diff --git a/src/main/scala/tilelink/Coalescing.scala b/src/main/scala/tilelink/Coalescing.scala
index 0cb7c91..2590884 100644
--- a/src/main/scala/tilelink/Coalescing.scala
+++ b/src/main/scala/tilelink/Coalescing.scala
@@ -216,10 +216,10 @@ class CoalShiftQueue[T <: Data]( gen: T,
   }
 
   when(io.queue.enq.fire) {
-    when(!io.queue.deq.fire) {
+    when(!shift) {
       used := (used << 1.U) | 1.U
     }
-  }.elsewhen(io.queue.deq.fire) {
+  }.elsewhen(shift) {
     used := used >> 1.U
   }
 
@@ -229,6 +229,7 @@ class CoalShiftQueue[T <: Data]( gen: T,
 
   assert(!flow, "flow-through is not implemented")
   if (flow) {
+    // FIXME old code
     when(io.queue.enq.valid) { io.queue.deq.valid := true.B }
     when(!valid(0)) { io.queue.deq.bits := io.queue.enq.bits }
   }
diff --git a/src/test/scala/coalescing/CoalescingUnitTest.scala b/src/test/scala/coalescing/CoalescingUnitTest.scala
index 66f6b9f..41017e5 100644
--- a/src/test/scala/coalescing/CoalescingUnitTest.scala
+++ b/src/test/scala/coalescing/CoalescingUnitTest.scala
@@ -375,9 +375,10 @@ class CoalShiftQueueTest extends AnyFlatSpec with ChiselScalatestTester {
     }
   }
 
-  it should "dequeue invalidated entries by itself" in {
+  it should "dequeue invalidated head on its own when allowShift" in {
     test(new CoalShiftQueue(gen = UInt(8.W), entries = 4)) { c =>
       c.io.invalidate.valid.poke(false.B)
+
       c.io.allowShift.poke(true.B)
 
       // prepare
@@ -399,19 +400,33 @@ class CoalShiftQueueTest extends AnyFlatSpec with ChiselScalatestTester {
       // invalidate two entries at head
       c.io.invalidate.valid.poke(true.B)
       c.io.invalidate.bits.poke(0x3.U)
+      c.io.queue.deq.ready.poke(false.B)
       // [ 0x56 | 0x34(inv) | 0x12(inv) ]
       c.clock.step()
-      // [ 0x56 | 0x34(inv) ]
+      //             [ 0x56 | 0x34(inv) ]
       c.io.invalidate.valid.poke(false.B)
       c.io.queue.deq.ready.poke(false.B)
       c.clock.step()
-      // [ 0x56 ]
+      //                         [ 0x56 ]
       c.io.queue.deq.ready.poke(true.B)
       c.io.queue.deq.valid.expect(true.B)
       c.io.queue.deq.bits.expect(0x56.U)
       c.clock.step()
       c.io.queue.deq.ready.poke(true.B)
       c.io.queue.deq.valid.expect(false.B)
+      c.clock.step()
+
+      // do one more enqueue-then-dequeue to see if used bit was properly cleared
+      c.io.queue.deq.ready.poke(false.B)
+      c.io.queue.enq.ready.expect(true.B)
+      c.io.queue.enq.valid.poke(true.B)
+      c.io.queue.enq.bits.poke(0x78.U)
+      c.clock.step()
+      // should dequeue right away
+      c.io.queue.enq.valid.poke(false.B)
+      c.io.queue.deq.ready.poke(true.B)
+      c.io.queue.deq.valid.expect(true.B)
+      c.io.queue.deq.bits.expect(0x78.U)
     }
   }