MIPS: Faster implementation of Math.exp()

src/mips/codegen-mips.cc

 // Copyright 2012 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 13 matching lines @@
 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "v8.h"
 
 #if defined(V8_TARGET_ARCH_MIPS)
 
 #include "codegen.h"
 #include "macro-assembler.h"
+#include "simulator-mips.h"
 
 namespace v8 {
 namespace internal {
 
-#define __ ACCESS_MASM(masm)
 
 UnaryMathFunction CreateTranscendentalFunction(TranscendentalCache::Type type) {
   switch (type) {
     case TranscendentalCache::SIN: return &sin;
     case TranscendentalCache::COS: return &cos;
     case TranscendentalCache::TAN: return &tan;
     case TranscendentalCache::LOG: return &log;
     default: UNIMPLEMENTED();
   }
   return NULL;
 }
 
 
+#define __ masm.
+
+
+#if defined(USE_SIMULATOR)
+byte* fast_exp_mips_machine_code = NULL;
+double fast_exp_simulator(double x) {
+  return Simulator::current(Isolate::Current())->CallFP(
+      fast_exp_mips_machine_code, x, 0);
+}
+#endif
+
+
+UnaryMathFunction CreateExpFunction() {
+  if (!CpuFeatures::IsSupported(FPU)) return &exp;
+  if (!FLAG_fast_math) return &exp;
+  size_t actual_size;
+  byte* buffer = static_cast<byte*>(OS::Allocate(1 * KB, &actual_size, true));
+  if (buffer == NULL) return &exp;
+  ExternalReference::InitializeMathExpData();
+
+  MacroAssembler masm(NULL, buffer, static_cast<int>(actual_size));
+
+  {
+    CpuFeatures::Scope use_fpu(FPU);
+    DoubleRegister input = f12;
+    DoubleRegister result = f0;
+    DoubleRegister double_scratch1 = f4;
+    DoubleRegister double_scratch2 = f6;
+    Register temp1 = t0;
+    Register temp2 = t1;
+    Register temp3 = t2;
+
+    if (!IsMipsSoftFloatABI) {
+      // Input value is in f12 anyway, nothing to do.
+    } else {
+      __ Move(input, a0, a1);
+    }
+    __ Push(temp3, temp2, temp1);
+    MathExpGenerator::EmitMathExp(
+        &masm, input, result, double_scratch1, double_scratch2,
+        temp1, temp2, temp3);
+    __ Pop(temp3, temp2, temp1);
+    if (!IsMipsSoftFloatABI) {
+      // Result is already in f0, nothing to do.
+    } else {
+      __ Move(a0, a1, result);
+    }
+    __ Ret();
+  }
+
+  CodeDesc desc;
+  masm.GetCode(&desc);
+
+  CPU::FlushICache(buffer, actual_size);
+  OS::ProtectCode(buffer, actual_size);
+
+#if !defined(USE_SIMULATOR)
+  return FUNCTION_CAST<UnaryMathFunction>(buffer);
+#else
+  fast_exp_mips_machine_code = buffer;
+  return &fast_exp_simulator;
+#endif
+}
+
+
+#undef __
+
+
 UnaryMathFunction CreateSqrtFunction() {
   return &sqrt;
 }
 
 // -------------------------------------------------------------------------
 // Platform-specific RuntimeCallHelper functions.
 
 void StubRuntimeCallHelper::BeforeCall(MacroAssembler* masm) const {
   masm->EnterFrame(StackFrame::INTERNAL);
   ASSERT(!masm->has_frame());
   masm->set_has_frame(true);
 }
 
 
 void StubRuntimeCallHelper::AfterCall(MacroAssembler* masm) const {
   masm->LeaveFrame(StackFrame::INTERNAL);
   ASSERT(masm->has_frame());
   masm->set_has_frame(false);
 }
 
 // -------------------------------------------------------------------------
 // Code generators
 
+#define __ ACCESS_MASM(masm)
+
 void ElementsTransitionGenerator::GenerateMapChangeElementsTransition(
     MacroAssembler* masm) {
   // ----------- S t a t e -------------
   //  -- a0    : value
   //  -- a1    : key
   //  -- a2    : receiver
   //  -- ra    : return address
   //  -- a3    : target map, scratch for subsequent call
   //  -- t0    : scratch (elements)
   // -----------------------------------
@@ skipping 354 matching lines @@
   __ Addu(at, string, at);
   __ lhu(result, MemOperand(at));
   __ jmp(&done);
   __ bind(&ascii);
   // Ascii string.
   __ Addu(at, string, index);
   __ lbu(result, MemOperand(at));
   __ bind(&done);
 }
 
