Refactored FlowGraphBuilder into a separate FlowGraph representation.

runtime/vm/flow_graph.cc

Index: runtime/vm/flow_graph.cc
diff --git a/runtime/vm/flow_graph.cc b/runtime/vm/flow_graph.cc
new file mode 100644
index 0000000000000000000000000000000000000000..f0f4dce5cfd6166d0aafcb45c00fa6ceca438bf9
--- /dev/null
+++ b/runtime/vm/flow_graph.cc
@@ -0,0 +1,472 @@
+// Copyright (c) 2012, 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/flow_graph.h"
+
+#include "vm/bit_vector.h"
+#include "vm/flow_graph_builder.h"
+#include "vm/longjump.h"
+
+namespace dart {
+
+FlowGraph::FlowGraph(const ParsedFunction& parsed_function)
+    : parsed_function_(parsed_function),
+      copied_parameter_count_(parsed_function.copied_parameter_count()),
+      // All parameters are copied if any parameter is.
+      non_copied_parameter_count_((copied_parameter_count_ == 0)
+          ? parsed_function.function().num_fixed_parameters()
+          : 0),
+      stack_local_count_(parsed_function.stack_local_count()),
+      graph_entry_(NULL),
+      preorder_(),
+      postorder_(),
+      reverse_postorder_(),
+      parent_(),
+      assigned_vars_(),
+      current_ssa_temp_index_(0) { }
+
+
+void FlowGraph::BuildGraph() {
+  FlowGraphBuilder builder(*this);
+  graph_entry_ = builder.BuildGraph();
+  ComputeOrders();
+}
+
+
+void FlowGraph::ComputeOrders() {
+  // Initialize state.
+  preorder_.TruncateTo(0);
+  postorder_.TruncateTo(0);
+  reverse_postorder_.TruncateTo(0);
+  parent_.TruncateTo(0);
+  assigned_vars_.TruncateTo(0);
+  // Perform a depth-first traversal of the graph to build preorder and
+  // postorder block orders.
+  graph_entry_->DiscoverBlocks(NULL,  // Entry block predecessor.
+                               &preorder_,
+                               &postorder_,
+                               &parent_,
+                               &assigned_vars_,
+                               variable_count(),
+                               non_copied_parameter_count());
+  // Number blocks in reverse postorder.
+  intptr_t block_count = postorder_.length();
+  for (intptr_t i = 0; i < block_count; ++i) {
+    postorder_[i]->set_block_id(block_count - i - 1);
+    reverse_postorder_.Add(postorder_[block_count - i - 1]);

[Kevin Millikin (Google), 2012/08/16 08:09:57]

I don't think a reversed copy of the other list adds much value.  I wonder if
anyone would complain if we removed it?

[zerny-google, 2012/08/16 11:52:27]

Differing to a later CL.

