Ben Pfaff [Sun, 24 May 2009 18:26:41 +0000 (11:26 -0700)]
New function parse_value() for parsing a value of specified width.
Occasionally a value of a given width needs to be parsed from syntax.
This commit adds a helper function for doing so and modifies a few pieces
of code to use it. Probably there are other places where it would be
useful that could not easily be found with "grep".
This commit also renames the range-parser code to value-parser and puts
the new function in there, as a natural generalization.
Ben Pfaff [Tue, 12 May 2009 03:08:19 +0000 (20:08 -0700)]
Make value_set_missing(), etc. tolerate values of width -1.
In some circumstances a value of width -1 crops up, e.g. when a case is
made from a dictionary that has had a variable deleted in the middle.
Such a value has no content at all. In the long run it should be possible
to get rid of these values entirely--their presence is a wart--but for now
the case and value code needs to tolerate them.
This fixes a segfault in the GUI when inserting a new case when the
datasheet case has a column with width -1 (due to deletion of a variable),
which was caused by case_set_missing() calling value_set_missing() for
the -1 width variable, which in turn was writing through an invalid
pointer.
Ben Pfaff [Mon, 11 May 2009 05:23:00 +0000 (22:23 -0700)]
Change "union value" to dynamically allocate long strings.
Until now, a single "union value" could hold a numeric value or a short
string value. A long string value (one longer than MAX_SHORT_STRING)
required a number of contiguous "union value"s. This situation was
inconvenient sometimes, because any occasion where a long string value
might be required (even if it was unlikely) required using dynamic
memory allocation.
With this change, a value of any type, regardless of whether it is numeric
or short or long string, occupies a single "union value". The internal
representation of short and long strings is now different, however: long
strings are now internally represented by a pointer to dynamically
allocated memory. This means that "union value"s must now be initialized
and uninitialized properly, to ensure that memory is properly allocated
and freed behind the scenese.
This change thus has a ripple effect on PSPP code that works with values.
In particular, code that deals with cases is greatly changed, because a
case now needs to know the type of each value that it contains. Thus, a
new concept called a "case prototype", which represents the type and
width of each value within a case, is introduced, and every place in PSPP
that creates a case must now create a corresponding prototype to go with
it. This is why this commit is so big.
As part of writing up this commit, it became clear that some code was poor
enough that it needed to be rewritten entirely. Therefore, CROSSTABS and
T-TEST are almost completely modified by this commit.
Ben Pfaff [Thu, 7 May 2009 05:58:01 +0000 (22:58 -0700)]
output: Add auxiliary data parameter to tab_dim.
Until now, the tab_dim function has not provided any way to pass auxiliary
data to the table dimensioning function. This commit adds this ability
and updates all the callers of tab_dim to do so.
Ben Pfaff [Tue, 5 May 2009 12:42:23 +0000 (05:42 -0700)]
model-checker: Add command-line parser for model checking options.
This adds a parser for command-line options to configure a set of
mc_options for running the model checker. It is used by an upcoming test
for the sparse_xarray. It might also make sense to break the datasheet
tests out of PSPP into a separate program using this parser.
Ben Pfaff [Tue, 5 May 2009 12:39:03 +0000 (05:39 -0700)]
Implement new command-line argument parser.
glibc has two option parsers, but neither one of them feels quite
right:
- getopt_long is simple, but not modular, in that there is no
easy way to make it accept multiple collections of options
supported by different modules.
- argp is more sophisticated and more complete, and hence more
complex. It still lacks one important feature for
modularity: there is no straightforward way for option groups
that are implemented independently to have separate auxiliary
data,
The parser implemented in this commit is meant to be simple and
modular. It is based internally on getopt_long.
The initial use for this option parser is for an upcoming commit of a test
program that has some of its own options and some from the model checker,
but it should also be appropriate for PSPP and PSPPIRE if anyone wants to
adapt them to use it.
Ben Pfaff [Tue, 5 May 2009 05:27:05 +0000 (22:27 -0700)]
model-checker: Revise advice on checking for duplicates.
Until now the documentation on the model checker has advised checking for
a duplicate state before checking for consistency, but in fact this can
cause bugs to be missed if only some paths to a given state yield
incorrect results. So revise the advice to check for consistency before
checking for a duplicate state.
