R 3.4.0 is released – with new speed upgrades and bug-fixes

Tal Galili

6 years ago

R 3.4.0 (codename “You Stupid Darkness”) was released 3 days ago. You can get the latest binaries version from here. (or the .tar.gz source code from here). The full list of bug fixes and new features is provided below.

As mentioned two months ago by David Smith, R 3.4.0 indicates several major changes aimed at improving the performance of R in various ways. These includes:

The JIT (‘Just In Time’) byte-code compiler is now enabled by default at its level 3. This means functions will be compiled on first or second use and top-level loops will be compiled and then run. (Thanks to Tomas Kalibera for extensive work to make this possible.) For now, the compiler will not compile code containing explicit calls to browser(): this is to support single stepping from the browser() call. JIT compilation can be disabled for the rest of the session using compiler::enableJIT(0) or by setting environment variable R_ENABLE_JIT to 0.
Matrix products now consistently bypass BLAS when the inputs have NaN/Inf values. Performance of the check of inputs has been improved. Performance when BLAS is used is improved for matrix/vector and vector/matrix multiplication (DGEMV is now used instead of DGEMM). One can now choose from alternative matrix product implementations via options(matprod = ). The “internal” implementation is not optimized for speed but consistent in precision with other summations in R (using long double accumulators where available). “blas” calls BLAS directly for best speed, but usually with undefined behavior for inputs with NaN/Inf.
Speedup in simplify2array() and hence sapply() and mapply() (for the case of names and common length #> 1), thanks to Suharto Anggono’s PR#17118.
Accumulating vectors in a loop is faster – Assigning to an element of a vector beyond the current length now over-allocates by a small fraction. The new vector is marked internally as growable, and the true length of the new vector is stored in the truelength field. This makes building up a vector result by assigning to the next element beyond the current length more efficient, though pre-allocating is still preferred. The implementation is subject to change and not intended to be used in packages at this time.
C-LEVEL FACILITIES have been extended.
Radix sort (which can be considered more efficient for some cases) is now chosen by method = “auto” for sort.int() for double vectors (and hence used for sort() for unclassed double vectors), excluding ‘long’ vectors. sort.int(method = “radix”) no longer rounds double vectors. The default method until R 3.2.0 was “shell”. A minimal comparison between the two shows that for very short vectors (100 values), “shell” would perform better. From a 1000 values, they are comparable, and for larger vectors – “radix” is doing 2-3 times faster (which is probably the use case for which we would care about more). More about this can be read in ?sort.int

#> 
#> set.seed(2017-04-24)
#> x  microbenchmark(shell = sort.int(x, method = "shell"), radix = sort.int(x, method = "radix"))
Unit: microseconds
  expr    min     lq     mean median     uq    max neval cld
 shell 15.775 16.606 17.80971 17.989 18.543 33.211   100  a 
 radix 32.657 34.595 35.67700 35.148 35.702 88.561   100   b
#> 
#> set.seed(2017-04-24)
#> x  microbenchmark(shell = sort.int(x, method = "shell"), radix = sort.int(x, method = "radix"))
Unit: microseconds
  expr    min     lq     mean median      uq    max neval cld
 shell 53.414 55.074 56.54395 56.182 57.0120 96.034   100   b
 radix 45.665 46.772 48.04222 47.325 48.1555 78.598   100  a 
#> 
#> set.seed(2017-04-24)
#> x  microbenchmark(shell = sort.int(x, method = "shell"), radix = sort.int(x, method = "radix"))
Unit: milliseconds
  expr      min       lq      mean    median        uq      max neval cld
 shell 93.33140 95.94478 107.75347 103.02756 115.33709 221.0800   100   b
 radix 38.18241 39.01516  46.47038  41.45722  47.49596 159.3518   100  a 
#> 
#>

More about the changes in R case be read at the nice post by David Smith, or in the list of changes given below.

