Apple recently updated the docs on dispatch_once to point out
that the storage for the dispatch_once_t must be static or global,
but not something that was ever used before as the implementation
doesn't use a memory barrier. So we drop the use and create the
semaphore when needed and use an atomic swap deal with any
threading races.
This commit adds a __wrap_memcpy function and a linker flag to use that
in place of memcpy for our Ruby gem C extension. This allows us to
always use the 2.2.5 version of memcpy, making it possible to use the
gem on distributions with pre-2.14 versions of glibc.
Before this change:
$ objdump -T protobuf_c.so | grep memcpy
0000000000000000 DF *UND* 0000000000000000 GLIBC_2.3.4 __memcpy_chk
0000000000000000 DF *UND* 0000000000000000 GLIBC_2.14 memcpy
After:
$ objdump -T protobuf_c.so | grep memcpy
0000000000000000 DF *UND* 0000000000000000 GLIBC_2.2.5 memcpy
0000000000000000 DF *UND* 0000000000000000 GLIBC_2.3.4 __memcpy_chk
0000000000042450 g DF .text 0000000000000005 Base __wrap_memcpy
This is based on gRPC's solution to a similar problem:
5098508d2d/src/core/lib/support/wrap_memcpy.c
This fixes issue #2783.
- Don't prune the extension registry as that can lead to failures when two
threads are racing.
- If adding the method fails, check and see if it already is bound to decide
the return result. Deals with threading races binding the methods.
--[name]_opt support depended on the plugin being register, and didn't support
working with just --[name]_out directive (where the plugin is found via the
users PATH. This extends the command line handing to allow --[name]_out to
be all it takes for the _opt directive to also be supported.
Fixes https://github.com/google/protobuf/issues/2712
- The Timestamp proto does not allow for negative nanos fields, so the seconds
must be shifted and a positive nanos then applied.
- Tweak the helpers on Duration to make it clear there is no "base" time
involved.
- Update the unittests for duration and timestamp to cover positive and
negative NSTimeIntervals and what their impact is on the protos.
* Update upb, fixes some bugs (including a hash table problem).
* Ruby: added a test for the previous hash table corruption.
Verified that this triggers the bug in the currently released
version.
* Ruby: bugfix for SEGV.
* Ruby: removed old code for dup'ing defs.
- If setting/clearing a repeated field/map that was objects, check the class
before checking the autocreator.
- Just to be paranoid, don’t mutate within copy/mutableCopy for the autocreated
classes to ensure there is less chance of issues if someone does something
really crazy threading wise.
- Some more tests for the internal AutocreatedArray/AutocreatedDictionary
classes to ensure things are working as expected.
- Add Xcode 8.2 to the full_mac_build.sh supported list.
Pull request #2517 caused cross compilations to start failing, because
the js_embed binary was being built to run on the target platform
instead of on the build machine. This change updates the Autotools build
to use the AX_PROG_CXX_FOR_BUILD macro to find a suitable compiler for
the build machine and always use that when building js_embed.
embed.cc: In function ‘std::string CEscape(const string&)’:
embed.cc:51:32: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (int i = 0; i < str.size(); ++i) {
^
Symbol is not yet available on older versions of Node.js and so this
test fails with them. This change just directly checks whether Symbol is
available before we try to use it.
