Implement range analysis for smi values.

runtime/vm/flow_graph_optimizer.cc

 // 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_optimizer.h"
 
 #include "vm/bit_vector.h"
 #include "vm/cha.h"
 #include "vm/flow_graph_builder.h"
 #include "vm/hash_map.h"
@@ skipping 1016 matching lines @@
     HandleEqualityCompare(this, comparison->AsEqualityCompare(), instr,
                           current_iterator());
   } else {
     ASSERT(comparison->IsStrictCompare());
     // Nothing to do.
   }
 }
 
 
 // SminessPropagator ensures that CheckSmis are eliminated across phis.
-class SminessPropagator {
+class SminessPropagator : public ValueObject {
  public:
   explicit SminessPropagator(FlowGraph* flow_graph)
       : flow_graph_(flow_graph),
         known_smis_(new BitVector(flow_graph_->current_ssa_temp_index())),
         rollback_checks_(10),
         in_worklist_(NULL),
         worklist_(0) { }
 
   void Propagate();
 
  private:
   void PropagateSminessRecursive(BlockEntryInstr* block);
   void AddToWorklist(PhiInstr* phi);
   PhiInstr* RemoveLastFromWorklist();
   void ProcessPhis();
 
   FlowGraph* flow_graph_;
 
   BitVector* known_smis_;
   GrowableArray<intptr_t> rollback_checks_;
 
   BitVector* in_worklist_;
   GrowableArray<PhiInstr*> worklist_;
+
+  DISALLOW_COPY_AND_ASSIGN(SminessPropagator);
 };
 
 
 void SminessPropagator::AddToWorklist(PhiInstr* phi) {
   if (in_worklist_ == NULL) {
     in_worklist_ = new BitVector(flow_graph_->current_ssa_temp_index());
   }
   if (!in_worklist_->Contains(phi->ssa_temp_index())) {
     in_worklist_->Add(phi->ssa_temp_index());
     worklist_.Add(phi);
@@ skipping 126 matching lines @@
   ProcessPhis();
 }
 
 
 void FlowGraphOptimizer::PropagateSminess() {
   SminessPropagator propagator(flow_graph_);
   propagator.Propagate();
 }
 
 
+// Range analysis for smi values.
+class RangeAnalysis : public ValueObject {
+ public:
+  explicit RangeAnalysis(FlowGraph* flow_graph) : flow_graph_(flow_graph) { }
+
+  // Infer ranges for all values and remove overflow checks from binary smi
+  // operations when proven redundant.
+  void Analyze();
+
+ private:
+  // Collect all values that were proven to be smi in smi_values_ array and all
+  // CheckSmi instructions in smi_check_ array.
+  void CollectSmiValues();
+
+  // Iterate over smi values and constraint them at branch successors.

[Kevin Millikin (Google), 2012/09/21 08:59:33]

constraint ==> constrain wherever used as a verb.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+  // Additionally constraint values after CheckSmi instructions.
+  void InsertConstraints();
+
+  // Iterate over uses of the given definition and discover branches that
+  // constraint it. Insert appropriate Constraint instructions at true
+  // and false successor and rename all dominated uses to refer to a
+  // Constraint instead of this definition.
+  void InsertConstraintsFor(Definition* defn);
+
+  // Create a constraint for defn, insert it after given instruction and
+  // rename all uses that are dominated by it.
+  ConstraintInstr* ConstraintValue(Definition* defn,

[Kevin Millikin (Google), 2012/09/21 08:59:33]