+
+static MemOperand ExpConstant(int index, Register base) {
+  return MemOperand(base, index * kDoubleSize);
+}
+
+
+void MathExpGenerator::EmitMathExp(MacroAssembler* masm,
+                                   DoubleRegister input,
+                                   DoubleRegister result,
+                                   DoubleRegister double_scratch1,
+                                   DoubleRegister double_scratch2,
+                                   Register temp1,
+                                   Register temp2,
+                                   Register temp3) {
+  ASSERT(!input.is(result));
+  ASSERT(!input.is(double_scratch1));
+  ASSERT(!input.is(double_scratch2));
+  ASSERT(!result.is(double_scratch1));
+  ASSERT(!result.is(double_scratch2));
+  ASSERT(!double_scratch1.is(double_scratch2));
+  ASSERT(!temp1.is(temp2));
+  ASSERT(!temp1.is(temp3));
+  ASSERT(!temp2.is(temp3));
+  ASSERT(ExternalReference::math_exp_constants(0).address() != NULL);
+
+  Label done;
+
+  __ li(temp3, Operand(ExternalReference::math_exp_constants(0)));
+
+  __ ldc1(double_scratch1, ExpConstant(0, temp3));
+  __ Move(result, kDoubleRegZero);
+  __ BranchF(&done, NULL, ge, double_scratch1, input);
+  __ ldc1(double_scratch2, ExpConstant(1, temp3));
+  __ ldc1(result, ExpConstant(2, temp3));
+  __ BranchF(&done, NULL, ge, input, double_scratch2);
+  __ ldc1(double_scratch1, ExpConstant(3, temp3));
+  __ ldc1(result, ExpConstant(4, temp3));
+  __ mul_d(double_scratch1, double_scratch1, input);
+  __ add_d(double_scratch1, double_scratch1, result);
+  __ Move(temp2, temp1, double_scratch1);
+  __ sub_d(double_scratch1, double_scratch1, result);
+  __ ldc1(result, ExpConstant(6, temp3));
+  __ ldc1(double_scratch2, ExpConstant(5, temp3));
+  __ mul_d(double_scratch1, double_scratch1, double_scratch2);
+  __ sub_d(double_scratch1, double_scratch1, input);
+  __ sub_d(result, result, double_scratch1);
+  __ mul_d(input, double_scratch1, double_scratch1);
+  __ mul_d(result, result, input);
+  __ srl(temp1, temp2, 11);
+  __ ldc1(double_scratch2, ExpConstant(7, temp3));
+  __ mul_d(result, result, double_scratch2);
+  __ sub_d(result, result, double_scratch1);
+  __ ldc1(double_scratch2, ExpConstant(8, temp3));
+  __ add_d(result, result, double_scratch2);
+  __ li(at, 0x7ff);
+  __ And(temp2, temp2, at);
+  __ Addu(temp1, temp1, Operand(0x3ff));
+  __ sll(temp1, temp1, 20);
+
+  // Must not call ExpConstant() after overwriting temp3!
+  __ li(temp3, Operand(ExternalReference::math_exp_log_table()));
+  __ sll(at, temp2, 3);
+  __ addu(at, at, temp3);
+  __ lw(at, MemOperand(at));
+  __ Addu(temp3, temp3, Operand(kPointerSize));
+  __ sll(temp2, temp2, 3);
+  __ addu(temp2, temp2, temp3);
+  __ lw(temp2, MemOperand(temp2));
+  __ Or(temp1, temp1, temp2);
+  __ Move(input, at, temp1);
+  __ mul_d(result, result, input);
+  __ bind(&done);
+}
+
+
 // nop(CODE_AGE_MARKER_NOP)
 static const uint32_t kCodeAgePatchFirstInstruction = 0x00010180;
 
 static byte* GetNoCodeAgeSequence(uint32_t* length) {
   // The sequence of instructions that is patched out for aging code is the
   // following boilerplate stack-building prologue that is found in FUNCTIONS
   static bool initialized = false;
   static uint32_t sequence[kNoCodeAgeSequenceLength];
   byte* byte_sequence = reinterpret_cast<byte*>(sequence);
   *length = kNoCodeAgeSequenceLength * Assembler::kInstrSize;
@@ skipping 79 matching lines @@
     patcher.masm()->dd(reinterpret_cast<uint32_t>(stub->instruction_start()));
   }
 }
 
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_MIPS