fix-gcc-10

[DiamonDinoia]

On my machine with gcc-10 I get the error:

/mnt/home/mbarbone/repos/xsimd/include/xsimd/types/../arch/./xsimd_avx512f.hpp:2740:59: error: '_mm512_cvtsi512_si32' was not declared in this scope; did you mean '_mm512_castsi512_si128'?
 2740 |                 return static_cast<T>(_mm512_cvtsi512_si32(self) & 0xFF);
      |                                       ~~~~~~~~~~~~~~~~~~~~^~~~~~
      |                                       _mm512_castsi512_si128
/mnt/home/mbarbone/repos/xsimd/include/xsimd/types/../arch/./xsimd_avx512f.hpp:2744:59: error: '_mm512_cvtsi512_si32' was not declared in this scope; did you mean '_mm512_castsi512_si128'?
 2744 |                 return static_cast<T>(_mm512_cvtsi512_si32(self) & 0xFFFF);
      |                                       ~~~~~~~~~~~~~~~~~~~~^~~~~~
      |                                       _mm512_castsi512_si128
/mnt/home/mbarbone/repos/xsimd/include/xsimd/types/../arch/./xsimd_avx512f.hpp:2748:59: error: '_mm512_cvtsi512_si32' was not declared in this scope; did you mean '_mm512_castsi512_si128'?
 2748 |                 return static_cast<T>(_mm512_cvtsi512_si32(self));
      |                                       ~~~~~~~~~~~~~~~~~~~~^~~~~~
      |                                       _mm512_castsi512_si128

This fixes it for me.

[serge-sans-paille]

Thanks. It would be great to have a CI reproducer first, could you have a look or do you want me to handle it?

[DiamonDinoia]

Is #1243 okay? https://github.com/DiamonDinoia/xsimd/actions/runs/20831227045/job/59845642039

[DiamonDinoia]

I could cherry pick the commit here in case you want to merge them in one go

[serge-sans-paille]

I could cherry pick the commit here in case you want to merge them in one go

yes, please do, first the ci commit then the fix :-)

[DiamonDinoia]

The fact that it breaks it is a good thing for me. I was chasing bugs in finufft with gcc-10. This might be the cause. flatironinstitute/finufft#780

[DiamonDinoia]

diff --git a/test/test_shuffle.cpp b/test/test_shuffle.cpp
index b082109..f500072 100644
--- a/test/test_shuffle.cpp
+++ b/test/test_shuffle.cpp
@@ -672,10 +672,8 @@ struct shuffle_test
             }
         };
 
-        std::array<value_type, size> ref_lo;
-        for (size_t i = 0; i < size; ++i)
-            ref_lo[i] = (i & 1) ? rhs[i / 2] : lhs[i / 2];
-        B b_ref_lo = B::load_unaligned(ref_lo.data());
+        // Use zip_lo as a stable reference for the expected interleave.
+        B b_ref_lo = xsimd::zip_lo(b_lhs, b_rhs);
 
         INFO("zip_lo");
         B b_res_lo = xsimd::shuffle(b_lhs, b_rhs, xsimd::make_batch_constant<mask_type, zip_lo_generator, arch_type>());
@@ -689,12 +687,8 @@ struct shuffle_test
             }
         };
 
-        std::array<value_type, size> ref_hi;
-        for (size_t i = 0; i < size; ++i)
-        {
-            ref_hi[i] = (i & 1) ? rhs[size / 2 + i / 2] : lhs[size / 2 + i / 2];
-        }
-        B b_ref_hi = B::load_unaligned(ref_hi.data());
+        // Use zip_hi as a stable reference for the expected interleave.
+        B b_ref_hi = xsimd::zip_hi(b_lhs, b_rhs);

the issue seems to be gcc-10 auto vectorization not xsimd explicit one.

Adding that this code too causes issues: https://marco.godbolt.org/z/bjo6o8TfG

[serge-sans-paille]

if that's a GCC issue, specific to gcc-10, we could add something specific, but is that worth the effort?

[DiamonDinoia]

gcc-10 has a problem with avx512 but there is also an issue that I introduced when changing the avx512 swizzle: flatironinstitute/finufft#780 (comment)

I should fix that. For gcc-10 if this does not pass CI, I am not sure what is best. I just need xsimd to work with gcc-10 as is not too old of a compiler otherwise I need a #error in finufft if avx512 is used to compile finufft.