+  }
+  // Link instructions backwards for optimized compilation.
+  // TODO(zerny): The builder should do this at construction time.
+  for (intptr_t i = 0; i < block_count; ++i) {
+    BlockEntryInstr* entry = postorder_[i];
+    Instruction* previous = entry;
+    for (ForwardInstructionIterator it(entry); !it.Done(); it.Advance()) {
+      Instruction* current = it.Current();
+      current->set_previous(previous);
+      previous = current;
+    }
+  }
+}
+
+
+void FlowGraph::ComputeSSA() {
+  GrowableArray<BitVector*> dominance_frontier;
+  ComputeDominators(&preorder_, &parent_, &dominance_frontier);
+  InsertPhis(preorder_, assigned_vars_, dominance_frontier);
+  GrowableArray<PhiInstr*> live_phis;
+  // Rename uses to reference inserted phis where appropriate.
+  // Collect phis that reach a non-environment use.
+  Rename(&live_phis);
+  // Propagate alive mark transitively from alive phis.
+  MarkLivePhis(&live_phis);
+}
+
+
+// Compute immediate dominators and the dominance frontier for each basic
+// block.  As a side effect of the algorithm, sets the immediate dominator
+// of each basic block.
+//
+// preorder: an input list of basic block entries in preorder.  The
+//     algorithm relies on the block ordering.
+//
+// parent: an input parameter encoding a depth-first spanning tree of
+//     the control flow graph.  The array maps the preorder block
+//     number of a block to the preorder block number of its spanning
+//     tree parent.
+//
+// dominance_frontier: an output parameter encoding the dominance frontier.
+//     The array maps the preorder block number of a block to the set of
+//     (preorder block numbers of) blocks in the dominance frontier.
+void FlowGraph::ComputeDominators(
+    GrowableArray<BlockEntryInstr*>* preorder,
+    GrowableArray<intptr_t>* parent,
+    GrowableArray<BitVector*>* dominance_frontier) {
+  // Use the SEMI-NCA algorithm to compute dominators.  This is a two-pass
+  // version of the Lengauer-Tarjan algorithm (LT is normally three passes)
+  // that eliminates a pass by using nearest-common ancestor (NCA) to
+  // compute immediate dominators from semidominators.  It also removes a
+  // level of indirection in the link-eval forest data structure.
+  //
+  // The algorithm is described in Georgiadis, Tarjan, and Werneck's
+  // "Finding Dominators in Practice".
+  // See http://www.cs.princeton.edu/~rwerneck/dominators/ .
+
+  // All arrays are maps between preorder basic-block numbers.
+  intptr_t size = parent->length();
+  GrowableArray<intptr_t> idom(size);  // Immediate dominator.
+  GrowableArray<intptr_t> semi(size);  // Semidominator.
+  GrowableArray<intptr_t> label(size);  // Label for link-eval forest.
+
+  // 1. First pass: compute semidominators as in Lengauer-Tarjan.
+  // Semidominators are computed from a depth-first spanning tree and are an
+  // approximation of immediate dominators.
+
+  // Use a link-eval data structure with path compression.  Implement path
+  // compression in place by mutating the parent array.  Each block has a
+  // label, which is the minimum block number on the compressed path.
+
+  // Initialize idom, semi, and label used by SEMI-NCA.  Initialize the
+  // dominance frontier output array.
+  for (intptr_t i = 0; i < size; ++i) {
+    idom.Add((*parent)[i]);
+    semi.Add(i);
+    label.Add(i);
+    dominance_frontier->Add(new BitVector(size));
+  }
+
+  // Loop over the blocks in reverse preorder (not including the graph
+  // entry).
+  for (intptr_t block_index = size - 1; block_index >= 1; --block_index) {
+    // Loop over the predecessors.
+    BlockEntryInstr* block = (*preorder)[block_index];
+    for (intptr_t i = 0, count = block->PredecessorCount(); i < count; ++i) {
+      BlockEntryInstr* pred = block->PredecessorAt(i);
+      ASSERT(pred != NULL);
+
+      // Look for the semidominator by ascending the semidominator path
+      // starting from pred.
+      intptr_t pred_index = pred->preorder_number();
+      intptr_t best = pred_index;
+      if (pred_index > block_index) {
+        CompressPath(block_index, pred_index, parent, &label);
