Implement fast case for strtod.

If there are few non-zero digits and the 10^exponent fits into a double then
we can compute the result using 1 (or 2) double operations.

BUG=
TEST=

Review URL: http://codereview.chromium.org/3584015

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5603 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/conversions.cc

@@ -354,7 +354,7 @@ static double InternalStringToInt(Iterator current, EndMark end, int radix) {
     ASSERT(buffer_pos < kBufferSize);
     buffer[buffer_pos] = '\0';
     Vector<char> buffer_vector(buffer, buffer_pos);
-    return sign ? -strtod(buffer_vector, NULL) : strtod(buffer_vector, NULL);
+    return sign ? -Strtod(buffer_vector, 0) : Strtod(buffer_vector, 0);
   }
 
   // The following code causes accumulating rounding error for numbers greater
@@ -653,7 +653,7 @@ static double InternalStringToDouble(Iterator current,
   ASSERT(buffer_pos < kBufferSize);
   buffer[buffer_pos] = '\0';
 
-  double converted = strtod(Vector<char>(buffer, buffer_pos), exponent);
+  double converted = Strtod(Vector<char>(buffer, buffer_pos), exponent);
   return sign? -converted: converted;
 }
 

src/strtod.cc

@@ -36,9 +36,47 @@
 namespace v8 {
 namespace internal {
 
+// 2^53 = 9007199254740992.
+// Any integer with at most 15 decimal digits will hence fit into a double
+// (which has a 53bit significand) without loss of precision.
+static const int kMaxExactDoubleIntegerDecimalDigits = 15;
+// 2^64 = 18446744073709551616
+// Any integer with at most 19 digits will hence fit into a 64bit datatype.
+static const int kMaxUint64DecimalDigits = 19;
+
+static const double exact_powers_of_ten[] = {
+  1.0,  // 10^0
+  10.0,
+  100.0,
+  1000.0,
+  10000.0,
+  100000.0,
+  1000000.0,
+  10000000.0,
+  100000000.0,
+  1000000000.0,
+  10000000000.0,  // 10^10
+  100000000000.0,
+  1000000000000.0,
+  10000000000000.0,
+  100000000000000.0,
+  1000000000000000.0,
+  10000000000000000.0,
+  100000000000000000.0,
+  1000000000000000000.0,
+  10000000000000000000.0,
+  100000000000000000000.0,  // 10^20
+  1000000000000000000000.0,
+  // 10^22 = 0x21e19e0c9bab2400000 = 0x878678326eac9 * 2^22
+  10000000000000000000000.0
+};
+
+static const int kExactPowersOfTenSize = ARRAY_SIZE(exact_powers_of_ten);
+
+
 extern "C" double gay_strtod(const char* s00, const char** se);
 
-double strtod(Vector<char> buffer, int exponent) {
+static double old_strtod(Vector<char> buffer, int exponent) {
   char gay_buffer[1024];
   Vector<char> gay_buffer_vector(gay_buffer, sizeof(gay_buffer));
   buffer.start()[buffer.length()] = '\0';
@@ -46,4 +84,63 @@ double strtod(Vector<char> buffer, int exponent) {
   return gay_strtod(gay_buffer, NULL);
 }
 
