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

- Use "leal" for subtraction of integer constant when non-constant input to
  subtract is used more than once.
- Use "incl", "decl", and "addl" when they are smaller/cheaper than their
  leal/addl/subl equivalant.

R=titzer@chromium.org

Review URL: https://codereview.chromium.org/738073002

Cr-Commit-Position: refs/heads/master@{#25439}

src/compiler/x64/code-generator-x64.cc

@@ -597,6 +597,12 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
     case kX64Lea:
       __ leaq(i.OutputRegister(), i.MemoryOperand());
       break;
+    case kX64Dec32:
+      __ decl(i.OutputRegister());
+      break;
+    case kX64Inc32:
+      __ incl(i.OutputRegister());
+      break;
     case kX64Push:
       if (HasImmediateInput(instr, 0)) {
         __ pushq(i.InputImmediate(0));

src/compiler/x64/instruction-codes-x64.h

@@ -75,6 +75,8 @@ namespace compiler {
   V(X64Movss)                      \
   V(X64Lea32)                      \
   V(X64Lea)                        \
+  V(X64Dec32)                      \
+  V(X64Inc32)                      \
   V(X64Push)                       \
   V(X64StoreWriteBarrier)
 

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

@@ -428,9 +428,33 @@ void InstructionSelector::VisitInt32Add(Node* node) {
   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);
@@ -462,6 +486,31 @@ void InstructionSelector::VisitInt32Sub(Node* 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);
   }
 }

test/unittests/compiler/x64/instruction-selector-x64-unittest.cc

@@ -244,14 +244,17 @@ TEST_F(InstructionSelectorTest, TruncateInt64ToInt32WithWord64Shr) {
 // Addition.
 
 
-TEST_F(InstructionSelectorTest, Int32AddWithInt32AddWithParameters) {
+TEST_F(InstructionSelectorTest, Int32AddWithInt32ParametersLea) {
   StreamBuilder m(this, kMachInt32, kMachInt32, kMachInt32);
   Node* const p0 = m.Parameter(0);
   Node* const p1 = m.Parameter(1);
   Node* const a0 = m.Int32Add(p0, p1);
-  m.Return(m.Int32Add(a0, p0));
+  USE(a0);
+  // Additional uses of input to add chooses lea
+  Node* const a1 = m.Int32Add(p0, p1);
+  m.Return(m.Int32Add(a0, a1));
   Stream s = m.Build();
-  ASSERT_EQ(2U, s.size());
+  ASSERT_EQ(3U, s.size());
   EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
   ASSERT_EQ(2U, s[0]->InputCount());
   EXPECT_EQ(s.ToVreg(p1), s.ToVreg(s[0]->InputAt(1)));
@@ -259,10 +262,12 @@ TEST_F(InstructionSelectorTest, Int32AddWithInt32AddWithParameters) {
 }
 
 
-TEST_F(InstructionSelectorTest, Int32AddConstantAsLea) {
+TEST_F(InstructionSelectorTest, Int32AddConstantAsLeaSingle) {
   StreamBuilder m(this, kMachInt32, kMachInt32);
   Node* const p0 = m.Parameter(0);
   Node* const c0 = m.Int32Constant(15);
+  // If there is only a single use of an add's input, still use lea and not add,
+  // it is faster.
   m.Return(m.Int32Add(p0, c0));
   Stream s = m.Build();
   ASSERT_EQ(1U, s.size());
@@ -274,10 +279,80 @@ TEST_F(InstructionSelectorTest, Int32AddConstantAsLea) {
 }
 
 
-TEST_F(InstructionSelectorTest, Int32AddCommutedConstantAsLea) {
+TEST_F(InstructionSelectorTest, Int32AddConstantAsInc) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(1);
+  // If there is only a single use of an add's input and the immediate constant
+  // for the add is 1, use inc.
+  m.Return(m.Int32Add(p0, c0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Inc32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+  ASSERT_EQ(1U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+}
+
+
+TEST_F(InstructionSelectorTest, Int32AddConstantAsDec) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(-1);
+  // If there is only a single use of an add's input and the immediate constant
+  // for the add is -11, use dec.
+  m.Return(m.Int32Add(p0, c0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Dec32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+  ASSERT_EQ(1U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+}
+
+
+TEST_F(InstructionSelectorTest, Int32AddConstantAsLeaDouble) {
   StreamBuilder m(this, kMachInt32, kMachInt32);
   Node* const p0 = m.Parameter(0);
   Node* const c0 = m.Int32Constant(15);
+  // A second use of an add's input uses lea
+  Node* const a0 = m.Int32Add(p0, c0);
+  USE(a0);
+  m.Return(m.Int32Add(p0, c0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
+  ASSERT_EQ(2U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+  EXPECT_TRUE(s[0]->InputAt(1)->IsImmediate());
+}
+
+
+TEST_F(InstructionSelectorTest, Int32AddCommutedConstantAsLeaSingle) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(15);
+  // If there is only a single use of an add's input, still use lea... it's
+  // generally faster than the add to reduce register pressure.
+  m.Return(m.Int32Add(c0, p0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
+  ASSERT_EQ(2U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+  EXPECT_TRUE(s[0]->InputAt(1)->IsImmediate());
+}
+
+
+TEST_F(InstructionSelectorTest, Int32AddCommutedConstantAsLeaDouble) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(15);
+  // A second use of an add's input uses lea
+  Node* const a0 = m.Int32Add(c0, p0);
+  USE(a0);
   m.Return(m.Int32Add(c0, p0));
   Stream s = m.Build();
   ASSERT_EQ(1U, s.size());
@@ -615,6 +690,38 @@ TEST_F(InstructionSelectorTest, Int32AddScaled8ShlWithConstant) {
 }
 
 
+TEST_F(InstructionSelectorTest, Int32SubConstantAsInc) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(-1);
+  // If there is only a single use of an add's input and the immediate constant
+  // for the add is 1, use inc.
+  m.Return(m.Int32Sub(p0, c0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Inc32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+  ASSERT_EQ(1U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+}
+
+
+TEST_F(InstructionSelectorTest, Int32SubConstantAsDec) {
+  StreamBuilder m(this, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const c0 = m.Int32Constant(1);
+  // If there is only a single use of an sub's input and the immediate constant
+  // for the add is 1, use dec.
+  m.Return(m.Int32Sub(p0, c0));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Dec32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+  ASSERT_EQ(1U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+}
+
+
 // -----------------------------------------------------------------------------
 // Multiplication.