Load: halve the cost of loading tag-heavy templates (0103)

[rmorozov]

Requested by Ruslan · project thread

Before: loading Load/many_tags took 44.15M instructions and 34,249 allocations. Most of that was the parser looking up keywords again at every precedence level for every token, copying tokens and parse results level by level, and copying Settings for every statement tag.

After: the same load takes 21.99M instructions (2.01x cheaper) and 16,559 allocations (-52%). Every other Load/ case is 31-51% cheaper, and Render/ is unchanged. Rendered output and error messages are byte-identical to master, with one exception that now matches Python (below).

How:

• The lexer classifies each symbol once and stores the result in Token::keyword; the parsers read it instead of calling GetKeyword. FindKeyword indexes by first character.
• The lexertk generator and token vector are reused across tags (LexBuffers), and tokens are built in place.
• Each precedence level keeps one result variable so NRVO applies, ParseUnaryPlusMinus calls the postfix and filter parsers only when the next token can start them, and token copies became references.
• StatementsParser holds const Settings& instead of a copy; statement infos are moved, not copied.
• Smaller: MarkMacroSpecialNames returns at once outside macros, attribute subscripts no longer allocate a constant node, IsMethodName binary-searches, FindBlockEnd skips plain characters, MatchTagAt gates raw/meta on the first letter.

Deliberate behaviour change (parity fix): a quoted keyword such as 'in' or 'if' used to be treated as the keyword, because GetKeyword read the token's source text. It is now a string, as in Python Jinja2, so {{ 1 'in' l }} is a syntax error and {% for a in l, 'if' %} renders. New cases in test/parity/cases/errors.py pin it.

bench/count.py --baseline (master before → this branch)

Benchmark	Instructions	Baseline	Change	Allocations	Bytes
Load/dict_ops	80,324	148,691	-45.98%	88 (-43)	11,001 (-5,576)
Load/expressions	105,233	206,111	-48.94%	103 (-43)	12,151 (-10,344)
Load/filters	163,195	274,433	-40.53%	153 (-75)	18,670 (-14,248)
Load/for_filter_if	69,984	135,397	-48.31%	80 (-43)	9,100 (-4,808)
Load/for_loop_vars	88,595	180,600	-50.94%	92 (-65)	10,965 (-8,472)
Load/for_range	33,542	56,665	-40.81%	47 (-9)	5,183 (-552)
Load/inheritance	59,016	113,566	-48.03%	68 (-39)	7,586 (-3,912)
Load/large_static	608,896	886,894	-31.35%	244 (-196)	76,399 (-15,576)
Load/macros	122,848	229,302	-46.43%	127 (-59)	18,221 (-11,208)
Load/many_tags	21,994,335	44,153,332	-50.19%	16,559 (-17,690)	1,801,775 (-2,953,352)
Load/mitsuhiko_table	124,182	237,726	-47.76%	133 (-51)	19,504 (-5,736)
Load/plain_text	9,584	9,676	-0.95%	18	1,610
Load/strings	156,142	294,038	-46.90%	142 (-82)	16,560 (-20,808)
Load/substitute	19,669	29,857	-34.12%	33 (-4)	2,963 (-216)

Render/ cases are all within -3.6%..+0.0% (no render code changed; for_loop_vars -3.6% comes from cheaper attribute subscripts).

Tasks

• 0103 marked done with the findings. Its hypothesis (each precedence level wraps operands in make_shared) was wrong; the profile showed keyword lookups and copies.
• Filed 0109 (load costs, round 2): the FullExpressionEvaluator wrapper, filter creation, string constant copies, the lexer itself, template destruction, expected churn and malloc_consolidate (reserving buffers made large_static 26% slower, so it was left out). First filed as 0107, renumbered after Split StringConverter::Filter into one member per mode (0061) #370 took 0107 (0108 is taken by perf: resolve include/extends/import targets once per render, with a Jinja2-exact option (0105) #376).

Process

• Roles: main session explored and implemented; verifier 1 round (ASan+UBSan, fuzz replay over 2,544 inputs, C++20 build, clang-format, output/error diff vs base). It found the quoted-keyword change, now pinned by corpus cases. Notes it raised: IsEqual on attribute subscripts compares null expr pointers (no practical effect); possible MSVC C4996 on 4-argument std::equal; clang-tidy 22 not run locally (CI checks it).
• master merged in to resolve the docs/tasks/README.md conflict from the renumbering, and again (a774747) before reporting final to the merge steward; tests pass on both merges.
• Red pushes in CI: 1. MSVC C++17 Release failed with C3493: a non-static local constexpr (charsCount in FindKeyword) used inside a capture-less lambda, which GCC, Clang and MSVC C++20 accept. Fixed by making it static constexpr (f770c6f).

🤖 Generated with Claude Code

https://claude.ai/code/session_01SGHFyqdC8NgXXsL6PatsQP