bit64 (version 4.0.5)

bit64-package: A S3 class for vectors of 64bit integers


Package 'bit64' provides fast serializable S3 atomic 64bit (signed) integers that can be used in vectors, matrices, arrays and data.frames. Methods are available for coercion from and to logicals, integers, doubles, characters and factors as well as many elementwise and summary functions.

Version 0.8 With 'integer64' vectors you can store very large integers at the expense of 64 bits, which is by factor 7 better than 'int64' from package 'int64'. Due to the smaller memory footprint, the atomic vector architecture and using only S3 instead of S4 classes, most operations are one to three orders of magnitude faster: Example speedups are 4x for serialization, 250x for adding, 900x for coercion and 2000x for object creation. Also 'integer64' avoids an ongoing (potentially infinite) penalty for garbage collection observed during existence of 'int64' objects (see code in example section).

Version 0.9 Package 'bit64' - which extends R with fast 64-bit integers - now has fast (single-threaded) implementations the most important univariate algorithmic operations (those based on hashing and sorting). We now have methods for 'match', ' 'quantile', 'median' and 'summary'. Regarding data management we also have novel generics 'unipos' (positions of the unique values), 'tiepos' ( positions of ties), 'keypos' (positions of foreign keys in a sorted dimension table) and derived methods 'as.factor' and 'as.ordered'. This 64- bit functionality is implemented carefully to be not slower than the respective 32-bit operations in Base R and also to avoid outlying waiting times observed with 'order', 'rank' and 'table' (speedup factors 20/16/200 respective). This increases the dataset size with wich we can work truly interactive. The speed is achieved by simple heuristic optimizers in high- level functions choosing the best from multiple low-level algorithms and further taking advantage of a novel caching if activated. In an example R session using a couple of these operations the 64-bit integers performed 22x faster than base 32-bit integers, hash-caching improved this to 24x, sortorder-caching was most efficient with 38x (caching hashing and sorting is not worth it with 32x at duplicated RAM consumption).


 # S3 method for integer64
 # S3 method for integer64
length(x) <- value
 # S3 method for integer64
print(x, quote=FALSE, …)
 # S3 method for integer64
str(object, vec.len  = strO$vec.len, give.head = TRUE, give.length = give.head, …)



length of vector using integer


an integer64 vector


an integer64 vector


an integer64 vector of values to be assigned


logical, indicating whether or not strings should be printed with surrounding quotes.


see str


see str


see str

further arguments to the NextMethod


integer64 returns a vector of 'integer64', i.e. a vector of double decorated with class 'integer64'.

Design considerations

64 bit integers are related to big data: we need them to overcome address space limitations. Therefore performance of the 64 bit integer type is critical. In the S language -- designed in 1975 -- atomic objects were defined to be vectors for a couple of good reasons: simplicity, option for implicit parallelization, good cache locality. In recent years many analytical databases have learnt that lesson: column based data bases provide superior performance for many applications, the result are products such as MonetDB, Sybase IQ, Vertica, Exasol, Ingres Vectorwise. If we introduce 64 bit integers not natively in Base R but as an external package, we should at least strive to make them as 'basic' as possible. Therefore the design choice of bit64 not only differs from int64, it is obvious: Like the other atomic types in Base R, we model data type 'integer64' as a contiguous atomic vector in memory, and we use the more basic S3 class system, not S4. Like package int64 we want our 'integer64' to be serializeable, therefore we also use an existing data type as the basis. Again the choice is obvious: R has only one 64 bit data type: doubles. By using doubles, integer64 inherits some functionality such as is.atomic, length, length<-, names, names<-, dim, dim<-, dimnames, dimnames.

