[SR-14528] Arm32/linux Performance regression

[hpux735]

Previous ID	SR-14528
Radar	None
Original Reporter	@hpux735
Type	Bug
Status	Closed
Resolution	Cannot Reproduce

Environment

Swift 5.3 on Arm32/linux with kernel 4.9 and 5.9.

Additional Detail from JIRA

Votes	1
Component/s	Compiler, Standard Library
Labels	Bug
Assignee	@colemancda
Priority	Medium

md5: cbcee298a1647106fe92626d16e134ab

Issue Description:

@Catfish-Man asked me to make a bug about this performance regression we're seeing between Swift 5.1 (we have had 5.1 deployed in production for quite some time) and Swift 5.3 (we just got it cross-compiling our binaries via Yocto/Bitbake).

It's a little hard to pin-down the cause. At first I thought it was strictly `futex` related, but I started a longer-running daemon and let it finish loading (where a lot of retain/release activity is) and those calls settled down. However, the performance issue remained.

I have several time profiles from `perf` available if desired, but they'll need to be used with a rootfs image containing the debug symbols (we have a link to this available, but won't share it on a public forum).

I have @colemancda assigned on our team to work on this bug full-time, but we're not really swift compiler engineers, so I'm sure he'll need some pointers to be maximally helpful

[typesanitizer]

Could you file a bug report through feedbackassistant.apple.com with the perf time profiles and the rootfs image link attached? That seems like it would be helpful in reproducing and measuring things.

[hpux735]

Ok, that's complete. I'm not sure how to link it here in a sensible way. The title of the report is:

`Performance regression between Swift 5.1 Swift 5.3 on Arm32`

[hpux735]

Here is a test tool we wrote for characterizing different factors that impact the regression. We're working on buttoning-up the readme now https://github.com/colemancda/SR-14528

[colemancda]

We have been able to build Swift 5.5.3 with Yocto and Buildroot without the performance penalty. It seems that Yocto was injecting some build settings that caused the underlying issue. The issue was never observed when cross compiling the Swift stdlib with Buildroot.

[colemancda]

No longer reproducible on Swift 5.5.3