wro.cpp
short

Catch2 3.16 fixes a bug that shipped with the feature

· english· audience: working-cpp· AI-generated, reviewed by Filip Sajdak

In short

  • Catch2 3.16.0 came out on 25 August. It adds no API and changes no language requirement.
  • catch_discover_tests has split test names containing semicolons since the first version of the script. That is now fixed.
  • Test registration is faster, at roughly 3000 tests per second instead of about 1000.
  • Benchmarks that return nothing now get an optimizer barrier, so their numbers may get worse.
  • There is no demo, because Compiler Explorer has nothing newer than the 3.0 previews.

Catch2 3.16.0 came out on 25 August, the same day as LLVM 23.1 and CMake 4.4.3. It adds no API and changes no language requirement. Most of it is performance work and CMake fixes, and it is worth a short read rather than a long one.

One entry stands out, and not because it is dramatic.

A bug with a long life

The CMake function catch_discover_tests enumerates your tests at build time so CTest can run them individually. Three fixes landed in it. In the second, the <target>_TESTS variable was not escaping test names. A name containing a semicolon got split into several partial names.

The release notes are direct about the age of it: “This bug has existed since the first version of the script.”

Semicolons are how CMake spells a list. A string containing one is a list of two things unless somebody escapes it. This detail is obvious once stated and invisible until it bites. The bug survived this long because it only affects people who put punctuation in test names. The failure mode is also a test that quietly does not run rather than an error. That combination, rare trigger plus silent failure, is what lets a bug reach adulthood. A test that vanishes from the list looks exactly like a test that passed.

The other two fixes in the same function are of a piece:

  • The variable was accumulating JSON fragments alongside test names, introduced in the previous release’s refactor.
  • TEST_PREFIX and TEST_SUFFIX were having leading and trailing whitespace stripped.

Speed

Test registration got materially faster. The notes put it at roughly 3000 tests registered per second, up from about 1000. The same release brings these changes:

  • a faster initial TEST_CASE registration
  • reduced overhead the first time BENCHMARK machinery appears in a translation unit
  • rewritten string matchers, where case-sensitive matching (the default) is described as significantly faster and case-insensitive is faster in most cases
  • faster compilation where several templated test-case macros use the same set of types

One change that affects measurements

Benchmarks that return nothing now get an optimizer barrier around the call.

This is a correctness fix dressed as a performance note. When a benchmarked function returns a value, the usual trick is to keep that value alive so the optimiser cannot delete the work. With no return value there is nothing to hold. A compiler is free to notice that two consecutive calls have no observable effect, and to merge or drop them. The barrier limits what can be moved between calls, so numbers from void-returning benchmarks should be more honest and, quite possibly, worse than they were.

If you track benchmark results over time and a void-returning case suddenly reports more work after upgrading, that is the fix, not a regression.

A warning for sharding

There is a new warning when sharding is used without a deterministic test order. That is the configuration where shards can overlap or miss tests entirely.

No demo for this one

Compiler Explorer carries Catch2, but nothing past the 3.0 previews, so there is no version there new enough to show any of the above. Everything here comes from the release notes rather than from something I ran, which is worth saying rather than implying otherwise.

Sources: Catch2 v3.16.0 release notes. Release date checked in the GitHub releases API on 28 August 2026 · Compiler Explorer library versions checked the same day.