Improve float conversion

include/clasp/core/bignum.h

@@ -86,6 +86,12 @@ class Bignum_O : public Integer_O {
   static Bignum_sp create(int64_t v) { return create_from_limbs((v < 0) ? -1 : 1, std::abs(v), true); }
 #endif
   static Bignum_sp create(uint64_t v) { return create_from_limbs(1, v, true); }
+  static Bignum_sp create(__uint128_t v) {
+    Bignum_sp b = create_from_limbs(2);
+    b->_limbs[0] = static_cast<mp_limb_t>(v);
+    b->_limbs[1] = static_cast<mp_limb_t>(v >> 64);
+    return b;
+  }
   static Bignum_sp create(double d) {
     // KLUDGE: there is no mpn function for converting from a double.
     // However, this conses, which we shouldn't need to do.
@@ -133,7 +139,26 @@ class Bignum_O : public Integer_O {
   virtual bool evenp_() const override { return !((this->limbs())[0] & 1); }
   virtual bool oddp_() const override { return (this->limbs())[0] & 1; }
 
-  template <std::integral integral> integral to_integral() const {
+  template <std::unsigned_integral integral> integral to_integral() const {
+    constexpr auto limb_width = sizeof(mp_limb_t) * 8;
+    constexpr auto integral_width = std::bit_width(std::numeric_limits<integral>::max());
+    constexpr auto max_limb_count = 1 + ((sizeof(integral) - 1) / sizeof(mp_limb_t));
+
+    const mp_limb_t* limbs = this->limbs();
+    mp_size_t len = this->length();
+
+    if ((len > -1) && ((limb_width * (len - 1) + std::bit_width(limbs[len - 1])) <= integral_width)) {
+      integral value = 0;
+      for (mp_size_t i = 0; i < length(); i++)
+        value |= (integral)limbs[i] << (i * limb_width);
+      return value;
+    }
+
+    TYPE_ERROR(this->asSmartPtr(), Cons_O::createList(cl::_sym_Integer_O, Integer_O::create(std::numeric_limits<integral>::min()),
+                                                      Integer_O::create(std::numeric_limits<integral>::max())));
+  }
+
+  template <std::signed_integral integral> integral to_integral() const {
     integral mn = std::numeric_limits<integral>::min();
     integral mx = std::numeric_limits<integral>::max();
     // First, if integral can only hold fixnums, conversion will always fail.

