Patchwork [v8,06/14] target-mips-ase-dsp: Add arithmetic instructions

login
register
mail settings
Submitter Jia Liu
Date Sept. 12, 2012, 2:01 a.m.
Message ID <1347415315-2180-7-git-send-email-proljc@gmail.com>
Download mbox | patch
Permalink /patch/183241/
State New
Headers show

Comments

Jia Liu - Sept. 12, 2012, 2:01 a.m.
Add MIPS ASE DSP Arithmetic instructions.

Signed-off-by: Jia Liu <proljc@gmail.com>
---
 target-mips/dsp_helper.c | 1833 ++++++++++++++++++++++++++++++++++++++++++++++
 target-mips/helper.h     |  126 ++++
 target-mips/translate.c  |  681 ++++++++++++++++-
 3 files changed, 2637 insertions(+), 3 deletions(-)
Aurelien Jarno - Sept. 18, 2012, 4:37 p.m.
On Wed, Sep 12, 2012 at 10:01:47AM +0800, Jia Liu wrote:
> Add MIPS ASE DSP Arithmetic instructions.
> 
> Signed-off-by: Jia Liu <proljc@gmail.com>
> ---
>  target-mips/dsp_helper.c | 1833 ++++++++++++++++++++++++++++++++++++++++++++++
>  target-mips/helper.h     |  126 ++++
>  target-mips/translate.c  |  681 ++++++++++++++++-
>  3 files changed, 2637 insertions(+), 3 deletions(-)
> 
> diff --git a/target-mips/dsp_helper.c b/target-mips/dsp_helper.c
> index 779f5c7..4f7cc9e 100644
> --- a/target-mips/dsp_helper.c
> +++ b/target-mips/dsp_helper.c
> @@ -1098,3 +1098,1836 @@ static inline int32_t mipsdsp_cmp_lt(uint32_t a, uint32_t b)
>      return a < b;
>  }
>  /*** MIPS DSP internal functions end ***/
> +
> +#define MIPSDSP_LHI 0xFFFFFFFF00000000ull
> +#define MIPSDSP_LLO 0x00000000FFFFFFFFull
> +#define MIPSDSP_HI  0xFFFF0000
> +#define MIPSDSP_LO  0x0000FFFF
> +#define MIPSDSP_Q3  0xFF000000
> +#define MIPSDSP_Q2  0x00FF0000
> +#define MIPSDSP_Q1  0x0000FF00
> +#define MIPSDSP_Q0  0x000000FF
> +
> +/** DSP Arithmetic Sub-class insns **/
> +target_ulong helper_addq_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    int16_t  rsh, rsl, rth, rtl, temph, templ;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    temph = mipsdsp_add_i16(env, rsh, rth);
> +    templ = mipsdsp_add_i16(env, rsl, rtl);
> +
> +    return (target_long)(int32_t)(((unsigned int)temph << 16) \
> +                                  | ((unsigned int)templ & 0xFFFF));
> +}
> +
> +target_ulong helper_addq_s_ph(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    int16_t rsh, rsl, rth, rtl, temph, templ;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    temph = mipsdsp_sat_add_i16(env, rsh, rth);
> +    templ = mipsdsp_sat_add_i16(env, rsl, rtl);
> +
> +    return (target_long)(int32_t)(((uint32_t)temph << 16) \
> +                                  | ((uint32_t)templ & 0xFFFF));
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_addq_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_add_i16(env, rs3, rt3);
> +    tempC = mipsdsp_add_i16(env, rs2, rt2);
> +    tempB = mipsdsp_add_i16(env, rs1, rt1);
> +    tempA = mipsdsp_add_i16(env, rs0, rt0);
> +
> +    return ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +        ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +}
> +
> +target_ulong helper_addq_s_qh(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sat_add_i16(env, rs3, rt3);
> +    tempC = mipsdsp_sat_add_i16(env, rs2, rt2);
> +    tempB = mipsdsp_sat_add_i16(env, rs1, rt1);
> +    tempA = mipsdsp_sat_add_i16(env, rs0, rt0);
> +
> +    return ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +        ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +}
> +#endif
> +
> +target_ulong helper_addq_s_w(CPUMIPSState *env,
> +                             target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    rd = mipsdsp_sat_add_i32(env, rs, rt);
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_addq_pw(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rs1, rs0;
> +    uint32_t rt1, rt0;
> +    uint32_t tempB, tempA;
> +
> +    rs1 = (rs >> 32) & MIPSDSP_LLO;
> +    rs0 = rs & MIPSDSP_LLO;
> +    rt1 = (rt >> 32) & MIPSDSP_LLO;
> +    rt0 = rt & MIPSDSP_LLO;
> +
> +    tempB = mipsdsp_add_i32(env, rs1, rt1);
> +    tempA = mipsdsp_add_i32(env, rs0, rt0);
> +
> +    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
> +}
> +
> +target_ulong helper_addq_s_pw(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rs1, rs0;
> +    uint32_t rt1, rt0;
> +    uint32_t tempB, tempA;
> +
> +    rs1 = (rs >> 32) & MIPSDSP_LLO;
> +    rs0 = rs & MIPSDSP_LLO;
> +    rt1 = (rt >> 32) & MIPSDSP_LLO;
> +    rt0 = rt & MIPSDSP_LLO;
> +
> +    tempB = mipsdsp_sat_add_i32(env, rs1, rt1);
> +    tempA = mipsdsp_sat_add_i32(env, rs0, rt0);
> +
> +    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
> +}
> +#endif
> +
> +target_ulong helper_addu_qb(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    uint8_t  rs0, rs1, rs2, rs3;
> +    uint8_t  rt0, rt1, rt2, rt3;
> +    uint8_t  temp0, temp1, temp2, temp3;
> +
> +    rs0 =  rs & MIPSDSP_Q0;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +
> +    temp0 = mipsdsp_add_u8(env, rs0, rt0);
> +    temp1 = mipsdsp_add_u8(env, rs1, rt1);
> +    temp2 = mipsdsp_add_u8(env, rs2, rt2);
> +    temp3 = mipsdsp_add_u8(env, rs3, rt3);
> +
> +    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
> +         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
> +         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
> +         ((uint32_t)temp0         & MIPSDSP_Q0);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addu_s_qb(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    uint8_t rs0, rs1, rs2, rs3;
> +    uint8_t rt0, rt1, rt2, rt3;
> +    uint8_t temp0, temp1, temp2, temp3;
> +
> +    rs0 =  rs & MIPSDSP_Q0;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +
> +    temp0 = mipsdsp_sat_add_u8(env, rs0, rt0);
> +    temp1 = mipsdsp_sat_add_u8(env, rs1, rt1);
> +    temp2 = mipsdsp_sat_add_u8(env, rs2, rt2);
> +    temp3 = mipsdsp_sat_add_u8(env, rs3, rt3);
> +
> +    rd = (((uint8_t)temp3 << 24) & MIPSDSP_Q3) |
> +         (((uint8_t)temp2 << 16) & MIPSDSP_Q2) |
> +         (((uint8_t)temp1 <<  8) & MIPSDSP_Q1) |
> +         ((uint8_t)temp0         & MIPSDSP_Q0);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_adduh_qb(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    uint8_t  rs0, rs1, rs2, rs3;
> +    uint8_t  rt0, rt1, rt2, rt3;
> +    uint8_t  temp0, temp1, temp2, temp3;
> +
> +    rs0 =  rs & MIPSDSP_Q0;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +
> +    temp0 = mipsdsp_rshift1_add_u8(rs0, rt0);
> +    temp1 = mipsdsp_rshift1_add_u8(rs1, rt1);
> +    temp2 = mipsdsp_rshift1_add_u8(rs2, rt2);
> +    temp3 = mipsdsp_rshift1_add_u8(rs3, rt3);
> +
> +    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
> +         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
> +         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
> +         ((uint32_t)temp0         & MIPSDSP_Q0);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_adduh_r_qb(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    uint8_t  rs0, rs1, rs2, rs3;
> +    uint8_t  rt0, rt1, rt2, rt3;
> +    uint8_t  temp0, temp1, temp2, temp3;
> +
> +    rs0 =  rs & MIPSDSP_Q0;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +
> +    temp0 = mipsdsp_rrshift1_add_u8(rs0, rt0);
> +    temp1 = mipsdsp_rrshift1_add_u8(rs1, rt1);
> +    temp2 = mipsdsp_rrshift1_add_u8(rs2, rt2);
> +    temp3 = mipsdsp_rrshift1_add_u8(rs3, rt3);
> +
> +    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
> +         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
> +         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
> +         ((uint32_t)temp0         & MIPSDSP_Q0);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addu_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl, rth, rtl, temph, templ;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +    temph = mipsdsp_add_u16(env, rsh, rth);
> +    templ = mipsdsp_add_u16(env, rsl, rtl);
> +    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addu_s_ph(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl, rth, rtl, temph, templ;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +    temph = mipsdsp_sat_add_u16(env, rsh, rth);
> +    templ = mipsdsp_sat_add_u16(env, rsl, rtl);
> +    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addqh_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int16_t rsh, rsl, rth, rtl, temph, templ;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    temph = mipsdsp_rshift1_add_q16(rsh, rth);
> +    templ = mipsdsp_rshift1_add_q16(rsl, rtl);
> +    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addqh_r_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int16_t rsh, rsl, rth, rtl, temph, templ;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    temph = mipsdsp_rrshift1_add_q16(rsh, rth);
> +    templ = mipsdsp_rrshift1_add_q16(rsl, rtl);
> +    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addqh_w(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +
> +    rd = mipsdsp_rshift1_add_q32(rs, rt);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_addqh_r_w(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +
> +    rd = mipsdsp_rrshift1_add_q32(rs, rt);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_addu_ob(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_add_u8(env, rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_addu_s_ob(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_sat_add_u8(env, rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_adduh_ob(target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_rshift1_add_u8(rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_adduh_r_ob(target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_rrshift1_add_u8(rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_addu_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_add_u16(env, rs3, rt3);
> +    tempC = mipsdsp_add_u16(env, rs2, rt2);
> +    tempB = mipsdsp_add_u16(env, rs1, rt1);
> +    tempA = mipsdsp_add_u16(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_addu_s_qh(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sat_add_u16(env, rs3, rt3);
> +    tempC = mipsdsp_sat_add_u16(env, rs2, rt2);
> +    tempB = mipsdsp_sat_add_u16(env, rs1, rt1);
> +    tempA = mipsdsp_sat_add_u16(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_subq_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl;
> +    uint16_t rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    tempB = mipsdsp_sub_i16(env, rsh, rth);
> +    tempA = mipsdsp_sub_i16(env, rsl, rtl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subq_s_ph(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl;
> +    uint16_t rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    tempB = mipsdsp_sat16_sub(env, rsh, rth);
> +    tempA = mipsdsp_sat16_sub(env, rsl, rtl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_subq_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sub_i16(env, rs3, rt3);
> +    tempC = mipsdsp_sub_i16(env, rs2, rt2);
> +    tempB = mipsdsp_sub_i16(env, rs1, rt1);
> +    tempA = mipsdsp_sub_i16(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +      ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +    return temp;
> +}
> +
> +target_ulong helper_subq_s_qh(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sat16_sub(env, rs3, rt3);
> +    tempC = mipsdsp_sat16_sub(env, rs2, rt2);
> +    tempB = mipsdsp_sat16_sub(env, rs1, rt1);
> +    tempA = mipsdsp_sat16_sub(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_subq_s_w(CPUMIPSState *env,
> +                             target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +
> +    rd = mipsdsp_sat32_sub(env, rs, rt);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_subq_pw(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rs1, rs0;
> +    uint32_t rt1, rt0;
> +    uint32_t tempB, tempA;
> +
> +    rs1 = (rs >> 32) & MIPSDSP_LLO;
> +    rs0 = rs & MIPSDSP_LLO;
> +    rt1 = (rt >> 32) & MIPSDSP_LLO;
> +    rt0 = rt & MIPSDSP_LLO;
> +
> +    tempB = mipsdsp_sub32(env, rs1, rt1);
> +    tempA = mipsdsp_sub32(env, rs0, rt0);
> +
> +    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
> +}
> +
> +target_ulong helper_subq_s_pw(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rs1, rs0;
> +    uint32_t rt1, rt0;
> +    uint32_t tempB, tempA;
> +
> +    rs1 = (rs >> 32) & MIPSDSP_LLO;
> +    rs0 = rs & MIPSDSP_LLO;
> +    rt1 = (rt >> 32) & MIPSDSP_LLO;
> +    rt0 = rt & MIPSDSP_LLO;
> +
> +    tempB = mipsdsp_sat32_sub(env, rs1, rt1);
> +    tempA = mipsdsp_sat32_sub(env, rs0, rt0);
> +
> +    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
> +}
> +#endif
> +
> +target_ulong helper_subu_qb(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs3, rs2, rs1, rs0;
> +    uint8_t rt3, rt2, rt1, rt0;
> +    uint8_t tempD, tempC, tempB, tempA;
> +    uint32_t rd;
> +
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs0 =  rs & MIPSDSP_Q0;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +
> +    tempD = mipsdsp_sub_u8(env, rs3, rt3);
> +    tempC = mipsdsp_sub_u8(env, rs2, rt2);
> +    tempB = mipsdsp_sub_u8(env, rs1, rt1);
> +    tempA = mipsdsp_sub_u8(env, rs0, rt0);
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subu_s_qb(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs3, rs2, rs1, rs0;
> +    uint8_t rt3, rt2, rt1, rt0;
> +    uint8_t tempD, tempC, tempB, tempA;
> +    uint32_t rd;
> +
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs0 =  rs & MIPSDSP_Q0;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +
> +    tempD = mipsdsp_satu8_sub(env, rs3, rt3);
> +    tempC = mipsdsp_satu8_sub(env, rs2, rt2);
> +    tempB = mipsdsp_satu8_sub(env, rs1, rt1);
> +    tempA = mipsdsp_satu8_sub(env, rs0, rt0);
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subuh_qb(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs3, rs2, rs1, rs0;
> +    uint8_t rt3, rt2, rt1, rt0;
> +    uint8_t tempD, tempC, tempB, tempA;
> +    uint32_t rd;
> +
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs0 =  rs & MIPSDSP_Q0;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +
> +    tempD = ((uint16_t)rs3 - (uint16_t)rt3) >> 1;
> +    tempC = ((uint16_t)rs2 - (uint16_t)rt2) >> 1;
> +    tempB = ((uint16_t)rs1 - (uint16_t)rt1) >> 1;
> +    tempA = ((uint16_t)rs0 - (uint16_t)rt0) >> 1;
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +
> +    return (target_ulong)rd;
> +}
> +
> +target_ulong helper_subuh_r_qb(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs3, rs2, rs1, rs0;
> +    uint8_t rt3, rt2, rt1, rt0;
> +    uint8_t tempD, tempC, tempB, tempA;
> +    uint32_t rd;
> +
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs0 =  rs & MIPSDSP_Q0;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +
> +    tempD = ((uint16_t)rs3 - (uint16_t)rt3 + 1) >> 1;
> +    tempC = ((uint16_t)rs2 - (uint16_t)rt2 + 1) >> 1;
> +    tempB = ((uint16_t)rs1 - (uint16_t)rt1 + 1) >> 1;
> +    tempA = ((uint16_t)rs0 - (uint16_t)rt0 + 1) >> 1;
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +
> +    return (target_ulong)rd;
> +}
> +
> +target_ulong helper_subu_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl, rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    tempB = mipsdsp_sub_u16_u16(env, rth, rsh);
> +    tempA = mipsdsp_sub_u16_u16(env, rtl, rsl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subu_s_ph(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl, rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    tempB = mipsdsp_satu16_sub_u16_u16(env, rth, rsh);
> +    tempA = mipsdsp_satu16_sub_u16_u16(env, rtl, rsl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subqh_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl;
> +    uint16_t rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +    tempB = mipsdsp_rshift1_sub_q16(rsh, rth);
> +    tempA = mipsdsp_rshift1_sub_q16(rsl, rtl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subqh_r_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rsh, rsl;
> +    uint16_t rth, rtl;
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +    tempB = mipsdsp_rrshift1_sub_q16(rsh, rth);
> +    tempA = mipsdsp_rrshift1_sub_q16(rsl, rtl);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subqh_w(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +
> +    rd = mipsdsp_rshift1_sub_q32(rs, rt);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_subqh_r_w(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +
> +    rd = mipsdsp_rrshift1_sub_q32(rs, rt);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_subu_ob(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_sub_u8(env, rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result += (uint64_t)temp[i] << (i * 8);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_subu_s_ob(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = mipsdsp_satu8_sub(env, rs_t[i], rt_t[i]);
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result += (uint64_t)temp[i] << (i * 8);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_subuh_ob(target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = ((uint16_t)rs_t[i] - (uint16_t)rt_t[i]) >> 1;
> +      }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_subuh_r_ob(target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint8_t rs_t[8], rt_t[8];
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
> +        temp[i] = ((uint16_t)rs_t[i] - (uint16_t)rt_t[i] + 1) >> 1;
> +    }
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_subu_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sub_u16_u16(env, rs3, rt3);
> +    tempC = mipsdsp_sub_u16_u16(env, rs2, rt2);
> +    tempB = mipsdsp_sub_u16_u16(env, rs1, rt1);
> +    tempA = mipsdsp_sub_u16_u16(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +      ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +    return temp;
> +}
> +
> +
> +target_ulong helper_subu_s_qh(CPUMIPSState *env,
> +                              target_ulong rs, target_ulong rt)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_satu16_sub_u16_u16(env, rs3, rt3);
> +    tempC = mipsdsp_satu16_sub_u16_u16(env, rs2, rt2);
> +    tempB = mipsdsp_satu16_sub_u16_u16(env, rs1, rt1);
> +    tempA = mipsdsp_satu16_sub_u16_u16(env, rs0, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_addsc(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint64_t temp, tempRs, tempRt;
> +    int32_t flag;
> +
> +    tempRs = (uint64_t)rs & MIPSDSP_LLO;
> +    tempRt = (uint64_t)rt & MIPSDSP_LLO;
> +
> +    temp = tempRs + tempRt;
> +    flag = (temp & 0x0100000000ull) >> 32;
> +    set_DSPControl_carryflag(env, flag);
> +
> +    return (target_long)(int32_t)(temp & MIPSDSP_LLO);
> +}
> +
> +target_ulong helper_addwc(CPUMIPSState *env, target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int32_t temp32, temp31;
> +    int64_t tempL;
> +
> +    tempL = (int32_t)rs + (int32_t)rt + get_DSPControl_carryflag(env);
> +    temp31 = (tempL >> 31) & 0x01;
> +    temp32 = (tempL >> 32) & 0x01;
> +
> +    if (temp31 != temp32) {
> +        set_DSPControl_overflow_flag(env, 1, 20);
> +    }
> +
> +    rd = tempL & MIPSDSP_LLO;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_modsub(target_ulong rs, target_ulong rt)
> +{
> +    int32_t decr;
> +    uint16_t lastindex;
> +    target_ulong rd;
> +
> +    decr = rt & MIPSDSP_Q0;
> +    lastindex = (rt >> 8) & MIPSDSP_LO;
> +
> +    if ((rs & MIPSDSP_LLO) == 0x00000000) {
> +        rd = (target_ulong)lastindex;
> +    } else {
> +        rd = rs - decr;
> +    }
> +
> +    return rd;
> +}
> +
> +target_ulong helper_raddu_w_qb(target_ulong rs)
> +{
> +    uint8_t  rs3, rs2, rs1, rs0;
> +    uint16_t temp;
> +    uint32_t rd;
> +
> +    rs3 = (rs & MIPSDSP_Q3) >> 24;
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs1 = (rs & MIPSDSP_Q1) >>  8;
> +    rs0 =  rs & MIPSDSP_Q0;
> +
> +    temp = (uint16_t)rs3 + (uint16_t)rs2 + (uint16_t)rs1 + (uint16_t)rs0;
> +    rd = temp;
> +
> +    return (target_ulong)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_raddu_l_ob(target_ulong rs)
> +{
> +    int i;
> +    uint16_t rs_t[8];
> +    uint64_t temp;
> +
> +    temp = 0;
> +
> +    for (i = 0; i < 8; i++) {
> +        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
> +        temp += (uint64_t)rs_t[i];
> +    }
> +
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_absq_s_qb(CPUMIPSState *env, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int8_t tempD, tempC, tempB, tempA;
> +
> +    tempD = (rt & MIPSDSP_Q3) >> 24;
> +    tempC = (rt & MIPSDSP_Q2) >> 16;
> +    tempB = (rt & MIPSDSP_Q1) >>  8;
> +    tempA =  rt & MIPSDSP_Q0;
> +
> +    rd = (((uint32_t)mipsdsp_sat_abs_u8 (env, tempD) << 24) & MIPSDSP_Q3) |
> +         (((uint32_t)mipsdsp_sat_abs_u8 (env, tempC) << 16) & MIPSDSP_Q2) |
