Add saturating add/sub intrinsics to RISC-V vector backend

Map Halide's saturating_add and saturating_sub operations to
RVV instructions vsadd/vsaddu and vssub/vssubu respectively.
These are commonly used in image processing for clamped arithmetic
and were missing from the RISC-V intrinsic tables.

Fixes #9184

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: rajeshgangam

src/CodeGen_RISCV.cpp

@@ -62,6 +62,8 @@ struct RISCVIntrinsic {
 const RISCVIntrinsic signed_intrinsic_defs[] = {
     {"vaadd", Type::Int, "halving_add", {Type::Int, Type::Int}, AddVLArg | RoundDown | Commutes},
     {"vaadd", Type::Int, "rounding_halving_add", {Type::Int, Type::Int}, AddVLArg | RoundUp | Commutes},
+    {"vsadd", Type::Int, "saturating_add", {Type::Int, Type::Int}, AddVLArg | Commutes},
+    {"vssub", Type::Int, "saturating_sub", {Type::Int, Type::Int}, AddVLArg},
     {"vwadd", {Type::Int, 2}, "widening_add", {Type::Int, Type::Int}, AddVLArg | MangleReturnType | Commutes},
     {"vwsub", {Type::Int, 2}, "widening_sub", {Type::Int, Type::Int}, AddVLArg | MangleReturnType},
     {"vwmul", {Type::Int, 2}, "widening_mul", {Type::Int, Type::Int}, AddVLArg | MangleReturnType | Commutes},
@@ -70,6 +72,8 @@ const RISCVIntrinsic signed_intrinsic_defs[] = {
 const RISCVIntrinsic unsigned_intrinsic_defs[] = {
     {"vaaddu", {Type::UInt}, "halving_add", {Type::UInt, Type::UInt}, AddVLArg | RoundDown | Commutes},
     {"vaaddu", {Type::UInt}, "rounding_halving_add", {Type::UInt, Type::UInt}, AddVLArg | RoundUp | Commutes},
+    {"vsaddu", {Type::UInt}, "saturating_add", {Type::UInt, Type::UInt}, AddVLArg | Commutes},
+    {"vssubu", {Type::UInt}, "saturating_sub", {Type::UInt, Type::UInt}, AddVLArg},
     {"vwaddu", {Type::UInt, 2}, "widening_add", {Type::UInt, Type::UInt}, AddVLArg | MangleReturnType | Commutes},
     {"vwsubu", {Type::UInt, 2}, "widening_sub", {Type::UInt, Type::UInt}, AddVLArg | MangleReturnType},
     {"vwmulu", {Type::UInt, 2}, "widening_mul", {Type::UInt, Type::UInt}, AddVLArg | MangleReturnType | Commutes},

test/correctness/simd_op_check_riscv.cpp

@@ -96,6 +96,10 @@ class SimdOpCheckRISCV : public SimdOpCheckTest {
         check("vaaddu.vv", lanes, halving_add(u_1, u_2));
         check("vaadd.vv", lanes, rounding_halving_add(i_1, i_2));
         check("vaaddu.vv", lanes, rounding_halving_add(u_1, u_2));
+        check("vsadd.vv", lanes, saturating_add(i_1, i_2));
+        check("vsaddu.vv", lanes, saturating_add(u_1, u_2));
+        check("vssub.vv", lanes, saturating_sub(i_1, i_2));
+        check("vssubu.vv", lanes, saturating_sub(u_1, u_2));
 
         // Widening intrinsics
         if (base_bit_width < 64) {