Collect edge count profiling data and reorder basic blocks.

runtime/vm/block_scheduler.cc

Index: runtime/vm/block_scheduler.cc
diff --git a/runtime/vm/block_scheduler.cc b/runtime/vm/block_scheduler.cc
new file mode 100644
index 0000000000000000000000000000000000000000..3dacd501cf45a2539442b4554f122a63732524b0
--- /dev/null
+++ b/runtime/vm/block_scheduler.cc
@@ -0,0 +1,217 @@
+// Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
+// for details. All rights reserved. Use of this source code is governed by a
+// BSD-style license that can be found in the LICENSE file.
+
+#include "vm/block_scheduler.h"
+
+#include "vm/allocation.h"
+#include "vm/flow_graph.h"
+
+namespace dart {
+
+// Compute the edge count at the deopt id of a TargetEntry or Goto.
+static intptr_t ComputeEdgeCount(const Code& unoptimized_code,
+                                 intptr_t deopt_id) {

[srdjan, 2013/08/14 00:28:24]

This seems like quite slow way to extract edge counts. What is the impact on
compilation time?

+  ASSERT(deopt_id != Isolate::kNoDeoptId);
+
+  // Intrinsified functions do not have edge counts, so give all edges equal
+  // weights.

[srdjan, 2013/08/14 00:28:24]

Some intrinsified functions should have edge counts as they are executed as part
of the slow path, however, those edgecounts are probably not
representative/reliable. What you are testing here is if this is an intrinsic
only code (no slow path).

+  if (unoptimized_code.pointer_offsets_length() == 0) return 1;
+
+  uword pc = unoptimized_code.GetPcForDeoptId(deopt_id, PcDescriptors::kDeopt);
+  Array& array = Array::Handle();
+  // Pointer offsets are sorted in decreasing order.  Find the first one
+  // after the deopt id's pc.
+  // TODO(kmillikin): Use a more reliable way to find the counter.
+  for (intptr_t j = unoptimized_code.pointer_offsets_length() - 1;
+       j >= 0;
+       --j) {
+    uword addr =
+        unoptimized_code.GetPointerOffsetAt(j) + unoptimized_code.EntryPoint();
+    if (addr > pc) {
+      array ^= *reinterpret_cast<RawObject**>(addr);
+      break;
+    }
+  }
+  ASSERT(!array.IsNull());
+  return Smi::Value(Smi::RawCast(array.At(0)));
+}
+
+
+// There is an edge from instruction->successor.  Set its weight (edge count
+// per function entry).
+static void SetEdgeWeight(Instruction* instruction,
+                          BlockEntryInstr* successor,
+                          const Code& unoptimized_code,
+                          intptr_t entry_count) {
+  intptr_t count = -1;
+  double weight = 0.0;

[Florian Schneider, 2013/08/12 11:30:09]

style-nit: I would not hoist these two declarations out of the if- and
else-blocks.

[Kevin Millikin (Google), 2013/08/13 11:14:57]

OK, done.

[srdjan, 2013/08/14 00:28:24]

Could you please upload your changes.

