Strtod fast-case that uses DiyFps and cached powers of ten.

This is a fixed version of r5677.
Review URL: http://codereview.chromium.org/3898007

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5686 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/cached-powers.cc

@@ -42,6 +42,11 @@ struct CachedPower {
 };
 
 static const CachedPower kCachedPowers[] = {
+  {V8_2PART_UINT64_C(0xfa8fd5a0, 081c0288), -1220, -348},
+  {V8_2PART_UINT64_C(0xbaaee17f, a23ebf76), -1193, -340},
+  {V8_2PART_UINT64_C(0x8b16fb20, 3055ac76), -1166, -332},
+  {V8_2PART_UINT64_C(0xcf42894a, 5dce35ea), -1140, -324},
+  {V8_2PART_UINT64_C(0x9a6bb0aa, 55653b2d), -1113, -316},
   {V8_2PART_UINT64_C(0xe61acf03, 3d1a45df), -1087, -308},
   {V8_2PART_UINT64_C(0xab70fe17, c79ac6ca), -1060, -300},
   {V8_2PART_UINT64_C(0xff77b1fc, bebcdc4f), -1034, -292},
@@ -129,24 +134,44 @@ static const CachedPower kCachedPowers[] = {
 static const int kCachedPowersLength = ARRAY_SIZE(kCachedPowers);
 static const int kCachedPowersOffset = -kCachedPowers[0].decimal_exponent;
 static const double kD_1_LOG2_10 = 0.30102999566398114;  //  1 / lg(10)
-static const int kCachedPowersDecimalDistance =
+const int PowersOfTenCache::kDecimalExponentDistance =
     kCachedPowers[1].decimal_exponent - kCachedPowers[0].decimal_exponent;
+const int PowersOfTenCache::kMinDecimalExponent =
+    kCachedPowers[0].decimal_exponent;
+const int PowersOfTenCache::kMaxDecimalExponent =
+    kCachedPowers[kCachedPowersLength - 1].decimal_exponent;
 
-void GetCachedPowerForBinaryExponentRange(int min_exponent,
-                                          int max_exponent,
-                                          DiyFp* power,
-                                          int* decimal_exponent) {
-    int kQ = DiyFp::kSignificandSize;
-    double k = ceiling((min_exponent + kQ - 1) * kD_1_LOG2_10);
-    int foo = kCachedPowersOffset;
-    int index =
-        (foo + static_cast<int>(k) - 1) / kCachedPowersDecimalDistance + 1;
-    ASSERT(0 <= index && index < kCachedPowersLength);
-    CachedPower cached_power = kCachedPowers[index];
-    ASSERT(min_exponent <= cached_power.binary_exponent);
-    ASSERT(cached_power.binary_exponent <= max_exponent);
-    *decimal_exponent = cached_power.decimal_exponent;
-    *power = DiyFp(cached_power.significand, cached_power.binary_exponent);
+void PowersOfTenCache::GetCachedPowerForBinaryExponentRange(
+    int min_exponent,
+    int max_exponent,
+    DiyFp* power,
+    int* decimal_exponent) {
+  int kQ = DiyFp::kSignificandSize;
+  double k = ceiling((min_exponent + kQ - 1) * kD_1_LOG2_10);
+  int foo = kCachedPowersOffset;
+  int index =
+      (foo + static_cast<int>(k) - 1) / kDecimalExponentDistance + 1;
+  ASSERT(0 <= index && index < kCachedPowersLength);
+  CachedPower cached_power = kCachedPowers[index];
+  ASSERT(min_exponent <= cached_power.binary_exponent);
+  ASSERT(cached_power.binary_exponent <= max_exponent);
+  *decimal_exponent = cached_power.decimal_exponent;
+  *power = DiyFp(cached_power.significand, cached_power.binary_exponent);
+}
+
+
+void PowersOfTenCache::GetCachedPowerForDecimalExponent(int requested_exponent,
+                                                        DiyFp* power,
+                                                        int* found_exponent) {
+  ASSERT(kMinDecimalExponent <= requested_exponent);
+  ASSERT(requested_exponent < kMaxDecimalExponent + kDecimalExponentDistance);
+  int index =
+      (requested_exponent + kCachedPowersOffset) / kDecimalExponentDistance;
+  CachedPower cached_power = kCachedPowers[index];
+  *power = DiyFp(cached_power.significand, cached_power.binary_exponent);
+  *found_exponent = cached_power.decimal_exponent;
+  ASSERT(*found_exponent <= requested_exponent);
+  ASSERT(requested_exponent < *found_exponent + kDecimalExponentDistance);
 }
 
