[VectorCombine] foldShuffleOfIntrinsics - failure to recognise repeated operands will reduce costs

[RKSimon]

define <8 x float> @src(<4 x float> %x0, <4 x float> %x1, <4 x float> %y0, <4 x float> %y1) {
  %l0 = call <4 x float> @llvm.fma.v4f32(<4 x float> %x0, <4 x float> %x0, <4 x float> zeroinitializer)
  %l1 = call <4 x float> @llvm.fma.v4f32(<4 x float> %x1, <4 x float> %x1, <4 x float> %l0)

  %h0 = call <4 x float> @llvm.fma.v4f32(<4 x float> %y0, <4 x float> %y0, <4 x float> zeroinitializer)
  %h1 = call <4 x float> @llvm.fma.v4f32(<4 x float> %y1, <4 x float> %y1, <4 x float> %h0)

  %res = shufflevector <4 x float> %l1, <4 x float> %h1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>

  ret <8 x float> %res
}

We fail to fold @src to @tgt as the cost analysis assume the shuffle of each operand of the fma are independent:

define <8 x float> @tgt(<4 x float> %x0, <4 x float> %x1, <4 x float> %y0, <4 x float> %y1) {
  %xy0 = shufflevector <4 x float> %x0, <4 x float> %y0, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %xy1 = shufflevector <4 x float> %x1, <4 x float> %y1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>

  %r0 = call <8 x float> @llvm.fma.v8f32(<8 x float> %xy0, <8 x float> %xy0, <8 x float> zeroinitializer)
  %r1 = call <8 x float> @llvm.fma.v8f32(<8 x float> %xy1, <8 x float> %xy1, <8 x float> %r0)

  ret <8 x float> %r1
}