Switch to the new compiler.  All tests pass\!  Need to submit bootstrap copies of the generated code used by the compiler itself.  These bootstrap copies are still built by the old compiler at the moment.

Pull TextBlob out of capnpc-capnp and into KJ as kj::StringTree.  Use it in the stringification code to make that code cleaner and avoid using iostreams.

Arrange for compiler to send to the code generator the complete contents of all files listed on the command line, all first-level dependencies of those files, and parent scopes of those dependencies, but nothing more.  This means it's sometimes possible to compile a schema file without parsing all transitive dependencies.  Also, in 'decode' mode, don't compile annotations, so that the files defining those annotations need not be opened at all.

Revert "Work around weird behavior in GCC 4.8 where it thinks this destructor's declaration and definition hase mismatching noexcept specs even though neither one has a noexcept spec at all."

This reverts commit c088dd95.

Reverted in favor of dwrensha's more complete fix.

Work around weird behavior in GCC 4.8 where it thinks this destructor's declaration and definition hase mismatching noexcept specs even though neither one has a noexcept spec at all.

Update Makefile.am to build new compiler.  Not bootstrapping yet, of course, since the C++ output plugin is not written.

Refactor compiler binary into a multi-tool with sub-commands like 'compile' and 'decode'.  The latter dynamically decodes binary input to text based on a provided schema file.

Fix compiler bugs.  Now test.capnp compiles to an identical schema to the old Haskell compiler, minus intentional changes.

Compiler actually running plugins!  Hooked it up to capnpc-capnp (the loopback plugin).  Working on getting the output identical to the old capnpc...  still a few bugs left.

Add utility code for really nice command-line option parsing and start using it in the compiler (still WIP).

typos

Futex-based mutex and once-init implementations for Linux.  These are measurably faster -- and much smaller -- than the pthread-based implementations.

Allow SchemaLoader to use a callback to lazily load schema nodes, including dependencies of loaded nodes.