A bunch of small improvements related to query translation #392
Conversation
- no hash calculations, - no overhead of hashed - if we want to change return type of the method to something more specific like IReadOnlyCollection or IReadOnlyList, then having arrays will help
- string operation improvements
this allows to create List more efficiently
This allow to have Count of processible requests in CommandFactory. Current API won't be hurt - new ctors provided and readonly field extends functionality so it is not a breaking change. True sequence is List<T> anyway.
- use ToArray(count)/ToList(capacity) if possible - if used only for foreach replaced with regular for with index - replaced some of usages with CollectionUtils.RangeToXxx()
more efficient collections operations
…e>>) is interlally accessible
+ small QueryHelper improvements
For every Query.All<T> one .MakeGenericMethod required and the result is only visited. The same root expressions are re-used many times during domain lifetime The number of items in the cache is finite and limit is the number of entities and interfaces in domain model (auxilary types included). In future we may change type of cache from concurrent dictionary to LRU cache and limit it to have only hot query roots represented
- ExpressionComparer uses zipping without Funcs, ValueTuple used instead - SubqueryDefalutResultRewriter uses type cast only when needed - Translator.ConstructQueryable is optimized to not have making generic type, it uses already given type by methodcallexpression. Additionally, by having overload that gets element type, we skip MethodCallExpression construction. - Small improvements in Translator.VisitSequence. Less operations for checking for entity set and no double search for type - typeof (Func<>) is cached within Translator - Small improvement for Union/Concat/Intersect/etc. operations by getting rid of extra enumerations - RecordSetHeader.Add(IReadOnlyList<Column>) substitutes counterpart with IEnumerable<Column> which produce less memory fluctuations. - RecordSetHeader.Add(Column) also doesn't use IEnumerable
Some types are more frequent to get accessor than others. For example - long, int and guid are frequent key values, entity version values so one of them will be in every query. - boolean fields are probably also frequent type amont model fields - DateTime/DateTimeOffset can be a time stamp for each row. - Double and decimal values are about money/amout of something/etc, they are more important than some other types. At the same time unsigned integers and Byte/SByte are, to me, used less frequently. It's a small optimization but difference between finding accessor on first and last check is ~2.5x and we always had to make 5-6 checks before reaching long which is, i believe, the most frequent key value type.
Avoid copying descriptors when it is possible by using IReadOnlyList<T> constructor parameter instead of params T[]. Constructors enumerate collections of descriptors, having option to pass List<T> and array of T is benifitial.
ReverseList uses backwards for, instead of true reverse
Even in case when we need to copy refs to some collection we don't need it, Reverse() copies refs to a buffer collection anyway.
|
Please consider following Return type covariance allows to avoid type casting in many cases. |
|
This PR is for 7.0. Do you want me to copy these changes to 7.0 or just merge it to master? |
I thought you are going to apply the PR's changes to master branch too. While I see now the |
|
Well, everything which is done in 6.0/7.0/7.1 happens to appear in master some time because of merges. We need this to propagate at least bug fixes from where they were originated (if version is still supported) to the newest versions. We have some enterprise customers (or customers of our customers) that still use 6.0 and if there is some issue which is very important for them, I'm expected to address it in 6.0. This PR contains no breaking changes of publicly available APIs, I didn't make drastic changes deliberately, even if they are performance beneficial. If your PR has some of such changes then I'd prefer to merge it to master. |
Changes of On the other hand, that usage is rare and trivial to convert |
|
It doesn't matter. Target runtime has no support for covariant returns. I will experiment with your suggestion in the version where it is supported by target runtime. |
No description provided.