Optimize qualified- / simple- ident? / keyword? / symbol?

Description

There are some optimizations that we can make for the simple-ident?, qualified-ident?, simple-symbol?, qualified-symbol?, simple-keyword?, and qualified-keyword? functions.

Two separate optimization can be done.

In places where boolean is used (the qualified- fns), we can instead use some? for a faster implementation:

can be converted to simply

Focusing on just V8 for now, for

and subtracting out the baseline cost for

this yields a speedup of 1.74.

Since we know that we are dealing with keywords or symbols when applying namespace, we can furthermore use (.-ns x).

Alone this gives a speedup of 2.21 and when combined with the previous optimization as

we see a combined speedup approaching infinity (the operation effectively becomes free, taking the same time as the baseline).

Note that we could perhaps use predicate-induced type inference along with a new macro for namespace that emits .-ns instead, but CLJS-2866 currently hinges on looking at the test in an if, while expressions like the above cause optimized js* output with short-circuiting &&.

These optimizations would be good because these functions are often used in spec-based generative code testing where performance can be an issue.

We have test coverage for these functions in cljs.predicates-test.

For reference here is the JavaScript currently emitted for qualified-symbol?

and here is what we get with the revisions

(The compactness of this revised implementation means that it might be further inlinable via Closure for more speedups for whole-program optimization.)

Environment

None

Activity

Show:
Mike Fikes
July 7, 2019, 7:14 PM

namespace and name intrinsics:

Mike Fikes
July 7, 2019, 10:59 PM

Predicate-induced type information flowing rightwards though and:

Mike Fikes
July 7, 2019, 11:21 PM

With the addition of the above two tickets and CLJS-3133, simple-keyword? produces the exact same code without having to modify it:

And qualified-keyword? is close:

compared to

Mike Fikes
July 7, 2019, 11:56 PM

Using some? as described in the ticket still seems relevant, as it enables && and elimination of unchecked ifs:

But, we don't need to use .-ns if we the other tickets in place.

Mike Fikes
July 7, 2019, 11:58 PM

My recommendation would be to wait and see if the other patches land, and if so, perhaps produce a simpler patch for this ticket that simply employs some? where appropriate.

Assignee

Unassigned

Reporter

Mike Fikes

Labels

Approval

None

Patch

Code

Priority

Major
Configure