R 4.3.0

CHANGES IN R 4.3.0

SIGNIFICANT USER-VISIBLE CHANGES:

• Calling && or || with LHS or (if evaluated) RHS of length greater
  than one is now always an error, with a report of the form

  'length = 4' in coercion to 'logical(1)'
  

  Environment variable _R_CHECK_LENGTH_1_LOGIC2_ no longer has any
  effect.

NEW FEATURES:

• The included BLAS sources have been updated to those shipped with
  LAPACK version 3.10.1. (This caused some platform-dependent
  changes to package check output.) And then to the sources from
  LAPACK version 3.11.0 (with changes only to double complex
  subroutines).

• The included LAPACK sources have been updated to include the four
  Fortran 90 routines rather than their Fortran 77 predecessors.
  This may give some different signs in SVDs or
  eigendecompositions.. (This completes the transition to LAPACK
  3.10.x begun in R 4.2.0.)

• The LAPACK sources have been updated to version 3.11.0. (No new
  subroutines have been added, so this almost entirely bug fixes:
  Those fixes do affect some computations with NaNs, including R's
  NA.)

• The parser now signals classed errors, notably in case of the
  pipe operator |>. The error object and message now give line and
  column numbers, mostly as proposed and provided by Duncan Murdoch
  in PR#18328.

• toeplitz() is now generalized for asymmetric cases, with a
  toeplitz2() variant.

• xy.coords() and xyz.coords() and consequently, e.g., plot(x,y, log = "y") now signal a classed warning about negative values
  of y (where log(.) is NA). Such a warning can be specifically
  suppressed or caught otherwise.

• Regular expression functions now check more thoroughly whether
  their inputs are valid strings (in their encoding, e.g. in
  UTF-8).

• The performance of grep(), sub(), gsub() and strsplit() has been
  improved, particularly with perl = TRUE and fixed = TRUE. Use of
  useBytes = TRUE for performance reasons should no longer be
  needed and is discouraged: it may lead to incorrect results.

• apropos() gains an argument dot_internals which is used by the
  completion (help(rcompgen)) engine to also see base internals
  such as .POSIXct().

• Support in tools::Rdiff() for comparing uncompressed PDF files is
  further reduced - see its help page.

• qqplot(x, y, ...) gains conf.level and conf.args arguments for
  computing and plotting a confidence band for the treatment
  function transforming the distribution of x into the distribution
  of y (Switzer, 1976, Biometrika). Contributed by Torsten
  Hothorn.

• Performance of package_dependencies() has been improved for cases
  when the number of dependencies is large.

• Strings newly created by gsub(), sub() and strsplit(), when any
  of the inputs is marked as "bytes", are also marked as "bytes".
  This reduces the risk of creating invalid strings and accidental
  substitution of bytes deemed invalid.

• Support for readLines(encoding = "bytes") has been added to allow
  processing special text files byte-by-byte, without creating
  invalid strings.

• iconv(from = "") now takes into account any declared encoding of
  the input elements and uses it in preference to the native
  encoding. This reduces the risk of accidental creation of
  invalid strings, particularly when different elements of the
  input have different encoding (including "bytes").

• Package repositories in getOption("repos") are now initialized
  from the repositories file when utils is loaded (if not already
  set, e.g., in .Rprofile). (From a report and patch proposal by
  Gabriel Becker in PR#18405.)

• compactPDF() gets a verbose option.

• type.convert() and hence read.table() get new option tryLogical = TRUE with back compatible default. When set to false, converts
  "F" or "T" columns to character.

• Added new unit prefixes "R" and "Q" for abbreviating
  (unrealistically large) sizes beyond 10^{27} in standard = "SI",
  thanks to Henrik Bengtsson's PR#18435.

• as.data.frame()'s default method now also works fine with atomic
  objects inheriting from classes such as "roman", "octmode" and
  "hexmode", such fulfilling the wish of PR#18421, by Benjamin
  Feakins.

• The as.data.frame.vector() utility now errors for wrong-length
  row.names. It warned for almost six years, with “Will be an
  error!”.

• sessionInfo() now also contains La_version() and reports codepage
  and timezone when relevant, in both print() and toLatex() methods
  which also get new option tzone for displaying timezone
  information when locale = FALSE.

