| Age | Commit message (Collapse) | Author |
|---|
|
|
|
Make kconfiglib.py consistent with the utilities.
|
|
Have BOOL/TRISTATE/... match _T_BOOL/_T_TRISTATE/... so that _set_type()
doesn't have to do any conversion in the common case. Conversion is
still needed for _T_DEF_BOOL/_T_DEF_TRISTATE/..., but those are rare.
Also make use of UNKNOWN being falsy.
Piggyback some rearranging of global constants so that related stuff
appears closer together, for readability.
|
|
|
|
Saves around 10 ms of startup time when they aren't, as measured with
'-X importtime'.
|
|
|
|
Make the exception argument to _KconfigIOError() harder to miss.
|
|
The current error message talks a lot about $srctree, but $srctree is
seldom the culprit in practice. More common is
'source "$(SOME_ENV_VAR)/foo"`, where SOME_ENV_VAR hasn't been set.
Include the complete 'source ...' line for missing Kconfig files, and
mention unset environment variables as a hint. Only mention $srctree
briefly.
Also shorten the message when a .config can't be a found a bit. This
message would usually only be seen when working directly with the
library.
|
|
|
|
Use a set comprehension.
|
|
|
|
Came up in https://github.com/ulfalizer/Kconfiglib/issues/67.
|
|
|
|
Due to an oversight, '# CONFIG_FOO is not set' with FOO undefined only
triggered a warning about assigning an undefined symbol if FOO was never
referenced inside the Kconfig files.
|
|
Previously, menuconfig.py only prompted for saving the configuration if
.config didn't exist or the user changed symbol values within the
interface.
Also make it prompt for save if Kconfig symbols have been added,
removed, or have had their defaults changed, provided it would make the
saved .config differ from the loaded one.
This usually won't matter for correctness, because loading an outdated
configuration performs an implicit olddefconfig, but it's less
confusing.
Also add a Kconfig.missing_syms attribute that records all assignments
to undefined symbols in the most recently loaded .config file. This is
needed to implement the check for whether the saved .config would be
different.
As an unrelated change, always prompt for saving if a .config has been
loaded from within the menuconfig interface. The intention is probably
often to save the configuration somewhere else, even if it isn't
modified.
|
|
Also add quotes around the filename.
|
|
The functions must be defined before they can be put into 'functions'.
|
|
Say what was expected in the error message.
|
|
Gets rid of some lookups.
A bigger hotspot is all the _next_line() calls though. Wonder if there's
some nice restructuring to avoid them. Getting rid of the _reuse_tokens
flag would be nice too.
|
|
self._line is only used for error reporting, and the empty string
returned at EOF would never be shown. Move the assignment to after
readline() and use a local variable to hold the line instead, to get rid
of some property lookups.
|
|
More common. isspace() returns False for empty strings.
|
|
Had missed the '(default: None)', and it doesn't hurt to point it out in
the description either.
Point out that filename=None is the default in the write_config()
'filename' description too, though it already had the '(default: None)'.
|
|
|
|
|
|
if .config is a symlink, then the intention is probably to overwrite the
target, but rename()ing the symlink to .config.old interferes with that.
Use the shutil.copyfile() fallback instead if .config is a symlink.
|
|
Make the previously obligatory 'filename' argument to load_config() and
write_config() default to None, and have that implement the behavior
you'd usually want: read/write either KCONFIG_CONFIG or ".config" if
unset, and read the 'option defconfig_list' configuration file if
KCONFIG_CONFIG/".config" doesn't exist.
For load_config(), filename=None also allows the configuration file to
be missing without raising an error. load_config() returns True if a
local configuration file was loaded, which is useful to check in the
menuconfig (if no local configuration file exists, we always want to
prompt for saving the configuration when exiting).
Also add a 'verbose' argument (default True) to load_config() and
read_config() that makes them print which files were read/written in
filename=None mode.
Also generalize olddefconfig.py and oldconfig.py to not require there to
already be a local configuration file. This was a bit silly for
olddefconfig.py in particular.
Remove the examples/defconfig.py script. It's a duplicate of
olddefconfig.py.
|
|
Add a default-True 'save_old' flag to write_config(). If 'save_old' is
True and an existing configuration file is being overwritten, a copy of
the old configuration file is saved to .<filename>.old (e.g.
.config.old) in the same directory.
