[llvm-dev] Porting OptBisect to New Pass Manager

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

[llvm-dev] Porting OptBisect to New Pass Manager

Bruce Hoult via llvm-dev
Hi Chandler,

I am now working on a bisecting tool to find mis-optimization on LLVM. I found OptBisect a very useful option and hope to make it work on the new pass manager. I have several questions about it.

1. Any plans to apply codegen stage with new pass manager?
    IIUC, new pass manager only works for opt stage. However the OptBisect option tries to accumulate pass counts through opt stage and codegen stage.
    Porting it to new pass manager means that, for now opt stage accumulation will use new pass manager and llc stage will still use legacy pass mechanism.
    May I ask if there is any plan to apply new pass manager for llc stage? If so, the OptBisect porting will also need to expend to codegen stage in the future.

2.  Do we want optnone support for new pass manager?
    I see OptBisect for each pass was first introduced by modifying optnone check. However new pass manager does not support it IIUC.
    We can either follow the legacy structure of skipFunction()/Module(), put skippability check in each single pass for new pass manager.
    Or since there is no optnone support, we now may also create label on each pass to determine it is skippable or not, and then do the check at pass manager level.

3. No region/basicblock/callgraphscc pass support for new pass manager?
    In legacy pass manager, OptBisect will be checked by skip() function on certain types of passes. like skipRegion(), skipBasicBlocks().
    But I did not see implementations of the passes I mentioned above in new passes. So any plan to have them any more?

4. Is PassInfoMixin the right place to put skipPass() function?
The OptBisect works like this: At the beginning of each pass runs, it will call a skipFunction()/Module()/... function under pass.h to check if -opt-bisect is set and how much is counter accumulating.
Seems all passes come from the structure PassInfoMixin API, and I see name() function in it which is useful of OptBisect.
Do you think it is the right place to put the skipPass() function, similar to where skipFunction()/Module() located in legacy pass manager?
    
I currently have a draft on OptBisect porting implementation, and I am willing to have people review it if you like.

Thanks for taking your time on those questions.

If anyone is also make similar effort on it, or have any suggestions of implementation, please let me know and I really appreciate :)

Best,
Zhizhou

_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Reply | Threaded
Open this post in threaded view
|

Re: [llvm-dev] Porting OptBisect to New Pass Manager

Bruce Hoult via llvm-dev


On 06/07/2018 12:26 AM, Zhizhou Yang via llvm-dev wrote:

> Hi Chandler,
>
> I am now working on a bisecting tool to find mis-optimization on LLVM.
> I found OptBisect a very useful option and hope to make it work on the
> new pass manager. I have several questions about it.
>
> 1. Any plans to apply codegen stage with new pass manager?
>     IIUC, new pass manager only works for opt stage. However the
> OptBisect option tries to accumulate pass counts through opt stage and
> codegen stage.
>   Porting it to new pass manager means that, for now opt stage
> accumulation will use new pass manager and llc stage will still use
> legacy pass mechanism.
>   May I ask if there is any plan to apply new pass manager for llc
> stage? If so, the OptBisect porting will also need to expend to
> codegen stage in the future.
>
> 2.  Do we want optnone support for new pass manager?
>     I see OptBisect for each pass was first introduced by modifying
> optnone check. However new pass manager does not support it IIUC.
>     We can either follow the legacy structure of
> skipFunction()/Module(), put skippability check in each single pass
> for new pass manager.
>     Or since there is no optnone support, we now may also create label
> on each pass to determine it is skippable or not, and then do the
> check at pass manager level.
>
> 3. No region/basicblock/callgraphscc pass support for new pass manager?
There is CallGraphSCC pass support in new pass manager.
See include/llvm/Analysis/CGSCCPassManager.h

There are no plans to implement region/BB passes afaik.

>     In legacy pass manager, OptBisect will be checked by skip()
> function on certain types of passes. like skipRegion(), skipBasicBlocks().
>     But I did not see implementations of the passes I mentioned above
> in new passes. So any plan to have them any more?
>
> 4. Is PassInfoMixin the right place to put skipPass() function?
> The OptBisect works like this: At the beginning of each pass runs, it
> will call a skipFunction()/Module()/... function under pass.h to check
> if -opt-bisect is set and how much is counter accumulating.
> Seems all passes come from the structure PassInfoMixin API, and I see
> name() function in it which is useful of OptBisect.
> Do you think it is the right place to put the skipPass() function,
> similar to where skipFunction()/Module() located in legacy pass manager?
PassInfoMixin was not intended to be a base for hierarchy of Passes, it
is merely a boilerplate helper.