• New function R_compiled_by() reports the C and Fortran compilers
  used to build R, if known.

• predict(<lm>, newdata = *) no longer unnecessarily creates an
  offset of all 0s.

• solve() for complex inputs now uses argument tol and by default
  checks for ‘computational singularity’ (as it long has done for
  numeric inputs).

• predict(<rank-deficient lm>, newdata=*) now obeys a new argument
  rankdeficient, with new default "warnif", warning only if there
  are non-estimable cases in newdata. Other options include
  rankdeficient = "NA", predicting NA for non-estimable newdata
  cases. This addresses PR#15072 by Russ Lenth and is based on his
  original proposal and discussions in PR#16158 also by David Firth
  and Elin Waring. Still somewhat experimental.

• Rgui console implementation now works better with the NVDA screen
  reader when the full blinking cursor is selected. The underlying
  improvements in cursor handling may help also other screen
  readers on Windows.

• The drop-field control in GraphApp can now be left with the TAB
  key and all controls can be navigated in the reverse order using
  the Shift+TAB key, improving accessibility of the Rgui
  configuration editor.

• qnorm(<very large negative>, log.p=TRUE) is now fully accurate
  (instead of to “only” minimally five digits).

• demo(error.catching) now also shows off withWarnings() and
  tryCatchWEMs().

• As an experimental feature the placeholder _ can now also be used
  in the rhs of a forward pipe |> expression as the first argument
  in an extraction call, such as _$coef. More generally, it can be
  used as the head of a chain of extractions, such as _$coef[[2]].

• Spaces in the environment variable used to choose the R session's
  temporary directory (TMPDIR, TMP and TEMP are tried in turn) are
  now fatal. (On Windows the ‘short path’ version of the path is
  tried and used if that does not contain a space.)

• all.equal.numeric() gets a new optional switch giveErr to return
  the numeric error as attribute. Relatedly,
  stopifnot(all.equal<some>(a, b, ..)) is as “smart” now, as
  stopifnot(all.equal(....)) has been already, thus allowing
  customized all.equal<Some>() wrappers.

• R on Windows is now able to work with path names longer than 260
  characters when these are enabled in the system (requires at
  least Windows 10 version 1607). Packages should be updated to
  work with long paths as well, instead of assuming PATH_MAX to be
  the maximum length. Custom front-ends and applications embedding
  R need to update their manifests if they wish to allow this
  feature. See
  https://blog.r-project.org/2023/03/07/path-length-limit-on-windows
  for more information.

• ‘Object not found’ and ‘Missing argument’ errors now give a more
  accurate error context. Patch provided by Lionel Henry in
  PR#18241.

• The @ operator is now an S3 generic. Based on contributions by
  Tomasz Kalinowski in PR#18482.

• New generic chooseOpsMethod() provides a mechanism for objects to
  resolve cases where two suitable methods are found for an Ops
  Group Generic. This supports experimenting with alternative
  object systems. Based on contributions by Tomasz Kalinowski in
  PR#18484.

• inherits(x, what) now accepts values other than a simple
  character vector for argument what. A new generic, nameOfClass(),
  is called to resolve the class name from what. This supports
  experimenting with alternative object systems. Based on
  contributions by Tomasz Kalinowski in PR#18485.

• Detection of BLAS/LAPACK in use (sessionInfo()) with FlexiBLAS
  now reports the current backend.

• The "data.frame" method for subset() now warns about extraneous
  arguments, typically catching the use of = instead of == in the
  subset expression.

• Calling a:b when numeric a or b is longer than one may now be
  made into an error by setting environment variable
  _R_CHECK_LENGTH_COLON_ to a true value, along the proposal in
  PR#18419 by Henrik Bengtsson.

• density(x, weights = *) now warns if automatic bandwidth
  selection happens without using weights; new optional warnWbw may
  suppress the warning. Prompted by Christoph Dalitz' PR#18490 and
  its discussants.

• rm(list = *) is faster and more readable thanks to Kevin Ushey's
  PR#18492.

• The plot.lm() function no longer produces a normal Q-Q plot for
  GLMs. Instead it plots a half-normal Q-Q plot of the absolute
  value of the standardized deviance residuals.