> +         (((uint32_t)mipsdsp_sat_abs_u8 (env, tempB) <<  8) & MIPSDSP_Q1) |
> +         ((uint32_t)mipsdsp_sat_abs_u8 (env, tempA) & MIPSDSP_Q0);
> +
> +    return (target_ulong)rd;
> +}
> +
> +target_ulong helper_absq_s_ph(CPUMIPSState *env, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int16_t tempA, tempB;
> +
> +    tempA = (rt & MIPSDSP_HI) >> 16;
> +    tempB =  rt & MIPSDSP_LO;
> +
> +    rd = ((uint32_t)(uint16_t)mipsdsp_sat_abs_u16 (env, tempA) << 16) |
> +        ((uint32_t)((uint16_t)mipsdsp_sat_abs_u16 (env, tempB)) & 0xFFFF);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_absq_s_w(CPUMIPSState *env, target_ulong rt)
> +{
> +    uint32_t rd;
> +    int32_t temp;
> +
> +    temp = rt;
> +    rd = mipsdsp_sat_abs_u32(env, temp);
> +
> +    return (target_ulong)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_absq_s_ob(CPUMIPSState *env, target_ulong rt)
> +{
> +    int8_t tempH, tempG, tempF, tempE;
> +    int8_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    tempH = (rt >> 56) & MIPSDSP_Q0;
> +    tempG = (rt >> 48) & MIPSDSP_Q0;
> +    tempF = (rt >> 40) & MIPSDSP_Q0;
> +    tempE = (rt >> 32) & MIPSDSP_Q0;
> +    tempD = (rt >> 24) & MIPSDSP_Q0;
> +    tempC = (rt >> 16) & MIPSDSP_Q0;
> +    tempB = (rt >> 8) & MIPSDSP_Q0;
> +    tempA = rt & MIPSDSP_Q0;
> +
> +    tempH = mipsdsp_sat_abs_u8(env, tempH);
> +    tempG = mipsdsp_sat_abs_u8(env, tempG);
> +    tempF = mipsdsp_sat_abs_u8(env, tempF);
> +    tempE = mipsdsp_sat_abs_u8(env, tempE);
> +    tempD = mipsdsp_sat_abs_u8(env, tempD);
> +    tempC = mipsdsp_sat_abs_u8(env, tempC);
> +    tempB = mipsdsp_sat_abs_u8(env, tempB);
> +    tempA = mipsdsp_sat_abs_u8(env, tempA);
> +
> +    temp = ((uint64_t)(uint8_t)tempH << 56) | ((uint64_t)(uint8_t)tempG << 48) |
> +        ((uint64_t)(uint8_t)tempF << 40) | ((uint64_t)(uint8_t)tempE << 32) |
> +        ((uint64_t)(uint8_t)tempD << 24) | ((uint64_t)(uint8_t)tempC << 16) |
> +        ((uint64_t)(uint8_t)tempB << 8) | (uint64_t)(uint8_t)tempA;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_absq_s_qh(CPUMIPSState *env, target_ulong rt)
> +{
> +    int16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    tempD = (rt >> 48) & MIPSDSP_LO;
> +    tempC = (rt >> 32) & MIPSDSP_LO;
> +    tempB = (rt >> 16) & MIPSDSP_LO;
> +    tempA = rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sat_abs_u16(env, tempD);
> +    tempC = mipsdsp_sat_abs_u16(env, tempC);
> +    tempB = mipsdsp_sat_abs_u16(env, tempB);
> +    tempA = mipsdsp_sat_abs_u16(env, tempA);
> +
> +    temp = ((uint64_t)(uint16_t)tempD << 48) | \
> +           ((uint64_t)(uint16_t)tempC << 32) | \
> +           ((uint64_t)(uint16_t)tempB << 16) | \
> +           (uint64_t)(uint16_t)tempA;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_absq_s_pw(CPUMIPSState *env, target_ulong rt)
> +{
> +    int32_t tempB, tempA;
> +    uint64_t temp;
> +
> +    tempB = (rt >> 32) & MIPSDSP_LLO;
> +    tempA = rt & MIPSDSP_LLO;
> +
> +    tempB = mipsdsp_sat_abs_u32(env, tempB);
> +    tempA = mipsdsp_sat_abs_u32(env, tempA);
> +
> +    temp = ((uint64_t)(uint32_t)tempB << 32) | (uint64_t)(uint32_t)tempA;
> +
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_precr_qb_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t  rs2, rs0, rt2, rt0;
> +    uint32_t rd;
> +
> +    rs2 = (rs & MIPSDSP_Q2) >> 16;
> +    rs0 =  rs & MIPSDSP_Q0;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rd = ((uint32_t)rs2 << 24) | ((uint32_t)rs0 << 16) |
> +         ((uint32_t)rt2 <<  8) | (uint32_t)rt0;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_precrq_qb_ph(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t tempD, tempC, tempB, tempA;
> +    uint32_t rd;
> +
> +    tempD = (rs & MIPSDSP_Q3) >> 24;
> +    tempC = (rs & MIPSDSP_Q1) >>  8;
> +    tempB = (rt & MIPSDSP_Q3) >> 24;
> +    tempA = (rt & MIPSDSP_Q1) >>  8;
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_precr_sra_ph_w(uint32_t sa, target_ulong rs,
> +                                   target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +
> +    tempB = ((int32_t)rt >> sa) & MIPSDSP_LO;
> +    tempA = ((int32_t)rs >> sa) & MIPSDSP_LO;
> +    rt = ((uint32_t)tempB << 16) | ((uint32_t)tempA & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rt;
> +}
> +
> +target_ulong helper_precr_sra_r_ph_w(uint32_t sa,
> +                                     target_ulong rs, target_ulong rt)
> +{
> +    uint64_t tempB, tempA;
> +
> +    /* If sa = 0, then (sa - 1) = -1 will case shift error, so we need else. */
> +    if (sa == 0) {
> +        tempB = (rt & MIPSDSP_LO) << 1;
> +        tempA = (rs & MIPSDSP_LO) << 1;
> +    } else {
> +        tempB = ((int32_t)rt >> (sa - 1)) + 1;
> +        tempA = ((int32_t)rs >> (sa - 1)) + 1;
> +    }
> +    rt = (((tempB >> 1) & MIPSDSP_LO) << 16) | ((tempA >> 1) & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rt;
> +}
> +
> +target_ulong helper_precrq_ph_w(target_ulong rs, target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    tempB = (rs & MIPSDSP_HI) >> 16;
> +    tempA = (rt & MIPSDSP_HI) >> 16;
> +    rd = ((uint32_t)tempB << 16) | ((uint32_t)tempA & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_precrq_rs_ph_w(CPUMIPSState *env,
> +                                   target_ulong rs, target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t rd;
> +
> +    tempB = mipsdsp_trunc16_sat16_round(env, rs);
> +    tempA = mipsdsp_trunc16_sat16_round(env, rt);
> +    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_precr_ob_qh(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs6, rs4, rs2, rs0;
> +    uint8_t rt6, rt4, rt2, rt0;
> +    uint64_t temp;
> +
> +    rs6 = (rs >> 48) & MIPSDSP_Q0;
> +    rs4 = (rs >> 32) & MIPSDSP_Q0;
> +    rs2 = (rs >> 16) & MIPSDSP_Q0;
> +    rs0 = rs & MIPSDSP_Q0;
> +    rt6 = (rt >> 48) & MIPSDSP_Q0;
> +    rt4 = (rt >> 32) & MIPSDSP_Q0;
> +    rt2 = (rt >> 16) & MIPSDSP_Q0;
> +    rt0 = rt & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)rs6 << 56) | ((uint64_t)rs4 << 48) |
> +           ((uint64_t)rs2 << 40) | ((uint64_t)rs0 << 32) |
> +           ((uint64_t)rt6 << 24) | ((uint64_t)rt4 << 16) |
> +           ((uint64_t)rt2 << 8) | (uint64_t)rt0;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_precr_sra_qh_pw(target_ulong rs, target_ulong rt,
> +                                    uint32_t sa)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t result;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    /* When sa = 0, we use rt2, rt0, rs2, rs0;
> +     * when sa != 0, we use rt3, rt1, rs3, rs1. */
> +    if (sa == 0) {
> +        tempD = rt2;
> +        tempC = rt0;
> +        tempB = rs2;
> +        tempA = rs0;
> +    } else {
> +        tempD = (int16_t)rt3 >> sa;
> +        tempC = (int16_t)rt1 >> sa;
> +        tempB = (int16_t)rs3 >> sa;
> +        tempA = (int16_t)rs1 >> sa;
> +    }
> +
> +    result = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +             ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return result;
> +}
> +
> +
> +target_ulong helper_precr_sra_r_qh_pw(target_ulong rs, target_ulong rt,
> +                                      uint32_t sa)
> +{
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t result;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    /* When sa = 0, we use rt2, rt0, rs2, rs0;
> +     * when sa != 0, we use rt3, rt1, rs3, rs1. */
> +    if (sa == 0) {
> +        tempD = rt2 << 1;
> +        tempC = rt0 << 1;
> +        tempB = rs2 << 1;
> +        tempA = rs0 << 1;
> +    } else {
> +        tempD = (((int16_t)rt3 >> sa) + 1) >> 1;
> +        tempC = (((int16_t)rt1 >> sa) + 1) >> 1;
> +        tempB = (((int16_t)rs3 >> sa) + 1) >> 1;
> +        tempA = (((int16_t)rs1 >> sa) + 1) >> 1;
> +    }
> +
> +    result = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +             ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return result;
> +}
> +
> +target_ulong helper_precrq_ob_qh(target_ulong rs, target_ulong rt)
> +{
> +    uint8_t rs6, rs4, rs2, rs0;
> +    uint8_t rt6, rt4, rt2, rt0;
> +    uint64_t temp;
> +
> +    rs6 = (rs >> 56) & MIPSDSP_Q0;
> +    rs4 = (rs >> 40) & MIPSDSP_Q0;
> +    rs2 = (rs >> 24) & MIPSDSP_Q0;
> +    rs0 = (rs >> 8) & MIPSDSP_Q0;
> +    rt6 = (rt >> 56) & MIPSDSP_Q0;
> +    rt4 = (rt >> 40) & MIPSDSP_Q0;
> +    rt2 = (rt >> 24) & MIPSDSP_Q0;
> +    rt0 = (rt >> 8) & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)rs6 << 56) | ((uint64_t)rs4 << 48) |
> +           ((uint64_t)rs2 << 40) | ((uint64_t)rs0 << 32) |
> +           ((uint64_t)rt6 << 24) | ((uint64_t)rt4 << 16) |
> +           ((uint64_t)rt2 << 8) | (uint64_t)rt0;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_precrq_qh_pw(target_ulong rs, target_ulong rt)
> +{
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    tempD = (rs >> 48) & MIPSDSP_LO;
> +    tempC = (rs >> 16) & MIPSDSP_LO;
> +    tempB = (rt >> 48) & MIPSDSP_LO;
> +    tempA = (rt >> 16) & MIPSDSP_LO;
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_precrq_rs_qh_pw(CPUMIPSState *env,
> +                                    target_ulong rs, target_ulong rt)
> +{
> +    uint32_t rs2, rs0;
> +    uint32_t rt2, rt0;
> +    uint16_t tempD, tempC, tempB, tempA;
> +    uint64_t temp;
> +
> +    rs2 = (rs >> 32) & MIPSDSP_LLO;
> +    rs0 = rs & MIPSDSP_LLO;
> +    rt2 = (rt >> 32) & MIPSDSP_LLO;
> +    rt0 = rt & MIPSDSP_LLO;
> +
> +    tempD = mipsdsp_trunc16_sat16_round(env, rs2);
> +    tempC = mipsdsp_trunc16_sat16_round(env, rs0);
> +    tempB = mipsdsp_trunc16_sat16_round(env, rt2);
> +    tempA = mipsdsp_trunc16_sat16_round(env, rt0);
> +
> +    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
> +           ((uint64_t)tempB << 16) | (uint64_t)tempA;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_precrq_pw_l(target_ulong rs, target_ulong rt)
> +{
> +    uint32_t tempB, tempA;
> +    uint64_t temp;
> +
> +    tempB = (rs >> 32) & MIPSDSP_LLO;
> +    tempA = (rt >> 32) & MIPSDSP_LLO;
> +
> +    temp = ((uint64_t)tempB << 32) | (uint64_t)tempA;
> +
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_precrqu_s_qb_ph(CPUMIPSState *env,
> +                                    target_ulong rs, target_ulong rt)
> +{
> +    uint8_t  tempD, tempC, tempB, tempA;
> +    uint16_t rsh, rsl, rth, rtl;
> +    uint32_t rd;
> +
> +    rsh = (rs & MIPSDSP_HI) >> 16;
> +    rsl =  rs & MIPSDSP_LO;
> +    rth = (rt & MIPSDSP_HI) >> 16;
> +    rtl =  rt & MIPSDSP_LO;
> +
> +    tempD = mipsdsp_sat8_reduce_precision(env, rsh);
> +    tempC = mipsdsp_sat8_reduce_precision(env, rsl);
> +    tempB = mipsdsp_sat8_reduce_precision(env, rth);
> +    tempA = mipsdsp_sat8_reduce_precision(env, rtl);
> +
> +    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
> +         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_precrqu_s_ob_qh(CPUMIPSState *env,
> +                                    target_ulong rs, target_ulong rt)
> +{
> +    int i;
> +    uint16_t rs3, rs2, rs1, rs0;
> +    uint16_t rt3, rt2, rt1, rt0;
> +    uint8_t temp[8];
> +    uint64_t result;
> +
> +    result = 0;
> +
> +    rs3 = (rs >> 48) & MIPSDSP_LO;
> +    rs2 = (rs >> 32) & MIPSDSP_LO;
> +    rs1 = (rs >> 16) & MIPSDSP_LO;
> +    rs0 = rs & MIPSDSP_LO;
> +    rt3 = (rt >> 48) & MIPSDSP_LO;
> +    rt2 = (rt >> 32) & MIPSDSP_LO;
> +    rt1 = (rt >> 16) & MIPSDSP_LO;
> +    rt0 = rt & MIPSDSP_LO;
> +
> +    temp[7] = mipsdsp_sat8_reduce_precision(env, rs3);
> +    temp[6] = mipsdsp_sat8_reduce_precision(env, rs2);
> +    temp[5] = mipsdsp_sat8_reduce_precision(env, rs1);
> +    temp[4] = mipsdsp_sat8_reduce_precision(env, rs0);
> +    temp[3] = mipsdsp_sat8_reduce_precision(env, rt3);
> +    temp[2] = mipsdsp_sat8_reduce_precision(env, rt2);
> +    temp[1] = mipsdsp_sat8_reduce_precision(env, rt1);
> +    temp[0] = mipsdsp_sat8_reduce_precision(env, rt0);
> +
> +    for (i = 0; i < 8; i++) {
> +        result |= (uint64_t)temp[i] << (8 * i);
> +    }
> +
> +    return result;
> +}
> +
> +target_ulong helper_preceq_pw_qhl(target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t tempBI, tempAI;
> +
> +    tempB = (rt >> 48) & MIPSDSP_LO;
> +    tempA = (rt >> 32) & MIPSDSP_LO;
> +
> +    tempBI = (uint32_t)tempB << 16;
> +    tempAI = (uint32_t)tempA << 16;
> +
> +    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
> +}
> +
> +target_ulong helper_preceq_pw_qhr(target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t tempBI, tempAI;
> +
> +    tempB = (rt >> 16) & MIPSDSP_LO;
> +    tempA = rt & MIPSDSP_LO;
> +
> +    tempBI = (uint32_t)tempB << 16;
> +    tempAI = (uint32_t)tempA << 16;
> +
> +    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
> +}
> +
> +target_ulong helper_preceq_pw_qhla(target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t tempBI, tempAI;
> +
> +    tempB = (rt >> 48) & MIPSDSP_LO;
> +    tempA = (rt >> 16) & MIPSDSP_LO;
> +
> +    tempBI = (uint32_t)tempB << 16;
> +    tempAI = (uint32_t)tempA << 16;
> +
> +    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
> +}
> +
> +target_ulong helper_preceq_pw_qhra(target_ulong rt)
> +{
> +    uint16_t tempB, tempA;
> +    uint32_t tempBI, tempAI;
> +
> +    tempB = (rt >> 32) & MIPSDSP_LO;
> +    tempA = rt & MIPSDSP_LO;
> +
> +    tempBI = (uint32_t)tempB << 16;
> +    tempAI = (uint32_t)tempA << 16;
> +
> +    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
> +}
> +
> +#endif
> +
> +target_ulong helper_precequ_ph_qbl(target_ulong rt)
> +{
> +    uint8_t  rt3, rt2;
> +    uint16_t tempB, tempA;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +
> +    tempB = (uint16_t)rt3 << 7;
> +    tempA = (uint16_t)rt2 << 7;
> +
> +    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
> +}
> +
> +target_ulong helper_precequ_ph_qbr(target_ulong rt)
> +{
> +    uint8_t  rt1, rt0;
> +    uint16_t tempB, tempA;
> +
> +    rt1 = (rt & MIPSDSP_Q1) >> 8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +    tempB = (uint16_t)rt1 << 7;
> +    tempA = (uint16_t)rt0 << 7;
> +
> +    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
> +}
> +
> +target_ulong helper_precequ_ph_qbla(target_ulong rt)
> +{
> +    uint8_t  rt3, rt1;
> +    uint16_t tempB, tempA;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +
> +    tempB = (uint16_t)rt3 << 7;
> +    tempA = (uint16_t)rt1 << 7;
> +
> +    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
> +}
> +
> +target_ulong helper_precequ_ph_qbra(target_ulong rt)
> +{
> +    uint8_t  rt2, rt0;
> +    uint16_t tempB, tempA;
> +
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt0 =  rt & MIPSDSP_Q0;
> +    tempB = (uint16_t)rt2 << 7;
> +    tempA = (uint16_t)rt0 << 7;
> +
> +    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_precequ_qh_obl(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint16_t tempDS, tempCS, tempBS, tempAS;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 56) & MIPSDSP_Q0;
> +    tempCC = (rt >> 48) & MIPSDSP_Q0;
> +    tempBC = (rt >> 40) & MIPSDSP_Q0;
> +    tempAC = (rt >> 32) & MIPSDSP_Q0;
> +
> +    tempDS = (uint16_t)tempDC << 7;
> +    tempCS = (uint16_t)tempCC << 7;
> +    tempBS = (uint16_t)tempBC << 7;
> +    tempAS = (uint16_t)tempAC << 7;
> +
> +    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
> +           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
> +    return temp;
> +}
> +
> +target_ulong helper_precequ_qh_obr(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint16_t tempDS, tempCS, tempBS, tempAS;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 24) & MIPSDSP_Q0;
> +    tempCC = (rt >> 16) & MIPSDSP_Q0;
> +    tempBC = (rt >> 8) & MIPSDSP_Q0;
> +    tempAC = rt & MIPSDSP_Q0;
> +
> +    tempDS = (uint16_t)tempDC << 7;
> +    tempCS = (uint16_t)tempCC << 7;
> +    tempBS = (uint16_t)tempBC << 7;
> +    tempAS = (uint16_t)tempAC << 7;
> +
> +    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
> +           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
> +    return temp;
> +}
> +
> +target_ulong helper_precequ_qh_obla(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint16_t tempDS, tempCS, tempBS, tempAS;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 56) & MIPSDSP_Q0;
> +    tempCC = (rt >> 40) & MIPSDSP_Q0;
> +    tempBC = (rt >> 24) & MIPSDSP_Q0;
> +    tempAC = (rt >> 8) & MIPSDSP_Q0;
> +
> +    tempDS = (uint16_t)tempDC << 7;
> +    tempCS = (uint16_t)tempCC << 7;
> +    tempBS = (uint16_t)tempBC << 7;
> +    tempAS = (uint16_t)tempAC << 7;
> +
> +    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
> +           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
> +    return temp;
> +}
> +
> +target_ulong helper_precequ_qh_obra(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint16_t tempDS, tempCS, tempBS, tempAS;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 48) & MIPSDSP_Q0;
> +    tempCC = (rt >> 32) & MIPSDSP_Q0;
> +    tempBC = (rt >> 16) & MIPSDSP_Q0;
> +    tempAC = rt & MIPSDSP_Q0;
> +
> +    tempDS = (uint16_t)tempDC << 7;
> +    tempCS = (uint16_t)tempCC << 7;
> +    tempBS = (uint16_t)tempBC << 7;
> +    tempAS = (uint16_t)tempAC << 7;
> +
> +    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
> +           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
> +    return temp;
> +}
> +#endif
> +
> +target_ulong helper_preceu_ph_qbl(target_ulong rt)
> +{
> +    uint8_t  rt3, rt2;
> +    uint32_t rd;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rd = ((uint32_t)(uint16_t)rt3 << 16) | \
> +         ((uint32_t)(uint16_t)rt2 & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_preceu_ph_qbr(target_ulong rt)
> +{
> +    uint8_t  rt1, rt0;
> +    uint32_t rd;
> +
> +    rt1 = (rt & MIPSDSP_Q1) >> 8;
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rd = ((uint32_t)(uint16_t)rt1 << 16) | \
> +         ((uint32_t)(uint16_t)rt0 & MIPSDSP_LO);
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_preceu_ph_qbla(target_ulong rt)
> +{
> +    uint8_t  rt3, rt1;
> +    uint32_t rd;
> +
> +    rt3 = (rt & MIPSDSP_Q3) >> 24;
> +    rt1 = (rt & MIPSDSP_Q1) >>  8;
> +    rd = ((uint32_t)(uint16_t)rt3 << 16) | \
> +         ((uint32_t)(uint16_t)rt1 & MIPSDSP_LO);
> +
> +    return (target_long)(int32_t)rd;
> +}
> +
> +target_ulong helper_preceu_ph_qbra(target_ulong rt)
> +{
> +    uint8_t  rt2, rt0;
> +    uint32_t rd;
> +
> +    rt2 = (rt & MIPSDSP_Q2) >> 16;
> +    rt0 =  rt & MIPSDSP_Q0;
> +    rd = ((uint32_t)(uint16_t)rt2 << 16) | \
> +         ((uint32_t)(uint16_t)rt0 & MIPSDSP_LO);
> +    return (target_long)(int32_t)rd;
> +}
> +
> +#if defined(TARGET_MIPS64)
> +target_ulong helper_preceu_qh_obl(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 56) & MIPSDSP_Q0;
> +    tempCC = (rt >> 48) & MIPSDSP_Q0;
> +    tempBC = (rt >> 40) & MIPSDSP_Q0;
> +    tempAC = (rt >> 32) & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
> +           ((uint64_t)(uint16_t)tempCC << 32) | \
> +           ((uint64_t)(uint16_t)tempBC << 16) | \
> +           (uint64_t)(uint16_t)tempAC;
> +    return temp;
> +}
> +
> +target_ulong helper_preceu_qh_obr(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 24) & MIPSDSP_Q0;
> +    tempCC = (rt >> 16) & MIPSDSP_Q0;
> +    tempBC = (rt >> 8) & MIPSDSP_Q0;
> +    tempAC = rt & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
> +           ((uint64_t)(uint16_t)tempCC << 32) | \
> +           ((uint64_t)(uint16_t)tempBC << 16) | \
> +           (uint64_t)(uint16_t)tempAC;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_preceu_qh_obla(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 56) & MIPSDSP_Q0;
> +    tempCC = (rt >> 40) & MIPSDSP_Q0;
> +    tempBC = (rt >> 24) & MIPSDSP_Q0;
> +    tempAC = (rt >> 8) & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
> +           ((uint64_t)(uint16_t)tempCC << 32) | \
> +           ((uint64_t)(uint16_t)tempBC << 16) | \
> +           (uint64_t)(uint16_t)tempAC;
> +
> +    return temp;
> +}
> +
> +target_ulong helper_preceu_qh_obra(target_ulong rt)
> +{
> +    uint8_t tempDC, tempCC, tempBC, tempAC;
> +    uint64_t temp;
> +
> +    tempDC = (rt >> 48) & MIPSDSP_Q0;
> +    tempCC = (rt >> 32) & MIPSDSP_Q0;
> +    tempBC = (rt >> 16) & MIPSDSP_Q0;
> +    tempAC = rt & MIPSDSP_Q0;
> +
> +    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
> +           ((uint64_t)(uint16_t)tempCC << 32) | \
> +           ((uint64_t)(uint16_t)tempBC << 16) | \
> +           (uint64_t)(uint16_t)tempAC;
> +
> +    return temp;
> +}
> +#endif
> +
> +#undef MIPSDSP_LHI
> +#undef MIPSDSP_LLO
> +#undef MIPSDSP_HI
> +#undef MIPSDSP_LO
> +#undef MIPSDSP_Q3
> +#undef MIPSDSP_Q2
> +#undef MIPSDSP_Q1
> +#undef MIPSDSP_Q0

