[clang] [llvm] [HLSL] Implement the f16tof32() intrinsic (PR #165860)

[Tim Corringham via llvm-commits]

================
@@ -3780,6 +3751,89 @@ bool SPIRVInstructionSelector::selectResourceNonUniformIndex(
   return true;
 }
 
+bool SPIRVInstructionSelector::selectF16ToF32(Register ResVReg,
+                                              const SPIRVType *ResType,
+                                              MachineInstr &I) const {
+  assert(I.getNumOperands() == 3);
+  assert(I.getOperand(0).isReg());
+  assert(I.getOperand(2).isReg());
+  Register SrcReg = I.getOperand(2).getReg();
+  const SPIRVType *SrcRegType = GR.getSPIRVTypeForVReg(SrcReg);
+  LLT SrcType = MRI->getType(SrcReg);
+  SPIRVType *SrcEltType = GR.getScalarOrVectorComponentType(SrcRegType);
+  SPIRVType *ResEltType = GR.getScalarOrVectorComponentType(ResType);
+  const TargetRegisterClass *SrcRegClass = GR.getRegClass(SrcEltType);
+  const TargetRegisterClass *ResRegClass = GR.getRegClass(ResEltType);
+  MachineIRBuilder MIRBuilder(I);
+  const SPIRVType *Vec2ResType =
+      GR.getOrCreateSPIRVVectorType(ResEltType, 2, MIRBuilder, false);
+  const TargetRegisterClass *Vec2RegClass = GR.getRegClass(Vec2ResType);
+
+  bool Result = true;
----------------
tcorringham wrote:

As you have probably noticed, I have made changes along the lines of farzonl's comment - moving the expansion of the vector spv_legacyf16tof32 to Clang Codegen and using a spv_unpackhalf2x16 instead.
This simplifies the spirv codegen significantly. I have included a comment along the lines of your suggestion in the clang codegen.

https://github.com/llvm/llvm-project/pull/165860