[turbofan] Optimize add operations to use 'leal' instruction on x64

Add MemoryOperandMatcher that recognizes node clusters in the form
[%r1 + %r2*SCALE + OFFSET] and explicit support in the x64 Int32Add
selector to use it to translate complex adds to 'leal' instructions.

R=titzer@chromium.org

Review URL: https://codereview.chromium.org/704713003

Cr-Commit-Position: refs/heads/master@{#25223}
git-svn-id: https://v8.googlecode.com/svn/branches/bleeding_edge@25223 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/compiler/node-matchers.h

@@ -5,6 +5,8 @@
 #ifndef V8_COMPILER_NODE_MATCHERS_H_
 #define V8_COMPILER_NODE_MATCHERS_H_
 
+#include "src/compiler/generic-node.h"
+#include "src/compiler/generic-node-inl.h"
 #include "src/compiler/node.h"
 #include "src/compiler/operator.h"
 #include "src/unique.h"
@@ -116,7 +118,7 @@ struct HeapObjectMatcher FINAL
 // right hand sides of a binary operation and can put constants on the right
 // if they appear on the left hand side of a commutative operation.
 template <typename Left, typename Right>
-struct BinopMatcher FINAL : public NodeMatcher {
+struct BinopMatcher : public NodeMatcher {
   explicit BinopMatcher(Node* node)
       : NodeMatcher(node), left_(InputAt(0)), right_(InputAt(1)) {
     if (HasProperty(Operator::kCommutative)) PutConstantOnRight();
@@ -128,12 +130,17 @@ struct BinopMatcher FINAL : public NodeMatcher {
   bool IsFoldable() const { return left().HasValue() && right().HasValue(); }
   bool LeftEqualsRight() const { return left().node() == right().node(); }
 
+ protected:
+  void SwapInputs() {
+    std::swap(left_, right_);
+    node()->ReplaceInput(0, left().node());
+    node()->ReplaceInput(1, right().node());
+  }
+
  private:
   void PutConstantOnRight() {
     if (left().HasValue() && !right().HasValue()) {
-      std::swap(left_, right_);
-      node()->ReplaceInput(0, left().node());
-      node()->ReplaceInput(1, right().node());
+      SwapInputs();
     }
   }
 
@@ -150,6 +157,189 @@ typedef BinopMatcher<UintPtrMatcher, UintPtrMatcher> UintPtrBinopMatcher;
 typedef BinopMatcher<Float64Matcher, Float64Matcher> Float64BinopMatcher;
 typedef BinopMatcher<NumberMatcher, NumberMatcher> NumberBinopMatcher;
 
