Chromium Code Reviews| Index: src/compiler/state-values-utils.cc |
| diff --git a/src/compiler/state-values-utils.cc b/src/compiler/state-values-utils.cc |
| index e8310d7d56ed95cb4ebaaa75b86ed56379ea9940..9d8196299e46847080e3c62ad3e6de80e9ddaec4 100644 |
| --- a/src/compiler/state-values-utils.cc |
| +++ b/src/compiler/state-values-utils.cc |
| @@ -4,10 +4,25 @@ |
| #include "src/compiler/state-values-utils.h" |
| +#include "src/bit-vector.h" |
| + |
| namespace v8 { |
| namespace internal { |
| namespace compiler { |
| +// A (Typed)StateValues node's has a bitmask specifying if its inputs are |
| +// represented sparsely. If the bitmask value is 0, then the inputs are not |
| +// sparse; otherwise, they should be interpreted as follows: |
|
Jarin
2016/12/08 07:59:43
This explanation should probably be in common-oper
Leszek Swirski
2016/12/08 15:44:31
Done.
|
| +// |
| +// * The bitmask represents which values are live, with 1 for live values |
| +// and 0 for dead (optimized out) values. |
| +// * The inputs to the node are the live values, in the order of the 1s from |
| +// least- to most-significant |
| +// * The top bit of the bitmask is a guard indicating the end of the values, |
| +// whether live or dead (and is not representative of a live node) |
|
Jarin (Google)
2016/12/07 15:16:44
If really want to use such a complex encoding sche
Leszek Swirski
2016/12/08 15:44:30
Done.
|
| +// |
| +// So, for N 1s in the bitmask, there are N - 1 inputs into the node. |
| + |
| StateValuesCache::StateValuesCache(JSGraph* js_graph) |
| : js_graph_(js_graph), |
| hash_map_(AreKeysEqual, ZoneHashMap::kDefaultHashMapCapacity, |
| @@ -47,6 +62,14 @@ bool StateValuesCache::IsKeysEqualToNode(StateValuesKey* key, Node* node) { |
| if (key->count != static_cast<size_t>(node->InputCount())) { |
| return false; |
| } |
| + |
| + DCHECK(node->opcode() == IrOpcode::kStateValues); |
| + uint32_t node_mask = OpParameter<uint32_t>(node); |
|
Jarin (Google)
2016/12/07 15:16:44
No OpParameter, please!
Leszek Swirski
2016/12/08 15:44:30
Done.
|
| + |
| + if (node_mask != key->mask) { |
| + return false; |
| + } |
| + |
| for (size_t i = 0; i < key->count; i++) { |
| if (key->values[i] != node->InputAt(static_cast<int>(i))) { |
| return false; |
| @@ -62,6 +85,9 @@ bool StateValuesCache::AreValueKeysEqual(StateValuesKey* key1, |
| if (key1->count != key2->count) { |
| return false; |
| } |
| + if (key1->mask != key2->mask) { |
| + return false; |
| + } |
| for (size_t i = 0; i < key1->count; i++) { |
| if (key1->values[i] != key2->values[i]) { |
| return false; |
| @@ -73,17 +99,15 @@ bool StateValuesCache::AreValueKeysEqual(StateValuesKey* key1, |
| Node* StateValuesCache::GetEmptyStateValues() { |
| if (empty_state_values_ == nullptr) { |
| - empty_state_values_ = graph()->NewNode(common()->StateValues(0)); |
| + empty_state_values_ = graph()->NewNode(common()->StateValues(0, 0u)); |
| } |
| return empty_state_values_; |
| } |
| - |
| -NodeVector* StateValuesCache::GetWorkingSpace(size_t level) { |
| - while (working_space_.size() <= level) { |
| - void* space = zone()->New(sizeof(NodeVector)); |
| - working_space_.push_back(new (space) |
| - NodeVector(kMaxInputCount, nullptr, zone())); |
| +StateValuesCache::WorkingBuffer& StateValuesCache::GetWorkingSpace( |
| + size_t level) { |
| + if (working_space_.size() <= level) { |
| + working_space_.resize(level + 1); |
| } |
| return working_space_[level]; |
| } |
| @@ -93,16 +117,16 @@ namespace { |
| int StateValuesHashKey(Node** nodes, size_t count) { |
| size_t hash = count; |
| for (size_t i = 0; i < count; i++) { |
| - hash = hash * 23 + nodes[i]->id(); |
