[turbofan] Rework Node guts to save space.

This reduces the storage per-Node storage from 7 words to 6 and per-edge
storage from 6 words to 4.
On average this is about 10%-15% space savings over the whole graph.

Remove the use of std::deque as the out-of-line storage for inputs.
Reduce size of Use links and use pointer arithmetic to find Node
from Use.

R=mstarzinger@chromium.org,jarin@chromium.org
BUG=

Review URL: https://codereview.chromium.org/1150923003

Cr-Commit-Position: refs/heads/master@{#28583}

test/cctest/compiler/test-control-reducer.cc

@@ -864,8 +864,8 @@ TEST(CMergeReduce_dead_rm1b) {
   Node* t = R.graph.NewNode(R.common.IfTrue(), R.start);
   Node* f = R.graph.NewNode(R.common.IfTrue(), R.start);
   for (int i = 0; i < 2; i++) {
-    Node* merge = R.graph.NewNode(R.common.Merge(3), R.dead, R.dead, R.dead);
     for (int j = i + 1; j < 3; j++) {
+      Node* merge = R.graph.NewNode(R.common.Merge(3), R.dead, R.dead, R.dead);
       merge->ReplaceInput(i, t);
       merge->ReplaceInput(j, f);
       R.ReduceMerge(merge, merge);

test/unittests/compiler/node-unittest.cc

@@ -233,6 +233,29 @@ TEST_F(NodeTest, TrimThenAppend) {
   EXPECT_THAT(node->inputs(), ElementsAre(n1, n3, n4, n5, n8, n9));
 }
 
+
+TEST_F(NodeTest, BigNodes) {
+  static const int kMaxSize = 512;
+  Node* inputs[kMaxSize];
+
+  Node* n0 = Node::New(zone(), 0, &kOp0, 0, nullptr, false);
+  Node* n1 = Node::New(zone(), 1, &kOp1, 1, &n0, false);
+
+  for (int i = 0; i < kMaxSize; i++) {
+    inputs[i] = i & 1 ? n0 : n1;
+  }
+
+  for (int size = 13; size <= kMaxSize; size += 9) {
+    Node* node = Node::New(zone(), 12345, &kOp0, size, inputs, false);
+    EXPECT_EQ(size, node->InputCount());
+
+    for (int i = 0; i < size; i++) {
+      EXPECT_EQ(inputs[i], node->InputAt(i));
+    }
+  }
+}
+
+
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8