[llvm-dev] Separate LoopVectorize LLVM pass

[Alex Susu via llvm-dev]

   Hello.
     I come back to this thread because I still have issues with my separate LoopVectorize 
(same code, but without relying on PassManagerBuilder and -O3) LLVM pass.
     More exactly, my LoopVectorize pass, when given the following reduction code:
       short SumReduce(short *C, short N) {
         short sum = 0;
         for (short i = 0; i < N; ++i) {
           sum += C[i];
         }
         return sum;
       }
     I get on the standard error of opt:
       LV: Checking a loop in "SumReduce" from test.c:15:3
       LV: Loop hints: force=? width=128 unroll=0
       LV: Found a loop: for.body
       LoopVectorize::canVectorizeInstrs(): PhiScev = {0,+,1}<nuw><nsw><%for.body>
       LoopVectorize::canVectorizeInstrs(): AR = {0,+,1}<nuw><nsw><%for.body>
       LV: Found an induction variable.
       LoopVectorize::canVectorizeInstrs(): PhiScev = %sum.02
       LV: PHI is not a poly recurrence.
       LV: PHI is not a poly recurrence.
       remark: <unknown>:0:0: loop not vectorized: value that could not be identified as 
reduction is used outside the loop
       LV: Found an unidentified PHI.  %sum.02 = phi i32 [ %add, %for.body ], [ 0, 
%for.body.preheader ]
       LV: Can't vectorize the instructions or CFG
       LV: Not vectorizing: Cannot prove legality.
       Exiting LoopVectorize::runOnFunction()

     Does anybody know what pass would be helpful for this issue (e.g. "PHI is not a poly 
recurrence."), besides the passes presented below already.?

   Thank you very much,
     Alex

On 4/17/2017 1:01 PM, Alex Susu wrote:
>   Hi.
>     I managed to find the reason why my own LoopVectorize pass does not vectorize.
>     Actually, the loop-simplify pass was getting executed, but because it did not run the
> passes loop-rotate and licm before it did not generate a proper loop preheader and the
> loops were not valid for vectorization.
>
>     To make my own LoopVectorize pass vectorize I got inspired from the arguments
> displayed on stdout of:
>       opt -debug-pass=Arguments -O3 ...
>     and looked at what other passes are being executed before and after -loop-vectorize .
>
>     So I chose the following:
>       opt -mem2reg -loop-rotate -licm -loop-unswitch -simplifycfg -domtree -basicaa -aa
> -instcombine -loops -loop-simplify \
>           -force-vector-width=128 -load MyLoopVectorize.so -my-loop-vectorize \
>           -loop-simplify -scalar-evolution -aa -loop-accesses -loop-load-elim -basicaa -aa
> -instcombine -scalar-evolution -demanded-bits -slp-vectorizer
>
>
>   Best regards,
>     Alex
>
> On 4/14/2017 11:04 PM, Alex Susu wrote:
>>   Hello.
>>     I am trying to create my own LoopVectorize.cpp pass as a separate pass from the LLVM
>> trunk, as described in http://llvm.org/docs/CMake.html#embedding-llvm-in-your-project. Did
>> anybody try something like this?
>>     I added close to the end of the .cpp file:
>>     /* this line seems to be required - it allows to run this pass
>>        as an embedded pass by giving opt -my-loop-vectorize */
>>     static RegisterPass<LoopVectorize> Z("my-loop-vectorize",
>>                                     "Write comments in source file.");
>>
>>     Note that I did NOT register my new pass in
>> llvm/lib/Transforms/IPO/PassManagerBuilder.cpp - the original LoopVectorize module is
>> registered there.
>>
>>     However, when I give:
>>       opt -debug -load NewLoopVectorize.so -my-loop-vectorize test.c
>>      it does not vectorize a loop that was normally vectorized by the original
>> LoopVectorize module in LLVM. My module gives:
>>         <<loop not vectorized: loop control flow is not understood by vectorizer
>>         LV: Not vectorizing: Cannot prove legality.>>
>>
>>     So I tried to explicitly run a few more passes before my new LoopVectorize to help it
>> (I think the most important one is -loop-simplify, which canonicalizes the loop):
>>       opt -debug -loop-simplify -mem2reg -lcssa -load NewLoopVectorize.so
>> -my-loop-vectorize test.c
>>     However, I do not see any line in the output of opt starting with "LoopSimplify" (when
>> I have success, when I give opt -O3, I get "LoopSimplify: Creating pre-header
>> for.body.preheader; LoopSimplify: Creating dedicated exit block for.end.loopexit" before
>> LoopVectorize) and my LoopVectorize still does not vectorize the loop.
>>
>>     Can anybody help me run explicitly the LoopSimplify pass (before my LoopVectorize
>> pass)?
>>
>>   Thank you. And wish you happy Easter holidays!
>>     Alex