The design of po4a seems to expect it to act as a build system,
running with a config file telling it which files to build. In
version 0.68-1, the po4a-gettextize tool has changed to error out
with a comment that it's not being run the way it expects.
See https://bugs.debian.org/1020821
po4a-updatepo works for now with 0.68-1, although it prints a deprecation
message.
Previously the config class had an operator bool and
it was a common pattern to use if(const config& = cfg.child(..)).
While this pattern was nice to use. It has severe drawbacks, in
particular it was unclear whether a function that took a config&
parameter allowed "invalid" configs, while most functions
did not, there were some that did. Furtheremore it lead to a few
buggy codes that were unconvered by this change (Not fixed though!),
in particular codes that tested local config objects that were
not references to being invalid, which could never be the case.
This commits replaces those with just `true` in order to not
change behaviour.
Some obvious cases were also removed including for example
things like `assert(config());` There is ony case in the ai code
that i'm not 100% sure of where one implementation of a virtual
function checked for an invalid config and another one that didn't.
With this, all code that check for a config child to be
present now uses config::optional_child which returns an object
that behaves similar to optional<(const) config&>, so it throws
on invalid dereferencing. But it also has operator[string] for
convinience, in particular to make is similary
easy to use the the previous `if (config& = .. child())`.
Also it has a tool DEBUG_CONFIG which tests whether all
optional_config values are checked before they are derefereneced.
Another method manditory_child was
added that throws when the key is not found, which replaces all
occurances of child() that did not check whether the result was
valid. This was neccecary (this= adding a new method instead of
renaming .child) to keep track of converted changes, and be sure
no occurances of child() were accidentally changed to the
throwing version.
We might want to rename one of mandatory_child or optional_child
to just child later. Not sure which one yet. I think it's better
to keep it in the current state (no config::child() ) for a while
though, so that people that currently used child() in their open
prs or other work get an error and not wrongly rely on the previous
behviour of config::child.
The interface of vconfig was not changed in this commit.
