Skip to content

partial c++20 module support#2516

Open
TroyKomodo wants to merge 12 commits intomozilla:mainfrom
TroyKomodo:troy/cxx20-modules
Open

partial c++20 module support#2516
TroyKomodo wants to merge 12 commits intomozilla:mainfrom
TroyKomodo:troy/cxx20-modules

Conversation

@TroyKomodo
Copy link
Contributor

@TroyKomodo TroyKomodo commented Dec 18, 2025

Adds partial support for c++20 modules when using clang.

Also detects when modules are used in MSVC / GCC and disable cache for those calls.

fixes #2216

partially addresses #2095

the clang c++20 module implementation (as well as msvc) is rather easy to parse and understand so adding support for it is fairly trivial. GCC however is a mess. An entirely different approach will be needed to get GCC module support which is far out side of the scope for this initial PR.

I will submit future PRs to get MSVC / GCC.

Test with nix

(pkgs.sccache.overrideAttrs (final: old: {
  version = "troy/cxx20-modules";
  cargoHash = "sha256-tCQrZ6cPGJrxHgeSJ4ZqZs6otWvVu6Ys3svAiwvmSDc=";
  src = fetchFromGitHub {
    owner = "troykomodo";
    repo = "sccache";
    rev = "troy/cxx20-modules";
    sha256 = "sha256-A0n/nsNPTHRC0Z6ERwCpoR1w3O/xl6d1yCw/5GeqNd4=";
  };
  cargoDeps = pkgs.rustPlatform.fetchCargoVendor {
    inherit (final) pname src version;
    hash = final.cargoHash;
  };
}))

@codecov-commenter
Copy link

codecov-commenter commented Dec 18, 2025

Codecov Report

❌ Patch coverage is 87.70764% with 37 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.39%. Comparing base (81954de) to head (10fc4f6).

Files with missing lines Patch % Lines
src/compiler/clang.rs 84.45% 23 Missing ⚠️
src/compiler/gcc.rs 85.54% 12 Missing ⚠️
src/compiler/compiler.rs 33.33% 2 Missing ⚠️
Additional details and impacted files
@@            Coverage Diff             @@
##             main    #2516      +/-   ##
==========================================
+ Coverage   73.16%   73.39%   +0.22%     
==========================================
  Files          68       68              
  Lines       36743    37208     +465     
==========================================
+ Hits        26884    27309     +425     
- Misses       9859     9899      +40     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

@ChuanqiXu9
Copy link

FYI ccache/ccache#1523

Copy link
Contributor

@mathstuf mathstuf left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The code looks good to me. I'll probably try to test it some time next week.

I can help out with GCC module map file parsing too (it will need to actually be parsed due to the $root meta-descriptor). However, the output path is also given by it…that may prove more difficult to "know".

@TroyKomodo
Copy link
Contributor Author

The code looks good to me. I'll probably try to test it some time next week.

I can help out with GCC module map file parsing too (it will need to actually be parsed due to the $root meta-descriptor). However, the output path is also given by it…that may prove more difficult to "know".

Thanks! It was my first time contributing to sccache so I'm not too familiar with the process.

I will fix up the clippy issues + windows test issues :p

@mathstuf
Copy link
Contributor

It looks like BMI-only compilations are not properly considered:

467:    actual-stdout> [2-1->6/8@76.9] Building CXX object 'CMakeFiles/CXXModules__export_from_ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi'          
467:    actual-stdout> FAILED: [code=254] CMakeFiles/CXXModules__export_from_ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi 
467:    actual-stdout> /home/boeckb/code/depot/group-devel/sccache/src/target/debug/sccache  /usr/lib64/ccache/clang++   -g -std=gnu++20 --precompile -MD -MT 'CMakeFiles/CXXModules__
export_from_ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi' -MF CMakeFiles/CXXModules__export_from_ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi.d @CMakeFiles/CXXModules__export_from_
ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi.modmap -o 'CMakeFiles/CXXModules__export_from_ninja@synth_cf4e03cd8b1e.dir/35a7889ff75f.bmi' -c /home/boeckb/misc/builds/cmake/build-cla
ng/Tests/RunCMake/CXXModulesCompile/exp-mods-from-ninja-ninja-install/lib/cxx/miu/importable.cxx
467:    actual-stdout> sccache: encountered fatal error
467:    actual-stdout> sccache: error: Failed to generate compile commands
467:    actual-stdout> sccache: caused by: Failed to generate compile commands
467:    actual-stdout> sccache: caused by: Missing object file output

This is from running CMake's CXXModules tests:

export SCCACHE_SERVER_PORT=23448 # unique to avoid ambient re-usage
export CXX="$(which sccache) $(which clang++)"
ctest -R CXXModules

@mathstuf
Copy link
Contributor

You can ignore the tests failing about JSON argument list mismatches; it is not robust against mutli-argument compilers like here. If you do the "symlink trick" (ln -s $(which sccache) clang++ && export CXX="$PWD/clang++", that probably avoids the issue).

