Translating an ARM iOS App to Intel macOS Using Bitcode | Hacker News Reader

Translating an ARM iOS App to Intel macOS Using Bitcode

highcaffeinecontent.com

148 pointssjuut23 comments

The intro makes it sound like this could be feasible for translating x86 Mac apps into ARM Mac apps.

> So, what if the Mac switched to using ARM chips instead of Intel? Well, as you can see, Apple could use Bitcode to translate every Bitcode-enabled app on the Mac App Store, without consulting developers, so it would be ready to go on day one.

But:

> Apple doesn't enable Bitcode for submissions to the Mac App Store today

Right. The flag to enable bitcode isn't even visible in Xcode for Mac apps. There's no such thing as Mac/x86 bitcode yet, in or out of the MAS. And:

> ARC appears to use some inline assembly, which means you'll need to disable ARC for a project for arm64-to-x86 translation to succeed right now.

So this approach will only work on Objective-C programs, written in a pre-2011 style -- which I'd bet is a vanishingly small fraction of the apps on the MAS today.

It's a neat exercise in what is possible for a simplest-possible-app, but it's hard to imagine Apple would actually use this approach. It'd require them to ship a couple big new features in Xcode, and then for everyone on the MAS to recompile their apps using those features. Easier all around to just add a new architecture in Xcode the normal way, and have developers compile fat binaries.

”ARC appears to use some inline assembly, which means you'll need to disable ARC for a project for arm64-to-x86 translation to succeed right now.”

That’s fairly easily fixed, certainly if you’re Apple. They already have the equivalent x86 code (Swift and modern Objective-C couldn't exist without it), so they could detect the specific inline ARM code fragment and replace it by the equivalent inline x86 code.

This seems like it would be very fragile.

That's exactly my thinking-- it might be feasible for Apple to use bitcode during an x86 to ARM transition, but what problem is that actually solving?

At best, Apple announces mandatory bitcode* for macOS App Store apps at WWDC in 2019; Apple announces the first ARM Macs in 2020 or 2021. So you've got a window of maybe a year or two of apps that don't have to be recompiled to support ARM, and most are likely still being updated since it hasn't been that long since they recompiled to supply Apple with a bitcode version.

This doesn't do anything to solve the big headache of an architecture transition, which are those legacy apps that will never be updated again-- you'd still need an emulation solution like Rosetta or Microsoft's x86 on ARM stuff to support those applications. At best, you're maybe working around the Adobes and Microsofts of the world that move too slowly to have a day-1 update when the new architecture comes out?

* I'll note that Bitcode wasn't mandatory on iOS when it was announced in 2015, and it's still not mandatory there today-- it seems unlikely that Mac Bitcode would be mandatory immediately when they announce it or in the foreseeable future.

Apple, having already forced PPC->X86 and now forcing X86->X64, clearly has no difficulty forcing an architecture update every 5 years, stranding apps in the past if they don’t update. Saying that they wouldn’t force X64->ARM64 due to migration issues they have repeatedly accepted makes no sense, given their history - and especially having done so regardless of Adobe’s inability to keep up.

They've been even more aggressive than you suggest. They promised 64-bit Carbon in 10.5, and included it in developer seeds, before finally yanking it from the final release. Adobe found out it was cancelled at the same time as everyone else: when Apple announced it at WWDC.

My recollection is that Apple did in fact ship a 64-bit HIToolbox library, but they didn't allow the general public to use it. I thought Adobe was one of the few people who were allowed to use it though, as a temporary measure until they rewrote Photoshop to stop using Carbon. I could be wrong though, that was a long time ago.

Both PPC->x86 and x86->x86_64 had (temporary) migration paths. The x86->x86_64 path was pretty easy; x86 apps were still supported for a rather long time. The PPC->x86 migration path wasn't as clean, but they did ship Rosetta for a couple of years before ditching it.

If they do Intel->ARM they'll need some sort of migration strategy. Would that involve a Rosetta 2.0?

The intel transition was 13 years ago? And the powerpc era lasted from 1992ish to 2005?

Wikipedia may be of some use here, I don’t know the exact dates of when OS X was released (with a Mac Classic emulator, how meta!) versus when Apple transitioned 6502->68k and 68k->ppc and ppc->x86. Sorry I can’t be of more help to your questions!

Bitcode for the MacOS App Store would make sense if Intel ever gets its head out of its backside and was able to ship the wide variety of processors Apple wants. Then they could be optimized at install time for the exact processor on your Mac.

I’m not sure it’s really worth it today but this support would give Apple air cover for a future intel-arm transition.

This is rather "embed intermediate representation, extract, run another target's back end, and link". I was hoping this would be explaining how to go from ARM by uplifting executable to LLVM IR and retargeting for Intel. This is notionally already possible with ELF and PE executables using McSema (https://github.com/trailofbits/mcsema).

If anyone is working on this please email me (address in profile), it's something we slated for a future research project and I'd love to talk.

>McSema and remill were developed and are maintained by Trail of Bits, funded by and used in research for DARPA and the US Department of Defense.

So?

Don’t tell them where the internet came from, there ain’t enough tinfoil in the world.

Static binary lifting is, in general, undecidable. Actually, correct disassembly is itself undecidable, and binary lifting builds on it. Doing it dynamically, like QEMU/tcg does, is doable.

It's really not too difficult to write a lifter, but it's time consuming. You have to take every instruction from the source instruction set and model its side effects in some IR. To make things simple, "whole binary" lifting tools seem to use a global CPU state (registers, flags, etc...) and so the IR you get is quite different from what you would obtain from a compiled program.

Sounds like we're reinventing the AS/400 piecemeal:

https://en.wikipedia.org/wiki/IBM_System_i#Instruction_set

http://ibmsystemsmag.com/blogs/you-and-i/archive/timi-protec...

I would be very surprised (and impressed!) if Apple managed to pull this off. With the bitcode transition on watchOS Apple had full control over the ABI and architecture that they were going to target: this isn't the case for x86. Sure, this works for the very simple case presented here, but I forsee it quickly breaking down for any non-trivial application. At the very least, it will not work correctly on every application.

Alternatives:

1. WebAssembly

2. A subset of LLVM-IR that's architecture independent. PNacl from Google used this across chip architectures

Neither of these would help migrate current Mac apps.

I meant, they are alternatives for Apple to consider, instead of reinventing the wheel on a cross architecture IR. Of course they won't help current Mac apps that don't have IR in those formats

I don’t think we really have a “cross architecture IR” yet.

Who's "we"? If you mean Apple, their IR (bitcode) is alomst a 100% the same as LLVM-IR, so it is not really Apple's. Aside from WebAssembly and PNacl cited in my parent comment, there's .net IR, JVM byte code, yada yada yada ..

The universe is teeming with cross architecture IRs. Eventually, LLVM will have a cross platform LLVM-IR, which will possibly be even more impactful than WebAssembly has been in it's short lifespan

Reply on news.ycombinator.com