 } }  // namespace v8::internal

src/cached-powers.h

@@ -33,10 +33,32 @@
 namespace v8 {
 namespace internal {
 
-void GetCachedPowerForBinaryExponentRange(int min_exponent,
-                                          int max_exponent,
-                                          DiyFp* power,
-                                          int* decimal_exponent);
+class PowersOfTenCache {
+ public:
+
+  // Not all powers of ten are cached. The decimal exponent of two neighboring
+  // cached numbers will differ by kDecimalExponentDistance.
+  static const int kDecimalExponentDistance;
+
+  static const int kMinDecimalExponent;
+  static const int kMaxDecimalExponent;
+
+  // Returns a cached power-of-ten with a binary exponent in the range
+  // [min_exponent; max_exponent] (boundaries included).
+  static void GetCachedPowerForBinaryExponentRange(int min_exponent,
+                                                   int max_exponent,
+                                                   DiyFp* power,
+                                                   int* decimal_exponent);
+
+  // Returns a cached power of ten x ~= 10^k such that
+  //   k <= decimal_exponent < k + kCachedPowersDecimalDistance.
+  // The given decimal_exponent must satisfy
+  //   kMinDecimalExponent <= requested_exponent, and
+  //   requested_exponent < kMaxDecimalExponent + kDecimalExponentDistance.
+  static void GetCachedPowerForDecimalExponent(int requested_exponent,
+                                               DiyFp* power,
+                                               int* found_exponent);
+};
 
 } }  // namespace v8::internal
 

src/double.h

@@ -45,10 +45,14 @@ class Double {
   static const uint64_t kSignificandMask =
       V8_2PART_UINT64_C(0x000FFFFF, FFFFFFFF);
   static const uint64_t kHiddenBit = V8_2PART_UINT64_C(0x00100000, 00000000);
+  static const int kPhysicalSignificandSize = 52;  // Excludes the hidden bit.
+  static const int kSignificandSize = 53;
 
   Double() : d64_(0) {}
   explicit Double(double d) : d64_(double_to_uint64(d)) {}
   explicit Double(uint64_t d64) : d64_(d64) {}
+  explicit Double(DiyFp diy_fp)
+    : d64_(DiyFpToUint64(diy_fp)) {}
 
   DiyFp AsDiyFp() const {
     ASSERT(!IsSpecial());
@@ -67,9 +71,9 @@ class Double {
       f <<= 1;
       e--;
     }
-    // Do the final shifts in one go. Don't forget the hidden bit (the '-1').
-    f <<= DiyFp::kSignificandSize - kSignificandSize - 1;
-    e -= DiyFp::kSignificandSize - kSignificandSize - 1;
+    // Do the final shifts in one go.
+    f <<= DiyFp::kSignificandSize - kSignificandSize;
+    e -= DiyFp::kSignificandSize - kSignificandSize;
     return DiyFp(f, e);
   }
 
@@ -82,7 +86,8 @@ class Double {
     if (IsDenormal()) return kDenormalExponent;
 
     uint64_t d64 = AsUint64();
-    int biased_e = static_cast<int>((d64 & kExponentMask) >> kSignificandSize);
+    int biased_e =
+        static_cast<int>((d64 & kExponentMask) >> kPhysicalSignificandSize);
     return biased_e - kExponentBias;
   }
 