Ben Pfaff [Tue, 5 May 2009 05:20:42 +0000 (22:20 -0700)]
model-checker: Add more progress functions.
The model checker supports "progress functions" that report the current
status of the model checking run. Until now the implementation only
exported a single progress function that printed a line of dots across
stderr. This commit moves the "fancy" progress function that was
previously part of the PSPP language code into the model checker itself
and adds an even more verbose progress function as well.
Ben Pfaff [Tue, 5 May 2009 05:33:48 +0000 (22:33 -0700)]
model-checker: Move summary printing function into model checker.
There is no reason that the model checker itself should not be able to
print a summary of its results. Until now, this code was buried in the
PSPP language code, but the model checker itself is a better place for it.
Ben Pfaff [Fri, 24 Apr 2009 04:09:12 +0000 (21:09 -0700)]
model-checker: Kill dependencies and move back to libpspp.
Commit 95b074ff3 "Moved the datasheet testing code out of
src/{libspp,data}" moved the model-checker implementation from libpspp
into language/tests because it depended on math/moments.h and
data/val-type.h, which violates the dependency structure of the PSPP
libraries.
However, now I want to use the model checker in a test that should not
need to use anything from language/tests, so this commit eliminates these
dependencies and moves the model checker back to src/libpspp.
Ben Pfaff [Tue, 5 May 2009 04:53:07 +0000 (21:53 -0700)]
sparse-array: Simplify code slightly.
Instead of checking whether the key is in range in each caller of
find_leaf_node, do it in find_leaf_node itself. This also allows checking
the cache before checking whether the key is in range, which might be an
optimization.
Ben Pfaff [Thu, 7 May 2009 03:22:09 +0000 (20:22 -0700)]
sparse-array: Improve iteration interface.
The sparse_array_scan function only supports iteration in the forward
direction and its interface is somewhat awkward. This commit replaces it
by four new functions that allow iteration in both forward and reverse
directions and have a more conventional interface.
Ben Pfaff [Tue, 5 May 2009 12:51:54 +0000 (05:51 -0700)]
sparse-array: Use __builtin_ctzl on GCC 4.0 or later, as an optimization.
This should be a worthwhile optimization in many cases, because
__builtin_ctzl compiles to a single machine instruction on x86, whereas
the generic implementation compiles to several.
Ben Pfaff [Fri, 24 Apr 2009 03:27:10 +0000 (20:27 -0700)]
range-set: Inline some simple functions.
Some of the range-set functions are very simple and worth inlining, so
move the definitions of those functions into range-set.h. This required
moving the definition of struct range_set and struct range_set_node into
the header. Some of the functions used internally by those functions had
to be moved too, and renamed as well since for internal use in range-set.c
their names did not respect the namespace.
Ben Pfaff [Fri, 24 Apr 2009 03:14:52 +0000 (20:14 -0700)]
range-set: Add test coverage for range_set_destroy(NULL).
"gcov -b" showed that range_set_destroy() was never called with a NULL
argument. There's no reason not to test that too (although of course it
is unlikely to be broken).
John Darrington [Sat, 16 May 2009 03:38:11 +0000 (11:38 +0800)]
Fix bug inserting rows and columns and rename state variable.
The code in psppire-data-editor was inspecting the variable
called "state" on the GtkWidget class whereas it should have
been looking at the PsppireSheet class.
To avoid any future confusion, PsppireSheet's "state" variable
has been renamed to select_status.
covariance-matrix.c (is_origin): Do not count a numeric variable as
the origin. Call is_origin only for categorical variables.
covariance-matrix.c (ordered_match_nodes): New function.
covariance-matrix.c (match_nodes): Use bitwise operators and
ordered_match_nodes for hashing instead of so many ifs.
covariance-matrix.c (hash_numeric_alpha): Use hash_value_short to hash
combinations of numeric and categorical variables.
covariance-matrix.c (get_value_from_subscript): New function to match
values and columns in a design matrix.
covariance-matrix.c: Removed unused design matrix struct sums from
struct covariance matrix. To get the means, use struct moments
instead.
covariance-matrix.c (covariance_accumulator_to_matrix): Use
design_matrix_set_element instead of covariance_matrix_insert. Drop
unused function covariance_matrix_insert.
covariance-matrix.c (covariance_accumulator_hash): Use hash_bytes instead of hsh_hash_bytes.
