Handle non-constant divisor in MathFloorOfDiv, on ia32/x64

Zheng Liu
zheng.z.liu@intel.com

Review URL: https://chromiumcodereview.appspot.com/11624022
Patch from Zheng Liu <zheng.z.liu@intel.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13289 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/ia32/lithium-codegen-ia32.cc

@@ -1212,7 +1212,7 @@ void LCodeGen::DoModI(LModI* instr) {
 
 
 void LCodeGen::DoDivI(LDivI* instr) {
-  if (instr->hydrogen()->HasPowerOf2Divisor()) {
+  if (!instr->is_flooring() && instr->hydrogen()->HasPowerOf2Divisor()) {
     Register dividend = ToRegister(instr->left());
     int32_t divisor =
         HConstant::cast(instr->hydrogen()->right())->Integer32Value();
@@ -1259,13 +1259,13 @@ void LCodeGen::DoDivI(LDivI* instr) {
 
   // Check for x / 0.
   Register right_reg = ToRegister(right);
-  if (instr->hydrogen()->CheckFlag(HValue::kCanBeDivByZero)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kCanBeDivByZero)) {
     __ test(right_reg, ToOperand(right));
     DeoptimizeIf(zero, instr->environment());
   }
 
   // Check for (0 / -x) that will produce negative zero.
-  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kBailoutOnMinusZero)) {
     Label left_not_zero;
     __ test(left_reg, Operand(left_reg));
     __ j(not_zero, &left_not_zero, Label::kNear);
@@ -1275,7 +1275,7 @@ void LCodeGen::DoDivI(LDivI* instr) {
   }
 
   // Check for (kMinInt / -1).
-  if (instr->hydrogen()->CheckFlag(HValue::kCanOverflow)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kCanOverflow)) {
     Label left_not_min_int;
     __ cmp(left_reg, kMinInt);
     __ j(not_zero, &left_not_min_int, Label::kNear);
@@ -1288,9 +1288,19 @@ void LCodeGen::DoDivI(LDivI* instr) {
   __ cdq();
   __ idiv(right_reg);
 
-  // Deoptimize if remainder is not 0.
-  __ test(edx, Operand(edx));
-  DeoptimizeIf(not_zero, instr->environment());
+  if (!instr->is_flooring()) {
+    // Deoptimize if remainder is not 0.
+    __ test(edx, Operand(edx));
+    DeoptimizeIf(not_zero, instr->environment());
+  } else {
+    Label done;
+    __ test(edx, edx);
+    __ j(zero, &done, Label::kNear);
+    __ xor_(edx, right_reg);
+    __ sar(edx, 31);
+    __ add(eax, edx);
+    __ bind(&done);
+  }
 }
 
 

src/ia32/lithium-ia32.cc

@@ -1305,12 +1305,31 @@ HValue* LChunkBuilder::SimplifiedDivisorForMathFloorOfDiv(HValue* divisor) {
     return constant_val->CopyToRepresentation(Representation::Integer32(),
                                               divisor->block()->zone());
   }
+  // A value with an integer representation does not need to be transformed.
+  if (divisor->representation().IsInteger32()) {
+    return divisor;
+  // A change from an integer32 can be replaced by the integer32 value.
+  } else if (divisor->IsChange() &&
+             HChange::cast(divisor)->from().IsInteger32()) {
+    return HChange::cast(divisor)->value();
+  }
   return NULL;
 }
 
 
 LInstruction* LChunkBuilder::DoMathFloorOfDiv(HMathFloorOfDiv* instr) {
   HValue* right = instr->right();
+  if (!right->IsConstant()) {
+    ASSERT(right->representation().IsInteger32());
+    // The temporary operand is necessary to ensure that right is not allocated
+    // into edx.
+    LOperand* temp = FixedTemp(edx);
+    LOperand* dividend = UseFixed(instr->left(), eax);
+    LOperand* divisor = UseRegister(instr->right());
+    LDivI* flooring_div = new(zone()) LDivI(dividend, divisor, temp);
+    return AssignEnvironment(DefineFixed(flooring_div, eax));
+  }
+
   ASSERT(right->IsConstant() && HConstant::cast(right)->HasInteger32Value());
   LOperand* divisor = chunk_->DefineConstantOperand(HConstant::cast(right));
   int32_t divisor_si = HConstant::cast(right)->Integer32Value();

src/ia32/lithium-ia32.h

@@ -565,6 +565,8 @@ class LDivI: public LTemplateInstruction<1, 2, 1> {
   LOperand* left() { return inputs_[0]; }
   LOperand* right() { return inputs_[1]; }
 
