search for: vectorizestorechains

Hi Frank, Thanks for working on this. Please look at vectorizeStoreChains. In this function we process all of the stores in the function in buckets of 16 elements because constructing consecutive stores is implemented using an O(n^2) algorithm. You can try to increase this threshold to 128 and see if it helps. I also agree with Renato and Chad that adding a flag to tel...

...%add29 = add nsw i32 %add27, %15 ret i32 %add29}* Shouldn't the SLP vectorizer vectorize above code? Am I missing out something? How does loop unrolling affect the final vectorization of code? I tried to debug above IR. Since there is no store in above IR, SLP vectorization doesn't call vectorizeStoreChains(). In vectorizeChainsInBlock(), as there is no phi node in above IR, it also fails to vectorize the IR. If it is perfectly legal to vectorize above IR, are we lacking code for it right now? Or its not possible to vectorize above code (as far as i know, it is perfectly possible to vectorize above c...

On 7 August 2014 17:33, Chad Rosier <mcrosier at codeaurora.org> wrote: > You might consider filing a bug (llvm.org/bugs) requesting a flag, but I > don't know if the code owners want to expose such a flag. I'm not sure that's a good idea as a raw access to that limit, as there are no guarantees that it'll stay the same. But maybe a flag turning some

...the chains, and it does > not support joining multiple trees. > > http://llvm.org/docs/doxygen/html/SLPVectorizer_8cpp_source.html#l02861 > > Thanks, > Nadav > > > I tried to debug above IR. Since there is no store in above IR, SLP > vectorization doesn't call > vectorizeStoreChains(). In vectorizeChainsInBlock(), as there is no phi > node in above IR, it > also fails to vectorize the IR. > > If it is perfectly legal to vectorize above IR, are we lacking code for it > right now? Or its not > possible to vectorize above code (as far as i know, it is perfectly...

...joining multiple trees. >> >> http://llvm.org/docs/doxygen/html/SLPVectorizer_8cpp_source.html#l02861 >> >> Thanks, >> Nadav >> >> >> I tried to debug above IR. Since there is no store in above IR, SLP >> vectorization doesn't call >> vectorizeStoreChains(). In vectorizeChainsInBlock(), as there is no phi >> node in above IR, it >> also fails to vectorize the IR. >> >> If it is perfectly legal to vectorize above IR, are we lacking code for >> it right now? Or its not >> possible to vectorize above code (as far as...

...Nadav > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I tried to debug above IR. Since there is no store in above > IR, SLP vectorization doesn't call > > > > > > > vectorizeStoreChains(). In vectorizeChainsInBlock(), as there > is no phi node in above IR, it > > > > > > > also fails to vectorize the IR. > > > > > > > > > > > > > > > > > > > > > If it is perfectly legal to vectorize above IR, ar...