new CRAN issues w. absl

[edzer]

We got this message some weeks ago:

Details

A recent CRAN submission (FORTLS) failed in my gcc-san checks because
pacakge s2 failed to install.  Specifically, for my enhanced gcc build
the installation error is

g++-15 -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=g
nu++17 -I"/home/hornik/tmp/R-d-gcc-xtra/include" -DNDEBUG -I../src -DSTRICT_R_HE
ADERS -I'/home/hornik/tmp/CRAN_special_gcc-san/Library/Rcpp/include' -I'/home/ho
rnik/tmp/CRAN_special_gcc-san/Library/wk/include' -I/usr/local/include -DUSE_TYP
E_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2 -Wall -pedantic -m
tune=native  -Wno-ignored-attributes -Wno-deprecated-declarations -Wno-stringop-
truncation   -c s2-cell-union.cpp -o s2-cell-union.o

...

/usr/include/absl/container/internal/raw_hash_set.h:4052:56:   in ‘constexpr’ expansion of ‘absl::debian7::container_internal::hash_policy_traits<absl::debian7::container_internal::FlatHashSetPolicy[<s2shapeutil::ShapeEdgeId>](s2shapeutil::ShapeEdgeId), void>::get_hash_slot_fn<absl::debian7::hash_internal::Hash[<s2shapeutil::ShapeEdgeId>](s2shapeutil::ShapeEdgeId) >()’
/usr/include/absl/container/internal/hash_policy_traits.h:163:54: error: ‘(absl::debian7::container_internal::TypeErasedApplyToSlotFn<absl::debian7::hash_internal::Hash[<s2shapeutil::ShapeEdgeId>](s2shapeutil::ShapeEdgeId), s2shapeutil::ShapeEdgeId> == 0)’ is not a constant expression
  163 |     return policy::template get_hash_slot_fn<Hash>() == nullptr
      |            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~

For my enhanced clang build I can installed but at runtime get

> # binary operations: difference, symmetric difference, intersection and union
> s2_difference(
+   "POLYGON ((0 0, 10 0, 10 10, 0 10, 0 0))",
+   "POLYGON ((5 5, 15 5, 15 15, 5 15, 5 5))",
+   # 32 bit platforms may need to set snap rounding
+   s2_options(snap = s2_snap_level(30))
+ )
/usr/include/absl/container/internal/raw_hash_set.h:991:46: runtime error: load 
of null pointer of type 'GenerationType' (aka 'unsigned char')
    #0 0x7bf2b780e5d1 in absl::debian7::container_internal::CommonFieldsGenerati
oninfoenabled::generation() const /usr/include/absl/container/internal/raw_hash_
set.h:991:46
    #1 0x7bf2b780e5d1 in bool absl::debian7::container_internal::HashSetResizeHe
lper::InitializeSlots<std::allocator<char>, 4ul, true, true, 4ul>(absl::debian7:
:container_internal::CommonFields&, std::allocator<char>, absl::debian7::contain
er_internal::ctrl_t, unsigned long, unsigned long) /usr/include/absl/container/i
nternal/raw_hash_set.h:2070:45
    #2 0x7bf2b780dc18 in absl::debian7::container_internal::raw_hash_set<absl::d
ebian7::container_internal::FlatHashSetPolicy<int>, absl::debian7::hash_internal
::Hash<int>, std::equal_to<int>, std::allocator<int>>::resize_impl(absl::debian7
::container_internal::CommonFields&, unsigned long, absl::debian7::container_int
ernal::HashtablezInfoHandle) /usr/include/absl/container/internal/raw_hash_set.h
:3662:23
    #3 0x7bf2b77e0845 in absl::debian7::flat_hash_set<int, absl::debian7::hash_i
nternal::Hash<int>, std::equal_to<int>, std::allocator<int>>::flat_hash_set(unsi
gned long, absl::debian7::hash_internal::Hash<int> const&, std::equal_to<int> co
nst&, std::allocator<int> const&) /usr/include/absl/container/flat_hash_set.h:172:15
    #4 0x7bf2b77e0845 in s2builder::AddExtraSites(MutableS2ShapeIndex const&) /h
ome/hornik/tmp/CRAN/s2.Rcheck/00_pkg_src/s2/src/s2/s2builder.cc:901:30

etc.

I also tried the effect of getting a newer libabls-dev from
experimental: this now has both gcc-san and clang-san fail to install.