+
+static Vector<char> TrimTrailingZeros(Vector<char> buffer) {
+  for (int i = buffer.length() - 1; i >= 0; --i) {
+    if (buffer[i] != '0') {
+      return Vector<char>(buffer.start(), i + 1);
+    }
+  }
+  return Vector<char>(buffer.start(), 0);
+}
+
+
+uint64_t ReadUint64(Vector<char> buffer) {
+  ASSERT(buffer.length() <= kMaxUint64DecimalDigits);
+  uint64_t result = 0;
+  for (int i = 0; i < buffer.length(); ++i) {
+    int digit = buffer[i] - '0';
+    ASSERT(0 <= digit && digit <= 9);
+    result = 10 * result + digit;
+  }
+  return result;
+}
+
+
+double Strtod(Vector<char> buffer, int exponent) {
+  Vector<char> trimmed = TrimTrailingZeros(buffer);
+  if (trimmed.length() == 0) return 0.0;
+  exponent += buffer.length() - trimmed.length();
+  if (trimmed.length() <= kMaxExactDoubleIntegerDecimalDigits) {
+    // The trimmed input fits into a double.
+    // If the 10^exponent (resp. 10^-exponent) fits into a double too then we
+    // can compute the result-double simply by multiplying (resp. dividing) the
+    // two numbers.
+    // This is possible because IEEE guarantees that floating-point operations
+    // return the best possible approximation.
+    if (exponent < 0 && -exponent < kExactPowersOfTenSize) {
+      // 10^-exponent fits into a double.
+      double buffer_d = static_cast<double>(ReadUint64(trimmed));
+      return buffer_d / exact_powers_of_ten[-exponent];
+    }
+    if (0 <= exponent && exponent < kExactPowersOfTenSize) {
+      // 10^exponent fits into a double.
+      double buffer_d = static_cast<double>(ReadUint64(trimmed));
+      return buffer_d * exact_powers_of_ten[exponent];
+    }
+    int remaining_digits =
+        kMaxExactDoubleIntegerDecimalDigits - trimmed.length();
+    if ((0 <= exponent) &&
+        (exponent - remaining_digits < kExactPowersOfTenSize)) {
+      // The trimmed string was short and we can multiply it with
+      // 10^remaining_digits. As a result the remaining exponent now fits
+      // into a double too.
+      double buffer_d = static_cast<double>(ReadUint64(trimmed));
+      buffer_d *= exact_powers_of_ten[remaining_digits];
+      return buffer_d * exact_powers_of_ten[exponent - remaining_digits];
+    }
+  }
+  return old_strtod(trimmed, exponent);
+}
+
 } }  // namespace v8::internal

src/strtod.h

@@ -33,7 +33,7 @@ namespace internal {
 
 // The buffer must only contain digits in the range [0-9]. It must not
 // contain a dot or a sign. It must not start with '0', and must not be empty.
-double strtod(Vector<char> buffer, int exponent);
+double Strtod(Vector<char> buffer, int exponent);
 
 } }  // namespace v8::internal
 

test/cctest/SConscript

@@ -70,6 +70,7 @@ SOURCES = {
     'test-sockets.cc',
     'test-spaces.cc',
     'test-strings.cc',
+    'test-strtod.cc',
     'test-thread-termination.cc',
     'test-threads.cc',
     'test-type-info.cc',

