Method PMADDUBSWOpImpl

FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp:3228–3268

}

Ref OpDispatchBuilder::PMADDUBSWOpImpl(IR::OpSize Size, Ref Src1, Ref Src2) {
  if (Size == OpSize::i64Bit) {
    const auto MultSize = Size << 1;
    // 64bit is more efficient

    // Src1 is unsigned
    auto Src1_16b = _VUXTL(MultSize, OpSize::i8Bit, Src1); // [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]

    // Src2 is signed
    auto Src2_16b = _VSXTL(MultSize, OpSize::i8Bit, Src2); // [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]

    auto ResMul_L = _VSMull(MultSize, OpSize::i16Bit, Src1_16b, Src2_16b);
    auto ResMul_H = _VSMull2(MultSize, OpSize::i16Bit, Src1_16b, Src2_16b);

    // Now add pairwise across the vector
    auto ResAdd = _VAddP(MultSize, OpSize::i32Bit, ResMul_L, ResMul_H);

    // Add saturate back down to 16bit
    return _VSQXTN(MultSize, OpSize::i32Bit, ResAdd);
  }

  // V{U,S}XTL{,2}/ and VUnZip{,2} can be optimized in this solution to save about one instruction.
  // We can up-front zero extend and sign extend the elements in-place.
  // This means extracting even and odd elements up-front so the unzips aren't required.
  // Requires implementing IR ops for BIC (vector, immediate) although.

  // Src1 is unsigned
  auto Src1_16b_L = _VUXTL(Size, OpSize::i8Bit, Src1); // [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]
  auto Src2_16b_L = _VSXTL(Size, OpSize::i8Bit, Src2); // [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]
  auto ResMul_L = _VMul(Size, OpSize::i16Bit, Src1_16b_L, Src2_16b_L);

  // Src2 is signed
  auto Src1_16b_H = _VUXTL2(Size, OpSize::i8Bit, Src1); // Offset to +64bits [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]
  auto Src2_16b_H = _VSXTL2(Size, OpSize::i8Bit, Src2); // Offset to +64bits [7:0 ], [15:8], [23:16], [31:24], [39:32], [47:40], [55:48], [63:56]
  auto ResMul_L_H = _VMul(Size, OpSize::i16Bit, Src1_16b_H, Src2_16b_H);

  auto TmpZip1 = _VUnZip(Size, OpSize::i16Bit, ResMul_L, ResMul_L_H);
  auto TmpZip2 = _VUnZip2(Size, OpSize::i16Bit, ResMul_L, ResMul_L_H);

  return _VSQAdd(Size, OpSize::i16Bit, TmpZip1, TmpZip2);
}

void OpDispatchBuilder::PMADDUBSW(OpcodeArgs) {
  const auto Size = OpSizeFromSrc(Op);