It looks like the absl version (system, 20240722) is too new for the s2geometry version we bundle. A solution would be to use the vendored (bundled) absl, but by default we pick up a system absl if it is new enough (>= 20230802); absl however does not aim for backward compatability and s2geometry wants particular versions of it, not >= some version. I've tried to replace s2geometry with 0.12.0 bundled with absl 20240116 but get a missing symbol error when linking s2.so (some absl module missing); I had to uncommend (remove) this line

s2/src/init.cpp

Line 23 in ad09e15

absl::SetFlag(&FLAGS_s2debug, false);

to get that running; I also tried s2geometry 0.13.1` with the absl it requires (20250814) but that led to other problems, e.g. here

s2/src/s2geography/constructor.h

Line 246 in ad09e15

err << error_.text();

where changing text() into message() didn't resolve the problem.

Two issues where I'm hanging now (WIP!):

• how does "our" absl_s2.pc lead to the long list of absl modules to link? (this needs to be fixed, somehow, to include all needed modules when upgrading s2geometry and bundled absl)
• should we go back from using the system absl by default, given that s2geometry essentially pins a version, and always use the bundled/vendored version?

@paleolimbot

[paleolimbot]

Sorry for being slow on all of this...linking Abseil is hard to do and Google and CRAN have completely different ideologies around how external libraries should work (Google expects the ability to pin all the versions and have user-provided builds, like pip, and CRAN expects the software to compile everywhere using mostly external libraries).

how does "our" absl_s2.pc lead to the long list of absl modules to link?

I settled on using pkg-config to resolve the iterdependencies between absl modules...in theory that only has to list the ones directly used by S2 and pkg-config (in theory) sorts out the recursive dependencies and linking order. I think this is slightly different on Windows and I might have solved that by just hard-coding the library list for the version of abseil provided by RTools.

should we go back from using the system absl by default, given that s2geometry essentially pins a version, and always use the bundled/vendored version?

The CMake build I think in practice has been working well for people (all Ubuntu 24.04 LTS users would have been using it) but it fails some CRAN checks around symbols (I believe the approach was referred to by CRAN maintainers as "clearly wrong" on some of our first attempts). This is another mismatch in Google vs. CRAN expectations.

In the short term I think we can probably find a way to make our vendored S2 work with the current version of Abseil in debian-testing. If you have a Dockerfile reproducer I can throw Claude at this, which I didn't have the last time I spent days trying to get Abseil to link everywhere.

In the longer term we have to do something that doesn't require (as many) custom changes to s2/Abseil specific to the R package because whenever something goes wrong it's difficult for me to find time to fix it (particularly for CRAN, which doesn't run a reproducible build system).

[paleolimbot]

Ah, sorry, I see from the email thread

FROM ghcr.io/r-devel/rcheckserver/debian

ENV R_VERSION 4.5.2

RUN curl -O https://cran.rstudio.com/src/base/R-4/R-${R_VERSION}.tar.gz
RUN tar -xzvf R-${R_VERSION}.tar.gz
RUN (cd R-${R_VERSION}; \
  ./configure \
  --prefix=/opt/R/${R_VERSION} \
  --enable-R-shlib \
  --enable-memory-profiling \
  --with-blas \
  --with-lapack; \
  make; \
  make install \
  )

I will give that a try.

[edzer]

To reproduce: I saved the Rconf script from Kurt (email Dec 29; below linked to as Rconf.txt, please rename) in /tmp, and made it executable by chmod +x /tmp/Rconf. I built the docker image with docker build . -t rdevel in the directory where the dockerfile resides as Dockerfile. Then I started it with docker run -v /tmp:/tmp -ti rdevel /bin/bash, which mounts local /tmp onto the image's /tmp. In the image, on the bash prompt:

Rconf.txt

cd R-4.5.2
R_SRC_DIR=. /tmp/Rconf -x
make
bin/R -e 'install.packages("s2", repos = "https://cloud.r-project.org/")'

gives, after installing Rcpp and wk:

Details

....
** R
** data
*** moving datasets to lazyload DB
** inst
** byte-compile and prepare package for lazy loading
** help
*** installing help indices
** building package indices
** testing if installed package can be loaded from temporary location
** checking absolute paths in shared objects and dynamic libraries
** testing if installed package can be loaded from final location
** testing if installed package keeps a record of temporary installation path
* DONE (wk)
* installing *source* package ‘s2’ ...
** this is package ‘s2’ version ‘1.1.9’
** package ‘s2’ successfully unpacked and MD5 sums checked
** using staged installation
Found OpenSSL pkg-config cflags and libs!
Testing compiler OpenSSL configuration using PKG_CFLAGS=
** Using PKG_CONFIG_PATH=:/tmp/RtmpZGwjxV/R.INSTALL3318225d8931/s2/tools/pkgconfig
** Using abseil-cpp from pkg-config
Testing Abseil configuration using PKG_CFLAGS= 
Using PKG_LIBS=-labsl_flags_internal -labsl_flags_marshalling -labsl_log_severity -labsl_flags_reflection -labsl_flags_private_handle_accessor -labsl_flags_commandlineflag -labsl_flags_commandlineflag_internal -labsl_flags_config -labsl_flags_program_name -labsl_raw_logging_internal -labsl_raw_logging_internal -labsl_raw_hash_set -labsl_hashtablez_sampler -labsl_throw_delegate -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_internal_check_op -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_internal_conditions -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_internal_message -labsl_examine_stack -labsl_log_internal_format -labsl_log_severity -labsl_log_severity -labsl_log_internal_proto -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_internal_nullguard -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_severity -labsl_log_internal_log_sink_set -labsl_log_internal_globals -labsl_log_severity -labsl_log_globals -labsl_hash -labsl_city -labsl_bad_variant_access -labsl_low_level_hash -labsl_log_severity -labsl_vlog_config_internal -labsl_log_internal_fnmatch -labsl_log_severity -labsl_log_sink -labsl_log_entry -labsl_log_severity -labsl_status -labsl_cord -labsl_cordz_info -labsl_cord_internal -labsl_cordz_functions -labsl_exponential_biased -labsl_cordz_handle -labsl_synchronization -labsl_graphcycles_internal -labsl_kernel_timeout_internal -labsl_time -labsl_civil_time -labsl_time_zone -labsl_crc_cord_state -labsl_crc32c -labsl_crc_internal -labsl_crc_cpu_detect -labsl_bad_optional_access -labsl_stacktrace -labsl_strerror -labsl_str_format_internal -labsl_symbolize -labsl_debugging_internal -labsl_demangle_internal -labsl_demangle_rust -labsl_decode_rust_punycode -labsl_utf8_for_code_point -labsl_malloc_internal -labsl_strings -labsl_strings_internal -labsl_string_view -labsl_base -Wl,--push-state,--as-needed -latomic -Wl,--pop-state -lrt -labsl_spinlock_wait -labsl_int128 -labsl_throw_delegate -labsl_raw_logging_internal -labsl_log_severity  -lssl -lcrypto 
Using PKG_CFLAGS= 
** libs
using C compiler: ‘gcc (Debian 15.2.0-12) 15.2.0’
using C++ compiler: ‘g++ (Debian 15.2.0-12) 15.2.0’
using C++17
g++ -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=gnu++17 -I"/R-4.5.2/include" -DNDEBUG -I../src -DSTRICT_R_HEADERS -I'/R-4.5.2/library/Rcpp/include' -I'/R-4.5.2/library/wk/include' -I/usr/local/include -DUSE_TYPE_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2   -c cpp-compat.cpp -o cpp-compat.o
g++ -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=gnu++17 -I"/R-4.5.2/include" -DNDEBUG -I../src -DSTRICT_R_HEADERS -I'/R-4.5.2/library/Rcpp/include' -I'/R-4.5.2/library/wk/include' -I/usr/local/include -DUSE_TYPE_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2   -c s2-accessors.cpp -o s2-accessors.o
g++ -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=gnu++17 -I"/R-4.5.2/include" -DNDEBUG -I../src -DSTRICT_R_HEADERS -I'/R-4.5.2/library/Rcpp/include' -I'/R-4.5.2/library/wk/include' -I/usr/local/include -DUSE_TYPE_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2   -c s2-bounds.cpp -o s2-bounds.o