| + hash = hash * 23 + (nodes[i] == nullptr ? 0 : nodes[i]->id()); |
| } |
| return static_cast<int>(hash & 0x7fffffff); |
| } |
| } // namespace |
| - |
| -Node* StateValuesCache::GetValuesNodeFromCache(Node** nodes, size_t count) { |
| - StateValuesKey key(count, nodes); |
| +Node* StateValuesCache::GetValuesNodeFromCache(Node** nodes, size_t count, |
| + uint32_t mask) { |
| + StateValuesKey key(count, mask, nodes); |
| int hash = StateValuesHashKey(nodes, count); |
| ZoneHashMap::Entry* lookup = |
| hash_map_.LookupOrInsert(&key, hash, ZoneAllocationPolicy(zone())); |
| @@ -110,8 +134,8 @@ Node* StateValuesCache::GetValuesNodeFromCache(Node** nodes, size_t count) { |
| Node* node; |
| if (lookup->value == nullptr) { |
| int input_count = static_cast<int>(count); |
| - node = graph()->NewNode(common()->StateValues(input_count), input_count, |
| - nodes); |
| + node = graph()->NewNode(common()->StateValues(input_count, mask), |
| + input_count, nodes); |
| NodeKey* new_key = new (zone()->New(sizeof(NodeKey))) NodeKey(node); |
| lookup->key = new_key; |
| lookup->value = node; |
| @@ -121,91 +145,175 @@ Node* StateValuesCache::GetValuesNodeFromCache(Node** nodes, size_t count) { |
| return node; |
| } |
| +Node* StateValuesCache::BuildTree(size_t& idx, Node** values, size_t count, |
| + const BitVector* liveness, size_t level) { |
| + WorkingBuffer& input_buffer = GetWorkingSpace(level); |
| + size_t input_count = 0; |
| + bool use_mask = false; |
| + uint32_t mask = 0; |
| + |
| + if (level == 0) { |
| + // Virtual inputs are the live inputs plus the implicit dead inputs, which |
| + // are implied by the liveness mask. |
| + size_t virtual_input_count = 0; |
| + while (idx < count && input_count < kMaxInputCount && |
| + (!use_mask || virtual_input_count < 31)) { |
|
Jarin (Google)
2016/12/07 15:16:44
Magic constants (31) defined at the top of the fun
Leszek Swirski
2016/12/08 15:44:31
Done (moved into the new sparse input mask type).
|
| + DCHECK_LE(idx, static_cast<size_t>(INT_MAX)); |
| + if (liveness == nullptr || liveness->Contains(static_cast<int>(idx))) { |
| + mask |= 1 << virtual_input_count; |
| + input_buffer[input_count++] = values[idx]; |
| + } else { |
| + use_mask = true; |
|
Jarin (Google)
2016/12/07 15:16:44
What improvement do you see from special casing th
Leszek Swirski
2016/12/08 15:44:31
I haven't measured it, but I've removed it since t
|
| + } |
| + virtual_input_count++; |
| -class StateValuesCache::ValueArrayIterator { |
| - public: |
| - ValueArrayIterator(Node** values, size_t count) |
| - : values_(values), count_(count), current_(0) {} |
| - |
| - void Advance() { |
| - if (!done()) { |
| - current_++; |
| + idx++; |
| } |
| - } |
| - |
| - bool done() { return current_ >= count_; } |
| - |
| - Node* node() { |
| - DCHECK(!done()); |
| - return values_[current_]; |
| - } |
| - |
| - private: |
| - Node** values_; |
| - size_t count_; |
| - size_t current_; |
| -}; |
| - |
| -Node* StateValuesCache::BuildTree(ValueArrayIterator* it, size_t max_height) { |
| - if (max_height == 0) { |
| - Node* node = it->node(); |
| - it->Advance(); |
| - return node; |
| + if (use_mask) { |
| + DCHECK(virtual_input_count < 32); |
| + mask |= 1 << virtual_input_count; |
| + } else { |
| + mask = 0; |
| + } |
| + } else { |
| + while (idx < count && input_count < kMaxInputCount) { |
| + if (count - idx < kMaxInputCount - input_count) { |
| + // If we have fewer values remaining than inputs remaining, dump the |
| + // remaining values into this node. |
| + |
| + // TODO(leszeks): We could optimise this further by counting remaining |
