> Unfortunately, it does not have any effect.

Show me your diff and the llvm-dis for cdm_wrapper.cc.

Here is the diff I tried: https://codereview.chromium.org/1950163002/diff/1/content_decryption_module.h
It fails with it.

I have just now realized the problem. Consider the CDM_EXPORT definition:

#if defined(CDM_IMPLEMENTATION)
#define CDM_EXPORT __attribute__((visibility("default")))
#else
#define CDM_EXPORT
#endif

Turns out, cdm_adapter.cc is compiled without CDM_IMPLEMENTATION set, and the visibility is not changed (here is where I got it wrong initially. I ought not to use their macro).

I had to modify my diff to make it pass:
https://codereview.chromium.org/1950163002/diff/20001/content_decryption_module.h

That said, the visibility does affect the checks, no problem here.
The issue is now that we need to add a number of new export macros to a number of classes. The original assumption that the visibility is set as we need does not hold.

There's still a chance that the visibility is set incorrectly. For instance, I am a bit puzzled why don't they set the symbol as visibility("default") in the code that uses the shared object.

Based on the offline discussion, we believe that setting visibility differently for the exporter and importer modules is incorrect. It seems to be a common pattern in media/ code:

1. CDM_EXPORT:
https://code.google.com/p/chromium/codesearch#chromium/src/media/cdm/api/content_decryption_module.h&q=CDM_EXPORT&sq=package:chromium&l=29&dr=CSs
#if defined(CDM_IMPLEMENTATION)
#define CDM_EXPORT __attribute__((visibility("default")))
#else
#define CDM_EXPORT
#endif

2. MEDIA_EXPORT
https://code.google.com/p/chromium/codesearch#chromium/src/media/base/media_export.h&q=MEDIA_EXPORT&sq=package:chromium&l=21&ct=rc&cd=1&dr=CSs
#if defined(MEDIA_IMPLEMENTATION)
#define MEDIA_EXPORT __attribute__((visibility("default")))
#else
#define MEDIA_EXPORT
#endif

3. GIN_EXPORT:
https://code.google.com/p/chromium/codesearch#chromium/src/gin/gin_export.h&q=visibility..default&sq=package:chromium&l=18&ct=rc&cd=4&dr=C
#if defined(GIN_IMPLEMENTATION)
#define GIN_EXPORT __attribute__((visibility("default")))
#else
#define GIN_EXPORT
#endif  // defined(GIN_IMPLEMENTATION)

4. IPC_EXPORT
https://code.google.com/p/chromium/codesearch#chromium/src/ipc/ipc_export.h&q=visibility..default&sq=package:chromium&l=28&ct=rc&cd=3&dr=C
#if defined(IPC_IMPLEMENTATION)
#define IPC_EXPORT __attribute__((visibility("default")))
#else
#define IPC_EXPORT
#endif

5. IPC_MOJO_EXPORT
https://code.google.com/p/chromium/codesearch#chromium/src/ipc/ipc_export.h&q=visibility..default&sq=package:chromium&l=34&dr=C
#if defined(IPC_MOJO_IMPLEMENTATION)
#define IPC_MOJO_EXPORT __attribute__((visibility("default")))
#else
#define IPC_MOJO_EXPORT
#endif

And many others: https://code.google.com/p/chromium/codesearch#search/&q=visibility..default&sq=package:chromium&type=cs

Initially, I believe that ifdef was motivated by Windows, where dllimport / dllexport attribute need to be set accordingly. In POSIX land, it seems that the correct settings are to use visibility("default") in both cases.

Peter, please, file a separate bug about the issue with such macros. Feel free to use the information I supplied and also put more specifics why could such inconsistent visibility lead to incorrect work of the linker.

Hi Peter,

resolving the issues with visibility is on our critical path for LTO launch. Please, let me know if you don't currently have the capacity to file an issue discussed above. In this case, I will try my best and file it myself.

Filed  bug 609564 

Thank you, Peter!

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/c5197d2ab04247a33e6934b576cdbd071ba562c2

commit c5197d2ab04247a33e6934b576cdbd071ba562c2
Author: krasin <krasin@google.com>
Date: Mon May 16 23:44:09 2016

Roll media/cdm/api deps.

This CL brings fixes for visibility of types used across Chromium->CDM bridge. See https://chromiumcodereview.appspot.com/1956123002/ and https://codereview.chromium.org/1964083003/

Also, removing now unnecessary lines in CFI blacklist.

BUG= 609564 ,  609175 ,  515347 

Review-Url: https://codereview.chromium.org/1959093004
Cr-Commit-Position: refs/heads/master@{#393981}

[modify] https://crrev.com/c5197d2ab04247a33e6934b576cdbd071ba562c2/DEPS
[modify] https://crrev.com/c5197d2ab04247a33e6934b576cdbd071ba562c2/tools/cfi/blacklist.txt

media_unittests are fixed per 'UBSanVptr Linux' buildbot: https://build.chromium.org/p/chromium.fyi/builders/UBSanVptr%20Linux/builds/276/steps/media_unittests

There're some other tests red on that bot, but media_unittests are green.