Errors are ignored, as the old configuration would usually just be a
nice-to-have, and not essential.
The same functionality could be added for minimal configuration files
and headers, but it's probably most useful for configuration files.
|
|
Consistent with elsewhere.
|
|
These type sanity checks have never hit once during development over the
years, and were inconsistently applied too. Remove them to simplify the
code.
Keep the InternalError exception for backwards compatibility, in case
something catches it.
|
|
If a symbol is removed from a Kconfig file, it makes sense to flag it as
changed, so that things that still (probably accidentally) depend on it
get rebuilt.
Saw a patch for the C tools with the same effect floating around, so
might as well add it already. The C tools had other brokeness as well
though...
|
|
Get rid of _indentation() and inline it into _parse_help(), adding some
simplifications and optimizations along the way.
Saves a few % of parsing of time. Help text parsing is surprisingly hot.
|
|
CONFIG_FOO=y and 'default y' and the like are more common than the n
versions, so test for them first.
Turning the tuples into sets would be even better on Python 3, as it
optimizes sets with constant keys into a LOAD_CONST, but it has a
performance penalty on Python 2.
|
|
Similar change to commit 4b23936 ("Speed up some token tests"), just for
more tests.
This especially helps for the "not a property line" case in
_parse_properties(), which checks against all the cases.
Saves 2-3% of parsing time together with the earlier change.
|
|
|
|
Turn some of the hotter membership test tuples into global sets, like
was already done for _TYPE_TOKENS. That saves some global lookups for
the tuple members and avoids repeatedly recreating tuples. It's 30%-50%
faster in a microbenchmark, even for two-element tuples (with global
lookups for the members).
|
|
Factors out some code.
Also use a quick 'is not UNKNOWN' test first inside it, which will
usually fail, since single-def symbols are more common. That avoids
building a tuple too.
|
|
Same deal as for the initial token, except we sometimes know that we're
dealing with the second token as well.
Inline _expect_nonconst_sym_and_eol() and _expect_str(), which are
single-use. That allows more specific error messages to be used as well.
Also tweak an outdated comment in _tokenize() re. None-termination.
Token fetching is more manual now.
|
|
|
|
Another possible optimization was missed in commit ab89ef6 ("Get rid of
_next_token() and _peek_token()"): The index of the initial token on a
line is known to be 0, so there's no need to check _tokens_i.
Also reads a bit clearer.
|
|
The warning for selecting/implying a choice symbol could be
misunderstood as saying that select/imply has no effect on choice
symbols in a particular case. Select/imply never has an effect on choice
symbols though.
Rephrase the warning to make it clearer.
|
|
The following cases were let through without a parse error (with the
extra tokens just being ignored):
- endif/endmenu/enchoice <extra tokens>
- default FOO <extra tokens> (though 'default FOO if' flagged an
error)
Make them generate an error.
|
|
These are pretty hot. Inline them to save a few % of parsing time.
They're pretty simple anyway.
_tokens_i was initialized to -1 to simplify the _next_token()
implementation. With _next_token() gone, initialize it to 0 instead,
which simplifies some other code.
|
|
|
|
The message has been in for three months now. Hopefully that was enough
for it to get noticed.
|
|
This is 30%-60% faster for both the matching and non-matching case, as
measured with timeit on Python 2.7 and 3.6, and saves at least a few
percent of total parsing time (and probably some evaluation time too).
isinstance(foo, tuple) is particularly slow.
Symbol and Choice instances are always created by us, so potential
subclassing shouldn't be a problem.
|
|
|
|
This allows some symbol values to be forced while running
all{def,no,yes,mod}config.py. See Documentation/kbuild/kconfig.txt in
the Linux kernel.
Add a helper function load_allconfig() to Kconfiglib to avoid code
duplication in the tools.
Also add functions for enabling/disabling the warning that's generated
when a symbol is assigned multiple times in a (set of) .config files and
the values differ. It should be disabled when merging the
KCONFIG_ALLCONFIG configuration file. Previously, only the warning
generated when the assigned values are identical could be disabled.
Disable all warnings related to assigning a symbol multiple times in
examples/merge_config.py as well.
|
|
|
|
|
|
For strings with no $ or \ in them (99.86% of all strings in the Linux
x86 Kconfigs), we can just find() the matching quote directly.
Saves a few % of tokenization time.
|
