diff --git a/include/fmt/format-inl.h b/include/fmt/format-inl.h
index 511c8d85..4f0d92a0 100644
--- a/include/fmt/format-inl.h
+++ b/include/fmt/format-inl.h
@@ -237,11 +237,7 @@ struct fp {
   template <typename Float> explicit FMT_CONSTEXPR fp(Float n) { assign(n); }
 
   template <typename Float>
-  using is_supported =
-      bool_constant<sizeof(Float) == sizeof(uint32_t) ||
-                    sizeof(Float) == sizeof(uint64_t) ||
-                    (sizeof(Float) == sizeof(uint128_t) &&
-                     std::numeric_limits<Float>::digits == 64)>;
+  using is_supported = bool_constant<std::numeric_limits<Float>::digits <= 64>;
 
   // Assigns d to this and return true iff predecessor is closer than successor.
   template <typename Float, FMT_ENABLE_IF(is_supported<Float>::value)>
diff --git a/include/fmt/format.h b/include/fmt/format.h
index 08c1f7d0..9e086669 100644
--- a/include/fmt/format.h
+++ b/include/fmt/format.h
@@ -288,9 +288,8 @@ template <typename Streambuf> class formatbuf : public Streambuf {
 };
 
 // Implementation of std::bit_cast for pre-C++20.
-template <typename To, typename From>
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) == sizeof(From))>
 FMT_CONSTEXPR20 auto bit_cast(const From& from) -> To {
-  static_assert(sizeof(To) == sizeof(From), "size mismatch");
 #ifdef __cpp_lib_bit_cast
   if (is_constant_evaluated()) return std::bit_cast<To>(from);
 #endif
@@ -369,28 +368,10 @@ using uint128_t = conditional_t<FMT_USE_INT128, uint128_opt, uint128_fallback>;
 
 #ifdef UINTPTR_MAX
 using uintptr_t = ::uintptr_t;
-inline auto to_uintptr(const void* p) -> uintptr_t {
-  return bit_cast<uintptr_t>(p);
-}
 #else
 using uintptr_t = uint128_t;
 #endif
 
-// A fallback for systems that lack uintptr_t.
-template <typename T = uint128_t> inline auto to_uintptr(const void* p) -> T {
-  constexpr auto size = static_cast<int>(sizeof(void*));
-  struct data_t {
-    unsigned char value[size];
-  } data = bit_cast<data_t>(p);
-  auto result = T();
-  if (const_check(is_big_endian())) {
-    for (int i = 0; i < size; ++i) result = (result << 8) | data.value[i];
-  } else {
-    for (int i = size - 1; i >= 0; --i) result = (result << 8) | data.value[i];
-  }
-  return result;
-}
-
 // Returns the largest possible value for type T. Same as
 // std::numeric_limits<T>::max() but shorter and not affected by the max macro.
 template <typename T> constexpr auto max_value() -> T {
@@ -403,6 +384,25 @@ template <typename T> constexpr auto num_bits() -> int {
 template <> constexpr auto num_bits<int128_opt>() -> int { return 128; }
 template <> constexpr auto num_bits<uint128_t>() -> int { return 128; }
 
+// A heterogeneous bit_cast used for converting 96-bit long double to uint128_t
+// and 128-bit pointers to uint128_fallback.
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) > sizeof(From))>
+inline auto bit_cast(const From& from) -> To {
+  constexpr auto size = static_cast<int>(sizeof(From) / sizeof(unsigned));
+  struct data_t {
+    unsigned value[size];
+  } data = bit_cast<data_t>(from);
+  auto result = To();
+  if (const_check(is_big_endian())) {
+    for (int i = 0; i < size; ++i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  } else {
+    for (int i = size - 1; i >= 0; --i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  }
+  return result;
+}
+
 FMT_INLINE void assume(bool condition) {
   (void)condition;
 #if FMT_HAS_BUILTIN(__builtin_assume) && !FMT_ICC_VERSION
@@ -1275,7 +1275,7 @@ template <> struct float_info<double> {
 template <typename T>
 struct float_info<T, enable_if_t<std::is_same<T, long double>::value &&
                                  std::numeric_limits<T>::digits == 64>> {
-  using carrier_uint = detail::uint128_opt;
+  using carrier_uint = detail::uint128_t;
   static const int significand_bits = 64;
   static const int exponent_bits = 15;
 };
@@ -1911,7 +1911,7 @@ FMT_CONSTEXPR auto write(OutputIt out, const Char* s,
     -> OutputIt {
   return check_cstring_type_spec(specs.type)
              ? write(out, basic_string_view<Char>(s), specs, {})
-             : write_ptr<Char>(out, to_uintptr(s), &specs);
+             : write_ptr<Char>(out, bit_cast<uintptr_t>(s), &specs);
 }
 
 template <typename Char, typename OutputIt>
@@ -2394,7 +2394,7 @@ auto write(OutputIt out, const T* value,
            const basic_format_specs<Char>& specs = {}, locale_ref = {})
     -> OutputIt {
   check_pointer_type_spec(specs.type, error_handler());
-  return write_ptr<Char>(out, to_uintptr(value), &specs);
+  return write_ptr<Char>(out, bit_cast<uintptr_t>(value), &specs);
 }
 
 // A write overload that handles implicit conversions.
diff --git a/test/format-impl-test.cc b/test/format-impl-test.cc
index 2cdc2899..41f736c6 100644
--- a/test/format-impl-test.cc
+++ b/test/format-impl-test.cc
@@ -340,16 +340,6 @@ TEST(format_impl_test, count_digits) {
   test_count_digits<uint64_t>();
 }
 
-TEST(format_impl_test, write_uintptr_fallback) {
-  std::string s;
-  fmt::detail::write_ptr<char>(
-      std::back_inserter(s),
-      fmt::detail::to_uintptr<fmt::detail::uint128_fallback>(
-          reinterpret_cast<void*>(0xface)),
-      nullptr);
-  EXPECT_EQ(s, "0xface");
-}
-
 #ifdef _WIN32
 #  include <windows.h>
 #endif
diff --git a/test/format-test.cc b/test/format-test.cc
index e7b23b7c..c7a468ad 100644
--- a/test/format-test.cc
+++ b/test/format-test.cc
@@ -1445,6 +1445,18 @@ TEST(format_test, format_pointer) {
   EXPECT_EQ("0x0", fmt::format("{}", nullptr));
 }
 
+TEST(format_test, write_uintptr_fallback) {
+  // Test that formatting a pointer by converting it to uint128_fallback works.
+  // This is needed to support systems without uintptr_t.
+  auto s = std::string();
+  fmt::detail::write_ptr<char>(
+      std::back_inserter(s),
+      fmt::detail::bit_cast<fmt::detail::uint128_fallback>(
+          reinterpret_cast<void*>(0xface)),
+      nullptr);
+  EXPECT_EQ(s, "0xface");
+}
+
 enum class color { red, green, blue };
 
 TEST(format_test, format_enum_class) {