+  bool is_flooring() { return hydrogen_value()->IsMathFloorOfDiv(); }
+
   DECLARE_CONCRETE_INSTRUCTION(DivI, "div-i")
   DECLARE_HYDROGEN_ACCESSOR(Div)
 };

src/x64/lithium-codegen-x64.cc

@@ -1144,7 +1144,7 @@ void LCodeGen::DoMathFloorOfDiv(LMathFloorOfDiv* instr) {
 
 
 void LCodeGen::DoDivI(LDivI* instr) {
-  if (instr->hydrogen()->HasPowerOf2Divisor()) {
+  if (!instr->is_flooring() && instr->hydrogen()->HasPowerOf2Divisor()) {
     Register dividend = ToRegister(instr->left());
     int32_t divisor =
         HConstant::cast(instr->hydrogen()->right())->Integer32Value();
@@ -1191,13 +1191,13 @@ void LCodeGen::DoDivI(LDivI* instr) {
 
   // Check for x / 0.
   Register right_reg = ToRegister(right);
-  if (instr->hydrogen()->CheckFlag(HValue::kCanBeDivByZero)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kCanBeDivByZero)) {
     __ testl(right_reg, right_reg);
     DeoptimizeIf(zero, instr->environment());
   }
 
   // Check for (0 / -x) that will produce negative zero.
-  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kBailoutOnMinusZero)) {
     Label left_not_zero;
     __ testl(left_reg, left_reg);
     __ j(not_zero, &left_not_zero, Label::kNear);
@@ -1207,7 +1207,7 @@ void LCodeGen::DoDivI(LDivI* instr) {
   }
 
   // Check for (kMinInt / -1).
-  if (instr->hydrogen()->CheckFlag(HValue::kCanOverflow)) {
+  if (instr->hydrogen_value()->CheckFlag(HValue::kCanOverflow)) {
     Label left_not_min_int;
     __ cmpl(left_reg, Immediate(kMinInt));
     __ j(not_zero, &left_not_min_int, Label::kNear);
@@ -1220,9 +1220,19 @@ void LCodeGen::DoDivI(LDivI* instr) {
   __ cdq();
   __ idivl(right_reg);
 
-  // Deoptimize if remainder is not 0.
-  __ testl(rdx, rdx);
-  DeoptimizeIf(not_zero, instr->environment());
+  if (!instr->is_flooring()) {
+    // Deoptimize if remainder is not 0.
+    __ testl(rdx, rdx);
+    DeoptimizeIf(not_zero, instr->environment());
+  } else {
+    Label done;
+    __ testl(rdx, rdx);
+    __ j(zero, &done, Label::kNear);
+    __ xorl(rdx, right_reg);
+    __ sarl(rdx, Immediate(31));
+    __ addl(rax, rdx);
+    __ bind(&done);
+  }
 }
 
 
@@ -1712,6 +1722,7 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
       break;
     case Token::DIV:
       __ divsd(left, right);
+      __ movaps(left, left);
       break;
     case Token::MOD:
       __ PrepareCallCFunction(2);

src/x64/lithium-x64.cc

@@ -1228,12 +1228,31 @@ HValue* LChunkBuilder::SimplifiedDivisorForMathFloorOfDiv(HValue* divisor) {
     return constant_val->CopyToRepresentation(Representation::Integer32(),
                                               divisor->block()->zone());
   }
+  // A value with an integer representation does not need to be transformed.
+  if (divisor->representation().IsInteger32()) {
+    return divisor;
+  // A change from an integer32 can be replaced by the integer32 value.
+  } else if (divisor->IsChange() &&
+             HChange::cast(divisor)->from().IsInteger32()) {
+    return HChange::cast(divisor)->value();
+  }
   return NULL;
 }
 
