[libc][i386] FPBit support for 96b long double (#115084)

`long double` is haunted on most architectures, but it is especially so on
i386-linux-gnu. While have 80b of significant data, on i386-linux-gnu this type
has 96b of storage.

Fixes for supporting printf family of conversions for `long double` on
i386-linux-gnu. This allows the libc-stdlib-tests and libc_stdio_unittests
ninja target tests to pass on i386-linux-gnu.

Fixes: #110894
Link: #93709
Co-authored-by: Michael Jones <[email protected]>

Author: nickdesaulniers

libc/src/__support/FPUtil/FPBits.h

@@ -127,7 +127,11 @@ template <> struct FPLayout<FPType::IEEE754_Binary128> {
 };
 
 template <> struct FPLayout<FPType::X86_Binary80> {
+#if __SIZEOF_LONG_DOUBLE__ == 12
+  using StorageType = UInt<__SIZEOF_LONG_DOUBLE__ * CHAR_BIT>;
+#else
   using StorageType = UInt128;
+#endif
   LIBC_INLINE_VAR static constexpr int SIGN_LEN = 1;
   LIBC_INLINE_VAR static constexpr int EXP_LEN = 15;
   LIBC_INLINE_VAR static constexpr int SIG_LEN = 64;

libc/src/__support/FPUtil/generic/sqrt_80_bit_long_double.h

@@ -21,7 +21,8 @@ namespace LIBC_NAMESPACE_DECL {
 namespace fputil {
 namespace x86 {
 
-LIBC_INLINE void normalize(int &exponent, UInt128 &mantissa) {
+LIBC_INLINE void normalize(int &exponent,
+                           FPBits<long double>::StorageType &mantissa) {
   const unsigned int shift = static_cast<unsigned int>(
       cpp::countl_zero(static_cast<uint64_t>(mantissa)) -
       (8 * sizeof(uint64_t) - 1 - FPBits<long double>::FRACTION_LEN));

libc/src/__support/big_int.h

@@ -469,7 +469,7 @@ struct BigInt {
                                                     !cpp::is_same_v<T, bool>>>
   LIBC_INLINE constexpr BigInt(T v) {
     constexpr size_t T_SIZE = sizeof(T) * CHAR_BIT;
-    const bool is_neg = Signed && (v < 0);
+    const bool is_neg = v < 0;
     for (size_t i = 0; i < WORD_COUNT; ++i) {
       if (v == 0) {
         extend(i, is_neg);
@@ -504,6 +504,12 @@ struct BigInt {
   // TODO: Reuse the Sign type.
   LIBC_INLINE constexpr bool is_neg() const { return SIGNED && get_msb(); }
 
+  template <size_t OtherBits, bool OtherSigned, typename OtherWordType>
+  LIBC_INLINE constexpr explicit
+  operator BigInt<OtherBits, OtherSigned, OtherWordType>() const {
+    return BigInt<OtherBits, OtherSigned, OtherWordType>(this);
+  }
+
   template <typename T> LIBC_INLINE constexpr explicit operator T() const {
     return to<T>();
   }
@@ -1058,6 +1064,8 @@ struct WordTypeSelector : cpp::type_identity<
 // Except if we request 16 or 32 bits explicitly.
 template <> struct WordTypeSelector<16> : cpp::type_identity<uint16_t> {};
 template <> struct WordTypeSelector<32> : cpp::type_identity<uint32_t> {};
+template <> struct WordTypeSelector<96> : cpp::type_identity<uint32_t> {};
+
 template <size_t Bits>
 using WordTypeSelectorT = typename WordTypeSelector<Bits>::type;
 } // namespace internal

libc/src/__support/float_to_string.h

@@ -373,23 +373,12 @@ LIBC_INLINE UInt<MID_INT_SIZE> get_table_negative_df(int exponent, size_t i) {
   return result;
 }
 
-LIBC_INLINE uint32_t fast_uint_mod_1e9(const UInt<MID_INT_SIZE> &val) {
-  // The formula for mult_const is:
-  //  1 + floor((2^(bits in target integer size + log_2(divider))) / divider)
-  // Where divider is 10^9 and target integer size is 128.
-  const UInt<MID_INT_SIZE> mult_const(
-      {0x31680A88F8953031u, 0x89705F4136B4A597u, 0});
-  const auto middle = (mult_const * val);
-  const uint64_t result = static_cast<uint64_t>(middle[2]);
-  const uint64_t shifted = result >> 29;
-  return static_cast<uint32_t>(static_cast<uint32_t>(val) -
-                               (EXP10_9 * shifted));
-}
-
 LIBC_INLINE uint32_t mul_shift_mod_1e9(const FPBits::StorageType mantissa,
                                        const UInt<MID_INT_SIZE> &large,
                                        const int32_t shift_amount) {
-  UInt<MID_INT_SIZE + FPBits::STORAGE_LEN> val(large);
+  // make sure the number of bits is always divisible by 64
+  UInt<internal::div_ceil(MID_INT_SIZE + FPBits::STORAGE_LEN, 64) * 64> val(
+      large);
   val = (val * mantissa) >> shift_amount;
   return static_cast<uint32_t>(
       val.div_uint_half_times_pow_2(static_cast<uint32_t>(EXP10_9), 0).value());

