diff --git a/examples/crossplatform/hello_world.mfk b/examples/crossplatform/hello_world.mfk
index c4deec52..a9430943 100644
--- a/examples/crossplatform/hello_world.mfk
+++ b/examples/crossplatform/hello_world.mfk
@@ -9,7 +9,7 @@ void main() {
     ensure_mixedcase()
 
 #if CBM_64 || CBM_264
-    set_bg_color(green)
+    set_bg_color(white)
 #endif
 #if CBM_64 || CBM_264 || ZX_SPECTRUM
     set_border(red)
diff --git a/src/main/scala/millfork/Main.scala b/src/main/scala/millfork/Main.scala
index 55f7acac..3eee9423 100644
--- a/src/main/scala/millfork/Main.scala
+++ b/src/main/scala/millfork/Main.scala
@@ -5,10 +5,10 @@ import java.nio.charset.StandardCharsets
 import java.nio.file.{Files, Paths}
 import java.util.Locale
 
-import millfork.assembly.m6809.opt.M6809OptimizationPresets
+import millfork.assembly.m6809.opt.{M6809OptimizationPresets, VeryLateM6809AssemblyOptimizations}
 import millfork.assembly.mos.AssemblyLine
 import millfork.assembly.mos.opt._
-import millfork.assembly.z80.opt.Z80OptimizationPresets
+import millfork.assembly.z80.opt.{VeryLateI80AssemblyOptimizations, Z80OptimizationPresets}
 import millfork.buildinfo.BuildInfo
 import millfork.cli.{CliParser, CliStatus}
 import millfork.compiler.LabelGenerator
@@ -273,7 +273,7 @@ object Main {
 
     // compile
     val assembler = new MosAssembler(program, env, platform)
-    val result = assembler.assemble(callGraph, assemblyOptimizations, options)
+    val result = assembler.assemble(callGraph, assemblyOptimizations, options, if (optLevel <= 1) (_,_) => Nil else VeryLateMosAssemblyOptimizations.All)
     options.log.assertNoErrors("Codegen failed")
     options.log.debug(f"Unoptimized code size: ${assembler.unoptimizedCodeSize}%5d B")
     options.log.debug(f"Optimized code size:   ${assembler.optimizedCodeSize}%5d B")
@@ -313,7 +313,7 @@ object Main {
 
     // compile
     val assembler = new Z80Assembler(program, env, platform)
-    val result = assembler.assemble(callGraph, assemblyOptimizations, options)
+    val result = assembler.assemble(callGraph, assemblyOptimizations, options, if (optLevel <= 1) (_,_) => Nil else VeryLateI80AssemblyOptimizations.All)
     options.log.assertNoErrors("Codegen failed")
     options.log.debug(f"Unoptimized code size: ${assembler.unoptimizedCodeSize}%5d B")
     options.log.debug(f"Optimized code size:   ${assembler.optimizedCodeSize}%5d B")
@@ -343,7 +343,7 @@ object Main {
 
     // compile
     val assembler = new M6809Assembler(program, env, platform)
-    val result = assembler.assemble(callGraph, assemblyOptimizations, options)
+    val result = assembler.assemble(callGraph, assemblyOptimizations, options, if (optLevel <= 1) (_,_) => Nil else VeryLateM6809AssemblyOptimizations.All)
     options.log.assertNoErrors("Codegen failed")
     options.log.debug(f"Unoptimized code size: ${assembler.unoptimizedCodeSize}%5d B")
     options.log.debug(f"Optimized code size:   ${assembler.optimizedCodeSize}%5d B")
@@ -376,7 +376,7 @@ object Main {
 
