[llvm] [X86][X86CompressEVEX] Do not compress when there is a compressed disp8 (PR #221923)

[via llvm-commits]

nishant-sachdeva-amd wrote:

Can we find a way to share the check with the code already present in `X86BaseInfo.h` ? 

```
/// True when an EVEX memory operand with raw byte displacement \p D and CDisp8
/// access size \p N encodes strictly smaller than the non-EVEX (VEX/legacy) form
/// would: EVEX spends a 1-byte scaled disp8 where the other form is forced to a
/// 4-byte disp32. This is the size-peer of X86MCCodeEmitter::isDispOrCDisp8 and
/// must stay in lockstep with it:
///   1. D outside [-128,127] -- else the non-EVEX form already uses a 1-byte
///                          disp8 (which spans -128..127), so the wider EVEX
///                          prefix would only grow the encoding.
///   2. D % N == 0        -- CDisp8 can only represent exact multiples of N.
///   3. isInt<8>(D / N)   -- the compressed displacement is a signed byte.
/// Direction-agnostic: used both by the VEX->EVEX promote pass and the
/// EVEX->VEX demotion veto in CompressEVEX.
inline bool cdisp8BeatsDisp32(int64_t D, int64_t N) {
  if (D >= -128 && D <= 127)
    return false;
  if (D % N != 0)
    return false;
  int64_t Scaled = D / N;
  return Scaled >= -128 && Scaled <= 127;
}

```

https://github.com/llvm/llvm-project/pull/221923