libmwemu 0.24.1

use crate::*;

// VPGATHERDD/VPGATHERDQ - Gather Packed Dword/Qword Values Using Dword Indices (AVX2)
//
// Gathers data from memory at locations specified by base address + dword indices.
// Uses a mask register to control which elements are loaded.
// The mask is also used to prevent page faults on unused elements.
//
// VPGATHERDD: Gather 8 packed dwords (32-bit) using 8 dword indices
// VPGATHERDQ: Gather 4 packed qwords (64-bit) using 4 dword indices (lower half)
//
// Opcodes (AVX2 - 256-bit YMM):
// VEX.256.66.0F38.W0 90 /r       VPGATHERDD ymm1, vm32y, ymm2
// VEX.256.66.0F38.W1 90 /r       VPGATHERDQ ymm1, vm32x, ymm2

const ALIGNED_ADDR: u64 = 0x3000;
const GATHER_DATA_ADDR: u64 = 0x4000;

// ============================================================================
// VPGATHERDD Tests - Gather 8 Dwords Using Dword Indices (256-bit)
// ============================================================================

#[test]
fn test_vpgatherdd_ymm0_base_ymm2_ymm1() {
    let mut emu = emu64();
    // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..64 {
        data[i * 4] = (i & 0xFF) as u8;
        data[i * 4 + 1] = ((i >> 8) & 0xFF) as u8;
        data[i * 4 + 2] = 0;
        data[i * 4 + 3] = 0;
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm3_base_ymm5_ymm4() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x5d, 0x90, 0x1c, 0x28, // VPGATHERDD YMM3, [RAX + YMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm6_base_ymm0_ymm7() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x45, 0x90, 0x34, 0x00, // VPGATHERDD YMM6, [RAX + YMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm9_base_ymm11_ymm10() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0x2d, 0x90, 0x0c, 0x58, // VPGATHERDD YMM9, [RAX + YMM11*1], YMM10
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm12_base_ymm14_ymm13() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0x15, 0x90, 0x24, 0x70, // VPGATHERDD YMM12, [RAX + YMM14*1], YMM13
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..64 {
        data[i * 4] = 0x12;
        data[i * 4 + 1] = 0x34;
        data[i * 4 + 2] = 0x56;
        data[i * 4 + 3] = 0x78;
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm0_base_ymm2_ymm1_scale2() {
    let mut emu = emu64();
    // VPGATHERDD with scale=2
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x50, // VPGATHERDD YMM0, [RAX + YMM2*2], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm0_base_ymm2_ymm1_scale4() {
    let mut emu = emu64();
    // VPGATHERDD with scale=4
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x90, // VPGATHERDD YMM0, [RAX + YMM2*4], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm0_base_ymm2_ymm1_scale8() {
    let mut emu = emu64();
    // VPGATHERDD with scale=8
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0xd0, // VPGATHERDD YMM0, [RAX + YMM2*8], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm3_rbx_base_ymm5_ymm4() {
    let mut emu = emu64();
    let code = [0x48, 0xbb];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x5d, 0x90, 0x1c, 0x2b, // VPGATHERDD YMM3, [RBX + YMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm6_rcx_base_ymm0_ymm7() {
    let mut emu = emu64();
    let code = [0x48, 0xb9];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x45, 0x90, 0x34, 0x01, // VPGATHERDD YMM6, [RCX + YMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_ymm9_rdx_base_ymm11_ymm10() {
    let mut emu = emu64();
    let code = [0x48, 0xba];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0x2d, 0x90, 0x0c, 0x5a, // VPGATHERDD YMM9, [RDX + YMM11*1], YMM10
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x33];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

// ============================================================================
// VPGATHERDQ Tests - Gather 4 Qwords Using Dword Indices (256-bit)
// ============================================================================

