[ggorman]

_complex_includes writes the extra complex arithmetic definitions a C++ language needs (def-complex) to complex_arith.h in the JIT cache directory, with open(..., 'w'), on every lowering of an Operator that uses a complex dtype. The directory is shared by every process compiling as the same user, and 'w' truncates the file before writing it, so with pytest -n auto one worker can truncate the header while another's compiler is reading it. The compile then fails on the first complex expression, e.g. no match for 'operator*' (operand types are 'int' and 'const std::complex<float>'); in DevitoPRO's CUDA backend the same race shows as user-defined literal operator not found for 1.0_if.

Changes:

• The header is keyed on its content (complex_arith_<sha1>.h), written to a private temporary file and renamed into place, and skipped when it already exists. A reader therefore sees either no file or a complete one, and two languages with different definitions no longer overwrite each other.
• test_complex_arith_header checks that two Operators share one header whose content is exactly def-complex.

Verification on main d7f53a3:

• 32 processes each JIT-compiling 30 distinct complex-valued language='CXX' Operators from one cache directory: main fails 7 of 960 compiles (twice, 7 each) with the error above; this patch 0 of 960, with a single complex_arith_<sha1>.h in the directory and no leftover temporaries.
• tests/test_dtypes.py: 47 passed, 5 skipped.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 95.77465% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.15%. Comparing base (16e6861) to head (9a332fc).

Files with missing lines	Patch %	Lines
devito/tools/os_helper.py	85.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3071      +/-   ##
==========================================
+ Coverage   84.13%   84.15%   +0.02%     
==========================================
  Files         258      258              
  Lines       57205    57267      +62     
  Branches     4889     4893       +4     
==========================================
+ Hits        48127    48191      +64     
+ Misses       8250     8249       -1     
+ Partials      828      827       -1     

Flag	Coverage Δ
pytest-gpu-aomp-amdgpuX	68.40% <17.39%> (-0.03%)	⬇️
pytest-gpu-gcc-	78.96% <85.91%> (+<0.01%)	⬆️
pytest-gpu-icx-	78.88% <85.91%> (+<0.01%)	⬆️
pytest-gpu-nvc-nvidiaX	68.98% <69.56%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 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.

[FabioLuporini]

I don't think this is the right way of fixing it, as we would have to change every possible place in which we construct shared header files accordingly, and there are quite a few across OSS and PRO

instead, I think Compiler's save_header (IIRC its name) should be made race-free

[ggorman]

Fair point, thanks both. Moved the fix into Compiler.save_header: it now writes to a temp file in the JIT dir and renames it into place, so every header saved through it (OSS and PRO: the f16 arithmetic, TMA cache, CUDA graph and alignment headers) gets the same guarantee. _complex_includes just calls save_header now and the sha1 is gone; since def-complex differs between C++ and CUDA/HIP, the header is keyed on the language name instead, the same way <lang>_f16_arith.h is in PRO (openmp_complex_arith.h, openacc_complex_arith.h, cuda_complex_arith.h, ...). The test is now a concurrency test of save_header itself in test_arch.py (fails on main, passes here), plus a check that the language-specific header is included with the full definitions.

[mloubout]

Needs rebase then gtg

[FabioLuporini]

GTG