Compare commits

...

6 commits

Author SHA1 Message Date
Lukas Brübach
cba909241b [CI] Cache the Windows compile results with sccache
A pull request that touches one file rebuilds it, but sccache can hand
back the objects for everything else, and the GitHub Actions cache keeps
them between runs, so the second build of a branch costs about what the
first one stored and nothing more.

sccache has to wrap the compiler, and cmake only wraps it for the
generators that run the compiler themselves: the Visual Studio
generator drives MSBuild, which would drop a launcher on the floor
without a word. The cache therefore goes on the Ninja job, and
compile.sh now refuses --sccache on any other generator instead of
pretending to cache, and prints the sccache statistics next to the
existing ccache ones so the hit rate is visible in the log.

For MSVC, a cacheable object has to carry its debug information
inside, which is /Z7 rather than /Zi; sccache cannot restore the
latter, it can only restore a whole object file. Ask CMake for the
format by name through CMP0141 and CMAKE_MSVC_DEBUG_INFORMATION_FORMAT
instead of hardcoding /Zi in the flags, where it would end up next to
/Z7, and keep the format as it was for every build that does not
cache. The linker still gets /DEBUG, so the release PDBs are written
either way, which is what that flag is there for.

The policy is set before project(), since that is when it takes
effect, and behind an if(POLICY) guard because it does not exist
before CMake 3.25 - where the old defaults remain the only option.
2026-10-14 08:14:20 +02:00
Lukas Brübach
1f8d252c68 [CI] Give the Ninja leg the same vcpkg scope as the Visual Studio leg 2026-10-14 08:14:20 +02:00
Lukas Brübach
28a5bd44e7 [CI] Measure Ninja against the Visual Studio generator on Windows
Ninja scales its job pool to the machine it runs on, and it lets a
compiler launcher through, which MSBuild does not; those two are the
reasons to expect it to beat the Visual Studio generator on a 4-core
runner. That is an expectation, not a measurement, so run it as an A/B
next to the Visual Studio job for a week: the extra matrix entry is
allowed to fail, keeps its own " (Ninja)" suffix in the job name, is
never uploaded to a release and never contributes PDBs, so it can
neither block a pull request nor ship anything.

Ninja needs cl.exe and the MSVC environment on PATH, which the Visual
Studio generator gets for free by locating the compiler itself, hence
the setup-msvc-dev step for the experimental job alone. Its build gets
neither the MSBuild switches nor a job count: Ninja rejects -p: options
and sizes its own pool, while the Visual Studio job keeps both. cmake
takes the generator platform from the environment, so it is unset for
everything that is not a Visual Studio generator, where it would
otherwise follow the generator around.

Delete this entry once the numbers are in: if Ninja wins, switch the
Windows matrix to it and fix the PDB paths with it.
2026-10-14 08:14:20 +02:00
Lukas Brübach
255c713ce4 [CI] Profile the MSBuild run this pull request measures 2026-10-14 08:14:19 +02:00
Lukas Brübach
af05b85f29 [CI] Build the Windows solution on more than one MSBuild node
cmake --build only passes a job count to MSBuild when it is asked for
one, and .ci/compile.sh never asked, so every Visual Studio generator
build ran single-node: projects were built one after another and MTT
only ever parallelized the files of whichever project happened to be
building. Ask for one node per core, overridable with
BUILD_PARALLEL_LEVEL, and fall back to a single node when the core
count is not exposed.

Alongside it set EnforceProcessCountAcrossBuilds and CL_MPCount, the
knobs Microsoft's build throughput write-up pairs with MTT: without
them the node count and MTT multiply into more cl.exe processes than
the runner has cores for.
2026-10-14 08:14:19 +02:00
Lukas Brübach
8ed6354595 [CI] Collect an MSBuild performance summary and binary log on request
The Windows job spends about 28 minutes inside one "Build project"
group, with nothing in it that says where the time goes: MSBuild runs
single-node because no --parallel is passed, so projects build one
after another and it is not visible which projects, tasks or phases are
responsible.

Add a --profile flag that makes the Windows build write
-flp:PerformanceSummary and -bl, repeat the summary in the job log so
the numbers need no download, and report the core count, which decides
how much parallelism is worth asking for. The binary log is uploaded as
an artifact only on the runs that asked for it.

