NEWS

fastmatch 1.1-8 (2026-01-17)

adjust to API changes up to R 4.6.0
add .SAFE argument to ctapply() which determines whether the index and value vectors are cached between calls to FUN (.SAFE=FALSE) or if a clean set of vectors is allocated for each call (.SAFE=TRUE). The latter is now the default, because cached vectors may not be assigned and thus are only safe to use with some functions FUN (see ?ctapply and #1). Previous versions behaved like .SAFE=FALSE. Note that .SAFE=FALSE is not supported in R 4.5.x, because SETLENGTH has been declared illegal, but only R 4.6.0 has growable vector support, so use R < 4.5.0 or R >= 4.6.0 if needed. In general, the performance gain from .SAFE=FALSE is only noticeable if X is big (~>1e7), INDEX creates a moderately large number of groups and FUN is fast enough to not dominate the cost. For example, ctapply(x, y, sum, .SAFE=FALSE) on 1e8 elements with average group length of 100 is ~25% faster than .SAFE=TRUE, but if the group length is too small or too big then either the call cost dominates or there are not enough allocations to benefit from the cache.
add %!fin% which is the negation of %fin%

use C99 version of flexible arrays to avoid false-positives in strict array sanitizers.
PROTECT fix when falling back to R's match and coercion happened before.
bugfix: fast.hash returned the result instead of the 'table' argument. (#9)

bug fix: -0 in table= would never be matched, because it was (correctly) hashed as 0, but final comparison was 0 to -0 in the table. (#12)

don't call XLENGTH() on NULL objects
if compiled with -DCHECKHASH fastmatch will re-compute the hash table every time and compare it to the existing table to verify its consistency. This can be used to detect incorrect use of fastmatch, i.e., cases where the object is modified and the hash table is not removed.

fix protection bug in case when fmatch() falls back to R's match() because of unsupported types (thanks to Tomáš Kalibera)

add fmatch.hash() which will create a hash table that can be used later with fmatch(). This can be used in cases where attaching the hash to the table implicitly is not reliable.
added ctapply() - a fast version of tapply()
added coalesce() - fast way of grouping unique values into contiguous groups (in linear time).
added %fin% - a fast version of %in%
fastmatch now supports long vectors. Note that the hash function is the same as in R and thus it uses at most 32-bits, hence long vectors can be used, but they must have less than 2^32 (~4e9) unique values.
bugfix: matching reals against a table that contains NA or NaNs would not match the position of those but return NA instead.
bugfix: fix crash when a newly unserialized hash table is used (since the table hash is not stored during serialization).

some R functions (such as subset assignment like x[1] <- 2) can create a new object (with possibly modified content) and copy all attributes including the hash cache. If the original object was used as a table in fmatch(), the hash cache will be copied into the modified object and thus its cache will be possibly out of sync with the object. fmatch() will now identify such cases and discard the hash to prevent errorneous results.

match() coerces POSIXlt objects into characters, but so far fmatch() performed the match on the actual objects. Now fmatch() coerces POSIXlt object into characters just like match(), but note that you will lose the ability to perform fast lookups if the table is a POSIXlt object -- please use POSIXct objects (much more efficient) or use as.character() on the POSIXlt object to create a table that you want to re-use.

bugfix: nomatch was ignored in the fastmatch implementation (thanks to Enrico Schumann for reporting)