perf: forget cached lookups per name, not per context (0158)

[rmorozov]

Requested by Ruslan · project thread

Before: a set, a loop body scope cleared after each pass, or a loop or macro slot bound or unbound started a new lookup epoch for the context. After that, every cached name lookup missed once, including lookups of names the change could not affect. On many_tags every root set did this. On chat_mistral 62 of the 71 misses per render came from these events.

After: those events drop only the cache entries of the names they touch. Other names stay cached. Rendered output does not change.

How:

• LookupCache groups its 128 entries into 32 buckets of 4, by name hash. NewSlot(hash) places a name expression in its name's bucket, so forgetting a name clears 4 keys. The keys sit in their own array, so this is two stores. The hit path does not change.
• Each scope records, in 32 bits in ScopeStack, the buckets of the names added to it. Clear and ExitScope forget those buckets without hashing names.
• A scope entered with names it did not record (a loop's kept body scope) starts a context epoch and forgets every bucket when it goes.
• BindSlot and UnbindSlots take the slot names' hashes.
• A static assert keeps InternalValueMap a node map, because the cache keeps pointers to scope values while other names are added.
• Second round, after the performance track asked for one: the loop and its lazy filter read the target's slot and hash once instead of after every item, and the first scope chunk no longer zero-fills bucket masks that Push sets anyway. This pays back the extra bucket clears per loop run.

Instruction counts after the second round (callgrind, Release, against 39dc889):

Better	Worse
many_tags -9.5%, macros -3.1%, chat_mistral -0.6%, for_range -0.4%, dict_ops -0.4%, config_file -0.3%, chat_qwen -0.2%, chat_llama -0.1%	inheritance +0.46%, substitute +0.41% (+9 instructions), mitsuhiko_table +0.24%, mitsuhiko_table_wide +0.22%, html_autoescape +0.14%

Allocations are unchanged. Every render case is now within the house bar of +0.5%.

Exceptions to the task's +0.3% bar:

• Render inheritance +0.46%: each nested context records and forgets buckets for the names it adds.
• Render substitute: +9 instructions.
• Load large_static +0.7% and substitute +0.6% (CI, first round): NewSlot computes the name's bucket and uses that bucket's counter. This costs a few instructions once per distinct name in a template, so it is a real cost and not layout.

Nested contexts still take a context epoch. A follow-up note on this is in docs/tasks/0158.

Roles: the verifier ran 1 round on the first commit. It found 5 clang-tidy hits and no correctness issue: the Debug cache cross-check passed on about 10k generated and hand-written templates, and the mutation checks were caught. ASan+UBSan, C++20 and preflight pass. The second-round commit passed 4747 tests in Debug (with the cache cross-check) and preflight. Red pushes so far: 0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AmFfzfwMghRTWPifzgmANC

---

Generated by Claude Code