Upgrading to R 3.4.0 on Windows

If you are using Windows you can easily upgrade to the latest version of R using the installr package. Simply run the following code in Rgui:

install.packages("installr") # install 
setInternet2(TRUE) # only for R versions older than 3.3.0
installr::updateR() # updating R.
# If you wish it to go faster, run: installr::updateR(T)

Running “updateR()” will detect if there is a new R version available, and if so it will download+install it (etc.). There is also a step by step tutorial (with screenshots) on how to upgrade R on Windows, using the installr package. If you only see the option to upgrade to an older version of R, then change your mirror or try again in a few hours (it usually take around 24 hours for all CRAN mirrors to get the latest version of R).

I try to keep the installr package updated and useful, so if you have any suggestions or remarks on the package – you are invited to open an issue in the github page.

CHANGES IN R 3.4.0

SIGNIFICANT USER-VISIBLE CHANGES

(Unix-alike) The default methods for download.file() and url() now choose "libcurl" except for file:// URLs. There will be small changes in the format and wording of messages, including in rare cases if an issue is a warning or an error. For example, when HTTP re-direction occurs, some messages refer to the final URL rather than the specified one.Those who use proxies should check that their settings are compatible (see ?download.file: the most commonly used forms work for both "internal" and "libcurl").
table() has been amended to be more internally consistent and become back compatible to R <= 2.7.2 again. Consequently, table(1:2, exclude = NULL) no longer contains a zero count for <NA#>, but useNA = "always" continues to do so.
summary.default() no longer rounds, but its print method does resulting in less extraneous rounding, notably of numbers in the ten thousands.
factor(x, exclude = L) behaves more rationally when x or L are character vectors. Further, exclude = <factor#> now behaves as documented for long.
Arithmetic, logic (&, |) and comparison (aka ‘relational’, e.g., <, ==) operations with arrays now behave consistently, notably for arrays of length zero.Arithmetic between length-1 arrays and longer non-arrays had silently dropped the array attributes and recycled. This now gives a warning and will signal an error in the future, as it has always for logic and comparison operations in these cases (e.g., compare matrix(1,1) + 2:3 and matrix(1,1) < 2:3).
The JIT (‘Just In Time’) byte-code compiler is now enabled by default at its level 3. This means functions will be compiled on first or second use and top-level loops will be compiled and then run. (Thanks to Tomas Kalibera for extensive work to make this possible.)For now, the compiler will not compile code containing explicit calls to browser(): this is to support single stepping from the browser() call.JIT compilation can be disabled for the rest of the session using compiler::enableJIT(0) or by setting environment variable R_ENABLE_JIT to 0.
xtabs() works more consistently with NAs, also in its result no longer setting them to 0. Further, a new logical option addNA allows to count NAs where appropriate. Additionally, for the case sparse = TRUE, the result’s dimnames are identical to the default case’s.
Matrix products now consistently bypass BLAS when the inputs have NaN/Inf values. Performance of the check of inputs has been improved. Performance when BLAS is used is improved for matrix/vector and vector/matrix multiplication (DGEMV is now used instead of DGEMM).One can now choose from alternative matrix product implementations via options(matprod = ). The "internal" implementation is not optimized for speed but consistent in precision with other summations in R (using long double accumulators where available). "blas" calls BLAS directly for best speed, but usually with undefined behavior for inputs with NaN/Inf.
factor() now uses order() to sort its levels, not sort.list(). This makes factor() support custom vector-like objects if methods for the appropriate generics are defined. This change has the side effect of making factor() succeed on empty or length-one non-atomic vector(-like) types (e.g., list), where it failed before.