+struct Int32AddMatcher : public Int32BinopMatcher {
+  explicit Int32AddMatcher(Node* node)
+      : Int32BinopMatcher(node), scale_exponent_(-1) {
+    PutScaledInputOnLeft();
+  }
+
+  bool HasScaledInput() const { return scale_exponent_ != -1; }
+  Node* ScaledInput() const {
+    DCHECK(HasScaledInput());
+    return left().node()->InputAt(0);
+  }
+  int ScaleExponent() const {
+    DCHECK(HasScaledInput());
+    return scale_exponent_;
+  }
+
+ private:
+  int GetInputScaleExponent(Node* node) const {
+    if (node->opcode() == IrOpcode::kWord32Shl) {
+      Int32BinopMatcher m(node);
+      if (m.right().HasValue()) {
+        int32_t value = m.right().Value();
+        if (value >= 0 && value <= 3) {
+          return value;
+        }
+      }
+    } else if (node->opcode() == IrOpcode::kInt32Mul) {
+      Int32BinopMatcher m(node);
+      if (m.right().HasValue()) {
+        int32_t value = m.right().Value();
+        if (value == 1) {
+          return 0;
+        } else if (value == 2) {
+          return 1;
+        } else if (value == 4) {
+          return 2;
+        } else if (value == 8) {
+          return 3;
+        }
+      }
+    }
+    return -1;
+  }
+
+  void PutScaledInputOnLeft() {
+    scale_exponent_ = GetInputScaleExponent(right().node());
+    if (scale_exponent_ >= 0) {
+      int left_scale_exponent = GetInputScaleExponent(left().node());
+      if (left_scale_exponent == -1) {
+        SwapInputs();
+      } else {
+        scale_exponent_ = left_scale_exponent;
+      }
+    } else {
+      scale_exponent_ = GetInputScaleExponent(left().node());
+      if (scale_exponent_ == -1) {
+        if (right().opcode() == IrOpcode::kInt32Add &&
+            left().opcode() != IrOpcode::kInt32Add) {
+          SwapInputs();
+        }
+      }
+    }
+  }
+
+  int scale_exponent_;
+};
+
+struct ScaledWithOffsetMatcher {
+  explicit ScaledWithOffsetMatcher(Node* node)
+      : matches_(false),
+        scaled_(NULL),
+        scale_exponent_(0),
+        offset_(NULL),
+        constant_(NULL) {
+    if (node->opcode() != IrOpcode::kInt32Add) return;
+
+    // The Int32AddMatcher canonicalizes the order of constants and scale
+    // factors that are used as inputs, so instead of enumerating all possible
+    // patterns by brute force, checking for node clusters using the following
+    // templates in the following order suffices to find all of the interesting
+    // cases (S = scaled input, O = offset input, C = constant input):
+    // (S + (O + C))
+    // (S + (O + O))
+    // (S + C)
+    // (S + O)
+    // ((S + C) + O)
+    // ((S + O) + C)
+    // ((O + C) + O)
+    // ((O + O) + C)
+    // (O + C)
+    // (O + O)
+    Int32AddMatcher base_matcher(node);
+    Node* left = base_matcher.left().node();
+    Node* right = base_matcher.right().node();
+    if (base_matcher.HasScaledInput() && left->OwnedBy(node)) {
+      scaled_ = base_matcher.ScaledInput();
+      scale_exponent_ = base_matcher.ScaleExponent();
+      if (right->opcode() == IrOpcode::kInt32Add && right->OwnedBy(node)) {
+        Int32AddMatcher right_matcher(right);
+        if (right_matcher.right().HasValue()) {
+          // (S + (O + C))
+          offset_ = right_matcher.left().node();
+          constant_ = right_matcher.right().node();
+        } else {
+          // (S + (O + O))
+          offset_ = right;
+        }
+      } else if (base_matcher.right().HasValue()) {
+        // (S + C)
+        constant_ = right;
+      } else {
+        // (S + O)
+        offset_ = right;
+      }
+    } else {
+      if (left->opcode() == IrOpcode::kInt32Add && left->OwnedBy(node)) {
+        Int32AddMatcher left_matcher(left);
+        Node* left_left = left_matcher.left().node();
+        Node* left_right = left_matcher.right().node();
+        if (left_matcher.HasScaledInput() && left_left->OwnedBy(left)) {
+          scaled_ = left_matcher.ScaledInput();
+          scale_exponent_ = left_matcher.ScaleExponent();
+          if (left_matcher.right().HasValue()) {
+            // ((S + C) + O)
+            constant_ = left_right;
+            offset_ = right;
+          } else if (base_matcher.right().HasValue()) {
+            // ((S + O) + C)
+            offset_ = left_right;
+            constant_ = right;
+          } else {
+            // (O + O)
+            scaled_ = left;
+            offset_ = right;
+          }
+        } else {
+          if (left_matcher.right().HasValue()) {
+            // ((O + C) + O)
+            scaled_ = left_left;
+            constant_ = left_right;
+            offset_ = right;
+          } else if (base_matcher.right().HasValue()) {
+            // ((O + O) + C)
+            scaled_ = left_left;
+            offset_ = left_right;
+            constant_ = right;
+          } else {
+            // (O + O)
+            scaled_ = left;
+            offset_ = right;
+          }
+        }
+      } else {
+        if (base_matcher.right().HasValue()) {
+          // (O + C)
+          offset_ = left;
+          constant_ = right;
+        } else {
+          // (O + O)
+          offset_ = left;
+          scaled_ = right;
+        }
+      }
+    }
+    matches_ = true;
+  }
+
+  bool matches() const { return matches_; }
+  Node* scaled() const { return scaled_; }
+  int scale_exponent() const { return scale_exponent_; }
+  Node* offset() const { return offset_; }
+  Node* constant() const { return constant_; }
+
+ private:
+  bool matches_;
+
+ protected:
+  Node* scaled_;
+  int scale_exponent_;
+  Node* offset_;
+  Node* constant_;
+};
+
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8