include/clasp/core/float_util.h

@@ -0,0 +1,165 @@
+#pragma once
+
+#include <fenv.h>
+
+namespace core {
+
+template <typename Float> struct float_convert {
+  static constexpr uint16_t significand_width = std::numeric_limits<Float>::digits;
+  static constexpr uint16_t exponent_width = std::bit_width((unsigned int)std::numeric_limits<Float>::max_exponent);
+  static constexpr uint16_t sign_width = 1;
+  static constexpr bool has_hidden_bit = ((sign_width + exponent_width + significand_width) % 8) != 0;
+  static constexpr uint16_t storage_width = sign_width + exponent_width + significand_width + ((has_hidden_bit) ? -1 : 0);
+  static constexpr int32_t exponent_bias = std::numeric_limits<Float>::max_exponent + significand_width - 2;
+  static constexpr int32_t max_exponent = std::numeric_limits<Float>::max_exponent - significand_width;
+  static constexpr int32_t min_exponent = 2 - exponent_bias - significand_width;
+  static constexpr int32_t min_normalized_exponent = 1 - exponent_bias;
+  using uint_t =
+      std::conditional_t<storage_width <= 8, uint8_t,
+                         std::conditional_t<storage_width <= 16, uint16_t,
+                                            std::conditional_t<storage_width <= 32, uint32_t,
+                                                               std::conditional_t<storage_width <= 64, uint64_t, __uint128_t>>>>;
+  static constexpr uint_t hidden_bit = has_hidden_bit ? uint_t{1} << (significand_width - 1) : uint_t{0};
+  static constexpr uint16_t exponent_shift = storage_width - sign_width - exponent_width;
+  static constexpr uint16_t sign_shift = storage_width - sign_width;
+  static constexpr uint_t significand_mask = (uint_t{1} << (significand_width + ((has_hidden_bit) ? -1 : 0))) - uint_t{1};
+  static constexpr uint_t payload_mask = (uint_t{1} << (significand_width + ((has_hidden_bit) ? -2 : -1))) - uint_t{1};
+  static constexpr int32_t non_finite_exponent = static_cast<int32_t>(((uint_t{1} << exponent_width) - uint_t{1}));
+  static constexpr uint_t exponent_mask = ((uint_t{1} << exponent_width) - uint_t{1}) << exponent_shift;
+  static constexpr uint_t sign_mask = ((uint_t{1} << sign_width) - uint_t{1}) << sign_shift;
+  static constexpr uint_t nan_type_mask = uint_t{1} << (significand_width + ((has_hidden_bit) ? -2 : -1));
+
+  enum class category { finite, quiet_nan, signaling_nan, infinity };
+
+  typedef union {
+    Float f;
+    uint_t b;
+  } convert_t;
+
+  static inline uint_t to_bits(Float f) {
+    convert_t convert = {.f = f};
+    return convert.b;
+  }
+
+  static inline Float from_bits(uint_t b) {
+    convert_t convert = {.b = b};
+    return convert.f;
+  }
+
+  struct quadruple {
+    category category;
+    uint_t significand;
+    int32_t exponent;
+    int16_t sign;
+  };
+
+  static quadruple to_quadruple(Float f) {
+    quadruple q;
+    uint_t b = to_bits(f);
+
+    q.significand = b & significand_mask;
+    q.exponent = static_cast<int32_t>((b & exponent_mask) >> exponent_shift);
+    q.sign = (b & sign_mask) ? int32_t{-1} : int32_t{1};
+
+    if (q.exponent != non_finite_exponent) {
+      q.category = category::finite;
+      if (q.exponent != 0) {
+        // Normal non-zero
+        q.significand |= hidden_bit;
+        q.exponent -= exponent_bias;
+      } else if (q.significand != 0) {
+        // Subnormal
+        int32_t shift = significand_width - std::bit_width(q.significand);
+        q.significand = q.significand << shift;
+        q.exponent = 1 - exponent_bias - shift;
+      }
+    } else if (q.significand == 0) {
+      q.category = category::infinity;
+    } else if (q.significand & nan_type_mask) {
+      q.category = category::quiet_nan;
+      q.significand &= ~nan_type_mask;
+    } else {
+      q.category = category::signaling_nan;
+    }
+
+    return q;
+  }
+
+  static Float from_quadruple(quadruple q) {
+    uint_t b = 0;
+
+    if (q.sign < 0)
+      b |= sign_mask;
+
+    switch (q.category) {
+    case category::infinity:
+      b |= exponent_mask;
+      break;
+    case category::quiet_nan:
+      b |= exponent_mask | nan_type_mask | (q.significand & payload_mask);
+      break;
+    case category::signaling_nan:
+      b |= exponent_mask | ((q.significand == 0) ? uint_t{1} : (q.significand & payload_mask));
+      break;
+    default:
+      if (q.significand != 0) {
+        uint_t significand = q.significand;
+        int32_t exponent = q.exponent;
+        int32_t shift = std::bit_width(significand) - significand_width;
+
+        // If we don't have enough bits then right shift.
+        if (shift < 0) {
+          significand = significand << -shift;
+          exponent += shift;
+          shift = 0;
+        }
+
+        // Check for subnormals and set the shift needed to normalize.
+        if ((exponent + shift) < min_normalized_exponent) {
+          shift = min_normalized_exponent - exponent;
+        }
+
+        // If we shift away all of the bits that is an underflow.
+        if (shift > std::bit_width(significand)) {
+          feraiseexcept(FE_UNDERFLOW);
+          // Return +/- zero if traps masked.
+          return from_bits(b);
+        }
+
+        // Round if we have extra bits.
+        if (shift > 0) {
+          if (significand & ((1 << shift) - 1))
+            feraiseexcept(FE_INEXACT);
+          significand = ((significand & (1 << (shift - 1))) ? (significand + (1 << shift)) : significand) >> shift;
+          exponent += shift;
+        }
+
+        // Check one more time to ensure rounding hasn't increased the width.
+        shift = std::max(static_cast<int>(std::bit_width(significand) - significand_width), 0);
+        significand = significand >> shift;
+        exponent += shift;
+
+        // Check for overflow.
+        if (exponent > max_exponent) {
+          feraiseexcept(FE_OVERFLOW);
+          // Return +/- infinity if traps masked.
+          return from_bits(b | exponent_mask);
+        }
+
+        if (std::bit_width(significand) < significand_width) {
+          // Subnormals
+          b |= significand & significand_mask;
+        } else {
+          // Normal
+          b |= (significand & significand_mask) |
+               ((static_cast<uint_t>(exponent + exponent_bias) << exponent_shift) & exponent_mask);
+        }
+      }
+      break;
+    }
+
+    return from_bits(b);
+  };
+};
+
+} // namespace core

