[gsoc2024][libc][math][c23] Implement _Float16 higher math functions

[overmighty]

This issue is part of the following Google Summer of Code 2024 project: https://summerofcode.withgoogle.com/programs/2024/projects/1vPVUj93.

• acosf16
• acoshf16
• acospif16
• asinf16
• asinhf16
• asinpif16
• atanf16
• atan2f16
• atan2pif16
• atanhf16
• atanpif16
• cbrtf16
• compoundnf16
• cosf16: [libc][math][c23] Add cosf16 function #118785
• coshf16: [libc][math][c23] Add sinhf16 and coshf16 C23 math functions #105947
• cospif16: [libc][math][c23] Add cospif16 function #113001
• dsqrtf16
• erff16
• erfcf16
• expf16: [libc][math][c23] Add expf16 C23 math function #100632
• exp10f16: [libc][math][c23] Add exp10f16 C23 math function #101588
• exp10m1f16: [libc][math][c23] Add exp10m1f16 C23 math function #105706
• exp2f16: [libc][math][c23] Add exp2f16 C23 math function #101217
• exp2m1f16: [libc][math][c23] Add exp2m1f16 C23 math function #105690
• expm1f16: [libc][math][c23] Add expm1f16 C23 math function #102387
• fmaf16
• f16sqrt (extension): [libc][math][c23] Add f16sqrt{,l,f128} C23 math functions #96642
• f16sqrtf (extension): [libc][math][c23] Add f16sqrtf C23 math function #95251
• f16sqrtl (extension): [libc][math][c23] Add f16sqrt{,l,f128} C23 math functions #96642
• f16sqrtf128: [libc][math][c23] Add f16sqrt{,l,f128} C23 math functions #96642
• fsqrtf16
• hypotf16
• lgammaf16
• logf16: [libc][math][c23] Add logf16 C23 math function #106072
• log10f16: [libc][math][c23] Add log10f16 C23 math function #106091
• log10p1f16
• log1pf16
• log2f16: [libc][math][c23] Add log2f16 C23 math function #106084
• log2p1f16
• logp1f16
• powf16
• pownf16
• powrf16
• rootnf16
• rsqrtf16
• sinf16: [libc][math][c23] Add sinf16 C23 math function #116674
• sincosf16 (extension)
• sinhf16: [libc][math][c23] Add sinhf16 and coshf16 C23 math functions #105947
• sinpif16: [libc] Add sinpif16 function #110994
• sqrtf16: [libc][math][c23] Add sqrtf16 C23 math function #112406
• tanf16: [libc][math][c23] Add tanf16 function #121018
• tanhf16: [libc][math][c23] Add tanhf16 C23 math function #106006
• tanpif16: [libc][math][c23] Add tanpif16 function #115183
• tgammaf16

[tgross35]

Are there any plans for LLVM to lower to these intrinsics, or does it already ever? Or is this just making the functions available for calling directly.

[overmighty]

Are there any plans for LLVM to lower to these intrinsics, or does it already ever? Or is this just making the functions available for calling directly.

Sorry for the late reply. I have been told there are plans to at some point have LLVM call LLVM libc math functions for computations internally, and later have LLVM lower to calls to these functions too.

[tgross35]

Lowering to these functions may be nice since it provides a way to get correctly rounded f16 results without correctly rounded f32 results (likely meaning performance wins on platforms where hardware f64 is slower than f32 or unavailable) but it does add some complexity. If this is done, I would appreciate the ability for frontends to specify which f16 symbols are available so it is possible to provide only a subset. For example, the frontend could configure that only sinf16, cosf16, and tanf16 are available and LLVM should use those where possible, but all other math operations will still lower to f32 versions with conversions.