From b05dcd6a9af3601d3c07d73e8b563204f0a5c69b Mon Sep 17 00:00:00 2001
From: Martien de Jong <martien.de-jong@xilinx.com>
Date: Mon, 9 Dec 2024 14:54:46 +0100
Subject: [PATCH] [AIE] fix resource off-by-one glitch

---
 llvm/lib/Target/AIE/AIEPostPipeliner.cpp | 20 +++++++++++---------
 llvm/lib/Target/AIE/AIEPostPipeliner.h   |  4 ++++
 2 files changed, 15 insertions(+), 9 deletions(-)

diff --git a/llvm/lib/Target/AIE/AIEPostPipeliner.cpp b/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
index 645e65dfa956..3d99c7bb25e2 100644
--- a/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
+++ b/llvm/lib/Target/AIE/AIEPostPipeliner.cpp
@@ -221,6 +221,7 @@ void PostPipeliner::computeForward() {
       const int NewEarliest = Me.Earliest + Dep.getSignedLatency();
       SInfo.Earliest = std::max(SInfo.Earliest, NewEarliest);
     }
+    Me.Slots = getSlotCounts(*SU.getInstr(), TII);
   }
 }
 
@@ -270,24 +271,25 @@ bool PostPipeliner::computeLoopCarriedParameters() {
     /* EMPTY */;
   }
 
-  // Adjust Earliest and Latest with resource requirements
+  // Adjust Earliest and Latest with resource requirements.
+  // FIXME: We do not account for negative latencies here. This can lead to
+  // suboptimality, but we only include true dependences, where negative
+  // latencies are rare.
   for (int K = 0; K < NInstr; K++) {
     auto &Me = Info[K];
-    SlotCounts ASlots;
+    SlotCounts ASlots(Me.Slots);
     for (int A : Me.Ancestors) {
-      auto &SU = DAG->SUnits[A];
-      ASlots += getSlotCounts(*SU.getInstr(), TII);
+      ASlots += Info[A].Slots;
     }
-    SlotCounts OSlots;
+    SlotCounts OSlots(Me.Slots);
     for (int O : Me.Offspring) {
-      auto &SU = DAG->SUnits[O];
-      OSlots += getSlotCounts(*SU.getInstr(), TII);
+      OSlots += Info[O].Slots;
     }
     LLVM_DEBUG(dbgs() << "SU" << K << " : " << Info[K].Earliest << " - "
                       << Info[K].Latest << " " << ASlots << " " << OSlots
                       << "\n");
-    Me.Earliest = std::max(Me.Earliest, ASlots.max());
-    Me.Latest = std::min(Me.Latest, -OSlots.max());
+    Me.Earliest = std::max(Me.Earliest, 0 + (ASlots.max() - 1));
+    Me.Latest = std::min(Me.Latest, -1 - (OSlots.max() - 1));
     LLVM_DEBUG(dbgs() << "    -> " << Info[K].Earliest << " - "
                       << Info[K].Latest << "\n");
   }
diff --git a/llvm/lib/Target/AIE/AIEPostPipeliner.h b/llvm/lib/Target/AIE/AIEPostPipeliner.h
index e16c654dcf9e..5fa8ca8d7f49 100644
--- a/llvm/lib/Target/AIE/AIEPostPipeliner.h
+++ b/llvm/lib/Target/AIE/AIEPostPipeliner.h
@@ -15,6 +15,7 @@
 #define LLVM_LIB_TARGET_AIE_AIEPOSTPIPELINER_H
 
 #include "AIEHazardRecognizer.h"
+#include "AIESlotCounts.h"
 #include "llvm/CodeGen/MachineScheduler.h"
 #include "llvm/CodeGen/ResourceScoreboard.h"
 #include <unordered_set>
@@ -64,6 +65,9 @@ class NodeInfo {
   int StaticEarliest = 0;
   int StaticLatest = -1;
 
+  // Slots necessary for this instruction.
+  SlotCounts Slots;
+
   // Record critical path components
   // The Pred/Succ that pushed my Earliest/Latest
   std::optional<int> LastEarliestPusher;