ConstraintValue ==> ConstrainValue

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+                                   Range* constraint,
+                                   Instruction* after);
+
+  // Replace uses of the definition def that are dominated by instruction dom
+  // with uses of other definition.
+  void RenameDominatedUses(Definition* def,
+                           Instruction* dom,
+                           Definition* other);
+
+  // Propagate range information until fix-point is reached.
+  void InferRanges();
+
+  // Walk the dominator tree, initialize ranges for smi values and place them
+  // to the worklist.
+  void InitializeRangesRecursive(BlockEntryInstr* block);
+
+  // Remove artificial Constraint instructions and replace them with actual
+  // unconstrained definitions.
+  void RemoveConstraints();
+
+  void CreateWorklists();
+
+  void AddToWorklist(Definition* value) {
+    const intptr_t index = value->ssa_temp_index();
+    if (!in_inactive_worklist_->Contains(index)) {
+      in_inactive_worklist_->Add(index);
+      inactive_worklist_->Add(value);
+    }
+  }
+
+  bool IsWorklistEmpty() const {
+    return active_worklist_->is_empty();
+  }
+
+  void SwapWorklists();
+
+  FlowGraph* flow_graph_;
+
+  GrowableArray<Definition*> smi_values_;  // Value that are known to be smi.
+  GrowableArray<CheckSmiInstr*> smi_checks_;  // All CheckSmi instructions.
+
+  // All Constraints inserted during InsertConstraints phase. They are treated
+  // as smi values.
+  GrowableArray<ConstraintInstr*> constraints_;
+
+  // Bitvector for a quick filtering of known smi values.
+  BitVector* smi_definitions_;
+
+  // Worklists using during range propagation.
+  ZoneGrowableArray<Definition*>* active_worklist_;
+  ZoneGrowableArray<Definition*>* inactive_worklist_;
+  BitVector* in_inactive_worklist_;
+
+  DISALLOW_COPY_AND_ASSIGN(RangeAnalysis);
+};
+
+
+void RangeAnalysis::Analyze() {
+  CollectSmiValues();
+  InsertConstraints();
+  InferRanges();
+  RemoveConstraints();
+}
+
+
+void RangeAnalysis::CollectSmiValues() {

[Kevin Millikin (Google), 2012/09/21 08:59:33]

It seems more straightforward to directly insert constraints and mark smi values
in the bitvector.  I don't think you need the lists smi_values_ and smi_checks_.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Agreed. I am going to commit as is and refactor it at a later stage. For now I
would like to keep this phases separated.

+  for (BlockIterator block_it = flow_graph_->reverse_postorder_iterator();
+       !block_it.Done();
+       block_it.Advance()) {
+    BlockEntryInstr* block = block_it.Current();
+    for (ForwardInstructionIterator instr_it(block);
+         !instr_it.Done();
+         instr_it.Advance()) {
+      Instruction* current = instr_it.Current();
+      Definition* defn = current->AsDefinition();
+      if (defn != NULL) {
+        if (defn->GetPropagatedCid() == kSmiCid) smi_values_.Add(defn);
+      } else if (current->IsCheckSmi()) {
+        smi_checks_.Add(current->AsCheckSmi());
+      }
+    }
+
+    JoinEntryInstr* join = block->AsJoinEntry();
+    if (join != NULL) {
+      for (PhiIterator phi_it(join); !phi_it.Done(); phi_it.Advance()) {
+        PhiInstr* current = phi_it.Current();
+        if (current->GetPropagatedCid() == kSmiCid) {
+          smi_values_.Add(current);
+        }
+      }
+    }
+  }
+}
+
+
+// Returns true if use is dominated by the given instruction.
+// Note: uses that occur at instruction itself are not dominated by it.
+static bool IsDominatedUse(Instruction* dom, Value* use) {
+  BlockEntryInstr* dom_block = dom->GetBlock();
+
+  Instruction* instr = use->instruction();
+
+  PhiInstr* phi = instr->AsPhi();
+  if (phi != NULL) {
+    return dom_block->Dominates(phi->block()->PredecessorAt(use->use_index()));
+  }
+
+  BlockEntryInstr* use_block = instr->GetBlock();
+  if (use_block == dom_block) {
+    // Fast path for the case of block entry.
+    if (dom_block == dom) return true;
+
+    for (Instruction* curr = dom->next(); curr != NULL; curr = curr->next()) {
+      if (curr == instr) return true;
+    }
+
+    return false;
+  }
+
+  return dom_block->Dominates(use_block);
+}
+
+
+void RangeAnalysis::RenameDominatedUses(Definition* def,
+                                        Instruction* dom,
+                                        Definition* other) {
+  Value* next_use = NULL;
+  Value* prev_use = NULL;
+  for (Value* use = def->input_use_list();
+       use != NULL;
+       use = next_use) {
+    next_use = use->next_use();
+
+    // Skip instructions that are outside of the graph and
+    // dead phis.
+    if (use->instruction()->WasEliminated() &&

