perf: optimize some conversions of unsigned ints to NSString

[armcknight]

Some work done in a tight loop on the profiler's sampling thread used inefficient methods to convert integers to strings. We sped this up by using snprintf.

Targets to optimize:

• boxing unsigned integers to NSNumbers and converting to NSString with -[NSNumber stringValue]
• boxing to NSNumber and calling sentry_formatHexAddress which called +[NSString stringWithFormat:]
• directly formatting an unsigned int to a string with [NSString stringWithFormat:@"%llu"

Observed a 75% reduction in cost of doing the conversions using Instruments, just in the profiler's sampling thread: https://www.notion.so/sentry/formatHexAddress-perf-optimization-6ce9e406e53b4d99aed2de591428599b?pvs=4

Converted a couple other usages in the SDK to also use the new function, so we get some other perf wins for free. There are still numerous usages of the older sentry_formatHexAddress(NSNumber *) that are using values from JSON-decoded NSDictionary instances in SentryCrashReportConverter, which could be a possible target for future optimizations by the Cocoa team.

TODO

• send a profile from current implementation using iOS-Swift, and then one from this changeset for the same operation, and compare them to make sure this doesn't break anything in subtle ways. tested the "fully blocking anr" in a manual transaction
  • main: https://sentry-sdks.sentry.io/profiling/profile/sentry-cocoa/d112e213125c483280ba883f15dd72b0/flamechart/?colorCoding=by+system+vs+application+frame&fov=0%2C0%2C14142140416%2C48&query=&sorting=call+order&tid=259&view=top+down
  • with changes: https://sentry-sdks.sentry.io/profiling/profile/sentry-cocoa/6496c43271564104b8a997de3ea5a645/flamechart/?colorCoding=by+system+vs+application+frame&fov=28102656%2C0%2C14030778368%2C48&query=&sorting=call+order&tid=259&view=top+down
    

#skip-changelog

[github-actions]

Performance metrics 🚀

	Plain	With Sentry	Diff
Startup time	1253.18 ms	1270.40 ms	17.22 ms
Size	20.76 KiB	434.65 KiB	413.88 KiB

Baseline results on branch: main

Startup times

Revision	Plain	With Sentry	Diff
c00eafe	1198.26 ms	1227.62 ms	29.36 ms
ecd9ecd	1241.28 ms	1260.35 ms	19.07 ms
8526e93	1228.24 ms	1239.14 ms	10.90 ms
2cf1b84	6265.25 ms	6277.98 ms	12.73 ms
efb2222	1256.44 ms	1278.90 ms	22.46 ms
aa27ac0	1214.02 ms	1227.42 ms	13.40 ms
d60f70a	1219.63 ms	1228.54 ms	8.91 ms
ea73af6	1241.69 ms	1252.96 ms	11.27 ms
9ef729b	1228.79 ms	1245.36 ms	16.57 ms
e79ead7	1233.39 ms	1255.52 ms	22.13 ms

App size

Revision	Plain	With Sentry	Diff
c00eafe	20.76 KiB	432.87 KiB	412.11 KiB
ecd9ecd	20.76 KiB	420.23 KiB	399.47 KiB
8526e93	20.76 KiB	420.22 KiB	399.47 KiB
2cf1b84	20.76 KiB	431.91 KiB	411.15 KiB
efb2222	20.76 KiB	424.45 KiB	403.69 KiB
aa27ac0	20.76 KiB	432.21 KiB	411.45 KiB
d60f70a	20.76 KiB	430.97 KiB	410.21 KiB
ea73af6	20.76 KiB	425.76 KiB	404.99 KiB
9ef729b	20.76 KiB	432.88 KiB	412.12 KiB
e79ead7	20.76 KiB	426.11 KiB	405.35 KiB

Previous results on branch: armcknight/perf/speed-up-profiler-hex-address-formatting

Startup times

Revision	Plain	With Sentry	Diff
69f3d5b	1243.90 ms	1250.70 ms	6.80 ms
b506df4	1271.62 ms	1273.61 ms	1.99 ms
5854c18	1245.02 ms	1256.52 ms	11.50 ms
8e69b81	1203.33 ms	1229.88 ms	26.55 ms
8b72bae	1220.41 ms	1257.62 ms	37.21 ms
f5bcb8f	1219.49 ms	1240.94 ms	21.45 ms

App size

Revision	Plain	With Sentry	Diff
69f3d5b	20.76 KiB	430.69 KiB	409.93 KiB
b506df4	20.76 KiB	433.00 KiB	412.24 KiB
5854c18	20.76 KiB	433.07 KiB	412.31 KiB
8e69b81	20.76 KiB	431.87 KiB	411.11 KiB
8b72bae	20.76 KiB	432.08 KiB	411.32 KiB
f5bcb8f	20.76 KiB	431.87 KiB	411.11 KiB

[philipphofmann]

I'm surprised that snprintf is much faster than sentry_formatHexAddress. Thanks for the improvement.

[armcknight]

I'm surprised that snprintf is much faster than sentry_formatHexAddress. Thanks for the improvement.

I think it comes down to all the extra allocations around NSStrings (even for the literal format string) and conversion of NSNumber back to an unsigned long long, both of which are avoided in the new version (and avoiding boxing the ull's to NSNumbers in the first place solely to pass them to this function is an extra benefit).

[philipphofmann]

@armcknight, what's stopping us from making this ready for review?

[armcknight]

@philipphofmann this one's ready for you now!

[philipphofmann]

Thanks, @armcknight, for this performance improvement 👏 😃 . The changelog needs some fixes, I believe. LGTM.

[philipphofmann]

m: I think that needs to move up a bit.

[armcknight]

Are you looking at ordering by PR number? I haven't really cared about it too much, but am happy to start.

[philipphofmann]

No, I thought the entry was in an already-released section, but I maybe was mistaken; sorry.

[armcknight]

The only CI failures were codecov, but only because the test coverage was not properly assessed for the header implementations. There is indeed test coverage for them.