At generation time, walk the file's dependencies to see what really contains
extensions so we can generate more minimal code that only links together the
roots that provided extensions. Gets a bunch of otherwise noop code out of
the call flow when the roots are +initialized.
This takes the code that was sitting in benchmarks/
already and makes it easier for language-specific
benchmarks to consume. Future PRs will enhance this
so that the language-specific benchmarks can report
metrics back that will be tracked over time in PerfKit.
Overview of changes:
- A new C#-specific command-line option, legacy_enum_values to revert to the old behavior
- When legacy_enum_values isn't specified, we strip the enum name as a prefix, and PascalCase the value name
- A new attribute within the C# code so that we can always tell the original in-proto name
Regenerating the C# code with legacy_enum_values leads to code which still compiles and works - but
there's more still to do.
NOTE: This is a binary breaking change as structure sizes have changed size
and/or order.
- Drop capturing field options, no other options were captured and other mobile
targeted languages don't try to capture this sort information (saved 8
bytes for every field defined (in static data and again in field descriptor
instance size data).
- No longer generate/compile in the messages/enums in descriptor.proto. If
developers need it, they should generate it and compile it in. Reduced the
overhead of the core library.
- Compute the number of has_bits actually needs to avoid over reserving.
- Let the boolean single fields store via a has_bit to avoid storage, makes
the common cases of the instance size smaller.
- Reorder some flags and down size the enums to contain the bits needed.
- Reorder the items in the structures to manually ensure they are are packed
better (especially when generating 64bit code - 8 bytes for every field,
16 bytes for every extension, instance sizes 8 bytes also).
- Split off the structure initialization so when the default is zero, the
generated static storage doesn't need to reserve the space. This is batched
at the message level, so all the fields for the message have to have zero
defaults to get the saves. By definition all proto3 syntax files fall into
this case but it also saves space for the proto2 that use the standard
defaults. (saves 8 bytes of static data for every field that had a zero
default)
- Don't track the enums defined by a message. Nothing in the runtime needs it
and it was just generation and runtime overhead. (saves 8 bytes per enum)
- Ensure EnumDescriptors are started up threadsafe in all cases.
- Split some of the Descriptor initialization into multiple methods so the
generated code isn't padded with lots of zero/nil args.
- Change how oneof info is feed to the runtime enabling us to generate less
static data (8 bytes saved per oneof for 64bit).
- Change how enum value informat is capture to pack the data and only decode
it if it ends up being needed. Avoids padding issues causing bloat of 64bit,
and removes the needs for extra pointers in addition to the data (just the
data and one pointer now).
Recently, descriptor.proto gained a GeneratedCodeInfo message, which means the generated code conflicts with our type.
Unfortunately this affects codegen as well, although this is a part of the public API which is very unlikely to affect hand-written code.
Generated code changes in next commit.
Added instructions on what tools to install to compile protobuf from
source. Removed the INSTALL.txt file because it's just a simple copy of
the autoconf documentation and confuses users.
Change-Id: I6fd8aa13495f1238fe5c62451b95ad480b1c4bed
Apple engineers have pointed out that OSSpinLocks are vulnerable to live locking
on iOS in cases of priority inversion:
. http://mjtsai.com/blog/2015/12/16/osspinlock-is-unsafe/
. https://lists.swift.org/pipermail/swift-dev/Week-of-Mon-20151214/000372.html
- Use a dispatch_semaphore_t within the extension registry.
- Use a dispatch_semaphore_t for protecting autocreation within messages.
- Drop the custom/internal GPBString class since we don't have really good
numbers to judge the locking replacements and it isn't required. We can
always bring it back with real data in the future.
- Hopefully complete the deps for other languages for the generated conformance proto sources.
- List the generated sources for cleanup by make's clean rules.
- Make the toplevel nuke the pyc files that can get created in the ObjC dir.
This is only thrown directly by JsonTokenizer, but surfaces from JsonParser as well. I've added doc comments to hopefully make everything clear.
The exception is actually thrown by the reader within JsonTokenizer, in anticipation of keeping track of the location within the document, but that change is not within this PR.
This includes all the well-known types except Any.
Some aspects are likely to require further work when the details of the JSON parsing expectations are hammered out in more detail. Some of these have "ignored" tests already.
Note that the choice *not* to use Json.NET was made for two reasons:
- Going from 0 dependencies to 1 dependency is a big hit, and there's not much benefit here
- Json.NET parses more leniently than we'd want; accommodating that would be nearly as much work as writing the tokenizer
This only really affects the JsonTokenizer, which could be replaced by Json.NET. The JsonParser code would be about the same length with Json.NET... but I wouldn't be as confident in it.
NS_ENUM changes defintion in Objective C++ based on the C++ spec being
compiled with, special case the one situation where it wouldn't support doing a
forward decl for the enum.
We still need the JSON representation, which relies on something like a DescriptorPool to fetch message types from based on the type URL. That will come a bit later.
(The DescriptorPool comment in this commit is just a note which will prove useful if we use DescriptorPool itself.)