     // compile
     val assembler = new Z80ToX86Crossassembler(program, env, platform)
-    val result = assembler.assemble(callGraph, assemblyOptimizations, options)
+    val result = assembler.assemble(callGraph, assemblyOptimizations, options, if (optLevel <= 1) (_,_) => Nil else VeryLateI80AssemblyOptimizations.All)
     options.log.assertNoErrors("Codegen failed")
     options.log.debug(f"Unoptimized code size: ${assembler.unoptimizedCodeSize}%5d B")
     options.log.debug(f"Optimized code size:   ${assembler.optimizedCodeSize}%5d B")
diff --git a/src/main/scala/millfork/assembly/m6809/opt/VeryLateM6809AssemblyOptimizations.scala b/src/main/scala/millfork/assembly/m6809/opt/VeryLateM6809AssemblyOptimizations.scala
new file mode 100644
index 00000000..af555b09
--- /dev/null
+++ b/src/main/scala/millfork/assembly/m6809/opt/VeryLateM6809AssemblyOptimizations.scala
@@ -0,0 +1,18 @@
+package millfork.assembly.m6809.opt
+
+import millfork.CompilationOptions
+import millfork.assembly.AssemblyOptimization
+import millfork.assembly.m6809.MLine
+import millfork.node.NiceFunctionProperty
+
+/**
+  * @author Karol Stasiak
+  */
+object VeryLateM6809AssemblyOptimizations {
+
+  def All(nice: Set[NiceFunctionProperty], options: CompilationOptions): Seq[AssemblyOptimization[MLine]] = {
+    val result = Seq.newBuilder[AssemblyOptimization[MLine]]
+    // TODO: add stuff here
+    result.result()
+  }
+}
diff --git a/src/main/scala/millfork/assembly/mos/opt/VeryLateMosAssemblyOptimizations.scala b/src/main/scala/millfork/assembly/mos/opt/VeryLateMosAssemblyOptimizations.scala
new file mode 100644
index 00000000..8162a2d8
--- /dev/null
+++ b/src/main/scala/millfork/assembly/mos/opt/VeryLateMosAssemblyOptimizations.scala
@@ -0,0 +1,59 @@
+package millfork.assembly.mos.opt
+
+import millfork.CompilationOptions
+import millfork.assembly.AssemblyOptimization
+import millfork.assembly.mos.{AssemblyLine, State}
+import millfork.compiler.mos.MosReturnDispatch
+import millfork.node.{MosNiceFunctionProperty, NiceFunctionProperty}
+import millfork.assembly.mos.Opcode._
+import millfork.assembly.mos.AddrMode._
+
+/**
+  * @author Karol Stasiak
+  */
+//noinspection ZeroIndexToHead
+object VeryLateMosAssemblyOptimizations {
+
+  val StoresOfImmediatesDifferingByOneViaX = new RuleBasedAssemblyOptimization("",
+    needsFlowInfo = FlowInfoRequirement.BothFlows,
+    (Elidable & HasOpcode(LDA) & HasAddrMode(Immediate) & MatchNumericImmediate(0) & HasIndex8) ~
+      (Elidable & HasOpcode(STA) & HasAddrModeIn(ZeroPage, Absolute) & DoesntMatterWhatItDoesWith(State.A, State.X)) ~
+      (Linear & Not(HasOpcode(LDA) & HasAddrMode(Immediate))).*.capture(10) ~
+      (Elidable & HasOpcode(LDA) & HasAddrMode(Immediate) & MatchNumericImmediate(1) & HasIndex8) ~
+      Where(ctx => ctx.get[Int](0).+(1).&(0xff) == ctx.get[Int](1)) ~
+      (Elidable & HasOpcode(STA) & HasAddrModeIn(ZeroPage, Absolute) & DoesntMatterWhatItDoesWith(State.A, State.X)) ~~> { code =>
+      code(0).copy(opcode = LDX) ::
+        code(1).copy(opcode = STX) :: (
+        code.drop(2).dropRight(2) ++ List(
+          AssemblyLine.implied(INX).pos(code(code.size - 2).source),
+          code.last.copy(opcode = STX)))
+    }
+  )
+
+  val StoresOfImmediatesDifferingByOneViaY = new RuleBasedAssemblyOptimization("",
+    needsFlowInfo = FlowInfoRequirement.BothFlows,
+    (Elidable & HasOpcode(LDA) & HasAddrMode(Immediate) & MatchNumericImmediate(0) & HasIndex8) ~
+      (Elidable & HasOpcode(STA) & HasAddrModeIn(ZeroPage, Absolute) & DoesntMatterWhatItDoesWith(State.A, State.Y)) ~
+      (Linear & Not(HasOpcode(LDA) & HasAddrMode(Immediate))).*.capture(10) ~
+      (Elidable & HasOpcode(LDA) & HasAddrMode(Immediate) & MatchNumericImmediate(1) & HasIndex8) ~
+      Where(ctx => ctx.get[Int](0).+(1).&(0xff) == ctx.get[Int](1)) ~
+      (Elidable & HasOpcode(STA) & HasAddrModeIn(ZeroPage, Absolute) & DoesntMatterWhatItDoesWith(State.A, State.Y)) ~~> { code =>
+      code(0).copy(opcode = LDY) ::
+        code(1).copy(opcode = STY) :: (
+        code.drop(2).dropRight(2) ++ List(
+          AssemblyLine.implied(INY).pos(code(code.size - 2).source),
+          code.last.copy(opcode = STY)))
+    }
+  )
+
+  def All(nice: Set[NiceFunctionProperty], options: CompilationOptions): Seq[AssemblyOptimization[AssemblyLine]] = {
+    val result = Seq.newBuilder[AssemblyOptimization[AssemblyLine]]
+    if (!nice(MosNiceFunctionProperty.DoesntChangeX)){
+      result += StoresOfImmediatesDifferingByOneViaX
+    }
+    if (!nice(MosNiceFunctionProperty.DoesntChangeY)){
+      result += StoresOfImmediatesDifferingByOneViaY
+    }
+    result.result()
+  }
+}
diff --git a/src/main/scala/millfork/assembly/z80/opt/VeryLateI80AssemblyOptimizations.scala b/src/main/scala/millfork/assembly/z80/opt/VeryLateI80AssemblyOptimizations.scala
new file mode 100644
index 00000000..c11e635f
--- /dev/null
+++ b/src/main/scala/millfork/assembly/z80/opt/VeryLateI80AssemblyOptimizations.scala
@@ -0,0 +1,19 @@
+package millfork.assembly.z80.opt
+
+import millfork.CompilationOptions
+import millfork.assembly.AssemblyOptimization
+import millfork.assembly.z80.ZLine
+import millfork.node.NiceFunctionProperty
+
+/**
+  * @author Karol Stasiak
+  */
+object VeryLateI80AssemblyOptimizations {
+
+  def All(nice: Set[NiceFunctionProperty], options: CompilationOptions): Seq[AssemblyOptimization[ZLine]] = {
+    val result = Seq.newBuilder[AssemblyOptimization[ZLine]]
+    // TODO: add stuff here
+    result.result()
+  }
+
+}
diff --git a/src/main/scala/millfork/output/AbstractAssembler.scala b/src/main/scala/millfork/output/AbstractAssembler.scala
index 01725dda..461faba0 100644
--- a/src/main/scala/millfork/output/AbstractAssembler.scala
+++ b/src/main/scala/millfork/output/AbstractAssembler.scala
@@ -195,7 +195,11 @@ abstract class AbstractAssembler[T <: AbstractCode](private val program: Program
 
