JLArrays 0.4 takes ~4.5 s to load (was 0.5 s), spent activating its methods

[IanButterworth]

Claude:

---

using JLArrays got about 9x slower with JLArrays v0.4.0 / GPUArrays v12.0.0.

Julia 1.13.1, macOS aarch64, fresh process, after precompiling:

	JLArrays 0.3.3 + GPUArrays 11.5.16	JLArrays 0.4.0 + GPUArrays 12.0.0
using JLArrays	0.47 to 0.54 s	4.52 to 4.64 s
allocated	52 MiB	1850 MiB
JLArrays precompile	0.8 s	4.8 s
invalidated method instances	198	1842

• Where: loading SIMD, AcceleratedKernels and GPUArrays first takes 0.07, 0.23 and 0.34 s. The rest is JLArrays' own image: @time_imports gives it 4.4 s with no compilation time.
• Profile: almost all of that is jl_activate_methods → jl_method_table_activate → type_morespecific_ (subtyping), i.e. inserting JLArrays' methods into the method tables. The new methods in sorting.jl, accumulate.jl and findall.jl (on AnyJLArray) are the likely cause.
• Invalidations: most of the new ones come from SIMD.jl, newly pulled in via AcceleratedKernels (!=, >=, > in simdvec.jl, about 1400 instances).

Repro:

using Pkg; Pkg.activate(temp=true); Pkg.add(name="JLArrays", version="0.4")
# then in a new process with that environment
@time using JLArrays

It shows up in Julia's TTFX CI: JLArrays/Matrix-Multiplication-On-Gpu-Like-Arrays load went from 0.9 s to 8.4 s on macOS when the job picked up these releases (perf.julialang.org).

Written with Claude Code (Claude Opus 5.5).

[IanButterworth]

Claude:

---

Same on Julia master, and JuliaLang/julia#62658 (fewer recorded backedges) cuts the invalidations but not the load time:

	using JLArrays (3 runs each)	allocated	invalidated method instances
nightly 1.14.0-DEV.3529 (beb5136a298)	5.65, 5.66, 5.81 s	1745 MiB	2338
JuliaLang/julia#62658 (1.14.0-DEV.3517)	5.61, 5.73, 5.79 s	1744 MiB	670

So the cost is in adding JLArrays' methods to the method tables, not in invalidating code.

Written with Claude Code (Claude Opus 5.5).