The workflow drives it from a build_profile matrix key that is empty by
default, so no build changes behaviour until someone sets it.
2026-10-14 08:14:19 +02:00
3 changed files with 183 additions and 11 deletions

View file

@ -14,11 +14,13 @@
# --debug or --release sets the build type ie CMAKE_BUILD_TYPE
# --ccache [<size>] uses ccache and shows stats, optionally provide size
# --evict-ccache <age> runs ccache eviction based on given age after build
# --sccache uses sccache instead, needs a generator that runs the compiler itself
# --dir <dir> sets the name of the build dir, default is "build"
# --cmake-generator <generator> sets CMAKE_GENERATOR as used by cmake
# --target-macos-version <version> sets the min os version - only used for macOS builds
# uses env: BUILDTYPE MAKE_INSTALL MAKE_PACKAGE PACKAGE_TYPE PACKAGE_SUFFIX MAKE_SERVER MAKE_NO_CLIENT MAKE_TEST USE_CCACHE CCACHE_SIZE CCACHE_EVICTION_AGE BUILD_DIR CMAKE_GENERATOR TARGET_MACOS_VERSION MAKE_SANITIZE MAKE_COVERAGE VCPKG_TARGET_MACOS_VERSION
# (correspond to args: --debug/--release --install --package <package type> --suffix <suffix> --server --test --sanitize --coverage --ccache <ccache_size> --dir <dir>)
# --profile writes an MSBuild performance summary and binary log, Windows only
# uses env: BUILDTYPE MAKE_INSTALL MAKE_PACKAGE PACKAGE_TYPE PACKAGE_SUFFIX MAKE_SERVER MAKE_NO_CLIENT MAKE_TEST USE_CCACHE CCACHE_SIZE CCACHE_EVICTION_AGE BUILD_DIR CMAKE_GENERATOR TARGET_MACOS_VERSION MAKE_SANITIZE MAKE_COVERAGE VCPKG_TARGET_MACOS_VERSION BUILD_PARALLEL_LEVEL BUILD_PROFILE USE_SCCACHE
# (correspond to args: --debug/--release --install --package <package type> --suffix <suffix> --server --test --sanitize --coverage --ccache <ccache_size> --dir <dir> --profile --sccache)
# exitcode: 1 for failure, 3 for invalid arguments
# Read arguments
@ -93,6 +95,10 @@ while [[ $# != 0 ]]; do
CCACHE_EVICTION_AGE=$1
shift
;;
'--sccache')
USE_SCCACHE=1
shift
;;
'--vcpkg')
USE_VCPKG=1
shift
@ -115,6 +121,10 @@ while [[ $# != 0 ]]; do
export CMAKE_GENERATOR=$1
shift
;;
'--profile')
BUILD_PROFILE=1
shift
;;
'--target-macos-version')
shift
if [[ $# == 0 ]]; then
@ -133,6 +143,12 @@ done
set -e
# cmake reads the generator platform from the environment, so a leftover platform
# follows the generator around; Ninja has no platform to select.
if [[ $CMAKE_GENERATOR != "Visual Studio"* ]]; then
unset CMAKE_GENERATOR_PLATFORM
fi
# Setup
./servatrice/check_schema_version.sh
if [[ ! $BUILDTYPE ]]; then
@ -158,6 +174,16 @@ export CMAKE_POLICY_VERSION_MINIMUM=3.10
# Add cmake flags
flags=("-DCMAKE_BUILD_TYPE=$BUILDTYPE")
if [[ $USE_CCACHE && $USE_SCCACHE ]]; then
echo "::error file=$0::ccache and sccache at the same time is not a supported combination"
exit 1
fi
if [[ $USE_SCCACHE && $CMAKE_GENERATOR != "Ninja"* ]]; then
# A compiler cache wraps the compiler, and only generators that run the compiler
# themselves are given a launcher to wrap: cmake does not pass one to MSBuild.
echo "::error file=$0::sccache needs a Ninja generator, not '$CMAKE_GENERATOR', which would ignore it"
exit 1
fi
if [[ $MAKE_SERVER ]]; then
flags+=("-DWITH_SERVER=1")
fi
@ -188,6 +214,10 @@ if [[ $USE_CCACHE ]]; then
ccache --max-size "$CCACHE_SIZE"
fi
fi
if [[ $USE_SCCACHE ]]; then
flags+=("-DUSE_SCCACHE=1")
flags+=("-DCMAKE_C_COMPILER_LAUNCHER=sccache" "-DCMAKE_CXX_COMPILER_LAUNCHER=sccache")
fi
if [[ $PACKAGE_TYPE ]]; then
flags+=("-DCPACK_GENERATOR=$PACKAGE_TYPE")
fi
@ -209,6 +239,22 @@ function ccachestatsverbose() {
fi
}
