Encode interpreter::SourcePositionTable as variable-length ints.

src/interpreter/source-position-table.cc

Index: src/interpreter/source-position-table.cc
diff --git a/src/interpreter/source-position-table.cc b/src/interpreter/source-position-table.cc
index 0b7c44e2d984a50eb8169af7daba6ce8af096854..922fb8a155afbc30d03ef901411ac1eb32f20260 100644
--- a/src/interpreter/source-position-table.cc
+++ b/src/interpreter/source-position-table.cc
@@ -4,7 +4,6 @@
 
 #include "src/interpreter/source-position-table.h"
 
-#include "src/assembler.h"
 #include "src/objects-inl.h"
 #include "src/objects.h"
 
@@ -12,29 +11,154 @@ namespace v8 {
 namespace internal {
 namespace interpreter {
 
-class IsStatementField : public BitField<bool, 0, 1> {};
-class SourcePositionField : public BitField<int, 1, 30> {};
+// We'll use a simple encoding scheme to record the source positions.
+// Conceptually, each position consists of:
+// - bytecode_offset: An integer index into the BytecodeArray
+// - source_position: An integer index into the source string.
+// - position type: Each position is either a statement or an expression.
+//
+// The basic idea for the encoding is to use a variable-length integer coding,
+// where each byte contains 7 bits of payload data, and 1 'more' bit that
+// determines whether additional bytes follow. Additionally:
+// - we record the difference from the previous position,
+// - we just stuff one bit for the type into the bytecode offset,
+// - we write least-significant bits first,
+// - negative numbers occur only rarely, so we use a denormalized
+//   most-significant byte (a byte with all zeros, which normally wouldn't make
+//   any sense)
+//   to encode a negative sign, so that we 'pay' nothing for positive numbers,
+//   but have to pay a full byte for negative integers.
+
+namespace {
+// Each byte is encoded as MoreBit | ValueBits.
+class MoreBit : public BitField8<bool, 7, 1> {};
+class ValueBits : public BitField8<int, 0, 7> {};
+
+// Bytecode offsets also include a bit for the position type in its least
+// significant digit.
+// Beware that the bytecode offset can be negative, and the default BitField
+// implementation might not do the right thing in this case.
+class TypeBit : public BitField<bool, 0, 1> {};
+class BytecodeOffsetBits : public BitField<int, 1, 31> {};
+
+// Helper: Add the offsets from 'other' to 'value'. Also set is_statement.
+void AddAndSetEntry(PositionTableEntry& value,
+                    const PositionTableEntry& other) {
+  value.bytecode_offset += other.bytecode_offset;
+  value.source_position += other.source_position;
+  value.is_statement = other.is_statement;
+}
+
+// Helper: Substract the offsets from 'other' from 'value'.
+void SubtractFromEntry(PositionTableEntry& value,
+                       const PositionTableEntry& other) {
+  value.bytecode_offset -= other.bytecode_offset;
+  value.source_position -= other.source_position;
+}
+
+// Helper: Encode an integer.
+void EncodeInt(ZoneVector<byte>& bytes, int value) {
+  DCHECK(abs(value) < (1 << 30));

[rmcilroy, 2016/02/22 12:38:40]

Is there a reason this DCHECK is needed?

[vogelheim, 2016/02/22 18:33:33]

Done.

+  bool sign = false;
+  if (value < 0) {
+    sign = true;
+    value = -value;
+  }
+
+  bool more;
+  do {
+    more = value > ValueBits::kMax;
+    bytes.push_back(MoreBit::encode(more || sign) |
+                    ValueBits::encode(value & ValueBits::kMax));
+    value >>= ValueBits::kSize;
+  } while (more);
+
+  if (sign) {
+    bytes.push_back(MoreBit::encode(false));
+  }
+}
+
+// Encode a PositionTableEntry.
+void EncodeEntry(ZoneVector<byte>& bytes, const PositionTableEntry& entry) {
+  EncodeInt(bytes, TypeBit::encode(entry.is_statement) |
+                       BytecodeOffsetBits::encode(entry.bytecode_offset &
+                                                  BytecodeOffsetBits::kMax));

[rmcilroy, 2016/02/22 12:38:40]

Could you add a DCHECK that entry.bytecode_offset < BytecodeOffsetBits::kMax

[vogelheim, 2016/02/22 18:33:33]

Done.