test/cctest/test-strtod.cc

+// Copyright 2006-2008 the V8 project authors. All rights reserved.
+
+#include <stdlib.h>
+
+#include "v8.h"
+
+#include "platform.h"
+#include "cctest.h"
+#include "strtod.h"
+
+using namespace v8::internal;
+
+static const int kBufferSize = 100;
+
+
+TEST(Strtod) {
+  char buffer[kBufferSize];
+  Vector<char> vector;
+
+  strncpy(buffer, "0", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+
+  CHECK_EQ(0.0, Strtod(vector, 1));
+  CHECK_EQ(0.0, Strtod(vector, 2));
+  CHECK_EQ(0.0, Strtod(vector, -2));
+  CHECK_EQ(0.0, Strtod(vector, -999));
+  CHECK_EQ(0.0, Strtod(vector, +999));
+
+  strncpy(buffer, "1", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(1.0, Strtod(vector, 0));
+  CHECK_EQ(10.0, Strtod(vector, 1));
+  CHECK_EQ(100.0, Strtod(vector, 2));
+  CHECK_EQ(1e20, Strtod(vector, 20));
+  CHECK_EQ(1e22, Strtod(vector, 22));
+  CHECK_EQ(1e23, Strtod(vector, 23));
+  CHECK_EQ(1e35, Strtod(vector, 35));
+  CHECK_EQ(1e36, Strtod(vector, 36));
+  CHECK_EQ(1e37, Strtod(vector, 37));
+  CHECK_EQ(1e-1, Strtod(vector, -1));
+  CHECK_EQ(1e-2, Strtod(vector, -2));
+  CHECK_EQ(1e-5, Strtod(vector, -5));
+  CHECK_EQ(1e-20, Strtod(vector, -20));
+  CHECK_EQ(1e-22, Strtod(vector, -22));
+  CHECK_EQ(1e-23, Strtod(vector, -23));
+  CHECK_EQ(1e-25, Strtod(vector, -25));
+  CHECK_EQ(1e-39, Strtod(vector, -39));
+
+  strncpy(buffer, "2", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(2.0, Strtod(vector, 0));
+  CHECK_EQ(20.0, Strtod(vector, 1));
+  CHECK_EQ(200.0, Strtod(vector, 2));
+  CHECK_EQ(2e20, Strtod(vector, 20));
+  CHECK_EQ(2e22, Strtod(vector, 22));
+  CHECK_EQ(2e23, Strtod(vector, 23));
+  CHECK_EQ(2e35, Strtod(vector, 35));
+  CHECK_EQ(2e36, Strtod(vector, 36));
+  CHECK_EQ(2e37, Strtod(vector, 37));
+  CHECK_EQ(2e-1, Strtod(vector, -1));
+  CHECK_EQ(2e-2, Strtod(vector, -2));
+  CHECK_EQ(2e-5, Strtod(vector, -5));
+  CHECK_EQ(2e-20, Strtod(vector, -20));
+  CHECK_EQ(2e-22, Strtod(vector, -22));
+  CHECK_EQ(2e-23, Strtod(vector, -23));
+  CHECK_EQ(2e-25, Strtod(vector, -25));
+  CHECK_EQ(2e-39, Strtod(vector, -39));
+
+  strncpy(buffer, "9", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(9.0, Strtod(vector, 0));
+  CHECK_EQ(90.0, Strtod(vector, 1));
+  CHECK_EQ(900.0, Strtod(vector, 2));
+  CHECK_EQ(9e20, Strtod(vector, 20));
+  CHECK_EQ(9e22, Strtod(vector, 22));
+  CHECK_EQ(9e23, Strtod(vector, 23));
+  CHECK_EQ(9e35, Strtod(vector, 35));
+  CHECK_EQ(9e36, Strtod(vector, 36));
+  CHECK_EQ(9e37, Strtod(vector, 37));
+  CHECK_EQ(9e-1, Strtod(vector, -1));
+  CHECK_EQ(9e-2, Strtod(vector, -2));
+  CHECK_EQ(9e-5, Strtod(vector, -5));
+  CHECK_EQ(9e-20, Strtod(vector, -20));
+  CHECK_EQ(9e-22, Strtod(vector, -22));
+  CHECK_EQ(9e-23, Strtod(vector, -23));
+  CHECK_EQ(9e-25, Strtod(vector, -25));
+  CHECK_EQ(9e-39, Strtod(vector, -39));
+
+  strncpy(buffer, "12345", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(12345.0, Strtod(vector, 0));
+  CHECK_EQ(123450.0, Strtod(vector, 1));
