feat: have cmake backend by JeffIrwin · Pull Request #1251 · fortran-lang/fpm

JeffIrwin · 2026-02-21T19:02:12Z

Summary

Adds fpm generate --cmake command to export fpm projects as CMake build files, enabling CMake-based builds for better IDE integration and deployment to environments without fpm.

Features

Generated CMakeLists.txt files for library, executable, and test targets
Dependency handling with automatic subdirectory inclusion
Full language support for Fortran (free/fixed form), C, and C++
Preprocessing with custom suffixes and per-dependency macros
Metapackage integration (MPI, OpenMP, BLAS, HDF5, NetCDF, stdlib, etc.)
Cross-platform tested on Linux, macOS, and Windows CI
Staleness detection warns when generated CMake files are out of sync
- Staleness warning has some false-positives
- Any change that causes fpm build to re-build will also cause a staleness warning
- Unfortunetaly, I think the only robust way to check cmake without false-positives is to entirely regenerate cmake in memory and compare it to the existing file. I fear that this is too expensive to do on every run of fpm build, fpm test, and fpm run

Test Coverage

Unit tests for core generation logic
Integration tests covering ~70 example packages
Self-hosting test: fpm builds itself using generated CMake

Performance

Optimized with hash tables for O(n) dependency resolution and source lookups.

Fixes #1240

Add a new subcommand that generates a CMakeLists.txt file from an fpm project. This allows building fpm projects with CMake for better IDE integration or deployment to environments where fpm is not available. The generated CMakeLists.txt includes: - Library target (if src/ sources exist) - Executable targets (from app/) - Test targets with enable_testing() and add_test() Limitations documented in help text: - fpm dependencies not auto-handled in CMake - Metapackages/registry deps not translated - Complex preprocessing may need adjustment Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Previously, `get_sources_for_exe` only included sources where `exe_name` matched the executable name. However, `exe_name` is only set for source files containing a `program` statement, not for helper modules. This caused CMake builds to fail for test executables like `fpm-test` that depend on multiple source files (e.g., main.f90 + testsuite.f90 + test_*.f90 modules). The fix changes the logic to: 1. Find the main program file by matching exe_name 2. Extract its directory path 3. Include ALL source files from that directory with the same scope This matches fpm's native build behavior where all sources in an executable's source directory are compiled together. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

…iority example

This reverts commit dbc6628.

Previously only detected fpm.toml changes. Now detects source file additions, deletions, renames, and content changes by hashing source file paths and content digests already computed by build_model. Changes: - Add compute_project_hash() that combines manifest + source digests - Move staleness check after build_model to access source digests - Update CMake header from "Manifest hash:" to "Project hash:" - Maintain backward compatibility with old format (triggers regeneration) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

…with my own personal branch

Wrap add_subdirectory calls in if(NOT TARGET) guards to prevent duplicate target definitions when packages have mutual dependencies. This allows CMake builds to succeed from either direction in circular dependency scenarios while maintaining correct behavior for normal and diamond dependencies. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Tests were only initializing compiler and build_dir fields, leaving other allocatable character fields unallocated. This caused runtime errors on macOS when build_model() passed these unallocated fields to new_compiler() as character(len=*) arguments. Fortran standard requires allocatable actual arguments to be allocated when passed to non-allocatable dummy arguments. macOS gfortran has stricter runtime checks and caught this violation. Added init_test_settings() helper that properly initializes all fields following the production build_settings() pattern. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

…ndencies

On Windows CI/CD, CMake was defaulting to Visual Studio generators which cannot find the GCC/gfortran compiler. Detect Windows/MSYS2 environments and explicitly use the MinGW Makefiles generator for Fortran support. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

When fpm generate cmake runs on Windows, it writes source file paths with Windows-style backslashes (e.g., .\.\src\lib.f) to CMakeLists.txt files. CMake interprets these backslashes as escape sequences, causing syntax errors like "Invalid character escape '\s'". This fix modifies the clean_path() function in cmake.f90 to call unix_path() before cleaning leading ./ prefixes. The unix_path() function converts all backslashes to forward slashes, which CMake accepts on all platforms including Windows. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Extends commit 238c65d to cover include directories, metapackage paths, and dependency subdirectory paths written to CMakeLists.txt files. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

High-priority fixes: - Propagate I/O errors from write_lines_to_file instead of silent failure - Check fpm generate exit code in test_cmake.sh before proceeding to cmake - Fix integer overflow in hash table using iand() instead of abs() - Optimize staleness check to read only first 5 lines for non-fpm CMake files Medium-priority fixes: - Use preprocessor config name as serialization key instead of positional index - Propagate errors from generate_dependency_cmake Low-priority fixes: - Replace O(n*m) source lookup with O(n) hash map in compute_project_hash Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Extract duplicated compiler-to-cpp-flag mapping into get_cpp_flag(), remove dead code (unused is_header_only, temp_dirs, FPM_SCOPE_EXAMPLE), fix portability bug using is_dir() instead of inquire(file=) for dirs, replace magic number with named project_hash_marker constant, handle C/C++ header files in language detection, detect dependency languages from sources instead of hardcoding Fortran, use original test name in add_test NAME, and fix test_cmake.sh cleanup of dependency files. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

GNU ld processes arguments left-to-right, so library flags (e.g. MPI's -lmpi_usempif08) must come after the object files that reference them. Previously, model%link_flags was prepended before the objects, causing "undefined reference" errors in metapackage_mpi builds. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

# Conflicts: # src/fpm/manifest/dependency.f90

Replaces O(n²) insertion sort + chain-hashing with O(n) XOR-based hash combination. This eliminates the sort_strings subroutine and hash table allocations while maintaining deterministic hashing. Uses XOR to combine per-file hashes (path + digest), which is safe since file paths are unique. Hash values will change for existing projects (one-time CMake regeneration cost). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

certik · 2026-02-26T01:05:07Z

Thanks for working on this, we need this. Is this PR working? Let us know when this is ready for review.