Our R level functions strictly follow the functional programming paragdim: no modification of arguments or other sideffects. Before version 0.93 we internally deviated from the strict paradigm in order to boost performance. Our C functions do not create new return values, instead we pass-in the memory to be returned as an argument. This gives us the freedom to apply the C-function to new or old vectors, which helps to avoid unnecessary memory allocation, unnecessary copying and unnessary garbage collection. Prior to 0.93 within our R functions we also deviated from conventional R programming by not using attr<- and attributes<- because they always did new memory allocation and copying in older R versions. If we wanted to set attributes of return values that we have freshly created, we instead used functions setattr and setattributes from package bit. From version 0.93 setattr is only used for manipulating cache objects, in ramsort.integer64 and sort.integer64 and in

Arithmetic precision and coercion

The fact that we introduce 64 bit long long integers -- without introducing 128-bit long doubles -- creates some subtle challenges: Unlike 32 bit integers, the integer64 are no longer a proper subset of double. If a binary arithmetic operation does involve a double and a integer, it is a no-brainer to return double without loss of information. If an integer64 meets a double, it is not trivial what type to return. Switching to integer64 limits our ability to represent very large numbers, switching to double limits our ability to distinguish x from x+1. Since the latter is the purpose of introducing 64 bit integers, we usually return integer64 from functions involving integer64, for example in c, cbind and rbind.

Different from Base R, our operators +, -, %/% and %% coerce their arguments to integer64 and always return integer64.

The multiplication operator * coerces its first argument to integer64 but allows its second argument to be also double: the second argument is internaly coerced to 'long double' and the result of the multiplication is returned as integer64.

The division / and power ^ operators also coerce their first argument to integer64 and coerce internally their second argument to 'long double', they return as double, like sqrt, log, log2 and log10 do.

argument1 op argument2 -> coerced1 op coerced2 -> result
integer64 + double -> integer64 + integer64 -> integer64
double + integer64 -> integer64 + integer64 -> integer64
integer64 - double -> integer64 - integer64 -> integer64
double - integer64 -> integer64 - integer64 -> integer64
integer64 %/% double -> integer64 %/% integer64 -> integer64
double %/% integer64 -> integer64 %/% integer64 -> integer64
integer64 %% double -> integer64 %% integer64 -> integer64
double %% integer64 -> integer64 %% integer64 -> integer64
integer64 * double -> integer64 * long double -> integer64
double * integer64 -> integer64 * integer64 -> integer64
integer64 / double -> integer64 / long double -> double
double / integer64 -> integer64 / long double -> double
integer64 ^ double -> integer64 / long double -> double
double ^ integer64 -> integer64 / long double -> double

Creating and testing S3 class 'integer64'

Our creator function integer64 takes an argument length, creates an atomic double vector of this length, attaches an S3 class attribute 'integer64' to it, and that's it. We simply rely on S3 method dispatch and interpret those 64bit elements as 'long long int'.

is.double currently returns TRUE for integer64 and might return FALSE in a later release. Consider is.double to have undefined behaviour and do query is.integer64 before querying is.double. The methods is.integer64 and is.vector both return TRUE for integer64. Note that we did not patch storage.mode and typeof, which both continue returning 'double' Like for 32 bit integer, mode returns 'numeric' and as.double) tries coercing to double). It is possible that 'integer64' becomes a vmode in package ff.

Further methods for creating integer64 are range which returns the range of the data type if calles without arguments, rep, seq.

For all available methods on integer64 vectors see the index below and the examples.

