[PATCH] D25652: [AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD).
Elena Demikhovsky via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 25 01:58:59 PDT 2016
delena added a comment.
In https://reviews.llvm.org/D25652#578226, @craig.topper wrote:
> Do you know of cases where passes are doing the wrong thing or is this paranoia?
I don't want to base any assumption to my paranoia. I'm adding Ayal and Farhana, who is working on optimization of interleaved memory accesses.
The question is what instruction is better by default VPERMT or VPERMI. I initially thought, that VPERMT is better inside loops, since VPERMI kills register with indices.
Using VPERMI will require reloading indices again and again for each iteration.
Ayal, Farhana, what do you think?
More information about the llvm-commits