+  CHECK_EQ(1234500.0, Strtod(vector, 2));
+  CHECK_EQ(12345e20, Strtod(vector, 20));
+  CHECK_EQ(12345e22, Strtod(vector, 22));
+  CHECK_EQ(12345e23, Strtod(vector, 23));
+  CHECK_EQ(12345e30, Strtod(vector, 30));
+  CHECK_EQ(12345e31, Strtod(vector, 31));
+  CHECK_EQ(12345e32, Strtod(vector, 32));
+  CHECK_EQ(12345e35, Strtod(vector, 35));
+  CHECK_EQ(12345e36, Strtod(vector, 36));
+  CHECK_EQ(12345e37, Strtod(vector, 37));
+  CHECK_EQ(12345e-1, Strtod(vector, -1));
+  CHECK_EQ(12345e-2, Strtod(vector, -2));
+  CHECK_EQ(12345e-5, Strtod(vector, -5));
+  CHECK_EQ(12345e-20, Strtod(vector, -20));
+  CHECK_EQ(12345e-22, Strtod(vector, -22));
+  CHECK_EQ(12345e-23, Strtod(vector, -23));
+  CHECK_EQ(12345e-25, Strtod(vector, -25));
+  CHECK_EQ(12345e-39, Strtod(vector, -39));
+
+  strncpy(buffer, "12345678901234", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(12345678901234.0, Strtod(vector, 0));
+  CHECK_EQ(123456789012340.0, Strtod(vector, 1));
+  CHECK_EQ(1234567890123400.0, Strtod(vector, 2));
+  CHECK_EQ(12345678901234e20, Strtod(vector, 20));
+  CHECK_EQ(12345678901234e22, Strtod(vector, 22));
+  CHECK_EQ(12345678901234e23, Strtod(vector, 23));
+  CHECK_EQ(12345678901234e30, Strtod(vector, 30));
+  CHECK_EQ(12345678901234e31, Strtod(vector, 31));
+  CHECK_EQ(12345678901234e32, Strtod(vector, 32));
+  CHECK_EQ(12345678901234e35, Strtod(vector, 35));
+  CHECK_EQ(12345678901234e36, Strtod(vector, 36));
+  CHECK_EQ(12345678901234e37, Strtod(vector, 37));
+  CHECK_EQ(12345678901234e-1, Strtod(vector, -1));
+  CHECK_EQ(12345678901234e-2, Strtod(vector, -2));
+  CHECK_EQ(12345678901234e-5, Strtod(vector, -5));
+  CHECK_EQ(12345678901234e-20, Strtod(vector, -20));
+  CHECK_EQ(12345678901234e-22, Strtod(vector, -22));
+  CHECK_EQ(12345678901234e-23, Strtod(vector, -23));
+  CHECK_EQ(12345678901234e-25, Strtod(vector, -25));
+  CHECK_EQ(12345678901234e-39, Strtod(vector, -39));
+
+  strncpy(buffer, "123456789012345", kBufferSize);
+  vector = Vector<char>(buffer, strlen(buffer));
+  CHECK_EQ(123456789012345.0, Strtod(vector, 0));
+  CHECK_EQ(1234567890123450.0, Strtod(vector, 1));
+  CHECK_EQ(12345678901234500.0, Strtod(vector, 2));
+  CHECK_EQ(123456789012345e20, Strtod(vector, 20));
+  CHECK_EQ(123456789012345e22, Strtod(vector, 22));
+  CHECK_EQ(123456789012345e23, Strtod(vector, 23));
+  CHECK_EQ(123456789012345e35, Strtod(vector, 35));
+  CHECK_EQ(123456789012345e36, Strtod(vector, 36));
+  CHECK_EQ(123456789012345e37, Strtod(vector, 37));
+  CHECK_EQ(123456789012345e39, Strtod(vector, 39));
+  CHECK_EQ(123456789012345e-1, Strtod(vector, -1));
+  CHECK_EQ(123456789012345e-2, Strtod(vector, -2));
+  CHECK_EQ(123456789012345e-5, Strtod(vector, -5));
+  CHECK_EQ(123456789012345e-20, Strtod(vector, -20));
+  CHECK_EQ(123456789012345e-22, Strtod(vector, -22));
+  CHECK_EQ(123456789012345e-23, Strtod(vector, -23));
+  CHECK_EQ(123456789012345e-25, Strtod(vector, -25));
+  CHECK_EQ(123456789012345e-39, Strtod(vector, -39));
+}