Index of implemented methods

creating,testing,printing see also description
NA_integer64_ NA_integer_ NA constant
integer64 integer create zero atomic vector
runif64 runif create random vector
rep.integer64 rep
seq.integer64 seq
is.integer64 is
is.integer inherited from Base R
is.vector.integer64 is.vector
identical.integer64 identical
length<-.integer64 length<-
length inherited from Base R
names<- inherited from Base R
names inherited from Base R
dim<- inherited from Base R
dim inherited from Base R
dimnames<- inherited from Base R
dimnames inherited from Base R
str inherited from Base R, does not print values correctly
print.integer64 print
str.integer64 str
coercing to integer64 see also
description as.integer64
generic as.integer64.bitstring as.bitstring
as.integer64.character character
as.integer64.double double
as.integer64.integer integer
as.integer64.integer64 integer64
as.integer64.logical logical
as.integer64.NULL NULL
coercing from integer64
see also description as.bitstring
as.bitstring generic as.bitstring.integer64
as.character as.double.integer64
as.double as.integer.integer64
as.integer as.logical.integer64
data structures see also description
c.integer64 c vector concatenate
cbind.integer64 cbind column bind
rbind.integer64 rbind row bind coerce atomic object to data.frame
data.frame inherited from Base R since we have coercion
subscripting see also
description [.integer64 [
vector and array extract [<-.integer64 [<-
vector and array assign [[.integer64 [[
scalar extract [[<-.integer64 [[<-
scalar assign binary operators
see also description +.integer64
+ returns integer64 -.integer64
- returns integer64 *.integer64
* returns integer64 ^.integer64
^ returns double /.integer64
/ returns double %/%.integer64
%/% returns integer64 %%.integer64
%% returns integer64
comparison operators see also description
==.integer64 ==
!=.integer64 !=
<.integer64 <
<=.integer64 <=
>.integer64 >
>=.integer64 >=
logical operators see also
description !.integer64 !
&.integer64 &
|.integer64 |
xor.integer64 xor
math functions
see also description returns logical format.integer64
format returns character abs.integer64
abs returns integer64 sign.integer64
sign returns integer64 log.integer64
log returns double log10.integer64
log10 returns double log2.integer64
log2 returns double sqrt.integer64
sqrt returns double ceiling.integer64
ceiling dummy returning its argument floor.integer64
floor dummy returning its argument trunc.integer64
trunc dummy returning its argument round.integer64
round dummy returning its argument signif.integer64
signif dummy returning its argument
cumulative functions see also description
cummin.integer64 cummin
cummax.integer64 cummax
cumsum.integer64 cumsum
cumprod.integer64 cumprod
diff.integer64 diff
summary functions see also
description range.integer64 range
min.integer64 min
max.integer64 max
sum.integer64 sum
mean.integer64 mean
prod.integer64 prod
all.integer64 all
any.integer64 any
algorithmically complex functions
see also description (caching) match.integer64
match position of x in table (h//o/so) %in%.integer64
%in% is x in table? (h//o/so) duplicated.integer64
duplicated is current element duplicate of previous one? (h//o/so) unique.integer64
unique (shorter) vector of unique values only (h/s/o/so) unipos.integer64
unipos positions corresponding to unique values (h/s/o/so) tiepos.integer64
tiepos positions of values that are tied (//o/so) keypos.integer64
keypos position of current value in sorted list of unique values (//o/so) as.factor.integer64
as.factor convert to (unordered) factor with sorted levels of previous values (//o/so) as.ordered.integer64
as.ordered convert to ordered factor with sorted levels of previous values (//o/so) table.integer64
table unique values and their frequencies (h/s/o/so) sort.integer64
sort sorted vector (/s/o/so) order.integer64
order positions of elements that would create sorted vector (//o/so) rank.integer64
rank (average) ranks of non-NAs, NAs kept in place (/s/o/so) quantile.integer64
quantile (existing) values at specified percentiles (/s/o/so) median.integer64
median (existing) value at percentile 0.5 (/s/o/so) summary.integer64
summary (/s/o/so) all.equal.integer64
all.equal test if two objects are (nearly) equal (/s/o/so)
helper functions see also description
minusclass minusclass removing class attritbute
plusclass plusclass inserting class attribute
binattr binattr define binary op behaviour
tested I/O functions see also
description read.table
inherited from Base R write.table
inherited from Base R serialize
inherited from Base R unserialize
inherited from Base R save
inherited from Base R load
inherited from Base R dput
inherited from Base R dget
inherited from Base R creating,testing,printing see also

Limitations inherited from implementing 64 bit integers via an external package

  • vector size of atomic vectors is still limited to .Machine$integer.max. However, external memory extending packages such as ff or bigmemory can extend their address space now with integer64. Having 64 bit integers also help with those not so obvious address issues that arise once we exchange data with SQL databases and datawarehouses, which use big integers as surrogate keys, e.g. on indexed primary key columns. This puts R into a relatively strong position compared to certain commercial statistical softwares, which sell database connectivity but neither have the range of 64 bit integers, nor have integers at all, nor have a single numeric data type in their macro-glue-language.

  • literals such as 123LL would require changes to Base R, up to then we need to write (and call) as.integer64(123L) or as.integer64(123) or as.integer64('123'). Only the latter allows to specify numbers beyond Base R's numeric data types and therefore is the recommended way to use -- using only one way may facilitate migrating code to literals at a later stage.

Limitations inherited from Base R, Core team, can you change this?

