Cap islpy below 2026.2 on main

[pbrubeck]

islpy 2026.2 removed the implicit conversion of a BasicSet to a Set. loopy 2025.2 — the newest release, and what loopy>2024.1 resolves to — calls BasicSet.make_disjoint() in count_insn_runs, so every flop count in the test suite fails:

AttributeError: 'islpy._isl.BasicSet' object has no attribute 'make_disjoint'

loopy converts explicitly on main, but taking loopy from git is not an option yet: its _AccessCheckMapper.map_subscript asks for the dependencies of every shape axis without guarding None, and PyOP2 builds args with an unknown leading axis (dat0 has shape=(None, 2) in wrap_to_reference_coords), so firedrake-check dies with DependencyMapperWithReductionInames encountered invalid foreign object: None.

So cap islpy at the last version that still converts. With islpy 2026.1 and loopy 2025.2, tests/tsfc and tests/firedrake/regression/test_locate_cell.py are green locally.

Lift the cap once loopy makes a release containing the explicit to_set(). #5378 does the same for release.

---

AI was used in preparing this contribution (Claude Code).

🤖 Generated with Claude Code

[connorjward]

Can there be an equivalent PR into release that pins islpy?

[pbrubeck]

Can there be an equivalent PR into release that pins islpy?

Sure, I'll open it.

[pbrubeck]

@connorjward It seems that the right fix is to cap islpy on main too. Do we want two separate PRs?

[connorjward]

@connorjward It seems that the right fix is to cap islpy on main too. Do we want two separate PRs?

Why is it needed on main? I would prefer on main to follow upstream branches.

[francesco-ballarin]

While this might be a good short-term solution, can you work with upstream islpy/loopy to fix the issue?

I had to open a PR on islpy (which then generated a new release 2026.2.1) to fix compatibility with nanobind 3.0, and if islpy gets pinned it won't have that fix.
(I can patch my way out of it, sure, but only for the short term)

[pbrubeck]

Why is it needed on main? I would prefer on main to follow upstream branches.

We cannot pin islpy and loopy to main. There is a chain of breakages:

Break 1 — loopy/statistics.py:1454 (loopy 2025.2) does set = set.make_disjoint() where set may be a BasicSet; that only worked via islpy's auto-promotion to Set, removed in islpy 2026.2.

Break 2 — pyop2/codegen/builder.py: shape = (None, arity) at line 50 (map arguments) and shape = (None, *arg.dim) at lines 861/879 (dat arguments), with Argument((None, 2), IntType, name="layers") at line 759. Those flow into pyop2/codegen/rep2loopy.py:797, which builds loopy.GlobalArg(name, dtype=dtype, shape=shape, strides=loopy.auto). loopy main then walks that shape in _AccessCheckMapper.map_subscript at loopy/check.py:771 — shape_deps.update(get_dependencies(shape_axis)) — and pymbolic raises on the None axis. check.py:1629 already guards this with if shape_axis is not None, and loopy/kernel/array.py:590 handles None axes too, so 771 is the only unguarded site. Adding the guard there is necessary but not sufficient: with it, tests/tsfc passes (369) but test_locate_cell still fails, so at least one further incompatibility sits behind it.

[connorjward]

Why is it needed on main? I would prefer on main to follow upstream branches.

We cannot pin islpy and loopy to main. There is a chain of breakages:

Break 1 — loopy/statistics.py:1454 (loopy 2025.2) does set = set.make_disjoint() where set may be a BasicSet; that only worked via islpy's auto-promotion to Set, removed in islpy 2026.2.
Break 2 — pyop2/codegen/builder.py: shape = (None, arity) at line 50 (map arguments) and shape = (None, *arg.dim) at lines 861/879 (dat arguments), with Argument((None, 2), IntType, name="layers") at line 759. Those flow into pyop2/codegen/rep2loopy.py:797, which builds loopy.GlobalArg(name, dtype=dtype, shape=shape, strides=loopy.auto). loopy main then walks that shape in _AccessCheckMapper.map_subscript at loopy/check.py:771 — shape_deps.update(get_dependencies(shape_axis)) — and pymbolic raises on the None axis. check.py:1629 already guards this with if shape_axis is not None, and loopy/kernel/array.py:590 handles None axes too, so 771 is the only unguarded site. Adding the guard there is necessary but not sufficient: with it, tests/tsfc passes (369) but test_locate_cell still fails, so at least one further incompatibility sits behind it.

I don't understand break 2

[connorjward]

I think pinning islpy is a non-starter. We have to address this more fundamentally. If loopy main gets things to pass then we should pursue getting a new loopy release made. If it doesn't pass then we should fix whatever fails.

[connorjward]

While this might be a good short-term solution, can you work with upstream islpy/loopy to fix the issue?

I had to open a PR on islpy (which then generated a new release 2026.2.1) to fix compatibility with nanobind 3.0, and if islpy gets pinned it won't have that fix. (I can patch my way out of it, sure, but only for the short term)

I've opened a loopy issue inducer/loopy#1039. Contrary to what I said before I think we do have to pin islpy because I don't know the loopy fix and we need CI back up.

[connorjward]

Closing this one as #5378 has just been merged into release and will be carried over ASAP.