Hello! > dropping the builtin as early as possible seems like it can only help us > optimize the code. Jakub > suggested in the PR that he liked this approach better than using > __builtin_shuffle in the header. > There is already some coverage in the testsuite (as I noticed when I tried to > restrict the > argument to [0, 3]...).
I think this approach is OK, the compilation can benefit from generic optimization paths on the tree level. There are some RTL optimizations in the combine pass, but these also apply with the new approach. > If this one is ok, I may add a few more (say shufps to begin with) later. Yes, the new approach is the way to go. > Bootstrap+regtest on x86_64-pc-linux-gnu. > > 2019-05-20 Marc Glisse <marc.gli...@inria.fr> > > PR rtl-optimization/43147 > * config/i386/i386.c (ix86_gimple_fold_builtin): Handle > IX86_BUILTIN_SHUFPD. LGTM. Thanks, Uros.
