[turbofan]: remove optimization of adds/subs to inc and dec

They generally cause regressions on most modern Intel chips. Replace them with
addl/subl.

Review URL: https://codereview.chromium.org/737153003

Cr-Commit-Position: refs/heads/master@{#25466}

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

@@ -427,29 +427,34 @@ void InstructionSelector::VisitInt32Add(Node* node) {
   // Try to match the Add to a leal pattern
   ScaledWithOffset32Matcher m(node);
   X64OperandGenerator g(this);
+  // It's possible to use a "leal", but it may not be smaller/cheaper. In the
+  // case that there are only two operands to the add and one of them isn't
+  // live, use a plain "addl".
   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()));
+      if (m.constant() == NULL) {
+        if (m.scaled() != NULL && m.scale_exponent() == 0) {
+          if (!IsLive(m.offset())) {
+            Emit(kX64Add32, g.DefineSameAsFirst(node),
+                 g.UseRegister(m.offset()), g.Use(m.scaled()));
             return;
-          } else if (value == -1) {
-            Emit(kX64Dec32, g.DefineSameAsFirst(node),
-                 g.UseRegister(m.offset()));
+          } else if (!IsLive(m.scaled())) {
+            Emit(kX64Add32, g.DefineSameAsFirst(node),
+                 g.UseRegister(m.scaled()), g.Use(m.offset()));
             return;
           }
         }
+      } else {
+        if (m.scale_exponent() == 0) {
+          if (m.scaled() == NULL || m.offset() == NULL) {
+            Node* non_constant = m.scaled() == NULL ? m.offset() : m.scaled();
+            if (!IsLive(non_constant)) {
+              Emit(kX64Add32, g.DefineSameAsFirst(node),
+                   g.UseRegister(non_constant), g.UseImmediate(m.constant()));
+              return;
+            }
+          }
+        }
       }
     }
 
@@ -487,21 +492,7 @@ void InstructionSelector::VisitInt32Sub(Node* node) {
     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 {
+      if (IsLive(m.left().node())) {
         // 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.

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

@@ -249,10 +249,9 @@ TEST_F(InstructionSelectorTest, Int32AddWithInt32ParametersLea) {
   Node* const p0 = m.Parameter(0);
   Node* const p1 = m.Parameter(1);
   Node* const a0 = m.Int32Add(p0, p1);
-  USE(a0);
   // Additional uses of input to add chooses lea
-  Node* const a1 = m.Int32Add(p0, p1);
-  m.Return(m.Int32Add(a0, a1));
+  Node* const a1 = m.Int32Div(p0, p1);
+  m.Return(m.Int32Div(a0, a1));
   Stream s = m.Build();
   ASSERT_EQ(3U, s.size());
   EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
@@ -266,20 +265,21 @@ 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));
+  // If there is only a single use of an add's input, use an "addl" not a
+  // "leal", it is faster.
+  Node* const v0 = m.Int32Add(p0, c0);
+  m.Return(v0);
   Stream s = m.Build();
   ASSERT_EQ(1U, s.size());
-  EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
-  EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
+  EXPECT_EQ(kX64Add32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, 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, Int32AddConstantAsInc) {
+TEST_F(InstructionSelectorTest, Int32AddConstantAsAdd) {
   StreamBuilder m(this, kMachInt32, kMachInt32);
   Node* const p0 = m.Parameter(0);
   Node* const c0 = m.Int32Constant(1);
@@ -288,26 +288,11 @@ TEST_F(InstructionSelectorTest, Int32AddConstantAsInc) {
   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(kX64Add32, s[0]->arch_opcode());
   EXPECT_EQ(kMode_None, s[0]->addressing_mode());
-  ASSERT_EQ(1U, s[0]->InputCount());
+  ASSERT_EQ(2U, s[0]->InputCount());
   EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+  EXPECT_TRUE(s[0]->InputAt(1)->IsImmediate());
 }
 
