prior shape (bare json list -> 'tools.build:cxxflags=[...]') leaked into
the consumer/rippled build via the conan-generated toolchain
(CMAKE_CXX_FLAGS_INIT). on macOS, where the build action doesn't
override CMAKE_CXX_FLAGS, this would silently apply the workaround flag
to rippled itself.
reshape the input to a json object keyed by Conan package pattern:
{"grpc/*":["-Wno-..."]}
action emits package-pattern scoped lines:
grpc/*:tools.build:cxxflags=["-Wno-..."]
flags only apply while building the matching dependency. consumer
toolchain stays clean. python validator rejects the consumer pattern
('&') and malformed shapes.
- restore the 8 .yml.disabled workflows back to .yml
- drop the gcc-13 gcov coverage matrix row; native clang-20 llvm-cov is
now the only codecov reporter
- update minimal_matrix indices to match
- remove the TEMP narrowing block that filtered to just the llvm-cov row
- log line says '7 configs (build x6 + clang-20 llvm-cov coverage)' to
reflect the new shape
reverts part of f8a30c528d - removing this line dropped the apple-clang
grpc workaround that the dependencies action used to apply
unconditionally before the matrix-driven refactor. the .disabled rename
is itself a TEMP measure being reverted before merge, so the field needs
to be in place when the macOS workflow comes back.
restore the conan_deps_cxxflags addition to xahau-ga-macos.yml.disabled.
prior commit framed it as dead code, but the .disabled rename is itself
a TEMP measure being reverted before merge - the field is needed for the
macOS workflow to keep building grpc 1.50.1 once it's re-enabled, since
the unconditional workaround was removed from the dependencies action.
- CodeCoverageLLVM.cmake: BASE_DIRECTORY was parsed but never used
(no analog to gcovr's -r in the llvm-cov commands we emit). dead arg.
- xahau-ga-macos.yml.disabled: revert the conan_deps_cxxflags addition.
the file is disabled so the edit was bit-rotting in unreachable code.
whoever revives the macOS workflow can wire up the field then.
two bugs caught in review:
- RippledSettings.cmake: the clang-only guard for coverage_tool=llvm ran
before coverage_test could auto-enable coverage. so
'-Dcoverage_tool=llvm -Dcoverage_test=Foo' on a gcc build slipped past
the guard and produced an instrumentation/tool mismatch. move the
guard after the auto-enable block.
- CodeCoverageLLVM.cmake: _find_llvm_cov_tools unconditionally
preferred xcrun's tools on APPLE, which on a homebrew clang-N build
would pair the user's clang with xcode's llvm-cov - exactly the
version mismatch the helper is trying to avoid. gate the xcrun branch
on CMAKE_CXX_COMPILER_ID == AppleClang.
new dependencies action input `conan_deps_cxxflags` (json list, default
'[]') drives `tools.build:cxxflags` in the conan profile. clearly named
to indicate the flags only apply to conan dependency builds, not the
rippled build itself.
removes the per-os/per-compiler hardcoded workaround (linux clang
conditional + unconditional macOS block) that used to set the same flag
in two places kept-in-sync by hand.
call sites:
- xahau-ga-nix.yml: clang-20 coverage row gets the
-Wno-missing-template-arg-list-after-template-kw workaround for grpc 1.50.1
- xahau-ga-macos.yml.disabled: same flag plumbed through (preserves
prior behaviour when re-enabled)
drop the workaround entries when grpc is bumped past the fix.
bare 'rippled' wasn't on PATH and the build dir isn't '.', so the
profile-collection step failed with 'No such file or directory'.
splice the resolved absolute path back into Cov_EXECUTABLE before
the run command consumes it.
renames every non-nix workflow to .yml.disabled so they stop firing on
PR pushes while we iterate on the llvm-cov coverage row. MUST be
reverted before merging to dev.
grpc 1.50.1 hits -Werror=missing-template-arg-list-after-template-kw on
clang-19+. macOS clang already had this workaround in the conan profile;
mirror it to the Linux branch, gated on compiler==clang.
prior guard checked base_ref against ['dev','candidate','release'] which
excluded the very PR we're iterating on. drop the guard - this branch
is for iteration only and must be reverted before merging anyway.
introduces native llvm source-based coverage as an alternative to the
existing gcov + gcovr pipeline. driven by a new -Dcoverage_tool cache
variable (default 'gcov' for backwards compat; 'llvm' enables
-fprofile-instr-generate / -fcoverage-mapping + llvm-profdata + llvm-cov).
cmake:
- RippledSettings.cmake: add coverage_tool with validation
- RippledInterface.cmake: split coverage compile/link flags by tool
- RippledCov.cmake: dispatch to the new helper when tool=llvm
- CodeCoverageLLVM.cmake (new): setup_target_for_coverage_llvm() driving
profraw -> profdata -> export with format-aware output (lcov/json/txt/html)
ci:
- new clang-20 llvm-cov coverage matrix row (apt.llvm.org bootstrap for
clang >= 19 since 24.04 default repos cap at clang-18)
- conditional install of llvm-N vs gcovr based on coverage_tool
- artifact + codecov upload generalised to coverage.lcov | coverage.xml
- coverage cmake-args switched to *_FLAGS_DEBUG so the build action's
stdlib flag isn't clobbered
temp (revert before merging to dev):
- matrix narrowed to just the llvm-cov row on non-main refs so we can
iterate without burning runners
- 'coverage-llm' added to push trigger for direct-push CI runs