  • identical with default parameters does not distinguish all bit-patterns of doubles. For testing purposes we provide a wrapper identical.integer64 that will distinguish all bit-patterns. It would be desireable to have a single call of identical handle both, double and integer64.

  • the colon operator : officially does not dispatches S3 methods, however, we have made it generic

         from <- lim.integer64()[1]
         to <- from+99

    As a limitation remains: it will only dispatch at its first argument from but not at its second to.

  • is.double does not dispatches S3 methods, However, we have made it generic and it will return FALSE on integer64.

  • c only dispatches c.integer64 if the first argument is integer64 and it does not recursively dispatch the proper method when called with argument recursive=TRUE Therefore


    does not work and for now you can only call

  • generic binary operators fail to dispatch *any* user-defined S3 method if the two arguments have two different S3 classes. For example we have two classes bit and bitwhich sparsely representing boolean vectors and we have methods &.bit and &.bitwhich. For an expression involving both as in bit & bitwhich, none of the two methods is dispatched. Instead a standard method is dispatched, which neither handles bit nor bitwhich. Although it lacks symmetry, the better choice would be to dispatch simply the method of the class of the first argument in case of class conflict. This choice would allow authors of extension packages providing coherent behaviour at least within their contributed classes. But as long as none of the package authors methods is dispatched, he cannot handle the conflicting classes at all.

  • unlist is not generic and if it were, we would face similar problems as with c()

  • vector with argument mode='integer64' cannot work without adjustment of Base R

  • as.vector with argument mode='integer64' cannot work without adjustment of Base R

  • is.vector does not dispatch its method is.vector.integer64

  • mode<- drops the class 'integer64' which is returned from as.integer64. Also it does not remove an existing class 'integer64' when assigning mode 'integer'.

  • storage.mode<- does not support external data types such as as.integer64

  • matrix does drop the 'integer64' class attribute.

  • array does drop the 'integer64' class attribute. In current R versions (1.15.1) this can be circumvented by activating the function as.vector.integer64 further down this file. However, the CRAN maintainer has requested to remove as.vector.integer64, even at the price of breaking previously working functionality of the package.

  • str does not print the values of integer64 correctly

further limitations

  • subscripting non-existing elements and subscripting with NAs is currently not supported. Such subscripting currently returns 9218868437227407266 instead of NA (the NA value of the underlying double code). Following the full R behaviour here would either destroy performance or require extensive C-coding.


Package: bit64
Type: Package
Version: 0.5.0
Date: 2011-12-12
License: GPL-2
LazyLoad: yes
Encoding: latin1

See Also

integer in base R


Run this code
message("Using integer64 in vector")
x <- integer64(8)    # create 64 bit vector
is.atomic(x)         # TRUE
is.integer64(x)      # TRUE
is.numeric(x)        # TRUE
is.integer(x)        # FALSE - debatable
is.double(x)         # FALSE - might change
x[] <- 1:2           # assigned value is recycled as usual
x[1:6]               # subscripting as usual
length(x) <- 13      # changing length as usual
rep(x, 2)            # replicate as usual
seq(as.integer64(1), 10)     # seq.integer64 is dispatched on first given argument
seq(to=as.integer64(10), 1)  # seq.integer64 is dispatched on first given argument
seq.integer64(along.with=x)  # or call seq.integer64 directly
# c.integer64 is dispatched only if *first* argument is integer64 ...
x <- c(x,runif(length(x), max=100)) 
# ... and coerces everything to integer64 - including double
names(x) <- letters  # use names as usual

message("Using integer64 in array - note that 'matrix' currently does not work")
message("as.vector.integer64 removed as requested by the CRAN maintainer")
message("as consequence 'array' also does not work anymore")
# }
<!-- %y <- array(as.integer64(NA), dim=c(3,4), dimnames=list(letters[1:3], LETTERS[1:4])) -->
# }
message("we still can create a matrix or array by assigning 'dim'")
y <- rep(as.integer64(NA), 12)
dim(y) <- c(3,4)
dimnames(y) <- list(letters[1:3], LETTERS[1:4])
y["a",] <- 1:2       # assigning as usual
y[1:2,-4]            # subscripting as usual
# cbind.integer64 dispatched on any argument and coerces everything to integer64
cbind(E=1:3, F=runif(3, 0, 100), G=c("-1","0","1"), y)

message("Using integer64 in data.frame")
d <- data.frame(x=x, y=runif(length(x), 0, 100))

message("Using integer64 with csv files")
fi64 <- tempfile()
write.csv(d, file=fi64, row.names=FALSE)
e <- read.csv(fi64, colClasses=c("integer64", NA))

message("Serializing and unserializing integer64")
dput(d, fi64)
e <- dget(fi64)
e <- d[,]
save(e, file=fi64)