All this part looks fine, but there I think there is too much copy and
paste for things that basically do:

- Split the argument into n parts of m bits
- Call a function on all parts
- Reassemble all the parts

If you want you might try to create a macro doing that. With a for loop
like you did in helper_subuh_r_ob(), and passing to the macro the values
of n, m and the function, you can probably use only one macro for most
of this.

> diff --git a/target-mips/helper.h b/target-mips/helper.h
> index 76fb451..cd4564b 100644
> --- a/target-mips/helper.h
> +++ b/target-mips/helper.h
> @@ -297,4 +297,130 @@ DEF_HELPER_0(rdhwr_ccres, tl)
>  DEF_HELPER_1(pmon, void, int)
>  DEF_HELPER_0(wait, void)
>  
> +/*** MIPS DSP ***/
> +/* DSP Arithmetic Sub-class insns */
> +DEF_HELPER_FLAGS_3(addq_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addq_s_ph, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(addq_qh, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addq_s_qh, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(addq_s_w, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(addq_pw, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addq_s_pw, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(addu_qb, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_s_qb, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(adduh_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(adduh_r_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_s_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(addqh_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(addqh_r_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(addqh_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(addqh_r_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(addu_ob, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_s_ob, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(adduh_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(adduh_r_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_qh, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addu_s_qh, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(subq_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subq_s_ph, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(subq_qh, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subq_s_qh, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(subq_s_w, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(subq_pw, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subq_s_pw, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(subu_qb, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_s_qb, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(subuh_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(subuh_r_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_s_ph, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(subqh_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(subqh_r_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(subqh_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(subqh_r_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(subu_ob, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_s_ob, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(subuh_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(subuh_r_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_qh, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(subu_s_qh, 0, tl, env, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(addsc, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_3(addwc, 0, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(modsub, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_1(raddu_w_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_1(raddu_l_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_2(absq_s_qb, 0, tl, env, tl)
> +DEF_HELPER_FLAGS_2(absq_s_ph, 0, tl, env, tl)
> +DEF_HELPER_FLAGS_2(absq_s_w, 0, tl, env, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_2(absq_s_ob, 0, tl, env, tl)
> +DEF_HELPER_FLAGS_2(absq_s_qh, 0, tl, env, tl)
> +DEF_HELPER_FLAGS_2(absq_s_pw, 0, tl, env, tl)
> +#endif
> +DEF_HELPER_FLAGS_2(precr_qb_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(precrq_qb_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(precr_sra_ph_w, TCG_CALL_CONST | TCG_CALL_PURE,
> +                   tl, i32, tl, tl)
> +DEF_HELPER_FLAGS_3(precr_sra_r_ph_w, TCG_CALL_CONST | TCG_CALL_PURE,
> +                   tl, i32, tl, tl)
> +DEF_HELPER_FLAGS_2(precrq_ph_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(precrq_rs_ph_w, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_2(precr_ob_qh, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(precr_sra_qh_pw,
> +                   TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl, i32)
> +DEF_HELPER_FLAGS_3(precr_sra_r_qh_pw,
> +                   TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl, i32)
> +DEF_HELPER_FLAGS_2(precrq_ob_qh, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_2(precrq_qh_pw, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +DEF_HELPER_FLAGS_3(precrq_rs_qh_pw,
> +                   TCG_CALL_CONST | TCG_CALL_PURE, tl, env, tl, tl)
> +DEF_HELPER_FLAGS_2(precrq_pw_l, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_3(precrqu_s_qb_ph, 0, tl, env, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_3(precrqu_s_ob_qh,
> +                   TCG_CALL_CONST | TCG_CALL_PURE, tl, env, tl, tl)
> +
> +DEF_HELPER_FLAGS_1(preceq_pw_qhl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceq_pw_qhr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceq_pw_qhla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceq_pw_qhra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_1(precequ_ph_qbl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_ph_qbr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_ph_qbla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_ph_qbra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_1(precequ_qh_obl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_qh_obr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_qh_obla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(precequ_qh_obra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#endif
> +DEF_HELPER_FLAGS_1(preceu_ph_qbl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_ph_qbr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_ph_qbla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_ph_qbra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#if defined(TARGET_MIPS64)
> +DEF_HELPER_FLAGS_1(preceu_qh_obl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_qh_obr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_qh_obla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +DEF_HELPER_FLAGS_1(preceu_qh_obra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
> +#endif
> +
>  #include "def-helper.h"
> diff --git a/target-mips/translate.c b/target-mips/translate.c
> index f7bb054..407b4ad 100644
> --- a/target-mips/translate.c
> +++ b/target-mips/translate.c
> @@ -316,6 +316,21 @@ enum {
>  
>      /* MIPS DSP Load */
>      OPC_LX_DSP         = 0x0A | OPC_SPECIAL3,
> +    /* MIPS DSP Arithmetic */
> +    OPC_ADDU_QB_DSP    = 0x10 | OPC_SPECIAL3,
> +#if defined(TARGET_MIPS64)
> +    OPC_ADDU_OB_DSP    = 0x14 | OPC_SPECIAL3,
> +#endif
> +    OPC_ABSQ_S_PH_DSP  = 0x12 | OPC_SPECIAL3,
> +#if defined(TARGET_MIPS64)
> +    OPC_ABSQ_S_QH_DSP  = 0x16 | OPC_SPECIAL3,
> +#endif
> +    /* OPC_ADDUH_QB_DSP is same as OPC_MULT_G_2E.  */
> +    /* OPC_ADDUH_QB_DSP   = 0x18 | OPC_SPECIAL3,  */
> +    OPC_CMPU_EQ_QB_DSP = 0x11 | OPC_SPECIAL3,
> +#if defined(TARGET_MIPS64)
> +    OPC_CMPU_EQ_OB_DSP = 0x15 | OPC_SPECIAL3,
> +#endif
>  };
>  
>  /* BSHFL opcodes */
> @@ -354,6 +369,144 @@ enum {
>  #endif
>  };
>  
> +#define MASK_ADDU_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_ADDQ_PH        = (0x0A << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDQ_S_PH      = (0x0E << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDQ_S_W       = (0x16 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDU_QB        = (0x00 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDU_S_QB      = (0x04 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDU_PH        = (0x08 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDU_S_PH      = (0x0C << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBQ_PH        = (0x0B << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBQ_S_PH      = (0x0F << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBQ_S_W       = (0x17 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBU_QB        = (0x01 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBU_S_QB      = (0x05 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBU_PH        = (0x09 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_SUBU_S_PH      = (0x0D << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDSC          = (0x10 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_ADDWC          = (0x11 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_MODSUB         = (0x12 << 6) | OPC_ADDU_QB_DSP,
> +    OPC_RADDU_W_QB     = (0x14 << 6) | OPC_ADDU_QB_DSP,
> +};
> +
> +#define OPC_ADDUH_QB_DSP OPC_MULT_G_2E
> +#define MASK_ADDUH_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_ADDUH_QB   = (0x00 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_ADDUH_R_QB = (0x02 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_ADDQH_PH   = (0x08 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_ADDQH_R_PH = (0x0A << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_ADDQH_W    = (0x10 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_ADDQH_R_W  = (0x12 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBUH_QB   = (0x01 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBUH_R_QB = (0x03 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBQH_PH   = (0x09 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBQH_R_PH = (0x0B << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBQH_W    = (0x11 << 6) | OPC_ADDUH_QB_DSP,
> +    OPC_SUBQH_R_W  = (0x13 << 6) | OPC_ADDUH_QB_DSP,
> +};
> +
> +#define MASK_ABSQ_S_PH(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_ABSQ_S_QB       = (0x01 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_ABSQ_S_PH       = (0x09 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_ABSQ_S_W        = (0x11 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQ_W_PHL    = (0x0C << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQ_W_PHR    = (0x0D << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQU_PH_QBL  = (0x04 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQU_PH_QBR  = (0x05 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQU_PH_QBLA = (0x06 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEQU_PH_QBRA = (0x07 << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEU_PH_QBL   = (0x1C << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEU_PH_QBR   = (0x1D << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEU_PH_QBLA  = (0x1E << 6) | OPC_ABSQ_S_PH_DSP,
> +    OPC_PRECEU_PH_QBRA  = (0x1F << 6) | OPC_ABSQ_S_PH_DSP,
> +};
> +
> +#define MASK_CMPU_EQ_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_PRECR_QB_PH      = (0x0D << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECRQ_QB_PH     = (0x0C << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECR_SRA_PH_W   = (0x1E << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECR_SRA_R_PH_W = (0x1F << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECRQ_PH_W      = (0x14 << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECRQ_RS_PH_W   = (0x15 << 6) | OPC_CMPU_EQ_QB_DSP,
> +    OPC_PRECRQU_S_QB_PH  = (0x0F << 6) | OPC_CMPU_EQ_QB_DSP,
> +};
> +
> +#if defined(TARGET_MIPS64)
> +#define MASK_ABSQ_S_QH(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_PRECEQ_L_PWL    = (0x14 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQ_L_PWR    = (0x15 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQ_PW_QHL   = (0x0C << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQ_PW_QHR   = (0x0D << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQ_PW_QHLA  = (0x0E << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQ_PW_QHRA  = (0x0F << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQU_QH_OBL  = (0x04 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQU_QH_OBR  = (0x05 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQU_QH_OBLA = (0x06 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEQU_QH_OBRA = (0x07 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEU_QH_OBL   = (0x1C << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEU_QH_OBR   = (0x1D << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEU_QH_OBLA  = (0x1E << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_PRECEU_QH_OBRA  = (0x1F << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_ABSQ_S_OB       = (0x01 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_ABSQ_S_PW       = (0x11 << 6) | OPC_ABSQ_S_QH_DSP,
> +    OPC_ABSQ_S_QH       = (0x09 << 6) | OPC_ABSQ_S_QH_DSP,
> +};
> +#endif
> +
> +#if defined(TARGET_MIPS64)
> +#define MASK_ADDU_OB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_RADDU_L_OB     = (0x14 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBQ_PW        = (0x13 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBQ_S_PW      = (0x17 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBQ_QH        = (0x0B << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBQ_S_QH      = (0x0F << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBU_OB        = (0x01 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBU_S_OB      = (0x05 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBU_QH        = (0x09 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBU_S_QH      = (0x0D << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBUH_OB       = (0x19 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_SUBUH_R_OB     = (0x1B << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDQ_PW        = (0x12 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDQ_S_PW      = (0x16 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDQ_QH        = (0x0A << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDQ_S_QH      = (0x0E << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDU_OB        = (0x00 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDU_S_OB      = (0x04 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDU_QH        = (0x08 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDU_S_QH      = (0x0C << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDUH_OB       = (0x18 << 6) | OPC_ADDU_OB_DSP,
> +    OPC_ADDUH_R_OB     = (0x1A << 6) | OPC_ADDU_OB_DSP,
> +};
> +#endif
> +
> +#if defined(TARGET_MIPS64)
> +#define MASK_CMPU_EQ_OB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
> +enum {
> +    /* MIPS DSP Arithmetic Sub-class */
> +    OPC_PRECR_OB_QH       = (0x0D << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECR_SRA_QH_PW   = (0x1E << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECR_SRA_R_QH_PW = (0x1F << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECRQ_OB_QH      = (0x0C << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECRQ_PW_L       = (0x1C << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECRQ_QH_PW      = (0x14 << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECRQ_RS_QH_PW   = (0x15 << 6) | OPC_CMPU_EQ_OB_DSP,
> +    OPC_PRECRQU_S_OB_QH   = (0x0F << 6) | OPC_CMPU_EQ_OB_DSP,
> +};
> +#endif
> +
>  /* Coprocessor 0 (rs field) */
>  #define MASK_CP0(op)       MASK_OP_MAJOR(op) | (op & (0x1F << 21))
>  
> @@ -12183,10 +12336,64 @@ static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
>              }
>              break;
>          case OPC_DIV_G_2E ... OPC_DIVU_G_2E:
> -        case OPC_MULT_G_2E ... OPC_MULTU_G_2E:
>          case OPC_MOD_G_2E ... OPC_MODU_G_2E:
> -            check_insn(env, ctx, INSN_LOONGSON2E);
> -            gen_loongson_integer(ctx, op1, rd, rs, rt);
> +        case OPC_MULT_G_2E ... OPC_MULTU_G_2E:
> +        /* OPC_MULT_G_2E, OPC_ADDUH_QB_DSP, OPC_MUL_PH_DSP have
> +         * the same mask and op1. */
> +            if ((env->insn_flags & ASE_DSPR2) && (op1 == OPC_MULT_G_2E)) {
> +                check_dspr2(ctx);
> +                op2 = MASK_ADDUH_QB(ctx->opcode);
> +                switch (op2) {
> +                case OPC_ADDUH_QB:
> +                    gen_helper_adduh_qb(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_ADDUH_R_QB:
> +                    gen_helper_adduh_r_qb(cpu_gpr[rd],
> +                                          cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_ADDQH_PH:
> +                    gen_helper_addqh_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_ADDQH_R_PH:
> +                    gen_helper_addqh_r_ph(cpu_gpr[rd], cpu_gpr[rs],
> +                                          cpu_gpr[rt]);
> +                    break;
> +                case OPC_ADDQH_W:
> +                    gen_helper_addqh_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_ADDQH_R_W:
> +                    gen_helper_addqh_r_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBUH_QB:
> +                    gen_helper_subuh_qb(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBUH_R_QB:
> +                    gen_helper_subuh_r_qb(cpu_gpr[rd],
> +                                          cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBQH_PH:
> +                    gen_helper_subqh_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBQH_R_PH:
> +                    gen_helper_subqh_r_ph(cpu_gpr[rd],
> +                                          cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBQH_W:
> +                    gen_helper_subqh_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                case OPC_SUBQH_R_W:
> +                    gen_helper_subqh_r_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                    break;
> +                default:
> +                    MIPS_INVAL("MASK ADDUH.QB");
> +                    generate_exception(ctx, EXCP_RI);
> +                    break;
> +                }
> +            } else if (env->insn_flags & INSN_LOONGSON2E) {
> +                gen_loongson_integer(ctx, op1, rd, rs, rt);
> +            } else {
> +                generate_exception(ctx, EXCP_RI);
> +            }
>              break;
>          case OPC_LX_DSP:
>              check_dsp(ctx);
> @@ -12240,6 +12447,224 @@ static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
>                  break;
>              }
>              break;
> +        case OPC_ABSQ_S_PH_DSP:
> +            op2 = MASK_ABSQ_S_PH(ctx->opcode);
> +            switch (op2) {
> +            case OPC_ABSQ_S_QB:
> +                check_dspr2(ctx);
> +                gen_helper_absq_s_qb(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            case OPC_ABSQ_S_PH:
> +                check_dsp(ctx);
> +                gen_helper_absq_s_ph(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            case OPC_ABSQ_S_W:
> +                check_dsp(ctx);
> +                gen_helper_absq_s_w(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQ_W_PHL:
> +                {
> +                    TCGv temp = tcg_temp_new();;
> +                    check_dsp(ctx);
> +                    tcg_gen_andi_tl(temp, cpu_gpr[rt], 0xFFFF0000);
> +                    tcg_gen_ext32s_tl(cpu_gpr[rd], temp);
> +                    tcg_temp_free(temp);
> +                    break;
> +                }
> +            case OPC_PRECEQ_W_PHR:
> +                {
> +                    TCGv temp = tcg_temp_new();;
> +                    check_dsp(ctx);
> +                    tcg_gen_andi_tl(temp, cpu_gpr[rt], 0x0000FFFF);
> +                    tcg_gen_shli_tl(temp, temp, 16);
> +                    tcg_gen_ext32s_tl(cpu_gpr[rd], temp);
> +                    tcg_temp_free(temp);
> +                    break;
> +                }
> +            case OPC_PRECEQU_PH_QBL:
> +                check_dsp(ctx);
> +                gen_helper_precequ_ph_qbl(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_PH_QBR:
> +                check_dsp(ctx);
> +                gen_helper_precequ_ph_qbr(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_PH_QBLA:
> +                check_dsp(ctx);
> +                gen_helper_precequ_ph_qbla(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_PH_QBRA:
> +                check_dsp(ctx);
> +                gen_helper_precequ_ph_qbra(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_PH_QBL:
> +                check_dsp(ctx);
> +                gen_helper_preceu_ph_qbl(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_PH_QBR:
> +                check_dsp(ctx);
> +                gen_helper_preceu_ph_qbr(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_PH_QBLA:
> +                check_dsp(ctx);
> +                gen_helper_preceu_ph_qbla(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_PH_QBRA:
> +                check_dsp(ctx);
> +                gen_helper_preceu_ph_qbra(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK ABSQ_S.PH");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;
> +        case OPC_ADDU_QB_DSP:
> +            op2 = MASK_ADDU_QB(ctx->opcode);
> +            switch (op2) {
> +            case OPC_ADDQ_PH:
> +                check_dsp(ctx);
> +                gen_helper_addq_ph(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_S_PH:
> +                check_dsp(ctx);
> +                gen_helper_addq_s_ph(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_S_W:
> +                check_dsp(ctx);
> +                gen_helper_addq_s_w(cpu_gpr[rd], cpu_env,
> +                                    cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_QB:
> +                check_dsp(ctx);
> +                gen_helper_addu_qb(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_S_QB:
> +                check_dsp(ctx);
> +                gen_helper_addu_s_qb(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_PH:
> +                check_dspr2(ctx);
> +                gen_helper_addu_ph(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_S_PH:
> +                check_dspr2(ctx);
> +                gen_helper_addu_s_ph(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_PH:
> +                check_dsp(ctx);
> +                gen_helper_subq_ph(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_S_PH:
> +                check_dsp(ctx);
> +                gen_helper_subq_s_ph(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_S_W:
> +                check_dsp(ctx);
> +                gen_helper_subq_s_w(cpu_gpr[rd], cpu_env,
> +                                    cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_QB:
> +                check_dsp(ctx);
> +                gen_helper_subu_qb(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_S_QB:
> +                check_dsp(ctx);
> +                gen_helper_subu_s_qb(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_PH:
> +                check_dspr2(ctx);
> +                gen_helper_subu_ph(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_S_PH:
> +                check_dspr2(ctx);
> +                gen_helper_subu_s_ph(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDSC:
> +                check_dsp(ctx);
> +                gen_helper_addsc(cpu_gpr[rd], cpu_env,
> +                                 cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDWC:
> +                check_dsp(ctx);
> +                gen_helper_addwc(cpu_gpr[rd], cpu_env,
> +                                 cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_MODSUB:
> +                check_dsp(ctx);
> +                gen_helper_modsub(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_RADDU_W_QB:
> +                check_dsp(ctx);
> +                gen_helper_raddu_w_qb(cpu_gpr[rd], cpu_gpr[rs]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK ADDU.QB");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;
> +        case OPC_CMPU_EQ_QB_DSP:
> +            op2 = MASK_CMPU_EQ_QB(ctx->opcode);
> +            switch (op2) {
> +            case OPC_PRECR_QB_PH:
> +                check_dspr2(ctx);
> +                gen_helper_precr_qb_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQ_QB_PH:
> +                check_dsp(ctx);
> +                gen_helper_precrq_qb_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECR_SRA_PH_W:
> +                check_dspr2(ctx);
> +                {
> +                    TCGv_i32 temp_rd = tcg_const_i32(rd);
> +                    gen_helper_precr_sra_ph_w(cpu_gpr[rt], temp_rd,
> +                                           cpu_gpr[rs], cpu_gpr[rt]);
> +                    tcg_temp_free_i32(temp_rd);
> +                    break;
> +                }
> +            case OPC_PRECR_SRA_R_PH_W:
> +                check_dspr2(ctx);
> +                {
> +                    TCGv_i32 temp_rd = tcg_const_i32(rd);
> +                    gen_helper_precr_sra_r_ph_w(cpu_gpr[rt], temp_rd,
> +                                            cpu_gpr[rs], cpu_gpr[rt]);
> +                    tcg_temp_free_i32(temp_rd);
> +                    break;
> +                }
> +            case OPC_PRECRQ_PH_W:
> +                check_dsp(ctx);
> +                gen_helper_precrq_ph_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQ_RS_PH_W:
> +                check_dsp(ctx);
> +                gen_helper_precrq_rs_ph_w(cpu_gpr[rd], cpu_env,
> +                                          cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQU_S_QB_PH:
> +                check_dsp(ctx);
> +                gen_helper_precrqu_s_qb_ph(cpu_gpr[rd], cpu_env,
> +                                           cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK CMPU.EQ.QB");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;
>  #if defined(TARGET_MIPS64)
>          case OPC_DEXTM ... OPC_DEXT:
>          case OPC_DINSM ... OPC_DINS:
> @@ -12260,6 +12685,256 @@ static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
>              gen_loongson_integer(ctx, op1, rd, rs, rt);
>              break;
>  #endif
> +#if defined(TARGET_MIPS64)
> +        case OPC_ABSQ_S_QH_DSP:
> +            op2 = MASK_ABSQ_S_QH(ctx->opcode);
> +            switch (op2) {
> +            case OPC_PRECEQ_L_PWL:
> +                check_dsp(ctx);
> +                tcg_gen_andi_tl(cpu_gpr[rd], cpu_gpr[rt], \
> +                                0xFFFFFFFF00000000ull);
> +                break;
> +            case OPC_PRECEQ_L_PWR:
> +                check_dsp(ctx);
> +                tcg_gen_andi_tl(cpu_gpr[rd], cpu_gpr[rt], 0xFFFFFFFFull);
> +                break;
> +            case OPC_PRECEQ_PW_QHL:
> +                check_dsp(ctx);
> +                gen_helper_preceq_pw_qhl(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQ_PW_QHR:
> +                check_dsp(ctx);
> +                gen_helper_preceq_pw_qhr(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQ_PW_QHLA:
> +                check_dsp(ctx);
> +                gen_helper_preceq_pw_qhla(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQ_PW_QHRA:
> +                check_dsp(ctx);
> +                gen_helper_preceq_pw_qhra(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_QH_OBL:
> +                check_dsp(ctx);
> +                gen_helper_precequ_qh_obl(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_QH_OBR:
> +                check_dsp(ctx);
> +                gen_helper_precequ_qh_obr(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_QH_OBLA:
> +                check_dsp(ctx);
> +                gen_helper_precequ_qh_obla(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEQU_QH_OBRA:
> +                check_dsp(ctx);
> +                gen_helper_precequ_qh_obra(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_QH_OBL:
> +                check_dsp(ctx);
> +                gen_helper_preceu_qh_obl(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_QH_OBR:
> +                check_dsp(ctx);
> +                gen_helper_preceu_qh_obr(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_QH_OBLA:
> +                check_dsp(ctx);
> +                gen_helper_preceu_qh_obla(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECEU_QH_OBRA:
> +                check_dsp(ctx);
> +                gen_helper_preceu_qh_obra(cpu_gpr[rd], cpu_gpr[rt]);
> +                break;
> +            case OPC_ABSQ_S_OB:
> +                check_dspr2(ctx);
> +                gen_helper_absq_s_ob(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            case OPC_ABSQ_S_PW:
> +                check_dsp(ctx);
> +                gen_helper_absq_s_pw(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            case OPC_ABSQ_S_QH:
> +                check_dsp(ctx);
> +                gen_helper_absq_s_qh(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK ABSQ_S.QH");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;
> +#endif
> +#if defined(TARGET_MIPS64)
> +        case OPC_ADDU_OB_DSP:
> +            op2 = MASK_ADDU_OB(ctx->opcode);
> +            switch (op2) {
> +            case OPC_RADDU_L_OB:
> +                check_dsp(ctx);
> +                gen_helper_raddu_l_ob(cpu_gpr[rd], cpu_gpr[rs]);
> +                break;
> +            case OPC_SUBQ_PW:
> +                check_dsp(ctx);
> +                gen_helper_subq_pw(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_S_PW:
> +                check_dsp(ctx);
> +                gen_helper_subq_s_pw(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_QH:
> +                check_dsp(ctx);
> +                gen_helper_subq_qh(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBQ_S_QH:
> +                check_dsp(ctx);
> +                gen_helper_subq_s_qh(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_OB:
> +                check_dsp(ctx);
> +                gen_helper_subu_ob(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_S_OB:
> +                check_dsp(ctx);
> +                gen_helper_subu_s_ob(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_QH:
> +                check_dspr2(ctx);
> +                gen_helper_subu_qh(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBU_S_QH:
> +                check_dspr2(ctx);
> +                gen_helper_subu_s_qh(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBUH_OB:
> +                check_dspr2(ctx);
> +                gen_helper_subuh_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_SUBUH_R_OB:
> +                check_dspr2(ctx);
> +                gen_helper_subuh_r_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_PW:
> +                check_dsp(ctx);
> +                gen_helper_addq_pw(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_S_PW:
> +                check_dsp(ctx);
> +                gen_helper_addq_s_pw(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_QH:
> +                check_dsp(ctx);
> +                gen_helper_addq_qh(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDQ_S_QH:
> +                check_dsp(ctx);
> +                gen_helper_addq_s_qh(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_OB:
> +                check_dsp(ctx);
> +                gen_helper_addu_ob(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_S_OB:
> +                check_dsp(ctx);
> +                gen_helper_addu_s_ob(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_QH:
> +                check_dspr2(ctx);
> +                gen_helper_addu_qh(cpu_gpr[rd], cpu_env,
> +                                   cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDU_S_QH:
> +                check_dspr2(ctx);
> +                gen_helper_addu_s_qh(cpu_gpr[rd], cpu_env,
> +                                     cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDUH_OB:
> +                check_dspr2(ctx);
> +                gen_helper_adduh_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_ADDUH_R_OB:
> +                check_dspr2(ctx);
> +                gen_helper_adduh_r_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK ADDU.OB");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;
> +#endif
> +#if defined(TARGET_MIPS64)
> +        case OPC_CMPU_EQ_OB_DSP:
> +            op2 = MASK_CMPU_EQ_OB(ctx->opcode);
> +            switch (op2) {
> +            case OPC_PRECR_OB_QH:
> +                check_dspr2(ctx);
> +                gen_helper_precr_ob_qh(cpu_gpr[rd], cpu_gpr[rs],
> +                                       cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECR_SRA_QH_PW:
> +                check_dspr2(ctx);
> +                {
> +                    TCGv_i32 sa_v = tcg_const_i32(rd);
> +                    gen_helper_precr_sra_qh_pw(cpu_gpr[rt], cpu_gpr[rs],
> +                                               cpu_gpr[rt], sa_v);
> +                    tcg_temp_free_i32(sa_v);
> +                    break;
> +                }
> +            case OPC_PRECR_SRA_R_QH_PW:
> +                check_dspr2(ctx);
> +                {
> +                    TCGv_i32 sa_v = tcg_const_i32(rd);
> +                    gen_helper_precr_sra_r_qh_pw(cpu_gpr[rt], cpu_gpr[rs],
> +                                                 cpu_gpr[rt], sa_v);
> +                    tcg_temp_free_i32(sa_v);
> +                    break;
> +                }
> +            case OPC_PRECRQ_OB_QH:
> +                check_dsp(ctx);
> +                gen_helper_precrq_ob_qh(cpu_gpr[rd], cpu_gpr[rs],
> +                                        cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQ_PW_L:
> +                check_dsp(ctx);
> +                gen_helper_precrq_pw_l(cpu_gpr[rd], cpu_gpr[rs],
> +                                       cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQ_QH_PW:
> +                check_dsp(ctx);
> +                gen_helper_precrq_qh_pw(cpu_gpr[rd], cpu_gpr[rs],
> +                                        cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQ_RS_QH_PW:
> +                check_dsp(ctx);
> +                gen_helper_precrq_rs_qh_pw(cpu_gpr[rd], cpu_env,
> +                                           cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            case OPC_PRECRQU_S_OB_QH:
> +                check_dsp(ctx);
> +                gen_helper_precrqu_s_ob_qh(cpu_gpr[rd], cpu_env,
> +                                           cpu_gpr[rs], cpu_gpr[rt]);
> +                break;
> +            default:            /* Invalid */
> +                MIPS_INVAL("MASK CMPU.EQ.OB");
> +                generate_exception(ctx, EXCP_RI);
> +                break;
> +            }
> +            break;

All these functions should check for rd, rs and rt being 0. For that I
think it's better to call a function with the needed arguments (like for 
example gen_loongson_integer()) and move all that code there. That way
you can add the check for rd, rs, rt in the common part, as well as the
check_dsp part.

