ARM64: Generate optimized code for Math.floor and Math.round with double outputs.

src/arm64/lithium-codegen-arm64.cc

 // Copyright 2013 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 "v8.h"
 
 #include "arm64/lithium-codegen-arm64.h"
 #include "arm64/lithium-gap-resolver-arm64.h"
 #include "code-stubs.h"
 #include "stub-cache.h"
@@ skipping 3743 matching lines @@
   Register temp1 = ToRegister(instr->temp1());
   Register temp2 = ToRegister(instr->temp2());
   Register temp3 = ToRegister(instr->temp3());
 
   MathExpGenerator::EmitMathExp(masm(), input, result,
                                 double_temp1, double_temp2,
                                 temp1, temp2, temp3);
 }
 
 
-void LCodeGen::DoMathFloor(LMathFloor* instr) {
-  // TODO(jbramley): If we could provide a double result, we could use frintm
-  // and produce a valid double result in a single instruction.
+void LCodeGen::DoMathFloorD(LMathFloorD* instr) {
+  DoubleRegister input = ToDoubleRegister(instr->value());
+  DoubleRegister result = ToDoubleRegister(instr->result());
+
+  __ Frintm(result, input);
+}
+
+
+void LCodeGen::DoMathFloorI(LMathFloorI* instr) {
   DoubleRegister input = ToDoubleRegister(instr->value());
   Register result = ToRegister(instr->result());
 
   if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
     DeoptimizeIfMinusZero(input, instr->environment());
   }
 
   __ Fcvtms(result, input);
 
   // Check that the result fits into a 32-bit integer.
@@ skipping 212 matching lines @@
     MathPowStub stub(isolate(), MathPowStub::INTEGER);
     __ CallStub(&stub);
   } else {
     ASSERT(exponent_type.IsDouble());
     MathPowStub stub(isolate(), MathPowStub::DOUBLE);
     __ CallStub(&stub);
   }
 }
 
 
-void LCodeGen::DoMathRound(LMathRound* instr) {
-  // TODO(jbramley): We could provide a double result here using frint.
+void LCodeGen::DoMathRoundD(LMathRoundD* instr) {
+  DoubleRegister input = ToDoubleRegister(instr->value());
+  DoubleRegister result = ToDoubleRegister(instr->result());
+  DoubleRegister scratch_d = double_scratch();
+
+  ASSERT(!AreAliased(input, result, scratch_d));
+
+  Label done;
+
+  __ Frinta(result, input);
+  __ Fcmp(input, 0.0);
+  __ Fccmp(result, input, ZFlag, lt);
+  // The result is correct if the input was in [-0, +infinity], or was a
+  // negative integral value.
+  __ B(eq, &done);
+
+  // Here the input is negative, non integral, with an exponent lower than 52.
+  // We do not have to worry about the 0.49999999999999994 (0x3fdfffffffffffff)
+  // case. So we can safely add 0.5.
+  __ Fmov(scratch_d, 0.5);
+  __ Fadd(result, input, scratch_d);
+  __ Frintm(result, result);
+  // The range [-0.5, -0.0[ yielded +0.0. Force the sign to negative.
+  __ Fabs(result, result);
+  __ Fneg(result, result);
+
+  __ Bind(&done);
+}
+
+
+void LCodeGen::DoMathRoundI(LMathRoundI* instr) {
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister temp1 = ToDoubleRegister(instr->temp1());
   Register result = ToRegister(instr->result());
   Label try_rounding;
   Label done;
 
   // Math.round() rounds to the nearest integer, with ties going towards
   // +infinity. This does not match any IEEE-754 rounding mode.
   //  - Infinities and NaNs are propagated unchanged, but cause deopts because
   //    they can't be represented as integers.
@@ skipping 18 matching lines @@
   __ Mov(result, 1);  // +0.5.
   // Remaining cases: [+0, +0.5[ or [-0.5, +0.5[, depending on
   // flag kBailoutOnMinusZero, will return 0 (xzr).
   __ Csel(result, result, xzr, eq);
   __ B(&done);
 
   __ Bind(&try_rounding);
   // Since we're providing a 32-bit result, we can implement ties-to-infinity by
   // adding 0.5 to the input, then taking the floor of the result. This does not
   // work for very large positive doubles because adding 0.5 would cause an
-  // intermediate rounding stage, so a different approach will be necessary if a
+  // intermediate rounding stage, so a different approach is necessary when a
   // double result is needed.
   __ Fadd(temp1, input, dot_five);
   __ Fcvtms(result, temp1);
 
   // Deopt if
   //  * the input was NaN
   //  * the result is not representable using a 32-bit integer.
   __ Fcmp(input, 0.0);
   __ Ccmp(result, Operand(result.W(), SXTW), NoFlag, vc);
   DeoptimizeIf(ne, instr->environment());
@@ skipping 1826 matching lines @@
   __ Ldr(result, FieldMemOperand(object, JSObject::kPropertiesOffset));
   // Index is equal to negated out of object property index plus 1.
   __ Sub(result, result, Operand::UntagSmiAndScale(index, kPointerSizeLog2));
   __ Ldr(result, FieldMemOperand(result,
                                  FixedArray::kHeaderSize - kPointerSize));
   __ Bind(deferred->exit());
   __ Bind(&done);
 }
 
 } }  // namespace v8::internal