[PATCH] D27395: Fixed store operation for a vector of i1.

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

You don't need load-modify-store for an unmasked store; we're allowed to clobber the padding bits, so you can do something like this:

  define void @store_vec_i1(<2 x i1>* %p, <2 x i1> %val) {
    ; following is equivalent to `store <2 x i1> %val, <2 x i1>* %p`
    %a0 = extractelement <2 x i1> val, i32 0
    %r0 = zext i1 %v1 to i8
    %a1 = extractelement <2 x i1> val, i32 1
    %b1 = zext i1 %v1 to i8
    %c1 = shl i8 %b1, 1
    %r1 = or i1 %r0, %c1
    %p1 = bitcast <2 x i1>* %p to i8*
    store i8 %r1, i8* %p1
  }


Repository:
  rL LLVM

https://reviews.llvm.org/D27395