> +#endif
>          default:            /* Invalid */
>              MIPS_INVAL("special3");
>              generate_exception(ctx, EXCP_RI);
> -- 
> 1.7.9.5
> 
> 
>

Patch

diff --git a/target-mips/dsp_helper.c b/target-mips/dsp_helper.c
index 779f5c7..4f7cc9e 100644
--- a/target-mips/dsp_helper.c
+++ b/target-mips/dsp_helper.c
@@ -1098,3 +1098,1836 @@  static inline int32_t mipsdsp_cmp_lt(uint32_t a, uint32_t b)
     return a < b;
 }
 /*** MIPS DSP internal functions end ***/
+
+#define MIPSDSP_LHI 0xFFFFFFFF00000000ull
+#define MIPSDSP_LLO 0x00000000FFFFFFFFull
+#define MIPSDSP_HI  0xFFFF0000
+#define MIPSDSP_LO  0x0000FFFF
+#define MIPSDSP_Q3  0xFF000000
+#define MIPSDSP_Q2  0x00FF0000
+#define MIPSDSP_Q1  0x0000FF00
+#define MIPSDSP_Q0  0x000000FF
+
+/** DSP Arithmetic Sub-class insns **/
+target_ulong helper_addq_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    int16_t  rsh, rsl, rth, rtl, temph, templ;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    temph = mipsdsp_add_i16(env, rsh, rth);
+    templ = mipsdsp_add_i16(env, rsl, rtl);
+
+    return (target_long)(int32_t)(((unsigned int)temph << 16) \
+                                  | ((unsigned int)templ & 0xFFFF));
+}
+
+target_ulong helper_addq_s_ph(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    int16_t rsh, rsl, rth, rtl, temph, templ;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    temph = mipsdsp_sat_add_i16(env, rsh, rth);
+    templ = mipsdsp_sat_add_i16(env, rsl, rtl);
+
+    return (target_long)(int32_t)(((uint32_t)temph << 16) \
+                                  | ((uint32_t)templ & 0xFFFF));
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_addq_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_add_i16(env, rs3, rt3);
+    tempC = mipsdsp_add_i16(env, rs2, rt2);
+    tempB = mipsdsp_add_i16(env, rs1, rt1);
+    tempA = mipsdsp_add_i16(env, rs0, rt0);
+
+    return ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+        ((uint64_t)tempB << 16) | (uint64_t)tempA;
+}
+
+target_ulong helper_addq_s_qh(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sat_add_i16(env, rs3, rt3);
+    tempC = mipsdsp_sat_add_i16(env, rs2, rt2);
+    tempB = mipsdsp_sat_add_i16(env, rs1, rt1);
+    tempA = mipsdsp_sat_add_i16(env, rs0, rt0);
+
+    return ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+        ((uint64_t)tempB << 16) | (uint64_t)tempA;
+}
+#endif
+
+target_ulong helper_addq_s_w(CPUMIPSState *env,
+                             target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    rd = mipsdsp_sat_add_i32(env, rs, rt);
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_addq_pw(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint32_t rs1, rs0;
+    uint32_t rt1, rt0;
+    uint32_t tempB, tempA;
+
+    rs1 = (rs >> 32) & MIPSDSP_LLO;
+    rs0 = rs & MIPSDSP_LLO;
+    rt1 = (rt >> 32) & MIPSDSP_LLO;
+    rt0 = rt & MIPSDSP_LLO;
+
+    tempB = mipsdsp_add_i32(env, rs1, rt1);
+    tempA = mipsdsp_add_i32(env, rs0, rt0);
+
+    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
+}
+
+target_ulong helper_addq_s_pw(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint32_t rs1, rs0;
+    uint32_t rt1, rt0;
+    uint32_t tempB, tempA;
+
+    rs1 = (rs >> 32) & MIPSDSP_LLO;
+    rs0 = rs & MIPSDSP_LLO;
+    rt1 = (rt >> 32) & MIPSDSP_LLO;
+    rt0 = rt & MIPSDSP_LLO;
+
+    tempB = mipsdsp_sat_add_i32(env, rs1, rt1);
+    tempA = mipsdsp_sat_add_i32(env, rs0, rt0);
+
+    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
+}
+#endif
+
+target_ulong helper_addu_qb(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    uint8_t  rs0, rs1, rs2, rs3;
+    uint8_t  rt0, rt1, rt2, rt3;
+    uint8_t  temp0, temp1, temp2, temp3;
+
+    rs0 =  rs & MIPSDSP_Q0;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+
+    rt0 =  rt & MIPSDSP_Q0;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+
+    temp0 = mipsdsp_add_u8(env, rs0, rt0);
+    temp1 = mipsdsp_add_u8(env, rs1, rt1);
+    temp2 = mipsdsp_add_u8(env, rs2, rt2);
+    temp3 = mipsdsp_add_u8(env, rs3, rt3);
+
+    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
+         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
+         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
+         ((uint32_t)temp0         & MIPSDSP_Q0);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addu_s_qb(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    uint8_t rs0, rs1, rs2, rs3;
+    uint8_t rt0, rt1, rt2, rt3;
+    uint8_t temp0, temp1, temp2, temp3;
+
+    rs0 =  rs & MIPSDSP_Q0;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+
+    rt0 =  rt & MIPSDSP_Q0;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+
+    temp0 = mipsdsp_sat_add_u8(env, rs0, rt0);
+    temp1 = mipsdsp_sat_add_u8(env, rs1, rt1);
+    temp2 = mipsdsp_sat_add_u8(env, rs2, rt2);
+    temp3 = mipsdsp_sat_add_u8(env, rs3, rt3);
+
+    rd = (((uint8_t)temp3 << 24) & MIPSDSP_Q3) |
+         (((uint8_t)temp2 << 16) & MIPSDSP_Q2) |
+         (((uint8_t)temp1 <<  8) & MIPSDSP_Q1) |
+         ((uint8_t)temp0         & MIPSDSP_Q0);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_adduh_qb(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    uint8_t  rs0, rs1, rs2, rs3;
+    uint8_t  rt0, rt1, rt2, rt3;
+    uint8_t  temp0, temp1, temp2, temp3;
+
+    rs0 =  rs & MIPSDSP_Q0;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+
+    rt0 =  rt & MIPSDSP_Q0;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+
+    temp0 = mipsdsp_rshift1_add_u8(rs0, rt0);
+    temp1 = mipsdsp_rshift1_add_u8(rs1, rt1);
+    temp2 = mipsdsp_rshift1_add_u8(rs2, rt2);
+    temp3 = mipsdsp_rshift1_add_u8(rs3, rt3);
+
+    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
+         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
+         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
+         ((uint32_t)temp0         & MIPSDSP_Q0);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_adduh_r_qb(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    uint8_t  rs0, rs1, rs2, rs3;
+    uint8_t  rt0, rt1, rt2, rt3;
+    uint8_t  temp0, temp1, temp2, temp3;
+
+    rs0 =  rs & MIPSDSP_Q0;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+
+    rt0 =  rt & MIPSDSP_Q0;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+
+    temp0 = mipsdsp_rrshift1_add_u8(rs0, rt0);
+    temp1 = mipsdsp_rrshift1_add_u8(rs1, rt1);
+    temp2 = mipsdsp_rrshift1_add_u8(rs2, rt2);
+    temp3 = mipsdsp_rrshift1_add_u8(rs3, rt3);
+
+    rd = (((uint32_t)temp3 << 24) & MIPSDSP_Q3) |
+         (((uint32_t)temp2 << 16) & MIPSDSP_Q2) |
+         (((uint32_t)temp1 <<  8) & MIPSDSP_Q1) |
+         ((uint32_t)temp0         & MIPSDSP_Q0);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addu_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl, rth, rtl, temph, templ;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+    temph = mipsdsp_add_u16(env, rsh, rth);
+    templ = mipsdsp_add_u16(env, rsl, rtl);
+    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addu_s_ph(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl, rth, rtl, temph, templ;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+    temph = mipsdsp_sat_add_u16(env, rsh, rth);
+    templ = mipsdsp_sat_add_u16(env, rsl, rtl);
+    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addqh_ph(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    int16_t rsh, rsl, rth, rtl, temph, templ;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    temph = mipsdsp_rshift1_add_q16(rsh, rth);
+    templ = mipsdsp_rshift1_add_q16(rsl, rtl);
+    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addqh_r_ph(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    int16_t rsh, rsl, rth, rtl, temph, templ;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    temph = mipsdsp_rrshift1_add_q16(rsh, rth);
+    templ = mipsdsp_rrshift1_add_q16(rsl, rtl);
+    rd = ((uint32_t)temph << 16) | ((uint32_t)templ & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addqh_w(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+
+    rd = mipsdsp_rshift1_add_q32(rs, rt);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_addqh_r_w(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+
+    rd = mipsdsp_rrshift1_add_q32(rs, rt);
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_addu_ob(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_add_u8(env, rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_addu_s_ob(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_sat_add_u8(env, rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_adduh_ob(target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_rshift1_add_u8(rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_adduh_r_ob(target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_rrshift1_add_u8(rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_addu_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_add_u16(env, rs3, rt3);
+    tempC = mipsdsp_add_u16(env, rs2, rt2);
+    tempB = mipsdsp_add_u16(env, rs1, rt1);
+    tempA = mipsdsp_add_u16(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return temp;
+}
+
+target_ulong helper_addu_s_qh(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sat_add_u16(env, rs3, rt3);
+    tempC = mipsdsp_sat_add_u16(env, rs2, rt2);
+    tempB = mipsdsp_sat_add_u16(env, rs1, rt1);
+    tempA = mipsdsp_sat_add_u16(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return temp;
+}
+#endif
+
+target_ulong helper_subq_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl;
+    uint16_t rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    tempB = mipsdsp_sub_i16(env, rsh, rth);
+    tempA = mipsdsp_sub_i16(env, rsl, rtl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subq_s_ph(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl;
+    uint16_t rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    tempB = mipsdsp_sat16_sub(env, rsh, rth);
+    tempA = mipsdsp_sat16_sub(env, rsl, rtl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_subq_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sub_i16(env, rs3, rt3);
+    tempC = mipsdsp_sub_i16(env, rs2, rt2);
+    tempB = mipsdsp_sub_i16(env, rs1, rt1);
+    tempA = mipsdsp_sub_i16(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+      ((uint64_t)tempB << 16) | (uint64_t)tempA;
+    return temp;
+}
+
+target_ulong helper_subq_s_qh(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sat16_sub(env, rs3, rt3);
+    tempC = mipsdsp_sat16_sub(env, rs2, rt2);
+    tempB = mipsdsp_sat16_sub(env, rs1, rt1);
+    tempA = mipsdsp_sat16_sub(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+    return temp;
+}
+#endif
+
+target_ulong helper_subq_s_w(CPUMIPSState *env,
+                             target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+
+    rd = mipsdsp_sat32_sub(env, rs, rt);
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_subq_pw(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint32_t rs1, rs0;
+    uint32_t rt1, rt0;
+    uint32_t tempB, tempA;
+
+    rs1 = (rs >> 32) & MIPSDSP_LLO;
+    rs0 = rs & MIPSDSP_LLO;
+    rt1 = (rt >> 32) & MIPSDSP_LLO;
+    rt0 = rt & MIPSDSP_LLO;
+
+    tempB = mipsdsp_sub32(env, rs1, rt1);
+    tempA = mipsdsp_sub32(env, rs0, rt0);
+
+    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
+}
+
+target_ulong helper_subq_s_pw(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint32_t rs1, rs0;
+    uint32_t rt1, rt0;
+    uint32_t tempB, tempA;
+
+    rs1 = (rs >> 32) & MIPSDSP_LLO;
+    rs0 = rs & MIPSDSP_LLO;
+    rt1 = (rt >> 32) & MIPSDSP_LLO;
+    rt0 = rt & MIPSDSP_LLO;
+
+    tempB = mipsdsp_sat32_sub(env, rs1, rt1);
+    tempA = mipsdsp_sat32_sub(env, rs0, rt0);
+
+    return ((uint64_t)tempB << 32) | (uint64_t)tempA;
+}
+#endif
+
+target_ulong helper_subu_qb(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint8_t rs3, rs2, rs1, rs0;
+    uint8_t rt3, rt2, rt1, rt0;
+    uint8_t tempD, tempC, tempB, tempA;
+    uint32_t rd;
+
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs0 =  rs & MIPSDSP_Q0;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt0 =  rt & MIPSDSP_Q0;
+
+    tempD = mipsdsp_sub_u8(env, rs3, rt3);
+    tempC = mipsdsp_sub_u8(env, rs2, rt2);
+    tempB = mipsdsp_sub_u8(env, rs1, rt1);
+    tempA = mipsdsp_sub_u8(env, rs0, rt0);
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subu_s_qb(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint8_t rs3, rs2, rs1, rs0;
+    uint8_t rt3, rt2, rt1, rt0;
+    uint8_t tempD, tempC, tempB, tempA;
+    uint32_t rd;
+
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs0 =  rs & MIPSDSP_Q0;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt0 =  rt & MIPSDSP_Q0;
+
+    tempD = mipsdsp_satu8_sub(env, rs3, rt3);
+    tempC = mipsdsp_satu8_sub(env, rs2, rt2);
+    tempB = mipsdsp_satu8_sub(env, rs1, rt1);
+    tempA = mipsdsp_satu8_sub(env, rs0, rt0);
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subuh_qb(target_ulong rs, target_ulong rt)
+{
+    uint8_t rs3, rs2, rs1, rs0;
+    uint8_t rt3, rt2, rt1, rt0;
+    uint8_t tempD, tempC, tempB, tempA;
+    uint32_t rd;
+
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs0 =  rs & MIPSDSP_Q0;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt0 =  rt & MIPSDSP_Q0;
+
