tsfc: cancel Jacobian products and reciprocal factors

[pbrubeck]

Description

Companion PRs (required):

• UFL: Cancel contractions of the Jacobian with its inverse FEniCS/ufl#496
• FIAT: gem: simplifications supporting Jacobian product cancellation fiat#255

Piola-mapped elements generate expressions of the form J * inv(J) that were not simplified, because JacobianInverse was expanded into individual matrix entries too early in form preprocessing. This PR enables the new UFL pipeline that preserves the Jacobian quantities as terminals, cancels IndexSum(J[a,k] * K[k,b] * factors, k) -> factors * delta_ab, eliminates the resulting Kronecker deltas, and cancels reciprocal determinant powers, before lowering geometry as usual:

• tsfc/ufl_utils.py: enable do_cancel_jacobian_products in compute_form_data, and apply the same simplifications in preprocess_expression (interpolation and point evaluation paths).
• tsfc/spectral.py: cancel reciprocal factors after replace_division.
• tsfc/driver.py: cancel reciprocal factors in dual evaluation, where the Piola pullback's detJ**2 (from UFL) meets the inverse basis transformation's (1/detJ)**2 (from FInAT); previously this product multiplied every term of the kernel.

Benchmark

Johnson-Mercier inner(u, v)*dx + inner(div(u), div(v))*dx, spectral mode, cold cache:

	baseline	this PR	change
2D flops	27020	18728	-31%
2D compile time	0.52 s	0.36 s	-31%
2D kernel size	47 kB	29 kB	-38%
3D flops	763527	444574	-42%
3D compile time	3.62 s	1.81 s	-51%
3D kernel size	323 kB	143 kB	-56%

On affine cells the div-div integrand now reaches TSFC as (1/detJ**2) J : rdiv(sigma), with all second-derivative geometry terms folded at the UFL level. Assembled JM matrices agree with the unoptimised pipeline to machine precision on perturbed simplices in 2D and 3D.

🤖 Generated with Claude Code