Avoid UB on clang

MariaDB · Apr 29, 2024 · 51200ac · 51200ac
1 parent 9eec158
commit 51200ac
Showing 1 changed file with 20 additions and 7 deletions.
diff --git a/mysys/crc32/crc32c_x86.cc b/mysys/crc32/crc32c_x86.cc
@@ -160,6 +160,21 @@ static inline __m128i load128(const char *buf)
 #define xor128(a, b) _mm_xor_epi64(a, b)
 #define and128(a, b) _mm_and_si128(a, b)
 
+/** Shift right by 384 bits */
+USE_VPCLMULQDQ
+static inline __m512i shrl512_384(__m512i a)
+{
+#if defined __GNUC__ && __GNUC__ >= 11
+ /* While technically incorrect, this would seem to translate into a
+ vextracti32x4 instruction, which actually outputs a ZMM register
+ (anything above the XMM range is cleared). */
+ return _mm512_castsi128_si512(_mm512_extracti64x2_epi64(a, 3));
+#else
+ /* On clang, this is needed in order to get a correct result. */
+ return _mm512_maskz_shuffle_i64x2(3, a, a, 3);
+#endif
+}
+
 USE_VPCLMULQDQ ATTRIBUTE_NOINLINE
 static unsigned crc32_avx512(unsigned crc, const char *buf, size_t size,
  const crc32_tab &tab)
@@ -224,13 +239,11 @@ static unsigned crc32_avx512(unsigned crc, const char *buf, size_t size,
  b896= _mm512_load_epi32(&tab.b896),
  b384= _mm512_load_epi32(&tab.b384);
 
- crc_out= _mm512_extracti64x2_epi64(m4, 3);
- __m512i m1;
- m1= xor3_512(_mm512_clmulepi64_epi128(m0, b896, 1),
- _mm512_clmulepi64_epi128(m0, b896, 0x10),
- _mm512_clmulepi64_epi128(m4, b384, 1));
+ __m512i m1= xor3_512(_mm512_clmulepi64_epi128(m0, b896, 1),
+ _mm512_clmulepi64_epi128(m0, b896, 0x10),
+ _mm512_clmulepi64_epi128(m4, b384, 1));
  m1= xor3_512(m1, _mm512_clmulepi64_epi128(m4, b384, 0x10),
- _mm512_castsi128_si512(crc_out));
+ shrl512_384(m4));
 
  __m256i m8=
  _mm512_castsi512_si256(_mm512_shuffle_i64x2(m1, m1, 0b01001110));
@@ -273,7 +286,7 @@ static unsigned crc32_avx512(unsigned crc, const char *buf, size_t size,
  __m512i crc512=
  xor3_512(_mm512_clmulepi64_epi128(m0, b384, 0x1),
  _mm512_clmulepi64_epi128(m0, b384, 0x10),
- _mm512_castsi128_si512(_mm512_extracti64x2_epi64(m0, 3)));
+ shrl512_384(m0));
  crc512= xor512(crc512, _mm512_shuffle_i64x2(crc512, crc512, 0b01001110));
  const __m256i crc256= _mm512_castsi512_si256(crc512);
  crc_out= xor128(_mm256_extracti64x2_epi64(crc256, 1),