+    tempD = ((uint16_t)rs3 - (uint16_t)rt3) >> 1;
+    tempC = ((uint16_t)rs2 - (uint16_t)rt2) >> 1;
+    tempB = ((uint16_t)rs1 - (uint16_t)rt1) >> 1;
+    tempA = ((uint16_t)rs0 - (uint16_t)rt0) >> 1;
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+
+    return (target_ulong)rd;
+}
+
+target_ulong helper_subuh_r_qb(target_ulong rs, target_ulong rt)
+{
+    uint8_t rs3, rs2, rs1, rs0;
+    uint8_t rt3, rt2, rt1, rt0;
+    uint8_t tempD, tempC, tempB, tempA;
+    uint32_t rd;
+
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs0 =  rs & MIPSDSP_Q0;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rt0 =  rt & MIPSDSP_Q0;
+
+    tempD = ((uint16_t)rs3 - (uint16_t)rt3 + 1) >> 1;
+    tempC = ((uint16_t)rs2 - (uint16_t)rt2 + 1) >> 1;
+    tempB = ((uint16_t)rs1 - (uint16_t)rt1 + 1) >> 1;
+    tempA = ((uint16_t)rs0 - (uint16_t)rt0 + 1) >> 1;
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+
+    return (target_ulong)rd;
+}
+
+target_ulong helper_subu_ph(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl, rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    tempB = mipsdsp_sub_u16_u16(env, rth, rsh);
+    tempA = mipsdsp_sub_u16_u16(env, rtl, rsl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subu_s_ph(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl, rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    tempB = mipsdsp_satu16_sub_u16_u16(env, rth, rsh);
+    tempA = mipsdsp_satu16_sub_u16_u16(env, rtl, rsl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subqh_ph(target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl;
+    uint16_t rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+    tempB = mipsdsp_rshift1_sub_q16(rsh, rth);
+    tempA = mipsdsp_rshift1_sub_q16(rsl, rtl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subqh_r_ph(target_ulong rs, target_ulong rt)
+{
+    uint16_t rsh, rsl;
+    uint16_t rth, rtl;
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+    tempB = mipsdsp_rrshift1_sub_q16(rsh, rth);
+    tempA = mipsdsp_rrshift1_sub_q16(rsl, rtl);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subqh_w(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+
+    rd = mipsdsp_rshift1_sub_q32(rs, rt);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_subqh_r_w(target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+
+    rd = mipsdsp_rrshift1_sub_q32(rs, rt);
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_subu_ob(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_sub_u8(env, rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result += (uint64_t)temp[i] << (i * 8);
+    }
+
+    return result;
+}
+
+target_ulong helper_subu_s_ob(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = mipsdsp_satu8_sub(env, rs_t[i], rt_t[i]);
+    }
+
+    for (i = 0; i < 8; i++) {
+        result += (uint64_t)temp[i] << (i * 8);
+    }
+
+    return result;
+}
+
+target_ulong helper_subuh_ob(target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = ((uint16_t)rs_t[i] - (uint16_t)rt_t[i]) >> 1;
+      }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_subuh_r_ob(target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint8_t rs_t[8], rt_t[8];
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        rt_t[i] = (rt >> (8 * i)) & MIPSDSP_Q0;
+        temp[i] = ((uint16_t)rs_t[i] - (uint16_t)rt_t[i] + 1) >> 1;
+    }
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_subu_qh(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sub_u16_u16(env, rs3, rt3);
+    tempC = mipsdsp_sub_u16_u16(env, rs2, rt2);
+    tempB = mipsdsp_sub_u16_u16(env, rs1, rt1);
+    tempA = mipsdsp_sub_u16_u16(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+      ((uint64_t)tempB << 16) | (uint64_t)tempA;
+    return temp;
+}
+
+
+target_ulong helper_subu_s_qh(CPUMIPSState *env,
+                              target_ulong rs, target_ulong rt)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_satu16_sub_u16_u16(env, rs3, rt3);
+    tempC = mipsdsp_satu16_sub_u16_u16(env, rs2, rt2);
+    tempB = mipsdsp_satu16_sub_u16_u16(env, rs1, rt1);
+    tempA = mipsdsp_satu16_sub_u16_u16(env, rs0, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+    return temp;
+}
+#endif
+
+target_ulong helper_addsc(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint64_t temp, tempRs, tempRt;
+    int32_t flag;
+
+    tempRs = (uint64_t)rs & MIPSDSP_LLO;
+    tempRt = (uint64_t)rt & MIPSDSP_LLO;
+
+    temp = tempRs + tempRt;
+    flag = (temp & 0x0100000000ull) >> 32;
+    set_DSPControl_carryflag(env, flag);
+
+    return (target_long)(int32_t)(temp & MIPSDSP_LLO);
+}
+
+target_ulong helper_addwc(CPUMIPSState *env, target_ulong rs, target_ulong rt)
+{
+    uint32_t rd;
+    int32_t temp32, temp31;
+    int64_t tempL;
+
+    tempL = (int32_t)rs + (int32_t)rt + get_DSPControl_carryflag(env);
+    temp31 = (tempL >> 31) & 0x01;
+    temp32 = (tempL >> 32) & 0x01;
+
+    if (temp31 != temp32) {
+        set_DSPControl_overflow_flag(env, 1, 20);
+    }
+
+    rd = tempL & MIPSDSP_LLO;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_modsub(target_ulong rs, target_ulong rt)
+{
+    int32_t decr;
+    uint16_t lastindex;
+    target_ulong rd;
+
+    decr = rt & MIPSDSP_Q0;
+    lastindex = (rt >> 8) & MIPSDSP_LO;
+
+    if ((rs & MIPSDSP_LLO) == 0x00000000) {
+        rd = (target_ulong)lastindex;
+    } else {
+        rd = rs - decr;
+    }
+
+    return rd;
+}
+
+target_ulong helper_raddu_w_qb(target_ulong rs)
+{
+    uint8_t  rs3, rs2, rs1, rs0;
+    uint16_t temp;
+    uint32_t rd;
+
+    rs3 = (rs & MIPSDSP_Q3) >> 24;
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs1 = (rs & MIPSDSP_Q1) >>  8;
+    rs0 =  rs & MIPSDSP_Q0;
+
+    temp = (uint16_t)rs3 + (uint16_t)rs2 + (uint16_t)rs1 + (uint16_t)rs0;
+    rd = temp;
+
+    return (target_ulong)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_raddu_l_ob(target_ulong rs)
+{
+    int i;
+    uint16_t rs_t[8];
+    uint64_t temp;
+
+    temp = 0;
+
+    for (i = 0; i < 8; i++) {
+        rs_t[i] = (rs >> (8 * i)) & MIPSDSP_Q0;
+        temp += (uint64_t)rs_t[i];
+    }
+
+    return temp;
+}
+#endif
+
+target_ulong helper_absq_s_qb(CPUMIPSState *env, target_ulong rt)
+{
+    uint32_t rd;
+    int8_t tempD, tempC, tempB, tempA;
+
+    tempD = (rt & MIPSDSP_Q3) >> 24;
+    tempC = (rt & MIPSDSP_Q2) >> 16;
+    tempB = (rt & MIPSDSP_Q1) >>  8;
+    tempA =  rt & MIPSDSP_Q0;
+
+    rd = (((uint32_t)mipsdsp_sat_abs_u8 (env, tempD) << 24) & MIPSDSP_Q3) |
+         (((uint32_t)mipsdsp_sat_abs_u8 (env, tempC) << 16) & MIPSDSP_Q2) |
+         (((uint32_t)mipsdsp_sat_abs_u8 (env, tempB) <<  8) & MIPSDSP_Q1) |
+         ((uint32_t)mipsdsp_sat_abs_u8 (env, tempA) & MIPSDSP_Q0);
+
+    return (target_ulong)rd;
+}
+
+target_ulong helper_absq_s_ph(CPUMIPSState *env, target_ulong rt)
+{
+    uint32_t rd;
+    int16_t tempA, tempB;
+
+    tempA = (rt & MIPSDSP_HI) >> 16;
+    tempB =  rt & MIPSDSP_LO;
+
+    rd = ((uint32_t)(uint16_t)mipsdsp_sat_abs_u16 (env, tempA) << 16) |
+        ((uint32_t)((uint16_t)mipsdsp_sat_abs_u16 (env, tempB)) & 0xFFFF);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_absq_s_w(CPUMIPSState *env, target_ulong rt)
+{
+    uint32_t rd;
+    int32_t temp;
+
+    temp = rt;
+    rd = mipsdsp_sat_abs_u32(env, temp);
+
+    return (target_ulong)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_absq_s_ob(CPUMIPSState *env, target_ulong rt)
+{
+    int8_t tempH, tempG, tempF, tempE;
+    int8_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    tempH = (rt >> 56) & MIPSDSP_Q0;
+    tempG = (rt >> 48) & MIPSDSP_Q0;
+    tempF = (rt >> 40) & MIPSDSP_Q0;
+    tempE = (rt >> 32) & MIPSDSP_Q0;
+    tempD = (rt >> 24) & MIPSDSP_Q0;
+    tempC = (rt >> 16) & MIPSDSP_Q0;
+    tempB = (rt >> 8) & MIPSDSP_Q0;
+    tempA = rt & MIPSDSP_Q0;
+
+    tempH = mipsdsp_sat_abs_u8(env, tempH);
+    tempG = mipsdsp_sat_abs_u8(env, tempG);
+    tempF = mipsdsp_sat_abs_u8(env, tempF);
+    tempE = mipsdsp_sat_abs_u8(env, tempE);
+    tempD = mipsdsp_sat_abs_u8(env, tempD);
+    tempC = mipsdsp_sat_abs_u8(env, tempC);
+    tempB = mipsdsp_sat_abs_u8(env, tempB);
+    tempA = mipsdsp_sat_abs_u8(env, tempA);
+
+    temp = ((uint64_t)(uint8_t)tempH << 56) | ((uint64_t)(uint8_t)tempG << 48) |
+        ((uint64_t)(uint8_t)tempF << 40) | ((uint64_t)(uint8_t)tempE << 32) |
+        ((uint64_t)(uint8_t)tempD << 24) | ((uint64_t)(uint8_t)tempC << 16) |
+        ((uint64_t)(uint8_t)tempB << 8) | (uint64_t)(uint8_t)tempA;
+
+    return temp;
+}
+
+target_ulong helper_absq_s_qh(CPUMIPSState *env, target_ulong rt)
+{
+    int16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    tempD = (rt >> 48) & MIPSDSP_LO;
+    tempC = (rt >> 32) & MIPSDSP_LO;
+    tempB = (rt >> 16) & MIPSDSP_LO;
+    tempA = rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sat_abs_u16(env, tempD);
+    tempC = mipsdsp_sat_abs_u16(env, tempC);
+    tempB = mipsdsp_sat_abs_u16(env, tempB);
+    tempA = mipsdsp_sat_abs_u16(env, tempA);
+
+    temp = ((uint64_t)(uint16_t)tempD << 48) | \
+           ((uint64_t)(uint16_t)tempC << 32) | \
+           ((uint64_t)(uint16_t)tempB << 16) | \
+           (uint64_t)(uint16_t)tempA;
+
+    return temp;
+}
+
+target_ulong helper_absq_s_pw(CPUMIPSState *env, target_ulong rt)
+{
+    int32_t tempB, tempA;
+    uint64_t temp;
+
+    tempB = (rt >> 32) & MIPSDSP_LLO;
+    tempA = rt & MIPSDSP_LLO;
+
+    tempB = mipsdsp_sat_abs_u32(env, tempB);
+    tempA = mipsdsp_sat_abs_u32(env, tempA);
+
+    temp = ((uint64_t)(uint32_t)tempB << 32) | (uint64_t)(uint32_t)tempA;
+
+    return temp;
+}
+#endif
+
+target_ulong helper_precr_qb_ph(target_ulong rs, target_ulong rt)
+{
+    uint8_t  rs2, rs0, rt2, rt0;
+    uint32_t rd;
+
+    rs2 = (rs & MIPSDSP_Q2) >> 16;
+    rs0 =  rs & MIPSDSP_Q0;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt0 =  rt & MIPSDSP_Q0;
+    rd = ((uint32_t)rs2 << 24) | ((uint32_t)rs0 << 16) |
+         ((uint32_t)rt2 <<  8) | (uint32_t)rt0;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_precrq_qb_ph(target_ulong rs, target_ulong rt)
+{
+    uint8_t tempD, tempC, tempB, tempA;
+    uint32_t rd;
+
+    tempD = (rs & MIPSDSP_Q3) >> 24;
+    tempC = (rs & MIPSDSP_Q1) >>  8;
+    tempB = (rt & MIPSDSP_Q3) >> 24;
+    tempA = (rt & MIPSDSP_Q1) >>  8;
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_precr_sra_ph_w(uint32_t sa, target_ulong rs,
+                                   target_ulong rt)
+{
+    uint16_t tempB, tempA;
+
+    tempB = ((int32_t)rt >> sa) & MIPSDSP_LO;
+    tempA = ((int32_t)rs >> sa) & MIPSDSP_LO;
+    rt = ((uint32_t)tempB << 16) | ((uint32_t)tempA & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rt;
+}
+
+target_ulong helper_precr_sra_r_ph_w(uint32_t sa,
+                                     target_ulong rs, target_ulong rt)
+{
+    uint64_t tempB, tempA;
+
+    /* If sa = 0, then (sa - 1) = -1 will case shift error, so we need else. */
+    if (sa == 0) {
+        tempB = (rt & MIPSDSP_LO) << 1;
+        tempA = (rs & MIPSDSP_LO) << 1;
+    } else {
+        tempB = ((int32_t)rt >> (sa - 1)) + 1;
+        tempA = ((int32_t)rs >> (sa - 1)) + 1;
+    }
+    rt = (((tempB >> 1) & MIPSDSP_LO) << 16) | ((tempA >> 1) & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rt;
+}
+
+target_ulong helper_precrq_ph_w(target_ulong rs, target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    tempB = (rs & MIPSDSP_HI) >> 16;
+    tempA = (rt & MIPSDSP_HI) >> 16;
+    rd = ((uint32_t)tempB << 16) | ((uint32_t)tempA & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_precrq_rs_ph_w(CPUMIPSState *env,
+                                   target_ulong rs, target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t rd;
+
+    tempB = mipsdsp_trunc16_sat16_round(env, rs);
+    tempA = mipsdsp_trunc16_sat16_round(env, rt);
+    rd = ((uint32_t)tempB << 16) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_precr_ob_qh(target_ulong rs, target_ulong rt)
+{
+    uint8_t rs6, rs4, rs2, rs0;
+    uint8_t rt6, rt4, rt2, rt0;
+    uint64_t temp;
+
+    rs6 = (rs >> 48) & MIPSDSP_Q0;
+    rs4 = (rs >> 32) & MIPSDSP_Q0;
+    rs2 = (rs >> 16) & MIPSDSP_Q0;
+    rs0 = rs & MIPSDSP_Q0;
+    rt6 = (rt >> 48) & MIPSDSP_Q0;
+    rt4 = (rt >> 32) & MIPSDSP_Q0;
+    rt2 = (rt >> 16) & MIPSDSP_Q0;
+    rt0 = rt & MIPSDSP_Q0;
+
+    temp = ((uint64_t)rs6 << 56) | ((uint64_t)rs4 << 48) |
+           ((uint64_t)rs2 << 40) | ((uint64_t)rs0 << 32) |
+           ((uint64_t)rt6 << 24) | ((uint64_t)rt4 << 16) |
+           ((uint64_t)rt2 << 8) | (uint64_t)rt0;
+
+    return temp;
+}
+
+target_ulong helper_precr_sra_qh_pw(target_ulong rs, target_ulong rt,
+                                    uint32_t sa)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t result;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    /* When sa = 0, we use rt2, rt0, rs2, rs0;
+     * when sa != 0, we use rt3, rt1, rs3, rs1. */
+    if (sa == 0) {
+        tempD = rt2;
+        tempC = rt0;
+        tempB = rs2;
+        tempA = rs0;
+    } else {
+        tempD = (int16_t)rt3 >> sa;
+        tempC = (int16_t)rt1 >> sa;
+        tempB = (int16_t)rs3 >> sa;
+        tempA = (int16_t)rs1 >> sa;
+    }
+
+    result = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+             ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return result;
+}
+
+
