[turbofan]: More optimizations to add and subtract operations on x64

src/compiler/x64/instruction-selector-x64.cc

 // Copyright 2014 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/compiler/generic-node-inl.h"
 #include "src/compiler/instruction-selector-impl.h"
 #include "src/compiler/node-matchers.h"
 
 namespace v8 {
 namespace internal {
@@ skipping 416 matching lines @@
 }
 
 }  // namespace
 
 
 void InstructionSelector::VisitInt32Add(Node* node) {
   // Try to match the Add to a leal pattern
   ScaledWithOffset32Matcher m(node);
   X64OperandGenerator g(this);
   if (m.matches() && (m.constant() == NULL || g.CanBeImmediate(m.constant()))) {
+    // The add can be represented as a "leal", but there may be a smaller
+    // representation that is better and no more expensive.
+    if (m.offset() != NULL) {
+      if (m.scaled() == NULL) {
+        if (!IsLive(m.offset())) {
+          // If the add is of the form (r1 + immediate) and the non-constant
+          // input to the add is owned by the add, then it doesn't need to be
+          // preserved across the operation, so use more compact,
+          // source-register-overwriting versions when they are available and
+          // smaller, e.g. "incl" and "decl".
+          int32_t value =
+              m.constant() == NULL ? 0 : OpParameter<int32_t>(m.constant());
+          if (value == 1) {
+            Emit(kX64Inc32, g.DefineSameAsFirst(node),
+                 g.UseRegister(m.offset()));
+            return;
+          } else if (value == -1) {
+            Emit(kX64Dec32, g.DefineSameAsFirst(node),
+                 g.UseRegister(m.offset()));
+            return;
+          }
+        }
+      }
+    }
+
     InstructionOperand* inputs[4];
     size_t input_count = 0;
-
     AddressingMode mode = GenerateMemoryOperandInputs(
         &g, m.scaled(), m.scale_exponent(), m.offset(), m.constant(), inputs,
         &input_count);
 
     DCHECK_NE(0, static_cast<int>(input_count));
     DCHECK_GE(arraysize(inputs), input_count);
 
     InstructionOperand* outputs[1];
     outputs[0] = g.DefineAsRegister(node);
 
@@ skipping 11 matching lines @@
   VisitBinop(this, node, kX64Add);
 }
 
 
 void InstructionSelector::VisitInt32Sub(Node* node) {
   X64OperandGenerator g(this);
   Int32BinopMatcher m(node);
   if (m.left().Is(0)) {
     Emit(kX64Neg32, g.DefineSameAsFirst(node), g.UseRegister(m.right().node()));
   } else {
+    if (m.right().HasValue() && g.CanBeImmediate(m.right().node())) {
+      // If the Non-constant input is owned by the subtract, using a "decl" or
+      // "incl" that overwrites that input is smaller and probably an overall
+      // win.
+      if (!IsLive(m.left().node())) {
+        if (m.right().Value() == 1) {
+          Emit(kX64Dec32, g.DefineSameAsFirst(node),
+               g.UseRegister(m.left().node()));
+          return;
+        }
+        if (m.right().Value() == -1) {
+          Emit(kX64Inc32, g.DefineSameAsFirst(node),
+               g.UseRegister(m.left().node()));
+          return;
+        }
+      } else {
+        // Special handling for subtraction of constants where the non-constant
+        // input is used elsewhere. To eliminate the gap move before the sub to
+        // copy the destination register, use a "leal" instead.
+        Emit(kX64Lea32 | AddressingModeField::encode(kMode_MRI),
+             g.DefineAsRegister(node), g.UseRegister(m.left().node()),
+             g.TempImmediate(-m.right().Value()));
+        return;
+      }
+    }
     VisitBinop(this, node, kX64Sub32);
   }
 }
 
 
 void InstructionSelector::VisitInt64Sub(Node* node) {
   X64OperandGenerator g(this);
   Int64BinopMatcher m(node);
   if (m.left().Is(0)) {
     Emit(kX64Neg, g.DefineSameAsFirst(node), g.UseRegister(m.right().node()));
@@ skipping 690 matching lines @@
     return MachineOperatorBuilder::kFloat64Floor |
            MachineOperatorBuilder::kFloat64Ceil |
            MachineOperatorBuilder::kFloat64RoundTruncate;
   }
   return MachineOperatorBuilder::kNoFlags;
 }
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8