Saagar Jha (@saagar@saagarjha.com)

Saagar Jha

04/07/2025 (replying to Saagar Jha)

Whatever it was I guess it wasn’t enough

Popover from Instruments after running the Processor Trace tool.

[Error]
Timestamp: (Before Run Started)
Message: Failed to stop recording session: Data Providers emitted errors: CPUTrace

1 replies →

1 replies

Longhorn

04/07/2025 (replying to Saagar Jha)

@saagar get-task-allow

Saagar Jha

04/07/2025 (replying to Saagar Jha)

lmao, ok. I guess this literally refuses to work if you don’t give the traced process get-task-allow which is definitely one of the choices of all time

I guess we can just let the https://developer.apple.com/forums/thread/776577 guy be confused why it doesn’t work or we can fix this: FB17150847

Screenshot of Binary Ninja showing decompilation of an unknowable function inside of libhwtrace.dylib that calls csops with CS_OPS_STATUS and checks for CS_GET_TASK_ALLOW before proceeding.

Saagar Jha

04/07/2025 (replying to Saagar Jha)

The process, of course, gets traced anyway because it don't care what the kernel thinks is impossible.

https://github.com/saagarjha/dotfiles/blob/master/fix_processor_trace.c

1 replies →

1 replies

tbodt

04/07/2025 (replying to Saagar Jha)

@saagar amfitrustedkeys eta wen

Saagar Jha

04/08/2025 (replying to tbodt)

@tbodt I thought you were working on it

tbodt

04/08/2025 (replying to Saagar Jha)

@saagar depends on how you define working

Saagar Jha

04/07/2025 (replying to Saagar Jha)

Now we’re getting somewhere. Not fast, mind you, but somewhere

Screenshot of Instruments running a Processor Trace against unxip on Saagar’s Software MacBook Air. The trace has been stopped after 5 seconds and is “Analyzing/Transferring Data”.

1 replies →

1 replies

Joachim

04/07/2025 (replying to Saagar Jha)

@saagar yeah, there is a *lot* of data in there.

Saagar Jha

04/07/2025 (replying to Saagar Jha)

CPU traces can be very large so Instruments helpfully provides an option asking you how much storage space you want to spare. For this 5 second trace I told it to use 10 GB. So far it has used 300 GB and it’s showing no signs of stopping (FB17151084)

1 replies →

1 replies

Jonathan Hendry

04/07/2025 (replying to Saagar Jha)

@saagar

Does it let you choose where to put it?

Saagar Jha

04/08/2025 (replying to Jonathan Hendry)

@jonhendry The intermediate stuff? No it goes in /var/folders or whatever. I don’t have the courage to try to save the resulting file

Jonathan Hendry

04/08/2025 (replying to Saagar Jha)

@saagar

So not something people with a small boot drive should attempt. If it's using that much disk it really ought to let you stuff it on an external drive.

Saagar Jha

04/07/2025 (replying to Saagar Jha)

It’s beautiful…

The analyzed trace, showing all tracks filled (but not expanded). The cursor is hovering over one track, actually one stack frame, showing that it is “specialized Array.init(_unsafeUninitializedCapacity:initializingWith:) (8.00ms)”

2 replies →

2 replies

James Atkinson

04/07/2025 (replying to Saagar Jha)

@saagar maybe @KacperHarasim can advise?

Pierre H.

04/07/2025 (replying to Saagar Jha)

@saagar isn’t it?!

Saagar Jha

04/07/2025 (replying to Saagar Jha)

Ever wondered system libraries have initializers? How about which ones are the most expensive? Which one is the first to touch the Objective-C runtime? Well here’s your answer in a single image

Screenshot of early dyld initialization, specifically libSystem_initializer calling _libc_initializer, _sanitizers_init, __malloc_init, _dyld_initializer, dyld4::APIs::_libdyld_initialize(), libdispatch_init, _libxpc_initializer in that order

4 replies →

4 replies

Adam Bell :catPOWER:

04/07/2025 (replying to Saagar Jha)

@saagar this is actually so sick

Ben Scheirman

04/07/2025 (replying to Saagar Jha)

@saagar do you mean "static initializers” ?

Saagar Jha

04/07/2025 (replying to Ben Scheirman)

@bens They’re morally kind of similar but I would shy away from calling them that because they are invoked differently

Pierre H.

04/07/2025 (replying to Saagar Jha)

@saagar you have no idea how familiar I am with this particular image.

Joachim

04/07/2025 (replying to Saagar Jha)

@saagar also you can now see how much time objc_msgSend actually takes in your app. And how often it was called.

Saagar Jha

04/08/2025 (replying to Joachim)

@cocoafrog Of course there’s always DTrace for that ;)

Joachim

04/08/2025 (replying to Saagar Jha)

@saagar not in iOS, iirc ;)

Saagar Jha

04/08/2025 (replying to Joachim)

@cocoafrog Yeah the iOS people are mean

Saagar Jha

04/07/2025 (replying to Saagar Jha)

Anyway I have actual work to do so I’ll play with this later. It’s very rough but also very cool and I expect it to be invaluable for any serious performance or debugging work. I hope to see it continue to improve :)

Saagar Jha

04/07/2025 (replying to Saagar Jha)

Also, fair warning: the entire process to generate the screenshots above created about 4 TB of disk traffic. Hopefully you have AppleCare

4 replies →

4 replies

Jevin Sweval

04/07/2025 (replying to Saagar Jha)

@saagar fffffffuuuuuuuuuuuu well that’s disappointing. Guess we need to figure out how to plug one of the 4U units in.

https://www.exostivlabs.com/exostiv-blade/

Kacper Harasim

04/07/2025 (replying to Jevin Sweval)

@jevinskie @saagar Developers can control how much data gets recorded using the recording options ("Limit to:"). We recommend starting with smaller traces, which should not consume as much RAM and load faster as well. We're devoted to continuing to optimize this experience and offering reasonable performance for larger traces over time.

Saagar Jha

04/08/2025 (replying to Kacper Harasim)

@KacperHarasim @jevinskie Curious what this number corresponds to

Jevin Sweval

04/07/2025 (replying to Saagar Jha)

@saagar is it tracev3 format? I wonder if some other compressor could be hooked in…
Any idea how the trace system avoids self-tracing/minimizes perturbation?

Pierre H.

04/07/2025 (replying to Jevin Sweval)

@jevinskie @saagar that’s not how tracing works. The CPU traces everything and then the trace is post processed. You know which asid is current and prune it