[turbofan] First version of loop peeling.

src/compiler/loop-peeling.cc

Index: src/compiler/loop-peeling.cc
diff --git a/src/compiler/loop-peeling.cc b/src/compiler/loop-peeling.cc
new file mode 100644
index 0000000000000000000000000000000000000000..28a78eb41b13807ff7ad55899c16e97299839c32
--- /dev/null
+++ b/src/compiler/loop-peeling.cc
@@ -0,0 +1,252 @@
+// Copyright 2015 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/compiler/common-operator.h"
+#include "src/compiler/graph.h"
+#include "src/compiler/loop-peeling.h"
+#include "src/compiler/node.h"
+#include "src/compiler/node-marker.h"
+#include "src/compiler/node-properties-inl.h"
+#include "src/zone.h"
+
+namespace v8 {
+namespace internal {
+namespace compiler {
+
+struct Peeling {
+  // Maps a node to its index in the {pairs} vector.
+  NodeMarker<size_t> node_map;
+  // The vector which contains the mapped nodes.
+  NodeVector* pairs;
+
+  Peeling(Graph* graph, Zone* tmp_zone, size_t max, NodeVector* p)
+      : node_map(graph, max), pairs(p) {}
+
+  Node* map(Node* node) {
+    if (node_map.Get(node) == 0) return node;
+    return pairs->at(node_map.Get(node));
+  }
+
+  void Insert(Node* original, Node* copy) {
+    node_map.Set(original, 1 + pairs->size());
+    pairs->push_back(original);
+    pairs->push_back(copy);
+  }
+
+  void CopyNodes(Graph* graph, Zone* tmp_zone, Node* dead, NodeRange nodes) {
+    NodeVector inputs(tmp_zone);
+    // Copy all the nodes first.
+    for (Node* node : nodes) {
+      inputs.clear();
+      for (Node* input : node->inputs()) inputs.push_back(map(input));
+      Insert(node, graph->NewNode(node->op(), node->InputCount(), &inputs[0]));
+    }
+
+    // Fix remaining inputs of the copies.
+    for (Node* original : nodes) {
+      Node* copy = pairs->at(node_map.Get(original));
+      for (int i = 0; i < copy->InputCount(); i++) {
+        copy->ReplaceInput(i, map(original->InputAt(i)));
+      }
+    }
+  }
+
+  bool Marked(Node* node) { return node_map.Get(node) > 0; }
+};
+
+
+class PeeledIterationImpl : public PeeledIteration {
+ public:
+  NodeVector node_pairs_;
+  explicit PeeledIterationImpl(Zone* zone) : node_pairs_(zone) {}
+};
+
+
+Node* PeeledIteration::map(Node* node) {
+  // TODO(turbofan): we use a simple linear search, since the peeled iteration
+  // is really only used in testing.
+  PeeledIterationImpl* impl = static_cast<PeeledIterationImpl*>(this);

[Benedikt Meurer, 2015/01/16 17:55:36]

How about avoiding the PeeledIterationImpl and just put the NodeVector into the
PeeledIteration class itself?

[titzer, 2015/01/19 09:47:12]

Was hoping to hide the implementation for now, since it is a linear search at
the moment and it's only used in testing. Still not sure if we need to have the
peeled iteration exposed in production code.

