[cfe-dev] ARMv7 float->int rounding

[Domagoj Saric via cfe-dev]

Hi everyone,
after years of 'trying&waiting' I still cannot get Clang to generate efficient 
code for float->int conversion on armv7 with either explicit round-to-nearest or 
'current'/'ambient' rounding mode:

std::int32_t round( float const floatingPointValue )
{
#if A
   return __builtin_lrintf( floatingPointValue );
#elif B
   std::int32_t integerValue;
   __asm__
   (
     "vcvtr.s32.f32 %0, %1" : "=w"( integerValue ) : "w"( floatingPointValue );
   );
   return integerValue;
#elif C
   return __builtin_arm_vcvtr_f( floatingPointValue, 0 );
#else // fallback
   return floatingPointValue + __builtin_copysignf( 0.5f, floatingPointValue );
#endif
}

A) 'fails'/is no good because https://llvm.org/bugs/show_bug.cgi?id=11544 
("Trivial math builtins not inlined") is still alive

B) crashes with an assertion ('why on earth' is clang distributed with 
assertions turned on?):
"error: couldn't allocate output register for constraint 'w'"

C) crashes with:
"fatal error: error in backend: Cannot select: intrinsic %llvm.arm.vcvtr"

...tested with Clang 3.6 from Android NDK r10e (latest) and Apple Clang from 
Xcode 7.2.1 (latest)...


Is there anything I can do to make Clang emit the vcvtr instruction?


ps. I stumbled on __builtin_arm_vcvtr_f by pure chance (it isn't documented 
anywhere, especially not its second parameter)...


-- 
Domagoj Saric
Software Architect
www.LittleEndian.com