NEW FEATURES

User errors such as integrate(f, 0:1, 2) are now caught.
Add signature argument to debug(), debugonce(), undebug() and isdebugged() for more conveniently debugging S3 and S4 methods. (Based on a patch by Gabe Becker.)
Add utils::debugcall() and utils::undebugcall() for debugging the function that would be called by evaluating the given expression. When the call is to an S4 generic or standard S3 generic, debugcall() debugs the method that would be dispatched. A number of internal utilities were added to support this, most notably utils::isS3stdGeneric(). (Based on a patch by Gabe Becker.)
Add utils::strcapture(). Given a character vector and a regular expression containing capture expressions, strcapture() will extract the captured tokens into a tabular data structure, typically a data.frame.
str() and strOptions() get a new option drop.deparse.attr with improved but changed default behaviour for expressions. For expressionobjects x, str(x) now may remove extraneous white space and truncate long lines.
str(<looooooooong_string#>) is no longer very slow; inspired by Mikko Korpela’s proposal in PR#16527.
str(x)‘s default method is more “accurate” and hence somewhat more generous in displaying character vectors; this will occasionally change Routputs (and need changes to some ‘*.Rout(.save)’ files).
For a classed integer vector such as x <- xtabs(~ c(1,9,9,9)), str(x) now shows both the class and "int", instead of only the latter.
isSymmetric(m) is much faster for large asymmetric matrices m via pre-tests and a new option tol1 (with which strict back compatibility is possible but not the default).
The result of eigen() now is of class "eigen" in the default case when eigenvectors are computed.
Zero-length date and date-time objects (of classes "POSIX[cl]?t") now print() “recognizably”.
xy.coords() and xyz.coords() get a new setLab option.
The method argument of sort.list(), order() and sort.int() gains an "auto" option (the default) which should behave the same as before when method was not supplied.
stopifnot(E, ..) now reports differences when E is a call to all.equal() and that is not true.
boxplot(<formula#>, *) gain optional arguments drop, sep, and lex.order to pass to split.default() which itself gains an argument lex.orderto pass to interaction() for more flexibility.
The plot() method for ppr() has enhanced default labels (xmin and main).
sample.int() gains an explicit useHash option (with a back compatible default).
identical() gains an ignore.srcref option which drops "srcref" and similar attributes when true (as by default).
diag(x, nrow = n) now preserves typeof(x), also for logical, integer and raw x (and as previously for complex and numeric).
smooth.spline() now allows direct specification of lambda, gets a hatvalues() method and keeps tol in the result, and optionally parts of the internal matrix computations.
addNA() is faster now, e.g. when applied twice. (Part of PR#16895.)
New option rstandard(<lm#>, type = "predicted") provides the “PRESS”–related leave-one-out cross-validation errors for linear models.
After seven years of deprecation, duplicated factor levels now produce a warning when printed and an error in levels<- instead of a warning.
Invalid factors, e.g., with duplicated levels (invalid but constructable) now give a warning when printed, via new function .valid.factor().
sessionInfo() has been updated for Apple’s change in OS naming as from ‘10.12’ (‘macOS Sierra’ vs ‘OS X El Capitan’).Its toLatex() method now includes the running component.
options(interrupt=) can be used to specify a default action for user interrupts. For now, if this option is not set and the error option is set, then an unhandled user interrupt invokes the error option. (This may be dropped in the future as interrupt conditions are not errorconditions.)
In most cases user interrupt handlers will be called with a "resume" restart available. Handlers can invoke this restart to resume computation. At the browser prompt the r command will invoke a "resume" restart if one is available. Some read operations cannot be resumed properly when interrupted and do not provide a "resume" restart.
Radix sort is now chosen by method = "auto" for sort.int() for double vectors (and hence used for sort() for unclassed double vectors), excluding ‘long’ vectors.sort.int(method = "radix") no longer rounds double vectors.
The default and data.frame methods for stack() preserve the names of empty elements in the levels of the ind column of the return value. Set the new drop argument to TRUE for the previous behavior.
Speedup in simplify2array() and hence sapply() and mapply() (for the case of names and common length #> 1), thanks to Suharto Anggono’s PR#17118.
table(x, exclude = NULL) now sets useNA = "ifany" (instead of "always"). Together with the bug fixes for this case, this recovers more consistent behaviour compatible to older versions of R. As a consequence, summary() for a logical vector no longer reports (zero) counts for NAwhen there are no NAs.
dump.frames() gets a new option include.GlobalEnv which allows to also dump the global environment, thanks to Andreas Kersting’s proposal in PR#17116.
system.time() now uses message() instead of cat() when terminated early, such that suppressMessages() has an effect; suggested by Ben Bolker.
citation() supports ‘inst/CITATION’ files from package source trees, with lib.loc pointing to the directory containing the package.
try() gains a new argument outFile with a default that can be modified via options(try.outFile = .), useful notably for Sweave.
The unexported low-level functions in package parallel for passing serialized R objects to and from forked children now support long vectors on 64-bit platforms. This removes some limits on higher-level functions such as mclapply() (but returning gigabyte results from forked processes via serialization should be avoided if at all possible).
Connections now print() without error even if invalid, e.g. after having been destroyed.
apropos() and find(simple.words = FALSE) no longer match object names starting with . which are known to be internal objects (such as .__S3MethodsTable__.).
Convenience function hasName() has been added; it is intended to replace the common idiom !is.null(x$name) without the usually unintended partial name matching.
strcapture() no longer fixes column names nor coerces strings to factors (suggested by Bill Dunlap).