@@ -156,12 +161,54 @@ class Double {
 
   double value() const { return uint64_to_double(d64_); }
 
+  // Returns the significand size for a given order of magnitude.
+  // If v = f*2^e with 2^p-1 <= f <= 2^p then p+e is v's order of magnitude.
+  // This function returns the number of significant binary digits v will have
+  // once its encoded into a double. In almost all cases this is equal to
+  // kSignificandSize. The only exception are denormals. They start with leading
+  // zeroes and their effective significand-size is hence smaller.
+  static int SignificandSizeForOrderOfMagnitude(int order) {
+    if (order >= (kDenormalExponent + kSignificandSize)) {
+      return kSignificandSize;
+    }
+    if (order <= kDenormalExponent) return 0;
+    return order - kDenormalExponent;
+  }
+
  private:
-  static const int kSignificandSize = 52;  // Excludes the hidden bit.
-  static const int kExponentBias = 0x3FF + kSignificandSize;
+  static const int kExponentBias = 0x3FF + kPhysicalSignificandSize;
   static const int kDenormalExponent = -kExponentBias + 1;
+  static const int kMaxExponent = 0x7FF - kExponentBias;
+  static const uint64_t kInfinity = V8_2PART_UINT64_C(0x7FF00000, 00000000);
 
-  uint64_t d64_;
+  const uint64_t d64_;
+
+  static uint64_t DiyFpToUint64(DiyFp diy_fp) {
+    uint64_t significand = diy_fp.f();
+    int exponent = diy_fp.e();
+    while (significand > kHiddenBit + kSignificandMask) {
+      significand >>= 1;
+      exponent++;
+    }
+    if (exponent >= kMaxExponent) {
+      return kInfinity;
+    }
+    if (exponent < kDenormalExponent) {
+      return 0;
+    }
+    while (exponent > kDenormalExponent && (significand & kHiddenBit) == 0) {
+      significand <<= 1;
+      exponent--;
+    }
+    uint64_t biased_exponent;
+    if (exponent == kDenormalExponent && (significand & kHiddenBit) == 0) {
+      biased_exponent = 0;
+    } else {
+      biased_exponent = static_cast<uint64_t>(exponent + kExponentBias);
+    }
+    return (significand & kSignificandMask) |
+        (biased_exponent << kPhysicalSignificandSize);
+  }
 };
 
 } }  // namespace v8::internal

src/fast-dtoa.cc

@@ -613,9 +613,10 @@ static bool Grisu3(double v,
      kMinimalTargetExponent - (w.e() + DiyFp::kSignificandSize);
   int ten_mk_maximal_binary_exponent =
      kMaximalTargetExponent - (w.e() + DiyFp::kSignificandSize);
-  GetCachedPowerForBinaryExponentRange(ten_mk_minimal_binary_exponent,
-                                       ten_mk_maximal_binary_exponent,
-                                       &ten_mk, &mk);
+  PowersOfTenCache::GetCachedPowerForBinaryExponentRange(
+      ten_mk_minimal_binary_exponent,
+      ten_mk_maximal_binary_exponent,
+      &ten_mk, &mk);
   ASSERT((kMinimalTargetExponent <= w.e() + ten_mk.e() +
           DiyFp::kSignificandSize) &&
          (kMaximalTargetExponent >= w.e() + ten_mk.e() +
@@ -671,9 +672,10 @@ static bool Grisu3Counted(double v,
      kMinimalTargetExponent - (w.e() + DiyFp::kSignificandSize);
   int ten_mk_maximal_binary_exponent =
      kMaximalTargetExponent - (w.e() + DiyFp::kSignificandSize);
-  GetCachedPowerForBinaryExponentRange(ten_mk_minimal_binary_exponent,
-                                       ten_mk_maximal_binary_exponent,
-                                       &ten_mk, &mk);
+  PowersOfTenCache::GetCachedPowerForBinaryExponentRange(
+      ten_mk_minimal_binary_exponent,
+      ten_mk_maximal_binary_exponent,
+      &ten_mk, &mk);
   ASSERT((kMinimalTargetExponent <= w.e() + ten_mk.e() +
           DiyFp::kSignificandSize) &&
          (kMaximalTargetExponent >= w.e() + ten_mk.e() +