function cpuinfo() {
# The number of cores decides how much build parallelism is useful; the hosted
# runners are not all sized alike, so read it off the machine instead of assuming.
case "$RUNNER_OS" in
Windows)
echo "cores: ${NUMBER_OF_PROCESSORS:-unknown}"
;;
macOS)
echo "cores: $(sysctl -n hw.ncpu)"
;;
*)
echo "cores: $(nproc 2>/dev/null || echo unknown)"
;;
esac
}
# Compile
if [[ $RUNNER_OS == macOS ]]; then
# QTDIR is needed for macOS since we actually only use the cached thin Qt binaries instead of the install-qt-action,
@ -293,9 +339,33 @@ if [[ $RUNNER_OS == macOS ]]; then
fi
elif [[ $RUNNER_OS == Windows ]]; then
# Enable MTT, see https://devblogs.microsoft.com/cppblog/improved-parallelism-in-msbuild/
# and https://devblogs.microsoft.com/cppblog/cpp-build-throughput-investigation-and-tune-up/#multitooltask-mtt
buildflags+=(-- -p:UseMultiToolTask=true -p:EnableClServerMode=true)
if [[ $CMAKE_GENERATOR == "Visual Studio"* ]]; then
# cmake only hands MSBuild a job count when it is asked for, and without it
# MSBuild builds a single node: one project at a time, whatever MTT does inside
# it. Ask for a node per core.
if [[ ! $BUILD_PARALLEL_LEVEL ]]; then
BUILD_PARALLEL_LEVEL="${NUMBER_OF_PROCESSORS:-1}"
fi
buildflags+=(--parallel "$BUILD_PARALLEL_LEVEL")
# Enable MTT, see https://devblogs.microsoft.com/cppblog/improved-parallelism-in-msbuild/
# and https://devblogs.microsoft.com/cppblog/cpp-build-throughput-investigation-and-tune-up/#multitooltask-mtt
# EnforceProcessCountAcrossBuilds makes the CL_MPCount ceiling apply across all
# the projects the nodes are building, instead of each node running its own
# cl.exe processes and /m and MTT multiplying into an oversubscribed machine.
buildflags+=(-- -p:UseMultiToolTask=true -p:EnableClServerMode=true -p:EnforceProcessCountAcrossBuilds=true -p:CL_MPCount="$BUILD_PARALLEL_LEVEL")
if [[ $BUILD_PROFILE ]]; then
# A performance summary says which projects and tasks cost what, the binary log
# keeps the full timeline; both are far too noisy to always produce. The files
# land next to the solution, i.e. in the build dir the workflow uploads. These
# join the MTT options above, which already carried the "--" separator.
buildflags+=("-flp:PerformanceSummary;v=q;LogFile=msbuild-perf.log" -bl:msbuild.binlog)
fi
else
# Ninja runs the compiler itself, so it neither needs the MSVC switches nor
# wants a job count: it sizes its own pool for the cores it finds.
echo "ninja $(ninja --version)"
fi
fi
if [[ $USE_CCACHE ]]; then
@ -311,10 +381,19 @@ cmake .. "${flags[@]}"
echo "::endgroup::"
echo "::group::Build project"
cpuinfo
echo "Running cmake --build with flags: ${buildflags[*]}"
cmake --build . "${buildflags[@]}"
echo "::endgroup::"
if [[ $BUILD_PROFILE && $RUNNER_OS == Windows && -f msbuild-perf.log ]]; then
# Repeat the summary in the job log so the numbers are readable without
# downloading the binary log, which is hundreds of megabytes.
echo "::group::MSBuild performance summary"
cat msbuild-perf.log
echo "::endgroup::"
fi
if [[ $USE_CCACHE ]]; then
if [[ $CCACHE_EVICTION_AGE ]]; then
echo "::group::evict ccache files older than $CCACHE_EVICTION_AGE"
@ -324,6 +403,10 @@ if [[ $USE_CCACHE ]]; then
echo "::group::Show ccache stats again"
ccachestatsverbose
echo "::endgroup::"
elif [[ $USE_SCCACHE ]]; then
echo "::group::Show sccache stats"
sccache --show-stats
echo "::endgroup::"
elif [[ $CCACHE_EVICTION_AGE ]]; then
echo "::error file=$0::ccache eviction is enabled while ccache is disabled!"
fi

