llvm dev - Mar 2011 - [LLVMdev] metadata to inform the optimizers that some code should stay unchanged

Hello LLVM,

I am working on some passes that perform code transformations. Since I am
interested in performance, I apply the O3 passes, right after my pass.
However, the optimization passes modify the code inserted by my pass in an
undesirable way. As far I know, there is no way to prevent the optimizers
from optimizing some regions of code. So what I intend to do is to attach
metadata to the instructions contained in basicblocks that I want to remain
unchanged, and to modify some of the optimization passes to be aware of the
metadata. For now, I am interested only in passes that would affect the
control flow graph, so for a start I modify the Simplify CFG pass and the
jump threading pass, but I will check which other passes might duplicate the
code, merge blocks etc.

Do you think the performance will drop significantly if some regions of code
are not optimized ? And do you consider this modification would be of any
use to the community?

Thank you.
Alexandra

-- 
View this message in context:
http://old.nabble.com/metadata-to-inform-the-optimizers-that-some-code-should-stay-unchanged-tp31039471p31039471.html
Sent from the LLVM - Dev mailing list archive at Nabble.com.

On Mar 1, 2011, at 2:10 AM, Xinfinity wrote:
> 
> 
> Hello LLVM,
> 
> I am working on some passes that perform code transformations. Since I am
> interested in performance, I apply the O3 passes, right after my pass.
> However, the optimization passes modify the code inserted by my pass in an
> undesirable way. As far I know, there is no way to prevent the optimizers
> from optimizing some regions of code. So what I intend to do is to attach
> metadata to the instructions contained in basicblocks that I want to remain
> unchanged, and to modify some of the optimization passes to be aware of the
> metadata.
IMO, this is not the right way to go.
> For now, I am interested only in passes that would affect the
> control flow graph, so for a start I modify the Simplify CFG pass and the
> jump threading pass, but I will check which other passes might duplicate
the
> code, merge blocks etc.
> 
> Do you think the performance will drop significantly if some regions of
code
> are not optimized ?
> And do you consider this modification would be of any
> use to the community?
> 
> Thank you.
> Alexandra
> 
> -- 
> View this message in context:
http://old.nabble.com/metadata-to-inform-the-optimizers-that-some-code-should-stay-unchanged-tp31039471p31039471.html
> Sent from the LLVM - Dev mailing list archive at Nabble.com.
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Xinfinity wrote:
>
>
> Hello LLVM,
>
> I am working on some passes that perform code transformations. Since I am
> interested in performance, I apply the O3 passes, right after my pass.
> However, the optimization passes modify the code inserted by my pass in an
> undesirable way. As far I know, there is no way to prevent the optimizers
> from optimizing some regions of code. So what I intend to do is to attach
> metadata to the instructions contained in basicblocks that I want to remain
> unchanged, and to modify some of the optimization passes to be aware of the
> metadata.
Please don't. Metadata is the wrong tool for anything where the metadata 
can't be discarded and correctness maintained.

Is there any reason you can't run opt -O3 first and your pass second? 
What is your pass actually doing and why does running the optimizations 
break it?

Nick

  For now, I am interested only in passes that would affect
the
> control flow graph, so for a start I modify the Simplify CFG pass and the
> jump threading pass, but I will check which other passes might duplicate
the
> code, merge blocks etc.
>
> Do you think the performance will drop significantly if some regions of
code
> are not optimized ? And do you consider this modification would be of any
> use to the community?
>
> Thank you.
> Alexandra
>

Nick Lewycky wrote:
> 
> Xinfinity wrote:
>>
>>
>> Hello LLVM,
>>
>> I am working on some passes that perform code transformations. Since I
am
>> interested in performance, I apply the O3 passes, right after my pass.
>> However, the optimization passes modify the code inserted by my pass in
>> an
>> undesirable way. As far I know, there is no way to prevent the
optimizers
>> from optimizing some regions of code. So what I intend to do is to
attach
>> metadata to the instructions contained in basicblocks that I want to
>> remain
>> unchanged, and to modify some of the optimization passes to be aware of
>> the
>> metadata.
> 
> Please don't. Metadata is the wrong tool for anything where the
metadata
> can't be discarded and correctness maintained.
> 
> Is there any reason you can't run opt -O3 first and your pass second? 
> What is your pass actually doing and why does running the optimizations 
> break it?
> 
> Nick
> 
> 
You are right, changing the optimizers is not a good idea, I will modify my
pass to make it O3 friendly. 
If I run opt -O3 first and my pas after, there is a significant drop of
performance. 
I use my pass to create multiple versions of regions of code and to insert a
selection mechanism that 
communicates with a runtime system. Callbacks to the runtime system must be
patched and they are inserted as inline asm to have a fixed structure and
size. O3 changes the position of the inline asm or duplicates them.
But it is better to work more on my pass, than to change all optimization
passes in LLVM. 

Thanks.

Alexandra

-- 
View this message in context:
http://old.nabble.com/metadata-to-inform-the-optimizers-that-some-code-should-stay-unchanged-tp31039471p31066087.html
Sent from the LLVM - Dev mailing list archive at Nabble.com.