* move internal _init and _shutdown functions in the Global
part of the code, as it is where they belong
* fix minor documentation stuff
SVN revision: 44730
_eina_stringshare_small_bucket_find() was receiving the plength with
size -1 since the first character is the same, but inside it was not
doing the same, thus it was always failing the first test based on
size.
SVN revision: 43634
Being able to indivually initialize individual modules was initially
"good", but at end it's putting complexities on users that would try
to "optimize" by doing just what they used, but in the end most people
would get them wrong, users would have to do lots of code and etc. At
the end it does not worth.
Most module init just register handful errors and log domains, so are
cheap. The exception is mempool users, that would dlopen() stuff, but
people that are concerned (embedded) can just compile those statically
in eina.
Since at the end any real application would use most of modules, we
actually end saving lots of function calls that would do nothing other
than increment a global counter.
I also did the init/shutdown use an array, making it easier to
maintain. The inital dependencies were analysed by a script I wrote, I
hope it's all right.
Please fix any breakages you find!
SVN revision: 42300
note that one can turn per module debug, for example:
EINA_LOG_LEVEL=4 EINA_LOG_LEVELS=eina_stringshare:0 ./bla
will enable level 4 (debug) for all modules except eina_stringshare
that is forced to 0 (just critical messages).
SVN revision: 42224
All these individual init functions are getting messy, some modules
lack them and it's easy to get inconsistent. Safety check needs error
and log, but these need safety checks as well, some modules (lalloc,
rbtree and others) use safety checks but provide no _init().
I want to know if we really gain something to init individual
modules. It should not be that expensive as init should not allocate
heavy resources and the recommendation is to call eina_init() so most
users will do that anyway.
If people agree I'll unmark all *_init() as EAPI and make them private
to eina lib.
SVN revision: 42214
safety checks will report null pointers and other error conditions on
public api's and can be disabled by compile time check.
note that in order to have these checks working we need to make
EINA_ARG_NONNULL() void, otherwise GCC can remove these checks since
they're known to be false.
This commit also make two minor changes:
* list and hash accessors and iterators are created even for empty
entities. This is correct in my point of view since NULL should
indicate error. Having these in were an optimziation, but not
worth it, these are not the most common case and hitting this path
is not of much cost.
* unmarked some parameters as nonnull, mainly on list and inlist.
SVN revision: 38327
This is a faster "add", if we know we're using a shared string we know
the node without any need to search it, just increment reference and
exit.
SVN revision: 37458
sizeof(Eina_Stringshare_Node) is now 24 bytes on 64bits platforms, but
str[] was pointing to before that, to the 20th byte, causing out of
bounds access.
Adding the padding will cause str[] to use the correct position. It
wastes 4 more bytes, like pre-optimizations, but it's just on big
machines.
SVN revision: 37305
mixing #ifdef'ed blocks inside code is bad, can lead to warnings if
some variables are not used and it's a pain to read.
instead, just define functions and always call them, choose their
implementation based on the ifdef macros. I opted to have 2
declarations, but one can go like other parts and #ifdef around the
function contents as well.
SVN revision: 37281
we don't need to use 'begin' flag (that takes a byte) just to see if
we're in the same memory block as the head, just do a pointer math.
SVN revision: 37270
trade off safety by speed, we will always assume str was previously
shared, like evas_stringshare_del() did and we can know with zero-cost
the number of references and can avoid strlen() too.
When references drop to zero we still have to do the hash, access the
bucket and then lookup the Red-Black tree, then walk the list of
nodes, but avoiding so would use more memory, unacceptable at this
point.
SVN revision: 37268
Vincent Torri
Gustavo Sverzut Barbieri
Jonathan Atton
Cedric BAIL
Sebastian Dransfeld
Carsten Haitzler