+target_ulong helper_precr_sra_r_qh_pw(target_ulong rs, target_ulong rt,
+                                      uint32_t sa)
+{
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t result;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    /* When sa = 0, we use rt2, rt0, rs2, rs0;
+     * when sa != 0, we use rt3, rt1, rs3, rs1. */
+    if (sa == 0) {
+        tempD = rt2 << 1;
+        tempC = rt0 << 1;
+        tempB = rs2 << 1;
+        tempA = rs0 << 1;
+    } else {
+        tempD = (((int16_t)rt3 >> sa) + 1) >> 1;
+        tempC = (((int16_t)rt1 >> sa) + 1) >> 1;
+        tempB = (((int16_t)rs3 >> sa) + 1) >> 1;
+        tempA = (((int16_t)rs1 >> sa) + 1) >> 1;
+    }
+
+    result = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+             ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return result;
+}
+
+target_ulong helper_precrq_ob_qh(target_ulong rs, target_ulong rt)
+{
+    uint8_t rs6, rs4, rs2, rs0;
+    uint8_t rt6, rt4, rt2, rt0;
+    uint64_t temp;
+
+    rs6 = (rs >> 56) & MIPSDSP_Q0;
+    rs4 = (rs >> 40) & MIPSDSP_Q0;
+    rs2 = (rs >> 24) & MIPSDSP_Q0;
+    rs0 = (rs >> 8) & MIPSDSP_Q0;
+    rt6 = (rt >> 56) & MIPSDSP_Q0;
+    rt4 = (rt >> 40) & MIPSDSP_Q0;
+    rt2 = (rt >> 24) & MIPSDSP_Q0;
+    rt0 = (rt >> 8) & MIPSDSP_Q0;
+
+    temp = ((uint64_t)rs6 << 56) | ((uint64_t)rs4 << 48) |
+           ((uint64_t)rs2 << 40) | ((uint64_t)rs0 << 32) |
+           ((uint64_t)rt6 << 24) | ((uint64_t)rt4 << 16) |
+           ((uint64_t)rt2 << 8) | (uint64_t)rt0;
+
+    return temp;
+}
+
+target_ulong helper_precrq_qh_pw(target_ulong rs, target_ulong rt)
+{
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    tempD = (rs >> 48) & MIPSDSP_LO;
+    tempC = (rs >> 16) & MIPSDSP_LO;
+    tempB = (rt >> 48) & MIPSDSP_LO;
+    tempA = (rt >> 16) & MIPSDSP_LO;
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return temp;
+}
+
+target_ulong helper_precrq_rs_qh_pw(CPUMIPSState *env,
+                                    target_ulong rs, target_ulong rt)
+{
+    uint32_t rs2, rs0;
+    uint32_t rt2, rt0;
+    uint16_t tempD, tempC, tempB, tempA;
+    uint64_t temp;
+
+    rs2 = (rs >> 32) & MIPSDSP_LLO;
+    rs0 = rs & MIPSDSP_LLO;
+    rt2 = (rt >> 32) & MIPSDSP_LLO;
+    rt0 = rt & MIPSDSP_LLO;
+
+    tempD = mipsdsp_trunc16_sat16_round(env, rs2);
+    tempC = mipsdsp_trunc16_sat16_round(env, rs0);
+    tempB = mipsdsp_trunc16_sat16_round(env, rt2);
+    tempA = mipsdsp_trunc16_sat16_round(env, rt0);
+
+    temp = ((uint64_t)tempD << 48) | ((uint64_t)tempC << 32) |
+           ((uint64_t)tempB << 16) | (uint64_t)tempA;
+
+    return temp;
+}
+
+target_ulong helper_precrq_pw_l(target_ulong rs, target_ulong rt)
+{
+    uint32_t tempB, tempA;
+    uint64_t temp;
+
+    tempB = (rs >> 32) & MIPSDSP_LLO;
+    tempA = (rt >> 32) & MIPSDSP_LLO;
+
+    temp = ((uint64_t)tempB << 32) | (uint64_t)tempA;
+
+    return temp;
+}
+#endif
+
+target_ulong helper_precrqu_s_qb_ph(CPUMIPSState *env,
+                                    target_ulong rs, target_ulong rt)
+{
+    uint8_t  tempD, tempC, tempB, tempA;
+    uint16_t rsh, rsl, rth, rtl;
+    uint32_t rd;
+
+    rsh = (rs & MIPSDSP_HI) >> 16;
+    rsl =  rs & MIPSDSP_LO;
+    rth = (rt & MIPSDSP_HI) >> 16;
+    rtl =  rt & MIPSDSP_LO;
+
+    tempD = mipsdsp_sat8_reduce_precision(env, rsh);
+    tempC = mipsdsp_sat8_reduce_precision(env, rsl);
+    tempB = mipsdsp_sat8_reduce_precision(env, rth);
+    tempA = mipsdsp_sat8_reduce_precision(env, rtl);
+
+    rd = ((uint32_t)tempD << 24) | ((uint32_t)tempC << 16) |
+         ((uint32_t)tempB <<  8) | (uint32_t)tempA;
+
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_precrqu_s_ob_qh(CPUMIPSState *env,
+                                    target_ulong rs, target_ulong rt)
+{
+    int i;
+    uint16_t rs3, rs2, rs1, rs0;
+    uint16_t rt3, rt2, rt1, rt0;
+    uint8_t temp[8];
+    uint64_t result;
+
+    result = 0;
+
+    rs3 = (rs >> 48) & MIPSDSP_LO;
+    rs2 = (rs >> 32) & MIPSDSP_LO;
+    rs1 = (rs >> 16) & MIPSDSP_LO;
+    rs0 = rs & MIPSDSP_LO;
+    rt3 = (rt >> 48) & MIPSDSP_LO;
+    rt2 = (rt >> 32) & MIPSDSP_LO;
+    rt1 = (rt >> 16) & MIPSDSP_LO;
+    rt0 = rt & MIPSDSP_LO;
+
+    temp[7] = mipsdsp_sat8_reduce_precision(env, rs3);
+    temp[6] = mipsdsp_sat8_reduce_precision(env, rs2);
+    temp[5] = mipsdsp_sat8_reduce_precision(env, rs1);
+    temp[4] = mipsdsp_sat8_reduce_precision(env, rs0);
+    temp[3] = mipsdsp_sat8_reduce_precision(env, rt3);
+    temp[2] = mipsdsp_sat8_reduce_precision(env, rt2);
+    temp[1] = mipsdsp_sat8_reduce_precision(env, rt1);
+    temp[0] = mipsdsp_sat8_reduce_precision(env, rt0);
+
+    for (i = 0; i < 8; i++) {
+        result |= (uint64_t)temp[i] << (8 * i);
+    }
+
+    return result;
+}
+
+target_ulong helper_preceq_pw_qhl(target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t tempBI, tempAI;
+
+    tempB = (rt >> 48) & MIPSDSP_LO;
+    tempA = (rt >> 32) & MIPSDSP_LO;
+
+    tempBI = (uint32_t)tempB << 16;
+    tempAI = (uint32_t)tempA << 16;
+
+    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
+}
+
+target_ulong helper_preceq_pw_qhr(target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t tempBI, tempAI;
+
+    tempB = (rt >> 16) & MIPSDSP_LO;
+    tempA = rt & MIPSDSP_LO;
+
+    tempBI = (uint32_t)tempB << 16;
+    tempAI = (uint32_t)tempA << 16;
+
+    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
+}
+
+target_ulong helper_preceq_pw_qhla(target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t tempBI, tempAI;
+
+    tempB = (rt >> 48) & MIPSDSP_LO;
+    tempA = (rt >> 16) & MIPSDSP_LO;
+
+    tempBI = (uint32_t)tempB << 16;
+    tempAI = (uint32_t)tempA << 16;
+
+    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
+}
+
+target_ulong helper_preceq_pw_qhra(target_ulong rt)
+{
+    uint16_t tempB, tempA;
+    uint32_t tempBI, tempAI;
+
+    tempB = (rt >> 32) & MIPSDSP_LO;
+    tempA = rt & MIPSDSP_LO;
+
+    tempBI = (uint32_t)tempB << 16;
+    tempAI = (uint32_t)tempA << 16;
+
+    return ((uint64_t)tempBI << 32) | ((uint64_t)tempAI);
+}
+
+#endif
+
+target_ulong helper_precequ_ph_qbl(target_ulong rt)
+{
+    uint8_t  rt3, rt2;
+    uint16_t tempB, tempA;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+
+    tempB = (uint16_t)rt3 << 7;
+    tempA = (uint16_t)rt2 << 7;
+
+    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
+}
+
+target_ulong helper_precequ_ph_qbr(target_ulong rt)
+{
+    uint8_t  rt1, rt0;
+    uint16_t tempB, tempA;
+
+    rt1 = (rt & MIPSDSP_Q1) >> 8;
+    rt0 =  rt & MIPSDSP_Q0;
+    tempB = (uint16_t)rt1 << 7;
+    tempA = (uint16_t)rt0 << 7;
+
+    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
+}
+
+target_ulong helper_precequ_ph_qbla(target_ulong rt)
+{
+    uint8_t  rt3, rt1;
+    uint16_t tempB, tempA;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+
+    tempB = (uint16_t)rt3 << 7;
+    tempA = (uint16_t)rt1 << 7;
+
+    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
+}
+
+target_ulong helper_precequ_ph_qbra(target_ulong rt)
+{
+    uint8_t  rt2, rt0;
+    uint16_t tempB, tempA;
+
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt0 =  rt & MIPSDSP_Q0;
+    tempB = (uint16_t)rt2 << 7;
+    tempA = (uint16_t)rt0 << 7;
+
+    return (target_long)(int32_t)(((uint32_t)tempB << 16) | (uint32_t)tempA);
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_precequ_qh_obl(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint16_t tempDS, tempCS, tempBS, tempAS;
+    uint64_t temp;
+
+    tempDC = (rt >> 56) & MIPSDSP_Q0;
+    tempCC = (rt >> 48) & MIPSDSP_Q0;
+    tempBC = (rt >> 40) & MIPSDSP_Q0;
+    tempAC = (rt >> 32) & MIPSDSP_Q0;
+
+    tempDS = (uint16_t)tempDC << 7;
+    tempCS = (uint16_t)tempCC << 7;
+    tempBS = (uint16_t)tempBC << 7;
+    tempAS = (uint16_t)tempAC << 7;
+
+    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
+           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
+    return temp;
+}
+
+target_ulong helper_precequ_qh_obr(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint16_t tempDS, tempCS, tempBS, tempAS;
+    uint64_t temp;
+
+    tempDC = (rt >> 24) & MIPSDSP_Q0;
+    tempCC = (rt >> 16) & MIPSDSP_Q0;
+    tempBC = (rt >> 8) & MIPSDSP_Q0;
+    tempAC = rt & MIPSDSP_Q0;
+
+    tempDS = (uint16_t)tempDC << 7;
+    tempCS = (uint16_t)tempCC << 7;
+    tempBS = (uint16_t)tempBC << 7;
+    tempAS = (uint16_t)tempAC << 7;
+
+    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
+           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
+    return temp;
+}
+
+target_ulong helper_precequ_qh_obla(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint16_t tempDS, tempCS, tempBS, tempAS;
+    uint64_t temp;
+
+    tempDC = (rt >> 56) & MIPSDSP_Q0;
+    tempCC = (rt >> 40) & MIPSDSP_Q0;
+    tempBC = (rt >> 24) & MIPSDSP_Q0;
+    tempAC = (rt >> 8) & MIPSDSP_Q0;
+
+    tempDS = (uint16_t)tempDC << 7;
+    tempCS = (uint16_t)tempCC << 7;
+    tempBS = (uint16_t)tempBC << 7;
+    tempAS = (uint16_t)tempAC << 7;
+
+    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
+           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
+    return temp;
+}
+
+target_ulong helper_precequ_qh_obra(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint16_t tempDS, tempCS, tempBS, tempAS;
+    uint64_t temp;
+
+    tempDC = (rt >> 48) & MIPSDSP_Q0;
+    tempCC = (rt >> 32) & MIPSDSP_Q0;
+    tempBC = (rt >> 16) & MIPSDSP_Q0;
+    tempAC = rt & MIPSDSP_Q0;
+
+    tempDS = (uint16_t)tempDC << 7;
+    tempCS = (uint16_t)tempCC << 7;
+    tempBS = (uint16_t)tempBC << 7;
+    tempAS = (uint16_t)tempAC << 7;
+
+    temp = ((uint64_t)tempDS << 48) | ((uint64_t)tempCS << 32) |
+           ((uint64_t)tempBS << 16) | (uint64_t)tempAS;
+    return temp;
+}
+#endif
+
+target_ulong helper_preceu_ph_qbl(target_ulong rt)
+{
+    uint8_t  rt3, rt2;
+    uint32_t rd;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rd = ((uint32_t)(uint16_t)rt3 << 16) | \
+         ((uint32_t)(uint16_t)rt2 & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_preceu_ph_qbr(target_ulong rt)
+{
+    uint8_t  rt1, rt0;
+    uint32_t rd;
+
+    rt1 = (rt & MIPSDSP_Q1) >> 8;
+    rt0 =  rt & MIPSDSP_Q0;
+    rd = ((uint32_t)(uint16_t)rt1 << 16) | \
+         ((uint32_t)(uint16_t)rt0 & MIPSDSP_LO);
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_preceu_ph_qbla(target_ulong rt)
+{
+    uint8_t  rt3, rt1;
+    uint32_t rd;
+
+    rt3 = (rt & MIPSDSP_Q3) >> 24;
+    rt1 = (rt & MIPSDSP_Q1) >>  8;
+    rd = ((uint32_t)(uint16_t)rt3 << 16) | \
+         ((uint32_t)(uint16_t)rt1 & MIPSDSP_LO);
+
+    return (target_long)(int32_t)rd;
+}
+
+target_ulong helper_preceu_ph_qbra(target_ulong rt)
+{
+    uint8_t  rt2, rt0;
+    uint32_t rd;
+
+    rt2 = (rt & MIPSDSP_Q2) >> 16;
+    rt0 =  rt & MIPSDSP_Q0;
+    rd = ((uint32_t)(uint16_t)rt2 << 16) | \
+         ((uint32_t)(uint16_t)rt0 & MIPSDSP_LO);
+    return (target_long)(int32_t)rd;
+}
+
+#if defined(TARGET_MIPS64)
+target_ulong helper_preceu_qh_obl(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint64_t temp;
+
+    tempDC = (rt >> 56) & MIPSDSP_Q0;
+    tempCC = (rt >> 48) & MIPSDSP_Q0;
+    tempBC = (rt >> 40) & MIPSDSP_Q0;
+    tempAC = (rt >> 32) & MIPSDSP_Q0;
+
+    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
+           ((uint64_t)(uint16_t)tempCC << 32) | \
+           ((uint64_t)(uint16_t)tempBC << 16) | \
+           (uint64_t)(uint16_t)tempAC;
+    return temp;
+}
+
+target_ulong helper_preceu_qh_obr(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint64_t temp;
+
+    tempDC = (rt >> 24) & MIPSDSP_Q0;
+    tempCC = (rt >> 16) & MIPSDSP_Q0;
+    tempBC = (rt >> 8) & MIPSDSP_Q0;
+    tempAC = rt & MIPSDSP_Q0;
+
+    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
+           ((uint64_t)(uint16_t)tempCC << 32) | \
+           ((uint64_t)(uint16_t)tempBC << 16) | \
+           (uint64_t)(uint16_t)tempAC;
+
+    return temp;
+}
+
+target_ulong helper_preceu_qh_obla(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint64_t temp;
+
+    tempDC = (rt >> 56) & MIPSDSP_Q0;
+    tempCC = (rt >> 40) & MIPSDSP_Q0;
+    tempBC = (rt >> 24) & MIPSDSP_Q0;
+    tempAC = (rt >> 8) & MIPSDSP_Q0;
+
+    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
+           ((uint64_t)(uint16_t)tempCC << 32) | \
+           ((uint64_t)(uint16_t)tempBC << 16) | \
+           (uint64_t)(uint16_t)tempAC;
+
+    return temp;
+}
+
+target_ulong helper_preceu_qh_obra(target_ulong rt)
+{
+    uint8_t tempDC, tempCC, tempBC, tempAC;
+    uint64_t temp;
+
+    tempDC = (rt >> 48) & MIPSDSP_Q0;
+    tempCC = (rt >> 32) & MIPSDSP_Q0;
+    tempBC = (rt >> 16) & MIPSDSP_Q0;
+    tempAC = rt & MIPSDSP_Q0;
+
+    temp = ((uint64_t)(uint16_t)tempDC << 48) | \
+           ((uint64_t)(uint16_t)tempCC << 32) | \
+           ((uint64_t)(uint16_t)tempBC << 16) | \
+           (uint64_t)(uint16_t)tempAC;
+
+    return temp;
+}
+#endif
+
+#undef MIPSDSP_LHI
+#undef MIPSDSP_LLO
+#undef MIPSDSP_HI
+#undef MIPSDSP_LO
+#undef MIPSDSP_Q3
+#undef MIPSDSP_Q2
+#undef MIPSDSP_Q1
+#undef MIPSDSP_Q0
diff --git a/target-mips/helper.h b/target-mips/helper.h
index 76fb451..cd4564b 100644
--- a/target-mips/helper.h
+++ b/target-mips/helper.h
@@ -297,4 +297,130 @@  DEF_HELPER_0(rdhwr_ccres, tl)
 DEF_HELPER_1(pmon, void, int)
 DEF_HELPER_0(wait, void)
 