• The print() method for class "summary.glm" no longer shows
  summary statistics for the deviance residuals by default. Its
  optional argument show.residuals can be used to show them if
  required.

• The tapply() function now accepts a data frame as its X argument,
  and allows INDEX to be a formula in that case. by.data.frame()
  similarly allows INDICES to be a formula.

• The performance of df[j] <- value (including for missing j) and
  write.table(df) has been improved for data frames df with a large
  number of columns. (Thanks to Gabriel Becker's PR#18500,
  PR#18503 and discussants, prompted by a report from Toby Dylan
  Hocking on the R-devel mailing list.)

• The matrix multiply operator %*% is now an S3 generic, belonging
  to new group generic matrixOps. From Tomasz Kalinowski's
  contribution in PR#18483.

• New function array2DF() to convert arrays to data frames,
  particularly useful for the list arrays created by tapply().

DATES and TIMES:

• On platforms where (non-UTC) datetimes before 1902 (or before
  1900 as with system functions on recent macOS) are guessed by
  extrapolating time zones from 1902-2037, there is a warning at
  the first use of extrapolation in a session. (As all time zones
  post 2037 are extrapolation, we do not warn on those.)

• (Platforms using --with-internal-tzone, including Windows and by
  default macOS). How years are printed in dates or date-times can
  be controlled by environment variable R_PAD_YEARS_BY_ZERO. The
  default remains to pad to 4 digits by zeroes, but setting value
  no gives no padding (as used by default by glibc).

• strftime() tries harder to determine the offset for the "%z"
  format, and succeeds on the mainstream R platforms.

• strftime() has a limit of 2048 bytes on the string produced -
  attempting to exceed this is an error. (Previously it silently
  truncated at 255 bytes.)

• sessionInfo() records (and by default prints) the system time
  zone as part of the locale information. Also, the source
  (system/internal) of the date-time conversion and printing
  functions.

• Objects of class "POSIXlt" created in this version of R always
  have 11 components: component zone is always set, and component
  gmtoff is set for times in UTC and usually set on the (almost
  all) platforms which have C-level support, otherwise is NA.

• There are comprehensive validity checks on the structure of
  objects of class "POSIXlt" when converting (including formatting
  and printing). (This avoids mis-conversions of hand-crafted
  objects.)

• There is some support for using the native date-time routines on
  macOS: this is only viable on recent versions (e.g. 12.6 and 13)
  and does get wrong some historical changes (before 1900, during
  WWII). Use of --with-internal-tzone remains the default.

• as.POSIXct(<numeric>) and as.POSIXlt(.) (without specifying
  origin) now work. So does as.Date(<numeric>).

• as.Date.POSIXct(., tz) now treats several tz values, notably
  "GMT" as equivalent to "UTC", proposed and improved by Michael
  Chirico and Joshua Ulrich in PR#17674.

• Experimental balancePOSIXlt() utility allows using “ragged” and
  or out-of-range "POSIXlt" objects more correctly, e.g., in
  subsetting and subassignments. Such objects are now documented.

  More experimentally, a "POSIXlt" object may have an attribute
  "balanced" indicating if it is known to be filled or fully
  balanced.

• Functions axis.Date() and axis.POSIXct() are rewritten to gain
  better default tick locations and better default formats by using
  prettyDate(). Thanks to Swetlana Herbrandt.

• The mapping of Windows' names for time zones to IANA's ‘Olson’
  names has been updated. When ICU is available (it is by
  default), it is used to get a mapping for the current region set
  in Windows. This can be overridden by setting environment
  variable TZ to the desired Olson name - see OlsonNames() for
  those currently available.

GRAPHICS:

• The graphics engine version, R_GE_version, has been bumped to 16
  and so packages that provide graphics devices should be
  reinstalled.

• The grDevices and grid packages have new functions for rendering
  typeset glyphs, primarily: grDevices::glyphInfo() and
  grid::grid.glyph().

  Rendering of typeset glyphs is only supported so far on the
  Cairo-based graphics devices and on the pdf() and quartz()
  devices.

• The defined behaviour for "clear" and "source" compositing
  operators (via grid::grid.group()) has been changed (to align
  better with simple interpretation of original Porter-Duff
  definitions).

