[llvm] [AMDGPU] Add half vector support for table-driven libcall optimization (PR #178638)

[Steffen Larsen via llvm-commits]

================
@@ -1400,148 +1402,97 @@ bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0,
              : (double)fpopr1->getValueAPF().convertToFloat();
   }
 
-  switch (FInfo.getId()) {
-  default : return false;
-
-  case AMDGPULibFunc::EI_ACOS:
-    Res0 = acos(opr0);
----------------
steffenlarsen wrote:

Aye, it's not perfect, though I suppose we do ensure a higher (or equally high) precision on these calculations by computing it as a double and then converting.

https://github.com/llvm/llvm-project/pull/178638