#[test]
fn test_vpgatherdq_ymm0_base_xmm2_ymm1() {
    let mut emu = emu64();
    // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..32 {
        data[i * 8] = (i & 0xFF) as u8;
        data[i * 8 + 1] = ((i >> 8) & 0xFF) as u8;
        data[i * 8 + 2] = 0;
        data[i * 8 + 3] = 0;
        data[i * 8 + 4] = 0;
        data[i * 8 + 5] = 0;
        data[i * 8 + 6] = 0;
        data[i * 8 + 7] = 0;
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm3_base_xmm5_ymm4() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xdd, 0x90, 0x1c, 0x28, // VPGATHERDQ YMM3, [RAX + XMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm6_base_xmm0_ymm7() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xc5, 0x90, 0x34, 0x00, // VPGATHERDQ YMM6, [RAX + XMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm9_base_xmm11_ymm10() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0xad, 0x90, 0x0c, 0x58, // VPGATHERDQ YMM9, [RAX + XMM11*1], YMM10
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm12_base_xmm14_ymm13() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0x95, 0x90, 0x24, 0x70, // VPGATHERDQ YMM12, [RAX + XMM14*1], YMM13
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..32 {
        data[i * 8] = 0x12;
        data[i * 8 + 1] = 0x34;
        data[i * 8 + 2] = 0x56;
        data[i * 8 + 3] = 0x78;
        data[i * 8 + 4] = 0x9A;
        data[i * 8 + 5] = 0xBC;
        data[i * 8 + 6] = 0xDE;
        data[i * 8 + 7] = 0xF0;
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm0_base_xmm2_ymm1_scale2() {
    let mut emu = emu64();
    // VPGATHERDQ with scale=2
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x50, // VPGATHERDQ YMM0, [RAX + XMM2*2], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm0_base_xmm2_ymm1_scale4() {
    let mut emu = emu64();
    // VPGATHERDQ with scale=4
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x90, // VPGATHERDQ YMM0, [RAX + XMM2*4], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC, 0xCC];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm0_base_xmm2_ymm1_scale8() {
    let mut emu = emu64();
    // VPGATHERDQ with scale=8
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0xd0, // VPGATHERDQ YMM0, [RAX + XMM2*8], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77, 0x77];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm3_rbx_base_xmm5_ymm4() {
    let mut emu = emu64();
    let code = [0x48, 0xbb];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xdd, 0x90, 0x1c, 0x2b, // VPGATHERDQ YMM3, [RBX + XMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_ymm6_rcx_base_xmm0_ymm7() {
    let mut emu = emu64();
    let code = [0x48, 0xb9];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xc5, 0x90, 0x34, 0x01, // VPGATHERDQ YMM6, [RCX + XMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22, 0x22];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

// ============================================================================
// Comprehensive tests
// ============================================================================

