Fixes

• Fixed bug where catch_discover_tests fails when no TEST_CASEs are present (#2962)
• Fixed Clang 19 -Wc++20-extensions warning (#2968)

Improvements

• Added std::initializer_list overloads for (Unordered)RangeEquals matcher (#2915, #2919)
• Added explicit casts to silence GCC's Wconversion (#2875)
• Made the use of builtin_constant_p tricks in assertion macros configurable (#2925)
  • It is used to prod GCC-like compilers into providing warnings for the asserted expressions, but the compilers miscompile it annoyingly often.
• Cleaned out Clang-Tidy's performance-enum-size warnings
• Added support for using from_range generator with iterators with value_type = const T (#2926)
  • This is not correct value_type typedef, but it is used in the wild and the change does not make the code meaningfully worse.

Fixes

• Fixed crash when stringifying pre-1970 (epoch) dates on Windows (#2944)

Miscellaneous

• Fixes and improvements for catch_discover_tests CMake helper
  • Removed redundant CTEST_FILE param when creating the indirection file for PRE_TEST discovery mode (#2936)
  • Rewrote the test discovery logic to use output from the JSON reporter
    • This means that catch_discover_tests now requires CMake 3.19 or newer
  • Added ADD_TAGS_AS_LABELS option. If specified, each CTest test will be labeled with corrensponding Catch2's test tag
• Bumped up the minimum required CMake version to build Catch2 to 3.16
• Meson build now provides option to avoid installing Catch2
• Bazel build is moved to Bzlmod.

Improvements

• Applied the JUnit reporter's optimization from last release to the SonarQube reporter
• Suppressed -Wuseless-cast in CHECK_THROWS_MATCHES (#2904)
• Standardize exit codes for various failures
  • Running no tests is now guaranteed to exit with 2 (without the --allow-running-no-tests flag)
  • All tests skipped is now always 4 (...)
  • Assertion failures are now always 42
  • and so on

Fixes

• Fixed out-of-bounds access when the arg parser encounters single - as an argument (#2905)

Miscellaneous

• Added catch_config_prefix_messages.hpp to meson build (#2903)
• catch_discover_tests now supports skipped tests (#2873)
  • You can get the old behaviour by calling catch_discover_tests with SKIP_IS_FAILURE option.

Improvements

• Slightly improved compile times of benchmarks
• Made the resolution estimation in benchmarks slightly more precise
• Added new test case macro, TEST_CASE_PERSISTENT_FIXTURE (#2885, #1602)
  • Unlike TEST_CASE_METHOD, the same underlying instance is used for all partial runs of that test case
• MASSIVELY improved performance of the JUnit reporter when handling successful assertions (#2897)
  • For 1 test case and 10M assertions, the new reporter runs 3x faster and uses up only 8 MB of memory, while the old one needs 7 GB of memory.
• Reworked how output redirects works.
  • Combining a reporter writing to stdout with capturing reporter no longer leads to the capturing reporter seeing all of the other reporter's output.
  • The file based redirect no longer opens up a new temporary file for each partial test case run, so it will not run out of temporary files when running many tests in single process.

Miscellaneous

• Better documentation for matchers on thrown exceptions (REQUIRE_THROWS_MATCHES)
• Improved catch_discover_tests's handling of environment paths (#2878)
  • It won't reorder paths in DL_PATHS or DYLD_FRAMEWORK_PATHS args
  • It won't overwrite the environment paths for test discovery

Fixes

• Fixed Windows ARM64 build by fixing the preprocessor condition guarding use _umul128 intrinsic.
• Fixed Windows ARM64EC build by removing intrinsic pragma it does not understand. (#2858)
  • Why doesn't the x64-emulation build mode understand x64 pragmas? Don't ask me, ask the MSVC guys.
• Fixed the JUnit reporter sometimes crashing when reporting a fatal error. (#1210, #2855)
  • The binary will still exit, but through the original error, rather than secondary error inside the reporter.
  • The underlying fix applies to all reporters, not just the JUnit one, but only JUnit was currently causing troubles.

Improvements

• Disable -Wnon-virtual-dtor in Decomposer and Matchers (#2854)
• precision in floating point stringmakers defaults to max_digits10.
  • This means that floating point values will be printed with enough precision to disambiguate any two floats.
• Column wrapping ignores ansi colour codes when calculating string width (#2833, #2849)
  • This makes the output much more readable when the provided messages contain colour codes.