View file

@ -385,6 +385,12 @@ jobs:
target: 10
runner: windows-2025 # https://github.com/actions/runner-images/blob/main/images/windows/Windows2025-VS2026-Readme.md
# Set build_profile: 1 to have the build write an MSBuild performance
# summary and a binary log; empty on every entry that runs by default.
# Enabled here on purpose: this pull request measures MSBuild
# parallelism, so its run is the one worth profiling.
build_profile: 1
cmake_generator: "Visual Studio 18 2026"
cmake_generator_platform: x64
make_package: 1
@ -395,11 +401,37 @@ jobs:
vcpkg_scope: 'vcpkg/windows10-x64-vs2026/'
vcpkg_upload: 1 # only one leg per vcpkg_scope hands its archives over
name: ${{ matrix.os }} ${{ matrix.target }}${{ matrix.soc == 'Intel' && ' Intel' || '' }}${{ matrix.type == 'Debug' && ' Debug' || '' }}
# A/B run: is Ninja faster than the Visual Studio generator on this runner?
# It is allowed to fail and is never published, see the steps that build
# a release below. Delete this entry once the answer is known.
- os: Windows
target: 10
runner: windows-2025
cmake_generator: Ninja
experiment: yes
make_package: 1
package_suffix: "-Win10-ninja"
qt_version: 6.11.*
qt_modules: qtimageformats qtmultimedia qtshadertools qtwebsockets
type: Release
use_sccache: 1
variant: ' (Ninja)'
# The same scope as the Visual Studio leg, so both halves of the A/B
# start from the same vcpkg cache and the comparison is about the
# generator alone. Without it this leg's prefix is empty, which makes
# the cache action restore the whole vcpkg/ namespace and write entries
# that no scope claims. vcpkg_upload is deliberately left unset, so only
# the Visual Studio leg hands its archives over.
vcpkg_scope: 'vcpkg/windows10-x64-vs2026/'
name: ${{ matrix.os }} ${{ matrix.target }}${{ matrix.soc == 'Intel' && ' Intel' || '' }}${{ matrix.type == 'Debug' && ' Debug' || '' }}${{ matrix.variant }}
needs: configure
runs-on: ${{ matrix.runner }}
continue-on-error: ${{ matrix.experiment == 'yes' }}
timeout-minutes: 100
env:
BUILD_PROFILE: '${{ matrix.build_profile }}'
CCACHE_DIR: ${{ github.workspace }}/.cache/
CCACHE_SIZE: 600M # space of all repo is 10Gi: https://docs.github.com/en/actions/using-workflows/caching-dependencies-to-speed-up-workflows#usage-limits-and-eviction-policy
@ -482,6 +514,28 @@ jobs:
modules: ${{ matrix.qt_modules }}
version: ${{ matrix.qt_version }}
# Ninja invokes cl.exe directly and needs the MSVC environment on PATH; the
# Visual Studio generator locates the compiler on its own and skips this.
- name: "[Windows/Ninja] Set up MSVC environment"
if: matrix.os == 'Windows' && matrix.cmake_generator == 'Ninja'
uses: TheMrMilchmann/setup-msvc-dev@v4.1.0
with:
arch: x64
# sccache is only useful for the Ninja job: a compiler cache has to wrap the
# compiler, and MSBuild is driven by cmake instead of running it, so a
# launcher handed to MSBuild is ignored without a word.
- name: "[Windows/Ninja] Set up sccache"
if: matrix.os == 'Windows' && matrix.cmake_generator == 'Ninja'
uses: mozilla-actions/sccache-action@v0.0.11
env:
SCCACHE_GHA_ENABLED: "true"
# A fork pull request can read the cache, but the token it runs with
# cannot write to the base branch's scope.
SCCACHE_GHA_RW_MODE: ${{ github.event.pull_request.head.repo.fork && 'READ_ONLY' || 'READ_WRITE' }}
# A cache that cannot be reached must not fail the build.
SCCACHE_IGNORE_SERVER_IO_ERROR: "1"
- name: "[Windows] Install NSIS"
if: matrix.os == 'Windows'
shell: bash
@ -530,8 +584,9 @@ jobs:
MACOS_CI_KEYCHAIN_PWD: ${{ secrets.PROD_MACOS_CI_KEYCHAIN_PWD }}
MAKE_PACKAGE: '${{ matrix.make_package }}'
PACKAGE_SUFFIX: '${{ matrix.package_suffix }}'
TARGET_MACOS_VERSION: ${{ matrix.override_target }}
TARGET_MACOS_VERSION: '${{ matrix.override_target }}'
USE_CCACHE: ${{ matrix.use_ccache }}
USE_SCCACHE: ${{ matrix.use_sccache }}
# Only pushes to the default branch may write the shared binary cache: the
# archives a run compiles are only as trustworthy as the code they were
# built from, and only merged code may populate the scope that master and
@ -690,8 +745,11 @@ jobs:
if-no-files-found: error
path: ${{ steps.build.outputs.path }}
# Ninja writes its PDBs next to the binaries instead of into a Release
# subdirectory, and the experimental job is not a source of release debug
# symbols, so it keeps out of both.
- name: "[Windows] Upload PDBs (Program Databases)"
if: matrix.os == 'Windows' && github.ref_type != 'tag'
if: matrix.os == 'Windows' && matrix.experiment != 'yes' && github.ref_type != 'tag'
uses: actions/upload-artifact@v7
with:
if-no-files-found: error
@ -701,8 +759,19 @@ jobs:
build/oracle/Release/*.pdb
build/servatrice/Release/*.pdb
# Written only when the matrix entry sets build_profile, so the artifact
# exists exactly on the runs that were asked to produce one.
- name: "[Windows] Upload MSBuild binary log"
if: always() && matrix.os == 'Windows' && env.BUILD_PROFILE == '1'
uses: actions/upload-artifact@v7
with:
if-no-files-found: error
name: ${{ steps.build.outputs.name }}-msbuild
path: build/msbuild.binlog
retention-days: 7
- name: "Upload to release"
if: needs.configure.outputs.tag != null && matrix.make_package == '1'
if: needs.configure.outputs.tag != null && matrix.make_package == '1' && matrix.experiment != 'yes'
id: upload_release
shell: bash
env:

