Introduce AST and refactor all parsers into an unified tokenizer (the name is not parser for historical reasons), replace instanceofs with TokenKind and type guards. Improved performance by 8%~117%, following are details for difference cases:

output

literal x 108,864 ops/sec ±0.76% (80 runs sampled) +66.79%
truncate x 118,342 ops/sec ±0.61% (81 runs sampled) +73.99%
date x 108,593 ops/sec ±1.17% (80 runs sampled) +74.70%
escape x 153,442 ops/sec ±1.00% (77 runs sampled) +85.92%
default x 127,896 ops/sec ±1.62% (73 runs sampled) +65.83%

tag

if x 110,173 ops/sec ±1.44% (81 runs sampled) +41.53%
unless x 98,669 ops/sec ±0.95% (75 runs sampled) +45.01%
for x 51,197 ops/sec ±0.94% (84 runs sampled) +117.51%
switch x 60,840 ops/sec ±0.64% (81 runs sampled) +75.66%
assign x 140,554 ops/sec ±1.16% (77 runs sampled) +76.76%
capture x 106,354 ops/sec ±1.19% (77 runs sampled) +36.83%
increment x 165,727 ops/sec ±1.15% (77 runs sampled) +46.86%
decrement x 179,814 ops/sec ±1.04% (75 runs sampled) +52.63%
tablerow x 57,498 ops/sec ±0.73% (77 runs sampled) +86.36%

demo

demo x 20,497 ops/sec ±4.61% (75 runs sampled) +100.30%

layout

cache=false x 4,670 ops/sec ±2.72% (76 runs sampled) +8.45%
cache=true x 11,940 ops/sec ±0.77% (80 runs sampled) +26.68%

memory

3.000 kB lorem-html before GC x 1.091 kB/tpl (1024 instances sampled) -80.93%
3.000 kB lorem-html after GC x -0.352 kB/tpl (1024 instances sampled)
2.115 kB todolist before GC x 16.636 kB/tpl (1024 instances sampled) +4.85%
2.115 kB todolist after GC x 14.225 kB/tpl (1024 instances sampled)