Remove debug_info option when no_debug_info option set

[bitnitdit]

No description provided.

[ferd]

I'm not quite sure why you would add no_debug_info rather than removing the bad options. Is there a use case you have in mind specifically here? Overrides or profiles-related in any way? I'm also a bit confused about why this apparently killed all existing test runs.

[bitnitdit]

Using rebar3 to create a foo release with the default setting, run $ rebar3 as prod release, the .beam files generated are all compiled with debug_info.

I want to remove the debug_info compile option. So in rebar.config file, adding the no_debug_info option in the prod profile as the following:

{profiles, [{prod, [{erl_opts, [no_debug_info]},
                    {relx, [{dev_mode, false},{include_erts, true}]}]
            }]
}.

then run $ rebar3 as prod release again, but this time the .beam files generated are all compiled with debug_info as well. That is to say, no_debug_info option in prod profile does not work.

The only way to remove the debug_info is to delete the first line, {erl_opts, [debug_info]}., in rebar.config file. But IMHO the profile settings should have higher priority.

[bitnitdit]

Just walk through the following lines,

erl_opts(Opts) ->
    RawErlOpts = filter_defines(?MODULE:get(Opts, erl_opts, []), []),
    Defines = [{d, list_to_atom(D)} ||
                  D <- ?MODULE:get(Opts, defines, [])],
    AllOpts = Defines ++ RawErlOpts,
    case proplists:is_defined(no_debug_info, AllOpts) of
        true ->
            [O || O <- AllOpts, O =/= no_debug_info];
        false ->
            [debug_info|AllOpts]
    end.

we can see that no_debug_info option does not do anything to any other options, its usage is only remove itself. So no_debug_info option is useless, I think it's not the designed semantic for it.

[ferd]

Since debug_info is the default behaviour, you could just remove debug_info from the default profile and it would all work.

[bitnitdit]

Yes, just remove debug_info from the default profile is working.

But what's the point of the no_debug_info option? Why we have a useless option no_debug_info there?

[ferd]

The no_debug_info turns off the rebar behaviour of inserting debug_info by default in compile options, basically.

[ferd]

To be clear, I'm not against the patch, but it would at least need to keep tests working.

[bitnitdit]

From the link https://travis-ci.org/erlang/rebar3/jobs/295167011:

... ...
%%% rebar_as_SUITE ==> as_with_task_args: OK

=ERROR REPORT==== 31-Oct-2017::05:08:23 ===
** Generic server rebar_log_meck terminating 
** Last message in was {'EXIT',<0.2858.0>,
                        {#Ref<0.688754590.3314286593.37531>,27911,
                         {'EXIT',
                          {undef,
                           [{rebar_log_meck_original,default_level,[],[]},
                            {rebar_log,default_level,[],[]},
                            {rebar3,log_level,0,
                             [{file,
                               "/home/travis/build/erlang/rebar3/_build/test/lib/rebar/src/rebar3.erl"},
                              {line,258}]},    
... ...

It seems like the failed test is caused by the function rebar_log_meck_original:default_level was undefined; not caused by this PR's code change.

In my local env, the $ ./bootstrap && ./rebar3 ct test can pass smoothly with the same code change.

I can not have any idea why the test failed on travis-ci. So sorry.

[ferd]

I'll try to look into it today then. That's odd.

[ferd]

So I have figured why tests explode...

Currently the behaviour is that:

• you have debug_info on (by default) and then turn it off with no_debug_info, you get no debug info
• you have debug_info on explicitly and then turn it off with no_debug_info, you keep getting debug info

The current rebar3's test suites depend on the later: we publish in the default profile without debug_info to keep the size of the executables small. We then use profile options to turn debug_info on for dialyzer or Common Test (and meck uses that debug info to work).

By modifying the feature the way this PR does it, when no_debug_info is used in the default profile, then follow-up profiles cannot ever enable it either.

This PR just happens to reverse the problem you experienced, and in doing so, breaks rebar3's test suite itself.

I need to think of what could be a good way to fix this to work well with profiles. So far I think the best would work by honoring whichever is the first found between no_debug_info or any of the debug_info-related entries in the list and removing the related terms in the trailing end of the erl_opts list to ensure nice playing with profiles. I'm not sure what corner-cases this may hide.

[bitnitdit]

@ferd , thanks!

Sorry for this PR is not good enough to cover the case you figured out. :-(

[ferd]

It's fine anyway. I think there's a good way forward from here anyway. I'll try to open up an alternative PR that you can try to see if it works well for you.

[ferd]

Attempt at #1656

[bitnitdit]

@ferd , the change in #1656 looks good to me. Thanks!

[bitnitdit]

This PR has been superseded by PR #1656. Close it.