g++ -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=gnu++17 -I"/R-4.5.2/include" -DNDEBUG -I../src -DSTRICT_R_HEADERS -I'/R-4.5.2/library/Rcpp/include' -I'/R-4.5.2/library/wk/include' -I/usr/local/include -DUSE_TYPE_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2   -c s2-cell.cpp -o s2-cell.o
g++ -fsanitize=address,undefined,bounds-strict -fno-omit-frame-pointer -std=gnu++17 -I"/R-4.5.2/include" -DNDEBUG -I../src -DSTRICT_R_HEADERS -I'/R-4.5.2/library/Rcpp/include' -I'/R-4.5.2/library/wk/include' -I/usr/local/include -DUSE_TYPE_CHECKING_STRICT -D_FORTIFY_SOURCE=3  -pthread -fpic  -g -O2   -c s2-cell-union.cpp -o s2-cell-union.o
In file included from /usr/include/absl/container/internal/raw_hash_set.h:213,
                 from /usr/include/absl/container/flat_hash_set.h:45,
                 from ../src/s2/s2closest_edge_query_base.h:33,
                 from ../src/s2/s2closest_edge_query.h:34,
                 from s2/s2shape_index_buffered_region.h:29,
                 from s2-cell-union.cpp:7:
/usr/include/absl/container/internal/hash_policy_traits.h: In instantiation of ‘static const absl::debian7::container_internal::PolicyFunctions& absl::debian7::container_internal::raw_hash_set<Policy, Hash, Eq, Alloc>::GetPolicyFunctions() [with Policy = absl::debian7::container_internal::FlatHashSetPolicy<s2shapeutil::ShapeEdgeId>; Hash = absl::debian7::hash_internal::Hash<s2shapeutil::ShapeEdgeId>; Eq = std::equal_to<s2shapeutil::ShapeEdgeId>; Alloc = std::allocator<s2shapeutil::ShapeEdgeId>]’:
/usr/include/absl/container/internal/raw_hash_set.h:2919:53:   required from ‘void absl::debian7::container_internal::raw_hash_set<Policy, Hash, Eq, Alloc>::clear() [with Policy = absl::debian7::container_internal::FlatHashSetPolicy<s2shapeutil::ShapeEdgeId>; Hash = absl::debian7::hash_internal::Hash<s2shapeutil::ShapeEdgeId>; Eq = std::equal_to<s2shapeutil::ShapeEdgeId>; Alloc = std::allocator<s2shapeutil::ShapeEdgeId>]’
 2919 |       ClearBackingArray(common(), GetPolicyFunctions(), /*reuse=*/cap < 128,
      |                                   ~~~~~~~~~~~~~~~~~~^~
../src/s2/s2closest_edge_query_base.h:560:22:   required from ‘void S2ClosestEdgeQueryBase<Distance>::FindClosestEdgesInternal(Target*, const Options&) [with Distance = S2MinDistance; Target = S2DistanceTarget<S2MinDistance>]’
  560 |   tested_edges_.clear();
      |   ~~~~~~~~~~~~~~~~~~~^~
../src/s2/s2closest_edge_query_base.h:537:3:   required from ‘void S2ClosestEdgeQueryBase<Distance>::FindClosestEdges(Target*, const Options&, std::vector<Result>*) [with Distance = S2MinDistance; Target = S2DistanceTarget<S2MinDistance>]’
  537 |   FindClosestEdgesInternal(target, options);
      |   ^~~~~~~~~~~~~~~~~~~~~~~~
../src/s2/s2closest_edge_query.h:409:25:   required from here
  409 |   base_.FindClosestEdges(target, options_, results);
      |   ~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/include/absl/container/internal/raw_hash_set.h:4052:56:   in ‘constexpr’ expansion of ‘absl::debian7::container_internal::hash_policy_traits<absl::debian7::container_internal::FlatHashSetPolicy<s2shapeutil::ShapeEdgeId>, void>::get_hash_slot_fn<absl::debian7::hash_internal::Hash<s2shapeutil::ShapeEdgeId> >()’
/usr/include/absl/container/internal/hash_policy_traits.h:163:54: error: ‘(absl::debian7::container_internal::TypeErasedApplyToSlotFn<absl::debian7::hash_internal::Hash<s2shapeutil::ShapeEdgeId>, s2shapeutil::ShapeEdgeId> == 0)’ is not a constant expression
  163 |     return Policy::template get_hash_slot_fn<Hash>() == nullptr
      |            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
make: *** [/R-4.5.2/etc/Makeconf:211: s2-cell-union.o] Error 1
ERROR: compilation failed for package ‘s2’
* removing ‘/R-4.5.2/library/s2’