 
@@ -317,10 +302,9 @@ TEST_F(InstructionSelectorTest, Int32AddConstantAsLeaDouble) {
   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));
+  m.Return(m.Int32Div(a0, p0));
   Stream s = m.Build();
-  ASSERT_EQ(1U, s.size());
+  ASSERT_EQ(2U, s.size());
   EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
   EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
   ASSERT_EQ(2U, s[0]->InputCount());
@@ -333,13 +317,12 @@ 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.
+  // If there is only a single use of an add's input, use "addl"
   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());
+  EXPECT_EQ(kX64Add32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, 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());
@@ -353,9 +336,9 @@ TEST_F(InstructionSelectorTest, Int32AddCommutedConstantAsLeaDouble) {
   // 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));
+  m.Return(m.Int32Div(a0, p0));
   Stream s = m.Build();
-  ASSERT_EQ(1U, s.size());
+  ASSERT_EQ(2U, s.size());
   EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
   EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
   ASSERT_EQ(2U, s[0]->InputCount());
@@ -364,6 +347,39 @@ TEST_F(InstructionSelectorTest, Int32AddCommutedConstantAsLeaDouble) {
 }
 
 
+TEST_F(InstructionSelectorTest, Int32AddSimpleAsAdd) {
+  StreamBuilder m(this, kMachInt32, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const p1 = m.Parameter(1);
+  // If one of the add's operands is only used once, use an "addl".
+  m.Return(m.Int32Add(p0, p1));
+  Stream s = m.Build();
+  ASSERT_EQ(1U, s.size());
+  EXPECT_EQ(kX64Add32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+  ASSERT_EQ(2U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+  EXPECT_EQ(s.ToVreg(p1), s.ToVreg(s[0]->InputAt(1)));
+}
+
+
+TEST_F(InstructionSelectorTest, Int32AddSimpleAsLea) {
+  StreamBuilder m(this, kMachInt32, kMachInt32, kMachInt32);
+  Node* const p0 = m.Parameter(0);
+  Node* const p1 = m.Parameter(1);
+  // If all of of the add's operands are used multiple times, use an "leal".
+  Node* const v1 = m.Int32Add(p0, p1);
+  m.Return(m.Int32Add(m.Int32Add(v1, p1), p0));
+  Stream s = m.Build();
+  ASSERT_EQ(3U, s.size());
+  EXPECT_EQ(kX64Lea32, s[0]->arch_opcode());
+  EXPECT_EQ(kMode_MR1, s[0]->addressing_mode());
+  ASSERT_EQ(2U, s[0]->InputCount());
+  EXPECT_EQ(s.ToVreg(p0), s.ToVreg(s[0]->InputAt(0)));
+  EXPECT_EQ(s.ToVreg(p1), s.ToVreg(s[0]->InputAt(1)));
+}
+
+
 TEST_F(InstructionSelectorTest, Int32AddScaled2Mul) {
   StreamBuilder m(this, kMachInt32, kMachInt32, kMachInt32);
   Node* const p0 = m.Parameter(0);
@@ -690,35 +706,38 @@ TEST_F(InstructionSelectorTest, Int32AddScaled8ShlWithConstant) {
 }
 
 
-TEST_F(InstructionSelectorTest, Int32SubConstantAsInc) {
+TEST_F(InstructionSelectorTest, Int32SubConstantAsSub) {
   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.
+  // If there is only a single use of on of the sub's non-constant input, use a
+  // "subl" instruction.
   m.Return(m.Int32Sub(p0, c0));
   Stream s = m.Build();
   ASSERT_EQ(1U, s.size());
-  EXPECT_EQ(kX64Inc32, s[0]->arch_opcode());
+  EXPECT_EQ(kX64Sub32, s[0]->arch_opcode());
   EXPECT_EQ(kMode_None, s[0]->addressing_mode());
-  ASSERT_EQ(1U, s[0]->InputCount());
+  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, Int32SubConstantAsDec) {
+TEST_F(InstructionSelectorTest, Int32SubConstantAsLea) {
   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));
+  Node* const c0 = m.Int32Constant(-1);
+  // If there are multiple uses of on of the sub's non-constant input, use a
+  // "leal" instruction.
+  Node* const v0 = m.Int32Sub(p0, c0);
+  m.Return(m.Int32Div(p0, v0));
   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());
+  ASSERT_EQ(2U, 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());
 }