True boolean datatype (no NAs), coercion from and to logicals, integers and integer subscripts; fast boolean operators and fast summary statistics. With 'bit' vectors you can store true binary booleans {FALSE,TRUE} at the expense of 1 bit only, on a 32 bit architecture this means factor 32 less RAM and ~ factor 32 more speed on boolean operations. Due to overhead of R calls, actual speed gain depends on the size of the vector: expect gains for vectors of size > 10000 elements. Even for one-time boolean operations it can pay-off to convert to bit, the pay-off is obvious, when such components are used more than once. Reading from and writing to bit is approximately as fast as accessing standard logicals - mostly due to R's time for memory allocation. The package allows to work with pre-allocated memory for return values by calling .Call() directly: when evaluating the speed of C-access with pre-allocated vector memory, coping from bit to logical requires only 70% of the time for copying from logical to logical; and copying from logical to bit comes at a performance penalty of 150%. the package now contains further classes for representing logical selections: 'bitwhich' for very skewed selections and 'ri' for selecting ranges of values for chunked processing. All three index classes can be used for subsetting 'ff' objects (ff-2.1-0 and higher).
CHANGES IN bit VERSION 1.1-14
BUG FIXES
o bit[i] and bit[i]<-v now check for non-positive integers
which prevents a segfault when bit[NA] or bit[NA]<-v
CHANGES IN bit VERSION 1.1-13
USER VISIBLE CHANGES
o logical NA is now mapped to bit FALSE as in ff booleans
o extractor function '[.bit' with positive numeric subscripts
(integer, double, bitwhich) now behaves like '[.logical' and returns
NA for out-of-bound requests and no element for 0
o extractor function '[[.bit' with positive numeric (integer, double,
bitwhich) subscripts now behaves like '[[.logical' and throws an error
for out-of-bound requests
o extractor function '[.bit' with range index subscripts (ri)
subscripts now behaves like '[[.bit' and throws an error
for out-of-bound requests
o assignment functions '[<-.bit' and '[[<-.bit' with positive numeric
(integer, double, bitwhich) subscripts now behave like '[<-.logical' and
'[[<-.logical' and silently increase vector length if necessary
o assignment function '[<-.bit' with range index subscripts (ri) now
behaves like '[[<-.bit' and silently increases vector length if necessary
o rlepack() is now a generic with a method for class 'integer'
o rleunpack() is now a generic with a method for class 'rlepack'
o unique.rlepack() now gives correct results for unordered sequences
o anyDuplicated.rlepack() now returns the position of the first
duplicate and gives correct results for unordered sequences
TUNING
o The package can now compiled with 64bit words instead of 32bit words,
since we only measured a minor speedup, we left 32bit as the default.
BUG FIXES
o extractor and assignment functions now check for legal (positive)
subscript bounds, hence illegally large subscripts or zero no longer
cause memory violations
CHANGES IN bit VERSION 1.1-12
NEW FEATURES
o function still.identical() has been moved to here from package bit64
o generic 'clone' and methods clone.default and clone.list have been moved to here from package ff
BUG FIXES
o bit[bitwhich] is now subscripting properly (VALGRIND)
o UBSAN should no longer complain about left shift of int
(although that never was a problem)
CHANGES IN bit VERSION 1.1-10
TUNING
o function 'vecseq' now calls C-code when calling with the default
parameters 'concat=TRUE, eval=TRUE' (wish of Matthew Dowle)
BUG FIXES
o all.bit no longer ignores TRUE values in the second and following words
(spotted by Nelson Chen)
CHANGES IN bit VERSION 1.1-9
NEW FEATURES
o new function 'repeat.time' for adaptive timing
CODE ORGANIZATION
o generics for sorting and ordering have been moved from 'ff' to 'bit'
CHANGES IN bit VERSION 1.1-7
USER VISIBLE CHANGES
o all calls to 'seq.int' have been replaced by 'seq_along' or 'seq_len'
o most calls to 'cat' have been replaced by 'message'
BUG FIXES
o chunk.default now works with chunk(from=2, to=3, by=1) thanks to Edwin de Jonge
CHANGES IN bit VERSION 1.1-5
NEW FEATURES
o new utility functions setattr() and setattributes() allow to set attributes
by reference (unlike attr()<- attributes()<- without copying the object)
o new utility unattr() returns copy of input with attributes removed
USER VISIBLE CHANGES
o certain operations like creating a bit object are even faster now: need
half the time and RAM through the use of setattr() instead of attr()<-
o [.bit now decorates its logical return vector with attr(,'vmode')='boolean',
i.e. we retain the information that there are no NAs.
BUG FIXES
o .onLoad() no longer calls installed.packages() which substantially
improves startup time (thanks to Brian Ripley)
CHANGES IN bit VERSION 1.1-2
USER VISIBLE CHANGES
o The package now has a namespace
CHANGES IN bit VERSION 1.1-1
USER VISIBLE CHANGES
o Function 'chunk' has been made generic, the default method
provides the previous behavior.
o New method to increase length of bitwhich objects.
o Added further coercion methods.
provides the previous behavior.
BUG FIXES
o as.bitwhich.ri now generates correct negative subscripts.
CHANGES IN bit VERSION 1.1-0
NEW FEATURES
o New class 'bitwhich' stores subscript positions in most efficient way:
TRUE for all()==TRUE, FALSE for !any()==TRUE. otherwise positive or
negative subscripts, whatever needs less RAM. Coercion functions and
logical operators are available, the latter being efficient for very
asymetric (skewed) distributions: selecting or exlcuding small factions
of the data.
o New class 'ri' (range index) allows to select ranges of positions for
chunked processing: all three classes 'bit', 'bitwhich' and 'ri' can be
used for subsetting 'ff' objects (ff-2.1.0 and higher).
o New c() method for 'bit' and 'bitwhich' objects which behaves like
c(logical).
o The bit methods sum(), any(), all(), min(), max(), range(), summary()
and which() now support a range argument that allows to restrict the
range of evaluation for chunked processing.
o New utilities for chunked processing: bbatch, repfromto, chunk, vecseq.
USER VISIBLE CHANGES
o reducing length of bit objects will now set hidden bits to FALSE,
such that subsequent length increase behaves consistent with bit
objects that had never been reduced in length: new bits are FALSE
o 'which' is no longer turned into a generic. Use 'bitwhich' instead,
or, 'as.which' if you need strictly positive subscripts.
o 'which.bit' has been renamed to 'as.which.bit'. It no longer has
parameter 'negative' and always returns positive subscripts (wish of
Stavros Macrakis). It now has second parameter 'range' in order to return
subscripts for chunked processing (note that the bitwhich representation
is not suitable for chunked processing). In order to facilitate coercion,
the return vector of 'as.which' now has class 'which'.
o the internal structure of a bit object has been changed to align with ff
ram objects: the bitlength of a bit object is no longer stored in
attr(bit, "n"), instead in attr(attr(bit, "physical"), "Length"),
which is accessible via physical(bit)$Length, but should be accessed
usually via length(bit).
o the semantics of 'min', 'max' and 'range' have been changed. They now
refer to the positions of TRUE in the bit vector (and thus are consistent
with bitwhich rather than with logical. The 'summary' method now returns
four elements c("FALSE"=, "TRUE"=, "Min."=, "Max."=).
BUG FIXES
o which.bit no longer returns integer() for a bit vector that has all TRUE
KNOWN PROBLEMS / TODOs
o NAs are mapped to TRUE in 'bit' and to FALSE in 'ff' booleans. Might be aligned
in a future release. Don't use bit if you have NAs - or map NAs explicitely.