PS-10122 (5.7) Optimize space skipping with neon

include/m_string.h

@@ -27,6 +27,9 @@
 #include "my_global.h"                          /* HAVE_* */
 
 #include <string.h>
+#if defined(__aarch64__)
+#include <arm_neon.h>
+#endif
 
 #define bfill please_use_memset_rather_than_bfill
 #ifdef bzero
@@ -340,6 +343,38 @@ static inline const uchar *skip_trailing_space(const uchar *ptr,size_t len)
     end--;
   return (end);
 }
+#elif defined(__aarch64__)
+static inline const uchar *skip_trailing_space(const uchar *ptr, size_t len)
+{
+  const uchar *end = ptr + len;
+  const size_t neno_vector_length = 16;
+  const uint8x16_t space_vec = vdupq_n_u8(0x20);
+  const uchar *current_end = end - neno_vector_length;
+  size_t vector_size = len >> 4;
+
+  while (vector_size > 0) {
+    uint8x16_t data_vec = vld1q_u8(current_end);
+    uint8x16_t result = vceqq_u8(data_vec, space_vec);
+
+    if (vminvq_u8(result) == 0) {
+      break;
+    }
+    vector_size--;
+    end -= neno_vector_length;
+    current_end -= neno_vector_length;
+  }
+
+  while (end - ptr >= 8) {
+    if (uint8korr(end-8) != 0x2020202020202020ULL) {
+      break;
+    }
+    end-= 8;
+  }
+  while (end > ptr && end[-1] == 0x20) {
+    end--;
+  }
+  return (end);
+}
 #else
 /*
   Reads 8 bytes at a time, ignoring alignment.

storage/innobase/row/row0mysql.cc

@@ -1195,10 +1195,7 @@ row_mysql_store_col_in_innobase_format(
 				break;
 			case 1:
 				/* space=0x20 */
-				while (col_len > 0
-				       && ptr[col_len - 1] == 0x20) {
-					col_len--;
-				}
+				col_len = skip_trailing_space(ptr, col_len) - ptr;
 			}
 		}
 	} else if (comp && type == DATA_MYSQL
@@ -1234,9 +1231,7 @@ row_mysql_store_col_in_innobase_format(
 		n_chars = dtype_get_len(dtype) / dtype_get_mbmaxlen(dtype);
 
 		/* Strip space padding. */
-		while (col_len > n_chars && ptr[col_len - 1] == 0x20) {
-			col_len--;
-		}
+		col_len = skip_trailing_space(ptr + n_chars, col_len - n_chars) - ptr;
 	} else if (!row_format_col) {
 		/* if mysql data is from a MySQL key value
 		since the length is always stored in 2 bytes,

strings/ctype-mb.c

@@ -28,6 +28,9 @@
 #include <my_global.h>
 #include "m_ctype.h"
 #include "m_string.h"
+#if defined(__aarch64__)
+#include <arm_neon.h>
+#endif
 
 
 size_t my_caseup_str_mb(const CHARSET_INFO *cs, char *str)
@@ -383,6 +386,80 @@ int my_wildcmp_mb(const CHARSET_INFO *cs,
 }
 
 
+#if defined(__aarch64__)
+size_t my_numchars_mb(const CHARSET_INFO *cs MY_ATTRIBUTE((unused)),
+		      const char *pos, const char *end)
+{
+  size_t count = 0;
+
+  if ((cs->state & MY_CS_UNICODE) && !(cs->state & MY_CS_NONASCII))
+  {
+    const uint8x16_t max_ascii = vdupq_n_u8(0x7F);
+    const size_t neon_vector_length = 16;
+
+    while (pos + neon_vector_length <= end)
+    {
+      uint8x16_t vec_src = vld1q_u8((const unsigned char *)pos);
+      // check if exists non-ascii character
+      // ascii -> non zero, non-ascii -> zero
+      uint8x16_t result = vcleq_u8(vec_src, max_ascii);
+      if (vminvq_u8(result) == 0)
+      {
+        // has non-ascii character
+        break;
+      }
+
+      count += neon_vector_length;
+      pos += neon_vector_length;
+    }
+  }
+
+  while (pos < end)
+  {
+    uint mb_len;
+    pos += (mb_len = my_ismbchar(cs, pos, end)) ? mb_len : 1;
+    count++;
+  }
+  return count;
+}
+
+
+size_t my_charpos_mb(const CHARSET_INFO *cs MY_ATTRIBUTE((unused)),
+		     const char *pos, const char *end, size_t length)
+{
+  const char *start = pos;
+
+  if ((cs->state & MY_CS_UNICODE) && !(cs->state & MY_CS_NONASCII))
+  {
+    const uint8x16_t max_ascii = vdupq_n_u8(0x7F);
+    const size_t neon_vector_length = 16;
+
+    while (length >= neon_vector_length && pos + neon_vector_length <= end)
+    {
+      uint8x16_t vec_src = vld1q_u8((const unsigned char *)pos);
+      // check if exists non-ascii character
+      // ascii -> non zero, non-ascii -> zero
+      uint8x16_t result = vcleq_u8(vec_src, max_ascii);
+      if (vminvq_u8(result) == 0)
+      {
+        // has non-ascii character
+        break;
+      }
+
+      length -= neon_vector_length;
+      pos += neon_vector_length;
+    }
+  }
+
+  while (length && pos < end)
+  {
+    uint mb_len;
+    pos += (mb_len = my_ismbchar(cs, pos, end)) ? mb_len : 1;
+    length--;
+  }
+  return (size_t) (length ? end+2-start : pos-start);
+}
+#else
 size_t my_numchars_mb(const CHARSET_INFO *cs MY_ATTRIBUTE((unused)),
 		      const char *pos, const char *end)
 {
@@ -410,6 +487,7 @@ size_t my_charpos_mb(const CHARSET_INFO *cs MY_ATTRIBUTE((unused)),
   }
   return (size_t) (length ? end+2-start : pos-start);
 }
+#endif
 
 
 size_t my_well_formed_len_mb(const CHARSET_INFO *cs, const char *b,