+        best = label[pred_index];
+      }
+
+      // Update the semidominator if we've found a better one.
+      semi[block_index] = Utils::Minimum(semi[block_index], semi[best]);
+    }
+
+    // Now use label for the semidominator.
+    label[block_index] = semi[block_index];
+  }
+
+  // 2. Compute the immediate dominators as the nearest common ancestor of
+  // spanning tree parent and semidominator, for all blocks except the entry.
+  for (intptr_t block_index = 1; block_index < size; ++block_index) {
+    intptr_t dom_index = idom[block_index];
+    while (dom_index > semi[block_index]) {
+      dom_index = idom[dom_index];
+    }
+    idom[block_index] = dom_index;
+    (*preorder)[block_index]->set_dominator((*preorder)[dom_index]);
+    (*preorder)[dom_index]->AddDominatedBlock((*preorder)[block_index]);
+  }
+
+  // 3. Now compute the dominance frontier for all blocks.  This is
+  // algorithm in "A Simple, Fast Dominance Algorithm" (Figure 5), which is
+  // attributed to a paper by Ferrante et al.  There is no bookkeeping
+  // required to avoid adding a block twice to the same block's dominance
+  // frontier because we use a set to represent the dominance frontier.
+  for (intptr_t block_index = 0; block_index < size; ++block_index) {
+    BlockEntryInstr* block = (*preorder)[block_index];
+    intptr_t count = block->PredecessorCount();
+    if (count <= 1) continue;
+    for (intptr_t i = 0; i < count; ++i) {
+      BlockEntryInstr* runner = block->PredecessorAt(i);
+      while (runner != block->dominator()) {
+        (*dominance_frontier)[runner->preorder_number()]->Add(block_index);
+        runner = runner->dominator();
+      }
+    }
+  }
+}
+
+
+void FlowGraph::CompressPath(intptr_t start_index,
+                             intptr_t current_index,
+                             GrowableArray<intptr_t>* parent,
+                             GrowableArray<intptr_t>* label) {
+  intptr_t next_index = (*parent)[current_index];
+  if (next_index > start_index) {
+    CompressPath(start_index, next_index, parent, label);
+    (*label)[current_index] =
+        Utils::Minimum((*label)[current_index], (*label)[next_index]);
+    (*parent)[current_index] = (*parent)[next_index];
+  }
+}
+
+
+void FlowGraph::InsertPhis(
+    const GrowableArray<BlockEntryInstr*>& preorder,
+    const GrowableArray<BitVector*>& assigned_vars,
+    const GrowableArray<BitVector*>& dom_frontier) {
+  const intptr_t block_count = preorder.length();
+  // Map preorder block number to the highest variable index that has a phi
+  // in that block.  Use it to avoid inserting multiple phis for the same
+  // variable.
+  GrowableArray<intptr_t> has_already(block_count);
+  // Map preorder block number to the highest variable index for which the
+  // block went on the worklist.  Use it to avoid adding the same block to
+  // the worklist more than once for the same variable.
+  GrowableArray<intptr_t> work(block_count);
+
+  // Initialize has_already and work.
+  for (intptr_t block_index = 0; block_index < block_count; ++block_index) {
+    has_already.Add(-1);
+    work.Add(-1);
+  }
+
+  // Insert phis for each variable in turn.
+  GrowableArray<BlockEntryInstr*> worklist;
+  for (intptr_t var_index = 0; var_index < variable_count(); ++var_index) {
+    // Add to the worklist each block containing an assignment.
+    for (intptr_t block_index = 0; block_index < block_count; ++block_index) {
+      if (assigned_vars[block_index]->Contains(var_index)) {
+        work[block_index] = var_index;
+        worklist.Add(preorder[block_index]);
+      }
+    }
+
+    while (!worklist.is_empty()) {
+      BlockEntryInstr* current = worklist.Last();
+      worklist.RemoveLast();
+      // Ensure a phi for each block in the dominance frontier of current.
+      for (BitVector::Iterator it(dom_frontier[current->preorder_number()]);
+           !it.Done();
+           it.Advance()) {
+        int index = it.Current();
+        if (has_already[index] < var_index) {
+          BlockEntryInstr* block = preorder[index];
+          ASSERT(block->IsJoinEntry());
+          block->AsJoinEntry()->InsertPhi(var_index, variable_count());
