Modernization Phase 2.0: modules scaffolding + compile-time baselines

[ptesavol]

Part 2 (C++ modules migration) begins. This phase adds the build scaffolding and — the main deliverable — the baseline numbers every later phase's compile-time claims will be measured against.

Scaffolding

• cmake/StreamrModules.cmake (canonical, synced to every package by sync-cmake-files.sh — same standalone-publishability policy as the other helpers): Ninja + non-AppleClang guards, CMP0155 NEW, streamr_add_module_library() (STATIC + FILE_SET CXX_MODULES rooted at the package's modules/), streamr_enable_imports() for import-using targets (tests), and the STREAMR_IMPORT_STD opt-in (OFF per the locked decision; requires the CMake-version-specific experimental UUID explicitly). Module scanning stays globally OFF — targets opt in via the helpers, so clangd's compile commands stay clean for all non-migrated code (the exact failure mode that bit us in Phase 1.1).
• bench.sh: clean / incremental [header] / trace modes measuring the root single-tree build (the workflow modules optimize — all 108 TUs: every test, generated proto, proxyclient). Prints host/compiler/parallelism with each run; trace builds a separate build-bench-trace/ tree (reuses build/vcpkg_installed) and feeds ClangBuildAnalyzer.

Baseline (macOS, Apple Silicon, clang 22.1.8, Debug, Ninja -j10)

Metric	Baseline
Clean build (108 TUs)	configure 8 s + build 102 s
Incremental: touch StreamID.hpp	10 TUs, 19 s
Incremental: touch SLogger.hpp	48 TUs, 62–70 s
CPU split (-ftime-trace)	parse 859 s vs codegen 61 s → 93% parsing

ClangBuildAnalyzer top expensive headers: DhtRpc.pb.h 96 s (33×), gtest.h 84 s (stays #include), SLogger.hpp 81 s (47×), Logger.hpp 66 s (50×), ConnectionManager.hpp 38 s (avg 9.5 s per inclusion), RpcCommunicator.hpp 35 s, Identifiers.hpp 31 s.

The data confirms the migration's premise: >90% of compile CPU is repeated header parsing, and the top of the list is exactly the .pb.h + streamr-header stack the module partitions will wrap. Template-instantiation costs (nlohmann ~22 s, std::format ~13 s, magic_enum ~13 s) are recorded too — modules won't move those, and post-migration numbers should be read with that in mind.

Linux/iOS baselines: to be captured before the Phase 2.4 checkpoint (noted in MODERNIZATION.md).

Next

Phase 2.1: the streamr-eventemitter canary (std-only, concept-heavy, single header) + streamr-json — first real module partitions, export(TARGETS) smoke test, and the clangd-experimental-modules check.

🤖 Generated with Claude Code