In some cases, the compiler may encode 16-bit "abbreviated" floating-point immediate values incorrectly. One set of instances where this issue occurs is in the fast-math implementations of the C library functions asinf, acosf and roundf. This causes results failures in code that relies on these operations. Affected code includes:

• Uses of library functions asinf, acosf, and roundf under the -ffast-math flag
• Uses of the "fast" intrinsics __builtin_c29_fast_asinf, __builtin_c29_fast_acosf, and __builtin_c29_fast_roundf

Workarounds:

If you are compiling without O3, you can remap the symbols to the original non-optimized fast math functions during linking using the --symbol_map command. For example:

--symbol_map=fast_acosf=asinf
--symbol_map=fast_acosf=acosf
--symbol_map=fast_acosf=roundf 

If you are compiling with O3 you can use one of the following approaches:

Map an alternative function to the original RTS library implementation with the correct function signature, this will result in calling the original RTS library implementation. For Example:

typedef float (*libcall)(float);
libcall RTS_roundf = roundf; 
libcall RTS__acosf = acosf; 
libcall RTS__asinf = asinf; 

The second option is to block optimization of the RTS call by wrapping the original RTS call in a non-optimized function. For Example:

float32_t __attribute__((optnone)) libcall(float32_t in) {
    return RTS_function(x);
}