[clang] [llvm] [mlir] [clang][NVPTX][MLIR][NVVM] Add overloaded fadd intrinsics (PR #217336)

[Durgadoss R via cfe-commits]

================
@@ -481,87 +483,34 @@ void NVVM::AddFOp::lowerAddFToLLVMIR(llvm::Value *argLHS, llvm::Value *argRHS,
                                      LLVM::ModuleTranslation &mt,
                                      llvm::IRBuilderBase &builder) {
   llvm::Type *opTypeLLVM = argLHS->getType();
-  bool isVectorOp = opTypeLLVM->isVectorTy();
   bool isSat = satMode != NVVM::SaturationMode::NONE;
 
-  // FIXME: Add intrinsics for add.rn.ftz.f16x2 and add.rn.ftz.f16 here when
-  // they are available.
-  static constexpr llvm::Intrinsic::ID f16IDs[] = {
-      llvm::Intrinsic::nvvm_add_rn_sat_f16,
-      llvm::Intrinsic::nvvm_add_rn_ftz_sat_f16,
-      llvm::Intrinsic::nvvm_add_rn_sat_v2f16,
-      llvm::Intrinsic::nvvm_add_rn_ftz_sat_v2f16,
-  };
-
-  static constexpr llvm::Intrinsic::ID f32IDs[] = {
-      llvm::Intrinsic::nvvm_add_rn_f, // default rounding mode RN
-      llvm::Intrinsic::nvvm_add_rn_f,
-      llvm::Intrinsic::nvvm_add_rm_f,
-      llvm::Intrinsic::nvvm_add_rp_f,
-      llvm::Intrinsic::nvvm_add_rz_f,
-      llvm::Intrinsic::nvvm_add_rn_sat_f, // default rounding mode RN
-      llvm::Intrinsic::nvvm_add_rn_sat_f,
-      llvm::Intrinsic::nvvm_add_rm_sat_f,
-      llvm::Intrinsic::nvvm_add_rp_sat_f,
-      llvm::Intrinsic::nvvm_add_rz_sat_f,
-      llvm::Intrinsic::nvvm_add_rn_ftz_f, // default rounding mode RN
-      llvm::Intrinsic::nvvm_add_rn_ftz_f,
-      llvm::Intrinsic::nvvm_add_rm_ftz_f,
-      llvm::Intrinsic::nvvm_add_rp_ftz_f,
-      llvm::Intrinsic::nvvm_add_rz_ftz_f,
-      llvm::Intrinsic::nvvm_add_rn_ftz_sat_f, // default rounding mode RN
-      llvm::Intrinsic::nvvm_add_rn_ftz_sat_f,
-      llvm::Intrinsic::nvvm_add_rm_ftz_sat_f,
-      llvm::Intrinsic::nvvm_add_rp_ftz_sat_f,
-      llvm::Intrinsic::nvvm_add_rz_ftz_sat_f,
-  };
-
-  static constexpr llvm::Intrinsic::ID f64IDs[] = {
-      llvm::Intrinsic::nvvm_add_rn_d, // default rounding mode RN
-      llvm::Intrinsic::nvvm_add_rn_d, llvm::Intrinsic::nvvm_add_rm_d,
-      llvm::Intrinsic::nvvm_add_rp_d, llvm::Intrinsic::nvvm_add_rz_d};
-
-  auto addIntrinsic = [&](llvm::Intrinsic::ID IID) -> llvm::Value * {
-    return createScalarizedIntrinsicCall(builder, IID, opTypeLLVM,
-                                         {argLHS, argRHS}, opTypeLLVM);
-  };
-
-  // f16 + f16 -> f16 / vector<2xf16> + vector<2xf16> -> vector<2xf16>
-  // FIXME: Allow lowering to add.rn.ftz.f16x2 and add.rn.ftz.f16 here when the
-  // intrinsics are available.
-  if (opTypeLLVM->getScalarType()->isHalfTy()) {
-    llvm::Value *result;
-    if (isSat) {
-      unsigned index = (isVectorOp << 1) | isFTZ;
-      result = addIntrinsic(f16IDs[index]);
-    } else {
-      result = builder.CreateFAdd(argLHS, argRHS);
-    }
-    mt.mapValue(res, result);
+  static constexpr llvm::Intrinsic::ID addIDs[2][2] = {
+      {llvm::Intrinsic::nvvm_fadd, llvm::Intrinsic::nvvm_fadd_sat},
+      {llvm::Intrinsic::nvvm_fadd_ftz, llvm::Intrinsic::nvvm_fadd_sat_ftz}};
+
+  static constexpr llvm::RoundingMode roundingModes[5] = {
+      llvm::RoundingMode::NearestTiesToEven,
+      llvm::RoundingMode::NearestTiesToEven, llvm::RoundingMode::TowardNegative,
+      llvm::RoundingMode::TowardPositive, llvm::RoundingMode::TowardZero};
----------------
durga4github wrote:

Dont we have an FP rounding mode enum in NVVM Dialect that we can use here? (instead of re-defining this array)

https://github.com/llvm/llvm-project/pull/217336