   protected def subbyte(c: Constant, index: Int, totalSize: Int): Constant = if (platform.isBigEndian) c.subbyteBe(index, totalSize) else c.subbyte(index)
 
-  def assemble(callGraph: CallGraph, unfilteredOptimizations: Seq[AssemblyOptimization[T]], options: CompilationOptions): AssemblerOutput = {
+  def assemble(
+                callGraph: CallGraph,
+                unfilteredOptimizations: Seq[AssemblyOptimization[T]],
+                options: CompilationOptions,
+                veryLateOptimizations: (Set[NiceFunctionProperty], CompilationOptions) => Seq[AssemblyOptimization[T]]): AssemblerOutput = {
     mem.programName = options.outputFileName.getOrElse("MILLFORK")
     val platform = options.platform
     val variableAllocators = platform.variableAllocators
@@ -277,7 +281,14 @@ abstract class AbstractAssembler[T <: AbstractCode](private val program: Program
           case None =>
             log.trace("Not inlining " + f, function.position)
             functionsThatCanBeCalledFromInlinedFunctions += function.name
-            compiledFunctions(f) = NormalCompiledFunction(function.declaredBank.getOrElse(platform.defaultCodeBank), code, function.address.isDefined, function.alignment)
+            val thisFunctionNiceProperties = niceFunctionProperties.filter(_._2.==(f)).map(_._1).toSet
+            val labelMapImm = labelMap.toMap
+            val niceFunctionPropertiesImm = niceFunctionProperties.toSet
+            val extraOptimizedCode = veryLateOptimizations(thisFunctionNiceProperties, options).foldLeft(code) { (c, opt) =>
+              val code = opt.optimize(function, c, OptimizationContext(options, labelMapImm, env.maybeGet[ThingInMemory]("__reg"), niceFunctionPropertiesImm))
+              if (code eq c) code else quickSimplify(code)
+            }
+            compiledFunctions(f) = NormalCompiledFunction(function.declaredBank.getOrElse(platform.defaultCodeBank), extraOptimizedCode, function.address.isDefined, function.alignment)
             optimizedCodeSize += code.map(_.sizeInBytes).sum
             if (options.flag(CompilationFlag.InterproceduralOptimization)) {
               gatherNiceFunctionProperties(options, niceFunctionProperties, function, code)
diff --git a/src/test/scala/millfork/test/emu/EmuI86Run.scala b/src/test/scala/millfork/test/emu/EmuI86Run.scala
index 09641e7c..dfc53dae 100644
--- a/src/test/scala/millfork/test/emu/EmuI86Run.scala
+++ b/src/test/scala/millfork/test/emu/EmuI86Run.scala
@@ -10,6 +10,7 @@ import javax.swing.SwingUtilities
 import millfork._
 import millfork.assembly.AssemblyOptimization
 import millfork.assembly.z80.ZLine
+import millfork.assembly.z80.opt.VeryLateI80AssemblyOptimizations
 import millfork.compiler.z80.Z80Compiler
 import millfork.compiler.{CompilationContext, LabelGenerator}
 import millfork.env.{Environment, InitializedArray, InitializedMemoryVariable, NormalFunction}
@@ -135,7 +136,7 @@ class EmuI86Run(nodeOptimizations: List[NodeOptimization], assemblyOptimizations
         val env2 = new Environment(None, "", CpuFamily.I80, options)
         env2.collectDeclarations(program, options)
         val assembler = new Z80ToX86Crossassembler(program, env2, platform)
-        val output = assembler.assemble(callGraph, assemblyOptimizations, options)
+        val output = assembler.assemble(callGraph, assemblyOptimizations, options, VeryLateI80AssemblyOptimizations.All)
         println(";;; compiled: -----------------")
         output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("////; DISCARD_")).foreach(println)
