Expanding to an imm is not always the best solution

[AZero13]

Take the following:

define i32 @test1_known_not_zero_mul(i32 %a) {
; CHECK-LABEL: test1_known_not_zero_mul:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mov w8, #4093 // =0xffd
; CHECK-NEXT:    mov w9, #12288 // =0x3000
; CHECK-NEXT:    movk w8, #65520, lsl #16
; CHECK-NEXT:    movk w9, #7, lsl #16
; CHECK-NEXT:    and w8, w0, w8
; CHECK-NEXT:    orr w0, w8, w9
; CHECK-NEXT:    ret
  %and = and i32 %a, -1044483
  %or = or disjoint i32 %and, 471040
  ret i32 %or
}

Now, here is a case where adding an instruction actually makes the code faster:

define i32 @test1_known_not_zero_mul(i32 %a) {
; CHECK-LABEL: test1_known_not_zero_mul:
; CHECK:       // %bb.0:
; CHECK-NEXT:    mul w9, w0, w0
; CHECK-NEXT:    mov w8, #12288 // =0x3000
; CHECK-NEXT:    movk w8, #7, lsl #16
; CHECK-NEXT:    and w9, w9, #0xfff03fff
; CHECK-NEXT:    orr w0, w9, w8
; CHECK-NEXT:    ret
  %mul = mul i32 %a, %a ; We set the second bit to 0.
  %and = and i32 %mul, -1044483
  %or = or disjoint i32 %and, 471040
  ret i32 %or
}

So, this means, we can just rewrite the first one, at minimum to:

; CHECK-NEXT:    mov w9, w0
; CHECK-NEXT:    mov w8, #12288 // =0x3000
; CHECK-NEXT:    movk w8, #7, lsl #16
; CHECK-NEXT:    and w9, w9, #0xfff03fff
; CHECK-NEXT:    orr w0, w9, w8
; CHECK-NEXT:    ret

Instead of decomposing the and. How can we do this?