diff --git a/bsnes/sfc/cpu/cpu.hpp b/bsnes/sfc/cpu/cpu.hpp
index 5542b053..54db38c8 100644
--- a/bsnes/sfc/cpu/cpu.hpp
+++ b/bsnes/sfc/cpu/cpu.hpp
@@ -122,6 +122,9 @@ private:
 
     uint2 autoJoypadPort0 = 0;
     uint2 autoJoypadPort1 = 0;
+
+    bool cpuLatch = false;
+    bool autoJoypadLatch = false;
   } status;
 
   struct IO {
diff --git a/bsnes/sfc/cpu/io.cpp b/bsnes/sfc/cpu/io.cpp
index 440085f5..a4d9a5ec 100644
--- a/bsnes/sfc/cpu/io.cpp
+++ b/bsnes/sfc/cpu/io.cpp
@@ -130,18 +130,20 @@ auto CPU::writeCPU(uint addr, uint8 data) -> void {
     //bit 0 is shared between JOYSER0 and JOYSER1:
     //strobing $4016.d0 affects both controller port latches.
     //$4017 bit 0 writes are ignored.
-    controllerPort1.device->latch(data & 1);
-    controllerPort2.device->latch(data & 1);
+    status.cpuLatch = data & 1;
+    controllerPort1.device->latch(status.autoJoypadLatch | status.cpuLatch);
+    controllerPort2.device->latch(status.autoJoypadLatch | status.cpuLatch);
     return;
 
   case 0x4200:  //NMITIMEN
     io.autoJoypadPoll = data & 1;
-    if(status.autoJoypadCounter < 2) {
+    if(status.autoJoypadCounter == 0) {
       // allow controller latches during this time
-      controllerPort1.device->latch(io.autoJoypadPoll);
-      controllerPort2.device->latch(io.autoJoypadPoll);
-    }else if (!io.autoJoypadPoll) {
-      status.autoJoypadCounter = 33; // Disable auto-joypad read
+      status.autoJoypadLatch = io.autoJoypadPoll;
+      controllerPort1.device->latch(status.autoJoypadLatch | status.cpuLatch);
+      controllerPort2.device->latch(status.autoJoypadLatch | status.cpuLatch);
+    } else if (!io.autoJoypadPoll && status.autoJoypadCounter >= 2) {
+      status.autoJoypadCounter = 33;
     }
 
     nmitimenUpdate(data);
diff --git a/bsnes/sfc/cpu/serialization.cpp b/bsnes/sfc/cpu/serialization.cpp
index 556e9a04..edeff5c9 100644
--- a/bsnes/sfc/cpu/serialization.cpp
+++ b/bsnes/sfc/cpu/serialization.cpp
@@ -48,6 +48,9 @@ auto CPU::serialize(serializer& s) -> void {
   s.integer(status.autoJoypadPort0);
   s.integer(status.autoJoypadPort1);
 
+  s.boolean(status.cpuLatch);
+  s.boolean(status.autoJoypadLatch);
+
   s.integer(io.wramAddress);
 
   s.boolean(io.hirqEnable);
diff --git a/bsnes/sfc/cpu/timing.cpp b/bsnes/sfc/cpu/timing.cpp
index 2273bac6..a325fa05 100644
--- a/bsnes/sfc/cpu/timing.cpp
+++ b/bsnes/sfc/cpu/timing.cpp
@@ -214,15 +214,9 @@ auto CPU::joypadEdge() -> void {
 
   if(status.autoJoypadCounter == 0) {
     //latch controller states on the first polling cycle
-    controllerPort1.device->latch(io.autoJoypadPoll);
-    controllerPort2.device->latch(io.autoJoypadPoll);
-  }
-
-  if(status.autoJoypadCounter == 1) {
-    //release latch and begin reading on the second cycle
-    controllerPort1.device->latch(0);
-    controllerPort2.device->latch(0);
-
+    status.autoJoypadLatch = io.autoJoypadPoll;
+    controllerPort1.device->latch(status.autoJoypadLatch | status.cpuLatch);
+    controllerPort2.device->latch(status.autoJoypadLatch | status.cpuLatch);
     if(io.autoJoypadPoll) {
       //shift registers are cleared to zero at start of auto-joypad polling
       io.joy1 = 0;
@@ -232,13 +226,20 @@ auto CPU::joypadEdge() -> void {
     }
   }
 
-  if(status.autoJoypadCounter >= 2) {
-    if (!io.autoJoypadPoll) {
-      // if auto-joypad polling is disabled at this point skip the rest of the polling
-      status.autoJoypadCounter = 33;
-      return;
-    }
+  if(status.autoJoypadCounter == 1) {
+    //release latch and begin reading on the second cycle
+    status.autoJoypadLatch = 0;
+    controllerPort1.device->latch(status.autoJoypadLatch | status.cpuLatch);
+    controllerPort2.device->latch(status.autoJoypadLatch | status.cpuLatch);
+  }
 
+  if(status.autoJoypadCounter != 1 && !io.autoJoypadPoll) {
+    // if auto-joypad polling is disabled at this point skip the rest of the polling
+    status.autoJoypadCounter = 33;
+    return;
+  }
+
+  if(status.autoJoypadCounter >= 2) {
     //sixteen bits are shifted into joy{1-4}, one bit per 256 clocks
     //the bits are read on one 128-clock cycle and written on the next
     if ((status.autoJoypadCounter & 1) == 0) {