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

[Durgadoss R via cfe-commits]

================
@@ -2063,43 +2071,62 @@ let Predicates = [doRsqrtOpt] in {
 // Add
 //
 
-def INT_NVVM_ADD_RN_SAT_F16 : F_MATH_2<"add.rn.sat.f16", B16, B16, B16, int_nvvm_add_rn_sat_f16>;
-def INT_NVVM_ADD_RN_FTZ_SAT_F16 : F_MATH_2<"add.rn.ftz.sat.f16", B16, B16, B16, int_nvvm_add_rn_ftz_sat_f16>;
-def INT_NVVM_ADD_RN_SAT_F16X2 : F_MATH_2<"add.rn.sat.f16x2", B32, B32, B32, int_nvvm_add_rn_sat_v2f16>;
-def INT_NVVM_ADD_RN_FTZ_SAT_F16X2 : F_MATH_2<"add.rn.ftz.sat.f16x2", B32, B32, B32, int_nvvm_add_rn_ftz_sat_v2f16>;
-
-def INT_NVVM_ADD_RN_FTZ_F : F_MATH_2<"add.rn.ftz.f32", B32, B32, B32, int_nvvm_add_rn_ftz_f>;
-def INT_NVVM_ADD_RN_SAT_FTZ_F : F_MATH_2<"add.rn.sat.ftz.f32", B32, B32, B32, int_nvvm_add_rn_ftz_sat_f>;
-def INT_NVVM_ADD_RN_F : F_MATH_2<"add.rn.f32", B32, B32, B32, int_nvvm_add_rn_f>;
-def INT_NVVM_ADD_RN_SAT_F : F_MATH_2<"add.rn.sat.f32", B32, B32, B32, int_nvvm_add_rn_sat_f>;
-def INT_NVVM_ADD_RZ_FTZ_F : F_MATH_2<"add.rz.ftz.f32", B32, B32, B32, int_nvvm_add_rz_ftz_f>;
-def INT_NVVM_ADD_RZ_SAT_FTZ_F : F_MATH_2<"add.rz.sat.ftz.f32", B32, B32, B32, int_nvvm_add_rz_ftz_sat_f>;
-def INT_NVVM_ADD_RZ_F : F_MATH_2<"add.rz.f32", B32, B32, B32, int_nvvm_add_rz_f>;
-def INT_NVVM_ADD_RZ_SAT_F : F_MATH_2<"add.rz.sat.f32", B32, B32, B32, int_nvvm_add_rz_sat_f>;
-def INT_NVVM_ADD_RM_FTZ_F : F_MATH_2<"add.rm.ftz.f32", B32, B32, B32, int_nvvm_add_rm_ftz_f>;
-def INT_NVVM_ADD_RM_SAT_FTZ_F : F_MATH_2<"add.rm.sat.ftz.f32", B32, B32, B32, int_nvvm_add_rm_ftz_sat_f>;
-def INT_NVVM_ADD_RM_F : F_MATH_2<"add.rm.f32", B32, B32, B32, int_nvvm_add_rm_f>;
-def INT_NVVM_ADD_RM_SAT_F : F_MATH_2<"add.rm.sat.f32", B32, B32, B32, int_nvvm_add_rm_sat_f>;
-def INT_NVVM_ADD_RP_FTZ_F : F_MATH_2<"add.rp.ftz.f32", B32, B32, B32, int_nvvm_add_rp_ftz_f>;
-def INT_NVVM_ADD_RP_SAT_FTZ_F : F_MATH_2<"add.rp.sat.ftz.f32", B32, B32, B32, int_nvvm_add_rp_ftz_sat_f>;
-def INT_NVVM_ADD_RP_F : F_MATH_2<"add.rp.f32", B32, B32, B32, int_nvvm_add_rp_f>;
-def INT_NVVM_ADD_RP_SAT_F : F_MATH_2<"add.rp.sat.f32", B32, B32, B32, int_nvvm_add_rp_sat_f>;
-
-def INT_NVVM_ADD_RN_D : F_MATH_2<"add.rn.f64", B64, B64, B64, int_nvvm_add_rn_d>;
-def INT_NVVM_ADD_RZ_D : F_MATH_2<"add.rz.f64", B64, B64, B64, int_nvvm_add_rz_d>;
-def INT_NVVM_ADD_RM_D : F_MATH_2<"add.rm.f64", B64, B64, B64, int_nvvm_add_rm_d>;
-def INT_NVVM_ADD_RP_D : F_MATH_2<"add.rp.f64", B64, B64, B64, int_nvvm_add_rp_d>;
+defvar BF16ArithPreds = [hasBF16Math, PTX78, SM90];
+
+class RndModeImm<string mode> : TImmLeaf<i32,
+  "return Imm == static_cast<int>(RoundingMode::" # mode # ");">;
+
+def rnd_rn_imm : RndModeImm<"NearestTiesToEven">;
+def rnd_rz_imm : RndModeImm<"TowardZero">;
+def rnd_rm_imm : RndModeImm<"TowardNegative">;
+def rnd_rp_imm : RndModeImm<"TowardPositive">;
+
+foreach t = [F16RT, F16X2RT] in
+  foreach ftz = ["", "_ftz"] in
+    foreach sat = ["", "_sat"] in
+      def INT_NVVM_ADD_RN # !toupper(ftz # sat # "_" # t.PtxType) :
+        F_MATH_2_RND_TY<!subst("_", ".", "add_rn" # ftz # sat # "_" # t.PtxType),
+                        t, !cast<Intrinsic>("int_nvvm_fadd" # sat # ftz),
----------------
durga4github wrote:

let us be consistent with the sat<->ftz ordering everywhere (across all usages)

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