Summary:
Adds a data structure that holds a series of values that can be *borrowed* for exclusive writing.
That means, that only a single consumer can write to any value owned by the data structure.
In addition, the data structure exposes read access via iteration over all contained values.
A typical use case would be a counter with thread-local values that are accumulated by readers in other parts of a programm. The design carefully avoids the use of atomics or locks for reading and writing. This approach avoids cache flushes and bus sync between cores.
Borrowing and returning a value go through a central lock to guarantee the consistency of the underlying data structure.
Values are allocated in a `std::forward_list`, which typically should avoid two values in the same cache line -- in that case, writing to one value would still cause cache flushing on other cores. An alternative approach would be to allocate values continuously on cache line boundaries (with padding between them). We can still change the code if the current approach turns out to be too naive (non-deterministic).
Reviewed By: SidharthGuglani
Differential Revision: D15535018
fbshipit-source-id: 212ac88bba9682a4c9d4326b46de0ee2fb5d9a7e
Summary:
Replace the *copy on write* vector with an atomic pointer to a linked list.
This allows to publish without locking a mutex, at the cost of the slower traversal of a linked list (a vector has better locality).
At the moment, the typical use case is to have one subscriber, meaning that the afforementioned slower traversal is not a problem.
Adding subscribers is implemented as atomic *compare and swap.*
Reviewed By: SidharthGuglani
Differential Revision: D15546964
fbshipit-source-id: 41bfa41f1ac6be5c9b6bf4288ea3271ee995877e
Summary:
We want to phase out usage of config pointers on nodes. Setting configs is no longer needed, as a config is unly used during construction.
Here we deprecate the setter, as it is no longer working as it used to (e.g. changing `useWebDefaults` after a node is constructed).
Reviewed By: SidharthGuglani
Differential Revision: D15416474
fbshipit-source-id: a2cc06cad0c5148cecce056ece5f141b3defe9a9
Summary: Removes a check introduced in D6969537, comparing `totalFlexGrowFactors` and `resolveFlexGrow` to both `0.0` *and* undefined.
Reviewed By: SidharthGuglani
Differential Revision: D15431425
fbshipit-source-id: 13c8f24e1bc8c49496097a6aa78e20ee5d3964a7
Summary:
`YGNode::setAndPropogateUseLegacyFlag` was only used for debugging purposes.
Here, we replace it with a free function in `Yoga.cpp`.
Now that we have events, the diffing functionality should go into a separate debugging package and be implemented in terms of an event listener. Let's do that as soon as we can support multiple listeners.
Reviewed By: SidharthGuglani
Differential Revision: D15316863
fbshipit-source-id: db929eba7c2de8aa1550e362dd2c175929c0070e
Summary:
In order to remove the config pointer from nodes, we have to keep track of whether the node is using web defaults.
This information fits into one bit that we can place in padding (i.e. no extra memory needed).
This allows us to get rid of config usage withing `YGNode` with some exceptions:
- `iterChildrenAfterCloningIfNeeded` -- this function will simply receive the configuration, or the cloning callback.
- `setAndPropogateUseLegacyFlag` -- will be removed in D15316863
- in `YGNode::reset` -- will go away utomatically once we remove the config pointer
Reviewed By: SidharthGuglani
Differential Revision: D15391536
fbshipit-source-id: 0fa0d0805c6862bd741fe4a7d9b637ed534f56a4
Summary: Using Mutex lock_guard mechanism when writing to subscribers and when accessing them in publish to make a copy
Reviewed By: davidaurelio
Differential Revision: D15391679
fbshipit-source-id: 16713ff28ce1762a5ca4c48c152897a92417e80b
Summary:
Adds one byte of private storage to `YGNode`, intended to be used by Yoga itself.
This is in previously unused alignment space, and won’t cause more memory to be allocated.
Reviewed By: SidharthGuglani
Differential Revision: D15296732
fbshipit-source-id: 3caf0a3cd506e4e324e51c31869c69be5781d476
Summary: Style bits had overlap, because `dimensionBit` was set with an incorrect increment.
