Extended Description

Adapted/reduced from the regression test in https://reviews.llvm.org/D111800 (that example could have further shuffle combining which would eliminate a shuffle and almost definitely be a win):

define <3 x i3> @​src(<3 x i3> %x, <3 x i3> %y, <3 x i3> %z) {
%a1 = sub <3 x i3> %x, %y
%a2 = sub <3 x i3> %x, %z
%r = shufflevector <3 x i3> %a1, <3 x i3> %a2, <3 x i32> <i32 0, i32 5, i32 3>
ret <3 x i3> %r
}

define <3 x i3> @​tgt(<3 x i3> %x, <3 x i3> %y, <3 x i3> %z) {
%s1 = shufflevector <3 x i3> %x, <3 x i3> %x, <3 x i32> <i32 0, i32 5, i32 3>
%s2 = shufflevector <3 x i3> %y, <3 x i3> %z, <3 x i32> <i32 0, i32 5, i32 3>
%r2 = sub <3 x i3> %s1, %s2
ret <3 x i3> %r2
}

https://alive2.llvm.org/ce/z/d8zB4D

Note: the shuffle of '%x' can be reduced to a unary shuffle; I left the mask as-is to show that we're re-using the incoming shuffle mask.

As long as the shuffle is no more expensive than the binop, this should be a win and would potentially allow subsequent combines.