Emulator

wrapping ZeldaAudioRenderer::StoreVPB with pragma optimize("",off) / on fixes the issue. Need to look a bit closer to see if it's dolphin's bug or actual msvc regression in the compiler shipped with vs 17.5 (cl 19.35.32215)

The build log for 5.0-18718 says it used MSBuild version 17.4.1+9a89d02ff and the build log for 5.0-18720 used MSBuild version 17.5.0+6f08c67f3 (there's actually a second 5.0-18720 log for some reason, not sure what happened there). So it's probably an MSVC regression (though I haven't myself confirmed that dolphin actually regressed in 5.0-18720).

The optimized version looks like the following:

void __fastcall DSP::HLE::ZeldaAudioRenderer::StoreVPB(
        DSP::HLE::ZeldaAudioRenderer *this,
        unsigned __int16 voice_id,
        unsigned __int16 *vpb)
{
  __int64 voice_id_; // rdi
  unsigned __int16 *Pointer; // rax MAPDST
  unsigned int m_flags; // ecx
  __int64 vpb_size; // rbx
  __int64 base_idx; // rdi
  unsigned __int64 loop_count; // rdx
  unsigned __int64 v12; // r8
  char *v13; // rax
  __m128i v14; // xmm0
  unsigned __int16 *v15; // rax
  __int16 v16; // cx

  voice_id_ = voice_id;
  Pointer = (unsigned __int16 *)DSP::HLE::HLEMemory_Get_Pointer(this->m_vpb_base_addr);
  m_flags = this->m_flags;
  vpb_size = ~(4 * (unsigned __int8)this->m_flags) & 0x40 | 0x80i64;
  base_idx = vpb_size * voice_id_;
  if ( (m_flags & 0x10) != 0 )
    DSP::HLE::ZeldaAudioRenderer::VPB::Compress((DSP::HLE::ZeldaAudioRenderer::VPB *)vpb);
  loop_count = vpb_size - 0x40;                 // loop_count is 40 or 80
  v12 = 0i64;
  if ( loop_count < 4
    || _isa_available < 2
    || &Pointer[base_idx] <= &vpb[loop_count - 1] && &Pointer[loop_count - 1 + base_idx] >= vpb )
  {
    v15 = vpb;
    do
    {
      v16 = *v15++;
      *(unsigned __int16 *)((char *)Pointer + 2 * base_idx - (_QWORD)vpb + (_QWORD)v15 - 2) = __ROR2__(v16, 8);
      --loop_count;
    }
    while ( loop_count );
  }
  else if ( loop_count < 0x20 )
  {
    v13 = (char *)Pointer + 2 * base_idx - (_QWORD)vpb;
    do
    {
      v14 = _mm_loadl_epi64((const __m128i *)vpb);
      vpb += 4;
      v12 += 4i64;
      *(_QWORD *)&v13[(_QWORD)vpb - 8] = _mm_shuffle_epi8(v14, (__m128i)_xmm).m128i_u64[0];
    }
    while ( v12 < loop_count );
  }
}

The problem appears to be that loop_count < 0x20 is wrong, or there should have been another branch for larger loop_count. I suspect v12 should have been incremented by 1 instead of 4 in this case.

Forcing it to take the non-sse path (setting __isa_available) produces the correct output.

https://developercommunity.visualstudio.com/t/VS-175-bad-codegen-optimizing-loop-with/10291620