From 6ca22a39e0482655a0115d405d7297bdc4d8e159 Mon Sep 17 00:00:00 2001
From: Hansung Kim <hansung_kim@berkeley.edu>
Date: Fri, 31 Mar 2023 19:40:10 -0700
Subject: [PATCH] Implement automatic dequeuing of invalid entries

---
 src/main/scala/tilelink/Coalescing.scala |  5 +++--
 src/test/scala/CoalescingUnitTest.scala  | 18 +++++++++++++++---
 2 files changed, 18 insertions(+), 5 deletions(-)

diff --git a/src/main/scala/tilelink/Coalescing.scala b/src/main/scala/tilelink/Coalescing.scala
index 3504061..95baf88 100644
--- a/src/main/scala/tilelink/Coalescing.scala
+++ b/src/main/scala/tilelink/Coalescing.scala
@@ -495,10 +495,11 @@ class CoalShiftQueue[T <: Data](
     else if (i == entries) false.B
     else Mux(io.invalidate(i), false.B, paddedValid(i))
 
+  val shift = io.deq.ready || (used =/= 0.U) && !valid(0)
   for (i <- 0 until entries) {
     val wdata = if (i == entries - 1) io.enq.bits else Mux(!used(i + 1), io.enq.bits, elts(i + 1))
     val wen = Mux(
-      io.deq.ready,
+      shift,
       (io.enq.fire && !paddedUsed(i + 1) && used(i)) || paddedValidAfterInvalidate(i + 1),
       // enqueue to the first empty slot above the top
       (io.enq.fire && paddedUsed(i - 1) && !used(i)) || !paddedValidAfterInvalidate(i)
@@ -506,7 +507,7 @@ class CoalShiftQueue[T <: Data](
     when(wen) { elts(i) := wdata }
 
     valid(i) := Mux(
-      io.deq.ready,
+      shift,
       (io.enq.fire && !paddedUsed(i + 1) && used(i)) || paddedValidAfterInvalidate(i + 1),
       // TODO: handle enqueueing to invalidated tail?
       (io.enq.fire && paddedUsed(i - 1) && !used(i)) || paddedValidAfterInvalidate(i)
diff --git a/src/test/scala/CoalescingUnitTest.scala b/src/test/scala/CoalescingUnitTest.scala
index d18041b..c267998 100644
--- a/src/test/scala/CoalescingUnitTest.scala
+++ b/src/test/scala/CoalescingUnitTest.scala
@@ -158,16 +158,28 @@ class CoalShiftQueueTest extends AnyFlatSpec with ChiselScalatestTester {
       c.io.enq.valid.poke(true.B)
       c.io.enq.bits.poke(0x34.U)
       c.clock.step()
+      c.io.deq.ready.poke(false.B)
+      c.io.enq.ready.expect(true.B)
+      c.io.enq.valid.poke(true.B)
+      c.io.enq.bits.poke(0x56.U)
+      c.clock.step()
       c.io.enq.valid.poke(false.B)
 
-      c.io.invalidate.poke(0x1.U)
+      // invalidate two entries at head
+      c.io.invalidate.poke(0x3.U)
       c.clock.step()
+      c.io.invalidate.poke(0x0.U)
       c.io.deq.ready.poke(false.B)
-      // 0x12 should be dequeued
+      // 0x12 should be dequeued now
+      c.clock.step()
+      // 0x34 should be dequeued now
       c.clock.step()
       c.io.deq.ready.poke(true.B)
       c.io.deq.valid.expect(true.B)
-      c.io.deq.bits.expect(0x34.U)
+      c.io.deq.bits.expect(0x56.U)
+      c.clock.step()
+      c.io.deq.ready.poke(true.B)
+      c.io.deq.valid.expect(false.B)
     }
   }
 }