It (roughly) doubles the size of the app bundle you upload to Apple. They then compile that down to create normal sized binaries, which are what are actually downloaded to the device.
Bear in mind the source->bitcode->binary process (I'm simplifying here) was previously all performed by LLVM on your computer anyway. It's just that now the final bitcode->binary step is done by Apple, to allow device-specific optimizations of the binary for the devices that are downloading it without you having to do all those device specific optimized builds yourself.
No idea. We'd need to do on-device profiling of 'plain' builds directly from XCode alongside the builds downloaded from the app store, but of course optimizations may only be possible on particular devices, and some apps would benefit more than others, so you'd have to do a huge amount of testing to track down the differences. Also Apple may be doing this due to planned advantages in the future, so testing done now might be pointless anyway.
Its just strange that Apple has not presented any actual numbers on this. Usually you'd expect them to talk about how XYZ is 90% faster, and 100x awesome and what not. It makes me think that they have something else in mind.
It's most likely so you don't have to update the binary when the next version of LLVM comes out. There are a lot of apps that don't get updated with major iOS versions and they would likely see improvement.
Conspiracy theories aside, wouldn't it take them more effort to ship a different version of LLVM for Xcode and another one that runs against bitcode?
