-
Notifications
You must be signed in to change notification settings - Fork 15.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
fix: missing -DPROTOBUF_USE_DLLS in pkg-config #12700
fix: missing -DPROTOBUF_USE_DLLS in pkg-config #12700
Conversation
When the protobuf libraries have been compiled as shared libraries the users of the library need to add `-DPROTOBUF_USE_DLLS` to their build line. Otherwise some symbols are missing.
Hi, I don't understand actually, why does Cheers, |
You may want to open a bug asking these questions. A comment in a PR is likely to get lost. Your questions make sense to me, but I have limited time to improve Protobuf, I just need it to work. |
…orms (#12983) Hi, It seems that until last year, the logic behind `PROTOBUF_USE_DLLS` was for Windows (MSCV) only. It was changed to all platforms here in 5a0887f Last month, the generated pkg config files were updated to reflect the protobuf build-time value of `PROTOBUF_USE_DLLS` as it was indeed noted that it changes the ABI. This was done in #12700 In the commit message it is mentionned that most likely we shall rather have a stable ABI. Finally in #12746 which at some point mentions https://issuetracker.google.com/issues/283987730#comment7 where a Google employee hits the linker issue: ``` undefined reference to `google::protobuf::internal::ThreadSafeArena::thread_cache_' ``` which denotes a mix of some .o or libs built `PROTOBUF_USE_DLLS` defined and some others build with `PROTOBUF_USE_DLLS` undefined, resulting in ABI incompatibilities. I also hit this issue while trying to include protobuf in a corporate environment using it's own proprietary build system in which it is expected that .a and .so use a compatible ABI. From my own understanding, ideally we should always use `thread_local` variables, but experience has shown that: - old iOS (iOS < 9) didn't seem to accept `thread_local`, leading to the `GOOGLE_PROTOBUF_NO_THREADLOCAL` macro later renamed `PROTOBUF_NO_THREADLOCAL` which allowed to disable this, but it is not set anywhere in the protobuf code base. Also I doubt you still want to support such old iOS now, so maybe you should consider removing all `PROTOBUF_NO_THREADLOCAL` related code paths (this pull request doesn't do this). - MSVC's DLL interface doesn't seem to accept exporting thread local variables (at least from what I understood, I know absolutely nothing about the Windows ecosystem), yet we can "hide" a thread local variable in a static function using a thread local variable. However in that case the access to TLS variable is not inlined, leading to worse performances, this hack shall be done only for Windows (actually when using MSVC) *AND* we build a shared library. - In all other cases, a classical `thread_local` shall be used, no matter if we build a static or a shared library. In particular on Linux which I guess is the target Google cares the more about for its own production. This pull request achieves this. Am I right in my conclusion ? Closes #12983 COPYBARA_INTEGRATE_REVIEW=#12983 from Romain-Geissler-1A:stable-abi-use-dll-non-windows dc23ff5 PiperOrigin-RevId: 538230923
…orms (protocolbuffers#12983) Hi, It seems that until last year, the logic behind `PROTOBUF_USE_DLLS` was for Windows (MSCV) only. It was changed to all platforms here in protocolbuffers@5a0887f Last month, the generated pkg config files were updated to reflect the protobuf build-time value of `PROTOBUF_USE_DLLS` as it was indeed noted that it changes the ABI. This was done in protocolbuffers#12700 In the commit message it is mentionned that most likely we shall rather have a stable ABI. Finally in protocolbuffers#12746 which at some point mentions https://issuetracker.google.com/issues/283987730#comment7 where a Google employee hits the linker issue: ``` undefined reference to `google::protobuf::internal::ThreadSafeArena::thread_cache_' ``` which denotes a mix of some .o or libs built `PROTOBUF_USE_DLLS` defined and some others build with `PROTOBUF_USE_DLLS` undefined, resulting in ABI incompatibilities. I also hit this issue while trying to include protobuf in a corporate environment using it's own proprietary build system in which it is expected that .a and .so use a compatible ABI. From my own understanding, ideally we should always use `thread_local` variables, but experience has shown that: - old iOS (iOS < 9) didn't seem to accept `thread_local`, leading to the `GOOGLE_PROTOBUF_NO_THREADLOCAL` macro later renamed `PROTOBUF_NO_THREADLOCAL` which allowed to disable this, but it is not set anywhere in the protobuf code base. Also I doubt you still want to support such old iOS now, so maybe you should consider removing all `PROTOBUF_NO_THREADLOCAL` related code paths (this pull request doesn't do this). - MSVC's DLL interface doesn't seem to accept exporting thread local variables (at least from what I understood, I know absolutely nothing about the Windows ecosystem), yet we can "hide" a thread local variable in a static function using a thread local variable. However in that case the access to TLS variable is not inlined, leading to worse performances, this hack shall be done only for Windows (actually when using MSVC) *AND* we build a shared library. - In all other cases, a classical `thread_local` shall be used, no matter if we build a static or a shared library. In particular on Linux which I guess is the target Google cares the more about for its own production. This pull request achieves this. Am I right in my conclusion ? Closes protocolbuffers#12983 COPYBARA_INTEGRATE_REVIEW=protocolbuffers#12983 from Romain-Geissler-1A:stable-abi-use-dll-non-windows dc23ff5 PiperOrigin-RevId: 538230923
When the protobuf libraries have been compiled as shared libraries the users of the library need to add
-DPROTOBUF_USE_DLLS
to their build line. Otherwise some symbols are missing.Fixes #12699
FWIW, I am not sure this is an ideal fix. It may be better to fix the headers such that no macros change the ABI.