorithm produces. The main benefit from sorting a vector is a data locality
that means less cache misses for each access. Reduces ESMStore::countRecords
perf cycles by 25%.
The resizing of LTEX store to the correct number of plugins was done in the load() method, but the load method won't be called if a plugin contains LAND records but doesn't contain LTEX records. For such plugins the Store<ESM::LandTexture>::search() function would then fail an assertion.
