clang x86_64 test failure - main.select - converting strings of numbers to compare against large bigint - loss of precision

As discovered on travis, clang test (simplified as below fail).

Simplified test

create table  t1 (  A_Id bigint(20) NOT NULL default '0',  A_UpdateBy char(10) NOT NULL default '',  A_UpdateDate bigint(20) NOT NULL default '0',  A_UpdateSerial int(11) NOT NULL default '0',  other_types bigint(20) NOT NULL default '0',  wss_type bigint(20) NOT NULL default '0');

INSERT INTO t1 VALUES (102935998719055004,'brade',1029359987,2,102935229116544068,102935229216544093);

select wss_type from t1 where wss_type ='102935229216544093';

This returns no results.

As discovered with cvicentiu, the error manifests in stored_field_cmp_to_item.

item_cmp_type returns a common factor of REAL between string and INT types and therefore adopts a double based equality test at the end of the stored_field_cmp_to_item function.

Unfortunately the results for the conversion from the ulonglong in item->val_real() and field->val_real() differ by ~80 resulting in no equally and the SQL statement returning the wrong result.

Options include:
a) pushing an error when users quote numbers in queries - probably too common
b) fix this such that where this string is meticulously parsed in my_strntoull10rnd_8bit, b1) the ability to convert this to an integer is determined. Special cases? exponents detected (these are tested that it still falls < CUTOFF so no action taken b2) large -ve exponents are rounded to 0. b3) Decimals are rounded (handle to DECIMAL type).
c) fudge the test case until it pops up again