+/*** MIPS DSP ***/
+/* DSP Arithmetic Sub-class insns */
+DEF_HELPER_FLAGS_3(addq_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addq_s_ph, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(addq_qh, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addq_s_qh, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(addq_s_w, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(addq_pw, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addq_s_pw, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(addu_qb, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addu_s_qb, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(adduh_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(adduh_r_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(addu_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addu_s_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(addqh_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(addqh_r_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(addqh_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(addqh_r_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(addu_ob, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addu_s_ob, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(adduh_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(adduh_r_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(addu_qh, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addu_s_qh, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(subq_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subq_s_ph, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(subq_qh, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subq_s_qh, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(subq_s_w, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(subq_pw, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subq_s_pw, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(subu_qb, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subu_s_qb, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(subuh_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(subuh_r_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(subu_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subu_s_ph, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(subqh_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(subqh_r_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(subqh_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(subqh_r_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(subu_ob, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subu_s_ob, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(subuh_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(subuh_r_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(subu_qh, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(subu_s_qh, 0, tl, env, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(addsc, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_3(addwc, 0, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(modsub, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_1(raddu_w_qb, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_1(raddu_l_ob, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#endif
+DEF_HELPER_FLAGS_2(absq_s_qb, 0, tl, env, tl)
+DEF_HELPER_FLAGS_2(absq_s_ph, 0, tl, env, tl)
+DEF_HELPER_FLAGS_2(absq_s_w, 0, tl, env, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_2(absq_s_ob, 0, tl, env, tl)
+DEF_HELPER_FLAGS_2(absq_s_qh, 0, tl, env, tl)
+DEF_HELPER_FLAGS_2(absq_s_pw, 0, tl, env, tl)
+#endif
+DEF_HELPER_FLAGS_2(precr_qb_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(precrq_qb_ph, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(precr_sra_ph_w, TCG_CALL_CONST | TCG_CALL_PURE,
+                   tl, i32, tl, tl)
+DEF_HELPER_FLAGS_3(precr_sra_r_ph_w, TCG_CALL_CONST | TCG_CALL_PURE,
+                   tl, i32, tl, tl)
+DEF_HELPER_FLAGS_2(precrq_ph_w, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(precrq_rs_ph_w, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_2(precr_ob_qh, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(precr_sra_qh_pw,
+                   TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl, i32)
+DEF_HELPER_FLAGS_3(precr_sra_r_qh_pw,
+                   TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl, i32)
+DEF_HELPER_FLAGS_2(precrq_ob_qh, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_2(precrq_qh_pw, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+DEF_HELPER_FLAGS_3(precrq_rs_qh_pw,
+                   TCG_CALL_CONST | TCG_CALL_PURE, tl, env, tl, tl)
+DEF_HELPER_FLAGS_2(precrq_pw_l, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl, tl)
+#endif
+DEF_HELPER_FLAGS_3(precrqu_s_qb_ph, 0, tl, env, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_3(precrqu_s_ob_qh,
+                   TCG_CALL_CONST | TCG_CALL_PURE, tl, env, tl, tl)
+
+DEF_HELPER_FLAGS_1(preceq_pw_qhl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceq_pw_qhr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceq_pw_qhla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceq_pw_qhra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#endif
+DEF_HELPER_FLAGS_1(precequ_ph_qbl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_ph_qbr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_ph_qbla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_ph_qbra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_1(precequ_qh_obl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_qh_obr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_qh_obla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(precequ_qh_obra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#endif
+DEF_HELPER_FLAGS_1(preceu_ph_qbl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_ph_qbr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_ph_qbla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_ph_qbra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#if defined(TARGET_MIPS64)
+DEF_HELPER_FLAGS_1(preceu_qh_obl, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_qh_obr, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_qh_obla, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+DEF_HELPER_FLAGS_1(preceu_qh_obra, TCG_CALL_CONST | TCG_CALL_PURE, tl, tl)
+#endif
+
 #include "def-helper.h"
diff --git a/target-mips/translate.c b/target-mips/translate.c
index f7bb054..407b4ad 100644
--- a/target-mips/translate.c
+++ b/target-mips/translate.c
@@ -316,6 +316,21 @@  enum {
 
     /* MIPS DSP Load */
     OPC_LX_DSP         = 0x0A | OPC_SPECIAL3,
+    /* MIPS DSP Arithmetic */
+    OPC_ADDU_QB_DSP    = 0x10 | OPC_SPECIAL3,
+#if defined(TARGET_MIPS64)
+    OPC_ADDU_OB_DSP    = 0x14 | OPC_SPECIAL3,
+#endif
+    OPC_ABSQ_S_PH_DSP  = 0x12 | OPC_SPECIAL3,
+#if defined(TARGET_MIPS64)
+    OPC_ABSQ_S_QH_DSP  = 0x16 | OPC_SPECIAL3,
+#endif
+    /* OPC_ADDUH_QB_DSP is same as OPC_MULT_G_2E.  */
+    /* OPC_ADDUH_QB_DSP   = 0x18 | OPC_SPECIAL3,  */
+    OPC_CMPU_EQ_QB_DSP = 0x11 | OPC_SPECIAL3,
+#if defined(TARGET_MIPS64)
+    OPC_CMPU_EQ_OB_DSP = 0x15 | OPC_SPECIAL3,
+#endif
 };
 
 /* BSHFL opcodes */
@@ -354,6 +369,144 @@  enum {
 #endif
 };
 
+#define MASK_ADDU_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_ADDQ_PH        = (0x0A << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDQ_S_PH      = (0x0E << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDQ_S_W       = (0x16 << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDU_QB        = (0x00 << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDU_S_QB      = (0x04 << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDU_PH        = (0x08 << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDU_S_PH      = (0x0C << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBQ_PH        = (0x0B << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBQ_S_PH      = (0x0F << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBQ_S_W       = (0x17 << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBU_QB        = (0x01 << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBU_S_QB      = (0x05 << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBU_PH        = (0x09 << 6) | OPC_ADDU_QB_DSP,
+    OPC_SUBU_S_PH      = (0x0D << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDSC          = (0x10 << 6) | OPC_ADDU_QB_DSP,
+    OPC_ADDWC          = (0x11 << 6) | OPC_ADDU_QB_DSP,
+    OPC_MODSUB         = (0x12 << 6) | OPC_ADDU_QB_DSP,
+    OPC_RADDU_W_QB     = (0x14 << 6) | OPC_ADDU_QB_DSP,
+};
+
+#define OPC_ADDUH_QB_DSP OPC_MULT_G_2E
+#define MASK_ADDUH_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_ADDUH_QB   = (0x00 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_ADDUH_R_QB = (0x02 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_ADDQH_PH   = (0x08 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_ADDQH_R_PH = (0x0A << 6) | OPC_ADDUH_QB_DSP,
+    OPC_ADDQH_W    = (0x10 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_ADDQH_R_W  = (0x12 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBUH_QB   = (0x01 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBUH_R_QB = (0x03 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBQH_PH   = (0x09 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBQH_R_PH = (0x0B << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBQH_W    = (0x11 << 6) | OPC_ADDUH_QB_DSP,
+    OPC_SUBQH_R_W  = (0x13 << 6) | OPC_ADDUH_QB_DSP,
+};
+
+#define MASK_ABSQ_S_PH(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_ABSQ_S_QB       = (0x01 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_ABSQ_S_PH       = (0x09 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_ABSQ_S_W        = (0x11 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQ_W_PHL    = (0x0C << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQ_W_PHR    = (0x0D << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQU_PH_QBL  = (0x04 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQU_PH_QBR  = (0x05 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQU_PH_QBLA = (0x06 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEQU_PH_QBRA = (0x07 << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEU_PH_QBL   = (0x1C << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEU_PH_QBR   = (0x1D << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEU_PH_QBLA  = (0x1E << 6) | OPC_ABSQ_S_PH_DSP,
+    OPC_PRECEU_PH_QBRA  = (0x1F << 6) | OPC_ABSQ_S_PH_DSP,
+};
+
+#define MASK_CMPU_EQ_QB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_PRECR_QB_PH      = (0x0D << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECRQ_QB_PH     = (0x0C << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECR_SRA_PH_W   = (0x1E << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECR_SRA_R_PH_W = (0x1F << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECRQ_PH_W      = (0x14 << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECRQ_RS_PH_W   = (0x15 << 6) | OPC_CMPU_EQ_QB_DSP,
+    OPC_PRECRQU_S_QB_PH  = (0x0F << 6) | OPC_CMPU_EQ_QB_DSP,
+};
+
+#if defined(TARGET_MIPS64)
+#define MASK_ABSQ_S_QH(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_PRECEQ_L_PWL    = (0x14 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQ_L_PWR    = (0x15 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQ_PW_QHL   = (0x0C << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQ_PW_QHR   = (0x0D << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQ_PW_QHLA  = (0x0E << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQ_PW_QHRA  = (0x0F << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQU_QH_OBL  = (0x04 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQU_QH_OBR  = (0x05 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQU_QH_OBLA = (0x06 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEQU_QH_OBRA = (0x07 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEU_QH_OBL   = (0x1C << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEU_QH_OBR   = (0x1D << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEU_QH_OBLA  = (0x1E << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_PRECEU_QH_OBRA  = (0x1F << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_ABSQ_S_OB       = (0x01 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_ABSQ_S_PW       = (0x11 << 6) | OPC_ABSQ_S_QH_DSP,
+    OPC_ABSQ_S_QH       = (0x09 << 6) | OPC_ABSQ_S_QH_DSP,
+};
+#endif
+
+#if defined(TARGET_MIPS64)
+#define MASK_ADDU_OB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_RADDU_L_OB     = (0x14 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBQ_PW        = (0x13 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBQ_S_PW      = (0x17 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBQ_QH        = (0x0B << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBQ_S_QH      = (0x0F << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBU_OB        = (0x01 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBU_S_OB      = (0x05 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBU_QH        = (0x09 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBU_S_QH      = (0x0D << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBUH_OB       = (0x19 << 6) | OPC_ADDU_OB_DSP,
+    OPC_SUBUH_R_OB     = (0x1B << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDQ_PW        = (0x12 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDQ_S_PW      = (0x16 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDQ_QH        = (0x0A << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDQ_S_QH      = (0x0E << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDU_OB        = (0x00 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDU_S_OB      = (0x04 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDU_QH        = (0x08 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDU_S_QH      = (0x0C << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDUH_OB       = (0x18 << 6) | OPC_ADDU_OB_DSP,
+    OPC_ADDUH_R_OB     = (0x1A << 6) | OPC_ADDU_OB_DSP,
+};
+#endif
+
+#if defined(TARGET_MIPS64)
+#define MASK_CMPU_EQ_OB(op) (MASK_SPECIAL3(op) | (op & (0x1F << 6)))
+enum {
+    /* MIPS DSP Arithmetic Sub-class */
+    OPC_PRECR_OB_QH       = (0x0D << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECR_SRA_QH_PW   = (0x1E << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECR_SRA_R_QH_PW = (0x1F << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECRQ_OB_QH      = (0x0C << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECRQ_PW_L       = (0x1C << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECRQ_QH_PW      = (0x14 << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECRQ_RS_QH_PW   = (0x15 << 6) | OPC_CMPU_EQ_OB_DSP,
+    OPC_PRECRQU_S_OB_QH   = (0x0F << 6) | OPC_CMPU_EQ_OB_DSP,
+};
+#endif
+
 /* Coprocessor 0 (rs field) */
 #define MASK_CP0(op)       MASK_OP_MAJOR(op) | (op & (0x1F << 21))
 