Reviewed By: SidharthGuglani
Differential Revision: D15335134
fbshipit-source-id: 370e1a73547d76b0e26bc6ab67acb96d33ddf180
Summary:
using shared_ptr for vector of subscribers
Further changes in commit stack support the mutiple subscribers in event system
Reviewed By: davidaurelio
Differential Revision: D15352512
fbshipit-source-id: fac7f4268abf9ca4277734aca2f21cd711eb7d6e
Summary:
Replaced global event subscriber with a vector of subscriber functions
Further changes in commit stack support the mutiple subscribers in event system
Reviewed By: davidaurelio
Differential Revision: D15352451
fbshipit-source-id: 7ca6f0943735bf1f76a906c23e15e14ae3c5f42c
Summary:
Adds YogaEventListener interface and it's implementation which will be used in flipper for events coming from Yoga
After this diff , we will start getting layout calculate events in flipper listener
Reviewed By: davidaurelio
Differential Revision: D15316928
fbshipit-source-id: da3a53374a52386037b553d460038d988b0162c2
Summary:
On MacOS, the following steps result in build errors:
```
mkdir build
cd build
cmake ..
make
```
The problem is that yogacore uses C++11 features (`constexpr`) but C++11 isn't specified in CMakeLists.txt.
This PR solves the poblem by adding the following code to the bottom of CMakeLists.txt:
```
set_target_properties(yogacore PROPERTIES CXX_STANDARD 11)
```
This solution was derived from https://stackoverflow.com/questions/45688522/how-to-enable-c17-in-cmake
Pull Request resolved: https://github.com/facebook/yoga/pull/887
Differential Revision: D15334282
Pulled By: davidaurelio
fbshipit-source-id: a599d8a8f555f7f9cd8dc333e12dc2387d5b2d7a
Summary:
Adds `LayoutPassStart` and `LayoutPassEnd` events.
The existing `NodeLayout` event in isolation is not as useful as it could be. Having events that mark start and end of a layout pass are a useful addition.
Differential Revision: D15305467
fbshipit-source-id: 14af6f65e698fb1e3112eb2ffd87a74d31df4840
Summary:
@public
`YGNodeGetInstanceCount` was only ever meant for tests, and caused data races in multi-threaded environments.
It was completely replaced with event-based counting for tests.
Here we remove public API around the counter, and the counter itself.
Reviewed By: SidharthGuglani
Differential Revision: D15174857
fbshipit-source-id: 228e85da565bac9e8485121e956a2e41910b11c8
Summary:
@public
replaces the global node counter with the event-based one for all tests.
Reviewed By: SidharthGuglani
Differential Revision: D15174856
fbshipit-source-id: f4401d502bdbaf3b6e4632a4d985aac260cb35a8
Summary:
@public
Test utility on top of the new event system that maintains a counter of instantiated nodes. Meant to replace the global node counter.
Reviewed By: SidharthGuglani
Differential Revision: D15174855
fbshipit-source-id: 6998472f95a09b8da652257a26596164bdcf43d6
Summary:
@public
Publish two events, `NodeAllocation` and `NodeDeallocation`, in the same places where the global node counter is changed.
Reviewed By: SidharthGuglani
Differential Revision: D15174858
fbshipit-source-id: 6e4e9add88513b9e987189ca5035d76da2a1de55
Summary:
@public
Reduces measure cache size to a number that is enough for 95% of nodes, according to our (FB-internal) measurements.
Node size: 776b -> 584b
Reviewed By: SidharthGuglani
Differential Revision: D15183567
fbshipit-source-id: 9ae8cc78074271a015e7618b931ba0356de87a0c
Summary: We conducted an experiment with different measure cache sizes. This has now been deallocatedi (D15183473). Remove the necessary APIs.
Reviewed By: SidharthGuglani
Differential Revision: D15183486
fbshipit-source-id: a38fa5a3ab0321c2521265f7d1cd6b495efd76cf
Summary:
@public
`YGConfig::YGConfig(YGConfig*)` was not initializing the same fields as the default constructors.
Here, we make the default constructor delegate to the more specialized one to remove duplication.
Reviewed By: SidharthGuglani
Differential Revision: D15164599
fbshipit-source-id: 27247709091b7664386057d09ac67d481877871f
Summary:
@public
We want to enable tooling, instrumentation, and statistics within Yoga without coupling these functionalities to our core code.