• Support for gradients, patterns, clipping paths, masks, groups,
  compositing operators, and affine transformations has been added
  to the quartz() device.

INSTALLATION on a UNIX-ALIKE:

• A system installation of generic LAPACK 3.10.0 or later will be
  preferred to the version in the R sources.

  configure option --with-lapack=no (equivalently --without-lapack)
  forces compilation of the internal LAPACK sources.

  If --with-lapack is not specified, a system liblapack is looked
  for and used if it reports version 3.10.0 or later and does not
  contain BLAS routines.

  Packages using LAPACK will need to be reinstalled if this changes
  to using an external library.

• On aarch64 Linux platforms using GCC, configure now defaults to
  -fPIC (instead of -fpic), as desired in PR#18326.

• configure now checks conversion of datetimes between POSIXlt and
  POSIXct around year 2020. Failure (which has been seen on
  platforms missing tzdata) is fatal.

• If configure option --with-valgrind-instrumentation is given
  value 1 or 2, option --with-system-valgrind-headers is now the
  default and ignored (with a warning). It is highly recommended
  that the system headers are installed alongside valgrind: they
  are part of its packaging on some Linux distributions and
  packaged separately (e.g. in the valgrind-devel RPM) on others.
  configure will give a warning if they are not found.

  The system headers will be required in a future release of R to
  build with valgrind instrumentation.

• libcurl 8.x is now accepted by configure: despite a change in
  major version number it changes neither API nor ABI.

INSTALLATION on WINDOWS:

• The makefiles and installer scripts for Windows have been
  tailored to Rtools43, an update of the Rtools42 toolchain. It is
  based on gcc 12 and newer versions of MinGW-W64, binutils and
  libraries. At this time R-devel can still be built using
  Rtools42 without changes, but when R-devel is installed via the
  installer, it will by default look for Rtools43.

• Old make targets rsync-extsoft and 32-bit ones that are no longer
  needed have been removed.

• Default builds (including for packages) no longer select C99.
  Thus the C standard used is the default for the compiler, which
  for the toolchain in Rtools43 is C17. (This is consistent with
  Unix builds.)

PACKAGE INSTALLATION:

• The default C++ standard has been changed to C++17 where
  available (which it is on all currently checked platforms): if
  not C++14 or C++11 is used if available otherwise C++ is not
  supported.

• USE_FC_LEN_T is the default: this uses the correct
  (compiler-dependent) prototypes for Fortran BLAS/LAPACK routines
  called from C/C++, and requires adjustment of many such calls -
  see ‘Writing R Extensions’ §6.6.1.

• There is initial support for C++23 as several compilers are now
  supporting -std=c++23 or -std=c++2b or similar. As for C++20,
  there no additional configure checks for C++23 features beyond a
  check that the compiler reports a __cplusplus value greater than
  that in the C++20 standard. C++ feature tests should be used.

• There is support for a package to indicate the version of the C
  standard which should be used to compile it, and for the
  installing user to specify this. In most cases R defaults to the
  C compiler's default standard which is C17 (a `bug-fix' of C11) -
  earlier versions of R or compilers may have defaulted to C99.

  Current options are:

  • USE_C17 Use a standard that is at most C17. The intention is to
    allow legacy packages to still be installed when later C
    standards become the default, including packages using new
    keywords as identifiers or with K&R-style function
    declarations. This will use C17 if available, falling back
    to C11.

  • USE_C90 Use the C90 (aka C89) standard. (As that standard did
    not require compilers to identify that version, all we can
    verify is that the compiler does not claim to be using a
    later standard. It may accept C99 features - for example
    clang accepts // to make comments.)

  • USE_C99 Use the C99 standard. This should be rarely needed - it
    avoids the few new features of C11/C17 which can be useful if
    a package assumes them if C17 is specified and they are not
    implemented.

  • USE_C23 Use C23 (or in future, later). Compiler/library support
    for C23 is still being implemented, but LLVM clang from
    15.0.0 and the upcoming GCC 13 have quite extensive support.

  These can be specified as part of the SystemRequirements field in
  the package's DESCRIPTION file or via options --use-C17 and so
  on of R CMD INSTALL and R CMD SHLIB.