include/clasp/core/foundation.h

@@ -51,46 +51,10 @@ THE SOFTWARE.
 
 namespace core {
 [[noreturn]] void lisp_throwLispError(const std::string& str);
-[[noreturn]] void lisp_floating_point_invalid_operation();
 [[noreturn]] void lisp_error_simple(const char* functionName, const char* fileName, int lineNumber, const string& fmt);
 [[noreturn]] void lisp_error_simple(const char* functionName, const char* fileName, int lineNumber, const std::string& str);
 void lisp_debugLogWrite(const char* fileName, const char* funcName, uint lineNumber, uint column, const std::string& message,
                         uint debugFlags = DEBUG_CPP_FUNCTION);
-
-template <typename Float> struct float_convert {
-  static constexpr uint16_t significand_width = std::numeric_limits<Float>::digits;
-  static constexpr uint16_t exponent_width = std::bit_width((unsigned int)std::numeric_limits<Float>::max_exponent);
-  static constexpr uint16_t sign_width = 1;
-  static constexpr bool has_hidden_bit = ((sign_width + exponent_width + significand_width) % 8) != 0;
-  static constexpr uint16_t storage_width = sign_width + exponent_width + significand_width + ((has_hidden_bit) ? -1 : 0);
-  static constexpr int32_t exponent_bias = std::numeric_limits<Float>::max_exponent + significand_width - 2;
-  using uint_t =
-      std::conditional_t<storage_width <= 8, uint8_t,
-                         std::conditional_t<storage_width <= 16, uint16_t,
-                                            std::conditional_t<storage_width <= 32, uint32_t,
-                                                               std::conditional_t<storage_width <= 64, uint64_t, __uint128_t>>>>;
-  static constexpr uint16_t exponent_shift = storage_width - sign_width - exponent_width;
-  static constexpr uint16_t sign_shift = storage_width - sign_width;
-  static constexpr uint_t significand_mask = (uint_t{1} << (significand_width + ((has_hidden_bit) ? -1 : 0))) - uint_t{1};
-  static constexpr uint_t exponent_mask = ((uint_t{1} << exponent_width) - uint_t{1}) << exponent_shift;
-  static constexpr uint_t sign_mask = ((uint_t{1} << sign_width) - uint_t{1}) << sign_shift;
-
-  typedef union {
-    Float f;
-    uint_t b;
-  } convert_t;
-
-  static inline uint_t to_bits(Float f) {
-    convert_t convert = {.f = f};
-    return convert.b;
-  }
-
-  static inline Float from_bits(uint_t b) {
-    convert_t convert = {.b = b};
-    return convert.f;
-  }
-};
-
 }; // namespace core
 
 template <typename Char> struct fmt::formatter<core::T_sp, Char> : fmt::formatter<fmt::basic_string_view<Char>> {

include/clasp/core/object.h

@@ -36,6 +36,7 @@ THE SOFTWARE.
 