+  EncodeInt(bytes, entry.source_position);
+}
+
+// Helper: Scan one value backwards.
+// *index is expected to point behind the current 'complete' value encoding
+// and will be set to point behind he previous one.
+void EncodeScanBackwards(ZoneVector<byte>& bytes, int* index) {
+  DCHECK(bytes.size() > 0);
+  DCHECK(*index > 0);
+
+  // Index is expected to point behind a 'complete' value encoding.
+  (*index)--;
+  DCHECK(!MoreBit::decode(bytes[*index]));
+
+  // Scan back until we either hit the beginning of our bytes, or another byte
+  // without 'more' bit.
+  while (*index > 0 && MoreBit::decode(bytes[(*index) - 1])) {
+    (*index)--;
+  }
+  DCHECK(*index >= 0);
+}
+
+// Helper: Decode an integer.
+void DecodeInt(const byte* bytes, int* index, int* v) {
+  byte current;
+  int n = 0;
+  int value = 0;
+  bool more;
+  do {
+    current = bytes[(*index)++];
+    value |= ValueBits::decode(current) << (n * ValueBits::kSize);
+    n++;
+    more = MoreBit::decode(current);
+  } while (more);
+
+  if (ValueBits::decode(current) == 0) {
+    value = -value;
+  }
+  *v = value;
+}
+
+void DecodeEntry(const byte* bytes, int* index, PositionTableEntry* entry) {
+  int tmp;
+  DecodeInt(bytes, index, &tmp);
+  entry->is_statement = TypeBit::decode(tmp);
+
+  // Directly shift tmp, because '>>' needs to be arithmetic shift in order to
+  // handle negative numbers properly.
+  entry->bytecode_offset = (tmp >> BytecodeOffsetBits::kShift);

[rmcilroy, 2016/02/22 12:38:40]

Personally I would just avoid using TypeBit / BytecodeOffsetBits entirely since
you need to do the shift explicitly here - I think that would be simpler and
possibly safer, WDYT?

[vogelheim, 2016/02/22 18:33:33]

Done.

+
+  DecodeInt(bytes, index, &entry->source_position);
+}
+
+}  // namespace
 
 void SourcePositionTableBuilder::AddStatementPosition(size_t bytecode_offset,
                                                       int source_position) {
-  int offset = static_cast<int>(bytecode_offset);
-  // If a position has already been assigned to this bytecode offset,
-  // do not reassign a new statement position.
-  if (CodeOffsetHasPosition(offset)) return;
-  uint32_t encoded = IsStatementField::encode(true) |
-                     SourcePositionField::encode(source_position);
-  entries_.push_back({offset, encoded});
+  AddEntry({static_cast<int>(bytecode_offset), source_position, true});
 }
 
 void SourcePositionTableBuilder::AddExpressionPosition(size_t bytecode_offset,
                                                        int source_position) {
-  int offset = static_cast<int>(bytecode_offset);
-  // If a position has already been assigned to this bytecode offset,
-  // do not reassign a new statement position.
-  if (CodeOffsetHasPosition(offset)) return;
-  uint32_t encoded = IsStatementField::encode(false) |
-                     SourcePositionField::encode(source_position);
-  entries_.push_back({offset, encoded});
+  AddEntry({static_cast<int>(bytecode_offset), source_position, false});
+}
+
+void SourcePositionTableBuilder::AddEntry(const PositionTableEntry& entry) {
+  if (!CodeOffsetHasPosition(entry.bytecode_offset)) {
+    PositionTableEntry tmp(entry);
+    SubtractFromEntry(tmp, previous_);
+    EncodeEntry(bytes_, tmp);
+    previous_ = entry;
+
+#ifdef ENABLE_SLOW_DCHECKS
+    raw_entries_.push_back(entry);
+#endif
+  }
 }
 
