Revert of [builtins] Introduce proper Float64Exp operator.

src/arm64/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 "src/arm64/codegen-arm64.h"
 
 #if V8_TARGET_ARCH_ARM64
 
 #include "src/arm64/simulator-arm64.h"
 #include "src/codegen.h"
 #include "src/macro-assembler.h"
 
 namespace v8 {
 namespace internal {
 
 #define __ ACCESS_MASM(masm)
 
+#if defined(USE_SIMULATOR)
+byte* fast_exp_arm64_machine_code = nullptr;
+double fast_exp_simulator(double x, Isolate* isolate) {
+  Simulator * simulator = Simulator::current(isolate);
+  Simulator::CallArgument args[] = {
+      Simulator::CallArgument(x),
+      Simulator::CallArgument::End()
+  };
+  return simulator->CallDouble(fast_exp_arm64_machine_code, args);
+}
+#endif
+
+
+UnaryMathFunctionWithIsolate CreateExpFunction(Isolate* isolate) {
+  // Use the Math.exp implemetation in MathExpGenerator::EmitMathExp() to create
+  // an AAPCS64-compliant exp() function. This will be faster than the C
+  // library's exp() function, but probably less accurate.
+  size_t actual_size;
+  byte* buffer =
+      static_cast<byte*>(base::OS::Allocate(1 * KB, &actual_size, true));
+  if (buffer == nullptr) return nullptr;
+
+  ExternalReference::InitializeMathExpData();
+  MacroAssembler masm(isolate, buffer, static_cast<int>(actual_size),
+                      CodeObjectRequired::kNo);
+  masm.SetStackPointer(csp);
+
+  // The argument will be in d0 on entry.
+  DoubleRegister input = d0;
+  // Use other caller-saved registers for all other values.
+  DoubleRegister result = d1;
+  DoubleRegister double_temp1 = d2;
+  DoubleRegister double_temp2 = d3;
+  Register temp1 = x10;
+  Register temp2 = x11;
+  Register temp3 = x12;
+
+  MathExpGenerator::EmitMathExp(&masm, input, result,
+                                double_temp1, double_temp2,
+                                temp1, temp2, temp3);
+  // Move the result to the return register.
+  masm.Fmov(d0, result);
+  masm.Ret();
+
+  CodeDesc desc;
+  masm.GetCode(&desc);
+  DCHECK(!RelocInfo::RequiresRelocation(desc));
+
+  Assembler::FlushICache(isolate, buffer, actual_size);
+  base::OS::ProtectCode(buffer, actual_size);
+
+#if !defined(USE_SIMULATOR)
+  return FUNCTION_CAST<UnaryMathFunctionWithIsolate>(buffer);
+#else
+  fast_exp_arm64_machine_code = buffer;
+  return &fast_exp_simulator;
+#endif
+}
+
+
 UnaryMathFunctionWithIsolate CreateSqrtFunction(Isolate* isolate) {
   return nullptr;
 }
 
 
 // -------------------------------------------------------------------------
 // Platform-specific RuntimeCallHelper functions.
 
 void StubRuntimeCallHelper::BeforeCall(MacroAssembler* masm) const {
   masm->EnterFrame(StackFrame::INTERNAL);
@@ skipping 415 matching lines @@
   __ TestAndBranchIfAnySet(result, kStringEncodingMask, &one_byte);
   // Two-byte string.
   __ Ldrh(result, MemOperand(string, index, SXTW, 1));
   __ B(&done);
   __ Bind(&one_byte);
   // One-byte string.
   __ Ldrb(result, MemOperand(string, index, SXTW));
   __ Bind(&done);
 }
 