View file

@ -8,8 +8,19 @@
# 3.16 required for Qt6 and target_precompile_headers()
cmake_minimum_required(VERSION 3.16)
# Let CMake pick the MSVC debug information format through
# CMAKE_MSVC_DEBUG_INFORMATION_FORMAT instead of writing it into the
# CMAKE_<LANG>_FLAGS_<CONFIG> defaults. The policy has to be settled before the
# first project() enables a language, and it does not exist before CMake 3.25,
# where the old defaults are the only thing available.
if(POLICY CMP0141)
cmake_policy(SET CMP0141 NEW)
endif()
# Use compiler cache (ccache)
option(USE_CCACHE "Cache the build results with ccache" ON)
# Use compiler cache (sccache), for generators that invoke the compiler themselves
option(USE_SCCACHE "Cache the build results with sccache" OFF)
# Treat warnings as errors (Debug builds only)
option(WARNING_AS_ERROR "Treat warnings as errors in debug builds" ON)
# Check for translation updates
@ -156,8 +167,17 @@ endif()
# Define proper compilation flags
if(MSVC)
# Disable Warning C4251, C++20 compatibility, Multi-threaded Builds, Warn Detection, Unwind Semantics, Debug Symbols
set(CMAKE_CXX_FLAGS "/wd4251 /Zc:__cplusplus /std:c++20 /permissive- /W4 /MP /EHsc /Zi")
# Debug information: a cache can only store and restore objects that carry it
# inside, so a cached build asks for /Z7 (Embedded); otherwise /Zi (ProgramDatabase).
# Either way the linker is told to write the PDB, in every configuration.
if(USE_SCCACHE)
set(CMAKE_MSVC_DEBUG_INFORMATION_FORMAT Embedded)
else()
set(CMAKE_MSVC_DEBUG_INFORMATION_FORMAT ProgramDatabase)
endif()
# Disable Warning C4251, C++20 compatibility, Multi-threaded Builds, Warn Detection, Unwind Semantics
set(CMAKE_CXX_FLAGS "/wd4251 /Zc:__cplusplus /std:c++20 /permissive- /W4 /MP /EHsc")
# Visual Studio: Maximum Optimization, Multi-threaded DLL
set(CMAKE_CXX_FLAGS_RELEASE "/Ox /MD")
# Visual Studio: No Optimization, Multi-threaded Debug DLL