FIND_IN_SET returns 0 when needle is longer in bytes than haystack token but equal under collation

Summary

Wrong result: FIND_IN_SET returns 0 when needle is longer in bytes than haystack token but equal under collation

Description

Under utf8mb4 accent-insensitive collations (e.g. utf8mb4_unicode_ci, utf8mb4_uca1400_ai_ci, utf8mb4_general_ci), 'ắ' = 'a' is true, but:

FIND_IN_SET('ắ', 'a')  -- returns 0 (wrong; expect 1)

Root cause in Item_func_find_in_set::val_int() (sql/item_func.cc):

if ((int) (buffer->length() - find->length()) >= 0)

{

  // ... strnncoll per comma-separated token ...

}

return 0;

If the entire list string is shorter in bytes than the needle, the function returns 0 without calling strnncoll. That is incorrect when the collation equates a multi-byte character to a shorter string.

FIND_IN_SET('ắ', 'a,b')  -- returns 1 (correct)

How to repeat

SET NAMES utf8mb4 COLLATE utf8mb4_unicode_ci;

SELECT 'ắ' = 'a' AS eq;                 -- 1

SELECT FIND_IN_SET('ắ', 'a') AS fis1;   -- Actual: 0; Expected: 1

SELECT FIND_IN_SET('ắ', 'a,b,c') AS fis2; -- Actual: 1; Expected: 1

SELECT FIND_IN_SET('a', 'ắ') AS fis3;   -- Actual: 1; Expected: 1

-- Also wrong on utf8mb4_general_ci / utf8mb4_uca1400_ai_ci with the same shape

SET NAMES utf8mb4 COLLATE utf8mb4_general_ci;

SELECT 'ắ' = 'a', FIND_IN_SET('ắ', 'a');  -- 1, 0

Observed on: 12.3.2-MariaDB

Actual result

FIND_IN_SET('ắ', 'a') = 0 while 'ắ' = 'a' is true.

Expected result

FIND_IN_SET('ắ', 'a') = 1 (needle equals the sole list element under the connection collation).