-
Notifications
You must be signed in to change notification settings - Fork 21.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
32% Faster Object#try #33747
32% Faster Object#try #33747
Conversation
04774bf
to
7a6d760
Compare
I follow the gain from inlining I don't see why
I suspect you'll find the saving is in avoiding the splat capture in |
Ahh, that makes sense. Thanks. My first "DRY" attempt of fixing this was to introduce a private method without the splat, but I like this approach better.
This version was easier for me to see that both versions are identical. I'll change it back to the original logic. |
Here’s the micro benchmark: ```ruby module ActiveSupport module NewTryable #:nodoc: def try(*a, &b) return unless a.empty? || respond_to?(a.first) return public_send(*a, &b) unless a.empty? return nil unless block_given? return instance_eval(&b) if b.arity == 0 yield self end def try!(*a, &b) return public_send(*a, &b) if !a.empty? return nil unless block_given? return instance_eval(&b) if b.arity == 0 yield self end end end module ActiveSupport module OldTryable #:nodoc: def try(*a, &b) try!(*a, &b) if a.empty? || respond_to?(a.first) end def try!(*a, &b) if a.empty? && block_given? if b.arity == 0 instance_eval(&b) else yield self end else public_send(*a, &b) end end end end class FooNew include ActiveSupport::NewTryable def foo end end class FooOld include ActiveSupport::OldTryable def foo end end foo_new = FooNew.new foo_old = FooOld.new require 'benchmark/ips' Benchmark.ips do |x| x.report("old") { foo_old.try(:foo) } x.report("new") { foo_new.try(:foo) } x.compare! end # Warming up -------------------------------------- # old 144.178k i/100ms # new 172.371k i/100ms # Calculating ------------------------------------- # old 2.181M (± 8.0%) i/s - 10.813M in 5.001419s # new 2.889M (± 7.7%) i/s - 14.479M in 5.051760s # Comparison: # new: 2888691.7 i/s # old: 2180740.7 i/s - 1.32x slower ``` Also reduces memory. On https://www.codetriage.com i’m seeing 1.5% fewer object allocations per request (in object count). Before: Total allocated: 1014475 bytes (8525 objects) After: Total allocated: 1015499 bytes (8389 objects)
7a6d760
to
ba7d126
Compare
Updated to use original logic inside the method. Still preserving the removal of the splat expansion and the reduction in method calls. |
There's definitely another 20% win by pulling out the method name into its own argument. Will open a followup PR |
Following up on #33747, this takes things a step further by pulling out the method name from the arguments array, letting us skip an allocation in the case where there are no arguments -- notably, this also no longer *requires* the splat to be an array, allowing us to benefit from optimizations in Jruby (and maybe MRI in the future) of skipping the array allocation entirely. Benchmark results: ``` Warming up -------------------------------------- old 179.987k i/100ms new 199.201k i/100ms Calculating ------------------------------------- old 3.029M (± 1.6%) i/s - 15.299M in 5.052417s new 3.657M (± 1.2%) i/s - 18.326M in 5.012648s Comparison: new: 3656620.7 i/s old: 3028848.3 i/s - 1.21x slower ```
Here’s the micro benchmark:
Also reduces memory. On https://www.codetriage.com i’m seeing 1.5% fewer object allocations per request (in object count).
Before:
Total allocated: 1014475 bytes (8525 objects)
After:
Total allocated: 1015499 bytes (8389 objects)
Why?
Most of the savings come from getting rid of an arg splat
*a
when callingtry!
from withintry
. This cause an extra array to be allocated. A tiny savings comes from removing the extra method call.Did you consider?
A logical question might be why not get rid of the splat in
public_send(*a, &b)
for the case where there's only a single symbol and for some reason it's not as big of a savings. No idea why.BTW if anyone shows up and wants to performance code golf, please use the microbenchmark to test your own ideas, don't post "did you consider ..." because that's kind of a jerk move to make someone else test your hypothesis for you (unless your in Rails core and you get a pass). Plus you'll get valuable benchmarking experience!