Have mono handle the vector as APIs that grow or shrink the vector type

tannergooding · tannergooding · commit 7b3f1325a03c · 2024-07-04T11:24:57.000-07:00
diff --git a/src/mono/mono/mini/interp/interp-internals.h b/src/mono/mono/mini/interp/interp-internals.h
@@ -31,6 +31,8 @@
 #define MINT_STACK_ALIGNMENT (2 * MINT_STACK_SLOT_SIZE)
 #define MINT_SIMD_ALIGNMENT (MINT_STACK_ALIGNMENT)
 #define SIZEOF_V128 16
+#define SIZEOF_V2 8
+#define SIZEOF_V3 12
 
 #define INTERP_STACK_SIZE (1024*1024)
 #define INTERP_REDZONE_SIZE (8*1024)
diff --git a/src/mono/mono/mini/interp/interp-simd-intrins.def b/src/mono/mono/mini/interp/interp-simd-intrins.def
@@ -71,6 +71,12 @@ INTERP_SIMD_INTRINSIC_P_PP (INTERP_SIMD_INTRINSIC_V128_R4_MULTIPLY, interp_v128_
 INTERP_SIMD_INTRINSIC_P_PP (INTERP_SIMD_INTRINSIC_V128_R4_DIVISION, interp_v128_r4_op_division, 231)
 
 INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V128_BITCAST, interp_v128_bitcast, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V128_TO_V2, interp_v128_to_v2, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V128_TO_V3, interp_v128_to_v3, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V2_TO_V128, interp_v2_to_v128, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V2_TO_V3, interp_v2_to_v3, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V3_TO_V128, interp_v3_to_v128, -1)
+INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V3_TO_V2, interp_v3_to_v2, -1)
 
 INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V128_I1_NEGATION, interp_v128_i1_op_negation, 97)
 INTERP_SIMD_INTRINSIC_P_P (INTERP_SIMD_INTRINSIC_V128_I2_NEGATION, interp_v128_i2_op_negation, 129)
diff --git a/src/mono/mono/mini/interp/interp-simd.c b/src/mono/mono/mini/interp/interp-simd.c
@@ -35,6 +35,58 @@ interp_v128_bitcast (gpointer res, gpointer v1)
 	*(v128_i1*)res = *(v128_i1*)v1;
 }
 
+// Vector2 AsVector2(Vector128<float> v1)
+static void
+interp_v128_to_v2 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V2);
+}
+
+// Vector3 AsVector3(Vector128<float> v1)
+static void
+interp_v128_to_v3 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V3);
+}
+
+// Vector128<float> AsVector128(Vector2 v1)
+static void
+interp_v2_to_v128 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V2);
+
+	guint32 *res_typed = (guint32*)res;
+	res_typed [2] = 0;
+	res_typed [3] = 0;
+}
+
+// Vector3 AsVector3(Vector2 v1)
+static void
+interp_v2_to_v3 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V2);
+
+	guint32 *res_typed = (guint32*)res;
+	res_typed [2] = 0;
+}
+
+// Vector128<float> AsVector128(Vector3 v1)
+static void
+interp_v3_to_v128 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V3);
+
+	guint32 *res_typed = (guint32*)res;
+	res_typed [3] = 0;
+}
+
+// Vector2 AsVector128(Vector3 v1)
+static void
+interp_v3_to_v2 (gpointer res, gpointer v1)
+{
+	memcpy(res, v1, SIZEOF_V2);
+}
+
 // op_Addition
 static void
 interp_v128_i1_op_addition (gpointer res, gpointer v1, gpointer v2)
diff --git a/src/mono/mono/mini/interp/simd-methods.def b/src/mono/mono/mini/interp/simd-methods.def
@@ -39,8 +39,11 @@ SIMD_METHOD(AsUInt16)
 SIMD_METHOD(AsUInt32)
 SIMD_METHOD(AsUInt64)
 SIMD_METHOD(AsVector)
-SIMD_METHOD(AsVector4)
 SIMD_METHOD(AsVector128)
+SIMD_METHOD(AsVector128Unsafe)
+SIMD_METHOD(AsVector2)
+SIMD_METHOD(AsVector3)
+SIMD_METHOD(AsVector4)
 SIMD_METHOD(ConditionalSelect)
 SIMD_METHOD(Create)
 SIMD_METHOD(CreateScalar)
diff --git a/src/mono/mono/mini/interp/transform-simd.c b/src/mono/mono/mini/interp/transform-simd.c
@@ -72,8 +72,11 @@ static guint16 sri_vector128_methods [] = {
 	SN_AsUInt32,
 	SN_AsUInt64,
 	SN_AsVector,
-	SN_AsVector4,
 	SN_AsVector128,
+	SN_AsVector128Unsafe,
+	SN_AsVector2,
+	SN_AsVector3,
+	SN_AsVector4,
 	SN_ConditionalSelect,
 	SN_Create,
 	SN_CreateScalar,
@@ -470,6 +473,9 @@ emit_sri_vector128 (TransformData *td, MonoMethod *cmethod, MonoMethodSignature
 		}
 		case SN_AsVector:
 		case SN_AsVector128:
+		case SN_AsVector128Unsafe:
+		case SN_AsVector2:
+		case SN_AsVector3:
 		case SN_AsVector4: {
 			if (!is_element_type_primitive (csignature->ret) || !is_element_type_primitive (csignature->params [0]))
 				return FALSE;
@@ -485,7 +491,42 @@ emit_sri_vector128 (TransformData *td, MonoMethod *cmethod, MonoMethodSignature
 				simd_intrins = INTERP_SIMD_INTRINSIC_V128_BITCAST;
 				break;
 			}
-			return FALSE;
+			
+			if ((ret_size != 8) && (ret_size != 12) && (ret_size != 16)) {
+				return FALSE;
+			}
+
+			if ((arg_size != 8) && (arg_size != 12) && (arg_size != 16)) {
+				return FALSE;
+			}
+
+			if (arg_size > ret_size) {
+				simd_opcode = MINT_SIMD_INTRINS_P_P;
+
+				if (ret_size == 8) {
+					if (arg_size == 16) {
+						simd_intrins = INTERP_SIMD_INTRINSIC_V128_TO_V2;
+					} else {
+						simd_intrins = INTERP_SIMD_INTRINSIC_V3_TO_V2;
+					}
+				} else {
+					simd_intrins = INTERP_SIMD_INTRINSIC_V128_TO_V3;
+				}
+				break;
+			} else {
+				simd_opcode = MINT_SIMD_INTRINS_P_P;
+
+				if (arg_size == 8) {
+					if (ret_size == 12) {
+						simd_intrins = INTERP_SIMD_INTRINSIC_V2_TO_V3;
+					} else {
+						simd_intrins = INTERP_SIMD_INTRINSIC_V2_TO_V128;
+					}
+				} else {
+					simd_intrins = INTERP_SIMD_INTRINSIC_V3_TO_V128;
+				}
+				break;
+			}
 		}
 		case SN_ConditionalSelect:
 			simd_opcode = MINT_SIMD_INTRINS_P_PPP;
diff --git a/src/mono/mono/mini/simd-intrinsics.c b/src/mono/mono/mini/simd-intrinsics.c
@@ -1196,6 +1196,9 @@ static guint16 sri_vector_methods [] = {
 	SN_AsUInt64,
 	SN_AsVector,
 	SN_AsVector128,
+	SN_AsVector128Unsafe,
+	SN_AsVector2,
+	SN_AsVector3,
 	SN_AsVector4,
 	SN_BitwiseAnd,
 	SN_BitwiseOr,
@@ -1640,7 +1643,11 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 	}
 	case SN_AsVector:
 	case SN_AsVector128:
-	case SN_AsVector4: {
+	case SN_AsVector128Unsafe:
+	case SN_AsVector2:
+	case SN_AsVector3:
+	case SN_AsVector4:
+	case SN_AsVector4Unsafe: {
 		if (!is_element_type_primitive (fsig->ret) || !is_element_type_primitive (fsig->params [0]))
 			return NULL;
 
