This places all of the tricky reflection logic of feature resolution into a single location that outputs a portable protobuf message. With this message, feature resolution becomes drastically simpler and easy to reimplement in other languages.
Follow-up changes will provide:
- An API to specify generator features from the CodeGenerator class for C++ plugins
- A CLI for building these default mappings
- Bazel/CMake rules to help embed these mappings in target languages (i.e. for runtimes and non-C++ plugins)
PiperOrigin-RevId: 559615720
This test validates that upb Python targets can be built successfully even if
Python is not installed locally.
I also updated our pinned upb version to pull in some recent fixes needed for
this test run.
PiperOrigin-RevId: 559504790
Disable the alignment check in 32-bit msvc.
This toolchain has a difference between "natural" and "required" alignment of
types and it can have the alignment of members of type `T` to be smaller than
`alignof(T)`.
Also, disable AnyTest.TestPackFromSerializationExceedsSizeLimit there because it can't allocate that much memory.
PiperOrigin-RevId: 559495544
This makes a few changes:
- It changes generated messages to reference message innards as a type in `__runtime` instead of branching on what fields should be there. That results in much less bifurcation in gencode and lets runtime-agnostic code reference raw message innards.
- It adds a generic mechanism for creating vtable-based mutators. These vtables point to thunks generated for interacting with C++ or upb fields. Right now, the design results in 2-word (msg+vtable) mutators for C++ and 3-word mutators (msg+arena+vtable) for UPB. See upb.rs for an explanation of the design options. I chose the `RawMessage+&Arena` design for mutator data as opposed to a `&MessageInner` design because it did not result in extra-indirection layout changes for message mutators. We could revisit this in the future with performance data, since this results in all field mutators being 3 words large instead of the register-friendly 2 words.
- And lastly, as a nearby change that touches on many of the same topics, it adds some extra SAFETY comments for Send/Sync in message gencode.
PiperOrigin-RevId: 559483437
This fixes Python/C++ and upb, and pushes the buggy behavior to pure python. There, it's very difficult to handle options on the bootstrapped proto with the current architecture. Future changes will attempt to address this more isolated issue.
PiperOrigin-RevId: 559450900
This will allow us to run all the upb tests from the protobuf repo, which
already uses this flag. I just had to remove a couple glob patterns that did
not match any files.
PiperOrigin-RevId: 559434154
If there is no local Python installation then the contents of this glob will be
empty. We want the build to succeed in that scenario, so this change will allow
that to happen.
PiperOrigin-RevId: 559407373
This will be used to seed feature resolution, which becomes just an edition lookup followed by proto merges. Runtime and generators that need to resolve their own features will be able to leverage this to avoid reimplementing all of the reflective logic that goes into the default calculation.
PiperOrigin-RevId: 559271634
This change adds test runs for GCC, Windows, and Mac. The 32-bit ASAN test
turned out to be somewhat difficult to migrate, so I left a TODO for that one.
I made sure to give each build environment its own cache prefix, so that we
don't end up with artifacts built by different toolchains conflicting with each
other in the same cache.
PiperOrigin-RevId: 559197655
1. getter function to access the upb_Map corresponding to a map proto field
2. conversion function to get a Go map from a upb_Map. The conversion function uses generics, so it must be a standalone function independent of `MessageAccessor`.
Accessing map fields is not as efficient as accessing other proto field types, because of two reasons:
- we need to make a copy of map values in Go, since the memory representations are different in UPB C
- we need to call C functions from Go, because the upb_Map struct fields are not exported and we can't access the bits directly, causing performance overhead
Luckily the map fields can be quite rare, so the performance overhead should not be too significant overall.
PiperOrigin-RevId: 559191359
The issue is that if another component (MediaPipe in this case) calls PackFrom, there'll be linker errors because kTypeGoogleApisComPrefix wasn't found.
PiperOrigin-RevId: 559132962
The Python comparison protocol requires that if an object doesn't know how to
compare itself to an object of a different type, it returns NotImplemented
rather than False. The interpreter will then try performing the comparison using
the other operand. This translates, for protos, to:
If a proto message doesn't know how to compare itself to an object of
non-message type, it returns NotImplemented. This way, the interpreter will then
try performing the comparison using the comparison methods of the other object,
which may know how to compare itself to a message. If not, then Python will
return the combined result (e.g., if both objects don't know how to perform
__eq__, then the equality operator `==` return false).
This change allows one to compare a proto with custom matchers such as mock.ANY
that the message doesn't know how to compare to, regardless of whether
mock.ANY is on the right-hand side or left-hand side of the equality (prior to
this change, it only worked with mock.ANY on the left-hand side).
See https://github.com/protocolbuffers/protobuf/issues/9173
PiperOrigin-RevId: 559056804
Remove message/message.h as a hdr from :message_internal
Lock down the visibility of :message_types to upb-only
message/types.h is not an internal header because the definitions there are part of the public surface. But the fact that it needs to exist is an implementation detail that should remain opaque, so make it a private dep.
PiperOrigin-RevId: 558960776
This new option will cause only the files that provide types to
be imported into the generated code. This helps reduce the inputs
that aren't needed (folks ignoring the protoc warnings) and for
things imported only for custom options as the objc generated code
doesn't capture those and thus the imports aren't needed in the
generated code.
Also do some general import cleanups.
PiperOrigin-RevId: 558867748
Implementing this method allows descriptors to be used in contexts that support
ABSL's formatting, most notably gunit's PrintToString and absl's StrFormat. The
implementations are done on the concrete descriptor type, which appears to be more
idiomatic for AbslStringify. Implementing them on pointers-to-descriptors would be
nicer for users who typically interact with pointers, but that would conflict with
AbslStringify's built-in support for pointers.
The implementation simply forwards to DebugString for simplicity.
PiperOrigin-RevId: 558854354
Arena enabled copy construction provides efficiency gains for the rather common construct where a message is currently copied by first creating a default initialized instance through the regular (arena) constructor, and then recursively filled from the source message using 'MergeFrom' or 'CopyFrom'.
Arena enabled copy construction is feature gated in port_def.inc and currently not enabled by default.
PiperOrigin-RevId: 558854125
It ended up being odd to construct the AccessorGenerators ahead of time given some specific Context<>: since each of the three methods on AccessorGenerators demands a different Context than the initial one so they shouldn't hold that Context anyway. It ended up just being a spooky parallel array to the Fields with an unenforced invariant that all methods would be called with a new Context<> but the same underlying FieldDescriptor later.
PiperOrigin-RevId: 558770643
The new naming will use our new vocabulary, and restrict feature access:
Unresolved source features - source-retention features that generators need to validate their own features. These should not be sent to the runtime or used for any decision-making.
Resolved source features - fully resolved source-retention features. These should be used to make codegen decisions, but shouldn't be sent to runtimes.
PiperOrigin-RevId: 558287302
This is a no-op change. Since Copybara will automatically format PRs with ClangFormat, this can lead to spurious changes that make CLs harder to review, as in cl/549058046. To avoid this, we format everything with clang-format in a single no-op change.
This CL also forces all pointer alignment to left pointers, for consistency (clang-format would normally deduce based on existing style in each file).
Also added `// clang-format off` comments in some places where PHP conventions lead to badly formatted code.
PiperOrigin-RevId: 558260317
Mike Kruskal
Eric Salo
Adam Cozzette
Protobuf Team Bot
Alyssa Haroldsen
Thomas Van Lenten
Protobuf Team Bot
Joshua Haberman
martijneken
Joshua Haberman
pizzud
Martijn Vels
Cong Liu