@@ -12183,10 +12336,64 @@  static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
             }
             break;
         case OPC_DIV_G_2E ... OPC_DIVU_G_2E:
-        case OPC_MULT_G_2E ... OPC_MULTU_G_2E:
         case OPC_MOD_G_2E ... OPC_MODU_G_2E:
-            check_insn(env, ctx, INSN_LOONGSON2E);
-            gen_loongson_integer(ctx, op1, rd, rs, rt);
+        case OPC_MULT_G_2E ... OPC_MULTU_G_2E:
+        /* OPC_MULT_G_2E, OPC_ADDUH_QB_DSP, OPC_MUL_PH_DSP have
+         * the same mask and op1. */
+            if ((env->insn_flags & ASE_DSPR2) && (op1 == OPC_MULT_G_2E)) {
+                check_dspr2(ctx);
+                op2 = MASK_ADDUH_QB(ctx->opcode);
+                switch (op2) {
+                case OPC_ADDUH_QB:
+                    gen_helper_adduh_qb(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_ADDUH_R_QB:
+                    gen_helper_adduh_r_qb(cpu_gpr[rd],
+                                          cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_ADDQH_PH:
+                    gen_helper_addqh_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_ADDQH_R_PH:
+                    gen_helper_addqh_r_ph(cpu_gpr[rd], cpu_gpr[rs],
+                                          cpu_gpr[rt]);
+                    break;
+                case OPC_ADDQH_W:
+                    gen_helper_addqh_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_ADDQH_R_W:
+                    gen_helper_addqh_r_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBUH_QB:
+                    gen_helper_subuh_qb(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBUH_R_QB:
+                    gen_helper_subuh_r_qb(cpu_gpr[rd],
+                                          cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBQH_PH:
+                    gen_helper_subqh_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBQH_R_PH:
+                    gen_helper_subqh_r_ph(cpu_gpr[rd],
+                                          cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBQH_W:
+                    gen_helper_subqh_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                case OPC_SUBQH_R_W:
+                    gen_helper_subqh_r_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                    break;
+                default:
+                    MIPS_INVAL("MASK ADDUH.QB");
+                    generate_exception(ctx, EXCP_RI);
+                    break;
+                }
+            } else if (env->insn_flags & INSN_LOONGSON2E) {
+                gen_loongson_integer(ctx, op1, rd, rs, rt);
+            } else {
+                generate_exception(ctx, EXCP_RI);
+            }
             break;
         case OPC_LX_DSP:
             check_dsp(ctx);
@@ -12240,6 +12447,224 @@  static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
                 break;
             }
             break;
+        case OPC_ABSQ_S_PH_DSP:
+            op2 = MASK_ABSQ_S_PH(ctx->opcode);
+            switch (op2) {
+            case OPC_ABSQ_S_QB:
+                check_dspr2(ctx);
+                gen_helper_absq_s_qb(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            case OPC_ABSQ_S_PH:
+                check_dsp(ctx);
+                gen_helper_absq_s_ph(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            case OPC_ABSQ_S_W:
+                check_dsp(ctx);
+                gen_helper_absq_s_w(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQ_W_PHL:
+                {
+                    TCGv temp = tcg_temp_new();;
+                    check_dsp(ctx);
+                    tcg_gen_andi_tl(temp, cpu_gpr[rt], 0xFFFF0000);
+                    tcg_gen_ext32s_tl(cpu_gpr[rd], temp);
+                    tcg_temp_free(temp);
+                    break;
+                }
+            case OPC_PRECEQ_W_PHR:
+                {
+                    TCGv temp = tcg_temp_new();;
+                    check_dsp(ctx);
+                    tcg_gen_andi_tl(temp, cpu_gpr[rt], 0x0000FFFF);
+                    tcg_gen_shli_tl(temp, temp, 16);
+                    tcg_gen_ext32s_tl(cpu_gpr[rd], temp);
+                    tcg_temp_free(temp);
+                    break;
+                }
+            case OPC_PRECEQU_PH_QBL:
+                check_dsp(ctx);
+                gen_helper_precequ_ph_qbl(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_PH_QBR:
+                check_dsp(ctx);
+                gen_helper_precequ_ph_qbr(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_PH_QBLA:
+                check_dsp(ctx);
+                gen_helper_precequ_ph_qbla(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_PH_QBRA:
+                check_dsp(ctx);
+                gen_helper_precequ_ph_qbra(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_PH_QBL:
+                check_dsp(ctx);
+                gen_helper_preceu_ph_qbl(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_PH_QBR:
+                check_dsp(ctx);
+                gen_helper_preceu_ph_qbr(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_PH_QBLA:
+                check_dsp(ctx);
+                gen_helper_preceu_ph_qbla(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_PH_QBRA:
+                check_dsp(ctx);
+                gen_helper_preceu_ph_qbra(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK ABSQ_S.PH");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
+        case OPC_ADDU_QB_DSP:
+            op2 = MASK_ADDU_QB(ctx->opcode);
+            switch (op2) {
+            case OPC_ADDQ_PH:
+                check_dsp(ctx);
+                gen_helper_addq_ph(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_S_PH:
+                check_dsp(ctx);
+                gen_helper_addq_s_ph(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_S_W:
+                check_dsp(ctx);
+                gen_helper_addq_s_w(cpu_gpr[rd], cpu_env,
+                                    cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_QB:
+                check_dsp(ctx);
+                gen_helper_addu_qb(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_S_QB:
+                check_dsp(ctx);
+                gen_helper_addu_s_qb(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_PH:
+                check_dspr2(ctx);
+                gen_helper_addu_ph(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_S_PH:
+                check_dspr2(ctx);
+                gen_helper_addu_s_ph(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_PH:
+                check_dsp(ctx);
+                gen_helper_subq_ph(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_S_PH:
+                check_dsp(ctx);
+                gen_helper_subq_s_ph(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_S_W:
+                check_dsp(ctx);
+                gen_helper_subq_s_w(cpu_gpr[rd], cpu_env,
+                                    cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_QB:
+                check_dsp(ctx);
+                gen_helper_subu_qb(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_S_QB:
+                check_dsp(ctx);
+                gen_helper_subu_s_qb(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_PH:
+                check_dspr2(ctx);
+                gen_helper_subu_ph(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_S_PH:
+                check_dspr2(ctx);
+                gen_helper_subu_s_ph(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDSC:
+                check_dsp(ctx);
+                gen_helper_addsc(cpu_gpr[rd], cpu_env,
+                                 cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDWC:
+                check_dsp(ctx);
+                gen_helper_addwc(cpu_gpr[rd], cpu_env,
+                                 cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_MODSUB:
+                check_dsp(ctx);
+                gen_helper_modsub(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_RADDU_W_QB:
+                check_dsp(ctx);
+                gen_helper_raddu_w_qb(cpu_gpr[rd], cpu_gpr[rs]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK ADDU.QB");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
+        case OPC_CMPU_EQ_QB_DSP:
+            op2 = MASK_CMPU_EQ_QB(ctx->opcode);
+            switch (op2) {
+            case OPC_PRECR_QB_PH:
+                check_dspr2(ctx);
+                gen_helper_precr_qb_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQ_QB_PH:
+                check_dsp(ctx);
+                gen_helper_precrq_qb_ph(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_PRECR_SRA_PH_W:
+                check_dspr2(ctx);
+                {
+                    TCGv_i32 temp_rd = tcg_const_i32(rd);
+                    gen_helper_precr_sra_ph_w(cpu_gpr[rt], temp_rd,
+                                           cpu_gpr[rs], cpu_gpr[rt]);
+                    tcg_temp_free_i32(temp_rd);
+                    break;
+                }
+            case OPC_PRECR_SRA_R_PH_W:
+                check_dspr2(ctx);
+                {
+                    TCGv_i32 temp_rd = tcg_const_i32(rd);
+                    gen_helper_precr_sra_r_ph_w(cpu_gpr[rt], temp_rd,
+                                            cpu_gpr[rs], cpu_gpr[rt]);
+                    tcg_temp_free_i32(temp_rd);
+                    break;
+                }
+            case OPC_PRECRQ_PH_W:
+                check_dsp(ctx);
+                gen_helper_precrq_ph_w(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQ_RS_PH_W:
+                check_dsp(ctx);
+                gen_helper_precrq_rs_ph_w(cpu_gpr[rd], cpu_env,
+                                          cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQU_S_QB_PH:
+                check_dsp(ctx);
+                gen_helper_precrqu_s_qb_ph(cpu_gpr[rd], cpu_env,
+                                           cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK CMPU.EQ.QB");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
 #if defined(TARGET_MIPS64)
         case OPC_DEXTM ... OPC_DEXT:
         case OPC_DINSM ... OPC_DINS:
@@ -12260,6 +12685,256 @@  static void decode_opc (CPUMIPSState *env, DisasContext *ctx, int *is_branch)
             gen_loongson_integer(ctx, op1, rd, rs, rt);
             break;
 #endif
+#if defined(TARGET_MIPS64)
+        case OPC_ABSQ_S_QH_DSP:
+            op2 = MASK_ABSQ_S_QH(ctx->opcode);
+            switch (op2) {
+            case OPC_PRECEQ_L_PWL:
+                check_dsp(ctx);
+                tcg_gen_andi_tl(cpu_gpr[rd], cpu_gpr[rt], \
+                                0xFFFFFFFF00000000ull);
+                break;
+            case OPC_PRECEQ_L_PWR:
+                check_dsp(ctx);
+                tcg_gen_andi_tl(cpu_gpr[rd], cpu_gpr[rt], 0xFFFFFFFFull);
+                break;
+            case OPC_PRECEQ_PW_QHL:
+                check_dsp(ctx);
+                gen_helper_preceq_pw_qhl(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQ_PW_QHR:
+                check_dsp(ctx);
+                gen_helper_preceq_pw_qhr(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQ_PW_QHLA:
+                check_dsp(ctx);
+                gen_helper_preceq_pw_qhla(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQ_PW_QHRA:
+                check_dsp(ctx);
+                gen_helper_preceq_pw_qhra(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_QH_OBL:
+                check_dsp(ctx);
+                gen_helper_precequ_qh_obl(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_QH_OBR:
+                check_dsp(ctx);
+                gen_helper_precequ_qh_obr(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_QH_OBLA:
+                check_dsp(ctx);
+                gen_helper_precequ_qh_obla(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEQU_QH_OBRA:
+                check_dsp(ctx);
+                gen_helper_precequ_qh_obra(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_QH_OBL:
+                check_dsp(ctx);
+                gen_helper_preceu_qh_obl(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_QH_OBR:
+                check_dsp(ctx);
+                gen_helper_preceu_qh_obr(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_QH_OBLA:
+                check_dsp(ctx);
+                gen_helper_preceu_qh_obla(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_PRECEU_QH_OBRA:
+                check_dsp(ctx);
+                gen_helper_preceu_qh_obra(cpu_gpr[rd], cpu_gpr[rt]);
+                break;
+            case OPC_ABSQ_S_OB:
+                check_dspr2(ctx);
+                gen_helper_absq_s_ob(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            case OPC_ABSQ_S_PW:
+                check_dsp(ctx);
+                gen_helper_absq_s_pw(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            case OPC_ABSQ_S_QH:
+                check_dsp(ctx);
+                gen_helper_absq_s_qh(cpu_gpr[rd], cpu_env, cpu_gpr[rt]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK ABSQ_S.QH");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
+#endif
+#if defined(TARGET_MIPS64)
+        case OPC_ADDU_OB_DSP:
+            op2 = MASK_ADDU_OB(ctx->opcode);
+            switch (op2) {
+            case OPC_RADDU_L_OB:
+                check_dsp(ctx);
+                gen_helper_raddu_l_ob(cpu_gpr[rd], cpu_gpr[rs]);
+                break;
+            case OPC_SUBQ_PW:
+                check_dsp(ctx);
+                gen_helper_subq_pw(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_S_PW:
+                check_dsp(ctx);
+                gen_helper_subq_s_pw(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_QH:
+                check_dsp(ctx);
+                gen_helper_subq_qh(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBQ_S_QH:
+                check_dsp(ctx);
+                gen_helper_subq_s_qh(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_OB:
+                check_dsp(ctx);
+                gen_helper_subu_ob(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_S_OB:
+                check_dsp(ctx);
+                gen_helper_subu_s_ob(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_QH:
+                check_dspr2(ctx);
+                gen_helper_subu_qh(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBU_S_QH:
+                check_dspr2(ctx);
+                gen_helper_subu_s_qh(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBUH_OB:
+                check_dspr2(ctx);
+                gen_helper_subuh_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_SUBUH_R_OB:
+                check_dspr2(ctx);
+                gen_helper_subuh_r_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_PW:
+                check_dsp(ctx);
+                gen_helper_addq_pw(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_S_PW:
+                check_dsp(ctx);
+                gen_helper_addq_s_pw(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_QH:
+                check_dsp(ctx);
+                gen_helper_addq_qh(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDQ_S_QH:
+                check_dsp(ctx);
+                gen_helper_addq_s_qh(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_OB:
+                check_dsp(ctx);
+                gen_helper_addu_ob(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_S_OB:
+                check_dsp(ctx);
+                gen_helper_addu_s_ob(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_QH:
+                check_dspr2(ctx);
+                gen_helper_addu_qh(cpu_gpr[rd], cpu_env,
+                                   cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDU_S_QH:
+                check_dspr2(ctx);
+                gen_helper_addu_s_qh(cpu_gpr[rd], cpu_env,
+                                     cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDUH_OB:
+                check_dspr2(ctx);
+                gen_helper_adduh_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_ADDUH_R_OB:
+                check_dspr2(ctx);
+                gen_helper_adduh_r_ob(cpu_gpr[rd], cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK ADDU.OB");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
+#endif
+#if defined(TARGET_MIPS64)
+        case OPC_CMPU_EQ_OB_DSP:
+            op2 = MASK_CMPU_EQ_OB(ctx->opcode);
+            switch (op2) {
+            case OPC_PRECR_OB_QH:
+                check_dspr2(ctx);
+                gen_helper_precr_ob_qh(cpu_gpr[rd], cpu_gpr[rs],
+                                       cpu_gpr[rt]);
+                break;
+            case OPC_PRECR_SRA_QH_PW:
+                check_dspr2(ctx);
+                {
+                    TCGv_i32 sa_v = tcg_const_i32(rd);
+                    gen_helper_precr_sra_qh_pw(cpu_gpr[rt], cpu_gpr[rs],
+                                               cpu_gpr[rt], sa_v);
+                    tcg_temp_free_i32(sa_v);
+                    break;
+                }
+            case OPC_PRECR_SRA_R_QH_PW:
+                check_dspr2(ctx);
+                {
+                    TCGv_i32 sa_v = tcg_const_i32(rd);
+                    gen_helper_precr_sra_r_qh_pw(cpu_gpr[rt], cpu_gpr[rs],
+                                                 cpu_gpr[rt], sa_v);
+                    tcg_temp_free_i32(sa_v);
+                    break;
+                }
+            case OPC_PRECRQ_OB_QH:
+                check_dsp(ctx);
+                gen_helper_precrq_ob_qh(cpu_gpr[rd], cpu_gpr[rs],
+                                        cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQ_PW_L:
+                check_dsp(ctx);
+                gen_helper_precrq_pw_l(cpu_gpr[rd], cpu_gpr[rs],
+                                       cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQ_QH_PW:
+                check_dsp(ctx);
+                gen_helper_precrq_qh_pw(cpu_gpr[rd], cpu_gpr[rs],
+                                        cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQ_RS_QH_PW:
+                check_dsp(ctx);
+                gen_helper_precrq_rs_qh_pw(cpu_gpr[rd], cpu_env,
+                                           cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            case OPC_PRECRQU_S_OB_QH:
+                check_dsp(ctx);
+                gen_helper_precrqu_s_ob_qh(cpu_gpr[rd], cpu_env,
+                                           cpu_gpr[rs], cpu_gpr[rt]);
+                break;
+            default:            /* Invalid */
+                MIPS_INVAL("MASK CMPU.EQ.OB");
+                generate_exception(ctx, EXCP_RI);
+                break;
+            }
+            break;
+#endif
         default:            /* Invalid */
             MIPS_INVAL("special3");
             generate_exception(ctx, EXCP_RI);