-
Notifications
You must be signed in to change notification settings - Fork 519
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
Make debug_info rules clear #1656
Conversation
Interesting. I still get the failures in CI. Will have to also investigate that one further. |
@ferd , the change looks good to me. Thanks! |
I found the problem; erl_opts are specifically merged in the profile-order so that if you go I'll rework the PR so it's done on the right direction. |
Current rebar3 uses debug_info rules where debug_info is added by default, and no_debug_info prevents the default from being added, and removes any explicit debug_info, if any. The problem is that if 'no_debug_info' is anywhere in the config for a run, it cannot be removed even with other profiles. additionally, no_debug_info ignores special tuples like {debug_info, {Mod, Data}} and {debug_info_key, Key}, which can be used to add debug info and encrypt it (in lieu of plain debug_info) respectively. This patch makes it so that the following rules are in place: - the last option seen takes priority, allowing profile overrides by the ordering rules the compiler expects - the overriden options shall be explicitly deleted to avoid confusing the compiler - any option related to debug info seen last cancels any no_debug_info that preceded it - any no_debug_info option seen last cancels all of the other debug_info options - if debug_info is seen last, it cancels out {debug_info_key, Key} - if {debug_info_key, Key} is seen last, it cancels out debug_info - All other options are left untouched in that context (defines can still be expanded and so on) This should allow proper profile rules to be followed. Note that erl_opt profile merging puts precedence on the *last* element of a list, to match the compiler.
e3f2b3b
to
bf7ec38
Compare
We got funny interactions since PR erlang#1656: - the last `debug_info`-related option seen in a list of options after profile merge is kept, allowing later profiles from overtaking earlier ones - if you go `rebar3 as a,b,c compile`, the options from profile A come before B, which come before C, so C's options win - overrides are applied in order of profile as well, giving a priority to a later profile than an earlier one - The values in overrides are prepended rather than suffixed to the existing list - this means if we have to overrides adding options, such as `default` adding `no_debug_info', and `dialyze` adding `debug_info`, the results are `[debug_info]` as `dialyze` is applied first, and then `[no_debug_info, debug_info]` as `default` overrides are applied - the final result is `no_debug_info` always winning when erl_opts are overriden specifically. only `debug_info` options are going to suffer this, and in the context of overrides. Other `erl_opts` should be fine. I'm not sure how that can be fixed at all. In the meanwhile, we can add support back while leaving the default to not having debug information. This is done by: - moving all `no_debug_info` options to the `prod` profile - forcing `prod` to be called by `./bootstrap` so that most people keep getting no debug info - anyone calling `rebar3 clean -a` and then rebuilding with `rebar3 escriptize` (i.e. rebar3 devs) get debug info going This is up for review and discussion.
We got funny interactions since PR erlang#1656: - the last `debug_info`-related option seen in a list of options after profile merge is kept, allowing later profiles from overtaking earlier ones - if you go `rebar3 as a,b,c compile`, the options from profile A come before B, which come before C, so C's options win - overrides are applied in order of profile as well, giving a priority to a later profile than an earlier one - The values in overrides are prepended rather than suffixed to the existing list - this means if we have to overrides adding options, such as `default` adding `no_debug_info', and `dialyze` adding `debug_info`, the results are `[debug_info]` as `dialyze` is applied first, and then `[no_debug_info, debug_info]` as `default` overrides are applied - the final result is `no_debug_info` always winning when erl_opts are overriden specifically. only `debug_info` options are going to suffer this, and in the context of overrides. Other `erl_opts` should be fine. I'm not sure how that can be fixed at all. In the meanwhile, we can add support back while leaving the default to not having debug information. This is done by: - moving all `no_debug_info` options to the `prod` profile - forcing `prod` to be called by `./bootstrap` so that most people keep getting no debug info - anyone calling `rebar3 clean -a` and then rebuilding with `rebar3 escriptize` (i.e. rebar3 devs) get debug info going This is up for review and discussion.
Current rebar3 uses debug_info rules where debug_info is added by
default, and no_debug_info prevents the default from being added, and
removes any explicit debug_info, if any. The problem is that if
'no_debug_info' is anywhere in the config for a run, it cannot be
removed even with other profiles. additionally, no_debug_info ignores
special tuples like {debug_info, {Mod, Data}} and {debug_info_key, Key},
which can be used to add debug info and encrypt it (in lieu of plain
debug_info) respectively.
This patch makes it so that the following rules are in place:
the ordering rules
not have to), the overriden options shall be explicitly deleted
that follows
debug_info options
still be expanded and so on)
This should allow proper profile rules to be followed. See #1653 for context.