+          has_already[index] = var_index;
+          if (work[index] < var_index) {
+            work[index] = var_index;
+            worklist.Add(block);
+          }
+        }
+      }
+    }
+  }
+}
+
+
+void FlowGraph::Rename(GrowableArray<PhiInstr*>* live_phis) {
+  // TODO(fschneider): Support catch-entry.
+  if (graph_entry_->SuccessorCount() > 1) {
+    Bailout("Catch-entry support in SSA.");
+  }
+
+  // Initialize start environment.
+  GrowableArray<Value*> start_env(variable_count());
+  for (intptr_t i = 0; i < parameter_count(); ++i) {
+    ParameterInstr* param = new ParameterInstr(i);
+    param->set_ssa_temp_index(alloc_ssa_temp_index());  // New SSA temp.
+    start_env.Add(new UseVal(param));
+  }
+
+  // All locals are initialized with #null.
+  Value* null_value = new ConstantVal(Object::ZoneHandle());
+  while (start_env.length() < variable_count()) {
+    start_env.Add(null_value);
+  }
+  graph_entry_->set_start_env(
+      new Environment(start_env, non_copied_parameter_count_));
+
+  BlockEntryInstr* normal_entry = graph_entry_->SuccessorAt(0);
+  ASSERT(normal_entry != NULL);  // Must have entry.
+  GrowableArray<Value*> env(variable_count());
+  env.AddArray(start_env);
+  RenameRecursive(normal_entry, &env, live_phis);
+}
+
+
+// Helper to a copy a value iff it is a UseVal.
+static Value* CopyValue(Value* value) {
+  return value->IsUse()
+      ? new UseVal(value->AsUse()->definition())
+      : value;
+}
+
+
+void FlowGraph::RenameRecursive(BlockEntryInstr* block_entry,
+                                GrowableArray<Value*>* env,
+                                GrowableArray<PhiInstr*>* live_phis) {
+  // 1. Process phis first.
+  if (block_entry->IsJoinEntry()) {
+    JoinEntryInstr* join = block_entry->AsJoinEntry();
+    if (join->phis() != NULL) {
+      for (intptr_t i = 0; i < join->phis()->length(); ++i) {
+        PhiInstr* phi = (*join->phis())[i];
+        if (phi != NULL) {
+          (*env)[i] = new UseVal(phi);
+          phi->set_ssa_temp_index(alloc_ssa_temp_index());  // New SSA temp.
+        }
+      }
+    }
+  }
+
+  // 2. Process normal instructions.
+  for (ForwardInstructionIterator it(block_entry); !it.Done(); it.Advance()) {
+    Instruction* current = it.Current();
+    // Attach current environment to the instruction. First, each instruction
+    // gets a full copy of the environment. Later we optimize this by
+    // eliminating unnecessary environments.
+    current->set_env(new Environment(*env, non_copied_parameter_count_));
+
+    // 2a. Handle uses:
+    // Update expression stack environment for each use.
+    // For each use of a LoadLocal or StoreLocal: Replace it with the value
+    // from the environment.
+    for (intptr_t i = current->InputCount() - 1; i >= 0; --i) {
+      Value* v = current->InputAt(i);
+      if (!v->IsUse()) continue;
+      // Update expression stack.
+      ASSERT(env->length() > variable_count());
+
+      Value* input_value = env->Last();
+      ASSERT(input_value->IsUse());
+      env->RemoveLast();
+
+      BindInstr* as_bind = v->AsUse()->definition()->AsBind();
+      if ((as_bind != NULL) &&
+          (as_bind->computation()->IsLoadLocal() ||
+           as_bind->computation()->IsStoreLocal())) {
+        // Assert exactly one use.
+        ASSERT(as_bind->use_list() == v);
+        ASSERT(as_bind->use_list()->next_use() == NULL);
+        // Remove the use, its defintion and copy the environment value.
+        v->RemoveFromUseList();
+        as_bind->RemoveFromGraph();
+        current->SetInputAt(i, CopyValue(input_value));
+      }
+    }
+
+    // Drop pushed arguments for calls.
+    for (intptr_t j = 0; j < current->ArgumentCount(); j++) {
+      env->RemoveLast();
+    }
+
+    // 2b. Handle LoadLocal and StoreLocal.
+    // For each LoadLocal: Remove it from the graph.
+    // For each StoreLocal: Remove it from the graph and update the environment.
+    BindInstr* bind = current->AsBind();
+    if (bind != NULL) {
+      LoadLocalComp* load = bind->computation()->AsLoadLocal();
+      StoreLocalComp* store = bind->computation()->AsStoreLocal();