[Florian Schneider, 2012/09/21 08:56:44]

Why do you need to skip those?

[Kevin Millikin (Google), 2012/09/21 08:59:33]

Can live phis be WasEliminated?  That's weird.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

I don't want to compute ranges for things that are not connected to the graph
and do not affect code generation.

I removed this WasEliminated check and instead now recompute UseLists before I
start range analysis to ensure that they contain only live values.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Our WasEliminated check just checks if previous() != NULL which is not true for
phis. 

In fact WasEliminated should be called IsLinked().

+        !(use->instruction()->IsPhi() &&
+          use->instruction()->AsPhi()->is_alive())) {
+      prev_use = use;
+      continue;
+    }
+
+    if (IsDominatedUse(dom, use)) {
+      if (prev_use != NULL) {
+        prev_use->set_next_use(next_use);
+      } else {
+        def->set_input_use_list(next_use);
+      }
+      use->set_definition(other);
+      use->AddToInputUseList();
+    } else {
+      prev_use = use;
+    }
+  }
+}
+
+
+// For a comparison operation return an operation for the equivalent flipped
+// comparison: a (op) b === b (op') a.
+static Token::Kind FlipComparison(Token::Kind op) {
+  switch (op) {
+    case Token::kEQ: return Token::kEQ;
+    case Token::kNE: return Token::kNE;
+    case Token::kLT: return Token::kGT;
+    case Token::kGT: return Token::kLT;
+    case Token::kLTE: return Token::kGTE;
+    case Token::kGTE: return Token::kLTE;
+    default:
+      UNREACHABLE();
+      return Token::kILLEGAL;
+  }
+}
+
+// For a comparison operation return an operation for the negated comparison:
+// !(a (op) b) === a (op') b
+static Token::Kind NegateComparison(Token::Kind op) {
+  switch (op) {
+    case Token::kEQ: return Token::kNE;
+    case Token::kNE: return Token::kEQ;
+    case Token::kLT: return Token::kGTE;
+    case Token::kGT: return Token::kLTE;
+    case Token::kLTE: return Token::kGT;
+    case Token::kGTE: return Token::kLT;
+    default:
+      UNREACHABLE();
+      return Token::kILLEGAL;
+  }
+}
+
+
+// Given a boundary (right operand) and a comparison operation return
+// a symbolic range constraint for the left operand of the comparison assuming
+// that it evaluated to true.
+// For example for the comparison a < b symbol a is constrained with range
+// [Smi::kMinValue, b - 1].
+static Range* ConstraintRange(Token::Kind op, Definition* boundary) {
+  switch (op) {
+    case Token::kEQ:
+      return new Range(RangeBoundary::FromDefinition(boundary),
+                       RangeBoundary::FromDefinition(boundary));
+    case Token::kNE:
+      return Range::Unknown();
+    case Token::kLT:
+      return new Range(RangeBoundary::MinSmi(),
+                       RangeBoundary::FromDefinition(boundary, -1));
+    case Token::kGT:
+      return new Range(RangeBoundary::FromDefinition(boundary, 1),
+                       RangeBoundary::MaxSmi());
+    case Token::kLTE:
+      return new Range(RangeBoundary::MinSmi(),
+                       RangeBoundary::FromDefinition(boundary));
+    case Token::kGTE:
+      return new Range(RangeBoundary::FromDefinition(boundary),
+                       RangeBoundary::MaxSmi());
+    default:
+      UNREACHABLE();
+      return Range::Unknown();
+  }
+}
+
+
+ConstraintInstr* RangeAnalysis::ConstraintValue(Definition* defn,