| + // live nodes, though this gets complicated with the 31 bit limit on the |
| + // mask. |
|
Jarin (Google)
2016/12/07 15:16:44
Maybe the remove the comment, the complexity budge
Leszek Swirski
2016/12/08 15:44:30
Kept the comment (shortened it a bit), after refac
|
| + |
| + // All previous inputs are live. |
| + mask = ((1 << input_count) - 1); |
| + |
| + // Add the remaining values as inputs. |
| + size_t virtual_input_count = input_count; |
| + while (idx < count) { |
| + DCHECK_LE(input_count, kMaxInputCount); |
| + DCHECK_LE(idx, static_cast<size_t>(INT_MAX)); |
| + DCHECK(!use_mask || virtual_input_count < 31); |
| + |
| + if (liveness == nullptr || |
| + liveness->Contains(static_cast<int>(idx))) { |
|
Jarin (Google)
2016/12/07 15:16:44
This looks very similar to the code above, perhaps
Leszek Swirski
2016/12/08 15:44:30
Done, I had a TODO for it at some point anyway.
|
| + mask |= 1 << virtual_input_count; |
| + input_buffer[input_count++] = values[idx]; |
| + } else { |
| + use_mask = true; |
| + } |
| + virtual_input_count++; |
| + |
| + idx++; |
| + } |
| + |
| + if (use_mask) { |
| + DCHECK(virtual_input_count < 32); |
| + mask |= 1 << virtual_input_count; |
| + } else { |
| + mask = 0; |
| + } |
| + } else { |
| + // Otherwise, add the values to a subtree and add that as an input. |
| + Node* subtree = BuildTree(idx, values, count, liveness, level - 1); |
| + input_buffer[input_count++] = subtree; |
| + } |
| + } |
| } |
| - DCHECK(!it->done()); |
| - NodeVector* buffer = GetWorkingSpace(max_height); |
| - size_t count = 0; |
| - for (; count < kMaxInputCount; count++) { |
| - if (it->done()) break; |
| - (*buffer)[count] = BuildTree(it, max_height - 1); |
| - } |
| - if (count == 1) { |
| - return (*buffer)[0]; |
| + if (input_count == 1 && !use_mask) { |
| + // Elide the StateValue node if there is only one input. |
| + return input_buffer[0]; |
| } else { |
| - return GetValuesNodeFromCache(&(buffer->front()), count); |
| + return GetValuesNodeFromCache(input_buffer.data(), input_count, mask); |
| } |
| } |
| - |
| -Node* StateValuesCache::GetNodeForValues(Node** values, size_t count) { |
| +Node* StateValuesCache::GetNodeForValues(Node** values, size_t count, |
| + const BitVector* liveness) { |
| #if DEBUG |
| for (size_t i = 0; i < count; i++) { |
| - DCHECK_NE(values[i]->opcode(), IrOpcode::kStateValues); |
| - DCHECK_NE(values[i]->opcode(), IrOpcode::kTypedStateValues); |
| + if (values[i] != nullptr) { |
| + DCHECK_NE(values[i]->opcode(), IrOpcode::kStateValues); |
| + DCHECK_NE(values[i]->opcode(), IrOpcode::kTypedStateValues); |
| + } |
| + } |
| + if (liveness != nullptr) { |
| + // Liveness can have extra bits for the stack or accumulator, which we |
| + // ignore here. |
| + DCHECK_LE(count, static_cast<size_t>(liveness->length())); |
| + |
| + for (size_t i = 0; i < count; i++) { |
| + if (liveness->Contains(static_cast<int>(i))) { |
| + DCHECK_NOT_NULL(values[i]); |
| + } |
| + } |
| } |
| #endif |
| if (count == 0) { |
| return GetEmptyStateValues(); |
| } |
| + |
| + // This is a worst-case tree height estimate, assuming that all values are |
| + // live. We could get a better estimate by counting zeroes in the liveness |
| + // vector, but there's no point -- any excess height in the tree will be |
| + // collapsed by the single-input elision at the end of BuildTree. |
| size_t height = 0; |
| - size_t max_nodes = 1; |
| - while (count > max_nodes) { |
| + size_t max_inputs = kMaxInputCount; |
| + while (count > max_inputs) { |
| height++; |
| - max_nodes *= kMaxInputCount; |
| + max_inputs *= kMaxInputCount; |
| } |