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

@@ -364,8 +364,79 @@ void InstructionSelector::VisitWord64Ror(Node* node) {
   VisitWord64Shift(this, node, kX64Ror);
 }
 
+namespace {
+
+AddressingMode GenerateMemoryOperandInputs(X64OperandGenerator* g, Node* scaled,
+                                           int scale_exponent, Node* offset,
+                                           Node* constant,
+                                           InstructionOperand* inputs[],
+                                           size_t* input_count) {
+  AddressingMode mode = kMode_MRI;
+  if (offset != NULL) {
+    inputs[(*input_count)++] = g->UseRegister(offset);
+    if (scaled != NULL) {
+      DCHECK(scale_exponent >= 0 && scale_exponent <= 3);
+      inputs[(*input_count)++] = g->UseRegister(scaled);
+      if (constant != NULL) {
+        inputs[(*input_count)++] = g->UseImmediate(constant);
+        static const AddressingMode kMRnI_modes[] = {kMode_MR1I, kMode_MR2I,
+                                                     kMode_MR4I, kMode_MR8I};
+        mode = kMRnI_modes[scale_exponent];
+      } else {
+        static const AddressingMode kMRn_modes[] = {kMode_MR1, kMode_MR2,
+                                                    kMode_MR4, kMode_MR8};
+        mode = kMRn_modes[scale_exponent];
+      }
+    } else {
+      DCHECK(constant != NULL);
+      inputs[(*input_count)++] = g->UseImmediate(constant);
+      mode = kMode_MRI;
+    }
+  } else {
+    DCHECK(scaled != NULL);
+    DCHECK(scale_exponent >= 0 && scale_exponent <= 3);
+    inputs[(*input_count)++] = g->UseRegister(scaled);
+    if (constant != NULL) {
+      inputs[(*input_count)++] = g->UseImmediate(constant);
+      static const AddressingMode kMnI_modes[] = {kMode_M1I, kMode_M2I,
+                                                  kMode_M4I, kMode_M8I};
+      mode = kMnI_modes[scale_exponent];
+    } else {
+      static const AddressingMode kMn_modes[] = {kMode_M1, kMode_M2, kMode_M4,
+                                                 kMode_M8};
+      mode = kMn_modes[scale_exponent];
+    }
+  }
+  return mode;
+}
+
+}  // namespace
+
 
 void InstructionSelector::VisitInt32Add(Node* node) {
+  // Try to match the Add to a leal pattern
+  ScaledWithOffsetMatcher m(node);
+  X64OperandGenerator g(this);
+  if (m.matches() && (m.constant() == NULL || g.CanBeImmediate(m.constant()))) {
+    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);
+
+    InstructionCode opcode = AddressingModeField::encode(mode) | kX64Lea32;
+
+    Emit(opcode, 1, outputs, input_count, inputs);
+    return;
+  }
+
   VisitBinop(this, node, kX64Add32);
 }
 

test/mjsunit/compiler/division-by-constant.js

@@ -101,6 +101,7 @@ function CreateTestValues() {
 
 // -----------------------------------------------------------------------------
 
+
 function TestDivisionLike(ref, construct, values, divisor) {
   // Define the function to test.
   var OptFun = new Function("dividend", construct(divisor));
@@ -111,12 +112,14 @@ function TestDivisionLike(ref, construct, values, divisor) {
   %OptimizeFunctionOnNextCall(OptFun);
   OptFun(13);
 
-  // Check results.
-  values.forEach(function(dividend) {
+function dude(dividend) {
     // Avoid deopt caused by overflow, we do not want to test this here.
     if (dividend === -2147483648 && divisor === -1) return;
     assertEquals(ref(dividend, divisor), OptFun(dividend));
-  });
+  }
+
+  // Check results.
+  values.forEach(dude);
 }
 
 function Test(ref, construct) {

test/unittests/unittests.gyp

@@ -49,7 +49,7 @@
         'compiler/js-operator-unittest.cc',
         'compiler/js-typed-lowering-unittest.cc',
         'compiler/machine-operator-reducer-unittest.cc',
-        'compiler/machine-operator-unittest.cc',
+        'compiler/node-matchers-unittest.cc',
         'compiler/node-test-utils.cc',
         'compiler/node-test-utils.h',
         'compiler/register-allocator-unittest.cc',