Fix type instability with many array references on Julia 1.11+

[vchuravy]

determine_eltype emits a single promote_type call with one eltype(array) argument per entry of ls.includedactualarrays, and that has one entry per array reference, not per array. On Julia ≥ 1.11, promote_type with 36 or more arguments is no longer inferred (Base.afoldl has max_args = 34), so the inferred element type is Any. That makes pick_vector_width, MM, __vload, vadd_nsw etc. dynamic, so every SIMD operation in the loop goes through runtime dispatch and allocates.

julia> Core.Compiler.return_type(promote_type, NTuple{35, Type{Float64}})
Type{Float64}

julia> Core.Compiler.return_type(promote_type, NTuple{36, Type{Float64}})
Any        # Julia 1.11.9 and 1.12.6; inferred for all lengths on 1.10

This PR builds a chain of two-argument promote_type calls instead, which is always inferred. The result is the same, since promote_type folds pairwise anyway.

Impact

We hit this in Trixi.jl: a hand-written @turbo flux-differencing kernel for MHD (flux_differencing_kernel_turbo! with nonconservative terms in src/solvers/dgsem_structured/dg_3d_turbo.jl) has about 40 array references per loop. On Julia 1.12 (Ryzen 7 7840U):

	before	after
one element	~31 ms, 8.6 MB allocated	5.1 µs, 0 bytes
full right-hand side, 512 elements	~18 s	5.5 ms

The results agree with the non-@turbo reference implementation to round-off. The existing sum_way_too_unrolled test in test/manyarrayrefs.jl (272 references) is affected too, but it only checks the result.

Tests

I added a type-stability check to test/manyarrayrefs.jl: sum_way_too_unrolled with 6×7 = 42 references must not allocate. On Julia 1.12.6 it fails on master (68560 bytes allocated) and passes with this change.

Pkg.test() on Julia 1.12.6 has one failure, test/dot.jl:12 (choose_order gives unroll 4 instead of the expected 8 for 32 registers). It is pre-existing: the released v0.12.174 gives the same result on this machine (AVX-512, Zen 4), and the code changed here doesn't affect choose_order. All other test groups pass.

A complementary change would be to deduplicate ls.includedactualarrays per array (e.g. addsetv! instead of push! in reconstruct_loopset.jl, as split_loops.jl already does). That would shorten the generated expression, but on its own it would not fix loops that reference 36+ distinct arrays, so I kept this PR minimal.

This PR was prepared with substantial assistance from Claude Code (Anthropic's AI coding assistant). I reviewed the changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MEJYUviAUshj5fZt8LTe8A

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.92%. Comparing base (369cebc) to head (ed42ca9).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #573      +/-   ##
==========================================
+ Coverage   83.81%   86.92%   +3.11%     
==========================================
  Files          40       40              
  Lines        9574     9684     +110     
==========================================
+ Hits         8024     8418     +394     
+ Misses       1550     1266     -284     

☔ 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.

[ranocha]

Thanks a lot!