[Bug c/33426] New: Support of #pragma ivdep

[Bug c/33426] New: Support of #pragma ivdep

[gjunk at sapience dot com]

When comparing run time on some (floating point) CPU intensive code, we find
that Intel's ICC (9.1) compiler runs significantly faster (30 - 50%) for some
programs - this seems to stem (large) part from ICC's ability to take advantage
of #pragma ivdep. There may be some other differences of course. Tests were run
using gcc 3.3.3 - re-compiling relevant key code with gcc 4.1.2 did not make
any real change. 

Wondering if there are any plans to support this ?

thanks

gene/


-- 
           Summary: Support of #pragma ivdep
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: gjunk at sapience dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2007-09-13 21:12 -------
Confirmed.  Zdenek, we should now be able to attach this to loop information
and preserve it until data dependence analysis, right?


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu dot
                   |                            |org, rakdver at gcc dot gnu
                   |                            |dot org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-09-13 21:12:49
               date|                            |
            Summary|Support of #pragma ivdep    |Support of #pragma ivdep


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2007-09-13 21:14 -------
What are the exact semantics of ivdep?
Please provide them before asking if we support it.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2007-09-13 21:21 -------
#pragma ivdep

Says the compiler may assume the following loop nest does not have loop carried
data dependenices.

Supported at least by SGI CC and the Intel compiler.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
   Last reconfirmed|2007-09-13 21:12:49         |2007-09-13 21:21:05
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2007-09-13 21:23 -------
http://techpubs.sgi.com/library/tpl/cgi-bin/getdoc.cgi/0650/bks/SGI_Developer/books/Pragmas/sgi_html/ch08.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[rguenth at gcc dot gnu dot org]

------- Comment #5 from rguenth at gcc dot gnu dot org  2007-09-13 21:25 -------
The SGI compiler restricts the pragma to innermost loops.  The Intel compiler
restricts the pragma to affect assumed data-dependencies only, that is, if
data dependence analysis statically can prove a data-dependency this
dependency is _not_ ignored.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|other                       |c


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug c/33426] Support of #pragma ivdep

[rguenth at gcc dot gnu dot org]

------- Comment #6 from rguenth at gcc dot gnu dot org  2007-09-13 21:26 -------
http://www.intel.com/cd/ids/developer/asmo-na/eng/65774.htm?page=4


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug other/33426] Support of #pragma ivdep

[pinskia at gcc dot gnu dot org]

------- Comment #7 from pinskia at gcc dot gnu dot org  2007-09-13 21:27 -------
http://gcc.gnu.org/ml/gcc-patches/2005-03/msg00166.html


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c                           |other


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug other/33426] Support of #pragma ivdep

[gjunk at sapience dot com]

------- Comment #8 from gjunk at sapience dot com  2007-09-13 22:21 -------
*** Bug 33427 has been marked as a duplicate of this bug. ***


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426

[Bug other/33426] Support of #pragma ivdep

[gjunk at sapience dot com]

------- Comment #9 from gjunk at sapience dot com  2007-09-19 21:34 -------
Thanks for feedback - not exactly sure where this leaves things - but if gcc
can deliver the same kind of floating point performance as icc it would be a
really good thing.

The 2005 article - wasn't clear to me what that meant other than this was
considered some time back - I assume this is a 2 step process.

   (1) making the info available (perhaps that is already done - is that what
that means?)

   (2) consuming the info so the optimization engine can do a better job -
which I suspect is work still remaining.

Thank you for all your help.

g/


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33426