FreeTDS Development Group

["Peter C. Norton" <spacey-freetds.org AT ssr.com>]

Group, 

I'm no good at alignment issues, so I'm looking for help here.  I've
got a 64-bit build of freetds-0.83.dev.20100902:

nortonp@ssqa1 11:54 /tmp/freetds/src/odbc/unittests $ ./array_out
odbctest
--------

connection parameters:
server:   'SYBQA1'
user:     'sa'
password: '????'
database: 'tempdb'
use tempdb
CREATE TABLE #odbc_test(i INT, t TEXT)
INSERT INTO #odbc_test(i, t) VALUES(1, 'arow number 0')
INSERT INTO #odbc_test(i, t) VALUES(2, 'brow number 13')
INSERT INTO #odbc_test(i, t) VALUES(3, 'crow number 26')
INSERT INTO #odbc_test(i, t) VALUES(4, 'drow number 39')
INSERT INTO #odbc_test(i, t) VALUES(5, 'erow number 52')
INSERT INTO #odbc_test(i, t) VALUES(6, 'frow number 65')
INSERT INTO #odbc_test(i, t) VALUES(7, 'grow number 78')
INSERT INTO #odbc_test(i, t) VALUES(8, 'hrow number 91')
INSERT INTO #odbc_test(i, t) VALUES(9, 'irow number 104')
INSERT INTO #odbc_test(i, t) VALUES(10, 'jrow number 117')
test line 149
test line 153
test line 159
Bus Error (core dumped)

freetds/src/odbc/unittests $ /proj/lang/SunOS.5.sparc/studio-12.1/bin/dbx 
array_out core
For information about new features see `help changes'
To remove this message, put `dbxenv suppress_startup_message 7.8' in your 
.dbxrc
Reading array_out
core file header read successfully
[etc...]
t@1 (l@1) program terminated by signal BUS (invalid address alignment)
Current function is query_test
   69                   ID_LENS(i) = 0;
(dbx) where
current thread: t@1
=>[1] query_test(expected = 0x100004fd8 "S", expected_status = 0x100005140 
"VVVVVVVVVV"), line 69 in "array_out.c"
  [2] main(argc = 1, argv = 0xffffffff7ffff758), line 160 in "array_out.c"

array_out.c looks like this:
     58 #define REC(f,n) (((char*)f)+rec_size*(n))
     59 #define DESCS(n) (rec ? (SQLCHAR*)REC(descs,n): (descs+(n)*desc_len))
     60 #define IDS(n) *(rec ? (SQLUINTEGER*)REC(ids,n) : &ids[n])
     61 #define ID_LENS(n) *(rec ? (SQLLEN*)REC(id_lens,n) : &id_lens[n])
     62 #define DESC_LENS(n) *(rec ? (SQLLEN*)REC(desc_lens,n) : 
&desc_lens[n])
     63
     64         processed = ARRAY_SIZE + 1;
     65         for (i = 0; i < ARRAY_SIZE; i++) {
     66                 statuses[i] = SQL_ROW_UPDATED;
     67                 IDS(i) = i * 132;
     68                 sprintf((char *) DESCS(i), "aaa");
>>>  69                 ID_LENS(i) = 0;
     70                 DESC_LENS(i) = -i;
     71         }
     72

The test has changed a bunch since 0.82, such that it doesn't compile
in 0.83 anymore, and I can't exactly reproduce our old 0.82 build
anymore (lots of other moving parts have moved on)...  I guess that
id_lens, being declared as follows:

        SQLLEN *id_lens, *desc_lens;

decomposes into:

typedef int             SQLINTEGER;

If I understand correctly, on 64-bit platforms the word size is 8
bytes, and empirically an int is 4 bytes and so is a (char *), so this
test is bound to cause an unaligned access on the second access, where
it's trying to access (id_lens[i] << 16) (unless I've got the
endianness backwards, please forgive me if so).

What's the right fix on this test?  Is there a macro to check for
alignment and cleverly return the desired byte, something fancier than
the code here: http://forums.sun.com/thread.jspa?threadID=5341322 ?

Thanks,

-Peter