  For further details see “Writing R Extensions” §1.2.5.

• (Windows) A src/Makefile.ucrt or src/Makefile.win file is now
  included after R_HOME/etcR_ARCH/Makeconf and so no longer needs
  to include that file itself. Installation of a package with such
  a file now uses a site Makevars file in the same way as a package
  with a src/Makevars.win file would.

• configure is now passed crucial variables such as CC and CFLAGS
  in its environment, as many packages were not setting them (as
  documented in ‘Writing R Extensions’ §1.2).

  This has most effect where configure is used to compile parts of
  the package - most often by cmake or libtool which obfuscate the
  actual compile commands used.

  Also used for configure.win and configure.ucrt on Windows.

FORTRAN FLAGS:

• The flag -fno-optimize-sibling-calls is no longer forced for
  gfortran 7 and later. It should no longer be needed now using
  ‘hidden’ character-length arguments when calling BLAS/LAPACK
  routines from C/C++ is the default even for packages. (Unless
  perhaps packages call Fortran code from C/C++ without using R's
  headers and without allowing for these arguments.)

C-LEVEL FACILITIES:

• The deprecated S-compatibility macros DOUBLE_* in
  R_ext/Constants.h (included by R.h) have been removed.

• The deprecated legacy typedefs of Sint and Sfloat in header R.h
  are no longer defined, and that header no longer includes header
  limits.h from C nor climits from C++.

• New macro CAD5R() is provided in Rinternals.h and used in a few
  places in the R sources.

• ALTREP now supports VECSXP vectors. Contributed by Gabor Csardi
  in PR#17620.

• The Rcomplex definition (in header R_ext/Complex.h) has been
  extended to prevent possible mis-compilation when interfacing
  with Fortran (PR#18430). The new definition causes compiler
  warnings with static initializers such as {1, 2}, which can be
  changed to {.r=1, .i=2}.

  Using the new definition from C++ depends on compiler extensions
  supporting C features that have not been incorporated into the
  C++ standards but are available in g++ and clang++: this may
  result in C++ compiler warnings but these have been worked around
  for recent versions of common compilers (GCC, Apple/LLVM clang,
  Intel).

  It is intended to change the inclusion of header R_ext/Complex.h
  by other R headers, so C/C++ code files which make use of
  Rcomplex should include that header explicitly.

UTILITIES:

• R CMD check does more checking of package .Rd files, warning
  about invalid email addresses and (some) invalid URIs and noting
  empty \item labels in description lists.

• R CMD check now also reports problems when reading package news
  in md (file NEWS.md) and (optionally) plain text (file NEWS)
  formats.

• _R_CHECK_TIMINGS_ defaults to a value from the environment even
  for R CMD check --as-cran; this allows for exceptionally fast or
  slow platforms.

  It now applies to checking PDF and HTML versions of the manuals,
  and ‘checking CRAN incoming feasibility’.

• R CMD check can optionally (but included in --as-cran) check
  whether HTML math rendering via KaTeX works for the package .Rd
  files.

• Non-interactive debugger invocations can be trapped by setting
  the environment variable _R_CHECK_BROWSER_NONINTERACTIVE_ to a
  true value. This is enabled by R CMD check --as-cran to detect
  the use of leftover browser() statements in the package.

• The use of sprintf and vsprintf from C/C++ has been deprecated in
  macOS 13 and is a known security risk. R CMD check now reports
  (on all platforms) if their use is found in compiled code:
  replace by snprintf or vsnprintf respectively. [NB: whether
  such calls get compiled into the package is platform-dependent.]

• Where recorded at installation, R CMD check reports the C and
  Fortran compilers used to build R.

  It reports the OS in use (if known, as given by osVersion) as
  well as that R was built for.

  It notes if a C++ standard was specified which is older than the
  current default: many packages have used C++11 to mean ‘not
  C++98’ - as C++11 is the minimum supported since R 4.0.0, that
  specification can probably be removed.

• R CMD INSTALL reports the compilers (and on macOS, the SDK) used,
  and this is copied to the output of R CMD check.

  Where a C++ standard is specified, it is reported.