#[test]
fn test_vpgatherdd_sequential_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..64 {
        let val = (i * 0x11) as u32;
        data[i * 4..i * 4 + 4].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_sequential_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..32usize {
        let val = (i as u64).wrapping_mul(0x1122334455667788u64);
        data[i * 8..i * 8 + 8].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_extended_regs() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0x3d, 0x90, 0x04, 0x40, // VPGATHERDD YMM8, [RAX + YMM8*1], YMM8
        0xc4, 0x62, 0x15, 0x90, 0x14, 0x70, // VPGATHERDD YMM10, [RAX + YMM14*1], YMM13
        0xc4, 0x62, 0x05, 0x90, 0x3c, 0x78, // VPGATHERDD YMM15, [RAX + YMM15*1], YMM15
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB, 0xAB];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_extended_regs() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0x62, 0xbd, 0x90, 0x04, 0x40, // VPGATHERDQ YMM8, [RAX + XMM8*1], YMM8
        0xc4, 0x62, 0x95, 0x90, 0x14, 0x70, // VPGATHERDQ YMM10, [RAX + XMM14*1], YMM13
        0xc4, 0x62, 0x85, 0x90, 0x3c, 0x78, // VPGATHERDQ YMM15, [RAX + XMM15*1], YMM15
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD, 0xCD];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_all_scales() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xc4, 0xe2, 0x65, 0x90, 0x1c, 0x68, // VPGATHERDD YMM3, [RAX + YMM5*2], YMM4
        0xc4, 0xe2, 0x45, 0x90, 0x34, 0x80, // VPGATHERDD YMM6, [RAX + YMM0*4], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99, 0x99];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_all_scales() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xc4, 0xe2, 0xe5, 0x90, 0x1c, 0x68, // VPGATHERDQ YMM3, [RAX + XMM5*2], YMM4
        0xc4, 0xe2, 0xc5, 0x90, 0x34, 0x80, // VPGATHERDQ YMM6, [RAX + XMM0*4], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88, 0x88];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_all_base_regs() {
    let mut emu = emu64();
    let code = [
        0x48, 0xb8, // MOV RAX, imm64
    ];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0x48, 0xbb, // MOV RBX, imm64
    ]);
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xc4, 0xe2, 0x65, 0x90, 0x1c, 0x2b, // VPGATHERDD YMM3, [RBX + YMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44, 0x44];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_all_base_regs() {
    let mut emu = emu64();
    let code = [
        0x48, 0xb8, // MOV RAX, imm64
    ];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0x48, 0xb9, // MOV RCX, imm64
    ]);
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xc4, 0xe2, 0xe5, 0x90, 0x1c, 0x29, // VPGATHERDQ YMM3, [RCX + XMM5*1], YMM4
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66, 0x66];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_zero_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    data[0..4].copy_from_slice(&0xDEADBEEFu32.to_le_bytes());
    for i in 1..64 {
        data[i * 4..i * 4 + 4].copy_from_slice(&(i as u32).to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_zero_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    data[0..8].copy_from_slice(&0xCAFEBABEDEADBEEFu64.to_le_bytes());
    for i in 1..32 {
        data[i * 8..i * 8 + 8].copy_from_slice(&(i as u64).to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_mixed_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..64 {
        let val = ((i * 7 + 13) & 0xFF) as u32;
        data[i * 4..i * 4 + 4].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_mixed_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..32 {
        let val = ((i * 11 + 7) & 0xFF) as u64;
        data[i * 8..i * 8 + 8].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_high_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..64 {
        let val = 0x80000000u32 | (i as u32);
        data[i * 4..i * 4 + 4].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_high_indices() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let mut data = vec![0u8; 256];
    for i in 0..32 {
        let val = 0x8000000000000000u64 | (i as u64);
        data[i * 8..i * 8 + 8].copy_from_slice(&val.to_le_bytes());
    }
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdd_chain_multiple() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0x75, 0x90, 0x04, 0x10, // VPGATHERDD YMM0, [RAX + YMM2*1], YMM1
        0xc4, 0xe2, 0x65, 0x90, 0x1c, 0x28, // VPGATHERDD YMM3, [RAX + YMM5*1], YMM4
        0xc4, 0xe2, 0x45, 0x90, 0x34, 0x00, // VPGATHERDD YMM6, [RAX + YMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A, 0x5A];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}

#[test]
fn test_vpgatherdq_chain_multiple() {
    let mut emu = emu64();
    let code = [0x48, 0xb8];
    let mut full_code = code.to_vec();
    full_code.extend_from_slice(&GATHER_DATA_ADDR.to_le_bytes());
    full_code.extend_from_slice(&[
        0xc4, 0xe2, 0xf5, 0x90, 0x04, 0x10, // VPGATHERDQ YMM0, [RAX + XMM2*1], YMM1
        0xc4, 0xe2, 0xe5, 0x90, 0x1c, 0x28, // VPGATHERDQ YMM3, [RAX + XMM5*1], YMM4
        0xc4, 0xe2, 0xc5, 0x90, 0x34, 0x00, // VPGATHERDQ YMM6, [RAX + XMM0*1], YMM7
        0xf4, // HLT
    ]);

    emu.load_code_bytes(&full_code);
    let data = vec![0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5, 0xA5];
    emu.maps.write_bytes_slice(GATHER_DATA_ADDR, &data);
    emu.run(None).unwrap();
}