[aleksisch]

Module cache format change: version 234 - existing caches are discarded once and rewritten.

Why. A warm load rebuilt one type node per holder, then the module's finalize walked the whole module to merge them back. On a warm dasLLAMA load that cost about a quarter of the run.

What changes.

• A type reference in a record names the node it points at; the reader returns the same pointer, so a served module comes back with exactly the sharing the cold compile left.
• The content-keyed type table, its recursive hash/compare and the deep copy per reference are gone.
• A module read from the cache carries the new fromModuleCache flag; ast_verify skips such a module, as its documentation already said.
• A served module's finalize skips the GC collect - it has nothing left to merge or free.
• tests/math/test_color_mapping.das compares tone_reinhard(1.0) with a tolerance: under fastmath it is 0.49999997, not 0.5 (the test came with Add linear PBR output and shared tone-mapping operators #4351).

Observable behavior.

• warm tutorials/dasLLAMA/02_chat.das: 1.97-2.06 s -> 1.45-1.55 s (25-26% faster); cache 66.3 MB -> 64.5 MB
• warm fio + match script: 890.6M -> 752.3M instructions (-15.5%)
• cold compile: unchanged within noise
• warm run requiring daslib/ast_verify: 11,304 false AST verify: lines would appear with shared types -> 0 (new test)

Where to look. AstSerializer::operator << ( TypeDeclPtr & ) in module_builtin_ast_serialize.cpp, and the skip in parseDaScriptEx (ast_parse.cpp).

Validation, claims, ledger

Validation

• Metrics were measured against dd236c1fee (the base before the last rebase), interleaved master/branch runs on a shared 32-core box: dasLLAMA warm 15 runs each (median 1.966 s -> 1.453 s), cold 3 + 6 runs each; instructions from callgrind on a warm require daslib/fio + require daslib/match script.
• Garbage check, local only: every test program compiled cold then warm with a diagnostic that ran the finalize collect on cache hits - 39,272 served module reads, 0 nodes left for the collect.
• Full suite and the --ser/--deser sweep passed locally on the pre-rebase tip; the post-rebase tip ran the module-cache tests and the full suite.
• Not run: the per-PR audit agents, make-pr/full preflight, and the full AOT suite - CI and #nightly cover them.

Claims - stated, not tested

• A served module's decode allocates only attached nodes. Verified by the garbage check above, not by a test in the tree; a new source of garbage on a cache hit would show as memory held until the module unloads, not as a failure.

Not done

• Macro-module simulation on a cache hit (fusion, debug info) is still the largest warm-load cost; untouched here.

#nightly

[aleksisch]

@borisbat this is fine. Earlier we made type-merging in serializing. Now we do it in module after the parsing. So we don't need to merge types again.

Although, earlier we merged types only, and now we merge types + location. It has benefits: we do not allocate anything after deserialization.