[Florian Schneider, 2012/09/21 08:56:44]

Suggested better name: InsertConstraintFor

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+                                                Range* constraint_range,
+                                                Instruction* after) {

[Florian Schneider, 2012/09/21 08:56:44]

Maybe assert that Definition defn dominates Instruction after.

+  // No need to constraint constants.

[Florian Schneider, 2012/09/21 08:56:44]

s/constraint/constrain/

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+  if (defn->IsConstant()) return NULL;
+
+  ConstraintInstr* constraint =
+      new ConstraintInstr(new Value(defn), constraint_range);
+  constraint->InsertAfter(after);
+  constraint->set_ssa_temp_index(flow_graph_->alloc_ssa_temp_index());
+  RenameDominatedUses(defn, after, constraint);
+  constraints_.Add(constraint);
+  constraint->value()->set_instruction(constraint);
+  constraint->value()->set_use_index(0);
+  constraint->value()->AddToInputUseList();
+  return constraint;
+}
+
+
+void RangeAnalysis::InsertConstraintsFor(Definition* defn) {
+  for (Value* use = defn->input_use_list();
+       use != NULL;
+       use = use->next_use()) {
+    if (use->instruction()->IsBranch()) {
+      BranchInstr* branch = use->instruction()->AsBranch();
+      RelationalOpInstr* rel_op = branch->comparison()->AsRelationalOp();
+      if ((rel_op != NULL) && (rel_op->operands_class_id() == kSmiCid)) {
+        // Found comparison of two smis. Constrain defn at true and false
+        // successors using the other operand as a boundary.
+        Definition* boundary =

[Kevin Millikin (Google), 2012/09/21 08:59:33]

Since you test use_index() anyway, it's clearer to skip arithmetic to flip the
operand:

if (use->use_index() == 0) {
  boundary = ...;
  op_kind = ...;
} else {
  boundary = ...;
  op_kind = ...;
}

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+            rel_op->InputAt(1 - use->use_index())->definition();
+
+        // ConstraintValue assumes that defn is left operand of a comparison
+        // if it is right operand flip the comparison.
+        const Token::Kind op_kind =
+            (use->use_index() == 0) ? rel_op->kind()
+                                    : FlipComparison(rel_op->kind());
+
+        // Constrain definition at the true successor.
+        ConstraintInstr* true_constraint =
+            ConstraintValue(defn,
+                            ConstraintRange(op_kind, boundary),
+                            branch->true_successor());
+        // Mark true_constraint an artificial use of boundary. This ensures
+        // that constraints range is recalculated if boundary's range changes.
+        if (true_constraint != NULL) true_constraint->AddDependency(boundary);
+
+        // Constrain definition with a negated condition at the false successor.
+        ConstraintInstr* false_constraint =
+            ConstraintValue(
+                defn,
+                ConstraintRange(NegateComparison(op_kind), boundary),
+                branch->false_successor());
+        // Mark false_constraint an artificial use of boundary. This ensures
+        // that constraints range is recalculated if boundary's range changes.
+        if (false_constraint != NULL) false_constraint->AddDependency(boundary);
+      }
+    }
+  }
+}
+
+
+void RangeAnalysis::InsertConstraints() {
+  for (intptr_t i = 0; i < smi_checks_.length(); i++) {
+    CheckSmiInstr* check = smi_checks_[i];
+    ConstraintInstr* constraint =
+        ConstraintValue(check->value()->definition(),
+                        Range::Unknown(),
+                        check);
+    InsertConstraintsFor(constraint);  // Constraint uses further.
+  }
+
+  for (intptr_t i = 0; i < smi_values_.length(); i++) {
+    InsertConstraintsFor(smi_values_[i]);
+  }
+}
+
+
+void RangeAnalysis::InitializeRangesRecursive(BlockEntryInstr* block) {
+  JoinEntryInstr* join = block->AsJoinEntry();

[Kevin Millikin (Google), 2012/09/21 08:59:33]

It also seems like we could do this in the same walk of the graph as inserting
constraints.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

See above comment. I am going to commit as is now.

