[swift-dev] [arc optimization] Why doesn't enum destructuring use guaranteed references?

[Robert Widmann]

Ran into this over the summer.  My understanding is that Michael Gottesman (CC’d) has been looking into pattern matching at +0.

The code in SILGenPattern needs to be reworked, but I found this problem intersects with the really old dynamic casting entry points in SILGen too which would seem to complicate a “simple fix”.  It’d be good to hear what plans, if any, there are to address this. It would be a real win to clean all of this up.

~Robert Widmann 

2017/12/28 14:27、Chris Lattner via swift-dev <swift-dev at swift.org>のメール:

> <moving to swift-dev, where most of the perf optimization people lurk>
> 
> This is a great question, I’m not sure what the answer is: maybe it is a simple case missed by the arc optimizer?
> 
> -Chris
> 
> 
> 
>> On Dec 27, 2017, at 9:19 PM, Félix Cloutier via swift-evolution <swift-evolution at swift.org> wrote:
>> 
>> Running this on my MBP with 10 as the parameter: https://gist.github.com/zneak/ae33bb970a08632cfb2925e2049f9e7a 
>> 
>> I get a runtime of about 10 seconds, 45% of which is spent in retain/release calls according to Instruments (!!), and at least half of that from Expr.count. Looking at the IR, Swift generously sprinkles retain/release calls through the outlined copy method:
>> 
>> `self` is retained at the beginning of `count`
>> The values that you get out of destructuring are retained
>> Of course, when you get `count` on these, they are retained again
>> 
>> Of course, Expr.count cannot modify itself or its subexpressions because the functions are not mutating; in fact, no function in that program can mutate an enum case. Why, then, is Swift retaining/releasing `self` and the values obtained from destructured patterns? They can't go away. Shouldn't we be getting guaranteed references instead of owning references?
>> 
>> That seems to hit pattern-matching-heavy programs with indirect cases pretty hard, and it's pretty frustrating because that seems to be about the nicest way to write this program, and there's no workaround from the developer's perspective. I don't think that this is a fatal flaw of refcounting, but unless I'm missing something, that's sub-par behavior.
>> 
>> Félix
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-evolution
> 
> _______________________________________________
> swift-dev mailing list
> swift-dev at swift.org
> https://lists.swift.org/mailman/listinfo/swift-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-dev/attachments/20171229/0497d257/attachment.html>