This commit introduces the foundations of a simple, global event system.
For the time being, we will only support a single subscriber. Should we require more than one, we can add support for it later.
Reviewed By: SidharthGuglani
Differential Revision: D15153678
fbshipit-source-id: 7d96f4c8def646a6a1b3908f946e7f81a6dba9c3
Summary:
@public
Makes bitfield getters/setters part of the bitfield ref template.
Since we introduced the tracking bit as template parameter in D14933022, every bitfield ref is an individual class anyway, and having function pointers doesn’t potentially lead to less code generation anyway.
Furthermore, this change can (in the absence of tracking bits) avoid less specialized templates dealing with refs, and to dynamic dispatch as a consequence.
Reviewed By: SidharthGuglani
Differential Revision: D15085495
fbshipit-source-id: 0dd70fa05e9d43a29e38a619cddb642c9ca3f7ab
Summary:
@public
In order to optimise property storage, we have to know how style properties are used in our apps.
Here, we add a bitmask that allows us to track which properties are set explicitely, and use that for our analysis.
Reviewed By: SidharthGuglani
Differential Revision: D14933022
fbshipit-source-id: 1ab8af562b14baba1d02057e527aa36d5c9a7823
Summary:
@public
The extra overload of `updateStyle` introduced in D15078961 can also handle `BitfieldRef`.
That means that we can remove the more specific implementation previously introduced for `BitfieldRef`
Reviewed By: SidharthGuglani
Differential Revision: D15081069
fbshipit-source-id: 98f1f3478627974c5273c85d268ca07350f303d7
Summary:
@public
Change style property accessors to return `Ref` instances instead of references to `CompactValue`.
This will allow to track assignments to properties later on, e.g. for instrumentation or dynamic property storage.
Reviewed By: SidharthGuglani
Differential Revision: D15078961
fbshipit-source-id: 259f05f7d30f093c04bf333c5bd4fb3601b8e933
Summary:
@public
Adds `YGStyle::ValueRepr` to make code depending on the actual type easier to write.
So far, we have treated `yoga::detail::CompactValue` as an implementation detail, and that’s what it’s supposed to stay.
React Native Fabric has one value conversion overload that depends on that type, though, and used `decltype(YGStyle{}.margin()[0])` until now.
That’s problematic for two reasons:
- we want to constrain the parameter of `operator[](...)` to enum types, making the `0` unsuitable
- we want to return the non-const overload of the operator to return a custom `Ref` type, which is not the type needed by Fabric.
Making the storage type explicit allows to write more forward-compatible code.
Reviewed By: SidharthGuglani
Differential Revision: D15078960
fbshipit-source-id: 932c27ef2f2cdc6ce965b79894268170f0ccdce5
Summary:
@public
Some `YGNode*` passed as `const YGNode*`, some const refs to sub-objects introduced.
This helps selecting the desired methods in more places, i.e. `const` overloads of accessors on `YGStyle`.
Reviewed By: SidharthGuglani
Differential Revision: D15078963
fbshipit-source-id: 5013721d6edcc68f42f4504f5c331da647a294bd
Summary:
@public
Having binary operators as member functions has disadvantages:
- the left hand side cannot be converted to `YGFloatOptional` implicitly (which we need for `YGStyle` refs)
- Operators are not necessarily commutative.
By moving these operators into free functions, and adding overloads for both variants if one operand is `float`, we get these properties.
Reviewed By: SidharthGuglani
Differential Revision: D15078962
fbshipit-source-id: 2e228a2ef90a8083c91788caa9eedfd4d140677f
Summary:
@public
Takes a const reference to the style of the printed node once, instead of using repeated calls to `node->getStyle()`.
Makes the code a bit shorter, and ensures that we are operating on `const YGStyle&`, which helps selecting the correct methods further up the stack.
Reviewed By: SidharthGuglani
Differential Revision: D14999094
fbshipit-source-id: 814f06b7e3179ac8cfb43d79fbec48ee4115d6e3
Summary:
@public
Introduces `YGNodeConstRef` as `const YGNode*`, i.e. a pointer to a constant `YGNode`.