• R CMD check's ‘checking compilation flags in Makevars’ has been
  relaxed to accept the use of flags such as -std=f2008 in
  PKG_FFLAGS.

• tools::buildVignettes() has a new argument skip, which is used by
  R CMD check to skip (and note) vignettes with unavailable
  \VignetteDepends (PR#18318).

• New generic .AtNames() added to enable class-specific completions
  after @. The formerly internal function findMatches() is now
  exported, mainly for use in methods for .DollarNames() and
  .AtNames().

DEPRECATED AND DEFUNCT:

• default.stringsAsFactors() is defunct.

• Calling as.data.frame.<class>() directly (for 12 atomic classes)
  is going to be formally deprecated, currently activated by
  setting the environment variable
  _R_CHECK_AS_DATA_FRAME_EXPLICIT_METHOD_ to non-empty, which also
  happens in R CMD check --as-cran.

BUG FIXES:

• Hashed environments with sizes less than 5 can now grow.
  (Reported to R-devel by Duncan Garmonsway.)

• as.character(<Rd>, deparse = TRUE) failed to re-escape curly
  braces in LaTeX-like text. (Reported by Hadley Wickham in
  PR#18324.)

• library() now passes its lib.loc argument when requiring Depends
  packages; reported (with fix) in PR#18331 by Mikael Jagan.

• R CMD Stangle: improved message about ‘Output’ files.

• head(x, n) and tail(x, n) now signal an error if n is not
  numeric, instead of incidentally “working” sometimes returning
  all of x. Reported and discussed by Colin Fay, in PR#18357.

• The "lm" method for summary() now gives the correct F-statistic
  when the model contains an offset. Reported in PR#18008.

• C() and `contrasts<-`() now preserve factor level names when
  given a function object (as opposed a function name which did
  preserve names). Reported in PR#17616.

• c(a = 1, 2)[[]] no longer matches 2 but rather signals a
  classed error. Reported and analysed by Davis Vaughan in
  PR#18367, a duplicate of PR#18004, by Jan Meis et al. For
  consistency, NULL[[]] is also erroneous now. x[[]] <- v gives an
  error of the same class "MissingSubscriptError".

• The relist() function of utils now supports NULL elements in the
  skeleton (PR#15854).

• ordered(levels = *) (missing x) now works analogously to factor(, ordered=TRUE); reported (with fix) by Achim Zeileis in PR#18389.

• User-defined Rd macro definitions can now span multiple lines,
  thanks to a patch from Duncan Murdoch. Previously, the Rd parser
  silently ignored everything after the first line.

• Plain-text help (tools::Rd2txt()) now preserves an initial blank
  line for text following description list items.

• tools::Rd2HTML() and tools::Rd2latex() no longer split \arguments
  and \value lists at Rd comments.

• tools::Rd2latex() now correctly handles optional text outside
  \items of argument lists as well as bracketed text at the
  beginning of sections, e.g., \value{[NULL]}.

• as.character(<POSIXt>) now behaves more in line with the methods
  for atomic vectors such as numbers, and is no longer influenced
  by options(). Ditto for as.character(<Date>). The
  as.character() method gets arguments digits and OutDec with
  defaults not depending on options(). Use of as.character(*, format = .) now warns.

• Similarly, the as.character.hexmode() and *.octmode() methods
  also behave as good citizen methods and back compatibility option
  keepStr = TRUE.

• The as.POSIXlt(<POSIXlt>) and as.POSIXct(<POSIXct>) default
  methods now do obey their tz argument, also in this case.

• as.POSIXlt(<Date>) now does apply a tz (time zone) argument, as
  does as.POSIXct(); partly suggested by Roland Fuß on the R-devel
  mailing list.

• as.Date.POSIXlt(x) now also works when the list components are of
  unequal length, aka “partially filled” or “ragged”.

• expand.model.frame() looked up variables in the wrong environment
  when applied to models fitted without data. Reported in
  PR#18414.

• time() now (also) uses the ts.eps = getOption("ts.eps") argument
  and thus by default rounds values very close to the start (or
  end) of a year. Based on a proposal by Andreï V. Kostyrka on
  R-help.

• Printing of a factanal() result with just one factor and sort = TRUE now works regularly, fixing PR#17863 by Timothy Bates,
  thanks to the ‘R Contributors’ working group.