@@ -1650,10 +1657,47 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 		MonoClass *arg_class = mono_class_from_mono_type_internal (fsig->params [0]);
 		int arg_size = mono_class_value_size (arg_class, NULL);
 
-		if (arg_size == ret_size)
+		if (arg_size == ret_size) {
 			return emit_simd_ins (cfg, klass, OP_XCAST, args [0]->dreg, -1);
+		}
 
-		return NULL;
+		if ((ret_size != 8) && (ret_size != 12) && (ret_size != 16)) {
+			return NULL;
+		}
+
+		if ((arg_size != 8) && (arg_size != 12) && (arg_size != 16)) {
+			return NULL;
+		}
+
+		bool isUnsafe = (id == SN_AsVector128Unsafe) || (id == SN_AsVector4Unsafe);
+
+		if (arg_size > ret_size) {
+#ifdef TARGET_ARM64
+			if (ret_size == 8) {
+				return emit_simd_ins_for_sig (cfg, klass, OP_XLOWER, 0, arg0_type, fsig, args);
+			}
+#endif
+			return emit_simd_ins (cfg, klass, OP_XCAST, args [0]->dreg, -1);
+		} else {
+#ifdef TARGET_ARM64
+			if (arg_size == 8) {
+				int op = isUnsafe ? OP_XWIDEN : OP_XWIDEN_UNSAFE;
+				return emit_simd_ins_for_sig (cfg, klass, op, 0, arg0_type, fsig, args);
+			}
+#endif
+			MonoInst *ins  = args [0];
+
+			if (!isUnsafe) {
+				MonoInst *zero = emit_xzero (cfg, klass);
+				if (arg_size == 8) {
+					ins = emit_vector_insert_element (cfg, klass, ins, MONO_TYPE_R4, zero, 2, FALSE);
+				}
+				if (ret_size == 16) {
+					ins = emit_vector_insert_element (cfg, klass, ins, MONO_TYPE_R4, zero, 3, FALSE);
+				}
+			}
+			return emit_simd_ins (cfg, klass, OP_XCAST, ins->dreg, -1);
+		}
 	}
 	case SN_Ceiling:
 	case SN_Floor: {
diff --git a/src/mono/mono/mini/simd-methods.h b/src/mono/mono/mini/simd-methods.h
@@ -82,10 +82,12 @@ METHOD(AsUInt32)
 METHOD(AsUInt64)
 METHOD(AsVector)
 METHOD(AsVector128)
+METHOD(AsVector128Unsafe)
 METHOD(AsVector2)
 METHOD(AsVector256)
 METHOD(AsVector3)
 METHOD(AsVector4)
+METHOD(AsVector4Unsafe)
 METHOD(BitwiseAnd)
 METHOD(BitwiseOr)
 METHOD(Create)