| - ValueArrayIterator it(values, count); |
| - |
| - Node* tree = BuildTree(&it, height); |
| + size_t idx = 0; |
| + Node* tree = BuildTree(idx, values, count, liveness, height); |
| // If the 'tree' is a single node, equip it with a StateValues wrapper. |
| - if (tree->opcode() != IrOpcode::kStateValues && |
| - tree->opcode() != IrOpcode::kTypedStateValues) { |
| - tree = GetValuesNodeFromCache(&tree, 1); |
| + if (tree->opcode() != IrOpcode::kStateValues) { |
| + tree = GetValuesNodeFromCache(&tree, 1, 0u); |
| } |
| +#if DEBUG |
| + { |
| + DCHECK_EQ(count, StateValuesAccess(tree).size()); |
| + int i; |
| + auto access = StateValuesAccess(tree); |
| + auto it = access.begin(); |
| + auto itend = access.end(); |
| + for (i = 0; it != itend; ++it, ++i) { |
| + if (liveness == nullptr || liveness->Contains(i)) { |
| + DCHECK((*it).node == values[i]); |
| + } else { |
| + DCHECK((*it).node == nullptr); |
| + } |
| + } |
| + DCHECK_EQ(static_cast<size_t>(i), count); |
| + } |
| +#endif |
| + |
| return tree; |
| } |
| +namespace { |
| + |
| +uint32_t GetStateValueMask(Node* node) { |
|
Jarin (Google)
2016/12/07 15:16:44
This should live somewhere in common-operator.h
Leszek Swirski
2016/12/08 15:44:30
Done.
|
| + if (node->opcode() == IrOpcode::kStateValues) { |
| + return OpParameter<uint32_t>(node); |
| + } else { |
| + DCHECK_EQ(node->opcode(), IrOpcode::kTypedStateValues); |
| + return OpParameter<TypedStateValueInfo>(node).mask(); |
| + } |
| +} |
| + |
| +} // namespace |
| StateValuesAccess::iterator::iterator(Node* node) : current_depth_(0) { |
| - // A hacky way initialize - just set the index before the node we want |
| - // to process and then advance to it. |
| stack_[current_depth_].node = node; |
| - stack_[current_depth_].index = -1; |
| - Advance(); |
| + stack_[current_depth_].index = 0; |
| + stack_[current_depth_].mask = GetStateValueMask(node); |
| + |
| + EnsureValid(); |
| } |
| @@ -215,11 +323,11 @@ StateValuesAccess::iterator::StatePos* StateValuesAccess::iterator::Top() { |
| return &(stack_[current_depth_]); |
| } |
| - |
| -void StateValuesAccess::iterator::Push(Node* node) { |
| +void StateValuesAccess::iterator::Push(Node* node, uint32_t mask) { |
| current_depth_++; |
| CHECK(current_depth_ < kMaxInlineDepth); |
| stack_[current_depth_].node = node; |
| + stack_[current_depth_].mask = mask; |
| stack_[current_depth_].index = 0; |
| } |
| @@ -234,37 +342,63 @@ bool StateValuesAccess::iterator::done() { return current_depth_ < 0; } |
| void StateValuesAccess::iterator::Advance() { |
| - // Advance the current index. |
| - Top()->index++; |
| + MoveToNextSibling(); |
| + EnsureValid(); |
| +} |
| - // Fix up the position to point to a valid node. |
| +void StateValuesAccess::iterator::MoveToNextSibling() { |
| + int mask = Top()->mask; |
| + if (mask == 0 || (mask & 0x1) == 1) { |
|
Jarin (Google)
2016/12/07 15:16:44
Nit: Why 0x1? On line 446, you say mask & 1. Perha
Leszek Swirski
2016/12/08 15:44:31
Completely refactored out into sparse input iterat
|
| + Top()->index++; |
| + } |
| + Top()->mask >>= 1; |
| +} |
| + |
| +void StateValuesAccess::iterator::EnsureValid() { |
| while (true) { |
| - // TODO(jarin): Factor to a separate method. |
| - Node* node = Top()->node; |
| + uint32_t mask = Top()->mask; |
| int index = Top()->index; |
| + Node* node = Top()->node; |
| - if (index >= node->InputCount()) { |
| - // Pop stack and move to the next sibling. |
| + if (mask != 0 && (mask & 0x1) == 0) { |
| + // We are on a valid (dead) node. |
|
Jarin (Google)
2016/12/07 15:16:44
dead -> optimized_out here and elsewhere
Leszek Swirski
2016/12/08 15:44:30
Done.