@TroyKomodo
Copy link
Contributor Author

Hmm, i think this is just me forgetting to change it somewhere. I will fix

@TroyKomodo
Copy link
Contributor Author

Okay so it seems i think we should always treat it as an object even if it really isnt...

@TroyKomodo

This comment was marked as outdated.

@sylvestre
Copy link
Collaborator

could you please add an integration test ? ie high level which integrates into cmake and uses C++ module.
see
tests/cmake/ tests/sccache_args.rs tests/sccache_cargo.rs tests/sccache_rustc.rs

@mathstuf
Copy link
Contributor

I tried the cmake tests and they ran fine for me but there were only 2...

There are dozens of test cases within each CTest-level test. These are the relevant ones, thanks!

Can you inspect the cache state (--show-stats) on an initial run and a subsequent run to see how well caching itself is working?

@TroyKomodo
Copy link
Contributor Author

could you please add an integration test ? ie high level which integrates into cmake and uses C++ module.

see

tests/cmake/ tests/sccache_args.rs tests/sccache_cargo.rs tests/sccache_rustc.rs

Yes!

@TroyKomodo

This comment was marked as outdated.

@TroyKomodo
Copy link
Contributor Author

@sylvestre
Copy link
Collaborator

are you sure ?

Run ${SCCACHE_PATH} --show-stats
Compile requests                      4
Compile requests executed             4
Cache hits                            2
Cache hits (C/C++)                    2
Cache misses                          2
Cache misses (C/C++)                  2
Cache hits rate                   50.00 %
Cache hits rate (C/C++)           50.00 %

@TroyKomodo
Copy link
Contributor Author

That seems right no? the first time we invoke the compiler not cached yet then we invoke again and its cached.

There are only 2 files to compile, so i expected it to be 4 requests 2 hit 2 misses.

@sylvestre
Copy link
Collaborator

oh yeah, my bad, i thought we were doing a reset

@TroyKomodo
Copy link
Contributor Author

oh yeah, my bad, i thought we were doing a reset

I just copied the other cmake test example, but I can reset the stats if that makes it easier to follow

@mathstuf
Copy link
Contributor

Gah, module compilation in the tests doesn't really happen without a setting. For clang, this should work:

-DCMake_TEST_MODULE_COMPILATION=named,compile_commands,collation,partitions,internal_partitions,export_bmi,install_bmi,shared,bmionly,build_database,import_std23

}

if !module_only_flag {
if let Some(module_output_path) = module_output_path {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

could you please move this into a function?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I can, however this function is ~550 lines, I think it would be better if in a followup PR we break up this function rather than doing a half job here by breaking up just this one conditional.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

i still want this to be a new function, sorry

@TroyKomodo
Copy link
Contributor Author

Gah, module compilation in the tests doesn't really happen without a setting. For clang, this should work:

-DCMake_TEST_MODULE_COMPILATION=named,compile_commands,collation,partitions,internal_partitions,export_bmi,install_bmi,shared,bmionly,build_database,import_std23

Ah, I see. Okay I will try again i think i need to setup my env to have support for import std.

@TroyKomodo
Copy link
Contributor Author

Gah, module compilation in the tests doesn't really happen without a setting. For clang, this should work:
-DCMake_TEST_MODULE_COMPILATION=named,compile_commands,collation,partitions,internal_partitions,export_bmi,install_bmi,shared,bmionly,build_database,import_std23

Ah, I see. Okay I will try again i think i need to setup my env to have support for import std.

