Correctly detect mixed-endian floating point format on ARM platform

Question

I recently ran into an issue with a third party library that was using the following code to test for mixed-endian floating-point format on ARM platforms:

#if defined(__arm__) && !(__ARM_EABI__)

This check was incorrectly detecting mixed-endian format on Android platforms but worked correctly on iOS platforms. After some research I found the debian ArmEabiPort document which contains the following in the GCC preprocessor macros for floating point section which says(emphasis mine):

When porting code to "armel", the following preprocessor macros are interesting:

• __VFP_FP__ means that the floating point format in use is that of the ARM VFP unit, which is native-endian IEEE-754.

• __MAVERICK__ means that the floating point format is that of the Cirrus Logic MaverickCrunch, which is also IEEE-754 and is always little-endian.

• __SOFTFP__ means that instead of floating point instructions, library calls are being generated for floating point math operations so that the code will run on a processor without an FPU.

__VFP_FP__ and __MAVERICK__ are mutually exclusive. If neither is set, that means the floating point format in use is the old mixed-endian 45670123 format of the FPA unit.

For our specific case updating the check to the following fixed the problem:

#if defined(__arm__) && !(__VFP_FP__)

and works on both Android and iOS platforms. Although from the document the more correct check would be:

#if defined(__arm__) && !(__VFP_FP__) && !(__MAVERICK__)

We would like to submit a patch back to the third-party but considering the old check did work for the person who submitted it and how difficult it is to find documentation I don't feel like I have enough information to get this correct.

Are there cases the last check misses? The debian document specifically covers gcc, how portable are these macros?

Update

Based on the helpful comments from artless noise the question boils down to:

The __ARM_EABI__ macro check was introduced as a patch by another user to check for mixed-endian floating point format. So apparently there are systems that this macro works for, what systems are those? Would the __VFP_FP__ macro cover those systems or do I need to take that macro into account as well to prevent breaking existing users when I submit a patch.

Answer 1

You'll also want to check to make sure you aren't on a soft-float toolchain, as the soft-float libraries don't have the mixed-endian issue:

#if defined(arm) && !defined(__SOFTFP__) && !defined(__VFP_FP__) && !defined(__MAVERICK__)

As to compiler support -- these should work in clang, but obviously, the commercial ARM compilers will do something else entirely.