perf: one access seam for InternalValue outside wave 2's files (0140 P0a)

[rmorozov]

Requested by Ruslan · project thread

Before: code all over src/ read InternalValue through how it is stored: GetIf<ListAdapter>, GetData(), std::get_if on the variant, switches on its index. P2 (the 32 B layout in docs/design/0140-internal-value-plan.md) would have had to edit every one of those sites.

After: outside internal_value.* and wave 2's files, values are read through one seam that P2 can keep. Behaviour is unchanged except for one parity fix: reverse of a dict.items() key now gives a string, as in Python. Before, it gave a list of characters.

The seam, all in src/internal_value.h:

• ValueKind / InternalValue::Kind() says what a value is.
• AsList / AsMap return non-owning ListRef / MapRef views. Overloads taking a temporary are deleted.
  • Visitors now receive the views, not the adapters. They get them as const lvalues, so that a const ListRef& overload beats a forwarding-reference template.
  • ToAdapter() makes the owning copy where a site keeps the list.
• AsStringView<CharT> and TakeString read and take strings. IsStringValue moved here from markup.h and value_methods.cpp.
• IMapAccessor::SetValue is const, so a value can hold its accessor as a const object.

Left for P0b, once wave 2 releases its files: the callers of GetIf<ListAdapter>/GetIf<MapAdapter>, GetStringView and NarrowStringView in expression_evaluator.* and statements.*, then GetData() itself.

How. Each site was converted mechanically, then recounted under Callgrind. The first count was +0.2..+6.6% Render, so three hot paths were reworked:

• IsNumber keeps per-alternative tests. A Kind() comparison cost +6.6% on expressions.
• OutStream::WriteValueTo tests int, string, host value and bool in turn. Strings meet at one WriteString, which GCC then keeps inline.
• ConvertToList returns a list's copy in place. Before, it copied twice, through std::optional.

Instruction counts (bench/count.py, Release, vs ec9e7b4; allocations and memory identical on every case):

case	change
Render/mitsuhiko_table	-1.56%
Render/mitsuhiko_table_wide	-1.76%
Render/filters	-0.88%
Render/for_range	-0.61%
Render/expressions	-0.46%
Render/large_static	+0.55%
Render/chat_qwen	+0.46%
Render/chat_llama	+0.36%
every other Render and Load case	within ±0.3%

Render/large_static's +0.55% is +150 instructions per render, about 2 per string write in WriteValueTo. It is an interim cost of the seam. P2, the 32 B layout, recovers it, together with #445's Load/plain_text +0.75%; the 0140 task file records both.

Parity. New corpus case filters.reverse_dict_item_key pins the reverse fix. filters.dictsort_pair_subscript pins a gap found on the way: p[0] of a dictsort pair is undefined. It is filed as docs/tasks/0159 and listed in divergences/filters.txt.

Process. Two verifier rounds ran.

• Round 1 found two clang-tidy errors (use-after-move in ParseAdjacentStrings, a misplaced NOLINT), both fixed.
• Round 2 found an unused include, removed.
• Both rounds also ran Debug, ASan+UBSan, a Clang 18 C++20 build and a 300-template differential against master and Python. The reverse fix was the only difference.

Red pushes so far: 0. scripts/preflight.sh passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_014gsczz1ZU7Lu24hymWrNJv