let
  pkgs = import <nixpkgs> {
    overlays = [
      (final: prev: {
        llvmPackages_21 = prev.llvmPackages_21 // {
          libcxx = prev.llvmPackages_21.libcxx.overrideAttrs (old: {
            postInstall = (old.postInstall or "") + ''
              substituteInPlace $out/lib/libc++.modules.json \
                --replace-fail '"../share' "\"$out/share"
            '';
          });
        };
      })
    ];
  };

  llvm = pkgs.llvmPackages_21;

  wrappedClangScanDeps = pkgs.writeShellScriptBin "clang-scan-deps" ''
    newargs=()
    inject_next=false
    for arg in "$@"; do
      if $inject_next; then
        newargs+=("$arg")
        newargs+=("-isystem" "${llvm.libcxx.dev}/include/c++/v1")
        inject_next=false
      elif [[ "$arg" == "--" ]]; then
        newargs+=("$arg")
        inject_next=true
      else
        newargs+=("$arg")
      fi
    done
    exec ${llvm.clang-tools}/bin/clang-scan-deps "''${newargs[@]}"
  '';

  sccache = "/home/troy/github/troy/sccache/result/bin/sccache";

  clangWrapper = pkgs.writeShellScriptBin "clang" ''
    exec ${sccache} "${llvm.libcxxClang}/bin/clang" "''${@}"
  '';
  clangxxWrapper = pkgs.writeShellScriptBin "clang++" ''
    exec ${sccache} "${llvm.libcxxClang}/bin/clang++" "''${@}"
  '';
in
pkgs.mkShell {
  hardeningDisable = [ "all" ];

  buildInputs = [
    pkgs.cmake
    pkgs.mold
    pkgs.ninja
    pkgs.openssl
    llvm.libcxxClang
    llvm.clang-tools
    wrappedClangScanDeps
  ];

  shellHook = ''
    export CC="${clangWrapper}/bin/clang"
    export CXX="${clangxxWrapper}/bin/clang++"
    export LD_LIBRARY_PATH="${llvm.libcxx.out}/lib:${pkgs.openssl.out}/lib:''${LD_LIBRARY_PATH:-}"
    export NIX_CFLAGS_COMPILE="-B${llvm.libcxx.out}/lib -isystem ${llvm.libcxx.dev}/include/c++/v1"
    export PATH="${wrappedClangScanDeps}/bin:''${PATH:-}"
  '';
}

A huge pain...

troy@troy-thinkpad ~/g/t/CMake (master) [SIGINT]> ctest -R CXXModules -j$(nproc) --test-dir=build   nix-shell 
Test project /home/troy/github/troy/CMake/build
Guessing configuration NoConfig
    Start 536: RunCMake.CXXModulesCompile
    Start 535: RunCMake.CXXModules
1/2 Test #535: RunCMake.CXXModules ..............   Passed   18.90 sec
2/2 Test #536: RunCMake.CXXModulesCompile .......   Passed  116.80 sec

100% tests passed, 0 tests failed out of 2

Label Time Summary:
CXXModules    = 135.70 sec*proc (2 tests)

Total Test time (real) = 116.81 sec
troy@troy-thinkpad ~/g/t/CMake (master)> ctest -R CXXModules -j$(nproc) --test-dir=build            nix-shell 
Test project /home/troy/github/troy/CMake/build
Guessing configuration NoConfig
    Start 536: RunCMake.CXXModulesCompile
    Start 535: RunCMake.CXXModules
1/2 Test #535: RunCMake.CXXModules ..............   Passed   19.94 sec
2/2 Test #536: RunCMake.CXXModulesCompile .......   Passed  104.62 sec

100% tests passed, 0 tests failed out of 2

Label Time Summary:
CXXModules    = 124.55 sec*proc (2 tests)

Total Test time (real) = 104.63 sec
troy@troy-thinkpad ~/g/t/CMake (master)>                                                            nix-shell 
Every 2.0s: /home/troy/github/troy/sccache/result/bin/sccache -s       troy-thinkpad: Mon Dec 22 12:47:11 2025

Compile requests                   1976
Compile requests executed           830
Cache hits                          756
Cache hits (C/C++)                  756
Cache misses                         72
Cache misses (C/C++)                 72
Cache hits rate                   91.30 %
Cache hits rate (C/C++)           91.30 %
Cache timeouts                        0
Cache read errors                     0
Forced recaches                       0
Cache write errors                    0
Cache errors                          0
Compilations                         72
Compilation failures                  2
Non-cacheable compilations            0
Non-cacheable calls                 244
Non-compilation calls               902
Unsupported compiler calls            0
Average cache write               0.001 s
Average compiler                  0.217 s
Average cache read hit            0.000 s
Failed distributed compilations       0

Non-cacheable reasons:
-E                                  244

Cache location                  Local disk: "/home/troy/github/troy/CMake/build/Tests/CMakeFiles/TestHome/.cac
he/sccache"
Use direct/preprocessor mode?   yes
Version (client)                0.12.0
Cache size                          155 MiB
Max cache size                       10 GiB



@mathstuf
Copy link
Contributor

If you could point me to where the code for the distributed cc lives so I can see how it works that would be great!