 #include <fmt/format.h>
 #include <fmt/ostream.h>
+#include <clasp/core/float_util.h>
 #include <clasp/core/newhash.h>
 #include <clasp/core/commonLispPackage.fwd.h>
 #include <clasp/core/corePackage.fwd.h>

src/core/bignum.cc

@@ -799,35 +799,39 @@ Number_sp Bignum_O::oneMinus_() const { return next_fadd(this->limbs(), this->le
 
 Number_sp Bignum_O::onePlus_() const { return next_fadd(this->limbs(), this->length(), 1); }
 
-double next_to_double(mp_size_t len, const mp_limb_t* limbs) {
-  // MPN does not seem to export a float conversion function,
-  // so we roll our own. FIXME: Could get bad with float rounding
-  // or otherwise as I'm making this up as I go.
+template <typename Float> Float limbs_to_float(mp_size_t len, const mp_limb_t* limbs) {
+  constexpr size_t limb_width = sizeof(mp_limb_t) * 8;
   mp_size_t size = std::abs(len);
+  struct float_convert<Float>::quadruple q = {
+    .category = float_convert<Float>::category::finite, .significand = 0, .exponent = (size - 1) * limb_width,
+    .sign = (len < 0) ? -1 : 1
+  };
+  size_t shift = float_convert<Float>::significand_width + 1;
+  size_t width = std::bit_width(limbs[size - 1]);
+
+  if (width >= shift) {
+    q.significand = limbs[size - 1] >> (width - shift);
+    q.exponent += width - shift;
+  } else {
+    q.significand = limbs[size - 1];
+    shift -= width;
+
+    for (mp_size_t i = size - 2; i > -1 && shift > 0; i--) {
+      size_t limb_shift = std::min(shift, limb_width);
+      q.significand = (q.significand << limb_shift) | (limbs[i] >> (limb_width - limb_shift));
+      q.exponent -= limb_shift;
+      shift -= limb_shift;
+    }
+  }
 
-  // There are no length zero bignums.
-  // If the bignum is only one long, we just convert directly.
-  if (len == 1)
-    return static_cast<double>(limbs[0]);
-  else if (len == -1)
-    return -static_cast<double>(limbs[0]);
-
-  // Otherwise, we just use the most significant two limbs.
-  // Assuming the float format has at most 128 bits of significand,
-  // the lower limbs ought to be irrelevant.
-  // (In fact one 64-bit number would probably be enough, but
-  //  it's possible the high bits of this number are almost all zero.)
-  double ultimate = static_cast<double>(limbs[size - 1]);
-  double penultimate = static_cast<double>(limbs[size - 2]);
-  double soon = penultimate + std::ldexp(ultimate, 64);
-  return std::ldexp(((len < 0) ? -soon : soon), 64 * (size - 2));
+  return float_convert<Float>::from_quadruple(q);
 }
 
-float Bignum_O::as_float_() const { return static_cast<float>(mpz_get_d(this->mpz().get_mpz_t())); }
+float Bignum_O::as_float_() const { return limbs_to_float<float>(this->length(), this->limbs()); }
 
-double Bignum_O::as_double_() const { return mpz_get_d(this->mpz().get_mpz_t()); }
+double Bignum_O::as_double_() const { return limbs_to_float<double>(this->length(), this->limbs()); }
 
-LongFloat Bignum_O::as_long_float_() const { return static_cast<LongFloat>(next_to_double(this->length(), this->limbs())); }
+LongFloat Bignum_O::as_long_float_() const { return limbs_to_float<LongFloat>(this->length(), this->limbs()); }
 
 DOCGROUP(clasp);
 CL_DEFUN int core__next_compare(Bignum_sp left, Bignum_sp right) {

src/core/foundation.cc

@@ -214,10 +214,6 @@ CL_DEFUN void core__dump_class_ids() { reg::dump_class_ids(); }
 
 SYMBOL_EXPORT_SC_(ClPkg,floating_point_invalid_operation);
 
-void lisp_floating_point_invalid_operation() {
-  lisp_error(cl::_sym_floating_point_invalid_operation,nil<core::T_O>());
-}
-
 void lisp_errorExpectedList(core::T_O* v) {
   T_sp tv((gctools::Tagged)v);
   TYPE_ERROR(tv, cl::_sym_list);