[Bug libstdc++/114958] use __builtin_shufflevector for std::experimental::simd split and concat (at least the common cases) to enable better optimizations

[mkretz at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114958

--- Comment #6 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
The last commit introduced a regression on i686 where __builtin_shufflevector
was producing MMX vectors (which can mess up the FPU). Untested patch which
resolves the issue:

libstdc++-v3/ChangeLog:

        PR libstdc++/114958
        * include/experimental/bits/simd.h (__as_vector): Don't use
        vector_size(8) on __i386__.
        (__vec_shuffle): Never return MMX vectors, widen to 16 bytes
        instead.
        (concat): Fix padding calculation to pick up widening logic from
        __as_vector.

diff --git a/libstdc++-v3/include/experimental/bits/simd.h
b/libstdc++-v3/include/experimental/bits/simd.h
index 6a6fd4f109d..63951df488c 100644
--- a/libstdc++-v3/include/experimental/bits/simd.h
+++ b/libstdc++-v3/include/experimental/bits/simd.h
@@ -1665,7 +1665,12 @@ __as_vector(_V __x)
          {
            static_assert(is_simd<_V>::value);
            using _Tp = typename _V::value_type;
+#ifdef __i386__
+           constexpr auto __bytes = sizeof(_Tp) == 8 ? 16 : sizeof(_Tp);
+           using _RV [[__gnu__::__vector_size__(__bytes)]] = _Tp;
+#else
            using _RV [[__gnu__::__vector_size__(sizeof(_Tp))]] = _Tp;
+#endif
            return _RV{__data(__x)};
          }
       }
@@ -2081,11 +2086,14 @@ __not(_Tp __a) noexcept
 // }}}
 // __vec_shuffle{{{
 template <typename _T0, typename _T1, typename _Fun, size_t... _Is>
-  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  _GLIBCXX_SIMD_INTRINSIC constexpr
+  __vector_type_t<typename _VectorTraits<_T0>::value_type, sizeof...(_Is)>
   __vec_shuffle(_T0 __x, _T1 __y, index_sequence<_Is...> __seq, _Fun
__idx_perm)
   {
     constexpr int _N0 = sizeof(__x) / sizeof(__x[0]);
     constexpr int _N1 = sizeof(__y) / sizeof(__y[0]);
+    using _Tp = typename _VectorTraits<_T0>::value_type;
+    using _RV [[maybe_unused]] = __vector_type_t<_Tp, sizeof...(_Is)>;
 #if __has_builtin(__builtin_shufflevector)
 #ifdef __clang__
     // Clang requires _T0 == _T1
@@ -2105,14 +2113,23 @@ __not(_Tp __a) noexcept
             });
     else
 #endif
-      return __builtin_shufflevector(__x, __y, [=] {
-              constexpr int __j = __idx_perm(_Is);
-              static_assert(__j < _N0 + _N1);
-              return __j;
-            }()...);
+      {
+       const auto __r = __builtin_shufflevector(__x, __y, [=] {
+                          constexpr int __j = __idx_perm(_Is);
+                          static_assert(__j < _N0 + _N1);
+                          return __j;
+                        }()...);
+#ifdef __i386__
+       if constexpr (sizeof(__r) == sizeof(_RV))
+         return __r;
+       else
+         return _RV {__r[_Is]...};
+#else
+       return __r;
+#endif
+      }
 #else
-    using _Tp = __remove_cvref_t<decltype(__x[0])>;
-    return __vector_type_t<_Tp, sizeof...(_Is)> {
+    return _RV {
       [=]() -> _Tp {
        constexpr int __j = __idx_perm(_Is);
        static_assert(__j < _N0 + _N1);
@@ -4393,9 +4410,9 @@ for (unsigned __j = 0; __j < __i; ++__j)
                __vec_shuffle(__as_vector(__xs)...,
std::make_index_sequence<_RW::_S_full_size>(),
                              [](int __i) {
                                constexpr int __sizes[2] =
{int(simd_size_v<_Tp, _As>)...};
-                               constexpr int __padding0
-                                 = sizeof(__vector_type_t<_Tp, __sizes[0]>) /
sizeof(_Tp)
-                                     - __sizes[0];
+                               constexpr int __vsizes[2]
+                                 = {int(sizeof(__as_vector(__xs)) /
sizeof(_Tp))...};
+                               constexpr int __padding0 = __vsizes[0] -
__sizes[0];
                                return __i >= _Np ? -1 : __i < __sizes[0] ? __i
: __i + __padding0;
                              })};
       }