+  if ((join != NULL) && (join->phis() != NULL)) {
+    for (intptr_t i = 0; i < join->phis()->length(); ++i) {
+      PhiInstr* phi = (*join->phis())[i];
+      if (phi == NULL) continue;
+      if (smi_definitions_->Contains(phi->ssa_temp_index())) {
+        phi->InferRange();
+        AddToWorklist(phi);
+      }
+    }
+  }
+
+  for (ForwardInstructionIterator it(block); !it.Done(); it.Advance()) {
+    Definition* defn = it.Current()->AsDefinition();
+    if ((defn != NULL) &&
+        (defn->ssa_temp_index() != -1) &&
+        smi_definitions_->Contains(defn->ssa_temp_index())) {
+      defn->InferRange();
+      AddToWorklist(defn);
+    }
+  }
+
+  for (intptr_t i = 0; i < block->dominated_blocks().length(); ++i) {
+    InitializeRangesRecursive(block->dominated_blocks()[i]);
+  }
+}
+
+
+void RangeAnalysis::CreateWorklists() {
+  active_worklist_ = new ZoneGrowableArray<Definition*>(10);
+  inactive_worklist_ = new ZoneGrowableArray<Definition*>(10);
+  in_inactive_worklist_ = new BitVector(
+      flow_graph_->current_ssa_temp_index());
+}
+
+
+void RangeAnalysis::SwapWorklists() {
+  ZoneGrowableArray<Definition*>* temp = active_worklist_;
+  active_worklist_ = inactive_worklist_;
+  inactive_worklist_ = temp;
+  inactive_worklist_->Clear();

[Kevin Millikin (Google), 2012/09/21 08:59:33]

This worklist is assumed empty here, right?

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

We don't pop anything from the worklist. We just process it from start til the
end, so it is not empty.

+  in_inactive_worklist_->Clear();
+}
+
+
+void RangeAnalysis::InferRanges() {
+  CreateWorklists();
+
+  // Initialize bitvector for quick filtering of smi values.
+  smi_definitions_ = new BitVector(flow_graph_->current_ssa_temp_index());
+  for (intptr_t i = 0; i < smi_values_.length(); i++) {
+    smi_definitions_->Add(smi_values_[i]->ssa_temp_index());
+  }
+  for (intptr_t i = 0; i < constraints_.length(); i++) {
+    smi_definitions_->Add(constraints_[i]->ssa_temp_index());
+  }
+
+  // Infer initial values of ranges.
+  InitializeRangesRecursive(flow_graph_->graph_entry());
+
+  // Active worklist is empty, inactive now contains all smi values.
+  SwapWorklists();
+
+  // Iterate until fix point is reached.
+  while (!IsWorklistEmpty()) {
+    for (intptr_t i = 0; i < active_worklist_->length(); i++) {
+      Definition* defn = (*active_worklist_)[i];
+      if (defn->InferRange()) {  // Update the range.
+        // Range change. Place all uses to the worklist.
+        for (Value* use = defn->input_use_list();
+             use != NULL;
+             use = use->next_use()) {
+          Definition* use_defn = use->instruction()->AsDefinition();
+          if ((use_defn != NULL) &&
+              (use_defn->ssa_temp_index() != -1) &&

[Florian Schneider, 2012/09/21 08:56:44]

Can we still have a Definition with an ssa_index == -1? We also have use_kind()
which has to be kValue in this case.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Yes, unfortunately we do. For example StoreIndexedInstr claims that it is a
definition with use_kind_ kValue but it has no SSA index because it obviously
produces no value.

