diff --git a/src/main/scala/millfork/OptimizationPresets.scala b/src/main/scala/millfork/OptimizationPresets.scala
index ff3a2bb8..f25739b0 100644
--- a/src/main/scala/millfork/OptimizationPresets.scala
+++ b/src/main/scala/millfork/OptimizationPresets.scala
@@ -154,6 +154,7 @@ object OptimizationPresets {
     AlwaysGoodOptimizations.PointlessMathFromFlow,
     AlwaysGoodOptimizations.PointlessMathFromFlow,
     AlwaysGoodOptimizations.PointlessMathFromFlow,
+    LaterOptimizations.CommutativeInPlaceModifications,
     AlwaysGoodOptimizations.LoadingOfJustWrittenValue,
     AlwaysGoodOptimizations.PointlessStackStore,
     AlwaysGoodOptimizations.OptimizeZeroComparisons,
diff --git a/src/main/scala/millfork/assembly/mos/opt/LaterOptimizations.scala b/src/main/scala/millfork/assembly/mos/opt/LaterOptimizations.scala
index badad500..0dd6150e 100644
--- a/src/main/scala/millfork/assembly/mos/opt/LaterOptimizations.scala
+++ b/src/main/scala/millfork/assembly/mos/opt/LaterOptimizations.scala
@@ -544,7 +544,23 @@ object LaterOptimizations {
 
   )
 
+  val CommutativeInPlaceModifications = new RuleBasedAssemblyOptimization("Commutative in-place modifications",
+    needsFlowInfo = FlowInfoRequirement.NoRequirement,
+
+    (Elidable & HasOpcode(LDY) & HasAddrModeIn(Immediate, Absolute, ZeroPage, LongAbsolute, AbsoluteX, ZeroPageX) & MatchAddrMode(0) & MatchParameter(1)) ~
+      (Elidable & HasOpcode(LDA) & HasAddrModeIn(AbsoluteY, IndexedY, LongIndexedY)) ~
+      (Elidable & HasOpcode(LDY) & HasAddrModeIn(Immediate, Absolute, ZeroPage, LongAbsolute, AbsoluteX, ZeroPageX)) ~
+      (Elidable & HasOpcodeIn(ORA, EOR, AND) & HasAddrModeIn(AbsoluteY, IndexedY, LongIndexedY)) ~
+      (Elidable & HasOpcode(LDY) & HasAddrModeIn(Immediate, Absolute, ZeroPage, LongAbsolute, AbsoluteX, ZeroPageX) & MatchAddrMode(0) & MatchParameter(1)) ~
+      (HasOpcodeIn(INY, DEY)).* ~
+      (HasOpcode(STA) & HasAddrModeIn(AbsoluteY, IndexedY, LongIndexedY)) ~~> { code =>
+      List(code(2), code(3).copy(opcode = LDA), code(5), code(1).copy(opcode = code(3).opcode)) ++ code.drop(5)
+    },
+
+  )
+
   val All = List(
+    CommutativeInPlaceModifications,
     DontUseIndexRegisters,
     DoubleLoadToDifferentRegisters,
     DoubleLoadToTheSameRegister,
diff --git a/src/test/scala/millfork/test/AssemblyOptimizationSuite.scala b/src/test/scala/millfork/test/AssemblyOptimizationSuite.scala
index cf96707c..c0a2b8e4 100644
--- a/src/test/scala/millfork/test/AssemblyOptimizationSuite.scala
+++ b/src/test/scala/millfork/test/AssemblyOptimizationSuite.scala
@@ -845,4 +845,30 @@ class AssemblyOptimizationSuite extends FunSuite with Matchers {
     }
     EmuUndocumentedRun(code)
   }
+
+  test("Optimize commutative in-place modifications using small arrays") {
+
+    val code =
+      """
+        |
+        |array output [4] @$c000
+        |
+        |void main() {
+        |    f($c000, 1, $c000, 2)
+        |    g(1, 2)
+        |}
+        |
+        |noinline void f(pointer p1, byte i1, pointer p2, byte i2) {
+        |    p1[i1] ^= p2[i2]
+        |}
+        |
+        |noinline void g(byte i1, byte i2) {
+        |    output[i1] ^= output[i2]
+        |}
+        |
+        |""".stripMargin
+    EmuCrossPlatformBenchmarkRun(Cpu.Mos)(code) { m =>
+
+    }
+  }
 }