From 5eef35ae2d128d1a16ffc6bfefd8557df511fd3d Mon Sep 17 00:00:00 2001
From: Evandro Menezes <e.menezes@samsung.com>
Date: Mon, 11 Dec 2017 21:09:27 +0000
Subject: [PATCH] [CodeGen] Improve the consistency of instruction fusion*

When either instruction in a fused pair has no other dependency, besides on
the other instruction, make sure that other instructions do not get
scheduled between them.  Additionally, avoid fusing an instruction more than
once along the same dependency chain.

Differential revision: https://reviews.llvm.org/D36704

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@320420 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/MacroFusion.cpp | 79 ++++++++++++++++++++++++++-----------
 1 file changed, 55 insertions(+), 24 deletions(-)

diff --git a/lib/CodeGen/MacroFusion.cpp b/lib/CodeGen/MacroFusion.cpp
index 13ddad59382..e7f426c469a 100644
--- a/lib/CodeGen/MacroFusion.cpp
+++ b/lib/CodeGen/MacroFusion.cpp
@@ -33,42 +33,74 @@ using namespace llvm;
 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
   cl::desc("Enable scheduling for macro fusion."), cl::init(true));
 
-static void fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
+static bool isHazard(const SDep &Dep) {
+  return Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output;
+}
+
+static bool fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
                                 SUnit &SecondSU) {
+  // Check that neither instr is already paired with another along the edge
+  // between them.
+  for (SDep &SI : FirstSU.Succs)
+    if (SI.isCluster())
+      return false;
+
+  for (SDep &SI : SecondSU.Preds)
+    if (SI.isCluster())
+      return false;
+  // Though the reachability checks above could be made more generic,
+  // perhaps as part of ScheduleDAGMI::addEdge(), since such edges are valid,
+  // the extra computation cost makes it less interesting in general cases.
+
   // Create a single weak edge between the adjacent instrs. The only effect is
   // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
-  DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster));
+  if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
+    return false;
 
-  // Adjust the latency between the anchor instr and its
-  // predecessors.
-  for (SDep &IDep : SecondSU.Preds)
-    if (IDep.getSUnit() == &FirstSU)
-      IDep.setLatency(0);
+  // Adjust the latency between both instrs.
+  for (SDep &SI : FirstSU.Succs)
+    if (SI.getSUnit() == &SecondSU)
+      SI.setLatency(0);
 
-  // Adjust the latency between the dependent instr and its
-  // predecessors.
-  for (SDep &IDep : FirstSU.Succs)
-    if (IDep.getSUnit() == &SecondSU)
-      IDep.setLatency(0);
+  for (SDep &SI : SecondSU.Preds)
+    if (SI.getSUnit() == &FirstSU)
+      SI.setLatency(0);
 
-  DEBUG(dbgs() << DAG.MF.getName() << "(): Macro fuse ";
+  DEBUG(dbgs() << "Macro fuse: ";
         FirstSU.print(dbgs(), &DAG); dbgs() << " - ";
         SecondSU.print(dbgs(), &DAG); dbgs() << " /  ";
         dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - " <<
                   DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n'; );
 
+  // Make data dependencies from the FirstSU also dependent on the SecondSU to
+  // prevent them from being scheduled between the FirstSU and the SecondSU.
   if (&SecondSU != &DAG.ExitSU)
-    // Make instructions dependent on FirstSU also dependent on SecondSU to
-    // prevent them from being scheduled between FirstSU and and SecondSU.
     for (const SDep &SI : FirstSU.Succs) {
-      if (SI.getSUnit() == &SecondSU)
+      SUnit *SU = SI.getSUnit();
+      if (SI.isWeak() || isHazard(SI) ||
+          SU == &DAG.ExitSU || SU == &SecondSU || SU->isPred(&SecondSU))
         continue;
-      DEBUG(dbgs() << "  Copy Succ ";
-            SI.getSUnit()->print(dbgs(), &DAG); dbgs() << '\n';);
-      DAG.addEdge(SI.getSUnit(), SDep(&SecondSU, SDep::Artificial));
+      DEBUG(dbgs() << "  Bind ";
+            SecondSU.print(dbgs(), &DAG); dbgs() << " - ";
+            SU->print(dbgs(), &DAG); dbgs() << '\n';);
+      DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
+    }
+
+  // Make the FirstSU also dependent on the dependencies of the SecondSU to
+  // prevent them from being scheduled between the FirstSU and the SecondSU.
+  if (&FirstSU != &DAG.EntrySU)
+    for (const SDep &SI : SecondSU.Preds) {
+      SUnit *SU = SI.getSUnit();
+      if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
+        continue;
+      DEBUG(dbgs() << "  Bind ";
+            SU->print(dbgs(), &DAG); dbgs() << " - ";
+            FirstSU.print(dbgs(), &DAG); dbgs() << '\n';);
+      DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
     }
 
   ++NumFused;
+  return true;
 }
 
 namespace {
@@ -116,9 +148,8 @@ bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
 
   // Explorer for fusion candidates among the dependencies of the anchor instr.
   for (SDep &Dep : AnchorSU.Preds) {
-    // Ignore dependencies that don't enforce ordering.
-    if (Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output ||
-        Dep.isWeak())
+    // Ignore dependencies other than data or strong ordering.
+    if (Dep.isWeak() || isHazard(Dep))
       continue;
 
     SUnit &DepSU = *Dep.getSUnit();
@@ -129,8 +160,8 @@ bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
     if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
       continue;
 
-    fuseInstructionPair(DAG, DepSU, AnchorSU);
-    return true;
+    if (fuseInstructionPair(DAG, DepSU, AnchorSU))
+      return true;
   }
 
   return false;