> I currently have a draft on OptBisect porting implementation, and I am
> willing to have people review it if you like.
>
> Thanks for taking your time on those questions.
>
> If anyone is also make similar effort on it, or have any suggestions
> of implementation, please let me know and I really appreciate :)
Just (finally! :-/ ) have sent out an RFC exactly on this topic:
   "RFC: Pass Execution Instrumentation interface"
It is currently work in progress and rather generic, so feel free to
comment/put more detail there.

regards,
   Fedor.

> Best,
> Zhizhou
>
>
> _______________________________________________
> LLVM Developers mailing list
> [hidden email]
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Reply | Threaded
Open this post in threaded view
|

Re: [llvm-dev] Porting OptBisect to New Pass Manager

Bruce Hoult via llvm-dev
Thanks for bringing this RFC up.

It seems that it will be exactly what we want for opt-bisect in new PM.

I will comment there.

On Wed, Jun 6, 2018 at 5:36 PM Fedor Sergeev via llvm-dev <[hidden email]> wrote:


On 06/07/2018 12:26 AM, Zhizhou Yang via llvm-dev wrote:
> Hi Chandler,
>
> I am now working on a bisecting tool to find mis-optimization on LLVM.
> I found OptBisect a very useful option and hope to make it work on the
> new pass manager. I have several questions about it.
>
> 1. Any plans to apply codegen stage with new pass manager?
>     IIUC, new pass manager only works for opt stage. However the
> OptBisect option tries to accumulate pass counts through opt stage and
> codegen stage.
>   Porting it to new pass manager means that, for now opt stage
> accumulation will use new pass manager and llc stage will still use
> legacy pass mechanism.
>   May I ask if there is any plan to apply new pass manager for llc
> stage? If so, the OptBisect porting will also need to expend to
> codegen stage in the future.
>
> 2.  Do we want optnone support for new pass manager?
>     I see OptBisect for each pass was first introduced by modifying
> optnone check. However new pass manager does not support it IIUC.
>     We can either follow the legacy structure of
> skipFunction()/Module(), put skippability check in each single pass
> for new pass manager.
>     Or since there is no optnone support, we now may also create label
> on each pass to determine it is skippable or not, and then do the
> check at pass manager level.
>
> 3. No region/basicblock/callgraphscc pass support for new pass manager?
There is CallGraphSCC pass support in new pass manager.
See include/llvm/Analysis/CGSCCPassManager.h

There are no plans to implement region/BB passes afaik.

>     In legacy pass manager, OptBisect will be checked by skip()
> function on certain types of passes. like skipRegion(), skipBasicBlocks().
>     But I did not see implementations of the passes I mentioned above
> in new passes. So any plan to have them any more?
>
> 4. Is PassInfoMixin the right place to put skipPass() function?
> The OptBisect works like this: At the beginning of each pass runs, it
> will call a skipFunction()/Module()/... function under pass.h to check
> if -opt-bisect is set and how much is counter accumulating.
> Seems all passes come from the structure PassInfoMixin API, and I see
> name() function in it which is useful of OptBisect.
> Do you think it is the right place to put the skipPass() function,
> similar to where skipFunction()/Module() located in legacy pass manager?
PassInfoMixin was not intended to be a base for hierarchy of Passes, it
is merely a boilerplate helper.

> I currently have a draft on OptBisect porting implementation, and I am
> willing to have people review it if you like.
>
> Thanks for taking your time on those questions.
>
> If anyone is also make similar effort on it, or have any suggestions
> of implementation, please let me know and I really appreciate :)
Just (finally! :-/ ) have sent out an RFC exactly on this topic:
   "RFC: Pass Execution Instrumentation interface"
It is currently work in progress and rather generic, so feel free to
comment/put more detail there.

regards,
   Fedor.

> Best,
> Zhizhou
>
>
> _______________________________________________
> LLVM Developers mailing list
> [hidden email]
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev