[LLVMdev] Loop unrolling opportunity in SPEC's libquantum with profile info
dnovillo at google.com
Thu Jan 16 10:09:38 PST 2014
On Thu, Jan 16, 2014 at 10:06 AM, Nadav Rotem <nrotem at apple.com> wrote:
>> If you do the trick of writing to a separate array, then the loop can
>> be vectorized.
> Do you mean that we should change the layout of the array? Or is there another trick?
No, sorry. I meant that if I change the code to instead do the xor
operation into a temp array, then I expect the loop to be vectorized.
But the overhead would likely negate the vectorization benefits
(haven't tried it, though). Incidentally, GCC gives up on vectorizing
this loop for essentially the same reason. It decides that the access
pattern is too complex to bother.
More information about the llvm-dev