A quick check with...
gfortran -ffast-math -funroll-loops -msse3 -O3 -fgraphite-identity
-ftree-vectorizer-verbose=2 air.f90 -o air >& air_cloog_legacy.txt
grep "LOOP VECTORIZED" air_cloog_legacy.txt | wc -l
5
gfortran -ffast-math -funroll-loops -msse3 -O3 -fgraphite-identity
-ftree-vectorizer-verbose=2 air.f90 -o air >& air_cloog_isl.txt
grep "LOOP VECTORIZED" air_cloog_isl.txt | wc -l
5
...for each build suggests that the 12% runtime performance penalty when using
cloog-isl isn't due to missed vectorizations.
Jack