JeffIrwin · 2026-02-26T01:51:57Z

Hi @certik, yes, this PR is working. Out of 72 example packages in the fpm repo, 71 build and run correctly with generated cmake. Only example_packages/features_with_dependency fails, but it does not work with fpm build either.

Perhaps we should discuss one concern before starting the review. As mentioned in the description, the cmake staleness detection has false positives. There are at least three options:

Revert the staleness detection and save it for a future PR
Accept some false positives in the current implementation
Take the more expensive but robust approach of entirely regenerating CMakeLists.txt in memory and compare against the existing file on every run of fpm build, fpm run, and fpm test

The current implementation uses fpm's digest of every source file to detect cmake staleness. Of course, most small changes to a single source file will not actually require regenerating CMakeLists.txt: https://github.com/JeffIrwin/fpm/blob/5f2eec967e93cdc71fb33eeebd2a83d49dc9c6b6/src/fpm_cmake_check.f90#L16-L18

JeffIrwin and others added 30 commits January 19, 2026 11:24

fix examples like hello_complex with multiple app executables

547d4b8

fix dependencies

d92c54b

fix help test for new generate command

e8b42b5

fix examples like hello_complex_2 with multiple exes and shared modules

679b62b

use the correct dependency path, important for local dependencies

34d28d2

fix examples like hello_fpm with dependencies but no lib of its own

d082a26

add cmake test script

3229b5d

detect c header files

44cf808

update cmake test script with list of failed cases

5e1bc43

fix c_includes example

0b153c5

fix c_main example

9e3eb97

detect file extensions consistent with the rest of fpm

2feddb9

fix c_main_preprocess example

1865a6e

fix circular example

49b958c

use fpm's existing target pruning infrastructure to fix dependency_pr…

2058b2b

…iority example

fix regression in c_includes caused by last commit

78078f9

fix features_with_dependency example

dbc6628

Revert "fix features_with_dependency example"

43399f3

This reverts commit dbc6628.

features_with_dependency is actually broken

247e6b7

fix fixed-form example

4c621f5

fix a few metapackage examples

f323816

fix mpi metapackage example

75d4884

fix linker warning about missing '--end-group' flag

4550fdd

fix c_header_only regression

50e6686

hdf5 example doesn't work yet, not sure why i thought it did

f6dfdfe

a few more examples have been fixed already by other changes

6abecc7

fix free-form example

783aaa6

fix link_executable example

3c85db0

JeffIrwin and others added 23 commits February 12, 2026 20:04

don't check for 'Manifest hash'; no need for backwards compatibility …

a94c252

…with my own personal branch

bump for ci

6149be1

run cmake tests in ci/cd

fa2b4e1

use the release binary for the cmake test

a9936ac

extend ci/cd block list with tests that work locally but require depe…

db4bc0f

…ndencies

use a stylistically consistent block list

b505bd9

fix comment in middle of bash line continuation

6336081

fix: normalize Windows backslashes in CMake include/dependency paths

af61d72

Extends commit 238c65d to cover include directories, metapackage paths, and dependency subdirectory paths written to CMakeLists.txt files. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

cleanup redundant './' file path strings

2af2801

Merge remote-tracking branch 'origin/main' into generate_cmake

9448830

Merge remote-tracking branch 'origin/main' into generate_cmake

c7ae421

# Conflicts: # src/fpm/manifest/dependency.f90

test fpm building itself using its own generated cmake

c04d756

remove my own personal CLAUDE.md file

87f71dd

remove unused is_in_list() function

dc30aec

JeffIrwin marked this pull request as ready for review February 21, 2026 22:11

JeffIrwin added 3 commits February 21, 2026 19:34

dry up duplicate summary logging code in cmake test script

cb1839d

remove unused variables

955805d

remove more unused variables

5f2eec9

JeffIrwin mentioned this pull request Feb 26, 2026

Have CMake and Make backends #69

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: have cmake backend#1251

feat: have cmake backend#1251
JeffIrwin wants to merge 95 commits intofortran-lang:mainfrom
JeffIrwin:generate_cmake

JeffIrwin commented Feb 21, 2026 •

edited

Loading

Uh oh!

certik commented Feb 26, 2026

Uh oh!

JeffIrwin commented Feb 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

JeffIrwin commented Feb 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Features

Test Coverage

Performance

Uh oh!

certik commented Feb 26, 2026

Uh oh!

JeffIrwin commented Feb 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

JeffIrwin commented Feb 21, 2026 •

edited

Loading