The downloaded source packages are in
	‘/tmp/RtmpwFLVLo/downloaded_packages’
Updating HTML index of packages in '.Library'
Making 'packages.html' ... done
Warning message:
In install.packages("s2", repos = "https://cloud.r-project.org/") :
  installation of package ‘s2’ had non-zero exit status

[paleolimbot]

Thank you! I successfully got this in a reproducible dockery thing. I'll plug away at this and see what I come up with.

[paleolimbot]

Just dumping some (unverified) LLM output:

The issue is that Debian's system abseil package has a bug with GCC 14 regarding constexpr evaluation in flat_hash_set. This affects multiple files in s2 and would require extensive changes to replace all uses.

This is not a formally reported bug I have a reference to - I was inferring from the error pattern. What's happening:

The issue: GCC 14 has stricter constexpr evaluation. Abseil's hash_policy_traits.h line 163 compares a function pointer to nullptr in a constexpr context. GCC 14 rejects this because comparing function pointers in constexpr is only fully supported in C++20+.

Debian's abseil: The absl::debian7:: namespace indicates Debian patches abseil to use an inline namespace for ABI compatibility. The version in the image appears to be older (pre-fix).

Upstream fix: This was fixed in abseil-cpp around version 20240116 or later. The fix changed how get_hash_slot_fn is evaluated.

• Wait for Debian to update their abseil package
• Use S2_FORCE_BUNDLED_ABSEIL=1 (which you said you can't)
• Request the CRAN check image maintainers update abseil
• Patch the Rconf script to use -std=gnu++20 for C++20 mode (but this may have other implications)

[aitap]

You've stumbled into a real GCC bug, not a C++20 problem: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71962

There doesn't seem to be a good solution to the problem, but maybe you could check for the presence of the compiler bug in your configure script and set S2_FORCE_BUNDLED_ABSEIL=1 based on that?

[paleolimbot]

Thank you! That makes sense as the build causing the issue is the sanitizer and that seems to be something that triggers the bug.

[cement-head]

@aitap Could you explain in a bit more detail how to alter my configure script (as per your suggests)?

EDIT: making sure the system has pkgconfig installed from with in the R 4.6.0 environment and both these packages (on Ubuntu/Debian installed: libssl-dev libabsl-dev), the development version of s2 from GitHub will compile and install on Ubuntu Jammy 22.04 LTS.

# install.packages("remotes")
remotes::install_github("r-spatial/s2")

[aitap]

@cement-head I had meant taking a reproducer example from the Bugzilla comments and checking whether it compiles without errors:

cat >test.cpp <<'EOF'
struct P {
  constexpr P(const int* p) : p(p) { }
  const int* p;
  explicit constexpr operator bool() const { return (bool)p; }
};

int main() {
  static constexpr int x{1};
  constexpr P p{&x};
  static_assert((bool)p, "");
}
EOF
if ! ${CXX17} ${CPPFLAGS} ${PKG_CFLAGS} ${CXX17FLAGS} ${CXX17STD} -c -o test.o test.cpp; then
 echo "This compiler suffers from GCC bug 71962, using bundled Abseil patched with a workaround"
 S2_FORCE_BUNDLED_ABSEIL=TRUE
fi
rm -f test.cpp test.o

but hopefully that's not actually necessary.

[paleolimbot]

I've spent a bit of time with this sanitizer bug in the context of adding sanitizer CI to the underlying s2geography, and unfortunately there is not much in the way of a workaround...Abseil doesn't care that there's a GCC bug in a sanitizer it doesn't use, and the offending code has been present for years. Our vendored-and-slighlty-patched Abseil plus vendored-and-slightly-patched s2geography seems to work but it also means we can't upgrade either which is unfortunate.

[barracuda156]

Cannot configure now with external abseil either. Looks like absl_s2 does not even exist anymore.

[paleolimbot]

absl_s2 was only ever a made up package name for our internal pkg-config file, and I don't think it was removed. Can you share more about your error?

[barracuda156]

@paleolimbot Turned out, it was a local issue with DYLD_LIBRARY_PATH somehow messing up with pkg-config behavior. It erred out during configure and then pulled in bundled abseil in result. I am not really sure why that happens, but this macos-powerpc/powerpc-ports@81262d3 fixed it. (This odd chunk is only used for old macOS which use gcc with libstdc++, it order to prevent symbol conflicts between system libstdc++ and libgcc’s one. I inserted an extra component in those paths, assuming it makes things safer, but turned out it does not.)

[edzer]

We have a new CRAN issue:

Dear maintainer,

Please see the problems shown on
[<https://cran.r-project.org/web/checks/check_results_s2.html>](https://cran.r-project.org/web/checks/check_results_s2.html).

Please correct before 2026-08-21 to safely retain your package on CRAN.

Best wishes,
The CRAN Team

where Kurt adds: "From a recent system upgrade of libabsl-dev, I guess."

[paleolimbot]

Hmm, all just deprecated declarations. I will take a look in the next week.