[Bug other/10944] alloc_page in ggc-page.c is slow

[Bug other/10944] alloc_page in ggc-page.c is slow

[steven at gcc dot gnu dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944



------- Additional Comments From steven at gcc dot gnu dot org  2003-10-12 20:34 -------
Discussed this with Zack, and the patch "is essentially what [he] had in mind."

Zack suggested there are still a few things that need to be done, however:
1) two-level free page caching:
   - after one collection, put all freepages on the per-order lists;
   - after the next collection, put all thepages that are still on
     the per-order lists on the global list;
   - and after a third collection release the pages still on the global
     list tothe OS.
   This probably takes some fiddling with the page_entry struct.  Also
   we'll have to see how this works in relation with page groups

2) scavenge pages from other per-order lists rather than go back to the
   OS for more.
   (Personally I think this won't be necessary if (1) is implemented properly)

[Bug other/10944] alloc_page in ggc-page.c is slow

[steven at gcc dot gnu dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|2003-06-01 18:58:30         |2003-10-13 14:37:14
               date|                            |


------- Additional Comments From steven at gcc dot gnu dot org  2003-10-13 14:37 -------
The new patch implements suggestion 1 from last nights list.

Andrew, you identified the loop over free_pages as a time consumer, but now that
it is gone, I get virtually no measurable time improvements (10ths of seconds on
>1min total compilation time) for the test case for PR8361.  Which (not
surprisingly) once again shows that GC is slow because with each call
ggc_collect we need to do more and more marking...

Are you absolutely, positively sure that this loop is a performance hog? 
Perhaps you can try the patch, it bootstrappes (c,objc,c++) and showed no
regressions on i686.  So far I haven't found convincing evidence that it
improves performance significantly except for the most pathological (artificial)
cases.

Sigh, back to waiting for a new GC strategy...

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2003-11-30 08:38 -------
I am going to push this (only because I have not have time to test the patch) and there is 
already a new GC strategy (for 3.4) but does not work with everything.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P1                          |P2
   Last reconfirmed|2003-10-13 14:37:14         |2003-11-30 08:38:20
               date|                            |
   Target Milestone|3.4                         |3.5


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944

[Bug other/10944] alloc_page in ggc-page.c is slow

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2003-12-20 14:10 -------
I have experimented with my patch on three different targets (i686,x86_64, and 
hpux) and have never seen any improvements with it even though it kills the 
supposed hot spot in the code.  I'm marking this as WAITING and I'll close it 
soonish, unless someone can test this on the reporter's platform (ppc-darwin) 
and show that it does help on that target. 
 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2003-12-23 01:10 -------
Steven can you wait like three weeks before closing because I go on vaction tommorrow for about 
two weeks and I cannot test the patch untill after I get back?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944

[Bug other/10944] alloc_page in ggc-page.c is slow

[dhazeghi at yahoo dot com]

------- Additional Comments From dhazeghi at yahoo dot com  2004-01-17 22:59 -------
Andrew, did you get a chance to test the patch for this?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-02-06 01:44 -------
I do not have time to test this, and it now this patch does not matter as Honza's work with the 
compiler's memory and compile time stuff and Steven did not get any benifit from it.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|                            |WONTFIX


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944

[Bug other/10944] alloc_page in ggc-page.c is slow

[zack@gcc.gnu.org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


zack@gcc.gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2003-05-23 21:52:36         |2003-06-01 18:58:30
               date|                            |





------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia@physics.uc.edu]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944



------- Additional Comments From pinskia@physics.uc.edu  2003-05-23 22:29 -------
Subject: Re: [Bug other/10944] alloc_page in ggc-page.c is slow

Yes it is spending 90% of the time in alloc_page in that list traversal.
I used Shikari (part of the CHUD tools) on Mac OS X for profiling the 
complier
10000 samples ever 1ms (10s in total).

I can find out the backtrace if you want it.

Thanks,
Andrew Pinski


On Friday, May 23, 2003, at 17:52 US/Eastern, zack@gcc.gnu.org wrote:
>
> The code in alloc_page is remarkably stupid, but the linked list 
> traversal
> should only occur when trying to allocate an object larger than a 
> single
> page, which is believed to happen almost never.  Can you please find 
> out
> whether it's really spending time in that list, or if some other area 
> is
> the cause of the problem?
>
> I'll take responsibility for this bug; I've been meaning to get us 
> per-order
> freelists for a long time.





------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia@physics.uc.edu]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944



------- Additional Comments From pinskia@physics.uc.edu  2003-05-23 22:40 -------
The backtraces are spread all over so that it is not just one place.



------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[zack@gcc.gnu.org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


zack@gcc.gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned@gcc.gnu.org      |zack@gcc.gnu.org
             Status|ASSIGNED                    |NEW





------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[zack@gcc.gnu.org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


zack@gcc.gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2003-05-23 20:19:39         |2003-05-23 21:52:36
               date|                            |


------- Additional Comments From zack@gcc.gnu.org  2003-05-23 21:52 -------
The code in alloc_page is remarkably stupid, but the linked list traversal
should only occur when trying to allocate an object larger than a single
page, which is believed to happen almost never.  Can you please find out
whether it's really spending time in that list, or if some other area is
the cause of the problem?

I'll take responsibility for this bug; I've been meaning to get us per-order
freelists for a long time.




------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia@physics.uc.edu]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


pinskia@physics.uc.edu changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2003-05-23 20:19:39
               date|                            |





------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug other/10944] alloc_page in ggc-page.c is slow

[pinskia@physics.uc.edu]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10944


pinskia@physics.uc.edu changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
OtherBugsDependingO|                            |8361
              nThis|                            |





------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.