 
 LInstruction* LChunkBuilder::DoMathFloorOfDiv(HMathFloorOfDiv* instr) {
   HValue* right = instr->right();
+  if (!right->IsConstant()) {
+    ASSERT(right->representation().IsInteger32());
+    // The temporary operand is necessary to ensure that right is not allocated
+    // into rdx.
+    LOperand* temp = FixedTemp(rdx);
+    LOperand* dividend = UseFixed(instr->left(), rax);
+    LOperand* divisor = UseRegister(instr->right());
+    LDivI* flooring_div = new(zone()) LDivI(dividend, divisor, temp);
+    return AssignEnvironment(DefineFixed(flooring_div, rax));
+  }
+
   ASSERT(right->IsConstant() && HConstant::cast(right)->HasInteger32Value());
   LOperand* divisor = chunk_->DefineConstantOperand(HConstant::cast(right));
   int32_t divisor_si = HConstant::cast(right)->Integer32Value();

src/x64/lithium-x64.h

@@ -573,6 +573,8 @@ class LDivI: public LTemplateInstruction<1, 2, 1> {
   LOperand* right() { return inputs_[1]; }
   LOperand* temp() { return temps_[0]; }
 
+  bool is_flooring() { return hydrogen_value()->IsMathFloorOfDiv(); }
+
   DECLARE_CONCRETE_INSTRUCTION(DivI, "div-i")
   DECLARE_HYDROGEN_ACCESSOR(Div)
 };

test/mjsunit/math-floor-of-div-nosudiv.js

@@ -184,10 +184,38 @@ test_div();
 %OptimizeFunctionOnNextCall(test_div);
 test_div();
 
+// Test for ia32/x64 flooring correctness.
+var values2 = [1, 3, 10, 99, 100, 101, 0x7fffffff];
+function test_div2() {
+  for (var i = 0; i < values2.length; i++) {
+    for (var j = 0; j < values2.length; j++) {
+      assertEquals(Math.floor(div((values2[i] | 0), (values2[j] | 0))),
+                   Math.floor((values2[i] | 0) / (values2[j] | 0)));
+      assertEquals(Math.floor(div(-(values2[i] | 0), (values2[j] | 0))),
+                   Math.floor(-(values2[i] | 0) / (values2[j] | 0)));
+      assertEquals(Math.floor(div((values2[i] | 0), -(values2[j] | 0))),
+                   Math.floor((values2[i] | 0) / -(values2[j] | 0)));
+      assertEquals(Math.floor(div(-(values2[i] | 0), -(values2[j] | 0))),
+                   Math.floor(-(values2[i] | 0) / -(values2[j] | 0)));
+    }
+  }
+}
+
+test_div2();
+%OptimizeFunctionOnNextCall(test_div2);
+test_div2();
+
+
 // Test for negative zero, overflow and division by 0.
 // Separate the tests to prevent deoptimizations from making the other optimized
 // test unreachable.
 
+// We box the value in an array to avoid constant propagation.
+var neg_one_in_array = [-1];
+var zero_in_array = [0];
+var min_int_in_array = [-2147483648];
+
+// Test for dividing by constant.
 function IsNegativeZero(x) {
   assertTrue(x == 0);  // Is 0 or -0.
   var y = 1 / x;
@@ -196,15 +224,12 @@ function IsNegativeZero(x) {
 }
 
 function test_div_deopt_minus_zero() {
-  var zero_in_array = [0];
   for (var i = 0; i < 2; ++i) {
     assertTrue(IsNegativeZero(Math.floor((zero_in_array[0] | 0) / -1)));
   }
 }
 
 function test_div_deopt_overflow() {
-  // We box the value in an array to avoid constant propagation.
-  var min_int_in_array = [-2147483648];
   for (var i = 0; i < 2; ++i) {
     // We use '| 0' to force the representation to int32.
     assertEquals(-min_int_in_array[0],
@@ -228,3 +253,36 @@ test_div_deopt_div_by_zero();
 test_div_deopt_minus_zero();
 test_div_deopt_overflow();
 test_div_deopt_div_by_zero();
+
+// Test for dividing by variable.
+function test_div_deopt_minus_zero_v() {
+  for (var i = 0; i < 2; ++i) {
+    assertTrue(IsNegativeZero(Math.floor((zero_in_array[0] | 0) /
+               neg_one_in_array[0])));
+  }
+}
+
+function test_div_deopt_overflow_v() {
+  for (var i = 0; i < 2; ++i) {
+    // We use '| 0' to force the representation to int32.
+    assertEquals(-min_int_in_array[0],
+                 Math.floor((min_int_in_array[0] | 0) / neg_one_in_array[0]));
+  }
+}
+
+function test_div_deopt_div_by_zero_v() {
+  for (var i = 0; i < 2; ++i) {
+    assertEquals(div(i, 0),
+                 Math.floor(i / zero_in_array[0]));
+  }
+}
+
+test_div_deopt_minus_zero_v();
+test_div_deopt_overflow_v();
+test_div_deopt_div_by_zero_v();
+%OptimizeFunctionOnNextCall(test_div_deopt_minus_zero_v);
+%OptimizeFunctionOnNextCall(test_div_deopt_overflow_v);
+%OptimizeFunctionOnNextCall(test_div_deopt_div_by_zero_v);
+test_div_deopt_minus_zero_v();
+test_div_deopt_overflow_v();
+test_div_deopt_div_by_zero_v();

test/mjsunit/math-floor-of-div.js

@@ -184,10 +184,38 @@ test_div();
 %OptimizeFunctionOnNextCall(test_div);
 test_div();
 