+              smi_definitions_->Contains(use_defn->ssa_temp_index())) {
+            AddToWorklist(use_defn);

[Kevin Millikin (Google), 2012/09/21 08:59:33]

The operation of the worklist is not clear.  Here it appears to contain
definitions that need to be analyzed due to one of their input ranges changing.

When initialized it appears to contain definitions whose range has just changed
and so need to have their uses analyzed.

I guess it must work, but it seems to me that the algorithm would just empty the
initial worklist without concluding anything new.

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Worklist contains definition for which InferRange should be invoked because one
of their inputs changed its range.

Initially we just place everything to the worklist because all ranges changed.

After that this loop will start placing only defs that depend on changed values
to the worklist.

When the fixpoint is reached nothing changes --- all InferRange return false ---
and worklist becomes empty.

+          }
+        }
+      }
+    }
+
+    // Active worklist has been processed. Inactive contains all values which
+    // can be affected by changed ranges.
+    SwapWorklists();
+  }
+}
+
+
+void RangeAnalysis::RemoveConstraints() {
+  for (intptr_t i = 0; i < constraints_.length(); i++) {
+    Definition* def = constraints_[i]->value()->definition();
+    // Some constraints might be constraining constraints. Unwind the chain of
+    // constraints until

[Florian Schneider, 2012/09/21 08:56:44]

Until what?

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+    while (def->IsConstraint()) {
+      def = def->AsConstraint()->value()->definition();
+    }
+    constraints_[i]->ReplaceUsesWith(def);
+    constraints_[i]->RemoveDependency();
+    constraints_[i]->RemoveFromGraph();
+  }
+}
+
+
+void FlowGraphOptimizer::InferSmiRanges() {
+  RangeAnalysis range_analysis(flow_graph_);
+  range_analysis.Analyze();
+}
+
+
 void FlowGraphTypePropagator::VisitBlocks() {
   ASSERT(current_iterator_ == NULL);
   for (intptr_t i = 0; i < block_order_.length(); ++i) {
     BlockEntryInstr* entry = block_order_[i];
     entry->Accept(this);
     ForwardInstructionIterator it(entry);
     current_iterator_ = &it;
     for (; !it.Done(); it.Advance()) {
       Instruction* current = it.Current();
       // No need to propagate the input types of the instruction, as long as
@@ skipping 1100 matching lines @@
 void ConstantPropagator::VisitCheckSmi(CheckSmiInstr* instr) {
   // Nothing to do. Has no value.
 }
 
 
 void ConstantPropagator::VisitConstant(ConstantInstr* instr) {
   SetValue(instr, instr->value());
 }
 
 
+void ConstantPropagator::VisitConstraint(ConstraintInstr* instr) {
+  // Nothing to do. Is not used outside of range analysis.

[Florian Schneider, 2012/09/21 08:56:44]

Maybe also make it UNREACHABLE()?

[Vyacheslav Egorov (Google), 2012/09/21 20:08:07]

Done.

+}
+
+
 void ConstantPropagator::VisitCheckEitherNonSmi(CheckEitherNonSmiInstr* instr) {
   // Nothing to do. Not a value.
 }
 
 
 void ConstantPropagator::VisitUnboxedDoubleBinaryOp(
     UnboxedDoubleBinaryOpInstr* instr) {
   const Object& left = instr->left()->definition()->constant_value();
   const Object& right = instr->right()->definition()->constant_value();
   if (IsNonConstant(left) || IsNonConstant(right)) {
@@ skipping 151 matching lines @@
        it.Advance()) {
     JoinEntryInstr* join = it.Current()->AsJoinEntry();
     if (join != NULL) join->EliminateUnreachablePhiInputs();
   }
 
   graph_->ComputeUseLists();
 }
 
 
 }  // namespace dart