+  for (size_t i = 0; i < impl->node_pairs_.size(); i += 2) {
+    if (impl->node_pairs_[i] == node) return impl->node_pairs_[i + 1];
+  }
+  return node;
+}
+
+
+PeeledIteration* LoopPeeler::Peel(Graph* graph, CommonOperatorBuilder* common,
+                                  LoopTree* loop_tree, LoopTree::Loop* loop,
+                                  Zone* tmp_zone) {
+  PeeledIterationImpl* iter = new (tmp_zone) PeeledIterationImpl(tmp_zone);
+  Peeling peeling(graph, tmp_zone, loop->TotalSize() * 2 + 2,
+                  &iter->node_pairs_);
+
+  //============================================================================
+  // Construct the peeled iteration.
+  //============================================================================
+  Node* dead = graph->NewNode(common->Dead());
+
+  // Map the loop header nodes to their entry values.
+  for (Node* node : loop_tree->HeaderNodes(loop)) {
+    // TODO(titzer): assuming loop entry at index 0.
+    peeling.Insert(node, node->InputAt(0));
+  }
+
+  // Copy all the nodes of loop body for the peeled iteration.
+  peeling.CopyNodes(graph, tmp_zone, dead, loop_tree->BodyNodes(loop));
+
+  //============================================================================
+  // Replace the entry to the loop with the output of the peeled iteration.
+  //============================================================================
+  Node* loop_node = loop_tree->GetLoopControl(loop);
+  Node* new_entry;
+  int backedges = loop_node->InputCount() - 1;
+  if (backedges > 1) {
+    // Multiple backedges from original loop, therefore multiple output edges
+    // from the peeled iteration.
+    NodeVector inputs(tmp_zone);
+    for (int i = 1; i < loop_node->InputCount(); i++) {
+      inputs.push_back(peeling.map(loop_node->InputAt(i)));
+    }
+    Node* merge =
+        graph->NewNode(common->Merge(backedges), backedges, &inputs[0]);
+
+    // Merge values from the multiple output edges of the peeled iteration.
+    for (Node* node : loop_tree->HeaderNodes(loop)) {
+      if (node->opcode() == IrOpcode::kLoop) continue;  // already done.
+      inputs.clear();
+      for (int i = 0; i < backedges; i++) {
+        inputs.push_back(peeling.map(node->InputAt(1 + i)));
+      }
+      for (Node* input : inputs) {
+        if (input != inputs[0]) {  // Non-redundant phi.
+          inputs.push_back(merge);
+          const Operator* op = common->ResizeMergeOrPhi(node->op(), backedges);
+          Node* phi = graph->NewNode(op, backedges + 1, &inputs[0]);
+          node->ReplaceInput(0, phi);
+          break;
+        }
+      }
+    }
+    new_entry = merge;
+  } else {
+    // Only one backedge, simply replace the input to loop with output of
+    // peeling.
+    for (Node* node : loop_tree->HeaderNodes(loop)) {
+      node->ReplaceInput(0, peeling.map(node->InputAt(0)));
+    }
+    new_entry = peeling.map(loop_node->InputAt(1));
+  }
+  loop_node->ReplaceInput(0, new_entry);
+
+  //============================================================================
+  // Find the loop exit region.
+  //============================================================================
+  NodeVector exits(tmp_zone);
+  Node* end = NULL;
+  for (Node* node : loop_tree->LoopNodes(loop)) {
+    for (Node* use : node->uses()) {
+      if (!loop_tree->Contains(loop, use)) {
+        if (node->opcode() == IrOpcode::kBranch &&
+            (use->opcode() == IrOpcode::kIfTrue ||
+             use->opcode() == IrOpcode::kIfFalse)) {
+          // This is a branch from inside the loop to outside the loop.
+          exits.push_back(use);
+        }
+      }
+    }
+  }
+
+  if (exits.size() == 0) return iter;  // no exits => NTL
+
+  if (exits.size() == 1) {
+    // Only one exit, so {end} is that exit.
+    end = exits[0];
+  } else {
+    // {end} should be the common merge from the exits.
+    NodeVector rets(tmp_zone);
+    for (Node* exit : exits) {
+      Node* found = NULL;
+      for (Node* use : exit->uses()) {
+        if (use->opcode() == IrOpcode::kMerge) {
+          found = use;
+          if (end == NULL) {
+            end = found;
+          } else {
+            CHECK_EQ(end, found);  // it should be unique!
+          }
+        } else if (use->opcode() == IrOpcode::kReturn) {
+          found = use;
+          rets.push_back(found);
+        }
+      }
+      // There should be a merge or a return for each exit.
+      CHECK_NE(NULL, found);
+    }
+    // Return nodes, the end merge, and the phis associated with the end merge
+    // must be duplicated as well.
+    for (Node* node : rets) exits.push_back(node);
+    if (end != NULL) {
+      exits.push_back(end);
+      for (Node* use : end->uses()) {
+        if (IrOpcode::IsPhiOpcode(use->opcode())) exits.push_back(use);
+      }
+    }
+  }
+
+  //============================================================================
+  // Duplicate the loop exit region and add a merge.
+  //============================================================================
+  NodeRange exit_range(&exits[0], &exits[0] + exits.size());
+  peeling.CopyNodes(graph, tmp_zone, dead, exit_range);
+
+  Node* merge = graph->NewNode(common->Merge(2), end, peeling.map(end));
+  end->ReplaceUses(merge);
+  merge->ReplaceInput(0, end);  // HULK SMASH!!
+
+  // Find and update all the edges into either the loop or exit region.
+  for (int i = 0; i < 2; i++) {
+    NodeRange range = i == 0 ? loop_tree->LoopNodes(loop) : exit_range;
+    ZoneVector<Edge> value_edges(tmp_zone);
+    ZoneVector<Edge> effect_edges(tmp_zone);
+
+    for (Node* node : range) {
+      // Gather value and effect edges from outside the region.
+      for (Edge edge : node->use_edges()) {
+        if (!peeling.Marked(edge.from())) {
+          // Edge from outside the loop into the region.
+          if (NodeProperties::IsValueEdge(edge) ||
+              NodeProperties::IsContextEdge(edge)) {
+            value_edges.push_back(edge);
+          } else if (NodeProperties::IsEffectEdge(edge)) {
+            effect_edges.push_back(edge);
+          } else {
+            // don't do anything for control edges.
+            // TODO(titzer): should update control edges to peeled?
+          }
+        }
+      }
+
+      // Update all the value and effect edges at once.
+      if (!value_edges.empty()) {
+        // TODO(titzer): machine type is wrong here.
+        Node* phi = graph->NewNode(common->Phi(kMachAnyTagged, 2), node,
+                                   peeling.map(node), merge);
+        for (Edge edge : value_edges) edge.UpdateTo(phi);
+        value_edges.clear();
+      }
+      if (!effect_edges.empty()) {
+        Node* effect_phi = graph->NewNode(common->EffectPhi(2), node,
+                                          peeling.map(node), merge);
+        for (Edge edge : effect_edges) edge.UpdateTo(effect_phi);
+        effect_edges.clear();
+      }
+    }
+  }
+
+  return iter;
+}
+
+}  // namespace compiler
+}  // namespace internal
+}  // namespace v8