strcapture() returns NA for non-matching values in x (suggested by Bill Dunlap).
source() gets new optional arguments, notably exprs; this is made use of in the new utility function withAutoprint().
sys.source() gets a new toplevel.env argument. This argument is useful for frameworks running package tests; contributed by Tomas Kalibera.
Sys.setFileTime() and file.copy(copy.date = TRUE) will set timestamps with fractions of seconds on platforms/filesystems which support this.
(Windows only.) file.info() now returns file timestamps including fractions of seconds; it has done so on other platforms since R 2.14.0. (NB: some filesystems do not record modification and access timestamps to sub-second resolution.)
The license check enabled by options(checkPackageLicense = TRUE) is now done when the package’s namespace is first loaded.
ppr() and supsmu() get an optional trace argument, and ppr(.., sm.method = ..spline) is no longer limited to sample size n <= 2500.
The POSIXct method for print() gets optional tz and usetz arguments, thanks to a report from Jennifer S. Lyon.
New function check_packages_in_dir_details() in package tools for analyzing package-check log files to obtain check details.
Package tools now exports function CRAN_package_db() for obtaining information about current packages in the CRAN package repository, and several functions for obtaining the check status of these packages.
The (default) Stangle driver Rtangle allows annotate to be a function and gets a new drop.evalFALSE option.
The default method for quantile(x, prob) should now be monotone in prob, even in border cases, see PR#16672.
bug.report() now tries to extract an email address from a BugReports field, and if there is none, from a Contacts field.
The format() and print() methods for object.size() results get new options standard and digits; notably, standard = "IEC" and standard = "SI" allow more standard (but less common) abbreviations than the default ones, e.g. for kilobytes. (From contributions by Henrik Bengtsson.)
If a reference class has a validity method, validObject will be called automatically from the default initialization method for reference classes.
tapply() gets new option default = NA allowing to change the previously hardcoded value.
read.dcf() now consistently interprets any ‘whitespace’ to be stripped to include newlines.
The maximum number of DLLs that can be loaded into R e.g. via dyn.load() can now be increased by setting the environment variable R_MAX_NUM_DLLS before starting R.
Assigning to an element of a vector beyond the current length now over-allocates by a small fraction. The new vector is marked internally as growable, and the true length of the new vector is stored in the truelength field. This makes building up a vector result by assigning to the next element beyond the current length more efficient, though pre-allocating is still preferred. The implementation is subject to change and not intended to be used in packages at this time.
Loading the parallel package namespace no longer sets or changes the .Random.seed, even if R_PARALLEL_PORT is unset.NB: This can break reproducibility of output, and did for a CRAN package.
Methods "wget" and "curl" for download.file() now give an R error rather than a non-zero return value when the external command has a non-zero status.
Encoding name "utf8" is mapped to "UTF-8". Many implementations of iconv accept "utf8", but not GNU libiconv (including the late 2016 version 1.15).
sessionInfo() shows the full paths to the library or executable files providing the BLAS/LAPACK implementations currently in use (not available on Windows).
The binning algorithm used by bandwidth selectors bw.ucv(), bw.bcv() and bw.SJ() switches to a version linear in the input size n for n #> nb/2. (The calculations are the same, but for larger n/nb it is worth doing the binning in advance.)
There is a new option PCRE_study which controls when grep(perl = TRUE) and friends ‘study’ the compiled pattern. Previously this was done for 11 or more input strings: it now defaults to 10 or more (but most examples need many more for the difference from studying to be noticeable).
grep(perl = TRUE) and friends can now make use of PCRE’s Just-In-Time mechanism, for PCRE #>= 8.20 on platforms where JIT is supported. It is used by default whenever the pattern is studied (see the previous item). (Based on a patch from Mikko Korpela.)This is controlled by a new option PCRE_use_JIT.Note that in general this makes little difference to the speed, and may take a little longer: its benefits are most evident on strings of thousands of characters. As a side effect it reduces the chances of C stack overflow in the PCRE library on very long strings (millions of characters, but see next item).
Warning: segfaults were seen using PCRE with JIT enabled on 64-bit Sparc builds.
There is a new option PCRE_limit_recursion for grep(perl = TRUE) and friends to set a recursion limit taking into account R‘s estimate of the remaining C stack space (or 10000 if that is not available). This reduces the chance of C stack overflow, but because it is conservative may report a non-match (with a warning) in examples that matched before. By default it is enabled if any input string has 1000 or more bytes. (PR#16757)
getGraphicsEvent() now works on X11(type = "cairo") devices. Thanks to Frederick Eaton (for reviving an earlier patch).
There is a new argument onIdle for getGraphicsEvent(), which allows an R function to be run whenever there are no pending graphics events. This is currently only supported on X11 devices. Thanks to Frederick Eaton.
The deriv() and similar functions now can compute derivatives of log1p(), sinpi() and similar one-argument functions, thanks to a contribution by Jerry Lewis.
median() gains a formal ... argument, so methods with extra arguments can be provided.
strwrap() reduces indent if it is more than half width rather than giving an error. (Suggested by Bill Dunlap.)
When the condition code in if(.) or while(.) is not of length one, an error instead of a warning may be triggered by setting an environment variable, see the help page.
Formatting and printing of bibliography entries (bibentry) is more flexible and better documented. Apart from setting options(citation.bibtex.max = 99) you can also use print(<citation#>, bibtex=TRUE) (or format(..)) to get the BibTeX entries in the case of more than one entry. This also affects citation(). Contributions to enable style = "html+bibtex" are welcome.

