Recommit rL305677: [CodeGen] Add generic MacroFusion pass

Use llvm::make_unique to avoid ambiguity with MSVC.

This patch adds a generic MacroFusion pass, that is used on X86 and
AArch64, which both define target-specific shouldScheduleAdjacent
functions. This generic pass should make it easier for other targets to
implement macro fusion and I intend to add macro fusion for ARM shortly.

Differential Revision: https://reviews.llvm.org/D34144

llvm-svn: 305690

llvm/include/llvm/CodeGen/MacroFusion.h

@@ -0,0 +1,41 @@
+//===- MacroFusion.h - Macro Fusion ------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains the definition of the DAG scheduling mutation to
+/// pair instructions back to back.
+//
+//===----------------------------------------------------------------------===//
+
+#include <functional>
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/CodeGen/MachineScheduler.h"
+
+namespace llvm {
+
+/// \brief Check if the instr pair, FirstMI and SecondMI, should be fused
+/// together. Given SecondMI, when FirstMI is unspecified, then check if
+/// SecondMI may be part of a fused pair at all.
+typedef std::function<bool(const TargetInstrInfo &TII,
+                           const TargetSubtargetInfo &TSI,
+                           const MachineInstr *FirstMI,
+                           const MachineInstr &SecondMI)>  ShouldSchedulePredTy;
+
+/// \brief Create a DAG scheduling mutation to pair instructions back to back
+/// for instructions that benefit according to the target-specific
+/// shouldScheduleAdjacent predicate function.
+std::unique_ptr<ScheduleDAGMutation>
+createMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent);
+
+/// \brief Create a DAG scheduling mutation to pair branch instructions with one
+/// of their predecessors back to back for instructions that benefit according
+/// to the target-specific shouldScheduleAdjacent predicate function.
+std::unique_ptr<ScheduleDAGMutation>
+createBranchMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent);
+
+} // end namespace llvm

llvm/lib/CodeGen/CMakeLists.txt

@@ -92,6 +92,7 @@ add_llvm_library(LLVMCodeGen
   PatchableFunction.cpp
   MIRPrinter.cpp
   MIRPrintingPass.cpp
+  MacroFusion.cpp
   OptimizePHIs.cpp
   ParallelCG.cpp
   PeepholeOptimizer.cpp

llvm/lib/CodeGen/MacroFusion.cpp

@@ -0,0 +1,150 @@
+//===- MacroFusion.cpp - Macro Fusion ----------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains the implementation of the DAG scheduling mutation
+/// to pair instructions back to back.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/MacroFusion.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Target/TargetInstrInfo.h"
+
+#define DEBUG_TYPE "misched"
+
+STATISTIC(NumFused, "Number of instr pairs fused");
+
+using namespace llvm;
+
+static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
+  cl::desc("Enable scheduling for macro fusion."), cl::init(true));
+
+namespace {
+
+static void fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
+                                SUnit &SecondSU) {
+  // 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));
+
+  // 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 the dependent instr and its
+  // predecessors.
+  for (SDep &IDep : FirstSU.Succs)
+    if (IDep.getSUnit() == &SecondSU)
+      IDep.setLatency(0);
+
+  DEBUG(dbgs() << DAG.MF.getName() << "(): 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'; );
+
+  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)
+        continue;
+      DEBUG(dbgs() << "  Copy Succ ";
+            SI.getSUnit()->print(dbgs(), &DAG); dbgs() << '\n';);
+      DAG.addEdge(SI.getSUnit(), SDep(&SecondSU, SDep::Artificial));
+    }
+
+  ++NumFused;
+}
+
+
+/// \brief Post-process the DAG to create cluster edges between instrs that may
+/// be fused by the processor into a single operation.
+class MacroFusion : public ScheduleDAGMutation {
+  ShouldSchedulePredTy shouldScheduleAdjacent;
+  bool FuseBlock;
+  bool scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU);
+
+public:
+  MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
+    : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
+
+  void apply(ScheduleDAGInstrs *DAGInstrs) override;
+};
+
+void MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) {
+  ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
+
+  if (FuseBlock)
+    // For each of the SUnits in the scheduling block, try to fuse the instr in
+    // it with one in its predecessors.
+    for (SUnit &ISU : DAG->SUnits)
+        scheduleAdjacentImpl(*DAG, ISU);
+
+  if (DAG->ExitSU.getInstr())
+    // Try to fuse the instr in the ExitSU with one in its predecessors.
+    scheduleAdjacentImpl(*DAG, DAG->ExitSU);
+}
+
+/// \brief Implement the fusion of instr pairs in the scheduling DAG,
+/// anchored at the instr in AnchorSU..
+bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
+  const MachineInstr &AnchorMI = *AnchorSU.getInstr();
+  const TargetInstrInfo &TII = *DAG.TII;
+  const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
+
+  // Check if the anchor instr may be fused.
+  if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
+    return false;
+
+  // 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())
+      continue;
+
+    SUnit &DepSU = *Dep.getSUnit();
+    if (DepSU.isBoundaryNode())
+      continue;
+
+    const MachineInstr *DepMI = DepSU.getInstr();
+    if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
+      continue;
+
+    fuseInstructionPair(DAG, DepSU, AnchorSU);
+    return true;
+  }
+
+  return false;
+}
+
+} // end anonymous namespace
+
+
+namespace llvm {
+
+std::unique_ptr<ScheduleDAGMutation>
+createMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent) {
+  if(EnableMacroFusion)
+    return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
+  return nullptr;
+}
+
+std::unique_ptr<ScheduleDAGMutation>
+createBranchMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent) {
+  if(EnableMacroFusion)
+    return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, false);
+  return nullptr;
+}
+
+} // end namespace llvm