Feature #14329
closedSpeedup `block.call` where `block` is passed block parameter.
Description
abstract¶
Speedup block.call
where block
is passed block parameter with a special object named "proxyblock" which responds to call
and invoke passed block.
background¶
Ruby 2.5 improved performance of passing a passed block parameter by lazy Proc creation ([Feature #14045]).
However, block.call
(block
is a passed block parameter) is not optimized and need to create a Proc
object immediately.
proposal¶
We need to make Proc creation lazily for performance. There are several way to achieve it, but I propose to use special object named "blockproxy" object.
This is a pseudo code to use it:
# block is given block parameter
block.call(1, 2, 3)
#=> translate at compile time
if block is not modified and
block is ISeq/IFunc block
tmp = blockproxy
else
tmp = block # create Proc and so on
end
tmp.call(1, 2, 3)
blockproxy.call
invoke given block if Proc#call
is not redefined, otherwise make a Proc and call Proc#call
as usual.
Advantage of this method is we can also use this technique with the safe navigation operator (block&.call
).
If block is not given, then tmp
will be nil
, and no method dispatched with &.
.
Note that this technique depends on the assumption "we can't access to the evaluated receiver just before method dispatch". We don't/can't access blockproxy
object at method dispatch, and no compatibility issue.
evaluation¶
Using https://github.com/k0kubun/benchmark_driver
prelude: |
def block_yield
yield
end
def bp_yield &b
yield
end
def bp_call &b
b.call
end
def bp_safe_call &b
b&.call
end
benchmark:
- block_yield{}
- bp_yield{}
- bp_call{}
- bp_safe_call{}
- bp_safe_call
Result:
Warming up --------------------------------------
block_yield{} 1.298M i/100ms
bp_yield{} 1.177M i/100ms
bp_call{} 447.723k i/100ms
bp_safe_call{} 413.261k i/100ms
bp_safe_call 2.955M i/100ms
Calculating -------------------------------------
trunk modified
block_yield{} 20.672M 20.808M i/s - 51.933M in 2.512265s 2.495806s
bp_yield{} 16.294M 16.220M i/s - 47.099M in 2.890459s 2.903797s
bp_call{} 5.626M 14.752M i/s - 17.909M in 3.182966s 1.213976s # x2.62
bp_safe_call{} 5.555M 14.557M i/s - 16.530M in 2.975892s 1.135586s # x2.62
bp_safe_call 31.339M 23.561M i/s - 118.184M in 3.771157s 5.016200s
The patch is here:
https://gist.github.com/ko1/d8a1a9d92075b27a8e95ca528cc57fd2