ios - Is there a way to compile for ARM rather than Thumb in Xcode 4?

Question

Apple is recommending to compiling for ARM rather than thumb if there are many floating point operations going on. My whole app is almost one big floating point operation.

Here's what they say in the iOS App Development Workflow Guide:

iOS devices support two instruction sets, ARM and Thumb. Xcode uses Thumb instructions by default because using Thumb typically reduces code size by about 35 percent relative to ARM. Applications that have extensive floating-point code might perform better if they use ARM instructions rather than Thumb. You can turn off Thumb for your application, so it compiles for ARM, by setting the Compile for Thumb build setting to No.

However, I cannot find any "Compile for Thumb" setting in my build settings. Did they rename it? Or is this unavailable now with Xcode 4?

Answer 1

First, the advice to not compile for the Thumb instruction set in order to improve floating point performance only really applies to the old ARMv6 devices.

ARMv7 hardware (iPhone 3G S and newer, including all iPads) uses the more efficient Thumb-2 instruction set, which does not suffer the same sort of floating point slowdowns. For ARMv7 builds, it is recommended in almost all cases that you build for Thumb. I provide a little more detail about this in my answer here.

This might be why this compiler setting is no longer exposed as a common option, because ARMv7 devices are the vast majority of iOS devices out there.

If you want to do this for just your ARMv6 builds, you can go to your build settings and mouse over the "Other C Flags" option. Click on the little plus button that appears to the right of this option and add a condition for the ARMv6 architecture. Do this again to create one for the ARMv7 architecture. Under the ARMv6 architecture, add the extra compiler flag of -mno-thumb (as Kevin suggests).

You should end up with something that looks like the following:

I do this in one of my applications, because I did see a performance boost on the older ARMv6 devices with that. However, another of my applications was slower when not building for Thumb on ARMv6, so you'll want to profile this first.

Additionally, there is currently a bug in the LLVM Compiler 3.0 that ships with Xcode 4.2 (which has since been fixed in 4.2.1, from what I hear) where floating point calculations are compiled wrong under Thumb for ARMv6. If you're using that particular version of Xcode, you'll need to do this for proper behavior on the older devices.