C-LEVEL FACILITIES

Entry points R_MakeExternalPtrFn and R_ExternalPtrFn are now declared in header ‘Rinternals.h’ to facilitate creating and retrieving an Rexternal pointer from a C function pointer without ISO C warnings about the conversion of function pointers.
There was an exception for the native Solaris C++ compiler to the dropping (in R 3.3.0) of legacy C++ headers from headers such as ‘R.h’ and ‘Rmath.h’ — this has now been removed. That compiler has strict C++98 compliance hence does not include extensions in its (non-legacy) C++ headers: some packages will need to request C++11 or replace non-C++98 calls such as lgamma: see §1.6.4 of ‘Writing R Extensions’.Because it is needed by about 70 CRAN packages, headers ‘R.h’ and ‘Rmath.h’ still declare
```
use namespace std;
```
when included on Solaris.
When included from C++, the R headers now use forms such as std::FILE directly rather than including the line
```
using std::FILE;
```
C++ code including these headers might be relying on the latter.
Headers ‘R_ext/BLAS.h’ and ‘R_ext/Lapack.h’ have many improved declarations including const for double-precision complex routines. Inter alia this avoids warnings when passing ‘string literal’ arguments from C++11 code.
Headers for Unix-only facilities ‘R_ext/GetX11Image.h’, ‘R_ext/QuartzDevice.h’ and ‘R_ext/eventloop.h’ are no longer installed on Windows.
No-longer-installed headers ‘GraphicsBase.h’, ‘RGraphics.h’, ‘Rmodules/RX11.h’ and ‘Rmodules/Rlapack.h’ which had a LGPL license no longer do so.
HAVE_UINTPTR_T is now defined where appropriate by Rconfig.h so that it can be included before Rinterface.h when CSTACK_DEFNS is defined and a C compiler (not C++) is in use. Rinterface.h now includes C header ‘stdint.h’ or C++11 header ‘cstdint’ where needed.
Package tools has a new function package_native_routine_registration_skeleton() to assist adding native-symbol registration to a package. See its help and §5.4.1 of ‘Writing R Extensions’ for how to use it. (At the time it was added it successfully automated adding registration to over 90% of CRAN packages which lacked it. Many of the failures were newly-detected bugs in the packages, e.g. 50 packages called entry points with varying numbers of arguments and 65 packages called entry points not in the package.)