+
+static MemOperand ExpConstant(Register base, int index) {
+  return MemOperand(base, index * kDoubleSize);
+}
+
+
+void MathExpGenerator::EmitMathExp(MacroAssembler* masm,
+                                   DoubleRegister input,
+                                   DoubleRegister result,
+                                   DoubleRegister double_temp1,
+                                   DoubleRegister double_temp2,
+                                   Register temp1,
+                                   Register temp2,
+                                   Register temp3) {
+  // TODO(jbramley): There are several instances where fnmsub could be used
+  // instead of fmul and fsub. Doing this changes the result, but since this is
+  // an estimation anyway, does it matter?
+
+  DCHECK(!AreAliased(input, result,
+                     double_temp1, double_temp2,
+                     temp1, temp2, temp3));
+  DCHECK(ExternalReference::math_exp_constants(0).address() != NULL);
+  DCHECK(!masm->serializer_enabled());  // External references not serializable.
+
+  Label done;
+  DoubleRegister double_temp3 = result;
+  Register constants = temp3;
+
+  // The algorithm used relies on some magic constants which are initialized in
+  // ExternalReference::InitializeMathExpData().
+
+  // Load the address of the start of the array.
+  __ Mov(constants, ExternalReference::math_exp_constants(0));
+
+  // We have to do a four-way split here:
+  //  - If input <= about -708.4, the output always rounds to zero.
+  //  - If input >= about 709.8, the output always rounds to +infinity.
+  //  - If the input is NaN, the output is NaN.
+  //  - Otherwise, the result needs to be calculated.
+  Label result_is_finite_non_zero;
+  // Assert that we can load offset 0 (the small input threshold) and offset 1
+  // (the large input threshold) with a single ldp.
+  DCHECK(kDRegSize == (ExpConstant(constants, 1).offset() -
+                              ExpConstant(constants, 0).offset()));
+  __ Ldp(double_temp1, double_temp2, ExpConstant(constants, 0));
+
+  __ Fcmp(input, double_temp1);
+  __ Fccmp(input, double_temp2, NoFlag, hi);
+  // At this point, the condition flags can be in one of five states:
+  //  NZCV
+  //  1000      -708.4 < input < 709.8    result = exp(input)
+  //  0110      input == 709.8            result = +infinity
+  //  0010      input > 709.8             result = +infinity
+  //  0011      input is NaN              result = input
+  //  0000      input <= -708.4           result = +0.0
+
+  // Continue the common case first. 'mi' tests N == 1.
+  __ B(&result_is_finite_non_zero, mi);
+
+  // TODO(jbramley): Consider adding a +infinity register for ARM64.
+  __ Ldr(double_temp2, ExpConstant(constants, 2));    // Synthesize +infinity.
+
+  // Select between +0.0 and +infinity. 'lo' tests C == 0.
+  __ Fcsel(result, fp_zero, double_temp2, lo);
+  // Select between {+0.0 or +infinity} and input. 'vc' tests V == 0.
+  __ Fcsel(result, result, input, vc);
+  __ B(&done);
+
+  // The rest is magic, as described in InitializeMathExpData().
+  __ Bind(&result_is_finite_non_zero);
+
+  // Assert that we can load offset 3 and offset 4 with a single ldp.
+  DCHECK(kDRegSize == (ExpConstant(constants, 4).offset() -
+                              ExpConstant(constants, 3).offset()));
+  __ Ldp(double_temp1, double_temp3, ExpConstant(constants, 3));
+  __ Fmadd(double_temp1, double_temp1, input, double_temp3);
+  __ Fmov(temp2.W(), double_temp1.S());
+  __ Fsub(double_temp1, double_temp1, double_temp3);
+
+  // Assert that we can load offset 5 and offset 6 with a single ldp.
+  DCHECK(kDRegSize == (ExpConstant(constants, 6).offset() -
+                              ExpConstant(constants, 5).offset()));
+  __ Ldp(double_temp2, double_temp3, ExpConstant(constants, 5));
+  // TODO(jbramley): Consider using Fnmsub here.
+  __ Fmul(double_temp1, double_temp1, double_temp2);
+  __ Fsub(double_temp1, double_temp1, input);
+
+  __ Fmul(double_temp2, double_temp1, double_temp1);
+  __ Fsub(double_temp3, double_temp3, double_temp1);
+  __ Fmul(double_temp3, double_temp3, double_temp2);
+
+  __ Mov(temp1.W(), Operand(temp2.W(), LSR, 11));
+
+  __ Ldr(double_temp2, ExpConstant(constants, 7));
+  // TODO(jbramley): Consider using Fnmsub here.
+  __ Fmul(double_temp3, double_temp3, double_temp2);
+  __ Fsub(double_temp3, double_temp3, double_temp1);
+
+  // The 8th constant is 1.0, so use an immediate move rather than a load.
+  // We can't generate a runtime assertion here as we would need to call Abort
+  // in the runtime and we don't have an Isolate when we generate this code.
+  __ Fmov(double_temp2, 1.0);
+  __ Fadd(double_temp3, double_temp3, double_temp2);
+
+  __ And(temp2, temp2, 0x7ff);
+  __ Add(temp1, temp1, 0x3ff);
+
+  // Do the final table lookup.
+  __ Mov(temp3, ExternalReference::math_exp_log_table());
+
+  __ Add(temp3, temp3, Operand(temp2, LSL, kDRegSizeLog2));
+  __ Ldp(temp2.W(), temp3.W(), MemOperand(temp3));
+  __ Orr(temp1.W(), temp3.W(), Operand(temp1.W(), LSL, 20));
+  __ Bfi(temp2, temp1, 32, 32);
+  __ Fmov(double_temp1, temp2);
+
+  __ Fmul(result, double_temp3, double_temp1);
+
+  __ Bind(&done);
+}
+
 #undef __
 
 }  // namespace internal
 }  // namespace v8
 
 #endif  // V8_TARGET_ARCH_ARM64