Missed Optimization: fold shift+add then `ashr 1` into smaller shift + add

[Leo0506]

define i32 @src(i32 %arg0) {
  %v0 = shl nsw i32 %arg0, 4
  %v1 = add nsw i32 %v0, 16
  %v2 = ashr exact i32 %v1, 1
  ret i32 %v2
}

define i32 @tgt(i32 %arg0) {
  %v0 = shl i32 %arg0, 3
  %v1 = add i32 %v0, 8
  ret i32 %v1
}

alive2: https://alive2.llvm.org/ce/z/bUptzx
godbolt: https://godbolt.org/z/zfarv13YK
Pattern Found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/ffmpeg/optimized/dnxhdenc.ll

[YLChenZ]

If x ≥ y and n ≥ y, (a << x + 2^n) >> y => a << (x - y) + 2^(n-y) should also be correct.

[manik-muk]

I'll take this issue!