Problem
Without route caching, the route table is rebuilt on every request. Each $routes->get() (or post(), add(), and so on) runs RouteCollection::create(), which costs about 1.9 µs per route. That's about 8 times more than the worst-case cost of matching a route.
Two steps in create() do work that is usually wasted:
- The placeholder loop.
create() runs str_ireplace(':' . $tag, ...) once for every registered placeholder (6 by default), even when $from has no : and nothing can match. That costs about 0.50 µs per route.
helper('array'). It's called on every create(). RouteCollection no longer uses any array-helper function, but the call still loads the helper for the app as a side effect. Repeat calls cost about 0.24 µs per route.
Measurements
PHP 8.3 on macOS, OPcache on, production mode, warm. Config construction cost was turned off to approximate the state after #10615. The routes are a realistic mix: a third static, a third with one placeholder, a third with two, spread across groups.
| Routes |
Build route table |
Match last route |
404 |
| 30 |
78 µs |
15 µs |
11 µs |
| 100 |
248 µs |
32 µs |
27 µs |
| 300 |
608 µs |
85 µs |
79 µs |
| 1000 |
2,030 µs |
263 µs |
252 µs |
Here's the cost breakdown of one $routes->get() call:
| Step |
Cost |
esc(strip_tags($from)) |
0.49 µs |
Placeholder str_ireplace loop (6 placeholders) |
0.50 µs |
helper('array') |
0.24 µs |
| Everything else |
~0.6 µs |
| Total |
~1.87 µs |
Proposal
- In
create(), run the placeholder loop only when the route key contains :. If it has no :, no str_ireplace can match, and no replacement can add a :. So the output is identical. That includes custom placeholders whose patterns contain :, such as (?:...).
- Call
helper('array') only on the first create() of each collection, using a private flag. The helper still loads at the same point as today, so apps that quietly depend on it keep working. A short comment will explain that the call stays only for this side effect.
Estimated savings are about 0.74 µs per route, roughly 40% of the build cost:
- 100 routes: about 75 µs per request
- 1000 routes: about 0.75 ms per request
Compatibility
- No change to public or protected APIs. The only new state is private.
- The route table comes out the same: same keys, handlers, names, and options.
- No changelog or upgrade-guide entry is planned.
Not changing
esc() stays in create(). Apps can replace esc() in app/Common.php, so skipping the call would quietly bypass a custom version on route paths. A faster esc() for the html context could be a separate issue.
- The matcher (
Router::checkRoutes()). A FastRoute-style matcher was prototyped. Without route caching, building its lookup structure on each request cost about as much as the linear scan it replaced. So the net gain was small at typical route counts.
Tests
Add cases to tests/system/Router/RouteCollectionTest.php that lock in today's behavior:
- A route with no placeholders keeps its exact key.
- A custom placeholder whose pattern contains
(?:...), mixed with built-in placeholders, still expands correctly.
Run RouteCollectionTest.php and RouterTest.php.
Related
Problem
Without route caching, the route table is rebuilt on every request. Each
$routes->get()(orpost(),add(), and so on) runsRouteCollection::create(), which costs about 1.9 µs per route. That's about 8 times more than the worst-case cost of matching a route.Two steps in
create()do work that is usually wasted:create()runsstr_ireplace(':' . $tag, ...)once for every registered placeholder (6 by default), even when$fromhas no:and nothing can match. That costs about 0.50 µs per route.helper('array'). It's called on everycreate().RouteCollectionno longer uses any array-helper function, but the call still loads the helper for the app as a side effect. Repeat calls cost about 0.24 µs per route.Measurements
PHP 8.3 on macOS, OPcache on, production mode, warm. Config construction cost was turned off to approximate the state after #10615. The routes are a realistic mix: a third static, a third with one placeholder, a third with two, spread across groups.
Here's the cost breakdown of one
$routes->get()call:esc(strip_tags($from))str_ireplaceloop (6 placeholders)helper('array')Proposal
create(), run the placeholder loop only when the route key contains:. If it has no:, nostr_ireplacecan match, and no replacement can add a:. So the output is identical. That includes custom placeholders whose patterns contain:, such as(?:...).helper('array')only on the firstcreate()of each collection, using a private flag. The helper still loads at the same point as today, so apps that quietly depend on it keep working. A short comment will explain that the call stays only for this side effect.Estimated savings are about 0.74 µs per route, roughly 40% of the build cost:
Compatibility
Not changing
esc()stays increate(). Apps can replaceesc()inapp/Common.php, so skipping the call would quietly bypass a custom version on route paths. A fasteresc()for thehtmlcontext could be a separate issue.Router::checkRoutes()). A FastRoute-style matcher was prototyped. Without route caching, building its lookup structure on each request cost about as much as the linear scan it replaced. So the net gain was small at typical route counts.Tests
Add cases to
tests/system/Router/RouteCollectionTest.phpthat lock in today's behavior:(?:...), mixed with built-in placeholders, still expands correctly.Run
RouteCollectionTest.phpandRouterTest.php.Related
Config/Filters.phpon every request #10616: filter discovery cost