Re: [PATCH] libstdc++: Format std::float16_t and std::bfloat16_t using respective std::to_chars overloads [PR126731]

On Thu, 20 Aug 2026 at 14:53, Tomasz Kaminski <[email protected]> wrote:
>
>
>
> On Thu, Aug 20, 2026 at 3:51 PM Tomasz Kaminski <[email protected]> wrote:
>>
>>
>>
>> On Thu, Aug 20, 2026 at 2:36 PM Jonathan Wakely <[email protected]> wrote:
>>>
>>> On Thu, 13 Aug 2026 at 10:36 +0200, Tomasz Kamiński wrote:
>>> >This patch removes the partial specialization for float16_t and bfloat16_t
>>> >that formatted the values by casting to float. In consequence an implicit
>>> >specialization of generic floating-point specialization is used, and
>>> >to_chars overloads for their types are used.
>>> >As explained in r13-3591-g0ae26533b3e268 (that adds corresponding to_chars
>>> >overload), the shortest string (produced when no precision is specified)
>>> >is shorter for above, than float.
>>> >
>>> >As formatter specializations were defined under the same condition as
>>> >corresponding to_chars overloads, the __formattable_float constrain on
>>> >formatter floating-point specialization makes it enable in same cases.
>>> >
>>> >This reverts parts of r14-3305-g6cf214b4fc97f5 (revereted partially
>>>
>>> "reverted"
>>>
>>> >in r14-3329-g27d0cfcb2b33de), leading to (necessary for correctness)
>>> >increase of the number template specializations. It also depends on
>>> >preserving values of above types (instead of float) in basic_format_arg
>>> >introduced r16-616-g9c9a7316adb996.
>>>
>>> Ah, so I added these specializations in r14-3305-g6cf214b4fc97f5 to
>>> support printing extended floating-point types in C++20, but that
>>> turned out to not work. So I adjusted the conditions for enabling them
>>> in r14-3329-g27d0cfcb2b33de, so that they are only enabled for C++23.
>>> But really we should have just removed the specializations again,
>>> because the __formattable_float specialization formats them properly
>>> since r16-616-g9c9a7316adb996.
>>>
>>> >The specializations for float32_t and float64_t are left unchanged,
>>> >as their to_chars overloads are also implemented in terms of casting
>>> >to float/double respectivelly.
>>> >
>>> >libstdc++-v3/ChangeLog:
>>> >
>>> >       PR libstdc++/126731
>>> >       * include/std/format (std::formatter<_Float16, _CharT>)
>>> >       (std::formatter<__format::__bflt16_t, _CharT>): Remove
>>> >       partial specializations.
>>> >       * testsuite/std/format/formatter/ext_float.cc: Test output
>>> >       for closest value to 1/10.
>>> >---
>>> >As far as I can tell this case will not be affected by P3505R4
>>> >"Fix the default floating-point representation in std::format",
>>> >as float(0.1f16) is 1.998p-4, while 0.1 f is 1.99999ap-4, so
>>> >only later could use 0.1 as shortest string.
>>> >
>>> >Testing on x86_64-linux. *format* passed in all standard modes,
>>> >-m32, debug and assertions. OK for trunk and 16?
>>> >(We cannot backport it to eariel standards due dependency described
>>> >above).
>>>
>>> Yes OK for trunk and gcc-16.
>>>
>>> If I understand correctly, I think we could remove them in gcc-14 and
>>> gcc-15, because they're redundant (the __formattable_float
>>> specialization can format them). But we would still get incorrect
>>> results, because basic_format_arg would convert them to float.
>>
>> We will get inconsistent results, depending on how you format them:
>>  * if (b)float16 is passed as argument to format call, then ye we will convert
>>    them to float, and get correct result
>
> I meant incorrect output in this case.
>>
>>  * if you call formatter<(b)float16> direclty (for example becaue you are writting
>>   formatter for a class with a member), you will get correct output
>>>
>>> So the extra specializations in gcc-14 and gcc-15 are not doing any harm, and
>>> it doesn't seem worth backporting it.
>>
>> They are keeping us consistently wrong; if we remove them, we will be inconsistent.
>> and wrong in the usual case.

Ah yes, so definitely worth keeping them on the older branches then.


