Merge pull request #133 from JunHe77:simd

PiperOrigin-RevId: 393681630

CMakeLists.txt

@@ -178,6 +178,15 @@ int main() {
   return _bzhi_u32(0, 1);
 }" SNAPPY_HAVE_BMI2)
 
+check_cxx_source_compiles("
+#include <arm_neon.h>
+int main() {
+  uint8_t val = 3, dup[8];
+  uint8x16_t v = vld1q_dup_u8(&val);
+  vst1q_u8(dup, v);
+  return 0;
+}" SNAPPY_HAVE_NEON)
+
 include(CheckSymbolExists)
 check_symbol_exists("mmap" "sys/mman.h" HAVE_FUNC_MMAP)
 check_symbol_exists("sysconf" "unistd.h" HAVE_FUNC_SYSCONF)

cmake/config.h.in

@@ -49,6 +49,9 @@
 /* Define to 1 if you target processors with BMI2+ and have <bmi2intrin.h>. */
 #cmakedefine01 SNAPPY_HAVE_BMI2
 
+/* Define to 1 if you target processors with NEON and have <arm_neon.h>. */
+#cmakedefine01 SNAPPY_HAVE_NEON
+
 /* Define to 1 if your processor stores words with the most significant byte
    first (like Motorola and SPARC, unlike Intel and VAX). */
 #cmakedefine01 SNAPPY_IS_BIG_ENDIAN

snappy-internal.h

@@ -33,13 +33,30 @@
 
 #include "snappy-stubs-internal.h"
 
+#if SNAPPY_HAVE_SSSE3
+// Please do not replace with <x86intrin.h> or with headers that assume more
+// advanced SSE versions without checking with all the OWNERS.
+#include <emmintrin.h>
+#include <tmmintrin.h>
+#endif
+
+#if SNAPPY_HAVE_NEON
+#include <arm_neon.h>
+#endif
+
+#if SNAPPY_HAVE_SSSE3 || SNAPPY_HAVE_NEON
+#define SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE 1
+#else
+#define SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE 0
+#endif
+
 namespace snappy {
 namespace internal {
 
 #if SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE
 #if SNAPPY_HAVE_SSSE3
 using V128 = __m128i;
-#else
+#elif SNAPPY_HAVE_NEON
 using V128 = uint8x16_t;
 #endif
 
@@ -72,7 +89,7 @@ inline V128 V128_Shuffle(V128 input, V128 shuffle_mask) {
 
 inline V128 V128_DupChar(char c) { return _mm_set1_epi8(c); }
 
-#else
+#elif SNAPPY_HAVE_NEON
 inline V128 V128_Load(const V128* src) {
   return vld1q_u8(reinterpret_cast<const uint8_t*>(src));
 }