Redundant zero-extension when bswap follows load

[dvulakh]

Example program:

type bigstring =
  (char, Bigarray.int8_unsigned_elt, Bigarray.c_layout) Bigarray.Array1.t

external unsafe_get_16
  :  bigstring
  -> pos:int64#
  -> int
  = "%caml_bigstring_get16u_indexed_by_int64#"

external swap16 : int -> int = "%bswap16"

let unsafe_get_16_swp t ~pos = unsafe_get_16 t ~pos |> swap16

cmm (post #3336):

 (+
   (<<
     (bswap_16
       (let ba_data/383 (load_mut int (+a t/380 8))
         (load_mut unsigned int16 (+ ba_data/383 pos/381))))
     1)
   1))

asm (post #3336):

movq	8(%rax), %rax
movzwq	(%rax,%rbx), %rax
xchg	%ah, %al
movzwq	%ax, %rax
leaq	1(%rax,%rax), %rax
ret

The second movzwq is a noop, because the bottom 16 bits of %rax were populated with a zero-extending load and then byte-swapped, so there is no reason to sign extend again.

Simplifying the code here needs to happen very late in the compiler: bswap remains an opaque operation that conceals the zero-extension from cmm onward, all the way until asm is emitted.