[interp] Add MINT_STELEM_VT_NOREF (dotnet#87161)

Based on instrumentation 0.2% of all mono_gc_wbarrier_value_copy_internal calls in a run of System.Runtime.Tests are from MINT_STELEM_VT. I suspect adding this NOREF version of the opcode will improve performance for data structures that store structs in arrays but don't contain any references - for example List<ValueTuple<...>> or Dictionary<int, int> - but it's hard to measure locally.

Author: kg

src/mono/mono/mini/interp/interp.c

@@ -6505,6 +6505,19 @@ MINT_IN_CASE(MINT_BRTRUE_I8_SP) ZEROP_SP(gint64, !=); MINT_IN_BREAK;
 			ip += 6;
 			MINT_IN_BREAK;
 		}
+		MINT_IN_CASE(MINT_STELEM_VT_NOREF) {
+			MonoArray *o = LOCAL_VAR (ip [1], MonoArray*);
+			NULL_CHECK (o);
+			guint32 aindex = LOCAL_VAR (ip [2], guint32);
+			if (aindex >= mono_array_length_internal (o))
+				THROW_EX (interp_get_exception_index_out_of_range (frame, ip), ip);
+
+			guint16 size = ip [5];
+			char *dst_addr = mono_array_addr_with_size_fast ((MonoArray *) o, size, aindex);
+			memcpy (dst_addr, locals + ip [3], size);
+			ip += 6;
+			MINT_IN_BREAK;
+		}
 		MINT_IN_CASE(MINT_CONV_OVF_I4_U4) {
 			gint32 val = LOCAL_VAR (ip [2], gint32);
 			if (val < 0)

src/mono/mono/mini/interp/mintops.def

@@ -419,6 +419,7 @@ OPDEF(MINT_STELEM_R4, "stelem.r4", 4, 0, 3, MintOpNoArgs)
 OPDEF(MINT_STELEM_R8, "stelem.r8", 4, 0, 3, MintOpNoArgs)
 OPDEF(MINT_STELEM_REF, "stelem.ref", 4, 0, 3, MintOpNoArgs)
 OPDEF(MINT_STELEM_VT, "stelem.vt", 6, 0, 3, MintOpTwoShorts)
+OPDEF(MINT_STELEM_VT_NOREF, "stelem.vt.noref", 6, 0, 3, MintOpTwoShorts)
 
 OPDEF(MINT_LDLEN, "ldlen", 3, 1, 1, MintOpNoArgs)
 

src/mono/mono/mini/interp/transform.c

@@ -7095,7 +7095,7 @@ generate_code (TransformData *td, MonoMethod *method, MonoMethodHeader *header,
 					int size = mono_class_value_size (klass, NULL);
 					g_assert (size < G_MAXUINT16);
 
-					handle_stelem (td, MINT_STELEM_VT);
+					handle_stelem (td, m_class_has_references (klass) ? MINT_STELEM_VT : MINT_STELEM_VT_NOREF);
 					td->last_ins->data [0] = get_data_item_index (td, klass);
 					td->last_ins->data [1] = GINT_TO_UINT16 (size);
 					break;

src/mono/wasm/runtime/jiterpreter-trace-generator.ts

@@ -3316,6 +3316,15 @@ function emit_arrayop(builder: WasmBuilder, frame: NativePointer, ip: MintOpcode
             builder.callImport("value_copy");
             return true;
         }
+        case MintOpcode.MINT_STELEM_VT_NOREF: {
+            const elementSize = getArgU16(ip, 5);
+            // dest
+            append_getelema1(builder, ip, objectOffset, indexOffset, elementSize);
+            // src
+            append_ldloca(builder, valueOffset, 0);
+            append_memmove_dest_src(builder, elementSize);
+            return true;
+        }
         default:
             return false;
     }