GLib's scribbling of freed memory is enabled on Arch so attempting to
read the deleted members of inst and book crash instead of reaturning
invalid results. These weren't really useful tests anyway.
The first attempt to fix this, a17bc85a, doesn't work because the
boost::date_time constructor gets enough information in most cases to
generate a date, just not the one we expect. This change looks for '-' in
the fourth position and if it's there assumes iso-extended format, otherwise
it assumes delimiter-less ISO without the 'T', i.e. %Y%m%d%H%M%S.
This allows direct conversion between Scheme numbers and gnc_numeric
without the performance or accuracy penalties arising from using doubles
as an intermediary.
This commit introduces a new feature flag:
GNC_FEATURE_GUID_FLAT_BAYESIAN. It signifies that the bayes import map
data are stored flat and by guid. Any time bayes import map data are
accessed, they are converted if necessary.
Before, it was necessary to remove '/' from tokens so that they won't be
divided up within kvp. Now that kvp doesn't parse tokens, it's okay to
pass '/', and it's better not to translate user-provided tokens if at all
possible.
The conversion assumed there were only three levels to bayes import
map kvp: IMAP token, user-supplied token, GUID/account name. In
actuality, since user-supplied tokens could have the delimiter in them,
there could be several. This fix takes that into account like so:
IMAP token, potentially several user-supplied tokens, GUID/account name.
The import map is undergoing two conversions at the same time: account names
to guids and an hierarchical representation to a flat representation in KVP.
The bayes data are stored in the KVP store. Before this commit, they are
stored under /import-map-bayes/<token>/<account guid>/count (where count
is the datum that "matters" in bayes matching).
The problem with this is that any token including the kvp delimiter
(currently '/') gets divided, and is not found correctly during bayes
kvp searching. The quickest solution to this is to replace all "/"
characters with some other character. That has been done, along with a
re-structuring of the bayes matching code to take advantage of c++
features to make the code more concise and readable.
Also modified some test functions to fix leaks and double-frees: the
same kvp value can't be in the kvp tree twice.
Also, when I added code to clean up after the tests, some things started
breaking due to double-delete. Apparently const_cast was hiding some
programming errors. Really? You don't say? When giving a GUID* to KvpValue,
the latter takes ownership of the former.
The nested representation was very noisy. Now, the string representation
shows one line per value with the full prefix which is also more
expressive than the old version.
The template avoids the need to cast to and from void*, and adds flexibility to
the targeted function's signature.
test-stuff.h defines a macro, "failure" which is used as an identifier
in the standard IO library, so I moved any inclusion of test-stuff.h to
the last include position so that "failure" wouldn't be defined before
the IO library was included.
Since Account.c is now Account.cpp, the function signatures look a bit
different internally. The tests rely on function signatures in error
messages. Instead of trying to figure out what the exact
function signature might be, I use a substring matching strategy to
ensure that the correct error was issued.
I.e., remove the shell invocation and with it the need to set the shebang.
Surprisingly this required some build-system modifications particularly
for cmake in order to correctly set the environment.
This will avoid a ninja-build from picking up a config.h generated by the autotools build
(in the root build directory). Picking up the wrong config.h may lead to all kinds of
subtle issues if the autotools run was done with different options than the cmake run.