John Darrington [Thu, 16 Apr 2009 13:06:31 +0000 (21:06 +0800)]
Improve the style handling of the column/row title buttons.
It seems that the style of the button wasn't `attached' before
it was used. Further, use of the "buttondefault" detail on
gtk_paint_box made wierd looking buttons with some themes.
I don't pretend to fully understand all this, but these changes
seem to be an improvement.
John Darrington [Mon, 13 Apr 2009 23:07:14 +0000 (07:07 +0800)]
Fixed bug in q2c when munging hyphenated strings.
Q2c wasn't terminating strings generated by its munge
function, resulting in garbage being placed into generated
code. Thanks to Michel Boaventura for reporting this.
Ben Pfaff [Thu, 9 Apr 2009 04:55:31 +0000 (21:55 -0700)]
Use Bob Jenkins lookup3 hash instead of FNV.
The Jenkins lookup3 hash is superior to FNV in collision resistance,
avalanching, and performance on systems that do not have fast
multiplication. It also provides a good way to combine the result of
a previous hashing step with the current hash, using its "basis" argument.
This commit replaces the PSPP implementation of FNV with the Jenkins
lookup3 hash and updates all the current users.
In addition, John Darrington pointed out that commit dd2e61b4a
"Make create_iconv() properly distinguish converters by name"
unintentionally introduced gratuitous hash collisions, by causing
all converters where tocode and fromcode were the same to hash to
value 0, and converters where tocode and fromcode were swapped to
hash to the same value as each other. Using the "basis" argument to
the Jenkins hash properly, instead of just attempting to combine
hash values with XOR, fixes this problem.
Ben Pfaff [Thu, 9 Apr 2009 04:39:22 +0000 (21:39 -0700)]
NPAR TESTS: Consistently order variables in summary statistics.
The set of variables in the NPAR TESTS specs structure was ordered
randomly, according to however the hash function happened to arrange them.
Sort them by variable name, instead, so that they always appear in
alphabetical order in, e.g., descriptive statistics output.
The particular hash function PSPP uses now tends to order variables
alphabetically anyhow. The next commit changes the PSPP hash functions,
so fixing this in advance prevents having to update any test output.
Ben Pfaff [Wed, 8 Apr 2009 04:20:47 +0000 (21:20 -0700)]
SET LOCALE: Don't use lex_tokstr() after skipping to next token.
The return value from lex_tokstr() is only valid until the next call to
lex_get() (or another function that changes the current token), so don't
advance past the token until we're done with its string value.
Ben Pfaff [Wed, 8 Apr 2009 04:15:40 +0000 (21:15 -0700)]
Make create_iconv() properly distinguish converters by name.
The code in create_iconv() assumed that every pair of different converters
had a different hash value. This is a bad assumption: eventually, we will
be unlucky, and two different converters will hash to the same value, and
we will get a bad conversion. So we have to compare (and store) the
names of the codes that each converters converts to and from.
Also, compute the hash value without making an extra copy of fromcode
and tocode.
Ben Pfaff [Wed, 8 Apr 2009 04:02:14 +0000 (21:02 -0700)]
Add "x" prefix to calls to plain malloc(), calloc(), strdup(), realloc().
In review commit 503f53bfdde "Read dictionary encoding from data files"
I noticed uses of plain strdup() (not xstrdup()). Some greps showed that
there were several other uses of strdup(), as well as calloc(), malloc(),
and realloc(), in the source tree. This commit adds "x" prefixes to each
of them, to ensure proper error handling.
Ben Pfaff [Wed, 8 Apr 2009 03:48:10 +0000 (20:48 -0700)]
Remove redundant test in lex_is_idn().
Commit 90f346cc0 "Made var_is_valid_name more permissive" made characters
with values above 127 valid in identifiers by allowing them in
lex_is_id1() and lex_is_idn(), but since the latter includes the former
in its test the addition there is redundant and can be eliminated.