Miscellaneous

• Conan support improvements
  • compatibility_cppstr is set to False. (#2860)
    • This means that Conan won't let you mix library and project with different C++ standard settings.
  • The implementation library CMake target name through Conan is properly set to Catch2::Catch2 (#2861)
• SelfTest target can be built through Bazel (#2857)

Fixes

• Fixed potential compilation error when asked to generate random integers whose type did not match std::(u)int*_t.
  • This manifested itself when generating random size_ts on MacOS
• Added missing outlined destructor causing Wdelete-incomplete when compiling against libstdc++ in C++23 mode (#2852)
• Fixed regression where decomposing assertion with const instance of std::foo_ordering would not compile

Improvements

• Reintroduced support for GCC 5 and 6 (#2836)
  • As with VS2017, if they start causing trouble again, they will be dropped again.
• Added workaround for targetting newest MacOS (Sonoma) using GCC (#2837, #2839)
• CATCH_CONFIG_DEFAULT_REPORTER can now be an arbitrary reporter spec
  • Previously it could only be a plain reporter name, so it was impossible to compile in custom arguments to the reporter.
• Improved performance of generating 64bit random integers by 20+%

Miscellaneous

• Significantly improved Conan in-tree recipe (#2831)
• DL_PATHS in catch_discover_tests now supports multiple arguments (#2852, #2736)
• Fixed preprocessor logic for checking whether we expect reproducible floating point results in tests.
• Improved the floating point tests structure to avoid Wunused when the reproducibility tests are disabled (#2845)

Fixes

• Fixed OOB access when computing filename tag (from the -# flag) for file without extension (#2798)
• Fixed the linking against log on Android to be PRIVATE (#2815)
• Fixed Wuseless-cast in benchmarking internals (#2823)

Improvements

• Restored compatibility with VS2017 (#2792, #2822)
  • The baseline for Catch2 is still C++14 with some reasonable workarounds for specific compilers, so if VS2017 starts acting up again, the support will be dropped again.
• Suppressed clang-tidy's bugprone-chained-comparison in assertions (#2801)
• Improved the static analysis mode to evaluate arguments to TEST_CASE and SECTION (#2817)
  • Clang-tidy should no longer warn about runtime arguments to these macros being unused in static analysis mode.
  • Clang-tidy can warn on issues involved arguments to these macros.
• Added support for literal-zero detectors based on consteval constructors
  • This is required for compiling REQUIRE((a <=> b) == 0) against MSVC's stdlib.
  • Sadly, MSVC still cannot compile this assertion as it does not implement C++20 correctly.
  • You can use clang-cl with MSVC's stdlib instead.
  • If for some godforsaken reasons you want to understand this better, read the two relevant commits: dc51386b9fd61f99ea9c660d01867e6ad489b403, and 0787132fc82a75e3fb255aa9484ca1dc1eff2a30.

Miscellaneous

• Disabled tests for FP random generator reproducibility on non-SSE2 x86 targets (#2796)
• Modified the in-tree Conan recipe to support Conan 2 (#2805)

Fixes

• Fixed -Wsubobject-linkage in the Console reporter (#2794)
• Fixed adding new CLI Options to lvalue parser using | (#2787)

Improvements

• Significantly improved performance of the CLI parsing.
  • This includes the cost of preparing the CLI parser, so Catch2's binaries start much faster.

Miscellaneous

• Added support for Bazel modules (#2781)
• Added CMake option to disable the build reproducibility settings (#2785)
• Added log library linking to the Meson build (#2784)

Improvements

• Introduced CATCH_CONFIG_PREFIX_MESSAGES to prefix only logging macros (#2544)
  • This means INFO, UNSCOPED_INFO, WARN and CAPTURE.
• Section hints in static analysis mode are now const
  • This prevents Clang-Tidy from complaining about misc-const-correctness.
• from_range generator supports C arrays and ranges that require ADL (#2737)
• Stringification support for std::optional now also includes std::nullopt (#2740)
• The Console reporter flushes output after writing benchmark runtime estimate.
  • This means that you can immediately see for how long the benchmark is expected to run.
• Added workaround to enable compilation with ICC 19.1 (#2551, #2766)
• Compiling Catch2 for XBox should work out of the box (#2772)
  • Catch2 should automatically disable getenv when compiled for XBox.
• Compiling Catch2 with exceptions disabled no longer triggers Wunused-function (#2726)
• random Generators for integral types are now reproducible across different platforms
  • Unlike <random>, Catch2's generators also support 1 byte integral types (char, bool, ...)
• random Generators for float and double are now reproducible across different platforms
  • long double varies across different platforms too much to be reproducible
  • This guarantee applies only to platforms with IEEE 754 floats.

Fixes

• UDL declaration inside Catch2 are now strictly conforming to the standard
  • operator "" _a is UB, operator ""_a is fine. Seriously.
• Fixed CAPTURE tests failing to compile in C++23 mode (#2744)
• Fixed missing include in catch_message.hpp (#2758)
• Fixed CHECK_ELSE suppressing failure from uncaught exceptions(#2723)

Miscellaneous

• The documentation for specifying which tests to run through commandline has been completely rewritten (#2738)
• Fixed installation when building Catch2 with meson (#2722, #2742)
• Fixed catch_discover_tests when using custom reporter and PRE_TEST discovery mode (#2747)
• catch_discover_tests supports multi-config CMake generator in PRE_TEST discovery mode (#2739, #2746)