+// Test for ia32/x64 flooring correctness.
+var values2 = [1, 3, 10, 99, 100, 101, 0x7fffffff];
+function test_div2() {
+  for (var i = 0; i < values2.length; i++) {
+    for (var j = 0; j < values2.length; j++) {
+      assertEquals(Math.floor(div((values2[i] | 0), (values2[j] | 0))),
+                   Math.floor((values2[i] | 0) / (values2[j] | 0)));
+      assertEquals(Math.floor(div(-(values2[i] | 0), (values2[j] | 0))),
+                   Math.floor(-(values2[i] | 0) / (values2[j] | 0)));
+      assertEquals(Math.floor(div((values2[i] | 0), -(values2[j] | 0))),
+                   Math.floor((values2[i] | 0) / -(values2[j] | 0)));
+      assertEquals(Math.floor(div(-(values2[i] | 0), -(values2[j] | 0))),
+                   Math.floor(-(values2[i] | 0) / -(values2[j] | 0)));
+    }
+  }
+}
+
+test_div2();
+%OptimizeFunctionOnNextCall(test_div2);
+test_div2();
+
+
 // Test for negative zero, overflow and division by 0.
 // Separate the tests to prevent deoptimizations from making the other optimized
 // test unreachable.
 
+// We box the value in an array to avoid constant propagation.
+var neg_one_in_array = [-1];
+var zero_in_array = [0];
+var min_int_in_array = [-2147483648];
+
+// Test for dividing by constant.
 function IsNegativeZero(x) {
   assertTrue(x == 0);  // Is 0 or -0.
   var y = 1 / x;
@@ -196,15 +224,12 @@ function IsNegativeZero(x) {
 }
 
 function test_div_deopt_minus_zero() {
-  var zero_in_array = [0];
   for (var i = 0; i < 2; ++i) {
     assertTrue(IsNegativeZero(Math.floor((zero_in_array[0] | 0) / -1)));
   }
 }
 
 function test_div_deopt_overflow() {
-  // We box the value in an array to avoid constant propagation.
-  var min_int_in_array = [-2147483648];
   for (var i = 0; i < 2; ++i) {
     // We use '| 0' to force the representation to int32.
     assertEquals(-min_int_in_array[0],
@@ -228,3 +253,36 @@ test_div_deopt_div_by_zero();
 test_div_deopt_minus_zero();
 test_div_deopt_overflow();
 test_div_deopt_div_by_zero();
+
+// Test for dividing by variable.
+function test_div_deopt_minus_zero_v() {
+  for (var i = 0; i < 2; ++i) {
+    assertTrue(IsNegativeZero(Math.floor((zero_in_array[0] | 0) /
+               neg_one_in_array[0])));
+  }
+}
+
+function test_div_deopt_overflow_v() {
+  for (var i = 0; i < 2; ++i) {
+    // We use '| 0' to force the representation to int32.
+    assertEquals(-min_int_in_array[0],
+                 Math.floor((min_int_in_array[0] | 0) / neg_one_in_array[0]));
+  }
+}
+
+function test_div_deopt_div_by_zero_v() {
+  for (var i = 0; i < 2; ++i) {
+    assertEquals(div(i, 0),
+                 Math.floor(i / zero_in_array[0]));
+  }
+}
+
+test_div_deopt_minus_zero_v();
+test_div_deopt_overflow_v();
+test_div_deopt_div_by_zero_v();
+%OptimizeFunctionOnNextCall(test_div_deopt_minus_zero_v);
+%OptimizeFunctionOnNextCall(test_div_deopt_overflow_v);
+%OptimizeFunctionOnNextCall(test_div_deopt_div_by_zero_v);
+test_div_deopt_minus_zero_v();
+test_div_deopt_overflow_v();
+test_div_deopt_div_by_zero_v();