• Printing 0-length objects of class "factor", "roman", "hexmode",
  "octmode", "person", "bibentry", or "citation" now prints
  something better, one of which fixes PR#18422, reported by
  Benjamin Feakins.

• Sys.timezone() queries timedatectl only if systemd is loaded;
  addressing a report by Jan Gorecki in PR#17421.

• The formula method of cor.test() had scoping problems when
  environment(formula) was not the calling environment; reported
  with a patch proposal by Mao Kobayashi in PR#18439.

• attach() of an environment with active bindings now preserves the
  active bindings. Reported by Kevin Ushey in PR#18425.

• BLAS detection now works also with system-provided libraries not
  available as regular files. This fixes detection of the
  Accelerate framework on macOS since Big Sur. Reported by David
  Novgorodsky.

• download.file() gives a helpful error message in case of an
  invalid download.file.method option, thanks to Colin Fay's report
  in PR#18455.

• Sporadic crashes of Rterm when using completion have been fixed.

• Rprof() is now more reliable. A livelock in thread
  initialization with too short sampling interval has been fixed on
  macOS. A deadlock in using the C runtime has been fixed on
  Windows. A potential deadlock has been prevented on Unix.

• Cursor placement in Rgui now works even after a fixed-width font
  is selected.

• Mandatory options (options()) are now set on startup so that
  saving and restoring them always works (PR#18372).

• Package installation, R CMD INSTALL or install.packages(*), now
  parses each of the <pkg>/R/*.R files individually instead of
  first concatenating and then parse()ing the large resulting file.
  This allows parser or syntax errors to be diagnosed with correct
  file names and line numbers, thanks to Simon Dedman's report and
  Bill Dunlap's patch in PR#17859.

  This does require syntactically self contained R source files
  now, fixing another inadvertent bug.

• predict.lm(<model with offset>) now finds the offset in the
  correct environment, thanks to André Gillibert's report and patch
  in PR#18456.

• getInitial(<formula>) now finds the selfStart model in the
  correct environment. (Reported by Ivan Krylov in PR#18368.)

• Fix for possible segfault when using recently-added graphics
  features, such as gradients, clipping paths, masks, and groups
  with pdf(file=NULL).

• class(m) <- class(m) no longer changes a matrix m by adding a
  class attribute.

• packageDate(pkg) now only warns once if there is no pkg.

• When ts() creates a multivariate time series, "mts", it also
  inherits from "array" now, and is.mts() is documented and
  stricter.

• Rd2txt() now preserves line breaks of \verb Rd content and from
  duplicated \cr. The former also fixes the rendering of verbatim
  output from Rd \Sexpr in plain-text help.

• uniroot(f, interval) should no longer wrongly converge outside
  the interval in some cases where abs(f(x)) == Inf for an x at the
  interval boundary, thanks to posts by Ben Bolker and Serguei
  Sokol on R-devel.

• Vectorized alpha handling in palette functions such as in gray(),
  rainbow(), or hcl.colors() works correctly now, thanks to Achim
  Zeileis' report and patch in PR#18476.

• Formatting and print()ing of bibentry objects has dropped the
  deprecated citation.bibtex.max argument, such that the bibtex
  argument's default for print.bibentry() depends directly on the
  citation.bibtex.max option, whereas in format.bibentry() the
  option no longer applies.

• Attempting to use a character string naming a foreign function
  entry point in a foreign function call in a package will now
  signal an error if the packages has called R_forceSymbols to
  specify that symbols must be used.

• An error in table() could permanently set options(warn=2)
  promoting all subsequent warnings to errors.

• The sigma() function gave misleading results for binary GLMs. A
  new method for objects of class "glm" returns the square root of
  the estimate of the dispersion parameter using the same
  calculation as summary.glm().

• bs() and ns() in the (typical) case of automatic knot
  construction, when some of the supposedly inner knots coincide
  with boundary knots, now moves them inside (with a warning),
  building on PR#18442 by Ben Bolker.

• R CMD on Windows now skips the site profile with --no-site-file
  and --vanilla even when R_PROFILE is set (PR#18512, from Kevin
  Ushey).