| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 8593.1 | | QUARRY::neth | Craig Neth | Mon Jan 27 1997 09:54 | 23 |
| >first an executable is created using -non_shared and -om options producing
>prog, then use pixie to create pixie.prog (executable ). Now execute pixie.prog
>to produce prog.Counts. Finally build prog again with -non_shared -om -WL,
>-om_ireorg_feedback option.
Almost!
Here are the steps:
1. create an executable using -non_shared BUT WITHOUT -om
2. use pixie to create prog.pixie (creates prog.Addrs)
3. execute prog.pixie to produce prog.Counts
4. Build program with -non_shared -om -WL,-om_ireorg_feedback,prog opt.
In V4.0, you can use the new -prof_gen option to handle steps 1&2, and the
new -prof_use_om_feedback flag to handle step 4.
You got the error message because you used -om on your 'first executable' -
since OM reorganizes the program, the program that you collected the
feedback data was reorganized, and that fouled up the mapping of the feedback
data to the original program that was being attempted with om_ireorg_feedback.
Craig
|
| 8593.2 | few more ? | HGOVC::XFMV01::ANANDV | | Thu Jan 30 1997 01:55 | 18 |
| hi,
thanx a lot.
I have few more questions regard to that.
When I used this flag on a program that did sorting of around
2000 integers, the performance improved greatly.
I have two questions
1) On what all types of programs should i expect performance
improvements ?
2) Are there any side effects of this flag i.e can it degrade
the performance in any case or effect the accuracy of data.
regds
anand
|
| 8593.3 | right -- No, Left -- NO, RIGHT -- NO, LEFT!!! | HYDRA::NEWMAN | Chuck Newman, 508/467-5499 (DTN 297), MRO1-3/F26 | Thu Jan 30 1997 09:39 | 12 |
| I too was intrigued by this, but unfortunately can't try it.
Several Digital -- oops, DIGITAL -- libraries aren't provided in
.a form but only in .so form, precluding linking -non_shared.
Link call_shared to save space
no, link non_shared because it runs faster
no, link call_shared because it runs the same speed
no, link non_shared because you can then do post link optimizations
no, link call_shared because you don't have all the .a files you need
Sigh...
|
| 8593.4 | More answers | SMURF::NETH | Craig Neth | Thu Jan 30 1997 11:17 | 28 |
| re: .3:
We're working on making -om work on -call_shared programs.
re: .2:
>1) On what all types of programs should i expect performance
> improvements ?
Largish, cpu-intensive programs get the most improvements.
> 2) Are there any side effects of this flag i.e can it degrade
> the performance in any case or effect the accuracy of data.
If -om breaks your program (i.e. causes bad answers) that's usually an
OM bug and we want to know about it.
Sometimes, OM will make the program run slower, especially if you've
already tuned the heck out of it with compiler switches and KAP.
For example, we have some Fortran programs we're looking at that
OM manages to slow down, because the compilers had already gotten
just about everything 'right', and OM's rearrangements actually hurt.
About 8-10% of the TPC-C numbers we're always bragging about are
the result of -om technology - OM really helps applications like
database servers.
Craig
|
| 8593.5 | more flags needed ? | HGOVC::XFMV01::ANANDV | | Mon Feb 03 1997 02:25 | 20 |
| hi,
I just ran the sorting program of 2000 integers.
with -om flag, without any flag and with O4 flag.
I expected the fastest time would be of the one using
-om flage, but the program with -O4 flag runs faster
(almost twice).
Is this expected behaviour ?
Should I include some more flags in the one where
I use -om flag.
regds
anand
|