INSTALLATION on a UNIX-ALIKE

readline headers (and not just the library) are required unless configuring with –with-readline=no.
configure now adds a compiler switch for C++11 code, even if the compiler supports C++11 by default. (This ensures that g++ 6.x uses C++11 mode and not its default mode of C++14 with ‘GNU extensions’.)The tests for C++11 compliance are now much more comprehensive. For gcc < 4.8, the tests from R 3.3.0 are used in order to maintain the same behaviour on Linux distributions with long-term support.
An alternative compiler for C++11 is now specified with CXX11, not CXX1X. Likewise C++11 flags are specified with CXX11FLAGS and the standard (e.g., -std=gnu++11 is specified with CXX11STD.
configure now tests for a C++14-compliant compiler by testing some basic features. This by default tries flags for the compiler specified by CXX11, but an alternative compiler, options and standard can be specified by variables CXX14, CXX14FLAGS and CXX14STD (e.g., -std=gnu++14).
There is a new macro CXXSTD to help specify the standard for C++ code, e.g. -std=c++98. This makes it easier to work with compilers which default to a later standard: for example, with CXX=g++6 CXXSTD=-std=c++98 configure will select commands for g++ 6.x which conform to C++11 and C++14 where specified but otherwise use C++98.
Support for the defunct IRIX and OSF/1 OSes and Alpha CPU has been removed.
configure checks that the compiler specified by $CXX $CXXFLAGS is able to compile C++ code.
configure checks for the required header ‘sys/select.h’ (or ‘sys/time.h’ on legacy systems) and system call select and aborts if they are not found.
If available, the POSIX 2008 system call utimensat will be used by Sys.setFileTime() and file.copy(copy.date = TRUE). This may result in slightly more accurate file times. (It is available on Linux and FreeBSD but not macOS.)
The minimum version requirement for libcurl has been reduced to 7.22.0, although at least 7.28.0 is preferred and earlier versions are little tested. (This is to support Debian 7 ‘Wheezy’ LTS and Ubuntu ‘Precise’ 12.04 LTS, although the latter is close to end-of-life.)
configure tests for a C++17-compliant compiler. The tests are experimental and subject to change in the future.

INCLUDED SOFTWARE

(Windows only) Tcl/Tk version 8.6.4 is now included in the binary builds. The ‘tcltk*.chm’ help file is no longer included; please consult the online help at http://www.tcl.tk/man/ instead.
The version of LAPACK included in the sources has been updated to 3.7.0: no new routines have been added to R.

PACKAGE INSTALLATION

There is support for compiling C++14 or C++17 code in packages on suitable platforms: see ‘Writing R Extensions’ for how to request this.
The order of flags when LinkingTo other packages has been changed so their include directories come earlier, before those specified in CPPFLAGS. This will only have an effect if non-system include directories are included with -I flags in CPPFLAGS (and so not the default -I/usr/local/include which is treated as a system include directory on most platforms).
Packages which register native routines for .C or .Fortran need to be re-installed for this version (unless installed with R-devel SVN revision r72375 or later).
Make variables with names containing CXX1X are deprecated in favour of those using CXX11, but for the time being are still made available viafile ‘etc/Makeconf’. Packages using them should be converted to the new forms and made dependent on R (#>= 3.4.0).

