Bug 14871 – Linker errors with 2.068.0-rc1

While testing 2.068.0-rc1 against some EMSI internal code as well as a development version of DCD I ran into some linking issues. It seems that the linker is trying to resolve some symbols whose names end with "__arrayZ" and "__assertFiZv". These symbol names seem to always be at module scope. Example error messages: /path_partially_redacted/ext/containers/src/containers/unrolledlist.d:445: undefined reference to `_D10containers12unrolledlist8__assertFiZv' bin/dcd-server.o: In function `_D3std12experimental9allocator15building_blocks15stats_collector232__T14StatsCollectorTS3std12experimental9allocator15building_blocks6region137__T6RegionTS3std12experimental9allocator10mallocator10MallocatorVki16VE3std8typecons43__T4FlagVAyaa13_67726f77446f776e7761726473Z4Flagi0Z6RegionVmi1024Vmi0Z14StatsCollector6defineFAyaAAyaXAya': msgpack-d/src/msgpack.d:(.text._D3std12experimental9allocator15building_blocks15stats_collector232__T14StatsCollectorTS3std12experimental9allocator15building_blocks6region137__T6RegionTS3std12experimental9allocator10mallocator10MallocatorVki16VE3std8typecons43__T4FlagVAyaa13_67726f77446f776e7761726473Z4Flagi0Z6RegionVmi1024Vmi0Z14StatsCollector6defineFAyaAAyaXAya+0x83): undefined reference to `_D3std12experimental9allocator15building_blocks15stats_collector7__arrayZ' It should be possible to reproduce part of the error with a checkout of DCD at this version: https://github.com/Hackerpilot/DCD/tree/7724ab9dc63221d383dfcfdf78429c570eeba60c (Remember to run `git submodule update --init --recursive after checking out)

Due to fix issue 14828 and related, 2.068.0-rc1 has introduced a change to the generation of internal symbols __array, __assert, and __unittest_fail. I have opposite stance for the change, but it's rejected. The PR that contains the discussion: https://github.com/D-Programming-Language/dmd/pull/4851 If you want to get more detailed information, please wait the reply from Walter or Martin.

Essentially the change was unnecessary to fix regressions, but just added to make compiler implementation 'simple'. I think it was not good decision.

The issues is that you're not compiling the std.experimental.allocator.building_blocks modules. The fix is straightforward. - containers/src/std/experimental/allocator/mallocator.d\ - containers/src/std/experimental/allocator/package.d\ - containers/src/std/experimental/allocator/common.d\ - containers/src/std/experimental/allocator/gc_allocator.d\ - containers/src/std/experimental/allocator/building_blocks/allocator_list.d\ - containers/src/std/experimental/allocator/typed.d\ + $(shell find containers/src/std/experimental/allocator -name "*.d")\ We fixed 2 regressions by using a simpler mechanism to emit helper functions. Now this requires you to link against all modules you use, where previously you could sometimes get away without linking. https://github.com/D-Programming-Language/dmd/pull/4851

I was able to get all of my code to build and link. There really needs to be something in the changelog that warns people about this.

> The issues is that you're not compiling the > std.experimental.allocator.building_blocks modules. WTF? I hope this will be fixed before 2.068 will be released?

I mean on compiler side. Because this kind of error its really stupid. And it breaks code.

Not code but, it makes no clue what exactly is wrong. I can see poor soul on SO asking what is wrong.

Just for statistics... I was hit by this compiler regression as well, in my own code-base, not experimental/allocator. Since we normally are quite vary about breaking code, I'm surprised this change was allowed in. I did manage to solve it on my own, was just about to raise a ticket but found the duplicate.

I open a compiler fix not to break any user project builds. https://github.com/D-Programming-Language/dmd/pull/4867

Please read the discussion here: https://github.com/D-Programming-Language/dmd/pull/4851 The general idea is, if your code imports A, then you must link with A.obj. Note that Kenji, Martin and I discussed this for a while and did not reach a consensus.

I saw that there is an addition to the changelog that will help people fix this problem. I think that warning is good enough, so I'm willing to close this issue.

Changlog comment is better than nothing

Resolved as "wontfix".