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

[Steffen Larsen via llvm-commits]

================
@@ -743,6 +743,38 @@ bool AMDGPULibCalls::fold(CallInst *CI) {
   return false;
 }
 
+static Constant *
+_Z4coshdgetConstantFloatVectorForArgType(LLVMContext &Ctx,
+                                         AMDGPULibFunc::EType ArgType,
+                                         ArrayRef<double> Values, Type *Ty) {
+  switch (ArgType) {
+  case AMDGPULibFunc::F16: {
+    SmallVector<uint16_t, 0> HalfIntValues;
+    for (double D : Values) {
+      APFloat APF16 = APFloat(D);
+      [[maybe_unused]] bool Unused;
+      APF16.convert(llvm::APFloat::IEEEhalf(),
+                    llvm::RoundingMode::NearestTiesToEven, &Unused);
+      uint16_t APF16Int = APF16.bitcastToAPInt().getZExtValue();
+      HalfIntValues.push_back(APF16Int);
+    }
+    ArrayRef<uint16_t> Tmp(HalfIntValues);
+    return ConstantDataVector::getFP(Ty->getScalarType(), Tmp);
+  }
+  case AMDGPULibFunc::F32: {
+    SmallVector<float, 0> FValues;
----------------
steffenlarsen wrote:

I didn't pay it no mind, but you are absolutely right. I took the liberty of adding a reservation as well.

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