test/cctest/test-strtod.cc

@@ -198,10 +198,58 @@ TEST(Strtod) {
   CHECK_EQ(1234e304, StrtodChar("0000000123400000", 299));
   CHECK_EQ(V8_INFINITY, StrtodChar("00000000180000000", 300));
   CHECK_EQ(17e307, StrtodChar("00000000170000000", 300));
+  CHECK_EQ(1.7976931348623157E+308, StrtodChar("17976931348623157", 292));
+  CHECK_EQ(1.7976931348623158E+308, StrtodChar("17976931348623158", 292));
+  CHECK_EQ(V8_INFINITY, StrtodChar("17976931348623159", 292));
 
   // The following number is the result of 89255.0/1e-22. Both floating-point
   // numbers can be accurately represented with doubles. However on Linux,x86
   // the floating-point stack is set to 80bits and the double-rounding
   // introduces an error.
   CHECK_EQ(89255e-22, StrtodChar("89255", -22));
+  CHECK_EQ(104110013277974872254e-225,
+           StrtodChar("104110013277974872254", -225));
+
+  CHECK_EQ(123456789e108, StrtodChar("123456789", 108));
+  CHECK_EQ(123456789e109, StrtodChar("123456789", 109));
+  CHECK_EQ(123456789e110, StrtodChar("123456789", 110));
+  CHECK_EQ(123456789e111, StrtodChar("123456789", 111));
+  CHECK_EQ(123456789e112, StrtodChar("123456789", 112));
+  CHECK_EQ(123456789e113, StrtodChar("123456789", 113));
+  CHECK_EQ(123456789e114, StrtodChar("123456789", 114));
+  CHECK_EQ(123456789e115, StrtodChar("123456789", 115));
+
+  CHECK_EQ(1234567890123456789012345e108,
+           StrtodChar("1234567890123456789012345", 108));
+  CHECK_EQ(1234567890123456789012345e109,
+           StrtodChar("1234567890123456789012345", 109));
+  CHECK_EQ(1234567890123456789012345e110,
+           StrtodChar("1234567890123456789012345", 110));
+  CHECK_EQ(1234567890123456789012345e111,
+           StrtodChar("1234567890123456789012345", 111));
+  CHECK_EQ(1234567890123456789012345e112,
+           StrtodChar("1234567890123456789012345", 112));
+  CHECK_EQ(1234567890123456789012345e113,
+           StrtodChar("1234567890123456789012345", 113));
+  CHECK_EQ(1234567890123456789012345e114,
+           StrtodChar("1234567890123456789012345", 114));
+  CHECK_EQ(1234567890123456789012345e115,
+           StrtodChar("1234567890123456789012345", 115));
+
+  CHECK_EQ(1234567890123456789052345e108,
+           StrtodChar("1234567890123456789052345", 108));
+  CHECK_EQ(1234567890123456789052345e109,
+           StrtodChar("1234567890123456789052345", 109));
+  CHECK_EQ(1234567890123456789052345e110,
+           StrtodChar("1234567890123456789052345", 110));
+  CHECK_EQ(1234567890123456789052345e111,
+           StrtodChar("1234567890123456789052345", 111));
+  CHECK_EQ(1234567890123456789052345e112,
+           StrtodChar("1234567890123456789052345", 112));
+  CHECK_EQ(1234567890123456789052345e113,
+           StrtodChar("1234567890123456789052345", 113));
+  CHECK_EQ(1234567890123456789052345e114,
+           StrtodChar("1234567890123456789052345", 114));
+  CHECK_EQ(1234567890123456789052345e115,
+           StrtodChar("1234567890123456789052345", 115));
 }