         println(";;; ---------------------------")
diff --git a/src/test/scala/millfork/test/emu/EmuM6809Run.scala b/src/test/scala/millfork/test/emu/EmuM6809Run.scala
index c5686d08..0a8f00b8 100644
--- a/src/test/scala/millfork/test/emu/EmuM6809Run.scala
+++ b/src/test/scala/millfork/test/emu/EmuM6809Run.scala
@@ -7,6 +7,7 @@ import fastparse.core.Parsed.{Failure, Success}
 import millfork._
 import millfork.assembly.AssemblyOptimization
 import millfork.assembly.m6809.MLine
+import millfork.assembly.m6809.opt.VeryLateM6809AssemblyOptimizations
 import millfork.compiler.m6809.M6809Compiler
 import millfork.compiler.{CompilationContext, LabelGenerator}
 import millfork.env.{Environment, InitializedArray, InitializedMemoryVariable, NormalFunction}
@@ -144,7 +145,7 @@ class EmuM6809Run(cpu: millfork.Cpu.Value, nodeOptimizations: List[NodeOptimizat
         val env2 = new Environment(None, "", CpuFamily.M6502, options)
         env2.collectDeclarations(program, options)
         val assembler = new M6809Assembler(program, env2, platform)
-        val output = assembler.assemble(callGraph, assemblyOptimizations, options)
+        val output = assembler.assemble(callGraph, assemblyOptimizations, options, VeryLateM6809AssemblyOptimizations.All)
         println(";;; compiled: -----------------")
         output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("; DISCARD_")).foreach(println)
         println(";;; ---------------------------")
diff --git a/src/test/scala/millfork/test/emu/EmuRun.scala b/src/test/scala/millfork/test/emu/EmuRun.scala
index a882fb01..8b98d9e8 100644
--- a/src/test/scala/millfork/test/emu/EmuRun.scala
+++ b/src/test/scala/millfork/test/emu/EmuRun.scala
@@ -9,6 +9,7 @@ import com.loomcom.symon.{Bus, Cpu, CpuState}
 import fastparse.core.Parsed.{Failure, Success}
 import millfork.assembly.AssemblyOptimization
 import millfork.assembly.mos.AssemblyLine
+import millfork.assembly.mos.opt.VeryLateMosAssemblyOptimizations
 import millfork.compiler.{CompilationContext, LabelGenerator}
 import millfork.compiler.mos.MosCompiler
 import millfork.env.{Environment, InitializedArray, InitializedMemoryVariable, NormalFunction}
@@ -221,7 +222,7 @@ class EmuRun(cpu: millfork.Cpu.Value, nodeOptimizations: List[NodeOptimization],
         val env2 = new Environment(None, "", CpuFamily.M6502, options)
         env2.collectDeclarations(program, options)
         val assembler = new MosAssembler(program, env2, platform)
-        val output = assembler.assemble(callGraph, assemblyOptimizations, options)
+        val output = assembler.assemble(callGraph, assemblyOptimizations, options, VeryLateMosAssemblyOptimizations.All)
         println(";;; compiled: -----------------")
         output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("; DISCARD_")).foreach(println)
         println(";;; ---------------------------")
diff --git a/src/test/scala/millfork/test/emu/EmuZ80Run.scala b/src/test/scala/millfork/test/emu/EmuZ80Run.scala
index 55675867..7e22c29f 100644
--- a/src/test/scala/millfork/test/emu/EmuZ80Run.scala
+++ b/src/test/scala/millfork/test/emu/EmuZ80Run.scala
@@ -18,6 +18,7 @@ import millfork.node.opt.NodeOptimization
 import millfork.output.{MemoryBank, Z80Assembler}
 import millfork.parser.{MosParser, PreprocessingResult, Preprocessor, Z80Parser}
 import millfork._
+import millfork.assembly.z80.opt.VeryLateI80AssemblyOptimizations
 import millfork.compiler.z80.Z80Compiler
 import org.scalatest.Matchers
 
