[pbrubeck]

Summary

Every loop generated by pyop3 included <petsc.h>, which took most of the time needed to compile a small kernel. Most loops use no PETSc: in the reynolds_robust_navier_stokes_hdiv demo, 78 of the 122 compiled loops have neither a matrix argument nor a local kernel. These are mostly the layout, map, sparsity and assignment loops.

The header is now only included for loops that need it:

 LoopyCodegenContext.finalize_kernel
   preambles = [stdio.h]
-  preambles += [petsc.h]
+  if subkernels or any argument is a Mat (OpaqueType)
+    preambles += [petsc.h]          # local kernels may use PetscScalar etc.

 with_attach_debugger(kernel)
+  preambles += [petscsys.h]         # like with_petsc_event and petsclog.h

Evidence

Compiling the generated loop idat_0[i] = i with the flags of pyop3 (mpicc, GCC 15):

	wall time
with <petsc.h>	0.14 s
without <petsc.h>	0.04 s

tests/firedrake/demos/test_demos_run.py -k reynolds, with fresh caches, run side by side:

• Before: 277 s cold, 234 s warm
  After: 268.5 s cold, 238 s warm

The warm difference is within the noise between runs.

Merge Danger

Door: two-way

Blast Radius: small

Only the preambles of generated loops change. A loop that uses a PETSc symbol without a matrix argument or a local kernel would now fail to compile. No such loop is generated in the demo. Cached kernels are keyed by their code, so they are recompiled once.

---

AI assistance: this PR was written with Claude Code (Claude Opus 5.5).

🤖 Generated with Claude Code