### A couple of unit tests follow hidden in a dontshow{} directive ###
# }
# }
message("== Differences between integer64 and int64 ==")

message("-- integer64 is atomic --")

message("-- The following performance numbers are measured under RWin64  --")
message("-- under RWin32 the advantage of integer64 over int64 is smaller --")

message("-- integer64 needs 7x/5x less RAM than int64 under 64/32 bit OS 
(and twice the RAM of integer as it should be) --")

message("-- integer64 creates 2000x/1300x faster than int64 under 64/32 bit OS
(and 3x the time of integer) --")
t32 <- system.time(integer(1e8))
t64 <- system.time(integer64(1e8))
#T64 <- system.time(int64(1e7))*10  # using 1e8 as above stalls our R on an i7 8 GB RAM Thinkpad

i32 <- sample(1e6)
d64 <- as.double(i32)

message("-- the following timings are rather conservative since timings
 of integer64 include garbage collection -- due to looped calls")
message("-- integer64 coerces 900x/100x faster than int64 
 under 64/32 bit OS (and 2x the time of coercing to integer) --")
t32 <- system.time(for(i in 1:1000)as.integer(d64))
t64 <- system.time(for(i in 1:1000)as.integer64(d64))
#T64 <- system.time(as.int64(d64))*1000
td64 <- system.time(for(i in 1:1000)as.double(i32))
t64 <- system.time(for(i in 1:1000)as.integer64(i32))
#T64 <- system.time(for(i in 1:10)as.int64(i32))*100

message("-- integer64 serializes 4x/0.8x faster than int64 
 under 64/32 bit OS (and less than 2x/6x the time of integer or double) --")
t32 <- system.time(for(i in 1:10)serialize(i32, NULL))
td64 <- system.time(for(i in 1:10)serialize(d64, NULL))
i64 <- as.integer64(i32); 
t64 <- system.time(for(i in 1:10)serialize(i64, NULL))
rm(i64); gc()
#I64 <- as.int64(i32); 
#T64 <- system.time(for(i in 1:10)serialize(I64, NULL))
#rm(I64); gc()

message("-- integer64 adds 250x/60x faster than int64
 under 64/32 bit OS (and less than 6x the time of integer or double) --")
td64 <- system.time(for(i in 1:100)d64+d64)
t32 <- system.time(for(i in 1:100)i32+i32)
i64 <- as.integer64(i32); 
t64 <- system.time(for(i in 1:100)i64+i64)
rm(i64); gc()
#I64 <- as.int64(i32); 
#T64 <- system.time(for(i in 1:10)I64+I64)*10
#rm(I64); gc()

message("-- integer64 sums 3x/0.2x faster than int64 
(and at about 5x/60X the time of integer and double) --")
td64 <- system.time(for(i in 1:100)sum(d64))
t32 <- system.time(for(i in 1:100)sum(i32))
i64 <- as.integer64(i32); 
t64 <- system.time(for(i in 1:100)sum(i64))
rm(i64); gc()
#I64 <- as.int64(i32); 
#T64 <- system.time(for(i in 1:100)sum(I64))
#rm(I64); gc()

message("-- integer64 diffs 5x/0.85x faster than integer and double
(int64 version 1.0 does not support diff) --")
td64 <- system.time(for(i in 1:10)diff(d64, lag=2L, differences=2L))
t32 <- system.time(for(i in 1:10)diff(i32, lag=2L, differences=2L))
i64 <- as.integer64(i32); 
t64 <- system.time(for(i in 1:10)diff(i64, lag=2L, differences=2L))
rm(i64); gc()