UTILITIES

Running R CMD check --as-cran with _R_CHECK_CRAN_INCOMING_REMOTE_ false now skips tests that require remote access. The remaining (local) tests typically run quickly compared to the remote tests.
R CMD build will now give priority to vignettes produced from files in the ‘vignettes’ directory over those in the ‘inst/doc’ directory, with a warning that the latter are being ignored.
R CMD config gains a –all option for printing names and values of all basic configure variables.It now knows about all the variables used for the C++98, C++11 and C++14 standards.
R CMD check now checks that output files in ‘inst/doc’ are newer than the source files in ‘vignettes’.
For consistency with other package subdirectories, files named ‘*.r’ in the ‘tests’ directory are now recognized as tests by R CMD check. (Wish of PR#17143.)
R CMD build and R CMD check now use the union of R_LIBS and .libPaths(). They may not be equivalent, e.g., when the latter is determined by R_PROFILE.
R CMD build now preserves dates when it copies files in preparing the tarball. (Previously on Windows it changed the dates on all files; on Unix, it changed some dates when installing vignettes.)
The new option R CMD check --no-stop-on-test-error allows running the remaining tests (under ‘tests/’) even if one gave an error.
Check customization via environment variables to detect side effects of .Call() and .External() calls which alter their arguments is described in §8 of the ‘R Internals’ manual.
R CMD check now checks any BugReports field to be non-empty and a suitable single URL.
R CMD check --as-cran now NOTEs if the package does not register its native routines or does not declare its intentions on (native) symbol search. (This will become a WARNING in due course.)

DEPRECATED AND DEFUNCT

(Windows only) Function setInternet2() is defunct.
Installation support for readline emulations based on editline (aka libedit) is deprecated.
Use of the C/C++ macro NO_C_HEADERS is defunct and silently ignored.
unix.time(), a traditional synonym for system.time(), has been deprecated.
structure(NULL, ..) is now deprecated as you cannot set attributes on NULL.
Header ‘Rconfig.h’ no longer defines SUPPORT_OPENMP; instead use _OPENMP (as documented for a long time).
(C-level Native routine registration.) The deprecated styles member of the R_CMethodDef and R_FortranMethodDef structures has been removed. Packages using these will need to be re-installed for R 3.4.0.
The deprecated support for PCRE versions older than 8.20 will be removed in R 3.4.1. (Versions 8.20–8.31 will still be accepted but remain deprecated.)