+  TargetEntryInstr* target = successor->AsTargetEntry();
+  if (target != NULL) {
+    count = ComputeEdgeCount(unoptimized_code, target->deopt_id());
+    if ((count >= 0) && (entry_count != 0)) {
+      weight = static_cast<double>(count) / static_cast<double>(entry_count);
+      target->set_edge_weight(weight);
+    }
+  } else {
+    GotoInstr* jump = instruction->AsGoto();
+    if (jump != NULL) {
+      count = ComputeEdgeCount(unoptimized_code, jump->deopt_id());
+      if ((count >= 0) && (entry_count != 0)) {
+        weight = static_cast<double>(count) / static_cast<double>(entry_count);
+        jump->set_edge_weight(weight);
+      }
+    }
+  }
+}
+
+
+void BlockScheduler::AssignEdgeWeights() const {
+  const Code& unoptimized_code = Code::Handle(
+      flow_graph()->parsed_function().function().unoptimized_code());
+
+  intptr_t entry_count =
+      ComputeEdgeCount(unoptimized_code,
+                       flow_graph()->graph_entry()->normal_entry()->deopt_id());
+  flow_graph()->graph_entry()->set_entry_count(entry_count);
+
+  for (BlockIterator it = flow_graph()->reverse_postorder_iterator();
+       !it.Done();
+       it.Advance()) {
+    BlockEntryInstr* block = it.Current();
+    Instruction* last = block->last_instruction();
+    for (intptr_t i = 0; i < last->SuccessorCount(); ++i) {
+      BlockEntryInstr* succ = last->SuccessorAt(i);
+      SetEdgeWeight(last, succ, unoptimized_code, entry_count);
+    }
+  }
+}
+
+
+// A weighted control-flow graph edge.
+struct Edge {
+  Edge(BlockEntryInstr* source, BlockEntryInstr* target, double weight)
+      : source(source), target(target), weight(weight) { }
+
+  static int LowestWeightFirst(const Edge* a, const Edge* b);
+
+  BlockEntryInstr* source;
+  BlockEntryInstr* target;
+  double weight;
+};
+
+
+// A linked list node in a chain of blocks.
+struct Link : public ZoneAllocated {
+  Link(BlockEntryInstr* block, Link* next) : block(block), next(next) { }
+
+  BlockEntryInstr* block;
+  Link* next;
+};
+
+
+// A chain of blocks with first and last pointers for fast concatenation and
+// a length to support adding a shorter chain's links to a longer chain.
+struct Chain : public ZoneAllocated {
+  explicit Chain(BlockEntryInstr* block)
+      : first(new Link(block, NULL)), last(first), length(1) { }
+
+  Link* first;
+  Link* last;
+  intptr_t length;
+};
+
+
+int Edge::LowestWeightFirst(const Edge* a, const Edge* b) {
+  return (a->weight < b->weight) ? -1 : (a->weight > b->weight);
+}
+
+
+// Combine two chains by adding the shorter chain's links to the longer
+// chain.
+static void Union(GrowableArray<Chain*>* chains,
+                  Chain* source_chain,
+                  Chain* target_chain) {
+  if (source_chain->length < target_chain->length) {
+    for (Link* link = source_chain->first; link != NULL; link = link->next) {
+      (*chains)[link->block->postorder_number()] = target_chain;
+    }
+    // Link the chains.
+    source_chain->last->next = target_chain->first;
+    // Update the state of the longer chain.
+    target_chain->first = source_chain->first;
+    target_chain->length += source_chain->length;
+  } else {
+    for (Link* link = target_chain->first; link != NULL; link = link->next) {
+      (*chains)[link->block->postorder_number()] = source_chain;
+    }
+    source_chain->last->next = target_chain->first;
+    source_chain->last = target_chain->last;
+    source_chain->length += target_chain->length;
+  }
+}
+
+
+void BlockScheduler::ReorderBlocks() const {
+  // Add every block to a chain of length 1 and compute a list of edges
+  // sorted by weight.
+  intptr_t block_count = flow_graph()->preorder().length();
+  GrowableArray<Edge> edges(2 * block_count);
+
+  // A map from a block's postorder number to the chain it is in.  Used to
+  // implement a simple (ordered) union-find data structure.  Chains are
+  // stored by pointer so that they are aliased (mutating one mutates all
+  // shared ones).  Find(n) is simply chains[n].
+  GrowableArray<Chain*> chains(block_count);
+
+  for (BlockIterator it = flow_graph()->postorder_iterator();
+       !it.Done();
+       it.Advance()) {
+    BlockEntryInstr* block = it.Current();
+    chains.Add(new Chain(block));
+
+    Instruction* last = block->last_instruction();
+    for (intptr_t i = 0; i < last->SuccessorCount(); ++i) {
+      BlockEntryInstr* succ = last->SuccessorAt(i);
+      double weight = 0.0;
+      if (succ->IsTargetEntry()) {
+        weight = succ->AsTargetEntry()->edge_weight();
+      } else if (last->IsGoto()) {
+        weight = last->AsGoto()->edge_weight();
+      }
+      edges.Add(Edge(block, succ, weight));
+    }
+  }
+
+  // Handle each edge in turn.  The edges are sorted by increasing weight.
+  edges.Sort(Edge::LowestWeightFirst);
+  while (!edges.is_empty()) {
+    Edge edge = edges.RemoveLast();
+    Chain* source_chain = chains[edge.source->postorder_number()];
+    Chain* target_chain = chains[edge.target->postorder_number()];
+
+    // If the source and target are already in the same chain or if the
+    // edge's source or target is not exposed at the appropriate end of a
+    // chain skip this edge.
+    if ((source_chain == target_chain) ||
+        (edge.source != source_chain->last->block) ||
+        (edge.target != target_chain->first->block)) {
+      continue;
+    }
+
+    Union(&chains, source_chain, target_chain);
+  }
+
+  // Build a new block order.  Emit each chain when its first block occurs
+  // in the original reverse postorder ordering (which gives a topological
+  // sort of the blocks).
+  for (intptr_t i = block_count - 1; i >= 0; --i) {
+    if (chains[i]->first->block == flow_graph()->postorder()[i]) {
+      for (Link* link = chains[i]->first; link != NULL; link = link->next) {
+        flow_graph()->codegen_block_order(true)->Add(link->block);
+      }
+    }
+  }
+}
+
+}  // namespace dart