Rerun benchmarks after reordering or u2 and u1 lowering. Seems to have caused performance regressions, even though the change sounds like it should be an improvement.

Author: chriselrod

benchmark/loadsharedlibs.jl

@@ -16,7 +16,7 @@ const LIBDIRECTCALLJIT = joinpath(LOOPVECBENCHDIR, "libdcjtest.so")
 # requires Clang with polly to build
 cfile = joinpath(LOOPVECBENCHDIR, "looptests.c")
 if !isfile(LIBCTEST) || mtime(cfile) > mtime(LIBCTEST)    
-    run(`clang -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -lm -shared -fPIC $cfile -o $LIBCTEST`)
+    run(`/usr/bin/clang -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -lm -shared -fPIC $cfile -o $LIBCTEST`)
     # run(`/usr/local/bin/clang -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -lm -mllvm -polly -mllvm -polly-vectorizer=stripmine -shared -fPIC $cfile -o $LIBCTEST`)
 end
 if !isfile(LIBICTEST) || mtime(cfile) > mtime(LIBICTEST)

benchmark/looptests.c

@@ -236,7 +236,7 @@ double randomaccess(double* restrict P, long* restrict basis, double* restrict c
   }
   return p;
 }
-double logdettriangle(double* T, long N){
+double logdettriangle(double* restrict T, long N){
   double ld = 0;
   for (long n = 0; n < N; n++){
     ld += log(T[n + n*N]);