We also use it for all style getters, which will avoid casts to `const YGNode*` in diffs up the stack.
We should use this pointer type for all functions that do not modify the underlying node.
Reviewed By: SidharthGuglani
Differential Revision: D14999095
fbshipit-source-id: 61cc53bb35e787a12ae12e70438d84c0a4983752
Summary:
@public
In order to encapsulate property access on `YGStyle`, as a first measure we wrap all fields with accessors.
This will e.g. enable dynamic property storage and instrumentation in the future.
All accessors have a `const` version that allows direct access via `const&`. For mutation, bit fields are wrapped with a custom reference object.
This style allows for the least amount of changes in client code. Property access simply needs appended parens, eg `style.direction` becomes `style.direction`.
Reviewed By: shergin
Differential Revision: D14999096
fbshipit-source-id: fbf29f7ddab520513d4618f5e70094c4f6330b30
Summary:
@public
Remove unnecessary `const` and `extern` specifiers from `Yoga.h`.
- Function declarations are `extern` by default
- The removed `const` specifiers for pass-by-valye parameters are only meaningful for the *definition* of functions, not for the declaration.
In this specific case, I found `const YGNodeRef` particularly confusing, as it is a `typedef` for a pointer type. `const` does not refer to the pointed-to object, but to the parameter itself, i.e. `const YGNodeRef` is `YGNode * const`, and not `const YGNode *`.
Reviewed By: SidharthGuglani
Differential Revision: D14999097
fbshipit-source-id: 8350870cb67f4a34722f796c4f4a2fc7dde41b99
Summary:
@public
This allows short methods defined in class declarations to occupy a single line.
The change makes class declarations more readable.
Reviewed By: SidharthGuglani
Differential Revision: D14950012
fbshipit-source-id: 1321949475184181c6cceb86613f730e430763e2
Summary:
We were using four edges for margin, padding and border. This diff changes the array size in YGLayout for margin, padding, border to reduce YGNode size and corresponding changes while we are setting values in YGLayout.
Reduces the YGNode size by 24 bytes
Reviewed By: davidaurelio
Differential Revision: D14892666
fbshipit-source-id: 94013d5183ee869901267c4c9941fd94fa05d848
Summary:
This directory was excluded from Black formatting in D14888843 when Black
was turned on for all of `xplat/`. Since there aren't a lot of `.py` files in
this directory, the opt-out is being removed, and the outstanding issues fixed.
Reviewed By: d16r
Differential Revision: D14889104
fbshipit-source-id: 440077c1efcb4c653151bca1da5636212978add5
Summary:
@public
compiler flags were pushed to C++14, but we don’t currently have any code that requires it. Setting to `-std=c++11` in order to fix the OSS build.
Reviewed By: SidharthGuglani
Differential Revision: D14833737
fbshipit-source-id: df2cd7da8c7124e89863c90d7b77bcf86c495618
Summary:
@public
Push version to *1.14.0.*
This release contains some bug fixes around `align-content`.
Reviewed By: colriot
Differential Revision: D14833417
fbshipit-source-id: f653d5fbb36f307c92b14c091c3206290256f036
Summary:
@public
Regenerating the “golden master” tests with chrome surfaced different bugs around `align-content`:
- a misunderstanding that values in `align-content` only applied *if there is only one line.* In fact, it applies *every time* a container is set to `flex-wrap: wrap`. Chrome had this wrong, and as such our tests were generated with incorrect parameters.
- empty children growing to the cross axis size of the container, even when `align-content` is different from `stretch`. This was implemented incorrectly in Chrome as well. Here, we fix it with an extra check.
Reviewed By: SidharthGuglani
Differential Revision: D14725402
fbshipit-source-id: a45bebdadb9c694dc0eb7e27cb52b3d247f81c50
Summary:
@public
In order to get out of pre-releases again, we move `YGSetUsedCachedEntries` from `Yoga.h` to `Yoga-internal.h`.
This way, it’s obvious that the function is not public, and we can remove it from future versions without breaking semver contracts.
Reviewed By: SidharthGuglani
Differential Revision: D14726029
fbshipit-source-id: 8d7e747e843bf5c4b5e1d4cfbfa37ca9d240dffb
