Handle negative input in inlined Math.round on Intel CPUs.

src/x64/lithium-codegen-x64.cc

Index: src/x64/lithium-codegen-x64.cc
diff --git a/src/x64/lithium-codegen-x64.cc b/src/x64/lithium-codegen-x64.cc
index 413bff03d7f9f485173e59db7ca9c684fe32c1dd..564e413a579408099acf4fd5f916c40d30f863ad 100644
--- a/src/x64/lithium-codegen-x64.cc
+++ b/src/x64/lithium-codegen-x64.cc
@@ -3486,7 +3486,7 @@ void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
     DeoptimizeIf(equal, instr->environment());
   } else {
     Label negative_sign, done;
-    // Deoptimize on negative inputs.
+    // Deoptimize on unordered.
     __ xorps(xmm_scratch, xmm_scratch);  // Zero the register.
     __ ucomisd(input_reg, xmm_scratch);
     DeoptimizeIf(parity_even, instr->environment());
@@ -3530,48 +3530,72 @@ void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
   const XMMRegister xmm_scratch = xmm0;
   Register output_reg = ToRegister(instr->result());
   XMMRegister input_reg = ToDoubleRegister(instr->value());
+  static int64_t one_half = V8_INT64_C(0x3FE0000000000000);  // 0.5
+  static int64_t minus_one_half = V8_INT64_C(0xBFE0000000000000);  // -0.5
 
-  Label done;
-  // xmm_scratch = 0.5
-  __ movq(kScratchRegister, V8_INT64_C(0x3FE0000000000000), RelocInfo::NONE64);
+  bool minus_zero_check =
+      instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero);
+
+  __ movq(kScratchRegister, one_half, RelocInfo::NONE64);
   __ movq(xmm_scratch, kScratchRegister);
-  Label below_half;
-  __ ucomisd(xmm_scratch, input_reg);
-  // If input_reg is NaN, this doesn't jump.
-  __ j(above, &below_half, Label::kNear);
-  // input = input + 0.5
-  // This addition might give a result that isn't the correct for
-  // rounding, due to loss of precision, but only for a number that's
-  // so big that the conversion below will overflow anyway.
-  __ addsd(xmm_scratch, input_reg);
-  // Compute Math.floor(input).
-  // Use truncating instruction (OK because input is positive).
-  __ cvttsd2si(output_reg, xmm_scratch);
-  // Overflow is signalled with minint.
-  __ cmpl(output_reg, Immediate(0x80000000));
-  DeoptimizeIf(equal, instr->environment());
-  __ jmp(&done);
 
-  __ bind(&below_half);
-  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
-    // Bailout if negative (including -0).
-    __ movq(output_reg, input_reg);
-    __ testq(output_reg, output_reg);
-    DeoptimizeIf(negative, instr->environment());
+  if (CpuFeatures::IsSupported(SSE4_1) && !minus_zero_check) {
+    CpuFeatures::Scope scope(SSE4_1);
+    __ addsd(xmm_scratch, input_reg);
+    __ roundsd(xmm_scratch, input_reg, Assembler::kRoundDown);
+    __ cvttsd2si(output_reg, xmm_scratch);
+    // Overflow is signalled with minint.
+    __ cmpl(output_reg, Immediate(0x80000000));
+    __ RecordComment("D2I conversion overflow");
+    DeoptimizeIf(equal, instr->environment());
   } else {
-    // Bailout if below -0.5, otherwise round to (positive) zero, even
-    // if negative.
-    // xmm_scrach = -0.5
-    __ movq(kScratchRegister,
-            V8_INT64_C(0xBFE0000000000000),
-            RelocInfo::NONE64);
+    Label done, round_to_zero, below_one_half, do_not_compensate;
+    __ ucomisd(xmm_scratch, input_reg);
+    __ j(above, &below_one_half);
+
+    // CVTTSD2SI rounds towards zero, since 0.5 <= x, we use floor(0.5 + x).
+    __ addsd(xmm_scratch, input_reg);
+    __ cvttsd2si(output_reg, xmm_scratch);
+    // Overflow is signalled with minint.
+    __ cmpl(output_reg, Immediate(0x80000000));
+    __ RecordComment("D2I conversion overflow");
+    DeoptimizeIf(equal, instr->environment());
+    __ jmp(&done);
+
+    __ bind(&below_one_half);
+    __ movq(kScratchRegister, minus_one_half, RelocInfo::NONE64);
     __ movq(xmm_scratch, kScratchRegister);
+    __ ucomisd(xmm_scratch, input_reg);
+    __ j(below_equal, &round_to_zero);
+
+    // CVTTSD2SI rounds towards zero, we use ceil(x - (-0.5)) and then
+    // compare and compensate.
+    __ subsd(input_reg, xmm_scratch);
+    __ cvttsd2si(output_reg, input_reg);
+    // Catch minint due to overflow, and to prevent overflow when compensating.
+    __ cmpl(output_reg, Immediate(0x80000000));
+    __ RecordComment("D2I conversion overflow");
+    DeoptimizeIf(equal, instr->environment());
+
+    __ cvtlsi2sd(xmm_scratch, output_reg);
     __ ucomisd(input_reg, xmm_scratch);
-    DeoptimizeIf(below, instr->environment());
-  }
-  __ xorl(output_reg, output_reg);
+    __ j(equal, &done, Label::kNear);
+    __ subl(output_reg, Immediate(1));
+    // No overflow because we already ruled out minint.
+    __ jmp(&done);
 
-  __ bind(&done);
+    __ bind(&round_to_zero);
+    // We return 0 for the input range [+0, 0.5[, or [-0.5, 0.5[ if
+    // we can ignore the difference between a result of -0 and +0.
+    if (minus_zero_check) {
+      __ movq(output_reg, input_reg);
+      __ testq(output_reg, output_reg);
+      __ RecordComment("Minus zero");
+      DeoptimizeIf(negative, instr->environment());
+    }
+    __ Set(output_reg, 0);
+    __ bind(&done);
+  }
 }