Re: [Open64-devel] Code Review request for fix bug974[CG]
Brought to you by:
ributzka,
suneeljain
Menu
▾
▴
Re: [Open64-devel] Code Review request for fix bug974[CG]
|
From: Nelson H. F. B. <be...@ma...> - 2012-05-05 17:48:46
|
Sun Chan <sun...@gm...> writes today:
>> IA64 alignment rule is no diff from x86, as far as I know.
No, x86 allows quite lax alignment (multiples of 2 or 4), but some
models may work more efficiently if alignment is to multiples of 8.
On IA-64, failure to align to a proper multiple produces a run-time
fixup that degrades performance, and also produces syslog messages,
like this:
--------------------- Kernel Begin ------------------------
WARNING: Unaligned Errors in these executables
emacs-x : 3 Time(s)
---------------------- Kernel End -------------------------
Recently, in the TeX Live 2012 build project, I found that an address
of an int was passed to an X11 library routine that picked it up as an
address of a long, producing run-time messages on IA-64 like this:
xdvi-xaw(5977): unaligned access to 0x60000fffffc7eb14, ip=0x20000000003cf500
We changed the type of the variable in the calling routine from int to
long int, and that eliminated the run-time complaints.
Here is a snippet from an IBM document ``Migrating C and C++
Applications to AIX 5L on IA-64 Partners in Development September,
2000'':
>> ...
>> AIX on PowerPC AIX 5L on IA-64 on Itanium
>> C and C++ 32-bit Source 64-bit Source ILP32 LP64
>> Data Type Size/Alignment Size/Alignment Size/Alignment Size/Alignment
>> char 1/1 1/1 1/1 1/1
>> short 2/2 2/2 2/2 2/2
>> int 4/4 4/4 4/4 4/4
>> long 4/4 8/8 4/4 8/8
>> longlong 8/8 8/8 8/8 8/8
>> pointer 4/4 8/8 4/4 8/8
>> float 4/4 4/4 4/4 4/4
>> double 8/4 8/4 8/8 8/8
>> longdouble 16/16 16/16 16/16 16/16
>>
>> *Note that long long has a size of 8 bytes and is aligned on 8 byte boundaries in both 32-bit and 64-bit modes.
>> Also note that integrals and floats are aligned according on their natural boundary (same as their size) which is
>> different than the rules for IA-32. Applications requiring the IA-32 alignment rules can use the
>> "#pragma align=ia64unix386". However the size of long double will remain 16 bytes even when the pragma is
>> used.
>> ...
Here is another snippet from the ``Intel IA-64 Architecture Software
Developer's Manual, Volume 1: IA-64 Application Architecture, January
2000, Order Number: 245317-001'':
>> ...
>> 3.2.2 Addressable Units and Alignment
>>
>> Memory can be addressed in units of 1, 2, 4, 8, 10 and 16 bytes. It is
>> recommended that all addressable units be stored on their naturally
>> aligned boundaries. Hardware and/or operating system software may have
>> support for unaligned accesses, possibly with some performance
>> cost. 10-byte floating-point values should be stored on 16-byte
>> aligned boundaries. Bits within larger units are always numbered from
>> 0 starting with the least-significant bit. Quantities loaded from
>> memory to general registers are always placed in the least-significant
>> portion of the register (loaded values are placed right justified in
>> the target general register). Instruction bundles (3 IA-64
>> instructions per bundle) are 16-byte units that are always aligned on
>> 16-byte boundaries.
>> ...
-------------------------------------------------------------------------------
- Nelson H. F. Beebe Tel: +1 801 581 5254 -
- University of Utah FAX: +1 801 581 4148 -
- Department of Mathematics, 110 LCB Internet e-mail: be...@ma... -
- 155 S 1400 E RM 233 be...@ac... be...@co... -
- Salt Lake City, UT 84112-0090, USA URL: http://www.math.utah.edu/~beebe/ -
-------------------------------------------------------------------------------
|