 void SourcePositionTableBuilder::RevertPosition(size_t bytecode_offset) {
@@ -42,41 +166,68 @@ void SourcePositionTableBuilder::RevertPosition(size_t bytecode_offset) {
   // If we already added a source position table entry, but the bytecode array
   // builder ended up not outputting a bytecode for the corresponding bytecode
   // offset, we have to remove that entry.
-  if (CodeOffsetHasPosition(offset)) entries_.pop_back();
+  if (CodeOffsetHasPosition(offset)) {
+    // Reverting means we need to discard 2 variable length ints. We also need
+    // to read them, in order to fixup the previous_*_ member variables.
+
+    // Scan backwards for 2 values.
+    int index = static_cast<int>(bytes_.size());
+    EncodeScanBackwards(bytes_, &index);
+    EncodeScanBackwards(bytes_, &index);
+
+    // Read from index to adjust previous_.
+    int tmp_index = index;
+    PositionTableEntry tmp_entry;
+    DecodeEntry(&*bytes_.begin(), &tmp_index, &tmp_entry);
+    SubtractFromEntry(previous_, tmp_entry);
+
+    bytes_.resize(index);
+
+#ifdef ENABLE_SLOW_DCHECKS
+    raw_entries_.pop_back();
+#endif
+  }
 }
 
-Handle<FixedArray> SourcePositionTableBuilder::ToFixedArray() {
-  int length = static_cast<int>(entries_.size());
-  Handle<FixedArray> table =
-      isolate_->factory()->NewFixedArray(length * 2, TENURED);
-  for (int i = 0; i < length; i++) {
-    table->set(i * 2, Smi::FromInt(entries_[i].bytecode_offset));
-    table->set(i * 2 + 1, Smi::FromInt(entries_[i].source_position_and_type));
+Handle<ByteArray> SourcePositionTableBuilder::ToSourcePositionTable() {
+  Handle<ByteArray> table = isolate_->factory()->NewByteArray(
+      static_cast<int>(bytes_.size()), TENURED);
+  if (bytes_.empty()) return table;
+
+  MemCopy(table->GetDataStartAddress(), &*bytes_.begin(), bytes_.size());
+
+#ifdef ENABLE_SLOW_DCHECKS
+  // Brute force testing: Record all positions and decode
+  // the entire table to verify they are identical.
+  auto raw = raw_entries_.begin();
+  for (SourcePositionTableIterator encoded(*table); !encoded.done();
+       encoded.Advance(), raw++) {
+    DCHECK(raw != raw_entries_.end());
+    DCHECK_EQ(encoded.bytecode_offset(), raw->bytecode_offset);
+    DCHECK_EQ(encoded.source_position(), raw->source_position);
+    DCHECK_EQ(encoded.is_statement(), raw->is_statement);
   }
+  DCHECK(raw == raw_entries_.end());
+#endif
+
   return table;
 }
 
-SourcePositionTableIterator::SourcePositionTableIterator(
-    BytecodeArray* bytecode_array)
-    : table_(bytecode_array->source_position_table()),
-      index_(0),
-      length_(table_->length()) {
-  DCHECK(table_->length() % 2 == 0);
+SourcePositionTableIterator::SourcePositionTableIterator(ByteArray* byte_array)
+    : table_(byte_array), index_(0), current_() {
   Advance();
 }
 
 void SourcePositionTableIterator::Advance() {
-  if (index_ < length_) {
-    int new_bytecode_offset = Smi::cast(table_->get(index_))->value();
-    // Bytecode offsets are in ascending order.
-    DCHECK(bytecode_offset_ < new_bytecode_offset || index_ == 0);
-    bytecode_offset_ = new_bytecode_offset;
-    uint32_t source_position_and_type =
-        static_cast<uint32_t>(Smi::cast(table_->get(index_ + 1))->value());
-    is_statement_ = IsStatementField::decode(source_position_and_type);
-    source_position_ = SourcePositionField::decode(source_position_and_type);
+  DCHECK(!done());
+  DCHECK(index_ >= 0 && index_ <= table_->length());
+  if (index_ == table_->length()) {
+    index_ = kDone;
+  } else {
+    PositionTableEntry tmp;
+    DecodeEntry(table_->GetDataStartAddress(), &index_, &tmp);
+    AddAndSetEntry(current_, tmp);
   }
-  index_ += 2;
 }
 
 }  // namespace interpreter