difference between polly stripmine, vector and unroll flags

Project Infrastructure LLVM Dev List Archives
1

Hello I m using following polly flags,

clang -O3 -mllvm -polly -mllvm -polly-vectorizer=stripmine

clang -O3 -mllvm -polly -mllvm -polly-vectorizer=polly

clang -O3 -mllvm -polly -mllvm -force-vector-interleave=32

I am getting different execution times for each 3. Can you please tell me the difference between the optimizations performed by these 3 flags?

Please help
Thank You

2

Hello I m using following polly flags,

clang -O3 -mllvm -polly -mllvm -polly-vectorizer=stripmine

This prepares instructions into vector-sized chunks
(-polly-prevect-width, 4 by default), but does not generate vector
instructions itself. The LoopVectorizer can generate vector
instructions from it.

clang -O3 -mllvm -polly -mllvm -polly-vectorizer=polly

Like -polly-vectorizer=stripmine, but generates vector instructions itself.

clang -O3 -mllvm -polly -mllvm -force-vector-interleave=32

This is a LoopVectorizer option. It does not vectorize, but interleave
(a special kind of unrolling).

Michael