*** Reason for rollback ***
Rolling back while investigating tsan reports.
*** Original change description ***
Remove a volatile read/write in SmallSortedMap.entrySet()
I don't know why this was volatile.
Volatile semantics are not strong enough for making this read-then-write a
transaction. You'd need a mutex or AtomicReference to achieve that.
***
PiperOrigin-RevId: 662300377
I don't know why this was volatile.
Volatile semantics are not strong enough for making this read-then-write a
transaction. You'd need a mutex or AtomicReference to achieve that.
Maybe it's so that if someone else has already made a EntrySet on another
thread, we can reuse that? But, it's probably cheaper to make a new EntrySet
probably than to do the volatile read-or-write (this is just a guess, may be wrong, I'm
not super experienced with the cost of volatiles).
The EntrySet is stateless (except for the state of its containing
SmallSortedMap) so it's fine to have more than one of them.
The commit introducing this doesn't explain why it's volatile, but the method
comment referring to "Similar to the AbstractMap implementation of {@code
keySet()}" may have some clues.
AbstractMap.keySet uses a transient keySet, not a volatile keySet.
Is it possible that this initial implementation mistook transient for volatile?
Anyway, it's unnecessary, let's rip it out and enjoy a little more performance.
PiperOrigin-RevId: 662071950
I've noted the rationale for this in the comment, and in the attached bug.
This should save about 4-8 bytes of memory (on the 64bit server JVM depending on if compressed object pointers are on) per object with extensions. And 4 bytes on Android.
PiperOrigin-RevId: 661900014
We mostly use generated pom files in our release currently, so we can delete all the files that aren't used and the tools to update them.
Note, java/bom/pom.xml java/pom.xml and java/protoc/pom.xml are all still used at release and java/kotlin/pom.xml is used for documentation so all of those need to stay for now.
PiperOrigin-RevId: 659664012
Store them as Object[], because you can't have a Entry[], because Entry[] has a generic 'this' over <K, V>, you get error "generic array creation" if you try to make "new Entry[]".
And if you try to cast Object[] to Entry[], you get ClassCastException. So I've just made it an object array that we cast when reading out of.
This should save memory and improve performance, because we have fewer pointers to chase.
Also fixed some warnings about unnecessary unchecked suppressions, and type-name should end with T.
PiperOrigin-RevId: 658585983
This should save a little memory. We've observed hundreds of such empty arrays
in some applications.
This affects non-empty messages with none of:
- repeated fields
- map fields
- fields to check if they're initialized
PiperOrigin-RevId: 658137927
This may save some alloactions if java's intrinsics aren't smart enough to avoid the roundtrip. But most JVMs probably have smart enough intrinsics, so this is probably not going to speed things up, just make the code look nicer.
PiperOrigin-RevId: 653445330
I believe they have intrinsics even for the long-form Integer.compareTo(Integer.valueOf(a), Integer.valueOf(b)) format which avoids the intermediate allocation.
So this probably won't make things faster, just makes the code a little cleaner.
Integer.compare was added in Java 1.7 which seems safe. Added to Android in SDK 19, which is less than the 21 minSDK supported by protobuf: 303239d74d
PiperOrigin-RevId: 653438420
We are seeing in profiles that this constructor is calling the other constructor.
It doesn't seem to be totally inlined, as the times spent in each constructor
stack frame are different.
In the common case of no builderParent, the constructor can be empty.
PiperOrigin-RevId: 653272368
Create WORKSPACE.bzlmod. Before building with Bzlmod resulted in use of full WORKSPACE.
Some repos are still there, but the file should eventually be empty.
Add dep to rules_kotlin 1.9.0. This was the first version available on BCR. It pushed upgrade of
rules_jvm_external to 6.0 and rules_java to 6.5.2 (keep 6.0.0 on Bazel 6.3.0).
Add missing maven and other deps to MODULE.bazel
CI changes:
Disable Bazel 6.4.0 with bzlmod. rules_jvm_external 6.0 use use_repo_rule, which is not supported by Bazel 6.
Add C++ build "Bazel7 with Bzlmod" enabled.
Add Java builds with "Bazel 7 with/without Bzlmod".
Fixes: https://github.com/protocolbuffers/protobuf/issues/17176
PiperOrigin-RevId: 652773197
These APIs were intended to support an internal-only mutable implementation of Java protobuf and should not actually be used in open source. Removal of these APIs shouldn't break anyone -- the equivalent immutable methods should be used instead.
PiperOrigin-RevId: 651833683
Looking at the code history, this test used to try two different ways to construct a proto containing map fields, and check that the resulting protos were the same. One of the ways was to obtain mutable `Map` objects for the fields, and modify those. When that possibility was removed from the proto API, the test code that used it was changed to use the other way,
map-modifying methods directly on the proto class. But the end result is that the test method here was just doing the same thing twice and checking that it got the same result.
PiperOrigin-RevId: 650639933
This is a performance optimisation that avoids us going through from
AbstractList.add(E) to LazyStringArrayList.add(int index, String) and then
having to call the index-based add in ArrayList, which has more bookkeeping
around moving elements across if necessary.
We can avoid that bookkeeping by adding this overload.
Also, group together overloads to squash a style warning.
PiperOrigin-RevId: 650089286
This should slightly speed things up.
Should be safe: these operations don't change the number of entries inside the loop.
PiperOrigin-RevId: 650087172
I've made a new name for the interface, ExtensionSerializer, so we can keep ExtensionWriter in the same place for backwards-compatibility for a little while.
PiperOrigin-RevId: 644172922
These classes are deprecated and will be removed in the next breaking change. Users should update gencode to >= 4.26.x which uses GeneratedMessage instead of GeneratedMessageV3.
Tested with //compatibility:java_conformance_v3.25.0 which builds the runtime against 3.25.0 gencode.
PiperOrigin-RevId: 644136172
This extends previous workaround for java features in unknown fields, to include features extensions that are known but use a mismatched descriptor.
This can happen when users bring their own descriptors via buildFrom.
PiperOrigin-RevId: 644013578
This was introduced by the previous fix for delimited inheritance, and was never released. This fix removes all getType() calls from crosslink, where it's not safe to inspect the message type, which is still a placeholder, until after crosslinking. Using the inferred type is not necessary since we treat messages and groups the same during crosslink.
PiperOrigin-RevId: 643394981
Mike Kruskal
Protobuf Team Bot
Mark Hansen
Deanna Garcia
zhangskz
Sandy Zhang
Éamonn McManus