+      if ((load != NULL) || (store != NULL)) {
+        intptr_t index;
+        if (store != NULL) {
+          index = store->local().BitIndexIn(non_copied_parameter_count_);
+          // Update renaming environment.
+          (*env)[index] = store->value();
+        } else {
+          // The graph construction ensures we do not have an unused LoadLocal
+          // computation.
+          ASSERT(bind->is_used());
+          index = load->local().BitIndexIn(non_copied_parameter_count_);
+
+          Value* value = (*env)[index];
+          if (value->IsUse()) {
+            PhiInstr* phi = value->AsUse()->definition()->AsPhi();
+            if ((phi != NULL) && !phi->is_alive()) {
+              phi->mark_alive();
+              live_phis->Add(phi);
+            }
+          }
+        }
+        // Update expression stack or remove from graph.
+        if (bind->is_used()) {
+          // Assert exactly one use.
+          ASSERT(bind->use_list() != NULL);
+          ASSERT(bind->use_list()->next_use() == NULL);
+          env->Add(CopyValue((*env)[index]));
+          // We remove load/store instructions when we find their use in 2a.
+        } else {
+          it.RemoveCurrentFromGraph();
+        }
+      } else {
+        // Not a load or store.
+        if (bind->is_used()) {
+          // Assign fresh SSA temporary and update expression stack.
+          bind->set_ssa_temp_index(alloc_ssa_temp_index());
+          env->Add(new UseVal(bind));
+        }
+      }
+    }
+
+    // 2c. Handle pushed argument.
+    PushArgumentInstr* push = current->AsPushArgument();
+    if (push != NULL) {
+      env->Add(new UseVal(push));
+    }
+  }
+
+  // 3. Process dominated blocks.
+  for (intptr_t i = 0; i < block_entry->dominated_blocks().length(); ++i) {
+    BlockEntryInstr* block = block_entry->dominated_blocks()[i];
+    GrowableArray<Value*> new_env(env->length());
+    new_env.AddArray(*env);
+    RenameRecursive(block, &new_env, live_phis);
+  }
+
+  // 4. Process successor block. We have edge-split form, so that only blocks
+  // with one successor can have a join block as successor.
+  if ((block_entry->last_instruction()->SuccessorCount() == 1) &&
+      block_entry->last_instruction()->SuccessorAt(0)->IsJoinEntry()) {
+    JoinEntryInstr* successor =
+        block_entry->last_instruction()->SuccessorAt(0)->AsJoinEntry();
+    intptr_t pred_index = successor->IndexOfPredecessor(block_entry);
+    ASSERT(pred_index >= 0);
+    if (successor->phis() != NULL) {
+      for (intptr_t i = 0; i < successor->phis()->length(); ++i) {
+        PhiInstr* phi = (*successor->phis())[i];
+        if (phi != NULL) {
+          // Rename input operand and make a copy if it is a UseVal.
+          phi->SetInputAt(pred_index, CopyValue((*env)[i]));
+        }
+      }
+    }
+  }
+}
+
+
+void FlowGraph::MarkLivePhis(GrowableArray<PhiInstr*>* live_phis) {
+  while (!live_phis->is_empty()) {
+    PhiInstr* phi = live_phis->Last();
+    live_phis->RemoveLast();
+    for (intptr_t i = 0; i < phi->InputCount(); i++) {
+      Value* val = phi->InputAt(i);
+      if (!val->IsUse()) continue;
+      PhiInstr* used_phi = val->AsUse()->definition()->AsPhi();
+      if ((used_phi != NULL) && !used_phi->is_alive()) {
+        used_phi->mark_alive();
+        live_phis->Add(used_phi);
+      }
+    }
+  }
+}
+
+
+void FlowGraph::Bailout(const char* reason) {
+  const char* kFormat = "FlowGraph Bailout: %s %s";
+  const char* function_name = parsed_function_.function().ToCString();
+  intptr_t len = OS::SNPrint(NULL, 0, kFormat, function_name, reason) + 1;
+  char* chars = Isolate::Current()->current_zone()->Alloc<char>(len);
+  OS::SNPrint(chars, len, kFormat, function_name, reason);
+  const Error& error = Error::Handle(
+      LanguageError::New(String::Handle(String::New(chars))));
+  Isolate::Current()->long_jump_base()->Jump(1, error);
+}
+}

[Kevin Millikin (Google), 2012/08/16 08:09:57]

There should be one or two lines of whitespace before the closing brace and it
should have a "// namespace dart" end-of-line comment.

[zerny-google, 2012/08/16 11:52:27]

Done.