X64 Crankshaft: Implement MathRound, MathFloor, MathSqrt, and MathPowHalf Una...

src/x64/lithium-codegen-x64.cc

 // Copyright 2011 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 2063 matching lines @@
   Abort("Unimplemented: %s", "DoDeferredMathAbsTaggedHeapNumber");
 }
 
 
 void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
   Abort("Unimplemented: %s", "DoMathAbs");
 }
 
 
 void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
-  Abort("Unimplemented: %s", "DoMathFloor");
+  XMMRegister xmm_scratch = xmm0;
+  Register output_reg = ToRegister(instr->result());
+  XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
+  __ xorpd(xmm_scratch, xmm_scratch);  // Zero the register.
+  __ ucomisd(input_reg, xmm_scratch);
+
+  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
+    DeoptimizeIf(below_equal, instr->environment());
+  } else {
+    DeoptimizeIf(below, instr->environment());
+  }
+
+  // Use truncating instruction (OK because input is positive).
+  __ cvttsd2si(output_reg, input_reg);

[Lasse Reichstein, 2011/02/18 14:13:47]

Use the "q" version, so you can properly handle min-int.

[William Hesse, 2011/02/18 14:33:28]

But then we have much more complicated tests, because integers not in the 32-bit
range are generated without overflow.  The output is an untagged int32.

+
+  // Overflow is signalled with minint.
+  __ cmpl(output_reg, Immediate(0x80000000));
+  DeoptimizeIf(equal, instr->environment());
 }
 
 
 void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
-  Abort("Unimplemented: %s", "DoMathRound");
+  const XMMRegister xmm_scratch = xmm0;
+  Register output_reg = ToRegister(instr->result());
+  XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
+
+  // xmm_scratch = 0.5
+  ExternalReference one_half = ExternalReference::address_of_one_half();
+  __ movq(kScratchRegister, one_half);
+  __ movsd(xmm_scratch, Operand(kScratchRegister, 0));

[Lasse Reichstein, 2011/02/18 14:13:47]

Why load from memory when we have 64-bit immediates. Just load the bit-pattern
into a scratch register and then to an XMM register.

[William Hesse, 2011/02/18 14:33:28]

Done.

+
+  // input = input + 0.5
+  __ addsd(input_reg, xmm_scratch);
+
+  // We need to return -0 for the input range [-0.5, 0[, otherwise
+  // compute Math.floor(value + 0.5).
+  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
+    __ ucomisd(input_reg, xmm_scratch);
+    DeoptimizeIf(below_equal, instr->environment());
+  } else {
+    // If we don't need to bailout on -0, we check only bailout
+    // on negative inputs.
+    __ xorpd(xmm_scratch, xmm_scratch);  // Zero the register.
+    __ ucomisd(input_reg, xmm_scratch);
+    DeoptimizeIf(below, instr->environment());
+  }
+
+  // Compute Math.floor(value + 0.5).
+  // Use truncating instruction (OK because input is positive).
+  __ cvttsd2si(output_reg, input_reg);
+
+  // Overflow is signalled with minint.
+  __ cmpl(output_reg, Immediate(0x80000000));
+  DeoptimizeIf(equal, instr->environment());
 }
 
 
 void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
-  Abort("Unimplemented: %s", "DoMathSqrt");
+  XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
+  ASSERT(ToDoubleRegister(instr->result()).is(input_reg));
+  __ sqrtsd(input_reg, input_reg);
 }
 
 
 void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
-  Abort("Unimplemented: %s", "DoMathPowHalf");
+  XMMRegister xmm_scratch = xmm0;
+  XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
+  ASSERT(ToDoubleRegister(instr->result()).is(input_reg));
+  __ xorpd(xmm_scratch, xmm_scratch);
+  __ addsd(input_reg, xmm_scratch);  // Convert -0 to +0.
+  __ sqrtsd(input_reg, input_reg);
 }
 
 
 void LCodeGen::DoPower(LPower* instr) {
   Abort("Unimplemented: %s", "DoPower");
 }
 
 
 void LCodeGen::DoMathLog(LUnaryMathOperation* instr) {
   Abort("Unimplemented: %s", "DoMathLog");
 }
 
 
 void LCodeGen::DoMathCos(LUnaryMathOperation* instr) {
   Abort("Unimplemented: %s", "DoMathCos");
 }
 
 
 void LCodeGen::DoMathSin(LUnaryMathOperation* instr) {
   Abort("Unimplemented: %s", "DoMathSin");
 }
 
 
 void LCodeGen::DoUnaryMathOperation(LUnaryMathOperation* instr) {
-  Abort("Unimplemented: %s", "DoUnaryMathOperation");
+  switch (instr->op()) {
+    case kMathAbs:
+      DoMathAbs(instr);
+      break;
+    case kMathFloor:
+      DoMathFloor(instr);
+      break;
+    case kMathRound:
+      DoMathRound(instr);
+      break;
+    case kMathSqrt:
+      DoMathSqrt(instr);
+      break;
+    case kMathPowHalf:
+      DoMathPowHalf(instr);
+      break;
+    case kMathCos:
+      DoMathCos(instr);
+      break;
+    case kMathSin:
+      DoMathSin(instr);
+      break;
+    case kMathLog:
+      DoMathLog(instr);
+      break;
+
+    default:
+      UNREACHABLE();
+  }
 }
 
 
 void LCodeGen::DoCallKeyed(LCallKeyed* instr) {
   Abort("Unimplemented: %s", "DoCallKeyed");
 }
 
 
 void LCodeGen::DoCallNamed(LCallNamed* instr) {
   ASSERT(ToRegister(instr->result()).is(rax));
@@ skipping 708 matching lines @@
   RegisterEnvironmentForDeoptimization(environment);
   ASSERT(osr_pc_offset_ == -1);
   osr_pc_offset_ = masm()->pc_offset();
 }
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_X64