tree c12a6d4f5e091d900d32f5db3e5419556e3d4240
parent 6a0e0722e7c7060de51a4b0f219a24162ad5a3f0
author Damien Neil <dneil@google.com> 1547832744 -0800
committer Damien Neil <dneil@google.com> 1548812026 +0000

internal/fileinit: generate reflect data structures from raw descriptors

This CL takes a significantly different approach to generating support
for protobuf reflection. The previous approach involved generating a
large number of Go literals to represent the reflection information.
While that approach was correct, it resulted in too much binary bloat.

The approach taken here initializes the reflection information from
the raw descriptor proto, which is a relatively dense representation
of the protobuf reflection information. In order to keep initialization
cost low, several measures were taken:
* At program init, the bare minimum is parsed in order to initialize
naming information for enums, messages, extensions, and services declared
in the file. This is done because those top-level declarations are often
relevant for registration.
* Only upon first are most of the other data structures for protobuf
reflection actually initialized.
* Instead of using proto.Unmarshal, a hand-written unmarshaler is used.
This allows us to avoid a dependendency on the descriptor proto and also
because the API for the descriptor proto is fundamentally non-performant
since it requires an allocation for every primitive field.

At a high-level, the new implementation lives in internal/fileinit.

Several changes were made to other parts of the repository:
* cmd/protoc-gen-go:
  * Stop compressing the raw descriptors. While compression does reduce
the size of the descriptors by approximately 2x, it is a pre-mature
optimization since the descriptors themselves are around 1% of the total
binary bloat that is due to generated protobufs.
  * Seeding protobuf reflection from the raw descriptor significantly
simplifies the generator implementation since it is no longer responsible
for constructing a tree of Go literals to represent the same information.
  * We remove the generation of the shadow types and instead call
protoimpl.MessageType.MessageOf. Unfortunately, this incurs an allocation
for every call to ProtoReflect since we need to allocate a tuple that wraps
a pointer to the message value, and a pointer to message type.
* internal/impl:
  * We add a MessageType.GoType field and make it required that it is
set prior to first use. This is done so that we can avoid calling
MessageType.init except for when it is actually needed. The allows code
to call (*FooMessage)(nil).ProtoReflect().Type() without fearing that the
init code will run, possibly triggering a recursive deadlock (where the
init code depends on getting the Type of some dependency which may be
declared within the same file).
* internal/cmd/generate-types:
  * The code to generate reflect/prototype/protofile_list_gen.go was copied
and altered to generated internal/fileinit.desc_list_gen.go.

At a high-level this CL adds significant technical complexity.
However, this is offset by several possible future changes:
* The prototype package can be drastically simplified. We can probably
reimplement internal/legacy to use internal/fileinit instead, allowing us
to drop another dependency on the prototype package. As a result, we can
probably delete most of the constructor types in that package.
* With the prototype package significantly pruned, and the fact that generated
code no longer depend on depends on that package, we can consider merging
what's left of prototype into protodesc.

Change-Id: I6090f023f2e1b6afaf62bd3ae883566242e30715
Reviewed-on: https://go-review.googlesource.com/c/158539
Reviewed-by: Herbie Ong <herbie@google.com>
Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>