really fix lshift for avx2

diff --git a/src/util/supervector/arch/x86/impl.cpp b/src/util/supervector/arch/x86/impl.cpp
index 855cc3fab1ddc5bce9f56c6399c538a07c5a3149..be1ee8fb9719a551144cf51cb36351355bd6458d 100644 (file)
--- a/src/util/supervector/arch/x86/impl.cpp
+++ b/src/util/supervector/arch/x86/impl.cpp
@@ -578,37 +578,37 @@ template <>
 really_inline SuperVector<32> SuperVector<32>::operator<<(uint8_t const N) const
 {
        switch(N) {
-       case 1: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 15)}; break;
-       case 2: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 14)}; break;
-       case 3: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 13)}; break;
-       case 4: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 12)}; break;
-       case 5: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 11)}; break;
-       case 6: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 10)}; break;
-       case 7: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 9)}; break;
-       case 8: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 8)}; break;
-       case 9: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 7)}; break;
-       case 10: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 6)}; break;
-       case 11: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 5)}; break;
-       case 12: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 4)}; break;
-       case 13: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 3)}; break;
-       case 14: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 2)}; break;
-       case 15: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 1)}; break;
+       case 1: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break;
+       case 2: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break;
+       case 3: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break;
+       case 4: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break;
+       case 5: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break;
+       case 6: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break;
+       case 7: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break;
+       case 8: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break;
+       case 9: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break;
+       case 10: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break;
+       case 11: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break;
+       case 12: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break;
+       case 13: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break;
+       case 14: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break;
+       case 15: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break;
        case 16: return {_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0))}; break;
-       case 17: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break;
-       case 18: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break;
-       case 19: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break;
-       case 20: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break;
-       case 21: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break;
-       case 22: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break;
-       case 23: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break;
-       case 24: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break;
-       case 25: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break;
-       case 26: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break;
-       case 27: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break;
-       case 28: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break;
-       case 29: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break;
-       case 30: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break;
-       case 31: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break;
+       case 17: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break;
+       case 18: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break;
+       case 19: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break;
+       case 20: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break;
+       case 21: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break;
+       case 22: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break;
+       case 23: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break;
+       case 24: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break;
+       case 25: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break;
+       case 26: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break;
+       case 27: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break;
+       case 28: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break;
+       case 29: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break;
+       case 30: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break;
+       case 31: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break;
        case 32: return Zeroes(); break;
        default: break;
        }