|
| + return; |
| + } |
| + |
| + if (mask == 1 || (mask == 0 && index >= node->InputCount())) { |
| + // We have hit the guard bit or exhausted our inputs. Pop the stack and |
| + // move to the next sibling. |
| Pop(); |
| if (done()) { |
| // Stack is exhausted, we have reached the end. |
| return; |
| } |
| - Top()->index++; |
| - } else if (node->InputAt(index)->opcode() == IrOpcode::kStateValues || |
| - node->InputAt(index)->opcode() == IrOpcode::kTypedStateValues) { |
| + MoveToNextSibling(); |
| + continue; |
| + } |
| + |
| + // At this point the value is known to be live and within our input nodes. |
| + Node* value_node = node->InputAt(Top()->index); |
| + |
| + if (value_node->opcode() == IrOpcode::kStateValues || |
| + value_node->opcode() == IrOpcode::kTypedStateValues) { |
| // Nested state, we need to push to the stack. |
| - Push(node->InputAt(index)); |
| - } else { |
| - // We are on a valid node, we can stop the iteration. |
| - return; |
| + Push(node->InputAt(index), GetStateValueMask(node->InputAt(index))); |
| + continue; |
| } |
| + |
| + // We are on a valid node, we can stop the iteration. |
| + return; |
| } |
| } |
| Node* StateValuesAccess::iterator::node() { |
| - return Top()->node->InputAt(Top()->index); |
| + if (Top()->mask != 0 && (Top()->mask & 0x1) == 0) { |
| + return nullptr; |
| + } else { |
| + return Top()->node->InputAt(Top()->index); |
| + } |
| } |
| @@ -274,8 +408,13 @@ MachineType StateValuesAccess::iterator::type() { |
| return MachineType::AnyTagged(); |
| } else { |
| DCHECK_EQ(IrOpcode::kTypedStateValues, state->opcode()); |
| - ZoneVector<MachineType> const* types = MachineTypesOf(state->op()); |
| - return (*types)[Top()->index]; |
| + |
| + if (Top()->mask != 0 && (Top()->mask & 0x1) == 0) { |
| + return MachineType::None(); |
| + } else { |
| + ZoneVector<MachineType> const* types = MachineTypesOf(state->op()); |
| + return (*types)[Top()->index]; |
| + } |
| } |
| } |
| @@ -300,14 +439,24 @@ StateValuesAccess::TypedNode StateValuesAccess::iterator::operator*() { |
| size_t StateValuesAccess::size() { |
| size_t count = 0; |
| - for (int i = 0; i < node_->InputCount(); i++) { |
| - if (node_->InputAt(i)->opcode() == IrOpcode::kStateValues || |
| - node_->InputAt(i)->opcode() == IrOpcode::kTypedStateValues) { |
| - count += StateValuesAccess(node_->InputAt(i)).size(); |
| - } else { |
| + uint32_t mask = GetStateValueMask(node_); |
| + |
| + int i = 0; |
| + while ((mask == 0 && i < node_->InputCount()) || (mask != 0 && mask != 1)) { |
| + if (mask != 0 && (mask & 1) == 0) { |
| count++; |
| + } else { |
| + if (node_->InputAt(i)->opcode() == IrOpcode::kStateValues || |
| + node_->InputAt(i)->opcode() == IrOpcode::kTypedStateValues) { |
| + count += StateValuesAccess(node_->InputAt(i)).size(); |
| + } else { |
| + count++; |
| + } |
| + i++; |
| } |
| + mask >>= 1; |
| } |
| + |
| return count; |
| } |