BUG FIXES

Getting or setting body() or formals() on non-functions for now signals a warning and may become an error for setting.
match(x, t), duplicated(x) and unique(x) work as documented for complex numbers with NAs or NaNs, where all those containing NA do match, whereas in the case of NaN‘s both real and imaginary parts must match, compatibly with how print() and format() work for complex numbers.
deparse(<complex#>, options = "digits17") prints more nicely now, mostly thanks to a suggestion by Richie Cotton.
Rotated symbols in plotmath expressions are now positioned correctly on x11(type = "Xlib"). (PR#16948)
as<-() avoids an infinite loop when a virtual class is interposed between a subclass and an actual superclass.
Fix level propagation in unlist() when the list contains zero-length lists or factors.
Fix S3 dispatch on S4 objects when the methods package is not attached.
Internal S4 dispatch sets .Generic in the method frame for consistency with standardGeneric(). (PR#16929)
Fix order(x, decreasing = TRUE) when x is an integer vector containing MAX_INT. Ported from a fix Matt Dowle made to data.table.
Fix caching by callNextMethod(), resolves PR#16973 and PR#16974.
grouping() puts NAs last, to be consistent with the default behavior of order().
Point mass limit cases: qpois(-2, 0) now gives NaN with a warning and qgeom(1, 1) is 0. (PR#16972)
table() no longer drops an "NaN" factor level, and better obeys exclude = <chr#>, thanks to Suharto Anggono’s patch for PR#16936. Also, in the case of exclude = NULL and NAs, these are tabulated correctly (again).Further, table(1:2, exclude = 1, useNA = "ifany") no longer erroneously reports <NA#> counts.Additionally, all cases of empty exclude are equivalent, and useNA is not overwritten when specified (as it was by exclude = NULL).
wilcox.test(x, conf.int=TRUE) no longer errors out in cases where the confidence interval is not available, such as for x = 0:2.
droplevels(f) now keeps <NA#> levels when present.
In integer arithmetic, NULL is now treated as integer(0) whereas it was previously treated as double(0).
The radix sort considers NA_real_ and NaN to be equivalent in rank (like the other sort algorithms).
When index.return=TRUE is passed to sort.int(), the radix sort treats NAs like sort.list() does (like the other sort algorithms).
When in tabulate(bin, nbin) length(bin) is larger than the maximal integer, the result is now of type double and hence no longer silently overflows to wrong values. (PR#17140)
as.character.factor() respects S4 inheritance when checking the type of its argument. (PR#17141)
The factor method for print() no longer sets the class of the factor to NULL, which would violate a basic constraint of an S4 object.
formatC(x, flag = f) allows two new flags, and signals an error for invalid flags also in the case of character formatting.
Reading from file("stdin") now also closes the connection and hence no longer leaks memory when reading from a full pipe, thanks to Gábor Csárdi, see thread starting at https://stat.ethz.ch/pipermail/r-devel/2016-November/073360.html.
Failure to create file in tempdir() for compressed pdf() graphics device no longer errors (then later segfaults). There is now a warning instead of error and compression is turned off for the device. Thanks to Alec Wysoker (PR#17191).
Asking for methods() on "|" returns only S3 methods. See https://stat.ethz.ch/pipermail/r-devel/2016-December/073476.html.
dev.capture() using Quartz Cocoa device (macOS) returned invalid components if the back-end chose to use ARGB instead of RGBA image format. (Reported by Noam Ross.)
seq("2", "5") now works too, equivalently to "2":"5" and seq.int().
seq.int(to = 1, by = 1) is now correct, other cases are integer (instead of double) when seq() is integer too, and the “non-finite” error messages are consistent between seq.default() and seq.int(), no longer mentioning NaN etc.
rep(x, times) and rep.int(x, times) now work when times is larger than the largest value representable in an integer vector. (PR#16932)
download.file(method = "libcurl") does not check for URL existence before attempting downloads; this is more robust to servers that do not support HEAD or range-based retrieval, but may create empty or incomplete files for aborted download requests.
Bandwidth selectors bw.ucv(), bw.bcv() and bw.SJ() now avoid integer overflow for large sample sizes.
str() no longer shows "list output truncated", in cases that list was not shown at all. Thanks to Neal Fultz (PR#17219)
Fix for cairo_pdf() (and svg() and cairo_ps()) when replaying a saved display list that contains a mix of grid and graphics output. (Report by Yihui Xie.)
The str() and as.hclust() methods for "dendrogram" now also work for deeply nested dendrograms thanks to non-recursive implementations by Bradley Broom.
sample() now uses two uniforms for added precision when the uniform generator is Knuth-TAOCP, Knuth-TAOCP-2002, or a user-defined generator and the population size is 2^25 or greater.
If a vignette in the ‘vignettes’ directory is listed in ‘.Rbuildignore’, R CMD build would not include it in the tarball, but would include it in the vignette database, leading to a check warning. (PR#17246)
tools::latexToUtf8() infinite looped on certain inputs. (PR#17138)
terms.formula() ignored argument names when determining whether two terms were identical. (PR#17235)
callNextMethod() was broken when called from a method that augments the formal arguments of a primitive generic.
Coercion of an S4 object to a vector during sub-assignment into a vector failed to dispatch through the as.vector() generic (often leading to a segfault).
Fix problems in command completion: Crash (PR#17222) and junk display in Windows, handling special characters in filenames on all systems.