Hello David!
I have recently started acquaintance with PGO in LLVM/clang and found
your e-mail thread:
http://lists.llvm.org/pipermail/llvm-dev/2016-May/099395.html . Here you
posted a nice list of optimizations that use profiling and of those
which could be using but don't. However that thread is about 2 years
old. Could you please kindly let me know if there were any significant
changes in this area since that time?
Thanks in advance!
Hello David!
I have recently started acquaintance with PGO in LLVM/clang and found
your e-mail thread:
[llvm-dev] About Clang llvm PGO . Here you
posted a nice list of optimizations that use profiling and of those
which could be using but don't. However that thread is about 2 years
old. Could you please kindly let me know if there were any significant
changes in this area since that time?
Yes, there were quite some changes since then. Here are some of the new
features:
* LLVM IR based PGO -- this is designed to maximize program performance.
The option to turn it on is -fprofile-generate/-fprofile-use
* value profiling support in PGO -- currently support indirect call target
profiling and memcpy/memset size profiling and optimizations
* Profile data is made available for inliner to use (enabled only for the
new pass manager: -fexperimental-new-pass-manager)
* Profile aware LICM is available -- implemented via a profile driven code
sinking pass
* Partial inlining is made profile aware; Graham Yu also added support for
multiple region function outlining (with PGO)
* BB layout heuristics are tuned with PGO
* hotness driven function layout optimization
There are pending work in the following area:
* profile aware loop vectorization, etc
* control heigh reduction optimization (Hiroshi is working on this)
ThinLTO also works well with PGO.
Hope this helps.
David
* What I can tell you is that there are many missing ones (that can benefit
*from profile): such as profile aware LICM (patch pending), speculative PRE,
loop unrolling, loop peeling, auto vectorization, inlining, function
splitting, function layout, function outlinling, profile driven size
optimization, induction variable optimization/strength reduction, stringOp
specialization/optimization/inlining, switch peeling/lowering etc. The
biggest profile user today include ralloc, BB layout, ifcvt, shrinkwrapping
etc, but there should be rooms to be improvement there too.
Hello David, thanks for detailed response!
Do you have any tests that you use to measure the PGO effectiveness? I have tested clang version 6.0 with the same sample that Jie Chen used in 2016 and actually both frontend-based PGO and IR-based make code run slower, see the average time:
clang++ -O3: 3.15 sec
clang++ -O3 and -fprofile-instr-use: 3.160 sec
clang++ -O3 and -fprofile-use: 3.180 sec
g++ (7.3.0) -O3: 3.640 sec
g++ (7.3.0) -O3 and -fprofile-use: 2.92 sec
Do you have any idea what can be wrong? Maybe there are some recommendations in which cases one should use PGO with clang and when it is better not to do it?
Thanks!
Victor, thanks for the experiment.
My suspicion is it is due to the remaining issues with block layout – especially with loop rotation (with PGO). Another problem is that tail dup is not happening after loop rotation which can limit the effectiveness of loop rotation.
I tried the internal option -mllvm -force-precise-rotation-cost and there is about 10% speedup with -fprofile-use. This option turns on more precise cost model when computing rotation strategy but it is not turned on by default.
+carrot who is working on this area.
thanks,
David
David, could you please clarify on which code did you gain 10% improvement? I have run numerous tests with and w/o this option and it looks like it has no effect on performance (I am talking of the old 2016 sample to be concrete). Maybe we could investigate it together? Just tell me where to start?
Victor, please file a bug tracking the issue. We can put relevant information there including test cases used in the experiment etc.
thanks,
David
Please let me know if I can help somehow.
Hello David and all involved =)
Hello David and all involved =)
> ThinLTO also works well with PGO.
Could you please let me know if there are any problems which prevent
using PGO with FullLTO?
There should not be any major problems -- except that the new pass manager
is needed for the inliner to be profile aware.
David