libc/src/__support/integer_literals.h

@@ -165,6 +165,10 @@ LIBC_INLINE constexpr T parse_with_prefix(const char *ptr) {
 
 } // namespace internal
 
+LIBC_INLINE constexpr UInt<96> operator""_u96(const char *x) {
+  return internal::parse_with_prefix<UInt<96>>(x);
+}
+
 LIBC_INLINE constexpr UInt128 operator""_u128(const char *x) {
   return internal::parse_with_prefix<UInt128>(x);
 }

libc/src/__support/str_to_float.h

@@ -206,7 +206,7 @@ eisel_lemire<long double>(ExpandedFloat<long double> init_num,
   using FPBits = typename fputil::FPBits<long double>;
   using StorageType = typename FPBits::StorageType;
 
-  StorageType mantissa = init_num.mantissa;
+  UInt128 mantissa = init_num.mantissa;
   int32_t exp10 = init_num.exponent;
 
   // Exp10 Range
@@ -225,7 +225,8 @@ eisel_lemire<long double>(ExpandedFloat<long double> init_num,
   }
 
   // Normalization
-  uint32_t clz = cpp::countl_zero<StorageType>(mantissa);
+  uint32_t clz = cpp::countl_zero(mantissa) -
+                 ((sizeof(UInt128) - sizeof(StorageType)) * CHAR_BIT);
   mantissa <<= clz;
 
   int32_t exp2 =
@@ -276,9 +277,8 @@ eisel_lemire<long double>(ExpandedFloat<long double> init_num,
   // Shifting to 65 bits for 80 bit floats and 113 bits for 128 bit floats
   uint32_t msb =
       static_cast<uint32_t>(final_approx_upper >> (FPBits::STORAGE_LEN - 1));
-  StorageType final_mantissa =
-      final_approx_upper >>
-      (msb + FPBits::STORAGE_LEN - (FPBits::FRACTION_LEN + 3));
+  UInt128 final_mantissa = final_approx_upper >> (msb + FPBits::STORAGE_LEN -
+                                                  (FPBits::FRACTION_LEN + 3));
   exp2 -= static_cast<uint32_t>(1 ^ msb); // same as !msb
 
   if (round == RoundDirection::Nearest) {
@@ -315,7 +315,7 @@ eisel_lemire<long double>(ExpandedFloat<long double> init_num,
   }
 
   ExpandedFloat<long double> output;
-  output.mantissa = final_mantissa;
+  output.mantissa = static_cast<StorageType>(final_mantissa);
   output.exponent = exp2;
   return output;
 }
@@ -558,7 +558,7 @@ clinger_fast_path(ExpandedFloat<T> init_num,
 
   FPBits result;
   T float_mantissa;
-  if constexpr (cpp::is_same_v<StorageType, UInt<128>>) {
+  if constexpr (is_big_int_v<StorageType> || sizeof(T) > sizeof(uint64_t)) {
     float_mantissa =
         (static_cast<T>(uint64_t(mantissa >> 64)) * static_cast<T>(0x1.0p64)) +
         static_cast<T>(uint64_t(mantissa));

libc/test/UnitTest/LibcTest.cpp

@@ -44,7 +44,6 @@ cpp::enable_if_t<(cpp::is_integral_v<T> && (sizeof(T) > sizeof(uint64_t))) ||
                      is_big_int_v<T>,
                  cpp::string>
 describeValue(T Value) {
-  static_assert(sizeof(T) % 8 == 0, "Unsupported size of UInt");
   const IntegerToString<T, radix::Hex::WithPrefix> buffer(Value);
   return buffer.view();
 }
@@ -242,6 +241,7 @@ TEST_SPECIALIZATION(__uint128_t);
 
 TEST_SPECIALIZATION(LIBC_NAMESPACE::Int<128>);
 
+TEST_SPECIALIZATION(LIBC_NAMESPACE::UInt<96>);
 TEST_SPECIALIZATION(LIBC_NAMESPACE::UInt<128>);
 TEST_SPECIALIZATION(LIBC_NAMESPACE::UInt<192>);
 TEST_SPECIALIZATION(LIBC_NAMESPACE::UInt<256>);