@@ -143,7 +144,7 @@ class EmuZ80Run(cpu: millfork.Cpu.Value, nodeOptimizations: List[NodeOptimizatio
         val env2 = new Environment(None, "", CpuFamily.I80, options)
         env2.collectDeclarations(program, options)
         val assembler = new Z80Assembler(program, env2, platform)
-        val output = assembler.assemble(callGraph, assemblyOptimizations, options)
+        val output = assembler.assemble(callGraph, assemblyOptimizations, options, VeryLateI80AssemblyOptimizations.All)
         println(";;; compiled: -----------------")
         output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("////; DISCARD_")).foreach(println)
         println(";;; ---------------------------")
diff --git a/src/test/scala/millfork/test/emu/ShouldNotCompile.scala b/src/test/scala/millfork/test/emu/ShouldNotCompile.scala
index 53040acd..03e585a6 100644
--- a/src/test/scala/millfork/test/emu/ShouldNotCompile.scala
+++ b/src/test/scala/millfork/test/emu/ShouldNotCompile.scala
@@ -95,17 +95,17 @@ object ShouldNotCompile extends Matchers {
           cpuFamily match {
             case CpuFamily.M6502 =>
               val assembler = new MosAssembler(program, env2, platform)
-              val output = assembler.assemble(callGraph, Nil, options)
+              val output = assembler.assemble(callGraph, Nil, options, (_, _) => Nil)
               output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("; DISCARD_")).foreach(println)
               fail("Failed: Compilation succeeded for 6502")
             case CpuFamily.I80 =>
               val assembler = new Z80Assembler(program, env2, platform)
-              val output = assembler.assemble(callGraph, Nil, options)
+              val output = assembler.assemble(callGraph, Nil, options, (_, _) => Nil)
               output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("; DISCARD_")).foreach(println)
               fail("Failed: Compilation succeeded for Z80")
             case CpuFamily.M6809 =>
               val assembler = new M6809Assembler(program, env2, platform)
-              val output = assembler.assemble(callGraph, Nil, options)
+              val output = assembler.assemble(callGraph, Nil, options, (_, _) => Nil)
               output.asm.takeWhile(s => !(s.startsWith(".") && s.contains("= $"))).filterNot(_.contains("; DISCARD_")).foreach(println)
               fail("Failed: Compilation succeeded for 6809")
             case _ =>