Look at src/dist/mod.rs. BuildResult would need built with the outputs. Inputs seem to be built into a tar file shipped via InputsReader. Chasing this down, it looks like src/compiler/c.rs's CInputsPackager needs to learn to ship module files and munge module mapper arguments/files as necessary.

@avikivity
Copy link
Contributor

Isn't this needed both for caching and for distributed compilation?

The inputs are needed for caching to check timestamps/sizes/hashes and for distribution to ship over the files. Note hashing the preprocessed output is insufficient.

The outputs are needed for caching to stash in the cache and for distribution so ship back.

For distribution, we might need to massage the command line, adding -fmodule-file for each input and output module to redirect them to out place in the filesystem.

@AJIOB
Copy link
Contributor

AJIOB commented Jan 17, 2026

Isn't this needed both for caching and for distributed compilation?

The inputs are needed for caching to check timestamps/sizes/hashes and for distribution to ship over the files. Note hashing the preprocessed output is insufficient.

The outputs are needed for caching to stash in the cache and for distribution so ship back.

For distribution, we might need to massage the command line, adding -fmodule-file for each input and output module to redirect them to out place in the filesystem.

IMO, we can think about distributed compiling support in additional PR

avikivity added a commit to avikivity/seastar that referenced this pull request Jan 18, 2026
sccache [1] combines ccache and distcc, and promises [2] to support
modules early. Support it via a more general --compiler-cache option,
replacing the current --ccache.

[1] https://github.com/mozilla/sccache
[2] mozilla/sccache#2516
avikivity added a commit to avikivity/seastar that referenced this pull request Jan 18, 2026
sccache [1] combines ccache and distcc, and promises [2] to support
modules early. Support it via a more general --compiler-cache option,
replacing the current --ccache.

The test workflow is adjusted for the new option name.

[1] https://github.com/mozilla/sccache
[2] mozilla/sccache#2516
@mathstuf
Copy link
Contributor

Note that an initial investigation to at least drop "unsupported" errors and/or TODO comments in places would be useful for further work. It could also help avoid confusion by anyone trying modules and distributed builds in the interim.

@TroyKomodo
Copy link
Contributor Author

Note that an initial investigation to at least drop "unsupported" errors and/or TODO comments in places would be useful for further work. It could also help avoid confusion by anyone trying modules and distributed builds in the interim.

I am still traveling so I haven't had time to try out the distributed compiler. I wish there were better tests than to setup a docker container and try run it. I wasn't even aware the distributed compiler was a thing until someone pointed it out in this PRs comments. Since all the unit tests passed... If I don't do the full impl, I will at minimum add unsupported errors for modules when using distributed compiler mode.

Sorry I haven't had time yet I should be fully settled down by Wednesday.

@mathstuf
Copy link
Contributor

There's no rush on my account at least; take your time.

@AJIOB
Copy link
Contributor

AJIOB commented Feb 4, 2026

Hi @TroyKomodo,

Do you have any updates?

Looks like we have conflicts.

@codspeed-hq
Copy link

codspeed-hq bot commented Feb 15, 2026

Merging this PR will not alter performance

✅ 62 untouched benchmarks
⏩ 2 skipped benchmarks1


Comparing TroyKomodo:troy/cxx20-modules (10fc4f6) with main (81954de)

Open in CodSpeed

Footnotes

  1. 2 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports.

@TroyKomodo
Copy link
Contributor Author

Okay! So i finally have gotten the time to look through this PR again. I think, this automatically works with dist mode since we add the PCMs to the extra_hash_files which are sent to the dist server automatically?

Could someone please provide me with a way to test this or someone test this for me, using the integration test in the PR.

Integration tests fail because this branch is from a fork.

@sylvestre
Copy link
Collaborator

Integration tests fail because this branch is from a fork.

why do you think that ?

Copy link
Contributor

@mathstuf mathstuf left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Seems fine as a stepping stone towards actual support. Just one minor request.

msvc_flag!("openmp:experimental", PassThrough),
msvc_flag!("permissive", PassThrough),
msvc_flag!("permissive-", PassThrough),
msvc_take_arg!("reference", OsString, Separated, TooHard),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The MSVC flags are very similar to Clang's. -interface and -internalPartition just need hashed like any other.

Can issues be opened and referenced from the comments for GCC and MSVC support?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

sccache doesn't support c++module well

7 participants