I noticed this too. Binaries produced from MSVC are a lot slower than clang build. Also, clang 16 builds are faster than clang 15.

Can you check if this is still the case with version 10.1? We reduced the number of allocations, which could explain some of those speed differences.

There are a little bit of differences in higher efforts, but lower efforts are more pronounced, but probably fast enough to not be noticed.

0x0.st/H7us.txt

There are a little bit of differences in higher efforts, but lower efforts are more pronounced, but probably fast enough to not be noticed.

0x0.st/H7us.txt

Would you be able to verify whether the build artefacts in https://github.com/libjxl/libjxl/actions/runs/8146246627 solve the issue?

v0.10.1 release and Github artifacts binaries comparison
bench_github.txt
Clang-cl build is faster at e1, but then is slightly slower and sometimes slightly faster. Maybe there are some commits that slowed it down.

I compared my local MSVC and Clang-Cl build with the same commit, and the result is consistent, with Clang-Cl being faster.
bench_local.txt
No additional compiler flags, just -DCMAKE_BUILD_TYPE=Release

What CPU are you running this on?

Intel Core i5-7200U

Edit: cmake configurations
cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DBUILD_TESTING=OFF -DJPEGXL_ENABLE_BENCHMARK=ON -DJPEGXL_ENABLE_OPENEXR=OFF -DJPEGXL_ENABLE_TCMALLOC=OFF -DJPEGXL_ENABLE_VIEWERS=OFF -DCMAKE_TOOLCHAIN_FILE=C:\\Users\\A\\scoop\\apps\\vcpkg\\current\\scripts\\buildsystems\\vcpkg.cmake -DVCPKG_TARGET_TRIPLET=x64-windows-static --fresh ..

There are a little bit of differences in higher efforts, but lower efforts are more pronounced, but probably fast enough to not be noticed.
0x0.st/H7us.txt

Would you be able to verify whether the build artefacts in https://github.com/libjxl/libjxl/actions/runs/8146246627 solve the issue?

Looking at the configure step of the Github actions, it seems like it was still being compiled by MSVC, not Clang-Cl. That explains this discrepancy.

Clang-cl build is faster at e1, but then is slightly slower and sometimes slightly faster. Maybe there are some commits that slowed it down.

https://github.com/libjxl/libjxl/actions/runs/8146246627/job/22264345808

-- The C compiler identification is MSVC 19.29.30154.0
-- The CXX compiler identification is MSVC 19.29.30154.0

If the build system is not specified, CMake is defaulted to MSBuild. MSBuild silently ignores the -DCMAKE_CXX_COMPILER flag, only Ninja accepts it.