message("-- integer64 subscripts 1000x/340x faster than int64
(and at the same speed / 10x slower as integer) --")
ts32 <- system.time(for(i in 1:1000)sample(1e6, 1e3))
t32<- system.time(for(i in 1:1000)i32[sample(1e6, 1e3)])
i64 <- as.integer64(i32); 
t64 <- system.time(for(i in 1:1000)i64[sample(1e6, 1e3)])
rm(i64); gc()
#I64 <- as.int64(i32); 
#T64 <- system.time(for(i in 1:100)I64[sample(1e6, 1e3)])*10
#rm(I64); gc()

message("-- integer64 assigns 200x/90x faster than int64
(and 50x/160x slower than integer) --")
ts32 <- system.time(for(i in 1:100)sample(1e6, 1e3))
t32 <- system.time(for(i in 1:100)i32[sample(1e6, 1e3)] <- 1:1e3)
i64 <- as.integer64(i32); 
i64 <- system.time(for(i in 1:100)i64[sample(1e6, 1e3)] <- 1:1e3)
rm(i64); gc()
#I64 <- as.int64(i32); 
#I64 <- system.time(for(i in 1:10)I64[sample(1e6, 1e3)] <- 1:1e3)*10
#rm(I64); gc()

tdfi32 <- system.time(dfi32 <- data.frame(a=i32, b=i32, c=i32))
tdfsi32 <- system.time(dfi32[1e6:1,])
fi32 <- tempfile()
tdfwi32 <- system.time(write.csv(dfi32, file=fi32, row.names=FALSE))
tdfri32 <- system.time(read.csv(fi32, colClasses=rep("integer", 3)))
rm(dfi32); gc()

i64 <- as.integer64(i32); 
tdfi64 <- system.time(dfi64 <- data.frame(a=i64, b=i64, c=i64))
tdfsi64 <- system.time(dfi64[1e6:1,])
fi64 <- tempfile()
tdfwi64 <- system.time(write.csv(dfi64, file=fi64, row.names=FALSE))
tdfri64 <- system.time(read.csv(fi64, colClasses=rep("integer64", 3)))
rm(i64, dfi64); gc()

#I64 <- as.int64(i32); 
#tdfI64 <- system.time(dfI64<-data.frame(a=I64, b=I64, c=I64))
#tdfsI64 <- system.time(dfI64[1e6:1,])
#fI64 <- tempfile()
#tdfwI64 <- system.time(write.csv(dfI64, file=fI64, row.names=FALSE))
#tdfrI64 <- system.time(read.csv(fI64, colClasses=rep("int64", 3)))
#rm(I64, dfI64); gc()

message("-- integer64 coerces 40x/6x faster to data.frame than int64
(and factor 1/9 slower than integer) --")
message("-- integer64 subscripts from data.frame 20x/2.5x faster than int64
 (and 3x/13x slower than integer) --")
message("-- integer64 csv writes about 2x/0.5x faster than int64
(and about 1.5x/5x slower than integer) --")
message("-- integer64 csv reads about 3x/1.5 faster than int64
(and about 2x slower than integer) --")

rm(i32, d64); gc()

message("-- investigating the impact on garbage collection: --")
message("-- the fragmented structure of int64 messes up R's RAM --")
message("-- and slows down R's gargbage collection just by existing --")

td32 <- double(21)
td32[1] <- system.time(d64 <- double(1e7))[3]
for (i in 2:11)td32[i] <- system.time(gc(), gcFirst=FALSE)[3]
for (i in 12:21)td32[i] <- system.time(gc(), gcFirst=FALSE)[3]

t64 <- double(21)
t64[1] <- system.time(i64 <- integer64(1e7))[3]
for (i in 2:11)t64[i] <- system.time(gc(), gcFirst=FALSE)[3]
for (i in 12:21)t64[i] <- system.time(gc(), gcFirst=FALSE)[3]

#T64 <- double(21)
#T64[1] <- system.time(I64 <- int64(1e7))[3]
#for (i in 2:11)T64[i] <- system.time(gc(), gcFirst=FALSE)[3]
#for (i in 12:21)T64[i] <- system.time(gc(), gcFirst=FALSE)[3]

#matplot(1:21, cbind(td32, t64, T64), pch=c("d","i","I"), log="y")
matplot(1:21, cbind(td32, t64), pch=c("d","i"), log="y")
# }
# }

Run the code above in your browser using DataLab