[Bug target/56309] conditional moves instead of compare and branch result in almost 2x slower code

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56309

--- Comment #29 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
FYI, I've posted patchset to only do tree if-conversion for vectorization and
not for scalar basic blocks (and also for conditional load/store support in
vectorized loops using AVX vmaskmov* and AVX2 masked gathers) in
http://gcc.gnu.org/ml/gcc-patches/2013-10/msg01373.html
I'd appreciate any benchmark results for that.

On the by-ref-O3.ii testcase there is no change, but by-val-O3.ii seems to be
consistently faster, around 20% on AMD 8354 and 35% on Intel 2600.