>>>
>>>
>>>
>>> > libstdc++-v3/include/std/format               | 44 -------------------
>>> > .../std/format/formatter/ext_float.cc         | 22 ++++++----
>>> > 2 files changed, 13 insertions(+), 53 deletions(-)
>>> >
>>> >diff --git a/libstdc++-v3/include/std/format b/libstdc++-v3/include/std/format
>>> >index 729cb89ec60..8d530e883c3 100644
>>> >--- a/libstdc++-v3/include/std/format
>>> >+++ b/libstdc++-v3/include/std/format
>>> >@@ -3003,28 +3003,6 @@ namespace __format
>>> >     };
>>> > #endif
>>> >
>>> >-#if defined(__STDCPP_FLOAT16_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
>>> >-  // Reuse __formatter_fp<C>::format<float, Out> for _Float16.
>>> >-  template<__format::__char _CharT>
>>> >-    struct formatter<_Float16, _CharT>
>>> >-    {
>>> >-      formatter() = default;
>>> >-
>>> >-      [[__gnu__::__always_inline__]]
>>> >-      constexpr typename basic_format_parse_context<_CharT>::iterator
>>> >-      parse(basic_format_parse_context<_CharT>& __pc)
>>> >-      { return _M_f.parse(__pc); }
>>> >-
>>> >-      template<typename _Out>
>>> >-      typename basic_format_context<_Out, _CharT>::iterator
>>> >-      format(_Float16 __u, basic_format_context<_Out, _CharT>& __fc) const
>>> >-      { return _M_f.format((float)__u, __fc); }
>>> >-
>>> >-    private:
>>> >-      __format::__formatter_fp<_CharT> _M_f;
>>> >-    };
>>> >-#endif
>>> >-
>>> > #if defined(__FLT32_DIG__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
>>> >   // Reuse __formatter_fp<C>::format<float, Out> for _Float32.
>>> >   template<__format::__char _CharT>
>>> >@@ -3114,28 +3092,6 @@ namespace __format
>>> >     };
>>> > #endif
>>> >
>>> >-#if defined(__STDCPP_BFLOAT16_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
>>> >-  // Reuse __formatter_fp<C>::format<float, Out> for bfloat16_t.
>>> >-  template<__format::__char _CharT>
>>> >-    struct formatter<__format::__bflt16_t, _CharT>
>>> >-    {
>>> >-      formatter() = default;
>>> >-
>>> >-      [[__gnu__::__always_inline__]]
>>> >-      constexpr typename basic_format_parse_context<_CharT>::iterator
>>> >-      parse(basic_format_parse_context<_CharT>& __pc)
>>> >-      { return _M_f.parse(__pc); }
>>> >-
>>> >-      template<typename _Out>
>>> >-      typename basic_format_context<_Out, _CharT>::iterator
>>> >-      format(__gnu_cxx::__bfloat16_t __u,
>>> >-             basic_format_context<_Out, _CharT>& __fc) const
>>> >-      { return _M_f.format((float)__u, __fc); }
>>> >-
>>> >-    private:
>>> >-      __format::__formatter_fp<_CharT> _M_f;
>>> >-    };
>>> >-#endif
>>> > #endif // __cpp_lib_to_chars
>>> >
>>> >   /** Format a pointer.
>>> >diff --git a/libstdc++-v3/testsuite/std/format/formatter/ext_float.cc b/libstdc++-v3/testsuite/std/format/formatter/ext_float.cc
>>> >index 82e696fa797..7b231b07006 100644
>>> >--- a/libstdc++-v3/testsuite/std/format/formatter/ext_float.cc
>>> >+++ b/libstdc++-v3/testsuite/std/format/formatter/ext_float.cc
>>> >@@ -4,10 +4,14 @@
>>> > #include <testsuite_hooks.h>
>>> >
>>> > template<typename T>
>>> >-bool format_float()
>>> >+void
>>> >+verify_output()
>>> > {
>>> >-    auto s = std::format("{:#} != {:<+7.3f}", (T)-0.0, (T)0.5);
>>> >-    return s == "-0. != +0.500 ";
>>> >+  auto s = std::format("{:#} != {:<+7.3f}", T(-0.0), T(0.5));
>>> >+  VERIFY( s == "-0. != +0.500 ");
>>> >+
>>> >+  s = std::format("{}", T(1)/T(10));
>>> >+  VERIFY( s == "0.1");
>>> > }
>>> >
>>> > #if __cplusplus > 202002L
>>> >@@ -23,7 +27,7 @@ test_float16()
>>> > {
>>> > #if __FLT16_DIG__
>>> >   if constexpr (formattable<_Float16>)
>>> >-    VERIFY( format_float<_Float16>() );
>>> >+    verify_output<_Float16>();
>>> >   else
>>> >     std::puts("Cannot format _Float16 on this target");
>>> > #endif
>>> >@@ -34,7 +38,7 @@ test_float32()
>>> > {
>>> > #if __FLT32_DIG__
>>> >   if constexpr (formattable<_Float32>)
>>> >-    VERIFY( format_float<_Float32>() );
>>> >+    verify_output<_Float32>();
>>> >   else
>>> >     std::puts("Cannot format _Float32 on this target");
>>> > #endif
>>> >@@ -45,7 +49,7 @@ test_float64()
>>> > {
>>> > #if __FLT64_DIG__
>>> >   if constexpr (formattable<_Float64>)
>>> >-    VERIFY( format_float<_Float64>() );
>>> >+    verify_output<_Float64>();
>>> >   else
>>> >     std::puts("Cannot format _Float64 on this target");
>>> > #endif
>>> >@@ -56,13 +60,13 @@ test_float128()
>>> > {
>>> > #ifdef __SIZEOF_FLOAT128__
>>> >   if constexpr (formattable<__float128>)
>>> >-    VERIFY( format_float<__float128>() );
>>> >+    verify_output<__float128>();
>>> >   else
>>> >     std::puts("Cannot format __float128 on this target");
>>> > #endif
>>> > #if __FLT128_DIG__
>>> >   if constexpr (formattable<_Float128>)
>>> >-    VERIFY( format_float<_Float128>() );
>>> >+    verify_output<_Float128>();
>>> >   else
>>> >     std::puts("Cannot format _Float128 on this target");
>>> > #endif
>>> >@@ -75,7 +79,7 @@ test_bfloat16()
>>> >   using bfloat16_t = decltype(0.0bf16);
>>> >
>>> >   if constexpr (formattable<bfloat16_t>)
>>> >-    VERIFY( format_float<bfloat16_t>() );
>>> >+    verify_output<bfloat16_t>();
>>> >   else
>>> >     std::puts("Cannot format bfloat16_t on this target");
>>> > #endif
>>> >--
>>> >2.55.0
>>> >
>>> >
>>>