perf(types): canonicalize acyclic type graphs bottom-up instead of by fixed point

[ggreif]

Fixes the long-standing complaint that test/bench/typtbl.mo takes over two
minutes to run.

Diagnosis

The type checker is not implicated: moc --check on that file takes 0.031s.
All the time is codegen, and it is cubic in the nesting depth:

depth	compile	implied exponent
256	1.32s
512	10.01s	n^2.92
768	36.83s	n^3.21
1024	94.53s	n^3.28

ExpGraph.combine finds the coarsest bisimulation by partition refinement,
restarting from the one-class partition and re-splitting every node on each
pass. A pass distinguishes one more level of structure, so a chain of n
distinct nodes needs n passes. typ_hash canonicalizes on every call, and
codegen names each generated @deserialize_go / @buffer_size function by the
hash of its type — so a type nested n deep is hashed once per level.

Fix

On an acyclic graph bisimilarity is just structural equality, which can be read
off bottom-up in a single topological pass. Cyclic types keep the fixed point,
where it is genuinely needed: type A = ?[A] and type B = [?B] denote the
same regular tree without being structurally equal.

depth	before	after
256	1.32s	0.76s
512	10.01s	0.55s
768	36.83s	0.97s
1024	94.53s	1.63s

test/bench/typtbl.mo as a whole: 2m16s → 2.5s.

Verification

• Output is unchanged. wasm2wat of typtbl.mo before and after is 390,978
  lines each, differing only in the embedded version string — not one
  instruction.
• dune runtest mo_types passes, including the monolist / polylist cases
  that exercise the cyclic path.
• No pessimisation. test/run (312 programs, cleaned before each run, quiet
  machine): 163.50s with the patch vs 164.49s without (real), 130.02s vs
  131.42s (user) — a wash, within noise.
• nix build .#common-tests .#gc-tests passes locally.

🤖 Generated with Claude Code

[github-actions]

Comparing from 01aee06 to f5836dc:
The produced WebAssembly code seems to be completely unchanged.
In terms of gas, no changes are observed in 5 tests.
In terms of size, no changes are observed in 5 tests.

[github-actions]

Cursor AI review

👍 APPROVE — looks safe to merge

Category	Assessment	Details
Summary	✅	Speeds up ExpGraph.combine by reading coarsest classes bottom-up on DAGs (topo_order + combine_acyclic) and keeping the existing fixed-point algorithm only when topo_order finds a cycle.
Code Quality	✅	Reuses rename, start_counting, and IM; combine_cyclic is the previous combine body unchanged. No new public API or speculative helpers.
Consistency	✅	Matches existing expGraph.ml style (ref cells, IM, polymorphic compare keys) and leaves expGraph.mli unchanged.
Correctness	✅	Traced canonicalize → typ_hash → compile_enhanced.ml / compile_classical.ml alias tags and @buffer_size/@deserialize_go names, plus ir_passes/show.ml and eq.ml. On a DAG, bottom-up (label, child classes) is the coarsest bisimulation; the 0↔root swap only permutes class ids so renumber still starts at node 0. Cyclic graphs still take the old path (typ_hash monolist/polylist).
Tests	✅	No .mo/.ok pairs in the diff, which matches a behavior-preserving hash. Cyclic fallback still covered by typ_hash.ml monolist/polylist; common acyclic types are the test/run path through typ_hash.
Changelog	✅	Internal compile-time optimization only (no language/prelude/CLI surface). Changelog.md is untouched, so no frozen-section edit.

Verdict

Decision: APPROVE
Risk: Low
Reason: Single-file DAG specialization of type-graph combine that computes the same partition as the existing fixed point, then relies on unchanged renumber; cyclic types keep the old algorithm. Hash-sensitive codegen/stable-alias uses stay observationally the same.

---

Generated for commit f5836dc