sve masked load issues (#1491)

* sve: unit.memory.exe * Fix fixed<0> in x86 arithmetic Co-authored-by: Joel Falcou <[email protected]>
jfalcou · Dec 19, 2022 · 635277d · 635277d
1 parent f5bf7da
commit 635277d
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 26 deletions.
diff --git a/include/eve/detail/function/simd/arm/sve/load.hpp b/include/eve/detail/function/simd/arm/sve/load.hpp
@@ -20,50 +20,46 @@ template<relative_conditional_expr C,
          typename N,
          simd_compatible_ptr<wide<T, N>> Ptr>
 EVE_FORCEINLINE wide<T, N>
-load_(EVE_SUPPORTS(sve_), C const& cond, safe_type const&, eve::as<wide<T, N>> const&, Ptr p)
-requires(sve_abi<abi_t<T, N>>)
+load_(EVE_SUPPORTS(sve_),
+  C const& cond, safe_type const& s, eve::as<wide<T, N>> const& tgt, Ptr p)
+  requires (!has_bundle_abi_v<wide<T, N>>)
 {
   auto ptr = unalign(p);
 
-  if constexpr(C::is_complete && C::is_inverted)
+  if constexpr ( C::has_alternative )
   {
-    if constexpr( N() == expected_cardinal_v<T> ) return svld1(sve_true<T>(), ptr);
-    else return svldff1(sve_true<T>(), ptr);
+    auto res = load(drop_alternative(cond), s, tgt, p);
+    return eve::replace_ignored(res, cond, cond.alternative);
   }
-  else
+  else if constexpr ( C::is_complete && !C::is_inverted ) return {};
+  else if constexpr ( has_aggregated_abi_v<wide<T, N>> )
   {
-    wide<T, N> loaded;
+    using half = wide<T, typename N::split_type>;
 
-    if constexpr(!C::is_complete)
-    {
-      auto mask = cond.mask(as<wide<T,N>>{});
-      if constexpr( N() == expected_cardinal_v<T> ) loaded = svld1(mask, ptr);
-      else                                          loaded = svldff1(mask, ptr);
-      if constexpr( C::has_alternative )
-      {
-        loaded = svsel(mask, loaded, wide<T,N>(cond.alternative));
-      }
-    }
-    else
-    {
-      if constexpr( C::has_alternative ) loaded = wide<T,N>(cond.alternative);
-    }
+    // condition does not matter - just presense matters
+    auto lo = load(cond, s, as<half>{}, ptr);
+    auto hi = load(cond, s, as<half>{}, ptr + half::size());
 
-    return loaded;
+    return {lo, hi};
   }
+  else if constexpr(C::is_complete && C::is_inverted && N() == expected_cardinal_v<T> )
+  {
+    return svld1(sve_true<T>(), ptr);
+  }
+  else return svldff1(sve_true<T>(), ptr);
 }
 
 //================================================================================================
 // Logical support
 //================================================================================================
-template<relative_conditional_expr C, typename T, typename N, data_source Pointer>
+template<relative_conditional_expr C, typename T, typename N,
+      simd_compatible_ptr<logical<wide<T, N>>> Pointer>
 EVE_FORCEINLINE logical<wide<T, N>>
                 load_(EVE_SUPPORTS(sve_),
                       C const& cond,
                       safe_type const&,
                       eve::as<logical<wide<T, N>>> const&,
                       Pointer ptr) noexcept
-requires(dereference_as<logical<T>, Pointer>::value && sve_abi<abi_t<T, N>>)
 {
   auto const c1     = map_alternative(cond, [](auto alt) { return alt.mask(); });
   auto const block  = load(c1, safe, eve::as<wide<T, N>> {}, ptr_cast<T const>(ptr));

diff --git a/include/eve/detail/function/simd/x86/arithmetic_compounds.hpp b/include/eve/detail/function/simd/x86/arithmetic_compounds.hpp
@@ -180,7 +180,8 @@ namespace eve::detail
         }
         else
         {
-          using htype  = wide<std::int64_t, fixed<N::value / 2>>;
+          static constexpr auto half_size = (N::value/2 > 0) ? N::value/2 : 1;
+          using htype  = wide<std::int64_t, fixed<half_size>>;
           htype mhi    = _mm_setr_epi32(-1, 0, -1, 0);
           htype mlo    = mhi;
           auto  lself  = _mm_srli_si128(self, 4);
@@ -202,7 +203,8 @@ namespace eve::detail
       else  if constexpr( c == category::uint16x8   ) self = _mm_mullo_epi16(self, other);
       else  if constexpr( c == category::int8x16 || c == category::uint8x16 )
       {
-        using htype  = wide<std::int16_t, fixed<N::value / 2>>;
+        static constexpr auto half_size = (N::value/2 > 0) ? N::value/2 : 1;
+        using htype  = wide<std::int16_t, fixed<half_size>>;
         htype mhi    = _mm_set1_epi16(0x00FF);
         htype mlo    = mhi;
         auto  lself  = _mm_srli_epi16(self, 8);