llvm dev - Feb 2017 - Make MachineLICM PostRA (or PreRa) work only on the inner loop for a loop nest

Hello.
     I am curious if I can make the MachineLICM transformation work only on the
inner loop
when it encounters a loop nest in the program.

     I've started examining the source code of the pass - 
http://llvm.org/docs/doxygen/html/MachineLICM_8cpp_source.html . I found
comments like
<<// If this is done before regalloc, only visit outer-most
preheader-sporting // loops>>.
     I'll come back if I find a solution.

   Thank you very much,
     Alex

Hello.
     I found a solution to this problem I asked about - make MachineLICM pre-RA
(or
post-RA, in principle) work only on the inner loop for a loop nest. Basically, I
had to
change the code of MachineLICM.cpp, by adding in method
       MachineLICM::HoistOutOfLoop(MachineDomTreeNode *HeaderN)
     in the loop:
       for (MachineDomTreeNode *Node : Scopes) {
         MachineBasicBlock *MBB = Node->getBlock();
         ...
       }
   some extra code that checks if MBB has name "vector.body", which is
the inner loop that
I care about, and in such case I set Preheader variable defined there to be the 
predecessor of vector.body, normally called "vector.ph".
     If you want I can provide the exact code patch.

     Since these are the only cases I care about, namely loops (or loop nests)
resulted
from the LoopVectorize LLVM module, this small patch serves well.

     So, I could not employ the command-line arguments of llc and had to adapt
the code of
the MachineLICM pass.

   Best regards,
     Alex


On 2/22/2017 9:05 AM, Alex Susu wrote:
>   Hello.
>     I am curious if I can make the MachineLICM transformation work only on
the inner loop
> when it encounters a loop nest in the program.
>
>     I've started examining the source code of the pass -
> http://llvm.org/docs/doxygen/html/MachineLICM_8cpp_source.html . I found
comments like
> <<// If this is done before regalloc, only visit outer-most
preheader-sporting // loops>>.
>     I'll come back if I find a solution.
>
>   Thank you very much,
>     Alex
>
>
>