grpc/doc/core/moving-to-c++.md

Moving gRPC core to C++

October 2017

ctiller, markdroth, vjpai

Background and Goal

gRPC core was originally written in C89 for several reasons (possibility of kernel integration, ease of wrapping, compiler support, etc). Over time, this was changed to C99 as all relevant compilers in active use came to support C99 effectively. Now, gRPC core is C++ (although the code is still idiomatically C code) with C linkage for public functions. Throughout all of these transitions, the public header files are committed to remain in C89.

The goal now is to make the gRPC core implementation true idiomatic C++ compatible with Google's C++ style guide.

Constraints

• No use of standard library if it requires link-time dependency
  • Standard library makes wrapping difficult/impossible and also reduces platform portability
  • This takes precedence over using C++ style guide
• Limited use of standard library if it does not require link-time dependency
• We can use things from std:: as long as they are header-only implementations.
• Since the standard library API does not specify whether any given part of the API is implemented header-only, the only way to know is to try using something and see if our tests fail.
• Since there is no guarantee that some header-only implementation in the standard library will remain header-only in the future, we should define our own API in lib/gprpp that is an alias for the thing we want to use in std:: and use the gprpp API in core. That way, if we later need to stop using the thing from std::, we can replace the alias with our own implementation.
• But lambdas are ok
• As are third-party libraries that meet our build requirements (such as many parts of abseil)
• There will be some C++ features that don't work
  • new and delete
  • pure virtual functions are not allowed because the message that prints out "Pure Virtual Function called" is part of the standard library
    • Make a #define GRPC_ABSTRACT {GPR_ASSERT(false);} instead of = 0;
• The sanity for making sure that we don't depend on libstdc++ is that at least some tests should explicitly not include it
  • Most tests can migrate to use gtest
    • There are tremendous # of code paths that can now be exposed to unit tests because of the use of gtest and C++
  • But at least some tests should not use gtest

Roadmap

• What should be the phases of getting code converted to idiomatic C++
  • Opportunistically do leaf code that other parts don't depend on
  • Spend a little time deciding how to do non-leaf stuff that isn't central or polymorphic (e.g., timer, call combiner)
  • For big central or polymorphic interfaces, actually do an API review (for things like transport, filter API, endpoint, closure, exec_ctx, ...) .
    • Core internal changes don't need a gRFC, but core surface changes do
    • But an API review should include at least a PR with the header change and tests to use it before it gets used more broadly
  • iomgr polling for POSIX is a gray area whether it's a leaf or central
• What is the schedule?
  • In Q4 2017, if some stuff happens opportunistically, great; otherwise ¯\_(ツ)_/¯
  • More updates as team time becomes available and committed to this project

Implications for C++ API and wrapped languages

• For C++ structs, switch to using when possible (e.g., Slice, ByteBuffer, ...)
• The C++ API implementation might directly start using grpc_transport_stream_op_batch rather than the core surface grpc_op.
• Can we get wrapped languages to a point where we can statically link C++? This will take a year in probability but that would allow the use of std::
  • Are there other environments that don't support std library, like maybe Android NDK?
    • Probably, that might push things out to 18 months